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_disable_pse.h"
  50 #include "opt_pmap.h"
  51 #endif
  52 #include "opt_msgbuf.h"
  53
  54 #include <sys/param.h>
  55 #include <sys/kernel.h>
  56 #include <sys/proc.h>
  57 #include <sys/msgbuf.h>
  58 #include <sys/vmmeter.h>
  59 #include <sys/mman.h>
  60 #include <sys/systm.h>
  61
  62 #include <vm/vm.h>
  63 #include <vm/vm_param.h>
  64 #include <sys/sysctl.h>
  65 #include <sys/lock.h>
  66 #include <vm/vm_kern.h>
  67 #include <vm/vm_page.h>
  68 #include <vm/vm_map.h>
  69 #include <vm/vm_object.h>
  70 #include <vm/vm_extern.h>
  71 #include <vm/vm_pageout.h>
  72 #include <vm/vm_pager.h>
  73 #include <vm/vm_zone.h>
  74
  75 #include <sys/user.h>
  76 #include <sys/thread2.h>
  77 #include <sys/spinlock2.h>
  78 #include <vm/vm_page2.h>
  79
  80 #include <machine/cputypes.h>
  81 #include <machine/md_var.h>
  82 #include <machine/specialreg.h>
  83 #include <machine/smp.h>
  84 #include <machine_base/apic/apicreg.h>
  85 #include <machine/globaldata.h>
  86 #include <machine/pmap.h>
  87 #include <machine/pmap_inval.h>
  88 #include <machine/inttypes.h>
  89
  90 #include <ddb/ddb.h>
  91
  92 #define PMAP_KEEP_PDIRS
  93 #ifndef PMAP_SHPGPERPROC
  94 #define PMAP_SHPGPERPROC 2000
  95 #endif
  96
  97 #if defined(DIAGNOSTIC)
  98 #define PMAP_DIAGNOSTIC
  99 #endif
 100
 101 #define MINPV 2048
 102
 103 /*
 104  * pmap debugging will report who owns a pv lock when blocking.
 105  */
 106 #ifdef PMAP_DEBUG
 107
 108 #define PMAP_DEBUG_DECL         ,const char *func, int lineno
 109 #define PMAP_DEBUG_ARGS         , __func__, __LINE__
 110 #define PMAP_DEBUG_COPY         , func, lineno
 111
 112 #define pv_get(pmap, pindex, pmarkp)    _pv_get(pmap, pindex, pmarkp    \
 113                                                         PMAP_DEBUG_ARGS)
 114 #define pv_lock(pv)                     _pv_lock(pv                     \
 115                                                         PMAP_DEBUG_ARGS)
 116 #define pv_hold_try(pv)                 _pv_hold_try(pv                 \
 117                                                         PMAP_DEBUG_ARGS)
 118 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp  \
 119                                                         PMAP_DEBUG_ARGS)
 120
 121 #define pv_free(pv, pvp)                _pv_free(pv, pvp PMAP_DEBUG_ARGS)
 122
 123 #else
 124
 125 #define PMAP_DEBUG_DECL
 126 #define PMAP_DEBUG_ARGS
 127 #define PMAP_DEBUG_COPY
 128
 129 #define pv_get(pmap, pindex, pmarkp)            _pv_get(pmap, pindex, pmarkp)
 130 #define pv_lock(pv)                     _pv_lock(pv)
 131 #define pv_hold_try(pv)                 _pv_hold_try(pv)
 132 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp)
 133 #define pv_free(pv, pvp)                _pv_free(pv, pvp)
 134
 135 #endif
 136
 137 /*
 138  * Get PDEs and PTEs for user/kernel address space
 139  */
 140 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
 141
 142 #define pmap_pde_v(pmap, pte)           ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 143 #define pmap_pte_w(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
 144 #define pmap_pte_m(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
 145 #define pmap_pte_u(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
 146 #define pmap_pte_v(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 147
 148 /*
 149  * Given a map and a machine independent protection code,
 150  * convert to a vax protection code.
 151  */
 152 #define pte_prot(m, p)          \
 153         (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
 154 static uint64_t protection_codes[PROTECTION_CODES_SIZE];
 155
 156 struct pmap kernel_pmap;
 157
 158 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
 159
 160 vm_paddr_t avail_start;         /* PA of first available physical page */
 161 vm_paddr_t avail_end;           /* PA of last available physical page */
 162 vm_offset_t virtual2_start;     /* cutout free area prior to kernel start */
 163 vm_offset_t virtual2_end;
 164 vm_offset_t virtual_start;      /* VA of first avail page (after kernel bss) */
 165 vm_offset_t virtual_end;        /* VA of last avail page (end of kernel AS) */
 166 vm_offset_t KvaStart;           /* VA start of KVA space */
 167 vm_offset_t KvaEnd;             /* VA end of KVA space (non-inclusive) */
 168 vm_offset_t KvaSize;            /* max size of kernel virtual address space */
 169 static boolean_t pmap_initialized = FALSE;      /* Has pmap_init completed? */
 170 //static int pgeflag;           /* PG_G or-in */
 171 //static int pseflag;           /* PG_PS or-in */
 172 uint64_t PatMsr;
 173
 174 static int ndmpdp;
 175 static vm_paddr_t dmaplimit;
 176 static int nkpt;
 177 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
 178
 179 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE];        /* PAT -> PG_ bits */
 180 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/    /* PAT -> PG_ bits */
 181
 182 static uint64_t KPTbase;
 183 static uint64_t KPTphys;
 184 static uint64_t KPDphys;        /* phys addr of kernel level 2 */
 185 static uint64_t KPDbase;        /* phys addr of kernel level 2 @ KERNBASE */
 186 uint64_t KPDPphys;      /* phys addr of kernel level 3 */
 187 uint64_t KPML4phys;     /* phys addr of kernel level 4 */
 188
 189 static uint64_t DMPDphys;       /* phys addr of direct mapped level 2 */
 190 static uint64_t DMPDPphys;      /* phys addr of direct mapped level 3 */
 191
 192 /*
 193  * Data for the pv entry allocation mechanism
 194  */
 195 static vm_zone_t pvzone;
 196 static struct vm_zone pvzone_store;
 197 static int pv_entry_max=0, pv_entry_high_water=0;
 198 static int pmap_pagedaemon_waken = 0;
 199 static struct pv_entry *pvinit;
 200
 201 /*
 202  * All those kernel PT submaps that BSD is so fond of
 203  */
 204 pt_entry_t *CMAP1 = NULL, *ptmmap;
 205 caddr_t CADDR1 = NULL, ptvmmap = NULL;
 206 static pt_entry_t *msgbufmap;
 207 struct msgbuf *msgbufp=NULL;
 208
 209 /*
 210  * PMAP default PG_* bits. Needed to be able to add
 211  * EPT/NPT pagetable pmap_bits for the VMM module
 212  */
 213 uint64_t pmap_bits_default[] = {
 214                 REGULAR_PMAP,                   /* TYPE_IDX             0 */
 215                 X86_PG_V,                       /* PG_V_IDX             1 */
 216                 X86_PG_RW,                      /* PG_RW_IDX            2 */
 217                 X86_PG_U,                       /* PG_U_IDX             3 */
 218                 X86_PG_A,                       /* PG_A_IDX             4 */
 219                 X86_PG_M,                       /* PG_M_IDX             5 */
 220                 X86_PG_PS,                      /* PG_PS_IDX3           6 */
 221                 X86_PG_G,                       /* PG_G_IDX             7 */
 222                 X86_PG_AVAIL1,                  /* PG_AVAIL1_IDX        8 */
 223                 X86_PG_AVAIL2,                  /* PG_AVAIL2_IDX        9 */
 224                 X86_PG_AVAIL3,                  /* PG_AVAIL3_IDX        10 */
 225                 X86_PG_NC_PWT | X86_PG_NC_PCD,  /* PG_N_IDX             11 */
 226                 X86_PG_NX,                      /* PG_NX_IDX            12 */
 227 };
 228 /*
 229  * Crashdump maps.
 230  */
 231 static pt_entry_t *pt_crashdumpmap;
 232 static caddr_t crashdumpmap;
 233
 234 static int pmap_debug = 0;
 235 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
 236     &pmap_debug, 0, "Debug pmap's");
 237 #ifdef PMAP_DEBUG2
 238 static int pmap_enter_debug = 0;
 239 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
 240     &pmap_enter_debug, 0, "Debug pmap_enter's");
 241 #endif
 242 static int pmap_yield_count = 64;
 243 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
 244     &pmap_yield_count, 0, "Yield during init_pt/release");
 245 static int pmap_mmu_optimize = 0;
 246 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
 247     &pmap_mmu_optimize, 0, "Share page table pages when possible");
 248 int pmap_fast_kernel_cpusync = 0;
 249 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
 250     &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
 251 int pmap_dynamic_delete = 0;
 252 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
 253     &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
 254 int pmap_lock_delay = 100;
 255 SYSCTL_INT(_machdep, OID_AUTO, pmap_lock_delay, CTLFLAG_RW,
 256     &pmap_lock_delay, 0, "Spin loops");
 257
 258 static int pmap_nx_enable = 0;
 259 /* needs manual TUNABLE in early probe, see below */
 260
 261 #define DISABLE_PSE
 262
 263 /* Standard user access funtions */
 264 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
 265     size_t *lencopied);
 266 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
 267 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
 268 extern int std_fubyte (const uint8_t *base);
 269 extern int std_subyte (uint8_t *base, uint8_t byte);
 270 extern int32_t std_fuword32 (const uint32_t *base);
 271 extern int64_t std_fuword64 (const uint64_t *base);
 272 extern int std_suword64 (uint64_t *base, uint64_t word);
 273 extern int std_suword32 (uint32_t *base, int word);
 274 extern uint32_t std_swapu32 (volatile uint32_t *base, uint32_t v);
 275 extern uint64_t std_swapu64 (volatile uint64_t *base, uint64_t v);
 276
 277 static void pv_hold(pv_entry_t pv);
 278 static int _pv_hold_try(pv_entry_t pv
 279                                 PMAP_DEBUG_DECL);
 280 static void pv_drop(pv_entry_t pv);
 281 static void _pv_lock(pv_entry_t pv
 282                                 PMAP_DEBUG_DECL);
 283 static void pv_unlock(pv_entry_t pv);
 284 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
 285                                 PMAP_DEBUG_DECL);
 286 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp
 287                                 PMAP_DEBUG_DECL);
 288 static void _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL);
 289 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex,
 290                                 vm_pindex_t **pmarkp, int *errorp);
 291 static void pv_put(pv_entry_t pv);
 292 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
 293 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
 294                       pv_entry_t *pvpp);
 295 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
 296                       pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
 297 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
 298                         pmap_inval_bulk_t *bulk, int destroy);
 299 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
 300 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
 301                         pmap_inval_bulk_t *bulk);
 302
 303 struct pmap_scan_info;
 304 static void pmap_remove_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 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
 309                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
 310                       pv_entry_t pt_pv, int sharept,
 311                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 312
 313 static void i386_protection_init (void);
 314 static void create_pagetables(vm_paddr_t *firstaddr);
 315 static void pmap_remove_all (vm_page_t m);
 316 static boolean_t pmap_testbit (vm_page_t m, int bit);
 317
 318 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
 319 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
 320
 321 static void pmap_pinit_defaults(struct pmap *pmap);
 322 static void pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark);
 323 static void pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark);
 324
 325 static unsigned pdir4mb;
 326
 327 static int
 328 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
 329 {
 330         if (pv1->pv_pindex < pv2->pv_pindex)
 331                 return(-1);
 332         if (pv1->pv_pindex > pv2->pv_pindex)
 333                 return(1);
 334         return(0);
 335 }
 336
 337 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
 338              pv_entry_compare, vm_pindex_t, pv_pindex);
 339
 340 static __inline
 341 void
 342 pmap_page_stats_adding(vm_page_t m)
 343 {
 344         globaldata_t gd = mycpu;
 345
 346         if (TAILQ_EMPTY(&m->md.pv_list)) {
 347                 ++gd->gd_vmtotal.t_arm;
 348         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 349                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 350                 ++gd->gd_vmtotal.t_armshr;
 351                 ++gd->gd_vmtotal.t_avmshr;
 352         } else {
 353                 ++gd->gd_vmtotal.t_avmshr;
 354         }
 355 }
 356
 357 static __inline
 358 void
 359 pmap_page_stats_deleting(vm_page_t m)
 360 {
 361         globaldata_t gd = mycpu;
 362
 363         if (TAILQ_EMPTY(&m->md.pv_list)) {
 364                 --gd->gd_vmtotal.t_arm;
 365         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 366                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 367                 --gd->gd_vmtotal.t_armshr;
 368                 --gd->gd_vmtotal.t_avmshr;
 369         } else {
 370                 --gd->gd_vmtotal.t_avmshr;
 371         }
 372 }
 373
 374 /*
 375  * Move the kernel virtual free pointer to the next
 376  * 2MB.  This is used to help improve performance
 377  * by using a large (2MB) page for much of the kernel
 378  * (.text, .data, .bss)
 379  */
 380 static
 381 vm_offset_t
 382 pmap_kmem_choose(vm_offset_t addr)
 383 {
 384         vm_offset_t newaddr = addr;
 385
 386         newaddr = roundup2(addr, NBPDR);
 387         return newaddr;
 388 }
 389
 390 /*
 391  * Returns the pindex of a page table entry (representing a terminal page).
 392  * There are NUPTE_TOTAL page table entries possible (a huge number)
 393  *
 394  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
 395  * We want to properly translate negative KVAs.
 396  */
 397 static __inline
 398 vm_pindex_t
 399 pmap_pte_pindex(vm_offset_t va)
 400 {
 401         return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
 402 }
 403
 404 /*
 405  * Returns the pindex of a page table.
 406  */
 407 static __inline
 408 vm_pindex_t
 409 pmap_pt_pindex(vm_offset_t va)
 410 {
 411         return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
 412 }
 413
 414 /*
 415  * Returns the pindex of a page directory.
 416  */
 417 static __inline
 418 vm_pindex_t
 419 pmap_pd_pindex(vm_offset_t va)
 420 {
 421         return (NUPTE_TOTAL + NUPT_TOTAL +
 422                 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
 423 }
 424
 425 static __inline
 426 vm_pindex_t
 427 pmap_pdp_pindex(vm_offset_t va)
 428 {
 429         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
 430                 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
 431 }
 432
 433 static __inline
 434 vm_pindex_t
 435 pmap_pml4_pindex(void)
 436 {
 437         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
 438 }
 439
 440 /*
 441  * Return various clipped indexes for a given VA
 442  *
 443  * Returns the index of a pt in a page directory, representing a page
 444  * table.
 445  */
 446 static __inline
 447 vm_pindex_t
 448 pmap_pt_index(vm_offset_t va)
 449 {
 450         return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
 451 }
 452
 453 /*
 454  * Returns the index of a pd in a page directory page, representing a page
 455  * directory.
 456  */
 457 static __inline
 458 vm_pindex_t
 459 pmap_pd_index(vm_offset_t va)
 460 {
 461         return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
 462 }
 463
 464 /*
 465  * Returns the index of a pdp in the pml4 table, representing a page
 466  * directory page.
 467  */
 468 static __inline
 469 vm_pindex_t
 470 pmap_pdp_index(vm_offset_t va)
 471 {
 472         return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
 473 }
 474
 475 /*
 476  * Locate the requested pt_entry
 477  */
 478 static __inline
 479 pv_entry_t
 480 pv_entry_lookup(pmap_t pmap, vm_pindex_t pindex)
 481 {
 482         pv_entry_t pv;
 483
 484         if (pindex < pmap_pt_pindex(0))
 485                 pv = pmap->pm_pvhint_pte;
 486         else if (pindex < pmap_pd_pindex(0))
 487                 pv = pmap->pm_pvhint_pt;
 488         else
 489                 pv = NULL;
 490         cpu_ccfence();
 491         if (pv == NULL || pv->pv_pmap != pmap) {
 492                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
 493                                                 pindex);
 494         } else if (pv->pv_pindex != pindex) {
 495                 pv = pv_entry_rb_tree_RB_LOOKUP_REL(&pmap->pm_pvroot,
 496                                                     pindex, pv);
 497         }
 498         return pv;
 499 }
 500
 501 /*
 502  * pmap_pte_quick:
 503  *
 504  *      Super fast pmap_pte routine best used when scanning the pv lists.
 505  *      This eliminates many course-grained invltlb calls.  Note that many of
 506  *      the pv list scans are across different pmaps and it is very wasteful
 507  *      to do an entire invltlb when checking a single mapping.
 508  */
 509 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
 510
 511 static
 512 pt_entry_t *
 513 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
 514 {
 515         return pmap_pte(pmap, va);
 516 }
 517
 518 /*
 519  * The placemarker hash must be broken up into four zones so lock
 520  * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
 521  *
 522  * Placemarkers are used to 'lock' page table indices that do not have
 523  * a pv_entry.  This allows the pmap to support managed and unmanaged
 524  * pages and shared page tables.
 525  */
 526 #define PM_PLACE_BASE   (PM_PLACEMARKS >> 2)
 527
 528 static __inline
 529 vm_pindex_t *
 530 pmap_placemarker_hash(pmap_t pmap, vm_pindex_t pindex)
 531 {
 532         int hi;
 533
 534         if (pindex < pmap_pt_pindex(0))         /* zone 0 - PTE */
 535                 hi = 0;
 536         else if (pindex < pmap_pd_pindex(0))    /* zone 1 - PT */
 537                 hi = PM_PLACE_BASE;
 538         else if (pindex < pmap_pdp_pindex(0))   /* zone 2 - PD */
 539                 hi = PM_PLACE_BASE << 1;
 540         else                                    /* zone 3 - PDP (and PML4E) */
 541                 hi = PM_PLACE_BASE | (PM_PLACE_BASE << 1);
 542         hi += pindex & (PM_PLACE_BASE - 1);
 543
 544         return (&pmap->pm_placemarks[hi]);
 545 }
 546
 547
 548 /*
 549  * Generic procedure to index a pte from a pt, pd, or pdp.
 550  *
 551  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
 552  *       a page table page index but is instead of PV lookup index.
 553  */
 554 static
 555 void *
 556 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
 557 {
 558         pt_entry_t *pte;
 559
 560         pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
 561         return(&pte[pindex]);
 562 }
 563
 564 /*
 565  * Return pointer to PDP slot in the PML4
 566  */
 567 static __inline
 568 pml4_entry_t *
 569 pmap_pdp(pmap_t pmap, vm_offset_t va)
 570 {
 571         return (&pmap->pm_pml4[pmap_pdp_index(va)]);
 572 }
 573
 574 /*
 575  * Return pointer to PD slot in the PDP given a pointer to the PDP
 576  */
 577 static __inline
 578 pdp_entry_t *
 579 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
 580 {
 581         pdp_entry_t *pd;
 582
 583         pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
 584         return (&pd[pmap_pd_index(va)]);
 585 }
 586
 587 /*
 588  * Return pointer to PD slot in the PDP.
 589  */
 590 static __inline
 591 pdp_entry_t *
 592 pmap_pd(pmap_t pmap, vm_offset_t va)
 593 {
 594         pml4_entry_t *pdp;
 595
 596         pdp = pmap_pdp(pmap, va);
 597         if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
 598                 return NULL;
 599         return (pmap_pdp_to_pd(*pdp, va));
 600 }
 601
 602 /*
 603  * Return pointer to PT slot in the PD given a pointer to the PD
 604  */
 605 static __inline
 606 pd_entry_t *
 607 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
 608 {
 609         pd_entry_t *pt;
 610
 611         pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
 612         return (&pt[pmap_pt_index(va)]);
 613 }
 614
 615 /*
 616  * Return pointer to PT slot in the PD
 617  *
 618  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
 619  *                   so we cannot lookup the PD via the PDP.  Instead we
 620  *                   must look it up via the pmap.
 621  */
 622 static __inline
 623 pd_entry_t *
 624 pmap_pt(pmap_t pmap, vm_offset_t va)
 625 {
 626         pdp_entry_t *pd;
 627         pv_entry_t pv;
 628         vm_pindex_t pd_pindex;
 629         vm_paddr_t phys;
 630
 631         if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
 632                 pd_pindex = pmap_pd_pindex(va);
 633                 spin_lock_shared(&pmap->pm_spin);
 634                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
 635                 if (pv == NULL || pv->pv_m == NULL) {
 636                         spin_unlock_shared(&pmap->pm_spin);
 637                         return NULL;
 638                 }
 639                 phys = VM_PAGE_TO_PHYS(pv->pv_m);
 640                 spin_unlock_shared(&pmap->pm_spin);
 641                 return (pmap_pd_to_pt(phys, va));
 642         } else {
 643                 pd = pmap_pd(pmap, va);
 644                 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
 645                          return NULL;
 646                 return (pmap_pd_to_pt(*pd, va));
 647         }
 648 }
 649
 650 /*
 651  * Return pointer to PTE slot in the PT given a pointer to the PT
 652  */
 653 static __inline
 654 pt_entry_t *
 655 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
 656 {
 657         pt_entry_t *pte;
 658
 659         pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
 660         return (&pte[pmap_pte_index(va)]);
 661 }
 662
 663 /*
 664  * Return pointer to PTE slot in the PT
 665  */
 666 static __inline
 667 pt_entry_t *
 668 pmap_pte(pmap_t pmap, vm_offset_t va)
 669 {
 670         pd_entry_t *pt;
 671
 672         pt = pmap_pt(pmap, va);
 673         if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
 674                  return NULL;
 675         if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
 676                 return ((pt_entry_t *)pt);
 677         return (pmap_pt_to_pte(*pt, va));
 678 }
 679
 680 /*
 681  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
 682  * the PT layer.  This will speed up core pmap operations considerably.
 683  *
 684  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
 685  *       must be in a known associated state (typically by being locked when
 686  *       the pmap spinlock isn't held).  We allow the race for that case.
 687  *
 688  * NOTE: pm_pvhint* is only accessed (read) with the spin-lock held, using
 689  *       cpu_ccfence() to prevent compiler optimizations from reloading the
 690  *       field.
 691  */
 692 static __inline
 693 void
 694 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
 695 {
 696         if (pindex < pmap_pt_pindex(0)) {
 697                 if (pv->pv_pmap)
 698                         pv->pv_pmap->pm_pvhint_pte = pv;
 699         } else if (pindex < pmap_pd_pindex(0)) {
 700                 if (pv->pv_pmap)
 701                         pv->pv_pmap->pm_pvhint_pt = pv;
 702         }
 703 }
 704
 705
 706 /*
 707  * Return address of PT slot in PD (KVM only)
 708  *
 709  * Cannot be used for user page tables because it might interfere with
 710  * the shared page-table-page optimization (pmap_mmu_optimize).
 711  */
 712 static __inline
 713 pd_entry_t *
 714 vtopt(vm_offset_t va)
 715 {
 716         uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
 717                                   NPML4EPGSHIFT)) - 1);
 718
 719         return (PDmap + ((va >> PDRSHIFT) & mask));
 720 }
 721
 722 /*
 723  * KVM - return address of PTE slot in PT
 724  */
 725 static __inline
 726 pt_entry_t *
 727 vtopte(vm_offset_t va)
 728 {
 729         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 730                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 731
 732         return (PTmap + ((va >> PAGE_SHIFT) & mask));
 733 }
 734
 735 /*
 736  * Returns the physical address translation from va for a user address.
 737  * (vm_paddr_t)-1 is returned on failure.
 738  */
 739 vm_paddr_t
 740 uservtophys(vm_offset_t va)
 741 {
 742         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 743                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 744         vm_paddr_t pa;
 745         pt_entry_t pte;
 746         pmap_t pmap;
 747
 748         pmap = vmspace_pmap(mycpu->gd_curthread->td_lwp->lwp_vmspace);
 749         pa = (vm_paddr_t)-1;
 750         if (va < VM_MAX_USER_ADDRESS) {
 751                 pte = kreadmem64(PTmap + ((va >> PAGE_SHIFT) & mask));
 752                 if (pte & pmap->pmap_bits[PG_V_IDX])
 753                         pa = (pte & PG_FRAME) | (va & PAGE_MASK);
 754         }
 755         return pa;
 756 }
 757
 758 static uint64_t
 759 allocpages(vm_paddr_t *firstaddr, long n)
 760 {
 761         uint64_t ret;
 762
 763         ret = *firstaddr;
 764         bzero((void *)ret, n * PAGE_SIZE);
 765         *firstaddr += n * PAGE_SIZE;
 766         return (ret);
 767 }
 768
 769 static
 770 void
 771 create_pagetables(vm_paddr_t *firstaddr)
 772 {
 773         long i;         /* must be 64 bits */
 774         long nkpt_base;
 775         long nkpt_phys;
 776         int j;
 777
 778         /*
 779          * We are running (mostly) V=P at this point
 780          *
 781          * Calculate NKPT - number of kernel page tables.  We have to
 782          * accomodoate prealloction of the vm_page_array, dump bitmap,
 783          * MSGBUF_SIZE, and other stuff.  Be generous.
 784          *
 785          * Maxmem is in pages.
 786          *
 787          * ndmpdp is the number of 1GB pages we wish to map.
 788          */
 789         ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
 790         if (ndmpdp < 4)         /* Minimum 4GB of dirmap */
 791                 ndmpdp = 4;
 792         KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
 793
 794         /*
 795          * Starting at the beginning of kvm (not KERNBASE).
 796          */
 797         nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
 798         nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
 799         nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
 800                        ndmpdp) + 511) / 512;
 801         nkpt_phys += 128;
 802
 803         /*
 804          * Starting at KERNBASE - map 2G worth of page table pages.
 805          * KERNBASE is offset -2G from the end of kvm.
 806          */
 807         nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
 808
 809         /*
 810          * Allocate pages
 811          */
 812         KPTbase = allocpages(firstaddr, nkpt_base);
 813         KPTphys = allocpages(firstaddr, nkpt_phys);
 814         KPML4phys = allocpages(firstaddr, 1);
 815         KPDPphys = allocpages(firstaddr, NKPML4E);
 816         KPDphys = allocpages(firstaddr, NKPDPE);
 817
 818         /*
 819          * Calculate the page directory base for KERNBASE,
 820          * that is where we start populating the page table pages.
 821          * Basically this is the end - 2.
 822          */
 823         KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
 824
 825         DMPDPphys = allocpages(firstaddr, NDMPML4E);
 826         if ((amd_feature & AMDID_PAGE1GB) == 0)
 827                 DMPDphys = allocpages(firstaddr, ndmpdp);
 828         dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
 829
 830         /*
 831          * Fill in the underlying page table pages for the area around
 832          * KERNBASE.  This remaps low physical memory to KERNBASE.
 833          *
 834          * Read-only from zero to physfree
 835          * XXX not fully used, underneath 2M pages
 836          */
 837         for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
 838                 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
 839                 ((pt_entry_t *)KPTbase)[i] |=
 840                     pmap_bits_default[PG_RW_IDX] |
 841                     pmap_bits_default[PG_V_IDX] |
 842                     pmap_bits_default[PG_G_IDX];
 843         }
 844
 845         /*
 846          * Now map the initial kernel page tables.  One block of page
 847          * tables is placed at the beginning of kernel virtual memory,
 848          * and another block is placed at KERNBASE to map the kernel binary,
 849          * data, bss, and initial pre-allocations.
 850          */
 851         for (i = 0; i < nkpt_base; i++) {
 852                 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
 853                 ((pd_entry_t *)KPDbase)[i] |=
 854                     pmap_bits_default[PG_RW_IDX] |
 855                     pmap_bits_default[PG_V_IDX];
 856         }
 857         for (i = 0; i < nkpt_phys; i++) {
 858                 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
 859                 ((pd_entry_t *)KPDphys)[i] |=
 860                     pmap_bits_default[PG_RW_IDX] |
 861                     pmap_bits_default[PG_V_IDX];
 862         }
 863
 864         /*
 865          * Map from zero to end of allocations using 2M pages as an
 866          * optimization.  This will bypass some of the KPTBase pages
 867          * above in the KERNBASE area.
 868          */
 869         for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
 870                 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
 871                 ((pd_entry_t *)KPDbase)[i] |=
 872                     pmap_bits_default[PG_RW_IDX] |
 873                     pmap_bits_default[PG_V_IDX] |
 874                     pmap_bits_default[PG_PS_IDX] |
 875                     pmap_bits_default[PG_G_IDX];
 876         }
 877
 878         /*
 879          * And connect up the PD to the PDP.  The kernel pmap is expected
 880          * to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
 881          */
 882         for (i = 0; i < NKPDPE; i++) {
 883                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
 884                                 KPDphys + (i << PAGE_SHIFT);
 885                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
 886                     pmap_bits_default[PG_RW_IDX] |
 887                     pmap_bits_default[PG_V_IDX] |
 888                     pmap_bits_default[PG_U_IDX];
 889         }
 890
 891         /*
 892          * Now set up the direct map space using either 2MB or 1GB pages
 893          * Preset PG_M and PG_A because demotion expects it.
 894          *
 895          * When filling in entries in the PD pages make sure any excess
 896          * entries are set to zero as we allocated enough PD pages
 897          */
 898         if ((amd_feature & AMDID_PAGE1GB) == 0) {
 899                 for (i = 0; i < NPDEPG * ndmpdp; i++) {
 900                         ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
 901                         ((pd_entry_t *)DMPDphys)[i] |=
 902                             pmap_bits_default[PG_RW_IDX] |
 903                             pmap_bits_default[PG_V_IDX] |
 904                             pmap_bits_default[PG_PS_IDX] |
 905                             pmap_bits_default[PG_G_IDX] |
 906                             pmap_bits_default[PG_M_IDX] |
 907                             pmap_bits_default[PG_A_IDX];
 908                 }
 909
 910                 /*
 911                  * And the direct map space's PDP
 912                  */
 913                 for (i = 0; i < ndmpdp; i++) {
 914                         ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
 915                                                         (i << PAGE_SHIFT);
 916                         ((pdp_entry_t *)DMPDPphys)[i] |=
 917                             pmap_bits_default[PG_RW_IDX] |
 918                             pmap_bits_default[PG_V_IDX] |
 919                             pmap_bits_default[PG_U_IDX];
 920                 }
 921         } else {
 922                 for (i = 0; i < ndmpdp; i++) {
 923                         ((pdp_entry_t *)DMPDPphys)[i] =
 924                                                 (vm_paddr_t)i << PDPSHIFT;
 925                         ((pdp_entry_t *)DMPDPphys)[i] |=
 926                             pmap_bits_default[PG_RW_IDX] |
 927                             pmap_bits_default[PG_V_IDX] |
 928                             pmap_bits_default[PG_PS_IDX] |
 929                             pmap_bits_default[PG_G_IDX] |
 930                             pmap_bits_default[PG_M_IDX] |
 931                             pmap_bits_default[PG_A_IDX];
 932                 }
 933         }
 934
 935         /* And recursively map PML4 to itself in order to get PTmap */
 936         ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
 937         ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
 938             pmap_bits_default[PG_RW_IDX] |
 939             pmap_bits_default[PG_V_IDX] |
 940             pmap_bits_default[PG_U_IDX];
 941
 942         /*
 943          * Connect the Direct Map slots up to the PML4
 944          */
 945         for (j = 0; j < NDMPML4E; ++j) {
 946                 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
 947                     (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
 948                     pmap_bits_default[PG_RW_IDX] |
 949                     pmap_bits_default[PG_V_IDX] |
 950                     pmap_bits_default[PG_U_IDX];
 951         }
 952
 953         /*
 954          * Connect the KVA slot up to the PML4
 955          */
 956         ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
 957         ((pdp_entry_t *)KPML4phys)[KPML4I] |=
 958             pmap_bits_default[PG_RW_IDX] |
 959             pmap_bits_default[PG_V_IDX] |
 960             pmap_bits_default[PG_U_IDX];
 961 }
 962
 963 /*
 964  *      Bootstrap the system enough to run with virtual memory.
 965  *
 966  *      On the i386 this is called after mapping has already been enabled
 967  *      and just syncs the pmap module with what has already been done.
 968  *      [We can't call it easily with mapping off since the kernel is not
 969  *      mapped with PA == VA, hence we would have to relocate every address
 970  *      from the linked base (virtual) address "KERNBASE" to the actual
 971  *      (physical) address starting relative to 0]
 972  */
 973 void
 974 pmap_bootstrap(vm_paddr_t *firstaddr)
 975 {
 976         vm_offset_t va;
 977         pt_entry_t *pte;
 978         int i;
 979
 980         KvaStart = VM_MIN_KERNEL_ADDRESS;
 981         KvaEnd = VM_MAX_KERNEL_ADDRESS;
 982         KvaSize = KvaEnd - KvaStart;
 983
 984         avail_start = *firstaddr;
 985
 986         /*
 987          * Create an initial set of page tables to run the kernel in.
 988          */
 989         create_pagetables(firstaddr);
 990
 991         virtual2_start = KvaStart;
 992         virtual2_end = PTOV_OFFSET;
 993
 994         virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
 995         virtual_start = pmap_kmem_choose(virtual_start);
 996
 997         virtual_end = VM_MAX_KERNEL_ADDRESS;
 998
 999         /* XXX do %cr0 as well */
1000         load_cr4(rcr4() | CR4_PGE | CR4_PSE);
1001         load_cr3(KPML4phys);
1002
1003         /*
1004          * Initialize protection array.
1005          */
1006         i386_protection_init();
1007
1008         /*
1009          * The kernel's pmap is statically allocated so we don't have to use
1010          * pmap_create, which is unlikely to work correctly at this part of
1011          * the boot sequence (XXX and which no longer exists).
1012          */
1013         kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
1014         kernel_pmap.pm_count = 1;
1015         CPUMASK_ASSALLONES(kernel_pmap.pm_active);
1016         RB_INIT(&kernel_pmap.pm_pvroot);
1017         spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
1018         for (i = 0; i < PM_PLACEMARKS; ++i)
1019                 kernel_pmap.pm_placemarks[i] = PM_NOPLACEMARK;
1020
1021         /*
1022          * Reserve some special page table entries/VA space for temporary
1023          * mapping of pages.
1024          */
1025 #define SYSMAP(c, p, v, n)      \
1026         v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
1027
1028         va = virtual_start;
1029         pte = vtopte(va);
1030
1031         /*
1032          * CMAP1/CMAP2 are used for zeroing and copying pages.
1033          */
1034         SYSMAP(caddr_t, CMAP1, CADDR1, 1)
1035
1036         /*
1037          * Crashdump maps.
1038          */
1039         SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
1040
1041         /*
1042          * ptvmmap is used for reading arbitrary physical pages via
1043          * /dev/mem.
1044          */
1045         SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
1046
1047         /*
1048          * msgbufp is used to map the system message buffer.
1049          * XXX msgbufmap is not used.
1050          */
1051         SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
1052                atop(round_page(MSGBUF_SIZE)))
1053
1054         virtual_start = va;
1055         virtual_start = pmap_kmem_choose(virtual_start);
1056
1057         *CMAP1 = 0;
1058
1059         /*
1060          * PG_G is terribly broken on SMP because we IPI invltlb's in some
1061          * cases rather then invl1pg.  Actually, I don't even know why it
1062          * works under UP because self-referential page table mappings
1063          */
1064 //      pgeflag = 0;
1065
1066 /*
1067  * Initialize the 4MB page size flag
1068  */
1069 //      pseflag = 0;
1070 /*
1071  * The 4MB page version of the initial
1072  * kernel page mapping.
1073  */
1074         pdir4mb = 0;
1075
1076 #if !defined(DISABLE_PSE)
1077         if (cpu_feature & CPUID_PSE) {
1078                 pt_entry_t ptditmp;
1079                 /*
1080                  * Note that we have enabled PSE mode
1081                  */
1082 //              pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
1083                 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
1084                 ptditmp &= ~(NBPDR - 1);
1085                 ptditmp |= pmap_bits_default[PG_V_IDX] |
1086                     pmap_bits_default[PG_RW_IDX] |
1087                     pmap_bits_default[PG_PS_IDX] |
1088                     pmap_bits_default[PG_U_IDX];
1089 //                  pgeflag;
1090                 pdir4mb = ptditmp;
1091         }
1092 #endif
1093         cpu_invltlb();
1094
1095         /* Initialize the PAT MSR */
1096         pmap_init_pat();
1097         pmap_pinit_defaults(&kernel_pmap);
1098
1099         TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1100                           &pmap_fast_kernel_cpusync);
1101
1102 }
1103
1104 /*
1105  * Setup the PAT MSR.
1106  */
1107 void
1108 pmap_init_pat(void)
1109 {
1110         uint64_t pat_msr;
1111         u_long cr0, cr4;
1112
1113         /*
1114          * Default values mapping PATi,PCD,PWT bits at system reset.
1115          * The default values effectively ignore the PATi bit by
1116          * repeating the encodings for 0-3 in 4-7, and map the PCD
1117          * and PWT bit combinations to the expected PAT types.
1118          */
1119         pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |        /* 000 */
1120                   PAT_VALUE(1, PAT_WRITE_THROUGH) |     /* 001 */
1121                   PAT_VALUE(2, PAT_UNCACHED) |          /* 010 */
1122                   PAT_VALUE(3, PAT_UNCACHEABLE) |       /* 011 */
1123                   PAT_VALUE(4, PAT_WRITE_BACK) |        /* 100 */
1124                   PAT_VALUE(5, PAT_WRITE_THROUGH) |     /* 101 */
1125                   PAT_VALUE(6, PAT_UNCACHED) |          /* 110 */
1126                   PAT_VALUE(7, PAT_UNCACHEABLE);        /* 111 */
1127         pat_pte_index[PAT_WRITE_BACK]   = 0;
1128         pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
1129         pat_pte_index[PAT_UNCACHED]     = X86_PG_NC_PCD;
1130         pat_pte_index[PAT_UNCACHEABLE]  = X86_PG_NC_PCD | X86_PG_NC_PWT;
1131         pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1132         pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1133
1134         if (cpu_feature & CPUID_PAT) {
1135                 /*
1136                  * If we support the PAT then set-up entries for
1137                  * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1138                  * 5 and 6.
1139                  */
1140                 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1141                           PAT_VALUE(5, PAT_WRITE_PROTECTED);
1142                 pat_msr = (pat_msr & ~PAT_MASK(6)) |
1143                           PAT_VALUE(6, PAT_WRITE_COMBINING);
1144                 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1145                 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PCD;
1146
1147                 /*
1148                  * Then enable the PAT
1149                  */
1150
1151                 /* Disable PGE. */
1152                 cr4 = rcr4();
1153                 load_cr4(cr4 & ~CR4_PGE);
1154
1155                 /* Disable caches (CD = 1, NW = 0). */
1156                 cr0 = rcr0();
1157                 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1158
1159                 /* Flushes caches and TLBs. */
1160                 wbinvd();
1161                 cpu_invltlb();
1162
1163                 /* Update PAT and index table. */
1164                 wrmsr(MSR_PAT, pat_msr);
1165
1166                 /* Flush caches and TLBs again. */
1167                 wbinvd();
1168                 cpu_invltlb();
1169
1170                 /* Restore caches and PGE. */
1171                 load_cr0(cr0);
1172                 load_cr4(cr4);
1173                 PatMsr = pat_msr;
1174         }
1175 }
1176
1177 /*
1178  * Set 4mb pdir for mp startup
1179  */
1180 void
1181 pmap_set_opt(void)
1182 {
1183         if (cpu_feature & CPUID_PSE) {
1184                 load_cr4(rcr4() | CR4_PSE);
1185                 if (pdir4mb && mycpu->gd_cpuid == 0) {  /* only on BSP */
1186                         cpu_invltlb();
1187                 }
1188         }
1189 }
1190
1191 /*
1192  *      Initialize the pmap module.
1193  *      Called by vm_init, to initialize any structures that the pmap
1194  *      system needs to map virtual memory.
1195  *      pmap_init has been enhanced to support in a fairly consistant
1196  *      way, discontiguous physical memory.
1197  */
1198 void
1199 pmap_init(void)
1200 {
1201         int i;
1202         int initial_pvs;
1203
1204         /*
1205          * Allocate memory for random pmap data structures.  Includes the
1206          * pv_head_table.
1207          */
1208
1209         for (i = 0; i < vm_page_array_size; i++) {
1210                 vm_page_t m;
1211
1212                 m = &vm_page_array[i];
1213                 TAILQ_INIT(&m->md.pv_list);
1214         }
1215
1216         /*
1217          * init the pv free list
1218          */
1219         initial_pvs = vm_page_array_size;
1220         if (initial_pvs < MINPV)
1221                 initial_pvs = MINPV;
1222         pvzone = &pvzone_store;
1223         pvinit = (void *)kmem_alloc(&kernel_map,
1224                                     initial_pvs * sizeof (struct pv_entry),
1225                                     VM_SUBSYS_PVENTRY);
1226         zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1227                   pvinit, initial_pvs);
1228
1229         /*
1230          * Now it is safe to enable pv_table recording.
1231          */
1232         pmap_initialized = TRUE;
1233 }
1234
1235 /*
1236  * Initialize the address space (zone) for the pv_entries.  Set a
1237  * high water mark so that the system can recover from excessive
1238  * numbers of pv entries.
1239  */
1240 void
1241 pmap_init2(void)
1242 {
1243         int shpgperproc = PMAP_SHPGPERPROC;
1244         int entry_max;
1245
1246         TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1247         pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1248         TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1249         pv_entry_high_water = 9 * (pv_entry_max / 10);
1250
1251         /*
1252          * Subtract out pages already installed in the zone (hack)
1253          */
1254         entry_max = pv_entry_max - vm_page_array_size;
1255         if (entry_max <= 0)
1256                 entry_max = 1;
1257
1258         zinitna(pvzone, NULL, 0, entry_max, ZONE_INTERRUPT);
1259
1260         /*
1261          * Enable dynamic deletion of empty higher-level page table pages
1262          * by default only if system memory is < 8GB (use 7GB for slop).
1263          * This can save a little memory, but imposes significant
1264          * performance overhead for things like bulk builds, and for programs
1265          * which do a lot of memory mapping and memory unmapping.
1266          */
1267         if (pmap_dynamic_delete < 0) {
1268                 if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1269                         pmap_dynamic_delete = 1;
1270                 else
1271                         pmap_dynamic_delete = 0;
1272         }
1273 }
1274
1275 /*
1276  * Typically used to initialize a fictitious page by vm/device_pager.c
1277  */
1278 void
1279 pmap_page_init(struct vm_page *m)
1280 {
1281         vm_page_init(m);
1282         TAILQ_INIT(&m->md.pv_list);
1283 }
1284
1285 /***************************************************
1286  * Low level helper routines.....
1287  ***************************************************/
1288
1289 /*
1290  * this routine defines the region(s) of memory that should
1291  * not be tested for the modified bit.
1292  */
1293 static __inline
1294 int
1295 pmap_track_modified(vm_pindex_t pindex)
1296 {
1297         vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1298         if ((va < clean_sva) || (va >= clean_eva))
1299                 return 1;
1300         else
1301                 return 0;
1302 }
1303
1304 /*
1305  * Extract the physical page address associated with the map/VA pair.
1306  * The page must be wired for this to work reliably.
1307  */
1308 vm_paddr_t
1309 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
1310 {
1311         vm_paddr_t rtval;
1312         pv_entry_t pt_pv;
1313         pt_entry_t *ptep;
1314
1315         rtval = 0;
1316         if (va >= VM_MAX_USER_ADDRESS) {
1317                 /*
1318                  * Kernel page directories might be direct-mapped and
1319                  * there is typically no PV tracking of pte's
1320                  */
1321                 pd_entry_t *pt;
1322
1323                 pt = pmap_pt(pmap, va);
1324                 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1325                         if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1326                                 rtval = *pt & PG_PS_FRAME;
1327                                 rtval |= va & PDRMASK;
1328                         } else {
1329                                 ptep = pmap_pt_to_pte(*pt, va);
1330                                 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1331                                         rtval = *ptep & PG_FRAME;
1332                                         rtval |= va & PAGE_MASK;
1333                                 }
1334                         }
1335                 }
1336                 if (handlep)
1337                         *handlep = NULL;
1338         } else {
1339                 /*
1340                  * User pages currently do not direct-map the page directory
1341                  * and some pages might not used managed PVs.  But all PT's
1342                  * will have a PV.
1343                  */
1344                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1345                 if (pt_pv) {
1346                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1347                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1348                                 rtval = *ptep & PG_FRAME;
1349                                 rtval |= va & PAGE_MASK;
1350                         }
1351                         if (handlep)
1352                                 *handlep = pt_pv;       /* locked until done */
1353                         else
1354                                 pv_put (pt_pv);
1355                 } else if (handlep) {
1356                         *handlep = NULL;
1357                 }
1358         }
1359         return rtval;
1360 }
1361
1362 void
1363 pmap_extract_done(void *handle)
1364 {
1365         if (handle)
1366                 pv_put((pv_entry_t)handle);
1367 }
1368
1369 /*
1370  * Similar to extract but checks protections, SMP-friendly short-cut for
1371  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1372  * fall-through to the real fault code.  Does not work with HVM page
1373  * tables.
1374  *
1375  * if busyp is NULL the returned page, if not NULL, is held (and not busied).
1376  *
1377  * If busyp is not NULL and this function sets *busyp non-zero, the returned
1378  * page is busied (and not held).
1379  *
1380  * If busyp is not NULL and this function sets *busyp to zero, the returned
1381  * page is held (and not busied).
1382  *
1383  * If VM_PROT_WRITE or VM_PROT_OVERRIDE_WRITE is set in prot, and the pte
1384  * is already writable, the returned page will be dirtied.  If the pte
1385  * is not already writable NULL is returned.  In otherwords, if either
1386  * bit is set and a vm_page_t is returned, any COW will already have happened
1387  * and that page can be written by the caller.
1388  *
1389  * WARNING! THE RETURNED PAGE IS ONLY HELD AND NOT SUITABLE FOR READING
1390  *          OR WRITING AS-IS.
1391  */
1392 vm_page_t
1393 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot, int *busyp)
1394 {
1395         if (pmap &&
1396             va < VM_MAX_USER_ADDRESS &&
1397             (pmap->pm_flags & PMAP_HVM) == 0) {
1398                 pv_entry_t pt_pv;
1399                 pv_entry_t pte_pv;
1400                 pt_entry_t *ptep;
1401                 pt_entry_t req;
1402                 vm_page_t m;
1403                 int error;
1404
1405                 req = pmap->pmap_bits[PG_V_IDX] |
1406                       pmap->pmap_bits[PG_U_IDX];
1407                 if (prot & (VM_PROT_WRITE | VM_PROT_OVERRIDE_WRITE))
1408                         req |= pmap->pmap_bits[PG_RW_IDX];
1409
1410                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1411                 if (pt_pv == NULL)
1412                         return (NULL);
1413                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1414                 if ((*ptep & req) != req) {
1415                         pv_put(pt_pv);
1416                         return (NULL);
1417                 }
1418                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), NULL, &error);
1419                 if (pte_pv && error == 0) {
1420                         m = pte_pv->pv_m;
1421                         if (prot & (VM_PROT_WRITE | VM_PROT_OVERRIDE_WRITE)) {
1422                                 /* interlocked by presence of pv_entry */
1423                                 vm_page_dirty(m);
1424                         }
1425                         if (busyp) {
1426                                 if (prot & VM_PROT_WRITE) {
1427                                         if (vm_page_busy_try(m, TRUE))
1428                                                 m = NULL;
1429                                         *busyp = 1;
1430                                 } else {
1431                                         vm_page_hold(m);
1432                                         *busyp = 0;
1433                                 }
1434                         } else {
1435                                 vm_page_hold(m);
1436                         }
1437                         pv_put(pte_pv);
1438                 } else if (pte_pv) {
1439                         pv_drop(pte_pv);
1440                         m = NULL;
1441                 } else {
1442                         /* error, since we didn't request a placemarker */
1443                         m = NULL;
1444                 }
1445                 pv_put(pt_pv);
1446                 return(m);
1447         } else {
1448                 return(NULL);
1449         }
1450 }
1451
1452 /*
1453  * Extract the physical page address associated kernel virtual address.
1454  */
1455 vm_paddr_t
1456 pmap_kextract(vm_offset_t va)
1457 {
1458         pd_entry_t pt;          /* pt entry in pd */
1459         vm_paddr_t pa;
1460
1461         if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1462                 pa = DMAP_TO_PHYS(va);
1463         } else {
1464                 pt = *vtopt(va);
1465                 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1466                         pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1467                 } else {
1468                         /*
1469                          * Beware of a concurrent promotion that changes the
1470                          * PDE at this point!  For example, vtopte() must not
1471                          * be used to access the PTE because it would use the
1472                          * new PDE.  It is, however, safe to use the old PDE
1473                          * because the page table page is preserved by the
1474                          * promotion.
1475                          */
1476                         pa = *pmap_pt_to_pte(pt, va);
1477                         pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1478                 }
1479         }
1480         return pa;
1481 }
1482
1483 /***************************************************
1484  * Low level mapping routines.....
1485  ***************************************************/
1486
1487 /*
1488  * Routine: pmap_kenter
1489  * Function:
1490  *      Add a wired page to the KVA
1491  *      NOTE! note that in order for the mapping to take effect -- you
1492  *      should do an invltlb after doing the pmap_kenter().
1493  */
1494 void
1495 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1496 {
1497         pt_entry_t *ptep;
1498         pt_entry_t npte;
1499
1500         npte = pa |
1501                kernel_pmap.pmap_bits[PG_RW_IDX] |
1502                kernel_pmap.pmap_bits[PG_V_IDX];
1503 //             pgeflag;
1504         ptep = vtopte(va);
1505 #if 1
1506         pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1507 #else
1508         /* FUTURE */
1509         if (*ptep)
1510                 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1511         else
1512                 *ptep = npte;
1513 #endif
1514 }
1515
1516 /*
1517  * Similar to pmap_kenter(), except we only invalidate the mapping on the
1518  * current CPU.  Returns 0 if the previous pte was 0, 1 if it wasn't
1519  * (caller can conditionalize calling smp_invltlb()).
1520  */
1521 int
1522 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1523 {
1524         pt_entry_t *ptep;
1525         pt_entry_t npte;
1526         int res;
1527
1528         npte = pa | kernel_pmap.pmap_bits[PG_RW_IDX] |
1529                     kernel_pmap.pmap_bits[PG_V_IDX];
1530         // npte |= pgeflag;
1531         ptep = vtopte(va);
1532 #if 1
1533         res = 1;
1534 #else
1535         /* FUTURE */
1536         res = (*ptep != 0);
1537 #endif
1538         atomic_swap_long(ptep, npte);
1539         cpu_invlpg((void *)va);
1540
1541         return res;
1542 }
1543
1544 /*
1545  * Enter addresses into the kernel pmap but don't bother
1546  * doing any tlb invalidations.  Caller will do a rollup
1547  * invalidation via pmap_rollup_inval().
1548  */
1549 int
1550 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1551 {
1552         pt_entry_t *ptep;
1553         pt_entry_t npte;
1554         int res;
1555
1556         npte = pa |
1557             kernel_pmap.pmap_bits[PG_RW_IDX] |
1558             kernel_pmap.pmap_bits[PG_V_IDX];
1559 //          pgeflag;
1560         ptep = vtopte(va);
1561 #if 1
1562         res = 1;
1563 #else
1564         /* FUTURE */
1565         res = (*ptep != 0);
1566 #endif
1567         atomic_swap_long(ptep, npte);
1568         cpu_invlpg((void *)va);
1569
1570         return res;
1571 }
1572
1573 /*
1574  * remove a page from the kernel pagetables
1575  */
1576 void
1577 pmap_kremove(vm_offset_t va)
1578 {
1579         pt_entry_t *ptep;
1580
1581         ptep = vtopte(va);
1582         pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1583 }
1584
1585 void
1586 pmap_kremove_quick(vm_offset_t va)
1587 {
1588         pt_entry_t *ptep;
1589
1590         ptep = vtopte(va);
1591         (void)pte_load_clear(ptep);
1592         cpu_invlpg((void *)va);
1593 }
1594
1595 /*
1596  * Remove addresses from the kernel pmap but don't bother
1597  * doing any tlb invalidations.  Caller will do a rollup
1598  * invalidation via pmap_rollup_inval().
1599  */
1600 void
1601 pmap_kremove_noinval(vm_offset_t va)
1602 {
1603         pt_entry_t *ptep;
1604
1605         ptep = vtopte(va);
1606         (void)pte_load_clear(ptep);
1607 }
1608
1609 /*
1610  * XXX these need to be recoded.  They are not used in any critical path.
1611  */
1612 void
1613 pmap_kmodify_rw(vm_offset_t va)
1614 {
1615         atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1616         cpu_invlpg((void *)va);
1617 }
1618
1619 /* NOT USED
1620 void
1621 pmap_kmodify_nc(vm_offset_t va)
1622 {
1623         atomic_set_long(vtopte(va), PG_N);
1624         cpu_invlpg((void *)va);
1625 }
1626 */
1627
1628 /*
1629  * Used to map a range of physical addresses into kernel virtual
1630  * address space during the low level boot, typically to map the
1631  * dump bitmap, message buffer, and vm_page_array.
1632  *
1633  * These mappings are typically made at some pointer after the end of the
1634  * kernel text+data.
1635  *
1636  * We could return PHYS_TO_DMAP(start) here and not allocate any
1637  * via (*virtp), but then kmem from userland and kernel dumps won't
1638  * have access to the related pointers.
1639  */
1640 vm_offset_t
1641 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1642 {
1643         vm_offset_t va;
1644         vm_offset_t va_start;
1645
1646         /*return PHYS_TO_DMAP(start);*/
1647
1648         va_start = *virtp;
1649         va = va_start;
1650
1651         while (start < end) {
1652                 pmap_kenter_quick(va, start);
1653                 va += PAGE_SIZE;
1654                 start += PAGE_SIZE;
1655         }
1656         *virtp = va;
1657         return va_start;
1658 }
1659
1660 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
1661
1662 /*
1663  * Remove the specified set of pages from the data and instruction caches.
1664  *
1665  * In contrast to pmap_invalidate_cache_range(), this function does not
1666  * rely on the CPU's self-snoop feature, because it is intended for use
1667  * when moving pages into a different cache domain.
1668  */
1669 void
1670 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1671 {
1672         vm_offset_t daddr, eva;
1673         int i;
1674
1675         if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1676             (cpu_feature & CPUID_CLFSH) == 0)
1677                 wbinvd();
1678         else {
1679                 cpu_mfence();
1680                 for (i = 0; i < count; i++) {
1681                         daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1682                         eva = daddr + PAGE_SIZE;
1683                         for (; daddr < eva; daddr += cpu_clflush_line_size)
1684                                 clflush(daddr);
1685                 }
1686                 cpu_mfence();
1687         }
1688 }
1689
1690 void
1691 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1692 {
1693         KASSERT((sva & PAGE_MASK) == 0,
1694             ("pmap_invalidate_cache_range: sva not page-aligned"));
1695         KASSERT((eva & PAGE_MASK) == 0,
1696             ("pmap_invalidate_cache_range: eva not page-aligned"));
1697
1698         if (cpu_feature & CPUID_SS) {
1699                 ; /* If "Self Snoop" is supported, do nothing. */
1700         } else {
1701                 /* Globally invalidate caches */
1702                 cpu_wbinvd_on_all_cpus();
1703         }
1704 }
1705
1706 /*
1707  * Invalidate the specified range of virtual memory on all cpus associated
1708  * with the pmap.
1709  */
1710 void
1711 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1712 {
1713         pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1714 }
1715
1716 /*
1717  * Add a list of wired pages to the kva.  This routine is used for temporary
1718  * kernel mappings such as those found in buffer cache buffer.  Page
1719  * modifications and accesses are not tracked or recorded.
1720  *
1721  * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1722  *       semantics as previous mappings may have been zerod without any
1723  *       invalidation.
1724  *
1725  * The page *must* be wired.
1726  */
1727 static __inline void
1728 _pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count, int doinval)
1729 {
1730         vm_offset_t end_va;
1731         vm_offset_t va;
1732
1733         end_va = beg_va + count * PAGE_SIZE;
1734
1735         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1736                 pt_entry_t pte;
1737                 pt_entry_t *ptep;
1738
1739                 ptep = vtopte(va);
1740                 pte = VM_PAGE_TO_PHYS(*m) |
1741                         kernel_pmap.pmap_bits[PG_RW_IDX] |
1742                         kernel_pmap.pmap_bits[PG_V_IDX] |
1743                         kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1744 //              pgeflag;
1745                 atomic_swap_long(ptep, pte);
1746                 m++;
1747         }
1748         if (doinval)
1749                 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1750 }
1751
1752 void
1753 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1754 {
1755         _pmap_qenter(beg_va, m, count, 1);
1756 }
1757
1758 void
1759 pmap_qenter_noinval(vm_offset_t beg_va, vm_page_t *m, int count)
1760 {
1761         _pmap_qenter(beg_va, m, count, 0);
1762 }
1763
1764 /*
1765  * This routine jerks page mappings from the kernel -- it is meant only
1766  * for temporary mappings such as those found in buffer cache buffers.
1767  * No recording modified or access status occurs.
1768  *
1769  * MPSAFE, INTERRUPT SAFE (cluster callback)
1770  */
1771 void
1772 pmap_qremove(vm_offset_t beg_va, int count)
1773 {
1774         vm_offset_t end_va;
1775         vm_offset_t va;
1776
1777         end_va = beg_va + count * PAGE_SIZE;
1778
1779         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1780                 pt_entry_t *pte;
1781
1782                 pte = vtopte(va);
1783                 (void)pte_load_clear(pte);
1784                 cpu_invlpg((void *)va);
1785         }
1786         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1787 }
1788
1789 /*
1790  * This routine removes temporary kernel mappings, only invalidating them
1791  * on the current cpu.  It should only be used under carefully controlled
1792  * conditions.
1793  */
1794 void
1795 pmap_qremove_quick(vm_offset_t beg_va, int count)
1796 {
1797         vm_offset_t end_va;
1798         vm_offset_t va;
1799
1800         end_va = beg_va + count * PAGE_SIZE;
1801
1802         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1803                 pt_entry_t *pte;
1804
1805                 pte = vtopte(va);
1806                 (void)pte_load_clear(pte);
1807                 cpu_invlpg((void *)va);
1808         }
1809 }
1810
1811 /*
1812  * This routine removes temporary kernel mappings *without* invalidating
1813  * the TLB.  It can only be used on permanent kva reservations such as those
1814  * found in buffer cache buffers, under carefully controlled circumstances.
1815  *
1816  * NOTE: Repopulating these KVAs requires unconditional invalidation.
1817  *       (pmap_qenter() does unconditional invalidation).
1818  */
1819 void
1820 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1821 {
1822         vm_offset_t end_va;
1823         vm_offset_t va;
1824
1825         end_va = beg_va + count * PAGE_SIZE;
1826
1827         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1828                 pt_entry_t *pte;
1829
1830                 pte = vtopte(va);
1831                 (void)pte_load_clear(pte);
1832         }
1833 }
1834
1835 /*
1836  * Create a new thread and optionally associate it with a (new) process.
1837  * NOTE! the new thread's cpu may not equal the current cpu.
1838  */
1839 void
1840 pmap_init_thread(thread_t td)
1841 {
1842         /* enforce pcb placement & alignment */
1843         td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1844         td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1845         td->td_savefpu = &td->td_pcb->pcb_save;
1846         td->td_sp = (char *)td->td_pcb; /* no -16 */
1847 }
1848
1849 /*
1850  * This routine directly affects the fork perf for a process.
1851  */
1852 void
1853 pmap_init_proc(struct proc *p)
1854 {
1855 }
1856
1857 static void
1858 pmap_pinit_defaults(struct pmap *pmap)
1859 {
1860         bcopy(pmap_bits_default, pmap->pmap_bits,
1861               sizeof(pmap_bits_default));
1862         bcopy(protection_codes, pmap->protection_codes,
1863               sizeof(protection_codes));
1864         bcopy(pat_pte_index, pmap->pmap_cache_bits,
1865               sizeof(pat_pte_index));
1866         pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1867         pmap->copyinstr = std_copyinstr;
1868         pmap->copyin = std_copyin;
1869         pmap->copyout = std_copyout;
1870         pmap->fubyte = std_fubyte;
1871         pmap->subyte = std_subyte;
1872         pmap->fuword32 = std_fuword32;
1873         pmap->fuword64 = std_fuword64;
1874         pmap->suword32 = std_suword32;
1875         pmap->suword64 = std_suword64;
1876         pmap->swapu32 = std_swapu32;
1877         pmap->swapu64 = std_swapu64;
1878 }
1879 /*
1880  * Initialize pmap0/vmspace0.
1881  *
1882  * On architectures where the kernel pmap is not integrated into the user
1883  * process pmap, this pmap represents the process pmap, not the kernel pmap.
1884  * kernel_pmap should be used to directly access the kernel_pmap.
1885  */
1886 void
1887 pmap_pinit0(struct pmap *pmap)
1888 {
1889         int i;
1890
1891         pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1892         pmap->pm_count = 1;
1893         CPUMASK_ASSZERO(pmap->pm_active);
1894         pmap->pm_pvhint_pt = NULL;
1895         pmap->pm_pvhint_pte = NULL;
1896         RB_INIT(&pmap->pm_pvroot);
1897         spin_init(&pmap->pm_spin, "pmapinit0");
1898         for (i = 0; i < PM_PLACEMARKS; ++i)
1899                 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1900         bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1901         pmap_pinit_defaults(pmap);
1902 }
1903
1904 /*
1905  * Initialize a preallocated and zeroed pmap structure,
1906  * such as one in a vmspace structure.
1907  */
1908 static void
1909 pmap_pinit_simple(struct pmap *pmap)
1910 {
1911         int i;
1912
1913         /*
1914          * Misc initialization
1915          */
1916         pmap->pm_count = 1;
1917         CPUMASK_ASSZERO(pmap->pm_active);
1918         pmap->pm_pvhint_pt = NULL;
1919         pmap->pm_pvhint_pte = NULL;
1920         pmap->pm_flags = PMAP_FLAG_SIMPLE;
1921
1922         pmap_pinit_defaults(pmap);
1923
1924         /*
1925          * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1926          * for this).
1927          */
1928         if (pmap->pm_pmlpv == NULL) {
1929                 RB_INIT(&pmap->pm_pvroot);
1930                 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1931                 spin_init(&pmap->pm_spin, "pmapinitsimple");
1932                 for (i = 0; i < PM_PLACEMARKS; ++i)
1933                         pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1934         }
1935 }
1936
1937 void
1938 pmap_pinit(struct pmap *pmap)
1939 {
1940         pv_entry_t pv;
1941         int j;
1942
1943         if (pmap->pm_pmlpv) {
1944                 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1945                         pmap_puninit(pmap);
1946                 }
1947         }
1948
1949         pmap_pinit_simple(pmap);
1950         pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1951
1952         /*
1953          * No need to allocate page table space yet but we do need a valid
1954          * page directory table.
1955          */
1956         if (pmap->pm_pml4 == NULL) {
1957                 pmap->pm_pml4 =
1958                     (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1959                                                         PAGE_SIZE,
1960                                                         VM_SUBSYS_PML4);
1961         }
1962
1963         /*
1964          * Allocate the page directory page, which wires it even though
1965          * it isn't being entered into some higher level page table (it
1966          * being the highest level).  If one is already cached we don't
1967          * have to do anything.
1968          */
1969         if ((pv = pmap->pm_pmlpv) == NULL) {
1970                 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1971                 pmap->pm_pmlpv = pv;
1972                 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1973                             VM_PAGE_TO_PHYS(pv->pv_m));
1974                 pv_put(pv);
1975
1976                 /*
1977                  * Install DMAP and KMAP.
1978                  */
1979                 for (j = 0; j < NDMPML4E; ++j) {
1980                         pmap->pm_pml4[DMPML4I + j] =
1981                             (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1982                             pmap->pmap_bits[PG_RW_IDX] |
1983                             pmap->pmap_bits[PG_V_IDX] |
1984                             pmap->pmap_bits[PG_U_IDX];
1985                 }
1986                 pmap->pm_pml4[KPML4I] = KPDPphys |
1987                     pmap->pmap_bits[PG_RW_IDX] |
1988                     pmap->pmap_bits[PG_V_IDX] |
1989                     pmap->pmap_bits[PG_U_IDX];
1990
1991                 /*
1992                  * install self-referential address mapping entry
1993                  */
1994                 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1995                     pmap->pmap_bits[PG_V_IDX] |
1996                     pmap->pmap_bits[PG_RW_IDX] |
1997                     pmap->pmap_bits[PG_A_IDX] |
1998                     pmap->pmap_bits[PG_M_IDX];
1999         } else {
2000                 KKASSERT(pv->pv_m->flags & PG_MAPPED);
2001                 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
2002         }
2003         KKASSERT(pmap->pm_pml4[255] == 0);
2004         KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
2005         KKASSERT(pv->pv_entry.rbe_left == NULL);
2006         KKASSERT(pv->pv_entry.rbe_right == NULL);
2007 }
2008
2009 /*
2010  * Clean up a pmap structure so it can be physically freed.  This routine
2011  * is called by the vmspace dtor function.  A great deal of pmap data is
2012  * left passively mapped to improve vmspace management so we have a bit
2013  * of cleanup work to do here.
2014  */
2015 void
2016 pmap_puninit(pmap_t pmap)
2017 {
2018         pv_entry_t pv;
2019         vm_page_t p;
2020
2021         KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
2022         if ((pv = pmap->pm_pmlpv) != NULL) {
2023                 if (pv_hold_try(pv) == 0)
2024                         pv_lock(pv);
2025                 KKASSERT(pv == pmap->pm_pmlpv);
2026                 p = pmap_remove_pv_page(pv);
2027                 pv_free(pv, NULL);
2028                 pv = NULL;      /* safety */
2029                 pmap_kremove((vm_offset_t)pmap->pm_pml4);
2030                 vm_page_busy_wait(p, FALSE, "pgpun");
2031                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
2032                 vm_page_unwire(p, 0);
2033                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2034
2035                 /*
2036                  * XXX eventually clean out PML4 static entries and
2037                  * use vm_page_free_zero()
2038                  */
2039                 vm_page_free(p);
2040                 pmap->pm_pmlpv = NULL;
2041         }
2042         if (pmap->pm_pml4) {
2043                 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
2044                 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
2045                 pmap->pm_pml4 = NULL;
2046         }
2047         KKASSERT(pmap->pm_stats.resident_count == 0);
2048         KKASSERT(pmap->pm_stats.wired_count == 0);
2049 }
2050
2051 /*
2052  * This function is now unused (used to add the pmap to the pmap_list)
2053  */
2054 void
2055 pmap_pinit2(struct pmap *pmap)
2056 {
2057 }
2058
2059 /*
2060  * This routine is called when various levels in the page table need to
2061  * be populated.  This routine cannot fail.
2062  *
2063  * This function returns two locked pv_entry's, one representing the
2064  * requested pv and one representing the requested pv's parent pv.  If
2065  * an intermediate page table does not exist it will be created, mapped,
2066  * wired, and the parent page table will be given an additional hold
2067  * count representing the presence of the child pv_entry.
2068  */
2069 static
2070 pv_entry_t
2071 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
2072 {
2073         pt_entry_t *ptep;
2074         pv_entry_t pv;
2075         pv_entry_t pvp;
2076         pt_entry_t v;
2077         vm_pindex_t pt_pindex;
2078         vm_page_t m;
2079         int isnew;
2080         int ispt;
2081
2082         /*
2083          * If the pv already exists and we aren't being asked for the
2084          * parent page table page we can just return it.  A locked+held pv
2085          * is returned.  The pv will also have a second hold related to the
2086          * pmap association that we don't have to worry about.
2087          */
2088         ispt = 0;
2089         pv = pv_alloc(pmap, ptepindex, &isnew);
2090         if (isnew == 0 && pvpp == NULL)
2091                 return(pv);
2092
2093         /*
2094          * Special case terminal PVs.  These are not page table pages so
2095          * no vm_page is allocated (the caller supplied the vm_page).  If
2096          * pvpp is non-NULL we are being asked to also removed the pt_pv
2097          * for this pv.
2098          *
2099          * Note that pt_pv's are only returned for user VAs. We assert that
2100          * a pt_pv is not being requested for kernel VAs.  The kernel
2101          * pre-wires all higher-level page tables so don't overload managed
2102          * higher-level page tables on top of it!
2103          */
2104         if (ptepindex < pmap_pt_pindex(0)) {
2105                 if (ptepindex >= NUPTE_USER) {
2106                         /* kernel manages this manually for KVM */
2107                         KKASSERT(pvpp == NULL);
2108                 } else {
2109                         KKASSERT(pvpp != NULL);
2110                         pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
2111                         pvp = pmap_allocpte(pmap, pt_pindex, NULL);
2112                         if (isnew)
2113                                 vm_page_wire_quick(pvp->pv_m);
2114                         *pvpp = pvp;
2115                 }
2116                 return(pv);
2117         }
2118
2119         /*
2120          * The kernel never uses managed PT/PD/PDP pages.
2121          */
2122         KKASSERT(pmap != &kernel_pmap);
2123
2124         /*
2125          * Non-terminal PVs allocate a VM page to represent the page table,
2126          * so we have to resolve pvp and calculate ptepindex for the pvp
2127          * and then for the page table entry index in the pvp for
2128          * fall-through.
2129          */
2130         if (ptepindex < pmap_pd_pindex(0)) {
2131                 /*
2132                  * pv is PT, pvp is PD
2133                  */
2134                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
2135                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
2136                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2137
2138                 /*
2139                  * PT index in PD
2140                  */
2141                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
2142                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
2143                 ispt = 1;
2144         } else if (ptepindex < pmap_pdp_pindex(0)) {
2145                 /*
2146                  * pv is PD, pvp is PDP
2147                  *
2148                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2149                  *                   the PD.
2150                  */
2151                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
2152                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2153
2154                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
2155                         KKASSERT(pvpp == NULL);
2156                         pvp = NULL;
2157                 } else {
2158                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
2159                 }
2160
2161                 /*
2162                  * PD index in PDP
2163                  */
2164                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
2165                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
2166         } else if (ptepindex < pmap_pml4_pindex()) {
2167                 /*
2168                  * pv is PDP, pvp is the root pml4 table
2169                  */
2170                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2171
2172                 /*
2173                  * PDP index in PML4
2174                  */
2175                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2176                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2177         } else {
2178                 /*
2179                  * pv represents the top-level PML4, there is no parent.
2180                  */
2181                 pvp = NULL;
2182         }
2183
2184         if (isnew == 0)
2185                 goto notnew;
2186
2187         /*
2188          * (isnew) is TRUE, pv is not terminal.
2189          *
2190          * (1) Add a wire count to the parent page table (pvp).
2191          * (2) Allocate a VM page for the page table.
2192          * (3) Enter the VM page into the parent page table.
2193          *
2194          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2195          */
2196         if (pvp)
2197                 vm_page_wire_quick(pvp->pv_m);
2198
2199         for (;;) {
2200                 m = vm_page_alloc(NULL, pv->pv_pindex,
2201                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2202                                   VM_ALLOC_INTERRUPT);
2203                 if (m)
2204                         break;
2205                 vm_wait(0);
2206         }
2207         vm_page_wire(m);        /* wire for mapping in parent */
2208         vm_page_unmanage(m);    /* m must be spinunlocked */
2209         pmap_zero_page(VM_PAGE_TO_PHYS(m));
2210         m->valid = VM_PAGE_BITS_ALL;
2211
2212         vm_page_spin_lock(m);
2213         pmap_page_stats_adding(m);
2214         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2215         pv->pv_m = m;
2216         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2217         vm_page_spin_unlock(m);
2218
2219         /*
2220          * (isnew) is TRUE, pv is not terminal.
2221          *
2222          * Wire the page into pvp.  Bump the resident_count for the pmap.
2223          * There is no pvp for the top level, address the pm_pml4[] array
2224          * directly.
2225          *
2226          * If the caller wants the parent we return it, otherwise
2227          * we just put it away.
2228          *
2229          * No interlock is needed for pte 0 -> non-zero.
2230          *
2231          * In the situation where *ptep is valid we might have an unmanaged
2232          * page table page shared from another page table which we need to
2233          * unshare before installing our private page table page.
2234          */
2235         if (pvp) {
2236                 v = VM_PAGE_TO_PHYS(m) |
2237                     (pmap->pmap_bits[PG_U_IDX] |
2238                      pmap->pmap_bits[PG_RW_IDX] |
2239                      pmap->pmap_bits[PG_V_IDX] |
2240                      pmap->pmap_bits[PG_A_IDX] |
2241                      pmap->pmap_bits[PG_M_IDX]);
2242                 ptep = pv_pte_lookup(pvp, ptepindex);
2243                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2244                         pt_entry_t pte;
2245
2246                         if (ispt == 0) {
2247                                 panic("pmap_allocpte: unexpected pte %p/%d",
2248                                       pvp, (int)ptepindex);
2249                         }
2250                         pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, v);
2251                         if (vm_page_unwire_quick(
2252                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2253                                 panic("pmap_allocpte: shared pgtable "
2254                                       "pg bad wirecount");
2255                         }
2256                 } else {
2257                         pt_entry_t pte;
2258
2259                         pte = atomic_swap_long(ptep, v);
2260                         if (pte != 0) {
2261                                 kprintf("install pgtbl mixup 0x%016jx "
2262                                         "old/new 0x%016jx/0x%016jx\n",
2263                                         (intmax_t)ptepindex, pte, v);
2264                         }
2265                 }
2266         }
2267         vm_page_wakeup(m);
2268
2269         /*
2270          * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2271          */
2272 notnew:
2273         if (pvp) {
2274                 KKASSERT(pvp->pv_m != NULL);
2275                 ptep = pv_pte_lookup(pvp, ptepindex);
2276                 v = VM_PAGE_TO_PHYS(pv->pv_m) |
2277                     (pmap->pmap_bits[PG_U_IDX] |
2278                      pmap->pmap_bits[PG_RW_IDX] |
2279                      pmap->pmap_bits[PG_V_IDX] |
2280                      pmap->pmap_bits[PG_A_IDX] |
2281                      pmap->pmap_bits[PG_M_IDX]);
2282                 if (*ptep != v) {
2283                         kprintf("mismatched upper level pt %016jx/%016jx\n",
2284                                 *ptep, v);
2285                 }
2286         }
2287         if (pvpp)
2288                 *pvpp = pvp;
2289         else if (pvp)
2290                 pv_put(pvp);
2291         return (pv);
2292 }
2293
2294 /*
2295  * This version of pmap_allocpte() checks for possible segment optimizations
2296  * that would allow page-table sharing.  It can be called for terminal
2297  * page or page table page ptepindex's.
2298  *
2299  * The function is called with page table page ptepindex's for fictitious
2300  * and unmanaged terminal pages.  That is, we don't want to allocate a
2301  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
2302  * for this case.
2303  *
2304  * This function can return a pv and *pvpp associated with the passed in pmap
2305  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
2306  * an unmanaged page table page will be entered into the pass in pmap.
2307  */
2308 static
2309 pv_entry_t
2310 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2311                   vm_map_entry_t entry, vm_offset_t va)
2312 {
2313         vm_object_t object;
2314         pmap_t obpmap;
2315         pmap_t *obpmapp;
2316         vm_offset_t b;
2317         pv_entry_t pte_pv;      /* in original or shared pmap */
2318         pv_entry_t pt_pv;       /* in original or shared pmap */
2319         pv_entry_t proc_pd_pv;  /* in original pmap */
2320         pv_entry_t proc_pt_pv;  /* in original pmap */
2321         pv_entry_t xpv;         /* PT in shared pmap */
2322         pd_entry_t *pt;         /* PT entry in PD of original pmap */
2323         pd_entry_t opte;        /* contents of *pt */
2324         pd_entry_t npte;        /* contents of *pt */
2325         vm_page_t m;
2326
2327         /*
2328          * Basic tests, require a non-NULL vm_map_entry, require proper
2329          * alignment and type for the vm_map_entry, require that the
2330          * underlying object already be allocated.
2331          *
2332          * We allow almost any type of object to use this optimization.
2333          * The object itself does NOT have to be sized to a multiple of the
2334          * segment size, but the memory mapping does.
2335          *
2336          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2337          *     won't work as expected.
2338          */
2339         if (entry == NULL ||
2340             pmap_mmu_optimize == 0 ||                   /* not enabled */
2341             (pmap->pm_flags & PMAP_HVM) ||              /* special pmap */
2342             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal or pt */
2343             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2344             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2345             entry->object.vm_object == NULL ||          /* needs VM object */
2346             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2347             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2348             (entry->offset & SEG_MASK) ||               /* must be aligned */
2349             (entry->start & SEG_MASK)) {
2350                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2351         }
2352
2353         /*
2354          * Make sure the full segment can be represented.
2355          */
2356         b = va & ~(vm_offset_t)SEG_MASK;
2357         if (b < entry->start || b + SEG_SIZE > entry->end)
2358                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2359
2360         /*
2361          * If the full segment can be represented dive the VM object's
2362          * shared pmap, allocating as required.
2363          */
2364         object = entry->object.vm_object;
2365
2366         if (entry->protection & VM_PROT_WRITE)
2367                 obpmapp = &object->md.pmap_rw;
2368         else
2369                 obpmapp = &object->md.pmap_ro;
2370
2371 #ifdef PMAP_DEBUG2
2372         if (pmap_enter_debug > 0) {
2373                 --pmap_enter_debug;
2374                 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2375                         "obpmapp %p %p\n",
2376                         va, entry->protection, object,
2377                         obpmapp, *obpmapp);
2378                 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2379                         entry, entry->start, entry->end);
2380         }
2381 #endif
2382
2383         /*
2384          * We allocate what appears to be a normal pmap but because portions
2385          * of this pmap are shared with other unrelated pmaps we have to
2386          * set pm_active to point to all cpus.
2387          *
2388          * XXX Currently using pmap_spin to interlock the update, can't use
2389          *     vm_object_hold/drop because the token might already be held
2390          *     shared OR exclusive and we don't know.
2391          */
2392         while ((obpmap = *obpmapp) == NULL) {
2393                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2394                 pmap_pinit_simple(obpmap);
2395                 pmap_pinit2(obpmap);
2396                 spin_lock(&pmap_spin);
2397                 if (*obpmapp != NULL) {
2398                         /*
2399                          * Handle race
2400                          */
2401                         spin_unlock(&pmap_spin);
2402                         pmap_release(obpmap);
2403                         pmap_puninit(obpmap);
2404                         kfree(obpmap, M_OBJPMAP);
2405                         obpmap = *obpmapp; /* safety */
2406                 } else {
2407                         obpmap->pm_active = smp_active_mask;
2408                         obpmap->pm_flags |= PMAP_SEGSHARED;
2409                         *obpmapp = obpmap;
2410                         spin_unlock(&pmap_spin);
2411                 }
2412         }
2413
2414         /*
2415          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2416          * pte/pt using the shared pmap from the object but also adjust
2417          * the process pmap's page table page as a side effect.
2418          */
2419
2420         /*
2421          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2422          * we will return.  This is true if ptepindex represents a terminal
2423          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2424          * the PD.
2425          */
2426         pt_pv = NULL;
2427         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2428 retry:
2429         if (ptepindex >= pmap_pt_pindex(0))
2430                 xpv = pte_pv;
2431         else
2432                 xpv = pt_pv;
2433
2434         /*
2435          * Resolve the PD in the process pmap so we can properly share the
2436          * page table page.  Lock order is bottom-up (leaf first)!
2437          *
2438          * NOTE: proc_pt_pv can be NULL.
2439          */
2440         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b), NULL);
2441         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2442 #ifdef PMAP_DEBUG2
2443         if (pmap_enter_debug > 0) {
2444                 --pmap_enter_debug;
2445                 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2446                         proc_pt_pv,
2447                         (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2448                         proc_pd_pv,
2449                         va);
2450         }
2451 #endif
2452
2453         /*
2454          * xpv is the page table page pv from the shared object
2455          * (for convenience), from above.
2456          *
2457          * Calculate the pte value for the PT to load into the process PD.
2458          * If we have to change it we must properly dispose of the previous
2459          * entry.
2460          */
2461         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2462         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2463                (pmap->pmap_bits[PG_U_IDX] |
2464                 pmap->pmap_bits[PG_RW_IDX] |
2465                 pmap->pmap_bits[PG_V_IDX] |
2466                 pmap->pmap_bits[PG_A_IDX] |
2467                 pmap->pmap_bits[PG_M_IDX]);
2468
2469         /*
2470          * Dispose of previous page table page if it was local to the
2471          * process pmap.  If the old pt is not empty we cannot dispose of it
2472          * until we clean it out.  This case should not arise very often so
2473          * it is not optimized.
2474          *
2475          * Leave pt_pv and pte_pv (in our object pmap) locked and intact
2476          * for the retry.
2477          */
2478         if (proc_pt_pv) {
2479                 pmap_inval_bulk_t bulk;
2480
2481                 if (proc_pt_pv->pv_m->wire_count != 1) {
2482                         pv_put(proc_pd_pv);
2483                         pv_put(proc_pt_pv);
2484                         pmap_remove(pmap,
2485                                     va & ~(vm_offset_t)SEG_MASK,
2486                                     (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2487                         goto retry;
2488                 }
2489
2490                 /*
2491                  * The release call will indirectly clean out *pt
2492                  */
2493                 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2494                 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2495                 pmap_inval_bulk_flush(&bulk);
2496                 proc_pt_pv = NULL;
2497                 /* relookup */
2498                 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2499         }
2500
2501         /*
2502          * Handle remaining cases.
2503          */
2504         if (*pt == 0) {
2505                 atomic_swap_long(pt, npte);
2506                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2507                 vm_page_wire_quick(proc_pd_pv->pv_m);   /* proc pd for sh pt */
2508                 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2509         } else if (*pt != npte) {
2510                 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2511
2512 #if 0
2513                 opte = pte_load_clear(pt);
2514                 KKASSERT(opte && opte != npte);
2515
2516                 *pt = npte;
2517 #endif
2518                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2519
2520                 /*
2521                  * Clean up opte, bump the wire_count for the process
2522                  * PD page representing the new entry if it was
2523                  * previously empty.
2524                  *
2525                  * If the entry was not previously empty and we have
2526                  * a PT in the proc pmap then opte must match that
2527                  * pt.  The proc pt must be retired (this is done
2528                  * later on in this procedure).
2529                  *
2530                  * NOTE: replacing valid pte, wire_count on proc_pd_pv
2531                  * stays the same.
2532                  */
2533                 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2534                 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2535                 if (vm_page_unwire_quick(m)) {
2536                         panic("pmap_allocpte_seg: "
2537                               "bad wire count %p",
2538                               m);
2539                 }
2540         }
2541
2542         /*
2543          * The existing process page table was replaced and must be destroyed
2544          * here.
2545          */
2546         if (proc_pd_pv)
2547                 pv_put(proc_pd_pv);
2548         if (pvpp)
2549                 *pvpp = pt_pv;
2550         else
2551                 pv_put(pt_pv);
2552
2553         return (pte_pv);
2554 }
2555
2556 /*
2557  * Release any resources held by the given physical map.
2558  *
2559  * Called when a pmap initialized by pmap_pinit is being released.  Should
2560  * only be called if the map contains no valid mappings.
2561  */
2562 struct pmap_release_info {
2563         pmap_t  pmap;
2564         int     retry;
2565         pv_entry_t pvp;
2566 };
2567
2568 static int pmap_release_callback(pv_entry_t pv, void *data);
2569
2570 void
2571 pmap_release(struct pmap *pmap)
2572 {
2573         struct pmap_release_info info;
2574
2575         KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2576                 ("pmap still active! %016jx",
2577                 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2578
2579         /*
2580          * There is no longer a pmap_list, if there were we would remove the
2581          * pmap from it here.
2582          */
2583
2584         /*
2585          * Pull pv's off the RB tree in order from low to high and release
2586          * each page.
2587          */
2588         info.pmap = pmap;
2589         do {
2590                 info.retry = 0;
2591                 info.pvp = NULL;
2592
2593                 spin_lock(&pmap->pm_spin);
2594                 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2595                         pmap_release_callback, &info);
2596                 spin_unlock(&pmap->pm_spin);
2597
2598                 if (info.pvp)
2599                         pv_put(info.pvp);
2600         } while (info.retry);
2601
2602
2603         /*
2604          * One resident page (the pml4 page) should remain.
2605          * No wired pages should remain.
2606          */
2607 #if 1
2608         if (pmap->pm_stats.resident_count !=
2609             ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1) ||
2610             pmap->pm_stats.wired_count != 0) {
2611                 kprintf("fatal pmap problem - pmap %p flags %08x "
2612                         "rescnt=%jd wirecnt=%jd\n",
2613                         pmap,
2614                         pmap->pm_flags,
2615                         pmap->pm_stats.resident_count,
2616                         pmap->pm_stats.wired_count);
2617                 tsleep(pmap, 0, "DEAD", 0);
2618         }
2619 #else
2620         KKASSERT(pmap->pm_stats.resident_count ==
2621                  ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2622         KKASSERT(pmap->pm_stats.wired_count == 0);
2623 #endif
2624 }
2625
2626 /*
2627  * Called from low to high.  We must cache the proper parent pv so we
2628  * can adjust its wired count.
2629  */
2630 static int
2631 pmap_release_callback(pv_entry_t pv, void *data)
2632 {
2633         struct pmap_release_info *info = data;
2634         pmap_t pmap = info->pmap;
2635         vm_pindex_t pindex;
2636         int r;
2637
2638         /*
2639          * Acquire a held and locked pv, check for release race
2640          */
2641         pindex = pv->pv_pindex;
2642         if (info->pvp == pv) {
2643                 spin_unlock(&pmap->pm_spin);
2644                 info->pvp = NULL;
2645         } else if (pv_hold_try(pv)) {
2646                 spin_unlock(&pmap->pm_spin);
2647         } else {
2648                 spin_unlock(&pmap->pm_spin);
2649                 pv_lock(pv);
2650                 pv_put(pv);
2651                 info->retry = 1;
2652                 spin_lock(&pmap->pm_spin);
2653
2654                 return -1;
2655         }
2656         KKASSERT(pv->pv_pmap == pmap && pindex == pv->pv_pindex);
2657
2658         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2659                 /*
2660                  * I am PTE, parent is PT
2661                  */
2662                 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2663                 pindex += NUPTE_TOTAL;
2664         } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2665                 /*
2666                  * I am PT, parent is PD
2667                  */
2668                 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2669                 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2670         } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2671                 /*
2672                  * I am PD, parent is PDP
2673                  */
2674                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2675                          NPDPEPGSHIFT;
2676                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2677         } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2678                 /*
2679                  * I am PDP, parent is PML4 (there's only one)
2680                  */
2681 #if 0
2682                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2683                            NUPD_TOTAL) >> NPML4EPGSHIFT;
2684                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2685 #endif
2686                 pindex = pmap_pml4_pindex();
2687         } else {
2688                 /*
2689                  * parent is NULL
2690                  */
2691                 if (info->pvp) {
2692                         pv_put(info->pvp);
2693                         info->pvp = NULL;
2694                 }
2695                 pindex = 0;
2696         }
2697         if (pindex) {
2698                 if (info->pvp && info->pvp->pv_pindex != pindex) {
2699                         pv_put(info->pvp);
2700                         info->pvp = NULL;
2701                 }
2702                 if (info->pvp == NULL)
2703                         info->pvp = pv_get(pmap, pindex, NULL);
2704         } else {
2705                 if (info->pvp) {
2706                         pv_put(info->pvp);
2707                         info->pvp = NULL;
2708                 }
2709         }
2710         r = pmap_release_pv(pv, info->pvp, NULL);
2711         spin_lock(&pmap->pm_spin);
2712
2713         return(r);
2714 }
2715
2716 /*
2717  * Called with held (i.e. also locked) pv.  This function will dispose of
2718  * the lock along with the pv.
2719  *
2720  * If the caller already holds the locked parent page table for pv it
2721  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2722  * pass NULL for pvp.
2723  */
2724 static int
2725 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2726 {
2727         vm_page_t p;
2728
2729         /*
2730          * The pmap is currently not spinlocked, pv is held+locked.
2731          * Remove the pv's page from its parent's page table.  The
2732          * parent's page table page's wire_count will be decremented.
2733          *
2734          * This will clean out the pte at any level of the page table.
2735          * If smp != 0 all cpus are affected.
2736          *
2737          * Do not tear-down recursively, its faster to just let the
2738          * release run its course.
2739          */
2740         pmap_remove_pv_pte(pv, pvp, bulk, 0);
2741
2742         /*
2743          * Terminal pvs are unhooked from their vm_pages.  Because
2744          * terminal pages aren't page table pages they aren't wired
2745          * by us, so we have to be sure not to unwire them either.
2746          */
2747         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2748                 pmap_remove_pv_page(pv);
2749                 goto skip;
2750         }
2751
2752         /*
2753          * We leave the top-level page table page cached, wired, and
2754          * mapped in the pmap until the dtor function (pmap_puninit())
2755          * gets called.
2756          *
2757          * Since we are leaving the top-level pv intact we need
2758          * to break out of what would otherwise be an infinite loop.
2759          */
2760         if (pv->pv_pindex == pmap_pml4_pindex()) {
2761                 pv_put(pv);
2762                 return(-1);
2763         }
2764
2765         /*
2766          * For page table pages (other than the top-level page),
2767          * remove and free the vm_page.  The representitive mapping
2768          * removed above by pmap_remove_pv_pte() did not undo the
2769          * last wire_count so we have to do that as well.
2770          */
2771         p = pmap_remove_pv_page(pv);
2772         vm_page_busy_wait(p, FALSE, "pmaprl");
2773         if (p->wire_count != 1) {
2774                 kprintf("p->wire_count was %016lx %d\n",
2775                         pv->pv_pindex, p->wire_count);
2776         }
2777         KKASSERT(p->wire_count == 1);
2778         KKASSERT(p->flags & PG_UNMANAGED);
2779
2780         vm_page_unwire(p, 0);
2781         KKASSERT(p->wire_count == 0);
2782
2783         vm_page_free(p);
2784 skip:
2785         pv_free(pv, pvp);
2786
2787         return 0;
2788 }
2789
2790 /*
2791  * This function will remove the pte associated with a pv from its parent.
2792  * Terminal pv's are supported.  All cpus specified by (bulk) are properly
2793  * invalidated.
2794  *
2795  * The wire count will be dropped on the parent page table.  The wire
2796  * count on the page being removed (pv->pv_m) from the parent page table
2797  * is NOT touched.  Note that terminal pages will not have any additional
2798  * wire counts while page table pages will have at least one representing
2799  * the mapping, plus others representing sub-mappings.
2800  *
2801  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2802  *       pages and user page table and terminal pages.
2803  *
2804  * NOTE: The pte being removed might be unmanaged, and the pv supplied might
2805  *       be freshly allocated and not imply that the pte is managed.  In this
2806  *       case pv->pv_m should be NULL.
2807  *
2808  * The pv must be locked.  The pvp, if supplied, must be locked.  All
2809  * supplied pv's will remain locked on return.
2810  *
2811  * XXX must lock parent pv's if they exist to remove pte XXX
2812  */
2813 static
2814 void
2815 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2816                    int destroy)
2817 {
2818         vm_pindex_t ptepindex = pv->pv_pindex;
2819         pmap_t pmap = pv->pv_pmap;
2820         vm_page_t p;
2821         int gotpvp = 0;
2822
2823         KKASSERT(pmap);
2824
2825         if (ptepindex == pmap_pml4_pindex()) {
2826                 /*
2827                  * We are the top level PML4E table, there is no parent.
2828                  */
2829                 p = pmap->pm_pmlpv->pv_m;
2830                 KKASSERT(pv->pv_m == p);        /* debugging */
2831         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2832                 /*
2833                  * Remove a PDP page from the PML4E.  This can only occur
2834                  * with user page tables.  We do not have to lock the
2835                  * pml4 PV so just ignore pvp.
2836                  */
2837                 vm_pindex_t pml4_pindex;
2838                 vm_pindex_t pdp_index;
2839                 pml4_entry_t *pdp;
2840
2841                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2842                 if (pvp == NULL) {
2843                         pml4_pindex = pmap_pml4_pindex();
2844                         pvp = pv_get(pv->pv_pmap, pml4_pindex, NULL);
2845                         KKASSERT(pvp);
2846                         gotpvp = 1;
2847                 }
2848
2849                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2850                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2851                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2852                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2853                 KKASSERT(pv->pv_m == p);        /* debugging */
2854         } else if (ptepindex >= pmap_pd_pindex(0)) {
2855                 /*
2856                  * Remove a PD page from the PDP
2857                  *
2858                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2859                  *                   of a simple pmap because it stops at
2860                  *                   the PD page.
2861                  */
2862                 vm_pindex_t pdp_pindex;
2863                 vm_pindex_t pd_index;
2864                 pdp_entry_t *pd;
2865
2866                 pd_index = ptepindex - pmap_pd_pindex(0);
2867
2868                 if (pvp == NULL) {
2869                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2870                                      (pd_index >> NPML4EPGSHIFT);
2871                         pvp = pv_get(pv->pv_pmap, pdp_pindex, NULL);
2872                         gotpvp = 1;
2873                 }
2874
2875                 if (pvp) {
2876                         pd = pv_pte_lookup(pvp, pd_index &
2877                                                 ((1ul << NPDPEPGSHIFT) - 1));
2878                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2879                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2880                         pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2881                 } else {
2882                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2883                         p = pv->pv_m;           /* degenerate test later */
2884                 }
2885                 KKASSERT(pv->pv_m == p);        /* debugging */
2886         } else if (ptepindex >= pmap_pt_pindex(0)) {
2887                 /*
2888                  *  Remove a PT page from the PD
2889                  */
2890                 vm_pindex_t pd_pindex;
2891                 vm_pindex_t pt_index;
2892                 pd_entry_t *pt;
2893
2894                 pt_index = ptepindex - pmap_pt_pindex(0);
2895
2896                 if (pvp == NULL) {
2897                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2898                                     (pt_index >> NPDPEPGSHIFT);
2899                         pvp = pv_get(pv->pv_pmap, pd_pindex, NULL);
2900                         KKASSERT(pvp);
2901                         gotpvp = 1;
2902                 }
2903
2904                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2905 #if 0
2906                 KASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0,
2907                         ("*pt unexpectedly invalid %016jx "
2908                          "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
2909                         *pt, gotpvp, ptepindex, pt_index, pv, pvp));
2910                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2911 #else
2912                 if ((*pt & pmap->pmap_bits[PG_V_IDX]) == 0) {
2913                         kprintf("*pt unexpectedly invalid %016jx "
2914                                 "gotpvp=%d ptepindex=%ld ptindex=%ld "
2915                                 "pv=%p pvp=%p\n",
2916                                 *pt, gotpvp, ptepindex, pt_index, pv, pvp);
2917                         tsleep(pt, 0, "DEAD", 0);
2918                         p = pv->pv_m;
2919                 } else {
2920                         p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2921                 }
2922 #endif
2923                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2924                 KKASSERT(pv->pv_m == p);        /* debugging */
2925         } else {
2926                 /*
2927                  * Remove a PTE from the PT page.  The PV might exist even if
2928                  * the PTE is not managed, in whichcase pv->pv_m should be
2929                  * NULL.
2930                  *
2931                  * NOTE: Userland pmaps manage the parent PT/PD/PDP page
2932                  *       table pages but the kernel_pmap does not.
2933                  *
2934                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2935                  *       pv is a pte_pv so we can safely lock pt_pv.
2936                  *
2937                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2938                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2939                  *       terminal ptes.
2940                  */
2941                 vm_pindex_t pt_pindex;
2942                 pt_entry_t *ptep;
2943                 pt_entry_t pte;
2944                 vm_offset_t va;
2945
2946                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2947                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2948
2949                 if (ptepindex >= NUPTE_USER) {
2950                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2951                         KKASSERT(pvp == NULL);
2952                         /* pvp remains NULL */
2953                 } else {
2954                         if (pvp == NULL) {
2955                                 pt_pindex = NUPTE_TOTAL +
2956                                             (ptepindex >> NPDPEPGSHIFT);
2957                                 pvp = pv_get(pv->pv_pmap, pt_pindex, NULL);
2958                                 KKASSERT(pvp);
2959                                 gotpvp = 1;
2960                         }
2961                         ptep = pv_pte_lookup(pvp, ptepindex &
2962                                                   ((1ul << NPDPEPGSHIFT) - 1));
2963                 }
2964                 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2965                 if (bulk == NULL)               /* XXX */
2966                         cpu_invlpg((void *)va); /* XXX */
2967
2968                 /*
2969                  * Now update the vm_page_t
2970                  */
2971                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
2972                     (pte & pmap->pmap_bits[PG_V_IDX])) {
2973                         /*
2974                          * Valid managed page, adjust (p).
2975                          */
2976                         if (pte & pmap->pmap_bits[PG_DEVICE_IDX]) {
2977                                 p = pv->pv_m;
2978                         } else {
2979                                 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2980                                 KKASSERT(pv->pv_m == p);
2981                         }
2982                         if (pte & pmap->pmap_bits[PG_M_IDX]) {
2983                                 if (pmap_track_modified(ptepindex))
2984                                         vm_page_dirty(p);
2985                         }
2986                         if (pte & pmap->pmap_bits[PG_A_IDX]) {
2987                                 vm_page_flag_set(p, PG_REFERENCED);
2988                         }
2989                 } else {
2990                         /*
2991                          * Unmanaged page, do not try to adjust the vm_page_t.
2992                          * pv could be freshly allocated for a pmap_enter(),
2993                          * replacing an unmanaged page with a managed one.
2994                          *
2995                          * pv->pv_m might reflect the new page and not the
2996                          * existing page.
2997                          *
2998                          * We could extract p from the physical address and
2999                          * adjust it but we explicitly do not for unmanaged
3000                          * pages.
3001                          */
3002                         p = NULL;
3003                 }
3004                 if (pte & pmap->pmap_bits[PG_W_IDX])
3005                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
3006                 if (pte & pmap->pmap_bits[PG_G_IDX])
3007                         cpu_invlpg((void *)va);
3008         }
3009
3010         /*
3011          * If requested, scrap the underlying pv->pv_m and the underlying
3012          * pv.  If this is a page-table-page we must also free the page.
3013          *
3014          * pvp must be returned locked.
3015          */
3016         if (destroy == 1) {
3017                 /*
3018                  * page table page (PT, PD, PDP, PML4), caller was responsible
3019                  * for testing wired_count.
3020                  */
3021                 KKASSERT(pv->pv_m->wire_count == 1);
3022                 p = pmap_remove_pv_page(pv);
3023                 pv_free(pv, pvp);
3024                 pv = NULL;
3025
3026                 vm_page_busy_wait(p, FALSE, "pgpun");
3027                 vm_page_unwire(p, 0);
3028                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
3029                 vm_page_free(p);
3030         } else if (destroy == 2) {
3031                 /*
3032                  * Normal page, remove from pmap and leave the underlying
3033                  * page untouched.
3034                  */
3035                 pmap_remove_pv_page(pv);
3036                 pv_free(pv, pvp);
3037                 pv = NULL;              /* safety */
3038         }
3039
3040         /*
3041          * If we acquired pvp ourselves then we are responsible for
3042          * recursively deleting it.
3043          */
3044         if (pvp && gotpvp) {
3045                 /*
3046                  * Recursively destroy higher-level page tables.
3047                  *
3048                  * This is optional.  If we do not, they will still
3049                  * be destroyed when the process exits.
3050                  *
3051                  * NOTE: Do not destroy pv_entry's with extra hold refs,
3052                  *       a caller may have unlocked it and intends to
3053                  *       continue to use it.
3054                  */
3055                 if (pmap_dynamic_delete &&
3056                     pvp->pv_m &&
3057                     pvp->pv_m->wire_count == 1 &&
3058                     (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
3059                     pvp->pv_pindex != pmap_pml4_pindex()) {
3060                         if (pmap_dynamic_delete == 2)
3061                                 kprintf("A %jd %08x\n", pvp->pv_pindex, pvp->pv_hold);
3062                         if (pmap != &kernel_pmap) {
3063                                 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
3064                                 pvp = NULL;     /* safety */
3065                         } else {
3066                                 kprintf("Attempt to remove kernel_pmap pindex "
3067                                         "%jd\n", pvp->pv_pindex);
3068                                 pv_put(pvp);
3069                         }
3070                 } else {
3071                         pv_put(pvp);
3072                 }
3073         }
3074 }
3075
3076 /*
3077  * Remove the vm_page association to a pv.  The pv must be locked.
3078  */
3079 static
3080 vm_page_t
3081 pmap_remove_pv_page(pv_entry_t pv)
3082 {
3083         vm_page_t m;
3084
3085         m = pv->pv_m;
3086         vm_page_spin_lock(m);
3087         KKASSERT(m && m == pv->pv_m);
3088         pv->pv_m = NULL;
3089         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3090         pmap_page_stats_deleting(m);
3091         if (TAILQ_EMPTY(&m->md.pv_list))
3092                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3093         vm_page_spin_unlock(m);
3094
3095         return(m);
3096 }
3097
3098 /*
3099  * Grow the number of kernel page table entries, if needed.
3100  *
3101  * This routine is always called to validate any address space
3102  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
3103  * space below KERNBASE.
3104  *
3105  * kernel_map must be locked exclusively by the caller.
3106  */
3107 void
3108 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
3109 {
3110         vm_paddr_t paddr;
3111         vm_offset_t ptppaddr;
3112         vm_page_t nkpg;
3113         pd_entry_t *pt, newpt;
3114         pdp_entry_t newpd;
3115         int update_kernel_vm_end;
3116
3117         /*
3118          * bootstrap kernel_vm_end on first real VM use
3119          */
3120         if (kernel_vm_end == 0) {
3121                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
3122                 nkpt = 0;
3123                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3124                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
3125                                         ~(PAGE_SIZE * NPTEPG - 1);
3126                         nkpt++;
3127                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
3128                                 kernel_vm_end = kernel_map.max_offset;
3129                                 break;
3130                         }
3131                 }
3132         }
3133
3134         /*
3135          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
3136          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
3137          * do not want to force-fill 128G worth of page tables.
3138          */
3139         if (kstart < KERNBASE) {
3140                 if (kstart > kernel_vm_end)
3141                         kstart = kernel_vm_end;
3142                 KKASSERT(kend <= KERNBASE);
3143                 update_kernel_vm_end = 1;
3144         } else {
3145                 update_kernel_vm_end = 0;
3146         }
3147
3148         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
3149         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
3150
3151         if (kend - 1 >= kernel_map.max_offset)
3152                 kend = kernel_map.max_offset;
3153
3154         while (kstart < kend) {
3155                 pt = pmap_pt(&kernel_pmap, kstart);
3156                 if (pt == NULL) {
3157                         /* We need a new PD entry */
3158                         nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3159                                              VM_ALLOC_NORMAL |
3160                                              VM_ALLOC_SYSTEM |
3161                                              VM_ALLOC_INTERRUPT);
3162                         if (nkpg == NULL) {
3163                                 panic("pmap_growkernel: no memory to grow "
3164                                       "kernel");
3165                         }
3166                         paddr = VM_PAGE_TO_PHYS(nkpg);
3167                         pmap_zero_page(paddr);
3168                         newpd = (pdp_entry_t)
3169                             (paddr |
3170                             kernel_pmap.pmap_bits[PG_V_IDX] |
3171                             kernel_pmap.pmap_bits[PG_RW_IDX] |
3172                             kernel_pmap.pmap_bits[PG_A_IDX] |
3173                             kernel_pmap.pmap_bits[PG_M_IDX]);
3174                         *pmap_pd(&kernel_pmap, kstart) = newpd;
3175                         continue; /* try again */
3176                 }
3177                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3178                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
3179                                  ~(PAGE_SIZE * NPTEPG - 1);
3180                         if (kstart - 1 >= kernel_map.max_offset) {
3181                                 kstart = kernel_map.max_offset;
3182                                 break;
3183                         }
3184                         continue;
3185                 }
3186
3187                 /*
3188                  * We need a new PT
3189                  *
3190                  * This index is bogus, but out of the way
3191                  */
3192                 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3193                                      VM_ALLOC_NORMAL |
3194                                      VM_ALLOC_SYSTEM |
3195                                      VM_ALLOC_INTERRUPT);
3196                 if (nkpg == NULL)
3197                         panic("pmap_growkernel: no memory to grow kernel");
3198
3199                 vm_page_wire(nkpg);
3200                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
3201                 pmap_zero_page(ptppaddr);
3202                 newpt = (pd_entry_t)(ptppaddr |
3203                                      kernel_pmap.pmap_bits[PG_V_IDX] |
3204                                      kernel_pmap.pmap_bits[PG_RW_IDX] |
3205                                      kernel_pmap.pmap_bits[PG_A_IDX] |
3206                                      kernel_pmap.pmap_bits[PG_M_IDX]);
3207                 atomic_swap_long(pmap_pt(&kernel_pmap, kstart), newpt);
3208
3209                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3210                           ~(PAGE_SIZE * NPTEPG - 1);
3211
3212                 if (kstart - 1 >= kernel_map.max_offset) {
3213                         kstart = kernel_map.max_offset;
3214                         break;
3215                 }
3216         }
3217
3218         /*
3219          * Only update kernel_vm_end for areas below KERNBASE.
3220          */
3221         if (update_kernel_vm_end && kernel_vm_end < kstart)
3222                 kernel_vm_end = kstart;
3223 }
3224
3225 /*
3226  *      Add a reference to the specified pmap.
3227  */
3228 void
3229 pmap_reference(pmap_t pmap)
3230 {
3231         if (pmap != NULL)
3232                 atomic_add_int(&pmap->pm_count, 1);
3233 }
3234
3235 /***************************************************
3236  * page management routines.
3237  ***************************************************/
3238
3239 /*
3240  * Hold a pv without locking it
3241  */
3242 static void
3243 pv_hold(pv_entry_t pv)
3244 {
3245         atomic_add_int(&pv->pv_hold, 1);
3246 }
3247
3248 /*
3249  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
3250  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
3251  * the pv properly.
3252  *
3253  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3254  * pv list via its page) must be held by the caller in order to stabilize
3255  * the pv.
3256  */
3257 static int
3258 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3259 {
3260         u_int count;
3261
3262         /*
3263          * Critical path shortcut expects pv to already have one ref
3264          * (for the pv->pv_pmap).
3265          */
3266         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3267 #ifdef PMAP_DEBUG
3268                 pv->pv_func = func;
3269                 pv->pv_line = lineno;
3270 #endif
3271                 return TRUE;
3272         }
3273
3274         for (;;) {
3275                 count = pv->pv_hold;
3276                 cpu_ccfence();
3277                 if ((count & PV_HOLD_LOCKED) == 0) {
3278                         if (atomic_cmpset_int(&pv->pv_hold, count,
3279                                               (count + 1) | PV_HOLD_LOCKED)) {
3280 #ifdef PMAP_DEBUG
3281                                 pv->pv_func = func;
3282                                 pv->pv_line = lineno;
3283 #endif
3284                                 return TRUE;
3285                         }
3286                 } else {
3287                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3288                                 return FALSE;
3289                 }
3290                 /* retry */
3291         }
3292 }
3293
3294 /*
3295  * Drop a previously held pv_entry which could not be locked, allowing its
3296  * destruction.
3297  *
3298  * Must not be called with a spinlock held as we might zfree() the pv if it
3299  * is no longer associated with a pmap and this was the last hold count.
3300  */
3301 static void
3302 pv_drop(pv_entry_t pv)
3303 {
3304         u_int count;
3305
3306         for (;;) {
3307                 count = pv->pv_hold;
3308                 cpu_ccfence();
3309                 KKASSERT((count & PV_HOLD_MASK) > 0);
3310                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3311                          (PV_HOLD_LOCKED | 1));
3312                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3313                         if ((count & PV_HOLD_MASK) == 1) {
3314 #ifdef PMAP_DEBUG2
3315                                 if (pmap_enter_debug > 0) {
3316                                         --pmap_enter_debug;
3317                                         kprintf("pv_drop: free pv %p\n", pv);
3318                                 }
3319 #endif
3320                                 KKASSERT(count == 1);
3321                                 KKASSERT(pv->pv_pmap == NULL);
3322                                 zfree(pvzone, pv);
3323                         }
3324                         return;
3325                 }
3326                 /* retry */
3327         }
3328 }
3329
3330 /*
3331  * Find or allocate the requested PV entry, returning a locked, held pv.
3332  *
3333  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3334  * for the caller and one representing the pmap and vm_page association.
3335  *
3336  * If (*isnew) is zero, the returned pv will have only one hold count.
3337  *
3338  * Since both associations can only be adjusted while the pv is locked,
3339  * together they represent just one additional hold.
3340  */
3341 static
3342 pv_entry_t
3343 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3344 {
3345         struct mdglobaldata *md = mdcpu;
3346         pv_entry_t pv;
3347         pv_entry_t pnew;
3348         int pmap_excl = 0;
3349
3350         pnew = NULL;
3351         if (md->gd_newpv) {
3352 #if 1
3353                 pnew = atomic_swap_ptr((void *)&md->gd_newpv, NULL);
3354 #else
3355                 crit_enter();
3356                 pnew = md->gd_newpv;    /* might race NULL */
3357                 md->gd_newpv = NULL;
3358                 crit_exit();
3359 #endif
3360         }
3361         if (pnew == NULL)
3362                 pnew = zalloc(pvzone);
3363
3364         spin_lock_shared(&pmap->pm_spin);
3365         for (;;) {
3366                 /*
3367                  * Shortcut cache
3368                  */
3369                 pv = pv_entry_lookup(pmap, pindex);
3370                 if (pv == NULL) {
3371                         vm_pindex_t *pmark;
3372
3373                         /*
3374                          * Requires exclusive pmap spinlock
3375                          */
3376                         if (pmap_excl == 0) {
3377                                 pmap_excl = 1;
3378                                 if (!spin_lock_upgrade_try(&pmap->pm_spin)) {
3379                                         spin_unlock_shared(&pmap->pm_spin);
3380                                         spin_lock(&pmap->pm_spin);
3381                                         continue;
3382                                 }
3383                         }
3384
3385                         /*
3386                          * We need to block if someone is holding our
3387                          * placemarker.  As long as we determine the
3388                          * placemarker has not been aquired we do not
3389                          * need to get it as acquision also requires
3390                          * the pmap spin lock.
3391                          *
3392                          * However, we can race the wakeup.
3393                          */
3394                         pmark = pmap_placemarker_hash(pmap, pindex);
3395
3396                         if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3397                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3398                                 tsleep_interlock(pmark, 0);
3399                                 if (((*pmark ^ pindex) &
3400                                      ~PM_PLACEMARK_WAKEUP) == 0) {
3401                                         spin_unlock(&pmap->pm_spin);
3402                                         tsleep(pmark, PINTERLOCKED, "pvplc", 0);
3403                                         spin_lock(&pmap->pm_spin);
3404                                 }
3405                                 continue;
3406                         }
3407
3408                         /*
3409                          * Setup the new entry
3410                          */
3411                         pnew->pv_pmap = pmap;
3412                         pnew->pv_pindex = pindex;
3413                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
3414 #ifdef PMAP_DEBUG
3415                         pnew->pv_func = func;
3416                         pnew->pv_line = lineno;
3417                         if (pnew->pv_line_lastfree > 0) {
3418                                 pnew->pv_line_lastfree =
3419                                                 -pnew->pv_line_lastfree;
3420                         }
3421 #endif
3422                         pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3423                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
3424                         spin_unlock(&pmap->pm_spin);
3425                         *isnew = 1;
3426
3427                         KKASSERT(pv == NULL);
3428                         return(pnew);
3429                 }
3430
3431                 /*
3432                  * We already have an entry, cleanup the staged pnew if
3433                  * we can get the lock, otherwise block and retry.
3434                  */
3435                 if (__predict_true(_pv_hold_try(pv PMAP_DEBUG_COPY))) {
3436                         if (pmap_excl)
3437                                 spin_unlock(&pmap->pm_spin);
3438                         else
3439                                 spin_unlock_shared(&pmap->pm_spin);
3440 #if 1
3441                         pnew = atomic_swap_ptr((void *)&md->gd_newpv, pnew);
3442                         if (pnew)
3443                                 zfree(pvzone, pnew);
3444 #else
3445                         crit_enter();
3446                         if (md->gd_newpv == NULL)
3447                                 md->gd_newpv = pnew;
3448                         else
3449                                 zfree(pvzone, pnew);
3450                         crit_exit();
3451 #endif
3452                         KKASSERT(pv->pv_pmap == pmap &&
3453                                  pv->pv_pindex == pindex);
3454                         *isnew = 0;
3455                         return(pv);
3456                 }
3457                 if (pmap_excl) {
3458                         spin_unlock(&pmap->pm_spin);
3459                         _pv_lock(pv PMAP_DEBUG_COPY);
3460                         pv_put(pv);
3461                         spin_lock(&pmap->pm_spin);
3462                 } else {
3463                         spin_unlock_shared(&pmap->pm_spin);
3464                         _pv_lock(pv PMAP_DEBUG_COPY);
3465                         pv_put(pv);
3466                         spin_lock_shared(&pmap->pm_spin);
3467                 }
3468         }
3469         /* NOT REACHED */
3470 }
3471
3472 /*
3473  * Find the requested PV entry, returning a locked+held pv or NULL
3474  */
3475 static
3476 pv_entry_t
3477 _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp PMAP_DEBUG_DECL)
3478 {
3479         pv_entry_t pv;
3480         int pmap_excl = 0;
3481
3482         spin_lock_shared(&pmap->pm_spin);
3483         for (;;) {
3484                 /*
3485                  * Shortcut cache
3486                  */
3487                 pv = pv_entry_lookup(pmap, pindex);
3488                 if (pv == NULL) {
3489                         /*
3490                          * Block if there is ANY placemarker.  If we are to
3491                          * return it, we must also aquire the spot, so we
3492                          * have to block even if the placemarker is held on
3493                          * a different address.
3494                          *
3495                          * OPTIMIZATION: If pmarkp is passed as NULL the
3496                          * caller is just probing (or looking for a real
3497                          * pv_entry), and in this case we only need to check
3498                          * to see if the placemarker matches pindex.
3499                          */
3500                         vm_pindex_t *pmark;
3501
3502                         /*
3503                          * Requires exclusive pmap spinlock
3504                          */
3505                         if (pmap_excl == 0) {
3506                                 pmap_excl = 1;
3507                                 if (!spin_lock_upgrade_try(&pmap->pm_spin)) {
3508                                         spin_unlock_shared(&pmap->pm_spin);
3509                                         spin_lock(&pmap->pm_spin);
3510                                         continue;
3511                                 }
3512                         }
3513
3514                         pmark = pmap_placemarker_hash(pmap, pindex);
3515
3516                         if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3517                             ((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3518                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3519                                 tsleep_interlock(pmark, 0);
3520                                 if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3521                                     ((*pmark ^ pindex) &
3522                                      ~PM_PLACEMARK_WAKEUP) == 0) {
3523                                         spin_unlock(&pmap->pm_spin);
3524                                         tsleep(pmark, PINTERLOCKED, "pvpld", 0);
3525                                         spin_lock(&pmap->pm_spin);
3526                                 }
3527                                 continue;
3528                         }
3529                         if (pmarkp) {
3530                                 if (atomic_swap_long(pmark, pindex) !=
3531                                     PM_NOPLACEMARK) {
3532                                         panic("_pv_get: pmark race");
3533                                 }
3534                                 *pmarkp = pmark;
3535                         }
3536                         spin_unlock(&pmap->pm_spin);
3537                         return NULL;
3538                 }
3539                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3540                         pv_cache(pv, pindex);
3541                         if (pmap_excl)
3542                                 spin_unlock(&pmap->pm_spin);
3543                         else
3544                                 spin_unlock_shared(&pmap->pm_spin);
3545                         KKASSERT(pv->pv_pmap == pmap &&
3546                                  pv->pv_pindex == pindex);
3547                         return(pv);
3548                 }
3549                 if (pmap_excl) {
3550                         spin_unlock(&pmap->pm_spin);
3551                         _pv_lock(pv PMAP_DEBUG_COPY);
3552                         pv_put(pv);
3553                         spin_lock(&pmap->pm_spin);
3554                 } else {
3555                         spin_unlock_shared(&pmap->pm_spin);
3556                         _pv_lock(pv PMAP_DEBUG_COPY);
3557                         pv_put(pv);
3558                         spin_lock_shared(&pmap->pm_spin);
3559                 }
3560         }
3561 }
3562
3563 /*
3564  * Lookup, hold, and attempt to lock (pmap,pindex).
3565  *
3566  * If the entry does not exist NULL is returned and *errorp is set to 0
3567  *
3568  * If the entry exists and could be successfully locked it is returned and
3569  * errorp is set to 0.
3570  *
3571  * If the entry exists but could NOT be successfully locked it is returned
3572  * held and *errorp is set to 1.
3573  *
3574  * If the entry is placemarked by someone else NULL is returned and *errorp
3575  * is set to 1.
3576  */
3577 static
3578 pv_entry_t
3579 pv_get_try(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp, int *errorp)
3580 {
3581         pv_entry_t pv;
3582
3583         spin_lock_shared(&pmap->pm_spin);
3584
3585         pv = pv_entry_lookup(pmap, pindex);
3586         if (pv == NULL) {
3587                 vm_pindex_t *pmark;
3588
3589                 pmark = pmap_placemarker_hash(pmap, pindex);
3590
3591                 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3592                         *errorp = 1;
3593                 } else if (pmarkp &&
3594                            atomic_cmpset_long(pmark, PM_NOPLACEMARK, pindex)) {
3595                         *errorp = 0;
3596                 } else {
3597                         /*
3598                          * Can't set a placemark with a NULL pmarkp, or if
3599                          * pmarkp is non-NULL but we failed to set our
3600                          * placemark.
3601                          */
3602                         *errorp = 1;
3603                 }
3604                 if (pmarkp)
3605                         *pmarkp = pmark;
3606                 spin_unlock_shared(&pmap->pm_spin);
3607
3608                 return NULL;
3609         }
3610
3611         /*
3612          * XXX This has problems if the lock is shared, why?
3613          */
3614         if (pv_hold_try(pv)) {
3615                 pv_cache(pv, pindex);   /* overwrite ok (shared lock) */
3616                 spin_unlock_shared(&pmap->pm_spin);
3617                 *errorp = 0;
3618                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3619                 return(pv);     /* lock succeeded */
3620         }
3621         spin_unlock_shared(&pmap->pm_spin);
3622         *errorp = 1;
3623
3624         return (pv);            /* lock failed */
3625 }
3626
3627 /*
3628  * Lock a held pv, keeping the hold count
3629  */
3630 static
3631 void
3632 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3633 {
3634         u_int count;
3635
3636         for (;;) {
3637                 count = pv->pv_hold;
3638                 cpu_ccfence();
3639                 if ((count & PV_HOLD_LOCKED) == 0) {
3640                         if (atomic_cmpset_int(&pv->pv_hold, count,
3641                                               count | PV_HOLD_LOCKED)) {
3642 #ifdef PMAP_DEBUG
3643                                 pv->pv_func = func;
3644                                 pv->pv_line = lineno;
3645 #endif
3646                                 return;
3647                         }
3648                         continue;
3649                 }
3650                 tsleep_interlock(pv, 0);
3651                 if (atomic_cmpset_int(&pv->pv_hold, count,
3652                                       count | PV_HOLD_WAITING)) {
3653 #ifdef PMAP_DEBUG2
3654                         if (pmap_enter_debug > 0) {
3655                                 --pmap_enter_debug;
3656                                 kprintf("pv waiting on %s:%d\n",
3657                                         pv->pv_func, pv->pv_line);
3658                         }
3659 #endif
3660                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
3661                 }
3662                 /* retry */
3663         }
3664 }
3665
3666 /*
3667  * Unlock a held and locked pv, keeping the hold count.
3668  */
3669 static
3670 void
3671 pv_unlock(pv_entry_t pv)
3672 {
3673         u_int count;
3674
3675         for (;;) {
3676                 count = pv->pv_hold;
3677                 cpu_ccfence();
3678                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3679                          (PV_HOLD_LOCKED | 1));
3680                 if (atomic_cmpset_int(&pv->pv_hold, count,
3681                                       count &
3682                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3683                         if (count & PV_HOLD_WAITING)
3684                                 wakeup(pv);
3685                         break;
3686                 }
3687         }
3688 }
3689
3690 /*
3691  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3692  * and the hold count drops to zero we will free it.
3693  *
3694  * Caller should not hold any spin locks.  We are protected from hold races
3695  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3696  * lock held.  A pv cannot be located otherwise.
3697  */
3698 static
3699 void
3700 pv_put(pv_entry_t pv)
3701 {
3702 #ifdef PMAP_DEBUG2
3703         if (pmap_enter_debug > 0) {
3704                 --pmap_enter_debug;
3705                 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3706         }
3707 #endif
3708
3709         /*
3710          * Normal put-aways must have a pv_m associated with the pv,
3711          * but allow the case where the pv has been destructed due
3712          * to pmap_dynamic_delete.
3713          */
3714         KKASSERT(pv->pv_pmap == NULL || pv->pv_m != NULL);
3715
3716         /*
3717          * Fast - shortcut most common condition
3718          */
3719         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3720                 return;
3721
3722         /*
3723          * Slow
3724          */
3725         pv_unlock(pv);
3726         pv_drop(pv);
3727 }
3728
3729 /*
3730  * Remove the pmap association from a pv, require that pv_m already be removed,
3731  * then unlock and drop the pv.  Any pte operations must have already been
3732  * completed.  This call may result in a last-drop which will physically free
3733  * the pv.
3734  *
3735  * Removing the pmap association entails an additional drop.
3736  *
3737  * pv must be exclusively locked on call and will be disposed of on return.
3738  */
3739 static
3740 void
3741 _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL)
3742 {
3743         pmap_t pmap;
3744
3745 #ifdef PMAP_DEBUG
3746         pv->pv_func_lastfree = func;
3747         pv->pv_line_lastfree = lineno;
3748 #endif
3749         KKASSERT(pv->pv_m == NULL);
3750         KKASSERT((pv->pv_hold & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
3751                   (PV_HOLD_LOCKED|1));
3752         if ((pmap = pv->pv_pmap) != NULL) {
3753                 spin_lock(&pmap->pm_spin);
3754                 KKASSERT(pv->pv_pmap == pmap);
3755                 if (pmap->pm_pvhint_pt == pv)
3756                         pmap->pm_pvhint_pt = NULL;
3757                 if (pmap->pm_pvhint_pte == pv)
3758                         pmap->pm_pvhint_pte = NULL;
3759                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3760                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3761                 pv->pv_pmap = NULL;
3762                 pv->pv_pindex = 0;
3763                 spin_unlock(&pmap->pm_spin);
3764
3765                 /*
3766                  * Try to shortcut three atomic ops, otherwise fall through
3767                  * and do it normally.  Drop two refs and the lock all in
3768                  * one go.
3769                  */
3770                 if (pvp)
3771                         vm_page_unwire_quick(pvp->pv_m);
3772                 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3773 #ifdef PMAP_DEBUG2
3774                         if (pmap_enter_debug > 0) {
3775                                 --pmap_enter_debug;
3776                                 kprintf("pv_free: free pv %p\n", pv);
3777                         }
3778 #endif
3779                         zfree(pvzone, pv);
3780                         return;
3781                 }
3782                 pv_drop(pv);    /* ref for pv_pmap */
3783         }
3784         pv_unlock(pv);
3785         pv_drop(pv);
3786 }
3787
3788 /*
3789  * This routine is very drastic, but can save the system
3790  * in a pinch.
3791  */
3792 void
3793 pmap_collect(void)
3794 {
3795         int i;
3796         vm_page_t m;
3797         static int warningdone=0;
3798
3799         if (pmap_pagedaemon_waken == 0)
3800                 return;
3801         pmap_pagedaemon_waken = 0;
3802         if (warningdone < 5) {
3803                 kprintf("pmap_collect: collecting pv entries -- "
3804                         "suggest increasing PMAP_SHPGPERPROC\n");
3805                 warningdone++;
3806         }
3807
3808         for (i = 0; i < vm_page_array_size; i++) {
3809                 m = &vm_page_array[i];
3810                 if (m->wire_count || m->hold_count)
3811                         continue;
3812                 if (vm_page_busy_try(m, TRUE) == 0) {
3813                         if (m->wire_count == 0 && m->hold_count == 0) {
3814                                 pmap_remove_all(m);
3815                         }
3816                         vm_page_wakeup(m);
3817                 }
3818         }
3819 }
3820
3821 /*
3822  * Scan the pmap for active page table entries and issue a callback.
3823  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3824  * its parent page table.
3825  *
3826  * pte_pv will be NULL if the page or page table is unmanaged.
3827  * pt_pv will point to the page table page containing the pte for the page.
3828  *
3829  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3830  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3831  *       process pmap's PD and page to the callback function.  This can be
3832  *       confusing because the pt_pv is really a pd_pv, and the target page
3833  *       table page is simply aliased by the pmap and not owned by it.
3834  *
3835  * It is assumed that the start and end are properly rounded to the page size.
3836  *
3837  * It is assumed that PD pages and above are managed and thus in the RB tree,
3838  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3839  */
3840 struct pmap_scan_info {
3841         struct pmap *pmap;
3842         vm_offset_t sva;
3843         vm_offset_t eva;
3844         vm_pindex_t sva_pd_pindex;
3845         vm_pindex_t eva_pd_pindex;
3846         void (*func)(pmap_t, struct pmap_scan_info *,
3847                      pv_entry_t, vm_pindex_t *, pv_entry_t,
3848                      int, vm_offset_t,
3849                      pt_entry_t *, void *);
3850         void *arg;
3851         pmap_inval_bulk_t bulk_core;
3852         pmap_inval_bulk_t *bulk;
3853         int count;
3854         int stop;
3855 };
3856
3857 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3858 static int pmap_scan_callback(pv_entry_t pv, void *data);
3859
3860 static void
3861 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3862 {
3863         struct pmap *pmap = info->pmap;
3864         pv_entry_t pd_pv;       /* A page directory PV */
3865         pv_entry_t pt_pv;       /* A page table PV */
3866         pv_entry_t pte_pv;      /* A page table entry PV */
3867         vm_pindex_t *pte_placemark;
3868         vm_pindex_t *pt_placemark;
3869         pt_entry_t *ptep;
3870         pt_entry_t oldpte;
3871         struct pv_entry dummy_pv;
3872
3873         info->stop = 0;
3874         if (pmap == NULL)
3875                 return;
3876         if (info->sva == info->eva)
3877                 return;
3878         if (smp_inval) {
3879                 info->bulk = &info->bulk_core;
3880                 pmap_inval_bulk_init(&info->bulk_core, pmap);
3881         } else {
3882                 info->bulk = NULL;
3883         }
3884
3885         /*
3886          * Hold the token for stability; if the pmap is empty we have nothing
3887          * to do.
3888          */
3889 #if 0
3890         if (pmap->pm_stats.resident_count == 0) {
3891                 return;
3892         }
3893 #endif
3894
3895         info->count = 0;
3896
3897         /*
3898          * Special handling for scanning one page, which is a very common
3899          * operation (it is?).
3900          *
3901          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3902          */
3903         if (info->sva + PAGE_SIZE == info->eva) {
3904                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3905                         /*
3906                          * Kernel mappings do not track wire counts on
3907                          * page table pages and only maintain pd_pv and
3908                          * pte_pv levels so pmap_scan() works.
3909                          */
3910                         pt_pv = NULL;
3911                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3912                                         &pte_placemark);
3913                         ptep = vtopte(info->sva);
3914                 } else {
3915                         /*
3916                          * User pages which are unmanaged will not have a
3917                          * pte_pv.  User page table pages which are unmanaged
3918                          * (shared from elsewhere) will also not have a pt_pv.
3919                          * The func() callback will pass both pte_pv and pt_pv
3920                          * as NULL in that case.
3921                          *
3922                          * We hold pte_placemark across the operation for
3923                          * unmanaged pages.
3924                          *
3925                          * WARNING!  We must hold pt_placemark across the
3926                          *           *ptep test to prevent misintepreting
3927                          *           a non-zero *ptep as a shared page
3928                          *           table page.  Hold it across the function
3929                          *           callback as well for SMP safety.
3930                          */
3931                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3932                                         &pte_placemark);
3933                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva),
3934                                         &pt_placemark);
3935                         if (pt_pv == NULL) {
3936                                 KKASSERT(pte_pv == NULL);
3937                                 pd_pv = pv_get(pmap,
3938                                                pmap_pd_pindex(info->sva),
3939                                                NULL);
3940                                 if (pd_pv) {
3941                                         ptep = pv_pte_lookup(pd_pv,
3942                                                     pmap_pt_index(info->sva));
3943                                         if (*ptep) {
3944                                                 info->func(pmap, info,
3945                                                      NULL, pt_placemark,
3946                                                      pd_pv, 1,
3947                                                      info->sva, ptep,
3948                                                      info->arg);
3949                                         } else {
3950                                                 pv_placemarker_wakeup(pmap,
3951                                                                   pt_placemark);
3952                                         }
3953                                         pv_put(pd_pv);
3954                                 } else {
3955                                         pv_placemarker_wakeup(pmap,
3956                                                               pt_placemark);
3957                                 }
3958                                 pv_placemarker_wakeup(pmap, pte_placemark);
3959                                 goto fast_skip;
3960                         }
3961                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3962                 }
3963
3964                 /*
3965                  * NOTE: *ptep can't be ripped out from under us if we hold
3966                  *       pte_pv (or pte_placemark) locked, but bits can
3967                  *       change.
3968                  */
3969                 oldpte = *ptep;
3970                 cpu_ccfence();
3971                 if (oldpte == 0) {
3972                         KKASSERT(pte_pv == NULL);
3973                         pv_placemarker_wakeup(pmap, pte_placemark);
3974                 } else if (pte_pv) {
3975                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3976                                            pmap->pmap_bits[PG_V_IDX])) ==
3977                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
3978                                  pmap->pmap_bits[PG_V_IDX]),
3979                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p",
3980                             *ptep, oldpte, info->sva, pte_pv));
3981                         info->func(pmap, info, pte_pv, NULL, pt_pv, 0,
3982                                    info->sva, ptep, info->arg);
3983                 } else {
3984                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3985                                            pmap->pmap_bits[PG_V_IDX])) ==
3986                             pmap->pmap_bits[PG_V_IDX],
3987                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
3988                             *ptep, oldpte, info->sva));
3989                         info->func(pmap, info, NULL, pte_placemark, pt_pv, 0,
3990                                    info->sva, ptep, info->arg);
3991                 }
3992                 if (pt_pv)
3993                         pv_put(pt_pv);
3994 fast_skip:
3995                 pmap_inval_bulk_flush(info->bulk);
3996                 return;
3997         }
3998
3999         /*
4000          * Nominal scan case, RB_SCAN() for PD pages and iterate from
4001          * there.
4002          *
4003          * WARNING! eva can overflow our standard ((N + mask) >> bits)
4004          *          bounds, resulting in a pd_pindex of 0.  To solve the
4005          *          problem we use an inclusive range.
4006          */
4007         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
4008         info->eva_pd_pindex = pmap_pd_pindex(info->eva - PAGE_SIZE);
4009
4010         if (info->sva >= VM_MAX_USER_ADDRESS) {
4011                 /*
4012                  * The kernel does not currently maintain any pv_entry's for
4013                  * higher-level page tables.
4014                  */
4015                 bzero(&dummy_pv, sizeof(dummy_pv));
4016                 dummy_pv.pv_pindex = info->sva_pd_pindex;
4017                 spin_lock(&pmap->pm_spin);
4018                 while (dummy_pv.pv_pindex <= info->eva_pd_pindex) {
4019                         pmap_scan_callback(&dummy_pv, info);
4020                         ++dummy_pv.pv_pindex;
4021                         if (dummy_pv.pv_pindex < info->sva_pd_pindex) /*wrap*/
4022                                 break;
4023                 }
4024                 spin_unlock(&pmap->pm_spin);
4025         } else {
4026                 /*
4027                  * User page tables maintain local PML4, PDP, and PD
4028                  * pv_entry's at the very least.  PT pv's might be
4029                  * unmanaged and thus not exist.  PTE pv's might be
4030                  * unmanaged and thus not exist.
4031                  */
4032                 spin_lock(&pmap->pm_spin);
4033                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot, pmap_scan_cmp,
4034                                          pmap_scan_callback, info);
4035                 spin_unlock(&pmap->pm_spin);
4036         }
4037         pmap_inval_bulk_flush(info->bulk);
4038 }
4039
4040 /*
4041  * WARNING! pmap->pm_spin held
4042  *
4043  * WARNING! eva can overflow our standard ((N + mask) >> bits)
4044  *          bounds, resulting in a pd_pindex of 0.  To solve the
4045  *          problem we use an inclusive range.
4046  */
4047 static int
4048 pmap_scan_cmp(pv_entry_t pv, void *data)
4049 {
4050         struct pmap_scan_info *info = data;
4051         if (pv->pv_pindex < info->sva_pd_pindex)
4052                 return(-1);
4053         if (pv->pv_pindex > info->eva_pd_pindex)
4054                 return(1);
4055         return(0);
4056 }
4057
4058 /*
4059  * pmap_scan() by PDs
4060  *
4061  * WARNING! pmap->pm_spin held
4062  */
4063 static int
4064 pmap_scan_callback(pv_entry_t pv, void *data)
4065 {
4066         struct pmap_scan_info *info = data;
4067         struct pmap *pmap = info->pmap;
4068         pv_entry_t pd_pv;       /* A page directory PV */
4069         pv_entry_t pt_pv;       /* A page table PV */
4070         vm_pindex_t *pt_placemark;
4071         pt_entry_t *ptep;
4072         pt_entry_t oldpte;
4073         vm_offset_t sva;
4074         vm_offset_t eva;
4075         vm_offset_t va_next;
4076         vm_pindex_t pd_pindex;
4077         int error;
4078
4079         /*
4080          * Stop if requested
4081          */
4082         if (info->stop)
4083                 return -1;
4084
4085         /*
4086          * Pull the PD pindex from the pv before releasing the spinlock.
4087          *
4088          * WARNING: pv is faked for kernel pmap scans.
4089          */
4090         pd_pindex = pv->pv_pindex;
4091         spin_unlock(&pmap->pm_spin);
4092         pv = NULL;      /* invalid after spinlock unlocked */
4093
4094         /*
4095          * Calculate the page range within the PD.  SIMPLE pmaps are
4096          * direct-mapped for the entire 2^64 address space.  Normal pmaps
4097          * reflect the user and kernel address space which requires
4098          * cannonicalization w/regards to converting pd_pindex's back
4099          * into addresses.
4100          */
4101         sva = (pd_pindex - pmap_pd_pindex(0)) << PDPSHIFT;
4102         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
4103             (sva & PML4_SIGNMASK)) {
4104                 sva |= PML4_SIGNMASK;
4105         }
4106         eva = sva + NBPDP;      /* can overflow */
4107         if (sva < info->sva)
4108                 sva = info->sva;
4109         if (eva < info->sva || eva > info->eva)
4110                 eva = info->eva;
4111
4112         /*
4113          * NOTE: kernel mappings do not track page table pages, only
4114          *       terminal pages.
4115          *
4116          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
4117          *       However, for the scan to be efficient we try to
4118          *       cache items top-down.
4119          */
4120         pd_pv = NULL;
4121         pt_pv = NULL;
4122
4123         for (; sva < eva; sva = va_next) {
4124                 if (info->stop)
4125                         break;
4126                 if (sva >= VM_MAX_USER_ADDRESS) {
4127                         if (pt_pv) {
4128                                 pv_put(pt_pv);
4129                                 pt_pv = NULL;
4130                         }
4131                         goto kernel_skip;
4132                 }
4133
4134                 /*
4135                  * PD cache, scan shortcut if it doesn't exist.
4136                  */
4137                 if (pd_pv == NULL) {
4138                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4139                 } else if (pd_pv->pv_pmap != pmap ||
4140                            pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
4141                         pv_put(pd_pv);
4142                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4143                 }
4144                 if (pd_pv == NULL) {
4145                         va_next = (sva + NBPDP) & ~PDPMASK;
4146                         if (va_next < sva)
4147                                 va_next = eva;
4148                         continue;
4149                 }
4150
4151                 /*
4152                  * PT cache
4153                  *
4154                  * NOTE: The cached pt_pv can be removed from the pmap when
4155                  *       pmap_dynamic_delete is enabled.
4156                  */
4157                 if (pt_pv && (pt_pv->pv_pmap != pmap ||
4158                               pt_pv->pv_pindex != pmap_pt_pindex(sva))) {
4159                         pv_put(pt_pv);
4160                         pt_pv = NULL;
4161                 }
4162                 if (pt_pv == NULL) {
4163                         pt_pv = pv_get_try(pmap, pmap_pt_pindex(sva),
4164                                            &pt_placemark, &error);
4165                         if (error) {
4166                                 pv_put(pd_pv);  /* lock order */
4167                                 pd_pv = NULL;
4168                                 if (pt_pv) {
4169                                         pv_lock(pt_pv);
4170                                         pv_put(pt_pv);
4171                                         pt_pv = NULL;
4172                                 } else {
4173                                         pv_placemarker_wait(pmap, pt_placemark);
4174                                 }
4175                                 va_next = sva;
4176                                 continue;
4177                         }
4178                         /* may have to re-check later if pt_pv is NULL here */
4179                 }
4180
4181                 /*
4182                  * If pt_pv is NULL we either have an shared page table
4183                  * page and must issue a callback specific to that case,
4184                  * or there is no page table page.
4185                  *
4186                  * Either way we can skip the page table page.
4187                  *
4188                  * WARNING! pt_pv can also be NULL due to a pv creation
4189                  *          race where we find it to be NULL and then
4190                  *          later see a pte_pv.  But its possible the pt_pv
4191                  *          got created inbetween the two operations, so
4192                  *          we must check.
4193                  */
4194                 if (pt_pv == NULL) {
4195                         /*
4196                          * Possible unmanaged (shared from another pmap)
4197                          * page table page.
4198                          *
4199                          * WARNING!  We must hold pt_placemark across the
4200                          *           *ptep test to prevent misintepreting
4201                          *           a non-zero *ptep as a shared page
4202                          *           table page.  Hold it across the function
4203                          *           callback as well for SMP safety.
4204                          */
4205                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
4206                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
4207                                 info->func(pmap, info, NULL, pt_placemark,
4208                                            pd_pv, 1,
4209                                            sva, ptep, info->arg);
4210                         } else {
4211                                 pv_placemarker_wakeup(pmap, pt_placemark);
4212                         }
4213
4214                         /*
4215                          * Done, move to next page table page.
4216                          */
4217                         va_next = (sva + NBPDR) & ~PDRMASK;
4218                         if (va_next < sva)
4219                                 va_next = eva;
4220                         continue;
4221                 }
4222
4223                 /*
4224                  * From this point in the loop testing pt_pv for non-NULL
4225                  * means we are in UVM, else if it is NULL we are in KVM.
4226                  *
4227                  * Limit our scan to either the end of the va represented
4228                  * by the current page table page, or to the end of the
4229                  * range being removed.
4230                  */
4231 kernel_skip:
4232                 va_next = (sva + NBPDR) & ~PDRMASK;
4233                 if (va_next < sva)
4234                         va_next = eva;
4235                 if (va_next > eva)
4236                         va_next = eva;
4237
4238                 /*
4239                  * Scan the page table for pages.  Some pages may not be
4240                  * managed (might not have a pv_entry).
4241                  *
4242                  * There is no page table management for kernel pages so
4243                  * pt_pv will be NULL in that case, but otherwise pt_pv
4244                  * is non-NULL, locked, and referenced.
4245                  */
4246
4247                 /*
4248                  * At this point a non-NULL pt_pv means a UVA, and a NULL
4249                  * pt_pv means a KVA.
4250                  */
4251                 if (pt_pv)
4252                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
4253                 else
4254                         ptep = vtopte(sva);
4255
4256                 while (sva < va_next) {
4257                         pv_entry_t pte_pv;
4258                         vm_pindex_t *pte_placemark;
4259
4260                         /*
4261                          * Yield every 64 pages, stop if requested.
4262                          */
4263                         if ((++info->count & 63) == 0)
4264                                 lwkt_user_yield();
4265                         if (info->stop)
4266                                 break;
4267
4268                         /*
4269                          * We can shortcut our scan if *ptep == 0.  This is
4270                          * an unlocked check.
4271                          */
4272                         if (*ptep == 0) {
4273                                 sva += PAGE_SIZE;
4274                                 ++ptep;
4275                                 continue;
4276                         }
4277                         cpu_ccfence();
4278
4279                         /*
4280                          * Acquire the related pte_pv, if any.  If *ptep == 0
4281                          * the related pte_pv should not exist, but if *ptep
4282                          * is not zero the pte_pv may or may not exist (e.g.
4283                          * will not exist for an unmanaged page).
4284                          *
4285                          * However a multitude of races are possible here
4286                          * so if we cannot lock definite state we clean out
4287                          * our cache and break the inner while() loop to
4288                          * force a loop up to the top of the for().
4289                          *
4290                          * XXX unlock/relock pd_pv, pt_pv, and re-test their
4291                          *     validity instead of looping up?
4292                          */
4293                         pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
4294                                             &pte_placemark, &error);
4295                         if (error) {
4296                                 pv_put(pd_pv);          /* lock order */
4297                                 pd_pv = NULL;
4298                                 if (pt_pv) {
4299                                         pv_put(pt_pv);  /* lock order */
4300                                         pt_pv = NULL;
4301                                 }
4302                                 if (pte_pv) {           /* block */
4303                                         pv_lock(pte_pv);
4304                                         pv_put(pte_pv);
4305                                         pte_pv = NULL;
4306                                 } else {
4307                                         pv_placemarker_wait(pmap,
4308                                                         pte_placemark);
4309                                 }
4310                                 va_next = sva;          /* retry */
4311                                 break;
4312                         }
4313
4314                         /*
4315                          * Reload *ptep after successfully locking the
4316                          * pindex.  If *ptep == 0 we had better NOT have a
4317                          * pte_pv.
4318                          */
4319                         cpu_ccfence();
4320                         oldpte = *ptep;
4321                         if (oldpte == 0) {
4322                                 if (pte_pv) {
4323                                         kprintf("Unexpected non-NULL pte_pv "
4324                                                 "%p pt_pv %p "
4325                                                 "*ptep = %016lx/%016lx\n",
4326                                                 pte_pv, pt_pv, *ptep, oldpte);
4327                                         panic("Unexpected non-NULL pte_pv");
4328                                 } else {
4329                                         pv_placemarker_wakeup(pmap, pte_placemark);
4330                                 }
4331                                 sva += PAGE_SIZE;
4332                                 ++ptep;
4333                                 continue;
4334                         }
4335
4336                         /*
4337                          * We can't hold pd_pv across the callback (because
4338                          * we don't pass it to the callback and the callback
4339                          * might deadlock)
4340                          */
4341                         if (pd_pv) {
4342                                 vm_page_wire_quick(pd_pv->pv_m);
4343                                 pv_unlock(pd_pv);
4344                         }
4345
4346                         /*
4347                          * Ready for the callback.  The locked pte_pv (if any)
4348                          * is consumed by the callback.  pte_pv will exist if
4349                          * the page is managed, and will not exist if it
4350                          * isn't.
4351                          */
4352                         if (oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4353                                 /*
4354                                  * Managed pte
4355                                  */
4356                                 KASSERT(pte_pv &&
4357                                          (oldpte & pmap->pmap_bits[PG_V_IDX]),
4358                                     ("badC *ptep %016lx/%016lx sva %016lx "
4359                                     "pte_pv %p",
4360                                     *ptep, oldpte, sva, pte_pv));
4361                                 /*
4362                                  * We must unlock pd_pv across the callback
4363                                  * to avoid deadlocks on any recursive
4364                                  * disposal.  Re-check that it still exists
4365                                  * after re-locking.
4366                                  *
4367                                  * Call target disposes of pte_pv and may
4368                                  * destroy but will not dispose of pt_pv.
4369                                  */
4370                                 info->func(pmap, info, pte_pv, NULL,
4371                                            pt_pv, 0,
4372                                            sva, ptep, info->arg);
4373                         } else {
4374                                 /*
4375                                  * Unmanaged pte
4376                                  *
4377                                  * We must unlock pd_pv across the callback
4378                                  * to avoid deadlocks on any recursive
4379                                  * disposal.  Re-check that it still exists
4380                                  * after re-locking.
4381                                  *
4382                                  * Call target disposes of pte_pv or
4383                                  * pte_placemark and may destroy but will
4384                                  * not dispose of pt_pv.
4385                                  */
4386                                 KASSERT(pte_pv == NULL &&
4387                                         (oldpte & pmap->pmap_bits[PG_V_IDX]),
4388                                     ("badD *ptep %016lx/%016lx sva %016lx "
4389                                     "pte_pv %p pte_pv->pv_m %p ",
4390                                      *ptep, oldpte, sva,
4391                                      pte_pv, (pte_pv ? pte_pv->pv_m : NULL)));
4392                                 if (pte_pv)
4393                                         kprintf("RaceD\n");
4394                                 if (pte_pv) {
4395                                         info->func(pmap, info,
4396                                                    pte_pv, NULL,
4397                                                    pt_pv, 0,
4398                                                    sva, ptep, info->arg);
4399                                 } else {
4400                                         info->func(pmap, info,
4401                                                    NULL, pte_placemark,
4402                                                    pt_pv, 0,
4403                                                    sva, ptep, info->arg);
4404                                 }
4405                         }
4406                         if (pd_pv) {
4407                                 pv_lock(pd_pv);
4408                                 vm_page_unwire_quick(pd_pv->pv_m);
4409                                 if (pd_pv->pv_pmap == NULL) {
4410                                         va_next = sva;          /* retry */
4411                                         break;
4412                                 }
4413                         }
4414
4415                         /*
4416                          * NOTE: The cached pt_pv can be removed from the
4417                          *       pmap when pmap_dynamic_delete is enabled,
4418                          *       which will cause ptep to become stale.
4419                          *
4420                          *       This also means that no pages remain under
4421                          *       the PT, so we can just break out of the inner
4422                          *       loop and let the outer loop clean everything
4423                          *       up.
4424                          */
4425                         if (pt_pv && pt_pv->pv_pmap != pmap)
4426                                 break;
4427                         pte_pv = NULL;
4428                         sva += PAGE_SIZE;
4429                         ++ptep;
4430                 }
4431         }
4432         if (pd_pv) {
4433                 pv_put(pd_pv);
4434                 pd_pv = NULL;
4435         }
4436         if (pt_pv) {
4437                 pv_put(pt_pv);
4438                 pt_pv = NULL;
4439         }
4440         if ((++info->count & 7) == 0)
4441                 lwkt_user_yield();
4442
4443         /*
4444          * Relock before returning.
4445          */
4446         spin_lock(&pmap->pm_spin);
4447         return (0);
4448 }
4449
4450 void
4451 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4452 {
4453         struct pmap_scan_info info;
4454
4455         info.pmap = pmap;
4456         info.sva = sva;
4457         info.eva = eva;
4458         info.func = pmap_remove_callback;
4459         info.arg = NULL;
4460         pmap_scan(&info, 1);
4461 #if 0
4462         cpu_invltlb();
4463         if (eva - sva < 1024*1024) {
4464                 while (sva < eva) {
4465                         cpu_invlpg((void *)sva);
4466                         sva += PAGE_SIZE;
4467                 }
4468         }
4469 #endif
4470 }
4471
4472 static void
4473 pmap_remove_noinval(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, 0);
4483 }
4484
4485 static void
4486 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4487                      pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4488                      pv_entry_t pt_pv, int sharept,
4489                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4490 {
4491         pt_entry_t pte;
4492
4493         if (pte_pv) {
4494                 /*
4495                  * Managed entry
4496                  *
4497                  * This will also drop pt_pv's wire_count. Note that
4498                  * terminal pages are not wired based on mmu presence.
4499                  *
4500                  * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4501                  */
4502                 KKASSERT(pte_pv->pv_m != NULL);
4503                 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4504                 pte_pv = NULL;  /* safety */
4505
4506                 /*
4507                  * Recursively destroy higher-level page tables.
4508                  *
4509                  * This is optional.  If we do not, they will still
4510                  * be destroyed when the process exits.
4511                  *
4512                  * NOTE: Do not destroy pv_entry's with extra hold refs,
4513                  *       a caller may have unlocked it and intends to
4514                  *       continue to use it.
4515                  */
4516                 if (pmap_dynamic_delete &&
4517                     pt_pv &&
4518                     pt_pv->pv_m &&
4519                     pt_pv->pv_m->wire_count == 1 &&
4520                     (pt_pv->pv_hold & PV_HOLD_MASK) == 2 &&
4521                     pt_pv->pv_pindex != pmap_pml4_pindex()) {
4522                         if (pmap_dynamic_delete == 2)
4523                                 kprintf("B %jd %08x\n", pt_pv->pv_pindex, pt_pv->pv_hold);
4524                         pv_hold(pt_pv); /* extra hold */
4525                         pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4526                         pv_lock(pt_pv); /* prior extra hold + relock */
4527                 }
4528         } else if (sharept == 0) {
4529                 /*
4530                  * Unmanaged pte (pte_placemark is non-NULL)
4531                  *
4532                  * pt_pv's wire_count is still bumped by unmanaged pages
4533                  * so we must decrement it manually.
4534                  *
4535                  * We have to unwire the target page table page.
4536                  */
4537                 pte = pmap_inval_bulk(info->bulk, va, ptep, 0);
4538                 if (pte & pmap->pmap_bits[PG_W_IDX])
4539                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4540                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4541                 if (vm_page_unwire_quick(pt_pv->pv_m))
4542                         panic("pmap_remove: insufficient wirecount");
4543                 pv_placemarker_wakeup(pmap, pte_placemark);
4544         } else {
4545                 /*
4546                  * Unmanaged page table (pt, pd, or pdp. Not pte) for
4547                  * a shared page table.
4548                  *
4549                  * pt_pv is actually the pd_pv for our pmap (not the shared
4550                  * object pmap).
4551                  *
4552                  * We have to unwire the target page table page and we
4553                  * have to unwire our page directory page.
4554                  *
4555                  * It is unclear how we can invalidate a segment so we
4556                  * invalidate -1 which invlidates the tlb.
4557                  */
4558                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4559                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4560                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4561                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4562                         panic("pmap_remove: shared pgtable1 bad wirecount");
4563                 if (vm_page_unwire_quick(pt_pv->pv_m))
4564                         panic("pmap_remove: shared pgtable2 bad wirecount");
4565                 pv_placemarker_wakeup(pmap, pte_placemark);
4566         }
4567 }
4568
4569 /*
4570  * Removes this physical page from all physical maps in which it resides.
4571  * Reflects back modify bits to the pager.
4572  *
4573  * This routine may not be called from an interrupt.
4574  */
4575 static
4576 void
4577 pmap_remove_all(vm_page_t m)
4578 {
4579         pv_entry_t pv;
4580         pmap_inval_bulk_t bulk;
4581
4582         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4583                 return;
4584
4585         vm_page_spin_lock(m);
4586         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4587                 KKASSERT(pv->pv_m == m);
4588                 if (pv_hold_try(pv)) {
4589                         vm_page_spin_unlock(m);
4590                 } else {
4591                         vm_page_spin_unlock(m);
4592                         pv_lock(pv);
4593                         pv_put(pv);
4594                         vm_page_spin_lock(m);
4595                         continue;
4596                 }
4597                 KKASSERT(pv->pv_pmap && pv->pv_m == m);
4598
4599                 /*
4600                  * Holding no spinlocks, pv is locked.  Once we scrap
4601                  * pv we can no longer use it as a list iterator (but
4602                  * we are doing a TAILQ_FIRST() so we are ok).
4603                  */
4604                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4605                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4606                 pv = NULL;      /* safety */
4607                 pmap_inval_bulk_flush(&bulk);
4608                 vm_page_spin_lock(m);
4609         }
4610         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4611         vm_page_spin_unlock(m);
4612 }
4613
4614 /*
4615  * Removes the page from a particular pmap
4616  */
4617 void
4618 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4619 {
4620         pv_entry_t pv;
4621         pmap_inval_bulk_t bulk;
4622
4623         if (!pmap_initialized)
4624                 return;
4625
4626 again:
4627         vm_page_spin_lock(m);
4628         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4629                 if (pv->pv_pmap != pmap)
4630                         continue;
4631                 KKASSERT(pv->pv_m == m);
4632                 if (pv_hold_try(pv)) {
4633                         vm_page_spin_unlock(m);
4634                 } else {
4635                         vm_page_spin_unlock(m);
4636                         pv_lock(pv);
4637                         pv_put(pv);
4638                         goto again;
4639                 }
4640                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
4641
4642                 /*
4643                  * Holding no spinlocks, pv is locked.  Once gone it can't
4644                  * be used as an iterator.  In fact, because we couldn't
4645                  * necessarily lock it atomically it may have moved within
4646                  * the list and ALSO cannot be used as an iterator.
4647                  */
4648                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4649                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4650                 pv = NULL;      /* safety */
4651                 pmap_inval_bulk_flush(&bulk);
4652                 goto again;
4653         }
4654         vm_page_spin_unlock(m);
4655 }
4656
4657 /*
4658  * Set the physical protection on the specified range of this map
4659  * as requested.  This function is typically only used for debug watchpoints
4660  * and COW pages.
4661  *
4662  * This function may not be called from an interrupt if the map is
4663  * not the kernel_pmap.
4664  *
4665  * NOTE!  For shared page table pages we just unmap the page.
4666  */
4667 void
4668 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4669 {
4670         struct pmap_scan_info info;
4671         /* JG review for NX */
4672
4673         if (pmap == NULL)
4674                 return;
4675         if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == VM_PROT_NONE) {
4676                 pmap_remove(pmap, sva, eva);
4677                 return;
4678         }
4679         if (prot & VM_PROT_WRITE)
4680                 return;
4681         info.pmap = pmap;
4682         info.sva = sva;
4683         info.eva = eva;
4684         info.func = pmap_protect_callback;
4685         info.arg = &prot;
4686         pmap_scan(&info, 1);
4687 }
4688
4689 static
4690 void
4691 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4692                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4693                       pv_entry_t pt_pv, int sharept,
4694                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4695 {
4696         pt_entry_t pbits;
4697         pt_entry_t cbits;
4698         pt_entry_t pte;
4699         vm_page_t m;
4700
4701 again:
4702         pbits = *ptep;
4703         cbits = pbits;
4704         if (pte_pv) {
4705                 KKASSERT(pte_pv->pv_m != NULL);
4706                 m = NULL;
4707                 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4708                         if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4709                                 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4710                                 KKASSERT(m == pte_pv->pv_m);
4711                                 vm_page_flag_set(m, PG_REFERENCED);
4712                         }
4713                         cbits &= ~pmap->pmap_bits[PG_A_IDX];
4714                 }
4715                 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4716                         if (pmap_track_modified(pte_pv->pv_pindex)) {
4717                                 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4718                                         if (m == NULL) {
4719                                                 m = PHYS_TO_VM_PAGE(pbits &
4720                                                                     PG_FRAME);
4721                                         }
4722                                         vm_page_dirty(m);
4723                                 }
4724                                 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4725                         }
4726                 }
4727         } else if (sharept) {
4728                 /*
4729                  * Unmanaged page table, pt_pv is actually the pd_pv
4730                  * for our pmap (not the object's shared pmap).
4731                  *
4732                  * When asked to protect something in a shared page table
4733                  * page we just unmap the page table page.  We have to
4734                  * invalidate the tlb in this situation.
4735                  *
4736                  * XXX Warning, shared page tables will not be used for
4737                  * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4738                  * so PHYS_TO_VM_PAGE() should be safe here.
4739                  */
4740                 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4741                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4742                         panic("pmap_protect: pgtable1 pg bad wirecount");
4743                 if (vm_page_unwire_quick(pt_pv->pv_m))
4744                         panic("pmap_protect: pgtable2 pg bad wirecount");
4745                 ptep = NULL;
4746         }
4747         /* else unmanaged page, adjust bits, no wire changes */
4748
4749         if (ptep) {
4750                 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4751 #ifdef PMAP_DEBUG2
4752                 if (pmap_enter_debug > 0) {
4753                         --pmap_enter_debug;
4754                         kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4755                                 "pt_pv=%p cbits=%08lx\n",
4756                                 va, ptep, pte_pv,
4757                                 pt_pv, cbits
4758                         );
4759                 }
4760 #endif
4761                 if (pbits != cbits) {
4762                         vm_offset_t xva;
4763
4764                         xva = (sharept) ? (vm_offset_t)-1 : va;
4765                         if (!pmap_inval_smp_cmpset(pmap, xva,
4766                                                    ptep, pbits, cbits)) {
4767                                 goto again;
4768                         }
4769                 }
4770         }
4771         if (pte_pv)
4772                 pv_put(pte_pv);
4773         else
4774                 pv_placemarker_wakeup(pmap, pte_placemark);
4775 }
4776
4777 /*
4778  * Insert the vm_page (m) at the virtual address (va), replacing any prior
4779  * mapping at that address.  Set protection and wiring as requested.
4780  *
4781  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4782  * possible.  If it is we enter the page into the appropriate shared pmap
4783  * hanging off the related VM object instead of the passed pmap, then we
4784  * share the page table page from the VM object's pmap into the current pmap.
4785  *
4786  * NOTE: This routine MUST insert the page into the pmap now, it cannot
4787  *       lazy-evaluate.
4788  *
4789  * NOTE: If (m) is PG_UNMANAGED it may also be a temporary fake vm_page_t.
4790  *       never record it.
4791  */
4792 void
4793 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4794            boolean_t wired, vm_map_entry_t entry)
4795 {
4796         pv_entry_t pt_pv;       /* page table */
4797         pv_entry_t pte_pv;      /* page table entry */
4798         vm_pindex_t *pte_placemark;
4799         pt_entry_t *ptep;
4800         vm_paddr_t opa;
4801         pt_entry_t origpte, newpte;
4802         vm_paddr_t pa;
4803
4804         if (pmap == NULL)
4805                 return;
4806         va = trunc_page(va);
4807 #ifdef PMAP_DIAGNOSTIC
4808         if (va >= KvaEnd)
4809                 panic("pmap_enter: toobig");
4810         if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4811                 panic("pmap_enter: invalid to pmap_enter page table "
4812                       "pages (va: 0x%lx)", va);
4813 #endif
4814         if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4815                 kprintf("Warning: pmap_enter called on UVA with "
4816                         "kernel_pmap\n");
4817 #ifdef DDB
4818                 db_print_backtrace();
4819 #endif
4820         }
4821         if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4822                 kprintf("Warning: pmap_enter called on KVA without"
4823                         "kernel_pmap\n");
4824 #ifdef DDB
4825                 db_print_backtrace();
4826 #endif
4827         }
4828
4829         /*
4830          * Get locked PV entries for our new page table entry (pte_pv or
4831          * pte_placemark) and for its parent page table (pt_pv).  We need
4832          * the parent so we can resolve the location of the ptep.
4833          *
4834          * Only hardware MMU actions can modify the ptep out from
4835          * under us.
4836          *
4837          * if (m) is fictitious or unmanaged we do not create a managing
4838          * pte_pv for it.  Any pre-existing page's management state must
4839          * match (avoiding code complexity).
4840          *
4841          * If the pmap is still being initialized we assume existing
4842          * page tables.
4843          *
4844          * Kernel mapppings do not track page table pages (i.e. pt_pv).
4845          *
4846          * WARNING! If replacing a managed mapping with an unmanaged mapping
4847          *          pte_pv will wind up being non-NULL and must be handled
4848          *          below.
4849          */
4850         if (pmap_initialized == FALSE) {
4851                 pte_pv = NULL;
4852                 pt_pv = NULL;
4853                 pte_placemark = NULL;
4854                 ptep = vtopte(va);
4855                 origpte = *ptep;
4856         } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4857                 pte_pv = pv_get(pmap, pmap_pte_pindex(va), &pte_placemark);
4858                 KKASSERT(pte_pv == NULL);
4859                 if (va >= VM_MAX_USER_ADDRESS) {
4860                         pt_pv = NULL;
4861                         ptep = vtopte(va);
4862                 } else {
4863                         pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4864                                                   NULL, entry, va);
4865                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4866                 }
4867                 origpte = *ptep;
4868                 cpu_ccfence();
4869                 KASSERT(origpte == 0 ||
4870                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4871                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4872         } else {
4873                 if (va >= VM_MAX_USER_ADDRESS) {
4874                         /*
4875                          * Kernel map, pv_entry-tracked.
4876                          */
4877                         pt_pv = NULL;
4878                         pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4879                         ptep = vtopte(va);
4880                 } else {
4881                         /*
4882                          * User map
4883                          */
4884                         pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4885                                                    &pt_pv, entry, va);
4886                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4887                 }
4888                 pte_placemark = NULL;   /* safety */
4889                 origpte = *ptep;
4890                 cpu_ccfence();
4891                 KASSERT(origpte == 0 ||
4892                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4893                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4894         }
4895
4896         pa = VM_PAGE_TO_PHYS(m);
4897         opa = origpte & PG_FRAME;
4898
4899         /*
4900          * Calculate the new PTE.  Note that pte_pv alone does not mean
4901          * the new pte_pv is managed, it could exist because the old pte
4902          * was managed even if the new one is not.
4903          */
4904         newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4905                  pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4906         if (wired)
4907                 newpte |= pmap->pmap_bits[PG_W_IDX];
4908         if (va < VM_MAX_USER_ADDRESS)
4909                 newpte |= pmap->pmap_bits[PG_U_IDX];
4910         if (pte_pv && (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) == 0)
4911                 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4912 //      if (pmap == &kernel_pmap)
4913 //              newpte |= pgeflag;
4914         newpte |= pmap->pmap_cache_bits[m->pat_mode];
4915         if (m->flags & PG_FICTITIOUS)
4916                 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4917
4918         /*
4919          * It is possible for multiple faults to occur in threaded
4920          * environments, the existing pte might be correct.
4921          */
4922         if (((origpte ^ newpte) &
4923             ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4924                           pmap->pmap_bits[PG_A_IDX])) == 0) {
4925                 goto done;
4926         }
4927
4928         /*
4929          * Ok, either the address changed or the protection or wiring
4930          * changed.
4931          *
4932          * Clear the current entry, interlocking the removal.  For managed
4933          * pte's this will also flush the modified state to the vm_page.
4934          * Atomic ops are mandatory in order to ensure that PG_M events are
4935          * not lost during any transition.
4936          *
4937          * WARNING: The caller has busied the new page but not the original
4938          *          vm_page which we are trying to replace.  Because we hold
4939          *          the pte_pv lock, but have not busied the page, PG bits
4940          *          can be cleared out from under us.
4941          */
4942         if (opa) {
4943                 if (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4944                         /*
4945                          * Old page was managed.  Expect pte_pv to exist.
4946                          * (it might also exist if the old page was unmanaged).
4947                          *
4948                          * NOTE: pt_pv won't exist for a kernel page
4949                          *       (managed or otherwise).
4950                          *
4951                          * NOTE: We may be reusing the pte_pv so we do not
4952                          *       destroy it in pmap_remove_pv_pte().
4953                          */
4954                         KKASSERT(pte_pv && pte_pv->pv_m);
4955                         if (prot & VM_PROT_NOSYNC) {
4956                                 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4957                         } else {
4958                                 pmap_inval_bulk_t bulk;
4959
4960                                 pmap_inval_bulk_init(&bulk, pmap);
4961                                 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4962                                 pmap_inval_bulk_flush(&bulk);
4963                         }
4964                         pmap_remove_pv_page(pte_pv);
4965                         /* will either set pte_pv->pv_m or pv_free() later */
4966                 } else {
4967                         /*
4968                          * Old page was not managed.  If we have a pte_pv
4969                          * it better not have a pv_m assigned to it.  If the
4970                          * new page is managed the pte_pv will be destroyed
4971                          * near the end (we need its interlock).
4972                          *
4973                          * NOTE: We leave the wire count on the PT page
4974                          *       intact for the followup enter, but adjust
4975                          *       the wired-pages count on the pmap.
4976                          */
4977                         KKASSERT(pte_pv == NULL);
4978                         if (prot & VM_PROT_NOSYNC) {
4979                                 /*
4980                                  * NOSYNC (no mmu sync) requested.
4981                                  */
4982                                 (void)pte_load_clear(ptep);
4983                                 cpu_invlpg((void *)va);
4984                         } else {
4985                                 /*
4986                                  * Nominal SYNC
4987                                  */
4988                                 pmap_inval_smp(pmap, va, 1, ptep, 0);
4989                         }
4990
4991                         /*
4992                          * We must adjust pm_stats manually for unmanaged
4993                          * pages.
4994                          */
4995                         if (pt_pv) {
4996                                 atomic_add_long(&pmap->pm_stats.
4997                                                 resident_count, -1);
4998                         }
4999                         if (origpte & pmap->pmap_bits[PG_W_IDX]) {
5000                                 atomic_add_long(&pmap->pm_stats.
5001                                                 wired_count, -1);
5002                         }
5003                 }
5004                 KKASSERT(*ptep == 0);
5005         }
5006
5007 #ifdef PMAP_DEBUG2
5008         if (pmap_enter_debug > 0) {
5009                 --pmap_enter_debug;
5010                 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
5011                         " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
5012                         va, m,
5013                         origpte, newpte, ptep,
5014                         pte_pv, pt_pv, opa, prot);
5015         }
5016 #endif
5017
5018         if ((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5019                 /*
5020                  * Entering an unmanaged page.  We must wire the pt_pv unless
5021                  * we retained the wiring from an unmanaged page we had
5022                  * removed (if we retained it via pte_pv that will go away
5023                  * soon).
5024                  */
5025                 if (pt_pv && (opa == 0 ||
5026                               (origpte & pmap->pmap_bits[PG_MANAGED_IDX]))) {
5027                         vm_page_wire_quick(pt_pv->pv_m);
5028                 }
5029                 if (wired)
5030                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
5031
5032                 /*
5033                  * Unmanaged pages need manual resident_count tracking.
5034                  */
5035                 if (pt_pv) {
5036                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.
5037                                         resident_count, 1);
5038                 }
5039                 if (newpte & pmap->pmap_bits[PG_RW_IDX])
5040                         vm_page_flag_set(m, PG_WRITEABLE);
5041         } else {
5042                 /*
5043                  * Entering a managed page.  Our pte_pv takes care of the
5044                  * PT wiring, so if we had removed an unmanaged page before
5045                  * we must adjust.
5046                  *
5047                  * We have to take care of the pmap wired count ourselves.
5048                  *
5049                  * Enter on the PV list if part of our managed memory.
5050                  */
5051                 KKASSERT(pte_pv && (pte_pv->pv_m == NULL || pte_pv->pv_m == m));
5052                 vm_page_spin_lock(m);
5053                 pte_pv->pv_m = m;
5054                 pmap_page_stats_adding(m);
5055                 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
5056                 vm_page_flag_set(m, PG_MAPPED);
5057                 if (newpte & pmap->pmap_bits[PG_RW_IDX])
5058                         vm_page_flag_set(m, PG_WRITEABLE);
5059                 vm_page_spin_unlock(m);
5060
5061                 if (pt_pv && opa &&
5062                     (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5063                         vm_page_unwire_quick(pt_pv->pv_m);
5064                 }
5065
5066                 /*
5067                  * Adjust pmap wired pages count for new entry.
5068                  */
5069                 if (wired) {
5070                         atomic_add_long(&pte_pv->pv_pmap->pm_stats.
5071                                         wired_count, 1);
5072                 }
5073         }
5074
5075         /*
5076          * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
5077          *
5078          * User VMAs do not because those will be zero->non-zero, so no
5079          * stale entries to worry about at this point.
5080          *
5081          * For KVM there appear to still be issues.  Theoretically we
5082          * should be able to scrap the interlocks entirely but we
5083          * get crashes.
5084          */
5085         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
5086                 pmap_inval_smp(pmap, va, 1, ptep, newpte);
5087         } else {
5088                 origpte = atomic_swap_long(ptep, newpte);
5089                 if (origpte & pmap->pmap_bits[PG_M_IDX]) {
5090                         kprintf("pmap [M] race @ %016jx\n", va);
5091                         atomic_set_long(ptep, pmap->pmap_bits[PG_M_IDX]);
5092                 }
5093                 if (pt_pv == NULL)
5094                         cpu_invlpg((void *)va);
5095         }
5096
5097         /*
5098          * Cleanup
5099          */
5100 done:
5101         KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
5102                  (m->flags & PG_MAPPED));
5103
5104         /*
5105          * Cleanup the pv entry, allowing other accessors.  If the new page
5106          * is not managed but we have a pte_pv (which was locking our
5107          * operation), we can free it now.  pte_pv->pv_m should be NULL.
5108          */
5109         if (pte_pv && (newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5110                 pv_free(pte_pv, pt_pv);
5111         } else if (pte_pv) {
5112                 pv_put(pte_pv);
5113         } else if (pte_placemark) {
5114                 pv_placemarker_wakeup(pmap, pte_placemark);
5115         }
5116         if (pt_pv)
5117                 pv_put(pt_pv);
5118 }
5119
5120 /*
5121  * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
5122  * This code also assumes that the pmap has no pre-existing entry for this
5123  * VA.
5124  *
5125  * This code currently may only be used on user pmaps, not kernel_pmap.
5126  */
5127 void
5128 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
5129 {
5130         pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
5131 }
5132
5133 /*
5134  * Make a temporary mapping for a physical address.  This is only intended
5135  * to be used for panic dumps.
5136  *
5137  * The caller is responsible for calling smp_invltlb().
5138  */
5139 void *
5140 pmap_kenter_temporary(vm_paddr_t pa, long i)
5141 {
5142         pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
5143         return ((void *)crashdumpmap);
5144 }
5145
5146 #define MAX_INIT_PT (96)
5147
5148 /*
5149  * This routine preloads the ptes for a given object into the specified pmap.
5150  * This eliminates the blast of soft faults on process startup and
5151  * immediately after an mmap.
5152  */
5153 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
5154
5155 void
5156 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
5157                     vm_object_t object, vm_pindex_t pindex,
5158                     vm_size_t size, int limit)
5159 {
5160         struct rb_vm_page_scan_info info;
5161         struct lwp *lp;
5162         vm_size_t psize;
5163
5164         /*
5165          * We can't preinit if read access isn't set or there is no pmap
5166          * or object.
5167          */
5168         if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
5169                 return;
5170
5171         /*
5172          * We can't preinit if the pmap is not the current pmap
5173          */
5174         lp = curthread->td_lwp;
5175         if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
5176                 return;
5177
5178         /*
5179          * Misc additional checks
5180          */
5181         psize = x86_64_btop(size);
5182
5183         if ((object->type != OBJT_VNODE) ||
5184                 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
5185                         (object->resident_page_count > MAX_INIT_PT))) {
5186                 return;
5187         }
5188
5189         if (pindex + psize > object->size) {
5190                 if (object->size < pindex)
5191                         return;
5192                 psize = object->size - pindex;
5193         }
5194
5195         if (psize == 0)
5196                 return;
5197
5198         /*
5199          * If everything is segment-aligned do not pre-init here.  Instead
5200          * allow the normal vm_fault path to pass a segment hint to
5201          * pmap_enter() which will then use an object-referenced shared
5202          * page table page.
5203          */
5204         if ((addr & SEG_MASK) == 0 &&
5205             (ctob(psize) & SEG_MASK) == 0 &&
5206             (ctob(pindex) & SEG_MASK) == 0) {
5207                 return;
5208         }
5209
5210         /*
5211          * Use a red-black scan to traverse the requested range and load
5212          * any valid pages found into the pmap.
5213          *
5214          * We cannot safely scan the object's memq without holding the
5215          * object token.
5216          */
5217         info.start_pindex = pindex;
5218         info.end_pindex = pindex + psize - 1;
5219         info.limit = limit;
5220         info.mpte = NULL;
5221         info.addr = addr;
5222         info.pmap = pmap;
5223         info.object = object;
5224
5225         /*
5226          * By using the NOLK scan, the callback function must be sure
5227          * to return -1 if the VM page falls out of the object.
5228          */
5229         vm_object_hold_shared(object);
5230         vm_page_rb_tree_RB_SCAN_NOLK(&object->rb_memq, rb_vm_page_scancmp,
5231                                      pmap_object_init_pt_callback, &info);
5232         vm_object_drop(object);
5233 }
5234
5235 static
5236 int
5237 pmap_object_init_pt_callback(vm_page_t p, void *data)
5238 {
5239         struct rb_vm_page_scan_info *info = data;
5240         vm_pindex_t rel_index;
5241         int hard_busy;
5242
5243         /*
5244          * don't allow an madvise to blow away our really
5245          * free pages allocating pv entries.
5246          */
5247         if ((info->limit & MAP_PREFAULT_MADVISE) &&
5248                 vmstats.v_free_count < vmstats.v_free_reserved) {
5249                     return(-1);
5250         }
5251
5252         /*
5253          * Ignore list markers and ignore pages we cannot instantly
5254          * busy (while holding the object token).
5255          */
5256         if (p->flags & PG_MARKER)
5257                 return 0;
5258         hard_busy = 0;
5259 again:
5260         if (hard_busy) {
5261                 if (vm_page_busy_try(p, TRUE))
5262                         return 0;
5263         } else {
5264                 if (vm_page_sbusy_try(p))
5265                         return 0;
5266         }
5267         if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
5268             (p->flags & PG_FICTITIOUS) == 0) {
5269                 if ((p->queue - p->pc) == PQ_CACHE) {
5270                         if (hard_busy == 0) {
5271                                 vm_page_sbusy_drop(p);
5272                                 hard_busy = 1;
5273                                 goto again;
5274                         }
5275                         vm_page_deactivate(p);
5276                 }
5277                 rel_index = p->pindex - info->start_pindex;
5278                 pmap_enter_quick(info->pmap,
5279                                  info->addr + x86_64_ptob(rel_index), p);
5280         }
5281         if (hard_busy)
5282                 vm_page_wakeup(p);
5283         else
5284                 vm_page_sbusy_drop(p);
5285
5286         /*
5287          * We are using an unlocked scan (that is, the scan expects its
5288          * current element to remain in the tree on return).  So we have
5289          * to check here and abort the scan if it isn't.
5290          */
5291         if (p->object != info->object)
5292                 return -1;
5293         lwkt_yield();
5294         return(0);
5295 }
5296
5297 /*
5298  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
5299  * address.
5300  *
5301  * Returns FALSE if it would be non-trivial or if a pte is already loaded
5302  * into the slot.
5303  *
5304  * XXX This is safe only because page table pages are not freed.
5305  */
5306 int
5307 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
5308 {
5309         pt_entry_t *pte;
5310
5311         /*spin_lock(&pmap->pm_spin);*/
5312         if ((pte = pmap_pte(pmap, addr)) != NULL) {
5313                 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
5314                         /*spin_unlock(&pmap->pm_spin);*/
5315                         return FALSE;
5316                 }
5317         }
5318         /*spin_unlock(&pmap->pm_spin);*/
5319         return TRUE;
5320 }
5321
5322 /*
5323  * Change the wiring attribute for a pmap/va pair.  The mapping must already
5324  * exist in the pmap.  The mapping may or may not be managed.  The wiring in
5325  * the page is not changed, the page is returned so the caller can adjust
5326  * its wiring (the page is not locked in any way).
5327  *
5328  * Wiring is not a hardware characteristic so there is no need to invalidate
5329  * TLB.  However, in an SMP environment we must use a locked bus cycle to
5330  * update the pte (if we are not using the pmap_inval_*() API that is)...
5331  * it's ok to do this for simple wiring changes.
5332  */
5333 vm_page_t
5334 pmap_unwire(pmap_t pmap, vm_offset_t va)
5335 {
5336         pt_entry_t *ptep;
5337         pv_entry_t pt_pv;
5338         vm_paddr_t pa;
5339         vm_page_t m;
5340
5341         if (pmap == NULL)
5342                 return NULL;
5343
5344         /*
5345          * Assume elements in the kernel pmap are stable
5346          */
5347         if (pmap == &kernel_pmap) {
5348                 if (pmap_pt(pmap, va) == 0)
5349                         return NULL;
5350                 ptep = pmap_pte_quick(pmap, va);
5351                 if (pmap_pte_v(pmap, ptep)) {
5352                         if (pmap_pte_w(pmap, ptep))
5353                                 atomic_add_long(&pmap->pm_stats.wired_count,-1);
5354                         atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5355                         pa = *ptep & PG_FRAME;
5356                         m = PHYS_TO_VM_PAGE(pa);
5357                 } else {
5358                         m = NULL;
5359                 }
5360         } else {
5361                 /*
5362                  * We can only [un]wire pmap-local pages (we cannot wire
5363                  * shared pages)
5364                  */
5365                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
5366                 if (pt_pv == NULL)
5367                         return NULL;
5368
5369                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
5370                 if ((*ptep & pmap->pmap_bits[PG_V_IDX]) == 0) {
5371                         pv_put(pt_pv);
5372                         return NULL;
5373                 }
5374
5375                 if (pmap_pte_w(pmap, ptep)) {
5376                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.wired_count,
5377                                         -1);
5378                 }
5379                 /* XXX else return NULL so caller doesn't unwire m ? */
5380
5381                 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5382
5383                 pa = *ptep & PG_FRAME;
5384                 m = PHYS_TO_VM_PAGE(pa);        /* held by wired count */
5385                 pv_put(pt_pv);
5386         }
5387         return m;
5388 }
5389
5390 /*
5391  * Copy the range specified by src_addr/len from the source map to
5392  * the range dst_addr/len in the destination map.
5393  *
5394  * This routine is only advisory and need not do anything.
5395  */
5396 void
5397 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
5398           vm_size_t len, vm_offset_t src_addr)
5399 {
5400 }
5401
5402 /*
5403  * pmap_zero_page:
5404  *
5405  *      Zero the specified physical page.
5406  *
5407  *      This function may be called from an interrupt and no locking is
5408  *      required.
5409  */
5410 void
5411 pmap_zero_page(vm_paddr_t phys)
5412 {
5413         vm_offset_t va = PHYS_TO_DMAP(phys);
5414
5415         pagezero((void *)va);
5416 }
5417
5418 /*
5419  * pmap_zero_page:
5420  *
5421  *      Zero part of a physical page by mapping it into memory and clearing
5422  *      its contents with bzero.
5423  *
5424  *      off and size may not cover an area beyond a single hardware page.
5425  */
5426 void
5427 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
5428 {
5429         vm_offset_t virt = PHYS_TO_DMAP(phys);
5430
5431         bzero((char *)virt + off, size);
5432 }
5433
5434 /*
5435  * pmap_copy_page:
5436  *
5437  *      Copy the physical page from the source PA to the target PA.
5438  *      This function may be called from an interrupt.  No locking
5439  *      is required.
5440  */
5441 void
5442 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
5443 {
5444         vm_offset_t src_virt, dst_virt;
5445
5446         src_virt = PHYS_TO_DMAP(src);
5447         dst_virt = PHYS_TO_DMAP(dst);
5448         bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
5449 }
5450
5451 /*
5452  * pmap_copy_page_frag:
5453  *
5454  *      Copy the physical page from the source PA to the target PA.
5455  *      This function may be called from an interrupt.  No locking
5456  *      is required.
5457  */
5458 void
5459 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
5460 {
5461         vm_offset_t src_virt, dst_virt;
5462
5463         src_virt = PHYS_TO_DMAP(src);
5464         dst_virt = PHYS_TO_DMAP(dst);
5465
5466         bcopy((char *)src_virt + (src & PAGE_MASK),
5467               (char *)dst_virt + (dst & PAGE_MASK),
5468               bytes);
5469 }
5470
5471 /*
5472  * Returns true if the pmap's pv is one of the first 16 pvs linked to from
5473  * this page.  This count may be changed upwards or downwards in the future;
5474  * it is only necessary that true be returned for a small subset of pmaps
5475  * for proper page aging.
5476  */
5477 boolean_t
5478 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
5479 {
5480         pv_entry_t pv;
5481         int loops = 0;
5482
5483         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5484                 return FALSE;
5485
5486         vm_page_spin_lock(m);
5487         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5488                 if (pv->pv_pmap == pmap) {
5489                         vm_page_spin_unlock(m);
5490                         return TRUE;
5491                 }
5492                 loops++;
5493                 if (loops >= 16)
5494                         break;
5495         }
5496         vm_page_spin_unlock(m);
5497         return (FALSE);
5498 }
5499
5500 /*
5501  * Remove all pages from specified address space this aids process exit
5502  * speeds.  Also, this code may be special cased for the current process
5503  * only.
5504  */
5505 void
5506 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5507 {
5508         pmap_remove_noinval(pmap, sva, eva);
5509         cpu_invltlb();
5510 }
5511
5512 /*
5513  * pmap_testbit tests bits in pte's note that the testbit/clearbit
5514  * routines are inline, and a lot of things compile-time evaluate.
5515  */
5516
5517 static
5518 boolean_t
5519 pmap_testbit(vm_page_t m, int bit)
5520 {
5521         pv_entry_t pv;
5522         pt_entry_t *pte;
5523         pmap_t pmap;
5524
5525         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5526                 return FALSE;
5527
5528         if (TAILQ_FIRST(&m->md.pv_list) == NULL)
5529                 return FALSE;
5530         vm_page_spin_lock(m);
5531         if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
5532                 vm_page_spin_unlock(m);
5533                 return FALSE;
5534         }
5535
5536         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5537 #if defined(PMAP_DIAGNOSTIC)
5538                 if (pv->pv_pmap == NULL) {
5539                         kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
5540                             pv->pv_pindex);
5541                         continue;
5542                 }
5543 #endif
5544                 pmap = pv->pv_pmap;
5545
5546                 /*
5547                  * If the bit being tested is the modified bit, then
5548                  * mark clean_map and ptes as never
5549                  * modified.
5550                  *
5551                  * WARNING!  Because we do not lock the pv, *pte can be in a
5552                  *           state of flux.  Despite this the value of *pte
5553                  *           will still be related to the vm_page in some way
5554                  *           because the pv cannot be destroyed as long as we
5555                  *           hold the vm_page spin lock.
5556                  */
5557                 if (bit == PG_A_IDX || bit == PG_M_IDX) {
5558                                 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5559                         if (!pmap_track_modified(pv->pv_pindex))
5560                                 continue;
5561                 }
5562
5563                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5564                 if (*pte & pmap->pmap_bits[bit]) {
5565                         vm_page_spin_unlock(m);
5566                         return TRUE;
5567                 }
5568         }
5569         vm_page_spin_unlock(m);
5570         return (FALSE);
5571 }
5572
5573 /*
5574  * This routine is used to modify bits in ptes.  Only one bit should be
5575  * specified.  PG_RW requires special handling.
5576  *
5577  * Caller must NOT hold any spin locks
5578  */
5579 static __inline
5580 void
5581 pmap_clearbit(vm_page_t m, int bit_index)
5582 {
5583         pv_entry_t pv;
5584         pt_entry_t *pte;
5585         pt_entry_t pbits;
5586         pmap_t pmap;
5587
5588         if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5589                 if (bit_index == PG_RW_IDX)
5590                         vm_page_flag_clear(m, PG_WRITEABLE);
5591                 return;
5592         }
5593
5594         /*
5595          * PG_M or PG_A case
5596          *
5597          * Loop over all current mappings setting/clearing as appropos If
5598          * setting RO do we need to clear the VAC?
5599          *
5600          * NOTE: When clearing PG_M we could also (not implemented) drop
5601          *       through to the PG_RW code and clear PG_RW too, forcing
5602          *       a fault on write to redetect PG_M for virtual kernels, but
5603          *       it isn't necessary since virtual kernels invalidate the
5604          *       pte when they clear the VPTE_M bit in their virtual page
5605          *       tables.
5606          *
5607          * NOTE: Does not re-dirty the page when clearing only PG_M.
5608          *
5609          * NOTE: Because we do not lock the pv, *pte can be in a state of
5610          *       flux.  Despite this the value of *pte is still somewhat
5611          *       related while we hold the vm_page spin lock.
5612          *
5613          *       *pte can be zero due to this race.  Since we are clearing
5614          *       bits we basically do no harm when this race occurs.
5615          */
5616         if (bit_index != PG_RW_IDX) {
5617                 vm_page_spin_lock(m);
5618                 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5619 #if defined(PMAP_DIAGNOSTIC)
5620                         if (pv->pv_pmap == NULL) {
5621                                 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5622                                     pv->pv_pindex);
5623                                 continue;
5624                         }
5625 #endif
5626                         pmap = pv->pv_pmap;
5627                         pte = pmap_pte_quick(pv->pv_pmap,
5628                                              pv->pv_pindex << PAGE_SHIFT);
5629                         pbits = *pte;
5630                         if (pbits & pmap->pmap_bits[bit_index])
5631                                 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5632                 }
5633                 vm_page_spin_unlock(m);
5634                 return;
5635         }
5636
5637         /*
5638          * Clear PG_RW.  Also clears PG_M and marks the page dirty if PG_M
5639          * was set.
5640          */
5641 restart:
5642         vm_page_spin_lock(m);
5643         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5644                 /*
5645                  * don't write protect pager mappings
5646                  */
5647                 if (!pmap_track_modified(pv->pv_pindex))
5648                         continue;
5649
5650 #if defined(PMAP_DIAGNOSTIC)
5651                 if (pv->pv_pmap == NULL) {
5652                         kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5653                                 pv->pv_pindex);
5654                         continue;
5655                 }
5656 #endif
5657                 pmap = pv->pv_pmap;
5658
5659                 /*
5660                  * Skip pages which do not have PG_RW set.
5661                  */
5662                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5663                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5664                         continue;
5665
5666                 /*
5667                  * We must lock the PV to be able to safely test the pte.
5668                  */
5669                 if (pv_hold_try(pv)) {
5670                         vm_page_spin_unlock(m);
5671                 } else {
5672                         vm_page_spin_unlock(m);
5673                         pv_lock(pv);    /* held, now do a blocking lock */
5674                         pv_put(pv);
5675                         goto restart;
5676                 }
5677
5678                 /*
5679                  * Reload pte after acquiring pv.
5680                  */
5681                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5682 #if 0
5683                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0) {
5684                         pv_put(pv);
5685                         goto restart;
5686                 }
5687 #endif
5688
5689                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
5690                 for (;;) {
5691                         pt_entry_t nbits;
5692
5693                         pbits = *pte;
5694                         cpu_ccfence();
5695                         nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5696                                           pmap->pmap_bits[PG_M_IDX]);
5697                         if (pmap_inval_smp_cmpset(pmap,
5698                                      ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5699                                      pte, pbits, nbits)) {
5700                                 break;
5701                         }
5702                         cpu_pause();
5703                 }
5704
5705                 /*
5706                  * If PG_M was found to be set while we were clearing PG_RW
5707                  * we also clear PG_M (done above) and mark the page dirty.
5708                  * Callers expect this behavior.
5709                  *
5710                  * we lost pv so it cannot be used as an iterator.  In fact,
5711                  * because we couldn't necessarily lock it atomically it may
5712                  * have moved within the list and ALSO cannot be used as an
5713                  * iterator.
5714                  */
5715                 vm_page_spin_lock(m);
5716                 if (pbits & pmap->pmap_bits[PG_M_IDX])
5717                         vm_page_dirty(m);
5718                 vm_page_spin_unlock(m);
5719                 pv_put(pv);
5720                 goto restart;
5721         }
5722         if (bit_index == PG_RW_IDX)
5723                 vm_page_flag_clear(m, PG_WRITEABLE);
5724         vm_page_spin_unlock(m);
5725 }
5726
5727 /*
5728  * Lower the permission for all mappings to a given page.
5729  *
5730  * Page must be busied by caller.  Because page is busied by caller this
5731  * should not be able to race a pmap_enter().
5732  */
5733 void
5734 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5735 {
5736         /* JG NX support? */
5737         if ((prot & VM_PROT_WRITE) == 0) {
5738                 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5739                         /*
5740                          * NOTE: pmap_clearbit(.. PG_RW) also clears
5741                          *       the PG_WRITEABLE flag in (m).
5742                          */
5743                         pmap_clearbit(m, PG_RW_IDX);
5744                 } else {
5745                         pmap_remove_all(m);
5746                 }
5747         }
5748 }
5749
5750 vm_paddr_t
5751 pmap_phys_address(vm_pindex_t ppn)
5752 {
5753         return (x86_64_ptob(ppn));
5754 }
5755
5756 /*
5757  * Return a count of reference bits for a page, clearing those bits.
5758  * It is not necessary for every reference bit to be cleared, but it
5759  * is necessary that 0 only be returned when there are truly no
5760  * reference bits set.
5761  *
5762  * XXX: The exact number of bits to check and clear is a matter that
5763  * should be tested and standardized at some point in the future for
5764  * optimal aging of shared pages.
5765  *
5766  * This routine may not block.
5767  */
5768 int
5769 pmap_ts_referenced(vm_page_t m)
5770 {
5771         pv_entry_t pv;
5772         pt_entry_t *pte;
5773         pmap_t pmap;
5774         int rtval = 0;
5775
5776         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5777                 return (rtval);
5778
5779         vm_page_spin_lock(m);
5780         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5781                 if (!pmap_track_modified(pv->pv_pindex))
5782                         continue;
5783                 pmap = pv->pv_pmap;
5784                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5785                 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5786                         atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5787                         rtval++;
5788                         if (rtval > 4)
5789                                 break;
5790                 }
5791         }
5792         vm_page_spin_unlock(m);
5793         return (rtval);
5794 }
5795
5796 /*
5797  *      pmap_is_modified:
5798  *
5799  *      Return whether or not the specified physical page was modified
5800  *      in any physical maps.
5801  */
5802 boolean_t
5803 pmap_is_modified(vm_page_t m)
5804 {
5805         boolean_t res;
5806
5807         res = pmap_testbit(m, PG_M_IDX);
5808         return (res);
5809 }
5810
5811 /*
5812  *      Clear the modify bits on the specified physical page.
5813  */
5814 void
5815 pmap_clear_modify(vm_page_t m)
5816 {
5817         pmap_clearbit(m, PG_M_IDX);
5818 }
5819
5820 /*
5821  *      pmap_clear_reference:
5822  *
5823  *      Clear the reference bit on the specified physical page.
5824  */
5825 void
5826 pmap_clear_reference(vm_page_t m)
5827 {
5828         pmap_clearbit(m, PG_A_IDX);
5829 }
5830
5831 /*
5832  * Miscellaneous support routines follow
5833  */
5834
5835 static
5836 void
5837 i386_protection_init(void)
5838 {
5839         uint64_t *kp;
5840         int prot;
5841
5842         /*
5843          * NX supported? (boot time loader.conf override only)
5844          */
5845         TUNABLE_INT_FETCH("machdep.pmap_nx_enable", &pmap_nx_enable);
5846         if (pmap_nx_enable == 0 || (amd_feature & AMDID_NX) == 0)
5847                 pmap_bits_default[PG_NX_IDX] = 0;
5848
5849         /*
5850          * 0 is basically read-only access, but also set the NX (no-execute)
5851          * bit when VM_PROT_EXECUTE is not specified.
5852          */
5853         kp = protection_codes;
5854         for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5855                 switch (prot) {
5856                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5857                         /*
5858                          * This case handled elsewhere
5859                          */
5860                         *kp++ = 0;
5861                         break;
5862                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5863                         /*
5864                          * Read-only is 0|NX
5865                          */
5866                         *kp++ = pmap_bits_default[PG_NX_IDX];
5867                         break;
5868                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5869                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5870                         /*
5871                          * Execute requires read access
5872                          */
5873                         *kp++ = 0;
5874                         break;
5875                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5876                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5877                         /*
5878                          * Write without execute is RW|NX
5879                          */
5880                         *kp++ = pmap_bits_default[PG_RW_IDX] |
5881                                 pmap_bits_default[PG_NX_IDX];
5882                         break;
5883                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5884                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5885                         /*
5886                          * Write with execute is RW
5887                          */
5888                         *kp++ = pmap_bits_default[PG_RW_IDX];
5889                         break;
5890                 }
5891         }
5892 }
5893
5894 /*
5895  * Map a set of physical memory pages into the kernel virtual
5896  * address space. Return a pointer to where it is mapped. This
5897  * routine is intended to be used for mapping device memory,
5898  * NOT real memory.
5899  *
5900  * NOTE: We can't use pgeflag unless we invalidate the pages one at
5901  *       a time.
5902  *
5903  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5904  *       work whether the cpu supports PAT or not.  The remaining PAT
5905  *       attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5906  *       supports PAT.
5907  */
5908 void *
5909 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5910 {
5911         return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5912 }
5913
5914 void *
5915 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5916 {
5917         return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5918 }
5919
5920 void *
5921 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5922 {
5923         return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5924 }
5925
5926 /*
5927  * Map a set of physical memory pages into the kernel virtual
5928  * address space. Return a pointer to where it is mapped. This
5929  * routine is intended to be used for mapping device memory,
5930  * NOT real memory.
5931  */
5932 void *
5933 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5934 {
5935         vm_offset_t va, tmpva, offset;
5936         pt_entry_t *pte;
5937         vm_size_t tmpsize;
5938
5939         offset = pa & PAGE_MASK;
5940         size = roundup(offset + size, PAGE_SIZE);
5941
5942         va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5943         if (va == 0)
5944                 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5945
5946         pa = pa & ~PAGE_MASK;
5947         for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5948                 pte = vtopte(tmpva);
5949                 *pte = pa |
5950                     kernel_pmap.pmap_bits[PG_RW_IDX] |
5951                     kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5952                     kernel_pmap.pmap_cache_bits[mode];
5953                 tmpsize -= PAGE_SIZE;
5954                 tmpva += PAGE_SIZE;
5955                 pa += PAGE_SIZE;
5956         }
5957         pmap_invalidate_range(&kernel_pmap, va, va + size);
5958         pmap_invalidate_cache_range(va, va + size);
5959
5960         return ((void *)(va + offset));
5961 }
5962
5963 void
5964 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5965 {
5966         vm_offset_t base, offset;
5967
5968         base = va & ~PAGE_MASK;
5969         offset = va & PAGE_MASK;
5970         size = roundup(offset + size, PAGE_SIZE);
5971         pmap_qremove(va, size >> PAGE_SHIFT);
5972         kmem_free(&kernel_map, base, size);
5973 }
5974
5975 /*
5976  * Sets the memory attribute for the specified page.
5977  */
5978 void
5979 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5980 {
5981
5982     m->pat_mode = ma;
5983
5984     /*
5985      * If "m" is a normal page, update its direct mapping.  This update
5986      * can be relied upon to perform any cache operations that are
5987      * required for data coherence.
5988      */
5989     if ((m->flags & PG_FICTITIOUS) == 0)
5990         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5991 }
5992
5993 /*
5994  * Change the PAT attribute on an existing kernel memory map.  Caller
5995  * must ensure that the virtual memory in question is not accessed
5996  * during the adjustment.
5997  */
5998 void
5999 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
6000 {
6001         pt_entry_t *pte;
6002         vm_offset_t base;
6003         int changed = 0;
6004
6005         if (va == 0)
6006                 panic("pmap_change_attr: va is NULL");
6007         base = trunc_page(va);
6008
6009         while (count) {
6010                 pte = vtopte(va);
6011                 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
6012                        kernel_pmap.pmap_cache_bits[mode];
6013                 --count;
6014                 va += PAGE_SIZE;
6015         }
6016
6017         changed = 1;    /* XXX: not optimal */
6018
6019         /*
6020          * Flush CPU caches if required to make sure any data isn't cached that
6021          * shouldn't be, etc.
6022          */
6023         if (changed) {
6024                 pmap_invalidate_range(&kernel_pmap, base, va);
6025                 pmap_invalidate_cache_range(base, va);
6026         }
6027 }
6028
6029 /*
6030  * perform the pmap work for mincore
6031  */
6032 int
6033 pmap_mincore(pmap_t pmap, vm_offset_t addr)
6034 {
6035         pt_entry_t *ptep, pte;
6036         vm_page_t m;
6037         int val = 0;
6038
6039         ptep = pmap_pte(pmap, addr);
6040
6041         if (ptep && (pte = *ptep) != 0) {
6042                 vm_offset_t pa;
6043
6044                 val = MINCORE_INCORE;
6045                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
6046                         goto done;
6047
6048                 pa = pte & PG_FRAME;
6049
6050                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
6051                         m = NULL;
6052                 else
6053                         m = PHYS_TO_VM_PAGE(pa);
6054
6055                 /*
6056                  * Modified by us
6057                  */
6058                 if (pte & pmap->pmap_bits[PG_M_IDX])
6059                         val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
6060                 /*
6061                  * Modified by someone
6062                  */
6063                 else if (m && (m->dirty || pmap_is_modified(m)))
6064                         val |= MINCORE_MODIFIED_OTHER;
6065                 /*
6066                  * Referenced by us
6067                  */
6068                 if (pte & pmap->pmap_bits[PG_A_IDX])
6069                         val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
6070
6071                 /*
6072                  * Referenced by someone
6073                  */
6074                 else if (m && ((m->flags & PG_REFERENCED) ||
6075                                 pmap_ts_referenced(m))) {
6076                         val |= MINCORE_REFERENCED_OTHER;
6077                         vm_page_flag_set(m, PG_REFERENCED);
6078                 }
6079         }
6080 done:
6081
6082         return val;
6083 }
6084
6085 /*
6086  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
6087  * vmspace will be ref'd and the old one will be deref'd.
6088  *
6089  * The vmspace for all lwps associated with the process will be adjusted
6090  * and cr3 will be reloaded if any lwp is the current lwp.
6091  *
6092  * The process must hold the vmspace->vm_map.token for oldvm and newvm
6093  */
6094 void
6095 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
6096 {
6097         struct vmspace *oldvm;
6098         struct lwp *lp;
6099
6100         oldvm = p->p_vmspace;
6101         if (oldvm != newvm) {
6102                 if (adjrefs)
6103                         vmspace_ref(newvm);
6104                 p->p_vmspace = newvm;
6105                 KKASSERT(p->p_nthreads == 1);
6106                 lp = RB_ROOT(&p->p_lwp_tree);
6107                 pmap_setlwpvm(lp, newvm);
6108                 if (adjrefs)
6109                         vmspace_rel(oldvm);
6110         }
6111 }
6112
6113 /*
6114  * Set the vmspace for a LWP.  The vmspace is almost universally set the
6115  * same as the process vmspace, but virtual kernels need to swap out contexts
6116  * on a per-lwp basis.
6117  *
6118  * Caller does not necessarily hold any vmspace tokens.  Caller must control
6119  * the lwp (typically be in the context of the lwp).  We use a critical
6120  * section to protect against statclock and hardclock (statistics collection).
6121  */
6122 void
6123 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
6124 {
6125         struct vmspace *oldvm;
6126         struct pmap *pmap;
6127
6128         oldvm = lp->lwp_vmspace;
6129
6130         if (oldvm != newvm) {
6131                 crit_enter();
6132                 KKASSERT((newvm->vm_refcnt & VM_REF_DELETED) == 0);
6133                 lp->lwp_vmspace = newvm;
6134                 if (curthread->td_lwp == lp) {
6135                         pmap = vmspace_pmap(newvm);
6136                         ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
6137                         if (pmap->pm_active_lock & CPULOCK_EXCL)
6138                                 pmap_interlock_wait(newvm);
6139 #if defined(SWTCH_OPTIM_STATS)
6140                         tlb_flush_count++;
6141 #endif
6142                         if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
6143                                 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
6144                         } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
6145                                 curthread->td_pcb->pcb_cr3 = KPML4phys;
6146                         } else {
6147                                 panic("pmap_setlwpvm: unknown pmap type\n");
6148                         }
6149                         load_cr3(curthread->td_pcb->pcb_cr3);
6150                         pmap = vmspace_pmap(oldvm);
6151                         ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
6152                                                mycpu->gd_cpuid);
6153                 }
6154                 crit_exit();
6155         }
6156 }
6157
6158 /*
6159  * Called when switching to a locked pmap, used to interlock against pmaps
6160  * undergoing modifications to prevent us from activating the MMU for the
6161  * target pmap until all such modifications have completed.  We have to do
6162  * this because the thread making the modifications has already set up its
6163  * SMP synchronization mask.
6164  *
6165  * This function cannot sleep!
6166  *
6167  * No requirements.
6168  */
6169 void
6170 pmap_interlock_wait(struct vmspace *vm)
6171 {
6172         struct pmap *pmap = &vm->vm_pmap;
6173
6174         if (pmap->pm_active_lock & CPULOCK_EXCL) {
6175                 crit_enter();
6176                 KKASSERT(curthread->td_critcount >= 2);
6177                 DEBUG_PUSH_INFO("pmap_interlock_wait");
6178                 while (pmap->pm_active_lock & CPULOCK_EXCL) {
6179                         cpu_ccfence();
6180                         lwkt_process_ipiq();
6181                 }
6182                 DEBUG_POP_INFO();
6183                 crit_exit();
6184         }
6185 }
6186
6187 vm_offset_t
6188 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
6189 {
6190
6191         if ((obj == NULL) || (size < NBPDR) ||
6192             ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
6193                 return addr;
6194         }
6195
6196         addr = roundup2(addr, NBPDR);
6197         return addr;
6198 }
6199
6200 /*
6201  * Used by kmalloc/kfree, page already exists at va
6202  */
6203 vm_page_t
6204 pmap_kvtom(vm_offset_t va)
6205 {
6206         pt_entry_t *ptep = vtopte(va);
6207
6208         KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
6209         return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
6210 }
6211
6212 /*
6213  * Initialize machine-specific shared page directory support.  This
6214  * is executed when a VM object is created.
6215  */
6216 void
6217 pmap_object_init(vm_object_t object)
6218 {
6219         object->md.pmap_rw = NULL;
6220         object->md.pmap_ro = NULL;
6221 }
6222
6223 /*
6224  * Clean up machine-specific shared page directory support.  This
6225  * is executed when a VM object is destroyed.
6226  */
6227 void
6228 pmap_object_free(vm_object_t object)
6229 {
6230         pmap_t pmap;
6231
6232         if ((pmap = object->md.pmap_rw) != NULL) {
6233                 object->md.pmap_rw = NULL;
6234                 pmap_remove_noinval(pmap,
6235                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6236                 CPUMASK_ASSZERO(pmap->pm_active);
6237                 pmap_release(pmap);
6238                 pmap_puninit(pmap);
6239                 kfree(pmap, M_OBJPMAP);
6240         }
6241         if ((pmap = object->md.pmap_ro) != NULL) {
6242                 object->md.pmap_ro = NULL;
6243                 pmap_remove_noinval(pmap,
6244                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6245                 CPUMASK_ASSZERO(pmap->pm_active);
6246                 pmap_release(pmap);
6247                 pmap_puninit(pmap);
6248                 kfree(pmap, M_OBJPMAP);
6249         }
6250 }
6251
6252 /*
6253  * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
6254  * VM page and issue a pginfo->callback.
6255  *
6256  * We are expected to dispose of any non-NULL pte_pv.
6257  */
6258 static
6259 void
6260 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
6261                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
6262                       pv_entry_t pt_pv, int sharept,
6263                       vm_offset_t va, pt_entry_t *ptep, void *arg)
6264 {
6265         struct pmap_pgscan_info *pginfo = arg;
6266         vm_page_t m;
6267
6268         if (pte_pv) {
6269                 /*
6270                  * Try to busy the page while we hold the pte_pv locked.
6271                  */
6272                 KKASSERT(pte_pv->pv_m);
6273                 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
6274                 if (vm_page_busy_try(m, TRUE) == 0) {
6275                         if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
6276                                 /*
6277                                  * The callback is issued with the pte_pv
6278                                  * unlocked and put away, and the pt_pv
6279                                  * unlocked.
6280                                  */
6281                                 pv_put(pte_pv);
6282                                 if (pt_pv) {
6283                                         vm_page_wire_quick(pt_pv->pv_m);
6284                                         pv_unlock(pt_pv);
6285                                 }
6286                                 if (pginfo->callback(pginfo, va, m) < 0)
6287                                         info->stop = 1;
6288                                 if (pt_pv) {
6289                                         pv_lock(pt_pv);
6290                                         vm_page_unwire_quick(pt_pv->pv_m);
6291                                 }
6292                         } else {
6293                                 vm_page_wakeup(m);
6294                                 pv_put(pte_pv);
6295                         }
6296                 } else {
6297                         ++pginfo->busycount;
6298                         pv_put(pte_pv);
6299                 }
6300         } else {
6301                 /*
6302                  * Shared page table or unmanaged page (sharept or !sharept)
6303                  */
6304                 pv_placemarker_wakeup(pmap, pte_placemark);
6305         }
6306 }
6307
6308 void
6309 pmap_pgscan(struct pmap_pgscan_info *pginfo)
6310 {
6311         struct pmap_scan_info info;
6312
6313         pginfo->offset = pginfo->beg_addr;
6314         info.pmap = pginfo->pmap;
6315         info.sva = pginfo->beg_addr;
6316         info.eva = pginfo->end_addr;
6317         info.func = pmap_pgscan_callback;
6318         info.arg = pginfo;
6319         pmap_scan(&info, 0);
6320         if (info.stop == 0)
6321                 pginfo->offset = pginfo->end_addr;
6322 }
6323
6324 /*
6325  * Wait for a placemarker that we do not own to clear.  The placemarker
6326  * in question is not necessarily set to the pindex we want, we may have
6327  * to wait on the element because we want to reserve it ourselves.
6328  *
6329  * NOTE: PM_PLACEMARK_WAKEUP sets a bit which is already set in
6330  *       PM_NOPLACEMARK, so it does not interfere with placemarks
6331  *       which have already been woken up.
6332  */
6333 static
6334 void
6335 pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark)
6336 {
6337         if (*pmark != PM_NOPLACEMARK) {
6338                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
6339                 tsleep_interlock(pmark, 0);
6340                 if (*pmark != PM_NOPLACEMARK)
6341                         tsleep(pmark, PINTERLOCKED, "pvplw", 0);
6342         }
6343 }
6344
6345 /*
6346  * Wakeup a placemarker that we own.  Replace the entry with
6347  * PM_NOPLACEMARK and issue a wakeup() if necessary.
6348  */
6349 static
6350 void
6351 pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark)
6352 {
6353         vm_pindex_t pindex;
6354
6355         pindex = atomic_swap_long(pmark, PM_NOPLACEMARK);
6356         KKASSERT(pindex != PM_NOPLACEMARK);
6357         if (pindex & PM_PLACEMARK_WAKEUP)
6358                 wakeup(pmark);
6359 }