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/sysref2.h>
  78 #include <sys/spinlock2.h>
  79 #include <vm/vm_page2.h>
  80
  81 #include <machine/cputypes.h>
  82 #include <machine/md_var.h>
  83 #include <machine/specialreg.h>
  84 #include <machine/smp.h>
  85 #include <machine_base/apic/apicreg.h>
  86 #include <machine/globaldata.h>
  87 #include <machine/pmap.h>
  88 #include <machine/pmap_inval.h>
  89 #include <machine/inttypes.h>
  90
  91 #include <ddb/ddb.h>
  92
  93 #define PMAP_KEEP_PDIRS
  94 #ifndef PMAP_SHPGPERPROC
  95 #define PMAP_SHPGPERPROC 2000
  96 #endif
  97
  98 #if defined(DIAGNOSTIC)
  99 #define PMAP_DIAGNOSTIC
 100 #endif
 101
 102 #define MINPV 2048
 103
 104 /*
 105  * pmap debugging will report who owns a pv lock when blocking.
 106  */
 107 #ifdef PMAP_DEBUG
 108
 109 #define PMAP_DEBUG_DECL         ,const char *func, int lineno
 110 #define PMAP_DEBUG_ARGS         , __func__, __LINE__
 111 #define PMAP_DEBUG_COPY         , func, lineno
 112
 113 #define pv_get(pmap, pindex, pmarkp)    _pv_get(pmap, pindex, pmarkp    \
 114                                                         PMAP_DEBUG_ARGS)
 115 #define pv_lock(pv)                     _pv_lock(pv                     \
 116                                                         PMAP_DEBUG_ARGS)
 117 #define pv_hold_try(pv)                 _pv_hold_try(pv                 \
 118                                                         PMAP_DEBUG_ARGS)
 119 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp  \
 120                                                         PMAP_DEBUG_ARGS)
 121
 122 #define pv_free(pv, pvp)                _pv_free(pv, pvp PMAP_DEBUG_ARGS)
 123
 124 #else
 125
 126 #define PMAP_DEBUG_DECL
 127 #define PMAP_DEBUG_ARGS
 128 #define PMAP_DEBUG_COPY
 129
 130 #define pv_get(pmap, pindex, pmarkp)            _pv_get(pmap, pindex, pmarkp)
 131 #define pv_lock(pv)                     _pv_lock(pv)
 132 #define pv_hold_try(pv)                 _pv_hold_try(pv)
 133 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp)
 134 #define pv_free(pv, pvp)                _pv_free(pv, pvp)
 135
 136 #endif
 137
 138 /*
 139  * Get PDEs and PTEs for user/kernel address space
 140  */
 141 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
 142
 143 #define pmap_pde_v(pmap, pte)           ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 144 #define pmap_pte_w(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
 145 #define pmap_pte_m(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
 146 #define pmap_pte_u(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
 147 #define pmap_pte_v(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 148
 149 /*
 150  * Given a map and a machine independent protection code,
 151  * convert to a vax protection code.
 152  */
 153 #define pte_prot(m, p)          \
 154         (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
 155 static uint64_t protection_codes[PROTECTION_CODES_SIZE];
 156
 157 struct pmap kernel_pmap;
 158
 159 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
 160
 161 vm_paddr_t avail_start;         /* PA of first available physical page */
 162 vm_paddr_t avail_end;           /* PA of last available physical page */
 163 vm_offset_t virtual2_start;     /* cutout free area prior to kernel start */
 164 vm_offset_t virtual2_end;
 165 vm_offset_t virtual_start;      /* VA of first avail page (after kernel bss) */
 166 vm_offset_t virtual_end;        /* VA of last avail page (end of kernel AS) */
 167 vm_offset_t KvaStart;           /* VA start of KVA space */
 168 vm_offset_t KvaEnd;             /* VA end of KVA space (non-inclusive) */
 169 vm_offset_t KvaSize;            /* max size of kernel virtual address space */
 170 static boolean_t pmap_initialized = FALSE;      /* Has pmap_init completed? */
 171 //static int pgeflag;           /* PG_G or-in */
 172 //static int pseflag;           /* PG_PS or-in */
 173 uint64_t PatMsr;
 174
 175 static int ndmpdp;
 176 static vm_paddr_t dmaplimit;
 177 static int nkpt;
 178 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
 179
 180 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE];        /* PAT -> PG_ bits */
 181 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/    /* PAT -> PG_ bits */
 182
 183 static uint64_t KPTbase;
 184 static uint64_t KPTphys;
 185 static uint64_t KPDphys;        /* phys addr of kernel level 2 */
 186 static uint64_t KPDbase;        /* phys addr of kernel level 2 @ KERNBASE */
 187 uint64_t KPDPphys;      /* phys addr of kernel level 3 */
 188 uint64_t KPML4phys;     /* phys addr of kernel level 4 */
 189
 190 static uint64_t DMPDphys;       /* phys addr of direct mapped level 2 */
 191 static uint64_t DMPDPphys;      /* phys addr of direct mapped level 3 */
 192
 193 /*
 194  * Data for the pv entry allocation mechanism
 195  */
 196 static vm_zone_t pvzone;
 197 static struct vm_zone pvzone_store;
 198 static int pv_entry_max=0, pv_entry_high_water=0;
 199 static int pmap_pagedaemon_waken = 0;
 200 static struct pv_entry *pvinit;
 201
 202 /*
 203  * All those kernel PT submaps that BSD is so fond of
 204  */
 205 pt_entry_t *CMAP1 = NULL, *ptmmap;
 206 caddr_t CADDR1 = NULL, ptvmmap = NULL;
 207 static pt_entry_t *msgbufmap;
 208 struct msgbuf *msgbufp=NULL;
 209
 210 /*
 211  * PMAP default PG_* bits. Needed to be able to add
 212  * EPT/NPT pagetable pmap_bits for the VMM module
 213  */
 214 uint64_t pmap_bits_default[] = {
 215                 REGULAR_PMAP,                   /* TYPE_IDX             0 */
 216                 X86_PG_V,                       /* PG_V_IDX             1 */
 217                 X86_PG_RW,                      /* PG_RW_IDX            2 */
 218                 X86_PG_U,                       /* PG_U_IDX             3 */
 219                 X86_PG_A,                       /* PG_A_IDX             4 */
 220                 X86_PG_M,                       /* PG_M_IDX             5 */
 221                 X86_PG_PS,                      /* PG_PS_IDX3           6 */
 222                 X86_PG_G,                       /* PG_G_IDX             7 */
 223                 X86_PG_AVAIL1,                  /* PG_AVAIL1_IDX        8 */
 224                 X86_PG_AVAIL2,                  /* PG_AVAIL2_IDX        9 */
 225                 X86_PG_AVAIL3,                  /* PG_AVAIL3_IDX        10 */
 226                 X86_PG_NC_PWT | X86_PG_NC_PCD,  /* PG_N_IDX             11 */
 227                 X86_PG_NX,                      /* PG_NX_IDX            12 */
 228 };
 229 /*
 230  * Crashdump maps.
 231  */
 232 static pt_entry_t *pt_crashdumpmap;
 233 static caddr_t crashdumpmap;
 234
 235 static int pmap_debug = 0;
 236 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
 237     &pmap_debug, 0, "Debug pmap's");
 238 #ifdef PMAP_DEBUG2
 239 static int pmap_enter_debug = 0;
 240 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
 241     &pmap_enter_debug, 0, "Debug pmap_enter's");
 242 #endif
 243 static int pmap_yield_count = 64;
 244 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
 245     &pmap_yield_count, 0, "Yield during init_pt/release");
 246 static int pmap_mmu_optimize = 0;
 247 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
 248     &pmap_mmu_optimize, 0, "Share page table pages when possible");
 249 int pmap_fast_kernel_cpusync = 0;
 250 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
 251     &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
 252 int pmap_dynamic_delete = 0;
 253 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
 254     &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
 255
 256 static int pmap_nx_enable = 0;
 257 /* needs manual TUNABLE in early probe, see below */
 258
 259 #define DISABLE_PSE
 260
 261 /* Standard user access funtions */
 262 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
 263     size_t *lencopied);
 264 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
 265 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
 266 extern int std_fubyte (const uint8_t *base);
 267 extern int std_subyte (uint8_t *base, uint8_t byte);
 268 extern int32_t std_fuword32 (const uint32_t *base);
 269 extern int64_t std_fuword64 (const uint64_t *base);
 270 extern int std_suword64 (uint64_t *base, uint64_t word);
 271 extern int std_suword32 (uint32_t *base, int word);
 272 extern uint32_t std_swapu32 (volatile uint32_t *base, uint32_t v);
 273 extern uint64_t std_swapu64 (volatile uint64_t *base, uint64_t v);
 274
 275 static void pv_hold(pv_entry_t pv);
 276 static int _pv_hold_try(pv_entry_t pv
 277                                 PMAP_DEBUG_DECL);
 278 static void pv_drop(pv_entry_t pv);
 279 static void _pv_lock(pv_entry_t pv
 280                                 PMAP_DEBUG_DECL);
 281 static void pv_unlock(pv_entry_t pv);
 282 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
 283                                 PMAP_DEBUG_DECL);
 284 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp
 285                                 PMAP_DEBUG_DECL);
 286 static void _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL);
 287 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex,
 288                                 vm_pindex_t **pmarkp, int *errorp);
 289 static void pv_put(pv_entry_t pv);
 290 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
 291 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
 292                       pv_entry_t *pvpp);
 293 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
 294                       pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
 295 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
 296                         pmap_inval_bulk_t *bulk, int destroy);
 297 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
 298 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
 299                         pmap_inval_bulk_t *bulk);
 300
 301 struct pmap_scan_info;
 302 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
 303                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
 304                       pv_entry_t pt_pv, int sharept,
 305                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 306 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
 307                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
 308                       pv_entry_t pt_pv, int sharept,
 309                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 310
 311 static void i386_protection_init (void);
 312 static void create_pagetables(vm_paddr_t *firstaddr);
 313 static void pmap_remove_all (vm_page_t m);
 314 static boolean_t pmap_testbit (vm_page_t m, int bit);
 315
 316 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
 317 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
 318
 319 static void pmap_pinit_defaults(struct pmap *pmap);
 320 static void pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark);
 321 static void pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark);
 322
 323 static unsigned pdir4mb;
 324
 325 static int
 326 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
 327 {
 328         if (pv1->pv_pindex < pv2->pv_pindex)
 329                 return(-1);
 330         if (pv1->pv_pindex > pv2->pv_pindex)
 331                 return(1);
 332         return(0);
 333 }
 334
 335 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
 336              pv_entry_compare, vm_pindex_t, pv_pindex);
 337
 338 static __inline
 339 void
 340 pmap_page_stats_adding(vm_page_t m)
 341 {
 342         globaldata_t gd = mycpu;
 343
 344         if (TAILQ_EMPTY(&m->md.pv_list)) {
 345                 ++gd->gd_vmtotal.t_arm;
 346         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 347                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 348                 ++gd->gd_vmtotal.t_armshr;
 349                 ++gd->gd_vmtotal.t_avmshr;
 350         } else {
 351                 ++gd->gd_vmtotal.t_avmshr;
 352         }
 353 }
 354
 355 static __inline
 356 void
 357 pmap_page_stats_deleting(vm_page_t m)
 358 {
 359         globaldata_t gd = mycpu;
 360
 361         if (TAILQ_EMPTY(&m->md.pv_list)) {
 362                 --gd->gd_vmtotal.t_arm;
 363         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 364                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 365                 --gd->gd_vmtotal.t_armshr;
 366                 --gd->gd_vmtotal.t_avmshr;
 367         } else {
 368                 --gd->gd_vmtotal.t_avmshr;
 369         }
 370 }
 371
 372 /*
 373  * Move the kernel virtual free pointer to the next
 374  * 2MB.  This is used to help improve performance
 375  * by using a large (2MB) page for much of the kernel
 376  * (.text, .data, .bss)
 377  */
 378 static
 379 vm_offset_t
 380 pmap_kmem_choose(vm_offset_t addr)
 381 {
 382         vm_offset_t newaddr = addr;
 383
 384         newaddr = roundup2(addr, NBPDR);
 385         return newaddr;
 386 }
 387
 388 /*
 389  * pmap_pte_quick:
 390  *
 391  *      Super fast pmap_pte routine best used when scanning the pv lists.
 392  *      This eliminates many course-grained invltlb calls.  Note that many of
 393  *      the pv list scans are across different pmaps and it is very wasteful
 394  *      to do an entire invltlb when checking a single mapping.
 395  */
 396 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
 397
 398 static
 399 pt_entry_t *
 400 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
 401 {
 402         return pmap_pte(pmap, va);
 403 }
 404
 405 /*
 406  * Returns the pindex of a page table entry (representing a terminal page).
 407  * There are NUPTE_TOTAL page table entries possible (a huge number)
 408  *
 409  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
 410  * We want to properly translate negative KVAs.
 411  */
 412 static __inline
 413 vm_pindex_t
 414 pmap_pte_pindex(vm_offset_t va)
 415 {
 416         return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
 417 }
 418
 419 /*
 420  * Returns the pindex of a page table.
 421  */
 422 static __inline
 423 vm_pindex_t
 424 pmap_pt_pindex(vm_offset_t va)
 425 {
 426         return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
 427 }
 428
 429 /*
 430  * Returns the pindex of a page directory.
 431  */
 432 static __inline
 433 vm_pindex_t
 434 pmap_pd_pindex(vm_offset_t va)
 435 {
 436         return (NUPTE_TOTAL + NUPT_TOTAL +
 437                 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
 438 }
 439
 440 static __inline
 441 vm_pindex_t
 442 pmap_pdp_pindex(vm_offset_t va)
 443 {
 444         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
 445                 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
 446 }
 447
 448 static __inline
 449 vm_pindex_t
 450 pmap_pml4_pindex(void)
 451 {
 452         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
 453 }
 454
 455 /*
 456  * Return various clipped indexes for a given VA
 457  *
 458  * Returns the index of a pt in a page directory, representing a page
 459  * table.
 460  */
 461 static __inline
 462 vm_pindex_t
 463 pmap_pt_index(vm_offset_t va)
 464 {
 465         return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
 466 }
 467
 468 /*
 469  * Returns the index of a pd in a page directory page, representing a page
 470  * directory.
 471  */
 472 static __inline
 473 vm_pindex_t
 474 pmap_pd_index(vm_offset_t va)
 475 {
 476         return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
 477 }
 478
 479 /*
 480  * Returns the index of a pdp in the pml4 table, representing a page
 481  * directory page.
 482  */
 483 static __inline
 484 vm_pindex_t
 485 pmap_pdp_index(vm_offset_t va)
 486 {
 487         return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
 488 }
 489
 490 /*
 491  * The placemarker hash must be broken up into four zones so lock
 492  * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
 493  *
 494  * Placemarkers are used to 'lock' page table indices that do not have
 495  * a pv_entry.  This allows the pmap to support managed and unmanaged
 496  * pages and shared page tables.
 497  */
 498 #define PM_PLACE_BASE   (PM_PLACEMARKS >> 2)
 499
 500 static __inline
 501 vm_pindex_t *
 502 pmap_placemarker_hash(pmap_t pmap, vm_pindex_t pindex)
 503 {
 504         int hi;
 505
 506         if (pindex < pmap_pt_pindex(0))         /* zone 0 - PTE */
 507                 hi = 0;
 508         else if (pindex < pmap_pd_pindex(0))    /* zone 1 - PT */
 509                 hi = PM_PLACE_BASE;
 510         else if (pindex < pmap_pdp_pindex(0))   /* zone 2 - PD */
 511                 hi = PM_PLACE_BASE << 1;
 512         else                                    /* zone 3 - PDP (and PML4E) */
 513                 hi = PM_PLACE_BASE | (PM_PLACE_BASE << 1);
 514         hi += pindex & (PM_PLACE_BASE - 1);
 515
 516         return (&pmap->pm_placemarks[hi]);
 517 }
 518
 519
 520 /*
 521  * Generic procedure to index a pte from a pt, pd, or pdp.
 522  *
 523  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
 524  *       a page table page index but is instead of PV lookup index.
 525  */
 526 static
 527 void *
 528 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
 529 {
 530         pt_entry_t *pte;
 531
 532         pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
 533         return(&pte[pindex]);
 534 }
 535
 536 /*
 537  * Return pointer to PDP slot in the PML4
 538  */
 539 static __inline
 540 pml4_entry_t *
 541 pmap_pdp(pmap_t pmap, vm_offset_t va)
 542 {
 543         return (&pmap->pm_pml4[pmap_pdp_index(va)]);
 544 }
 545
 546 /*
 547  * Return pointer to PD slot in the PDP given a pointer to the PDP
 548  */
 549 static __inline
 550 pdp_entry_t *
 551 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
 552 {
 553         pdp_entry_t *pd;
 554
 555         pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
 556         return (&pd[pmap_pd_index(va)]);
 557 }
 558
 559 /*
 560  * Return pointer to PD slot in the PDP.
 561  */
 562 static __inline
 563 pdp_entry_t *
 564 pmap_pd(pmap_t pmap, vm_offset_t va)
 565 {
 566         pml4_entry_t *pdp;
 567
 568         pdp = pmap_pdp(pmap, va);
 569         if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
 570                 return NULL;
 571         return (pmap_pdp_to_pd(*pdp, va));
 572 }
 573
 574 /*
 575  * Return pointer to PT slot in the PD given a pointer to the PD
 576  */
 577 static __inline
 578 pd_entry_t *
 579 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
 580 {
 581         pd_entry_t *pt;
 582
 583         pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
 584         return (&pt[pmap_pt_index(va)]);
 585 }
 586
 587 /*
 588  * Return pointer to PT slot in the PD
 589  *
 590  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
 591  *                   so we cannot lookup the PD via the PDP.  Instead we
 592  *                   must look it up via the pmap.
 593  */
 594 static __inline
 595 pd_entry_t *
 596 pmap_pt(pmap_t pmap, vm_offset_t va)
 597 {
 598         pdp_entry_t *pd;
 599         pv_entry_t pv;
 600         vm_pindex_t pd_pindex;
 601         vm_paddr_t phys;
 602
 603         if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
 604                 pd_pindex = pmap_pd_pindex(va);
 605                 spin_lock_shared(&pmap->pm_spin);
 606                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
 607                 if (pv == NULL || pv->pv_m == NULL) {
 608                         spin_unlock_shared(&pmap->pm_spin);
 609                         return NULL;
 610                 }
 611                 phys = VM_PAGE_TO_PHYS(pv->pv_m);
 612                 spin_unlock_shared(&pmap->pm_spin);
 613                 return (pmap_pd_to_pt(phys, va));
 614         } else {
 615                 pd = pmap_pd(pmap, va);
 616                 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
 617                          return NULL;
 618                 return (pmap_pd_to_pt(*pd, va));
 619         }
 620 }
 621
 622 /*
 623  * Return pointer to PTE slot in the PT given a pointer to the PT
 624  */
 625 static __inline
 626 pt_entry_t *
 627 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
 628 {
 629         pt_entry_t *pte;
 630
 631         pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
 632         return (&pte[pmap_pte_index(va)]);
 633 }
 634
 635 /*
 636  * Return pointer to PTE slot in the PT
 637  */
 638 static __inline
 639 pt_entry_t *
 640 pmap_pte(pmap_t pmap, vm_offset_t va)
 641 {
 642         pd_entry_t *pt;
 643
 644         pt = pmap_pt(pmap, va);
 645         if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
 646                  return NULL;
 647         if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
 648                 return ((pt_entry_t *)pt);
 649         return (pmap_pt_to_pte(*pt, va));
 650 }
 651
 652 /*
 653  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
 654  * the PT layer.  This will speed up core pmap operations considerably.
 655  *
 656  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
 657  *       must be in a known associated state (typically by being locked when
 658  *       the pmap spinlock isn't held).  We allow the race for that case.
 659  *
 660  * NOTE: pm_pvhint is only accessed (read) with the spin-lock held, using
 661  *       cpu_ccfence() to prevent compiler optimizations from reloading the
 662  *       field.
 663  */
 664 static __inline
 665 void
 666 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
 667 {
 668         if (pindex >= pmap_pt_pindex(0) && pindex < pmap_pd_pindex(0)) {
 669                 if (pv->pv_pmap)
 670                         pv->pv_pmap->pm_pvhint = pv;
 671         }
 672 }
 673
 674
 675 /*
 676  * Return address of PT slot in PD (KVM only)
 677  *
 678  * Cannot be used for user page tables because it might interfere with
 679  * the shared page-table-page optimization (pmap_mmu_optimize).
 680  */
 681 static __inline
 682 pd_entry_t *
 683 vtopt(vm_offset_t va)
 684 {
 685         uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
 686                                   NPML4EPGSHIFT)) - 1);
 687
 688         return (PDmap + ((va >> PDRSHIFT) & mask));
 689 }
 690
 691 /*
 692  * KVM - return address of PTE slot in PT
 693  */
 694 static __inline
 695 pt_entry_t *
 696 vtopte(vm_offset_t va)
 697 {
 698         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 699                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 700
 701         return (PTmap + ((va >> PAGE_SHIFT) & mask));
 702 }
 703
 704 /*
 705  * Returns the physical address translation from va for a user address.
 706  * (vm_paddr_t)-1 is returned on failure.
 707  */
 708 vm_paddr_t
 709 uservtophys(vm_offset_t va)
 710 {
 711         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 712                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 713         vm_paddr_t pa;
 714         pt_entry_t pte;
 715         pmap_t pmap;
 716
 717         pmap = vmspace_pmap(mycpu->gd_curthread->td_lwp->lwp_vmspace);
 718         pa = (vm_paddr_t)-1;
 719         if (va < VM_MAX_USER_ADDRESS) {
 720                 pte = kreadmem64(PTmap + ((va >> PAGE_SHIFT) & mask));
 721                 if (pte & pmap->pmap_bits[PG_V_IDX])
 722                         pa = (pte & PG_FRAME) | (va & PAGE_MASK);
 723         }
 724         return pa;
 725 }
 726
 727 static uint64_t
 728 allocpages(vm_paddr_t *firstaddr, long n)
 729 {
 730         uint64_t ret;
 731
 732         ret = *firstaddr;
 733         bzero((void *)ret, n * PAGE_SIZE);
 734         *firstaddr += n * PAGE_SIZE;
 735         return (ret);
 736 }
 737
 738 static
 739 void
 740 create_pagetables(vm_paddr_t *firstaddr)
 741 {
 742         long i;         /* must be 64 bits */
 743         long nkpt_base;
 744         long nkpt_phys;
 745         int j;
 746
 747         /*
 748          * We are running (mostly) V=P at this point
 749          *
 750          * Calculate NKPT - number of kernel page tables.  We have to
 751          * accomodoate prealloction of the vm_page_array, dump bitmap,
 752          * MSGBUF_SIZE, and other stuff.  Be generous.
 753          *
 754          * Maxmem is in pages.
 755          *
 756          * ndmpdp is the number of 1GB pages we wish to map.
 757          */
 758         ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
 759         if (ndmpdp < 4)         /* Minimum 4GB of dirmap */
 760                 ndmpdp = 4;
 761         KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
 762
 763         /*
 764          * Starting at the beginning of kvm (not KERNBASE).
 765          */
 766         nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
 767         nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
 768         nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
 769                        ndmpdp) + 511) / 512;
 770         nkpt_phys += 128;
 771
 772         /*
 773          * Starting at KERNBASE - map 2G worth of page table pages.
 774          * KERNBASE is offset -2G from the end of kvm.
 775          */
 776         nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
 777
 778         /*
 779          * Allocate pages
 780          */
 781         KPTbase = allocpages(firstaddr, nkpt_base);
 782         KPTphys = allocpages(firstaddr, nkpt_phys);
 783         KPML4phys = allocpages(firstaddr, 1);
 784         KPDPphys = allocpages(firstaddr, NKPML4E);
 785         KPDphys = allocpages(firstaddr, NKPDPE);
 786
 787         /*
 788          * Calculate the page directory base for KERNBASE,
 789          * that is where we start populating the page table pages.
 790          * Basically this is the end - 2.
 791          */
 792         KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
 793
 794         DMPDPphys = allocpages(firstaddr, NDMPML4E);
 795         if ((amd_feature & AMDID_PAGE1GB) == 0)
 796                 DMPDphys = allocpages(firstaddr, ndmpdp);
 797         dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
 798
 799         /*
 800          * Fill in the underlying page table pages for the area around
 801          * KERNBASE.  This remaps low physical memory to KERNBASE.
 802          *
 803          * Read-only from zero to physfree
 804          * XXX not fully used, underneath 2M pages
 805          */
 806         for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
 807                 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
 808                 ((pt_entry_t *)KPTbase)[i] |=
 809                     pmap_bits_default[PG_RW_IDX] |
 810                     pmap_bits_default[PG_V_IDX] |
 811                     pmap_bits_default[PG_G_IDX];
 812         }
 813
 814         /*
 815          * Now map the initial kernel page tables.  One block of page
 816          * tables is placed at the beginning of kernel virtual memory,
 817          * and another block is placed at KERNBASE to map the kernel binary,
 818          * data, bss, and initial pre-allocations.
 819          */
 820         for (i = 0; i < nkpt_base; i++) {
 821                 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
 822                 ((pd_entry_t *)KPDbase)[i] |=
 823                     pmap_bits_default[PG_RW_IDX] |
 824                     pmap_bits_default[PG_V_IDX];
 825         }
 826         for (i = 0; i < nkpt_phys; i++) {
 827                 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
 828                 ((pd_entry_t *)KPDphys)[i] |=
 829                     pmap_bits_default[PG_RW_IDX] |
 830                     pmap_bits_default[PG_V_IDX];
 831         }
 832
 833         /*
 834          * Map from zero to end of allocations using 2M pages as an
 835          * optimization.  This will bypass some of the KPTBase pages
 836          * above in the KERNBASE area.
 837          */
 838         for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
 839                 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
 840                 ((pd_entry_t *)KPDbase)[i] |=
 841                     pmap_bits_default[PG_RW_IDX] |
 842                     pmap_bits_default[PG_V_IDX] |
 843                     pmap_bits_default[PG_PS_IDX] |
 844                     pmap_bits_default[PG_G_IDX];
 845         }
 846
 847         /*
 848          * And connect up the PD to the PDP.  The kernel pmap is expected
 849          * to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
 850          */
 851         for (i = 0; i < NKPDPE; i++) {
 852                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
 853                                 KPDphys + (i << PAGE_SHIFT);
 854                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
 855                     pmap_bits_default[PG_RW_IDX] |
 856                     pmap_bits_default[PG_V_IDX] |
 857                     pmap_bits_default[PG_U_IDX];
 858         }
 859
 860         /*
 861          * Now set up the direct map space using either 2MB or 1GB pages
 862          * Preset PG_M and PG_A because demotion expects it.
 863          *
 864          * When filling in entries in the PD pages make sure any excess
 865          * entries are set to zero as we allocated enough PD pages
 866          */
 867         if ((amd_feature & AMDID_PAGE1GB) == 0) {
 868                 for (i = 0; i < NPDEPG * ndmpdp; i++) {
 869                         ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
 870                         ((pd_entry_t *)DMPDphys)[i] |=
 871                             pmap_bits_default[PG_RW_IDX] |
 872                             pmap_bits_default[PG_V_IDX] |
 873                             pmap_bits_default[PG_PS_IDX] |
 874                             pmap_bits_default[PG_G_IDX] |
 875                             pmap_bits_default[PG_M_IDX] |
 876                             pmap_bits_default[PG_A_IDX];
 877                 }
 878
 879                 /*
 880                  * And the direct map space's PDP
 881                  */
 882                 for (i = 0; i < ndmpdp; i++) {
 883                         ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
 884                                                         (i << PAGE_SHIFT);
 885                         ((pdp_entry_t *)DMPDPphys)[i] |=
 886                             pmap_bits_default[PG_RW_IDX] |
 887                             pmap_bits_default[PG_V_IDX] |
 888                             pmap_bits_default[PG_U_IDX];
 889                 }
 890         } else {
 891                 for (i = 0; i < ndmpdp; i++) {
 892                         ((pdp_entry_t *)DMPDPphys)[i] =
 893                                                 (vm_paddr_t)i << PDPSHIFT;
 894                         ((pdp_entry_t *)DMPDPphys)[i] |=
 895                             pmap_bits_default[PG_RW_IDX] |
 896                             pmap_bits_default[PG_V_IDX] |
 897                             pmap_bits_default[PG_PS_IDX] |
 898                             pmap_bits_default[PG_G_IDX] |
 899                             pmap_bits_default[PG_M_IDX] |
 900                             pmap_bits_default[PG_A_IDX];
 901                 }
 902         }
 903
 904         /* And recursively map PML4 to itself in order to get PTmap */
 905         ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
 906         ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
 907             pmap_bits_default[PG_RW_IDX] |
 908             pmap_bits_default[PG_V_IDX] |
 909             pmap_bits_default[PG_U_IDX];
 910
 911         /*
 912          * Connect the Direct Map slots up to the PML4
 913          */
 914         for (j = 0; j < NDMPML4E; ++j) {
 915                 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
 916                     (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
 917                     pmap_bits_default[PG_RW_IDX] |
 918                     pmap_bits_default[PG_V_IDX] |
 919                     pmap_bits_default[PG_U_IDX];
 920         }
 921
 922         /*
 923          * Connect the KVA slot up to the PML4
 924          */
 925         ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
 926         ((pdp_entry_t *)KPML4phys)[KPML4I] |=
 927             pmap_bits_default[PG_RW_IDX] |
 928             pmap_bits_default[PG_V_IDX] |
 929             pmap_bits_default[PG_U_IDX];
 930 }
 931
 932 /*
 933  *      Bootstrap the system enough to run with virtual memory.
 934  *
 935  *      On the i386 this is called after mapping has already been enabled
 936  *      and just syncs the pmap module with what has already been done.
 937  *      [We can't call it easily with mapping off since the kernel is not
 938  *      mapped with PA == VA, hence we would have to relocate every address
 939  *      from the linked base (virtual) address "KERNBASE" to the actual
 940  *      (physical) address starting relative to 0]
 941  */
 942 void
 943 pmap_bootstrap(vm_paddr_t *firstaddr)
 944 {
 945         vm_offset_t va;
 946         pt_entry_t *pte;
 947         int i;
 948
 949         KvaStart = VM_MIN_KERNEL_ADDRESS;
 950         KvaEnd = VM_MAX_KERNEL_ADDRESS;
 951         KvaSize = KvaEnd - KvaStart;
 952
 953         avail_start = *firstaddr;
 954
 955         /*
 956          * Create an initial set of page tables to run the kernel in.
 957          */
 958         create_pagetables(firstaddr);
 959
 960         virtual2_start = KvaStart;
 961         virtual2_end = PTOV_OFFSET;
 962
 963         virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
 964         virtual_start = pmap_kmem_choose(virtual_start);
 965
 966         virtual_end = VM_MAX_KERNEL_ADDRESS;
 967
 968         /* XXX do %cr0 as well */
 969         load_cr4(rcr4() | CR4_PGE | CR4_PSE);
 970         load_cr3(KPML4phys);
 971
 972         /*
 973          * Initialize protection array.
 974          */
 975         i386_protection_init();
 976
 977         /*
 978          * The kernel's pmap is statically allocated so we don't have to use
 979          * pmap_create, which is unlikely to work correctly at this part of
 980          * the boot sequence (XXX and which no longer exists).
 981          */
 982         kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
 983         kernel_pmap.pm_count = 1;
 984         CPUMASK_ASSALLONES(kernel_pmap.pm_active);
 985         RB_INIT(&kernel_pmap.pm_pvroot);
 986         spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
 987         for (i = 0; i < PM_PLACEMARKS; ++i)
 988                 kernel_pmap.pm_placemarks[i] = PM_NOPLACEMARK;
 989
 990         /*
 991          * Reserve some special page table entries/VA space for temporary
 992          * mapping of pages.
 993          */
 994 #define SYSMAP(c, p, v, n)      \
 995         v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
 996
 997         va = virtual_start;
 998         pte = vtopte(va);
 999
1000         /*
1001          * CMAP1/CMAP2 are used for zeroing and copying pages.
1002          */
1003         SYSMAP(caddr_t, CMAP1, CADDR1, 1)
1004
1005         /*
1006          * Crashdump maps.
1007          */
1008         SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
1009
1010         /*
1011          * ptvmmap is used for reading arbitrary physical pages via
1012          * /dev/mem.
1013          */
1014         SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
1015
1016         /*
1017          * msgbufp is used to map the system message buffer.
1018          * XXX msgbufmap is not used.
1019          */
1020         SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
1021                atop(round_page(MSGBUF_SIZE)))
1022
1023         virtual_start = va;
1024         virtual_start = pmap_kmem_choose(virtual_start);
1025
1026         *CMAP1 = 0;
1027
1028         /*
1029          * PG_G is terribly broken on SMP because we IPI invltlb's in some
1030          * cases rather then invl1pg.  Actually, I don't even know why it
1031          * works under UP because self-referential page table mappings
1032          */
1033 //      pgeflag = 0;
1034
1035 /*
1036  * Initialize the 4MB page size flag
1037  */
1038 //      pseflag = 0;
1039 /*
1040  * The 4MB page version of the initial
1041  * kernel page mapping.
1042  */
1043         pdir4mb = 0;
1044
1045 #if !defined(DISABLE_PSE)
1046         if (cpu_feature & CPUID_PSE) {
1047                 pt_entry_t ptditmp;
1048                 /*
1049                  * Note that we have enabled PSE mode
1050                  */
1051 //              pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
1052                 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
1053                 ptditmp &= ~(NBPDR - 1);
1054                 ptditmp |= pmap_bits_default[PG_V_IDX] |
1055                     pmap_bits_default[PG_RW_IDX] |
1056                     pmap_bits_default[PG_PS_IDX] |
1057                     pmap_bits_default[PG_U_IDX];
1058 //                  pgeflag;
1059                 pdir4mb = ptditmp;
1060         }
1061 #endif
1062         cpu_invltlb();
1063
1064         /* Initialize the PAT MSR */
1065         pmap_init_pat();
1066         pmap_pinit_defaults(&kernel_pmap);
1067
1068         TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1069                           &pmap_fast_kernel_cpusync);
1070
1071 }
1072
1073 /*
1074  * Setup the PAT MSR.
1075  */
1076 void
1077 pmap_init_pat(void)
1078 {
1079         uint64_t pat_msr;
1080         u_long cr0, cr4;
1081
1082         /*
1083          * Default values mapping PATi,PCD,PWT bits at system reset.
1084          * The default values effectively ignore the PATi bit by
1085          * repeating the encodings for 0-3 in 4-7, and map the PCD
1086          * and PWT bit combinations to the expected PAT types.
1087          */
1088         pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |        /* 000 */
1089                   PAT_VALUE(1, PAT_WRITE_THROUGH) |     /* 001 */
1090                   PAT_VALUE(2, PAT_UNCACHED) |          /* 010 */
1091                   PAT_VALUE(3, PAT_UNCACHEABLE) |       /* 011 */
1092                   PAT_VALUE(4, PAT_WRITE_BACK) |        /* 100 */
1093                   PAT_VALUE(5, PAT_WRITE_THROUGH) |     /* 101 */
1094                   PAT_VALUE(6, PAT_UNCACHED) |          /* 110 */
1095                   PAT_VALUE(7, PAT_UNCACHEABLE);        /* 111 */
1096         pat_pte_index[PAT_WRITE_BACK]   = 0;
1097         pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
1098         pat_pte_index[PAT_UNCACHED]     = X86_PG_NC_PCD;
1099         pat_pte_index[PAT_UNCACHEABLE]  = X86_PG_NC_PCD | X86_PG_NC_PWT;
1100         pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1101         pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1102
1103         if (cpu_feature & CPUID_PAT) {
1104                 /*
1105                  * If we support the PAT then set-up entries for
1106                  * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1107                  * 5 and 6.
1108                  */
1109                 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1110                           PAT_VALUE(5, PAT_WRITE_PROTECTED);
1111                 pat_msr = (pat_msr & ~PAT_MASK(6)) |
1112                           PAT_VALUE(6, PAT_WRITE_COMBINING);
1113                 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1114                 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PCD;
1115
1116                 /*
1117                  * Then enable the PAT
1118                  */
1119
1120                 /* Disable PGE. */
1121                 cr4 = rcr4();
1122                 load_cr4(cr4 & ~CR4_PGE);
1123
1124                 /* Disable caches (CD = 1, NW = 0). */
1125                 cr0 = rcr0();
1126                 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1127
1128                 /* Flushes caches and TLBs. */
1129                 wbinvd();
1130                 cpu_invltlb();
1131
1132                 /* Update PAT and index table. */
1133                 wrmsr(MSR_PAT, pat_msr);
1134
1135                 /* Flush caches and TLBs again. */
1136                 wbinvd();
1137                 cpu_invltlb();
1138
1139                 /* Restore caches and PGE. */
1140                 load_cr0(cr0);
1141                 load_cr4(cr4);
1142                 PatMsr = pat_msr;
1143         }
1144 }
1145
1146 /*
1147  * Set 4mb pdir for mp startup
1148  */
1149 void
1150 pmap_set_opt(void)
1151 {
1152         if (cpu_feature & CPUID_PSE) {
1153                 load_cr4(rcr4() | CR4_PSE);
1154                 if (pdir4mb && mycpu->gd_cpuid == 0) {  /* only on BSP */
1155                         cpu_invltlb();
1156                 }
1157         }
1158 }
1159
1160 /*
1161  *      Initialize the pmap module.
1162  *      Called by vm_init, to initialize any structures that the pmap
1163  *      system needs to map virtual memory.
1164  *      pmap_init has been enhanced to support in a fairly consistant
1165  *      way, discontiguous physical memory.
1166  */
1167 void
1168 pmap_init(void)
1169 {
1170         int i;
1171         int initial_pvs;
1172
1173         /*
1174          * Allocate memory for random pmap data structures.  Includes the
1175          * pv_head_table.
1176          */
1177
1178         for (i = 0; i < vm_page_array_size; i++) {
1179                 vm_page_t m;
1180
1181                 m = &vm_page_array[i];
1182                 TAILQ_INIT(&m->md.pv_list);
1183         }
1184
1185         /*
1186          * init the pv free list
1187          */
1188         initial_pvs = vm_page_array_size;
1189         if (initial_pvs < MINPV)
1190                 initial_pvs = MINPV;
1191         pvzone = &pvzone_store;
1192         pvinit = (void *)kmem_alloc(&kernel_map,
1193                                     initial_pvs * sizeof (struct pv_entry),
1194                                     VM_SUBSYS_PVENTRY);
1195         zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1196                   pvinit, initial_pvs);
1197
1198         /*
1199          * Now it is safe to enable pv_table recording.
1200          */
1201         pmap_initialized = TRUE;
1202 }
1203
1204 /*
1205  * Initialize the address space (zone) for the pv_entries.  Set a
1206  * high water mark so that the system can recover from excessive
1207  * numbers of pv entries.
1208  */
1209 void
1210 pmap_init2(void)
1211 {
1212         int shpgperproc = PMAP_SHPGPERPROC;
1213         int entry_max;
1214
1215         TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1216         pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1217         TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1218         pv_entry_high_water = 9 * (pv_entry_max / 10);
1219
1220         /*
1221          * Subtract out pages already installed in the zone (hack)
1222          */
1223         entry_max = pv_entry_max - vm_page_array_size;
1224         if (entry_max <= 0)
1225                 entry_max = 1;
1226
1227         zinitna(pvzone, NULL, 0, entry_max, ZONE_INTERRUPT);
1228
1229         /*
1230          * Enable dynamic deletion of empty higher-level page table pages
1231          * by default only if system memory is < 8GB (use 7GB for slop).
1232          * This can save a little memory, but imposes significant
1233          * performance overhead for things like bulk builds, and for programs
1234          * which do a lot of memory mapping and memory unmapping.
1235          */
1236         if (pmap_dynamic_delete < 0) {
1237                 if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1238                         pmap_dynamic_delete = 1;
1239                 else
1240                         pmap_dynamic_delete = 0;
1241         }
1242 }
1243
1244 /*
1245  * Typically used to initialize a fictitious page by vm/device_pager.c
1246  */
1247 void
1248 pmap_page_init(struct vm_page *m)
1249 {
1250         vm_page_init(m);
1251         TAILQ_INIT(&m->md.pv_list);
1252 }
1253
1254 /***************************************************
1255  * Low level helper routines.....
1256  ***************************************************/
1257
1258 /*
1259  * this routine defines the region(s) of memory that should
1260  * not be tested for the modified bit.
1261  */
1262 static __inline
1263 int
1264 pmap_track_modified(vm_pindex_t pindex)
1265 {
1266         vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1267         if ((va < clean_sva) || (va >= clean_eva))
1268                 return 1;
1269         else
1270                 return 0;
1271 }
1272
1273 /*
1274  * Extract the physical page address associated with the map/VA pair.
1275  * The page must be wired for this to work reliably.
1276  */
1277 vm_paddr_t
1278 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
1279 {
1280         vm_paddr_t rtval;
1281         pv_entry_t pt_pv;
1282         pt_entry_t *ptep;
1283
1284         rtval = 0;
1285         if (va >= VM_MAX_USER_ADDRESS) {
1286                 /*
1287                  * Kernel page directories might be direct-mapped and
1288                  * there is typically no PV tracking of pte's
1289                  */
1290                 pd_entry_t *pt;
1291
1292                 pt = pmap_pt(pmap, va);
1293                 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1294                         if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1295                                 rtval = *pt & PG_PS_FRAME;
1296                                 rtval |= va & PDRMASK;
1297                         } else {
1298                                 ptep = pmap_pt_to_pte(*pt, va);
1299                                 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1300                                         rtval = *ptep & PG_FRAME;
1301                                         rtval |= va & PAGE_MASK;
1302                                 }
1303                         }
1304                 }
1305                 if (handlep)
1306                         *handlep = NULL;
1307         } else {
1308                 /*
1309                  * User pages currently do not direct-map the page directory
1310                  * and some pages might not used managed PVs.  But all PT's
1311                  * will have a PV.
1312                  */
1313                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1314                 if (pt_pv) {
1315                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1316                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1317                                 rtval = *ptep & PG_FRAME;
1318                                 rtval |= va & PAGE_MASK;
1319                         }
1320                         if (handlep)
1321                                 *handlep = pt_pv;       /* locked until done */
1322                         else
1323                                 pv_put (pt_pv);
1324                 } else if (handlep) {
1325                         *handlep = NULL;
1326                 }
1327         }
1328         return rtval;
1329 }
1330
1331 void
1332 pmap_extract_done(void *handle)
1333 {
1334         if (handle)
1335                 pv_put((pv_entry_t)handle);
1336 }
1337
1338 /*
1339  * Similar to extract but checks protections, SMP-friendly short-cut for
1340  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1341  * fall-through to the real fault code.  Does not work with HVM page
1342  * tables.
1343  *
1344  * if busyp is NULL the returned page, if not NULL, is held (and not busied).
1345  *
1346  * If busyp is not NULL and this function sets *busyp non-zero, the returned
1347  * page is busied (and not held).
1348  *
1349  * If busyp is not NULL and this function sets *busyp to zero, the returned
1350  * page is held (and not busied).
1351  *
1352  * If VM_PROT_WRITE or VM_PROT_OVERRIDE_WRITE is set in prot, and the pte
1353  * is already writable, the returned page will be dirtied.  If the pte
1354  * is not already writable NULL is returned.  In otherwords, if either
1355  * bit is set and a vm_page_t is returned, any COW will already have happened
1356  * and that page can be written by the caller.
1357  *
1358  * WARNING! THE RETURNED PAGE IS ONLY HELD AND NOT SUITABLE FOR READING
1359  *          OR WRITING AS-IS.
1360  */
1361 vm_page_t
1362 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot, int *busyp)
1363 {
1364         if (pmap &&
1365             va < VM_MAX_USER_ADDRESS &&
1366             (pmap->pm_flags & PMAP_HVM) == 0) {
1367                 pv_entry_t pt_pv;
1368                 pv_entry_t pte_pv;
1369                 pt_entry_t *ptep;
1370                 pt_entry_t req;
1371                 vm_page_t m;
1372                 int error;
1373
1374                 req = pmap->pmap_bits[PG_V_IDX] |
1375                       pmap->pmap_bits[PG_U_IDX];
1376                 if (prot & (VM_PROT_WRITE | VM_PROT_OVERRIDE_WRITE))
1377                         req |= pmap->pmap_bits[PG_RW_IDX];
1378
1379                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1380                 if (pt_pv == NULL)
1381                         return (NULL);
1382                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1383                 if ((*ptep & req) != req) {
1384                         pv_put(pt_pv);
1385                         return (NULL);
1386                 }
1387                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), NULL, &error);
1388                 if (pte_pv && error == 0) {
1389                         m = pte_pv->pv_m;
1390                         if (prot & (VM_PROT_WRITE | VM_PROT_OVERRIDE_WRITE)) {
1391                                 /* interlocked by presence of pv_entry */
1392                                 vm_page_dirty(m);
1393                         }
1394                         if (busyp) {
1395                                 if (prot & VM_PROT_WRITE) {
1396                                         if (vm_page_busy_try(m, TRUE))
1397                                                 m = NULL;
1398                                         *busyp = 1;
1399                                 } else {
1400                                         vm_page_hold(m);
1401                                         *busyp = 0;
1402                                 }
1403                         } else {
1404                                 vm_page_hold(m);
1405                         }
1406                         pv_put(pte_pv);
1407                 } else if (pte_pv) {
1408                         pv_drop(pte_pv);
1409                         m = NULL;
1410                 } else {
1411                         /* error, since we didn't request a placemarker */
1412                         m = NULL;
1413                 }
1414                 pv_put(pt_pv);
1415                 return(m);
1416         } else {
1417                 return(NULL);
1418         }
1419 }
1420
1421 /*
1422  * Extract the physical page address associated kernel virtual address.
1423  */
1424 vm_paddr_t
1425 pmap_kextract(vm_offset_t va)
1426 {
1427         pd_entry_t pt;          /* pt entry in pd */
1428         vm_paddr_t pa;
1429
1430         if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1431                 pa = DMAP_TO_PHYS(va);
1432         } else {
1433                 pt = *vtopt(va);
1434                 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1435                         pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1436                 } else {
1437                         /*
1438                          * Beware of a concurrent promotion that changes the
1439                          * PDE at this point!  For example, vtopte() must not
1440                          * be used to access the PTE because it would use the
1441                          * new PDE.  It is, however, safe to use the old PDE
1442                          * because the page table page is preserved by the
1443                          * promotion.
1444                          */
1445                         pa = *pmap_pt_to_pte(pt, va);
1446                         pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1447                 }
1448         }
1449         return pa;
1450 }
1451
1452 /***************************************************
1453  * Low level mapping routines.....
1454  ***************************************************/
1455
1456 /*
1457  * Routine: pmap_kenter
1458  * Function:
1459  *      Add a wired page to the KVA
1460  *      NOTE! note that in order for the mapping to take effect -- you
1461  *      should do an invltlb after doing the pmap_kenter().
1462  */
1463 void
1464 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1465 {
1466         pt_entry_t *ptep;
1467         pt_entry_t npte;
1468
1469         npte = pa |
1470                kernel_pmap.pmap_bits[PG_RW_IDX] |
1471                kernel_pmap.pmap_bits[PG_V_IDX];
1472 //             pgeflag;
1473         ptep = vtopte(va);
1474 #if 1
1475         pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1476 #else
1477         /* FUTURE */
1478         if (*ptep)
1479                 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1480         else
1481                 *ptep = npte;
1482 #endif
1483 }
1484
1485 /*
1486  * Similar to pmap_kenter(), except we only invalidate the mapping on the
1487  * current CPU.  Returns 0 if the previous pte was 0, 1 if it wasn't
1488  * (caller can conditionalize calling smp_invltlb()).
1489  */
1490 int
1491 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1492 {
1493         pt_entry_t *ptep;
1494         pt_entry_t npte;
1495         int res;
1496
1497         npte = pa | kernel_pmap.pmap_bits[PG_RW_IDX] |
1498                     kernel_pmap.pmap_bits[PG_V_IDX];
1499         // npte |= pgeflag;
1500         ptep = vtopte(va);
1501 #if 1
1502         res = 1;
1503 #else
1504         /* FUTURE */
1505         res = (*ptep != 0);
1506 #endif
1507         atomic_swap_long(ptep, npte);
1508         cpu_invlpg((void *)va);
1509
1510         return res;
1511 }
1512
1513 /*
1514  * Enter addresses into the kernel pmap but don't bother
1515  * doing any tlb invalidations.  Caller will do a rollup
1516  * invalidation via pmap_rollup_inval().
1517  */
1518 int
1519 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1520 {
1521         pt_entry_t *ptep;
1522         pt_entry_t npte;
1523         int res;
1524
1525         npte = pa |
1526             kernel_pmap.pmap_bits[PG_RW_IDX] |
1527             kernel_pmap.pmap_bits[PG_V_IDX];
1528 //          pgeflag;
1529         ptep = vtopte(va);
1530 #if 1
1531         res = 1;
1532 #else
1533         /* FUTURE */
1534         res = (*ptep != 0);
1535 #endif
1536         atomic_swap_long(ptep, npte);
1537         cpu_invlpg((void *)va);
1538
1539         return res;
1540 }
1541
1542 /*
1543  * remove a page from the kernel pagetables
1544  */
1545 void
1546 pmap_kremove(vm_offset_t va)
1547 {
1548         pt_entry_t *ptep;
1549
1550         ptep = vtopte(va);
1551         pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1552 }
1553
1554 void
1555 pmap_kremove_quick(vm_offset_t va)
1556 {
1557         pt_entry_t *ptep;
1558
1559         ptep = vtopte(va);
1560         (void)pte_load_clear(ptep);
1561         cpu_invlpg((void *)va);
1562 }
1563
1564 /*
1565  * Remove addresses from the kernel pmap but don't bother
1566  * doing any tlb invalidations.  Caller will do a rollup
1567  * invalidation via pmap_rollup_inval().
1568  */
1569 void
1570 pmap_kremove_noinval(vm_offset_t va)
1571 {
1572         pt_entry_t *ptep;
1573
1574         ptep = vtopte(va);
1575         (void)pte_load_clear(ptep);
1576 }
1577
1578 /*
1579  * XXX these need to be recoded.  They are not used in any critical path.
1580  */
1581 void
1582 pmap_kmodify_rw(vm_offset_t va)
1583 {
1584         atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1585         cpu_invlpg((void *)va);
1586 }
1587
1588 /* NOT USED
1589 void
1590 pmap_kmodify_nc(vm_offset_t va)
1591 {
1592         atomic_set_long(vtopte(va), PG_N);
1593         cpu_invlpg((void *)va);
1594 }
1595 */
1596
1597 /*
1598  * Used to map a range of physical addresses into kernel virtual
1599  * address space during the low level boot, typically to map the
1600  * dump bitmap, message buffer, and vm_page_array.
1601  *
1602  * These mappings are typically made at some pointer after the end of the
1603  * kernel text+data.
1604  *
1605  * We could return PHYS_TO_DMAP(start) here and not allocate any
1606  * via (*virtp), but then kmem from userland and kernel dumps won't
1607  * have access to the related pointers.
1608  */
1609 vm_offset_t
1610 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1611 {
1612         vm_offset_t va;
1613         vm_offset_t va_start;
1614
1615         /*return PHYS_TO_DMAP(start);*/
1616
1617         va_start = *virtp;
1618         va = va_start;
1619
1620         while (start < end) {
1621                 pmap_kenter_quick(va, start);
1622                 va += PAGE_SIZE;
1623                 start += PAGE_SIZE;
1624         }
1625         *virtp = va;
1626         return va_start;
1627 }
1628
1629 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
1630
1631 /*
1632  * Remove the specified set of pages from the data and instruction caches.
1633  *
1634  * In contrast to pmap_invalidate_cache_range(), this function does not
1635  * rely on the CPU's self-snoop feature, because it is intended for use
1636  * when moving pages into a different cache domain.
1637  */
1638 void
1639 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1640 {
1641         vm_offset_t daddr, eva;
1642         int i;
1643
1644         if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1645             (cpu_feature & CPUID_CLFSH) == 0)
1646                 wbinvd();
1647         else {
1648                 cpu_mfence();
1649                 for (i = 0; i < count; i++) {
1650                         daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1651                         eva = daddr + PAGE_SIZE;
1652                         for (; daddr < eva; daddr += cpu_clflush_line_size)
1653                                 clflush(daddr);
1654                 }
1655                 cpu_mfence();
1656         }
1657 }
1658
1659 void
1660 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1661 {
1662         KASSERT((sva & PAGE_MASK) == 0,
1663             ("pmap_invalidate_cache_range: sva not page-aligned"));
1664         KASSERT((eva & PAGE_MASK) == 0,
1665             ("pmap_invalidate_cache_range: eva not page-aligned"));
1666
1667         if (cpu_feature & CPUID_SS) {
1668                 ; /* If "Self Snoop" is supported, do nothing. */
1669         } else {
1670                 /* Globally invalidate caches */
1671                 cpu_wbinvd_on_all_cpus();
1672         }
1673 }
1674
1675 /*
1676  * Invalidate the specified range of virtual memory on all cpus associated
1677  * with the pmap.
1678  */
1679 void
1680 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1681 {
1682         pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1683 }
1684
1685 /*
1686  * Add a list of wired pages to the kva.  This routine is used for temporary
1687  * kernel mappings such as those found in buffer cache buffer.  Page
1688  * modifications and accesses are not tracked or recorded.
1689  *
1690  * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1691  *       semantics as previous mappings may have been zerod without any
1692  *       invalidation.
1693  *
1694  * The page *must* be wired.
1695  */
1696 static __inline void
1697 _pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count, int doinval)
1698 {
1699         vm_offset_t end_va;
1700         vm_offset_t va;
1701
1702         end_va = beg_va + count * PAGE_SIZE;
1703
1704         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1705                 pt_entry_t pte;
1706                 pt_entry_t *ptep;
1707
1708                 ptep = vtopte(va);
1709                 pte = VM_PAGE_TO_PHYS(*m) |
1710                         kernel_pmap.pmap_bits[PG_RW_IDX] |
1711                         kernel_pmap.pmap_bits[PG_V_IDX] |
1712                         kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1713 //              pgeflag;
1714                 atomic_swap_long(ptep, pte);
1715                 m++;
1716         }
1717         if (doinval)
1718                 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1719 }
1720
1721 void
1722 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1723 {
1724         _pmap_qenter(beg_va, m, count, 1);
1725 }
1726
1727 void
1728 pmap_qenter_noinval(vm_offset_t beg_va, vm_page_t *m, int count)
1729 {
1730         _pmap_qenter(beg_va, m, count, 0);
1731 }
1732
1733 /*
1734  * This routine jerks page mappings from the kernel -- it is meant only
1735  * for temporary mappings such as those found in buffer cache buffers.
1736  * No recording modified or access status occurs.
1737  *
1738  * MPSAFE, INTERRUPT SAFE (cluster callback)
1739  */
1740 void
1741 pmap_qremove(vm_offset_t beg_va, int count)
1742 {
1743         vm_offset_t end_va;
1744         vm_offset_t va;
1745
1746         end_va = beg_va + count * PAGE_SIZE;
1747
1748         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1749                 pt_entry_t *pte;
1750
1751                 pte = vtopte(va);
1752                 (void)pte_load_clear(pte);
1753                 cpu_invlpg((void *)va);
1754         }
1755         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1756 }
1757
1758 /*
1759  * This routine removes temporary kernel mappings, only invalidating them
1760  * on the current cpu.  It should only be used under carefully controlled
1761  * conditions.
1762  */
1763 void
1764 pmap_qremove_quick(vm_offset_t beg_va, int count)
1765 {
1766         vm_offset_t end_va;
1767         vm_offset_t va;
1768
1769         end_va = beg_va + count * PAGE_SIZE;
1770
1771         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1772                 pt_entry_t *pte;
1773
1774                 pte = vtopte(va);
1775                 (void)pte_load_clear(pte);
1776                 cpu_invlpg((void *)va);
1777         }
1778 }
1779
1780 /*
1781  * This routine removes temporary kernel mappings *without* invalidating
1782  * the TLB.  It can only be used on permanent kva reservations such as those
1783  * found in buffer cache buffers, under carefully controlled circumstances.
1784  *
1785  * NOTE: Repopulating these KVAs requires unconditional invalidation.
1786  *       (pmap_qenter() does unconditional invalidation).
1787  */
1788 void
1789 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1790 {
1791         vm_offset_t end_va;
1792         vm_offset_t va;
1793
1794         end_va = beg_va + count * PAGE_SIZE;
1795
1796         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1797                 pt_entry_t *pte;
1798
1799                 pte = vtopte(va);
1800                 (void)pte_load_clear(pte);
1801         }
1802 }
1803
1804 /*
1805  * Create a new thread and optionally associate it with a (new) process.
1806  * NOTE! the new thread's cpu may not equal the current cpu.
1807  */
1808 void
1809 pmap_init_thread(thread_t td)
1810 {
1811         /* enforce pcb placement & alignment */
1812         td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1813         td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1814         td->td_savefpu = &td->td_pcb->pcb_save;
1815         td->td_sp = (char *)td->td_pcb; /* no -16 */
1816 }
1817
1818 /*
1819  * This routine directly affects the fork perf for a process.
1820  */
1821 void
1822 pmap_init_proc(struct proc *p)
1823 {
1824 }
1825
1826 static void
1827 pmap_pinit_defaults(struct pmap *pmap)
1828 {
1829         bcopy(pmap_bits_default, pmap->pmap_bits,
1830               sizeof(pmap_bits_default));
1831         bcopy(protection_codes, pmap->protection_codes,
1832               sizeof(protection_codes));
1833         bcopy(pat_pte_index, pmap->pmap_cache_bits,
1834               sizeof(pat_pte_index));
1835         pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1836         pmap->copyinstr = std_copyinstr;
1837         pmap->copyin = std_copyin;
1838         pmap->copyout = std_copyout;
1839         pmap->fubyte = std_fubyte;
1840         pmap->subyte = std_subyte;
1841         pmap->fuword32 = std_fuword32;
1842         pmap->fuword64 = std_fuword64;
1843         pmap->suword32 = std_suword32;
1844         pmap->suword64 = std_suword64;
1845         pmap->swapu32 = std_swapu32;
1846         pmap->swapu64 = std_swapu64;
1847 }
1848 /*
1849  * Initialize pmap0/vmspace0.
1850  *
1851  * On architectures where the kernel pmap is not integrated into the user
1852  * process pmap, this pmap represents the process pmap, not the kernel pmap.
1853  * kernel_pmap should be used to directly access the kernel_pmap.
1854  */
1855 void
1856 pmap_pinit0(struct pmap *pmap)
1857 {
1858         int i;
1859
1860         pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1861         pmap->pm_count = 1;
1862         CPUMASK_ASSZERO(pmap->pm_active);
1863         pmap->pm_pvhint = NULL;
1864         RB_INIT(&pmap->pm_pvroot);
1865         spin_init(&pmap->pm_spin, "pmapinit0");
1866         for (i = 0; i < PM_PLACEMARKS; ++i)
1867                 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1868         bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1869         pmap_pinit_defaults(pmap);
1870 }
1871
1872 /*
1873  * Initialize a preallocated and zeroed pmap structure,
1874  * such as one in a vmspace structure.
1875  */
1876 static void
1877 pmap_pinit_simple(struct pmap *pmap)
1878 {
1879         int i;
1880
1881         /*
1882          * Misc initialization
1883          */
1884         pmap->pm_count = 1;
1885         CPUMASK_ASSZERO(pmap->pm_active);
1886         pmap->pm_pvhint = NULL;
1887         pmap->pm_flags = PMAP_FLAG_SIMPLE;
1888
1889         pmap_pinit_defaults(pmap);
1890
1891         /*
1892          * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1893          * for this).
1894          */
1895         if (pmap->pm_pmlpv == NULL) {
1896                 RB_INIT(&pmap->pm_pvroot);
1897                 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1898                 spin_init(&pmap->pm_spin, "pmapinitsimple");
1899                 for (i = 0; i < PM_PLACEMARKS; ++i)
1900                         pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1901         }
1902 }
1903
1904 void
1905 pmap_pinit(struct pmap *pmap)
1906 {
1907         pv_entry_t pv;
1908         int j;
1909
1910         if (pmap->pm_pmlpv) {
1911                 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1912                         pmap_puninit(pmap);
1913                 }
1914         }
1915
1916         pmap_pinit_simple(pmap);
1917         pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1918
1919         /*
1920          * No need to allocate page table space yet but we do need a valid
1921          * page directory table.
1922          */
1923         if (pmap->pm_pml4 == NULL) {
1924                 pmap->pm_pml4 =
1925                     (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1926                                                         PAGE_SIZE,
1927                                                         VM_SUBSYS_PML4);
1928         }
1929
1930         /*
1931          * Allocate the page directory page, which wires it even though
1932          * it isn't being entered into some higher level page table (it
1933          * being the highest level).  If one is already cached we don't
1934          * have to do anything.
1935          */
1936         if ((pv = pmap->pm_pmlpv) == NULL) {
1937                 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1938                 pmap->pm_pmlpv = pv;
1939                 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1940                             VM_PAGE_TO_PHYS(pv->pv_m));
1941                 pv_put(pv);
1942
1943                 /*
1944                  * Install DMAP and KMAP.
1945                  */
1946                 for (j = 0; j < NDMPML4E; ++j) {
1947                         pmap->pm_pml4[DMPML4I + j] =
1948                             (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1949                             pmap->pmap_bits[PG_RW_IDX] |
1950                             pmap->pmap_bits[PG_V_IDX] |
1951                             pmap->pmap_bits[PG_U_IDX];
1952                 }
1953                 pmap->pm_pml4[KPML4I] = KPDPphys |
1954                     pmap->pmap_bits[PG_RW_IDX] |
1955                     pmap->pmap_bits[PG_V_IDX] |
1956                     pmap->pmap_bits[PG_U_IDX];
1957
1958                 /*
1959                  * install self-referential address mapping entry
1960                  */
1961                 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1962                     pmap->pmap_bits[PG_V_IDX] |
1963                     pmap->pmap_bits[PG_RW_IDX] |
1964                     pmap->pmap_bits[PG_A_IDX] |
1965                     pmap->pmap_bits[PG_M_IDX];
1966         } else {
1967                 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1968                 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1969         }
1970         KKASSERT(pmap->pm_pml4[255] == 0);
1971         KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1972         KKASSERT(pv->pv_entry.rbe_left == NULL);
1973         KKASSERT(pv->pv_entry.rbe_right == NULL);
1974 }
1975
1976 /*
1977  * Clean up a pmap structure so it can be physically freed.  This routine
1978  * is called by the vmspace dtor function.  A great deal of pmap data is
1979  * left passively mapped to improve vmspace management so we have a bit
1980  * of cleanup work to do here.
1981  */
1982 void
1983 pmap_puninit(pmap_t pmap)
1984 {
1985         pv_entry_t pv;
1986         vm_page_t p;
1987
1988         KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1989         if ((pv = pmap->pm_pmlpv) != NULL) {
1990                 if (pv_hold_try(pv) == 0)
1991                         pv_lock(pv);
1992                 KKASSERT(pv == pmap->pm_pmlpv);
1993                 p = pmap_remove_pv_page(pv);
1994                 pv_free(pv, NULL);
1995                 pv = NULL;      /* safety */
1996                 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1997                 vm_page_busy_wait(p, FALSE, "pgpun");
1998                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1999                 vm_page_unwire(p, 0);
2000                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2001
2002                 /*
2003                  * XXX eventually clean out PML4 static entries and
2004                  * use vm_page_free_zero()
2005                  */
2006                 vm_page_free(p);
2007                 pmap->pm_pmlpv = NULL;
2008         }
2009         if (pmap->pm_pml4) {
2010                 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
2011                 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
2012                 pmap->pm_pml4 = NULL;
2013         }
2014         KKASSERT(pmap->pm_stats.resident_count == 0);
2015         KKASSERT(pmap->pm_stats.wired_count == 0);
2016 }
2017
2018 /*
2019  * This function is now unused (used to add the pmap to the pmap_list)
2020  */
2021 void
2022 pmap_pinit2(struct pmap *pmap)
2023 {
2024 }
2025
2026 /*
2027  * This routine is called when various levels in the page table need to
2028  * be populated.  This routine cannot fail.
2029  *
2030  * This function returns two locked pv_entry's, one representing the
2031  * requested pv and one representing the requested pv's parent pv.  If
2032  * an intermediate page table does not exist it will be created, mapped,
2033  * wired, and the parent page table will be given an additional hold
2034  * count representing the presence of the child pv_entry.
2035  */
2036 static
2037 pv_entry_t
2038 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
2039 {
2040         pt_entry_t *ptep;
2041         pv_entry_t pv;
2042         pv_entry_t pvp;
2043         pt_entry_t v;
2044         vm_pindex_t pt_pindex;
2045         vm_page_t m;
2046         int isnew;
2047         int ispt;
2048
2049         /*
2050          * If the pv already exists and we aren't being asked for the
2051          * parent page table page we can just return it.  A locked+held pv
2052          * is returned.  The pv will also have a second hold related to the
2053          * pmap association that we don't have to worry about.
2054          */
2055         ispt = 0;
2056         pv = pv_alloc(pmap, ptepindex, &isnew);
2057         if (isnew == 0 && pvpp == NULL)
2058                 return(pv);
2059
2060         /*
2061          * Special case terminal PVs.  These are not page table pages so
2062          * no vm_page is allocated (the caller supplied the vm_page).  If
2063          * pvpp is non-NULL we are being asked to also removed the pt_pv
2064          * for this pv.
2065          *
2066          * Note that pt_pv's are only returned for user VAs. We assert that
2067          * a pt_pv is not being requested for kernel VAs.  The kernel
2068          * pre-wires all higher-level page tables so don't overload managed
2069          * higher-level page tables on top of it!
2070          */
2071         if (ptepindex < pmap_pt_pindex(0)) {
2072                 if (ptepindex >= NUPTE_USER) {
2073                         /* kernel manages this manually for KVM */
2074                         KKASSERT(pvpp == NULL);
2075                 } else {
2076                         KKASSERT(pvpp != NULL);
2077                         pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
2078                         pvp = pmap_allocpte(pmap, pt_pindex, NULL);
2079                         if (isnew)
2080                                 vm_page_wire_quick(pvp->pv_m);
2081                         *pvpp = pvp;
2082                 }
2083                 return(pv);
2084         }
2085
2086         /*
2087          * The kernel never uses managed PT/PD/PDP pages.
2088          */
2089         KKASSERT(pmap != &kernel_pmap);
2090
2091         /*
2092          * Non-terminal PVs allocate a VM page to represent the page table,
2093          * so we have to resolve pvp and calculate ptepindex for the pvp
2094          * and then for the page table entry index in the pvp for
2095          * fall-through.
2096          */
2097         if (ptepindex < pmap_pd_pindex(0)) {
2098                 /*
2099                  * pv is PT, pvp is PD
2100                  */
2101                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
2102                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
2103                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2104
2105                 /*
2106                  * PT index in PD
2107                  */
2108                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
2109                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
2110                 ispt = 1;
2111         } else if (ptepindex < pmap_pdp_pindex(0)) {
2112                 /*
2113                  * pv is PD, pvp is PDP
2114                  *
2115                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2116                  *                   the PD.
2117                  */
2118                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
2119                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2120
2121                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
2122                         KKASSERT(pvpp == NULL);
2123                         pvp = NULL;
2124                 } else {
2125                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
2126                 }
2127
2128                 /*
2129                  * PD index in PDP
2130                  */
2131                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
2132                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
2133         } else if (ptepindex < pmap_pml4_pindex()) {
2134                 /*
2135                  * pv is PDP, pvp is the root pml4 table
2136                  */
2137                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2138
2139                 /*
2140                  * PDP index in PML4
2141                  */
2142                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2143                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2144         } else {
2145                 /*
2146                  * pv represents the top-level PML4, there is no parent.
2147                  */
2148                 pvp = NULL;
2149         }
2150
2151         if (isnew == 0)
2152                 goto notnew;
2153
2154         /*
2155          * (isnew) is TRUE, pv is not terminal.
2156          *
2157          * (1) Add a wire count to the parent page table (pvp).
2158          * (2) Allocate a VM page for the page table.
2159          * (3) Enter the VM page into the parent page table.
2160          *
2161          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2162          */
2163         if (pvp)
2164                 vm_page_wire_quick(pvp->pv_m);
2165
2166         for (;;) {
2167                 m = vm_page_alloc(NULL, pv->pv_pindex,
2168                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2169                                   VM_ALLOC_INTERRUPT);
2170                 if (m)
2171                         break;
2172                 vm_wait(0);
2173         }
2174         vm_page_wire(m);        /* wire for mapping in parent */
2175         vm_page_unmanage(m);    /* m must be spinunlocked */
2176         pmap_zero_page(VM_PAGE_TO_PHYS(m));
2177         m->valid = VM_PAGE_BITS_ALL;
2178
2179         vm_page_spin_lock(m);
2180         pmap_page_stats_adding(m);
2181         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2182         pv->pv_m = m;
2183         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2184         vm_page_spin_unlock(m);
2185
2186         /*
2187          * (isnew) is TRUE, pv is not terminal.
2188          *
2189          * Wire the page into pvp.  Bump the resident_count for the pmap.
2190          * There is no pvp for the top level, address the pm_pml4[] array
2191          * directly.
2192          *
2193          * If the caller wants the parent we return it, otherwise
2194          * we just put it away.
2195          *
2196          * No interlock is needed for pte 0 -> non-zero.
2197          *
2198          * In the situation where *ptep is valid we might have an unmanaged
2199          * page table page shared from another page table which we need to
2200          * unshare before installing our private page table page.
2201          */
2202         if (pvp) {
2203                 v = VM_PAGE_TO_PHYS(m) |
2204                     (pmap->pmap_bits[PG_U_IDX] |
2205                      pmap->pmap_bits[PG_RW_IDX] |
2206                      pmap->pmap_bits[PG_V_IDX] |
2207                      pmap->pmap_bits[PG_A_IDX] |
2208                      pmap->pmap_bits[PG_M_IDX]);
2209                 ptep = pv_pte_lookup(pvp, ptepindex);
2210                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2211                         pt_entry_t pte;
2212
2213                         if (ispt == 0) {
2214                                 panic("pmap_allocpte: unexpected pte %p/%d",
2215                                       pvp, (int)ptepindex);
2216                         }
2217                         pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, v);
2218                         if (vm_page_unwire_quick(
2219                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2220                                 panic("pmap_allocpte: shared pgtable "
2221                                       "pg bad wirecount");
2222                         }
2223                 } else {
2224                         pt_entry_t pte;
2225
2226                         pte = atomic_swap_long(ptep, v);
2227                         if (pte != 0) {
2228                                 kprintf("install pgtbl mixup 0x%016jx "
2229                                         "old/new 0x%016jx/0x%016jx\n",
2230                                         (intmax_t)ptepindex, pte, v);
2231                         }
2232                 }
2233         }
2234         vm_page_wakeup(m);
2235
2236         /*
2237          * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2238          */
2239 notnew:
2240         if (pvp) {
2241                 KKASSERT(pvp->pv_m != NULL);
2242                 ptep = pv_pte_lookup(pvp, ptepindex);
2243                 v = VM_PAGE_TO_PHYS(pv->pv_m) |
2244                     (pmap->pmap_bits[PG_U_IDX] |
2245                      pmap->pmap_bits[PG_RW_IDX] |
2246                      pmap->pmap_bits[PG_V_IDX] |
2247                      pmap->pmap_bits[PG_A_IDX] |
2248                      pmap->pmap_bits[PG_M_IDX]);
2249                 if (*ptep != v) {
2250                         kprintf("mismatched upper level pt %016jx/%016jx\n",
2251                                 *ptep, v);
2252                 }
2253         }
2254         if (pvpp)
2255                 *pvpp = pvp;
2256         else if (pvp)
2257                 pv_put(pvp);
2258         return (pv);
2259 }
2260
2261 /*
2262  * This version of pmap_allocpte() checks for possible segment optimizations
2263  * that would allow page-table sharing.  It can be called for terminal
2264  * page or page table page ptepindex's.
2265  *
2266  * The function is called with page table page ptepindex's for fictitious
2267  * and unmanaged terminal pages.  That is, we don't want to allocate a
2268  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
2269  * for this case.
2270  *
2271  * This function can return a pv and *pvpp associated with the passed in pmap
2272  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
2273  * an unmanaged page table page will be entered into the pass in pmap.
2274  */
2275 static
2276 pv_entry_t
2277 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2278                   vm_map_entry_t entry, vm_offset_t va)
2279 {
2280         vm_object_t object;
2281         pmap_t obpmap;
2282         pmap_t *obpmapp;
2283         vm_offset_t b;
2284         pv_entry_t pte_pv;      /* in original or shared pmap */
2285         pv_entry_t pt_pv;       /* in original or shared pmap */
2286         pv_entry_t proc_pd_pv;  /* in original pmap */
2287         pv_entry_t proc_pt_pv;  /* in original pmap */
2288         pv_entry_t xpv;         /* PT in shared pmap */
2289         pd_entry_t *pt;         /* PT entry in PD of original pmap */
2290         pd_entry_t opte;        /* contents of *pt */
2291         pd_entry_t npte;        /* contents of *pt */
2292         vm_page_t m;
2293
2294         /*
2295          * Basic tests, require a non-NULL vm_map_entry, require proper
2296          * alignment and type for the vm_map_entry, require that the
2297          * underlying object already be allocated.
2298          *
2299          * We allow almost any type of object to use this optimization.
2300          * The object itself does NOT have to be sized to a multiple of the
2301          * segment size, but the memory mapping does.
2302          *
2303          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2304          *     won't work as expected.
2305          */
2306         if (entry == NULL ||
2307             pmap_mmu_optimize == 0 ||                   /* not enabled */
2308             (pmap->pm_flags & PMAP_HVM) ||              /* special pmap */
2309             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal or pt */
2310             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2311             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2312             entry->object.vm_object == NULL ||          /* needs VM object */
2313             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2314             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2315             (entry->offset & SEG_MASK) ||               /* must be aligned */
2316             (entry->start & SEG_MASK)) {
2317                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2318         }
2319
2320         /*
2321          * Make sure the full segment can be represented.
2322          */
2323         b = va & ~(vm_offset_t)SEG_MASK;
2324         if (b < entry->start || b + SEG_SIZE > entry->end)
2325                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2326
2327         /*
2328          * If the full segment can be represented dive the VM object's
2329          * shared pmap, allocating as required.
2330          */
2331         object = entry->object.vm_object;
2332
2333         if (entry->protection & VM_PROT_WRITE)
2334                 obpmapp = &object->md.pmap_rw;
2335         else
2336                 obpmapp = &object->md.pmap_ro;
2337
2338 #ifdef PMAP_DEBUG2
2339         if (pmap_enter_debug > 0) {
2340                 --pmap_enter_debug;
2341                 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2342                         "obpmapp %p %p\n",
2343                         va, entry->protection, object,
2344                         obpmapp, *obpmapp);
2345                 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2346                         entry, entry->start, entry->end);
2347         }
2348 #endif
2349
2350         /*
2351          * We allocate what appears to be a normal pmap but because portions
2352          * of this pmap are shared with other unrelated pmaps we have to
2353          * set pm_active to point to all cpus.
2354          *
2355          * XXX Currently using pmap_spin to interlock the update, can't use
2356          *     vm_object_hold/drop because the token might already be held
2357          *     shared OR exclusive and we don't know.
2358          */
2359         while ((obpmap = *obpmapp) == NULL) {
2360                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2361                 pmap_pinit_simple(obpmap);
2362                 pmap_pinit2(obpmap);
2363                 spin_lock(&pmap_spin);
2364                 if (*obpmapp != NULL) {
2365                         /*
2366                          * Handle race
2367                          */
2368                         spin_unlock(&pmap_spin);
2369                         pmap_release(obpmap);
2370                         pmap_puninit(obpmap);
2371                         kfree(obpmap, M_OBJPMAP);
2372                         obpmap = *obpmapp; /* safety */
2373                 } else {
2374                         obpmap->pm_active = smp_active_mask;
2375                         obpmap->pm_flags |= PMAP_SEGSHARED;
2376                         *obpmapp = obpmap;
2377                         spin_unlock(&pmap_spin);
2378                 }
2379         }
2380
2381         /*
2382          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2383          * pte/pt using the shared pmap from the object but also adjust
2384          * the process pmap's page table page as a side effect.
2385          */
2386
2387         /*
2388          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2389          * we will return.  This is true if ptepindex represents a terminal
2390          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2391          * the PD.
2392          */
2393         pt_pv = NULL;
2394         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2395 retry:
2396         if (ptepindex >= pmap_pt_pindex(0))
2397                 xpv = pte_pv;
2398         else
2399                 xpv = pt_pv;
2400
2401         /*
2402          * Resolve the PD in the process pmap so we can properly share the
2403          * page table page.  Lock order is bottom-up (leaf first)!
2404          *
2405          * NOTE: proc_pt_pv can be NULL.
2406          */
2407         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b), NULL);
2408         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2409 #ifdef PMAP_DEBUG2
2410         if (pmap_enter_debug > 0) {
2411                 --pmap_enter_debug;
2412                 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2413                         proc_pt_pv,
2414                         (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2415                         proc_pd_pv,
2416                         va);
2417         }
2418 #endif
2419
2420         /*
2421          * xpv is the page table page pv from the shared object
2422          * (for convenience), from above.
2423          *
2424          * Calculate the pte value for the PT to load into the process PD.
2425          * If we have to change it we must properly dispose of the previous
2426          * entry.
2427          */
2428         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2429         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2430                (pmap->pmap_bits[PG_U_IDX] |
2431                 pmap->pmap_bits[PG_RW_IDX] |
2432                 pmap->pmap_bits[PG_V_IDX] |
2433                 pmap->pmap_bits[PG_A_IDX] |
2434                 pmap->pmap_bits[PG_M_IDX]);
2435
2436         /*
2437          * Dispose of previous page table page if it was local to the
2438          * process pmap.  If the old pt is not empty we cannot dispose of it
2439          * until we clean it out.  This case should not arise very often so
2440          * it is not optimized.
2441          *
2442          * Leave pt_pv and pte_pv (in our object pmap) locked and intact
2443          * for the retry.
2444          */
2445         if (proc_pt_pv) {
2446                 pmap_inval_bulk_t bulk;
2447
2448                 if (proc_pt_pv->pv_m->wire_count != 1) {
2449                         pv_put(proc_pd_pv);
2450                         pv_put(proc_pt_pv);
2451                         pmap_remove(pmap,
2452                                     va & ~(vm_offset_t)SEG_MASK,
2453                                     (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2454                         goto retry;
2455                 }
2456
2457                 /*
2458                  * The release call will indirectly clean out *pt
2459                  */
2460                 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2461                 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2462                 pmap_inval_bulk_flush(&bulk);
2463                 proc_pt_pv = NULL;
2464                 /* relookup */
2465                 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2466         }
2467
2468         /*
2469          * Handle remaining cases.
2470          */
2471         if (*pt == 0) {
2472                 atomic_swap_long(pt, npte);
2473                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2474                 vm_page_wire_quick(proc_pd_pv->pv_m);   /* proc pd for sh pt */
2475                 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2476         } else if (*pt != npte) {
2477                 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2478
2479 #if 0
2480                 opte = pte_load_clear(pt);
2481                 KKASSERT(opte && opte != npte);
2482
2483                 *pt = npte;
2484 #endif
2485                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2486
2487                 /*
2488                  * Clean up opte, bump the wire_count for the process
2489                  * PD page representing the new entry if it was
2490                  * previously empty.
2491                  *
2492                  * If the entry was not previously empty and we have
2493                  * a PT in the proc pmap then opte must match that
2494                  * pt.  The proc pt must be retired (this is done
2495                  * later on in this procedure).
2496                  *
2497                  * NOTE: replacing valid pte, wire_count on proc_pd_pv
2498                  * stays the same.
2499                  */
2500                 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2501                 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2502                 if (vm_page_unwire_quick(m)) {
2503                         panic("pmap_allocpte_seg: "
2504                               "bad wire count %p",
2505                               m);
2506                 }
2507         }
2508
2509         /*
2510          * The existing process page table was replaced and must be destroyed
2511          * here.
2512          */
2513         if (proc_pd_pv)
2514                 pv_put(proc_pd_pv);
2515         if (pvpp)
2516                 *pvpp = pt_pv;
2517         else
2518                 pv_put(pt_pv);
2519
2520         return (pte_pv);
2521 }
2522
2523 /*
2524  * Release any resources held by the given physical map.
2525  *
2526  * Called when a pmap initialized by pmap_pinit is being released.  Should
2527  * only be called if the map contains no valid mappings.
2528  */
2529 struct pmap_release_info {
2530         pmap_t  pmap;
2531         int     retry;
2532         pv_entry_t pvp;
2533 };
2534
2535 static int pmap_release_callback(pv_entry_t pv, void *data);
2536
2537 void
2538 pmap_release(struct pmap *pmap)
2539 {
2540         struct pmap_release_info info;
2541
2542         KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2543                 ("pmap still active! %016jx",
2544                 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2545
2546         /*
2547          * There is no longer a pmap_list, if there were we would remove the
2548          * pmap from it here.
2549          */
2550
2551         /*
2552          * Pull pv's off the RB tree in order from low to high and release
2553          * each page.
2554          */
2555         info.pmap = pmap;
2556         do {
2557                 info.retry = 0;
2558                 info.pvp = NULL;
2559
2560                 spin_lock(&pmap->pm_spin);
2561                 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2562                         pmap_release_callback, &info);
2563                 spin_unlock(&pmap->pm_spin);
2564
2565                 if (info.pvp)
2566                         pv_put(info.pvp);
2567         } while (info.retry);
2568
2569
2570         /*
2571          * One resident page (the pml4 page) should remain.
2572          * No wired pages should remain.
2573          */
2574 #if 1
2575         if (pmap->pm_stats.resident_count !=
2576             ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1) ||
2577             pmap->pm_stats.wired_count != 0) {
2578                 kprintf("fatal pmap problem - pmap %p flags %08x "
2579                         "rescnt=%jd wirecnt=%jd\n",
2580                         pmap,
2581                         pmap->pm_flags,
2582                         pmap->pm_stats.resident_count,
2583                         pmap->pm_stats.wired_count);
2584                 tsleep(pmap, 0, "DEAD", 0);
2585         }
2586 #else
2587         KKASSERT(pmap->pm_stats.resident_count ==
2588                  ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2589         KKASSERT(pmap->pm_stats.wired_count == 0);
2590 #endif
2591 }
2592
2593 /*
2594  * Called from low to high.  We must cache the proper parent pv so we
2595  * can adjust its wired count.
2596  */
2597 static int
2598 pmap_release_callback(pv_entry_t pv, void *data)
2599 {
2600         struct pmap_release_info *info = data;
2601         pmap_t pmap = info->pmap;
2602         vm_pindex_t pindex;
2603         int r;
2604
2605         /*
2606          * Acquire a held and locked pv, check for release race
2607          */
2608         pindex = pv->pv_pindex;
2609         if (info->pvp == pv) {
2610                 spin_unlock(&pmap->pm_spin);
2611                 info->pvp = NULL;
2612         } else if (pv_hold_try(pv)) {
2613                 spin_unlock(&pmap->pm_spin);
2614         } else {
2615                 spin_unlock(&pmap->pm_spin);
2616                 pv_lock(pv);
2617                 pv_put(pv);
2618                 info->retry = 1;
2619                 spin_lock(&pmap->pm_spin);
2620
2621                 return -1;
2622         }
2623         KKASSERT(pv->pv_pmap == pmap && pindex == pv->pv_pindex);
2624
2625         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2626                 /*
2627                  * I am PTE, parent is PT
2628                  */
2629                 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2630                 pindex += NUPTE_TOTAL;
2631         } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2632                 /*
2633                  * I am PT, parent is PD
2634                  */
2635                 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2636                 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2637         } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2638                 /*
2639                  * I am PD, parent is PDP
2640                  */
2641                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2642                          NPDPEPGSHIFT;
2643                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2644         } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2645                 /*
2646                  * I am PDP, parent is PML4 (there's only one)
2647                  */
2648 #if 0
2649                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2650                            NUPD_TOTAL) >> NPML4EPGSHIFT;
2651                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2652 #endif
2653                 pindex = pmap_pml4_pindex();
2654         } else {
2655                 /*
2656                  * parent is NULL
2657                  */
2658                 if (info->pvp) {
2659                         pv_put(info->pvp);
2660                         info->pvp = NULL;
2661                 }
2662                 pindex = 0;
2663         }
2664         if (pindex) {
2665                 if (info->pvp && info->pvp->pv_pindex != pindex) {
2666                         pv_put(info->pvp);
2667                         info->pvp = NULL;
2668                 }
2669                 if (info->pvp == NULL)
2670                         info->pvp = pv_get(pmap, pindex, NULL);
2671         } else {
2672                 if (info->pvp) {
2673                         pv_put(info->pvp);
2674                         info->pvp = NULL;
2675                 }
2676         }
2677         r = pmap_release_pv(pv, info->pvp, NULL);
2678         spin_lock(&pmap->pm_spin);
2679
2680         return(r);
2681 }
2682
2683 /*
2684  * Called with held (i.e. also locked) pv.  This function will dispose of
2685  * the lock along with the pv.
2686  *
2687  * If the caller already holds the locked parent page table for pv it
2688  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2689  * pass NULL for pvp.
2690  */
2691 static int
2692 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2693 {
2694         vm_page_t p;
2695
2696         /*
2697          * The pmap is currently not spinlocked, pv is held+locked.
2698          * Remove the pv's page from its parent's page table.  The
2699          * parent's page table page's wire_count will be decremented.
2700          *
2701          * This will clean out the pte at any level of the page table.
2702          * If smp != 0 all cpus are affected.
2703          *
2704          * Do not tear-down recursively, its faster to just let the
2705          * release run its course.
2706          */
2707         pmap_remove_pv_pte(pv, pvp, bulk, 0);
2708
2709         /*
2710          * Terminal pvs are unhooked from their vm_pages.  Because
2711          * terminal pages aren't page table pages they aren't wired
2712          * by us, so we have to be sure not to unwire them either.
2713          */
2714         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2715                 pmap_remove_pv_page(pv);
2716                 goto skip;
2717         }
2718
2719         /*
2720          * We leave the top-level page table page cached, wired, and
2721          * mapped in the pmap until the dtor function (pmap_puninit())
2722          * gets called.
2723          *
2724          * Since we are leaving the top-level pv intact we need
2725          * to break out of what would otherwise be an infinite loop.
2726          */
2727         if (pv->pv_pindex == pmap_pml4_pindex()) {
2728                 pv_put(pv);
2729                 return(-1);
2730         }
2731
2732         /*
2733          * For page table pages (other than the top-level page),
2734          * remove and free the vm_page.  The representitive mapping
2735          * removed above by pmap_remove_pv_pte() did not undo the
2736          * last wire_count so we have to do that as well.
2737          */
2738         p = pmap_remove_pv_page(pv);
2739         vm_page_busy_wait(p, FALSE, "pmaprl");
2740         if (p->wire_count != 1) {
2741                 kprintf("p->wire_count was %016lx %d\n",
2742                         pv->pv_pindex, p->wire_count);
2743         }
2744         KKASSERT(p->wire_count == 1);
2745         KKASSERT(p->flags & PG_UNMANAGED);
2746
2747         vm_page_unwire(p, 0);
2748         KKASSERT(p->wire_count == 0);
2749
2750         vm_page_free(p);
2751 skip:
2752         pv_free(pv, pvp);
2753
2754         return 0;
2755 }
2756
2757 /*
2758  * This function will remove the pte associated with a pv from its parent.
2759  * Terminal pv's are supported.  All cpus specified by (bulk) are properly
2760  * invalidated.
2761  *
2762  * The wire count will be dropped on the parent page table.  The wire
2763  * count on the page being removed (pv->pv_m) from the parent page table
2764  * is NOT touched.  Note that terminal pages will not have any additional
2765  * wire counts while page table pages will have at least one representing
2766  * the mapping, plus others representing sub-mappings.
2767  *
2768  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2769  *       pages and user page table and terminal pages.
2770  *
2771  * NOTE: The pte being removed might be unmanaged, and the pv supplied might
2772  *       be freshly allocated and not imply that the pte is managed.  In this
2773  *       case pv->pv_m should be NULL.
2774  *
2775  * The pv must be locked.  The pvp, if supplied, must be locked.  All
2776  * supplied pv's will remain locked on return.
2777  *
2778  * XXX must lock parent pv's if they exist to remove pte XXX
2779  */
2780 static
2781 void
2782 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2783                    int destroy)
2784 {
2785         vm_pindex_t ptepindex = pv->pv_pindex;
2786         pmap_t pmap = pv->pv_pmap;
2787         vm_page_t p;
2788         int gotpvp = 0;
2789
2790         KKASSERT(pmap);
2791
2792         if (ptepindex == pmap_pml4_pindex()) {
2793                 /*
2794                  * We are the top level PML4E table, there is no parent.
2795                  */
2796                 p = pmap->pm_pmlpv->pv_m;
2797                 KKASSERT(pv->pv_m == p);        /* debugging */
2798         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2799                 /*
2800                  * Remove a PDP page from the PML4E.  This can only occur
2801                  * with user page tables.  We do not have to lock the
2802                  * pml4 PV so just ignore pvp.
2803                  */
2804                 vm_pindex_t pml4_pindex;
2805                 vm_pindex_t pdp_index;
2806                 pml4_entry_t *pdp;
2807
2808                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2809                 if (pvp == NULL) {
2810                         pml4_pindex = pmap_pml4_pindex();
2811                         pvp = pv_get(pv->pv_pmap, pml4_pindex, NULL);
2812                         KKASSERT(pvp);
2813                         gotpvp = 1;
2814                 }
2815
2816                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2817                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2818                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2819                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2820                 KKASSERT(pv->pv_m == p);        /* debugging */
2821         } else if (ptepindex >= pmap_pd_pindex(0)) {
2822                 /*
2823                  * Remove a PD page from the PDP
2824                  *
2825                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2826                  *                   of a simple pmap because it stops at
2827                  *                   the PD page.
2828                  */
2829                 vm_pindex_t pdp_pindex;
2830                 vm_pindex_t pd_index;
2831                 pdp_entry_t *pd;
2832
2833                 pd_index = ptepindex - pmap_pd_pindex(0);
2834
2835                 if (pvp == NULL) {
2836                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2837                                      (pd_index >> NPML4EPGSHIFT);
2838                         pvp = pv_get(pv->pv_pmap, pdp_pindex, NULL);
2839                         gotpvp = 1;
2840                 }
2841
2842                 if (pvp) {
2843                         pd = pv_pte_lookup(pvp, pd_index &
2844                                                 ((1ul << NPDPEPGSHIFT) - 1));
2845                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2846                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2847                         pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2848                 } else {
2849                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2850                         p = pv->pv_m;           /* degenerate test later */
2851                 }
2852                 KKASSERT(pv->pv_m == p);        /* debugging */
2853         } else if (ptepindex >= pmap_pt_pindex(0)) {
2854                 /*
2855                  *  Remove a PT page from the PD
2856                  */
2857                 vm_pindex_t pd_pindex;
2858                 vm_pindex_t pt_index;
2859                 pd_entry_t *pt;
2860
2861                 pt_index = ptepindex - pmap_pt_pindex(0);
2862
2863                 if (pvp == NULL) {
2864                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2865                                     (pt_index >> NPDPEPGSHIFT);
2866                         pvp = pv_get(pv->pv_pmap, pd_pindex, NULL);
2867                         KKASSERT(pvp);
2868                         gotpvp = 1;
2869                 }
2870
2871                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2872 #if 0
2873                 KASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0,
2874                         ("*pt unexpectedly invalid %016jx "
2875                          "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
2876                         *pt, gotpvp, ptepindex, pt_index, pv, pvp));
2877                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2878 #else
2879                 if ((*pt & pmap->pmap_bits[PG_V_IDX]) == 0) {
2880                         kprintf("*pt unexpectedly invalid %016jx "
2881                                 "gotpvp=%d ptepindex=%ld ptindex=%ld "
2882                                 "pv=%p pvp=%p\n",
2883                                 *pt, gotpvp, ptepindex, pt_index, pv, pvp);
2884                         tsleep(pt, 0, "DEAD", 0);
2885                         p = pv->pv_m;
2886                 } else {
2887                         p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2888                 }
2889 #endif
2890                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2891                 KKASSERT(pv->pv_m == p);        /* debugging */
2892         } else {
2893                 /*
2894                  * Remove a PTE from the PT page.  The PV might exist even if
2895                  * the PTE is not managed, in whichcase pv->pv_m should be
2896                  * NULL.
2897                  *
2898                  * NOTE: Userland pmaps manage the parent PT/PD/PDP page
2899                  *       table pages but the kernel_pmap does not.
2900                  *
2901                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2902                  *       pv is a pte_pv so we can safely lock pt_pv.
2903                  *
2904                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2905                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2906                  *       terminal ptes.
2907                  */
2908                 vm_pindex_t pt_pindex;
2909                 pt_entry_t *ptep;
2910                 pt_entry_t pte;
2911                 vm_offset_t va;
2912
2913                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2914                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2915
2916                 if (ptepindex >= NUPTE_USER) {
2917                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2918                         KKASSERT(pvp == NULL);
2919                         /* pvp remains NULL */
2920                 } else {
2921                         if (pvp == NULL) {
2922                                 pt_pindex = NUPTE_TOTAL +
2923                                             (ptepindex >> NPDPEPGSHIFT);
2924                                 pvp = pv_get(pv->pv_pmap, pt_pindex, NULL);
2925                                 KKASSERT(pvp);
2926                                 gotpvp = 1;
2927                         }
2928                         ptep = pv_pte_lookup(pvp, ptepindex &
2929                                                   ((1ul << NPDPEPGSHIFT) - 1));
2930                 }
2931                 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2932                 if (bulk == NULL)               /* XXX */
2933                         cpu_invlpg((void *)va); /* XXX */
2934
2935                 /*
2936                  * Now update the vm_page_t
2937                  */
2938                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
2939                     (pte & pmap->pmap_bits[PG_V_IDX])) {
2940                         /*
2941                          * Valid managed page, adjust (p).
2942                          */
2943                         if (pte & pmap->pmap_bits[PG_DEVICE_IDX]) {
2944                                 p = pv->pv_m;
2945                         } else {
2946                                 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2947                                 KKASSERT(pv->pv_m == p);
2948                         }
2949                         if (pte & pmap->pmap_bits[PG_M_IDX]) {
2950                                 if (pmap_track_modified(ptepindex))
2951                                         vm_page_dirty(p);
2952                         }
2953                         if (pte & pmap->pmap_bits[PG_A_IDX]) {
2954                                 vm_page_flag_set(p, PG_REFERENCED);
2955                         }
2956                 } else {
2957                         /*
2958                          * Unmanaged page, do not try to adjust the vm_page_t.
2959                          * pv could be freshly allocated for a pmap_enter(),
2960                          * replacing an unmanaged page with a managed one.
2961                          *
2962                          * pv->pv_m might reflect the new page and not the
2963                          * existing page.
2964                          *
2965                          * We could extract p from the physical address and
2966                          * adjust it but we explicitly do not for unmanaged
2967                          * pages.
2968                          */
2969                         p = NULL;
2970                 }
2971                 if (pte & pmap->pmap_bits[PG_W_IDX])
2972                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
2973                 if (pte & pmap->pmap_bits[PG_G_IDX])
2974                         cpu_invlpg((void *)va);
2975         }
2976
2977         /*
2978          * If requested, scrap the underlying pv->pv_m and the underlying
2979          * pv.  If this is a page-table-page we must also free the page.
2980          *
2981          * pvp must be returned locked.
2982          */
2983         if (destroy == 1) {
2984                 /*
2985                  * page table page (PT, PD, PDP, PML4), caller was responsible
2986                  * for testing wired_count.
2987                  */
2988                 KKASSERT(pv->pv_m->wire_count == 1);
2989                 p = pmap_remove_pv_page(pv);
2990                 pv_free(pv, pvp);
2991                 pv = NULL;
2992
2993                 vm_page_busy_wait(p, FALSE, "pgpun");
2994                 vm_page_unwire(p, 0);
2995                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2996                 vm_page_free(p);
2997         } else if (destroy == 2) {
2998                 /*
2999                  * Normal page, remove from pmap and leave the underlying
3000                  * page untouched.
3001                  */
3002                 pmap_remove_pv_page(pv);
3003                 pv_free(pv, pvp);
3004                 pv = NULL;              /* safety */
3005         }
3006
3007         /*
3008          * If we acquired pvp ourselves then we are responsible for
3009          * recursively deleting it.
3010          */
3011         if (pvp && gotpvp) {
3012                 /*
3013                  * Recursively destroy higher-level page tables.
3014                  *
3015                  * This is optional.  If we do not, they will still
3016                  * be destroyed when the process exits.
3017                  *
3018                  * NOTE: Do not destroy pv_entry's with extra hold refs,
3019                  *       a caller may have unlocked it and intends to
3020                  *       continue to use it.
3021                  */
3022                 if (pmap_dynamic_delete &&
3023                     pvp->pv_m &&
3024                     pvp->pv_m->wire_count == 1 &&
3025                     (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
3026                     pvp->pv_pindex != pmap_pml4_pindex()) {
3027                         if (pmap_dynamic_delete == 2)
3028                                 kprintf("A %jd %08x\n", pvp->pv_pindex, pvp->pv_hold);
3029                         if (pmap != &kernel_pmap) {
3030                                 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
3031                                 pvp = NULL;     /* safety */
3032                         } else {
3033                                 kprintf("Attempt to remove kernel_pmap pindex "
3034                                         "%jd\n", pvp->pv_pindex);
3035                                 pv_put(pvp);
3036                         }
3037                 } else {
3038                         pv_put(pvp);
3039                 }
3040         }
3041 }
3042
3043 /*
3044  * Remove the vm_page association to a pv.  The pv must be locked.
3045  */
3046 static
3047 vm_page_t
3048 pmap_remove_pv_page(pv_entry_t pv)
3049 {
3050         vm_page_t m;
3051
3052         m = pv->pv_m;
3053         vm_page_spin_lock(m);
3054         KKASSERT(m && m == pv->pv_m);
3055         pv->pv_m = NULL;
3056         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3057         pmap_page_stats_deleting(m);
3058         if (TAILQ_EMPTY(&m->md.pv_list))
3059                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3060         vm_page_spin_unlock(m);
3061
3062         return(m);
3063 }
3064
3065 /*
3066  * Grow the number of kernel page table entries, if needed.
3067  *
3068  * This routine is always called to validate any address space
3069  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
3070  * space below KERNBASE.
3071  *
3072  * kernel_map must be locked exclusively by the caller.
3073  */
3074 void
3075 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
3076 {
3077         vm_paddr_t paddr;
3078         vm_offset_t ptppaddr;
3079         vm_page_t nkpg;
3080         pd_entry_t *pt, newpt;
3081         pdp_entry_t newpd;
3082         int update_kernel_vm_end;
3083
3084         /*
3085          * bootstrap kernel_vm_end on first real VM use
3086          */
3087         if (kernel_vm_end == 0) {
3088                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
3089                 nkpt = 0;
3090                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3091                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
3092                                         ~(PAGE_SIZE * NPTEPG - 1);
3093                         nkpt++;
3094                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
3095                                 kernel_vm_end = kernel_map.max_offset;
3096                                 break;
3097                         }
3098                 }
3099         }
3100
3101         /*
3102          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
3103          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
3104          * do not want to force-fill 128G worth of page tables.
3105          */
3106         if (kstart < KERNBASE) {
3107                 if (kstart > kernel_vm_end)
3108                         kstart = kernel_vm_end;
3109                 KKASSERT(kend <= KERNBASE);
3110                 update_kernel_vm_end = 1;
3111         } else {
3112                 update_kernel_vm_end = 0;
3113         }
3114
3115         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
3116         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
3117
3118         if (kend - 1 >= kernel_map.max_offset)
3119                 kend = kernel_map.max_offset;
3120
3121         while (kstart < kend) {
3122                 pt = pmap_pt(&kernel_pmap, kstart);
3123                 if (pt == NULL) {
3124                         /* We need a new PD entry */
3125                         nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3126                                              VM_ALLOC_NORMAL |
3127                                              VM_ALLOC_SYSTEM |
3128                                              VM_ALLOC_INTERRUPT);
3129                         if (nkpg == NULL) {
3130                                 panic("pmap_growkernel: no memory to grow "
3131                                       "kernel");
3132                         }
3133                         paddr = VM_PAGE_TO_PHYS(nkpg);
3134                         pmap_zero_page(paddr);
3135                         newpd = (pdp_entry_t)
3136                             (paddr |
3137                             kernel_pmap.pmap_bits[PG_V_IDX] |
3138                             kernel_pmap.pmap_bits[PG_RW_IDX] |
3139                             kernel_pmap.pmap_bits[PG_A_IDX] |
3140                             kernel_pmap.pmap_bits[PG_M_IDX]);
3141                         *pmap_pd(&kernel_pmap, kstart) = newpd;
3142                         continue; /* try again */
3143                 }
3144                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3145                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
3146                                  ~(PAGE_SIZE * NPTEPG - 1);
3147                         if (kstart - 1 >= kernel_map.max_offset) {
3148                                 kstart = kernel_map.max_offset;
3149                                 break;
3150                         }
3151                         continue;
3152                 }
3153
3154                 /*
3155                  * We need a new PT
3156                  *
3157                  * This index is bogus, but out of the way
3158                  */
3159                 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3160                                      VM_ALLOC_NORMAL |
3161                                      VM_ALLOC_SYSTEM |
3162                                      VM_ALLOC_INTERRUPT);
3163                 if (nkpg == NULL)
3164                         panic("pmap_growkernel: no memory to grow kernel");
3165
3166                 vm_page_wire(nkpg);
3167                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
3168                 pmap_zero_page(ptppaddr);
3169                 newpt = (pd_entry_t)(ptppaddr |
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                 atomic_swap_long(pmap_pt(&kernel_pmap, kstart), newpt);
3175
3176                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3177                           ~(PAGE_SIZE * NPTEPG - 1);
3178
3179                 if (kstart - 1 >= kernel_map.max_offset) {
3180                         kstart = kernel_map.max_offset;
3181                         break;
3182                 }
3183         }
3184
3185         /*
3186          * Only update kernel_vm_end for areas below KERNBASE.
3187          */
3188         if (update_kernel_vm_end && kernel_vm_end < kstart)
3189                 kernel_vm_end = kstart;
3190 }
3191
3192 /*
3193  *      Add a reference to the specified pmap.
3194  */
3195 void
3196 pmap_reference(pmap_t pmap)
3197 {
3198         if (pmap != NULL)
3199                 atomic_add_int(&pmap->pm_count, 1);
3200 }
3201
3202 /***************************************************
3203  * page management routines.
3204  ***************************************************/
3205
3206 /*
3207  * Hold a pv without locking it
3208  */
3209 static void
3210 pv_hold(pv_entry_t pv)
3211 {
3212         atomic_add_int(&pv->pv_hold, 1);
3213 }
3214
3215 /*
3216  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
3217  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
3218  * the pv properly.
3219  *
3220  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3221  * pv list via its page) must be held by the caller in order to stabilize
3222  * the pv.
3223  */
3224 static int
3225 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3226 {
3227         u_int count;
3228
3229         /*
3230          * Critical path shortcut expects pv to already have one ref
3231          * (for the pv->pv_pmap).
3232          */
3233         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3234 #ifdef PMAP_DEBUG
3235                 pv->pv_func = func;
3236                 pv->pv_line = lineno;
3237 #endif
3238                 return TRUE;
3239         }
3240
3241         for (;;) {
3242                 count = pv->pv_hold;
3243                 cpu_ccfence();
3244                 if ((count & PV_HOLD_LOCKED) == 0) {
3245                         if (atomic_cmpset_int(&pv->pv_hold, count,
3246                                               (count + 1) | PV_HOLD_LOCKED)) {
3247 #ifdef PMAP_DEBUG
3248                                 pv->pv_func = func;
3249                                 pv->pv_line = lineno;
3250 #endif
3251                                 return TRUE;
3252                         }
3253                 } else {
3254                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3255                                 return FALSE;
3256                 }
3257                 /* retry */
3258         }
3259 }
3260
3261 /*
3262  * Drop a previously held pv_entry which could not be locked, allowing its
3263  * destruction.
3264  *
3265  * Must not be called with a spinlock held as we might zfree() the pv if it
3266  * is no longer associated with a pmap and this was the last hold count.
3267  */
3268 static void
3269 pv_drop(pv_entry_t pv)
3270 {
3271         u_int count;
3272
3273         for (;;) {
3274                 count = pv->pv_hold;
3275                 cpu_ccfence();
3276                 KKASSERT((count & PV_HOLD_MASK) > 0);
3277                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3278                          (PV_HOLD_LOCKED | 1));
3279                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3280                         if ((count & PV_HOLD_MASK) == 1) {
3281 #ifdef PMAP_DEBUG2
3282                                 if (pmap_enter_debug > 0) {
3283                                         --pmap_enter_debug;
3284                                         kprintf("pv_drop: free pv %p\n", pv);
3285                                 }
3286 #endif
3287                                 KKASSERT(count == 1);
3288                                 KKASSERT(pv->pv_pmap == NULL);
3289                                 zfree(pvzone, pv);
3290                         }
3291                         return;
3292                 }
3293                 /* retry */
3294         }
3295 }
3296
3297 /*
3298  * Find or allocate the requested PV entry, returning a locked, held pv.
3299  *
3300  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3301  * for the caller and one representing the pmap and vm_page association.
3302  *
3303  * If (*isnew) is zero, the returned pv will have only one hold count.
3304  *
3305  * Since both associations can only be adjusted while the pv is locked,
3306  * together they represent just one additional hold.
3307  */
3308 static
3309 pv_entry_t
3310 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3311 {
3312         struct mdglobaldata *md = mdcpu;
3313         pv_entry_t pv;
3314         pv_entry_t pnew;
3315         int pmap_excl = 0;
3316
3317         pnew = NULL;
3318         if (md->gd_newpv) {
3319 #if 0
3320                 pnew = atomic_swap_ptr((void *)&md->gd_newpv, NULL);
3321 #else
3322                 crit_enter();
3323                 pnew = md->gd_newpv;    /* might race NULL */
3324                 md->gd_newpv = NULL;
3325                 crit_exit();
3326 #endif
3327         }
3328         if (pnew == NULL)
3329                 pnew = zalloc(pvzone);
3330
3331         spin_lock_shared(&pmap->pm_spin);
3332         for (;;) {
3333                 /*
3334                  * Shortcut cache
3335                  */
3336                 pv = pmap->pm_pvhint;
3337                 cpu_ccfence();
3338                 if (pv == NULL ||
3339                     pv->pv_pmap != pmap ||
3340                     pv->pv_pindex != pindex) {
3341                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3342                                                         pindex);
3343                 }
3344                 if (pv == NULL) {
3345                         vm_pindex_t *pmark;
3346
3347                         /*
3348                          * Requires exclusive pmap spinlock
3349                          */
3350                         if (pmap_excl == 0) {
3351                                 pmap_excl = 1;
3352                                 if (!spin_lock_upgrade_try(&pmap->pm_spin)) {
3353                                         spin_unlock_shared(&pmap->pm_spin);
3354                                         spin_lock(&pmap->pm_spin);
3355                                         continue;
3356                                 }
3357                         }
3358
3359                         /*
3360                          * We need to block if someone is holding our
3361                          * placemarker.  As long as we determine the
3362                          * placemarker has not been aquired we do not
3363                          * need to get it as acquision also requires
3364                          * the pmap spin lock.
3365                          *
3366                          * However, we can race the wakeup.
3367                          */
3368                         pmark = pmap_placemarker_hash(pmap, pindex);
3369
3370                         if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3371                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3372                                 tsleep_interlock(pmark, 0);
3373                                 if (((*pmark ^ pindex) &
3374                                      ~PM_PLACEMARK_WAKEUP) == 0) {
3375                                         spin_unlock(&pmap->pm_spin);
3376                                         tsleep(pmark, PINTERLOCKED, "pvplc", 0);
3377                                         spin_lock(&pmap->pm_spin);
3378                                 }
3379                                 continue;
3380                         }
3381
3382                         /*
3383                          * Setup the new entry
3384                          */
3385                         pnew->pv_pmap = pmap;
3386                         pnew->pv_pindex = pindex;
3387                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
3388 #ifdef PMAP_DEBUG
3389                         pnew->pv_func = func;
3390                         pnew->pv_line = lineno;
3391                         if (pnew->pv_line_lastfree > 0) {
3392                                 pnew->pv_line_lastfree =
3393                                                 -pnew->pv_line_lastfree;
3394                         }
3395 #endif
3396                         pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3397                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
3398                         spin_unlock(&pmap->pm_spin);
3399                         *isnew = 1;
3400
3401                         KKASSERT(pv == NULL);
3402                         return(pnew);
3403                 }
3404
3405                 /*
3406                  * We already have an entry, cleanup the staged pnew if
3407                  * we can get the lock, otherwise block and retry.
3408                  */
3409                 if (__predict_true(_pv_hold_try(pv PMAP_DEBUG_COPY))) {
3410                         if (pmap_excl)
3411                                 spin_unlock(&pmap->pm_spin);
3412                         else
3413                                 spin_unlock_shared(&pmap->pm_spin);
3414 #if 0
3415                         pnew = atomic_swap_ptr((void *)&md->gd_newpv, pnew);
3416                         if (pnew)
3417                                 zfree(pvzone, pnew);
3418 #else
3419                         crit_enter();
3420                         if (md->gd_newpv == NULL)
3421                                 md->gd_newpv = pnew;
3422                         else
3423                                 zfree(pvzone, pnew);
3424                         crit_exit();
3425 #endif
3426                         KKASSERT(pv->pv_pmap == pmap &&
3427                                  pv->pv_pindex == pindex);
3428                         *isnew = 0;
3429                         return(pv);
3430                 }
3431                 if (pmap_excl) {
3432                         spin_unlock(&pmap->pm_spin);
3433                         _pv_lock(pv PMAP_DEBUG_COPY);
3434                         pv_put(pv);
3435                         spin_lock(&pmap->pm_spin);
3436                 } else {
3437                         spin_unlock_shared(&pmap->pm_spin);
3438                         _pv_lock(pv PMAP_DEBUG_COPY);
3439                         pv_put(pv);
3440                         spin_lock_shared(&pmap->pm_spin);
3441                 }
3442         }
3443         /* NOT REACHED */
3444 }
3445
3446 /*
3447  * Find the requested PV entry, returning a locked+held pv or NULL
3448  */
3449 static
3450 pv_entry_t
3451 _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp PMAP_DEBUG_DECL)
3452 {
3453         pv_entry_t pv;
3454         int pmap_excl = 0;
3455
3456         spin_lock_shared(&pmap->pm_spin);
3457         for (;;) {
3458                 /*
3459                  * Shortcut cache
3460                  */
3461                 pv = pmap->pm_pvhint;
3462                 cpu_ccfence();
3463                 if (pv == NULL ||
3464                     pv->pv_pmap != pmap ||
3465                     pv->pv_pindex != pindex) {
3466                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3467                                                         pindex);
3468                 }
3469                 if (pv == NULL) {
3470                         /*
3471                          * Block if there is ANY placemarker.  If we are to
3472                          * return it, we must also aquire the spot, so we
3473                          * have to block even if the placemarker is held on
3474                          * a different address.
3475                          *
3476                          * OPTIMIZATION: If pmarkp is passed as NULL the
3477                          * caller is just probing (or looking for a real
3478                          * pv_entry), and in this case we only need to check
3479                          * to see if the placemarker matches pindex.
3480                          */
3481                         vm_pindex_t *pmark;
3482
3483                         /*
3484                          * Requires exclusive pmap spinlock
3485                          */
3486                         if (pmap_excl == 0) {
3487                                 pmap_excl = 1;
3488                                 if (!spin_lock_upgrade_try(&pmap->pm_spin)) {
3489                                         spin_unlock_shared(&pmap->pm_spin);
3490                                         spin_lock(&pmap->pm_spin);
3491                                         continue;
3492                                 }
3493                         }
3494
3495                         pmark = pmap_placemarker_hash(pmap, pindex);
3496
3497                         if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3498                             ((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3499                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3500                                 tsleep_interlock(pmark, 0);
3501                                 if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3502                                     ((*pmark ^ pindex) &
3503                                      ~PM_PLACEMARK_WAKEUP) == 0) {
3504                                         spin_unlock(&pmap->pm_spin);
3505                                         tsleep(pmark, PINTERLOCKED, "pvpld", 0);
3506                                         spin_lock(&pmap->pm_spin);
3507                                 }
3508                                 continue;
3509                         }
3510                         if (pmarkp) {
3511                                 if (atomic_swap_long(pmark, pindex) !=
3512                                     PM_NOPLACEMARK) {
3513                                         panic("_pv_get: pmark race");
3514                                 }
3515                                 *pmarkp = pmark;
3516                         }
3517                         spin_unlock(&pmap->pm_spin);
3518                         return NULL;
3519                 }
3520                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3521                         pv_cache(pv, pindex);
3522                         if (pmap_excl)
3523                                 spin_unlock(&pmap->pm_spin);
3524                         else
3525                                 spin_unlock_shared(&pmap->pm_spin);
3526                         KKASSERT(pv->pv_pmap == pmap &&
3527                                  pv->pv_pindex == pindex);
3528                         return(pv);
3529                 }
3530                 if (pmap_excl) {
3531                         spin_unlock(&pmap->pm_spin);
3532                         _pv_lock(pv PMAP_DEBUG_COPY);
3533                         pv_put(pv);
3534                         spin_lock(&pmap->pm_spin);
3535                 } else {
3536                         spin_unlock_shared(&pmap->pm_spin);
3537                         _pv_lock(pv PMAP_DEBUG_COPY);
3538                         pv_put(pv);
3539                         spin_lock_shared(&pmap->pm_spin);
3540                 }
3541         }
3542 }
3543
3544 /*
3545  * Lookup, hold, and attempt to lock (pmap,pindex).
3546  *
3547  * If the entry does not exist NULL is returned and *errorp is set to 0
3548  *
3549  * If the entry exists and could be successfully locked it is returned and
3550  * errorp is set to 0.
3551  *
3552  * If the entry exists but could NOT be successfully locked it is returned
3553  * held and *errorp is set to 1.
3554  *
3555  * If the entry is placemarked by someone else NULL is returned and *errorp
3556  * is set to 1.
3557  */
3558 static
3559 pv_entry_t
3560 pv_get_try(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp, int *errorp)
3561 {
3562         pv_entry_t pv;
3563
3564         spin_lock_shared(&pmap->pm_spin);
3565
3566         pv = pmap->pm_pvhint;
3567         cpu_ccfence();
3568         if (pv == NULL ||
3569             pv->pv_pmap != pmap ||
3570             pv->pv_pindex != pindex) {
3571                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3572         }
3573
3574         if (pv == NULL) {
3575                 vm_pindex_t *pmark;
3576
3577                 pmark = pmap_placemarker_hash(pmap, pindex);
3578
3579                 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3580                         *errorp = 1;
3581                 } else if (pmarkp &&
3582                            atomic_cmpset_long(pmark, PM_NOPLACEMARK, pindex)) {
3583                         *errorp = 0;
3584                 } else {
3585                         /*
3586                          * Can't set a placemark with a NULL pmarkp, or if
3587                          * pmarkp is non-NULL but we failed to set our
3588                          * placemark.
3589                          */
3590                         *errorp = 1;
3591                 }
3592                 if (pmarkp)
3593                         *pmarkp = pmark;
3594                 spin_unlock_shared(&pmap->pm_spin);
3595
3596                 return NULL;
3597         }
3598
3599         /*
3600          * XXX This has problems if the lock is shared, why?
3601          */
3602         if (pv_hold_try(pv)) {
3603                 pv_cache(pv, pindex);   /* overwrite ok (shared lock) */
3604                 spin_unlock_shared(&pmap->pm_spin);
3605                 *errorp = 0;
3606                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3607                 return(pv);     /* lock succeeded */
3608         }
3609         spin_unlock_shared(&pmap->pm_spin);
3610         *errorp = 1;
3611
3612         return (pv);            /* lock failed */
3613 }
3614
3615 /*
3616  * Lock a held pv, keeping the hold count
3617  */
3618 static
3619 void
3620 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3621 {
3622         u_int count;
3623
3624         for (;;) {
3625                 count = pv->pv_hold;
3626                 cpu_ccfence();
3627                 if ((count & PV_HOLD_LOCKED) == 0) {
3628                         if (atomic_cmpset_int(&pv->pv_hold, count,
3629                                               count | PV_HOLD_LOCKED)) {
3630 #ifdef PMAP_DEBUG
3631                                 pv->pv_func = func;
3632                                 pv->pv_line = lineno;
3633 #endif
3634                                 return;
3635                         }
3636                         continue;
3637                 }
3638                 tsleep_interlock(pv, 0);
3639                 if (atomic_cmpset_int(&pv->pv_hold, count,
3640                                       count | PV_HOLD_WAITING)) {
3641 #ifdef PMAP_DEBUG2
3642                         if (pmap_enter_debug > 0) {
3643                                 --pmap_enter_debug;
3644                                 kprintf("pv waiting on %s:%d\n",
3645                                         pv->pv_func, pv->pv_line);
3646                         }
3647 #endif
3648                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
3649                 }
3650                 /* retry */
3651         }
3652 }
3653
3654 /*
3655  * Unlock a held and locked pv, keeping the hold count.
3656  */
3657 static
3658 void
3659 pv_unlock(pv_entry_t pv)
3660 {
3661         u_int count;
3662
3663         for (;;) {
3664                 count = pv->pv_hold;
3665                 cpu_ccfence();
3666                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3667                          (PV_HOLD_LOCKED | 1));
3668                 if (atomic_cmpset_int(&pv->pv_hold, count,
3669                                       count &
3670                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3671                         if (count & PV_HOLD_WAITING)
3672                                 wakeup(pv);
3673                         break;
3674                 }
3675         }
3676 }
3677
3678 /*
3679  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3680  * and the hold count drops to zero we will free it.
3681  *
3682  * Caller should not hold any spin locks.  We are protected from hold races
3683  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3684  * lock held.  A pv cannot be located otherwise.
3685  */
3686 static
3687 void
3688 pv_put(pv_entry_t pv)
3689 {
3690 #ifdef PMAP_DEBUG2
3691         if (pmap_enter_debug > 0) {
3692                 --pmap_enter_debug;
3693                 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3694         }
3695 #endif
3696
3697         /*
3698          * Normal put-aways must have a pv_m associated with the pv,
3699          * but allow the case where the pv has been destructed due
3700          * to pmap_dynamic_delete.
3701          */
3702         KKASSERT(pv->pv_pmap == NULL || pv->pv_m != NULL);
3703
3704         /*
3705          * Fast - shortcut most common condition
3706          */
3707         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3708                 return;
3709
3710         /*
3711          * Slow
3712          */
3713         pv_unlock(pv);
3714         pv_drop(pv);
3715 }
3716
3717 /*
3718  * Remove the pmap association from a pv, require that pv_m already be removed,
3719  * then unlock and drop the pv.  Any pte operations must have already been
3720  * completed.  This call may result in a last-drop which will physically free
3721  * the pv.
3722  *
3723  * Removing the pmap association entails an additional drop.
3724  *
3725  * pv must be exclusively locked on call and will be disposed of on return.
3726  */
3727 static
3728 void
3729 _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL)
3730 {
3731         pmap_t pmap;
3732
3733 #ifdef PMAP_DEBUG
3734         pv->pv_func_lastfree = func;
3735         pv->pv_line_lastfree = lineno;
3736 #endif
3737         KKASSERT(pv->pv_m == NULL);
3738         KKASSERT((pv->pv_hold & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
3739                   (PV_HOLD_LOCKED|1));
3740         if ((pmap = pv->pv_pmap) != NULL) {
3741                 spin_lock(&pmap->pm_spin);
3742                 KKASSERT(pv->pv_pmap == pmap);
3743                 if (pmap->pm_pvhint == pv)
3744                         pmap->pm_pvhint = NULL;
3745                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3746                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3747                 pv->pv_pmap = NULL;
3748                 pv->pv_pindex = 0;
3749                 spin_unlock(&pmap->pm_spin);
3750
3751                 /*
3752                  * Try to shortcut three atomic ops, otherwise fall through
3753                  * and do it normally.  Drop two refs and the lock all in
3754                  * one go.
3755                  */
3756                 if (pvp)
3757                         vm_page_unwire_quick(pvp->pv_m);
3758                 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3759 #ifdef PMAP_DEBUG2
3760                         if (pmap_enter_debug > 0) {
3761                                 --pmap_enter_debug;
3762                                 kprintf("pv_free: free pv %p\n", pv);
3763                         }
3764 #endif
3765                         zfree(pvzone, pv);
3766                         return;
3767                 }
3768                 pv_drop(pv);    /* ref for pv_pmap */
3769         }
3770         pv_unlock(pv);
3771         pv_drop(pv);
3772 }
3773
3774 /*
3775  * This routine is very drastic, but can save the system
3776  * in a pinch.
3777  */
3778 void
3779 pmap_collect(void)
3780 {
3781         int i;
3782         vm_page_t m;
3783         static int warningdone=0;
3784
3785         if (pmap_pagedaemon_waken == 0)
3786                 return;
3787         pmap_pagedaemon_waken = 0;
3788         if (warningdone < 5) {
3789                 kprintf("pmap_collect: collecting pv entries -- "
3790                         "suggest increasing PMAP_SHPGPERPROC\n");
3791                 warningdone++;
3792         }
3793
3794         for (i = 0; i < vm_page_array_size; i++) {
3795                 m = &vm_page_array[i];
3796                 if (m->wire_count || m->hold_count)
3797                         continue;
3798                 if (vm_page_busy_try(m, TRUE) == 0) {
3799                         if (m->wire_count == 0 && m->hold_count == 0) {
3800                                 pmap_remove_all(m);
3801                         }
3802                         vm_page_wakeup(m);
3803                 }
3804         }
3805 }
3806
3807 /*
3808  * Scan the pmap for active page table entries and issue a callback.
3809  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3810  * its parent page table.
3811  *
3812  * pte_pv will be NULL if the page or page table is unmanaged.
3813  * pt_pv will point to the page table page containing the pte for the page.
3814  *
3815  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3816  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3817  *       process pmap's PD and page to the callback function.  This can be
3818  *       confusing because the pt_pv is really a pd_pv, and the target page
3819  *       table page is simply aliased by the pmap and not owned by it.
3820  *
3821  * It is assumed that the start and end are properly rounded to the page size.
3822  *
3823  * It is assumed that PD pages and above are managed and thus in the RB tree,
3824  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3825  */
3826 struct pmap_scan_info {
3827         struct pmap *pmap;
3828         vm_offset_t sva;
3829         vm_offset_t eva;
3830         vm_pindex_t sva_pd_pindex;
3831         vm_pindex_t eva_pd_pindex;
3832         void (*func)(pmap_t, struct pmap_scan_info *,
3833                      pv_entry_t, vm_pindex_t *, pv_entry_t,
3834                      int, vm_offset_t,
3835                      pt_entry_t *, void *);
3836         void *arg;
3837         pmap_inval_bulk_t bulk_core;
3838         pmap_inval_bulk_t *bulk;
3839         int count;
3840         int stop;
3841 };
3842
3843 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3844 static int pmap_scan_callback(pv_entry_t pv, void *data);
3845
3846 static void
3847 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3848 {
3849         struct pmap *pmap = info->pmap;
3850         pv_entry_t pd_pv;       /* A page directory PV */
3851         pv_entry_t pt_pv;       /* A page table PV */
3852         pv_entry_t pte_pv;      /* A page table entry PV */
3853         vm_pindex_t *pte_placemark;
3854         vm_pindex_t *pt_placemark;
3855         pt_entry_t *ptep;
3856         pt_entry_t oldpte;
3857         struct pv_entry dummy_pv;
3858
3859         info->stop = 0;
3860         if (pmap == NULL)
3861                 return;
3862         if (info->sva == info->eva)
3863                 return;
3864         if (smp_inval) {
3865                 info->bulk = &info->bulk_core;
3866                 pmap_inval_bulk_init(&info->bulk_core, pmap);
3867         } else {
3868                 info->bulk = NULL;
3869         }
3870
3871         /*
3872          * Hold the token for stability; if the pmap is empty we have nothing
3873          * to do.
3874          */
3875 #if 0
3876         if (pmap->pm_stats.resident_count == 0) {
3877                 return;
3878         }
3879 #endif
3880
3881         info->count = 0;
3882
3883         /*
3884          * Special handling for scanning one page, which is a very common
3885          * operation (it is?).
3886          *
3887          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3888          */
3889         if (info->sva + PAGE_SIZE == info->eva) {
3890                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3891                         /*
3892                          * Kernel mappings do not track wire counts on
3893                          * page table pages and only maintain pd_pv and
3894                          * pte_pv levels so pmap_scan() works.
3895                          */
3896                         pt_pv = NULL;
3897                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3898                                         &pte_placemark);
3899                         ptep = vtopte(info->sva);
3900                 } else {
3901                         /*
3902                          * User pages which are unmanaged will not have a
3903                          * pte_pv.  User page table pages which are unmanaged
3904                          * (shared from elsewhere) will also not have a pt_pv.
3905                          * The func() callback will pass both pte_pv and pt_pv
3906                          * as NULL in that case.
3907                          *
3908                          * We hold pte_placemark across the operation for
3909                          * unmanaged pages.
3910                          *
3911                          * WARNING!  We must hold pt_placemark across the
3912                          *           *ptep test to prevent misintepreting
3913                          *           a non-zero *ptep as a shared page
3914                          *           table page.  Hold it across the function
3915                          *           callback as well for SMP safety.
3916                          */
3917                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3918                                         &pte_placemark);
3919                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva),
3920                                         &pt_placemark);
3921                         if (pt_pv == NULL) {
3922                                 KKASSERT(pte_pv == NULL);
3923                                 pd_pv = pv_get(pmap,
3924                                                pmap_pd_pindex(info->sva),
3925                                                NULL);
3926                                 if (pd_pv) {
3927                                         ptep = pv_pte_lookup(pd_pv,
3928                                                     pmap_pt_index(info->sva));
3929                                         if (*ptep) {
3930                                                 info->func(pmap, info,
3931                                                      NULL, pt_placemark,
3932                                                      pd_pv, 1,
3933                                                      info->sva, ptep,
3934                                                      info->arg);
3935                                         } else {
3936                                                 pv_placemarker_wakeup(pmap,
3937                                                                   pt_placemark);
3938                                         }
3939                                         pv_put(pd_pv);
3940                                 } else {
3941                                         pv_placemarker_wakeup(pmap,
3942                                                               pt_placemark);
3943                                 }
3944                                 pv_placemarker_wakeup(pmap, pte_placemark);
3945                                 goto fast_skip;
3946                         }
3947                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3948                 }
3949
3950                 /*
3951                  * NOTE: *ptep can't be ripped out from under us if we hold
3952                  *       pte_pv (or pte_placemark) locked, but bits can
3953                  *       change.
3954                  */
3955                 oldpte = *ptep;
3956                 cpu_ccfence();
3957                 if (oldpte == 0) {
3958                         KKASSERT(pte_pv == NULL);
3959                         pv_placemarker_wakeup(pmap, pte_placemark);
3960                 } else if (pte_pv) {
3961                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3962                                            pmap->pmap_bits[PG_V_IDX])) ==
3963                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
3964                                  pmap->pmap_bits[PG_V_IDX]),
3965                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p",
3966                             *ptep, oldpte, info->sva, pte_pv));
3967                         info->func(pmap, info, pte_pv, NULL, pt_pv, 0,
3968                                    info->sva, ptep, info->arg);
3969                 } else {
3970                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3971                                            pmap->pmap_bits[PG_V_IDX])) ==
3972                             pmap->pmap_bits[PG_V_IDX],
3973                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
3974                             *ptep, oldpte, info->sva));
3975                         info->func(pmap, info, NULL, pte_placemark, pt_pv, 0,
3976                                    info->sva, ptep, info->arg);
3977                 }
3978                 if (pt_pv)
3979                         pv_put(pt_pv);
3980 fast_skip:
3981                 pmap_inval_bulk_flush(info->bulk);
3982                 return;
3983         }
3984
3985         /*
3986          * Nominal scan case, RB_SCAN() for PD pages and iterate from
3987          * there.
3988          *
3989          * WARNING! eva can overflow our standard ((N + mask) >> bits)
3990          *          bounds, resulting in a pd_pindex of 0.  To solve the
3991          *          problem we use an inclusive range.
3992          */
3993         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3994         info->eva_pd_pindex = pmap_pd_pindex(info->eva - PAGE_SIZE);
3995
3996         if (info->sva >= VM_MAX_USER_ADDRESS) {
3997                 /*
3998                  * The kernel does not currently maintain any pv_entry's for
3999                  * higher-level page tables.
4000                  */
4001                 bzero(&dummy_pv, sizeof(dummy_pv));
4002                 dummy_pv.pv_pindex = info->sva_pd_pindex;
4003                 spin_lock(&pmap->pm_spin);
4004                 while (dummy_pv.pv_pindex <= info->eva_pd_pindex) {
4005                         pmap_scan_callback(&dummy_pv, info);
4006                         ++dummy_pv.pv_pindex;
4007                         if (dummy_pv.pv_pindex < info->sva_pd_pindex) /*wrap*/
4008                                 break;
4009                 }
4010                 spin_unlock(&pmap->pm_spin);
4011         } else {
4012                 /*
4013                  * User page tables maintain local PML4, PDP, and PD
4014                  * pv_entry's at the very least.  PT pv's might be
4015                  * unmanaged and thus not exist.  PTE pv's might be
4016                  * unmanaged and thus not exist.
4017                  */
4018                 spin_lock(&pmap->pm_spin);
4019                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot, pmap_scan_cmp,
4020                                          pmap_scan_callback, info);
4021                 spin_unlock(&pmap->pm_spin);
4022         }
4023         pmap_inval_bulk_flush(info->bulk);
4024 }
4025
4026 /*
4027  * WARNING! pmap->pm_spin held
4028  *
4029  * WARNING! eva can overflow our standard ((N + mask) >> bits)
4030  *          bounds, resulting in a pd_pindex of 0.  To solve the
4031  *          problem we use an inclusive range.
4032  */
4033 static int
4034 pmap_scan_cmp(pv_entry_t pv, void *data)
4035 {
4036         struct pmap_scan_info *info = data;
4037         if (pv->pv_pindex < info->sva_pd_pindex)
4038                 return(-1);
4039         if (pv->pv_pindex > info->eva_pd_pindex)
4040                 return(1);
4041         return(0);
4042 }
4043
4044 /*
4045  * pmap_scan() by PDs
4046  *
4047  * WARNING! pmap->pm_spin held
4048  */
4049 static int
4050 pmap_scan_callback(pv_entry_t pv, void *data)
4051 {
4052         struct pmap_scan_info *info = data;
4053         struct pmap *pmap = info->pmap;
4054         pv_entry_t pd_pv;       /* A page directory PV */
4055         pv_entry_t pt_pv;       /* A page table PV */
4056         vm_pindex_t *pt_placemark;
4057         pt_entry_t *ptep;
4058         pt_entry_t oldpte;
4059         vm_offset_t sva;
4060         vm_offset_t eva;
4061         vm_offset_t va_next;
4062         vm_pindex_t pd_pindex;
4063         int error;
4064
4065         /*
4066          * Stop if requested
4067          */
4068         if (info->stop)
4069                 return -1;
4070
4071         /*
4072          * Pull the PD pindex from the pv before releasing the spinlock.
4073          *
4074          * WARNING: pv is faked for kernel pmap scans.
4075          */
4076         pd_pindex = pv->pv_pindex;
4077         spin_unlock(&pmap->pm_spin);
4078         pv = NULL;      /* invalid after spinlock unlocked */
4079
4080         /*
4081          * Calculate the page range within the PD.  SIMPLE pmaps are
4082          * direct-mapped for the entire 2^64 address space.  Normal pmaps
4083          * reflect the user and kernel address space which requires
4084          * cannonicalization w/regards to converting pd_pindex's back
4085          * into addresses.
4086          */
4087         sva = (pd_pindex - pmap_pd_pindex(0)) << PDPSHIFT;
4088         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
4089             (sva & PML4_SIGNMASK)) {
4090                 sva |= PML4_SIGNMASK;
4091         }
4092         eva = sva + NBPDP;      /* can overflow */
4093         if (sva < info->sva)
4094                 sva = info->sva;
4095         if (eva < info->sva || eva > info->eva)
4096                 eva = info->eva;
4097
4098         /*
4099          * NOTE: kernel mappings do not track page table pages, only
4100          *       terminal pages.
4101          *
4102          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
4103          *       However, for the scan to be efficient we try to
4104          *       cache items top-down.
4105          */
4106         pd_pv = NULL;
4107         pt_pv = NULL;
4108
4109         for (; sva < eva; sva = va_next) {
4110                 if (info->stop)
4111                         break;
4112                 if (sva >= VM_MAX_USER_ADDRESS) {
4113                         if (pt_pv) {
4114                                 pv_put(pt_pv);
4115                                 pt_pv = NULL;
4116                         }
4117                         goto kernel_skip;
4118                 }
4119
4120                 /*
4121                  * PD cache, scan shortcut if it doesn't exist.
4122                  */
4123                 if (pd_pv == NULL) {
4124                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4125                 } else if (pd_pv->pv_pmap != pmap ||
4126                            pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
4127                         pv_put(pd_pv);
4128                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4129                 }
4130                 if (pd_pv == NULL) {
4131                         va_next = (sva + NBPDP) & ~PDPMASK;
4132                         if (va_next < sva)
4133                                 va_next = eva;
4134                         continue;
4135                 }
4136
4137                 /*
4138                  * PT cache
4139                  *
4140                  * NOTE: The cached pt_pv can be removed from the pmap when
4141                  *       pmap_dynamic_delete is enabled.
4142                  */
4143                 if (pt_pv && (pt_pv->pv_pmap != pmap ||
4144                               pt_pv->pv_pindex != pmap_pt_pindex(sva))) {
4145                         pv_put(pt_pv);
4146                         pt_pv = NULL;
4147                 }
4148                 if (pt_pv == NULL) {
4149                         pt_pv = pv_get_try(pmap, pmap_pt_pindex(sva),
4150                                            &pt_placemark, &error);
4151                         if (error) {
4152                                 pv_put(pd_pv);  /* lock order */
4153                                 pd_pv = NULL;
4154                                 if (pt_pv) {
4155                                         pv_lock(pt_pv);
4156                                         pv_put(pt_pv);
4157                                         pt_pv = NULL;
4158                                 } else {
4159                                         pv_placemarker_wait(pmap, pt_placemark);
4160                                 }
4161                                 va_next = sva;
4162                                 continue;
4163                         }
4164                         /* may have to re-check later if pt_pv is NULL here */
4165                 }
4166
4167                 /*
4168                  * If pt_pv is NULL we either have an shared page table
4169                  * page and must issue a callback specific to that case,
4170                  * or there is no page table page.
4171                  *
4172                  * Either way we can skip the page table page.
4173                  *
4174                  * WARNING! pt_pv can also be NULL due to a pv creation
4175                  *          race where we find it to be NULL and then
4176                  *          later see a pte_pv.  But its possible the pt_pv
4177                  *          got created inbetween the two operations, so
4178                  *          we must check.
4179                  */
4180                 if (pt_pv == NULL) {
4181                         /*
4182                          * Possible unmanaged (shared from another pmap)
4183                          * page table page.
4184                          *
4185                          * WARNING!  We must hold pt_placemark across the
4186                          *           *ptep test to prevent misintepreting
4187                          *           a non-zero *ptep as a shared page
4188                          *           table page.  Hold it across the function
4189                          *           callback as well for SMP safety.
4190                          */
4191                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
4192                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
4193                                 info->func(pmap, info, NULL, pt_placemark,
4194                                            pd_pv, 1,
4195                                            sva, ptep, info->arg);
4196                         } else {
4197                                 pv_placemarker_wakeup(pmap, pt_placemark);
4198                         }
4199
4200                         /*
4201                          * Done, move to next page table page.
4202                          */
4203                         va_next = (sva + NBPDR) & ~PDRMASK;
4204                         if (va_next < sva)
4205                                 va_next = eva;
4206                         continue;
4207                 }
4208
4209                 /*
4210                  * From this point in the loop testing pt_pv for non-NULL
4211                  * means we are in UVM, else if it is NULL we are in KVM.
4212                  *
4213                  * Limit our scan to either the end of the va represented
4214                  * by the current page table page, or to the end of the
4215                  * range being removed.
4216                  */
4217 kernel_skip:
4218                 va_next = (sva + NBPDR) & ~PDRMASK;
4219                 if (va_next < sva)
4220                         va_next = eva;
4221                 if (va_next > eva)
4222                         va_next = eva;
4223
4224                 /*
4225                  * Scan the page table for pages.  Some pages may not be
4226                  * managed (might not have a pv_entry).
4227                  *
4228                  * There is no page table management for kernel pages so
4229                  * pt_pv will be NULL in that case, but otherwise pt_pv
4230                  * is non-NULL, locked, and referenced.
4231                  */
4232
4233                 /*
4234                  * At this point a non-NULL pt_pv means a UVA, and a NULL
4235                  * pt_pv means a KVA.
4236                  */
4237                 if (pt_pv)
4238                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
4239                 else
4240                         ptep = vtopte(sva);
4241
4242                 while (sva < va_next) {
4243                         pv_entry_t pte_pv;
4244                         vm_pindex_t *pte_placemark;
4245
4246                         /*
4247                          * Yield every 64 pages, stop if requested.
4248                          */
4249                         if ((++info->count & 63) == 0)
4250                                 lwkt_user_yield();
4251                         if (info->stop)
4252                                 break;
4253
4254                         /*
4255                          * We can shortcut our scan if *ptep == 0.  This is
4256                          * an unlocked check.
4257                          */
4258                         if (*ptep == 0) {
4259                                 sva += PAGE_SIZE;
4260                                 ++ptep;
4261                                 continue;
4262                         }
4263                         cpu_ccfence();
4264
4265                         /*
4266                          * Acquire the related pte_pv, if any.  If *ptep == 0
4267                          * the related pte_pv should not exist, but if *ptep
4268                          * is not zero the pte_pv may or may not exist (e.g.
4269                          * will not exist for an unmanaged page).
4270                          *
4271                          * However a multitude of races are possible here
4272                          * so if we cannot lock definite state we clean out
4273                          * our cache and break the inner while() loop to
4274                          * force a loop up to the top of the for().
4275                          *
4276                          * XXX unlock/relock pd_pv, pt_pv, and re-test their
4277                          *     validity instead of looping up?
4278                          */
4279                         pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
4280                                             &pte_placemark, &error);
4281                         if (error) {
4282                                 pv_put(pd_pv);          /* lock order */
4283                                 pd_pv = NULL;
4284                                 if (pt_pv) {
4285                                         pv_put(pt_pv);  /* lock order */
4286                                         pt_pv = NULL;
4287                                 }
4288                                 if (pte_pv) {           /* block */
4289                                         pv_lock(pte_pv);
4290                                         pv_put(pte_pv);
4291                                         pte_pv = NULL;
4292                                 } else {
4293                                         pv_placemarker_wait(pmap,
4294                                                         pte_placemark);
4295                                 }
4296                                 va_next = sva;          /* retry */
4297                                 break;
4298                         }
4299
4300                         /*
4301                          * Reload *ptep after successfully locking the
4302                          * pindex.  If *ptep == 0 we had better NOT have a
4303                          * pte_pv.
4304                          */
4305                         cpu_ccfence();
4306                         oldpte = *ptep;
4307                         if (oldpte == 0) {
4308                                 if (pte_pv) {
4309                                         kprintf("Unexpected non-NULL pte_pv "
4310                                                 "%p pt_pv %p "
4311                                                 "*ptep = %016lx/%016lx\n",
4312                                                 pte_pv, pt_pv, *ptep, oldpte);
4313                                         panic("Unexpected non-NULL pte_pv");
4314                                 } else {
4315                                         pv_placemarker_wakeup(pmap, pte_placemark);
4316                                 }
4317                                 sva += PAGE_SIZE;
4318                                 ++ptep;
4319                                 continue;
4320                         }
4321
4322                         /*
4323                          * We can't hold pd_pv across the callback (because
4324                          * we don't pass it to the callback and the callback
4325                          * might deadlock)
4326                          */
4327                         if (pd_pv) {
4328                                 vm_page_wire_quick(pd_pv->pv_m);
4329                                 pv_unlock(pd_pv);
4330                         }
4331
4332                         /*
4333                          * Ready for the callback.  The locked pte_pv (if any)
4334                          * is consumed by the callback.  pte_pv will exist if
4335                          * the page is managed, and will not exist if it
4336                          * isn't.
4337                          */
4338                         if (oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4339                                 /*
4340                                  * Managed pte
4341                                  */
4342                                 KASSERT(pte_pv &&
4343                                          (oldpte & pmap->pmap_bits[PG_V_IDX]),
4344                                     ("badC *ptep %016lx/%016lx sva %016lx "
4345                                     "pte_pv %p",
4346                                     *ptep, oldpte, sva, pte_pv));
4347                                 /*
4348                                  * We must unlock pd_pv across the callback
4349                                  * to avoid deadlocks on any recursive
4350                                  * disposal.  Re-check that it still exists
4351                                  * after re-locking.
4352                                  *
4353                                  * Call target disposes of pte_pv and may
4354                                  * destroy but will not dispose of pt_pv.
4355                                  */
4356                                 info->func(pmap, info, pte_pv, NULL,
4357                                            pt_pv, 0,
4358                                            sva, ptep, info->arg);
4359                         } else {
4360                                 /*
4361                                  * Unmanaged pte
4362                                  *
4363                                  * We must unlock pd_pv across the callback
4364                                  * to avoid deadlocks on any recursive
4365                                  * disposal.  Re-check that it still exists
4366                                  * after re-locking.
4367                                  *
4368                                  * Call target disposes of pte_pv or
4369                                  * pte_placemark and may destroy but will
4370                                  * not dispose of pt_pv.
4371                                  */
4372                                 KASSERT(pte_pv == NULL &&
4373                                         (oldpte & pmap->pmap_bits[PG_V_IDX]),
4374                                     ("badD *ptep %016lx/%016lx sva %016lx "
4375                                     "pte_pv %p pte_pv->pv_m %p ",
4376                                      *ptep, oldpte, sva,
4377                                      pte_pv, (pte_pv ? pte_pv->pv_m : NULL)));
4378                                 if (pte_pv)
4379                                         kprintf("RaceD\n");
4380                                 if (pte_pv) {
4381                                         info->func(pmap, info,
4382                                                    pte_pv, NULL,
4383                                                    pt_pv, 0,
4384                                                    sva, ptep, info->arg);
4385                                 } else {
4386                                         info->func(pmap, info,
4387                                                    NULL, pte_placemark,
4388                                                    pt_pv, 0,
4389                                                    sva, ptep, info->arg);
4390                                 }
4391                         }
4392                         if (pd_pv) {
4393                                 pv_lock(pd_pv);
4394                                 vm_page_unwire_quick(pd_pv->pv_m);
4395                                 if (pd_pv->pv_pmap == NULL) {
4396                                         va_next = sva;          /* retry */
4397                                         break;
4398                                 }
4399                         }
4400
4401                         /*
4402                          * NOTE: The cached pt_pv can be removed from the
4403                          *       pmap when pmap_dynamic_delete is enabled,
4404                          *       which will cause ptep to become stale.
4405                          *
4406                          *       This also means that no pages remain under
4407                          *       the PT, so we can just break out of the inner
4408                          *       loop and let the outer loop clean everything
4409                          *       up.
4410                          */
4411                         if (pt_pv && pt_pv->pv_pmap != pmap)
4412                                 break;
4413                         pte_pv = NULL;
4414                         sva += PAGE_SIZE;
4415                         ++ptep;
4416                 }
4417         }
4418         if (pd_pv) {
4419                 pv_put(pd_pv);
4420                 pd_pv = NULL;
4421         }
4422         if (pt_pv) {
4423                 pv_put(pt_pv);
4424                 pt_pv = NULL;
4425         }
4426         if ((++info->count & 7) == 0)
4427                 lwkt_user_yield();
4428
4429         /*
4430          * Relock before returning.
4431          */
4432         spin_lock(&pmap->pm_spin);
4433         return (0);
4434 }
4435
4436 void
4437 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4438 {
4439         struct pmap_scan_info info;
4440
4441         info.pmap = pmap;
4442         info.sva = sva;
4443         info.eva = eva;
4444         info.func = pmap_remove_callback;
4445         info.arg = NULL;
4446         pmap_scan(&info, 1);
4447 #if 0
4448         cpu_invltlb();
4449         if (eva - sva < 1024*1024) {
4450                 while (sva < eva) {
4451                         cpu_invlpg((void *)sva);
4452                         sva += PAGE_SIZE;
4453                 }
4454         }
4455 #endif
4456 }
4457
4458 static void
4459 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4460 {
4461         struct pmap_scan_info info;
4462
4463         info.pmap = pmap;
4464         info.sva = sva;
4465         info.eva = eva;
4466         info.func = pmap_remove_callback;
4467         info.arg = NULL;
4468         pmap_scan(&info, 0);
4469 }
4470
4471 static void
4472 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4473                      pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4474                      pv_entry_t pt_pv, int sharept,
4475                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4476 {
4477         pt_entry_t pte;
4478
4479         if (pte_pv) {
4480                 /*
4481                  * Managed entry
4482                  *
4483                  * This will also drop pt_pv's wire_count. Note that
4484                  * terminal pages are not wired based on mmu presence.
4485                  *
4486                  * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4487                  */
4488                 KKASSERT(pte_pv->pv_m != NULL);
4489                 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4490                 pte_pv = NULL;  /* safety */
4491
4492                 /*
4493                  * Recursively destroy higher-level page tables.
4494                  *
4495                  * This is optional.  If we do not, they will still
4496                  * be destroyed when the process exits.
4497                  *
4498                  * NOTE: Do not destroy pv_entry's with extra hold refs,
4499                  *       a caller may have unlocked it and intends to
4500                  *       continue to use it.
4501                  */
4502                 if (pmap_dynamic_delete &&
4503                     pt_pv &&
4504                     pt_pv->pv_m &&
4505                     pt_pv->pv_m->wire_count == 1 &&
4506                     (pt_pv->pv_hold & PV_HOLD_MASK) == 2 &&
4507                     pt_pv->pv_pindex != pmap_pml4_pindex()) {
4508                         if (pmap_dynamic_delete == 2)
4509                                 kprintf("B %jd %08x\n", pt_pv->pv_pindex, pt_pv->pv_hold);
4510                         pv_hold(pt_pv); /* extra hold */
4511                         pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4512                         pv_lock(pt_pv); /* prior extra hold + relock */
4513                 }
4514         } else if (sharept == 0) {
4515                 /*
4516                  * Unmanaged pte (pte_placemark is non-NULL)
4517                  *
4518                  * pt_pv's wire_count is still bumped by unmanaged pages
4519                  * so we must decrement it manually.
4520                  *
4521                  * We have to unwire the target page table page.
4522                  */
4523                 pte = pmap_inval_bulk(info->bulk, va, ptep, 0);
4524                 if (pte & pmap->pmap_bits[PG_W_IDX])
4525                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4526                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4527                 if (vm_page_unwire_quick(pt_pv->pv_m))
4528                         panic("pmap_remove: insufficient wirecount");
4529                 pv_placemarker_wakeup(pmap, pte_placemark);
4530         } else {
4531                 /*
4532                  * Unmanaged page table (pt, pd, or pdp. Not pte) for
4533                  * a shared page table.
4534                  *
4535                  * pt_pv is actually the pd_pv for our pmap (not the shared
4536                  * object pmap).
4537                  *
4538                  * We have to unwire the target page table page and we
4539                  * have to unwire our page directory page.
4540                  *
4541                  * It is unclear how we can invalidate a segment so we
4542                  * invalidate -1 which invlidates the tlb.
4543                  */
4544                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4545                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4546                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4547                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4548                         panic("pmap_remove: shared pgtable1 bad wirecount");
4549                 if (vm_page_unwire_quick(pt_pv->pv_m))
4550                         panic("pmap_remove: shared pgtable2 bad wirecount");
4551                 pv_placemarker_wakeup(pmap, pte_placemark);
4552         }
4553 }
4554
4555 /*
4556  * Removes this physical page from all physical maps in which it resides.
4557  * Reflects back modify bits to the pager.
4558  *
4559  * This routine may not be called from an interrupt.
4560  */
4561 static
4562 void
4563 pmap_remove_all(vm_page_t m)
4564 {
4565         pv_entry_t pv;
4566         pmap_inval_bulk_t bulk;
4567
4568         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4569                 return;
4570
4571         vm_page_spin_lock(m);
4572         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4573                 KKASSERT(pv->pv_m == m);
4574                 if (pv_hold_try(pv)) {
4575                         vm_page_spin_unlock(m);
4576                 } else {
4577                         vm_page_spin_unlock(m);
4578                         pv_lock(pv);
4579                         pv_put(pv);
4580                         vm_page_spin_lock(m);
4581                         continue;
4582                 }
4583                 KKASSERT(pv->pv_pmap && pv->pv_m == m);
4584
4585                 /*
4586                  * Holding no spinlocks, pv is locked.  Once we scrap
4587                  * pv we can no longer use it as a list iterator (but
4588                  * we are doing a TAILQ_FIRST() so we are ok).
4589                  */
4590                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4591                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4592                 pv = NULL;      /* safety */
4593                 pmap_inval_bulk_flush(&bulk);
4594                 vm_page_spin_lock(m);
4595         }
4596         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4597         vm_page_spin_unlock(m);
4598 }
4599
4600 /*
4601  * Removes the page from a particular pmap
4602  */
4603 void
4604 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4605 {
4606         pv_entry_t pv;
4607         pmap_inval_bulk_t bulk;
4608
4609         if (!pmap_initialized)
4610                 return;
4611
4612 again:
4613         vm_page_spin_lock(m);
4614         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4615                 if (pv->pv_pmap != pmap)
4616                         continue;
4617                 KKASSERT(pv->pv_m == m);
4618                 if (pv_hold_try(pv)) {
4619                         vm_page_spin_unlock(m);
4620                 } else {
4621                         vm_page_spin_unlock(m);
4622                         pv_lock(pv);
4623                         pv_put(pv);
4624                         goto again;
4625                 }
4626                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
4627
4628                 /*
4629                  * Holding no spinlocks, pv is locked.  Once gone it can't
4630                  * be used as an iterator.  In fact, because we couldn't
4631                  * necessarily lock it atomically it may have moved within
4632                  * the list and ALSO cannot be used as an iterator.
4633                  */
4634                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4635                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4636                 pv = NULL;      /* safety */
4637                 pmap_inval_bulk_flush(&bulk);
4638                 goto again;
4639         }
4640         vm_page_spin_unlock(m);
4641 }
4642
4643 /*
4644  * Set the physical protection on the specified range of this map
4645  * as requested.  This function is typically only used for debug watchpoints
4646  * and COW pages.
4647  *
4648  * This function may not be called from an interrupt if the map is
4649  * not the kernel_pmap.
4650  *
4651  * NOTE!  For shared page table pages we just unmap the page.
4652  */
4653 void
4654 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4655 {
4656         struct pmap_scan_info info;
4657         /* JG review for NX */
4658
4659         if (pmap == NULL)
4660                 return;
4661         if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == VM_PROT_NONE) {
4662                 pmap_remove(pmap, sva, eva);
4663                 return;
4664         }
4665         if (prot & VM_PROT_WRITE)
4666                 return;
4667         info.pmap = pmap;
4668         info.sva = sva;
4669         info.eva = eva;
4670         info.func = pmap_protect_callback;
4671         info.arg = &prot;
4672         pmap_scan(&info, 1);
4673 }
4674
4675 static
4676 void
4677 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4678                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4679                       pv_entry_t pt_pv, int sharept,
4680                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4681 {
4682         pt_entry_t pbits;
4683         pt_entry_t cbits;
4684         pt_entry_t pte;
4685         vm_page_t m;
4686
4687 again:
4688         pbits = *ptep;
4689         cbits = pbits;
4690         if (pte_pv) {
4691                 KKASSERT(pte_pv->pv_m != NULL);
4692                 m = NULL;
4693                 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4694                         if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4695                                 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4696                                 KKASSERT(m == pte_pv->pv_m);
4697                                 vm_page_flag_set(m, PG_REFERENCED);
4698                         }
4699                         cbits &= ~pmap->pmap_bits[PG_A_IDX];
4700                 }
4701                 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4702                         if (pmap_track_modified(pte_pv->pv_pindex)) {
4703                                 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4704                                         if (m == NULL) {
4705                                                 m = PHYS_TO_VM_PAGE(pbits &
4706                                                                     PG_FRAME);
4707                                         }
4708                                         vm_page_dirty(m);
4709                                 }
4710                                 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4711                         }
4712                 }
4713         } else if (sharept) {
4714                 /*
4715                  * Unmanaged page table, pt_pv is actually the pd_pv
4716                  * for our pmap (not the object's shared pmap).
4717                  *
4718                  * When asked to protect something in a shared page table
4719                  * page we just unmap the page table page.  We have to
4720                  * invalidate the tlb in this situation.
4721                  *
4722                  * XXX Warning, shared page tables will not be used for
4723                  * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4724                  * so PHYS_TO_VM_PAGE() should be safe here.
4725                  */
4726                 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4727                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4728                         panic("pmap_protect: pgtable1 pg bad wirecount");
4729                 if (vm_page_unwire_quick(pt_pv->pv_m))
4730                         panic("pmap_protect: pgtable2 pg bad wirecount");
4731                 ptep = NULL;
4732         }
4733         /* else unmanaged page, adjust bits, no wire changes */
4734
4735         if (ptep) {
4736                 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4737 #ifdef PMAP_DEBUG2
4738                 if (pmap_enter_debug > 0) {
4739                         --pmap_enter_debug;
4740                         kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4741                                 "pt_pv=%p cbits=%08lx\n",
4742                                 va, ptep, pte_pv,
4743                                 pt_pv, cbits
4744                         );
4745                 }
4746 #endif
4747                 if (pbits != cbits) {
4748                         vm_offset_t xva;
4749
4750                         xva = (sharept) ? (vm_offset_t)-1 : va;
4751                         if (!pmap_inval_smp_cmpset(pmap, xva,
4752                                                    ptep, pbits, cbits)) {
4753                                 goto again;
4754                         }
4755                 }
4756         }
4757         if (pte_pv)
4758                 pv_put(pte_pv);
4759         else
4760                 pv_placemarker_wakeup(pmap, pte_placemark);
4761 }
4762
4763 /*
4764  * Insert the vm_page (m) at the virtual address (va), replacing any prior
4765  * mapping at that address.  Set protection and wiring as requested.
4766  *
4767  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4768  * possible.  If it is we enter the page into the appropriate shared pmap
4769  * hanging off the related VM object instead of the passed pmap, then we
4770  * share the page table page from the VM object's pmap into the current pmap.
4771  *
4772  * NOTE: This routine MUST insert the page into the pmap now, it cannot
4773  *       lazy-evaluate.
4774  *
4775  * NOTE: If (m) is PG_UNMANAGED it may also be a temporary fake vm_page_t.
4776  *       never record it.
4777  */
4778 void
4779 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4780            boolean_t wired, vm_map_entry_t entry)
4781 {
4782         pv_entry_t pt_pv;       /* page table */
4783         pv_entry_t pte_pv;      /* page table entry */
4784         vm_pindex_t *pte_placemark;
4785         pt_entry_t *ptep;
4786         vm_paddr_t opa;
4787         pt_entry_t origpte, newpte;
4788         vm_paddr_t pa;
4789
4790         if (pmap == NULL)
4791                 return;
4792         va = trunc_page(va);
4793 #ifdef PMAP_DIAGNOSTIC
4794         if (va >= KvaEnd)
4795                 panic("pmap_enter: toobig");
4796         if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4797                 panic("pmap_enter: invalid to pmap_enter page table "
4798                       "pages (va: 0x%lx)", va);
4799 #endif
4800         if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4801                 kprintf("Warning: pmap_enter called on UVA with "
4802                         "kernel_pmap\n");
4803 #ifdef DDB
4804                 db_print_backtrace();
4805 #endif
4806         }
4807         if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4808                 kprintf("Warning: pmap_enter called on KVA without"
4809                         "kernel_pmap\n");
4810 #ifdef DDB
4811                 db_print_backtrace();
4812 #endif
4813         }
4814
4815         /*
4816          * Get locked PV entries for our new page table entry (pte_pv or
4817          * pte_placemark) and for its parent page table (pt_pv).  We need
4818          * the parent so we can resolve the location of the ptep.
4819          *
4820          * Only hardware MMU actions can modify the ptep out from
4821          * under us.
4822          *
4823          * if (m) is fictitious or unmanaged we do not create a managing
4824          * pte_pv for it.  Any pre-existing page's management state must
4825          * match (avoiding code complexity).
4826          *
4827          * If the pmap is still being initialized we assume existing
4828          * page tables.
4829          *
4830          * Kernel mapppings do not track page table pages (i.e. pt_pv).
4831          *
4832          * WARNING! If replacing a managed mapping with an unmanaged mapping
4833          *          pte_pv will wind up being non-NULL and must be handled
4834          *          below.
4835          */
4836         if (pmap_initialized == FALSE) {
4837                 pte_pv = NULL;
4838                 pt_pv = NULL;
4839                 pte_placemark = NULL;
4840                 ptep = vtopte(va);
4841                 origpte = *ptep;
4842         } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4843                 pte_pv = pv_get(pmap, pmap_pte_pindex(va), &pte_placemark);
4844                 KKASSERT(pte_pv == NULL);
4845                 if (va >= VM_MAX_USER_ADDRESS) {
4846                         pt_pv = NULL;
4847                         ptep = vtopte(va);
4848                 } else {
4849                         pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4850                                                   NULL, entry, va);
4851                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4852                 }
4853                 origpte = *ptep;
4854                 cpu_ccfence();
4855                 KASSERT(origpte == 0 ||
4856                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4857                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4858         } else {
4859                 if (va >= VM_MAX_USER_ADDRESS) {
4860                         /*
4861                          * Kernel map, pv_entry-tracked.
4862                          */
4863                         pt_pv = NULL;
4864                         pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4865                         ptep = vtopte(va);
4866                 } else {
4867                         /*
4868                          * User map
4869                          */
4870                         pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4871                                                    &pt_pv, entry, va);
4872                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4873                 }
4874                 pte_placemark = NULL;   /* safety */
4875                 origpte = *ptep;
4876                 cpu_ccfence();
4877                 KASSERT(origpte == 0 ||
4878                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4879                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4880         }
4881
4882         pa = VM_PAGE_TO_PHYS(m);
4883         opa = origpte & PG_FRAME;
4884
4885         /*
4886          * Calculate the new PTE.  Note that pte_pv alone does not mean
4887          * the new pte_pv is managed, it could exist because the old pte
4888          * was managed even if the new one is not.
4889          */
4890         newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4891                  pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4892         if (wired)
4893                 newpte |= pmap->pmap_bits[PG_W_IDX];
4894         if (va < VM_MAX_USER_ADDRESS)
4895                 newpte |= pmap->pmap_bits[PG_U_IDX];
4896         if (pte_pv && (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) == 0)
4897                 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4898 //      if (pmap == &kernel_pmap)
4899 //              newpte |= pgeflag;
4900         newpte |= pmap->pmap_cache_bits[m->pat_mode];
4901         if (m->flags & PG_FICTITIOUS)
4902                 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4903
4904         /*
4905          * It is possible for multiple faults to occur in threaded
4906          * environments, the existing pte might be correct.
4907          */
4908         if (((origpte ^ newpte) &
4909             ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4910                           pmap->pmap_bits[PG_A_IDX])) == 0) {
4911                 goto done;
4912         }
4913
4914         /*
4915          * Ok, either the address changed or the protection or wiring
4916          * changed.
4917          *
4918          * Clear the current entry, interlocking the removal.  For managed
4919          * pte's this will also flush the modified state to the vm_page.
4920          * Atomic ops are mandatory in order to ensure that PG_M events are
4921          * not lost during any transition.
4922          *
4923          * WARNING: The caller has busied the new page but not the original
4924          *          vm_page which we are trying to replace.  Because we hold
4925          *          the pte_pv lock, but have not busied the page, PG bits
4926          *          can be cleared out from under us.
4927          */
4928         if (opa) {
4929                 if (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4930                         /*
4931                          * Old page was managed.  Expect pte_pv to exist.
4932                          * (it might also exist if the old page was unmanaged).
4933                          *
4934                          * NOTE: pt_pv won't exist for a kernel page
4935                          *       (managed or otherwise).
4936                          *
4937                          * NOTE: We may be reusing the pte_pv so we do not
4938                          *       destroy it in pmap_remove_pv_pte().
4939                          */
4940                         KKASSERT(pte_pv && pte_pv->pv_m);
4941                         if (prot & VM_PROT_NOSYNC) {
4942                                 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4943                         } else {
4944                                 pmap_inval_bulk_t bulk;
4945
4946                                 pmap_inval_bulk_init(&bulk, pmap);
4947                                 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4948                                 pmap_inval_bulk_flush(&bulk);
4949                         }
4950                         pmap_remove_pv_page(pte_pv);
4951                         /* will either set pte_pv->pv_m or pv_free() later */
4952                 } else {
4953                         /*
4954                          * Old page was not managed.  If we have a pte_pv
4955                          * it better not have a pv_m assigned to it.  If the
4956                          * new page is managed the pte_pv will be destroyed
4957                          * near the end (we need its interlock).
4958                          *
4959                          * NOTE: We leave the wire count on the PT page
4960                          *       intact for the followup enter, but adjust
4961                          *       the wired-pages count on the pmap.
4962                          */
4963                         KKASSERT(pte_pv == NULL);
4964                         if (prot & VM_PROT_NOSYNC) {
4965                                 /*
4966                                  * NOSYNC (no mmu sync) requested.
4967                                  */
4968                                 (void)pte_load_clear(ptep);
4969                                 cpu_invlpg((void *)va);
4970                         } else {
4971                                 /*
4972                                  * Nominal SYNC
4973                                  */
4974                                 pmap_inval_smp(pmap, va, 1, ptep, 0);
4975                         }
4976
4977                         /*
4978                          * We must adjust pm_stats manually for unmanaged
4979                          * pages.
4980                          */
4981                         if (pt_pv) {
4982                                 atomic_add_long(&pmap->pm_stats.
4983                                                 resident_count, -1);
4984                         }
4985                         if (origpte & pmap->pmap_bits[PG_W_IDX]) {
4986                                 atomic_add_long(&pmap->pm_stats.
4987                                                 wired_count, -1);
4988                         }
4989                 }
4990                 KKASSERT(*ptep == 0);
4991         }
4992
4993 #ifdef PMAP_DEBUG2
4994         if (pmap_enter_debug > 0) {
4995                 --pmap_enter_debug;
4996                 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4997                         " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4998                         va, m,
4999                         origpte, newpte, ptep,
5000                         pte_pv, pt_pv, opa, prot);
5001         }
5002 #endif
5003
5004         if ((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5005                 /*
5006                  * Entering an unmanaged page.  We must wire the pt_pv unless
5007                  * we retained the wiring from an unmanaged page we had
5008                  * removed (if we retained it via pte_pv that will go away
5009                  * soon).
5010                  */
5011                 if (pt_pv && (opa == 0 ||
5012                               (origpte & pmap->pmap_bits[PG_MANAGED_IDX]))) {
5013                         vm_page_wire_quick(pt_pv->pv_m);
5014                 }
5015                 if (wired)
5016                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
5017
5018                 /*
5019                  * Unmanaged pages need manual resident_count tracking.
5020                  */
5021                 if (pt_pv) {
5022                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.
5023                                         resident_count, 1);
5024                 }
5025                 if (newpte & pmap->pmap_bits[PG_RW_IDX])
5026                         vm_page_flag_set(m, PG_WRITEABLE);
5027         } else {
5028                 /*
5029                  * Entering a managed page.  Our pte_pv takes care of the
5030                  * PT wiring, so if we had removed an unmanaged page before
5031                  * we must adjust.
5032                  *
5033                  * We have to take care of the pmap wired count ourselves.
5034                  *
5035                  * Enter on the PV list if part of our managed memory.
5036                  */
5037                 KKASSERT(pte_pv && (pte_pv->pv_m == NULL || pte_pv->pv_m == m));
5038                 vm_page_spin_lock(m);
5039                 pte_pv->pv_m = m;
5040                 pmap_page_stats_adding(m);
5041                 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
5042                 vm_page_flag_set(m, PG_MAPPED);
5043                 if (newpte & pmap->pmap_bits[PG_RW_IDX])
5044                         vm_page_flag_set(m, PG_WRITEABLE);
5045                 vm_page_spin_unlock(m);
5046
5047                 if (pt_pv && opa &&
5048                     (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5049                         vm_page_unwire_quick(pt_pv->pv_m);
5050                 }
5051
5052                 /*
5053                  * Adjust pmap wired pages count for new entry.
5054                  */
5055                 if (wired) {
5056                         atomic_add_long(&pte_pv->pv_pmap->pm_stats.
5057                                         wired_count, 1);
5058                 }
5059         }
5060
5061         /*
5062          * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
5063          *
5064          * User VMAs do not because those will be zero->non-zero, so no
5065          * stale entries to worry about at this point.
5066          *
5067          * For KVM there appear to still be issues.  Theoretically we
5068          * should be able to scrap the interlocks entirely but we
5069          * get crashes.
5070          */
5071         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
5072                 pmap_inval_smp(pmap, va, 1, ptep, newpte);
5073         } else {
5074                 origpte = atomic_swap_long(ptep, newpte);
5075                 if (origpte & pmap->pmap_bits[PG_M_IDX]) {
5076                         kprintf("pmap [M] race @ %016jx\n", va);
5077                         atomic_set_long(ptep, pmap->pmap_bits[PG_M_IDX]);
5078                 }
5079                 if (pt_pv == NULL)
5080                         cpu_invlpg((void *)va);
5081         }
5082
5083         /*
5084          * Cleanup
5085          */
5086 done:
5087         KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
5088                  (m->flags & PG_MAPPED));
5089
5090         /*
5091          * Cleanup the pv entry, allowing other accessors.  If the new page
5092          * is not managed but we have a pte_pv (which was locking our
5093          * operation), we can free it now.  pte_pv->pv_m should be NULL.
5094          */
5095         if (pte_pv && (newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5096                 pv_free(pte_pv, pt_pv);
5097         } else if (pte_pv) {
5098                 pv_put(pte_pv);
5099         } else if (pte_placemark) {
5100                 pv_placemarker_wakeup(pmap, pte_placemark);
5101         }
5102         if (pt_pv)
5103                 pv_put(pt_pv);
5104 }
5105
5106 /*
5107  * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
5108  * This code also assumes that the pmap has no pre-existing entry for this
5109  * VA.
5110  *
5111  * This code currently may only be used on user pmaps, not kernel_pmap.
5112  */
5113 void
5114 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
5115 {
5116         pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
5117 }
5118
5119 /*
5120  * Make a temporary mapping for a physical address.  This is only intended
5121  * to be used for panic dumps.
5122  *
5123  * The caller is responsible for calling smp_invltlb().
5124  */
5125 void *
5126 pmap_kenter_temporary(vm_paddr_t pa, long i)
5127 {
5128         pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
5129         return ((void *)crashdumpmap);
5130 }
5131
5132 #define MAX_INIT_PT (96)
5133
5134 /*
5135  * This routine preloads the ptes for a given object into the specified pmap.
5136  * This eliminates the blast of soft faults on process startup and
5137  * immediately after an mmap.
5138  */
5139 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
5140
5141 void
5142 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
5143                     vm_object_t object, vm_pindex_t pindex,
5144                     vm_size_t size, int limit)
5145 {
5146         struct rb_vm_page_scan_info info;
5147         struct lwp *lp;
5148         vm_size_t psize;
5149
5150         /*
5151          * We can't preinit if read access isn't set or there is no pmap
5152          * or object.
5153          */
5154         if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
5155                 return;
5156
5157         /*
5158          * We can't preinit if the pmap is not the current pmap
5159          */
5160         lp = curthread->td_lwp;
5161         if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
5162                 return;
5163
5164         /*
5165          * Misc additional checks
5166          */
5167         psize = x86_64_btop(size);
5168
5169         if ((object->type != OBJT_VNODE) ||
5170                 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
5171                         (object->resident_page_count > MAX_INIT_PT))) {
5172                 return;
5173         }
5174
5175         if (pindex + psize > object->size) {
5176                 if (object->size < pindex)
5177                         return;
5178                 psize = object->size - pindex;
5179         }
5180
5181         if (psize == 0)
5182                 return;
5183
5184         /*
5185          * If everything is segment-aligned do not pre-init here.  Instead
5186          * allow the normal vm_fault path to pass a segment hint to
5187          * pmap_enter() which will then use an object-referenced shared
5188          * page table page.
5189          */
5190         if ((addr & SEG_MASK) == 0 &&
5191             (ctob(psize) & SEG_MASK) == 0 &&
5192             (ctob(pindex) & SEG_MASK) == 0) {
5193                 return;
5194         }
5195
5196         /*
5197          * Use a red-black scan to traverse the requested range and load
5198          * any valid pages found into the pmap.
5199          *
5200          * We cannot safely scan the object's memq without holding the
5201          * object token.
5202          */
5203         info.start_pindex = pindex;
5204         info.end_pindex = pindex + psize - 1;
5205         info.limit = limit;
5206         info.mpte = NULL;
5207         info.addr = addr;
5208         info.pmap = pmap;
5209
5210         vm_object_hold_shared(object);
5211         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
5212                                 pmap_object_init_pt_callback, &info);
5213         vm_object_drop(object);
5214 }
5215
5216 static
5217 int
5218 pmap_object_init_pt_callback(vm_page_t p, void *data)
5219 {
5220         struct rb_vm_page_scan_info *info = data;
5221         vm_pindex_t rel_index;
5222
5223         /*
5224          * don't allow an madvise to blow away our really
5225          * free pages allocating pv entries.
5226          */
5227         if ((info->limit & MAP_PREFAULT_MADVISE) &&
5228                 vmstats.v_free_count < vmstats.v_free_reserved) {
5229                     return(-1);
5230         }
5231
5232         /*
5233          * Ignore list markers and ignore pages we cannot instantly
5234          * busy (while holding the object token).
5235          */
5236         if (p->flags & PG_MARKER)
5237                 return 0;
5238         if (vm_page_busy_try(p, TRUE))
5239                 return 0;
5240         if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
5241             (p->flags & PG_FICTITIOUS) == 0) {
5242                 if ((p->queue - p->pc) == PQ_CACHE)
5243                         vm_page_deactivate(p);
5244                 rel_index = p->pindex - info->start_pindex;
5245                 pmap_enter_quick(info->pmap,
5246                                  info->addr + x86_64_ptob(rel_index), p);
5247         }
5248         vm_page_wakeup(p);
5249         lwkt_yield();
5250         return(0);
5251 }
5252
5253 /*
5254  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
5255  * address.
5256  *
5257  * Returns FALSE if it would be non-trivial or if a pte is already loaded
5258  * into the slot.
5259  *
5260  * XXX This is safe only because page table pages are not freed.
5261  */
5262 int
5263 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
5264 {
5265         pt_entry_t *pte;
5266
5267         /*spin_lock(&pmap->pm_spin);*/
5268         if ((pte = pmap_pte(pmap, addr)) != NULL) {
5269                 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
5270                         /*spin_unlock(&pmap->pm_spin);*/
5271                         return FALSE;
5272                 }
5273         }
5274         /*spin_unlock(&pmap->pm_spin);*/
5275         return TRUE;
5276 }
5277
5278 /*
5279  * Change the wiring attribute for a pmap/va pair.  The mapping must already
5280  * exist in the pmap.  The mapping may or may not be managed.  The wiring in
5281  * the page is not changed, the page is returned so the caller can adjust
5282  * its wiring (the page is not locked in any way).
5283  *
5284  * Wiring is not a hardware characteristic so there is no need to invalidate
5285  * TLB.  However, in an SMP environment we must use a locked bus cycle to
5286  * update the pte (if we are not using the pmap_inval_*() API that is)...
5287  * it's ok to do this for simple wiring changes.
5288  */
5289 vm_page_t
5290 pmap_unwire(pmap_t pmap, vm_offset_t va)
5291 {
5292         pt_entry_t *ptep;
5293         pv_entry_t pt_pv;
5294         vm_paddr_t pa;
5295         vm_page_t m;
5296
5297         if (pmap == NULL)
5298                 return NULL;
5299
5300         /*
5301          * Assume elements in the kernel pmap are stable
5302          */
5303         if (pmap == &kernel_pmap) {
5304                 if (pmap_pt(pmap, va) == 0)
5305                         return NULL;
5306                 ptep = pmap_pte_quick(pmap, va);
5307                 if (pmap_pte_v(pmap, ptep)) {
5308                         if (pmap_pte_w(pmap, ptep))
5309                                 atomic_add_long(&pmap->pm_stats.wired_count,-1);
5310                         atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5311                         pa = *ptep & PG_FRAME;
5312                         m = PHYS_TO_VM_PAGE(pa);
5313                 } else {
5314                         m = NULL;
5315                 }
5316         } else {
5317                 /*
5318                  * We can only [un]wire pmap-local pages (we cannot wire
5319                  * shared pages)
5320                  */
5321                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
5322                 if (pt_pv == NULL)
5323                         return NULL;
5324
5325                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
5326                 if ((*ptep & pmap->pmap_bits[PG_V_IDX]) == 0) {
5327                         pv_put(pt_pv);
5328                         return NULL;
5329                 }
5330
5331                 if (pmap_pte_w(pmap, ptep)) {
5332                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.wired_count,
5333                                         -1);
5334                 }
5335                 /* XXX else return NULL so caller doesn't unwire m ? */
5336
5337                 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5338
5339                 pa = *ptep & PG_FRAME;
5340                 m = PHYS_TO_VM_PAGE(pa);        /* held by wired count */
5341                 pv_put(pt_pv);
5342         }
5343         return m;
5344 }
5345
5346 /*
5347  * Copy the range specified by src_addr/len from the source map to
5348  * the range dst_addr/len in the destination map.
5349  *
5350  * This routine is only advisory and need not do anything.
5351  */
5352 void
5353 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
5354           vm_size_t len, vm_offset_t src_addr)
5355 {
5356 }
5357
5358 /*
5359  * pmap_zero_page:
5360  *
5361  *      Zero the specified physical page.
5362  *
5363  *      This function may be called from an interrupt and no locking is
5364  *      required.
5365  */
5366 void
5367 pmap_zero_page(vm_paddr_t phys)
5368 {
5369         vm_offset_t va = PHYS_TO_DMAP(phys);
5370
5371         pagezero((void *)va);
5372 }
5373
5374 /*
5375  * pmap_zero_page:
5376  *
5377  *      Zero part of a physical page by mapping it into memory and clearing
5378  *      its contents with bzero.
5379  *
5380  *      off and size may not cover an area beyond a single hardware page.
5381  */
5382 void
5383 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
5384 {
5385         vm_offset_t virt = PHYS_TO_DMAP(phys);
5386
5387         bzero((char *)virt + off, size);
5388 }
5389
5390 /*
5391  * pmap_copy_page:
5392  *
5393  *      Copy the physical page from the source PA to the target PA.
5394  *      This function may be called from an interrupt.  No locking
5395  *      is required.
5396  */
5397 void
5398 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
5399 {
5400         vm_offset_t src_virt, dst_virt;
5401
5402         src_virt = PHYS_TO_DMAP(src);
5403         dst_virt = PHYS_TO_DMAP(dst);
5404         bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
5405 }
5406
5407 /*
5408  * pmap_copy_page_frag:
5409  *
5410  *      Copy the physical page from the source PA to the target PA.
5411  *      This function may be called from an interrupt.  No locking
5412  *      is required.
5413  */
5414 void
5415 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
5416 {
5417         vm_offset_t src_virt, dst_virt;
5418
5419         src_virt = PHYS_TO_DMAP(src);
5420         dst_virt = PHYS_TO_DMAP(dst);
5421
5422         bcopy((char *)src_virt + (src & PAGE_MASK),
5423               (char *)dst_virt + (dst & PAGE_MASK),
5424               bytes);
5425 }
5426
5427 /*
5428  * Returns true if the pmap's pv is one of the first 16 pvs linked to from
5429  * this page.  This count may be changed upwards or downwards in the future;
5430  * it is only necessary that true be returned for a small subset of pmaps
5431  * for proper page aging.
5432  */
5433 boolean_t
5434 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
5435 {
5436         pv_entry_t pv;
5437         int loops = 0;
5438
5439         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5440                 return FALSE;
5441
5442         vm_page_spin_lock(m);
5443         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5444                 if (pv->pv_pmap == pmap) {
5445                         vm_page_spin_unlock(m);
5446                         return TRUE;
5447                 }
5448                 loops++;
5449                 if (loops >= 16)
5450                         break;
5451         }
5452         vm_page_spin_unlock(m);
5453         return (FALSE);
5454 }
5455
5456 /*
5457  * Remove all pages from specified address space this aids process exit
5458  * speeds.  Also, this code may be special cased for the current process
5459  * only.
5460  */
5461 void
5462 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5463 {
5464         pmap_remove_noinval(pmap, sva, eva);
5465         cpu_invltlb();
5466 }
5467
5468 /*
5469  * pmap_testbit tests bits in pte's note that the testbit/clearbit
5470  * routines are inline, and a lot of things compile-time evaluate.
5471  */
5472 static
5473 boolean_t
5474 pmap_testbit(vm_page_t m, int bit)
5475 {
5476         pv_entry_t pv;
5477         pt_entry_t *pte;
5478         pmap_t pmap;
5479
5480         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5481                 return FALSE;
5482
5483         if (TAILQ_FIRST(&m->md.pv_list) == NULL)
5484                 return FALSE;
5485         vm_page_spin_lock(m);
5486         if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
5487                 vm_page_spin_unlock(m);
5488                 return FALSE;
5489         }
5490
5491         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5492
5493 #if defined(PMAP_DIAGNOSTIC)
5494                 if (pv->pv_pmap == NULL) {
5495                         kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
5496                             pv->pv_pindex);
5497                         continue;
5498                 }
5499 #endif
5500                 pmap = pv->pv_pmap;
5501
5502                 /*
5503                  * If the bit being tested is the modified bit, then
5504                  * mark clean_map and ptes as never
5505                  * modified.
5506                  *
5507                  * WARNING!  Because we do not lock the pv, *pte can be in a
5508                  *           state of flux.  Despite this the value of *pte
5509                  *           will still be related to the vm_page in some way
5510                  *           because the pv cannot be destroyed as long as we
5511                  *           hold the vm_page spin lock.
5512                  */
5513                 if (bit == PG_A_IDX || bit == PG_M_IDX) {
5514                                 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5515                         if (!pmap_track_modified(pv->pv_pindex))
5516                                 continue;
5517                 }
5518
5519                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5520                 if (*pte & pmap->pmap_bits[bit]) {
5521                         vm_page_spin_unlock(m);
5522                         return TRUE;
5523                 }
5524         }
5525         vm_page_spin_unlock(m);
5526         return (FALSE);
5527 }
5528
5529 /*
5530  * This routine is used to modify bits in ptes.  Only one bit should be
5531  * specified.  PG_RW requires special handling.
5532  *
5533  * Caller must NOT hold any spin locks
5534  */
5535 static __inline
5536 void
5537 pmap_clearbit(vm_page_t m, int bit_index)
5538 {
5539         pv_entry_t pv;
5540         pt_entry_t *pte;
5541         pt_entry_t pbits;
5542         pmap_t pmap;
5543
5544         if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5545                 if (bit_index == PG_RW_IDX)
5546                         vm_page_flag_clear(m, PG_WRITEABLE);
5547                 return;
5548         }
5549
5550         /*
5551          * PG_M or PG_A case
5552          *
5553          * Loop over all current mappings setting/clearing as appropos If
5554          * setting RO do we need to clear the VAC?
5555          *
5556          * NOTE: When clearing PG_M we could also (not implemented) drop
5557          *       through to the PG_RW code and clear PG_RW too, forcing
5558          *       a fault on write to redetect PG_M for virtual kernels, but
5559          *       it isn't necessary since virtual kernels invalidate the
5560          *       pte when they clear the VPTE_M bit in their virtual page
5561          *       tables.
5562          *
5563          * NOTE: Does not re-dirty the page when clearing only PG_M.
5564          *
5565          * NOTE: Because we do not lock the pv, *pte can be in a state of
5566          *       flux.  Despite this the value of *pte is still somewhat
5567          *       related while we hold the vm_page spin lock.
5568          *
5569          *       *pte can be zero due to this race.  Since we are clearing
5570          *       bits we basically do no harm when this race occurs.
5571          */
5572         if (bit_index != PG_RW_IDX) {
5573                 vm_page_spin_lock(m);
5574                 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5575 #if defined(PMAP_DIAGNOSTIC)
5576                         if (pv->pv_pmap == NULL) {
5577                                 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5578                                     pv->pv_pindex);
5579                                 continue;
5580                         }
5581 #endif
5582                         pmap = pv->pv_pmap;
5583                         pte = pmap_pte_quick(pv->pv_pmap,
5584                                              pv->pv_pindex << PAGE_SHIFT);
5585                         pbits = *pte;
5586                         if (pbits & pmap->pmap_bits[bit_index])
5587                                 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5588                 }
5589                 vm_page_spin_unlock(m);
5590                 return;
5591         }
5592
5593         /*
5594          * Clear PG_RW.  Also clears PG_M and marks the page dirty if PG_M
5595          * was set.
5596          */
5597 restart:
5598         vm_page_spin_lock(m);
5599         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5600                 /*
5601                  * don't write protect pager mappings
5602                  */
5603                 if (!pmap_track_modified(pv->pv_pindex))
5604                         continue;
5605
5606 #if defined(PMAP_DIAGNOSTIC)
5607                 if (pv->pv_pmap == NULL) {
5608                         kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5609                                 pv->pv_pindex);
5610                         continue;
5611                 }
5612 #endif
5613                 pmap = pv->pv_pmap;
5614
5615                 /*
5616                  * Skip pages which do not have PG_RW set.
5617                  */
5618                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5619                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5620                         continue;
5621
5622                 /*
5623                  * We must lock the PV to be able to safely test the pte.
5624                  */
5625                 if (pv_hold_try(pv)) {
5626                         vm_page_spin_unlock(m);
5627                 } else {
5628                         vm_page_spin_unlock(m);
5629                         pv_lock(pv);    /* held, now do a blocking lock */
5630                         pv_put(pv);
5631                         goto restart;
5632                 }
5633
5634                 /*
5635                  * Reload pte after acquiring pv.
5636                  */
5637                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5638 #if 0
5639                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0) {
5640                         pv_put(pv);
5641                         goto restart;
5642                 }
5643 #endif
5644
5645                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
5646                 for (;;) {
5647                         pt_entry_t nbits;
5648
5649                         pbits = *pte;
5650                         cpu_ccfence();
5651                         nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5652                                           pmap->pmap_bits[PG_M_IDX]);
5653                         if (pmap_inval_smp_cmpset(pmap,
5654                                      ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5655                                      pte, pbits, nbits)) {
5656                                 break;
5657                         }
5658                         cpu_pause();
5659                 }
5660
5661                 /*
5662                  * If PG_M was found to be set while we were clearing PG_RW
5663                  * we also clear PG_M (done above) and mark the page dirty.
5664                  * Callers expect this behavior.
5665                  *
5666                  * we lost pv so it cannot be used as an iterator.  In fact,
5667                  * because we couldn't necessarily lock it atomically it may
5668                  * have moved within the list and ALSO cannot be used as an
5669                  * iterator.
5670                  */
5671                 vm_page_spin_lock(m);
5672                 if (pbits & pmap->pmap_bits[PG_M_IDX])
5673                         vm_page_dirty(m);
5674                 vm_page_spin_unlock(m);
5675                 pv_put(pv);
5676                 goto restart;
5677         }
5678         if (bit_index == PG_RW_IDX)
5679                 vm_page_flag_clear(m, PG_WRITEABLE);
5680         vm_page_spin_unlock(m);
5681 }
5682
5683 /*
5684  * Lower the permission for all mappings to a given page.
5685  *
5686  * Page must be busied by caller.  Because page is busied by caller this
5687  * should not be able to race a pmap_enter().
5688  */
5689 void
5690 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5691 {
5692         /* JG NX support? */
5693         if ((prot & VM_PROT_WRITE) == 0) {
5694                 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5695                         /*
5696                          * NOTE: pmap_clearbit(.. PG_RW) also clears
5697                          *       the PG_WRITEABLE flag in (m).
5698                          */
5699                         pmap_clearbit(m, PG_RW_IDX);
5700                 } else {
5701                         pmap_remove_all(m);
5702                 }
5703         }
5704 }
5705
5706 vm_paddr_t
5707 pmap_phys_address(vm_pindex_t ppn)
5708 {
5709         return (x86_64_ptob(ppn));
5710 }
5711
5712 /*
5713  * Return a count of reference bits for a page, clearing those bits.
5714  * It is not necessary for every reference bit to be cleared, but it
5715  * is necessary that 0 only be returned when there are truly no
5716  * reference bits set.
5717  *
5718  * XXX: The exact number of bits to check and clear is a matter that
5719  * should be tested and standardized at some point in the future for
5720  * optimal aging of shared pages.
5721  *
5722  * This routine may not block.
5723  */
5724 int
5725 pmap_ts_referenced(vm_page_t m)
5726 {
5727         pv_entry_t pv;
5728         pt_entry_t *pte;
5729         pmap_t pmap;
5730         int rtval = 0;
5731
5732         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5733                 return (rtval);
5734
5735         vm_page_spin_lock(m);
5736         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5737                 if (!pmap_track_modified(pv->pv_pindex))
5738                         continue;
5739                 pmap = pv->pv_pmap;
5740                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5741                 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5742                         atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5743                         rtval++;
5744                         if (rtval > 4)
5745                                 break;
5746                 }
5747         }
5748         vm_page_spin_unlock(m);
5749         return (rtval);
5750 }
5751
5752 /*
5753  *      pmap_is_modified:
5754  *
5755  *      Return whether or not the specified physical page was modified
5756  *      in any physical maps.
5757  */
5758 boolean_t
5759 pmap_is_modified(vm_page_t m)
5760 {
5761         boolean_t res;
5762
5763         res = pmap_testbit(m, PG_M_IDX);
5764         return (res);
5765 }
5766
5767 /*
5768  *      Clear the modify bits on the specified physical page.
5769  */
5770 void
5771 pmap_clear_modify(vm_page_t m)
5772 {
5773         pmap_clearbit(m, PG_M_IDX);
5774 }
5775
5776 /*
5777  *      pmap_clear_reference:
5778  *
5779  *      Clear the reference bit on the specified physical page.
5780  */
5781 void
5782 pmap_clear_reference(vm_page_t m)
5783 {
5784         pmap_clearbit(m, PG_A_IDX);
5785 }
5786
5787 /*
5788  * Miscellaneous support routines follow
5789  */
5790
5791 static
5792 void
5793 i386_protection_init(void)
5794 {
5795         uint64_t *kp;
5796         int prot;
5797
5798         /*
5799          * NX supported? (boot time loader.conf override only)
5800          */
5801         TUNABLE_INT_FETCH("machdep.pmap_nx_enable", &pmap_nx_enable);
5802         if (pmap_nx_enable == 0 || (amd_feature & AMDID_NX) == 0)
5803                 pmap_bits_default[PG_NX_IDX] = 0;
5804
5805         /*
5806          * 0 is basically read-only access, but also set the NX (no-execute)
5807          * bit when VM_PROT_EXECUTE is not specified.
5808          */
5809         kp = protection_codes;
5810         for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5811                 switch (prot) {
5812                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5813                         /*
5814                          * This case handled elsewhere
5815                          */
5816                         *kp++ = 0;
5817                         break;
5818                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5819                         /*
5820                          * Read-only is 0|NX
5821                          */
5822                         *kp++ = pmap_bits_default[PG_NX_IDX];
5823                         break;
5824                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5825                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5826                         /*
5827                          * Execute requires read access
5828                          */
5829                         *kp++ = 0;
5830                         break;
5831                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5832                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5833                         /*
5834                          * Write without execute is RW|NX
5835                          */
5836                         *kp++ = pmap_bits_default[PG_RW_IDX] |
5837                                 pmap_bits_default[PG_NX_IDX];
5838                         break;
5839                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5840                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5841                         /*
5842                          * Write with execute is RW
5843                          */
5844                         *kp++ = pmap_bits_default[PG_RW_IDX];
5845                         break;
5846                 }
5847         }
5848 }
5849
5850 /*
5851  * Map a set of physical memory pages into the kernel virtual
5852  * address space. Return a pointer to where it is mapped. This
5853  * routine is intended to be used for mapping device memory,
5854  * NOT real memory.
5855  *
5856  * NOTE: We can't use pgeflag unless we invalidate the pages one at
5857  *       a time.
5858  *
5859  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5860  *       work whether the cpu supports PAT or not.  The remaining PAT
5861  *       attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5862  *       supports PAT.
5863  */
5864 void *
5865 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5866 {
5867         return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5868 }
5869
5870 void *
5871 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5872 {
5873         return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5874 }
5875
5876 void *
5877 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5878 {
5879         return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5880 }
5881
5882 /*
5883  * Map a set of physical memory pages into the kernel virtual
5884  * address space. Return a pointer to where it is mapped. This
5885  * routine is intended to be used for mapping device memory,
5886  * NOT real memory.
5887  */
5888 void *
5889 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5890 {
5891         vm_offset_t va, tmpva, offset;
5892         pt_entry_t *pte;
5893         vm_size_t tmpsize;
5894
5895         offset = pa & PAGE_MASK;
5896         size = roundup(offset + size, PAGE_SIZE);
5897
5898         va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5899         if (va == 0)
5900                 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5901
5902         pa = pa & ~PAGE_MASK;
5903         for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5904                 pte = vtopte(tmpva);
5905                 *pte = pa |
5906                     kernel_pmap.pmap_bits[PG_RW_IDX] |
5907                     kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5908                     kernel_pmap.pmap_cache_bits[mode];
5909                 tmpsize -= PAGE_SIZE;
5910                 tmpva += PAGE_SIZE;
5911                 pa += PAGE_SIZE;
5912         }
5913         pmap_invalidate_range(&kernel_pmap, va, va + size);
5914         pmap_invalidate_cache_range(va, va + size);
5915
5916         return ((void *)(va + offset));
5917 }
5918
5919 void
5920 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5921 {
5922         vm_offset_t base, offset;
5923
5924         base = va & ~PAGE_MASK;
5925         offset = va & PAGE_MASK;
5926         size = roundup(offset + size, PAGE_SIZE);
5927         pmap_qremove(va, size >> PAGE_SHIFT);
5928         kmem_free(&kernel_map, base, size);
5929 }
5930
5931 /*
5932  * Sets the memory attribute for the specified page.
5933  */
5934 void
5935 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5936 {
5937
5938     m->pat_mode = ma;
5939
5940     /*
5941      * If "m" is a normal page, update its direct mapping.  This update
5942      * can be relied upon to perform any cache operations that are
5943      * required for data coherence.
5944      */
5945     if ((m->flags & PG_FICTITIOUS) == 0)
5946         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5947 }
5948
5949 /*
5950  * Change the PAT attribute on an existing kernel memory map.  Caller
5951  * must ensure that the virtual memory in question is not accessed
5952  * during the adjustment.
5953  */
5954 void
5955 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5956 {
5957         pt_entry_t *pte;
5958         vm_offset_t base;
5959         int changed = 0;
5960
5961         if (va == 0)
5962                 panic("pmap_change_attr: va is NULL");
5963         base = trunc_page(va);
5964
5965         while (count) {
5966                 pte = vtopte(va);
5967                 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5968                        kernel_pmap.pmap_cache_bits[mode];
5969                 --count;
5970                 va += PAGE_SIZE;
5971         }
5972
5973         changed = 1;    /* XXX: not optimal */
5974
5975         /*
5976          * Flush CPU caches if required to make sure any data isn't cached that
5977          * shouldn't be, etc.
5978          */
5979         if (changed) {
5980                 pmap_invalidate_range(&kernel_pmap, base, va);
5981                 pmap_invalidate_cache_range(base, va);
5982         }
5983 }
5984
5985 /*
5986  * perform the pmap work for mincore
5987  */
5988 int
5989 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5990 {
5991         pt_entry_t *ptep, pte;
5992         vm_page_t m;
5993         int val = 0;
5994
5995         ptep = pmap_pte(pmap, addr);
5996
5997         if (ptep && (pte = *ptep) != 0) {
5998                 vm_offset_t pa;
5999
6000                 val = MINCORE_INCORE;
6001                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
6002                         goto done;
6003
6004                 pa = pte & PG_FRAME;
6005
6006                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
6007                         m = NULL;
6008                 else
6009                         m = PHYS_TO_VM_PAGE(pa);
6010
6011                 /*
6012                  * Modified by us
6013                  */
6014                 if (pte & pmap->pmap_bits[PG_M_IDX])
6015                         val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
6016                 /*
6017                  * Modified by someone
6018                  */
6019                 else if (m && (m->dirty || pmap_is_modified(m)))
6020                         val |= MINCORE_MODIFIED_OTHER;
6021                 /*
6022                  * Referenced by us
6023                  */
6024                 if (pte & pmap->pmap_bits[PG_A_IDX])
6025                         val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
6026
6027                 /*
6028                  * Referenced by someone
6029                  */
6030                 else if (m && ((m->flags & PG_REFERENCED) ||
6031                                 pmap_ts_referenced(m))) {
6032                         val |= MINCORE_REFERENCED_OTHER;
6033                         vm_page_flag_set(m, PG_REFERENCED);
6034                 }
6035         }
6036 done:
6037
6038         return val;
6039 }
6040
6041 /*
6042  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
6043  * vmspace will be ref'd and the old one will be deref'd.
6044  *
6045  * The vmspace for all lwps associated with the process will be adjusted
6046  * and cr3 will be reloaded if any lwp is the current lwp.
6047  *
6048  * The process must hold the vmspace->vm_map.token for oldvm and newvm
6049  */
6050 void
6051 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
6052 {
6053         struct vmspace *oldvm;
6054         struct lwp *lp;
6055
6056         oldvm = p->p_vmspace;
6057         if (oldvm != newvm) {
6058                 if (adjrefs)
6059                         vmspace_ref(newvm);
6060                 p->p_vmspace = newvm;
6061                 KKASSERT(p->p_nthreads == 1);
6062                 lp = RB_ROOT(&p->p_lwp_tree);
6063                 pmap_setlwpvm(lp, newvm);
6064                 if (adjrefs)
6065                         vmspace_rel(oldvm);
6066         }
6067 }
6068
6069 /*
6070  * Set the vmspace for a LWP.  The vmspace is almost universally set the
6071  * same as the process vmspace, but virtual kernels need to swap out contexts
6072  * on a per-lwp basis.
6073  *
6074  * Caller does not necessarily hold any vmspace tokens.  Caller must control
6075  * the lwp (typically be in the context of the lwp).  We use a critical
6076  * section to protect against statclock and hardclock (statistics collection).
6077  */
6078 void
6079 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
6080 {
6081         struct vmspace *oldvm;
6082         struct pmap *pmap;
6083
6084         oldvm = lp->lwp_vmspace;
6085
6086         if (oldvm != newvm) {
6087                 crit_enter();
6088                 KKASSERT((newvm->vm_refcnt & VM_REF_DELETED) == 0);
6089                 lp->lwp_vmspace = newvm;
6090                 if (curthread->td_lwp == lp) {
6091                         pmap = vmspace_pmap(newvm);
6092                         ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
6093                         if (pmap->pm_active_lock & CPULOCK_EXCL)
6094                                 pmap_interlock_wait(newvm);
6095 #if defined(SWTCH_OPTIM_STATS)
6096                         tlb_flush_count++;
6097 #endif
6098                         if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
6099                                 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
6100                         } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
6101                                 curthread->td_pcb->pcb_cr3 = KPML4phys;
6102                         } else {
6103                                 panic("pmap_setlwpvm: unknown pmap type\n");
6104                         }
6105                         load_cr3(curthread->td_pcb->pcb_cr3);
6106                         pmap = vmspace_pmap(oldvm);
6107                         ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
6108                                                mycpu->gd_cpuid);
6109                 }
6110                 crit_exit();
6111         }
6112 }
6113
6114 /*
6115  * Called when switching to a locked pmap, used to interlock against pmaps
6116  * undergoing modifications to prevent us from activating the MMU for the
6117  * target pmap until all such modifications have completed.  We have to do
6118  * this because the thread making the modifications has already set up its
6119  * SMP synchronization mask.
6120  *
6121  * This function cannot sleep!
6122  *
6123  * No requirements.
6124  */
6125 void
6126 pmap_interlock_wait(struct vmspace *vm)
6127 {
6128         struct pmap *pmap = &vm->vm_pmap;
6129
6130         if (pmap->pm_active_lock & CPULOCK_EXCL) {
6131                 crit_enter();
6132                 KKASSERT(curthread->td_critcount >= 2);
6133                 DEBUG_PUSH_INFO("pmap_interlock_wait");
6134                 while (pmap->pm_active_lock & CPULOCK_EXCL) {
6135                         cpu_ccfence();
6136                         lwkt_process_ipiq();
6137                 }
6138                 DEBUG_POP_INFO();
6139                 crit_exit();
6140         }
6141 }
6142
6143 vm_offset_t
6144 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
6145 {
6146
6147         if ((obj == NULL) || (size < NBPDR) ||
6148             ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
6149                 return addr;
6150         }
6151
6152         addr = roundup2(addr, NBPDR);
6153         return addr;
6154 }
6155
6156 /*
6157  * Used by kmalloc/kfree, page already exists at va
6158  */
6159 vm_page_t
6160 pmap_kvtom(vm_offset_t va)
6161 {
6162         pt_entry_t *ptep = vtopte(va);
6163
6164         KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
6165         return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
6166 }
6167
6168 /*
6169  * Initialize machine-specific shared page directory support.  This
6170  * is executed when a VM object is created.
6171  */
6172 void
6173 pmap_object_init(vm_object_t object)
6174 {
6175         object->md.pmap_rw = NULL;
6176         object->md.pmap_ro = NULL;
6177 }
6178
6179 /*
6180  * Clean up machine-specific shared page directory support.  This
6181  * is executed when a VM object is destroyed.
6182  */
6183 void
6184 pmap_object_free(vm_object_t object)
6185 {
6186         pmap_t pmap;
6187
6188         if ((pmap = object->md.pmap_rw) != NULL) {
6189                 object->md.pmap_rw = NULL;
6190                 pmap_remove_noinval(pmap,
6191                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6192                 CPUMASK_ASSZERO(pmap->pm_active);
6193                 pmap_release(pmap);
6194                 pmap_puninit(pmap);
6195                 kfree(pmap, M_OBJPMAP);
6196         }
6197         if ((pmap = object->md.pmap_ro) != NULL) {
6198                 object->md.pmap_ro = NULL;
6199                 pmap_remove_noinval(pmap,
6200                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6201                 CPUMASK_ASSZERO(pmap->pm_active);
6202                 pmap_release(pmap);
6203                 pmap_puninit(pmap);
6204                 kfree(pmap, M_OBJPMAP);
6205         }
6206 }
6207
6208 /*
6209  * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
6210  * VM page and issue a pginfo->callback.
6211  *
6212  * We are expected to dispose of any non-NULL pte_pv.
6213  */
6214 static
6215 void
6216 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
6217                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
6218                       pv_entry_t pt_pv, int sharept,
6219                       vm_offset_t va, pt_entry_t *ptep, void *arg)
6220 {
6221         struct pmap_pgscan_info *pginfo = arg;
6222         vm_page_t m;
6223
6224         if (pte_pv) {
6225                 /*
6226                  * Try to busy the page while we hold the pte_pv locked.
6227                  */
6228                 KKASSERT(pte_pv->pv_m);
6229                 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
6230                 if (vm_page_busy_try(m, TRUE) == 0) {
6231                         if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
6232                                 /*
6233                                  * The callback is issued with the pte_pv
6234                                  * unlocked and put away, and the pt_pv
6235                                  * unlocked.
6236                                  */
6237                                 pv_put(pte_pv);
6238                                 if (pt_pv) {
6239                                         vm_page_wire_quick(pt_pv->pv_m);
6240                                         pv_unlock(pt_pv);
6241                                 }
6242                                 if (pginfo->callback(pginfo, va, m) < 0)
6243                                         info->stop = 1;
6244                                 if (pt_pv) {
6245                                         pv_lock(pt_pv);
6246                                         vm_page_unwire_quick(pt_pv->pv_m);
6247                                 }
6248                         } else {
6249                                 vm_page_wakeup(m);
6250                                 pv_put(pte_pv);
6251                         }
6252                 } else {
6253                         ++pginfo->busycount;
6254                         pv_put(pte_pv);
6255                 }
6256         } else {
6257                 /*
6258                  * Shared page table or unmanaged page (sharept or !sharept)
6259                  */
6260                 pv_placemarker_wakeup(pmap, pte_placemark);
6261         }
6262 }
6263
6264 void
6265 pmap_pgscan(struct pmap_pgscan_info *pginfo)
6266 {
6267         struct pmap_scan_info info;
6268
6269         pginfo->offset = pginfo->beg_addr;
6270         info.pmap = pginfo->pmap;
6271         info.sva = pginfo->beg_addr;
6272         info.eva = pginfo->end_addr;
6273         info.func = pmap_pgscan_callback;
6274         info.arg = pginfo;
6275         pmap_scan(&info, 0);
6276         if (info.stop == 0)
6277                 pginfo->offset = pginfo->end_addr;
6278 }
6279
6280 /*
6281  * Wait for a placemarker that we do not own to clear.  The placemarker
6282  * in question is not necessarily set to the pindex we want, we may have
6283  * to wait on the element because we want to reserve it ourselves.
6284  *
6285  * NOTE: PM_PLACEMARK_WAKEUP sets a bit which is already set in
6286  *       PM_NOPLACEMARK, so it does not interfere with placemarks
6287  *       which have already been woken up.
6288  */
6289 static
6290 void
6291 pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark)
6292 {
6293         if (*pmark != PM_NOPLACEMARK) {
6294                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
6295                 tsleep_interlock(pmark, 0);
6296                 if (*pmark != PM_NOPLACEMARK)
6297                         tsleep(pmark, PINTERLOCKED, "pvplw", 0);
6298         }
6299 }
6300
6301 /*
6302  * Wakeup a placemarker that we own.  Replace the entry with
6303  * PM_NOPLACEMARK and issue a wakeup() if necessary.
6304  */
6305 static
6306 void
6307 pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark)
6308 {
6309         vm_pindex_t pindex;
6310
6311         pindex = atomic_swap_long(pmark, PM_NOPLACEMARK);
6312         KKASSERT(pindex != PM_NOPLACEMARK);
6313         if (pindex & PM_PLACEMARK_WAKEUP)
6314                 wakeup(pmark);
6315 }