sys/platform/pc64/x86_64/pmap.c

   1 /*
   2  * Copyright (c) 1991 Regents of the University of California.
   3  * Copyright (c) 1994 John S. Dyson
   4  * Copyright (c) 1994 David Greenman
   5  * Copyright (c) 2003 Peter Wemm
   6  * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
   7  * Copyright (c) 2008, 2009 The DragonFly Project.
   8  * Copyright (c) 2008, 2009 Jordan Gordeev.
   9  * Copyright (c) 2011-2017 Matthew Dillon
  10  * All rights reserved.
  11  *
  12  * This code is derived from software contributed to Berkeley by
  13  * the Systems Programming Group of the University of Utah Computer
  14  * Science Department and William Jolitz of UUNET Technologies Inc.
  15  *
  16  * Redistribution and use in source and binary forms, with or without
  17  * modification, are permitted provided that the following conditions
  18  * are met:
  19  * 1. Redistributions of source code must retain the above copyright
  20  *    notice, this list of conditions and the following disclaimer.
  21  * 2. Redistributions in binary form must reproduce the above copyright
  22  *    notice, this list of conditions and the following disclaimer in the
  23  *    documentation and/or other materials provided with the distribution.
  24  * 3. All advertising materials mentioning features or use of this software
  25  *    must display the following acknowledgement:
  26  *      This product includes software developed by the University of
  27  *      California, Berkeley and its contributors.
  28  * 4. Neither the name of the University nor the names of its contributors
  29  *    may be used to endorse or promote products derived from this software
  30  *    without specific prior written permission.
  31  *
  32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  42  * SUCH DAMAGE.
  43  */
  44 /*
  45  * Manage physical address maps for x86-64 systems.
  46  */
  47
  48 #if 0 /* JG */
  49 #include "opt_pmap.h"
  50 #endif
  51 #include "opt_msgbuf.h"
  52
  53 #include <sys/param.h>
  54 #include <sys/kernel.h>
  55 #include <sys/proc.h>
  56 #include <sys/msgbuf.h>
  57 #include <sys/vmmeter.h>
  58 #include <sys/mman.h>
  59 #include <sys/systm.h>
  60
  61 #include <vm/vm.h>
  62 #include <vm/vm_param.h>
  63 #include <sys/sysctl.h>
  64 #include <sys/lock.h>
  65 #include <vm/vm_kern.h>
  66 #include <vm/vm_page.h>
  67 #include <vm/vm_map.h>
  68 #include <vm/vm_object.h>
  69 #include <vm/vm_extern.h>
  70 #include <vm/vm_pageout.h>
  71 #include <vm/vm_pager.h>
  72 #include <vm/vm_zone.h>
  73
  74 #include <sys/user.h>
  75 #include <sys/thread2.h>
  76 #include <sys/spinlock2.h>
  77 #include <vm/vm_page2.h>
  78
  79 #include <machine/cputypes.h>
  80 #include <machine/md_var.h>
  81 #include <machine/specialreg.h>
  82 #include <machine/smp.h>
  83 #include <machine_base/apic/apicreg.h>
  84 #include <machine/globaldata.h>
  85 #include <machine/pmap.h>
  86 #include <machine/pmap_inval.h>
  87 #include <machine/inttypes.h>
  88
  89 #include <ddb/ddb.h>
  90
  91 #define PMAP_KEEP_PDIRS
  92 #ifndef PMAP_SHPGPERPROC
  93 #define PMAP_SHPGPERPROC 2000
  94 #endif
  95
  96 #if defined(DIAGNOSTIC)
  97 #define PMAP_DIAGNOSTIC
  98 #endif
  99
 100 #define MINPV 2048
 101
 102 /*
 103  * pmap debugging will report who owns a pv lock when blocking.
 104  */
 105 #ifdef PMAP_DEBUG
 106
 107 #define PMAP_DEBUG_DECL         ,const char *func, int lineno
 108 #define PMAP_DEBUG_ARGS         , __func__, __LINE__
 109 #define PMAP_DEBUG_COPY         , func, lineno
 110
 111 #define pv_get(pmap, pindex, pmarkp)    _pv_get(pmap, pindex, pmarkp    \
 112                                                         PMAP_DEBUG_ARGS)
 113 #define pv_lock(pv)                     _pv_lock(pv                     \
 114                                                         PMAP_DEBUG_ARGS)
 115 #define pv_hold_try(pv)                 _pv_hold_try(pv                 \
 116                                                         PMAP_DEBUG_ARGS)
 117 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp  \
 118                                                         PMAP_DEBUG_ARGS)
 119
 120 #define pv_free(pv, pvp)                _pv_free(pv, pvp PMAP_DEBUG_ARGS)
 121
 122 #else
 123
 124 #define PMAP_DEBUG_DECL
 125 #define PMAP_DEBUG_ARGS
 126 #define PMAP_DEBUG_COPY
 127
 128 #define pv_get(pmap, pindex, pmarkp)            _pv_get(pmap, pindex, pmarkp)
 129 #define pv_lock(pv)                     _pv_lock(pv)
 130 #define pv_hold_try(pv)                 _pv_hold_try(pv)
 131 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp)
 132 #define pv_free(pv, pvp)                _pv_free(pv, pvp)
 133
 134 #endif
 135
 136 /*
 137  * Get PDEs and PTEs for user/kernel address space
 138  */
 139 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
 140
 141 #define pmap_pde_v(pmap, pte)           ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 142 #define pmap_pte_w(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
 143 #define pmap_pte_m(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
 144 #define pmap_pte_u(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
 145 #define pmap_pte_v(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 146
 147 /*
 148  * Given a map and a machine independent protection code,
 149  * convert to a vax protection code.
 150  */
 151 #define pte_prot(m, p)          \
 152         (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
 153 static uint64_t protection_codes[PROTECTION_CODES_SIZE];
 154
 155 struct pmap kernel_pmap;
 156
 157 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
 158
 159 vm_paddr_t avail_start;         /* PA of first available physical page */
 160 vm_paddr_t avail_end;           /* PA of last available physical page */
 161 vm_offset_t virtual2_start;     /* cutout free area prior to kernel start */
 162 vm_offset_t virtual2_end;
 163 vm_offset_t virtual_start;      /* VA of first avail page (after kernel bss) */
 164 vm_offset_t virtual_end;        /* VA of last avail page (end of kernel AS) */
 165 vm_offset_t KvaStart;           /* VA start of KVA space */
 166 vm_offset_t KvaEnd;             /* VA end of KVA space (non-inclusive) */
 167 vm_offset_t KvaSize;            /* max size of kernel virtual address space */
 168 static boolean_t pmap_initialized = FALSE;      /* Has pmap_init completed? */
 169 //static int pgeflag;           /* PG_G or-in */
 170 uint64_t PatMsr;
 171
 172 static int ndmpdp;
 173 static vm_paddr_t dmaplimit;
 174 static int nkpt;
 175 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
 176
 177 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE];        /* PAT -> PG_ bits */
 178 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/    /* PAT -> PG_ bits */
 179
 180 static uint64_t KPTbase;
 181 static uint64_t KPTphys;
 182 static uint64_t KPDphys;        /* phys addr of kernel level 2 */
 183 static uint64_t KPDbase;        /* phys addr of kernel level 2 @ KERNBASE */
 184 uint64_t KPDPphys;      /* phys addr of kernel level 3 */
 185 uint64_t KPML4phys;     /* phys addr of kernel level 4 */
 186
 187 static uint64_t DMPDphys;       /* phys addr of direct mapped level 2 */
 188 static uint64_t DMPDPphys;      /* phys addr of direct mapped level 3 */
 189
 190 /*
 191  * Data for the pv entry allocation mechanism
 192  */
 193 static vm_zone_t pvzone;
 194 static struct vm_zone pvzone_store;
 195 static vm_pindex_t pv_entry_max=0, pv_entry_high_water=0;
 196 static int pmap_pagedaemon_waken = 0;
 197 static struct pv_entry *pvinit;
 198
 199 /*
 200  * All those kernel PT submaps that BSD is so fond of
 201  */
 202 pt_entry_t *CMAP1 = NULL, *ptmmap;
 203 caddr_t CADDR1 = NULL, ptvmmap = NULL;
 204 static pt_entry_t *msgbufmap;
 205 struct msgbuf *msgbufp=NULL;
 206
 207 /*
 208  * PMAP default PG_* bits. Needed to be able to add
 209  * EPT/NPT pagetable pmap_bits for the VMM module
 210  */
 211 uint64_t pmap_bits_default[] = {
 212                 REGULAR_PMAP,                   /* TYPE_IDX             0 */
 213                 X86_PG_V,                       /* PG_V_IDX             1 */
 214                 X86_PG_RW,                      /* PG_RW_IDX            2 */
 215                 X86_PG_U,                       /* PG_U_IDX             3 */
 216                 X86_PG_A,                       /* PG_A_IDX             4 */
 217                 X86_PG_M,                       /* PG_M_IDX             5 */
 218                 X86_PG_PS,                      /* PG_PS_IDX3           6 */
 219                 X86_PG_G,                       /* PG_G_IDX             7 */
 220                 X86_PG_AVAIL1,                  /* PG_AVAIL1_IDX        8 */
 221                 X86_PG_AVAIL2,                  /* PG_AVAIL2_IDX        9 */
 222                 X86_PG_AVAIL3,                  /* PG_AVAIL3_IDX        10 */
 223                 X86_PG_NC_PWT | X86_PG_NC_PCD,  /* PG_N_IDX             11 */
 224                 X86_PG_NX,                      /* PG_NX_IDX            12 */
 225 };
 226 /*
 227  * Crashdump maps.
 228  */
 229 static pt_entry_t *pt_crashdumpmap;
 230 static caddr_t crashdumpmap;
 231
 232 static int pmap_debug = 0;
 233 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
 234     &pmap_debug, 0, "Debug pmap's");
 235 #ifdef PMAP_DEBUG2
 236 static int pmap_enter_debug = 0;
 237 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
 238     &pmap_enter_debug, 0, "Debug pmap_enter's");
 239 #endif
 240 static int pmap_yield_count = 64;
 241 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
 242     &pmap_yield_count, 0, "Yield during init_pt/release");
 243 static int pmap_mmu_optimize = 0;
 244 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
 245     &pmap_mmu_optimize, 0, "Share page table pages when possible");
 246 int pmap_fast_kernel_cpusync = 0;
 247 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
 248     &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
 249 int pmap_dynamic_delete = 0;
 250 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
 251     &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
 252 int pmap_lock_delay = 100;
 253 SYSCTL_INT(_machdep, OID_AUTO, pmap_lock_delay, CTLFLAG_RW,
 254     &pmap_lock_delay, 0, "Spin loops");
 255
 256 static int pmap_nx_enable = 0;
 257 /* needs manual TUNABLE in early probe, see below */
 258
 259 /* Standard user access funtions */
 260 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
 261     size_t *lencopied);
 262 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
 263 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
 264 extern int std_fubyte (const uint8_t *base);
 265 extern int std_subyte (uint8_t *base, uint8_t byte);
 266 extern int32_t std_fuword32 (const uint32_t *base);
 267 extern int64_t std_fuword64 (const uint64_t *base);
 268 extern int std_suword64 (uint64_t *base, uint64_t word);
 269 extern int std_suword32 (uint32_t *base, int word);
 270 extern uint32_t std_swapu32 (volatile uint32_t *base, uint32_t v);
 271 extern uint64_t std_swapu64 (volatile uint64_t *base, uint64_t v);
 272
 273 static void pv_hold(pv_entry_t pv);
 274 static int _pv_hold_try(pv_entry_t pv
 275                                 PMAP_DEBUG_DECL);
 276 static void pv_drop(pv_entry_t pv);
 277 static void _pv_lock(pv_entry_t pv
 278                                 PMAP_DEBUG_DECL);
 279 static void pv_unlock(pv_entry_t pv);
 280 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
 281                                 PMAP_DEBUG_DECL);
 282 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp
 283                                 PMAP_DEBUG_DECL);
 284 static void _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL);
 285 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex,
 286                                 vm_pindex_t **pmarkp, int *errorp);
 287 static void pv_put(pv_entry_t pv);
 288 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
 289 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
 290                       pv_entry_t *pvpp);
 291 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
 292                       pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
 293 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
 294                         pmap_inval_bulk_t *bulk, int destroy);
 295 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
 296 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
 297                         pmap_inval_bulk_t *bulk);
 298
 299 struct pmap_scan_info;
 300 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
 301                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
 302                       pv_entry_t pt_pv, int sharept,
 303                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 304 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
 305                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
 306                       pv_entry_t pt_pv, int sharept,
 307                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 308
 309 static void i386_protection_init (void);
 310 static void create_pagetables(vm_paddr_t *firstaddr);
 311 static void pmap_remove_all (vm_page_t m);
 312 static boolean_t pmap_testbit (vm_page_t m, int bit);
 313
 314 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
 315 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
 316
 317 static void pmap_pinit_defaults(struct pmap *pmap);
 318 static void pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark);
 319 static void pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark);
 320
 321 static int
 322 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
 323 {
 324         if (pv1->pv_pindex < pv2->pv_pindex)
 325                 return(-1);
 326         if (pv1->pv_pindex > pv2->pv_pindex)
 327                 return(1);
 328         return(0);
 329 }
 330
 331 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
 332              pv_entry_compare, vm_pindex_t, pv_pindex);
 333
 334 static __inline
 335 void
 336 pmap_page_stats_adding(vm_page_t m)
 337 {
 338         globaldata_t gd = mycpu;
 339
 340         if (TAILQ_EMPTY(&m->md.pv_list)) {
 341                 ++gd->gd_vmtotal.t_arm;
 342         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 343                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 344                 ++gd->gd_vmtotal.t_armshr;
 345                 ++gd->gd_vmtotal.t_avmshr;
 346         } else {
 347                 ++gd->gd_vmtotal.t_avmshr;
 348         }
 349 }
 350
 351 static __inline
 352 void
 353 pmap_page_stats_deleting(vm_page_t m)
 354 {
 355         globaldata_t gd = mycpu;
 356
 357         if (TAILQ_EMPTY(&m->md.pv_list)) {
 358                 --gd->gd_vmtotal.t_arm;
 359         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 360                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 361                 --gd->gd_vmtotal.t_armshr;
 362                 --gd->gd_vmtotal.t_avmshr;
 363         } else {
 364                 --gd->gd_vmtotal.t_avmshr;
 365         }
 366 }
 367
 368 /*
 369  * This is an ineligent crowbar to prevent heavily threaded programs
 370  * from creating long live-locks in the pmap code when pmap_mmu_optimize
 371  * is enabled.  Without it a pmap-local page table page can wind up being
 372  * constantly created and destroyed (without injury, but also without
 373  * progress) as the optimization tries to switch to the object's shared page
 374  * table page.
 375  */
 376 static __inline void
 377 pmap_softwait(pmap_t pmap)
 378 {
 379         while (pmap->pm_softhold) {
 380                 tsleep_interlock(&pmap->pm_softhold, 0);
 381                 if (pmap->pm_softhold)
 382                         tsleep(&pmap->pm_softhold, PINTERLOCKED, "mmopt", 0);
 383         }
 384 }
 385
 386 static __inline void
 387 pmap_softhold(pmap_t pmap)
 388 {
 389         while (atomic_swap_int(&pmap->pm_softhold, 1) == 1) {
 390                 tsleep_interlock(&pmap->pm_softhold, 0);
 391                 if (atomic_swap_int(&pmap->pm_softhold, 1) == 1)
 392                         tsleep(&pmap->pm_softhold, PINTERLOCKED, "mmopt", 0);
 393         }
 394 }
 395
 396 static __inline void
 397 pmap_softdone(pmap_t pmap)
 398 {
 399         atomic_swap_int(&pmap->pm_softhold, 0);
 400         wakeup(&pmap->pm_softhold);
 401 }
 402
 403 /*
 404  * Move the kernel virtual free pointer to the next
 405  * 2MB.  This is used to help improve performance
 406  * by using a large (2MB) page for much of the kernel
 407  * (.text, .data, .bss)
 408  */
 409 static
 410 vm_offset_t
 411 pmap_kmem_choose(vm_offset_t addr)
 412 {
 413         vm_offset_t newaddr = addr;
 414
 415         newaddr = roundup2(addr, NBPDR);
 416         return newaddr;
 417 }
 418
 419 /*
 420  * Returns the pindex of a page table entry (representing a terminal page).
 421  * There are NUPTE_TOTAL page table entries possible (a huge number)
 422  *
 423  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
 424  * We want to properly translate negative KVAs.
 425  */
 426 static __inline
 427 vm_pindex_t
 428 pmap_pte_pindex(vm_offset_t va)
 429 {
 430         return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
 431 }
 432
 433 /*
 434  * Returns the pindex of a page table.
 435  */
 436 static __inline
 437 vm_pindex_t
 438 pmap_pt_pindex(vm_offset_t va)
 439 {
 440         return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
 441 }
 442
 443 /*
 444  * Returns the pindex of a page directory.
 445  */
 446 static __inline
 447 vm_pindex_t
 448 pmap_pd_pindex(vm_offset_t va)
 449 {
 450         return (NUPTE_TOTAL + NUPT_TOTAL +
 451                 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
 452 }
 453
 454 static __inline
 455 vm_pindex_t
 456 pmap_pdp_pindex(vm_offset_t va)
 457 {
 458         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
 459                 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
 460 }
 461
 462 static __inline
 463 vm_pindex_t
 464 pmap_pml4_pindex(void)
 465 {
 466         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
 467 }
 468
 469 /*
 470  * Return various clipped indexes for a given VA
 471  *
 472  * Returns the index of a pt in a page directory, representing a page
 473  * table.
 474  */
 475 static __inline
 476 vm_pindex_t
 477 pmap_pt_index(vm_offset_t va)
 478 {
 479         return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
 480 }
 481
 482 /*
 483  * Returns the index of a pd in a page directory page, representing a page
 484  * directory.
 485  */
 486 static __inline
 487 vm_pindex_t
 488 pmap_pd_index(vm_offset_t va)
 489 {
 490         return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
 491 }
 492
 493 /*
 494  * Returns the index of a pdp in the pml4 table, representing a page
 495  * directory page.
 496  */
 497 static __inline
 498 vm_pindex_t
 499 pmap_pdp_index(vm_offset_t va)
 500 {
 501         return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
 502 }
 503
 504 /*
 505  * Locate the requested pt_entry
 506  */
 507 static __inline
 508 pv_entry_t
 509 pv_entry_lookup(pmap_t pmap, vm_pindex_t pindex)
 510 {
 511         pv_entry_t pv;
 512
 513         if (pindex < pmap_pt_pindex(0))
 514                 pv = pmap->pm_pvhint_pte;
 515         else if (pindex < pmap_pd_pindex(0))
 516                 pv = pmap->pm_pvhint_pt;
 517         else
 518                 pv = NULL;
 519         cpu_ccfence();
 520         if (pv == NULL || pv->pv_pmap != pmap) {
 521                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
 522                                                 pindex);
 523         } else if (pv->pv_pindex != pindex) {
 524                 pv = pv_entry_rb_tree_RB_LOOKUP_REL(&pmap->pm_pvroot,
 525                                                     pindex, pv);
 526         }
 527         return pv;
 528 }
 529
 530 /*
 531  * pmap_pte_quick:
 532  *
 533  *      Super fast pmap_pte routine best used when scanning the pv lists.
 534  *      This eliminates many course-grained invltlb calls.  Note that many of
 535  *      the pv list scans are across different pmaps and it is very wasteful
 536  *      to do an entire invltlb when checking a single mapping.
 537  */
 538 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
 539
 540 static
 541 pt_entry_t *
 542 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
 543 {
 544         return pmap_pte(pmap, va);
 545 }
 546
 547 /*
 548  * The placemarker hash must be broken up into four zones so lock
 549  * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
 550  *
 551  * Placemarkers are used to 'lock' page table indices that do not have
 552  * a pv_entry.  This allows the pmap to support managed and unmanaged
 553  * pages and shared page tables.
 554  */
 555 #define PM_PLACE_BASE   (PM_PLACEMARKS >> 2)
 556
 557 static __inline
 558 vm_pindex_t *
 559 pmap_placemarker_hash(pmap_t pmap, vm_pindex_t pindex)
 560 {
 561         int hi;
 562
 563         if (pindex < pmap_pt_pindex(0))         /* zone 0 - PTE */
 564                 hi = 0;
 565         else if (pindex < pmap_pd_pindex(0))    /* zone 1 - PT */
 566                 hi = PM_PLACE_BASE;
 567         else if (pindex < pmap_pdp_pindex(0))   /* zone 2 - PD */
 568                 hi = PM_PLACE_BASE << 1;
 569         else                                    /* zone 3 - PDP (and PML4E) */
 570                 hi = PM_PLACE_BASE | (PM_PLACE_BASE << 1);
 571         hi += pindex & (PM_PLACE_BASE - 1);
 572
 573         return (&pmap->pm_placemarks[hi]);
 574 }
 575
 576
 577 /*
 578  * Generic procedure to index a pte from a pt, pd, or pdp.
 579  *
 580  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
 581  *       a page table page index but is instead of PV lookup index.
 582  */
 583 static
 584 void *
 585 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
 586 {
 587         pt_entry_t *pte;
 588
 589         pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
 590         return(&pte[pindex]);
 591 }
 592
 593 /*
 594  * Return pointer to PDP slot in the PML4
 595  */
 596 static __inline
 597 pml4_entry_t *
 598 pmap_pdp(pmap_t pmap, vm_offset_t va)
 599 {
 600         return (&pmap->pm_pml4[pmap_pdp_index(va)]);
 601 }
 602
 603 /*
 604  * Return pointer to PD slot in the PDP given a pointer to the PDP
 605  */
 606 static __inline
 607 pdp_entry_t *
 608 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
 609 {
 610         pdp_entry_t *pd;
 611
 612         pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
 613         return (&pd[pmap_pd_index(va)]);
 614 }
 615
 616 /*
 617  * Return pointer to PD slot in the PDP.
 618  */
 619 static __inline
 620 pdp_entry_t *
 621 pmap_pd(pmap_t pmap, vm_offset_t va)
 622 {
 623         pml4_entry_t *pdp;
 624
 625         pdp = pmap_pdp(pmap, va);
 626         if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
 627                 return NULL;
 628         return (pmap_pdp_to_pd(*pdp, va));
 629 }
 630
 631 /*
 632  * Return pointer to PT slot in the PD given a pointer to the PD
 633  */
 634 static __inline
 635 pd_entry_t *
 636 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
 637 {
 638         pd_entry_t *pt;
 639
 640         pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
 641         return (&pt[pmap_pt_index(va)]);
 642 }
 643
 644 /*
 645  * Return pointer to PT slot in the PD
 646  *
 647  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
 648  *                   so we cannot lookup the PD via the PDP.  Instead we
 649  *                   must look it up via the pmap.
 650  */
 651 static __inline
 652 pd_entry_t *
 653 pmap_pt(pmap_t pmap, vm_offset_t va)
 654 {
 655         pdp_entry_t *pd;
 656         pv_entry_t pv;
 657         vm_pindex_t pd_pindex;
 658         vm_paddr_t phys;
 659
 660         if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
 661                 pd_pindex = pmap_pd_pindex(va);
 662                 spin_lock_shared(&pmap->pm_spin);
 663                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
 664                 if (pv == NULL || pv->pv_m == NULL) {
 665                         spin_unlock_shared(&pmap->pm_spin);
 666                         return NULL;
 667                 }
 668                 phys = VM_PAGE_TO_PHYS(pv->pv_m);
 669                 spin_unlock_shared(&pmap->pm_spin);
 670                 return (pmap_pd_to_pt(phys, va));
 671         } else {
 672                 pd = pmap_pd(pmap, va);
 673                 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
 674                          return NULL;
 675                 return (pmap_pd_to_pt(*pd, va));
 676         }
 677 }
 678
 679 /*
 680  * Return pointer to PTE slot in the PT given a pointer to the PT
 681  */
 682 static __inline
 683 pt_entry_t *
 684 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
 685 {
 686         pt_entry_t *pte;
 687
 688         pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
 689         return (&pte[pmap_pte_index(va)]);
 690 }
 691
 692 /*
 693  * Return pointer to PTE slot in the PT
 694  */
 695 static __inline
 696 pt_entry_t *
 697 pmap_pte(pmap_t pmap, vm_offset_t va)
 698 {
 699         pd_entry_t *pt;
 700
 701         pt = pmap_pt(pmap, va);
 702         if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
 703                  return NULL;
 704         if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
 705                 return ((pt_entry_t *)pt);
 706         return (pmap_pt_to_pte(*pt, va));
 707 }
 708
 709 /*
 710  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
 711  * the PT layer.  This will speed up core pmap operations considerably.
 712  *
 713  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
 714  *       must be in a known associated state (typically by being locked when
 715  *       the pmap spinlock isn't held).  We allow the race for that case.
 716  *
 717  * NOTE: pm_pvhint* is only accessed (read) with the spin-lock held, using
 718  *       cpu_ccfence() to prevent compiler optimizations from reloading the
 719  *       field.
 720  */
 721 static __inline
 722 void
 723 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
 724 {
 725         if (pindex < pmap_pt_pindex(0)) {
 726                 if (pv->pv_pmap)
 727                         pv->pv_pmap->pm_pvhint_pte = pv;
 728         } else if (pindex < pmap_pd_pindex(0)) {
 729                 if (pv->pv_pmap)
 730                         pv->pv_pmap->pm_pvhint_pt = pv;
 731         }
 732 }
 733
 734
 735 /*
 736  * Return address of PT slot in PD (KVM only)
 737  *
 738  * Cannot be used for user page tables because it might interfere with
 739  * the shared page-table-page optimization (pmap_mmu_optimize).
 740  */
 741 static __inline
 742 pd_entry_t *
 743 vtopt(vm_offset_t va)
 744 {
 745         uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
 746                                   NPML4EPGSHIFT)) - 1);
 747
 748         return (PDmap + ((va >> PDRSHIFT) & mask));
 749 }
 750
 751 /*
 752  * KVM - return address of PTE slot in PT
 753  */
 754 static __inline
 755 pt_entry_t *
 756 vtopte(vm_offset_t va)
 757 {
 758         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 759                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 760
 761         return (PTmap + ((va >> PAGE_SHIFT) & mask));
 762 }
 763
 764 /*
 765  * Returns the physical address translation from va for a user address.
 766  * (vm_paddr_t)-1 is returned on failure.
 767  */
 768 vm_paddr_t
 769 uservtophys(vm_offset_t va)
 770 {
 771         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 772                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 773         vm_paddr_t pa;
 774         pt_entry_t pte;
 775         pmap_t pmap;
 776
 777         pmap = vmspace_pmap(mycpu->gd_curthread->td_lwp->lwp_vmspace);
 778         pa = (vm_paddr_t)-1;
 779         if (va < VM_MAX_USER_ADDRESS) {
 780                 pte = kreadmem64(PTmap + ((va >> PAGE_SHIFT) & mask));
 781                 if (pte & pmap->pmap_bits[PG_V_IDX])
 782                         pa = (pte & PG_FRAME) | (va & PAGE_MASK);
 783         }
 784         return pa;
 785 }
 786
 787 static uint64_t
 788 allocpages(vm_paddr_t *firstaddr, long n)
 789 {
 790         uint64_t ret;
 791
 792         ret = *firstaddr;
 793         bzero((void *)ret, n * PAGE_SIZE);
 794         *firstaddr += n * PAGE_SIZE;
 795         return (ret);
 796 }
 797
 798 static
 799 void
 800 create_pagetables(vm_paddr_t *firstaddr)
 801 {
 802         long i;         /* must be 64 bits */
 803         long nkpt_base;
 804         long nkpt_phys;
 805         int j;
 806
 807         /*
 808          * We are running (mostly) V=P at this point
 809          *
 810          * Calculate NKPT - number of kernel page tables.  We have to
 811          * accomodoate prealloction of the vm_page_array, dump bitmap,
 812          * MSGBUF_SIZE, and other stuff.  Be generous.
 813          *
 814          * Maxmem is in pages.
 815          *
 816          * ndmpdp is the number of 1GB pages we wish to map.
 817          */
 818         ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
 819         if (ndmpdp < 4)         /* Minimum 4GB of dirmap */
 820                 ndmpdp = 4;
 821         KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
 822
 823         /*
 824          * Starting at the beginning of kvm (not KERNBASE).
 825          */
 826         nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
 827         nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
 828         nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
 829                        ndmpdp) + 511) / 512;
 830         nkpt_phys += 128;
 831
 832         /*
 833          * Starting at KERNBASE - map 2G worth of page table pages.
 834          * KERNBASE is offset -2G from the end of kvm.
 835          */
 836         nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
 837
 838         /*
 839          * Allocate pages
 840          */
 841         KPTbase = allocpages(firstaddr, nkpt_base);
 842         KPTphys = allocpages(firstaddr, nkpt_phys);
 843         KPML4phys = allocpages(firstaddr, 1);
 844         KPDPphys = allocpages(firstaddr, NKPML4E);
 845         KPDphys = allocpages(firstaddr, NKPDPE);
 846
 847         /*
 848          * Calculate the page directory base for KERNBASE,
 849          * that is where we start populating the page table pages.
 850          * Basically this is the end - 2.
 851          */
 852         KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
 853
 854         DMPDPphys = allocpages(firstaddr, NDMPML4E);
 855         if ((amd_feature & AMDID_PAGE1GB) == 0)
 856                 DMPDphys = allocpages(firstaddr, ndmpdp);
 857         dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
 858
 859         /*
 860          * Fill in the underlying page table pages for the area around
 861          * KERNBASE.  This remaps low physical memory to KERNBASE.
 862          *
 863          * Read-only from zero to physfree
 864          * XXX not fully used, underneath 2M pages
 865          */
 866         for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
 867                 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
 868                 ((pt_entry_t *)KPTbase)[i] |=
 869                     pmap_bits_default[PG_RW_IDX] |
 870                     pmap_bits_default[PG_V_IDX] |
 871                     pmap_bits_default[PG_G_IDX];
 872         }
 873
 874         /*
 875          * Now map the initial kernel page tables.  One block of page
 876          * tables is placed at the beginning of kernel virtual memory,
 877          * and another block is placed at KERNBASE to map the kernel binary,
 878          * data, bss, and initial pre-allocations.
 879          */
 880         for (i = 0; i < nkpt_base; i++) {
 881                 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
 882                 ((pd_entry_t *)KPDbase)[i] |=
 883                     pmap_bits_default[PG_RW_IDX] |
 884                     pmap_bits_default[PG_V_IDX];
 885         }
 886         for (i = 0; i < nkpt_phys; i++) {
 887                 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
 888                 ((pd_entry_t *)KPDphys)[i] |=
 889                     pmap_bits_default[PG_RW_IDX] |
 890                     pmap_bits_default[PG_V_IDX];
 891         }
 892
 893         /*
 894          * Map from zero to end of allocations using 2M pages as an
 895          * optimization.  This will bypass some of the KPTBase pages
 896          * above in the KERNBASE area.
 897          */
 898         for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
 899                 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
 900                 ((pd_entry_t *)KPDbase)[i] |=
 901                     pmap_bits_default[PG_RW_IDX] |
 902                     pmap_bits_default[PG_V_IDX] |
 903                     pmap_bits_default[PG_PS_IDX] |
 904                     pmap_bits_default[PG_G_IDX];
 905         }
 906
 907         /*
 908          * And connect up the PD to the PDP.  The kernel pmap is expected
 909          * to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
 910          */
 911         for (i = 0; i < NKPDPE; i++) {
 912                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
 913                                 KPDphys + (i << PAGE_SHIFT);
 914                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
 915                     pmap_bits_default[PG_RW_IDX] |
 916                     pmap_bits_default[PG_V_IDX] |
 917                     pmap_bits_default[PG_U_IDX];
 918         }
 919
 920         /*
 921          * Now set up the direct map space using either 2MB or 1GB pages
 922          * Preset PG_M and PG_A because demotion expects it.
 923          *
 924          * When filling in entries in the PD pages make sure any excess
 925          * entries are set to zero as we allocated enough PD pages
 926          */
 927         if ((amd_feature & AMDID_PAGE1GB) == 0) {
 928                 for (i = 0; i < NPDEPG * ndmpdp; i++) {
 929                         ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
 930                         ((pd_entry_t *)DMPDphys)[i] |=
 931                             pmap_bits_default[PG_RW_IDX] |
 932                             pmap_bits_default[PG_V_IDX] |
 933                             pmap_bits_default[PG_PS_IDX] |
 934                             pmap_bits_default[PG_G_IDX] |
 935                             pmap_bits_default[PG_M_IDX] |
 936                             pmap_bits_default[PG_A_IDX];
 937                 }
 938
 939                 /*
 940                  * And the direct map space's PDP
 941                  */
 942                 for (i = 0; i < ndmpdp; i++) {
 943                         ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
 944                                                         (i << PAGE_SHIFT);
 945                         ((pdp_entry_t *)DMPDPphys)[i] |=
 946                             pmap_bits_default[PG_RW_IDX] |
 947                             pmap_bits_default[PG_V_IDX] |
 948                             pmap_bits_default[PG_U_IDX];
 949                 }
 950         } else {
 951                 for (i = 0; i < ndmpdp; i++) {
 952                         ((pdp_entry_t *)DMPDPphys)[i] =
 953                                                 (vm_paddr_t)i << PDPSHIFT;
 954                         ((pdp_entry_t *)DMPDPphys)[i] |=
 955                             pmap_bits_default[PG_RW_IDX] |
 956                             pmap_bits_default[PG_V_IDX] |
 957                             pmap_bits_default[PG_PS_IDX] |
 958                             pmap_bits_default[PG_G_IDX] |
 959                             pmap_bits_default[PG_M_IDX] |
 960                             pmap_bits_default[PG_A_IDX];
 961                 }
 962         }
 963
 964         /* And recursively map PML4 to itself in order to get PTmap */
 965         ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
 966         ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
 967             pmap_bits_default[PG_RW_IDX] |
 968             pmap_bits_default[PG_V_IDX] |
 969             pmap_bits_default[PG_U_IDX];
 970
 971         /*
 972          * Connect the Direct Map slots up to the PML4
 973          */
 974         for (j = 0; j < NDMPML4E; ++j) {
 975                 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
 976                     (DMPDPphys + ((vm_paddr_t)j << PAGE_SHIFT)) |
 977                     pmap_bits_default[PG_RW_IDX] |
 978                     pmap_bits_default[PG_V_IDX] |
 979                     pmap_bits_default[PG_U_IDX];
 980         }
 981
 982         /*
 983          * Connect the KVA slot up to the PML4
 984          */
 985         ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
 986         ((pdp_entry_t *)KPML4phys)[KPML4I] |=
 987             pmap_bits_default[PG_RW_IDX] |
 988             pmap_bits_default[PG_V_IDX] |
 989             pmap_bits_default[PG_U_IDX];
 990 }
 991
 992 /*
 993  *      Bootstrap the system enough to run with virtual memory.
 994  *
 995  *      On the i386 this is called after mapping has already been enabled
 996  *      and just syncs the pmap module with what has already been done.
 997  *      [We can't call it easily with mapping off since the kernel is not
 998  *      mapped with PA == VA, hence we would have to relocate every address
 999  *      from the linked base (virtual) address "KERNBASE" to the actual
1000  *      (physical) address starting relative to 0]
1001  */
1002 void
1003 pmap_bootstrap(vm_paddr_t *firstaddr)
1004 {
1005         vm_offset_t va;
1006         pt_entry_t *pte;
1007         int i;
1008
1009         KvaStart = VM_MIN_KERNEL_ADDRESS;
1010         KvaEnd = VM_MAX_KERNEL_ADDRESS;
1011         KvaSize = KvaEnd - KvaStart;
1012
1013         avail_start = *firstaddr;
1014
1015         /*
1016          * Create an initial set of page tables to run the kernel in.
1017          */
1018         create_pagetables(firstaddr);
1019
1020         virtual2_start = KvaStart;
1021         virtual2_end = PTOV_OFFSET;
1022
1023         virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
1024         virtual_start = pmap_kmem_choose(virtual_start);
1025
1026         virtual_end = VM_MAX_KERNEL_ADDRESS;
1027
1028         /* XXX do %cr0 as well */
1029         load_cr4(rcr4() | CR4_PGE | CR4_PSE);
1030         load_cr3(KPML4phys);
1031
1032         /*
1033          * Initialize protection array.
1034          */
1035         i386_protection_init();
1036
1037         /*
1038          * The kernel's pmap is statically allocated so we don't have to use
1039          * pmap_create, which is unlikely to work correctly at this part of
1040          * the boot sequence (XXX and which no longer exists).
1041          */
1042         kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
1043         kernel_pmap.pm_count = 1;
1044         CPUMASK_ASSALLONES(kernel_pmap.pm_active);
1045         RB_INIT(&kernel_pmap.pm_pvroot);
1046         spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
1047         for (i = 0; i < PM_PLACEMARKS; ++i)
1048                 kernel_pmap.pm_placemarks[i] = PM_NOPLACEMARK;
1049
1050         /*
1051          * Reserve some special page table entries/VA space for temporary
1052          * mapping of pages.
1053          */
1054 #define SYSMAP(c, p, v, n)      \
1055         v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
1056
1057         va = virtual_start;
1058         pte = vtopte(va);
1059
1060         /*
1061          * CMAP1/CMAP2 are used for zeroing and copying pages.
1062          */
1063         SYSMAP(caddr_t, CMAP1, CADDR1, 1)
1064
1065         /*
1066          * Crashdump maps.
1067          */
1068         SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
1069
1070         /*
1071          * ptvmmap is used for reading arbitrary physical pages via
1072          * /dev/mem.
1073          */
1074         SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
1075
1076         /*
1077          * msgbufp is used to map the system message buffer.
1078          * XXX msgbufmap is not used.
1079          */
1080         SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
1081                atop(round_page(MSGBUF_SIZE)))
1082
1083         virtual_start = va;
1084         virtual_start = pmap_kmem_choose(virtual_start);
1085
1086         *CMAP1 = 0;
1087
1088         /*
1089          * PG_G is terribly broken on SMP because we IPI invltlb's in some
1090          * cases rather then invl1pg.  Actually, I don't even know why it
1091          * works under UP because self-referential page table mappings
1092          */
1093 //      pgeflag = 0;
1094
1095         cpu_invltlb();
1096
1097         /* Initialize the PAT MSR */
1098         pmap_init_pat();
1099         pmap_pinit_defaults(&kernel_pmap);
1100
1101         TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1102                           &pmap_fast_kernel_cpusync);
1103
1104 }
1105
1106 /*
1107  * Setup the PAT MSR.
1108  */
1109 void
1110 pmap_init_pat(void)
1111 {
1112         uint64_t pat_msr;
1113         u_long cr0, cr4;
1114
1115         /*
1116          * Default values mapping PATi,PCD,PWT bits at system reset.
1117          * The default values effectively ignore the PATi bit by
1118          * repeating the encodings for 0-3 in 4-7, and map the PCD
1119          * and PWT bit combinations to the expected PAT types.
1120          */
1121         pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |        /* 000 */
1122                   PAT_VALUE(1, PAT_WRITE_THROUGH) |     /* 001 */
1123                   PAT_VALUE(2, PAT_UNCACHED) |          /* 010 */
1124                   PAT_VALUE(3, PAT_UNCACHEABLE) |       /* 011 */
1125                   PAT_VALUE(4, PAT_WRITE_BACK) |        /* 100 */
1126                   PAT_VALUE(5, PAT_WRITE_THROUGH) |     /* 101 */
1127                   PAT_VALUE(6, PAT_UNCACHED) |          /* 110 */
1128                   PAT_VALUE(7, PAT_UNCACHEABLE);        /* 111 */
1129         pat_pte_index[PAT_WRITE_BACK]   = 0;
1130         pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
1131         pat_pte_index[PAT_UNCACHED]     = X86_PG_NC_PCD;
1132         pat_pte_index[PAT_UNCACHEABLE]  = X86_PG_NC_PCD | X86_PG_NC_PWT;
1133         pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1134         pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1135
1136         if (cpu_feature & CPUID_PAT) {
1137                 /*
1138                  * If we support the PAT then set-up entries for
1139                  * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1140                  * 5 and 6.
1141                  */
1142                 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1143                           PAT_VALUE(5, PAT_WRITE_PROTECTED);
1144                 pat_msr = (pat_msr & ~PAT_MASK(6)) |
1145                           PAT_VALUE(6, PAT_WRITE_COMBINING);
1146                 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1147                 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PCD;
1148
1149                 /*
1150                  * Then enable the PAT
1151                  */
1152
1153                 /* Disable PGE. */
1154                 cr4 = rcr4();
1155                 load_cr4(cr4 & ~CR4_PGE);
1156
1157                 /* Disable caches (CD = 1, NW = 0). */
1158                 cr0 = rcr0();
1159                 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1160
1161                 /* Flushes caches and TLBs. */
1162                 wbinvd();
1163                 cpu_invltlb();
1164
1165                 /* Update PAT and index table. */
1166                 wrmsr(MSR_PAT, pat_msr);
1167
1168                 /* Flush caches and TLBs again. */
1169                 wbinvd();
1170                 cpu_invltlb();
1171
1172                 /* Restore caches and PGE. */
1173                 load_cr0(cr0);
1174                 load_cr4(cr4);
1175                 PatMsr = pat_msr;
1176         }
1177 }
1178
1179 /*
1180  * Set 4mb pdir for mp startup
1181  */
1182 void
1183 pmap_set_opt(void)
1184 {
1185         if (cpu_feature & CPUID_PSE) {
1186                 load_cr4(rcr4() | CR4_PSE);
1187                 if (mycpu->gd_cpuid == 0)       /* only on BSP */
1188                         cpu_invltlb();
1189         }
1190 }
1191
1192 /*
1193  *      Initialize the pmap module.
1194  *      Called by vm_init, to initialize any structures that the pmap
1195  *      system needs to map virtual memory.
1196  *      pmap_init has been enhanced to support in a fairly consistant
1197  *      way, discontiguous physical memory.
1198  */
1199 void
1200 pmap_init(void)
1201 {
1202         vm_pindex_t initial_pvs;
1203         vm_pindex_t i;
1204
1205         /*
1206          * Allocate memory for random pmap data structures.  Includes the
1207          * pv_head_table.
1208          */
1209
1210         for (i = 0; i < vm_page_array_size; i++) {
1211                 vm_page_t m;
1212
1213                 m = &vm_page_array[i];
1214                 TAILQ_INIT(&m->md.pv_list);
1215         }
1216
1217         /*
1218          * init the pv free list
1219          */
1220         initial_pvs = vm_page_array_size;
1221         if (initial_pvs < MINPV)
1222                 initial_pvs = MINPV;
1223         pvzone = &pvzone_store;
1224         pvinit = (void *)kmem_alloc(&kernel_map,
1225                                     initial_pvs * sizeof (struct pv_entry),
1226                                     VM_SUBSYS_PVENTRY);
1227         zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1228                   pvinit, initial_pvs);
1229
1230         /*
1231          * Now it is safe to enable pv_table recording.
1232          */
1233         pmap_initialized = TRUE;
1234 }
1235
1236 /*
1237  * Initialize the address space (zone) for the pv_entries.  Set a
1238  * high water mark so that the system can recover from excessive
1239  * numbers of pv entries.
1240  */
1241 void
1242 pmap_init2(void)
1243 {
1244         vm_pindex_t shpgperproc = PMAP_SHPGPERPROC;
1245         vm_pindex_t entry_max;
1246
1247         TUNABLE_LONG_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1248         pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1249         TUNABLE_LONG_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1250         pv_entry_high_water = 9 * (pv_entry_max / 10);
1251
1252         /*
1253          * Subtract out pages already installed in the zone (hack)
1254          */
1255         entry_max = pv_entry_max - vm_page_array_size;
1256         if (entry_max <= 0)
1257                 entry_max = 1;
1258
1259         zinitna(pvzone, NULL, 0, entry_max, ZONE_INTERRUPT);
1260
1261         /*
1262          * Enable dynamic deletion of empty higher-level page table pages
1263          * by default only if system memory is < 8GB (use 7GB for slop).
1264          * This can save a little memory, but imposes significant
1265          * performance overhead for things like bulk builds, and for programs
1266          * which do a lot of memory mapping and memory unmapping.
1267          */
1268         if (pmap_dynamic_delete < 0) {
1269                 if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1270                         pmap_dynamic_delete = 1;
1271                 else
1272                         pmap_dynamic_delete = 0;
1273         }
1274 }
1275
1276 /*
1277  * Typically used to initialize a fictitious page by vm/device_pager.c
1278  */
1279 void
1280 pmap_page_init(struct vm_page *m)
1281 {
1282         vm_page_init(m);
1283         TAILQ_INIT(&m->md.pv_list);
1284 }
1285
1286 /***************************************************
1287  * Low level helper routines.....
1288  ***************************************************/
1289
1290 /*
1291  * this routine defines the region(s) of memory that should
1292  * not be tested for the modified bit.
1293  */
1294 static __inline
1295 int
1296 pmap_track_modified(vm_pindex_t pindex)
1297 {
1298         vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1299         if ((va < clean_sva) || (va >= clean_eva))
1300                 return 1;
1301         else
1302                 return 0;
1303 }
1304
1305 /*
1306  * Extract the physical page address associated with the map/VA pair.
1307  * The page must be wired for this to work reliably.
1308  */
1309 vm_paddr_t
1310 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
1311 {
1312         vm_paddr_t rtval;
1313         pv_entry_t pt_pv;
1314         pt_entry_t *ptep;
1315
1316         rtval = 0;
1317         if (va >= VM_MAX_USER_ADDRESS) {
1318                 /*
1319                  * Kernel page directories might be direct-mapped and
1320                  * there is typically no PV tracking of pte's
1321                  */
1322                 pd_entry_t *pt;
1323
1324                 pt = pmap_pt(pmap, va);
1325                 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1326                         if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1327                                 rtval = *pt & PG_PS_FRAME;
1328                                 rtval |= va & PDRMASK;
1329                         } else {
1330                                 ptep = pmap_pt_to_pte(*pt, va);
1331                                 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1332                                         rtval = *ptep & PG_FRAME;
1333                                         rtval |= va & PAGE_MASK;
1334                                 }
1335                         }
1336                 }
1337                 if (handlep)
1338                         *handlep = NULL;
1339         } else {
1340                 /*
1341                  * User pages currently do not direct-map the page directory
1342                  * and some pages might not used managed PVs.  But all PT's
1343                  * will have a PV.
1344                  */
1345                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1346                 if (pt_pv) {
1347                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1348                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1349                                 rtval = *ptep & PG_FRAME;
1350                                 rtval |= va & PAGE_MASK;
1351                         }
1352                         if (handlep)
1353                                 *handlep = pt_pv;       /* locked until done */
1354                         else
1355                                 pv_put (pt_pv);
1356                 } else if (handlep) {
1357                         *handlep = NULL;
1358                 }
1359         }
1360         return rtval;
1361 }
1362
1363 void
1364 pmap_extract_done(void *handle)
1365 {
1366         if (handle)
1367                 pv_put((pv_entry_t)handle);
1368 }
1369
1370 /*
1371  * Similar to extract but checks protections, SMP-friendly short-cut for
1372  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1373  * fall-through to the real fault code.  Does not work with HVM page
1374  * tables.
1375  *
1376  * if busyp is NULL the returned page, if not NULL, is held (and not busied).
1377  *
1378  * If busyp is not NULL and this function sets *busyp non-zero, the returned
1379  * page is busied (and not held).
1380  *
1381  * If busyp is not NULL and this function sets *busyp to zero, the returned
1382  * page is held (and not busied).
1383  *
1384  * If VM_PROT_WRITE is set in prot, and the pte is already writable, the
1385  * returned page will be dirtied.  If the pte is not already writable NULL
1386  * is returned.  In otherwords, if the bit is set and a vm_page_t is returned,
1387  * any COW will already have happened and that page can be written by the
1388  * caller.
1389  *
1390  * WARNING! THE RETURNED PAGE IS ONLY HELD AND NOT SUITABLE FOR READING
1391  *          OR WRITING AS-IS.
1392  */
1393 vm_page_t
1394 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot, int *busyp)
1395 {
1396         if (pmap &&
1397             va < VM_MAX_USER_ADDRESS &&
1398             (pmap->pm_flags & PMAP_HVM) == 0) {
1399                 pv_entry_t pt_pv;
1400                 pv_entry_t pte_pv;
1401                 pt_entry_t *ptep;
1402                 pt_entry_t req;
1403                 vm_page_t m;
1404                 int error;
1405
1406                 req = pmap->pmap_bits[PG_V_IDX] |
1407                       pmap->pmap_bits[PG_U_IDX];
1408                 if (prot & VM_PROT_WRITE)
1409                         req |= pmap->pmap_bits[PG_RW_IDX];
1410
1411                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1412                 if (pt_pv == NULL)
1413                         return (NULL);
1414                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1415                 if ((*ptep & req) != req) {
1416                         pv_put(pt_pv);
1417                         return (NULL);
1418                 }
1419                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), NULL, &error);
1420                 if (pte_pv && error == 0) {
1421                         m = pte_pv->pv_m;
1422                         if (prot & VM_PROT_WRITE) {
1423                                 /* interlocked by presence of pv_entry */
1424                                 vm_page_dirty(m);
1425                         }
1426                         if (busyp) {
1427                                 if (prot & VM_PROT_WRITE) {
1428                                         if (vm_page_busy_try(m, TRUE))
1429                                                 m = NULL;
1430                                         *busyp = 1;
1431                                 } else {
1432                                         vm_page_hold(m);
1433                                         *busyp = 0;
1434                                 }
1435                         } else {
1436                                 vm_page_hold(m);
1437                         }
1438                         pv_put(pte_pv);
1439                 } else if (pte_pv) {
1440                         pv_drop(pte_pv);
1441                         m = NULL;
1442                 } else {
1443                         /* error, since we didn't request a placemarker */
1444                         m = NULL;
1445                 }
1446                 pv_put(pt_pv);
1447                 return(m);
1448         } else {
1449                 return(NULL);
1450         }
1451 }
1452
1453 /*
1454  * Extract the physical page address associated kernel virtual address.
1455  */
1456 vm_paddr_t
1457 pmap_kextract(vm_offset_t va)
1458 {
1459         pd_entry_t pt;          /* pt entry in pd */
1460         vm_paddr_t pa;
1461
1462         if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1463                 pa = DMAP_TO_PHYS(va);
1464         } else {
1465                 pt = *vtopt(va);
1466                 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1467                         pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1468                 } else {
1469                         /*
1470                          * Beware of a concurrent promotion that changes the
1471                          * PDE at this point!  For example, vtopte() must not
1472                          * be used to access the PTE because it would use the
1473                          * new PDE.  It is, however, safe to use the old PDE
1474                          * because the page table page is preserved by the
1475                          * promotion.
1476                          */
1477                         pa = *pmap_pt_to_pte(pt, va);
1478                         pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1479                 }
1480         }
1481         return pa;
1482 }
1483
1484 /***************************************************
1485  * Low level mapping routines.....
1486  ***************************************************/
1487
1488 /*
1489  * Routine: pmap_kenter
1490  * Function:
1491  *      Add a wired page to the KVA
1492  *      NOTE! note that in order for the mapping to take effect -- you
1493  *      should do an invltlb after doing the pmap_kenter().
1494  */
1495 void
1496 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1497 {
1498         pt_entry_t *ptep;
1499         pt_entry_t npte;
1500
1501         npte = pa |
1502                kernel_pmap.pmap_bits[PG_RW_IDX] |
1503                kernel_pmap.pmap_bits[PG_V_IDX];
1504 //             pgeflag;
1505         ptep = vtopte(va);
1506 #if 1
1507         pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1508 #else
1509         /* FUTURE */
1510         if (*ptep)
1511                 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1512         else
1513                 *ptep = npte;
1514 #endif
1515 }
1516
1517 /*
1518  * Similar to pmap_kenter(), except we only invalidate the mapping on the
1519  * current CPU.  Returns 0 if the previous pte was 0, 1 if it wasn't
1520  * (caller can conditionalize calling smp_invltlb()).
1521  */
1522 int
1523 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1524 {
1525         pt_entry_t *ptep;
1526         pt_entry_t npte;
1527         int res;
1528
1529         npte = pa | kernel_pmap.pmap_bits[PG_RW_IDX] |
1530                     kernel_pmap.pmap_bits[PG_V_IDX];
1531         // npte |= pgeflag;
1532         ptep = vtopte(va);
1533 #if 1
1534         res = 1;
1535 #else
1536         /* FUTURE */
1537         res = (*ptep != 0);
1538 #endif
1539         atomic_swap_long(ptep, npte);
1540         cpu_invlpg((void *)va);
1541
1542         return res;
1543 }
1544
1545 /*
1546  * Enter addresses into the kernel pmap but don't bother
1547  * doing any tlb invalidations.  Caller will do a rollup
1548  * invalidation via pmap_rollup_inval().
1549  */
1550 int
1551 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1552 {
1553         pt_entry_t *ptep;
1554         pt_entry_t npte;
1555         int res;
1556
1557         npte = pa |
1558             kernel_pmap.pmap_bits[PG_RW_IDX] |
1559             kernel_pmap.pmap_bits[PG_V_IDX];
1560 //          pgeflag;
1561         ptep = vtopte(va);
1562 #if 1
1563         res = 1;
1564 #else
1565         /* FUTURE */
1566         res = (*ptep != 0);
1567 #endif
1568         atomic_swap_long(ptep, npte);
1569         cpu_invlpg((void *)va);
1570
1571         return res;
1572 }
1573
1574 /*
1575  * remove a page from the kernel pagetables
1576  */
1577 void
1578 pmap_kremove(vm_offset_t va)
1579 {
1580         pt_entry_t *ptep;
1581
1582         ptep = vtopte(va);
1583         pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1584 }
1585
1586 void
1587 pmap_kremove_quick(vm_offset_t va)
1588 {
1589         pt_entry_t *ptep;
1590
1591         ptep = vtopte(va);
1592         (void)pte_load_clear(ptep);
1593         cpu_invlpg((void *)va);
1594 }
1595
1596 /*
1597  * Remove addresses from the kernel pmap but don't bother
1598  * doing any tlb invalidations.  Caller will do a rollup
1599  * invalidation via pmap_rollup_inval().
1600  */
1601 void
1602 pmap_kremove_noinval(vm_offset_t va)
1603 {
1604         pt_entry_t *ptep;
1605
1606         ptep = vtopte(va);
1607         (void)pte_load_clear(ptep);
1608 }
1609
1610 /*
1611  * XXX these need to be recoded.  They are not used in any critical path.
1612  */
1613 void
1614 pmap_kmodify_rw(vm_offset_t va)
1615 {
1616         atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1617         cpu_invlpg((void *)va);
1618 }
1619
1620 /* NOT USED
1621 void
1622 pmap_kmodify_nc(vm_offset_t va)
1623 {
1624         atomic_set_long(vtopte(va), PG_N);
1625         cpu_invlpg((void *)va);
1626 }
1627 */
1628
1629 /*
1630  * Used to map a range of physical addresses into kernel virtual
1631  * address space during the low level boot, typically to map the
1632  * dump bitmap, message buffer, and vm_page_array.
1633  *
1634  * These mappings are typically made at some pointer after the end of the
1635  * kernel text+data.
1636  *
1637  * We could return PHYS_TO_DMAP(start) here and not allocate any
1638  * via (*virtp), but then kmem from userland and kernel dumps won't
1639  * have access to the related pointers.
1640  */
1641 vm_offset_t
1642 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1643 {
1644         vm_offset_t va;
1645         vm_offset_t va_start;
1646
1647         /*return PHYS_TO_DMAP(start);*/
1648
1649         va_start = *virtp;
1650         va = va_start;
1651
1652         while (start < end) {
1653                 pmap_kenter_quick(va, start);
1654                 va += PAGE_SIZE;
1655                 start += PAGE_SIZE;
1656         }
1657         *virtp = va;
1658         return va_start;
1659 }
1660
1661 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
1662
1663 /*
1664  * Remove the specified set of pages from the data and instruction caches.
1665  *
1666  * In contrast to pmap_invalidate_cache_range(), this function does not
1667  * rely on the CPU's self-snoop feature, because it is intended for use
1668  * when moving pages into a different cache domain.
1669  */
1670 void
1671 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1672 {
1673         vm_offset_t daddr, eva;
1674         int i;
1675
1676         if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1677             (cpu_feature & CPUID_CLFSH) == 0)
1678                 wbinvd();
1679         else {
1680                 cpu_mfence();
1681                 for (i = 0; i < count; i++) {
1682                         daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1683                         eva = daddr + PAGE_SIZE;
1684                         for (; daddr < eva; daddr += cpu_clflush_line_size)
1685                                 clflush(daddr);
1686                 }
1687                 cpu_mfence();
1688         }
1689 }
1690
1691 void
1692 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1693 {
1694         KASSERT((sva & PAGE_MASK) == 0,
1695             ("pmap_invalidate_cache_range: sva not page-aligned"));
1696         KASSERT((eva & PAGE_MASK) == 0,
1697             ("pmap_invalidate_cache_range: eva not page-aligned"));
1698
1699         if (cpu_feature & CPUID_SS) {
1700                 ; /* If "Self Snoop" is supported, do nothing. */
1701         } else {
1702                 /* Globally invalidate caches */
1703                 cpu_wbinvd_on_all_cpus();
1704         }
1705 }
1706
1707 /*
1708  * Invalidate the specified range of virtual memory on all cpus associated
1709  * with the pmap.
1710  */
1711 void
1712 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1713 {
1714         pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1715 }
1716
1717 /*
1718  * Add a list of wired pages to the kva.  This routine is used for temporary
1719  * kernel mappings such as those found in buffer cache buffer.  Page
1720  * modifications and accesses are not tracked or recorded.
1721  *
1722  * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1723  *       semantics as previous mappings may have been zerod without any
1724  *       invalidation.
1725  *
1726  * The page *must* be wired.
1727  */
1728 static __inline void
1729 _pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count, int doinval)
1730 {
1731         vm_offset_t end_va;
1732         vm_offset_t va;
1733
1734         end_va = beg_va + count * PAGE_SIZE;
1735
1736         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1737                 pt_entry_t pte;
1738                 pt_entry_t *ptep;
1739
1740                 ptep = vtopte(va);
1741                 pte = VM_PAGE_TO_PHYS(*m) |
1742                         kernel_pmap.pmap_bits[PG_RW_IDX] |
1743                         kernel_pmap.pmap_bits[PG_V_IDX] |
1744                         kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1745 //              pgeflag;
1746                 atomic_swap_long(ptep, pte);
1747                 m++;
1748         }
1749         if (doinval)
1750                 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1751 }
1752
1753 void
1754 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1755 {
1756         _pmap_qenter(beg_va, m, count, 1);
1757 }
1758
1759 void
1760 pmap_qenter_noinval(vm_offset_t beg_va, vm_page_t *m, int count)
1761 {
1762         _pmap_qenter(beg_va, m, count, 0);
1763 }
1764
1765 /*
1766  * This routine jerks page mappings from the kernel -- it is meant only
1767  * for temporary mappings such as those found in buffer cache buffers.
1768  * No recording modified or access status occurs.
1769  *
1770  * MPSAFE, INTERRUPT SAFE (cluster callback)
1771  */
1772 void
1773 pmap_qremove(vm_offset_t beg_va, int count)
1774 {
1775         vm_offset_t end_va;
1776         vm_offset_t va;
1777
1778         end_va = beg_va + count * PAGE_SIZE;
1779
1780         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1781                 pt_entry_t *pte;
1782
1783                 pte = vtopte(va);
1784                 (void)pte_load_clear(pte);
1785                 cpu_invlpg((void *)va);
1786         }
1787         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1788 }
1789
1790 /*
1791  * This routine removes temporary kernel mappings, only invalidating them
1792  * on the current cpu.  It should only be used under carefully controlled
1793  * conditions.
1794  */
1795 void
1796 pmap_qremove_quick(vm_offset_t beg_va, int count)
1797 {
1798         vm_offset_t end_va;
1799         vm_offset_t va;
1800
1801         end_va = beg_va + count * PAGE_SIZE;
1802
1803         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1804                 pt_entry_t *pte;
1805
1806                 pte = vtopte(va);
1807                 (void)pte_load_clear(pte);
1808                 cpu_invlpg((void *)va);
1809         }
1810 }
1811
1812 /*
1813  * This routine removes temporary kernel mappings *without* invalidating
1814  * the TLB.  It can only be used on permanent kva reservations such as those
1815  * found in buffer cache buffers, under carefully controlled circumstances.
1816  *
1817  * NOTE: Repopulating these KVAs requires unconditional invalidation.
1818  *       (pmap_qenter() does unconditional invalidation).
1819  */
1820 void
1821 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1822 {
1823         vm_offset_t end_va;
1824         vm_offset_t va;
1825
1826         end_va = beg_va + count * PAGE_SIZE;
1827
1828         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1829                 pt_entry_t *pte;
1830
1831                 pte = vtopte(va);
1832                 (void)pte_load_clear(pte);
1833         }
1834 }
1835
1836 /*
1837  * Create a new thread and optionally associate it with a (new) process.
1838  * NOTE! the new thread's cpu may not equal the current cpu.
1839  */
1840 void
1841 pmap_init_thread(thread_t td)
1842 {
1843         /* enforce pcb placement & alignment */
1844         td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1845         td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1846         td->td_savefpu = &td->td_pcb->pcb_save;
1847         td->td_sp = (char *)td->td_pcb; /* no -16 */
1848 }
1849
1850 /*
1851  * This routine directly affects the fork perf for a process.
1852  */
1853 void
1854 pmap_init_proc(struct proc *p)
1855 {
1856 }
1857
1858 static void
1859 pmap_pinit_defaults(struct pmap *pmap)
1860 {
1861         bcopy(pmap_bits_default, pmap->pmap_bits,
1862               sizeof(pmap_bits_default));
1863         bcopy(protection_codes, pmap->protection_codes,
1864               sizeof(protection_codes));
1865         bcopy(pat_pte_index, pmap->pmap_cache_bits,
1866               sizeof(pat_pte_index));
1867         pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1868         pmap->copyinstr = std_copyinstr;
1869         pmap->copyin = std_copyin;
1870         pmap->copyout = std_copyout;
1871         pmap->fubyte = std_fubyte;
1872         pmap->subyte = std_subyte;
1873         pmap->fuword32 = std_fuword32;
1874         pmap->fuword64 = std_fuword64;
1875         pmap->suword32 = std_suword32;
1876         pmap->suword64 = std_suword64;
1877         pmap->swapu32 = std_swapu32;
1878         pmap->swapu64 = std_swapu64;
1879 }
1880 /*
1881  * Initialize pmap0/vmspace0.
1882  *
1883  * On architectures where the kernel pmap is not integrated into the user
1884  * process pmap, this pmap represents the process pmap, not the kernel pmap.
1885  * kernel_pmap should be used to directly access the kernel_pmap.
1886  */
1887 void
1888 pmap_pinit0(struct pmap *pmap)
1889 {
1890         int i;
1891
1892         pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1893         pmap->pm_count = 1;
1894         CPUMASK_ASSZERO(pmap->pm_active);
1895         pmap->pm_pvhint_pt = NULL;
1896         pmap->pm_pvhint_pte = NULL;
1897         RB_INIT(&pmap->pm_pvroot);
1898         spin_init(&pmap->pm_spin, "pmapinit0");
1899         for (i = 0; i < PM_PLACEMARKS; ++i)
1900                 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1901         bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1902         pmap_pinit_defaults(pmap);
1903 }
1904
1905 /*
1906  * Initialize a preallocated and zeroed pmap structure,
1907  * such as one in a vmspace structure.
1908  */
1909 static void
1910 pmap_pinit_simple(struct pmap *pmap)
1911 {
1912         int i;
1913
1914         /*
1915          * Misc initialization
1916          */
1917         pmap->pm_count = 1;
1918         CPUMASK_ASSZERO(pmap->pm_active);
1919         pmap->pm_pvhint_pt = NULL;
1920         pmap->pm_pvhint_pte = NULL;
1921         pmap->pm_flags = PMAP_FLAG_SIMPLE;
1922
1923         pmap_pinit_defaults(pmap);
1924
1925         /*
1926          * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1927          * for this).
1928          */
1929         if (pmap->pm_pmlpv == NULL) {
1930                 RB_INIT(&pmap->pm_pvroot);
1931                 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1932                 spin_init(&pmap->pm_spin, "pmapinitsimple");
1933                 for (i = 0; i < PM_PLACEMARKS; ++i)
1934                         pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1935         }
1936 }
1937
1938 void
1939 pmap_pinit(struct pmap *pmap)
1940 {
1941         pv_entry_t pv;
1942         int j;
1943
1944         if (pmap->pm_pmlpv) {
1945                 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1946                         pmap_puninit(pmap);
1947                 }
1948         }
1949
1950         pmap_pinit_simple(pmap);
1951         pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1952
1953         /*
1954          * No need to allocate page table space yet but we do need a valid
1955          * page directory table.
1956          */
1957         if (pmap->pm_pml4 == NULL) {
1958                 pmap->pm_pml4 =
1959                     (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1960                                                         PAGE_SIZE,
1961                                                         VM_SUBSYS_PML4);
1962         }
1963
1964         /*
1965          * Allocate the page directory page, which wires it even though
1966          * it isn't being entered into some higher level page table (it
1967          * being the highest level).  If one is already cached we don't
1968          * have to do anything.
1969          */
1970         if ((pv = pmap->pm_pmlpv) == NULL) {
1971                 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1972                 pmap->pm_pmlpv = pv;
1973                 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1974                             VM_PAGE_TO_PHYS(pv->pv_m));
1975                 pv_put(pv);
1976
1977                 /*
1978                  * Install DMAP and KMAP.
1979                  */
1980                 for (j = 0; j < NDMPML4E; ++j) {
1981                         pmap->pm_pml4[DMPML4I + j] =
1982                             (DMPDPphys + ((vm_paddr_t)j << PAGE_SHIFT)) |
1983                             pmap->pmap_bits[PG_RW_IDX] |
1984                             pmap->pmap_bits[PG_V_IDX] |
1985                             pmap->pmap_bits[PG_U_IDX];
1986                 }
1987                 pmap->pm_pml4[KPML4I] = KPDPphys |
1988                     pmap->pmap_bits[PG_RW_IDX] |
1989                     pmap->pmap_bits[PG_V_IDX] |
1990                     pmap->pmap_bits[PG_U_IDX];
1991
1992                 /*
1993                  * install self-referential address mapping entry
1994                  */
1995                 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1996                     pmap->pmap_bits[PG_V_IDX] |
1997                     pmap->pmap_bits[PG_RW_IDX] |
1998                     pmap->pmap_bits[PG_A_IDX] |
1999                     pmap->pmap_bits[PG_M_IDX];
2000         } else {
2001                 KKASSERT(pv->pv_m->flags & PG_MAPPED);
2002                 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
2003         }
2004         KKASSERT(pmap->pm_pml4[255] == 0);
2005         KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
2006         KKASSERT(pv->pv_entry.rbe_left == NULL);
2007         KKASSERT(pv->pv_entry.rbe_right == NULL);
2008 }
2009
2010 /*
2011  * Clean up a pmap structure so it can be physically freed.  This routine
2012  * is called by the vmspace dtor function.  A great deal of pmap data is
2013  * left passively mapped to improve vmspace management so we have a bit
2014  * of cleanup work to do here.
2015  */
2016 void
2017 pmap_puninit(pmap_t pmap)
2018 {
2019         pv_entry_t pv;
2020         vm_page_t p;
2021
2022         KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
2023         if ((pv = pmap->pm_pmlpv) != NULL) {
2024                 if (pv_hold_try(pv) == 0)
2025                         pv_lock(pv);
2026                 KKASSERT(pv == pmap->pm_pmlpv);
2027                 p = pmap_remove_pv_page(pv);
2028                 pv_free(pv, NULL);
2029                 pv = NULL;      /* safety */
2030                 pmap_kremove((vm_offset_t)pmap->pm_pml4);
2031                 vm_page_busy_wait(p, FALSE, "pgpun");
2032                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
2033                 vm_page_unwire(p, 0);
2034                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2035
2036                 /*
2037                  * XXX eventually clean out PML4 static entries and
2038                  * use vm_page_free_zero()
2039                  */
2040                 vm_page_free(p);
2041                 pmap->pm_pmlpv = NULL;
2042         }
2043         if (pmap->pm_pml4) {
2044                 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
2045                 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
2046                 pmap->pm_pml4 = NULL;
2047         }
2048         KKASSERT(pmap->pm_stats.resident_count == 0);
2049         KKASSERT(pmap->pm_stats.wired_count == 0);
2050 }
2051
2052 /*
2053  * This function is now unused (used to add the pmap to the pmap_list)
2054  */
2055 void
2056 pmap_pinit2(struct pmap *pmap)
2057 {
2058 }
2059
2060 /*
2061  * This routine is called when various levels in the page table need to
2062  * be populated.  This routine cannot fail.
2063  *
2064  * This function returns two locked pv_entry's, one representing the
2065  * requested pv and one representing the requested pv's parent pv.  If
2066  * an intermediate page table does not exist it will be created, mapped,
2067  * wired, and the parent page table will be given an additional hold
2068  * count representing the presence of the child pv_entry.
2069  */
2070 static
2071 pv_entry_t
2072 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
2073 {
2074         pt_entry_t *ptep;
2075         pv_entry_t pv;
2076         pv_entry_t pvp;
2077         pt_entry_t v;
2078         vm_pindex_t pt_pindex;
2079         vm_page_t m;
2080         int isnew;
2081         int ispt;
2082
2083         /*
2084          * If the pv already exists and we aren't being asked for the
2085          * parent page table page we can just return it.  A locked+held pv
2086          * is returned.  The pv will also have a second hold related to the
2087          * pmap association that we don't have to worry about.
2088          */
2089         ispt = 0;
2090         pv = pv_alloc(pmap, ptepindex, &isnew);
2091         if (isnew == 0 && pvpp == NULL)
2092                 return(pv);
2093
2094         /*
2095          * Special case terminal PVs.  These are not page table pages so
2096          * no vm_page is allocated (the caller supplied the vm_page).  If
2097          * pvpp is non-NULL we are being asked to also removed the pt_pv
2098          * for this pv.
2099          *
2100          * Note that pt_pv's are only returned for user VAs. We assert that
2101          * a pt_pv is not being requested for kernel VAs.  The kernel
2102          * pre-wires all higher-level page tables so don't overload managed
2103          * higher-level page tables on top of it!
2104          */
2105         if (ptepindex < pmap_pt_pindex(0)) {
2106                 if (ptepindex >= NUPTE_USER) {
2107                         /* kernel manages this manually for KVM */
2108                         KKASSERT(pvpp == NULL);
2109                 } else {
2110                         KKASSERT(pvpp != NULL);
2111                         pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
2112                         pvp = pmap_allocpte(pmap, pt_pindex, NULL);
2113                         if (isnew)
2114                                 vm_page_wire_quick(pvp->pv_m);
2115                         *pvpp = pvp;
2116                 }
2117                 return(pv);
2118         }
2119
2120         /*
2121          * The kernel never uses managed PT/PD/PDP pages.
2122          */
2123         KKASSERT(pmap != &kernel_pmap);
2124
2125         /*
2126          * Non-terminal PVs allocate a VM page to represent the page table,
2127          * so we have to resolve pvp and calculate ptepindex for the pvp
2128          * and then for the page table entry index in the pvp for
2129          * fall-through.
2130          */
2131         if (ptepindex < pmap_pd_pindex(0)) {
2132                 /*
2133                  * pv is PT, pvp is PD
2134                  */
2135                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
2136                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
2137                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2138
2139                 /*
2140                  * PT index in PD
2141                  */
2142                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
2143                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
2144                 ispt = 1;
2145         } else if (ptepindex < pmap_pdp_pindex(0)) {
2146                 /*
2147                  * pv is PD, pvp is PDP
2148                  *
2149                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2150                  *                   the PD.
2151                  */
2152                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
2153                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2154
2155                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
2156                         KKASSERT(pvpp == NULL);
2157                         pvp = NULL;
2158                 } else {
2159                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
2160                 }
2161
2162                 /*
2163                  * PD index in PDP
2164                  */
2165                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
2166                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
2167         } else if (ptepindex < pmap_pml4_pindex()) {
2168                 /*
2169                  * pv is PDP, pvp is the root pml4 table
2170                  */
2171                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2172
2173                 /*
2174                  * PDP index in PML4
2175                  */
2176                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2177                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2178         } else {
2179                 /*
2180                  * pv represents the top-level PML4, there is no parent.
2181                  */
2182                 pvp = NULL;
2183         }
2184
2185         if (isnew == 0)
2186                 goto notnew;
2187
2188         /*
2189          * (isnew) is TRUE, pv is not terminal.
2190          *
2191          * (1) Add a wire count to the parent page table (pvp).
2192          * (2) Allocate a VM page for the page table.
2193          * (3) Enter the VM page into the parent page table.
2194          *
2195          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2196          */
2197         if (pvp)
2198                 vm_page_wire_quick(pvp->pv_m);
2199
2200         for (;;) {
2201                 m = vm_page_alloc(NULL, pv->pv_pindex,
2202                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2203                                   VM_ALLOC_INTERRUPT);
2204                 if (m)
2205                         break;
2206                 vm_wait(0);
2207         }
2208         vm_page_wire(m);        /* wire for mapping in parent */
2209         vm_page_unmanage(m);    /* m must be spinunlocked */
2210         pmap_zero_page(VM_PAGE_TO_PHYS(m));
2211         m->valid = VM_PAGE_BITS_ALL;
2212
2213         vm_page_spin_lock(m);
2214         pmap_page_stats_adding(m);
2215         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2216         pv->pv_m = m;
2217         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2218         vm_page_spin_unlock(m);
2219
2220         /*
2221          * (isnew) is TRUE, pv is not terminal.
2222          *
2223          * Wire the page into pvp.  Bump the resident_count for the pmap.
2224          * There is no pvp for the top level, address the pm_pml4[] array
2225          * directly.
2226          *
2227          * If the caller wants the parent we return it, otherwise
2228          * we just put it away.
2229          *
2230          * No interlock is needed for pte 0 -> non-zero.
2231          *
2232          * In the situation where *ptep is valid we might have an unmanaged
2233          * page table page shared from another page table which we need to
2234          * unshare before installing our private page table page.
2235          */
2236         if (pvp) {
2237                 v = VM_PAGE_TO_PHYS(m) |
2238                     (pmap->pmap_bits[PG_U_IDX] |
2239                      pmap->pmap_bits[PG_RW_IDX] |
2240                      pmap->pmap_bits[PG_V_IDX] |
2241                      pmap->pmap_bits[PG_A_IDX] |
2242                      pmap->pmap_bits[PG_M_IDX]);
2243                 ptep = pv_pte_lookup(pvp, ptepindex);
2244                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2245                         pt_entry_t pte;
2246
2247                         if (ispt == 0) {
2248                                 panic("pmap_allocpte: unexpected pte %p/%d",
2249                                       pvp, (int)ptepindex);
2250                         }
2251                         pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, v);
2252                         if (vm_page_unwire_quick(
2253                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2254                                 panic("pmap_allocpte: shared pgtable "
2255                                       "pg bad wirecount");
2256                         }
2257                 } else {
2258                         pt_entry_t pte;
2259
2260                         pte = atomic_swap_long(ptep, v);
2261                         if (pte != 0) {
2262                                 kprintf("install pgtbl mixup 0x%016jx "
2263                                         "old/new 0x%016jx/0x%016jx\n",
2264                                         (intmax_t)ptepindex, pte, v);
2265                         }
2266                 }
2267         }
2268         vm_page_wakeup(m);
2269
2270         /*
2271          * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2272          */
2273 notnew:
2274         if (pvp) {
2275                 KKASSERT(pvp->pv_m != NULL);
2276                 ptep = pv_pte_lookup(pvp, ptepindex);
2277                 v = VM_PAGE_TO_PHYS(pv->pv_m) |
2278                     (pmap->pmap_bits[PG_U_IDX] |
2279                      pmap->pmap_bits[PG_RW_IDX] |
2280                      pmap->pmap_bits[PG_V_IDX] |
2281                      pmap->pmap_bits[PG_A_IDX] |
2282                      pmap->pmap_bits[PG_M_IDX]);
2283                 if (*ptep != v) {
2284                         kprintf("mismatched upper level pt %016jx/%016jx\n",
2285                                 *ptep, v);
2286                 }
2287         }
2288         if (pvpp)
2289                 *pvpp = pvp;
2290         else if (pvp)
2291                 pv_put(pvp);
2292         return (pv);
2293 }
2294
2295 /*
2296  * This version of pmap_allocpte() checks for possible segment optimizations
2297  * that would allow page-table sharing.  It can be called for terminal
2298  * page or page table page ptepindex's.
2299  *
2300  * The function is called with page table page ptepindex's for fictitious
2301  * and unmanaged terminal pages.  That is, we don't want to allocate a
2302  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
2303  * for this case.
2304  *
2305  * This function can return a pv and *pvpp associated with the passed in pmap
2306  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
2307  * an unmanaged page table page will be entered into the pass in pmap.
2308  */
2309 static
2310 pv_entry_t
2311 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2312                   vm_map_entry_t entry, vm_offset_t va)
2313 {
2314         vm_object_t object;
2315         pmap_t obpmap;
2316         pmap_t *obpmapp;
2317         vm_pindex_t *pt_placemark;
2318         vm_offset_t b;
2319         pv_entry_t pte_pv;      /* in original or shared pmap */
2320         pv_entry_t pt_pv;       /* in original or shared pmap */
2321         pv_entry_t proc_pd_pv;  /* in original pmap */
2322         pv_entry_t proc_pt_pv;  /* in original pmap */
2323         pv_entry_t xpv;         /* PT in shared pmap */
2324         pd_entry_t *pt;         /* PT entry in PD of original pmap */
2325         pd_entry_t opte;        /* contents of *pt */
2326         pd_entry_t npte;        /* contents of *pt */
2327         vm_page_t m;
2328         int softhold;
2329
2330         /*
2331          * Basic tests, require a non-NULL vm_map_entry, require proper
2332          * alignment and type for the vm_map_entry, require that the
2333          * underlying object already be allocated.
2334          *
2335          * We allow almost any type of object to use this optimization.
2336          * The object itself does NOT have to be sized to a multiple of the
2337          * segment size, but the memory mapping does.
2338          *
2339          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2340          *     won't work as expected.
2341          */
2342         if (entry == NULL ||
2343             pmap_mmu_optimize == 0 ||                   /* not enabled */
2344             (pmap->pm_flags & PMAP_HVM) ||              /* special pmap */
2345             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal or pt */
2346             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2347             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2348             entry->object.vm_object == NULL ||          /* needs VM object */
2349             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2350             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2351             (entry->offset & SEG_MASK) ||               /* must be aligned */
2352             (entry->start & SEG_MASK)) {
2353                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2354         }
2355
2356         /*
2357          * Make sure the full segment can be represented.
2358          */
2359         b = va & ~(vm_offset_t)SEG_MASK;
2360         if (b < entry->start || b + SEG_SIZE > entry->end)
2361                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2362
2363         /*
2364          * If the full segment can be represented dive the VM object's
2365          * shared pmap, allocating as required.
2366          */
2367         object = entry->object.vm_object;
2368
2369         if (entry->protection & VM_PROT_WRITE)
2370                 obpmapp = &object->md.pmap_rw;
2371         else
2372                 obpmapp = &object->md.pmap_ro;
2373
2374 #ifdef PMAP_DEBUG2
2375         if (pmap_enter_debug > 0) {
2376                 --pmap_enter_debug;
2377                 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2378                         "obpmapp %p %p\n",
2379                         va, entry->protection, object,
2380                         obpmapp, *obpmapp);
2381                 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2382                         entry, entry->start, entry->end);
2383         }
2384 #endif
2385
2386         /*
2387          * We allocate what appears to be a normal pmap but because portions
2388          * of this pmap are shared with other unrelated pmaps we have to
2389          * set pm_active to point to all cpus.
2390          *
2391          * XXX Currently using pmap_spin to interlock the update, can't use
2392          *     vm_object_hold/drop because the token might already be held
2393          *     shared OR exclusive and we don't know.
2394          */
2395         while ((obpmap = *obpmapp) == NULL) {
2396                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2397                 pmap_pinit_simple(obpmap);
2398                 pmap_pinit2(obpmap);
2399                 spin_lock(&pmap_spin);
2400                 if (*obpmapp != NULL) {
2401                         /*
2402                          * Handle race
2403                          */
2404                         spin_unlock(&pmap_spin);
2405                         pmap_release(obpmap);
2406                         pmap_puninit(obpmap);
2407                         kfree(obpmap, M_OBJPMAP);
2408                         obpmap = *obpmapp; /* safety */
2409                 } else {
2410                         obpmap->pm_active = smp_active_mask;
2411                         obpmap->pm_flags |= PMAP_SEGSHARED;
2412                         *obpmapp = obpmap;
2413                         spin_unlock(&pmap_spin);
2414                 }
2415         }
2416
2417         /*
2418          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2419          * pte/pt using the shared pmap from the object but also adjust
2420          * the process pmap's page table page as a side effect.
2421          */
2422
2423         /*
2424          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2425          * we will return.  This is true if ptepindex represents a terminal
2426          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2427          * the PD.
2428          */
2429         pt_pv = NULL;
2430         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2431         softhold = 0;
2432 retry:
2433         if (ptepindex >= pmap_pt_pindex(0))
2434                 xpv = pte_pv;
2435         else
2436                 xpv = pt_pv;
2437
2438         /*
2439          * Resolve the PD in the process pmap so we can properly share the
2440          * page table page.  Lock order is bottom-up (leaf first)!
2441          *
2442          * NOTE: proc_pt_pv can be NULL.
2443          */
2444         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b), &pt_placemark);
2445         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2446 #ifdef PMAP_DEBUG2
2447         if (pmap_enter_debug > 0) {
2448                 --pmap_enter_debug;
2449                 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2450                         proc_pt_pv,
2451                         (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2452                         proc_pd_pv,
2453                         va);
2454         }
2455 #endif
2456
2457         /*
2458          * xpv is the page table page pv from the shared object
2459          * (for convenience), from above.
2460          *
2461          * Calculate the pte value for the PT to load into the process PD.
2462          * If we have to change it we must properly dispose of the previous
2463          * entry.
2464          */
2465         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2466         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2467                (pmap->pmap_bits[PG_U_IDX] |
2468                 pmap->pmap_bits[PG_RW_IDX] |
2469                 pmap->pmap_bits[PG_V_IDX] |
2470                 pmap->pmap_bits[PG_A_IDX] |
2471                 pmap->pmap_bits[PG_M_IDX]);
2472
2473         /*
2474          * Dispose of previous page table page if it was local to the
2475          * process pmap.  If the old pt is not empty we cannot dispose of it
2476          * until we clean it out.  This case should not arise very often so
2477          * it is not optimized.
2478          *
2479          * Leave pt_pv and pte_pv (in our object pmap) locked and intact
2480          * for the retry.
2481          */
2482         if (proc_pt_pv) {
2483                 pmap_inval_bulk_t bulk;
2484
2485                 if (proc_pt_pv->pv_m->wire_count != 1) {
2486                         /*
2487                          * The page table has a bunch of stuff in it
2488                          * which we have to scrap.
2489                          */
2490                         if (softhold == 0) {
2491                                 softhold = 1;
2492                                 pmap_softhold(pmap);
2493                         }
2494                         pv_put(proc_pd_pv);
2495                         pv_put(proc_pt_pv);
2496                         pmap_remove(pmap,
2497                                     va & ~(vm_offset_t)SEG_MASK,
2498                                     (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2499                 } else {
2500                         /*
2501                          * The page table is empty and can be destroyed.
2502                          * However, doing so leaves the pt slot unlocked,
2503                          * so we have to loop-up to handle any races until
2504                          * we get a NULL proc_pt_pv and a proper pt_placemark.
2505                          */
2506                         pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2507                         pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2508                         pmap_inval_bulk_flush(&bulk);
2509                         pv_put(proc_pd_pv);
2510                 }
2511                 goto retry;
2512         }
2513
2514         /*
2515          * Handle remaining cases.  We are holding pt_placemark to lock
2516          * the page table page in the primary pmap while we manipulate
2517          * it.
2518          */
2519         if (*pt == 0) {
2520                 atomic_swap_long(pt, npte);
2521                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2522                 vm_page_wire_quick(proc_pd_pv->pv_m);   /* proc pd for sh pt */
2523                 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2524         } else if (*pt != npte) {
2525                 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2526
2527 #if 0
2528                 opte = pte_load_clear(pt);
2529                 KKASSERT(opte && opte != npte);
2530
2531                 *pt = npte;
2532 #endif
2533                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2534
2535                 /*
2536                  * Clean up opte, bump the wire_count for the process
2537                  * PD page representing the new entry if it was
2538                  * previously empty.
2539                  *
2540                  * If the entry was not previously empty and we have
2541                  * a PT in the proc pmap then opte must match that
2542                  * pt.  The proc pt must be retired (this is done
2543                  * later on in this procedure).
2544                  *
2545                  * NOTE: replacing valid pte, wire_count on proc_pd_pv
2546                  * stays the same.
2547                  */
2548                 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2549                 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2550                 if (vm_page_unwire_quick(m)) {
2551                         panic("pmap_allocpte_seg: "
2552                               "bad wire count %p",
2553                               m);
2554                 }
2555         }
2556
2557         if (softhold)
2558                 pmap_softdone(pmap);
2559
2560         /*
2561          * Remove our earmark on the page table page.
2562          */
2563         pv_placemarker_wakeup(pmap, pt_placemark);
2564
2565         /*
2566          * The existing process page table was replaced and must be destroyed
2567          * here.
2568          */
2569         if (proc_pd_pv)
2570                 pv_put(proc_pd_pv);
2571         if (pvpp)
2572                 *pvpp = pt_pv;
2573         else
2574                 pv_put(pt_pv);
2575         return (pte_pv);
2576 }
2577
2578 /*
2579  * Release any resources held by the given physical map.
2580  *
2581  * Called when a pmap initialized by pmap_pinit is being released.  Should
2582  * only be called if the map contains no valid mappings.
2583  */
2584 struct pmap_release_info {
2585         pmap_t  pmap;
2586         int     retry;
2587         pv_entry_t pvp;
2588 };
2589
2590 static int pmap_release_callback(pv_entry_t pv, void *data);
2591
2592 void
2593 pmap_release(struct pmap *pmap)
2594 {
2595         struct pmap_release_info info;
2596
2597         KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2598                 ("pmap still active! %016jx",
2599                 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2600
2601         /*
2602          * There is no longer a pmap_list, if there were we would remove the
2603          * pmap from it here.
2604          */
2605
2606         /*
2607          * Pull pv's off the RB tree in order from low to high and release
2608          * each page.
2609          */
2610         info.pmap = pmap;
2611         do {
2612                 info.retry = 0;
2613                 info.pvp = NULL;
2614
2615                 spin_lock(&pmap->pm_spin);
2616                 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2617                         pmap_release_callback, &info);
2618                 spin_unlock(&pmap->pm_spin);
2619
2620                 if (info.pvp)
2621                         pv_put(info.pvp);
2622         } while (info.retry);
2623
2624
2625         /*
2626          * One resident page (the pml4 page) should remain.
2627          * No wired pages should remain.
2628          */
2629 #if 1
2630         if (pmap->pm_stats.resident_count !=
2631             ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1) ||
2632             pmap->pm_stats.wired_count != 0) {
2633                 kprintf("fatal pmap problem - pmap %p flags %08x "
2634                         "rescnt=%jd wirecnt=%jd\n",
2635                         pmap,
2636                         pmap->pm_flags,
2637                         pmap->pm_stats.resident_count,
2638                         pmap->pm_stats.wired_count);
2639                 tsleep(pmap, 0, "DEAD", 0);
2640         }
2641 #else
2642         KKASSERT(pmap->pm_stats.resident_count ==
2643                  ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2644         KKASSERT(pmap->pm_stats.wired_count == 0);
2645 #endif
2646 }
2647
2648 /*
2649  * Called from low to high.  We must cache the proper parent pv so we
2650  * can adjust its wired count.
2651  */
2652 static int
2653 pmap_release_callback(pv_entry_t pv, void *data)
2654 {
2655         struct pmap_release_info *info = data;
2656         pmap_t pmap = info->pmap;
2657         vm_pindex_t pindex;
2658         int r;
2659
2660         /*
2661          * Acquire a held and locked pv, check for release race
2662          */
2663         pindex = pv->pv_pindex;
2664         if (info->pvp == pv) {
2665                 spin_unlock(&pmap->pm_spin);
2666                 info->pvp = NULL;
2667         } else if (pv_hold_try(pv)) {
2668                 spin_unlock(&pmap->pm_spin);
2669         } else {
2670                 spin_unlock(&pmap->pm_spin);
2671                 pv_lock(pv);
2672                 pv_put(pv);
2673                 info->retry = 1;
2674                 spin_lock(&pmap->pm_spin);
2675
2676                 return -1;
2677         }
2678         KKASSERT(pv->pv_pmap == pmap && pindex == pv->pv_pindex);
2679
2680         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2681                 /*
2682                  * I am PTE, parent is PT
2683                  */
2684                 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2685                 pindex += NUPTE_TOTAL;
2686         } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2687                 /*
2688                  * I am PT, parent is PD
2689                  */
2690                 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2691                 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2692         } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2693                 /*
2694                  * I am PD, parent is PDP
2695                  */
2696                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2697                          NPDPEPGSHIFT;
2698                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2699         } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2700                 /*
2701                  * I am PDP, parent is PML4 (there's only one)
2702                  */
2703 #if 0
2704                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2705                            NUPD_TOTAL) >> NPML4EPGSHIFT;
2706                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2707 #endif
2708                 pindex = pmap_pml4_pindex();
2709         } else {
2710                 /*
2711                  * parent is NULL
2712                  */
2713                 if (info->pvp) {
2714                         pv_put(info->pvp);
2715                         info->pvp = NULL;
2716                 }
2717                 pindex = 0;
2718         }
2719         if (pindex) {
2720                 if (info->pvp && info->pvp->pv_pindex != pindex) {
2721                         pv_put(info->pvp);
2722                         info->pvp = NULL;
2723                 }
2724                 if (info->pvp == NULL)
2725                         info->pvp = pv_get(pmap, pindex, NULL);
2726         } else {
2727                 if (info->pvp) {
2728                         pv_put(info->pvp);
2729                         info->pvp = NULL;
2730                 }
2731         }
2732         r = pmap_release_pv(pv, info->pvp, NULL);
2733         spin_lock(&pmap->pm_spin);
2734
2735         return(r);
2736 }
2737
2738 /*
2739  * Called with held (i.e. also locked) pv.  This function will dispose of
2740  * the lock along with the pv.
2741  *
2742  * If the caller already holds the locked parent page table for pv it
2743  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2744  * pass NULL for pvp.
2745  */
2746 static int
2747 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2748 {
2749         vm_page_t p;
2750
2751         /*
2752          * The pmap is currently not spinlocked, pv is held+locked.
2753          * Remove the pv's page from its parent's page table.  The
2754          * parent's page table page's wire_count will be decremented.
2755          *
2756          * This will clean out the pte at any level of the page table.
2757          * If smp != 0 all cpus are affected.
2758          *
2759          * Do not tear-down recursively, its faster to just let the
2760          * release run its course.
2761          */
2762         pmap_remove_pv_pte(pv, pvp, bulk, 0);
2763
2764         /*
2765          * Terminal pvs are unhooked from their vm_pages.  Because
2766          * terminal pages aren't page table pages they aren't wired
2767          * by us, so we have to be sure not to unwire them either.
2768          */
2769         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2770                 pmap_remove_pv_page(pv);
2771                 goto skip;
2772         }
2773
2774         /*
2775          * We leave the top-level page table page cached, wired, and
2776          * mapped in the pmap until the dtor function (pmap_puninit())
2777          * gets called.
2778          *
2779          * Since we are leaving the top-level pv intact we need
2780          * to break out of what would otherwise be an infinite loop.
2781          */
2782         if (pv->pv_pindex == pmap_pml4_pindex()) {
2783                 pv_put(pv);
2784                 return(-1);
2785         }
2786
2787         /*
2788          * For page table pages (other than the top-level page),
2789          * remove and free the vm_page.  The representitive mapping
2790          * removed above by pmap_remove_pv_pte() did not undo the
2791          * last wire_count so we have to do that as well.
2792          */
2793         p = pmap_remove_pv_page(pv);
2794         vm_page_busy_wait(p, FALSE, "pmaprl");
2795         if (p->wire_count != 1) {
2796                 kprintf("p->wire_count was %016lx %d\n",
2797                         pv->pv_pindex, p->wire_count);
2798         }
2799         KKASSERT(p->wire_count == 1);
2800         KKASSERT(p->flags & PG_UNMANAGED);
2801
2802         vm_page_unwire(p, 0);
2803         KKASSERT(p->wire_count == 0);
2804
2805         vm_page_free(p);
2806 skip:
2807         pv_free(pv, pvp);
2808
2809         return 0;
2810 }
2811
2812 /*
2813  * This function will remove the pte associated with a pv from its parent.
2814  * Terminal pv's are supported.  All cpus specified by (bulk) are properly
2815  * invalidated.
2816  *
2817  * The wire count will be dropped on the parent page table.  The wire
2818  * count on the page being removed (pv->pv_m) from the parent page table
2819  * is NOT touched.  Note that terminal pages will not have any additional
2820  * wire counts while page table pages will have at least one representing
2821  * the mapping, plus others representing sub-mappings.
2822  *
2823  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2824  *       pages and user page table and terminal pages.
2825  *
2826  * NOTE: The pte being removed might be unmanaged, and the pv supplied might
2827  *       be freshly allocated and not imply that the pte is managed.  In this
2828  *       case pv->pv_m should be NULL.
2829  *
2830  * The pv must be locked.  The pvp, if supplied, must be locked.  All
2831  * supplied pv's will remain locked on return.
2832  *
2833  * XXX must lock parent pv's if they exist to remove pte XXX
2834  */
2835 static
2836 void
2837 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2838                    int destroy)
2839 {
2840         vm_pindex_t ptepindex = pv->pv_pindex;
2841         pmap_t pmap = pv->pv_pmap;
2842         vm_page_t p;
2843         int gotpvp = 0;
2844
2845         KKASSERT(pmap);
2846
2847         if (ptepindex == pmap_pml4_pindex()) {
2848                 /*
2849                  * We are the top level PML4E table, there is no parent.
2850                  */
2851                 p = pmap->pm_pmlpv->pv_m;
2852                 KKASSERT(pv->pv_m == p);        /* debugging */
2853         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2854                 /*
2855                  * Remove a PDP page from the PML4E.  This can only occur
2856                  * with user page tables.  We do not have to lock the
2857                  * pml4 PV so just ignore pvp.
2858                  */
2859                 vm_pindex_t pml4_pindex;
2860                 vm_pindex_t pdp_index;
2861                 pml4_entry_t *pdp;
2862
2863                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2864                 if (pvp == NULL) {
2865                         pml4_pindex = pmap_pml4_pindex();
2866                         pvp = pv_get(pv->pv_pmap, pml4_pindex, NULL);
2867                         KKASSERT(pvp);
2868                         gotpvp = 1;
2869                 }
2870
2871                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2872                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2873                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2874                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2875                 KKASSERT(pv->pv_m == p);        /* debugging */
2876         } else if (ptepindex >= pmap_pd_pindex(0)) {
2877                 /*
2878                  * Remove a PD page from the PDP
2879                  *
2880                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2881                  *                   of a simple pmap because it stops at
2882                  *                   the PD page.
2883                  */
2884                 vm_pindex_t pdp_pindex;
2885                 vm_pindex_t pd_index;
2886                 pdp_entry_t *pd;
2887
2888                 pd_index = ptepindex - pmap_pd_pindex(0);
2889
2890                 if (pvp == NULL) {
2891                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2892                                      (pd_index >> NPML4EPGSHIFT);
2893                         pvp = pv_get(pv->pv_pmap, pdp_pindex, NULL);
2894                         gotpvp = 1;
2895                 }
2896
2897                 if (pvp) {
2898                         pd = pv_pte_lookup(pvp, pd_index &
2899                                                 ((1ul << NPDPEPGSHIFT) - 1));
2900                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2901                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2902                         pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2903                 } else {
2904                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2905                         p = pv->pv_m;           /* degenerate test later */
2906                 }
2907                 KKASSERT(pv->pv_m == p);        /* debugging */
2908         } else if (ptepindex >= pmap_pt_pindex(0)) {
2909                 /*
2910                  *  Remove a PT page from the PD
2911                  */
2912                 vm_pindex_t pd_pindex;
2913                 vm_pindex_t pt_index;
2914                 pd_entry_t *pt;
2915
2916                 pt_index = ptepindex - pmap_pt_pindex(0);
2917
2918                 if (pvp == NULL) {
2919                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2920                                     (pt_index >> NPDPEPGSHIFT);
2921                         pvp = pv_get(pv->pv_pmap, pd_pindex, NULL);
2922                         KKASSERT(pvp);
2923                         gotpvp = 1;
2924                 }
2925
2926                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2927 #if 0
2928                 KASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0,
2929                         ("*pt unexpectedly invalid %016jx "
2930                          "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
2931                         *pt, gotpvp, ptepindex, pt_index, pv, pvp));
2932                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2933 #else
2934                 if ((*pt & pmap->pmap_bits[PG_V_IDX]) == 0) {
2935                         kprintf("*pt unexpectedly invalid %016jx "
2936                                 "gotpvp=%d ptepindex=%ld ptindex=%ld "
2937                                 "pv=%p pvp=%p\n",
2938                                 *pt, gotpvp, ptepindex, pt_index, pv, pvp);
2939                         tsleep(pt, 0, "DEAD", 0);
2940                         p = pv->pv_m;
2941                 } else {
2942                         p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2943                 }
2944 #endif
2945                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2946                 KKASSERT(pv->pv_m == p);        /* debugging */
2947         } else {
2948                 /*
2949                  * Remove a PTE from the PT page.  The PV might exist even if
2950                  * the PTE is not managed, in whichcase pv->pv_m should be
2951                  * NULL.
2952                  *
2953                  * NOTE: Userland pmaps manage the parent PT/PD/PDP page
2954                  *       table pages but the kernel_pmap does not.
2955                  *
2956                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2957                  *       pv is a pte_pv so we can safely lock pt_pv.
2958                  *
2959                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2960                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2961                  *       terminal ptes.
2962                  */
2963                 vm_pindex_t pt_pindex;
2964                 pt_entry_t *ptep;
2965                 pt_entry_t pte;
2966                 vm_offset_t va;
2967
2968                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2969                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2970
2971                 if (ptepindex >= NUPTE_USER) {
2972                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2973                         KKASSERT(pvp == NULL);
2974                         /* pvp remains NULL */
2975                 } else {
2976                         if (pvp == NULL) {
2977                                 pt_pindex = NUPTE_TOTAL +
2978                                             (ptepindex >> NPDPEPGSHIFT);
2979                                 pvp = pv_get(pv->pv_pmap, pt_pindex, NULL);
2980                                 KKASSERT(pvp);
2981                                 gotpvp = 1;
2982                         }
2983                         ptep = pv_pte_lookup(pvp, ptepindex &
2984                                                   ((1ul << NPDPEPGSHIFT) - 1));
2985                 }
2986                 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2987                 if (bulk == NULL)               /* XXX */
2988                         cpu_invlpg((void *)va); /* XXX */
2989
2990                 /*
2991                  * Now update the vm_page_t
2992                  */
2993                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
2994                     (pte & pmap->pmap_bits[PG_V_IDX])) {
2995                         /*
2996                          * Valid managed page, adjust (p).
2997                          */
2998                         if (pte & pmap->pmap_bits[PG_DEVICE_IDX]) {
2999                                 p = pv->pv_m;
3000                         } else {
3001                                 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
3002                                 KKASSERT(pv->pv_m == p);
3003                         }
3004                         if (pte & pmap->pmap_bits[PG_M_IDX]) {
3005                                 if (pmap_track_modified(ptepindex))
3006                                         vm_page_dirty(p);
3007                         }
3008                         if (pte & pmap->pmap_bits[PG_A_IDX]) {
3009                                 vm_page_flag_set(p, PG_REFERENCED);
3010                         }
3011                 } else {
3012                         /*
3013                          * Unmanaged page, do not try to adjust the vm_page_t.
3014                          * pv could be freshly allocated for a pmap_enter(),
3015                          * replacing an unmanaged page with a managed one.
3016                          *
3017                          * pv->pv_m might reflect the new page and not the
3018                          * existing page.
3019                          *
3020                          * We could extract p from the physical address and
3021                          * adjust it but we explicitly do not for unmanaged
3022                          * pages.
3023                          */
3024                         p = NULL;
3025                 }
3026                 if (pte & pmap->pmap_bits[PG_W_IDX])
3027                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
3028                 if (pte & pmap->pmap_bits[PG_G_IDX])
3029                         cpu_invlpg((void *)va);
3030         }
3031
3032         /*
3033          * If requested, scrap the underlying pv->pv_m and the underlying
3034          * pv.  If this is a page-table-page we must also free the page.
3035          *
3036          * pvp must be returned locked.
3037          */
3038         if (destroy == 1) {
3039                 /*
3040                  * page table page (PT, PD, PDP, PML4), caller was responsible
3041                  * for testing wired_count.
3042                  */
3043                 KKASSERT(pv->pv_m->wire_count == 1);
3044                 p = pmap_remove_pv_page(pv);
3045                 pv_free(pv, pvp);
3046                 pv = NULL;
3047
3048                 vm_page_busy_wait(p, FALSE, "pgpun");
3049                 vm_page_unwire(p, 0);
3050                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
3051                 vm_page_free(p);
3052         } else if (destroy == 2) {
3053                 /*
3054                  * Normal page, remove from pmap and leave the underlying
3055                  * page untouched.
3056                  */
3057                 pmap_remove_pv_page(pv);
3058                 pv_free(pv, pvp);
3059                 pv = NULL;              /* safety */
3060         }
3061
3062         /*
3063          * If we acquired pvp ourselves then we are responsible for
3064          * recursively deleting it.
3065          */
3066         if (pvp && gotpvp) {
3067                 /*
3068                  * Recursively destroy higher-level page tables.
3069                  *
3070                  * This is optional.  If we do not, they will still
3071                  * be destroyed when the process exits.
3072                  *
3073                  * NOTE: Do not destroy pv_entry's with extra hold refs,
3074                  *       a caller may have unlocked it and intends to
3075                  *       continue to use it.
3076                  */
3077                 if (pmap_dynamic_delete &&
3078                     pvp->pv_m &&
3079                     pvp->pv_m->wire_count == 1 &&
3080                     (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
3081                     pvp->pv_pindex != pmap_pml4_pindex()) {
3082                         if (pmap_dynamic_delete == 2)
3083                                 kprintf("A %jd %08x\n", pvp->pv_pindex, pvp->pv_hold);
3084                         if (pmap != &kernel_pmap) {
3085                                 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
3086                                 pvp = NULL;     /* safety */
3087                         } else {
3088                                 kprintf("Attempt to remove kernel_pmap pindex "
3089                                         "%jd\n", pvp->pv_pindex);
3090                                 pv_put(pvp);
3091                         }
3092                 } else {
3093                         pv_put(pvp);
3094                 }
3095         }
3096 }
3097
3098 /*
3099  * Remove the vm_page association to a pv.  The pv must be locked.
3100  */
3101 static
3102 vm_page_t
3103 pmap_remove_pv_page(pv_entry_t pv)
3104 {
3105         vm_page_t m;
3106
3107         m = pv->pv_m;
3108         vm_page_spin_lock(m);
3109         KKASSERT(m && m == pv->pv_m);
3110         pv->pv_m = NULL;
3111         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3112         pmap_page_stats_deleting(m);
3113         if (TAILQ_EMPTY(&m->md.pv_list))
3114                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3115         vm_page_spin_unlock(m);
3116
3117         return(m);
3118 }
3119
3120 /*
3121  * Grow the number of kernel page table entries, if needed.
3122  *
3123  * This routine is always called to validate any address space
3124  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
3125  * space below KERNBASE.
3126  *
3127  * kernel_map must be locked exclusively by the caller.
3128  */
3129 void
3130 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
3131 {
3132         vm_paddr_t paddr;
3133         vm_offset_t ptppaddr;
3134         vm_page_t nkpg;
3135         pd_entry_t *pt, newpt;
3136         pdp_entry_t newpd;
3137         int update_kernel_vm_end;
3138
3139         /*
3140          * bootstrap kernel_vm_end on first real VM use
3141          */
3142         if (kernel_vm_end == 0) {
3143                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
3144                 nkpt = 0;
3145                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3146                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
3147                                         ~(PAGE_SIZE * NPTEPG - 1);
3148                         nkpt++;
3149                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
3150                                 kernel_vm_end = kernel_map.max_offset;
3151                                 break;
3152                         }
3153                 }
3154         }
3155
3156         /*
3157          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
3158          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
3159          * do not want to force-fill 128G worth of page tables.
3160          */
3161         if (kstart < KERNBASE) {
3162                 if (kstart > kernel_vm_end)
3163                         kstart = kernel_vm_end;
3164                 KKASSERT(kend <= KERNBASE);
3165                 update_kernel_vm_end = 1;
3166         } else {
3167                 update_kernel_vm_end = 0;
3168         }
3169
3170         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
3171         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
3172
3173         if (kend - 1 >= kernel_map.max_offset)
3174                 kend = kernel_map.max_offset;
3175
3176         while (kstart < kend) {
3177                 pt = pmap_pt(&kernel_pmap, kstart);
3178                 if (pt == NULL) {
3179                         /* We need a new PD entry */
3180                         nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3181                                              VM_ALLOC_NORMAL |
3182                                              VM_ALLOC_SYSTEM |
3183                                              VM_ALLOC_INTERRUPT);
3184                         if (nkpg == NULL) {
3185                                 panic("pmap_growkernel: no memory to grow "
3186                                       "kernel");
3187                         }
3188                         paddr = VM_PAGE_TO_PHYS(nkpg);
3189                         pmap_zero_page(paddr);
3190                         newpd = (pdp_entry_t)
3191                             (paddr |
3192                             kernel_pmap.pmap_bits[PG_V_IDX] |
3193                             kernel_pmap.pmap_bits[PG_RW_IDX] |
3194                             kernel_pmap.pmap_bits[PG_A_IDX] |
3195                             kernel_pmap.pmap_bits[PG_M_IDX]);
3196                         *pmap_pd(&kernel_pmap, kstart) = newpd;
3197                         continue; /* try again */
3198                 }
3199                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3200                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
3201                                  ~(PAGE_SIZE * NPTEPG - 1);
3202                         if (kstart - 1 >= kernel_map.max_offset) {
3203                                 kstart = kernel_map.max_offset;
3204                                 break;
3205                         }
3206                         continue;
3207                 }
3208
3209                 /*
3210                  * We need a new PT
3211                  *
3212                  * This index is bogus, but out of the way
3213                  */
3214                 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3215                                      VM_ALLOC_NORMAL |
3216                                      VM_ALLOC_SYSTEM |
3217                                      VM_ALLOC_INTERRUPT);
3218                 if (nkpg == NULL)
3219                         panic("pmap_growkernel: no memory to grow kernel");
3220
3221                 vm_page_wire(nkpg);
3222                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
3223                 pmap_zero_page(ptppaddr);
3224                 newpt = (pd_entry_t)(ptppaddr |
3225                                      kernel_pmap.pmap_bits[PG_V_IDX] |
3226                                      kernel_pmap.pmap_bits[PG_RW_IDX] |
3227                                      kernel_pmap.pmap_bits[PG_A_IDX] |
3228                                      kernel_pmap.pmap_bits[PG_M_IDX]);
3229                 atomic_swap_long(pmap_pt(&kernel_pmap, kstart), newpt);
3230
3231                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3232                           ~(PAGE_SIZE * NPTEPG - 1);
3233
3234                 if (kstart - 1 >= kernel_map.max_offset) {
3235                         kstart = kernel_map.max_offset;
3236                         break;
3237                 }
3238         }
3239
3240         /*
3241          * Only update kernel_vm_end for areas below KERNBASE.
3242          */
3243         if (update_kernel_vm_end && kernel_vm_end < kstart)
3244                 kernel_vm_end = kstart;
3245 }
3246
3247 /*
3248  *      Add a reference to the specified pmap.
3249  */
3250 void
3251 pmap_reference(pmap_t pmap)
3252 {
3253         if (pmap != NULL)
3254                 atomic_add_int(&pmap->pm_count, 1);
3255 }
3256
3257 /***************************************************
3258  * page management routines.
3259  ***************************************************/
3260
3261 /*
3262  * Hold a pv without locking it
3263  */
3264 static void
3265 pv_hold(pv_entry_t pv)
3266 {
3267         atomic_add_int(&pv->pv_hold, 1);
3268 }
3269
3270 /*
3271  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
3272  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
3273  * the pv properly.
3274  *
3275  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3276  * pv list via its page) must be held by the caller in order to stabilize
3277  * the pv.
3278  */
3279 static int
3280 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3281 {
3282         u_int count;
3283
3284         /*
3285          * Critical path shortcut expects pv to already have one ref
3286          * (for the pv->pv_pmap).
3287          */
3288         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3289 #ifdef PMAP_DEBUG
3290                 pv->pv_func = func;
3291                 pv->pv_line = lineno;
3292 #endif
3293                 return TRUE;
3294         }
3295
3296         for (;;) {
3297                 count = pv->pv_hold;
3298                 cpu_ccfence();
3299                 if ((count & PV_HOLD_LOCKED) == 0) {
3300                         if (atomic_cmpset_int(&pv->pv_hold, count,
3301                                               (count + 1) | PV_HOLD_LOCKED)) {
3302 #ifdef PMAP_DEBUG
3303                                 pv->pv_func = func;
3304                                 pv->pv_line = lineno;
3305 #endif
3306                                 return TRUE;
3307                         }
3308                 } else {
3309                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3310                                 return FALSE;
3311                 }
3312                 /* retry */
3313         }
3314 }
3315
3316 /*
3317  * Drop a previously held pv_entry which could not be locked, allowing its
3318  * destruction.
3319  *
3320  * Must not be called with a spinlock held as we might zfree() the pv if it
3321  * is no longer associated with a pmap and this was the last hold count.
3322  */
3323 static void
3324 pv_drop(pv_entry_t pv)
3325 {
3326         u_int count;
3327
3328         for (;;) {
3329                 count = pv->pv_hold;
3330                 cpu_ccfence();
3331                 KKASSERT((count & PV_HOLD_MASK) > 0);
3332                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3333                          (PV_HOLD_LOCKED | 1));
3334                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3335                         if ((count & PV_HOLD_MASK) == 1) {
3336 #ifdef PMAP_DEBUG2
3337                                 if (pmap_enter_debug > 0) {
3338                                         --pmap_enter_debug;
3339                                         kprintf("pv_drop: free pv %p\n", pv);
3340                                 }
3341 #endif
3342                                 KKASSERT(count == 1);
3343                                 KKASSERT(pv->pv_pmap == NULL);
3344                                 zfree(pvzone, pv);
3345                         }
3346                         return;
3347                 }
3348                 /* retry */
3349         }
3350 }
3351
3352 /*
3353  * Find or allocate the requested PV entry, returning a locked, held pv.
3354  *
3355  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3356  * for the caller and one representing the pmap and vm_page association.
3357  *
3358  * If (*isnew) is zero, the returned pv will have only one hold count.
3359  *
3360  * Since both associations can only be adjusted while the pv is locked,
3361  * together they represent just one additional hold.
3362  */
3363 static
3364 pv_entry_t
3365 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3366 {
3367         struct mdglobaldata *md = mdcpu;
3368         pv_entry_t pv;
3369         pv_entry_t pnew;
3370         int pmap_excl = 0;
3371
3372         pnew = NULL;
3373         if (md->gd_newpv) {
3374 #if 1
3375                 pnew = atomic_swap_ptr((void *)&md->gd_newpv, NULL);
3376 #else
3377                 crit_enter();
3378                 pnew = md->gd_newpv;    /* might race NULL */
3379                 md->gd_newpv = NULL;
3380                 crit_exit();
3381 #endif
3382         }
3383         if (pnew == NULL)
3384                 pnew = zalloc(pvzone);
3385
3386         spin_lock_shared(&pmap->pm_spin);
3387         for (;;) {
3388                 /*
3389                  * Shortcut cache
3390                  */
3391                 pv = pv_entry_lookup(pmap, pindex);
3392                 if (pv == NULL) {
3393                         vm_pindex_t *pmark;
3394
3395                         /*
3396                          * Requires exclusive pmap spinlock
3397                          */
3398                         if (pmap_excl == 0) {
3399                                 pmap_excl = 1;
3400                                 if (!spin_lock_upgrade_try(&pmap->pm_spin)) {
3401                                         spin_unlock_shared(&pmap->pm_spin);
3402                                         spin_lock(&pmap->pm_spin);
3403                                         continue;
3404                                 }
3405                         }
3406
3407                         /*
3408                          * We need to block if someone is holding our
3409                          * placemarker.  As long as we determine the
3410                          * placemarker has not been aquired we do not
3411                          * need to get it as acquision also requires
3412                          * the pmap spin lock.
3413                          *
3414                          * However, we can race the wakeup.
3415                          */
3416                         pmark = pmap_placemarker_hash(pmap, pindex);
3417
3418                         if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3419                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3420                                 tsleep_interlock(pmark, 0);
3421                                 if (((*pmark ^ pindex) &
3422                                      ~PM_PLACEMARK_WAKEUP) == 0) {
3423                                         spin_unlock(&pmap->pm_spin);
3424                                         tsleep(pmark, PINTERLOCKED, "pvplc", 0);
3425                                         spin_lock(&pmap->pm_spin);
3426                                 }
3427                                 continue;
3428                         }
3429
3430                         /*
3431                          * Setup the new entry
3432                          */
3433                         pnew->pv_pmap = pmap;
3434                         pnew->pv_pindex = pindex;
3435                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
3436 #ifdef PMAP_DEBUG
3437                         pnew->pv_func = func;
3438                         pnew->pv_line = lineno;
3439                         if (pnew->pv_line_lastfree > 0) {
3440                                 pnew->pv_line_lastfree =
3441                                                 -pnew->pv_line_lastfree;
3442                         }
3443 #endif
3444                         pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3445                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
3446                         spin_unlock(&pmap->pm_spin);
3447                         *isnew = 1;
3448
3449                         KKASSERT(pv == NULL);
3450                         return(pnew);
3451                 }
3452
3453                 /*
3454                  * We already have an entry, cleanup the staged pnew if
3455                  * we can get the lock, otherwise block and retry.
3456                  */
3457                 if (__predict_true(_pv_hold_try(pv PMAP_DEBUG_COPY))) {
3458                         if (pmap_excl)
3459                                 spin_unlock(&pmap->pm_spin);
3460                         else
3461                                 spin_unlock_shared(&pmap->pm_spin);
3462 #if 1
3463                         pnew = atomic_swap_ptr((void *)&md->gd_newpv, pnew);
3464                         if (pnew)
3465                                 zfree(pvzone, pnew);
3466 #else
3467                         crit_enter();
3468                         if (md->gd_newpv == NULL)
3469                                 md->gd_newpv = pnew;
3470                         else
3471                                 zfree(pvzone, pnew);
3472                         crit_exit();
3473 #endif
3474                         KKASSERT(pv->pv_pmap == pmap &&
3475                                  pv->pv_pindex == pindex);
3476                         *isnew = 0;
3477                         return(pv);
3478                 }
3479                 if (pmap_excl) {
3480                         spin_unlock(&pmap->pm_spin);
3481                         _pv_lock(pv PMAP_DEBUG_COPY);
3482                         pv_put(pv);
3483                         spin_lock(&pmap->pm_spin);
3484                 } else {
3485                         spin_unlock_shared(&pmap->pm_spin);
3486                         _pv_lock(pv PMAP_DEBUG_COPY);
3487                         pv_put(pv);
3488                         spin_lock_shared(&pmap->pm_spin);
3489                 }
3490         }
3491         /* NOT REACHED */
3492 }
3493
3494 /*
3495  * Find the requested PV entry, returning a locked+held pv or NULL
3496  */
3497 static
3498 pv_entry_t
3499 _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp PMAP_DEBUG_DECL)
3500 {
3501         pv_entry_t pv;
3502         int pmap_excl = 0;
3503
3504         spin_lock_shared(&pmap->pm_spin);
3505         for (;;) {
3506                 /*
3507                  * Shortcut cache
3508                  */
3509                 pv = pv_entry_lookup(pmap, pindex);
3510                 if (pv == NULL) {
3511                         /*
3512                          * Block if there is ANY placemarker.  If we are to
3513                          * return it, we must also aquire the spot, so we
3514                          * have to block even if the placemarker is held on
3515                          * a different address.
3516                          *
3517                          * OPTIMIZATION: If pmarkp is passed as NULL the
3518                          * caller is just probing (or looking for a real
3519                          * pv_entry), and in this case we only need to check
3520                          * to see if the placemarker matches pindex.
3521                          */
3522                         vm_pindex_t *pmark;
3523
3524                         /*
3525                          * Requires exclusive pmap spinlock
3526                          */
3527                         if (pmap_excl == 0) {
3528                                 pmap_excl = 1;
3529                                 if (!spin_lock_upgrade_try(&pmap->pm_spin)) {
3530                                         spin_unlock_shared(&pmap->pm_spin);
3531                                         spin_lock(&pmap->pm_spin);
3532                                         continue;
3533                                 }
3534                         }
3535
3536                         pmark = pmap_placemarker_hash(pmap, pindex);
3537
3538                         if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3539                             ((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3540                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3541                                 tsleep_interlock(pmark, 0);
3542                                 if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3543                                     ((*pmark ^ pindex) &
3544                                      ~PM_PLACEMARK_WAKEUP) == 0) {
3545                                         spin_unlock(&pmap->pm_spin);
3546                                         tsleep(pmark, PINTERLOCKED, "pvpld", 0);
3547                                         spin_lock(&pmap->pm_spin);
3548                                 }
3549                                 continue;
3550                         }
3551                         if (pmarkp) {
3552                                 if (atomic_swap_long(pmark, pindex) !=
3553                                     PM_NOPLACEMARK) {
3554                                         panic("_pv_get: pmark race");
3555                                 }
3556                                 *pmarkp = pmark;
3557                         }
3558                         spin_unlock(&pmap->pm_spin);
3559                         return NULL;
3560                 }
3561                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3562                         pv_cache(pv, pindex);
3563                         if (pmap_excl)
3564                                 spin_unlock(&pmap->pm_spin);
3565                         else
3566                                 spin_unlock_shared(&pmap->pm_spin);
3567                         KKASSERT(pv->pv_pmap == pmap &&
3568                                  pv->pv_pindex == pindex);
3569                         return(pv);
3570                 }
3571                 if (pmap_excl) {
3572                         spin_unlock(&pmap->pm_spin);
3573                         _pv_lock(pv PMAP_DEBUG_COPY);
3574                         pv_put(pv);
3575                         spin_lock(&pmap->pm_spin);
3576                 } else {
3577                         spin_unlock_shared(&pmap->pm_spin);
3578                         _pv_lock(pv PMAP_DEBUG_COPY);
3579                         pv_put(pv);
3580                         spin_lock_shared(&pmap->pm_spin);
3581                 }
3582         }
3583 }
3584
3585 /*
3586  * Lookup, hold, and attempt to lock (pmap,pindex).
3587  *
3588  * If the entry does not exist NULL is returned and *errorp is set to 0
3589  *
3590  * If the entry exists and could be successfully locked it is returned and
3591  * errorp is set to 0.
3592  *
3593  * If the entry exists but could NOT be successfully locked it is returned
3594  * held and *errorp is set to 1.
3595  *
3596  * If the entry is placemarked by someone else NULL is returned and *errorp
3597  * is set to 1.
3598  */
3599 static
3600 pv_entry_t
3601 pv_get_try(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp, int *errorp)
3602 {
3603         pv_entry_t pv;
3604
3605         spin_lock_shared(&pmap->pm_spin);
3606
3607         pv = pv_entry_lookup(pmap, pindex);
3608         if (pv == NULL) {
3609                 vm_pindex_t *pmark;
3610
3611                 pmark = pmap_placemarker_hash(pmap, pindex);
3612
3613                 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3614                         *errorp = 1;
3615                 } else if (pmarkp &&
3616                            atomic_cmpset_long(pmark, PM_NOPLACEMARK, pindex)) {
3617                         *errorp = 0;
3618                 } else {
3619                         /*
3620                          * Can't set a placemark with a NULL pmarkp, or if
3621                          * pmarkp is non-NULL but we failed to set our
3622                          * placemark.
3623                          */
3624                         *errorp = 1;
3625                 }
3626                 if (pmarkp)
3627                         *pmarkp = pmark;
3628                 spin_unlock_shared(&pmap->pm_spin);
3629
3630                 return NULL;
3631         }
3632
3633         /*
3634          * XXX This has problems if the lock is shared, why?
3635          */
3636         if (pv_hold_try(pv)) {
3637                 pv_cache(pv, pindex);   /* overwrite ok (shared lock) */
3638                 spin_unlock_shared(&pmap->pm_spin);
3639                 *errorp = 0;
3640                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3641                 return(pv);     /* lock succeeded */
3642         }
3643         spin_unlock_shared(&pmap->pm_spin);
3644         *errorp = 1;
3645
3646         return (pv);            /* lock failed */
3647 }
3648
3649 /*
3650  * Lock a held pv, keeping the hold count
3651  */
3652 static
3653 void
3654 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3655 {
3656         u_int count;
3657
3658         for (;;) {
3659                 count = pv->pv_hold;
3660                 cpu_ccfence();
3661                 if ((count & PV_HOLD_LOCKED) == 0) {
3662                         if (atomic_cmpset_int(&pv->pv_hold, count,
3663                                               count | PV_HOLD_LOCKED)) {
3664 #ifdef PMAP_DEBUG
3665                                 pv->pv_func = func;
3666                                 pv->pv_line = lineno;
3667 #endif
3668                                 return;
3669                         }
3670                         continue;
3671                 }
3672                 tsleep_interlock(pv, 0);
3673                 if (atomic_cmpset_int(&pv->pv_hold, count,
3674                                       count | PV_HOLD_WAITING)) {
3675 #ifdef PMAP_DEBUG2
3676                         if (pmap_enter_debug > 0) {
3677                                 --pmap_enter_debug;
3678                                 kprintf("pv waiting on %s:%d\n",
3679                                         pv->pv_func, pv->pv_line);
3680                         }
3681 #endif
3682                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
3683                 }
3684                 /* retry */
3685         }
3686 }
3687
3688 /*
3689  * Unlock a held and locked pv, keeping the hold count.
3690  */
3691 static
3692 void
3693 pv_unlock(pv_entry_t pv)
3694 {
3695         u_int count;
3696
3697         for (;;) {
3698                 count = pv->pv_hold;
3699                 cpu_ccfence();
3700                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3701                          (PV_HOLD_LOCKED | 1));
3702                 if (atomic_cmpset_int(&pv->pv_hold, count,
3703                                       count &
3704                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3705                         if (count & PV_HOLD_WAITING)
3706                                 wakeup(pv);
3707                         break;
3708                 }
3709         }
3710 }
3711
3712 /*
3713  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3714  * and the hold count drops to zero we will free it.
3715  *
3716  * Caller should not hold any spin locks.  We are protected from hold races
3717  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3718  * lock held.  A pv cannot be located otherwise.
3719  */
3720 static
3721 void
3722 pv_put(pv_entry_t pv)
3723 {
3724 #ifdef PMAP_DEBUG2
3725         if (pmap_enter_debug > 0) {
3726                 --pmap_enter_debug;
3727                 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3728         }
3729 #endif
3730
3731         /*
3732          * Normal put-aways must have a pv_m associated with the pv,
3733          * but allow the case where the pv has been destructed due
3734          * to pmap_dynamic_delete.
3735          */
3736         KKASSERT(pv->pv_pmap == NULL || pv->pv_m != NULL);
3737
3738         /*
3739          * Fast - shortcut most common condition
3740          */
3741         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3742                 return;
3743
3744         /*
3745          * Slow
3746          */
3747         pv_unlock(pv);
3748         pv_drop(pv);
3749 }
3750
3751 /*
3752  * Remove the pmap association from a pv, require that pv_m already be removed,
3753  * then unlock and drop the pv.  Any pte operations must have already been
3754  * completed.  This call may result in a last-drop which will physically free
3755  * the pv.
3756  *
3757  * Removing the pmap association entails an additional drop.
3758  *
3759  * pv must be exclusively locked on call and will be disposed of on return.
3760  */
3761 static
3762 void
3763 _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL)
3764 {
3765         pmap_t pmap;
3766
3767 #ifdef PMAP_DEBUG
3768         pv->pv_func_lastfree = func;
3769         pv->pv_line_lastfree = lineno;
3770 #endif
3771         KKASSERT(pv->pv_m == NULL);
3772         KKASSERT((pv->pv_hold & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
3773                   (PV_HOLD_LOCKED|1));
3774         if ((pmap = pv->pv_pmap) != NULL) {
3775                 spin_lock(&pmap->pm_spin);
3776                 KKASSERT(pv->pv_pmap == pmap);
3777                 if (pmap->pm_pvhint_pt == pv)
3778                         pmap->pm_pvhint_pt = NULL;
3779                 if (pmap->pm_pvhint_pte == pv)
3780                         pmap->pm_pvhint_pte = NULL;
3781                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3782                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3783                 pv->pv_pmap = NULL;
3784                 pv->pv_pindex = 0;
3785                 spin_unlock(&pmap->pm_spin);
3786
3787                 /*
3788                  * Try to shortcut three atomic ops, otherwise fall through
3789                  * and do it normally.  Drop two refs and the lock all in
3790                  * one go.
3791                  */
3792                 if (pvp)
3793                         vm_page_unwire_quick(pvp->pv_m);
3794                 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3795 #ifdef PMAP_DEBUG2
3796                         if (pmap_enter_debug > 0) {
3797                                 --pmap_enter_debug;
3798                                 kprintf("pv_free: free pv %p\n", pv);
3799                         }
3800 #endif
3801                         zfree(pvzone, pv);
3802                         return;
3803                 }
3804                 pv_drop(pv);    /* ref for pv_pmap */
3805         }
3806         pv_unlock(pv);
3807         pv_drop(pv);
3808 }
3809
3810 /*
3811  * This routine is very drastic, but can save the system
3812  * in a pinch.
3813  */
3814 void
3815 pmap_collect(void)
3816 {
3817         int i;
3818         vm_page_t m;
3819         static int warningdone=0;
3820
3821         if (pmap_pagedaemon_waken == 0)
3822                 return;
3823         pmap_pagedaemon_waken = 0;
3824         if (warningdone < 5) {
3825                 kprintf("pmap_collect: collecting pv entries -- "
3826                         "suggest increasing PMAP_SHPGPERPROC\n");
3827                 warningdone++;
3828         }
3829
3830         for (i = 0; i < vm_page_array_size; i++) {
3831                 m = &vm_page_array[i];
3832                 if (m->wire_count || m->hold_count)
3833                         continue;
3834                 if (vm_page_busy_try(m, TRUE) == 0) {
3835                         if (m->wire_count == 0 && m->hold_count == 0) {
3836                                 pmap_remove_all(m);
3837                         }
3838                         vm_page_wakeup(m);
3839                 }
3840         }
3841 }
3842
3843 /*
3844  * Scan the pmap for active page table entries and issue a callback.
3845  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3846  * its parent page table.
3847  *
3848  * pte_pv will be NULL if the page or page table is unmanaged.
3849  * pt_pv will point to the page table page containing the pte for the page.
3850  *
3851  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3852  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3853  *       process pmap's PD and page to the callback function.  This can be
3854  *       confusing because the pt_pv is really a pd_pv, and the target page
3855  *       table page is simply aliased by the pmap and not owned by it.
3856  *
3857  * It is assumed that the start and end are properly rounded to the page size.
3858  *
3859  * It is assumed that PD pages and above are managed and thus in the RB tree,
3860  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3861  */
3862 struct pmap_scan_info {
3863         struct pmap *pmap;
3864         vm_offset_t sva;
3865         vm_offset_t eva;
3866         vm_pindex_t sva_pd_pindex;
3867         vm_pindex_t eva_pd_pindex;
3868         void (*func)(pmap_t, struct pmap_scan_info *,
3869                      pv_entry_t, vm_pindex_t *, pv_entry_t,
3870                      int, vm_offset_t,
3871                      pt_entry_t *, void *);
3872         void *arg;
3873         pmap_inval_bulk_t bulk_core;
3874         pmap_inval_bulk_t *bulk;
3875         int count;
3876         int stop;
3877 };
3878
3879 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3880 static int pmap_scan_callback(pv_entry_t pv, void *data);
3881
3882 static void
3883 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3884 {
3885         struct pmap *pmap = info->pmap;
3886         pv_entry_t pd_pv;       /* A page directory PV */
3887         pv_entry_t pt_pv;       /* A page table PV */
3888         pv_entry_t pte_pv;      /* A page table entry PV */
3889         vm_pindex_t *pte_placemark;
3890         vm_pindex_t *pt_placemark;
3891         pt_entry_t *ptep;
3892         pt_entry_t oldpte;
3893         struct pv_entry dummy_pv;
3894
3895         info->stop = 0;
3896         if (pmap == NULL)
3897                 return;
3898         if (info->sva == info->eva)
3899                 return;
3900         if (smp_inval) {
3901                 info->bulk = &info->bulk_core;
3902                 pmap_inval_bulk_init(&info->bulk_core, pmap);
3903         } else {
3904                 info->bulk = NULL;
3905         }
3906
3907         /*
3908          * Hold the token for stability; if the pmap is empty we have nothing
3909          * to do.
3910          */
3911 #if 0
3912         if (pmap->pm_stats.resident_count == 0) {
3913                 return;
3914         }
3915 #endif
3916
3917         info->count = 0;
3918
3919         /*
3920          * Special handling for scanning one page, which is a very common
3921          * operation (it is?).
3922          *
3923          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3924          */
3925         if (info->sva + PAGE_SIZE == info->eva) {
3926                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3927                         /*
3928                          * Kernel mappings do not track wire counts on
3929                          * page table pages and only maintain pd_pv and
3930                          * pte_pv levels so pmap_scan() works.
3931                          */
3932                         pt_pv = NULL;
3933                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3934                                         &pte_placemark);
3935                         ptep = vtopte(info->sva);
3936                 } else {
3937                         /*
3938                          * User pages which are unmanaged will not have a
3939                          * pte_pv.  User page table pages which are unmanaged
3940                          * (shared from elsewhere) will also not have a pt_pv.
3941                          * The func() callback will pass both pte_pv and pt_pv
3942                          * as NULL in that case.
3943                          *
3944                          * We hold pte_placemark across the operation for
3945                          * unmanaged pages.
3946                          *
3947                          * WARNING!  We must hold pt_placemark across the
3948                          *           *ptep test to prevent misintepreting
3949                          *           a non-zero *ptep as a shared page
3950                          *           table page.  Hold it across the function
3951                          *           callback as well for SMP safety.
3952                          */
3953                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3954                                         &pte_placemark);
3955                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva),
3956                                         &pt_placemark);
3957                         if (pt_pv == NULL) {
3958                                 KKASSERT(pte_pv == NULL);
3959                                 pd_pv = pv_get(pmap,
3960                                                pmap_pd_pindex(info->sva),
3961                                                NULL);
3962                                 if (pd_pv) {
3963                                         ptep = pv_pte_lookup(pd_pv,
3964                                                     pmap_pt_index(info->sva));
3965                                         if (*ptep) {
3966                                                 info->func(pmap, info,
3967                                                      NULL, pt_placemark,
3968                                                      pd_pv, 1,
3969                                                      info->sva, ptep,
3970                                                      info->arg);
3971                                         } else {
3972                                                 pv_placemarker_wakeup(pmap,
3973                                                                   pt_placemark);
3974                                         }
3975                                         pv_put(pd_pv);
3976                                 } else {
3977                                         pv_placemarker_wakeup(pmap,
3978                                                               pt_placemark);
3979                                 }
3980                                 pv_placemarker_wakeup(pmap, pte_placemark);
3981                                 goto fast_skip;
3982                         }
3983                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3984                 }
3985
3986                 /*
3987                  * NOTE: *ptep can't be ripped out from under us if we hold
3988                  *       pte_pv (or pte_placemark) locked, but bits can
3989                  *       change.
3990                  */
3991                 oldpte = *ptep;
3992                 cpu_ccfence();
3993                 if (oldpte == 0) {
3994                         KKASSERT(pte_pv == NULL);
3995                         pv_placemarker_wakeup(pmap, pte_placemark);
3996                 } else if (pte_pv) {
3997                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3998                                            pmap->pmap_bits[PG_V_IDX])) ==
3999                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
4000                                  pmap->pmap_bits[PG_V_IDX]),
4001                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p",
4002                             *ptep, oldpte, info->sva, pte_pv));
4003                         info->func(pmap, info, pte_pv, NULL, pt_pv, 0,
4004                                    info->sva, ptep, info->arg);
4005                 } else {
4006                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
4007                                            pmap->pmap_bits[PG_V_IDX])) ==
4008                             pmap->pmap_bits[PG_V_IDX],
4009                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
4010                             *ptep, oldpte, info->sva));
4011                         info->func(pmap, info, NULL, pte_placemark, pt_pv, 0,
4012                                    info->sva, ptep, info->arg);
4013                 }
4014                 if (pt_pv)
4015                         pv_put(pt_pv);
4016 fast_skip:
4017                 pmap_inval_bulk_flush(info->bulk);
4018                 return;
4019         }
4020
4021         /*
4022          * Nominal scan case, RB_SCAN() for PD pages and iterate from
4023          * there.
4024          *
4025          * WARNING! eva can overflow our standard ((N + mask) >> bits)
4026          *          bounds, resulting in a pd_pindex of 0.  To solve the
4027          *          problem we use an inclusive range.
4028          */
4029         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
4030         info->eva_pd_pindex = pmap_pd_pindex(info->eva - PAGE_SIZE);
4031
4032         if (info->sva >= VM_MAX_USER_ADDRESS) {
4033                 /*
4034                  * The kernel does not currently maintain any pv_entry's for
4035                  * higher-level page tables.
4036                  */
4037                 bzero(&dummy_pv, sizeof(dummy_pv));
4038                 dummy_pv.pv_pindex = info->sva_pd_pindex;
4039                 spin_lock(&pmap->pm_spin);
4040                 while (dummy_pv.pv_pindex <= info->eva_pd_pindex) {
4041                         pmap_scan_callback(&dummy_pv, info);
4042                         ++dummy_pv.pv_pindex;
4043                         if (dummy_pv.pv_pindex < info->sva_pd_pindex) /*wrap*/
4044                                 break;
4045                 }
4046                 spin_unlock(&pmap->pm_spin);
4047         } else {
4048                 /*
4049                  * User page tables maintain local PML4, PDP, and PD
4050                  * pv_entry's at the very least.  PT pv's might be
4051                  * unmanaged and thus not exist.  PTE pv's might be
4052                  * unmanaged and thus not exist.
4053                  */
4054                 spin_lock(&pmap->pm_spin);
4055                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot, pmap_scan_cmp,
4056                                          pmap_scan_callback, info);
4057                 spin_unlock(&pmap->pm_spin);
4058         }
4059         pmap_inval_bulk_flush(info->bulk);
4060 }
4061
4062 /*
4063  * WARNING! pmap->pm_spin held
4064  *
4065  * WARNING! eva can overflow our standard ((N + mask) >> bits)
4066  *          bounds, resulting in a pd_pindex of 0.  To solve the
4067  *          problem we use an inclusive range.
4068  */
4069 static int
4070 pmap_scan_cmp(pv_entry_t pv, void *data)
4071 {
4072         struct pmap_scan_info *info = data;
4073         if (pv->pv_pindex < info->sva_pd_pindex)
4074                 return(-1);
4075         if (pv->pv_pindex > info->eva_pd_pindex)
4076                 return(1);
4077         return(0);
4078 }
4079
4080 /*
4081  * pmap_scan() by PDs
4082  *
4083  * WARNING! pmap->pm_spin held
4084  */
4085 static int
4086 pmap_scan_callback(pv_entry_t pv, void *data)
4087 {
4088         struct pmap_scan_info *info = data;
4089         struct pmap *pmap = info->pmap;
4090         pv_entry_t pd_pv;       /* A page directory PV */
4091         pv_entry_t pt_pv;       /* A page table PV */
4092         vm_pindex_t *pt_placemark;
4093         pt_entry_t *ptep;
4094         pt_entry_t oldpte;
4095         vm_offset_t sva;
4096         vm_offset_t eva;
4097         vm_offset_t va_next;
4098         vm_pindex_t pd_pindex;
4099         int error;
4100
4101         /*
4102          * Stop if requested
4103          */
4104         if (info->stop)
4105                 return -1;
4106
4107         /*
4108          * Pull the PD pindex from the pv before releasing the spinlock.
4109          *
4110          * WARNING: pv is faked for kernel pmap scans.
4111          */
4112         pd_pindex = pv->pv_pindex;
4113         spin_unlock(&pmap->pm_spin);
4114         pv = NULL;      /* invalid after spinlock unlocked */
4115
4116         /*
4117          * Calculate the page range within the PD.  SIMPLE pmaps are
4118          * direct-mapped for the entire 2^64 address space.  Normal pmaps
4119          * reflect the user and kernel address space which requires
4120          * cannonicalization w/regards to converting pd_pindex's back
4121          * into addresses.
4122          */
4123         sva = (pd_pindex - pmap_pd_pindex(0)) << PDPSHIFT;
4124         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
4125             (sva & PML4_SIGNMASK)) {
4126                 sva |= PML4_SIGNMASK;
4127         }
4128         eva = sva + NBPDP;      /* can overflow */
4129         if (sva < info->sva)
4130                 sva = info->sva;
4131         if (eva < info->sva || eva > info->eva)
4132                 eva = info->eva;
4133
4134         /*
4135          * NOTE: kernel mappings do not track page table pages, only
4136          *       terminal pages.
4137          *
4138          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
4139          *       However, for the scan to be efficient we try to
4140          *       cache items top-down.
4141          */
4142         pd_pv = NULL;
4143         pt_pv = NULL;
4144
4145         for (; sva < eva; sva = va_next) {
4146                 if (info->stop)
4147                         break;
4148                 if (sva >= VM_MAX_USER_ADDRESS) {
4149                         if (pt_pv) {
4150                                 pv_put(pt_pv);
4151                                 pt_pv = NULL;
4152                         }
4153                         goto kernel_skip;
4154                 }
4155
4156                 /*
4157                  * PD cache, scan shortcut if it doesn't exist.
4158                  */
4159                 if (pd_pv == NULL) {
4160                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4161                 } else if (pd_pv->pv_pmap != pmap ||
4162                            pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
4163                         pv_put(pd_pv);
4164                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4165                 }
4166                 if (pd_pv == NULL) {
4167                         va_next = (sva + NBPDP) & ~PDPMASK;
4168                         if (va_next < sva)
4169                                 va_next = eva;
4170                         continue;
4171                 }
4172
4173                 /*
4174                  * PT cache
4175                  *
4176                  * NOTE: The cached pt_pv can be removed from the pmap when
4177                  *       pmap_dynamic_delete is enabled.
4178                  */
4179                 if (pt_pv && (pt_pv->pv_pmap != pmap ||
4180                               pt_pv->pv_pindex != pmap_pt_pindex(sva))) {
4181                         pv_put(pt_pv);
4182                         pt_pv = NULL;
4183                 }
4184                 if (pt_pv == NULL) {
4185                         pt_pv = pv_get_try(pmap, pmap_pt_pindex(sva),
4186                                            &pt_placemark, &error);
4187                         if (error) {
4188                                 pv_put(pd_pv);  /* lock order */
4189                                 pd_pv = NULL;
4190                                 if (pt_pv) {
4191                                         pv_lock(pt_pv);
4192                                         pv_put(pt_pv);
4193                                         pt_pv = NULL;
4194                                 } else {
4195                                         pv_placemarker_wait(pmap, pt_placemark);
4196                                 }
4197                                 va_next = sva;
4198                                 continue;
4199                         }
4200                         /* may have to re-check later if pt_pv is NULL here */
4201                 }
4202
4203                 /*
4204                  * If pt_pv is NULL we either have an shared page table
4205                  * page and must issue a callback specific to that case,
4206                  * or there is no page table page.
4207                  *
4208                  * Either way we can skip the page table page.
4209                  *
4210                  * WARNING! pt_pv can also be NULL due to a pv creation
4211                  *          race where we find it to be NULL and then
4212                  *          later see a pte_pv.  But its possible the pt_pv
4213                  *          got created inbetween the two operations, so
4214                  *          we must check.
4215                  */
4216                 if (pt_pv == NULL) {
4217                         /*
4218                          * Possible unmanaged (shared from another pmap)
4219                          * page table page.
4220                          *
4221                          * WARNING!  We must hold pt_placemark across the
4222                          *           *ptep test to prevent misintepreting
4223                          *           a non-zero *ptep as a shared page
4224                          *           table page.  Hold it across the function
4225                          *           callback as well for SMP safety.
4226                          */
4227                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
4228                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
4229                                 info->func(pmap, info, NULL, pt_placemark,
4230                                            pd_pv, 1,
4231                                            sva, ptep, info->arg);
4232                         } else {
4233                                 pv_placemarker_wakeup(pmap, pt_placemark);
4234                         }
4235
4236                         /*
4237                          * Done, move to next page table page.
4238                          */
4239                         va_next = (sva + NBPDR) & ~PDRMASK;
4240                         if (va_next < sva)
4241                                 va_next = eva;
4242                         continue;
4243                 }
4244
4245                 /*
4246                  * From this point in the loop testing pt_pv for non-NULL
4247                  * means we are in UVM, else if it is NULL we are in KVM.
4248                  *
4249                  * Limit our scan to either the end of the va represented
4250                  * by the current page table page, or to the end of the
4251                  * range being removed.
4252                  */
4253 kernel_skip:
4254                 va_next = (sva + NBPDR) & ~PDRMASK;
4255                 if (va_next < sva)
4256                         va_next = eva;
4257                 if (va_next > eva)
4258                         va_next = eva;
4259
4260                 /*
4261                  * Scan the page table for pages.  Some pages may not be
4262                  * managed (might not have a pv_entry).
4263                  *
4264                  * There is no page table management for kernel pages so
4265                  * pt_pv will be NULL in that case, but otherwise pt_pv
4266                  * is non-NULL, locked, and referenced.
4267                  */
4268
4269                 /*
4270                  * At this point a non-NULL pt_pv means a UVA, and a NULL
4271                  * pt_pv means a KVA.
4272                  */
4273                 if (pt_pv)
4274                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
4275                 else
4276                         ptep = vtopte(sva);
4277
4278                 while (sva < va_next) {
4279                         pv_entry_t pte_pv;
4280                         vm_pindex_t *pte_placemark;
4281
4282                         /*
4283                          * Yield every 64 pages, stop if requested.
4284                          */
4285                         if ((++info->count & 63) == 0)
4286                                 lwkt_user_yield();
4287                         if (info->stop)
4288                                 break;
4289
4290                         /*
4291                          * We can shortcut our scan if *ptep == 0.  This is
4292                          * an unlocked check.
4293                          */
4294                         if (*ptep == 0) {
4295                                 sva += PAGE_SIZE;
4296                                 ++ptep;
4297                                 continue;
4298                         }
4299                         cpu_ccfence();
4300
4301                         /*
4302                          * Acquire the related pte_pv, if any.  If *ptep == 0
4303                          * the related pte_pv should not exist, but if *ptep
4304                          * is not zero the pte_pv may or may not exist (e.g.
4305                          * will not exist for an unmanaged page).
4306                          *
4307                          * However a multitude of races are possible here
4308                          * so if we cannot lock definite state we clean out
4309                          * our cache and break the inner while() loop to
4310                          * force a loop up to the top of the for().
4311                          *
4312                          * XXX unlock/relock pd_pv, pt_pv, and re-test their
4313                          *     validity instead of looping up?
4314                          */
4315                         pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
4316                                             &pte_placemark, &error);
4317                         if (error) {
4318                                 pv_put(pd_pv);          /* lock order */
4319                                 pd_pv = NULL;
4320                                 if (pt_pv) {
4321                                         pv_put(pt_pv);  /* lock order */
4322                                         pt_pv = NULL;
4323                                 }
4324                                 if (pte_pv) {           /* block */
4325                                         pv_lock(pte_pv);
4326                                         pv_put(pte_pv);
4327                                         pte_pv = NULL;
4328                                 } else {
4329                                         pv_placemarker_wait(pmap,
4330                                                         pte_placemark);
4331                                 }
4332                                 va_next = sva;          /* retry */
4333                                 break;
4334                         }
4335
4336                         /*
4337                          * Reload *ptep after successfully locking the
4338                          * pindex.  If *ptep == 0 we had better NOT have a
4339                          * pte_pv.
4340                          */
4341                         cpu_ccfence();
4342                         oldpte = *ptep;
4343                         if (oldpte == 0) {
4344                                 if (pte_pv) {
4345                                         kprintf("Unexpected non-NULL pte_pv "
4346                                                 "%p pt_pv %p "
4347                                                 "*ptep = %016lx/%016lx\n",
4348                                                 pte_pv, pt_pv, *ptep, oldpte);
4349                                         panic("Unexpected non-NULL pte_pv");
4350                                 } else {
4351                                         pv_placemarker_wakeup(pmap, pte_placemark);
4352                                 }
4353                                 sva += PAGE_SIZE;
4354                                 ++ptep;
4355                                 continue;
4356                         }
4357
4358                         /*
4359                          * We can't hold pd_pv across the callback (because
4360                          * we don't pass it to the callback and the callback
4361                          * might deadlock)
4362                          */
4363                         if (pd_pv) {
4364                                 vm_page_wire_quick(pd_pv->pv_m);
4365                                 pv_unlock(pd_pv);
4366                         }
4367
4368                         /*
4369                          * Ready for the callback.  The locked pte_pv (if any)
4370                          * is consumed by the callback.  pte_pv will exist if
4371                          * the page is managed, and will not exist if it
4372                          * isn't.
4373                          */
4374                         if (oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4375                                 /*
4376                                  * Managed pte
4377                                  */
4378                                 KASSERT(pte_pv &&
4379                                          (oldpte & pmap->pmap_bits[PG_V_IDX]),
4380                                     ("badC *ptep %016lx/%016lx sva %016lx "
4381                                     "pte_pv %p",
4382                                     *ptep, oldpte, sva, pte_pv));
4383                                 /*
4384                                  * We must unlock pd_pv across the callback
4385                                  * to avoid deadlocks on any recursive
4386                                  * disposal.  Re-check that it still exists
4387                                  * after re-locking.
4388                                  *
4389                                  * Call target disposes of pte_pv and may
4390                                  * destroy but will not dispose of pt_pv.
4391                                  */
4392                                 info->func(pmap, info, pte_pv, NULL,
4393                                            pt_pv, 0,
4394                                            sva, ptep, info->arg);
4395                         } else {
4396                                 /*
4397                                  * Unmanaged pte
4398                                  *
4399                                  * We must unlock pd_pv across the callback
4400                                  * to avoid deadlocks on any recursive
4401                                  * disposal.  Re-check that it still exists
4402                                  * after re-locking.
4403                                  *
4404                                  * Call target disposes of pte_pv or
4405                                  * pte_placemark and may destroy but will
4406                                  * not dispose of pt_pv.
4407                                  */
4408                                 KASSERT(pte_pv == NULL &&
4409                                         (oldpte & pmap->pmap_bits[PG_V_IDX]),
4410                                     ("badD *ptep %016lx/%016lx sva %016lx "
4411                                     "pte_pv %p pte_pv->pv_m %p ",
4412                                      *ptep, oldpte, sva,
4413                                      pte_pv, (pte_pv ? pte_pv->pv_m : NULL)));
4414                                 if (pte_pv)
4415                                         kprintf("RaceD\n");
4416                                 if (pte_pv) {
4417                                         info->func(pmap, info,
4418                                                    pte_pv, NULL,
4419                                                    pt_pv, 0,
4420                                                    sva, ptep, info->arg);
4421                                 } else {
4422                                         info->func(pmap, info,
4423                                                    NULL, pte_placemark,
4424                                                    pt_pv, 0,
4425                                                    sva, ptep, info->arg);
4426                                 }
4427                         }
4428                         if (pd_pv) {
4429                                 pv_lock(pd_pv);
4430                                 vm_page_unwire_quick(pd_pv->pv_m);
4431                                 if (pd_pv->pv_pmap == NULL) {
4432                                         va_next = sva;          /* retry */
4433                                         break;
4434                                 }
4435                         }
4436
4437                         /*
4438                          * NOTE: The cached pt_pv can be removed from the
4439                          *       pmap when pmap_dynamic_delete is enabled,
4440                          *       which will cause ptep to become stale.
4441                          *
4442                          *       This also means that no pages remain under
4443                          *       the PT, so we can just break out of the inner
4444                          *       loop and let the outer loop clean everything
4445                          *       up.
4446                          */
4447                         if (pt_pv && pt_pv->pv_pmap != pmap)
4448                                 break;
4449                         pte_pv = NULL;
4450                         sva += PAGE_SIZE;
4451                         ++ptep;
4452                 }
4453         }
4454         if (pd_pv) {
4455                 pv_put(pd_pv);
4456                 pd_pv = NULL;
4457         }
4458         if (pt_pv) {
4459                 pv_put(pt_pv);
4460                 pt_pv = NULL;
4461         }
4462         if ((++info->count & 7) == 0)
4463                 lwkt_user_yield();
4464
4465         /*
4466          * Relock before returning.
4467          */
4468         spin_lock(&pmap->pm_spin);
4469         return (0);
4470 }
4471
4472 void
4473 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4474 {
4475         struct pmap_scan_info info;
4476
4477         info.pmap = pmap;
4478         info.sva = sva;
4479         info.eva = eva;
4480         info.func = pmap_remove_callback;
4481         info.arg = NULL;
4482         pmap_scan(&info, 1);
4483 #if 0
4484         cpu_invltlb();
4485         if (eva - sva < 1024*1024) {
4486                 while (sva < eva) {
4487                         cpu_invlpg((void *)sva);
4488                         sva += PAGE_SIZE;
4489                 }
4490         }
4491 #endif
4492 }
4493
4494 static void
4495 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4496 {
4497         struct pmap_scan_info info;
4498
4499         info.pmap = pmap;
4500         info.sva = sva;
4501         info.eva = eva;
4502         info.func = pmap_remove_callback;
4503         info.arg = NULL;
4504         pmap_scan(&info, 0);
4505 }
4506
4507 static void
4508 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4509                      pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4510                      pv_entry_t pt_pv, int sharept,
4511                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4512 {
4513         pt_entry_t pte;
4514
4515         if (pte_pv) {
4516                 /*
4517                  * Managed entry
4518                  *
4519                  * This will also drop pt_pv's wire_count. Note that
4520                  * terminal pages are not wired based on mmu presence.
4521                  *
4522                  * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4523                  */
4524                 KKASSERT(pte_pv->pv_m != NULL);
4525                 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4526                 pte_pv = NULL;  /* safety */
4527
4528                 /*
4529                  * Recursively destroy higher-level page tables.
4530                  *
4531                  * This is optional.  If we do not, they will still
4532                  * be destroyed when the process exits.
4533                  *
4534                  * NOTE: Do not destroy pv_entry's with extra hold refs,
4535                  *       a caller may have unlocked it and intends to
4536                  *       continue to use it.
4537                  */
4538                 if (pmap_dynamic_delete &&
4539                     pt_pv &&
4540                     pt_pv->pv_m &&
4541                     pt_pv->pv_m->wire_count == 1 &&
4542                     (pt_pv->pv_hold & PV_HOLD_MASK) == 2 &&
4543                     pt_pv->pv_pindex != pmap_pml4_pindex()) {
4544                         if (pmap_dynamic_delete == 2)
4545                                 kprintf("B %jd %08x\n", pt_pv->pv_pindex, pt_pv->pv_hold);
4546                         pv_hold(pt_pv); /* extra hold */
4547                         pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4548                         pv_lock(pt_pv); /* prior extra hold + relock */
4549                 }
4550         } else if (sharept == 0) {
4551                 /*
4552                  * Unmanaged pte (pte_placemark is non-NULL)
4553                  *
4554                  * pt_pv's wire_count is still bumped by unmanaged pages
4555                  * so we must decrement it manually.
4556                  *
4557                  * We have to unwire the target page table page.
4558                  */
4559                 pte = pmap_inval_bulk(info->bulk, va, ptep, 0);
4560                 if (pte & pmap->pmap_bits[PG_W_IDX])
4561                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4562                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4563                 if (vm_page_unwire_quick(pt_pv->pv_m))
4564                         panic("pmap_remove: insufficient wirecount");
4565                 pv_placemarker_wakeup(pmap, pte_placemark);
4566         } else {
4567                 /*
4568                  * Unmanaged page table (pt, pd, or pdp. Not pte) for
4569                  * a shared page table.
4570                  *
4571                  * pt_pv is actually the pd_pv for our pmap (not the shared
4572                  * object pmap).
4573                  *
4574                  * We have to unwire the target page table page and we
4575                  * have to unwire our page directory page.
4576                  *
4577                  * It is unclear how we can invalidate a segment so we
4578                  * invalidate -1 which invlidates the tlb.
4579                  */
4580                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4581                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4582                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4583                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4584                         panic("pmap_remove: shared pgtable1 bad wirecount");
4585                 if (vm_page_unwire_quick(pt_pv->pv_m))
4586                         panic("pmap_remove: shared pgtable2 bad wirecount");
4587                 pv_placemarker_wakeup(pmap, pte_placemark);
4588         }
4589 }
4590
4591 /*
4592  * Removes this physical page from all physical maps in which it resides.
4593  * Reflects back modify bits to the pager.
4594  *
4595  * This routine may not be called from an interrupt.
4596  */
4597 static
4598 void
4599 pmap_remove_all(vm_page_t m)
4600 {
4601         pv_entry_t pv;
4602         pmap_inval_bulk_t bulk;
4603
4604         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4605                 return;
4606
4607         vm_page_spin_lock(m);
4608         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4609                 KKASSERT(pv->pv_m == m);
4610                 if (pv_hold_try(pv)) {
4611                         vm_page_spin_unlock(m);
4612                 } else {
4613                         vm_page_spin_unlock(m);
4614                         pv_lock(pv);
4615                         pv_put(pv);
4616                         vm_page_spin_lock(m);
4617                         continue;
4618                 }
4619                 KKASSERT(pv->pv_pmap && pv->pv_m == m);
4620
4621                 /*
4622                  * Holding no spinlocks, pv is locked.  Once we scrap
4623                  * pv we can no longer use it as a list iterator (but
4624                  * we are doing a TAILQ_FIRST() so we are ok).
4625                  */
4626                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4627                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4628                 pv = NULL;      /* safety */
4629                 pmap_inval_bulk_flush(&bulk);
4630                 vm_page_spin_lock(m);
4631         }
4632         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4633         vm_page_spin_unlock(m);
4634 }
4635
4636 /*
4637  * Removes the page from a particular pmap
4638  */
4639 void
4640 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4641 {
4642         pv_entry_t pv;
4643         pmap_inval_bulk_t bulk;
4644
4645         if (!pmap_initialized)
4646                 return;
4647
4648 again:
4649         vm_page_spin_lock(m);
4650         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4651                 if (pv->pv_pmap != pmap)
4652                         continue;
4653                 KKASSERT(pv->pv_m == m);
4654                 if (pv_hold_try(pv)) {
4655                         vm_page_spin_unlock(m);
4656                 } else {
4657                         vm_page_spin_unlock(m);
4658                         pv_lock(pv);
4659                         pv_put(pv);
4660                         goto again;
4661                 }
4662                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
4663
4664                 /*
4665                  * Holding no spinlocks, pv is locked.  Once gone it can't
4666                  * be used as an iterator.  In fact, because we couldn't
4667                  * necessarily lock it atomically it may have moved within
4668                  * the list and ALSO cannot be used as an iterator.
4669                  */
4670                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4671                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4672                 pv = NULL;      /* safety */
4673                 pmap_inval_bulk_flush(&bulk);
4674                 goto again;
4675         }
4676         vm_page_spin_unlock(m);
4677 }
4678
4679 /*
4680  * Set the physical protection on the specified range of this map
4681  * as requested.  This function is typically only used for debug watchpoints
4682  * and COW pages.
4683  *
4684  * This function may not be called from an interrupt if the map is
4685  * not the kernel_pmap.
4686  *
4687  * NOTE!  For shared page table pages we just unmap the page.
4688  */
4689 void
4690 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4691 {
4692         struct pmap_scan_info info;
4693         /* JG review for NX */
4694
4695         if (pmap == NULL)
4696                 return;
4697         if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == VM_PROT_NONE) {
4698                 pmap_remove(pmap, sva, eva);
4699                 return;
4700         }
4701         if (prot & VM_PROT_WRITE)
4702                 return;
4703         info.pmap = pmap;
4704         info.sva = sva;
4705         info.eva = eva;
4706         info.func = pmap_protect_callback;
4707         info.arg = &prot;
4708         pmap_scan(&info, 1);
4709 }
4710
4711 static
4712 void
4713 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4714                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4715                       pv_entry_t pt_pv, int sharept,
4716                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4717 {
4718         pt_entry_t pbits;
4719         pt_entry_t cbits;
4720         pt_entry_t pte;
4721         vm_page_t m;
4722
4723 again:
4724         pbits = *ptep;
4725         cbits = pbits;
4726         if (pte_pv) {
4727                 KKASSERT(pte_pv->pv_m != NULL);
4728                 m = NULL;
4729                 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4730                         if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4731                                 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4732                                 KKASSERT(m == pte_pv->pv_m);
4733                                 vm_page_flag_set(m, PG_REFERENCED);
4734                         }
4735                         cbits &= ~pmap->pmap_bits[PG_A_IDX];
4736                 }
4737                 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4738                         if (pmap_track_modified(pte_pv->pv_pindex)) {
4739                                 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4740                                         if (m == NULL) {
4741                                                 m = PHYS_TO_VM_PAGE(pbits &
4742                                                                     PG_FRAME);
4743                                         }
4744                                         vm_page_dirty(m);
4745                                 }
4746                                 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4747                         }
4748                 }
4749         } else if (sharept) {
4750                 /*
4751                  * Unmanaged page table, pt_pv is actually the pd_pv
4752                  * for our pmap (not the object's shared pmap).
4753                  *
4754                  * When asked to protect something in a shared page table
4755                  * page we just unmap the page table page.  We have to
4756                  * invalidate the tlb in this situation.
4757                  *
4758                  * XXX Warning, shared page tables will not be used for
4759                  * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4760                  * so PHYS_TO_VM_PAGE() should be safe here.
4761                  */
4762                 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4763                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4764                         panic("pmap_protect: pgtable1 pg bad wirecount");
4765                 if (vm_page_unwire_quick(pt_pv->pv_m))
4766                         panic("pmap_protect: pgtable2 pg bad wirecount");
4767                 ptep = NULL;
4768         }
4769         /* else unmanaged page, adjust bits, no wire changes */
4770
4771         if (ptep) {
4772                 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4773 #ifdef PMAP_DEBUG2
4774                 if (pmap_enter_debug > 0) {
4775                         --pmap_enter_debug;
4776                         kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4777                                 "pt_pv=%p cbits=%08lx\n",
4778                                 va, ptep, pte_pv,
4779                                 pt_pv, cbits
4780                         );
4781                 }
4782 #endif
4783                 if (pbits != cbits) {
4784                         vm_offset_t xva;
4785
4786                         xva = (sharept) ? (vm_offset_t)-1 : va;
4787                         if (!pmap_inval_smp_cmpset(pmap, xva,
4788                                                    ptep, pbits, cbits)) {
4789                                 goto again;
4790                         }
4791                 }
4792         }
4793         if (pte_pv)
4794                 pv_put(pte_pv);
4795         else
4796                 pv_placemarker_wakeup(pmap, pte_placemark);
4797 }
4798
4799 /*
4800  * Insert the vm_page (m) at the virtual address (va), replacing any prior
4801  * mapping at that address.  Set protection and wiring as requested.
4802  *
4803  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4804  * possible.  If it is we enter the page into the appropriate shared pmap
4805  * hanging off the related VM object instead of the passed pmap, then we
4806  * share the page table page from the VM object's pmap into the current pmap.
4807  *
4808  * NOTE: This routine MUST insert the page into the pmap now, it cannot
4809  *       lazy-evaluate.
4810  *
4811  * NOTE: If (m) is PG_UNMANAGED it may also be a temporary fake vm_page_t.
4812  *       never record it.
4813  */
4814 void
4815 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4816            boolean_t wired, vm_map_entry_t entry)
4817 {
4818         pv_entry_t pt_pv;       /* page table */
4819         pv_entry_t pte_pv;      /* page table entry */
4820         vm_pindex_t *pte_placemark;
4821         pt_entry_t *ptep;
4822         vm_paddr_t opa;
4823         pt_entry_t origpte, newpte;
4824         vm_paddr_t pa;
4825
4826         if (pmap == NULL)
4827                 return;
4828         va = trunc_page(va);
4829 #ifdef PMAP_DIAGNOSTIC
4830         if (va >= KvaEnd)
4831                 panic("pmap_enter: toobig");
4832         if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4833                 panic("pmap_enter: invalid to pmap_enter page table "
4834                       "pages (va: 0x%lx)", va);
4835 #endif
4836         if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4837                 kprintf("Warning: pmap_enter called on UVA with "
4838                         "kernel_pmap\n");
4839 #ifdef DDB
4840                 db_print_backtrace();
4841 #endif
4842         }
4843         if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4844                 kprintf("Warning: pmap_enter called on KVA without"
4845                         "kernel_pmap\n");
4846 #ifdef DDB
4847                 db_print_backtrace();
4848 #endif
4849         }
4850
4851         /*
4852          * Get locked PV entries for our new page table entry (pte_pv or
4853          * pte_placemark) and for its parent page table (pt_pv).  We need
4854          * the parent so we can resolve the location of the ptep.
4855          *
4856          * Only hardware MMU actions can modify the ptep out from
4857          * under us.
4858          *
4859          * if (m) is fictitious or unmanaged we do not create a managing
4860          * pte_pv for it.  Any pre-existing page's management state must
4861          * match (avoiding code complexity).
4862          *
4863          * If the pmap is still being initialized we assume existing
4864          * page tables.
4865          *
4866          * Kernel mapppings do not track page table pages (i.e. pt_pv).
4867          *
4868          * WARNING! If replacing a managed mapping with an unmanaged mapping
4869          *          pte_pv will wind up being non-NULL and must be handled
4870          *          below.
4871          */
4872         if (pmap_initialized == FALSE) {
4873                 pte_pv = NULL;
4874                 pt_pv = NULL;
4875                 pte_placemark = NULL;
4876                 ptep = vtopte(va);
4877                 origpte = *ptep;
4878         } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4879                 pmap_softwait(pmap);
4880                 pte_pv = pv_get(pmap, pmap_pte_pindex(va), &pte_placemark);
4881                 KKASSERT(pte_pv == NULL);
4882                 if (va >= VM_MAX_USER_ADDRESS) {
4883                         pt_pv = NULL;
4884                         ptep = vtopte(va);
4885                 } else {
4886                         pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4887                                                   NULL, entry, va);
4888                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4889                 }
4890                 origpte = *ptep;
4891                 cpu_ccfence();
4892                 KASSERT(origpte == 0 ||
4893                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4894                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4895         } else {
4896                 pmap_softwait(pmap);
4897                 if (va >= VM_MAX_USER_ADDRESS) {
4898                         /*
4899                          * Kernel map, pv_entry-tracked.
4900                          */
4901                         pt_pv = NULL;
4902                         pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4903                         ptep = vtopte(va);
4904                 } else {
4905                         /*
4906                          * User map
4907                          */
4908                         pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4909                                                    &pt_pv, entry, va);
4910                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4911                 }
4912                 pte_placemark = NULL;   /* safety */
4913                 origpte = *ptep;
4914                 cpu_ccfence();
4915                 KASSERT(origpte == 0 ||
4916                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4917                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4918         }
4919
4920         pa = VM_PAGE_TO_PHYS(m);
4921         opa = origpte & PG_FRAME;
4922
4923         /*
4924          * Calculate the new PTE.  Note that pte_pv alone does not mean
4925          * the new pte_pv is managed, it could exist because the old pte
4926          * was managed even if the new one is not.
4927          */
4928         newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4929                  pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4930         if (wired)
4931                 newpte |= pmap->pmap_bits[PG_W_IDX];
4932         if (va < VM_MAX_USER_ADDRESS)
4933                 newpte |= pmap->pmap_bits[PG_U_IDX];
4934         if (pte_pv && (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) == 0)
4935                 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4936 //      if (pmap == &kernel_pmap)
4937 //              newpte |= pgeflag;
4938         newpte |= pmap->pmap_cache_bits[m->pat_mode];
4939         if (m->flags & PG_FICTITIOUS)
4940                 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4941
4942         /*
4943          * It is possible for multiple faults to occur in threaded
4944          * environments, the existing pte might be correct.
4945          */
4946         if (((origpte ^ newpte) &
4947             ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4948                           pmap->pmap_bits[PG_A_IDX])) == 0) {
4949                 goto done;
4950         }
4951
4952         /*
4953          * Ok, either the address changed or the protection or wiring
4954          * changed.
4955          *
4956          * Clear the current entry, interlocking the removal.  For managed
4957          * pte's this will also flush the modified state to the vm_page.
4958          * Atomic ops are mandatory in order to ensure that PG_M events are
4959          * not lost during any transition.
4960          *
4961          * WARNING: The caller has busied the new page but not the original
4962          *          vm_page which we are trying to replace.  Because we hold
4963          *          the pte_pv lock, but have not busied the page, PG bits
4964          *          can be cleared out from under us.
4965          */
4966         if (opa) {
4967                 if (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4968                         /*
4969                          * Old page was managed.  Expect pte_pv to exist.
4970                          * (it might also exist if the old page was unmanaged).
4971                          *
4972                          * NOTE: pt_pv won't exist for a kernel page
4973                          *       (managed or otherwise).
4974                          *
4975                          * NOTE: We may be reusing the pte_pv so we do not
4976                          *       destroy it in pmap_remove_pv_pte().
4977                          */
4978                         KKASSERT(pte_pv && pte_pv->pv_m);
4979                         if (prot & VM_PROT_NOSYNC) {
4980                                 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4981                         } else {
4982                                 pmap_inval_bulk_t bulk;
4983
4984                                 pmap_inval_bulk_init(&bulk, pmap);
4985                                 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4986                                 pmap_inval_bulk_flush(&bulk);
4987                         }
4988                         pmap_remove_pv_page(pte_pv);
4989                         /* will either set pte_pv->pv_m or pv_free() later */
4990                 } else {
4991                         /*
4992                          * Old page was not managed.  If we have a pte_pv
4993                          * it better not have a pv_m assigned to it.  If the
4994                          * new page is managed the pte_pv will be destroyed
4995                          * near the end (we need its interlock).
4996                          *
4997                          * NOTE: We leave the wire count on the PT page
4998                          *       intact for the followup enter, but adjust
4999                          *       the wired-pages count on the pmap.
5000                          */
5001                         KKASSERT(pte_pv == NULL);
5002                         if (prot & VM_PROT_NOSYNC) {
5003                                 /*
5004                                  * NOSYNC (no mmu sync) requested.
5005                                  */
5006                                 (void)pte_load_clear(ptep);
5007                                 cpu_invlpg((void *)va);
5008                         } else {
5009                                 /*
5010                                  * Nominal SYNC
5011                                  */
5012                                 pmap_inval_smp(pmap, va, 1, ptep, 0);
5013                         }
5014
5015                         /*
5016                          * We must adjust pm_stats manually for unmanaged
5017                          * pages.
5018                          */
5019                         if (pt_pv) {
5020                                 atomic_add_long(&pmap->pm_stats.
5021                                                 resident_count, -1);
5022                         }
5023                         if (origpte & pmap->pmap_bits[PG_W_IDX]) {
5024                                 atomic_add_long(&pmap->pm_stats.
5025                                                 wired_count, -1);
5026                         }
5027                 }
5028                 KKASSERT(*ptep == 0);
5029         }
5030
5031 #ifdef PMAP_DEBUG2
5032         if (pmap_enter_debug > 0) {
5033                 --pmap_enter_debug;
5034                 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
5035                         " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
5036                         va, m,
5037                         origpte, newpte, ptep,
5038                         pte_pv, pt_pv, opa, prot);
5039         }
5040 #endif
5041
5042         if ((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5043                 /*
5044                  * Entering an unmanaged page.  We must wire the pt_pv unless
5045                  * we retained the wiring from an unmanaged page we had
5046                  * removed (if we retained it via pte_pv that will go away
5047                  * soon).
5048                  */
5049                 if (pt_pv && (opa == 0 ||
5050                               (origpte & pmap->pmap_bits[PG_MANAGED_IDX]))) {
5051                         vm_page_wire_quick(pt_pv->pv_m);
5052                 }
5053                 if (wired)
5054                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
5055
5056                 /*
5057                  * Unmanaged pages need manual resident_count tracking.
5058                  */
5059                 if (pt_pv) {
5060                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.
5061                                         resident_count, 1);
5062                 }
5063                 if (newpte & pmap->pmap_bits[PG_RW_IDX])
5064                         vm_page_flag_set(m, PG_WRITEABLE);
5065         } else {
5066                 /*
5067                  * Entering a managed page.  Our pte_pv takes care of the
5068                  * PT wiring, so if we had removed an unmanaged page before
5069                  * we must adjust.
5070                  *
5071                  * We have to take care of the pmap wired count ourselves.
5072                  *
5073                  * Enter on the PV list if part of our managed memory.
5074                  */
5075                 KKASSERT(pte_pv && (pte_pv->pv_m == NULL || pte_pv->pv_m == m));
5076                 vm_page_spin_lock(m);
5077                 pte_pv->pv_m = m;
5078                 pmap_page_stats_adding(m);
5079                 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
5080                 vm_page_flag_set(m, PG_MAPPED);
5081                 if (newpte & pmap->pmap_bits[PG_RW_IDX])
5082                         vm_page_flag_set(m, PG_WRITEABLE);
5083                 vm_page_spin_unlock(m);
5084
5085                 if (pt_pv && opa &&
5086                     (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5087                         vm_page_unwire_quick(pt_pv->pv_m);
5088                 }
5089
5090                 /*
5091                  * Adjust pmap wired pages count for new entry.
5092                  */
5093                 if (wired) {
5094                         atomic_add_long(&pte_pv->pv_pmap->pm_stats.
5095                                         wired_count, 1);
5096                 }
5097         }
5098
5099         /*
5100          * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
5101          *
5102          * User VMAs do not because those will be zero->non-zero, so no
5103          * stale entries to worry about at this point.
5104          *
5105          * For KVM there appear to still be issues.  Theoretically we
5106          * should be able to scrap the interlocks entirely but we
5107          * get crashes.
5108          */
5109         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
5110                 pmap_inval_smp(pmap, va, 1, ptep, newpte);
5111         } else {
5112                 origpte = atomic_swap_long(ptep, newpte);
5113                 if (origpte & pmap->pmap_bits[PG_M_IDX]) {
5114                         kprintf("pmap [M] race @ %016jx\n", va);
5115                         atomic_set_long(ptep, pmap->pmap_bits[PG_M_IDX]);
5116                 }
5117                 if (pt_pv == NULL)
5118                         cpu_invlpg((void *)va);
5119         }
5120
5121         /*
5122          * Cleanup
5123          */
5124 done:
5125         KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
5126                  (m->flags & PG_MAPPED));
5127
5128         /*
5129          * Cleanup the pv entry, allowing other accessors.  If the new page
5130          * is not managed but we have a pte_pv (which was locking our
5131          * operation), we can free it now.  pte_pv->pv_m should be NULL.
5132          */
5133         if (pte_pv && (newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5134                 pv_free(pte_pv, pt_pv);
5135         } else if (pte_pv) {
5136                 pv_put(pte_pv);
5137         } else if (pte_placemark) {
5138                 pv_placemarker_wakeup(pmap, pte_placemark);
5139         }
5140         if (pt_pv)
5141                 pv_put(pt_pv);
5142 }
5143
5144 /*
5145  * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
5146  * This code also assumes that the pmap has no pre-existing entry for this
5147  * VA.
5148  *
5149  * This code currently may only be used on user pmaps, not kernel_pmap.
5150  */
5151 void
5152 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
5153 {
5154         pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
5155 }
5156
5157 /*
5158  * Make a temporary mapping for a physical address.  This is only intended
5159  * to be used for panic dumps.
5160  *
5161  * The caller is responsible for calling smp_invltlb().
5162  */
5163 void *
5164 pmap_kenter_temporary(vm_paddr_t pa, long i)
5165 {
5166         pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
5167         return ((void *)crashdumpmap);
5168 }
5169
5170 #define MAX_INIT_PT (96)
5171
5172 /*
5173  * This routine preloads the ptes for a given object into the specified pmap.
5174  * This eliminates the blast of soft faults on process startup and
5175  * immediately after an mmap.
5176  */
5177 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
5178
5179 void
5180 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
5181                     vm_object_t object, vm_pindex_t pindex,
5182                     vm_size_t size, int limit)
5183 {
5184         struct rb_vm_page_scan_info info;
5185         struct lwp *lp;
5186         vm_size_t psize;
5187
5188         /*
5189          * We can't preinit if read access isn't set or there is no pmap
5190          * or object.
5191          */
5192         if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
5193                 return;
5194
5195         /*
5196          * We can't preinit if the pmap is not the current pmap
5197          */
5198         lp = curthread->td_lwp;
5199         if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
5200                 return;
5201
5202         /*
5203          * Misc additional checks
5204          */
5205         psize = x86_64_btop(size);
5206
5207         if ((object->type != OBJT_VNODE) ||
5208                 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
5209                         (object->resident_page_count > MAX_INIT_PT))) {
5210                 return;
5211         }
5212
5213         if (pindex + psize > object->size) {
5214                 if (object->size < pindex)
5215                         return;
5216                 psize = object->size - pindex;
5217         }
5218
5219         if (psize == 0)
5220                 return;
5221
5222         /*
5223          * If everything is segment-aligned do not pre-init here.  Instead
5224          * allow the normal vm_fault path to pass a segment hint to
5225          * pmap_enter() which will then use an object-referenced shared
5226          * page table page.
5227          */
5228         if ((addr & SEG_MASK) == 0 &&
5229             (ctob(psize) & SEG_MASK) == 0 &&
5230             (ctob(pindex) & SEG_MASK) == 0) {
5231                 return;
5232         }
5233
5234         /*
5235          * Use a red-black scan to traverse the requested range and load
5236          * any valid pages found into the pmap.
5237          *
5238          * We cannot safely scan the object's memq without holding the
5239          * object token.
5240          */
5241         info.start_pindex = pindex;
5242         info.end_pindex = pindex + psize - 1;
5243         info.limit = limit;
5244         info.mpte = NULL;
5245         info.addr = addr;
5246         info.pmap = pmap;
5247         info.object = object;
5248
5249         /*
5250          * By using the NOLK scan, the callback function must be sure
5251          * to return -1 if the VM page falls out of the object.
5252          */
5253         vm_object_hold_shared(object);
5254         vm_page_rb_tree_RB_SCAN_NOLK(&object->rb_memq, rb_vm_page_scancmp,
5255                                      pmap_object_init_pt_callback, &info);
5256         vm_object_drop(object);
5257 }
5258
5259 static
5260 int
5261 pmap_object_init_pt_callback(vm_page_t p, void *data)
5262 {
5263         struct rb_vm_page_scan_info *info = data;
5264         vm_pindex_t rel_index;
5265         int hard_busy;
5266
5267         /*
5268          * don't allow an madvise to blow away our really
5269          * free pages allocating pv entries.
5270          */
5271         if ((info->limit & MAP_PREFAULT_MADVISE) &&
5272                 vmstats.v_free_count < vmstats.v_free_reserved) {
5273                     return(-1);
5274         }
5275
5276         /*
5277          * Ignore list markers and ignore pages we cannot instantly
5278          * busy (while holding the object token).
5279          */
5280         if (p->flags & PG_MARKER)
5281                 return 0;
5282         hard_busy = 0;
5283 again:
5284         if (hard_busy) {
5285                 if (vm_page_busy_try(p, TRUE))
5286                         return 0;
5287         } else {
5288                 if (vm_page_sbusy_try(p))
5289                         return 0;
5290         }
5291         if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
5292             (p->flags & PG_FICTITIOUS) == 0) {
5293                 if ((p->queue - p->pc) == PQ_CACHE) {
5294                         if (hard_busy == 0) {
5295                                 vm_page_sbusy_drop(p);
5296                                 hard_busy = 1;
5297                                 goto again;
5298                         }
5299                         vm_page_deactivate(p);
5300                 }
5301                 rel_index = p->pindex - info->start_pindex;
5302                 pmap_enter_quick(info->pmap,
5303                                  info->addr + x86_64_ptob(rel_index), p);
5304         }
5305         if (hard_busy)
5306                 vm_page_wakeup(p);
5307         else
5308                 vm_page_sbusy_drop(p);
5309
5310         /*
5311          * We are using an unlocked scan (that is, the scan expects its
5312          * current element to remain in the tree on return).  So we have
5313          * to check here and abort the scan if it isn't.
5314          */
5315         if (p->object != info->object)
5316                 return -1;
5317         lwkt_yield();
5318         return(0);
5319 }
5320
5321 /*
5322  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
5323  * address.
5324  *
5325  * Returns FALSE if it would be non-trivial or if a pte is already loaded
5326  * into the slot.
5327  *
5328  * XXX This is safe only because page table pages are not freed.
5329  */
5330 int
5331 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
5332 {
5333         pt_entry_t *pte;
5334
5335         /*spin_lock(&pmap->pm_spin);*/
5336         if ((pte = pmap_pte(pmap, addr)) != NULL) {
5337                 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
5338                         /*spin_unlock(&pmap->pm_spin);*/
5339                         return FALSE;
5340                 }
5341         }
5342         /*spin_unlock(&pmap->pm_spin);*/
5343         return TRUE;
5344 }
5345
5346 /*
5347  * Change the wiring attribute for a pmap/va pair.  The mapping must already
5348  * exist in the pmap.  The mapping may or may not be managed.  The wiring in
5349  * the page is not changed, the page is returned so the caller can adjust
5350  * its wiring (the page is not locked in any way).
5351  *
5352  * Wiring is not a hardware characteristic so there is no need to invalidate
5353  * TLB.  However, in an SMP environment we must use a locked bus cycle to
5354  * update the pte (if we are not using the pmap_inval_*() API that is)...
5355  * it's ok to do this for simple wiring changes.
5356  */
5357 vm_page_t
5358 pmap_unwire(pmap_t pmap, vm_offset_t va)
5359 {
5360         pt_entry_t *ptep;
5361         pv_entry_t pt_pv;
5362         vm_paddr_t pa;
5363         vm_page_t m;
5364
5365         if (pmap == NULL)
5366                 return NULL;
5367
5368         /*
5369          * Assume elements in the kernel pmap are stable
5370          */
5371         if (pmap == &kernel_pmap) {
5372                 if (pmap_pt(pmap, va) == 0)
5373                         return NULL;
5374                 ptep = pmap_pte_quick(pmap, va);
5375                 if (pmap_pte_v(pmap, ptep)) {
5376                         if (pmap_pte_w(pmap, ptep))
5377                                 atomic_add_long(&pmap->pm_stats.wired_count,-1);
5378                         atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5379                         pa = *ptep & PG_FRAME;
5380                         m = PHYS_TO_VM_PAGE(pa);
5381                 } else {
5382                         m = NULL;
5383                 }
5384         } else {
5385                 /*
5386                  * We can only [un]wire pmap-local pages (we cannot wire
5387                  * shared pages)
5388                  */
5389                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
5390                 if (pt_pv == NULL)
5391                         return NULL;
5392
5393                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
5394                 if ((*ptep & pmap->pmap_bits[PG_V_IDX]) == 0) {
5395                         pv_put(pt_pv);
5396                         return NULL;
5397                 }
5398
5399                 if (pmap_pte_w(pmap, ptep)) {
5400                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.wired_count,
5401                                         -1);
5402                 }
5403                 /* XXX else return NULL so caller doesn't unwire m ? */
5404
5405                 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5406
5407                 pa = *ptep & PG_FRAME;
5408                 m = PHYS_TO_VM_PAGE(pa);        /* held by wired count */
5409                 pv_put(pt_pv);
5410         }
5411         return m;
5412 }
5413
5414 /*
5415  * Copy the range specified by src_addr/len from the source map to
5416  * the range dst_addr/len in the destination map.
5417  *
5418  * This routine is only advisory and need not do anything.
5419  */
5420 void
5421 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
5422           vm_size_t len, vm_offset_t src_addr)
5423 {
5424 }
5425
5426 /*
5427  * pmap_zero_page:
5428  *
5429  *      Zero the specified physical page.
5430  *
5431  *      This function may be called from an interrupt and no locking is
5432  *      required.
5433  */
5434 void
5435 pmap_zero_page(vm_paddr_t phys)
5436 {
5437         vm_offset_t va = PHYS_TO_DMAP(phys);
5438
5439         pagezero((void *)va);
5440 }
5441
5442 /*
5443  * pmap_zero_page:
5444  *
5445  *      Zero part of a physical page by mapping it into memory and clearing
5446  *      its contents with bzero.
5447  *
5448  *      off and size may not cover an area beyond a single hardware page.
5449  */
5450 void
5451 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
5452 {
5453         vm_offset_t virt = PHYS_TO_DMAP(phys);
5454
5455         bzero((char *)virt + off, size);
5456 }
5457
5458 /*
5459  * pmap_copy_page:
5460  *
5461  *      Copy the physical page from the source PA to the target PA.
5462  *      This function may be called from an interrupt.  No locking
5463  *      is required.
5464  */
5465 void
5466 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
5467 {
5468         vm_offset_t src_virt, dst_virt;
5469
5470         src_virt = PHYS_TO_DMAP(src);
5471         dst_virt = PHYS_TO_DMAP(dst);
5472         bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
5473 }
5474
5475 /*
5476  * pmap_copy_page_frag:
5477  *
5478  *      Copy the physical page from the source PA to the target PA.
5479  *      This function may be called from an interrupt.  No locking
5480  *      is required.
5481  */
5482 void
5483 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
5484 {
5485         vm_offset_t src_virt, dst_virt;
5486
5487         src_virt = PHYS_TO_DMAP(src);
5488         dst_virt = PHYS_TO_DMAP(dst);
5489
5490         bcopy((char *)src_virt + (src & PAGE_MASK),
5491               (char *)dst_virt + (dst & PAGE_MASK),
5492               bytes);
5493 }
5494
5495 /*
5496  * Returns true if the pmap's pv is one of the first 16 pvs linked to from
5497  * this page.  This count may be changed upwards or downwards in the future;
5498  * it is only necessary that true be returned for a small subset of pmaps
5499  * for proper page aging.
5500  */
5501 boolean_t
5502 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
5503 {
5504         pv_entry_t pv;
5505         int loops = 0;
5506
5507         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5508                 return FALSE;
5509
5510         vm_page_spin_lock(m);
5511         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5512                 if (pv->pv_pmap == pmap) {
5513                         vm_page_spin_unlock(m);
5514                         return TRUE;
5515                 }
5516                 loops++;
5517                 if (loops >= 16)
5518                         break;
5519         }
5520         vm_page_spin_unlock(m);
5521         return (FALSE);
5522 }
5523
5524 /*
5525  * Remove all pages from specified address space this aids process exit
5526  * speeds.  Also, this code may be special cased for the current process
5527  * only.
5528  */
5529 void
5530 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5531 {
5532         pmap_remove_noinval(pmap, sva, eva);
5533         cpu_invltlb();
5534 }
5535
5536 /*
5537  * pmap_testbit tests bits in pte's note that the testbit/clearbit
5538  * routines are inline, and a lot of things compile-time evaluate.
5539  */
5540
5541 static
5542 boolean_t
5543 pmap_testbit(vm_page_t m, int bit)
5544 {
5545         pv_entry_t pv;
5546         pt_entry_t *pte;
5547         pmap_t pmap;
5548
5549         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5550                 return FALSE;
5551
5552         if (TAILQ_FIRST(&m->md.pv_list) == NULL)
5553                 return FALSE;
5554         vm_page_spin_lock(m);
5555         if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
5556                 vm_page_spin_unlock(m);
5557                 return FALSE;
5558         }
5559
5560         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5561 #if defined(PMAP_DIAGNOSTIC)
5562                 if (pv->pv_pmap == NULL) {
5563                         kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
5564                             pv->pv_pindex);
5565                         continue;
5566                 }
5567 #endif
5568                 pmap = pv->pv_pmap;
5569
5570                 /*
5571                  * If the bit being tested is the modified bit, then
5572                  * mark clean_map and ptes as never
5573                  * modified.
5574                  *
5575                  * WARNING!  Because we do not lock the pv, *pte can be in a
5576                  *           state of flux.  Despite this the value of *pte
5577                  *           will still be related to the vm_page in some way
5578                  *           because the pv cannot be destroyed as long as we
5579                  *           hold the vm_page spin lock.
5580                  */
5581                 if (bit == PG_A_IDX || bit == PG_M_IDX) {
5582                                 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5583                         if (!pmap_track_modified(pv->pv_pindex))
5584                                 continue;
5585                 }
5586
5587                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5588                 if (*pte & pmap->pmap_bits[bit]) {
5589                         vm_page_spin_unlock(m);
5590                         return TRUE;
5591                 }
5592         }
5593         vm_page_spin_unlock(m);
5594         return (FALSE);
5595 }
5596
5597 /*
5598  * This routine is used to modify bits in ptes.  Only one bit should be
5599  * specified.  PG_RW requires special handling.
5600  *
5601  * Caller must NOT hold any spin locks
5602  */
5603 static __inline
5604 void
5605 pmap_clearbit(vm_page_t m, int bit_index)
5606 {
5607         pv_entry_t pv;
5608         pt_entry_t *pte;
5609         pt_entry_t pbits;
5610         pmap_t pmap;
5611
5612         if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5613                 if (bit_index == PG_RW_IDX)
5614                         vm_page_flag_clear(m, PG_WRITEABLE);
5615                 return;
5616         }
5617
5618         /*
5619          * PG_M or PG_A case
5620          *
5621          * Loop over all current mappings setting/clearing as appropos If
5622          * setting RO do we need to clear the VAC?
5623          *
5624          * NOTE: When clearing PG_M we could also (not implemented) drop
5625          *       through to the PG_RW code and clear PG_RW too, forcing
5626          *       a fault on write to redetect PG_M for virtual kernels, but
5627          *       it isn't necessary since virtual kernels invalidate the
5628          *       pte when they clear the VPTE_M bit in their virtual page
5629          *       tables.
5630          *
5631          * NOTE: Does not re-dirty the page when clearing only PG_M.
5632          *
5633          * NOTE: Because we do not lock the pv, *pte can be in a state of
5634          *       flux.  Despite this the value of *pte is still somewhat
5635          *       related while we hold the vm_page spin lock.
5636          *
5637          *       *pte can be zero due to this race.  Since we are clearing
5638          *       bits we basically do no harm when this race occurs.
5639          */
5640         if (bit_index != PG_RW_IDX) {
5641                 vm_page_spin_lock(m);
5642                 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5643 #if defined(PMAP_DIAGNOSTIC)
5644                         if (pv->pv_pmap == NULL) {
5645                                 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5646                                     pv->pv_pindex);
5647                                 continue;
5648                         }
5649 #endif
5650                         pmap = pv->pv_pmap;
5651                         pte = pmap_pte_quick(pv->pv_pmap,
5652                                              pv->pv_pindex << PAGE_SHIFT);
5653                         pbits = *pte;
5654                         if (pbits & pmap->pmap_bits[bit_index])
5655                                 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5656                 }
5657                 vm_page_spin_unlock(m);
5658                 return;
5659         }
5660
5661         /*
5662          * Clear PG_RW.  Also clears PG_M and marks the page dirty if PG_M
5663          * was set.
5664          */
5665 restart:
5666         vm_page_spin_lock(m);
5667         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5668                 /*
5669                  * don't write protect pager mappings
5670                  */
5671                 if (!pmap_track_modified(pv->pv_pindex))
5672                         continue;
5673
5674 #if defined(PMAP_DIAGNOSTIC)
5675                 if (pv->pv_pmap == NULL) {
5676                         kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5677                                 pv->pv_pindex);
5678                         continue;
5679                 }
5680 #endif
5681                 pmap = pv->pv_pmap;
5682
5683                 /*
5684                  * Skip pages which do not have PG_RW set.
5685                  */
5686                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5687                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5688                         continue;
5689
5690                 /*
5691                  * We must lock the PV to be able to safely test the pte.
5692                  */
5693                 if (pv_hold_try(pv)) {
5694                         vm_page_spin_unlock(m);
5695                 } else {
5696                         vm_page_spin_unlock(m);
5697                         pv_lock(pv);    /* held, now do a blocking lock */
5698                         pv_put(pv);
5699                         goto restart;
5700                 }
5701
5702                 /*
5703                  * Reload pte after acquiring pv.
5704                  */
5705                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5706 #if 0
5707                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0) {
5708                         pv_put(pv);
5709                         goto restart;
5710                 }
5711 #endif
5712
5713                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
5714                 for (;;) {
5715                         pt_entry_t nbits;
5716
5717                         pbits = *pte;
5718                         cpu_ccfence();
5719                         nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5720                                           pmap->pmap_bits[PG_M_IDX]);
5721                         if (pmap_inval_smp_cmpset(pmap,
5722                                      ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5723                                      pte, pbits, nbits)) {
5724                                 break;
5725                         }
5726                         cpu_pause();
5727                 }
5728
5729                 /*
5730                  * If PG_M was found to be set while we were clearing PG_RW
5731                  * we also clear PG_M (done above) and mark the page dirty.
5732                  * Callers expect this behavior.
5733                  *
5734                  * we lost pv so it cannot be used as an iterator.  In fact,
5735                  * because we couldn't necessarily lock it atomically it may
5736                  * have moved within the list and ALSO cannot be used as an
5737                  * iterator.
5738                  */
5739                 vm_page_spin_lock(m);
5740                 if (pbits & pmap->pmap_bits[PG_M_IDX])
5741                         vm_page_dirty(m);
5742                 vm_page_spin_unlock(m);
5743                 pv_put(pv);
5744                 goto restart;
5745         }
5746         if (bit_index == PG_RW_IDX)
5747                 vm_page_flag_clear(m, PG_WRITEABLE);
5748         vm_page_spin_unlock(m);
5749 }
5750
5751 /*
5752  * Lower the permission for all mappings to a given page.
5753  *
5754  * Page must be busied by caller.  Because page is busied by caller this
5755  * should not be able to race a pmap_enter().
5756  */
5757 void
5758 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5759 {
5760         /* JG NX support? */
5761         if ((prot & VM_PROT_WRITE) == 0) {
5762                 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5763                         /*
5764                          * NOTE: pmap_clearbit(.. PG_RW) also clears
5765                          *       the PG_WRITEABLE flag in (m).
5766                          */
5767                         pmap_clearbit(m, PG_RW_IDX);
5768                 } else {
5769                         pmap_remove_all(m);
5770                 }
5771         }
5772 }
5773
5774 vm_paddr_t
5775 pmap_phys_address(vm_pindex_t ppn)
5776 {
5777         return (x86_64_ptob(ppn));
5778 }
5779
5780 /*
5781  * Return a count of reference bits for a page, clearing those bits.
5782  * It is not necessary for every reference bit to be cleared, but it
5783  * is necessary that 0 only be returned when there are truly no
5784  * reference bits set.
5785  *
5786  * XXX: The exact number of bits to check and clear is a matter that
5787  * should be tested and standardized at some point in the future for
5788  * optimal aging of shared pages.
5789  *
5790  * This routine may not block.
5791  */
5792 int
5793 pmap_ts_referenced(vm_page_t m)
5794 {
5795         pv_entry_t pv;
5796         pt_entry_t *pte;
5797         pmap_t pmap;
5798         int rtval = 0;
5799
5800         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5801                 return (rtval);
5802
5803         vm_page_spin_lock(m);
5804         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5805                 if (!pmap_track_modified(pv->pv_pindex))
5806                         continue;
5807                 pmap = pv->pv_pmap;
5808                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5809                 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5810                         atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5811                         rtval++;
5812                         if (rtval > 4)
5813                                 break;
5814                 }
5815         }
5816         vm_page_spin_unlock(m);
5817         return (rtval);
5818 }
5819
5820 /*
5821  *      pmap_is_modified:
5822  *
5823  *      Return whether or not the specified physical page was modified
5824  *      in any physical maps.
5825  */
5826 boolean_t
5827 pmap_is_modified(vm_page_t m)
5828 {
5829         boolean_t res;
5830
5831         res = pmap_testbit(m, PG_M_IDX);
5832         return (res);
5833 }
5834
5835 /*
5836  *      Clear the modify bits on the specified physical page.
5837  */
5838 void
5839 pmap_clear_modify(vm_page_t m)
5840 {
5841         pmap_clearbit(m, PG_M_IDX);
5842 }
5843
5844 /*
5845  *      pmap_clear_reference:
5846  *
5847  *      Clear the reference bit on the specified physical page.
5848  */
5849 void
5850 pmap_clear_reference(vm_page_t m)
5851 {
5852         pmap_clearbit(m, PG_A_IDX);
5853 }
5854
5855 /*
5856  * Miscellaneous support routines follow
5857  */
5858
5859 static
5860 void
5861 i386_protection_init(void)
5862 {
5863         uint64_t *kp;
5864         int prot;
5865
5866         /*
5867          * NX supported? (boot time loader.conf override only)
5868          */
5869         TUNABLE_INT_FETCH("machdep.pmap_nx_enable", &pmap_nx_enable);
5870         if (pmap_nx_enable == 0 || (amd_feature & AMDID_NX) == 0)
5871                 pmap_bits_default[PG_NX_IDX] = 0;
5872
5873         /*
5874          * 0 is basically read-only access, but also set the NX (no-execute)
5875          * bit when VM_PROT_EXECUTE is not specified.
5876          */
5877         kp = protection_codes;
5878         for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5879                 switch (prot) {
5880                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5881                         /*
5882                          * This case handled elsewhere
5883                          */
5884                         *kp++ = 0;
5885                         break;
5886                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5887                         /*
5888                          * Read-only is 0|NX
5889                          */
5890                         *kp++ = pmap_bits_default[PG_NX_IDX];
5891                         break;
5892                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5893                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5894                         /*
5895                          * Execute requires read access
5896                          */
5897                         *kp++ = 0;
5898                         break;
5899                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5900                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5901                         /*
5902                          * Write without execute is RW|NX
5903                          */
5904                         *kp++ = pmap_bits_default[PG_RW_IDX] |
5905                                 pmap_bits_default[PG_NX_IDX];
5906                         break;
5907                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5908                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5909                         /*
5910                          * Write with execute is RW
5911                          */
5912                         *kp++ = pmap_bits_default[PG_RW_IDX];
5913                         break;
5914                 }
5915         }
5916 }
5917
5918 /*
5919  * Map a set of physical memory pages into the kernel virtual
5920  * address space. Return a pointer to where it is mapped. This
5921  * routine is intended to be used for mapping device memory,
5922  * NOT real memory.
5923  *
5924  * NOTE: We can't use pgeflag unless we invalidate the pages one at
5925  *       a time.
5926  *
5927  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5928  *       work whether the cpu supports PAT or not.  The remaining PAT
5929  *       attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5930  *       supports PAT.
5931  */
5932 void *
5933 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5934 {
5935         return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5936 }
5937
5938 void *
5939 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5940 {
5941         return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5942 }
5943
5944 void *
5945 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5946 {
5947         return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5948 }
5949
5950 /*
5951  * Map a set of physical memory pages into the kernel virtual
5952  * address space. Return a pointer to where it is mapped. This
5953  * routine is intended to be used for mapping device memory,
5954  * NOT real memory.
5955  */
5956 void *
5957 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5958 {
5959         vm_offset_t va, tmpva, offset;
5960         pt_entry_t *pte;
5961         vm_size_t tmpsize;
5962
5963         offset = pa & PAGE_MASK;
5964         size = roundup(offset + size, PAGE_SIZE);
5965
5966         va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5967         if (va == 0)
5968                 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5969
5970         pa = pa & ~PAGE_MASK;
5971         for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5972                 pte = vtopte(tmpva);
5973                 *pte = pa |
5974                     kernel_pmap.pmap_bits[PG_RW_IDX] |
5975                     kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5976                     kernel_pmap.pmap_cache_bits[mode];
5977                 tmpsize -= PAGE_SIZE;
5978                 tmpva += PAGE_SIZE;
5979                 pa += PAGE_SIZE;
5980         }
5981         pmap_invalidate_range(&kernel_pmap, va, va + size);
5982         pmap_invalidate_cache_range(va, va + size);
5983
5984         return ((void *)(va + offset));
5985 }
5986
5987 void
5988 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5989 {
5990         vm_offset_t base, offset;
5991
5992         base = va & ~PAGE_MASK;
5993         offset = va & PAGE_MASK;
5994         size = roundup(offset + size, PAGE_SIZE);
5995         pmap_qremove(va, size >> PAGE_SHIFT);
5996         kmem_free(&kernel_map, base, size);
5997 }
5998
5999 /*
6000  * Sets the memory attribute for the specified page.
6001  */
6002 void
6003 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
6004 {
6005
6006     m->pat_mode = ma;
6007
6008     /*
6009      * If "m" is a normal page, update its direct mapping.  This update
6010      * can be relied upon to perform any cache operations that are
6011      * required for data coherence.
6012      */
6013     if ((m->flags & PG_FICTITIOUS) == 0)
6014         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
6015 }
6016
6017 /*
6018  * Change the PAT attribute on an existing kernel memory map.  Caller
6019  * must ensure that the virtual memory in question is not accessed
6020  * during the adjustment.
6021  */
6022 void
6023 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
6024 {
6025         pt_entry_t *pte;
6026         vm_offset_t base;
6027         int changed = 0;
6028
6029         if (va == 0)
6030                 panic("pmap_change_attr: va is NULL");
6031         base = trunc_page(va);
6032
6033         while (count) {
6034                 pte = vtopte(va);
6035                 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
6036                        kernel_pmap.pmap_cache_bits[mode];
6037                 --count;
6038                 va += PAGE_SIZE;
6039         }
6040
6041         changed = 1;    /* XXX: not optimal */
6042
6043         /*
6044          * Flush CPU caches if required to make sure any data isn't cached that
6045          * shouldn't be, etc.
6046          */
6047         if (changed) {
6048                 pmap_invalidate_range(&kernel_pmap, base, va);
6049                 pmap_invalidate_cache_range(base, va);
6050         }
6051 }
6052
6053 /*
6054  * perform the pmap work for mincore
6055  */
6056 int
6057 pmap_mincore(pmap_t pmap, vm_offset_t addr)
6058 {
6059         pt_entry_t *ptep, pte;
6060         vm_page_t m;
6061         int val = 0;
6062
6063         ptep = pmap_pte(pmap, addr);
6064
6065         if (ptep && (pte = *ptep) != 0) {
6066                 vm_offset_t pa;
6067
6068                 val = MINCORE_INCORE;
6069                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
6070                         goto done;
6071
6072                 pa = pte & PG_FRAME;
6073
6074                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
6075                         m = NULL;
6076                 else
6077                         m = PHYS_TO_VM_PAGE(pa);
6078
6079                 /*
6080                  * Modified by us
6081                  */
6082                 if (pte & pmap->pmap_bits[PG_M_IDX])
6083                         val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
6084                 /*
6085                  * Modified by someone
6086                  */
6087                 else if (m && (m->dirty || pmap_is_modified(m)))
6088                         val |= MINCORE_MODIFIED_OTHER;
6089                 /*
6090                  * Referenced by us
6091                  */
6092                 if (pte & pmap->pmap_bits[PG_A_IDX])
6093                         val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
6094
6095                 /*
6096                  * Referenced by someone
6097                  */
6098                 else if (m && ((m->flags & PG_REFERENCED) ||
6099                                 pmap_ts_referenced(m))) {
6100                         val |= MINCORE_REFERENCED_OTHER;
6101                         vm_page_flag_set(m, PG_REFERENCED);
6102                 }
6103         }
6104 done:
6105
6106         return val;
6107 }
6108
6109 /*
6110  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
6111  * vmspace will be ref'd and the old one will be deref'd.
6112  *
6113  * The vmspace for all lwps associated with the process will be adjusted
6114  * and cr3 will be reloaded if any lwp is the current lwp.
6115  *
6116  * The process must hold the vmspace->vm_map.token for oldvm and newvm
6117  */
6118 void
6119 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
6120 {
6121         struct vmspace *oldvm;
6122         struct lwp *lp;
6123
6124         oldvm = p->p_vmspace;
6125         if (oldvm != newvm) {
6126                 if (adjrefs)
6127                         vmspace_ref(newvm);
6128                 p->p_vmspace = newvm;
6129                 KKASSERT(p->p_nthreads == 1);
6130                 lp = RB_ROOT(&p->p_lwp_tree);
6131                 pmap_setlwpvm(lp, newvm);
6132                 if (adjrefs)
6133                         vmspace_rel(oldvm);
6134         }
6135 }
6136
6137 /*
6138  * Set the vmspace for a LWP.  The vmspace is almost universally set the
6139  * same as the process vmspace, but virtual kernels need to swap out contexts
6140  * on a per-lwp basis.
6141  *
6142  * Caller does not necessarily hold any vmspace tokens.  Caller must control
6143  * the lwp (typically be in the context of the lwp).  We use a critical
6144  * section to protect against statclock and hardclock (statistics collection).
6145  */
6146 void
6147 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
6148 {
6149         struct vmspace *oldvm;
6150         struct pmap *pmap;
6151
6152         oldvm = lp->lwp_vmspace;
6153
6154         if (oldvm != newvm) {
6155                 crit_enter();
6156                 KKASSERT((newvm->vm_refcnt & VM_REF_DELETED) == 0);
6157                 lp->lwp_vmspace = newvm;
6158                 if (curthread->td_lwp == lp) {
6159                         pmap = vmspace_pmap(newvm);
6160                         ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
6161                         if (pmap->pm_active_lock & CPULOCK_EXCL)
6162                                 pmap_interlock_wait(newvm);
6163 #if defined(SWTCH_OPTIM_STATS)
6164                         tlb_flush_count++;
6165 #endif
6166                         if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
6167                                 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
6168                         } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
6169                                 curthread->td_pcb->pcb_cr3 = KPML4phys;
6170                         } else {
6171                                 panic("pmap_setlwpvm: unknown pmap type\n");
6172                         }
6173                         load_cr3(curthread->td_pcb->pcb_cr3);
6174                         pmap = vmspace_pmap(oldvm);
6175                         ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
6176                                                mycpu->gd_cpuid);
6177                 }
6178                 crit_exit();
6179         }
6180 }
6181
6182 /*
6183  * Called when switching to a locked pmap, used to interlock against pmaps
6184  * undergoing modifications to prevent us from activating the MMU for the
6185  * target pmap until all such modifications have completed.  We have to do
6186  * this because the thread making the modifications has already set up its
6187  * SMP synchronization mask.
6188  *
6189  * This function cannot sleep!
6190  *
6191  * No requirements.
6192  */
6193 void
6194 pmap_interlock_wait(struct vmspace *vm)
6195 {
6196         struct pmap *pmap = &vm->vm_pmap;
6197
6198         if (pmap->pm_active_lock & CPULOCK_EXCL) {
6199                 crit_enter();
6200                 KKASSERT(curthread->td_critcount >= 2);
6201                 DEBUG_PUSH_INFO("pmap_interlock_wait");
6202                 while (pmap->pm_active_lock & CPULOCK_EXCL) {
6203                         cpu_ccfence();
6204                         lwkt_process_ipiq();
6205                 }
6206                 DEBUG_POP_INFO();
6207                 crit_exit();
6208         }
6209 }
6210
6211 vm_offset_t
6212 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
6213 {
6214
6215         if ((obj == NULL) || (size < NBPDR) ||
6216             ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
6217                 return addr;
6218         }
6219
6220         addr = roundup2(addr, NBPDR);
6221         return addr;
6222 }
6223
6224 /*
6225  * Used by kmalloc/kfree, page already exists at va
6226  */
6227 vm_page_t
6228 pmap_kvtom(vm_offset_t va)
6229 {
6230         pt_entry_t *ptep = vtopte(va);
6231
6232         KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
6233         return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
6234 }
6235
6236 /*
6237  * Initialize machine-specific shared page directory support.  This
6238  * is executed when a VM object is created.
6239  */
6240 void
6241 pmap_object_init(vm_object_t object)
6242 {
6243         object->md.pmap_rw = NULL;
6244         object->md.pmap_ro = NULL;
6245 }
6246
6247 /*
6248  * Clean up machine-specific shared page directory support.  This
6249  * is executed when a VM object is destroyed.
6250  */
6251 void
6252 pmap_object_free(vm_object_t object)
6253 {
6254         pmap_t pmap;
6255
6256         if ((pmap = object->md.pmap_rw) != NULL) {
6257                 object->md.pmap_rw = NULL;
6258                 pmap_remove_noinval(pmap,
6259                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6260                 CPUMASK_ASSZERO(pmap->pm_active);
6261                 pmap_release(pmap);
6262                 pmap_puninit(pmap);
6263                 kfree(pmap, M_OBJPMAP);
6264         }
6265         if ((pmap = object->md.pmap_ro) != NULL) {
6266                 object->md.pmap_ro = NULL;
6267                 pmap_remove_noinval(pmap,
6268                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6269                 CPUMASK_ASSZERO(pmap->pm_active);
6270                 pmap_release(pmap);
6271                 pmap_puninit(pmap);
6272                 kfree(pmap, M_OBJPMAP);
6273         }
6274 }
6275
6276 /*
6277  * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
6278  * VM page and issue a pginfo->callback.
6279  *
6280  * We are expected to dispose of any non-NULL pte_pv.
6281  */
6282 static
6283 void
6284 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
6285                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
6286                       pv_entry_t pt_pv, int sharept,
6287                       vm_offset_t va, pt_entry_t *ptep, void *arg)
6288 {
6289         struct pmap_pgscan_info *pginfo = arg;
6290         vm_page_t m;
6291
6292         if (pte_pv) {
6293                 /*
6294                  * Try to busy the page while we hold the pte_pv locked.
6295                  */
6296                 KKASSERT(pte_pv->pv_m);
6297                 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
6298                 if (vm_page_busy_try(m, TRUE) == 0) {
6299                         if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
6300                                 /*
6301                                  * The callback is issued with the pte_pv
6302                                  * unlocked and put away, and the pt_pv
6303                                  * unlocked.
6304                                  */
6305                                 pv_put(pte_pv);
6306                                 if (pt_pv) {
6307                                         vm_page_wire_quick(pt_pv->pv_m);
6308                                         pv_unlock(pt_pv);
6309                                 }
6310                                 if (pginfo->callback(pginfo, va, m) < 0)
6311                                         info->stop = 1;
6312                                 if (pt_pv) {
6313                                         pv_lock(pt_pv);
6314                                         vm_page_unwire_quick(pt_pv->pv_m);
6315                                 }
6316                         } else {
6317                                 vm_page_wakeup(m);
6318                                 pv_put(pte_pv);
6319                         }
6320                 } else {
6321                         ++pginfo->busycount;
6322                         pv_put(pte_pv);
6323                 }
6324         } else {
6325                 /*
6326                  * Shared page table or unmanaged page (sharept or !sharept)
6327                  */
6328                 pv_placemarker_wakeup(pmap, pte_placemark);
6329         }
6330 }
6331
6332 void
6333 pmap_pgscan(struct pmap_pgscan_info *pginfo)
6334 {
6335         struct pmap_scan_info info;
6336
6337         pginfo->offset = pginfo->beg_addr;
6338         info.pmap = pginfo->pmap;
6339         info.sva = pginfo->beg_addr;
6340         info.eva = pginfo->end_addr;
6341         info.func = pmap_pgscan_callback;
6342         info.arg = pginfo;
6343         pmap_scan(&info, 0);
6344         if (info.stop == 0)
6345                 pginfo->offset = pginfo->end_addr;
6346 }
6347
6348 /*
6349  * Wait for a placemarker that we do not own to clear.  The placemarker
6350  * in question is not necessarily set to the pindex we want, we may have
6351  * to wait on the element because we want to reserve it ourselves.
6352  *
6353  * NOTE: PM_PLACEMARK_WAKEUP sets a bit which is already set in
6354  *       PM_NOPLACEMARK, so it does not interfere with placemarks
6355  *       which have already been woken up.
6356  */
6357 static
6358 void
6359 pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark)
6360 {
6361         if (*pmark != PM_NOPLACEMARK) {
6362                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
6363                 tsleep_interlock(pmark, 0);
6364                 if (*pmark != PM_NOPLACEMARK)
6365                         tsleep(pmark, PINTERLOCKED, "pvplw", 0);
6366         }
6367 }
6368
6369 /*
6370  * Wakeup a placemarker that we own.  Replace the entry with
6371  * PM_NOPLACEMARK and issue a wakeup() if necessary.
6372  */
6373 static
6374 void
6375 pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark)
6376 {
6377         vm_pindex_t pindex;
6378
6379         pindex = atomic_swap_long(pmark, PM_NOPLACEMARK);
6380         KKASSERT(pindex != PM_NOPLACEMARK);
6381         if (pindex & PM_PLACEMARK_WAKEUP)
6382                 wakeup(pmark);
6383 }