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) 2011-2012 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)            _pv_get(pmap, pindex            \
 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 #else
 123
 124 #define PMAP_DEBUG_DECL
 125 #define PMAP_DEBUG_ARGS
 126 #define PMAP_DEBUG_COPY
 127
 128 #define pv_get(pmap, pindex)            _pv_get(pmap, pindex)
 129 #define pv_lock(pv)                     _pv_lock(pv)
 130 #define pv_hold_try(pv)                 _pv_hold_try(pv)
 131 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp)
 132
 133 #endif
 134
 135 /*
 136  * Get PDEs and PTEs for user/kernel address space
 137  */
 138 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
 139
 140 #define pmap_pde_v(pmap, pte)           ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 141 #define pmap_pte_w(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
 142 #define pmap_pte_m(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
 143 #define pmap_pte_u(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
 144 #define pmap_pte_v(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 145
 146 /*
 147  * Given a map and a machine independent protection code,
 148  * convert to a vax protection code.
 149  */
 150 #define pte_prot(m, p)          \
 151         (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
 152 static int protection_codes[PROTECTION_CODES_SIZE];
 153
 154 struct pmap kernel_pmap;
 155 static TAILQ_HEAD(,pmap)        pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
 156
 157 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
 158
 159 vm_paddr_t avail_start;         /* PA of first available physical page */
 160 vm_paddr_t avail_end;           /* PA of last available physical page */
 161 vm_offset_t virtual2_start;     /* cutout free area prior to kernel start */
 162 vm_offset_t virtual2_end;
 163 vm_offset_t virtual_start;      /* VA of first avail page (after kernel bss) */
 164 vm_offset_t virtual_end;        /* VA of last avail page (end of kernel AS) */
 165 vm_offset_t KvaStart;           /* VA start of KVA space */
 166 vm_offset_t KvaEnd;             /* VA end of KVA space (non-inclusive) */
 167 vm_offset_t KvaSize;            /* max size of kernel virtual address space */
 168 static boolean_t pmap_initialized = FALSE;      /* Has pmap_init completed? */
 169 //static int pgeflag;           /* PG_G or-in */
 170 //static int pseflag;           /* PG_PS or-in */
 171 uint64_t PatMsr;
 172
 173 static int ndmpdp;
 174 static vm_paddr_t dmaplimit;
 175 static int nkpt;
 176 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
 177
 178 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE];        /* PAT -> PG_ bits */
 179 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/    /* PAT -> PG_ bits */
 180
 181 static uint64_t KPTbase;
 182 static uint64_t KPTphys;
 183 static uint64_t KPDphys;        /* phys addr of kernel level 2 */
 184 static uint64_t KPDbase;        /* phys addr of kernel level 2 @ KERNBASE */
 185 uint64_t KPDPphys;      /* phys addr of kernel level 3 */
 186 uint64_t KPML4phys;     /* phys addr of kernel level 4 */
 187
 188 static uint64_t DMPDphys;       /* phys addr of direct mapped level 2 */
 189 static uint64_t DMPDPphys;      /* phys addr of direct mapped level 3 */
 190
 191 /*
 192  * Data for the pv entry allocation mechanism
 193  */
 194 static vm_zone_t pvzone;
 195 static struct vm_zone pvzone_store;
 196 static struct vm_object pvzone_obj;
 197 static int pv_entry_max=0, pv_entry_high_water=0;
 198 static int pmap_pagedaemon_waken = 0;
 199 static struct pv_entry *pvinit;
 200
 201 /*
 202  * All those kernel PT submaps that BSD is so fond of
 203  */
 204 pt_entry_t *CMAP1 = NULL, *ptmmap;
 205 caddr_t CADDR1 = NULL, ptvmmap = NULL;
 206 static pt_entry_t *msgbufmap;
 207 struct msgbuf *msgbufp=NULL;
 208
 209 /*
 210  * PMAP default PG_* bits. Needed to be able to add
 211  * EPT/NPT pagetable pmap_bits for the VMM module
 212  */
 213 uint64_t pmap_bits_default[] = {
 214                 REGULAR_PMAP,                                   /* TYPE_IDX             0 */
 215                 X86_PG_V,                                       /* PG_V_IDX             1 */
 216                 X86_PG_RW,                                      /* PG_RW_IDX            2 */
 217                 X86_PG_U,                                       /* PG_U_IDX             3 */
 218                 X86_PG_A,                                       /* PG_A_IDX             4 */
 219                 X86_PG_M,                                       /* PG_M_IDX             5 */
 220                 X86_PG_PS,                                      /* PG_PS_IDX3           6 */
 221                 X86_PG_G,                                       /* PG_G_IDX             7 */
 222                 X86_PG_AVAIL1,                                  /* PG_AVAIL1_IDX        8 */
 223                 X86_PG_AVAIL2,                                  /* PG_AVAIL2_IDX        9 */
 224                 X86_PG_AVAIL3,                                  /* PG_AVAIL3_IDX        10 */
 225                 X86_PG_NC_PWT | X86_PG_NC_PCD,                  /* PG_N_IDX     11 */
 226 };
 227 /*
 228  * Crashdump maps.
 229  */
 230 static pt_entry_t *pt_crashdumpmap;
 231 static caddr_t crashdumpmap;
 232
 233 static int pmap_debug = 0;
 234 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
 235     &pmap_debug, 0, "Debug pmap's");
 236 #ifdef PMAP_DEBUG2
 237 static int pmap_enter_debug = 0;
 238 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
 239     &pmap_enter_debug, 0, "Debug pmap_enter's");
 240 #endif
 241 static int pmap_yield_count = 64;
 242 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
 243     &pmap_yield_count, 0, "Yield during init_pt/release");
 244 static int pmap_mmu_optimize = 0;
 245 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
 246     &pmap_mmu_optimize, 0, "Share page table pages when possible");
 247 int pmap_fast_kernel_cpusync = 0;
 248 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
 249     &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
 250
 251 #define DISABLE_PSE
 252
 253 /* Standard user access funtions */
 254 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
 255     size_t *lencopied);
 256 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
 257 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
 258 extern int std_fubyte (const void *base);
 259 extern int std_subyte (void *base, int byte);
 260 extern long std_fuword (const void *base);
 261 extern int std_suword (void *base, long word);
 262 extern int std_suword32 (void *base, int word);
 263
 264 static void pv_hold(pv_entry_t pv);
 265 static int _pv_hold_try(pv_entry_t pv
 266                                 PMAP_DEBUG_DECL);
 267 static void pv_drop(pv_entry_t pv);
 268 static void _pv_lock(pv_entry_t pv
 269                                 PMAP_DEBUG_DECL);
 270 static void pv_unlock(pv_entry_t pv);
 271 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
 272                                 PMAP_DEBUG_DECL);
 273 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
 274                                 PMAP_DEBUG_DECL);
 275 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
 276 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
 277 static void pv_put(pv_entry_t pv);
 278 static void pv_free(pv_entry_t pv, pv_entry_t pvp, int putaway);
 279 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
 280 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
 281                       pv_entry_t *pvpp);
 282 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
 283                       pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
 284 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
 285                         pmap_inval_bulk_t *bulk, int destroy);
 286 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
 287 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
 288                         pmap_inval_bulk_t *bulk);
 289
 290 struct pmap_scan_info;
 291 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
 292                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
 293                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 294 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
 295                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
 296                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 297
 298 static void i386_protection_init (void);
 299 static void create_pagetables(vm_paddr_t *firstaddr);
 300 static void pmap_remove_all (vm_page_t m);
 301 static boolean_t pmap_testbit (vm_page_t m, int bit);
 302
 303 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
 304 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
 305
 306 static void pmap_pinit_defaults(struct pmap *pmap);
 307
 308 static unsigned pdir4mb;
 309
 310 static int
 311 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
 312 {
 313         if (pv1->pv_pindex < pv2->pv_pindex)
 314                 return(-1);
 315         if (pv1->pv_pindex > pv2->pv_pindex)
 316                 return(1);
 317         return(0);
 318 }
 319
 320 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
 321              pv_entry_compare, vm_pindex_t, pv_pindex);
 322
 323 static __inline
 324 void
 325 pmap_page_stats_adding(vm_page_t m)
 326 {
 327         globaldata_t gd = mycpu;
 328
 329         if (TAILQ_EMPTY(&m->md.pv_list)) {
 330                 ++gd->gd_vmtotal.t_arm;
 331         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 332                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 333                 ++gd->gd_vmtotal.t_armshr;
 334                 ++gd->gd_vmtotal.t_avmshr;
 335         } else {
 336                 ++gd->gd_vmtotal.t_avmshr;
 337         }
 338 }
 339
 340 static __inline
 341 void
 342 pmap_page_stats_deleting(vm_page_t m)
 343 {
 344         globaldata_t gd = mycpu;
 345
 346         if (TAILQ_EMPTY(&m->md.pv_list)) {
 347                 --gd->gd_vmtotal.t_arm;
 348         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 349                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 350                 --gd->gd_vmtotal.t_armshr;
 351                 --gd->gd_vmtotal.t_avmshr;
 352         } else {
 353                 --gd->gd_vmtotal.t_avmshr;
 354         }
 355 }
 356
 357 /*
 358  * Move the kernel virtual free pointer to the next
 359  * 2MB.  This is used to help improve performance
 360  * by using a large (2MB) page for much of the kernel
 361  * (.text, .data, .bss)
 362  */
 363 static
 364 vm_offset_t
 365 pmap_kmem_choose(vm_offset_t addr)
 366 {
 367         vm_offset_t newaddr = addr;
 368
 369         newaddr = roundup2(addr, NBPDR);
 370         return newaddr;
 371 }
 372
 373 /*
 374  * pmap_pte_quick:
 375  *
 376  *      Super fast pmap_pte routine best used when scanning the pv lists.
 377  *      This eliminates many course-grained invltlb calls.  Note that many of
 378  *      the pv list scans are across different pmaps and it is very wasteful
 379  *      to do an entire invltlb when checking a single mapping.
 380  */
 381 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
 382
 383 static
 384 pt_entry_t *
 385 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
 386 {
 387         return pmap_pte(pmap, va);
 388 }
 389
 390 /*
 391  * Returns the pindex of a page table entry (representing a terminal page).
 392  * There are NUPTE_TOTAL page table entries possible (a huge number)
 393  *
 394  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
 395  * We want to properly translate negative KVAs.
 396  */
 397 static __inline
 398 vm_pindex_t
 399 pmap_pte_pindex(vm_offset_t va)
 400 {
 401         return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
 402 }
 403
 404 /*
 405  * Returns the pindex of a page table.
 406  */
 407 static __inline
 408 vm_pindex_t
 409 pmap_pt_pindex(vm_offset_t va)
 410 {
 411         return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
 412 }
 413
 414 /*
 415  * Returns the pindex of a page directory.
 416  */
 417 static __inline
 418 vm_pindex_t
 419 pmap_pd_pindex(vm_offset_t va)
 420 {
 421         return (NUPTE_TOTAL + NUPT_TOTAL +
 422                 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
 423 }
 424
 425 static __inline
 426 vm_pindex_t
 427 pmap_pdp_pindex(vm_offset_t va)
 428 {
 429         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
 430                 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
 431 }
 432
 433 static __inline
 434 vm_pindex_t
 435 pmap_pml4_pindex(void)
 436 {
 437         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
 438 }
 439
 440 /*
 441  * Return various clipped indexes for a given VA
 442  *
 443  * Returns the index of a pt in a page directory, representing a page
 444  * table.
 445  */
 446 static __inline
 447 vm_pindex_t
 448 pmap_pt_index(vm_offset_t va)
 449 {
 450         return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
 451 }
 452
 453 /*
 454  * Returns the index of a pd in a page directory page, representing a page
 455  * directory.
 456  */
 457 static __inline
 458 vm_pindex_t
 459 pmap_pd_index(vm_offset_t va)
 460 {
 461         return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
 462 }
 463
 464 /*
 465  * Returns the index of a pdp in the pml4 table, representing a page
 466  * directory page.
 467  */
 468 static __inline
 469 vm_pindex_t
 470 pmap_pdp_index(vm_offset_t va)
 471 {
 472         return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
 473 }
 474
 475 /*
 476  * Generic procedure to index a pte from a pt, pd, or pdp.
 477  *
 478  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
 479  *       a page table page index but is instead of PV lookup index.
 480  */
 481 static
 482 void *
 483 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
 484 {
 485         pt_entry_t *pte;
 486
 487         pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
 488         return(&pte[pindex]);
 489 }
 490
 491 /*
 492  * Return pointer to PDP slot in the PML4
 493  */
 494 static __inline
 495 pml4_entry_t *
 496 pmap_pdp(pmap_t pmap, vm_offset_t va)
 497 {
 498         return (&pmap->pm_pml4[pmap_pdp_index(va)]);
 499 }
 500
 501 /*
 502  * Return pointer to PD slot in the PDP given a pointer to the PDP
 503  */
 504 static __inline
 505 pdp_entry_t *
 506 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
 507 {
 508         pdp_entry_t *pd;
 509
 510         pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
 511         return (&pd[pmap_pd_index(va)]);
 512 }
 513
 514 /*
 515  * Return pointer to PD slot in the PDP.
 516  */
 517 static __inline
 518 pdp_entry_t *
 519 pmap_pd(pmap_t pmap, vm_offset_t va)
 520 {
 521         pml4_entry_t *pdp;
 522
 523         pdp = pmap_pdp(pmap, va);
 524         if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
 525                 return NULL;
 526         return (pmap_pdp_to_pd(*pdp, va));
 527 }
 528
 529 /*
 530  * Return pointer to PT slot in the PD given a pointer to the PD
 531  */
 532 static __inline
 533 pd_entry_t *
 534 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
 535 {
 536         pd_entry_t *pt;
 537
 538         pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
 539         return (&pt[pmap_pt_index(va)]);
 540 }
 541
 542 /*
 543  * Return pointer to PT slot in the PD
 544  *
 545  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
 546  *                   so we cannot lookup the PD via the PDP.  Instead we
 547  *                   must look it up via the pmap.
 548  */
 549 static __inline
 550 pd_entry_t *
 551 pmap_pt(pmap_t pmap, vm_offset_t va)
 552 {
 553         pdp_entry_t *pd;
 554         pv_entry_t pv;
 555         vm_pindex_t pd_pindex;
 556
 557         if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
 558                 pd_pindex = pmap_pd_pindex(va);
 559                 spin_lock(&pmap->pm_spin);
 560                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
 561                 spin_unlock(&pmap->pm_spin);
 562                 if (pv == NULL || pv->pv_m == NULL)
 563                         return NULL;
 564                 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
 565         } else {
 566                 pd = pmap_pd(pmap, va);
 567                 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
 568                          return NULL;
 569                 return (pmap_pd_to_pt(*pd, va));
 570         }
 571 }
 572
 573 /*
 574  * Return pointer to PTE slot in the PT given a pointer to the PT
 575  */
 576 static __inline
 577 pt_entry_t *
 578 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
 579 {
 580         pt_entry_t *pte;
 581
 582         pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
 583         return (&pte[pmap_pte_index(va)]);
 584 }
 585
 586 /*
 587  * Return pointer to PTE slot in the PT
 588  */
 589 static __inline
 590 pt_entry_t *
 591 pmap_pte(pmap_t pmap, vm_offset_t va)
 592 {
 593         pd_entry_t *pt;
 594
 595         pt = pmap_pt(pmap, va);
 596         if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
 597                  return NULL;
 598         if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
 599                 return ((pt_entry_t *)pt);
 600         return (pmap_pt_to_pte(*pt, va));
 601 }
 602
 603 /*
 604  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
 605  * the PT layer.  This will speed up core pmap operations considerably.
 606  *
 607  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
 608  *       must be in a known associated state (typically by being locked when
 609  *       the pmap spinlock isn't held).  We allow the race for that case.
 610  */
 611 static __inline
 612 void
 613 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
 614 {
 615         if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
 616                 pv->pv_pmap->pm_pvhint = pv;
 617 }
 618
 619
 620 /*
 621  * Return address of PT slot in PD (KVM only)
 622  *
 623  * Cannot be used for user page tables because it might interfere with
 624  * the shared page-table-page optimization (pmap_mmu_optimize).
 625  */
 626 static __inline
 627 pd_entry_t *
 628 vtopt(vm_offset_t va)
 629 {
 630         uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
 631                                   NPML4EPGSHIFT)) - 1);
 632
 633         return (PDmap + ((va >> PDRSHIFT) & mask));
 634 }
 635
 636 /*
 637  * KVM - return address of PTE slot in PT
 638  */
 639 static __inline
 640 pt_entry_t *
 641 vtopte(vm_offset_t va)
 642 {
 643         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 644                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 645
 646         return (PTmap + ((va >> PAGE_SHIFT) & mask));
 647 }
 648
 649 static uint64_t
 650 allocpages(vm_paddr_t *firstaddr, long n)
 651 {
 652         uint64_t ret;
 653
 654         ret = *firstaddr;
 655         bzero((void *)ret, n * PAGE_SIZE);
 656         *firstaddr += n * PAGE_SIZE;
 657         return (ret);
 658 }
 659
 660 static
 661 void
 662 create_pagetables(vm_paddr_t *firstaddr)
 663 {
 664         long i;         /* must be 64 bits */
 665         long nkpt_base;
 666         long nkpt_phys;
 667         int j;
 668
 669         /*
 670          * We are running (mostly) V=P at this point
 671          *
 672          * Calculate NKPT - number of kernel page tables.  We have to
 673          * accomodoate prealloction of the vm_page_array, dump bitmap,
 674          * MSGBUF_SIZE, and other stuff.  Be generous.
 675          *
 676          * Maxmem is in pages.
 677          *
 678          * ndmpdp is the number of 1GB pages we wish to map.
 679          */
 680         ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
 681         if (ndmpdp < 4)         /* Minimum 4GB of dirmap */
 682                 ndmpdp = 4;
 683         KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
 684
 685         /*
 686          * Starting at the beginning of kvm (not KERNBASE).
 687          */
 688         nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
 689         nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
 690         nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
 691                        ndmpdp) + 511) / 512;
 692         nkpt_phys += 128;
 693
 694         /*
 695          * Starting at KERNBASE - map 2G worth of page table pages.
 696          * KERNBASE is offset -2G from the end of kvm.
 697          */
 698         nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
 699
 700         /*
 701          * Allocate pages
 702          */
 703         KPTbase = allocpages(firstaddr, nkpt_base);
 704         KPTphys = allocpages(firstaddr, nkpt_phys);
 705         KPML4phys = allocpages(firstaddr, 1);
 706         KPDPphys = allocpages(firstaddr, NKPML4E);
 707         KPDphys = allocpages(firstaddr, NKPDPE);
 708
 709         /*
 710          * Calculate the page directory base for KERNBASE,
 711          * that is where we start populating the page table pages.
 712          * Basically this is the end - 2.
 713          */
 714         KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
 715
 716         DMPDPphys = allocpages(firstaddr, NDMPML4E);
 717         if ((amd_feature & AMDID_PAGE1GB) == 0)
 718                 DMPDphys = allocpages(firstaddr, ndmpdp);
 719         dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
 720
 721         /*
 722          * Fill in the underlying page table pages for the area around
 723          * KERNBASE.  This remaps low physical memory to KERNBASE.
 724          *
 725          * Read-only from zero to physfree
 726          * XXX not fully used, underneath 2M pages
 727          */
 728         for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
 729                 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
 730                 ((pt_entry_t *)KPTbase)[i] |=
 731                     pmap_bits_default[PG_RW_IDX] |
 732                     pmap_bits_default[PG_V_IDX] |
 733                     pmap_bits_default[PG_G_IDX];
 734         }
 735
 736         /*
 737          * Now map the initial kernel page tables.  One block of page
 738          * tables is placed at the beginning of kernel virtual memory,
 739          * and another block is placed at KERNBASE to map the kernel binary,
 740          * data, bss, and initial pre-allocations.
 741          */
 742         for (i = 0; i < nkpt_base; i++) {
 743                 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
 744                 ((pd_entry_t *)KPDbase)[i] |=
 745                     pmap_bits_default[PG_RW_IDX] |
 746                     pmap_bits_default[PG_V_IDX];
 747         }
 748         for (i = 0; i < nkpt_phys; i++) {
 749                 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
 750                 ((pd_entry_t *)KPDphys)[i] |=
 751                     pmap_bits_default[PG_RW_IDX] |
 752                     pmap_bits_default[PG_V_IDX];
 753         }
 754
 755         /*
 756          * Map from zero to end of allocations using 2M pages as an
 757          * optimization.  This will bypass some of the KPTBase pages
 758          * above in the KERNBASE area.
 759          */
 760         for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
 761                 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
 762                 ((pd_entry_t *)KPDbase)[i] |=
 763                     pmap_bits_default[PG_RW_IDX] |
 764                     pmap_bits_default[PG_V_IDX] |
 765                     pmap_bits_default[PG_PS_IDX] |
 766                     pmap_bits_default[PG_G_IDX];
 767         }
 768
 769         /*
 770          * And connect up the PD to the PDP.  The kernel pmap is expected
 771          * to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
 772          */
 773         for (i = 0; i < NKPDPE; i++) {
 774                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
 775                                 KPDphys + (i << PAGE_SHIFT);
 776                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
 777                     pmap_bits_default[PG_RW_IDX] |
 778                     pmap_bits_default[PG_V_IDX] |
 779                     pmap_bits_default[PG_U_IDX];
 780         }
 781
 782         /*
 783          * Now set up the direct map space using either 2MB or 1GB pages
 784          * Preset PG_M and PG_A because demotion expects it.
 785          *
 786          * When filling in entries in the PD pages make sure any excess
 787          * entries are set to zero as we allocated enough PD pages
 788          */
 789         if ((amd_feature & AMDID_PAGE1GB) == 0) {
 790                 for (i = 0; i < NPDEPG * ndmpdp; i++) {
 791                         ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
 792                         ((pd_entry_t *)DMPDphys)[i] |=
 793                             pmap_bits_default[PG_RW_IDX] |
 794                             pmap_bits_default[PG_V_IDX] |
 795                             pmap_bits_default[PG_PS_IDX] |
 796                             pmap_bits_default[PG_G_IDX] |
 797                             pmap_bits_default[PG_M_IDX] |
 798                             pmap_bits_default[PG_A_IDX];
 799                 }
 800
 801                 /*
 802                  * And the direct map space's PDP
 803                  */
 804                 for (i = 0; i < ndmpdp; i++) {
 805                         ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
 806                                                         (i << PAGE_SHIFT);
 807                         ((pdp_entry_t *)DMPDPphys)[i] |=
 808                             pmap_bits_default[PG_RW_IDX] |
 809                             pmap_bits_default[PG_V_IDX] |
 810                             pmap_bits_default[PG_U_IDX];
 811                 }
 812         } else {
 813                 for (i = 0; i < ndmpdp; i++) {
 814                         ((pdp_entry_t *)DMPDPphys)[i] =
 815                                                 (vm_paddr_t)i << PDPSHIFT;
 816                         ((pdp_entry_t *)DMPDPphys)[i] |=
 817                             pmap_bits_default[PG_RW_IDX] |
 818                             pmap_bits_default[PG_V_IDX] |
 819                             pmap_bits_default[PG_PS_IDX] |
 820                             pmap_bits_default[PG_G_IDX] |
 821                             pmap_bits_default[PG_M_IDX] |
 822                             pmap_bits_default[PG_A_IDX];
 823                 }
 824         }
 825
 826         /* And recursively map PML4 to itself in order to get PTmap */
 827         ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
 828         ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
 829             pmap_bits_default[PG_RW_IDX] |
 830             pmap_bits_default[PG_V_IDX] |
 831             pmap_bits_default[PG_U_IDX];
 832
 833         /*
 834          * Connect the Direct Map slots up to the PML4
 835          */
 836         for (j = 0; j < NDMPML4E; ++j) {
 837                 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
 838                     (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
 839                     pmap_bits_default[PG_RW_IDX] |
 840                     pmap_bits_default[PG_V_IDX] |
 841                     pmap_bits_default[PG_U_IDX];
 842         }
 843
 844         /*
 845          * Connect the KVA slot up to the PML4
 846          */
 847         ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
 848         ((pdp_entry_t *)KPML4phys)[KPML4I] |=
 849             pmap_bits_default[PG_RW_IDX] |
 850             pmap_bits_default[PG_V_IDX] |
 851             pmap_bits_default[PG_U_IDX];
 852 }
 853
 854 /*
 855  *      Bootstrap the system enough to run with virtual memory.
 856  *
 857  *      On the i386 this is called after mapping has already been enabled
 858  *      and just syncs the pmap module with what has already been done.
 859  *      [We can't call it easily with mapping off since the kernel is not
 860  *      mapped with PA == VA, hence we would have to relocate every address
 861  *      from the linked base (virtual) address "KERNBASE" to the actual
 862  *      (physical) address starting relative to 0]
 863  */
 864 void
 865 pmap_bootstrap(vm_paddr_t *firstaddr)
 866 {
 867         vm_offset_t va;
 868         pt_entry_t *pte;
 869
 870         KvaStart = VM_MIN_KERNEL_ADDRESS;
 871         KvaEnd = VM_MAX_KERNEL_ADDRESS;
 872         KvaSize = KvaEnd - KvaStart;
 873
 874         avail_start = *firstaddr;
 875
 876         /*
 877          * Create an initial set of page tables to run the kernel in.
 878          */
 879         create_pagetables(firstaddr);
 880
 881         virtual2_start = KvaStart;
 882         virtual2_end = PTOV_OFFSET;
 883
 884         virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
 885         virtual_start = pmap_kmem_choose(virtual_start);
 886
 887         virtual_end = VM_MAX_KERNEL_ADDRESS;
 888
 889         /* XXX do %cr0 as well */
 890         load_cr4(rcr4() | CR4_PGE | CR4_PSE);
 891         load_cr3(KPML4phys);
 892
 893         /*
 894          * Initialize protection array.
 895          */
 896         i386_protection_init();
 897
 898         /*
 899          * The kernel's pmap is statically allocated so we don't have to use
 900          * pmap_create, which is unlikely to work correctly at this part of
 901          * the boot sequence (XXX and which no longer exists).
 902          */
 903         kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
 904         kernel_pmap.pm_count = 1;
 905         CPUMASK_ASSALLONES(kernel_pmap.pm_active);
 906         RB_INIT(&kernel_pmap.pm_pvroot);
 907         spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
 908         lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
 909
 910         /*
 911          * Reserve some special page table entries/VA space for temporary
 912          * mapping of pages.
 913          */
 914 #define SYSMAP(c, p, v, n)      \
 915         v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
 916
 917         va = virtual_start;
 918         pte = vtopte(va);
 919
 920         /*
 921          * CMAP1/CMAP2 are used for zeroing and copying pages.
 922          */
 923         SYSMAP(caddr_t, CMAP1, CADDR1, 1)
 924
 925         /*
 926          * Crashdump maps.
 927          */
 928         SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
 929
 930         /*
 931          * ptvmmap is used for reading arbitrary physical pages via
 932          * /dev/mem.
 933          */
 934         SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
 935
 936         /*
 937          * msgbufp is used to map the system message buffer.
 938          * XXX msgbufmap is not used.
 939          */
 940         SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
 941                atop(round_page(MSGBUF_SIZE)))
 942
 943         virtual_start = va;
 944         virtual_start = pmap_kmem_choose(virtual_start);
 945
 946         *CMAP1 = 0;
 947
 948         /*
 949          * PG_G is terribly broken on SMP because we IPI invltlb's in some
 950          * cases rather then invl1pg.  Actually, I don't even know why it
 951          * works under UP because self-referential page table mappings
 952          */
 953 //      pgeflag = 0;
 954
 955 /*
 956  * Initialize the 4MB page size flag
 957  */
 958 //      pseflag = 0;
 959 /*
 960  * The 4MB page version of the initial
 961  * kernel page mapping.
 962  */
 963         pdir4mb = 0;
 964
 965 #if !defined(DISABLE_PSE)
 966         if (cpu_feature & CPUID_PSE) {
 967                 pt_entry_t ptditmp;
 968                 /*
 969                  * Note that we have enabled PSE mode
 970                  */
 971 //              pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
 972                 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
 973                 ptditmp &= ~(NBPDR - 1);
 974                 ptditmp |= pmap_bits_default[PG_V_IDX] |
 975                     pmap_bits_default[PG_RW_IDX] |
 976                     pmap_bits_default[PG_PS_IDX] |
 977                     pmap_bits_default[PG_U_IDX];
 978 //                  pgeflag;
 979                 pdir4mb = ptditmp;
 980         }
 981 #endif
 982         cpu_invltlb();
 983
 984         /* Initialize the PAT MSR */
 985         pmap_init_pat();
 986         pmap_pinit_defaults(&kernel_pmap);
 987
 988         TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
 989                           &pmap_fast_kernel_cpusync);
 990
 991 }
 992
 993 /*
 994  * Setup the PAT MSR.
 995  */
 996 void
 997 pmap_init_pat(void)
 998 {
 999         uint64_t pat_msr;
1000         u_long cr0, cr4;
1001
1002         /*
1003          * Default values mapping PATi,PCD,PWT bits at system reset.
1004          * The default values effectively ignore the PATi bit by
1005          * repeating the encodings for 0-3 in 4-7, and map the PCD
1006          * and PWT bit combinations to the expected PAT types.
1007          */
1008         pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |        /* 000 */
1009                   PAT_VALUE(1, PAT_WRITE_THROUGH) |     /* 001 */
1010                   PAT_VALUE(2, PAT_UNCACHED) |          /* 010 */
1011                   PAT_VALUE(3, PAT_UNCACHEABLE) |       /* 011 */
1012                   PAT_VALUE(4, PAT_WRITE_BACK) |        /* 100 */
1013                   PAT_VALUE(5, PAT_WRITE_THROUGH) |     /* 101 */
1014                   PAT_VALUE(6, PAT_UNCACHED) |          /* 110 */
1015                   PAT_VALUE(7, PAT_UNCACHEABLE);        /* 111 */
1016         pat_pte_index[PAT_WRITE_BACK]   = 0;
1017         pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
1018         pat_pte_index[PAT_UNCACHED]     = X86_PG_NC_PCD;
1019         pat_pte_index[PAT_UNCACHEABLE]  = X86_PG_NC_PCD | X86_PG_NC_PWT;
1020         pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1021         pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1022
1023         if (cpu_feature & CPUID_PAT) {
1024                 /*
1025                  * If we support the PAT then set-up entries for
1026                  * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1027                  * 4 and 5.
1028                  */
1029                 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1030                           PAT_VALUE(4, PAT_WRITE_PROTECTED);
1031                 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1032                           PAT_VALUE(5, PAT_WRITE_COMBINING);
1033                 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1034                 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1035
1036                 /*
1037                  * Then enable the PAT
1038                  */
1039
1040                 /* Disable PGE. */
1041                 cr4 = rcr4();
1042                 load_cr4(cr4 & ~CR4_PGE);
1043
1044                 /* Disable caches (CD = 1, NW = 0). */
1045                 cr0 = rcr0();
1046                 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1047
1048                 /* Flushes caches and TLBs. */
1049                 wbinvd();
1050                 cpu_invltlb();
1051
1052                 /* Update PAT and index table. */
1053                 wrmsr(MSR_PAT, pat_msr);
1054
1055                 /* Flush caches and TLBs again. */
1056                 wbinvd();
1057                 cpu_invltlb();
1058
1059                 /* Restore caches and PGE. */
1060                 load_cr0(cr0);
1061                 load_cr4(cr4);
1062                 PatMsr = pat_msr;
1063         }
1064 }
1065
1066 /*
1067  * Set 4mb pdir for mp startup
1068  */
1069 void
1070 pmap_set_opt(void)
1071 {
1072         if (cpu_feature & CPUID_PSE) {
1073                 load_cr4(rcr4() | CR4_PSE);
1074                 if (pdir4mb && mycpu->gd_cpuid == 0) {  /* only on BSP */
1075                         cpu_invltlb();
1076                 }
1077         }
1078 }
1079
1080 /*
1081  *      Initialize the pmap module.
1082  *      Called by vm_init, to initialize any structures that the pmap
1083  *      system needs to map virtual memory.
1084  *      pmap_init has been enhanced to support in a fairly consistant
1085  *      way, discontiguous physical memory.
1086  */
1087 void
1088 pmap_init(void)
1089 {
1090         int i;
1091         int initial_pvs;
1092
1093         /*
1094          * Allocate memory for random pmap data structures.  Includes the
1095          * pv_head_table.
1096          */
1097
1098         for (i = 0; i < vm_page_array_size; i++) {
1099                 vm_page_t m;
1100
1101                 m = &vm_page_array[i];
1102                 TAILQ_INIT(&m->md.pv_list);
1103         }
1104
1105         /*
1106          * init the pv free list
1107          */
1108         initial_pvs = vm_page_array_size;
1109         if (initial_pvs < MINPV)
1110                 initial_pvs = MINPV;
1111         pvzone = &pvzone_store;
1112         pvinit = (void *)kmem_alloc(&kernel_map,
1113                                     initial_pvs * sizeof (struct pv_entry),
1114                                     VM_SUBSYS_PVENTRY);
1115         zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1116                   pvinit, initial_pvs);
1117
1118         /*
1119          * Now it is safe to enable pv_table recording.
1120          */
1121         pmap_initialized = TRUE;
1122 }
1123
1124 /*
1125  * Initialize the address space (zone) for the pv_entries.  Set a
1126  * high water mark so that the system can recover from excessive
1127  * numbers of pv entries.
1128  */
1129 void
1130 pmap_init2(void)
1131 {
1132         int shpgperproc = PMAP_SHPGPERPROC;
1133         int entry_max;
1134
1135         TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1136         pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1137         TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1138         pv_entry_high_water = 9 * (pv_entry_max / 10);
1139
1140         /*
1141          * Subtract out pages already installed in the zone (hack)
1142          */
1143         entry_max = pv_entry_max - vm_page_array_size;
1144         if (entry_max <= 0)
1145                 entry_max = 1;
1146
1147         zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT);
1148 }
1149
1150 /*
1151  * Typically used to initialize a fictitious page by vm/device_pager.c
1152  */
1153 void
1154 pmap_page_init(struct vm_page *m)
1155 {
1156         vm_page_init(m);
1157         TAILQ_INIT(&m->md.pv_list);
1158 }
1159
1160 /***************************************************
1161  * Low level helper routines.....
1162  ***************************************************/
1163
1164 /*
1165  * this routine defines the region(s) of memory that should
1166  * not be tested for the modified bit.
1167  */
1168 static __inline
1169 int
1170 pmap_track_modified(vm_pindex_t pindex)
1171 {
1172         vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1173         if ((va < clean_sva) || (va >= clean_eva))
1174                 return 1;
1175         else
1176                 return 0;
1177 }
1178
1179 /*
1180  * Extract the physical page address associated with the map/VA pair.
1181  * The page must be wired for this to work reliably.
1182  *
1183  * XXX for the moment we're using pv_find() instead of pv_get(), as
1184  *     callers might be expecting non-blocking operation.
1185  */
1186 vm_paddr_t
1187 pmap_extract(pmap_t pmap, vm_offset_t va)
1188 {
1189         vm_paddr_t rtval;
1190         pv_entry_t pt_pv;
1191         pt_entry_t *ptep;
1192
1193         rtval = 0;
1194         if (va >= VM_MAX_USER_ADDRESS) {
1195                 /*
1196                  * Kernel page directories might be direct-mapped and
1197                  * there is typically no PV tracking of pte's
1198                  */
1199                 pd_entry_t *pt;
1200
1201                 pt = pmap_pt(pmap, va);
1202                 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1203                         if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1204                                 rtval = *pt & PG_PS_FRAME;
1205                                 rtval |= va & PDRMASK;
1206                         } else {
1207                                 ptep = pmap_pt_to_pte(*pt, va);
1208                                 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1209                                         rtval = *ptep & PG_FRAME;
1210                                         rtval |= va & PAGE_MASK;
1211                                 }
1212                         }
1213                 }
1214         } else {
1215                 /*
1216                  * User pages currently do not direct-map the page directory
1217                  * and some pages might not used managed PVs.  But all PT's
1218                  * will have a PV.
1219                  */
1220                 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1221                 if (pt_pv) {
1222                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1223                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1224                                 rtval = *ptep & PG_FRAME;
1225                                 rtval |= va & PAGE_MASK;
1226                         }
1227                         pv_drop(pt_pv);
1228                 }
1229         }
1230         return rtval;
1231 }
1232
1233 /*
1234  * Similar to extract but checks protections, SMP-friendly short-cut for
1235  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1236  * fall-through to the real fault code.
1237  *
1238  * The returned page, if not NULL, is held (and not busied).
1239  */
1240 vm_page_t
1241 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1242 {
1243         if (pmap && va < VM_MAX_USER_ADDRESS) {
1244                 pv_entry_t pt_pv;
1245                 pv_entry_t pte_pv;
1246                 pt_entry_t *ptep;
1247                 pt_entry_t req;
1248                 vm_page_t m;
1249                 int error;
1250
1251                 req = pmap->pmap_bits[PG_V_IDX] |
1252                       pmap->pmap_bits[PG_U_IDX];
1253                 if (prot & VM_PROT_WRITE)
1254                         req |= pmap->pmap_bits[PG_RW_IDX];
1255
1256                 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1257                 if (pt_pv == NULL)
1258                         return (NULL);
1259                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1260                 if ((*ptep & req) != req) {
1261                         pv_drop(pt_pv);
1262                         return (NULL);
1263                 }
1264                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), &error);
1265                 if (pte_pv && error == 0) {
1266                         m = pte_pv->pv_m;
1267                         vm_page_hold(m);
1268                         if (prot & VM_PROT_WRITE)
1269                                 vm_page_dirty(m);
1270                         pv_put(pte_pv);
1271                 } else if (pte_pv) {
1272                         pv_drop(pte_pv);
1273                         m = NULL;
1274                 } else {
1275                         m = NULL;
1276                 }
1277                 pv_drop(pt_pv);
1278                 return(m);
1279         } else {
1280                 return(NULL);
1281         }
1282 }
1283
1284 /*
1285  * Extract the physical page address associated kernel virtual address.
1286  */
1287 vm_paddr_t
1288 pmap_kextract(vm_offset_t va)
1289 {
1290         pd_entry_t pt;          /* pt entry in pd */
1291         vm_paddr_t pa;
1292
1293         if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1294                 pa = DMAP_TO_PHYS(va);
1295         } else {
1296                 pt = *vtopt(va);
1297                 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1298                         pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1299                 } else {
1300                         /*
1301                          * Beware of a concurrent promotion that changes the
1302                          * PDE at this point!  For example, vtopte() must not
1303                          * be used to access the PTE because it would use the
1304                          * new PDE.  It is, however, safe to use the old PDE
1305                          * because the page table page is preserved by the
1306                          * promotion.
1307                          */
1308                         pa = *pmap_pt_to_pte(pt, va);
1309                         pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1310                 }
1311         }
1312         return pa;
1313 }
1314
1315 /***************************************************
1316  * Low level mapping routines.....
1317  ***************************************************/
1318
1319 /*
1320  * Routine: pmap_kenter
1321  * Function:
1322  *      Add a wired page to the KVA
1323  *      NOTE! note that in order for the mapping to take effect -- you
1324  *      should do an invltlb after doing the pmap_kenter().
1325  */
1326 void
1327 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1328 {
1329         pt_entry_t *ptep;
1330         pt_entry_t npte;
1331
1332         npte = pa |
1333             kernel_pmap.pmap_bits[PG_RW_IDX] |
1334             kernel_pmap.pmap_bits[PG_V_IDX];
1335 //          pgeflag;
1336         ptep = vtopte(va);
1337 #if 1
1338         pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1339 #else
1340         /* FUTURE */
1341         if (*ptep)
1342                 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1343         else
1344                 *ptep = npte;
1345 #endif
1346 }
1347
1348 /*
1349  * Similar to pmap_kenter(), except we only invalidate the mapping on the
1350  * current CPU.  Returns 0 if the previous pte was 0, 1 if it wasn't
1351  * (caller can conditionalize calling smp_invltlb()).
1352  */
1353 int
1354 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1355 {
1356         pt_entry_t *ptep;
1357         pt_entry_t npte;
1358         int res;
1359
1360         npte = pa |
1361             kernel_pmap.pmap_bits[PG_RW_IDX] |
1362             kernel_pmap.pmap_bits[PG_V_IDX];
1363 //          pgeflag;
1364         ptep = vtopte(va);
1365 #if 1
1366         res = 1;
1367 #else
1368         /* FUTURE */
1369         res = (*ptep != 0);
1370 #endif
1371         *ptep = npte;
1372         cpu_invlpg((void *)va);
1373
1374         return res;
1375 }
1376
1377 /*
1378  * Enter addresses into the kernel pmap but don't bother
1379  * doing any tlb invalidations.  Caller will do a rollup
1380  * invalidation via pmap_rollup_inval().
1381  */
1382 int
1383 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1384 {
1385         pt_entry_t *ptep;
1386         pt_entry_t npte;
1387         int res;
1388
1389         npte = pa |
1390             kernel_pmap.pmap_bits[PG_RW_IDX] |
1391             kernel_pmap.pmap_bits[PG_V_IDX];
1392 //          pgeflag;
1393         ptep = vtopte(va);
1394 #if 1
1395         res = 1;
1396 #else
1397         /* FUTURE */
1398         res = (*ptep != 0);
1399 #endif
1400         *ptep = npte;
1401         cpu_invlpg((void *)va);
1402
1403         return res;
1404 }
1405
1406 /*
1407  * remove a page from the kernel pagetables
1408  */
1409 void
1410 pmap_kremove(vm_offset_t va)
1411 {
1412         pt_entry_t *ptep;
1413
1414         ptep = vtopte(va);
1415         pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1416 }
1417
1418 void
1419 pmap_kremove_quick(vm_offset_t va)
1420 {
1421         pt_entry_t *ptep;
1422
1423         ptep = vtopte(va);
1424         (void)pte_load_clear(ptep);
1425         cpu_invlpg((void *)va);
1426 }
1427
1428 /*
1429  * Remove addresses from the kernel pmap but don't bother
1430  * doing any tlb invalidations.  Caller will do a rollup
1431  * invalidation via pmap_rollup_inval().
1432  */
1433 void
1434 pmap_kremove_noinval(vm_offset_t va)
1435 {
1436         pt_entry_t *ptep;
1437
1438         ptep = vtopte(va);
1439         (void)pte_load_clear(ptep);
1440 }
1441
1442 /*
1443  * XXX these need to be recoded.  They are not used in any critical path.
1444  */
1445 void
1446 pmap_kmodify_rw(vm_offset_t va)
1447 {
1448         atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1449         cpu_invlpg((void *)va);
1450 }
1451
1452 /* NOT USED
1453 void
1454 pmap_kmodify_nc(vm_offset_t va)
1455 {
1456         atomic_set_long(vtopte(va), PG_N);
1457         cpu_invlpg((void *)va);
1458 }
1459 */
1460
1461 /*
1462  * Used to map a range of physical addresses into kernel virtual
1463  * address space during the low level boot, typically to map the
1464  * dump bitmap, message buffer, and vm_page_array.
1465  *
1466  * These mappings are typically made at some pointer after the end of the
1467  * kernel text+data.
1468  *
1469  * We could return PHYS_TO_DMAP(start) here and not allocate any
1470  * via (*virtp), but then kmem from userland and kernel dumps won't
1471  * have access to the related pointers.
1472  */
1473 vm_offset_t
1474 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1475 {
1476         vm_offset_t va;
1477         vm_offset_t va_start;
1478
1479         /*return PHYS_TO_DMAP(start);*/
1480
1481         va_start = *virtp;
1482         va = va_start;
1483
1484         while (start < end) {
1485                 pmap_kenter_quick(va, start);
1486                 va += PAGE_SIZE;
1487                 start += PAGE_SIZE;
1488         }
1489         *virtp = va;
1490         return va_start;
1491 }
1492
1493 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
1494
1495 /*
1496  * Remove the specified set of pages from the data and instruction caches.
1497  *
1498  * In contrast to pmap_invalidate_cache_range(), this function does not
1499  * rely on the CPU's self-snoop feature, because it is intended for use
1500  * when moving pages into a different cache domain.
1501  */
1502 void
1503 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1504 {
1505         vm_offset_t daddr, eva;
1506         int i;
1507
1508         if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1509             (cpu_feature & CPUID_CLFSH) == 0)
1510                 wbinvd();
1511         else {
1512                 cpu_mfence();
1513                 for (i = 0; i < count; i++) {
1514                         daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1515                         eva = daddr + PAGE_SIZE;
1516                         for (; daddr < eva; daddr += cpu_clflush_line_size)
1517                                 clflush(daddr);
1518                 }
1519                 cpu_mfence();
1520         }
1521 }
1522
1523 void
1524 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1525 {
1526         KASSERT((sva & PAGE_MASK) == 0,
1527             ("pmap_invalidate_cache_range: sva not page-aligned"));
1528         KASSERT((eva & PAGE_MASK) == 0,
1529             ("pmap_invalidate_cache_range: eva not page-aligned"));
1530
1531         if (cpu_feature & CPUID_SS) {
1532                 ; /* If "Self Snoop" is supported, do nothing. */
1533         } else {
1534                 /* Globally invalidate caches */
1535                 cpu_wbinvd_on_all_cpus();
1536         }
1537 }
1538
1539 /*
1540  * Invalidate the specified range of virtual memory on all cpus associated
1541  * with the pmap.
1542  */
1543 void
1544 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1545 {
1546         pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1547 }
1548
1549 /*
1550  * Add a list of wired pages to the kva.  This routine is used for temporary
1551  * kernel mappings such as those found in buffer cache buffer.  Page
1552  * modifications and accesses are not tracked or recorded.
1553  *
1554  * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1555  *       semantics as previous mappings may have been zerod without any
1556  *       invalidation.
1557  *
1558  * The page *must* be wired.
1559  */
1560 void
1561 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1562 {
1563         vm_offset_t end_va;
1564         vm_offset_t va;
1565
1566         end_va = beg_va + count * PAGE_SIZE;
1567
1568         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1569                 pt_entry_t *pte;
1570
1571                 pte = vtopte(va);
1572                 *pte = VM_PAGE_TO_PHYS(*m) |
1573                     kernel_pmap.pmap_bits[PG_RW_IDX] |
1574                     kernel_pmap.pmap_bits[PG_V_IDX] |
1575                     kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1576 //              pgeflag;
1577                 m++;
1578         }
1579         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1580 }
1581
1582 /*
1583  * This routine jerks page mappings from the kernel -- it is meant only
1584  * for temporary mappings such as those found in buffer cache buffers.
1585  * No recording modified or access status occurs.
1586  *
1587  * MPSAFE, INTERRUPT SAFE (cluster callback)
1588  */
1589 void
1590 pmap_qremove(vm_offset_t beg_va, int count)
1591 {
1592         vm_offset_t end_va;
1593         vm_offset_t va;
1594
1595         end_va = beg_va + count * PAGE_SIZE;
1596
1597         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1598                 pt_entry_t *pte;
1599
1600                 pte = vtopte(va);
1601                 (void)pte_load_clear(pte);
1602                 cpu_invlpg((void *)va);
1603         }
1604         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1605 }
1606
1607 /*
1608  * This routine removes temporary kernel mappings, only invalidating them
1609  * on the current cpu.  It should only be used under carefully controlled
1610  * conditions.
1611  */
1612 void
1613 pmap_qremove_quick(vm_offset_t beg_va, int count)
1614 {
1615         vm_offset_t end_va;
1616         vm_offset_t va;
1617
1618         end_va = beg_va + count * PAGE_SIZE;
1619
1620         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1621                 pt_entry_t *pte;
1622
1623                 pte = vtopte(va);
1624                 (void)pte_load_clear(pte);
1625                 cpu_invlpg((void *)va);
1626         }
1627 }
1628
1629 /*
1630  * This routine removes temporary kernel mappings *without* invalidating
1631  * the TLB.  It can only be used on permanent kva reservations such as those
1632  * found in buffer cache buffers, under carefully controlled circumstances.
1633  *
1634  * NOTE: Repopulating these KVAs requires unconditional invalidation.
1635  *       (pmap_qenter() does unconditional invalidation).
1636  */
1637 void
1638 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1639 {
1640         vm_offset_t end_va;
1641         vm_offset_t va;
1642
1643         end_va = beg_va + count * PAGE_SIZE;
1644
1645         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1646                 pt_entry_t *pte;
1647
1648                 pte = vtopte(va);
1649                 (void)pte_load_clear(pte);
1650         }
1651 }
1652
1653 /*
1654  * Create a new thread and optionally associate it with a (new) process.
1655  * NOTE! the new thread's cpu may not equal the current cpu.
1656  */
1657 void
1658 pmap_init_thread(thread_t td)
1659 {
1660         /* enforce pcb placement & alignment */
1661         td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1662         td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1663         td->td_savefpu = &td->td_pcb->pcb_save;
1664         td->td_sp = (char *)td->td_pcb; /* no -16 */
1665 }
1666
1667 /*
1668  * This routine directly affects the fork perf for a process.
1669  */
1670 void
1671 pmap_init_proc(struct proc *p)
1672 {
1673 }
1674
1675 static void
1676 pmap_pinit_defaults(struct pmap *pmap)
1677 {
1678         bcopy(pmap_bits_default, pmap->pmap_bits,
1679               sizeof(pmap_bits_default));
1680         bcopy(protection_codes, pmap->protection_codes,
1681               sizeof(protection_codes));
1682         bcopy(pat_pte_index, pmap->pmap_cache_bits,
1683               sizeof(pat_pte_index));
1684         pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1685         pmap->copyinstr = std_copyinstr;
1686         pmap->copyin = std_copyin;
1687         pmap->copyout = std_copyout;
1688         pmap->fubyte = std_fubyte;
1689         pmap->subyte = std_subyte;
1690         pmap->fuword = std_fuword;
1691         pmap->suword = std_suword;
1692         pmap->suword32 = std_suword32;
1693 }
1694 /*
1695  * Initialize pmap0/vmspace0.  This pmap is not added to pmap_list because
1696  * it, and IdlePTD, represents the template used to update all other pmaps.
1697  *
1698  * On architectures where the kernel pmap is not integrated into the user
1699  * process pmap, this pmap represents the process pmap, not the kernel pmap.
1700  * kernel_pmap should be used to directly access the kernel_pmap.
1701  */
1702 void
1703 pmap_pinit0(struct pmap *pmap)
1704 {
1705         pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1706         pmap->pm_count = 1;
1707         CPUMASK_ASSZERO(pmap->pm_active);
1708         pmap->pm_pvhint = NULL;
1709         RB_INIT(&pmap->pm_pvroot);
1710         spin_init(&pmap->pm_spin, "pmapinit0");
1711         lwkt_token_init(&pmap->pm_token, "pmap_tok");
1712         bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1713         pmap_pinit_defaults(pmap);
1714 }
1715
1716 /*
1717  * Initialize a preallocated and zeroed pmap structure,
1718  * such as one in a vmspace structure.
1719  */
1720 static void
1721 pmap_pinit_simple(struct pmap *pmap)
1722 {
1723         /*
1724          * Misc initialization
1725          */
1726         pmap->pm_count = 1;
1727         CPUMASK_ASSZERO(pmap->pm_active);
1728         pmap->pm_pvhint = NULL;
1729         pmap->pm_flags = PMAP_FLAG_SIMPLE;
1730
1731         pmap_pinit_defaults(pmap);
1732
1733         /*
1734          * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1735          * for this).
1736          */
1737         if (pmap->pm_pmlpv == NULL) {
1738                 RB_INIT(&pmap->pm_pvroot);
1739                 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1740                 spin_init(&pmap->pm_spin, "pmapinitsimple");
1741                 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1742         }
1743 }
1744
1745 void
1746 pmap_pinit(struct pmap *pmap)
1747 {
1748         pv_entry_t pv;
1749         int j;
1750
1751         if (pmap->pm_pmlpv) {
1752                 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1753                         pmap_puninit(pmap);
1754                 }
1755         }
1756
1757         pmap_pinit_simple(pmap);
1758         pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1759
1760         /*
1761          * No need to allocate page table space yet but we do need a valid
1762          * page directory table.
1763          */
1764         if (pmap->pm_pml4 == NULL) {
1765                 pmap->pm_pml4 =
1766                     (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1767                                                         PAGE_SIZE,
1768                                                         VM_SUBSYS_PML4);
1769         }
1770
1771         /*
1772          * Allocate the page directory page, which wires it even though
1773          * it isn't being entered into some higher level page table (it
1774          * being the highest level).  If one is already cached we don't
1775          * have to do anything.
1776          */
1777         if ((pv = pmap->pm_pmlpv) == NULL) {
1778                 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1779                 pmap->pm_pmlpv = pv;
1780                 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1781                             VM_PAGE_TO_PHYS(pv->pv_m));
1782                 pv_put(pv);
1783
1784                 /*
1785                  * Install DMAP and KMAP.
1786                  */
1787                 for (j = 0; j < NDMPML4E; ++j) {
1788                         pmap->pm_pml4[DMPML4I + j] =
1789                             (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1790                             pmap->pmap_bits[PG_RW_IDX] |
1791                             pmap->pmap_bits[PG_V_IDX] |
1792                             pmap->pmap_bits[PG_U_IDX];
1793                 }
1794                 pmap->pm_pml4[KPML4I] = KPDPphys |
1795                     pmap->pmap_bits[PG_RW_IDX] |
1796                     pmap->pmap_bits[PG_V_IDX] |
1797                     pmap->pmap_bits[PG_U_IDX];
1798
1799                 /*
1800                  * install self-referential address mapping entry
1801                  */
1802                 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1803                     pmap->pmap_bits[PG_V_IDX] |
1804                     pmap->pmap_bits[PG_RW_IDX] |
1805                     pmap->pmap_bits[PG_A_IDX] |
1806                     pmap->pmap_bits[PG_M_IDX];
1807         } else {
1808                 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1809                 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1810         }
1811         KKASSERT(pmap->pm_pml4[255] == 0);
1812         KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1813         KKASSERT(pv->pv_entry.rbe_left == NULL);
1814         KKASSERT(pv->pv_entry.rbe_right == NULL);
1815 }
1816
1817 /*
1818  * Clean up a pmap structure so it can be physically freed.  This routine
1819  * is called by the vmspace dtor function.  A great deal of pmap data is
1820  * left passively mapped to improve vmspace management so we have a bit
1821  * of cleanup work to do here.
1822  */
1823 void
1824 pmap_puninit(pmap_t pmap)
1825 {
1826         pv_entry_t pv;
1827         vm_page_t p;
1828
1829         KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1830         if ((pv = pmap->pm_pmlpv) != NULL) {
1831                 if (pv_hold_try(pv) == 0)
1832                         pv_lock(pv);
1833                 KKASSERT(pv == pmap->pm_pmlpv);
1834                 p = pmap_remove_pv_page(pv);
1835                 pv_free(pv, NULL, 1);
1836                 pv = NULL;      /* safety */
1837                 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1838                 vm_page_busy_wait(p, FALSE, "pgpun");
1839                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1840                 vm_page_unwire(p, 0);
1841                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1842
1843                 /*
1844                  * XXX eventually clean out PML4 static entries and
1845                  * use vm_page_free_zero()
1846                  */
1847                 vm_page_free(p);
1848                 pmap->pm_pmlpv = NULL;
1849         }
1850         if (pmap->pm_pml4) {
1851                 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1852                 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1853                 pmap->pm_pml4 = NULL;
1854         }
1855         KKASSERT(pmap->pm_stats.resident_count == 0);
1856         KKASSERT(pmap->pm_stats.wired_count == 0);
1857 }
1858
1859 /*
1860  * Wire in kernel global address entries.  To avoid a race condition
1861  * between pmap initialization and pmap_growkernel, this procedure
1862  * adds the pmap to the master list (which growkernel scans to update),
1863  * then copies the template.
1864  */
1865 void
1866 pmap_pinit2(struct pmap *pmap)
1867 {
1868         spin_lock(&pmap_spin);
1869         TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1870         spin_unlock(&pmap_spin);
1871 }
1872
1873 /*
1874  * This routine is called when various levels in the page table need to
1875  * be populated.  This routine cannot fail.
1876  *
1877  * This function returns two locked pv_entry's, one representing the
1878  * requested pv and one representing the requested pv's parent pv.  If
1879  * an intermediate page table does not exist it will be created, mapped,
1880  * wired, and the parent page table will be given an additional hold
1881  * count representing the presence of the child pv_entry.
1882  */
1883 static
1884 pv_entry_t
1885 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1886 {
1887         pt_entry_t *ptep;
1888         pv_entry_t pv;
1889         pv_entry_t pvp;
1890         vm_pindex_t pt_pindex;
1891         vm_page_t m;
1892         int isnew;
1893         int ispt;
1894
1895         /*
1896          * If the pv already exists and we aren't being asked for the
1897          * parent page table page we can just return it.  A locked+held pv
1898          * is returned.  The pv will also have a second hold related to the
1899          * pmap association that we don't have to worry about.
1900          */
1901         ispt = 0;
1902         pv = pv_alloc(pmap, ptepindex, &isnew);
1903         if (isnew == 0 && pvpp == NULL)
1904                 return(pv);
1905
1906         /*
1907          * Special case terminal PVs.  These are not page table pages so
1908          * no vm_page is allocated (the caller supplied the vm_page).  If
1909          * pvpp is non-NULL we are being asked to also removed the pt_pv
1910          * for this pv.
1911          *
1912          * Note that pt_pv's are only returned for user VAs. We assert that
1913          * a pt_pv is not being requested for kernel VAs.  The kernel
1914          * pre-wires all higher-level page tables so don't overload managed
1915          * higher-level page tables on top of it!
1916          */
1917         if (ptepindex < pmap_pt_pindex(0)) {
1918                 if (ptepindex >= NUPTE_USER) {
1919                         /* kernel manages this manually for KVM */
1920                         KKASSERT(pvpp == NULL);
1921                 } else {
1922                         KKASSERT(pvpp != NULL);
1923                         pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1924                         pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1925                         if (isnew) {
1926                                 vm_page_wire_quick(pvp->pv_m);
1927                                 if (pvpp)
1928                                         *pvpp = pvp;
1929                                 else
1930                                         pv_put(pvp);
1931                         } else {
1932                                 *pvpp = pvp;
1933                         }
1934                 }
1935                 return(pv);
1936         }
1937
1938         /*
1939          * Non-terminal PVs allocate a VM page to represent the page table,
1940          * so we have to resolve pvp and calculate ptepindex for the pvp
1941          * and then for the page table entry index in the pvp for
1942          * fall-through.
1943          */
1944         if (ptepindex < pmap_pd_pindex(0)) {
1945                 /*
1946                  * pv is PT, pvp is PD
1947                  */
1948                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1949                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1950                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1951                 if (!isnew)
1952                         goto notnew;
1953
1954                 /*
1955                  * PT index in PD
1956                  */
1957                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1958                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1959                 ispt = 1;
1960         } else if (ptepindex < pmap_pdp_pindex(0)) {
1961                 /*
1962                  * pv is PD, pvp is PDP
1963                  *
1964                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1965                  *                   the PD.
1966                  */
1967                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1968                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1969
1970                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1971                         KKASSERT(pvpp == NULL);
1972                         pvp = NULL;
1973                 } else {
1974                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
1975                 }
1976                 if (!isnew)
1977                         goto notnew;
1978
1979                 /*
1980                  * PD index in PDP
1981                  */
1982                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1983                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1984         } else if (ptepindex < pmap_pml4_pindex()) {
1985                 /*
1986                  * pv is PDP, pvp is the root pml4 table
1987                  */
1988                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1989                 if (!isnew)
1990                         goto notnew;
1991
1992                 /*
1993                  * PDP index in PML4
1994                  */
1995                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1996                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1997         } else {
1998                 /*
1999                  * pv represents the top-level PML4, there is no parent.
2000                  */
2001                 pvp = NULL;
2002                 if (!isnew)
2003                         goto notnew;
2004         }
2005
2006         /*
2007          * (isnew) is TRUE, pv is not terminal.
2008          *
2009          * (1) Add a wire count to the parent page table (pvp).
2010          * (2) Allocate a VM page for the page table.
2011          * (3) Enter the VM page into the parent page table.
2012          *
2013          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2014          */
2015         if (pvp)
2016                 vm_page_wire_quick(pvp->pv_m);
2017
2018         for (;;) {
2019                 m = vm_page_alloc(NULL, pv->pv_pindex,
2020                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2021                                   VM_ALLOC_INTERRUPT);
2022                 if (m)
2023                         break;
2024                 vm_wait(0);
2025         }
2026         vm_page_spin_lock(m);
2027         pmap_page_stats_adding(m);
2028         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2029         pv->pv_m = m;
2030         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2031         vm_page_spin_unlock(m);
2032         vm_page_unmanage(m);    /* m must be spinunlocked */
2033
2034         pmap_zero_page(VM_PAGE_TO_PHYS(m));
2035         m->valid = VM_PAGE_BITS_ALL;
2036         vm_page_wire(m);        /* wire for mapping in parent */
2037
2038         /*
2039          * Wire the page into pvp.  Bump the resident_count for the pmap.
2040          * There is no pvp for the top level, address the pm_pml4[] array
2041          * directly.
2042          *
2043          * If the caller wants the parent we return it, otherwise
2044          * we just put it away.
2045          *
2046          * No interlock is needed for pte 0 -> non-zero.
2047          *
2048          * In the situation where *ptep is valid we might have an unmanaged
2049          * page table page shared from another page table which we need to
2050          * unshare before installing our private page table page.
2051          */
2052         if (pvp) {
2053                 ptep = pv_pte_lookup(pvp, ptepindex);
2054                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2055                         pt_entry_t pte;
2056
2057                         if (ispt == 0) {
2058                                 panic("pmap_allocpte: unexpected pte %p/%d",
2059                                       pvp, (int)ptepindex);
2060                         }
2061                         pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
2062                         if (vm_page_unwire_quick(
2063                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2064                                 panic("pmap_allocpte: shared pgtable "
2065                                       "pg bad wirecount");
2066                         }
2067                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
2068                 }
2069                 *ptep = VM_PAGE_TO_PHYS(m) |
2070                         (pmap->pmap_bits[PG_U_IDX] |
2071                          pmap->pmap_bits[PG_RW_IDX] |
2072                          pmap->pmap_bits[PG_V_IDX] |
2073                          pmap->pmap_bits[PG_A_IDX] |
2074                          pmap->pmap_bits[PG_M_IDX]);
2075         }
2076         vm_page_wakeup(m);
2077 notnew:
2078         if (pvpp)
2079                 *pvpp = pvp;
2080         else if (pvp)
2081                 pv_put(pvp);
2082         return (pv);
2083 }
2084
2085 /*
2086  * This version of pmap_allocpte() checks for possible segment optimizations
2087  * that would allow page-table sharing.  It can be called for terminal
2088  * page or page table page ptepindex's.
2089  *
2090  * The function is called with page table page ptepindex's for fictitious
2091  * and unmanaged terminal pages.  That is, we don't want to allocate a
2092  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
2093  * for this case.
2094  *
2095  * This function can return a pv and *pvpp associated with the passed in pmap
2096  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
2097  * an unmanaged page table page will be entered into the pass in pmap.
2098  */
2099 static
2100 pv_entry_t
2101 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2102                   vm_map_entry_t entry, vm_offset_t va)
2103 {
2104         vm_object_t object;
2105         pmap_t obpmap;
2106         pmap_t *obpmapp;
2107         vm_offset_t b;
2108         pv_entry_t pte_pv;      /* in original or shared pmap */
2109         pv_entry_t pt_pv;       /* in original or shared pmap */
2110         pv_entry_t proc_pd_pv;  /* in original pmap */
2111         pv_entry_t proc_pt_pv;  /* in original pmap */
2112         pv_entry_t xpv;         /* PT in shared pmap */
2113         pd_entry_t *pt;         /* PT entry in PD of original pmap */
2114         pd_entry_t opte;        /* contents of *pt */
2115         pd_entry_t npte;        /* contents of *pt */
2116         vm_page_t m;
2117
2118 retry:
2119         /*
2120          * Basic tests, require a non-NULL vm_map_entry, require proper
2121          * alignment and type for the vm_map_entry, require that the
2122          * underlying object already be allocated.
2123          *
2124          * We allow almost any type of object to use this optimization.
2125          * The object itself does NOT have to be sized to a multiple of the
2126          * segment size, but the memory mapping does.
2127          *
2128          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2129          *     won't work as expected.
2130          */
2131         if (entry == NULL ||
2132             pmap_mmu_optimize == 0 ||                   /* not enabled */
2133             (pmap->pm_flags & PMAP_HVM) ||              /* special pmap */
2134             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal or pt */
2135             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2136             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2137             entry->object.vm_object == NULL ||          /* needs VM object */
2138             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2139             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2140             (entry->offset & SEG_MASK) ||               /* must be aligned */
2141             (entry->start & SEG_MASK)) {
2142                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2143         }
2144
2145         /*
2146          * Make sure the full segment can be represented.
2147          */
2148         b = va & ~(vm_offset_t)SEG_MASK;
2149         if (b < entry->start || b + SEG_SIZE > entry->end)
2150                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2151
2152         /*
2153          * If the full segment can be represented dive the VM object's
2154          * shared pmap, allocating as required.
2155          */
2156         object = entry->object.vm_object;
2157
2158         if (entry->protection & VM_PROT_WRITE)
2159                 obpmapp = &object->md.pmap_rw;
2160         else
2161                 obpmapp = &object->md.pmap_ro;
2162
2163 #ifdef PMAP_DEBUG2
2164         if (pmap_enter_debug > 0) {
2165                 --pmap_enter_debug;
2166                 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2167                         "obpmapp %p %p\n",
2168                         va, entry->protection, object,
2169                         obpmapp, *obpmapp);
2170                 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2171                         entry, entry->start, entry->end);
2172         }
2173 #endif
2174
2175         /*
2176          * We allocate what appears to be a normal pmap but because portions
2177          * of this pmap are shared with other unrelated pmaps we have to
2178          * set pm_active to point to all cpus.
2179          *
2180          * XXX Currently using pmap_spin to interlock the update, can't use
2181          *     vm_object_hold/drop because the token might already be held
2182          *     shared OR exclusive and we don't know.
2183          */
2184         while ((obpmap = *obpmapp) == NULL) {
2185                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2186                 pmap_pinit_simple(obpmap);
2187                 pmap_pinit2(obpmap);
2188                 spin_lock(&pmap_spin);
2189                 if (*obpmapp != NULL) {
2190                         /*
2191                          * Handle race
2192                          */
2193                         spin_unlock(&pmap_spin);
2194                         pmap_release(obpmap);
2195                         pmap_puninit(obpmap);
2196                         kfree(obpmap, M_OBJPMAP);
2197                         obpmap = *obpmapp; /* safety */
2198                 } else {
2199                         obpmap->pm_active = smp_active_mask;
2200                         obpmap->pm_flags |= PMAP_SEGSHARED;
2201                         *obpmapp = obpmap;
2202                         spin_unlock(&pmap_spin);
2203                 }
2204         }
2205
2206         /*
2207          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2208          * pte/pt using the shared pmap from the object but also adjust
2209          * the process pmap's page table page as a side effect.
2210          */
2211
2212         /*
2213          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2214          * we will return.  This is true if ptepindex represents a terminal
2215          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2216          * the PD.
2217          */
2218         pt_pv = NULL;
2219         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2220         if (ptepindex >= pmap_pt_pindex(0))
2221                 xpv = pte_pv;
2222         else
2223                 xpv = pt_pv;
2224
2225         /*
2226          * Resolve the PD in the process pmap so we can properly share the
2227          * page table page.  Lock order is bottom-up (leaf first)!
2228          *
2229          * NOTE: proc_pt_pv can be NULL.
2230          */
2231         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2232         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2233 #ifdef PMAP_DEBUG2
2234         if (pmap_enter_debug > 0) {
2235                 --pmap_enter_debug;
2236                 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2237                         proc_pt_pv,
2238                         (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2239                         proc_pd_pv,
2240                         va);
2241         }
2242 #endif
2243
2244         /*
2245          * xpv is the page table page pv from the shared object
2246          * (for convenience), from above.
2247          *
2248          * Calculate the pte value for the PT to load into the process PD.
2249          * If we have to change it we must properly dispose of the previous
2250          * entry.
2251          */
2252         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2253         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2254                (pmap->pmap_bits[PG_U_IDX] |
2255                 pmap->pmap_bits[PG_RW_IDX] |
2256                 pmap->pmap_bits[PG_V_IDX] |
2257                 pmap->pmap_bits[PG_A_IDX] |
2258                 pmap->pmap_bits[PG_M_IDX]);
2259
2260         /*
2261          * Dispose of previous page table page if it was local to the
2262          * process pmap.  If the old pt is not empty we cannot dispose of it
2263          * until we clean it out.  This case should not arise very often so
2264          * it is not optimized.
2265          */
2266         if (proc_pt_pv) {
2267                 pmap_inval_bulk_t bulk;
2268
2269                 if (proc_pt_pv->pv_m->wire_count != 1) {
2270                         pv_put(proc_pd_pv);
2271                         pv_put(proc_pt_pv);
2272                         pv_put(pt_pv);
2273                         pv_put(pte_pv);
2274                         pmap_remove(pmap,
2275                                     va & ~(vm_offset_t)SEG_MASK,
2276                                     (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2277                         goto retry;
2278                 }
2279
2280                 /*
2281                  * The release call will indirectly clean out *pt
2282                  */
2283                 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2284                 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2285                 pmap_inval_bulk_flush(&bulk);
2286                 proc_pt_pv = NULL;
2287                 /* relookup */
2288                 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2289         }
2290
2291         /*
2292          * Handle remaining cases.
2293          */
2294         if (*pt == 0) {
2295                 *pt = npte;
2296                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2297                 vm_page_wire_quick(proc_pd_pv->pv_m);   /* proc pd for sh pt */
2298                 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2299         } else if (*pt != npte) {
2300                 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2301
2302 #if 0
2303                 opte = pte_load_clear(pt);
2304                 KKASSERT(opte && opte != npte);
2305
2306                 *pt = npte;
2307 #endif
2308                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2309
2310                 /*
2311                  * Clean up opte, bump the wire_count for the process
2312                  * PD page representing the new entry if it was
2313                  * previously empty.
2314                  *
2315                  * If the entry was not previously empty and we have
2316                  * a PT in the proc pmap then opte must match that
2317                  * pt.  The proc pt must be retired (this is done
2318                  * later on in this procedure).
2319                  *
2320                  * NOTE: replacing valid pte, wire_count on proc_pd_pv
2321                  * stays the same.
2322                  */
2323                 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2324                 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2325                 if (vm_page_unwire_quick(m)) {
2326                         panic("pmap_allocpte_seg: "
2327                               "bad wire count %p",
2328                               m);
2329                 }
2330         }
2331
2332         /*
2333          * The existing process page table was replaced and must be destroyed
2334          * here.
2335          */
2336         if (proc_pd_pv)
2337                 pv_put(proc_pd_pv);
2338         if (pvpp)
2339                 *pvpp = pt_pv;
2340         else
2341                 pv_put(pt_pv);
2342
2343         return (pte_pv);
2344 }
2345
2346 /*
2347  * Release any resources held by the given physical map.
2348  *
2349  * Called when a pmap initialized by pmap_pinit is being released.  Should
2350  * only be called if the map contains no valid mappings.
2351  *
2352  * Caller must hold pmap->pm_token
2353  */
2354 struct pmap_release_info {
2355         pmap_t  pmap;
2356         int     retry;
2357         pv_entry_t pvp;
2358 };
2359
2360 static int pmap_release_callback(pv_entry_t pv, void *data);
2361
2362 void
2363 pmap_release(struct pmap *pmap)
2364 {
2365         struct pmap_release_info info;
2366
2367         KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2368                 ("pmap still active! %016jx",
2369                 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2370
2371         spin_lock(&pmap_spin);
2372         TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
2373         spin_unlock(&pmap_spin);
2374
2375         /*
2376          * Pull pv's off the RB tree in order from low to high and release
2377          * each page.
2378          */
2379         info.pmap = pmap;
2380         do {
2381                 info.retry = 0;
2382                 info.pvp = NULL;
2383
2384                 spin_lock(&pmap->pm_spin);
2385                 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2386                         pmap_release_callback, &info);
2387                 spin_unlock(&pmap->pm_spin);
2388
2389                 if (info.pvp)
2390                         pv_put(info.pvp);
2391         } while (info.retry);
2392
2393
2394         /*
2395          * One resident page (the pml4 page) should remain.
2396          * No wired pages should remain.
2397          */
2398         KKASSERT(pmap->pm_stats.resident_count ==
2399                  ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2400
2401         KKASSERT(pmap->pm_stats.wired_count == 0);
2402 }
2403
2404 /*
2405  * Called from low to high.  We must cache the proper parent pv so we
2406  * can adjust its wired count.
2407  */
2408 static int
2409 pmap_release_callback(pv_entry_t pv, void *data)
2410 {
2411         struct pmap_release_info *info = data;
2412         pmap_t pmap = info->pmap;
2413         vm_pindex_t pindex;
2414         int r;
2415
2416         if (info->pvp == pv) {
2417                 spin_unlock(&pmap->pm_spin);
2418                 info->pvp = NULL;
2419         } else if (pv_hold_try(pv)) {
2420                 spin_unlock(&pmap->pm_spin);
2421         } else {
2422                 spin_unlock(&pmap->pm_spin);
2423                 pv_lock(pv);
2424         }
2425         if (pv->pv_pmap != pmap) {
2426                 pv_put(pv);
2427                 spin_lock(&pmap->pm_spin);
2428                 info->retry = 1;
2429                 return(-1);
2430         }
2431
2432         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2433                 /*
2434                  * parent is PT
2435                  */
2436                 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2437                 pindex += NUPTE_TOTAL;
2438         } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2439                 /*
2440                  * parent is PD
2441                  */
2442                 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2443                 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2444         } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2445                 /*
2446                  * parent is PDP
2447                  */
2448                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2449                          NPDPEPGSHIFT;
2450                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2451         } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2452                 /*
2453                  * parent is PML4 (there's only one)
2454                  */
2455 #if 0
2456                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2457                            NUPD_TOTAL) >> NPML4EPGSHIFT;
2458                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2459 #endif
2460                 pindex = pmap_pml4_pindex();
2461         } else {
2462                 /*
2463                  * parent is NULL
2464                  */
2465                 if (info->pvp) {
2466                         pv_put(info->pvp);
2467                         info->pvp = NULL;
2468                 }
2469                 pindex = 0;
2470         }
2471         if (pindex) {
2472                 if (info->pvp && info->pvp->pv_pindex != pindex) {
2473                         pv_put(info->pvp);
2474                         info->pvp = NULL;
2475                 }
2476                 if (info->pvp == NULL)
2477                         info->pvp = pv_get(pmap, pindex);
2478         } else {
2479                 if (info->pvp) {
2480                         pv_put(info->pvp);
2481                         info->pvp = NULL;
2482                 }
2483         }
2484         r = pmap_release_pv(pv, info->pvp, NULL);
2485         spin_lock(&pmap->pm_spin);
2486         return(r);
2487 }
2488
2489 /*
2490  * Called with held (i.e. also locked) pv.  This function will dispose of
2491  * the lock along with the pv.
2492  *
2493  * If the caller already holds the locked parent page table for pv it
2494  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2495  * pass NULL for pvp.
2496  */
2497 static int
2498 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2499 {
2500         vm_page_t p;
2501
2502         /*
2503          * The pmap is currently not spinlocked, pv is held+locked.
2504          * Remove the pv's page from its parent's page table.  The
2505          * parent's page table page's wire_count will be decremented.
2506          *
2507          * This will clean out the pte at any level of the page table.
2508          * If smp != 0 all cpus are affected.
2509          *
2510          * Do not tear-down recursively, its faster to just let the
2511          * release run its course.
2512          */
2513         pmap_remove_pv_pte(pv, pvp, bulk, 0);
2514
2515         /*
2516          * Terminal pvs are unhooked from their vm_pages.  Because
2517          * terminal pages aren't page table pages they aren't wired
2518          * by us, so we have to be sure not to unwire them either.
2519          */
2520         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2521                 pmap_remove_pv_page(pv);
2522                 goto skip;
2523         }
2524
2525         /*
2526          * We leave the top-level page table page cached, wired, and
2527          * mapped in the pmap until the dtor function (pmap_puninit())
2528          * gets called.
2529          *
2530          * Since we are leaving the top-level pv intact we need
2531          * to break out of what would otherwise be an infinite loop.
2532          */
2533         if (pv->pv_pindex == pmap_pml4_pindex()) {
2534                 pv_put(pv);
2535                 return(-1);
2536         }
2537
2538         /*
2539          * For page table pages (other than the top-level page),
2540          * remove and free the vm_page.  The representitive mapping
2541          * removed above by pmap_remove_pv_pte() did not undo the
2542          * last wire_count so we have to do that as well.
2543          */
2544         p = pmap_remove_pv_page(pv);
2545         vm_page_busy_wait(p, FALSE, "pmaprl");
2546         if (p->wire_count != 1) {
2547                 kprintf("p->wire_count was %016lx %d\n",
2548                         pv->pv_pindex, p->wire_count);
2549         }
2550         KKASSERT(p->wire_count == 1);
2551         KKASSERT(p->flags & PG_UNMANAGED);
2552
2553         vm_page_unwire(p, 0);
2554         KKASSERT(p->wire_count == 0);
2555
2556         vm_page_free(p);
2557 skip:
2558         pv_free(pv, pvp, 1);
2559
2560         return 0;
2561 }
2562
2563 /*
2564  * This function will remove the pte associated with a pv from its parent.
2565  * Terminal pv's are supported.  All cpus specified by (bulk) are properly
2566  * invalidated.
2567  *
2568  * The wire count will be dropped on the parent page table.  The wire
2569  * count on the page being removed (pv->pv_m) from the parent page table
2570  * is NOT touched.  Note that terminal pages will not have any additional
2571  * wire counts while page table pages will have at least one representing
2572  * the mapping, plus others representing sub-mappings.
2573  *
2574  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2575  *       pages and user page table and terminal pages.
2576  *
2577  * The pv must be locked.  The pvp, if supplied, must be locked.  All
2578  * supplied pv's will remain locked on return.
2579  *
2580  * XXX must lock parent pv's if they exist to remove pte XXX
2581  */
2582 static
2583 void
2584 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2585                    int destroy)
2586 {
2587         vm_pindex_t ptepindex = pv->pv_pindex;
2588         pmap_t pmap = pv->pv_pmap;
2589         vm_page_t p;
2590         int gotpvp = 0;
2591
2592         KKASSERT(pmap);
2593
2594         if (ptepindex == pmap_pml4_pindex()) {
2595                 /*
2596                  * We are the top level pml4 table, there is no parent.
2597                  */
2598                 p = pmap->pm_pmlpv->pv_m;
2599         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2600                 /*
2601                  * Remove a PDP page from the pml4e.  This can only occur
2602                  * with user page tables.  We do not have to lock the
2603                  * pml4 PV so just ignore pvp.
2604                  */
2605                 vm_pindex_t pml4_pindex;
2606                 vm_pindex_t pdp_index;
2607                 pml4_entry_t *pdp;
2608
2609                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2610                 if (pvp == NULL) {
2611                         pml4_pindex = pmap_pml4_pindex();
2612                         pvp = pv_get(pv->pv_pmap, pml4_pindex);
2613                         KKASSERT(pvp);
2614                         gotpvp = 1;
2615                 }
2616                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2617                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2618                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2619                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2620         } else if (ptepindex >= pmap_pd_pindex(0)) {
2621                 /*
2622                  * Remove a PD page from the pdp
2623                  *
2624                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2625                  *                   of a simple pmap because it stops at
2626                  *                   the PD page.
2627                  */
2628                 vm_pindex_t pdp_pindex;
2629                 vm_pindex_t pd_index;
2630                 pdp_entry_t *pd;
2631
2632                 pd_index = ptepindex - pmap_pd_pindex(0);
2633
2634                 if (pvp == NULL) {
2635                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2636                                      (pd_index >> NPML4EPGSHIFT);
2637                         pvp = pv_get(pv->pv_pmap, pdp_pindex);
2638                         gotpvp = 1;
2639                 }
2640                 if (pvp) {
2641                         pd = pv_pte_lookup(pvp, pd_index &
2642                                                 ((1ul << NPDPEPGSHIFT) - 1));
2643                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2644                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2645                         pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2646                 } else {
2647                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2648                         p = pv->pv_m;           /* degenerate test later */
2649                 }
2650         } else if (ptepindex >= pmap_pt_pindex(0)) {
2651                 /*
2652                  *  Remove a PT page from the pd
2653                  */
2654                 vm_pindex_t pd_pindex;
2655                 vm_pindex_t pt_index;
2656                 pd_entry_t *pt;
2657
2658                 pt_index = ptepindex - pmap_pt_pindex(0);
2659
2660                 if (pvp == NULL) {
2661                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2662                                     (pt_index >> NPDPEPGSHIFT);
2663                         pvp = pv_get(pv->pv_pmap, pd_pindex);
2664                         KKASSERT(pvp);
2665                         gotpvp = 1;
2666                 }
2667                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2668                 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2669                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2670                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2671         } else {
2672                 /*
2673                  * Remove a PTE from the PT page
2674                  *
2675                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2676                  *       pv is a pte_pv so we can safely lock pt_pv.
2677                  *
2678                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2679                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2680                  *       terminal ptes.
2681                  */
2682                 vm_pindex_t pt_pindex;
2683                 pt_entry_t *ptep;
2684                 pt_entry_t pte;
2685                 vm_offset_t va;
2686
2687                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2688                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2689
2690                 if (ptepindex >= NUPTE_USER) {
2691                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2692                         KKASSERT(pvp == NULL);
2693                 } else {
2694                         if (pvp == NULL) {
2695                                 pt_pindex = NUPTE_TOTAL +
2696                                             (ptepindex >> NPDPEPGSHIFT);
2697                                 pvp = pv_get(pv->pv_pmap, pt_pindex);
2698                                 KKASSERT(pvp);
2699                                 gotpvp = 1;
2700                         }
2701                         ptep = pv_pte_lookup(pvp, ptepindex &
2702                                                   ((1ul << NPDPEPGSHIFT) - 1));
2703                 }
2704                 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2705                 if (bulk == NULL)               /* XXX */
2706                         cpu_invlpg((void *)va); /* XXX */
2707
2708                 /*
2709                  * Now update the vm_page_t
2710                  */
2711                 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) !=
2712                     (pmap->pmap_bits[PG_MANAGED_IDX]|pmap->pmap_bits[PG_V_IDX])) {
2713                         kprintf("remove_pte badpte %016lx %016lx %d\n",
2714                                 pte, pv->pv_pindex,
2715                                 pv->pv_pindex < pmap_pt_pindex(0));
2716                 }
2717                 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2718                 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2719                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2720                         p = pv->pv_m;
2721                 else
2722                         p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2723                 /* p = pv->pv_m; */
2724
2725                 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2726                         if (pmap_track_modified(ptepindex))
2727                                 vm_page_dirty(p);
2728                 }
2729                 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2730                         vm_page_flag_set(p, PG_REFERENCED);
2731                 }
2732                 if (pte & pmap->pmap_bits[PG_W_IDX])
2733                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
2734                 if (pte & pmap->pmap_bits[PG_G_IDX])
2735                         cpu_invlpg((void *)va);
2736         }
2737         KKASSERT(pv->pv_m == p);        /* XXX remove me later */
2738
2739         /*
2740          * If requested, scrap the underlying pv->pv_m and the underlying
2741          * pv.  If this is a page-table-page we must also free the page.
2742          *
2743          * pvp must be returned locked.
2744          */
2745         if (destroy == 1) {
2746                 /*
2747                  * page table page (PT, PD, PDP, PML4), caller was responsible
2748                  * for testing wired_count.
2749                  */
2750                 vm_page_t p;
2751
2752                 KKASSERT(pv->pv_m->wire_count == 1);
2753                 p = pmap_remove_pv_page(pv);
2754                 pv_free(pv, pvp, 1);
2755                 pv = NULL;
2756
2757                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
2758                 vm_page_busy_wait(p, FALSE, "pgpun");
2759                 vm_page_unwire(p, 0);
2760                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2761                 vm_page_free(p);
2762         } else if (destroy == 2) {
2763                 /*
2764                  * Normal page (leave page untouched)
2765                  */
2766                 pmap_remove_pv_page(pv);
2767                 pv_free(pv, pvp, 1);
2768                 pv = NULL;              /* safety */
2769         }
2770
2771         /*
2772          * If we acquired pvp ourselves then we are responsible for
2773          * recursively deleting it.
2774          */
2775         if (pvp && gotpvp) {
2776                 /*
2777                  * Recursively destroy higher-level page tables.
2778                  *
2779                  * This is optional.  If we do not, they will still
2780                  * be destroyed when the process exits.
2781                  */
2782                 if (pvp->pv_m &&
2783                     pvp->pv_m->wire_count == 1 &&
2784                     pvp->pv_pindex != pmap_pml4_pindex()) {
2785                         if (pmap != &kernel_pmap) {
2786                                 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
2787                                 pvp = NULL;     /* safety */
2788                         } else {
2789                                 kprintf("Attempt to remove kernel_pmap pindex "
2790                                         "%jd\n", pvp->pv_pindex);
2791                                 pv_put(pvp);
2792                         }
2793                 } else {
2794                         pv_put(pvp);
2795                 }
2796         }
2797 }
2798
2799 /*
2800  * Remove the vm_page association to a pv.  The pv must be locked.
2801  */
2802 static
2803 vm_page_t
2804 pmap_remove_pv_page(pv_entry_t pv)
2805 {
2806         vm_page_t m;
2807
2808         m = pv->pv_m;
2809         KKASSERT(m);
2810         vm_page_spin_lock(m);
2811         pv->pv_m = NULL;
2812         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2813         pmap_page_stats_deleting(m);
2814         /*
2815         if (m->object)
2816                 atomic_add_int(&m->object->agg_pv_list_count, -1);
2817         */
2818         if (TAILQ_EMPTY(&m->md.pv_list))
2819                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2820         vm_page_spin_unlock(m);
2821
2822         return(m);
2823 }
2824
2825 /*
2826  * Grow the number of kernel page table entries, if needed.
2827  *
2828  * This routine is always called to validate any address space
2829  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
2830  * space below KERNBASE.
2831  */
2832 void
2833 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2834 {
2835         vm_paddr_t paddr;
2836         vm_offset_t ptppaddr;
2837         vm_page_t nkpg;
2838         pd_entry_t *pt, newpt;
2839         pdp_entry_t newpd;
2840         int update_kernel_vm_end;
2841
2842         /*
2843          * bootstrap kernel_vm_end on first real VM use
2844          */
2845         if (kernel_vm_end == 0) {
2846                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2847                 nkpt = 0;
2848                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2849                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2850                                         ~(PAGE_SIZE * NPTEPG - 1);
2851                         nkpt++;
2852                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2853                                 kernel_vm_end = kernel_map.max_offset;
2854                                 break;
2855                         }
2856                 }
2857         }
2858
2859         /*
2860          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
2861          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
2862          * do not want to force-fill 128G worth of page tables.
2863          */
2864         if (kstart < KERNBASE) {
2865                 if (kstart > kernel_vm_end)
2866                         kstart = kernel_vm_end;
2867                 KKASSERT(kend <= KERNBASE);
2868                 update_kernel_vm_end = 1;
2869         } else {
2870                 update_kernel_vm_end = 0;
2871         }
2872
2873         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2874         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2875
2876         if (kend - 1 >= kernel_map.max_offset)
2877                 kend = kernel_map.max_offset;
2878
2879         while (kstart < kend) {
2880                 pt = pmap_pt(&kernel_pmap, kstart);
2881                 if (pt == NULL) {
2882                         /* We need a new PD entry */
2883                         nkpg = vm_page_alloc(NULL, nkpt,
2884                                              VM_ALLOC_NORMAL |
2885                                              VM_ALLOC_SYSTEM |
2886                                              VM_ALLOC_INTERRUPT);
2887                         if (nkpg == NULL) {
2888                                 panic("pmap_growkernel: no memory to grow "
2889                                       "kernel");
2890                         }
2891                         paddr = VM_PAGE_TO_PHYS(nkpg);
2892                         pmap_zero_page(paddr);
2893                         newpd = (pdp_entry_t)
2894                             (paddr |
2895                             kernel_pmap.pmap_bits[PG_V_IDX] |
2896                             kernel_pmap.pmap_bits[PG_RW_IDX] |
2897                             kernel_pmap.pmap_bits[PG_A_IDX] |
2898                             kernel_pmap.pmap_bits[PG_M_IDX]);
2899                         *pmap_pd(&kernel_pmap, kstart) = newpd;
2900                         nkpt++;
2901                         continue; /* try again */
2902                 }
2903                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2904                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
2905                                  ~(PAGE_SIZE * NPTEPG - 1);
2906                         if (kstart - 1 >= kernel_map.max_offset) {
2907                                 kstart = kernel_map.max_offset;
2908                                 break;
2909                         }
2910                         continue;
2911                 }
2912
2913                 /*
2914                  * We need a new PT
2915                  *
2916                  * This index is bogus, but out of the way
2917                  */
2918                 nkpg = vm_page_alloc(NULL, nkpt,
2919                                      VM_ALLOC_NORMAL |
2920                                      VM_ALLOC_SYSTEM |
2921                                      VM_ALLOC_INTERRUPT);
2922                 if (nkpg == NULL)
2923                         panic("pmap_growkernel: no memory to grow kernel");
2924
2925                 vm_page_wire(nkpg);
2926                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2927                 pmap_zero_page(ptppaddr);
2928                 newpt = (pd_entry_t) (ptppaddr |
2929                     kernel_pmap.pmap_bits[PG_V_IDX] |
2930                     kernel_pmap.pmap_bits[PG_RW_IDX] |
2931                     kernel_pmap.pmap_bits[PG_A_IDX] |
2932                     kernel_pmap.pmap_bits[PG_M_IDX]);
2933                 *pmap_pt(&kernel_pmap, kstart) = newpt;
2934                 nkpt++;
2935
2936                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2937                           ~(PAGE_SIZE * NPTEPG - 1);
2938
2939                 if (kstart - 1 >= kernel_map.max_offset) {
2940                         kstart = kernel_map.max_offset;
2941                         break;
2942                 }
2943         }
2944
2945         /*
2946          * Only update kernel_vm_end for areas below KERNBASE.
2947          */
2948         if (update_kernel_vm_end && kernel_vm_end < kstart)
2949                 kernel_vm_end = kstart;
2950 }
2951
2952 /*
2953  *      Add a reference to the specified pmap.
2954  */
2955 void
2956 pmap_reference(pmap_t pmap)
2957 {
2958         if (pmap != NULL) {
2959                 lwkt_gettoken(&pmap->pm_token);
2960                 ++pmap->pm_count;
2961                 lwkt_reltoken(&pmap->pm_token);
2962         }
2963 }
2964
2965 /***************************************************
2966  * page management routines.
2967  ***************************************************/
2968
2969 /*
2970  * Hold a pv without locking it
2971  */
2972 static void
2973 pv_hold(pv_entry_t pv)
2974 {
2975         atomic_add_int(&pv->pv_hold, 1);
2976 }
2977
2978 /*
2979  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
2980  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
2981  * the pv properly.
2982  *
2983  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2984  * pv list via its page) must be held by the caller.
2985  */
2986 static int
2987 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2988 {
2989         u_int count;
2990
2991         /*
2992          * Critical path shortcut expects pv to already have one ref
2993          * (for the pv->pv_pmap).
2994          */
2995         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
2996 #ifdef PMAP_DEBUG
2997                 pv->pv_func = func;
2998                 pv->pv_line = lineno;
2999 #endif
3000                 return TRUE;
3001         }
3002
3003         for (;;) {
3004                 count = pv->pv_hold;
3005                 cpu_ccfence();
3006                 if ((count & PV_HOLD_LOCKED) == 0) {
3007                         if (atomic_cmpset_int(&pv->pv_hold, count,
3008                                               (count + 1) | PV_HOLD_LOCKED)) {
3009 #ifdef PMAP_DEBUG
3010                                 pv->pv_func = func;
3011                                 pv->pv_line = lineno;
3012 #endif
3013                                 return TRUE;
3014                         }
3015                 } else {
3016                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3017                                 return FALSE;
3018                 }
3019                 /* retry */
3020         }
3021 }
3022
3023 /*
3024  * Drop a previously held pv_entry which could not be locked, allowing its
3025  * destruction.
3026  *
3027  * Must not be called with a spinlock held as we might zfree() the pv if it
3028  * is no longer associated with a pmap and this was the last hold count.
3029  */
3030 static void
3031 pv_drop(pv_entry_t pv)
3032 {
3033         u_int count;
3034
3035         for (;;) {
3036                 count = pv->pv_hold;
3037                 cpu_ccfence();
3038                 KKASSERT((count & PV_HOLD_MASK) > 0);
3039                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3040                          (PV_HOLD_LOCKED | 1));
3041                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3042                         if ((count & PV_HOLD_MASK) == 1) {
3043 #ifdef PMAP_DEBUG2
3044                                 if (pmap_enter_debug > 0) {
3045                                         --pmap_enter_debug;
3046                                         kprintf("pv_drop: free pv %p\n", pv);
3047                                 }
3048 #endif
3049                                 KKASSERT(count == 1);
3050                                 KKASSERT(pv->pv_pmap == NULL);
3051                                 zfree(pvzone, pv);
3052                         }
3053                         return;
3054                 }
3055                 /* retry */
3056         }
3057 }
3058
3059 /*
3060  * Find or allocate the requested PV entry, returning a locked, held pv.
3061  *
3062  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3063  * for the caller and one representing the pmap and vm_page association.
3064  *
3065  * If (*isnew) is zero, the returned pv will have only one hold count.
3066  *
3067  * Since both associations can only be adjusted while the pv is locked,
3068  * together they represent just one additional hold.
3069  */
3070 static
3071 pv_entry_t
3072 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3073 {
3074         pv_entry_t pv;
3075         pv_entry_t pnew = NULL;
3076
3077         spin_lock(&pmap->pm_spin);
3078         for (;;) {
3079                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3080                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3081                                                         pindex);
3082                 }
3083                 if (pv == NULL) {
3084                         if (pnew == NULL) {
3085                                 spin_unlock(&pmap->pm_spin);
3086                                 pnew = zalloc(pvzone);
3087                                 spin_lock(&pmap->pm_spin);
3088                                 continue;
3089                         }
3090                         pnew->pv_pmap = pmap;
3091                         pnew->pv_pindex = pindex;
3092                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
3093 #ifdef PMAP_DEBUG
3094                         pnew->pv_func = func;
3095                         pnew->pv_line = lineno;
3096 #endif
3097                         pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3098                         ++pmap->pm_generation;
3099                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
3100                         spin_unlock(&pmap->pm_spin);
3101                         *isnew = 1;
3102                         return(pnew);
3103                 }
3104                 if (pnew) {
3105                         spin_unlock(&pmap->pm_spin);
3106                         zfree(pvzone, pnew);
3107                         pnew = NULL;
3108                         spin_lock(&pmap->pm_spin);
3109                         continue;
3110                 }
3111                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3112                         spin_unlock(&pmap->pm_spin);
3113                 } else {
3114                         spin_unlock(&pmap->pm_spin);
3115                         _pv_lock(pv PMAP_DEBUG_COPY);
3116                 }
3117                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3118                         *isnew = 0;
3119                         return(pv);
3120                 }
3121                 pv_put(pv);
3122                 spin_lock(&pmap->pm_spin);
3123         }
3124 }
3125
3126 /*
3127  * Find the requested PV entry, returning a locked+held pv or NULL
3128  */
3129 static
3130 pv_entry_t
3131 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
3132 {
3133         pv_entry_t pv;
3134
3135         spin_lock(&pmap->pm_spin);
3136         for (;;) {
3137                 /*
3138                  * Shortcut cache
3139                  */
3140                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3141                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3142                                                         pindex);
3143                 }
3144                 if (pv == NULL) {
3145                         spin_unlock(&pmap->pm_spin);
3146                         return NULL;
3147                 }
3148                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3149                         spin_unlock(&pmap->pm_spin);
3150                 } else {
3151                         spin_unlock(&pmap->pm_spin);
3152                         _pv_lock(pv PMAP_DEBUG_COPY);
3153                 }
3154                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3155                         pv_cache(pv, pindex);
3156                         return(pv);
3157                 }
3158                 pv_put(pv);
3159                 spin_lock(&pmap->pm_spin);
3160         }
3161 }
3162
3163 /*
3164  * Lookup, hold, and attempt to lock (pmap,pindex).
3165  *
3166  * If the entry does not exist NULL is returned and *errorp is set to 0
3167  *
3168  * If the entry exists and could be successfully locked it is returned and
3169  * errorp is set to 0.
3170  *
3171  * If the entry exists but could NOT be successfully locked it is returned
3172  * held and *errorp is set to 1.
3173  */
3174 static
3175 pv_entry_t
3176 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
3177 {
3178         pv_entry_t pv;
3179
3180         spin_lock_shared(&pmap->pm_spin);
3181         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3182                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3183         if (pv == NULL) {
3184                 spin_unlock_shared(&pmap->pm_spin);
3185                 *errorp = 0;
3186                 return NULL;
3187         }
3188         if (pv_hold_try(pv)) {
3189                 pv_cache(pv, pindex);
3190                 spin_unlock_shared(&pmap->pm_spin);
3191                 *errorp = 0;
3192                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3193                 return(pv);     /* lock succeeded */
3194         }
3195         spin_unlock_shared(&pmap->pm_spin);
3196         *errorp = 1;
3197         return (pv);            /* lock failed */
3198 }
3199
3200 /*
3201  * Find the requested PV entry, returning a held pv or NULL
3202  */
3203 static
3204 pv_entry_t
3205 pv_find(pmap_t pmap, vm_pindex_t pindex)
3206 {
3207         pv_entry_t pv;
3208
3209         spin_lock_shared(&pmap->pm_spin);
3210
3211         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3212                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3213         if (pv == NULL) {
3214                 spin_unlock_shared(&pmap->pm_spin);
3215                 return NULL;
3216         }
3217         pv_hold(pv);
3218         pv_cache(pv, pindex);
3219         spin_unlock_shared(&pmap->pm_spin);
3220         return(pv);
3221 }
3222
3223 /*
3224  * Lock a held pv, keeping the hold count
3225  */
3226 static
3227 void
3228 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3229 {
3230         u_int count;
3231
3232         for (;;) {
3233                 count = pv->pv_hold;
3234                 cpu_ccfence();
3235                 if ((count & PV_HOLD_LOCKED) == 0) {
3236                         if (atomic_cmpset_int(&pv->pv_hold, count,
3237                                               count | PV_HOLD_LOCKED)) {
3238 #ifdef PMAP_DEBUG
3239                                 pv->pv_func = func;
3240                                 pv->pv_line = lineno;
3241 #endif
3242                                 return;
3243                         }
3244                         continue;
3245                 }
3246                 tsleep_interlock(pv, 0);
3247                 if (atomic_cmpset_int(&pv->pv_hold, count,
3248                                       count | PV_HOLD_WAITING)) {
3249 #ifdef PMAP_DEBUG
3250                         kprintf("pv waiting on %s:%d\n",
3251                                         pv->pv_func, pv->pv_line);
3252 #endif
3253                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
3254                 }
3255                 /* retry */
3256         }
3257 }
3258
3259 /*
3260  * Unlock a held and locked pv, keeping the hold count.
3261  */
3262 static
3263 void
3264 pv_unlock(pv_entry_t pv)
3265 {
3266         u_int count;
3267
3268         for (;;) {
3269                 count = pv->pv_hold;
3270                 cpu_ccfence();
3271                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3272                          (PV_HOLD_LOCKED | 1));
3273                 if (atomic_cmpset_int(&pv->pv_hold, count,
3274                                       count &
3275                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3276                         if (count & PV_HOLD_WAITING)
3277                                 wakeup(pv);
3278                         break;
3279                 }
3280         }
3281 }
3282
3283 /*
3284  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3285  * and the hold count drops to zero we will free it.
3286  *
3287  * Caller should not hold any spin locks.  We are protected from hold races
3288  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3289  * lock held.  A pv cannot be located otherwise.
3290  */
3291 static
3292 void
3293 pv_put(pv_entry_t pv)
3294 {
3295 #ifdef PMAP_DEBUG2
3296         if (pmap_enter_debug > 0) {
3297                 --pmap_enter_debug;
3298                 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3299         }
3300 #endif
3301
3302         /*
3303          * Fast - shortcut most common condition
3304          */
3305         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3306                 return;
3307
3308         /*
3309          * Slow
3310          */
3311         pv_unlock(pv);
3312         pv_drop(pv);
3313 }
3314
3315 /*
3316  * Remove the pmap association from a pv, require that pv_m already be removed,
3317  * then unlock and drop the pv.  Any pte operations must have already been
3318  * completed.  This call may result in a last-drop which will physically free
3319  * the pv.
3320  *
3321  * Removing the pmap association entails an additional drop.
3322  *
3323  * pv must be exclusively locked on call and will be disposed of on return.
3324  */
3325 static
3326 void
3327 pv_free(pv_entry_t pv, pv_entry_t pvp, int putaway)
3328 {
3329         pmap_t pmap;
3330
3331         KKASSERT(pv->pv_m == NULL);
3332         KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3333         if ((pmap = pv->pv_pmap) != NULL) {
3334                 spin_lock(&pmap->pm_spin);
3335                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3336                 ++pmap->pm_generation;
3337                 if (pmap->pm_pvhint == pv)
3338                         pmap->pm_pvhint = NULL;
3339                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3340                 pv->pv_pmap = NULL;
3341                 pv->pv_pindex = 0;
3342                 spin_unlock(&pmap->pm_spin);
3343
3344                 /*
3345                  * Try to shortcut three atomic ops, otherwise fall through
3346                  * and do it normally.  Drop two refs and the lock all in
3347                  * one go.
3348                  */
3349                 if (putaway &&
3350                     atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3351 #ifdef PMAP_DEBUG2
3352                         if (pmap_enter_debug > 0) {
3353                                 --pmap_enter_debug;
3354                                 kprintf("pv_free: free pv %p\n", pv);
3355                         }
3356 #endif
3357                         zfree(pvzone, pv);
3358                         if (pvp)
3359                                 vm_page_unwire_quick(pvp->pv_m);
3360                         return;
3361                 }
3362                 pv_drop(pv);    /* ref for pv_pmap */
3363         }
3364         if (putaway)
3365                 pv_put(pv);
3366         if (pvp)
3367                 vm_page_unwire_quick(pvp->pv_m);
3368 }
3369
3370 /*
3371  * This routine is very drastic, but can save the system
3372  * in a pinch.
3373  */
3374 void
3375 pmap_collect(void)
3376 {
3377         int i;
3378         vm_page_t m;
3379         static int warningdone=0;
3380
3381         if (pmap_pagedaemon_waken == 0)
3382                 return;
3383         pmap_pagedaemon_waken = 0;
3384         if (warningdone < 5) {
3385                 kprintf("pmap_collect: collecting pv entries -- "
3386                         "suggest increasing PMAP_SHPGPERPROC\n");
3387                 warningdone++;
3388         }
3389
3390         for (i = 0; i < vm_page_array_size; i++) {
3391                 m = &vm_page_array[i];
3392                 if (m->wire_count || m->hold_count)
3393                         continue;
3394                 if (vm_page_busy_try(m, TRUE) == 0) {
3395                         if (m->wire_count == 0 && m->hold_count == 0) {
3396                                 pmap_remove_all(m);
3397                         }
3398                         vm_page_wakeup(m);
3399                 }
3400         }
3401 }
3402
3403 /*
3404  * Scan the pmap for active page table entries and issue a callback.
3405  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3406  * its parent page table.
3407  *
3408  * pte_pv will be NULL if the page or page table is unmanaged.
3409  * pt_pv will point to the page table page containing the pte for the page.
3410  *
3411  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3412  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3413  *       process pmap's PD and page to the callback function.  This can be
3414  *       confusing because the pt_pv is really a pd_pv, and the target page
3415  *       table page is simply aliased by the pmap and not owned by it.
3416  *
3417  * It is assumed that the start and end are properly rounded to the page size.
3418  *
3419  * It is assumed that PD pages and above are managed and thus in the RB tree,
3420  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3421  */
3422 struct pmap_scan_info {
3423         struct pmap *pmap;
3424         vm_offset_t sva;
3425         vm_offset_t eva;
3426         vm_pindex_t sva_pd_pindex;
3427         vm_pindex_t eva_pd_pindex;
3428         void (*func)(pmap_t, struct pmap_scan_info *,
3429                      pv_entry_t, pv_entry_t, int, vm_offset_t,
3430                      pt_entry_t *, void *);
3431         void *arg;
3432         pmap_inval_bulk_t bulk_core;
3433         pmap_inval_bulk_t *bulk;
3434         int count;
3435         int stop;
3436 };
3437
3438 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3439 static int pmap_scan_callback(pv_entry_t pv, void *data);
3440
3441 static void
3442 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3443 {
3444         struct pmap *pmap = info->pmap;
3445         pv_entry_t pd_pv;       /* A page directory PV */
3446         pv_entry_t pt_pv;       /* A page table PV */
3447         pv_entry_t pte_pv;      /* A page table entry PV */
3448         pt_entry_t *ptep;
3449         pt_entry_t oldpte;
3450         struct pv_entry dummy_pv;
3451         int generation;
3452
3453         info->stop = 0;
3454         if (pmap == NULL)
3455                 return;
3456         if (smp_inval) {
3457                 info->bulk = &info->bulk_core;
3458                 pmap_inval_bulk_init(&info->bulk_core, pmap);
3459         } else {
3460                 info->bulk = NULL;
3461         }
3462
3463         /*
3464          * Hold the token for stability; if the pmap is empty we have nothing
3465          * to do.
3466          */
3467         lwkt_gettoken(&pmap->pm_token);
3468 #if 0
3469         if (pmap->pm_stats.resident_count == 0) {
3470                 lwkt_reltoken(&pmap->pm_token);
3471                 return;
3472         }
3473 #endif
3474
3475         info->count = 0;
3476
3477 again:
3478         /*
3479          * Special handling for scanning one page, which is a very common
3480          * operation (it is?).
3481          *
3482          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3483          */
3484         if (info->sva + PAGE_SIZE == info->eva) {
3485                 generation = pmap->pm_generation;
3486                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3487                         /*
3488                          * Kernel mappings do not track wire counts on
3489                          * page table pages and only maintain pd_pv and
3490                          * pte_pv levels so pmap_scan() works.
3491                          */
3492                         pt_pv = NULL;
3493                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3494                         ptep = vtopte(info->sva);
3495                 } else {
3496                         /*
3497                          * User pages which are unmanaged will not have a
3498                          * pte_pv.  User page table pages which are unmanaged
3499                          * (shared from elsewhere) will also not have a pt_pv.
3500                          * The func() callback will pass both pte_pv and pt_pv
3501                          * as NULL in that case.
3502                          */
3503                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3504                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3505                         if (pt_pv == NULL) {
3506                                 KKASSERT(pte_pv == NULL);
3507                                 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3508                                 if (pd_pv) {
3509                                         ptep = pv_pte_lookup(pd_pv,
3510                                                     pmap_pt_index(info->sva));
3511                                         if (*ptep) {
3512                                                 info->func(pmap, info,
3513                                                      NULL, pd_pv, 1,
3514                                                      info->sva, ptep,
3515                                                      info->arg);
3516                                         }
3517                                         pv_put(pd_pv);
3518                                 }
3519                                 goto fast_skip;
3520                         }
3521                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3522                 }
3523
3524                 /*
3525                  * NOTE: *ptep can't be ripped out from under us if we hold
3526                  *       pte_pv locked, but bits can change.  However, there is
3527                  *       a race where another thread may be inserting pte_pv
3528                  *       and setting *ptep just after our pte_pv lookup fails.
3529                  *
3530                  *       In this situation we can end up with a NULL pte_pv
3531                  *       but find that we have a managed *ptep.  We explicitly
3532                  *       check for this race.
3533                  */
3534                 oldpte = *ptep;
3535                 cpu_ccfence();
3536                 if (oldpte == 0) {
3537                         /*
3538                          * Unlike the pv_find() case below we actually
3539                          * acquired a locked pv in this case so any
3540                          * race should have been resolved.  It is expected
3541                          * to not exist.
3542                          */
3543                         KKASSERT(pte_pv == NULL);
3544                 } else if (pte_pv) {
3545                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3546                                            pmap->pmap_bits[PG_V_IDX])) ==
3547                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
3548                                  pmap->pmap_bits[PG_V_IDX]),
3549                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3550                              "generation %d/%d",
3551                             *ptep, oldpte, info->sva, pte_pv,
3552                             generation, pmap->pm_generation));
3553                         info->func(pmap, info, pte_pv, pt_pv, 0,
3554                                    info->sva, ptep, info->arg);
3555                 } else {
3556                         /*
3557                          * Check for insertion race
3558                          */
3559                         if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3560                             pt_pv) {
3561                                 pte_pv = pv_find(pmap,
3562                                                  pmap_pte_pindex(info->sva));
3563                                 if (pte_pv) {
3564                                         pv_drop(pte_pv);
3565                                         pv_put(pt_pv);
3566                                         kprintf("pmap_scan: RACE1 "
3567                                                 "%016jx, %016lx\n",
3568                                                 info->sva, oldpte);
3569                                         goto again;
3570                                 }
3571                         }
3572
3573                         /*
3574                          * Didn't race
3575                          */
3576                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3577                                            pmap->pmap_bits[PG_V_IDX])) ==
3578                             pmap->pmap_bits[PG_V_IDX],
3579                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3580                              "generation %d/%d",
3581                             *ptep, oldpte, info->sva,
3582                             generation, pmap->pm_generation));
3583                         info->func(pmap, info, NULL, pt_pv, 0,
3584                             info->sva, ptep, info->arg);
3585                 }
3586                 if (pt_pv)
3587                         pv_put(pt_pv);
3588 fast_skip:
3589                 pmap_inval_bulk_flush(info->bulk);
3590                 lwkt_reltoken(&pmap->pm_token);
3591                 return;
3592         }
3593
3594         /*
3595          * Nominal scan case, RB_SCAN() for PD pages and iterate from
3596          * there.
3597          */
3598         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3599         info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3600
3601         if (info->sva >= VM_MAX_USER_ADDRESS) {
3602                 /*
3603                  * The kernel does not currently maintain any pv_entry's for
3604                  * higher-level page tables.
3605                  */
3606                 bzero(&dummy_pv, sizeof(dummy_pv));
3607                 dummy_pv.pv_pindex = info->sva_pd_pindex;
3608                 spin_lock(&pmap->pm_spin);
3609                 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3610                         pmap_scan_callback(&dummy_pv, info);
3611                         ++dummy_pv.pv_pindex;
3612                 }
3613                 spin_unlock(&pmap->pm_spin);
3614         } else {
3615                 /*
3616                  * User page tables maintain local PML4, PDP, and PD
3617                  * pv_entry's at the very least.  PT pv's might be
3618                  * unmanaged and thus not exist.  PTE pv's might be
3619                  * unmanaged and thus not exist.
3620                  */
3621                 spin_lock(&pmap->pm_spin);
3622                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3623                         pmap_scan_cmp, pmap_scan_callback, info);
3624                 spin_unlock(&pmap->pm_spin);
3625         }
3626         pmap_inval_bulk_flush(info->bulk);
3627         lwkt_reltoken(&pmap->pm_token);
3628 }
3629
3630 /*
3631  * WARNING! pmap->pm_spin held
3632  */
3633 static int
3634 pmap_scan_cmp(pv_entry_t pv, void *data)
3635 {
3636         struct pmap_scan_info *info = data;
3637         if (pv->pv_pindex < info->sva_pd_pindex)
3638                 return(-1);
3639         if (pv->pv_pindex >= info->eva_pd_pindex)
3640                 return(1);
3641         return(0);
3642 }
3643
3644 /*
3645  * WARNING! pmap->pm_spin held
3646  */
3647 static int
3648 pmap_scan_callback(pv_entry_t pv, void *data)
3649 {
3650         struct pmap_scan_info *info = data;
3651         struct pmap *pmap = info->pmap;
3652         pv_entry_t pd_pv;       /* A page directory PV */
3653         pv_entry_t pt_pv;       /* A page table PV */
3654         pv_entry_t pte_pv;      /* A page table entry PV */
3655         pt_entry_t *ptep;
3656         pt_entry_t oldpte;
3657         vm_offset_t sva;
3658         vm_offset_t eva;
3659         vm_offset_t va_next;
3660         vm_pindex_t pd_pindex;
3661         int error;
3662         int generation;
3663
3664         /*
3665          * Stop if requested
3666          */
3667         if (info->stop)
3668                 return -1;
3669
3670         /*
3671          * Pull the PD pindex from the pv before releasing the spinlock.
3672          *
3673          * WARNING: pv is faked for kernel pmap scans.
3674          */
3675         pd_pindex = pv->pv_pindex;
3676         spin_unlock(&pmap->pm_spin);
3677         pv = NULL;      /* invalid after spinlock unlocked */
3678
3679         /*
3680          * Calculate the page range within the PD.  SIMPLE pmaps are
3681          * direct-mapped for the entire 2^64 address space.  Normal pmaps
3682          * reflect the user and kernel address space which requires
3683          * cannonicalization w/regards to converting pd_pindex's back
3684          * into addresses.
3685          */
3686         sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3687         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3688             (sva & PML4_SIGNMASK)) {
3689                 sva |= PML4_SIGNMASK;
3690         }
3691         eva = sva + NBPDP;      /* can overflow */
3692         if (sva < info->sva)
3693                 sva = info->sva;
3694         if (eva < info->sva || eva > info->eva)
3695                 eva = info->eva;
3696
3697         /*
3698          * NOTE: kernel mappings do not track page table pages, only
3699          *       terminal pages.
3700          *
3701          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3702          *       However, for the scan to be efficient we try to
3703          *       cache items top-down.
3704          */
3705         pd_pv = NULL;
3706         pt_pv = NULL;
3707
3708         for (; sva < eva; sva = va_next) {
3709                 if (info->stop)
3710                         break;
3711                 if (sva >= VM_MAX_USER_ADDRESS) {
3712                         if (pt_pv) {
3713                                 pv_put(pt_pv);
3714                                 pt_pv = NULL;
3715                         }
3716                         goto kernel_skip;
3717                 }
3718
3719                 /*
3720                  * PD cache (degenerate case if we skip).  It is possible
3721                  * for the PD to not exist due to races.  This is ok.
3722                  */
3723                 if (pd_pv == NULL) {
3724                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3725                 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3726                         pv_put(pd_pv);
3727                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3728                 }
3729                 if (pd_pv == NULL) {
3730                         va_next = (sva + NBPDP) & ~PDPMASK;
3731                         if (va_next < sva)
3732                                 va_next = eva;
3733                         continue;
3734                 }
3735
3736                 /*
3737                  * PT cache
3738                  */
3739                 if (pt_pv == NULL) {
3740                         vm_page_wire_quick(pd_pv->pv_m);
3741                         pv_unlock(pd_pv);
3742                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3743                         pv_lock(pd_pv);
3744                         vm_page_unwire_quick(pd_pv->pv_m);
3745                 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3746                         vm_page_wire_quick(pd_pv->pv_m);
3747                         pv_unlock(pd_pv);
3748                         pv_put(pt_pv);
3749                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3750                         pv_lock(pd_pv);
3751                         vm_page_unwire_quick(pd_pv->pv_m);
3752                 }
3753
3754                 /*
3755                  * If pt_pv is NULL we either have an shared page table
3756                  * page and must issue a callback specific to that case,
3757                  * or there is no page table page.
3758                  *
3759                  * Either way we can skip the page table page.
3760                  */
3761                 if (pt_pv == NULL) {
3762                         /*
3763                          * Possible unmanaged (shared from another pmap)
3764                          * page table page.
3765                          */
3766                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3767                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3768                                 info->func(pmap, info, NULL, pd_pv, 1,
3769                                            sva, ptep, info->arg);
3770                         }
3771
3772                         /*
3773                          * Done, move to next page table page.
3774                          */
3775                         va_next = (sva + NBPDR) & ~PDRMASK;
3776                         if (va_next < sva)
3777                                 va_next = eva;
3778                         continue;
3779                 }
3780
3781                 /*
3782                  * From this point in the loop testing pt_pv for non-NULL
3783                  * means we are in UVM, else if it is NULL we are in KVM.
3784                  *
3785                  * Limit our scan to either the end of the va represented
3786                  * by the current page table page, or to the end of the
3787                  * range being removed.
3788                  */
3789 kernel_skip:
3790                 va_next = (sva + NBPDR) & ~PDRMASK;
3791                 if (va_next < sva)
3792                         va_next = eva;
3793                 if (va_next > eva)
3794                         va_next = eva;
3795
3796                 /*
3797                  * Scan the page table for pages.  Some pages may not be
3798                  * managed (might not have a pv_entry).
3799                  *
3800                  * There is no page table management for kernel pages so
3801                  * pt_pv will be NULL in that case, but otherwise pt_pv
3802                  * is non-NULL, locked, and referenced.
3803                  */
3804
3805                 /*
3806                  * At this point a non-NULL pt_pv means a UVA, and a NULL
3807                  * pt_pv means a KVA.
3808                  */
3809                 if (pt_pv)
3810                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3811                 else
3812                         ptep = vtopte(sva);
3813
3814                 while (sva < va_next) {
3815                         /*
3816                          * Yield every 64 pages, stop if requested.
3817                          */
3818                         if ((++info->count & 63) == 0)
3819                                 lwkt_user_yield();
3820                         if (info->stop)
3821                                 break;
3822
3823                         /*
3824                          * Check if pt_pv has been lost (probably due to
3825                          * a remove of the underlying pages).
3826                          */
3827                         if (pt_pv && pt_pv->pv_pmap == NULL)
3828                                 break;
3829
3830                         /*
3831                          * Acquire the related pte_pv, if any.  If *ptep == 0
3832                          * the related pte_pv should not exist, but if *ptep
3833                          * is not zero the pte_pv may or may not exist (e.g.
3834                          * will not exist for an unmanaged page).
3835                          *
3836                          * However a multitude of races are possible here.
3837                          *
3838                          * In addition, the (pt_pv, pte_pv) lock order is
3839                          * backwards, so we have to be careful in aquiring
3840                          * a properly locked pte_pv.
3841                          */
3842                         generation = pmap->pm_generation;
3843                         if (pt_pv) {
3844                                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3845                                                     &error);
3846                                 if (error) {
3847                                         if (pd_pv) {
3848                                                 vm_page_wire_quick(pd_pv->pv_m);
3849                                                 pv_unlock(pd_pv);
3850                                         }
3851                                         vm_page_wire_quick(pt_pv->pv_m);
3852                                         pv_unlock(pt_pv);/* must be non-NULL */
3853                                         pv_lock(pte_pv); /* safe to block now */
3854                                         pv_put(pte_pv);
3855                                         pte_pv = NULL;
3856                                         pv_lock(pt_pv);
3857                                         vm_page_unwire_quick(pt_pv->pv_m);
3858
3859                                         /*
3860                                          * pt_pv reloaded, need new ptep
3861                                          */
3862                                         KKASSERT(pt_pv != NULL);
3863                                         ptep = pv_pte_lookup(pt_pv,
3864                                                         pmap_pte_index(sva));
3865                                         if (pd_pv) {
3866                                                 pv_lock(pd_pv);
3867                                                 vm_page_unwire_quick(pd_pv->pv_m);
3868                                         }
3869                                         continue;
3870                                 }
3871                         } else {
3872                                 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3873                         }
3874
3875                         /*
3876                          * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3877                          */
3878                         oldpte = *ptep;
3879                         if (oldpte == 0) {
3880                                 if (pte_pv) {
3881                                         kprintf("Unexpected non-NULL pte_pv "
3882                                                 "%p pt_pv %p "
3883                                                 "*ptep = %016lx/%016lx\n",
3884                                                 pte_pv, pt_pv, *ptep, oldpte);
3885                                         panic("Unexpected non-NULL pte_pv");
3886                                 }
3887                                 sva += PAGE_SIZE;
3888                                 ++ptep;
3889                                 continue;
3890                         }
3891
3892                         /*
3893                          * Ready for the callback.  The locked pte_pv (if any)
3894                          * is consumed by the callback.  pte_pv will exist if
3895                          * the page is managed, and will not exist if it
3896                          * isn't.
3897                          */
3898                         if (pte_pv) {
3899                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3900                                     (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3901                                     ("badC *ptep %016lx/%016lx sva %016lx "
3902                                     "pte_pv %p pm_generation %d/%d",
3903                                     *ptep, oldpte, sva, pte_pv,
3904                                     generation, pmap->pm_generation));
3905                                 /*
3906                                  * We must unlock pd_pv across the callback
3907                                  * to avoid deadlocks on any recursive
3908                                  * disposal.  Re-check that it still exists
3909                                  * after re-locking.
3910                                  */
3911                                 if (pd_pv)
3912                                         pv_unlock(pd_pv);
3913                                 info->func(pmap, info, pte_pv, pt_pv, 0,
3914                                            sva, ptep, info->arg);
3915                                 if (pd_pv) {
3916                                         pv_lock(pd_pv);
3917                                         if (pd_pv->pv_pmap == NULL) {
3918                                                 pv_put(pd_pv);
3919                                                 pd_pv = NULL;
3920                                         }
3921                                 }
3922                         } else {
3923                                 /*
3924                                  * Check for insertion race.  Since there is no
3925                                  * pte_pv to guard us it is possible for us
3926                                  * to race another thread doing an insertion.
3927                                  * Our lookup misses the pte_pv but our *ptep
3928                                  * check sees the inserted pte.
3929                                  *
3930                                  * XXX panic case seems to occur within a
3931                                  * vm_fork() of /bin/sh, which frankly
3932                                  * shouldn't happen since no other threads
3933                                  * should be inserting to our pmap in that
3934                                  * situation.  Removing, possibly.  Inserting,
3935                                  * shouldn't happen.
3936                                  */
3937                                 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3938                                     pt_pv) {
3939                                         pte_pv = pv_find(pmap,
3940                                                          pmap_pte_pindex(sva));
3941                                         if (pte_pv) {
3942                                                 pv_drop(pte_pv);
3943                                                 kprintf("pmap_scan: RACE2 "
3944                                                         "%016jx, %016lx\n",
3945                                                         sva, oldpte);
3946                                                 continue;
3947                                         }
3948                                 }
3949
3950                                 /*
3951                                  * Didn't race
3952                                  *
3953                                  * We must unlock pd_pv across the callback
3954                                  * to avoid deadlocks on any recursive
3955                                  * disposal.  Re-check that it still exists
3956                                  * after re-locking.
3957                                  */
3958                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3959                                     pmap->pmap_bits[PG_V_IDX],
3960                                     ("badD *ptep %016lx/%016lx sva %016lx "
3961                                     "pte_pv NULL pm_generation %d/%d",
3962                                      *ptep, oldpte, sva,
3963                                      generation, pmap->pm_generation));
3964                                 if (pd_pv)
3965                                         pv_unlock(pd_pv);
3966                                 info->func(pmap, info, NULL, pt_pv, 0,
3967                                            sva, ptep, info->arg);
3968                                 if (pd_pv) {
3969                                         pv_lock(pd_pv);
3970                                         if (pd_pv->pv_pmap == NULL) {
3971                                                 pv_put(pd_pv);
3972                                                 pd_pv = NULL;
3973                                         }
3974                                 }
3975                         }
3976                         pte_pv = NULL;
3977                         sva += PAGE_SIZE;
3978                         ++ptep;
3979                 }
3980         }
3981         if (pd_pv) {
3982                 pv_put(pd_pv);
3983                 pd_pv = NULL;
3984         }
3985         if (pt_pv) {
3986                 pv_put(pt_pv);
3987                 pt_pv = NULL;
3988         }
3989         if ((++info->count & 7) == 0)
3990                 lwkt_user_yield();
3991
3992         /*
3993          * Relock before returning.
3994          */
3995         spin_lock(&pmap->pm_spin);
3996         return (0);
3997 }
3998
3999 void
4000 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4001 {
4002         struct pmap_scan_info info;
4003
4004         info.pmap = pmap;
4005         info.sva = sva;
4006         info.eva = eva;
4007         info.func = pmap_remove_callback;
4008         info.arg = NULL;
4009         pmap_scan(&info, 1);
4010 }
4011
4012 static void
4013 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4014 {
4015         struct pmap_scan_info info;
4016
4017         info.pmap = pmap;
4018         info.sva = sva;
4019         info.eva = eva;
4020         info.func = pmap_remove_callback;
4021         info.arg = NULL;
4022         pmap_scan(&info, 0);
4023 }
4024
4025 static void
4026 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4027                      pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4028                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4029 {
4030         pt_entry_t pte;
4031
4032         if (pte_pv) {
4033                 /*
4034                  * This will also drop pt_pv's wire_count. Note that
4035                  * terminal pages are not wired based on mmu presence.
4036                  *
4037                  * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4038                  */
4039                 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4040                 pte_pv = NULL;  /* safety */
4041
4042                 /*
4043                  * Recursively destroy higher-level page tables.
4044                  *
4045                  * This is optional.  If we do not, they will still
4046                  * be destroyed when the process exits.
4047                  */
4048                 if (pt_pv && pt_pv->pv_m && pt_pv->pv_m->wire_count == 1 &&
4049                     pt_pv->pv_pindex != pmap_pml4_pindex()) {
4050                         pv_hold(pt_pv);
4051                         pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4052                         pv_lock(pt_pv);
4053                 }
4054         } else if (sharept == 0) {
4055                 /*
4056                  * Unmanaged page table (pt, pd, or pdp. Not pte).
4057                  *
4058                  * pt_pv's wire_count is still bumped by unmanaged pages
4059                  * so we must decrement it manually.
4060                  *
4061                  * We have to unwire the target page table page.
4062                  *
4063                  * It is unclear how we can invalidate a segment so we
4064                  * invalidate -1 which invlidates the tlb.
4065                  */
4066                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4067                 if (pte & pmap->pmap_bits[PG_W_IDX])
4068                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4069                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4070                 if (vm_page_unwire_quick(pt_pv->pv_m))
4071                         panic("pmap_remove: insufficient wirecount");
4072         } else {
4073                 /*
4074                  * Unmanaged page table (pt, pd, or pdp. Not pte) for
4075                  * a shared page table.
4076                  *
4077                  * pt_pv is actually the pd_pv for our pmap (not the shared
4078                  * object pmap).
4079                  *
4080                  * We have to unwire the target page table page and we
4081                  * have to unwire our page directory page.
4082                  *
4083                  * It is unclear how we can invalidate a segment so we
4084                  * invalidate -1 which invlidates the tlb.
4085                  */
4086                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4087                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4088                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4089                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4090                         panic("pmap_remove: shared pgtable1 bad wirecount");
4091                 if (vm_page_unwire_quick(pt_pv->pv_m))
4092                         panic("pmap_remove: shared pgtable2 bad wirecount");
4093         }
4094 }
4095
4096 /*
4097  * Removes this physical page from all physical maps in which it resides.
4098  * Reflects back modify bits to the pager.
4099  *
4100  * This routine may not be called from an interrupt.
4101  */
4102 static
4103 void
4104 pmap_remove_all(vm_page_t m)
4105 {
4106         pv_entry_t pv;
4107         pmap_inval_bulk_t bulk;
4108
4109         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4110                 return;
4111
4112         vm_page_spin_lock(m);
4113         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4114                 KKASSERT(pv->pv_m == m);
4115                 if (pv_hold_try(pv)) {
4116                         vm_page_spin_unlock(m);
4117                 } else {
4118                         vm_page_spin_unlock(m);
4119                         pv_lock(pv);
4120                 }
4121                 if (pv->pv_m != m) {
4122                         pv_put(pv);
4123                         vm_page_spin_lock(m);
4124                         continue;
4125                 }
4126
4127                 /*
4128                  * Holding no spinlocks, pv is locked.
4129                  */
4130                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4131                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4132                 pv = NULL;      /* safety */
4133                 pmap_inval_bulk_flush(&bulk);
4134 #if 0
4135                 pmap_remove_pv_page(pv);
4136                 pv_free(pv, 1);
4137 #endif
4138                 vm_page_spin_lock(m);
4139         }
4140         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4141         vm_page_spin_unlock(m);
4142 }
4143
4144 /*
4145  * Removes the page from a particular pmap
4146  */
4147 void
4148 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4149 {
4150         pv_entry_t pv;
4151         pmap_inval_bulk_t bulk;
4152
4153         if (!pmap_initialized)
4154                 return;
4155
4156 again:
4157         vm_page_spin_lock(m);
4158         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4159                 if (pv->pv_pmap != pmap)
4160                         continue;
4161                 KKASSERT(pv->pv_m == m);
4162                 if (pv_hold_try(pv)) {
4163                         vm_page_spin_unlock(m);
4164                 } else {
4165                         vm_page_spin_unlock(m);
4166                         pv_lock(pv);
4167                 }
4168                 if (pv->pv_m != m) {
4169                         pv_put(pv);
4170                         goto again;
4171                 }
4172
4173                 /*
4174                  * Holding no spinlocks, pv is locked.
4175                  */
4176                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4177                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4178                 pv = NULL;      /* safety */
4179                 pmap_inval_bulk_flush(&bulk);
4180 #if 0
4181                 pmap_remove_pv_page(pv);
4182                 pv_free(pv, 1);
4183 #endif
4184                 goto again;
4185         }
4186         vm_page_spin_unlock(m);
4187 }
4188
4189 /*
4190  * Set the physical protection on the specified range of this map
4191  * as requested.  This function is typically only used for debug watchpoints
4192  * and COW pages.
4193  *
4194  * This function may not be called from an interrupt if the map is
4195  * not the kernel_pmap.
4196  *
4197  * NOTE!  For shared page table pages we just unmap the page.
4198  */
4199 void
4200 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4201 {
4202         struct pmap_scan_info info;
4203         /* JG review for NX */
4204
4205         if (pmap == NULL)
4206                 return;
4207         if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
4208                 pmap_remove(pmap, sva, eva);
4209                 return;
4210         }
4211         if (prot & VM_PROT_WRITE)
4212                 return;
4213         info.pmap = pmap;
4214         info.sva = sva;
4215         info.eva = eva;
4216         info.func = pmap_protect_callback;
4217         info.arg = &prot;
4218         pmap_scan(&info, 1);
4219 }
4220
4221 static
4222 void
4223 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4224                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4225                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4226 {
4227         pt_entry_t pbits;
4228         pt_entry_t cbits;
4229         pt_entry_t pte;
4230         vm_page_t m;
4231
4232 again:
4233         pbits = *ptep;
4234         cbits = pbits;
4235         if (pte_pv) {
4236                 m = NULL;
4237                 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4238                         if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4239                                 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4240                                 KKASSERT(m == pte_pv->pv_m);
4241                                 vm_page_flag_set(m, PG_REFERENCED);
4242                         }
4243                         cbits &= ~pmap->pmap_bits[PG_A_IDX];
4244                 }
4245                 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4246                         if (pmap_track_modified(pte_pv->pv_pindex)) {
4247                                 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4248                                         if (m == NULL) {
4249                                                 m = PHYS_TO_VM_PAGE(pbits &
4250                                                                     PG_FRAME);
4251                                         }
4252                                         vm_page_dirty(m);
4253                                 }
4254                                 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4255                         }
4256                 }
4257         } else if (sharept) {
4258                 /*
4259                  * Unmanaged page table, pt_pv is actually the pd_pv
4260                  * for our pmap (not the object's shared pmap).
4261                  *
4262                  * When asked to protect something in a shared page table
4263                  * page we just unmap the page table page.  We have to
4264                  * invalidate the tlb in this situation.
4265                  *
4266                  * XXX Warning, shared page tables will not be used for
4267                  * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4268                  * so PHYS_TO_VM_PAGE() should be safe here.
4269                  */
4270                 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4271                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4272                         panic("pmap_protect: pgtable1 pg bad wirecount");
4273                 if (vm_page_unwire_quick(pt_pv->pv_m))
4274                         panic("pmap_protect: pgtable2 pg bad wirecount");
4275                 ptep = NULL;
4276         }
4277         /* else unmanaged page, adjust bits, no wire changes */
4278
4279         if (ptep) {
4280                 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4281 #ifdef PMAP_DEBUG2
4282                 if (pmap_enter_debug > 0) {
4283                         --pmap_enter_debug;
4284                         kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4285                                 "pt_pv=%p cbits=%08lx\n",
4286                                 va, ptep, pte_pv,
4287                                 pt_pv, cbits
4288                         );
4289                 }
4290 #endif
4291                 if (pbits != cbits) {
4292                         if (!pmap_inval_smp_cmpset(pmap, (vm_offset_t)-1,
4293                                                    ptep, pbits, cbits)) {
4294                                 goto again;
4295                         }
4296                 }
4297         }
4298         if (pte_pv)
4299                 pv_put(pte_pv);
4300 }
4301
4302 /*
4303  * Insert the vm_page (m) at the virtual address (va), replacing any prior
4304  * mapping at that address.  Set protection and wiring as requested.
4305  *
4306  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4307  * possible.  If it is we enter the page into the appropriate shared pmap
4308  * hanging off the related VM object instead of the passed pmap, then we
4309  * share the page table page from the VM object's pmap into the current pmap.
4310  *
4311  * NOTE: This routine MUST insert the page into the pmap now, it cannot
4312  *       lazy-evaluate.
4313  */
4314 void
4315 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4316            boolean_t wired, vm_map_entry_t entry)
4317 {
4318         pv_entry_t pt_pv;       /* page table */
4319         pv_entry_t pte_pv;      /* page table entry */
4320         pt_entry_t *ptep;
4321         vm_paddr_t opa;
4322         pt_entry_t origpte, newpte;
4323         vm_paddr_t pa;
4324
4325         if (pmap == NULL)
4326                 return;
4327         va = trunc_page(va);
4328 #ifdef PMAP_DIAGNOSTIC
4329         if (va >= KvaEnd)
4330                 panic("pmap_enter: toobig");
4331         if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4332                 panic("pmap_enter: invalid to pmap_enter page table "
4333                       "pages (va: 0x%lx)", va);
4334 #endif
4335         if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4336                 kprintf("Warning: pmap_enter called on UVA with "
4337                         "kernel_pmap\n");
4338 #ifdef DDB
4339                 db_print_backtrace();
4340 #endif
4341         }
4342         if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4343                 kprintf("Warning: pmap_enter called on KVA without"
4344                         "kernel_pmap\n");
4345 #ifdef DDB
4346                 db_print_backtrace();
4347 #endif
4348         }
4349
4350         /*
4351          * Get locked PV entries for our new page table entry (pte_pv)
4352          * and for its parent page table (pt_pv).  We need the parent
4353          * so we can resolve the location of the ptep.
4354          *
4355          * Only hardware MMU actions can modify the ptep out from
4356          * under us.
4357          *
4358          * if (m) is fictitious or unmanaged we do not create a managing
4359          * pte_pv for it.  Any pre-existing page's management state must
4360          * match (avoiding code complexity).
4361          *
4362          * If the pmap is still being initialized we assume existing
4363          * page tables.
4364          *
4365          * Kernel mapppings do not track page table pages (i.e. pt_pv).
4366          */
4367         if (pmap_initialized == FALSE) {
4368                 pte_pv = NULL;
4369                 pt_pv = NULL;
4370                 ptep = vtopte(va);
4371                 origpte = *ptep;
4372         } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4373                 pte_pv = NULL;
4374                 if (va >= VM_MAX_USER_ADDRESS) {
4375                         pt_pv = NULL;
4376                         ptep = vtopte(va);
4377                 } else {
4378                         pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4379                                                   NULL, entry, va);
4380                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4381                 }
4382                 origpte = *ptep;
4383                 cpu_ccfence();
4384                 KASSERT(origpte == 0 ||
4385                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4386                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4387         } else {
4388                 if (va >= VM_MAX_USER_ADDRESS) {
4389                         /*
4390                          * Kernel map, pv_entry-tracked.
4391                          */
4392                         pt_pv = NULL;
4393                         pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4394                         ptep = vtopte(va);
4395                 } else {
4396                         /*
4397                          * User map
4398                          */
4399                         pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4400                                                    &pt_pv, entry, va);
4401                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4402                 }
4403                 origpte = *ptep;
4404                 cpu_ccfence();
4405                 KASSERT(origpte == 0 ||
4406                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4407                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4408         }
4409
4410         pa = VM_PAGE_TO_PHYS(m);
4411         opa = origpte & PG_FRAME;
4412
4413         newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4414                  pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4415         if (wired)
4416                 newpte |= pmap->pmap_bits[PG_W_IDX];
4417         if (va < VM_MAX_USER_ADDRESS)
4418                 newpte |= pmap->pmap_bits[PG_U_IDX];
4419         if (pte_pv)
4420                 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4421 //      if (pmap == &kernel_pmap)
4422 //              newpte |= pgeflag;
4423         newpte |= pmap->pmap_cache_bits[m->pat_mode];
4424         if (m->flags & PG_FICTITIOUS)
4425                 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4426
4427         /*
4428          * It is possible for multiple faults to occur in threaded
4429          * environments, the existing pte might be correct.
4430          */
4431         if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4432             pmap->pmap_bits[PG_A_IDX])) == 0)
4433                 goto done;
4434
4435         /*
4436          * Ok, either the address changed or the protection or wiring
4437          * changed.
4438          *
4439          * Clear the current entry, interlocking the removal.  For managed
4440          * pte's this will also flush the modified state to the vm_page.
4441          * Atomic ops are mandatory in order to ensure that PG_M events are
4442          * not lost during any transition.
4443          *
4444          * WARNING: The caller has busied the new page but not the original
4445          *          vm_page which we are trying to replace.  Because we hold
4446          *          the pte_pv lock, but have not busied the page, PG bits
4447          *          can be cleared out from under us.
4448          */
4449         if (opa) {
4450                 if (pte_pv) {
4451                         /*
4452                          * pt_pv won't exist for a kernel page (managed or
4453                          * otherwise).
4454                          */
4455                         if (prot & VM_PROT_NOSYNC) {
4456                                 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4457                         } else {
4458                                 pmap_inval_bulk_t bulk;
4459
4460                                 pmap_inval_bulk_init(&bulk, pmap);
4461                                 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4462                                 pmap_inval_bulk_flush(&bulk);
4463                         }
4464                         if (pte_pv->pv_m)
4465                                 pmap_remove_pv_page(pte_pv);
4466                 } else if (prot & VM_PROT_NOSYNC) {
4467                         /*
4468                          * Unmanaged page, NOSYNC (no mmu sync) requested.
4469                          *
4470                          * Leave wire count on PT page intact.
4471                          */
4472                         (void)pte_load_clear(ptep);
4473                         cpu_invlpg((void *)va);
4474                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
4475                 } else {
4476                         /*
4477                          * Unmanaged page, normal enter.
4478                          *
4479                          * Leave wire count on PT page intact.
4480                          */
4481                         pmap_inval_smp(pmap, va, 1, ptep, 0);
4482                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
4483                 }
4484                 KKASSERT(*ptep == 0);
4485         }
4486
4487 #ifdef PMAP_DEBUG2
4488         if (pmap_enter_debug > 0) {
4489                 --pmap_enter_debug;
4490                 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4491                         " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4492                         va, m,
4493                         origpte, newpte, ptep,
4494                         pte_pv, pt_pv, opa, prot);
4495         }
4496 #endif
4497
4498         if (pte_pv) {
4499                 /*
4500                  * Enter on the PV list if part of our managed memory.
4501                  * Wiring of the PT page is already handled.
4502                  */
4503                 KKASSERT(pte_pv->pv_m == NULL);
4504                 vm_page_spin_lock(m);
4505                 pte_pv->pv_m = m;
4506                 pmap_page_stats_adding(m);
4507                 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4508                 vm_page_flag_set(m, PG_MAPPED);
4509                 vm_page_spin_unlock(m);
4510         } else if (pt_pv && opa == 0) {
4511                 /*
4512                  * We have to adjust the wire count on the PT page ourselves
4513                  * for unmanaged entries.  If opa was non-zero we retained
4514                  * the existing wire count from the removal.
4515                  */
4516                 vm_page_wire_quick(pt_pv->pv_m);
4517         }
4518
4519         /*
4520          * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4521          *
4522          * User VMAs do not because those will be zero->non-zero, so no
4523          * stale entries to worry about at this point.
4524          *
4525          * For KVM there appear to still be issues.  Theoretically we
4526          * should be able to scrap the interlocks entirely but we
4527          * get crashes.
4528          */
4529         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4530                 pmap_inval_smp(pmap, va, 1, ptep, newpte);
4531         } else {
4532                 *(volatile pt_entry_t *)ptep = newpte;
4533                 if (pt_pv == NULL)
4534                         cpu_invlpg((void *)va);
4535         }
4536
4537         if (wired) {
4538                 if (pte_pv) {
4539                         atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4540                                         1);
4541                 } else {
4542                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
4543                 }
4544         }
4545         if (newpte & pmap->pmap_bits[PG_RW_IDX])
4546                 vm_page_flag_set(m, PG_WRITEABLE);
4547
4548         /*
4549          * Unmanaged pages need manual resident_count tracking.
4550          */
4551         if (pte_pv == NULL && pt_pv)
4552                 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4553
4554         /*
4555          * Cleanup
4556          */
4557 done:
4558         KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4559                  (m->flags & PG_MAPPED));
4560
4561         /*
4562          * Cleanup the pv entry, allowing other accessors.
4563          */
4564         if (pte_pv)
4565                 pv_put(pte_pv);
4566         if (pt_pv)
4567                 pv_put(pt_pv);
4568 }
4569
4570 /*
4571  * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4572  * This code also assumes that the pmap has no pre-existing entry for this
4573  * VA.
4574  *
4575  * This code currently may only be used on user pmaps, not kernel_pmap.
4576  */
4577 void
4578 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4579 {
4580         pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4581 }
4582
4583 /*
4584  * Make a temporary mapping for a physical address.  This is only intended
4585  * to be used for panic dumps.
4586  *
4587  * The caller is responsible for calling smp_invltlb().
4588  */
4589 void *
4590 pmap_kenter_temporary(vm_paddr_t pa, long i)
4591 {
4592         pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4593         return ((void *)crashdumpmap);
4594 }
4595
4596 #define MAX_INIT_PT (96)
4597
4598 /*
4599  * This routine preloads the ptes for a given object into the specified pmap.
4600  * This eliminates the blast of soft faults on process startup and
4601  * immediately after an mmap.
4602  */
4603 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4604
4605 void
4606 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4607                     vm_object_t object, vm_pindex_t pindex,
4608                     vm_size_t size, int limit)
4609 {
4610         struct rb_vm_page_scan_info info;
4611         struct lwp *lp;
4612         vm_size_t psize;
4613
4614         /*
4615          * We can't preinit if read access isn't set or there is no pmap
4616          * or object.
4617          */
4618         if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4619                 return;
4620
4621         /*
4622          * We can't preinit if the pmap is not the current pmap
4623          */
4624         lp = curthread->td_lwp;
4625         if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4626                 return;
4627
4628         /*
4629          * Misc additional checks
4630          */
4631         psize = x86_64_btop(size);
4632
4633         if ((object->type != OBJT_VNODE) ||
4634                 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4635                         (object->resident_page_count > MAX_INIT_PT))) {
4636                 return;
4637         }
4638
4639         if (pindex + psize > object->size) {
4640                 if (object->size < pindex)
4641                         return;
4642                 psize = object->size - pindex;
4643         }
4644
4645         if (psize == 0)
4646                 return;
4647
4648         /*
4649          * If everything is segment-aligned do not pre-init here.  Instead
4650          * allow the normal vm_fault path to pass a segment hint to
4651          * pmap_enter() which will then use an object-referenced shared
4652          * page table page.
4653          */
4654         if ((addr & SEG_MASK) == 0 &&
4655             (ctob(psize) & SEG_MASK) == 0 &&
4656             (ctob(pindex) & SEG_MASK) == 0) {
4657                 return;
4658         }
4659
4660         /*
4661          * Use a red-black scan to traverse the requested range and load
4662          * any valid pages found into the pmap.
4663          *
4664          * We cannot safely scan the object's memq without holding the
4665          * object token.
4666          */
4667         info.start_pindex = pindex;
4668         info.end_pindex = pindex + psize - 1;
4669         info.limit = limit;
4670         info.mpte = NULL;
4671         info.addr = addr;
4672         info.pmap = pmap;
4673
4674         vm_object_hold_shared(object);
4675         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4676                                 pmap_object_init_pt_callback, &info);
4677         vm_object_drop(object);
4678 }
4679
4680 static
4681 int
4682 pmap_object_init_pt_callback(vm_page_t p, void *data)
4683 {
4684         struct rb_vm_page_scan_info *info = data;
4685         vm_pindex_t rel_index;
4686
4687         /*
4688          * don't allow an madvise to blow away our really
4689          * free pages allocating pv entries.
4690          */
4691         if ((info->limit & MAP_PREFAULT_MADVISE) &&
4692                 vmstats.v_free_count < vmstats.v_free_reserved) {
4693                     return(-1);
4694         }
4695
4696         /*
4697          * Ignore list markers and ignore pages we cannot instantly
4698          * busy (while holding the object token).
4699          */
4700         if (p->flags & PG_MARKER)
4701                 return 0;
4702         if (vm_page_busy_try(p, TRUE))
4703                 return 0;
4704         if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4705             (p->flags & PG_FICTITIOUS) == 0) {
4706                 if ((p->queue - p->pc) == PQ_CACHE)
4707                         vm_page_deactivate(p);
4708                 rel_index = p->pindex - info->start_pindex;
4709                 pmap_enter_quick(info->pmap,
4710                                  info->addr + x86_64_ptob(rel_index), p);
4711         }
4712         vm_page_wakeup(p);
4713         lwkt_yield();
4714         return(0);
4715 }
4716
4717 /*
4718  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4719  * address.
4720  *
4721  * Returns FALSE if it would be non-trivial or if a pte is already loaded
4722  * into the slot.
4723  *
4724  * XXX This is safe only because page table pages are not freed.
4725  */
4726 int
4727 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4728 {
4729         pt_entry_t *pte;
4730
4731         /*spin_lock(&pmap->pm_spin);*/
4732         if ((pte = pmap_pte(pmap, addr)) != NULL) {
4733                 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4734                         /*spin_unlock(&pmap->pm_spin);*/
4735                         return FALSE;
4736                 }
4737         }
4738         /*spin_unlock(&pmap->pm_spin);*/
4739         return TRUE;
4740 }
4741
4742 /*
4743  * Change the wiring attribute for a pmap/va pair.  The mapping must already
4744  * exist in the pmap.  The mapping may or may not be managed.
4745  */
4746 void
4747 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
4748                    vm_map_entry_t entry)
4749 {
4750         pt_entry_t *ptep;
4751         pv_entry_t pv;
4752
4753         if (pmap == NULL)
4754                 return;
4755         lwkt_gettoken(&pmap->pm_token);
4756         pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL, entry, va);
4757         ptep = pv_pte_lookup(pv, pmap_pte_index(va));
4758
4759         if (wired && !pmap_pte_w(pmap, ptep))
4760                 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4761         else if (!wired && pmap_pte_w(pmap, ptep))
4762                 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4763
4764         /*
4765          * Wiring is not a hardware characteristic so there is no need to
4766          * invalidate TLB.  However, in an SMP environment we must use
4767          * a locked bus cycle to update the pte (if we are not using
4768          * the pmap_inval_*() API that is)... it's ok to do this for simple
4769          * wiring changes.
4770          */
4771         if (wired)
4772                 atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4773         else
4774                 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4775         pv_put(pv);
4776         lwkt_reltoken(&pmap->pm_token);
4777 }
4778
4779
4780
4781 /*
4782  * Copy the range specified by src_addr/len from the source map to
4783  * the range dst_addr/len in the destination map.
4784  *
4785  * This routine is only advisory and need not do anything.
4786  */
4787 void
4788 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4789           vm_size_t len, vm_offset_t src_addr)
4790 {
4791 }
4792
4793 /*
4794  * pmap_zero_page:
4795  *
4796  *      Zero the specified physical page.
4797  *
4798  *      This function may be called from an interrupt and no locking is
4799  *      required.
4800  */
4801 void
4802 pmap_zero_page(vm_paddr_t phys)
4803 {
4804         vm_offset_t va = PHYS_TO_DMAP(phys);
4805
4806         pagezero((void *)va);
4807 }
4808
4809 /*
4810  * pmap_zero_page:
4811  *
4812  *      Zero part of a physical page by mapping it into memory and clearing
4813  *      its contents with bzero.
4814  *
4815  *      off and size may not cover an area beyond a single hardware page.
4816  */
4817 void
4818 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4819 {
4820         vm_offset_t virt = PHYS_TO_DMAP(phys);
4821
4822         bzero((char *)virt + off, size);
4823 }
4824
4825 /*
4826  * pmap_copy_page:
4827  *
4828  *      Copy the physical page from the source PA to the target PA.
4829  *      This function may be called from an interrupt.  No locking
4830  *      is required.
4831  */
4832 void
4833 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4834 {
4835         vm_offset_t src_virt, dst_virt;
4836
4837         src_virt = PHYS_TO_DMAP(src);
4838         dst_virt = PHYS_TO_DMAP(dst);
4839         bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4840 }
4841
4842 /*
4843  * pmap_copy_page_frag:
4844  *
4845  *      Copy the physical page from the source PA to the target PA.
4846  *      This function may be called from an interrupt.  No locking
4847  *      is required.
4848  */
4849 void
4850 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4851 {
4852         vm_offset_t src_virt, dst_virt;
4853
4854         src_virt = PHYS_TO_DMAP(src);
4855         dst_virt = PHYS_TO_DMAP(dst);
4856
4857         bcopy((char *)src_virt + (src & PAGE_MASK),
4858               (char *)dst_virt + (dst & PAGE_MASK),
4859               bytes);
4860 }
4861
4862 /*
4863  * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4864  * this page.  This count may be changed upwards or downwards in the future;
4865  * it is only necessary that true be returned for a small subset of pmaps
4866  * for proper page aging.
4867  */
4868 boolean_t
4869 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4870 {
4871         pv_entry_t pv;
4872         int loops = 0;
4873
4874         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4875                 return FALSE;
4876
4877         vm_page_spin_lock(m);
4878         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4879                 if (pv->pv_pmap == pmap) {
4880                         vm_page_spin_unlock(m);
4881                         return TRUE;
4882                 }
4883                 loops++;
4884                 if (loops >= 16)
4885                         break;
4886         }
4887         vm_page_spin_unlock(m);
4888         return (FALSE);
4889 }
4890
4891 /*
4892  * Remove all pages from specified address space this aids process exit
4893  * speeds.  Also, this code may be special cased for the current process
4894  * only.
4895  */
4896 void
4897 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4898 {
4899         pmap_remove_noinval(pmap, sva, eva);
4900         cpu_invltlb();
4901 }
4902
4903 /*
4904  * pmap_testbit tests bits in pte's note that the testbit/clearbit
4905  * routines are inline, and a lot of things compile-time evaluate.
4906  */
4907 static
4908 boolean_t
4909 pmap_testbit(vm_page_t m, int bit)
4910 {
4911         pv_entry_t pv;
4912         pt_entry_t *pte;
4913         pmap_t pmap;
4914
4915         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4916                 return FALSE;
4917
4918         if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4919                 return FALSE;
4920         vm_page_spin_lock(m);
4921         if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4922                 vm_page_spin_unlock(m);
4923                 return FALSE;
4924         }
4925
4926         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4927
4928 #if defined(PMAP_DIAGNOSTIC)
4929                 if (pv->pv_pmap == NULL) {
4930                         kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4931                             pv->pv_pindex);
4932                         continue;
4933                 }
4934 #endif
4935                 pmap = pv->pv_pmap;
4936
4937                 /*
4938                  * If the bit being tested is the modified bit, then
4939                  * mark clean_map and ptes as never
4940                  * modified.
4941                  *
4942                  * WARNING!  Because we do not lock the pv, *pte can be in a
4943                  *           state of flux.  Despite this the value of *pte
4944                  *           will still be related to the vm_page in some way
4945                  *           because the pv cannot be destroyed as long as we
4946                  *           hold the vm_page spin lock.
4947                  */
4948                 if (bit == PG_A_IDX || bit == PG_M_IDX) {
4949                                 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
4950                         if (!pmap_track_modified(pv->pv_pindex))
4951                                 continue;
4952                 }
4953
4954                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4955                 if (*pte & pmap->pmap_bits[bit]) {
4956                         vm_page_spin_unlock(m);
4957                         return TRUE;
4958                 }
4959         }
4960         vm_page_spin_unlock(m);
4961         return (FALSE);
4962 }
4963
4964 /*
4965  * This routine is used to modify bits in ptes.  Only one bit should be
4966  * specified.  PG_RW requires special handling.
4967  *
4968  * Caller must NOT hold any spin locks
4969  */
4970 static __inline
4971 void
4972 pmap_clearbit(vm_page_t m, int bit_index)
4973 {
4974         pv_entry_t pv;
4975         pt_entry_t *pte;
4976         pt_entry_t pbits;
4977         pmap_t pmap;
4978
4979         if (bit_index == PG_RW_IDX)
4980                 vm_page_flag_clear(m, PG_WRITEABLE);
4981         if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
4982                 return;
4983         }
4984
4985         /*
4986          * PG_M or PG_A case
4987          *
4988          * Loop over all current mappings setting/clearing as appropos If
4989          * setting RO do we need to clear the VAC?
4990          *
4991          * NOTE: When clearing PG_M we could also (not implemented) drop
4992          *       through to the PG_RW code and clear PG_RW too, forcing
4993          *       a fault on write to redetect PG_M for virtual kernels, but
4994          *       it isn't necessary since virtual kernels invalidate the
4995          *       pte when they clear the VPTE_M bit in their virtual page
4996          *       tables.
4997          *
4998          * NOTE: Does not re-dirty the page when clearing only PG_M.
4999          *
5000          * NOTE: Because we do not lock the pv, *pte can be in a state of
5001          *       flux.  Despite this the value of *pte is still somewhat
5002          *       related while we hold the vm_page spin lock.
5003          *
5004          *       *pte can be zero due to this race.  Since we are clearing
5005          *       bits we basically do no harm when this race  ccurs.
5006          */
5007         if (bit_index != PG_RW_IDX) {
5008                 vm_page_spin_lock(m);
5009                 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5010 #if defined(PMAP_DIAGNOSTIC)
5011                         if (pv->pv_pmap == NULL) {
5012                                 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5013                                     pv->pv_pindex);
5014                                 continue;
5015                         }
5016 #endif
5017                         pmap = pv->pv_pmap;
5018                         pte = pmap_pte_quick(pv->pv_pmap,
5019                                              pv->pv_pindex << PAGE_SHIFT);
5020                         pbits = *pte;
5021                         if (pbits & pmap->pmap_bits[bit_index])
5022                                 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5023                 }
5024                 vm_page_spin_unlock(m);
5025                 return;
5026         }
5027
5028         /*
5029          * Clear PG_RW.  Also clears PG_M and marks the page dirty if PG_M
5030          * was set.
5031          */
5032 restart:
5033         vm_page_spin_lock(m);
5034         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5035                 /*
5036                  * don't write protect pager mappings
5037                  */
5038                 if (!pmap_track_modified(pv->pv_pindex))
5039                         continue;
5040
5041 #if defined(PMAP_DIAGNOSTIC)
5042                 if (pv->pv_pmap == NULL) {
5043                         kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5044                             pv->pv_pindex);
5045                         continue;
5046                 }
5047 #endif
5048                 pmap = pv->pv_pmap;
5049                 /*
5050                  * Skip pages which do not have PG_RW set.
5051                  */
5052                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5053                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5054                         continue;
5055
5056                 /*
5057                  * Lock the PV
5058                  */
5059                 if (pv_hold_try(pv)) {
5060                         vm_page_spin_unlock(m);
5061                 } else {
5062                         vm_page_spin_unlock(m);
5063                         pv_lock(pv);    /* held, now do a blocking lock */
5064                 }
5065                 if (pv->pv_pmap != pmap || pv->pv_m != m) {
5066                         pv_put(pv);     /* and release */
5067                         goto restart;   /* anything could have happened */
5068                 }
5069                 KKASSERT(pv->pv_pmap == pmap);
5070                 for (;;) {
5071                         pt_entry_t nbits;
5072
5073                         pbits = *pte;
5074                         cpu_ccfence();
5075                         nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5076                                           pmap->pmap_bits[PG_M_IDX]);
5077                         if (pmap_inval_smp_cmpset(pmap,
5078                                      ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5079                                      pte, pbits, nbits)) {
5080                                 break;
5081                         }
5082                         cpu_pause();
5083                 }
5084                 vm_page_spin_lock(m);
5085
5086                 /*
5087                  * If PG_M was found to be set while we were clearing PG_RW
5088                  * we also clear PG_M (done above) and mark the page dirty.
5089                  * Callers expect this behavior.
5090                  */
5091                 if (pbits & pmap->pmap_bits[PG_M_IDX])
5092                         vm_page_dirty(m);
5093                 pv_put(pv);
5094         }
5095         vm_page_spin_unlock(m);
5096 }
5097
5098 /*
5099  * Lower the permission for all mappings to a given page.
5100  *
5101  * Page must be busied by caller.  Because page is busied by caller this
5102  * should not be able to race a pmap_enter().
5103  */
5104 void
5105 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5106 {
5107         /* JG NX support? */
5108         if ((prot & VM_PROT_WRITE) == 0) {
5109                 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5110                         /*
5111                          * NOTE: pmap_clearbit(.. PG_RW) also clears
5112                          *       the PG_WRITEABLE flag in (m).
5113                          */
5114                         pmap_clearbit(m, PG_RW_IDX);
5115                 } else {
5116                         pmap_remove_all(m);
5117                 }
5118         }
5119 }
5120
5121 vm_paddr_t
5122 pmap_phys_address(vm_pindex_t ppn)
5123 {
5124         return (x86_64_ptob(ppn));
5125 }
5126
5127 /*
5128  * Return a count of reference bits for a page, clearing those bits.
5129  * It is not necessary for every reference bit to be cleared, but it
5130  * is necessary that 0 only be returned when there are truly no
5131  * reference bits set.
5132  *
5133  * XXX: The exact number of bits to check and clear is a matter that
5134  * should be tested and standardized at some point in the future for
5135  * optimal aging of shared pages.
5136  *
5137  * This routine may not block.
5138  */
5139 int
5140 pmap_ts_referenced(vm_page_t m)
5141 {
5142         pv_entry_t pv;
5143         pt_entry_t *pte;
5144         pmap_t pmap;
5145         int rtval = 0;
5146
5147         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5148                 return (rtval);
5149
5150         vm_page_spin_lock(m);
5151         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5152                 if (!pmap_track_modified(pv->pv_pindex))
5153                         continue;
5154                 pmap = pv->pv_pmap;
5155                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5156                 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5157                         atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5158                         rtval++;
5159                         if (rtval > 4)
5160                                 break;
5161                 }
5162         }
5163         vm_page_spin_unlock(m);
5164         return (rtval);
5165 }
5166
5167 /*
5168  *      pmap_is_modified:
5169  *
5170  *      Return whether or not the specified physical page was modified
5171  *      in any physical maps.
5172  */
5173 boolean_t
5174 pmap_is_modified(vm_page_t m)
5175 {
5176         boolean_t res;
5177
5178         res = pmap_testbit(m, PG_M_IDX);
5179         return (res);
5180 }
5181
5182 /*
5183  *      Clear the modify bits on the specified physical page.
5184  */
5185 void
5186 pmap_clear_modify(vm_page_t m)
5187 {
5188         pmap_clearbit(m, PG_M_IDX);
5189 }
5190
5191 /*
5192  *      pmap_clear_reference:
5193  *
5194  *      Clear the reference bit on the specified physical page.
5195  */
5196 void
5197 pmap_clear_reference(vm_page_t m)
5198 {
5199         pmap_clearbit(m, PG_A_IDX);
5200 }
5201
5202 /*
5203  * Miscellaneous support routines follow
5204  */
5205
5206 static
5207 void
5208 i386_protection_init(void)
5209 {
5210         int *kp, prot;
5211
5212         /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit  */
5213         kp = protection_codes;
5214         for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5215                 switch (prot) {
5216                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5217                         /*
5218                          * Read access is also 0. There isn't any execute bit,
5219                          * so just make it readable.
5220                          */
5221                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5222                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5223                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5224                         *kp++ = 0;
5225                         break;
5226                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5227                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5228                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5229                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5230                         *kp++ = pmap_bits_default[PG_RW_IDX];
5231                         break;
5232                 }
5233         }
5234 }
5235
5236 /*
5237  * Map a set of physical memory pages into the kernel virtual
5238  * address space. Return a pointer to where it is mapped. This
5239  * routine is intended to be used for mapping device memory,
5240  * NOT real memory.
5241  *
5242  * NOTE: We can't use pgeflag unless we invalidate the pages one at
5243  *       a time.
5244  *
5245  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5246  *       work whether the cpu supports PAT or not.  The remaining PAT
5247  *       attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5248  *       supports PAT.
5249  */
5250 void *
5251 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5252 {
5253         return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5254 }
5255
5256 void *
5257 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5258 {
5259         return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5260 }
5261
5262 void *
5263 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5264 {
5265         return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5266 }
5267
5268 /*
5269  * Map a set of physical memory pages into the kernel virtual
5270  * address space. Return a pointer to where it is mapped. This
5271  * routine is intended to be used for mapping device memory,
5272  * NOT real memory.
5273  */
5274 void *
5275 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5276 {
5277         vm_offset_t va, tmpva, offset;
5278         pt_entry_t *pte;
5279         vm_size_t tmpsize;
5280
5281         offset = pa & PAGE_MASK;
5282         size = roundup(offset + size, PAGE_SIZE);
5283
5284         va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5285         if (va == 0)
5286                 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5287
5288         pa = pa & ~PAGE_MASK;
5289         for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5290                 pte = vtopte(tmpva);
5291                 *pte = pa |
5292                     kernel_pmap.pmap_bits[PG_RW_IDX] |
5293                     kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5294                     kernel_pmap.pmap_cache_bits[mode];
5295                 tmpsize -= PAGE_SIZE;
5296                 tmpva += PAGE_SIZE;
5297                 pa += PAGE_SIZE;
5298         }
5299         pmap_invalidate_range(&kernel_pmap, va, va + size);
5300         pmap_invalidate_cache_range(va, va + size);
5301
5302         return ((void *)(va + offset));
5303 }
5304
5305 void
5306 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5307 {
5308         vm_offset_t base, offset;
5309
5310         base = va & ~PAGE_MASK;
5311         offset = va & PAGE_MASK;
5312         size = roundup(offset + size, PAGE_SIZE);
5313         pmap_qremove(va, size >> PAGE_SHIFT);
5314         kmem_free(&kernel_map, base, size);
5315 }
5316
5317 /*
5318  * Sets the memory attribute for the specified page.
5319  */
5320 void
5321 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5322 {
5323
5324     m->pat_mode = ma;
5325
5326     /*
5327      * If "m" is a normal page, update its direct mapping.  This update
5328      * can be relied upon to perform any cache operations that are
5329      * required for data coherence.
5330      */
5331     if ((m->flags & PG_FICTITIOUS) == 0)
5332         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5333 }
5334
5335 /*
5336  * Change the PAT attribute on an existing kernel memory map.  Caller
5337  * must ensure that the virtual memory in question is not accessed
5338  * during the adjustment.
5339  */
5340 void
5341 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5342 {
5343         pt_entry_t *pte;
5344         vm_offset_t base;
5345         int changed = 0;
5346
5347         if (va == 0)
5348                 panic("pmap_change_attr: va is NULL");
5349         base = trunc_page(va);
5350
5351         while (count) {
5352                 pte = vtopte(va);
5353                 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5354                        kernel_pmap.pmap_cache_bits[mode];
5355                 --count;
5356                 va += PAGE_SIZE;
5357         }
5358
5359         changed = 1;    /* XXX: not optimal */
5360
5361         /*
5362          * Flush CPU caches if required to make sure any data isn't cached that
5363          * shouldn't be, etc.
5364          */
5365         if (changed) {
5366                 pmap_invalidate_range(&kernel_pmap, base, va);
5367                 pmap_invalidate_cache_range(base, va);
5368         }
5369 }
5370
5371 /*
5372  * perform the pmap work for mincore
5373  */
5374 int
5375 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5376 {
5377         pt_entry_t *ptep, pte;
5378         vm_page_t m;
5379         int val = 0;
5380
5381         lwkt_gettoken(&pmap->pm_token);
5382         ptep = pmap_pte(pmap, addr);
5383
5384         if (ptep && (pte = *ptep) != 0) {
5385                 vm_offset_t pa;
5386
5387                 val = MINCORE_INCORE;
5388                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5389                         goto done;
5390
5391                 pa = pte & PG_FRAME;
5392
5393                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5394                         m = NULL;
5395                 else
5396                         m = PHYS_TO_VM_PAGE(pa);
5397
5398                 /*
5399                  * Modified by us
5400                  */
5401                 if (pte & pmap->pmap_bits[PG_M_IDX])
5402                         val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5403                 /*
5404                  * Modified by someone
5405                  */
5406                 else if (m && (m->dirty || pmap_is_modified(m)))
5407                         val |= MINCORE_MODIFIED_OTHER;
5408                 /*
5409                  * Referenced by us
5410                  */
5411                 if (pte & pmap->pmap_bits[PG_A_IDX])
5412                         val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5413
5414                 /*
5415                  * Referenced by someone
5416                  */
5417                 else if (m && ((m->flags & PG_REFERENCED) ||
5418                                 pmap_ts_referenced(m))) {
5419                         val |= MINCORE_REFERENCED_OTHER;
5420                         vm_page_flag_set(m, PG_REFERENCED);
5421                 }
5422         }
5423 done:
5424         lwkt_reltoken(&pmap->pm_token);
5425
5426         return val;
5427 }
5428
5429 /*
5430  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
5431  * vmspace will be ref'd and the old one will be deref'd.
5432  *
5433  * The vmspace for all lwps associated with the process will be adjusted
5434  * and cr3 will be reloaded if any lwp is the current lwp.
5435  *
5436  * The process must hold the vmspace->vm_map.token for oldvm and newvm
5437  */
5438 void
5439 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5440 {
5441         struct vmspace *oldvm;
5442         struct lwp *lp;
5443
5444         oldvm = p->p_vmspace;
5445         if (oldvm != newvm) {
5446                 if (adjrefs)
5447                         vmspace_ref(newvm);
5448                 p->p_vmspace = newvm;
5449                 KKASSERT(p->p_nthreads == 1);
5450                 lp = RB_ROOT(&p->p_lwp_tree);
5451                 pmap_setlwpvm(lp, newvm);
5452                 if (adjrefs)
5453                         vmspace_rel(oldvm);
5454         }
5455 }
5456
5457 /*
5458  * Set the vmspace for a LWP.  The vmspace is almost universally set the
5459  * same as the process vmspace, but virtual kernels need to swap out contexts
5460  * on a per-lwp basis.
5461  *
5462  * Caller does not necessarily hold any vmspace tokens.  Caller must control
5463  * the lwp (typically be in the context of the lwp).  We use a critical
5464  * section to protect against statclock and hardclock (statistics collection).
5465  */
5466 void
5467 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5468 {
5469         struct vmspace *oldvm;
5470         struct pmap *pmap;
5471
5472         oldvm = lp->lwp_vmspace;
5473
5474         if (oldvm != newvm) {
5475                 crit_enter();
5476                 lp->lwp_vmspace = newvm;
5477                 if (curthread->td_lwp == lp) {
5478                         pmap = vmspace_pmap(newvm);
5479                         ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5480                         if (pmap->pm_active_lock & CPULOCK_EXCL)
5481                                 pmap_interlock_wait(newvm);
5482 #if defined(SWTCH_OPTIM_STATS)
5483                         tlb_flush_count++;
5484 #endif
5485                         if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5486                                 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5487                         } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5488                                 curthread->td_pcb->pcb_cr3 = KPML4phys;
5489                         } else {
5490                                 panic("pmap_setlwpvm: unknown pmap type\n");
5491                         }
5492                         load_cr3(curthread->td_pcb->pcb_cr3);
5493                         pmap = vmspace_pmap(oldvm);
5494                         ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5495                                                mycpu->gd_cpuid);
5496                 }
5497                 crit_exit();
5498         }
5499 }
5500
5501 /*
5502  * Called when switching to a locked pmap, used to interlock against pmaps
5503  * undergoing modifications to prevent us from activating the MMU for the
5504  * target pmap until all such modifications have completed.  We have to do
5505  * this because the thread making the modifications has already set up its
5506  * SMP synchronization mask.
5507  *
5508  * This function cannot sleep!
5509  *
5510  * No requirements.
5511  */
5512 void
5513 pmap_interlock_wait(struct vmspace *vm)
5514 {
5515         struct pmap *pmap = &vm->vm_pmap;
5516
5517         if (pmap->pm_active_lock & CPULOCK_EXCL) {
5518                 crit_enter();
5519                 KKASSERT(curthread->td_critcount >= 2);
5520                 DEBUG_PUSH_INFO("pmap_interlock_wait");
5521                 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5522                         cpu_ccfence();
5523                         lwkt_process_ipiq();
5524                 }
5525                 DEBUG_POP_INFO();
5526                 crit_exit();
5527         }
5528 }
5529
5530 vm_offset_t
5531 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5532 {
5533
5534         if ((obj == NULL) || (size < NBPDR) ||
5535             ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5536                 return addr;
5537         }
5538
5539         addr = roundup2(addr, NBPDR);
5540         return addr;
5541 }
5542
5543 /*
5544  * Used by kmalloc/kfree, page already exists at va
5545  */
5546 vm_page_t
5547 pmap_kvtom(vm_offset_t va)
5548 {
5549         pt_entry_t *ptep = vtopte(va);
5550
5551         KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5552         return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5553 }
5554
5555 /*
5556  * Initialize machine-specific shared page directory support.  This
5557  * is executed when a VM object is created.
5558  */
5559 void
5560 pmap_object_init(vm_object_t object)
5561 {
5562         object->md.pmap_rw = NULL;
5563         object->md.pmap_ro = NULL;
5564 }
5565
5566 /*
5567  * Clean up machine-specific shared page directory support.  This
5568  * is executed when a VM object is destroyed.
5569  */
5570 void
5571 pmap_object_free(vm_object_t object)
5572 {
5573         pmap_t pmap;
5574
5575         if ((pmap = object->md.pmap_rw) != NULL) {
5576                 object->md.pmap_rw = NULL;
5577                 pmap_remove_noinval(pmap,
5578                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5579                 CPUMASK_ASSZERO(pmap->pm_active);
5580                 pmap_release(pmap);
5581                 pmap_puninit(pmap);
5582                 kfree(pmap, M_OBJPMAP);
5583         }
5584         if ((pmap = object->md.pmap_ro) != NULL) {
5585                 object->md.pmap_ro = NULL;
5586                 pmap_remove_noinval(pmap,
5587                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5588                 CPUMASK_ASSZERO(pmap->pm_active);
5589                 pmap_release(pmap);
5590                 pmap_puninit(pmap);
5591                 kfree(pmap, M_OBJPMAP);
5592         }
5593 }
5594
5595 /*
5596  * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
5597  * VM page and issue a pginfo->callback.
5598  *
5599  * We are expected to dispose of any non-NULL pte_pv.
5600  */
5601 static
5602 void
5603 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
5604                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
5605                       vm_offset_t va, pt_entry_t *ptep, void *arg)
5606 {
5607         struct pmap_pgscan_info *pginfo = arg;
5608         vm_page_t m;
5609
5610         if (pte_pv) {
5611                 /*
5612                  * Try to busy the page while we hold the pte_pv locked.
5613                  */
5614                 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
5615                 if (vm_page_busy_try(m, TRUE) == 0) {
5616                         if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
5617                                 /*
5618                                  * The callback is issued with the pte_pv
5619                                  * unlocked and put away, and the pt_pv
5620                                  * unlocked.
5621                                  */
5622                                 pv_put(pte_pv);
5623                                 if (pt_pv)
5624                                         pv_unlock(pt_pv);
5625                                 if (pginfo->callback(pginfo, va, m) < 0)
5626                                         info->stop = 1;
5627                                 if (pt_pv)
5628                                         pv_lock(pt_pv);
5629                         } else {
5630                                 vm_page_wakeup(m);
5631                                 pv_put(pte_pv);
5632                         }
5633                 } else {
5634                         ++pginfo->busycount;
5635                         pv_put(pte_pv);
5636                 }
5637         } else if (sharept) {
5638                 /* shared page table */
5639         } else {
5640                 /* else unmanaged page */
5641         }
5642 }
5643
5644 void
5645 pmap_pgscan(struct pmap_pgscan_info *pginfo)
5646 {
5647         struct pmap_scan_info info;
5648
5649         pginfo->offset = pginfo->beg_addr;
5650         info.pmap = pginfo->pmap;
5651         info.sva = pginfo->beg_addr;
5652         info.eva = pginfo->end_addr;
5653         info.func = pmap_pgscan_callback;
5654         info.arg = pginfo;
5655         pmap_scan(&info, 0);
5656         if (info.stop == 0)
5657                 pginfo->offset = pginfo->end_addr;
5658 }