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.
1914          */
1915         if (ptepindex < pmap_pt_pindex(0)) {
1916                 if (ptepindex >= NUPTE_USER)
1917                         KKASSERT(pvpp == NULL);
1918                 else
1919                         KKASSERT(pvpp != NULL);
1920                 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1921                 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1922                 if (isnew) {
1923                         vm_page_wire_quick(pvp->pv_m);
1924                         if (pvpp)
1925                                 *pvpp = pvp;
1926                         else
1927                                 pv_put(pvp);
1928                 } else {
1929                         *pvpp = pvp;
1930                 }
1931                 return(pv);
1932         }
1933
1934         /*
1935          * Non-terminal PVs allocate a VM page to represent the page table,
1936          * so we have to resolve pvp and calculate ptepindex for the pvp
1937          * and then for the page table entry index in the pvp for
1938          * fall-through.
1939          */
1940         if (ptepindex < pmap_pd_pindex(0)) {
1941                 /*
1942                  * pv is PT, pvp is PD
1943                  */
1944                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1945                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1946                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1947                 if (!isnew)
1948                         goto notnew;
1949
1950                 /*
1951                  * PT index in PD
1952                  */
1953                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1954                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1955                 ispt = 1;
1956         } else if (ptepindex < pmap_pdp_pindex(0)) {
1957                 /*
1958                  * pv is PD, pvp is PDP
1959                  *
1960                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1961                  *                   the PD.
1962                  */
1963                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1964                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1965
1966                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1967                         KKASSERT(pvpp == NULL);
1968                         pvp = NULL;
1969                 } else {
1970                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
1971                 }
1972                 if (!isnew)
1973                         goto notnew;
1974
1975                 /*
1976                  * PD index in PDP
1977                  */
1978                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1979                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1980         } else if (ptepindex < pmap_pml4_pindex()) {
1981                 /*
1982                  * pv is PDP, pvp is the root pml4 table
1983                  */
1984                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1985                 if (!isnew)
1986                         goto notnew;
1987
1988                 /*
1989                  * PDP index in PML4
1990                  */
1991                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1992                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1993         } else {
1994                 /*
1995                  * pv represents the top-level PML4, there is no parent.
1996                  */
1997                 pvp = NULL;
1998                 if (!isnew)
1999                         goto notnew;
2000         }
2001
2002         /*
2003          * (isnew) is TRUE, pv is not terminal.
2004          *
2005          * (1) Add a wire count to the parent page table (pvp).
2006          * (2) Allocate a VM page for the page table.
2007          * (3) Enter the VM page into the parent page table.
2008          *
2009          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2010          */
2011         if (pvp)
2012                 vm_page_wire_quick(pvp->pv_m);
2013
2014         for (;;) {
2015                 m = vm_page_alloc(NULL, pv->pv_pindex,
2016                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2017                                   VM_ALLOC_INTERRUPT);
2018                 if (m)
2019                         break;
2020                 vm_wait(0);
2021         }
2022         vm_page_spin_lock(m);
2023         pmap_page_stats_adding(m);
2024         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2025         pv->pv_m = m;
2026         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2027         vm_page_spin_unlock(m);
2028         vm_page_unmanage(m);    /* m must be spinunlocked */
2029
2030         pmap_zero_page(VM_PAGE_TO_PHYS(m));
2031         m->valid = VM_PAGE_BITS_ALL;
2032         vm_page_wire(m);        /* wire for mapping in parent */
2033
2034         /*
2035          * Wire the page into pvp.  Bump the resident_count for the pmap.
2036          * There is no pvp for the top level, address the pm_pml4[] array
2037          * directly.
2038          *
2039          * If the caller wants the parent we return it, otherwise
2040          * we just put it away.
2041          *
2042          * No interlock is needed for pte 0 -> non-zero.
2043          *
2044          * In the situation where *ptep is valid we might have an unmanaged
2045          * page table page shared from another page table which we need to
2046          * unshare before installing our private page table page.
2047          */
2048         if (pvp) {
2049                 ptep = pv_pte_lookup(pvp, ptepindex);
2050                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2051                         pt_entry_t pte;
2052
2053                         if (ispt == 0) {
2054                                 panic("pmap_allocpte: unexpected pte %p/%d",
2055                                       pvp, (int)ptepindex);
2056                         }
2057                         pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
2058                         if (vm_page_unwire_quick(
2059                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2060                                 panic("pmap_allocpte: shared pgtable "
2061                                       "pg bad wirecount");
2062                         }
2063                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
2064                 }
2065                 *ptep = VM_PAGE_TO_PHYS(m) |
2066                         (pmap->pmap_bits[PG_U_IDX] |
2067                          pmap->pmap_bits[PG_RW_IDX] |
2068                          pmap->pmap_bits[PG_V_IDX] |
2069                          pmap->pmap_bits[PG_A_IDX] |
2070                          pmap->pmap_bits[PG_M_IDX]);
2071         }
2072         vm_page_wakeup(m);
2073 notnew:
2074         if (pvpp)
2075                 *pvpp = pvp;
2076         else if (pvp)
2077                 pv_put(pvp);
2078         return (pv);
2079 }
2080
2081 /*
2082  * This version of pmap_allocpte() checks for possible segment optimizations
2083  * that would allow page-table sharing.  It can be called for terminal
2084  * page or page table page ptepindex's.
2085  *
2086  * The function is called with page table page ptepindex's for fictitious
2087  * and unmanaged terminal pages.  That is, we don't want to allocate a
2088  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
2089  * for this case.
2090  *
2091  * This function can return a pv and *pvpp associated with the passed in pmap
2092  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
2093  * an unmanaged page table page will be entered into the pass in pmap.
2094  */
2095 static
2096 pv_entry_t
2097 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2098                   vm_map_entry_t entry, vm_offset_t va)
2099 {
2100         vm_object_t object;
2101         pmap_t obpmap;
2102         pmap_t *obpmapp;
2103         vm_offset_t b;
2104         pv_entry_t pte_pv;      /* in original or shared pmap */
2105         pv_entry_t pt_pv;       /* in original or shared pmap */
2106         pv_entry_t proc_pd_pv;  /* in original pmap */
2107         pv_entry_t proc_pt_pv;  /* in original pmap */
2108         pv_entry_t xpv;         /* PT in shared pmap */
2109         pd_entry_t *pt;         /* PT entry in PD of original pmap */
2110         pd_entry_t opte;        /* contents of *pt */
2111         pd_entry_t npte;        /* contents of *pt */
2112         vm_page_t m;
2113
2114 retry:
2115         /*
2116          * Basic tests, require a non-NULL vm_map_entry, require proper
2117          * alignment and type for the vm_map_entry, require that the
2118          * underlying object already be allocated.
2119          *
2120          * We allow almost any type of object to use this optimization.
2121          * The object itself does NOT have to be sized to a multiple of the
2122          * segment size, but the memory mapping does.
2123          *
2124          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2125          *     won't work as expected.
2126          */
2127         if (entry == NULL ||
2128             pmap_mmu_optimize == 0 ||                   /* not enabled */
2129             (pmap->pm_flags & PMAP_HVM) ||              /* special pmap */
2130             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal or pt */
2131             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2132             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2133             entry->object.vm_object == NULL ||          /* needs VM object */
2134             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2135             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2136             (entry->offset & SEG_MASK) ||               /* must be aligned */
2137             (entry->start & SEG_MASK)) {
2138                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2139         }
2140
2141         /*
2142          * Make sure the full segment can be represented.
2143          */
2144         b = va & ~(vm_offset_t)SEG_MASK;
2145         if (b < entry->start || b + SEG_SIZE > entry->end)
2146                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2147
2148         /*
2149          * If the full segment can be represented dive the VM object's
2150          * shared pmap, allocating as required.
2151          */
2152         object = entry->object.vm_object;
2153
2154         if (entry->protection & VM_PROT_WRITE)
2155                 obpmapp = &object->md.pmap_rw;
2156         else
2157                 obpmapp = &object->md.pmap_ro;
2158
2159 #ifdef PMAP_DEBUG2
2160         if (pmap_enter_debug > 0) {
2161                 --pmap_enter_debug;
2162                 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2163                         "obpmapp %p %p\n",
2164                         va, entry->protection, object,
2165                         obpmapp, *obpmapp);
2166                 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2167                         entry, entry->start, entry->end);
2168         }
2169 #endif
2170
2171         /*
2172          * We allocate what appears to be a normal pmap but because portions
2173          * of this pmap are shared with other unrelated pmaps we have to
2174          * set pm_active to point to all cpus.
2175          *
2176          * XXX Currently using pmap_spin to interlock the update, can't use
2177          *     vm_object_hold/drop because the token might already be held
2178          *     shared OR exclusive and we don't know.
2179          */
2180         while ((obpmap = *obpmapp) == NULL) {
2181                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2182                 pmap_pinit_simple(obpmap);
2183                 pmap_pinit2(obpmap);
2184                 spin_lock(&pmap_spin);
2185                 if (*obpmapp != NULL) {
2186                         /*
2187                          * Handle race
2188                          */
2189                         spin_unlock(&pmap_spin);
2190                         pmap_release(obpmap);
2191                         pmap_puninit(obpmap);
2192                         kfree(obpmap, M_OBJPMAP);
2193                         obpmap = *obpmapp; /* safety */
2194                 } else {
2195                         obpmap->pm_active = smp_active_mask;
2196                         obpmap->pm_flags |= PMAP_SEGSHARED;
2197                         *obpmapp = obpmap;
2198                         spin_unlock(&pmap_spin);
2199                 }
2200         }
2201
2202         /*
2203          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2204          * pte/pt using the shared pmap from the object but also adjust
2205          * the process pmap's page table page as a side effect.
2206          */
2207
2208         /*
2209          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2210          * we will return.  This is true if ptepindex represents a terminal
2211          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2212          * the PD.
2213          */
2214         pt_pv = NULL;
2215         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2216         if (ptepindex >= pmap_pt_pindex(0))
2217                 xpv = pte_pv;
2218         else
2219                 xpv = pt_pv;
2220
2221         /*
2222          * Resolve the PD in the process pmap so we can properly share the
2223          * page table page.  Lock order is bottom-up (leaf first)!
2224          *
2225          * NOTE: proc_pt_pv can be NULL.
2226          */
2227         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2228         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2229 #ifdef PMAP_DEBUG2
2230         if (pmap_enter_debug > 0) {
2231                 --pmap_enter_debug;
2232                 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2233                         proc_pt_pv,
2234                         (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2235                         proc_pd_pv,
2236                         va);
2237         }
2238 #endif
2239
2240         /*
2241          * xpv is the page table page pv from the shared object
2242          * (for convenience), from above.
2243          *
2244          * Calculate the pte value for the PT to load into the process PD.
2245          * If we have to change it we must properly dispose of the previous
2246          * entry.
2247          */
2248         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2249         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2250                (pmap->pmap_bits[PG_U_IDX] |
2251                 pmap->pmap_bits[PG_RW_IDX] |
2252                 pmap->pmap_bits[PG_V_IDX] |
2253                 pmap->pmap_bits[PG_A_IDX] |
2254                 pmap->pmap_bits[PG_M_IDX]);
2255
2256         /*
2257          * Dispose of previous page table page if it was local to the
2258          * process pmap.  If the old pt is not empty we cannot dispose of it
2259          * until we clean it out.  This case should not arise very often so
2260          * it is not optimized.
2261          */
2262         if (proc_pt_pv) {
2263                 pmap_inval_bulk_t bulk;
2264
2265                 if (proc_pt_pv->pv_m->wire_count != 1) {
2266                         pv_put(proc_pd_pv);
2267                         pv_put(proc_pt_pv);
2268                         pv_put(pt_pv);
2269                         pv_put(pte_pv);
2270                         pmap_remove(pmap,
2271                                     va & ~(vm_offset_t)SEG_MASK,
2272                                     (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2273                         goto retry;
2274                 }
2275
2276                 /*
2277                  * The release call will indirectly clean out *pt
2278                  */
2279                 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2280                 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2281                 pmap_inval_bulk_flush(&bulk);
2282                 proc_pt_pv = NULL;
2283                 /* relookup */
2284                 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2285         }
2286
2287         /*
2288          * Handle remaining cases.
2289          */
2290         if (*pt == 0) {
2291                 *pt = npte;
2292                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2293                 vm_page_wire_quick(proc_pd_pv->pv_m);   /* proc pd for sh pt */
2294                 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2295         } else if (*pt != npte) {
2296                 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2297
2298 #if 0
2299                 opte = pte_load_clear(pt);
2300                 KKASSERT(opte && opte != npte);
2301
2302                 *pt = npte;
2303 #endif
2304                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2305
2306                 /*
2307                  * Clean up opte, bump the wire_count for the process
2308                  * PD page representing the new entry if it was
2309                  * previously empty.
2310                  *
2311                  * If the entry was not previously empty and we have
2312                  * a PT in the proc pmap then opte must match that
2313                  * pt.  The proc pt must be retired (this is done
2314                  * later on in this procedure).
2315                  *
2316                  * NOTE: replacing valid pte, wire_count on proc_pd_pv
2317                  * stays the same.
2318                  */
2319                 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2320                 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2321                 if (vm_page_unwire_quick(m)) {
2322                         panic("pmap_allocpte_seg: "
2323                               "bad wire count %p",
2324                               m);
2325                 }
2326         }
2327
2328         /*
2329          * The existing process page table was replaced and must be destroyed
2330          * here.
2331          */
2332         if (proc_pd_pv)
2333                 pv_put(proc_pd_pv);
2334         if (pvpp)
2335                 *pvpp = pt_pv;
2336         else
2337                 pv_put(pt_pv);
2338
2339         return (pte_pv);
2340 }
2341
2342 /*
2343  * Release any resources held by the given physical map.
2344  *
2345  * Called when a pmap initialized by pmap_pinit is being released.  Should
2346  * only be called if the map contains no valid mappings.
2347  *
2348  * Caller must hold pmap->pm_token
2349  */
2350 struct pmap_release_info {
2351         pmap_t  pmap;
2352         int     retry;
2353         pv_entry_t pvp;
2354 };
2355
2356 static int pmap_release_callback(pv_entry_t pv, void *data);
2357
2358 void
2359 pmap_release(struct pmap *pmap)
2360 {
2361         struct pmap_release_info info;
2362
2363         KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2364                 ("pmap still active! %016jx",
2365                 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2366
2367         spin_lock(&pmap_spin);
2368         TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
2369         spin_unlock(&pmap_spin);
2370
2371         /*
2372          * Pull pv's off the RB tree in order from low to high and release
2373          * each page.
2374          */
2375         info.pmap = pmap;
2376         do {
2377                 info.retry = 0;
2378                 info.pvp = NULL;
2379
2380                 spin_lock(&pmap->pm_spin);
2381                 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2382                         pmap_release_callback, &info);
2383                 spin_unlock(&pmap->pm_spin);
2384
2385                 if (info.pvp)
2386                         pv_put(info.pvp);
2387         } while (info.retry);
2388
2389
2390         /*
2391          * One resident page (the pml4 page) should remain.
2392          * No wired pages should remain.
2393          */
2394         KKASSERT(pmap->pm_stats.resident_count ==
2395                  ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2396
2397         KKASSERT(pmap->pm_stats.wired_count == 0);
2398 }
2399
2400 /*
2401  * Called from low to high.  We must cache the proper parent pv so we
2402  * can adjust its wired count.
2403  */
2404 static int
2405 pmap_release_callback(pv_entry_t pv, void *data)
2406 {
2407         struct pmap_release_info *info = data;
2408         pmap_t pmap = info->pmap;
2409         vm_pindex_t pindex;
2410         int r;
2411
2412         if (info->pvp == pv) {
2413                 spin_unlock(&pmap->pm_spin);
2414                 info->pvp = NULL;
2415         } else if (pv_hold_try(pv)) {
2416                 spin_unlock(&pmap->pm_spin);
2417         } else {
2418                 spin_unlock(&pmap->pm_spin);
2419                 pv_lock(pv);
2420         }
2421         if (pv->pv_pmap != pmap) {
2422                 pv_put(pv);
2423                 spin_lock(&pmap->pm_spin);
2424                 info->retry = 1;
2425                 return(-1);
2426         }
2427
2428         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2429                 /*
2430                  * parent is PT
2431                  */
2432                 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2433                 pindex += NUPTE_TOTAL;
2434         } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2435                 /*
2436                  * parent is PD
2437                  */
2438                 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2439                 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2440         } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2441                 /*
2442                  * parent is PDP
2443                  */
2444                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2445                          NPDPEPGSHIFT;
2446                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2447         } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2448                 /*
2449                  * parent is PML4 (there's only one)
2450                  */
2451 #if 0
2452                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2453                            NUPD_TOTAL) >> NPML4EPGSHIFT;
2454                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2455 #endif
2456                 pindex = pmap_pml4_pindex();
2457         } else {
2458                 /*
2459                  * parent is NULL
2460                  */
2461                 if (info->pvp) {
2462                         pv_put(info->pvp);
2463                         info->pvp = NULL;
2464                 }
2465                 pindex = 0;
2466         }
2467         if (pindex) {
2468                 if (info->pvp && info->pvp->pv_pindex != pindex) {
2469                         pv_put(info->pvp);
2470                         info->pvp = NULL;
2471                 }
2472                 if (info->pvp == NULL)
2473                         info->pvp = pv_get(pmap, pindex);
2474         } else {
2475                 if (info->pvp) {
2476                         pv_put(info->pvp);
2477                         info->pvp = NULL;
2478                 }
2479         }
2480         r = pmap_release_pv(pv, info->pvp, NULL);
2481         spin_lock(&pmap->pm_spin);
2482         return(r);
2483 }
2484
2485 /*
2486  * Called with held (i.e. also locked) pv.  This function will dispose of
2487  * the lock along with the pv.
2488  *
2489  * If the caller already holds the locked parent page table for pv it
2490  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2491  * pass NULL for pvp.
2492  */
2493 static int
2494 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2495 {
2496         vm_page_t p;
2497
2498         /*
2499          * The pmap is currently not spinlocked, pv is held+locked.
2500          * Remove the pv's page from its parent's page table.  The
2501          * parent's page table page's wire_count will be decremented.
2502          *
2503          * This will clean out the pte at any level of the page table.
2504          * If smp != 0 all cpus are affected.
2505          *
2506          * Do not tear-down recursively, its faster to just let the
2507          * release run its course.
2508          */
2509         pmap_remove_pv_pte(pv, pvp, bulk, 0);
2510
2511         /*
2512          * Terminal pvs are unhooked from their vm_pages.  Because
2513          * terminal pages aren't page table pages they aren't wired
2514          * by us, so we have to be sure not to unwire them either.
2515          */
2516         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2517                 pmap_remove_pv_page(pv);
2518                 goto skip;
2519         }
2520
2521         /*
2522          * We leave the top-level page table page cached, wired, and
2523          * mapped in the pmap until the dtor function (pmap_puninit())
2524          * gets called.
2525          *
2526          * Since we are leaving the top-level pv intact we need
2527          * to break out of what would otherwise be an infinite loop.
2528          */
2529         if (pv->pv_pindex == pmap_pml4_pindex()) {
2530                 pv_put(pv);
2531                 return(-1);
2532         }
2533
2534         /*
2535          * For page table pages (other than the top-level page),
2536          * remove and free the vm_page.  The representitive mapping
2537          * removed above by pmap_remove_pv_pte() did not undo the
2538          * last wire_count so we have to do that as well.
2539          */
2540         p = pmap_remove_pv_page(pv);
2541         vm_page_busy_wait(p, FALSE, "pmaprl");
2542         if (p->wire_count != 1) {
2543                 kprintf("p->wire_count was %016lx %d\n",
2544                         pv->pv_pindex, p->wire_count);
2545         }
2546         KKASSERT(p->wire_count == 1);
2547         KKASSERT(p->flags & PG_UNMANAGED);
2548
2549         vm_page_unwire(p, 0);
2550         KKASSERT(p->wire_count == 0);
2551
2552         vm_page_free(p);
2553 skip:
2554         pv_free(pv, pvp, 1);
2555
2556         return 0;
2557 }
2558
2559 /*
2560  * This function will remove the pte associated with a pv from its parent.
2561  * Terminal pv's are supported.  All cpus specified by (bulk) are properly
2562  * invalidated.
2563  *
2564  * The wire count will be dropped on the parent page table.  The wire
2565  * count on the page being removed (pv->pv_m) from the parent page table
2566  * is NOT touched.  Note that terminal pages will not have any additional
2567  * wire counts while page table pages will have at least one representing
2568  * the mapping, plus others representing sub-mappings.
2569  *
2570  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2571  *       pages and user page table and terminal pages.
2572  *
2573  * The pv must be locked.  The pvp, if supplied, must be locked.  All
2574  * supplied pv's will remain locked on return.
2575  *
2576  * XXX must lock parent pv's if they exist to remove pte XXX
2577  */
2578 static
2579 void
2580 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2581                    int destroy)
2582 {
2583         vm_pindex_t ptepindex = pv->pv_pindex;
2584         pmap_t pmap = pv->pv_pmap;
2585         vm_page_t p;
2586         int gotpvp = 0;
2587
2588         KKASSERT(pmap);
2589
2590         if (ptepindex == pmap_pml4_pindex()) {
2591                 /*
2592                  * We are the top level pml4 table, there is no parent.
2593                  */
2594                 p = pmap->pm_pmlpv->pv_m;
2595         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2596                 /*
2597                  * Remove a PDP page from the pml4e.  This can only occur
2598                  * with user page tables.  We do not have to lock the
2599                  * pml4 PV so just ignore pvp.
2600                  */
2601                 vm_pindex_t pml4_pindex;
2602                 vm_pindex_t pdp_index;
2603                 pml4_entry_t *pdp;
2604
2605                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2606                 if (pvp == NULL) {
2607                         pml4_pindex = pmap_pml4_pindex();
2608                         pvp = pv_get(pv->pv_pmap, pml4_pindex);
2609                         KKASSERT(pvp);
2610                         gotpvp = 1;
2611                 }
2612                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2613                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2614                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2615                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2616         } else if (ptepindex >= pmap_pd_pindex(0)) {
2617                 /*
2618                  * Remove a PD page from the pdp
2619                  *
2620                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2621                  *                   of a simple pmap because it stops at
2622                  *                   the PD page.
2623                  */
2624                 vm_pindex_t pdp_pindex;
2625                 vm_pindex_t pd_index;
2626                 pdp_entry_t *pd;
2627
2628                 pd_index = ptepindex - pmap_pd_pindex(0);
2629
2630                 if (pvp == NULL) {
2631                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2632                                      (pd_index >> NPML4EPGSHIFT);
2633                         pvp = pv_get(pv->pv_pmap, pdp_pindex);
2634                         gotpvp = 1;
2635                 }
2636                 if (pvp) {
2637                         pd = pv_pte_lookup(pvp, pd_index &
2638                                                 ((1ul << NPDPEPGSHIFT) - 1));
2639                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2640                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2641                         pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2642                 } else {
2643                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2644                         p = pv->pv_m;           /* degenerate test later */
2645                 }
2646         } else if (ptepindex >= pmap_pt_pindex(0)) {
2647                 /*
2648                  *  Remove a PT page from the pd
2649                  */
2650                 vm_pindex_t pd_pindex;
2651                 vm_pindex_t pt_index;
2652                 pd_entry_t *pt;
2653
2654                 pt_index = ptepindex - pmap_pt_pindex(0);
2655
2656                 if (pvp == NULL) {
2657                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2658                                     (pt_index >> NPDPEPGSHIFT);
2659                         pvp = pv_get(pv->pv_pmap, pd_pindex);
2660                         KKASSERT(pvp);
2661                         gotpvp = 1;
2662                 }
2663                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2664                 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2665                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2666                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2667         } else {
2668                 /*
2669                  * Remove a PTE from the PT page
2670                  *
2671                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2672                  *       pv is a pte_pv so we can safely lock pt_pv.
2673                  *
2674                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2675                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2676                  *       terminal ptes.
2677                  */
2678                 vm_pindex_t pt_pindex;
2679                 pt_entry_t *ptep;
2680                 pt_entry_t pte;
2681                 vm_offset_t va;
2682
2683                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2684                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2685
2686                 if (ptepindex >= NUPTE_USER) {
2687                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2688                         KKASSERT(pvp == NULL);
2689                 } else {
2690                         if (pvp == NULL) {
2691                                 pt_pindex = NUPTE_TOTAL +
2692                                             (ptepindex >> NPDPEPGSHIFT);
2693                                 pvp = pv_get(pv->pv_pmap, pt_pindex);
2694                                 KKASSERT(pvp);
2695                                 gotpvp = 1;
2696                         }
2697                         ptep = pv_pte_lookup(pvp, ptepindex &
2698                                                   ((1ul << NPDPEPGSHIFT) - 1));
2699                 }
2700                 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2701                 if (bulk == NULL)               /* XXX */
2702                         cpu_invlpg((void *)va); /* XXX */
2703
2704                 /*
2705                  * Now update the vm_page_t
2706                  */
2707                 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) !=
2708                     (pmap->pmap_bits[PG_MANAGED_IDX]|pmap->pmap_bits[PG_V_IDX])) {
2709                         kprintf("remove_pte badpte %016lx %016lx %d\n",
2710                                 pte, pv->pv_pindex,
2711                                 pv->pv_pindex < pmap_pt_pindex(0));
2712                 }
2713                 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2714                 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2715                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2716                         p = pv->pv_m;
2717                 else
2718                         p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2719                 /* p = pv->pv_m; */
2720
2721                 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2722                         if (pmap_track_modified(ptepindex))
2723                                 vm_page_dirty(p);
2724                 }
2725                 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2726                         vm_page_flag_set(p, PG_REFERENCED);
2727                 }
2728                 if (pte & pmap->pmap_bits[PG_W_IDX])
2729                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
2730                 if (pte & pmap->pmap_bits[PG_G_IDX])
2731                         cpu_invlpg((void *)va);
2732         }
2733         KKASSERT(pv->pv_m == p);        /* XXX remove me later */
2734
2735         /*
2736          * If requested, scrap the underlying pv->pv_m and the underlying
2737          * pv.  If this is a page-table-page we must also free the page.
2738          *
2739          * pvp must be returned locked.
2740          */
2741         if (destroy == 1) {
2742                 /*
2743                  * page table page (PT, PD, PDP, PML4), caller was responsible
2744                  * for testing wired_count.
2745                  */
2746                 vm_page_t p;
2747
2748                 KKASSERT(pv->pv_m->wire_count == 1);
2749                 p = pmap_remove_pv_page(pv);
2750                 pv_free(pv, pvp, 1);
2751                 pv = NULL;
2752
2753                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
2754                 vm_page_busy_wait(p, FALSE, "pgpun");
2755                 vm_page_unwire(p, 0);
2756                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2757                 vm_page_free(p);
2758         } else if (destroy == 2) {
2759                 /*
2760                  * Normal page (leave page untouched)
2761                  */
2762                 pmap_remove_pv_page(pv);
2763                 pv_free(pv, pvp, 1);
2764                 pv = NULL;              /* safety */
2765         }
2766
2767         /*
2768          * If we acquired pvp ourselves then we are responsible for
2769          * recursively deleting it.
2770          */
2771         if (pvp && gotpvp) {
2772                 /*
2773                  * Recursively destroy higher-level page tables.
2774                  *
2775                  * This is optional.  If we do not, they will still
2776                  * be destroyed when the process exits.
2777                  */
2778                 if (pvp->pv_m &&
2779                     pvp->pv_m->wire_count == 1 &&
2780                     pvp->pv_pindex != pmap_pml4_pindex()) {
2781                         pmap_remove_pv_pte(pvp, NULL, bulk, 1);
2782                         pvp = NULL;     /* safety */
2783                 } else {
2784                         pv_put(pvp);
2785                 }
2786         }
2787 }
2788
2789 /*
2790  * Remove the vm_page association to a pv.  The pv must be locked.
2791  */
2792 static
2793 vm_page_t
2794 pmap_remove_pv_page(pv_entry_t pv)
2795 {
2796         vm_page_t m;
2797
2798         m = pv->pv_m;
2799         KKASSERT(m);
2800         vm_page_spin_lock(m);
2801         pv->pv_m = NULL;
2802         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2803         pmap_page_stats_deleting(m);
2804         /*
2805         if (m->object)
2806                 atomic_add_int(&m->object->agg_pv_list_count, -1);
2807         */
2808         if (TAILQ_EMPTY(&m->md.pv_list))
2809                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2810         vm_page_spin_unlock(m);
2811
2812         return(m);
2813 }
2814
2815 /*
2816  * Grow the number of kernel page table entries, if needed.
2817  *
2818  * This routine is always called to validate any address space
2819  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
2820  * space below KERNBASE.
2821  */
2822 void
2823 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2824 {
2825         vm_paddr_t paddr;
2826         vm_offset_t ptppaddr;
2827         vm_page_t nkpg;
2828         pd_entry_t *pt, newpt;
2829         pdp_entry_t newpd;
2830         int update_kernel_vm_end;
2831
2832         /*
2833          * bootstrap kernel_vm_end on first real VM use
2834          */
2835         if (kernel_vm_end == 0) {
2836                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2837                 nkpt = 0;
2838                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2839                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2840                                         ~(PAGE_SIZE * NPTEPG - 1);
2841                         nkpt++;
2842                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2843                                 kernel_vm_end = kernel_map.max_offset;
2844                                 break;
2845                         }
2846                 }
2847         }
2848
2849         /*
2850          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
2851          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
2852          * do not want to force-fill 128G worth of page tables.
2853          */
2854         if (kstart < KERNBASE) {
2855                 if (kstart > kernel_vm_end)
2856                         kstart = kernel_vm_end;
2857                 KKASSERT(kend <= KERNBASE);
2858                 update_kernel_vm_end = 1;
2859         } else {
2860                 update_kernel_vm_end = 0;
2861         }
2862
2863         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2864         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2865
2866         if (kend - 1 >= kernel_map.max_offset)
2867                 kend = kernel_map.max_offset;
2868
2869         while (kstart < kend) {
2870                 pt = pmap_pt(&kernel_pmap, kstart);
2871                 if (pt == NULL) {
2872                         /* We need a new PDP entry */
2873                         nkpg = vm_page_alloc(NULL, nkpt,
2874                                              VM_ALLOC_NORMAL |
2875                                              VM_ALLOC_SYSTEM |
2876                                              VM_ALLOC_INTERRUPT);
2877                         if (nkpg == NULL) {
2878                                 panic("pmap_growkernel: no memory to grow "
2879                                       "kernel");
2880                         }
2881                         paddr = VM_PAGE_TO_PHYS(nkpg);
2882                         pmap_zero_page(paddr);
2883                         newpd = (pdp_entry_t)
2884                             (paddr |
2885                             kernel_pmap.pmap_bits[PG_V_IDX] |
2886                             kernel_pmap.pmap_bits[PG_RW_IDX] |
2887                             kernel_pmap.pmap_bits[PG_A_IDX] |
2888                             kernel_pmap.pmap_bits[PG_M_IDX]);
2889                         *pmap_pd(&kernel_pmap, kstart) = newpd;
2890                         nkpt++;
2891                         continue; /* try again */
2892                 }
2893                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2894                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
2895                                  ~(PAGE_SIZE * NPTEPG - 1);
2896                         if (kstart - 1 >= kernel_map.max_offset) {
2897                                 kstart = kernel_map.max_offset;
2898                                 break;
2899                         }
2900                         continue;
2901                 }
2902
2903                 /*
2904                  * This index is bogus, but out of the way
2905                  */
2906                 nkpg = vm_page_alloc(NULL, nkpt,
2907                                      VM_ALLOC_NORMAL |
2908                                      VM_ALLOC_SYSTEM |
2909                                      VM_ALLOC_INTERRUPT);
2910                 if (nkpg == NULL)
2911                         panic("pmap_growkernel: no memory to grow kernel");
2912
2913                 vm_page_wire(nkpg);
2914                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2915                 pmap_zero_page(ptppaddr);
2916                 newpt = (pd_entry_t) (ptppaddr |
2917                     kernel_pmap.pmap_bits[PG_V_IDX] |
2918                     kernel_pmap.pmap_bits[PG_RW_IDX] |
2919                     kernel_pmap.pmap_bits[PG_A_IDX] |
2920                     kernel_pmap.pmap_bits[PG_M_IDX]);
2921                 *pmap_pt(&kernel_pmap, kstart) = newpt;
2922                 nkpt++;
2923
2924                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2925                           ~(PAGE_SIZE * NPTEPG - 1);
2926
2927                 if (kstart - 1 >= kernel_map.max_offset) {
2928                         kstart = kernel_map.max_offset;
2929                         break;
2930                 }
2931         }
2932
2933         /*
2934          * Only update kernel_vm_end for areas below KERNBASE.
2935          */
2936         if (update_kernel_vm_end && kernel_vm_end < kstart)
2937                 kernel_vm_end = kstart;
2938 }
2939
2940 /*
2941  *      Add a reference to the specified pmap.
2942  */
2943 void
2944 pmap_reference(pmap_t pmap)
2945 {
2946         if (pmap != NULL) {
2947                 lwkt_gettoken(&pmap->pm_token);
2948                 ++pmap->pm_count;
2949                 lwkt_reltoken(&pmap->pm_token);
2950         }
2951 }
2952
2953 /***************************************************
2954  * page management routines.
2955  ***************************************************/
2956
2957 /*
2958  * Hold a pv without locking it
2959  */
2960 static void
2961 pv_hold(pv_entry_t pv)
2962 {
2963         atomic_add_int(&pv->pv_hold, 1);
2964 }
2965
2966 /*
2967  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
2968  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
2969  * the pv properly.
2970  *
2971  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2972  * pv list via its page) must be held by the caller.
2973  */
2974 static int
2975 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2976 {
2977         u_int count;
2978
2979         /*
2980          * Critical path shortcut expects pv to already have one ref
2981          * (for the pv->pv_pmap).
2982          */
2983         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
2984 #ifdef PMAP_DEBUG
2985                 pv->pv_func = func;
2986                 pv->pv_line = lineno;
2987 #endif
2988                 return TRUE;
2989         }
2990
2991         for (;;) {
2992                 count = pv->pv_hold;
2993                 cpu_ccfence();
2994                 if ((count & PV_HOLD_LOCKED) == 0) {
2995                         if (atomic_cmpset_int(&pv->pv_hold, count,
2996                                               (count + 1) | PV_HOLD_LOCKED)) {
2997 #ifdef PMAP_DEBUG
2998                                 pv->pv_func = func;
2999                                 pv->pv_line = lineno;
3000 #endif
3001                                 return TRUE;
3002                         }
3003                 } else {
3004                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3005                                 return FALSE;
3006                 }
3007                 /* retry */
3008         }
3009 }
3010
3011 /*
3012  * Drop a previously held pv_entry which could not be locked, allowing its
3013  * destruction.
3014  *
3015  * Must not be called with a spinlock held as we might zfree() the pv if it
3016  * is no longer associated with a pmap and this was the last hold count.
3017  */
3018 static void
3019 pv_drop(pv_entry_t pv)
3020 {
3021         u_int count;
3022
3023         for (;;) {
3024                 count = pv->pv_hold;
3025                 cpu_ccfence();
3026                 KKASSERT((count & PV_HOLD_MASK) > 0);
3027                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3028                          (PV_HOLD_LOCKED | 1));
3029                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3030                         if ((count & PV_HOLD_MASK) == 1) {
3031 #ifdef PMAP_DEBUG2
3032                                 if (pmap_enter_debug > 0) {
3033                                         --pmap_enter_debug;
3034                                         kprintf("pv_drop: free pv %p\n", pv);
3035                                 }
3036 #endif
3037                                 KKASSERT(count == 1);
3038                                 KKASSERT(pv->pv_pmap == NULL);
3039                                 zfree(pvzone, pv);
3040                         }
3041                         return;
3042                 }
3043                 /* retry */
3044         }
3045 }
3046
3047 /*
3048  * Find or allocate the requested PV entry, returning a locked, held pv.
3049  *
3050  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3051  * for the caller and one representing the pmap and vm_page association.
3052  *
3053  * If (*isnew) is zero, the returned pv will have only one hold count.
3054  *
3055  * Since both associations can only be adjusted while the pv is locked,
3056  * together they represent just one additional hold.
3057  */
3058 static
3059 pv_entry_t
3060 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3061 {
3062         pv_entry_t pv;
3063         pv_entry_t pnew = NULL;
3064
3065         spin_lock(&pmap->pm_spin);
3066         for (;;) {
3067                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3068                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3069                                                         pindex);
3070                 }
3071                 if (pv == NULL) {
3072                         if (pnew == NULL) {
3073                                 spin_unlock(&pmap->pm_spin);
3074                                 pnew = zalloc(pvzone);
3075                                 spin_lock(&pmap->pm_spin);
3076                                 continue;
3077                         }
3078                         pnew->pv_pmap = pmap;
3079                         pnew->pv_pindex = pindex;
3080                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
3081 #ifdef PMAP_DEBUG
3082                         pnew->pv_func = func;
3083                         pnew->pv_line = lineno;
3084 #endif
3085                         pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3086                         ++pmap->pm_generation;
3087                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
3088                         spin_unlock(&pmap->pm_spin);
3089                         *isnew = 1;
3090                         return(pnew);
3091                 }
3092                 if (pnew) {
3093                         spin_unlock(&pmap->pm_spin);
3094                         zfree(pvzone, pnew);
3095                         pnew = NULL;
3096                         spin_lock(&pmap->pm_spin);
3097                         continue;
3098                 }
3099                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3100                         spin_unlock(&pmap->pm_spin);
3101                 } else {
3102                         spin_unlock(&pmap->pm_spin);
3103                         _pv_lock(pv PMAP_DEBUG_COPY);
3104                 }
3105                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3106                         *isnew = 0;
3107                         return(pv);
3108                 }
3109                 pv_put(pv);
3110                 spin_lock(&pmap->pm_spin);
3111         }
3112 }
3113
3114 /*
3115  * Find the requested PV entry, returning a locked+held pv or NULL
3116  */
3117 static
3118 pv_entry_t
3119 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
3120 {
3121         pv_entry_t pv;
3122
3123         spin_lock(&pmap->pm_spin);
3124         for (;;) {
3125                 /*
3126                  * Shortcut cache
3127                  */
3128                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3129                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3130                                                         pindex);
3131                 }
3132                 if (pv == NULL) {
3133                         spin_unlock(&pmap->pm_spin);
3134                         return NULL;
3135                 }
3136                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3137                         spin_unlock(&pmap->pm_spin);
3138                 } else {
3139                         spin_unlock(&pmap->pm_spin);
3140                         _pv_lock(pv PMAP_DEBUG_COPY);
3141                 }
3142                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3143                         pv_cache(pv, pindex);
3144                         return(pv);
3145                 }
3146                 pv_put(pv);
3147                 spin_lock(&pmap->pm_spin);
3148         }
3149 }
3150
3151 /*
3152  * Lookup, hold, and attempt to lock (pmap,pindex).
3153  *
3154  * If the entry does not exist NULL is returned and *errorp is set to 0
3155  *
3156  * If the entry exists and could be successfully locked it is returned and
3157  * errorp is set to 0.
3158  *
3159  * If the entry exists but could NOT be successfully locked it is returned
3160  * held and *errorp is set to 1.
3161  */
3162 static
3163 pv_entry_t
3164 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
3165 {
3166         pv_entry_t pv;
3167
3168         spin_lock_shared(&pmap->pm_spin);
3169         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3170                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3171         if (pv == NULL) {
3172                 spin_unlock_shared(&pmap->pm_spin);
3173                 *errorp = 0;
3174                 return NULL;
3175         }
3176         if (pv_hold_try(pv)) {
3177                 pv_cache(pv, pindex);
3178                 spin_unlock_shared(&pmap->pm_spin);
3179                 *errorp = 0;
3180                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3181                 return(pv);     /* lock succeeded */
3182         }
3183         spin_unlock_shared(&pmap->pm_spin);
3184         *errorp = 1;
3185         return (pv);            /* lock failed */
3186 }
3187
3188 /*
3189  * Find the requested PV entry, returning a held pv or NULL
3190  */
3191 static
3192 pv_entry_t
3193 pv_find(pmap_t pmap, vm_pindex_t pindex)
3194 {
3195         pv_entry_t pv;
3196
3197         spin_lock_shared(&pmap->pm_spin);
3198
3199         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3200                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3201         if (pv == NULL) {
3202                 spin_unlock_shared(&pmap->pm_spin);
3203                 return NULL;
3204         }
3205         pv_hold(pv);
3206         pv_cache(pv, pindex);
3207         spin_unlock_shared(&pmap->pm_spin);
3208         return(pv);
3209 }
3210
3211 /*
3212  * Lock a held pv, keeping the hold count
3213  */
3214 static
3215 void
3216 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3217 {
3218         u_int count;
3219
3220         for (;;) {
3221                 count = pv->pv_hold;
3222                 cpu_ccfence();
3223                 if ((count & PV_HOLD_LOCKED) == 0) {
3224                         if (atomic_cmpset_int(&pv->pv_hold, count,
3225                                               count | PV_HOLD_LOCKED)) {
3226 #ifdef PMAP_DEBUG
3227                                 pv->pv_func = func;
3228                                 pv->pv_line = lineno;
3229 #endif
3230                                 return;
3231                         }
3232                         continue;
3233                 }
3234                 tsleep_interlock(pv, 0);
3235                 if (atomic_cmpset_int(&pv->pv_hold, count,
3236                                       count | PV_HOLD_WAITING)) {
3237 #ifdef PMAP_DEBUG
3238                         kprintf("pv waiting on %s:%d\n",
3239                                         pv->pv_func, pv->pv_line);
3240 #endif
3241                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
3242                 }
3243                 /* retry */
3244         }
3245 }
3246
3247 /*
3248  * Unlock a held and locked pv, keeping the hold count.
3249  */
3250 static
3251 void
3252 pv_unlock(pv_entry_t pv)
3253 {
3254         u_int count;
3255
3256         for (;;) {
3257                 count = pv->pv_hold;
3258                 cpu_ccfence();
3259                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3260                          (PV_HOLD_LOCKED | 1));
3261                 if (atomic_cmpset_int(&pv->pv_hold, count,
3262                                       count &
3263                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3264                         if (count & PV_HOLD_WAITING)
3265                                 wakeup(pv);
3266                         break;
3267                 }
3268         }
3269 }
3270
3271 /*
3272  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3273  * and the hold count drops to zero we will free it.
3274  *
3275  * Caller should not hold any spin locks.  We are protected from hold races
3276  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3277  * lock held.  A pv cannot be located otherwise.
3278  */
3279 static
3280 void
3281 pv_put(pv_entry_t pv)
3282 {
3283 #ifdef PMAP_DEBUG2
3284         if (pmap_enter_debug > 0) {
3285                 --pmap_enter_debug;
3286                 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3287         }
3288 #endif
3289
3290         /*
3291          * Fast - shortcut most common condition
3292          */
3293         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3294                 return;
3295
3296         /*
3297          * Slow
3298          */
3299         pv_unlock(pv);
3300         pv_drop(pv);
3301 }
3302
3303 /*
3304  * Remove the pmap association from a pv, require that pv_m already be removed,
3305  * then unlock and drop the pv.  Any pte operations must have already been
3306  * completed.  This call may result in a last-drop which will physically free
3307  * the pv.
3308  *
3309  * Removing the pmap association entails an additional drop.
3310  *
3311  * pv must be exclusively locked on call and will be disposed of on return.
3312  */
3313 static
3314 void
3315 pv_free(pv_entry_t pv, pv_entry_t pvp, int putaway)
3316 {
3317         pmap_t pmap;
3318
3319         KKASSERT(pv->pv_m == NULL);
3320         KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3321         if ((pmap = pv->pv_pmap) != NULL) {
3322                 spin_lock(&pmap->pm_spin);
3323                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3324                 ++pmap->pm_generation;
3325                 if (pmap->pm_pvhint == pv)
3326                         pmap->pm_pvhint = NULL;
3327                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3328                 pv->pv_pmap = NULL;
3329                 pv->pv_pindex = 0;
3330                 spin_unlock(&pmap->pm_spin);
3331
3332                 /*
3333                  * Try to shortcut three atomic ops, otherwise fall through
3334                  * and do it normally.  Drop two refs and the lock all in
3335                  * one go.
3336                  */
3337                 if (putaway &&
3338                     atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3339 #ifdef PMAP_DEBUG2
3340                         if (pmap_enter_debug > 0) {
3341                                 --pmap_enter_debug;
3342                                 kprintf("pv_free: free pv %p\n", pv);
3343                         }
3344 #endif
3345                         zfree(pvzone, pv);
3346                         if (pvp)
3347                                 vm_page_unwire_quick(pvp->pv_m);
3348                         return;
3349                 }
3350                 pv_drop(pv);    /* ref for pv_pmap */
3351         }
3352         if (putaway)
3353                 pv_put(pv);
3354         if (pvp)
3355                 vm_page_unwire_quick(pvp->pv_m);
3356 }
3357
3358 /*
3359  * This routine is very drastic, but can save the system
3360  * in a pinch.
3361  */
3362 void
3363 pmap_collect(void)
3364 {
3365         int i;
3366         vm_page_t m;
3367         static int warningdone=0;
3368
3369         if (pmap_pagedaemon_waken == 0)
3370                 return;
3371         pmap_pagedaemon_waken = 0;
3372         if (warningdone < 5) {
3373                 kprintf("pmap_collect: collecting pv entries -- "
3374                         "suggest increasing PMAP_SHPGPERPROC\n");
3375                 warningdone++;
3376         }
3377
3378         for (i = 0; i < vm_page_array_size; i++) {
3379                 m = &vm_page_array[i];
3380                 if (m->wire_count || m->hold_count)
3381                         continue;
3382                 if (vm_page_busy_try(m, TRUE) == 0) {
3383                         if (m->wire_count == 0 && m->hold_count == 0) {
3384                                 pmap_remove_all(m);
3385                         }
3386                         vm_page_wakeup(m);
3387                 }
3388         }
3389 }
3390
3391 /*
3392  * Scan the pmap for active page table entries and issue a callback.
3393  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3394  * its parent page table.
3395  *
3396  * pte_pv will be NULL if the page or page table is unmanaged.
3397  * pt_pv will point to the page table page containing the pte for the page.
3398  *
3399  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3400  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3401  *       process pmap's PD and page to the callback function.  This can be
3402  *       confusing because the pt_pv is really a pd_pv, and the target page
3403  *       table page is simply aliased by the pmap and not owned by it.
3404  *
3405  * It is assumed that the start and end are properly rounded to the page size.
3406  *
3407  * It is assumed that PD pages and above are managed and thus in the RB tree,
3408  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3409  */
3410 struct pmap_scan_info {
3411         struct pmap *pmap;
3412         vm_offset_t sva;
3413         vm_offset_t eva;
3414         vm_pindex_t sva_pd_pindex;
3415         vm_pindex_t eva_pd_pindex;
3416         void (*func)(pmap_t, struct pmap_scan_info *,
3417                      pv_entry_t, pv_entry_t, int, vm_offset_t,
3418                      pt_entry_t *, void *);
3419         void *arg;
3420         pmap_inval_bulk_t bulk_core;
3421         pmap_inval_bulk_t *bulk;
3422         int count;
3423         int stop;
3424 };
3425
3426 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3427 static int pmap_scan_callback(pv_entry_t pv, void *data);
3428
3429 static void
3430 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3431 {
3432         struct pmap *pmap = info->pmap;
3433         pv_entry_t pd_pv;       /* A page directory PV */
3434         pv_entry_t pt_pv;       /* A page table PV */
3435         pv_entry_t pte_pv;      /* A page table entry PV */
3436         pt_entry_t *ptep;
3437         pt_entry_t oldpte;
3438         struct pv_entry dummy_pv;
3439         int generation;
3440
3441         info->stop = 0;
3442         if (pmap == NULL)
3443                 return;
3444         if (smp_inval) {
3445                 info->bulk = &info->bulk_core;
3446                 pmap_inval_bulk_init(&info->bulk_core, pmap);
3447         } else {
3448                 info->bulk = NULL;
3449         }
3450
3451         /*
3452          * Hold the token for stability; if the pmap is empty we have nothing
3453          * to do.
3454          */
3455         lwkt_gettoken(&pmap->pm_token);
3456 #if 0
3457         if (pmap->pm_stats.resident_count == 0) {
3458                 lwkt_reltoken(&pmap->pm_token);
3459                 return;
3460         }
3461 #endif
3462
3463         info->count = 0;
3464
3465 again:
3466         /*
3467          * Special handling for scanning one page, which is a very common
3468          * operation (it is?).
3469          *
3470          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3471          */
3472         if (info->sva + PAGE_SIZE == info->eva) {
3473                 generation = pmap->pm_generation;
3474                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3475                         /*
3476                          * Kernel mappings do not track wire counts on
3477                          * page table pages and only maintain pd_pv and
3478                          * pte_pv levels so pmap_scan() works.
3479                          */
3480                         pt_pv = NULL;
3481                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3482                         ptep = vtopte(info->sva);
3483                 } else {
3484                         /*
3485                          * User pages which are unmanaged will not have a
3486                          * pte_pv.  User page table pages which are unmanaged
3487                          * (shared from elsewhere) will also not have a pt_pv.
3488                          * The func() callback will pass both pte_pv and pt_pv
3489                          * as NULL in that case.
3490                          */
3491                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3492                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3493                         if (pt_pv == NULL) {
3494                                 KKASSERT(pte_pv == NULL);
3495                                 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3496                                 if (pd_pv) {
3497                                         ptep = pv_pte_lookup(pd_pv,
3498                                                     pmap_pt_index(info->sva));
3499                                         if (*ptep) {
3500                                                 info->func(pmap, info,
3501                                                      NULL, pd_pv, 1,
3502                                                      info->sva, ptep,
3503                                                      info->arg);
3504                                         }
3505                                         pv_put(pd_pv);
3506                                 }
3507                                 goto fast_skip;
3508                         }
3509                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3510                 }
3511
3512                 /*
3513                  * NOTE: *ptep can't be ripped out from under us if we hold
3514                  *       pte_pv locked, but bits can change.  However, there is
3515                  *       a race where another thread may be inserting pte_pv
3516                  *       and setting *ptep just after our pte_pv lookup fails.
3517                  *
3518                  *       In this situation we can end up with a NULL pte_pv
3519                  *       but find that we have a managed *ptep.  We explicitly
3520                  *       check for this race.
3521                  */
3522                 oldpte = *ptep;
3523                 cpu_ccfence();
3524                 if (oldpte == 0) {
3525                         /*
3526                          * Unlike the pv_find() case below we actually
3527                          * acquired a locked pv in this case so any
3528                          * race should have been resolved.  It is expected
3529                          * to not exist.
3530                          */
3531                         KKASSERT(pte_pv == NULL);
3532                 } else if (pte_pv) {
3533                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3534                                            pmap->pmap_bits[PG_V_IDX])) ==
3535                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
3536                                  pmap->pmap_bits[PG_V_IDX]),
3537                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3538                              "generation %d/%d",
3539                             *ptep, oldpte, info->sva, pte_pv,
3540                             generation, pmap->pm_generation));
3541                         info->func(pmap, info, pte_pv, pt_pv, 0,
3542                                    info->sva, ptep, info->arg);
3543                 } else {
3544                         /*
3545                          * Check for insertion race
3546                          */
3547                         if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3548                             pt_pv) {
3549                                 pte_pv = pv_find(pmap,
3550                                                  pmap_pte_pindex(info->sva));
3551                                 if (pte_pv) {
3552                                         pv_drop(pte_pv);
3553                                         pv_put(pt_pv);
3554                                         kprintf("pmap_scan: RACE1 "
3555                                                 "%016jx, %016lx\n",
3556                                                 info->sva, oldpte);
3557                                         goto again;
3558                                 }
3559                         }
3560
3561                         /*
3562                          * Didn't race
3563                          */
3564                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3565                                            pmap->pmap_bits[PG_V_IDX])) ==
3566                             pmap->pmap_bits[PG_V_IDX],
3567                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3568                              "generation %d/%d",
3569                             *ptep, oldpte, info->sva,
3570                             generation, pmap->pm_generation));
3571                         info->func(pmap, info, NULL, pt_pv, 0,
3572                             info->sva, ptep, info->arg);
3573                 }
3574                 if (pt_pv)
3575                         pv_put(pt_pv);
3576 fast_skip:
3577                 pmap_inval_bulk_flush(info->bulk);
3578                 lwkt_reltoken(&pmap->pm_token);
3579                 return;
3580         }
3581
3582         /*
3583          * Nominal scan case, RB_SCAN() for PD pages and iterate from
3584          * there.
3585          */
3586         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3587         info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3588
3589         if (info->sva >= VM_MAX_USER_ADDRESS) {
3590                 /*
3591                  * The kernel does not currently maintain any pv_entry's for
3592                  * higher-level page tables.
3593                  */
3594                 bzero(&dummy_pv, sizeof(dummy_pv));
3595                 dummy_pv.pv_pindex = info->sva_pd_pindex;
3596                 spin_lock(&pmap->pm_spin);
3597                 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3598                         pmap_scan_callback(&dummy_pv, info);
3599                         ++dummy_pv.pv_pindex;
3600                 }
3601                 spin_unlock(&pmap->pm_spin);
3602         } else {
3603                 /*
3604                  * User page tables maintain local PML4, PDP, and PD
3605                  * pv_entry's at the very least.  PT pv's might be
3606                  * unmanaged and thus not exist.  PTE pv's might be
3607                  * unmanaged and thus not exist.
3608                  */
3609                 spin_lock(&pmap->pm_spin);
3610                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3611                         pmap_scan_cmp, pmap_scan_callback, info);
3612                 spin_unlock(&pmap->pm_spin);
3613         }
3614         pmap_inval_bulk_flush(info->bulk);
3615         lwkt_reltoken(&pmap->pm_token);
3616 }
3617
3618 /*
3619  * WARNING! pmap->pm_spin held
3620  */
3621 static int
3622 pmap_scan_cmp(pv_entry_t pv, void *data)
3623 {
3624         struct pmap_scan_info *info = data;
3625         if (pv->pv_pindex < info->sva_pd_pindex)
3626                 return(-1);
3627         if (pv->pv_pindex >= info->eva_pd_pindex)
3628                 return(1);
3629         return(0);
3630 }
3631
3632 /*
3633  * WARNING! pmap->pm_spin held
3634  */
3635 static int
3636 pmap_scan_callback(pv_entry_t pv, void *data)
3637 {
3638         struct pmap_scan_info *info = data;
3639         struct pmap *pmap = info->pmap;
3640         pv_entry_t pd_pv;       /* A page directory PV */
3641         pv_entry_t pt_pv;       /* A page table PV */
3642         pv_entry_t pte_pv;      /* A page table entry PV */
3643         pt_entry_t *ptep;
3644         pt_entry_t oldpte;
3645         vm_offset_t sva;
3646         vm_offset_t eva;
3647         vm_offset_t va_next;
3648         vm_pindex_t pd_pindex;
3649         int error;
3650         int generation;
3651
3652         /*
3653          * Stop if requested
3654          */
3655         if (info->stop)
3656                 return -1;
3657
3658         /*
3659          * Pull the PD pindex from the pv before releasing the spinlock.
3660          *
3661          * WARNING: pv is faked for kernel pmap scans.
3662          */
3663         pd_pindex = pv->pv_pindex;
3664         spin_unlock(&pmap->pm_spin);
3665         pv = NULL;      /* invalid after spinlock unlocked */
3666
3667         /*
3668          * Calculate the page range within the PD.  SIMPLE pmaps are
3669          * direct-mapped for the entire 2^64 address space.  Normal pmaps
3670          * reflect the user and kernel address space which requires
3671          * cannonicalization w/regards to converting pd_pindex's back
3672          * into addresses.
3673          */
3674         sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3675         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3676             (sva & PML4_SIGNMASK)) {
3677                 sva |= PML4_SIGNMASK;
3678         }
3679         eva = sva + NBPDP;      /* can overflow */
3680         if (sva < info->sva)
3681                 sva = info->sva;
3682         if (eva < info->sva || eva > info->eva)
3683                 eva = info->eva;
3684
3685         /*
3686          * NOTE: kernel mappings do not track page table pages, only
3687          *       terminal pages.
3688          *
3689          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3690          *       However, for the scan to be efficient we try to
3691          *       cache items top-down.
3692          */
3693         pd_pv = NULL;
3694         pt_pv = NULL;
3695
3696         for (; sva < eva; sva = va_next) {
3697                 if (info->stop)
3698                         break;
3699                 if (sva >= VM_MAX_USER_ADDRESS) {
3700                         if (pt_pv) {
3701                                 pv_put(pt_pv);
3702                                 pt_pv = NULL;
3703                         }
3704                         goto kernel_skip;
3705                 }
3706
3707                 /*
3708                  * PD cache (degenerate case if we skip).  It is possible
3709                  * for the PD to not exist due to races.  This is ok.
3710                  */
3711                 if (pd_pv == NULL) {
3712                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3713                 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3714                         pv_put(pd_pv);
3715                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3716                 }
3717                 if (pd_pv == NULL) {
3718                         va_next = (sva + NBPDP) & ~PDPMASK;
3719                         if (va_next < sva)
3720                                 va_next = eva;
3721                         continue;
3722                 }
3723
3724                 /*
3725                  * PT cache
3726                  */
3727                 if (pt_pv == NULL) {
3728                         vm_page_wire_quick(pd_pv->pv_m);
3729                         pv_unlock(pd_pv);
3730                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3731                         pv_lock(pd_pv);
3732                         vm_page_unwire_quick(pd_pv->pv_m);
3733                 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3734                         vm_page_wire_quick(pd_pv->pv_m);
3735                         pv_unlock(pd_pv);
3736                         pv_put(pt_pv);
3737                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3738                         pv_lock(pd_pv);
3739                         vm_page_unwire_quick(pd_pv->pv_m);
3740                 }
3741
3742                 /*
3743                  * If pt_pv is NULL we either have an shared page table
3744                  * page and must issue a callback specific to that case,
3745                  * or there is no page table page.
3746                  *
3747                  * Either way we can skip the page table page.
3748                  */
3749                 if (pt_pv == NULL) {
3750                         /*
3751                          * Possible unmanaged (shared from another pmap)
3752                          * page table page.
3753                          */
3754                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3755                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3756                                 info->func(pmap, info, NULL, pd_pv, 1,
3757                                            sva, ptep, info->arg);
3758                         }
3759
3760                         /*
3761                          * Done, move to next page table page.
3762                          */
3763                         va_next = (sva + NBPDR) & ~PDRMASK;
3764                         if (va_next < sva)
3765                                 va_next = eva;
3766                         continue;
3767                 }
3768
3769                 /*
3770                  * From this point in the loop testing pt_pv for non-NULL
3771                  * means we are in UVM, else if it is NULL we are in KVM.
3772                  *
3773                  * Limit our scan to either the end of the va represented
3774                  * by the current page table page, or to the end of the
3775                  * range being removed.
3776                  */
3777 kernel_skip:
3778                 va_next = (sva + NBPDR) & ~PDRMASK;
3779                 if (va_next < sva)
3780                         va_next = eva;
3781                 if (va_next > eva)
3782                         va_next = eva;
3783
3784                 /*
3785                  * Scan the page table for pages.  Some pages may not be
3786                  * managed (might not have a pv_entry).
3787                  *
3788                  * There is no page table management for kernel pages so
3789                  * pt_pv will be NULL in that case, but otherwise pt_pv
3790                  * is non-NULL, locked, and referenced.
3791                  */
3792
3793                 /*
3794                  * At this point a non-NULL pt_pv means a UVA, and a NULL
3795                  * pt_pv means a KVA.
3796                  */
3797                 if (pt_pv)
3798                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3799                 else
3800                         ptep = vtopte(sva);
3801
3802                 while (sva < va_next) {
3803                         /*
3804                          * Yield every 64 pages, stop if requested.
3805                          */
3806                         if ((++info->count & 63) == 0)
3807                                 lwkt_user_yield();
3808                         if (info->stop)
3809                                 break;
3810
3811                         /*
3812                          * Check if pt_pv has been lost (probably due to
3813                          * a remove of the underlying pages).
3814                          */
3815                         if (pt_pv && pt_pv->pv_pmap == NULL)
3816                                 break;
3817
3818                         /*
3819                          * Acquire the related pte_pv, if any.  If *ptep == 0
3820                          * the related pte_pv should not exist, but if *ptep
3821                          * is not zero the pte_pv may or may not exist (e.g.
3822                          * will not exist for an unmanaged page).
3823                          *
3824                          * However a multitude of races are possible here.
3825                          *
3826                          * In addition, the (pt_pv, pte_pv) lock order is
3827                          * backwards, so we have to be careful in aquiring
3828                          * a properly locked pte_pv.
3829                          */
3830                         generation = pmap->pm_generation;
3831                         if (pt_pv) {
3832                                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3833                                                     &error);
3834                                 if (error) {
3835                                         if (pd_pv) {
3836                                                 vm_page_wire_quick(pd_pv->pv_m);
3837                                                 pv_unlock(pd_pv);
3838                                         }
3839                                         vm_page_wire_quick(pt_pv->pv_m);
3840                                         pv_unlock(pt_pv);/* must be non-NULL */
3841                                         pv_lock(pte_pv); /* safe to block now */
3842                                         pv_put(pte_pv);
3843                                         pte_pv = NULL;
3844                                         pv_lock(pt_pv);
3845                                         vm_page_unwire_quick(pt_pv->pv_m);
3846
3847                                         /*
3848                                          * pt_pv reloaded, need new ptep
3849                                          */
3850                                         KKASSERT(pt_pv != NULL);
3851                                         ptep = pv_pte_lookup(pt_pv,
3852                                                         pmap_pte_index(sva));
3853                                         if (pd_pv) {
3854                                                 pv_lock(pd_pv);
3855                                                 vm_page_unwire_quick(pd_pv->pv_m);
3856                                         }
3857                                         continue;
3858                                 }
3859                         } else {
3860                                 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3861                         }
3862
3863                         /*
3864                          * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3865                          */
3866                         oldpte = *ptep;
3867                         if (oldpte == 0) {
3868                                 if (pte_pv) {
3869                                         kprintf("Unexpected non-NULL pte_pv "
3870                                                 "%p pt_pv %p "
3871                                                 "*ptep = %016lx/%016lx\n",
3872                                                 pte_pv, pt_pv, *ptep, oldpte);
3873                                         panic("Unexpected non-NULL pte_pv");
3874                                 }
3875                                 sva += PAGE_SIZE;
3876                                 ++ptep;
3877                                 continue;
3878                         }
3879
3880                         /*
3881                          * Ready for the callback.  The locked pte_pv (if any)
3882                          * is consumed by the callback.  pte_pv will exist if
3883                          * the page is managed, and will not exist if it
3884                          * isn't.
3885                          */
3886                         if (pte_pv) {
3887                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3888                                     (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3889                                     ("badC *ptep %016lx/%016lx sva %016lx "
3890                                     "pte_pv %p pm_generation %d/%d",
3891                                     *ptep, oldpte, sva, pte_pv,
3892                                     generation, pmap->pm_generation));
3893                                 /*
3894                                  * We must unlock pd_pv across the callback
3895                                  * to avoid deadlocks on any recursive
3896                                  * disposal.  Re-check that it still exists
3897                                  * after re-locking.
3898                                  */
3899                                 if (pd_pv)
3900                                         pv_unlock(pd_pv);
3901                                 info->func(pmap, info, pte_pv, pt_pv, 0,
3902                                            sva, ptep, info->arg);
3903                                 if (pd_pv) {
3904                                         pv_lock(pd_pv);
3905                                         if (pd_pv->pv_pmap == NULL) {
3906                                                 pv_put(pd_pv);
3907                                                 pd_pv = NULL;
3908                                         }
3909                                 }
3910                         } else {
3911                                 /*
3912                                  * Check for insertion race.  Since there is no
3913                                  * pte_pv to guard us it is possible for us
3914                                  * to race another thread doing an insertion.
3915                                  * Our lookup misses the pte_pv but our *ptep
3916                                  * check sees the inserted pte.
3917                                  *
3918                                  * XXX panic case seems to occur within a
3919                                  * vm_fork() of /bin/sh, which frankly
3920                                  * shouldn't happen since no other threads
3921                                  * should be inserting to our pmap in that
3922                                  * situation.  Removing, possibly.  Inserting,
3923                                  * shouldn't happen.
3924                                  */
3925                                 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3926                                     pt_pv) {
3927                                         pte_pv = pv_find(pmap,
3928                                                          pmap_pte_pindex(sva));
3929                                         if (pte_pv) {
3930                                                 pv_drop(pte_pv);
3931                                                 kprintf("pmap_scan: RACE2 "
3932                                                         "%016jx, %016lx\n",
3933                                                         sva, oldpte);
3934                                                 continue;
3935                                         }
3936                                 }
3937
3938                                 /*
3939                                  * Didn't race
3940                                  *
3941                                  * We must unlock pd_pv across the callback
3942                                  * to avoid deadlocks on any recursive
3943                                  * disposal.  Re-check that it still exists
3944                                  * after re-locking.
3945                                  */
3946                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3947                                     pmap->pmap_bits[PG_V_IDX],
3948                                     ("badD *ptep %016lx/%016lx sva %016lx "
3949                                     "pte_pv NULL pm_generation %d/%d",
3950                                      *ptep, oldpte, sva,
3951                                      generation, pmap->pm_generation));
3952                                 if (pd_pv)
3953                                         pv_unlock(pd_pv);
3954                                 info->func(pmap, info, NULL, pt_pv, 0,
3955                                            sva, ptep, info->arg);
3956                                 if (pd_pv) {
3957                                         pv_lock(pd_pv);
3958                                         if (pd_pv->pv_pmap == NULL) {
3959                                                 pv_put(pd_pv);
3960                                                 pd_pv = NULL;
3961                                         }
3962                                 }
3963                         }
3964                         pte_pv = NULL;
3965                         sva += PAGE_SIZE;
3966                         ++ptep;
3967                 }
3968         }
3969         if (pd_pv) {
3970                 pv_put(pd_pv);
3971                 pd_pv = NULL;
3972         }
3973         if (pt_pv) {
3974                 pv_put(pt_pv);
3975                 pt_pv = NULL;
3976         }
3977         if ((++info->count & 7) == 0)
3978                 lwkt_user_yield();
3979
3980         /*
3981          * Relock before returning.
3982          */
3983         spin_lock(&pmap->pm_spin);
3984         return (0);
3985 }
3986
3987 void
3988 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3989 {
3990         struct pmap_scan_info info;
3991
3992         info.pmap = pmap;
3993         info.sva = sva;
3994         info.eva = eva;
3995         info.func = pmap_remove_callback;
3996         info.arg = NULL;
3997         pmap_scan(&info, 1);
3998 }
3999
4000 static void
4001 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4002 {
4003         struct pmap_scan_info info;
4004
4005         info.pmap = pmap;
4006         info.sva = sva;
4007         info.eva = eva;
4008         info.func = pmap_remove_callback;
4009         info.arg = NULL;
4010         pmap_scan(&info, 0);
4011 }
4012
4013 static void
4014 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4015                      pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4016                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4017 {
4018         pt_entry_t pte;
4019
4020         if (pte_pv) {
4021                 /*
4022                  * This will also drop pt_pv's wire_count. Note that
4023                  * terminal pages are not wired based on mmu presence.
4024                  *
4025                  * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4026                  */
4027                 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4028                 pte_pv = NULL;  /* safety */
4029
4030                 /*
4031                  * Recursively destroy higher-level page tables.
4032                  *
4033                  * This is optional.  If we do not, they will still
4034                  * be destroyed when the process exits.
4035                  */
4036                 if (pt_pv && pt_pv->pv_m && pt_pv->pv_m->wire_count == 1 &&
4037                     pt_pv->pv_pindex != pmap_pml4_pindex()) {
4038                         pv_hold(pt_pv);
4039                         pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4040                         pv_lock(pt_pv);
4041                 }
4042         } else if (sharept == 0) {
4043                 /*
4044                  * Unmanaged page table (pt, pd, or pdp. Not pte).
4045                  *
4046                  * pt_pv's wire_count is still bumped by unmanaged pages
4047                  * so we must decrement it manually.
4048                  *
4049                  * We have to unwire the target page table page.
4050                  *
4051                  * It is unclear how we can invalidate a segment so we
4052                  * invalidate -1 which invlidates the tlb.
4053                  */
4054                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4055                 if (pte & pmap->pmap_bits[PG_W_IDX])
4056                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4057                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4058                 if (vm_page_unwire_quick(pt_pv->pv_m))
4059                         panic("pmap_remove: insufficient wirecount");
4060         } else {
4061                 /*
4062                  * Unmanaged page table (pt, pd, or pdp. Not pte) for
4063                  * a shared page table.
4064                  *
4065                  * pt_pv is actually the pd_pv for our pmap (not the shared
4066                  * object pmap).
4067                  *
4068                  * We have to unwire the target page table page and we
4069                  * have to unwire our page directory page.
4070                  *
4071                  * It is unclear how we can invalidate a segment so we
4072                  * invalidate -1 which invlidates the tlb.
4073                  */
4074                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4075                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4076                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4077                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4078                         panic("pmap_remove: shared pgtable1 bad wirecount");
4079                 if (vm_page_unwire_quick(pt_pv->pv_m))
4080                         panic("pmap_remove: shared pgtable2 bad wirecount");
4081         }
4082 }
4083
4084 /*
4085  * Removes this physical page from all physical maps in which it resides.
4086  * Reflects back modify bits to the pager.
4087  *
4088  * This routine may not be called from an interrupt.
4089  */
4090 static
4091 void
4092 pmap_remove_all(vm_page_t m)
4093 {
4094         pv_entry_t pv;
4095         pmap_inval_bulk_t bulk;
4096
4097         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4098                 return;
4099
4100         vm_page_spin_lock(m);
4101         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4102                 KKASSERT(pv->pv_m == m);
4103                 if (pv_hold_try(pv)) {
4104                         vm_page_spin_unlock(m);
4105                 } else {
4106                         vm_page_spin_unlock(m);
4107                         pv_lock(pv);
4108                 }
4109                 if (pv->pv_m != m) {
4110                         pv_put(pv);
4111                         vm_page_spin_lock(m);
4112                         continue;
4113                 }
4114
4115                 /*
4116                  * Holding no spinlocks, pv is locked.
4117                  */
4118                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4119                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4120                 pv = NULL;      /* safety */
4121                 pmap_inval_bulk_flush(&bulk);
4122 #if 0
4123                 pmap_remove_pv_page(pv);
4124                 pv_free(pv, 1);
4125 #endif
4126                 vm_page_spin_lock(m);
4127         }
4128         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4129         vm_page_spin_unlock(m);
4130 }
4131
4132 /*
4133  * Removes the page from a particular pmap
4134  */
4135 void
4136 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4137 {
4138         pv_entry_t pv;
4139         pmap_inval_bulk_t bulk;
4140
4141         if (!pmap_initialized)
4142                 return;
4143
4144 again:
4145         vm_page_spin_lock(m);
4146         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4147                 if (pv->pv_pmap != pmap)
4148                         continue;
4149                 KKASSERT(pv->pv_m == m);
4150                 if (pv_hold_try(pv)) {
4151                         vm_page_spin_unlock(m);
4152                 } else {
4153                         vm_page_spin_unlock(m);
4154                         pv_lock(pv);
4155                 }
4156                 if (pv->pv_m != m) {
4157                         pv_put(pv);
4158                         goto again;
4159                 }
4160
4161                 /*
4162                  * Holding no spinlocks, pv is locked.
4163                  */
4164                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4165                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4166                 pv = NULL;      /* safety */
4167                 pmap_inval_bulk_flush(&bulk);
4168 #if 0
4169                 pmap_remove_pv_page(pv);
4170                 pv_free(pv, 1);
4171 #endif
4172                 goto again;
4173         }
4174         vm_page_spin_unlock(m);
4175 }
4176
4177 /*
4178  * Set the physical protection on the specified range of this map
4179  * as requested.  This function is typically only used for debug watchpoints
4180  * and COW pages.
4181  *
4182  * This function may not be called from an interrupt if the map is
4183  * not the kernel_pmap.
4184  *
4185  * NOTE!  For shared page table pages we just unmap the page.
4186  */
4187 void
4188 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4189 {
4190         struct pmap_scan_info info;
4191         /* JG review for NX */
4192
4193         if (pmap == NULL)
4194                 return;
4195         if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
4196                 pmap_remove(pmap, sva, eva);
4197                 return;
4198         }
4199         if (prot & VM_PROT_WRITE)
4200                 return;
4201         info.pmap = pmap;
4202         info.sva = sva;
4203         info.eva = eva;
4204         info.func = pmap_protect_callback;
4205         info.arg = &prot;
4206         pmap_scan(&info, 1);
4207 }
4208
4209 static
4210 void
4211 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4212                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4213                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4214 {
4215         pt_entry_t pbits;
4216         pt_entry_t cbits;
4217         pt_entry_t pte;
4218         vm_page_t m;
4219
4220 again:
4221         pbits = *ptep;
4222         cbits = pbits;
4223         if (pte_pv) {
4224                 m = NULL;
4225                 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4226                         if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4227                                 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4228                                 KKASSERT(m == pte_pv->pv_m);
4229                                 vm_page_flag_set(m, PG_REFERENCED);
4230                         }
4231                         cbits &= ~pmap->pmap_bits[PG_A_IDX];
4232                 }
4233                 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4234                         if (pmap_track_modified(pte_pv->pv_pindex)) {
4235                                 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4236                                         if (m == NULL) {
4237                                                 m = PHYS_TO_VM_PAGE(pbits &
4238                                                                     PG_FRAME);
4239                                         }
4240                                         vm_page_dirty(m);
4241                                 }
4242                                 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4243                         }
4244                 }
4245         } else if (sharept) {
4246                 /*
4247                  * Unmanaged page table, pt_pv is actually the pd_pv
4248                  * for our pmap (not the object's shared pmap).
4249                  *
4250                  * When asked to protect something in a shared page table
4251                  * page we just unmap the page table page.  We have to
4252                  * invalidate the tlb in this situation.
4253                  *
4254                  * XXX Warning, shared page tables will not be used for
4255                  * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4256                  * so PHYS_TO_VM_PAGE() should be safe here.
4257                  */
4258                 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4259                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4260                         panic("pmap_protect: pgtable1 pg bad wirecount");
4261                 if (vm_page_unwire_quick(pt_pv->pv_m))
4262                         panic("pmap_protect: pgtable2 pg bad wirecount");
4263                 ptep = NULL;
4264         }
4265         /* else unmanaged page, adjust bits, no wire changes */
4266
4267         if (ptep) {
4268                 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4269 #ifdef PMAP_DEBUG2
4270                 if (pmap_enter_debug > 0) {
4271                         --pmap_enter_debug;
4272                         kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4273                                 "pt_pv=%p cbits=%08lx\n",
4274                                 va, ptep, pte_pv,
4275                                 pt_pv, cbits
4276                         );
4277                 }
4278 #endif
4279                 if (pbits != cbits) {
4280                         if (!pmap_inval_smp_cmpset(pmap, (vm_offset_t)-1,
4281                                                    ptep, pbits, cbits)) {
4282                                 goto again;
4283                         }
4284                 }
4285         }
4286         if (pte_pv)
4287                 pv_put(pte_pv);
4288 }
4289
4290 /*
4291  * Insert the vm_page (m) at the virtual address (va), replacing any prior
4292  * mapping at that address.  Set protection and wiring as requested.
4293  *
4294  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4295  * possible.  If it is we enter the page into the appropriate shared pmap
4296  * hanging off the related VM object instead of the passed pmap, then we
4297  * share the page table page from the VM object's pmap into the current pmap.
4298  *
4299  * NOTE: This routine MUST insert the page into the pmap now, it cannot
4300  *       lazy-evaluate.
4301  */
4302 void
4303 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4304            boolean_t wired, vm_map_entry_t entry)
4305 {
4306         pv_entry_t pt_pv;       /* page table */
4307         pv_entry_t pte_pv;      /* page table entry */
4308         pt_entry_t *ptep;
4309         vm_paddr_t opa;
4310         pt_entry_t origpte, newpte;
4311         vm_paddr_t pa;
4312
4313         if (pmap == NULL)
4314                 return;
4315         va = trunc_page(va);
4316 #ifdef PMAP_DIAGNOSTIC
4317         if (va >= KvaEnd)
4318                 panic("pmap_enter: toobig");
4319         if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4320                 panic("pmap_enter: invalid to pmap_enter page table "
4321                       "pages (va: 0x%lx)", va);
4322 #endif
4323         if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4324                 kprintf("Warning: pmap_enter called on UVA with "
4325                         "kernel_pmap\n");
4326 #ifdef DDB
4327                 db_print_backtrace();
4328 #endif
4329         }
4330         if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4331                 kprintf("Warning: pmap_enter called on KVA without"
4332                         "kernel_pmap\n");
4333 #ifdef DDB
4334                 db_print_backtrace();
4335 #endif
4336         }
4337
4338         /*
4339          * Get locked PV entries for our new page table entry (pte_pv)
4340          * and for its parent page table (pt_pv).  We need the parent
4341          * so we can resolve the location of the ptep.
4342          *
4343          * Only hardware MMU actions can modify the ptep out from
4344          * under us.
4345          *
4346          * if (m) is fictitious or unmanaged we do not create a managing
4347          * pte_pv for it.  Any pre-existing page's management state must
4348          * match (avoiding code complexity).
4349          *
4350          * If the pmap is still being initialized we assume existing
4351          * page tables.
4352          *
4353          * Kernel mapppings do not track page table pages (i.e. pt_pv).
4354          */
4355         if (pmap_initialized == FALSE) {
4356                 pte_pv = NULL;
4357                 pt_pv = NULL;
4358                 ptep = vtopte(va);
4359                 origpte = *ptep;
4360         } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4361                 pte_pv = NULL;
4362                 if (va >= VM_MAX_USER_ADDRESS) {
4363                         pt_pv = NULL;
4364                         ptep = vtopte(va);
4365                 } else {
4366                         pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4367                                                   NULL, entry, va);
4368                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4369                 }
4370                 origpte = *ptep;
4371                 cpu_ccfence();
4372                 KASSERT(origpte == 0 ||
4373                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4374                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4375         } else {
4376                 if (va >= VM_MAX_USER_ADDRESS) {
4377                         /*
4378                          * Kernel map, pv_entry-tracked.
4379                          */
4380                         pt_pv = NULL;
4381                         pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4382                         ptep = vtopte(va);
4383                 } else {
4384                         /*
4385                          * User map
4386                          */
4387                         pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4388                                                    &pt_pv, entry, va);
4389                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4390                 }
4391                 origpte = *ptep;
4392                 cpu_ccfence();
4393                 KASSERT(origpte == 0 ||
4394                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4395                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4396         }
4397
4398         pa = VM_PAGE_TO_PHYS(m);
4399         opa = origpte & PG_FRAME;
4400
4401         newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4402                  pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4403         if (wired)
4404                 newpte |= pmap->pmap_bits[PG_W_IDX];
4405         if (va < VM_MAX_USER_ADDRESS)
4406                 newpte |= pmap->pmap_bits[PG_U_IDX];
4407         if (pte_pv)
4408                 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4409 //      if (pmap == &kernel_pmap)
4410 //              newpte |= pgeflag;
4411         newpte |= pmap->pmap_cache_bits[m->pat_mode];
4412         if (m->flags & PG_FICTITIOUS)
4413                 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4414
4415         /*
4416          * It is possible for multiple faults to occur in threaded
4417          * environments, the existing pte might be correct.
4418          */
4419         if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4420             pmap->pmap_bits[PG_A_IDX])) == 0)
4421                 goto done;
4422
4423         /*
4424          * Ok, either the address changed or the protection or wiring
4425          * changed.
4426          *
4427          * Clear the current entry, interlocking the removal.  For managed
4428          * pte's this will also flush the modified state to the vm_page.
4429          * Atomic ops are mandatory in order to ensure that PG_M events are
4430          * not lost during any transition.
4431          *
4432          * WARNING: The caller has busied the new page but not the original
4433          *          vm_page which we are trying to replace.  Because we hold
4434          *          the pte_pv lock, but have not busied the page, PG bits
4435          *          can be cleared out from under us.
4436          */
4437         if (opa) {
4438                 if (pte_pv) {
4439                         /*
4440                          * pt_pv won't exist for a kernel page (managed or
4441                          * otherwise).
4442                          */
4443                         if (prot & VM_PROT_NOSYNC) {
4444                                 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4445                         } else {
4446                                 pmap_inval_bulk_t bulk;
4447
4448                                 pmap_inval_bulk_init(&bulk, pmap);
4449                                 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4450                                 pmap_inval_bulk_flush(&bulk);
4451                         }
4452                         if (pte_pv->pv_m)
4453                                 pmap_remove_pv_page(pte_pv);
4454                 } else if (prot & VM_PROT_NOSYNC) {
4455                         /*
4456                          * Unmanaged page, NOSYNC (no mmu sync) requested.
4457                          *
4458                          * Leave wire count on PT page intact.
4459                          */
4460                         (void)pte_load_clear(ptep);
4461                         cpu_invlpg((void *)va);
4462                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
4463                 } else {
4464                         /*
4465                          * Unmanaged page, normal enter.
4466                          *
4467                          * Leave wire count on PT page intact.
4468                          */
4469                         pmap_inval_smp(pmap, va, 1, ptep, 0);
4470                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
4471                 }
4472                 KKASSERT(*ptep == 0);
4473         }
4474
4475 #ifdef PMAP_DEBUG2
4476         if (pmap_enter_debug > 0) {
4477                 --pmap_enter_debug;
4478                 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4479                         " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4480                         va, m,
4481                         origpte, newpte, ptep,
4482                         pte_pv, pt_pv, opa, prot);
4483         }
4484 #endif
4485
4486         if (pte_pv) {
4487                 /*
4488                  * Enter on the PV list if part of our managed memory.
4489                  * Wiring of the PT page is already handled.
4490                  */
4491                 KKASSERT(pte_pv->pv_m == NULL);
4492                 vm_page_spin_lock(m);
4493                 pte_pv->pv_m = m;
4494                 pmap_page_stats_adding(m);
4495                 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4496                 vm_page_flag_set(m, PG_MAPPED);
4497                 vm_page_spin_unlock(m);
4498         } else if (pt_pv && opa == 0) {
4499                 /*
4500                  * We have to adjust the wire count on the PT page ourselves
4501                  * for unmanaged entries.  If opa was non-zero we retained
4502                  * the existing wire count from the removal.
4503                  */
4504                 vm_page_wire_quick(pt_pv->pv_m);
4505         }
4506
4507         /*
4508          * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4509          *
4510          * User VMAs do not because those will be zero->non-zero, so no
4511          * stale entries to worry about at this point.
4512          *
4513          * For KVM there appear to still be issues.  Theoretically we
4514          * should be able to scrap the interlocks entirely but we
4515          * get crashes.
4516          */
4517         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4518                 pmap_inval_smp(pmap, va, 1, ptep, newpte);
4519         } else {
4520                 *(volatile pt_entry_t *)ptep = newpte;
4521                 if (pt_pv == NULL)
4522                         cpu_invlpg((void *)va);
4523         }
4524
4525         if (wired) {
4526                 if (pte_pv) {
4527                         atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4528                                         1);
4529                 } else {
4530                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
4531                 }
4532         }
4533         if (newpte & pmap->pmap_bits[PG_RW_IDX])
4534                 vm_page_flag_set(m, PG_WRITEABLE);
4535
4536         /*
4537          * Unmanaged pages need manual resident_count tracking.
4538          */
4539         if (pte_pv == NULL && pt_pv)
4540                 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4541
4542         /*
4543          * Cleanup
4544          */
4545 done:
4546         KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4547                  (m->flags & PG_MAPPED));
4548
4549         /*
4550          * Cleanup the pv entry, allowing other accessors.
4551          */
4552         if (pte_pv)
4553                 pv_put(pte_pv);
4554         if (pt_pv)
4555                 pv_put(pt_pv);
4556 }
4557
4558 /*
4559  * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4560  * This code also assumes that the pmap has no pre-existing entry for this
4561  * VA.
4562  *
4563  * This code currently may only be used on user pmaps, not kernel_pmap.
4564  */
4565 void
4566 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4567 {
4568         pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4569 }
4570
4571 /*
4572  * Make a temporary mapping for a physical address.  This is only intended
4573  * to be used for panic dumps.
4574  *
4575  * The caller is responsible for calling smp_invltlb().
4576  */
4577 void *
4578 pmap_kenter_temporary(vm_paddr_t pa, long i)
4579 {
4580         pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4581         return ((void *)crashdumpmap);
4582 }
4583
4584 #define MAX_INIT_PT (96)
4585
4586 /*
4587  * This routine preloads the ptes for a given object into the specified pmap.
4588  * This eliminates the blast of soft faults on process startup and
4589  * immediately after an mmap.
4590  */
4591 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4592
4593 void
4594 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4595                     vm_object_t object, vm_pindex_t pindex,
4596                     vm_size_t size, int limit)
4597 {
4598         struct rb_vm_page_scan_info info;
4599         struct lwp *lp;
4600         vm_size_t psize;
4601
4602         /*
4603          * We can't preinit if read access isn't set or there is no pmap
4604          * or object.
4605          */
4606         if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4607                 return;
4608
4609         /*
4610          * We can't preinit if the pmap is not the current pmap
4611          */
4612         lp = curthread->td_lwp;
4613         if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4614                 return;
4615
4616         /*
4617          * Misc additional checks
4618          */
4619         psize = x86_64_btop(size);
4620
4621         if ((object->type != OBJT_VNODE) ||
4622                 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4623                         (object->resident_page_count > MAX_INIT_PT))) {
4624                 return;
4625         }
4626
4627         if (pindex + psize > object->size) {
4628                 if (object->size < pindex)
4629                         return;
4630                 psize = object->size - pindex;
4631         }
4632
4633         if (psize == 0)
4634                 return;
4635
4636         /*
4637          * If everything is segment-aligned do not pre-init here.  Instead
4638          * allow the normal vm_fault path to pass a segment hint to
4639          * pmap_enter() which will then use an object-referenced shared
4640          * page table page.
4641          */
4642         if ((addr & SEG_MASK) == 0 &&
4643             (ctob(psize) & SEG_MASK) == 0 &&
4644             (ctob(pindex) & SEG_MASK) == 0) {
4645                 return;
4646         }
4647
4648         /*
4649          * Use a red-black scan to traverse the requested range and load
4650          * any valid pages found into the pmap.
4651          *
4652          * We cannot safely scan the object's memq without holding the
4653          * object token.
4654          */
4655         info.start_pindex = pindex;
4656         info.end_pindex = pindex + psize - 1;
4657         info.limit = limit;
4658         info.mpte = NULL;
4659         info.addr = addr;
4660         info.pmap = pmap;
4661
4662         vm_object_hold_shared(object);
4663         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4664                                 pmap_object_init_pt_callback, &info);
4665         vm_object_drop(object);
4666 }
4667
4668 static
4669 int
4670 pmap_object_init_pt_callback(vm_page_t p, void *data)
4671 {
4672         struct rb_vm_page_scan_info *info = data;
4673         vm_pindex_t rel_index;
4674
4675         /*
4676          * don't allow an madvise to blow away our really
4677          * free pages allocating pv entries.
4678          */
4679         if ((info->limit & MAP_PREFAULT_MADVISE) &&
4680                 vmstats.v_free_count < vmstats.v_free_reserved) {
4681                     return(-1);
4682         }
4683
4684         /*
4685          * Ignore list markers and ignore pages we cannot instantly
4686          * busy (while holding the object token).
4687          */
4688         if (p->flags & PG_MARKER)
4689                 return 0;
4690         if (vm_page_busy_try(p, TRUE))
4691                 return 0;
4692         if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4693             (p->flags & PG_FICTITIOUS) == 0) {
4694                 if ((p->queue - p->pc) == PQ_CACHE)
4695                         vm_page_deactivate(p);
4696                 rel_index = p->pindex - info->start_pindex;
4697                 pmap_enter_quick(info->pmap,
4698                                  info->addr + x86_64_ptob(rel_index), p);
4699         }
4700         vm_page_wakeup(p);
4701         lwkt_yield();
4702         return(0);
4703 }
4704
4705 /*
4706  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4707  * address.
4708  *
4709  * Returns FALSE if it would be non-trivial or if a pte is already loaded
4710  * into the slot.
4711  *
4712  * XXX This is safe only because page table pages are not freed.
4713  */
4714 int
4715 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4716 {
4717         pt_entry_t *pte;
4718
4719         /*spin_lock(&pmap->pm_spin);*/
4720         if ((pte = pmap_pte(pmap, addr)) != NULL) {
4721                 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4722                         /*spin_unlock(&pmap->pm_spin);*/
4723                         return FALSE;
4724                 }
4725         }
4726         /*spin_unlock(&pmap->pm_spin);*/
4727         return TRUE;
4728 }
4729
4730 /*
4731  * Change the wiring attribute for a pmap/va pair.  The mapping must already
4732  * exist in the pmap.  The mapping may or may not be managed.
4733  */
4734 void
4735 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
4736                    vm_map_entry_t entry)
4737 {
4738         pt_entry_t *ptep;
4739         pv_entry_t pv;
4740
4741         if (pmap == NULL)
4742                 return;
4743         lwkt_gettoken(&pmap->pm_token);
4744         pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL, entry, va);
4745         ptep = pv_pte_lookup(pv, pmap_pte_index(va));
4746
4747         if (wired && !pmap_pte_w(pmap, ptep))
4748                 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4749         else if (!wired && pmap_pte_w(pmap, ptep))
4750                 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4751
4752         /*
4753          * Wiring is not a hardware characteristic so there is no need to
4754          * invalidate TLB.  However, in an SMP environment we must use
4755          * a locked bus cycle to update the pte (if we are not using
4756          * the pmap_inval_*() API that is)... it's ok to do this for simple
4757          * wiring changes.
4758          */
4759         if (wired)
4760                 atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4761         else
4762                 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4763         pv_put(pv);
4764         lwkt_reltoken(&pmap->pm_token);
4765 }
4766
4767
4768
4769 /*
4770  * Copy the range specified by src_addr/len from the source map to
4771  * the range dst_addr/len in the destination map.
4772  *
4773  * This routine is only advisory and need not do anything.
4774  */
4775 void
4776 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4777           vm_size_t len, vm_offset_t src_addr)
4778 {
4779 }
4780
4781 /*
4782  * pmap_zero_page:
4783  *
4784  *      Zero the specified physical page.
4785  *
4786  *      This function may be called from an interrupt and no locking is
4787  *      required.
4788  */
4789 void
4790 pmap_zero_page(vm_paddr_t phys)
4791 {
4792         vm_offset_t va = PHYS_TO_DMAP(phys);
4793
4794         pagezero((void *)va);
4795 }
4796
4797 /*
4798  * pmap_zero_page:
4799  *
4800  *      Zero part of a physical page by mapping it into memory and clearing
4801  *      its contents with bzero.
4802  *
4803  *      off and size may not cover an area beyond a single hardware page.
4804  */
4805 void
4806 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4807 {
4808         vm_offset_t virt = PHYS_TO_DMAP(phys);
4809
4810         bzero((char *)virt + off, size);
4811 }
4812
4813 /*
4814  * pmap_copy_page:
4815  *
4816  *      Copy the physical page from the source PA to the target PA.
4817  *      This function may be called from an interrupt.  No locking
4818  *      is required.
4819  */
4820 void
4821 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4822 {
4823         vm_offset_t src_virt, dst_virt;
4824
4825         src_virt = PHYS_TO_DMAP(src);
4826         dst_virt = PHYS_TO_DMAP(dst);
4827         bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4828 }
4829
4830 /*
4831  * pmap_copy_page_frag:
4832  *
4833  *      Copy the physical page from the source PA to the target PA.
4834  *      This function may be called from an interrupt.  No locking
4835  *      is required.
4836  */
4837 void
4838 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4839 {
4840         vm_offset_t src_virt, dst_virt;
4841
4842         src_virt = PHYS_TO_DMAP(src);
4843         dst_virt = PHYS_TO_DMAP(dst);
4844
4845         bcopy((char *)src_virt + (src & PAGE_MASK),
4846               (char *)dst_virt + (dst & PAGE_MASK),
4847               bytes);
4848 }
4849
4850 /*
4851  * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4852  * this page.  This count may be changed upwards or downwards in the future;
4853  * it is only necessary that true be returned for a small subset of pmaps
4854  * for proper page aging.
4855  */
4856 boolean_t
4857 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4858 {
4859         pv_entry_t pv;
4860         int loops = 0;
4861
4862         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4863                 return FALSE;
4864
4865         vm_page_spin_lock(m);
4866         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4867                 if (pv->pv_pmap == pmap) {
4868                         vm_page_spin_unlock(m);
4869                         return TRUE;
4870                 }
4871                 loops++;
4872                 if (loops >= 16)
4873                         break;
4874         }
4875         vm_page_spin_unlock(m);
4876         return (FALSE);
4877 }
4878
4879 /*
4880  * Remove all pages from specified address space this aids process exit
4881  * speeds.  Also, this code may be special cased for the current process
4882  * only.
4883  */
4884 void
4885 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4886 {
4887         pmap_remove_noinval(pmap, sva, eva);
4888         cpu_invltlb();
4889 }
4890
4891 /*
4892  * pmap_testbit tests bits in pte's note that the testbit/clearbit
4893  * routines are inline, and a lot of things compile-time evaluate.
4894  */
4895 static
4896 boolean_t
4897 pmap_testbit(vm_page_t m, int bit)
4898 {
4899         pv_entry_t pv;
4900         pt_entry_t *pte;
4901         pmap_t pmap;
4902
4903         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4904                 return FALSE;
4905
4906         if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4907                 return FALSE;
4908         vm_page_spin_lock(m);
4909         if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4910                 vm_page_spin_unlock(m);
4911                 return FALSE;
4912         }
4913
4914         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4915
4916 #if defined(PMAP_DIAGNOSTIC)
4917                 if (pv->pv_pmap == NULL) {
4918                         kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4919                             pv->pv_pindex);
4920                         continue;
4921                 }
4922 #endif
4923                 pmap = pv->pv_pmap;
4924
4925                 /*
4926                  * If the bit being tested is the modified bit, then
4927                  * mark clean_map and ptes as never
4928                  * modified.
4929                  *
4930                  * WARNING!  Because we do not lock the pv, *pte can be in a
4931                  *           state of flux.  Despite this the value of *pte
4932                  *           will still be related to the vm_page in some way
4933                  *           because the pv cannot be destroyed as long as we
4934                  *           hold the vm_page spin lock.
4935                  */
4936                 if (bit == PG_A_IDX || bit == PG_M_IDX) {
4937                                 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
4938                         if (!pmap_track_modified(pv->pv_pindex))
4939                                 continue;
4940                 }
4941
4942                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4943                 if (*pte & pmap->pmap_bits[bit]) {
4944                         vm_page_spin_unlock(m);
4945                         return TRUE;
4946                 }
4947         }
4948         vm_page_spin_unlock(m);
4949         return (FALSE);
4950 }
4951
4952 /*
4953  * This routine is used to modify bits in ptes.  Only one bit should be
4954  * specified.  PG_RW requires special handling.
4955  *
4956  * Caller must NOT hold any spin locks
4957  */
4958 static __inline
4959 void
4960 pmap_clearbit(vm_page_t m, int bit_index)
4961 {
4962         pv_entry_t pv;
4963         pt_entry_t *pte;
4964         pt_entry_t pbits;
4965         pmap_t pmap;
4966
4967         if (bit_index == PG_RW_IDX)
4968                 vm_page_flag_clear(m, PG_WRITEABLE);
4969         if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
4970                 return;
4971         }
4972
4973         /*
4974          * PG_M or PG_A case
4975          *
4976          * Loop over all current mappings setting/clearing as appropos If
4977          * setting RO do we need to clear the VAC?
4978          *
4979          * NOTE: When clearing PG_M we could also (not implemented) drop
4980          *       through to the PG_RW code and clear PG_RW too, forcing
4981          *       a fault on write to redetect PG_M for virtual kernels, but
4982          *       it isn't necessary since virtual kernels invalidate the
4983          *       pte when they clear the VPTE_M bit in their virtual page
4984          *       tables.
4985          *
4986          * NOTE: Does not re-dirty the page when clearing only PG_M.
4987          *
4988          * NOTE: Because we do not lock the pv, *pte can be in a state of
4989          *       flux.  Despite this the value of *pte is still somewhat
4990          *       related while we hold the vm_page spin lock.
4991          *
4992          *       *pte can be zero due to this race.  Since we are clearing
4993          *       bits we basically do no harm when this race  ccurs.
4994          */
4995         if (bit_index != PG_RW_IDX) {
4996                 vm_page_spin_lock(m);
4997                 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4998 #if defined(PMAP_DIAGNOSTIC)
4999                         if (pv->pv_pmap == NULL) {
5000                                 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5001                                     pv->pv_pindex);
5002                                 continue;
5003                         }
5004 #endif
5005                         pmap = pv->pv_pmap;
5006                         pte = pmap_pte_quick(pv->pv_pmap,
5007                                              pv->pv_pindex << PAGE_SHIFT);
5008                         pbits = *pte;
5009                         if (pbits & pmap->pmap_bits[bit_index])
5010                                 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5011                 }
5012                 vm_page_spin_unlock(m);
5013                 return;
5014         }
5015
5016         /*
5017          * Clear PG_RW.  Also clears PG_M and marks the page dirty if PG_M
5018          * was set.
5019          */
5020 restart:
5021         vm_page_spin_lock(m);
5022         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5023                 /*
5024                  * don't write protect pager mappings
5025                  */
5026                 if (!pmap_track_modified(pv->pv_pindex))
5027                         continue;
5028
5029 #if defined(PMAP_DIAGNOSTIC)
5030                 if (pv->pv_pmap == NULL) {
5031                         kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5032                             pv->pv_pindex);
5033                         continue;
5034                 }
5035 #endif
5036                 pmap = pv->pv_pmap;
5037                 /*
5038                  * Skip pages which do not have PG_RW set.
5039                  */
5040                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5041                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5042                         continue;
5043
5044                 /*
5045                  * Lock the PV
5046                  */
5047                 if (pv_hold_try(pv)) {
5048                         vm_page_spin_unlock(m);
5049                 } else {
5050                         vm_page_spin_unlock(m);
5051                         pv_lock(pv);    /* held, now do a blocking lock */
5052                 }
5053                 if (pv->pv_pmap != pmap || pv->pv_m != m) {
5054                         pv_put(pv);     /* and release */
5055                         goto restart;   /* anything could have happened */
5056                 }
5057                 KKASSERT(pv->pv_pmap == pmap);
5058                 for (;;) {
5059                         pt_entry_t nbits;
5060
5061                         pbits = *pte;
5062                         cpu_ccfence();
5063                         nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5064                                           pmap->pmap_bits[PG_M_IDX]);
5065                         if (pmap_inval_smp_cmpset(pmap,
5066                                      ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5067                                      pte, pbits, nbits)) {
5068                                 break;
5069                         }
5070                         cpu_pause();
5071                 }
5072                 vm_page_spin_lock(m);
5073
5074                 /*
5075                  * If PG_M was found to be set while we were clearing PG_RW
5076                  * we also clear PG_M (done above) and mark the page dirty.
5077                  * Callers expect this behavior.
5078                  */
5079                 if (pbits & pmap->pmap_bits[PG_M_IDX])
5080                         vm_page_dirty(m);
5081                 pv_put(pv);
5082         }
5083         vm_page_spin_unlock(m);
5084 }
5085
5086 /*
5087  * Lower the permission for all mappings to a given page.
5088  *
5089  * Page must be busied by caller.  Because page is busied by caller this
5090  * should not be able to race a pmap_enter().
5091  */
5092 void
5093 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5094 {
5095         /* JG NX support? */
5096         if ((prot & VM_PROT_WRITE) == 0) {
5097                 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5098                         /*
5099                          * NOTE: pmap_clearbit(.. PG_RW) also clears
5100                          *       the PG_WRITEABLE flag in (m).
5101                          */
5102                         pmap_clearbit(m, PG_RW_IDX);
5103                 } else {
5104                         pmap_remove_all(m);
5105                 }
5106         }
5107 }
5108
5109 vm_paddr_t
5110 pmap_phys_address(vm_pindex_t ppn)
5111 {
5112         return (x86_64_ptob(ppn));
5113 }
5114
5115 /*
5116  * Return a count of reference bits for a page, clearing those bits.
5117  * It is not necessary for every reference bit to be cleared, but it
5118  * is necessary that 0 only be returned when there are truly no
5119  * reference bits set.
5120  *
5121  * XXX: The exact number of bits to check and clear is a matter that
5122  * should be tested and standardized at some point in the future for
5123  * optimal aging of shared pages.
5124  *
5125  * This routine may not block.
5126  */
5127 int
5128 pmap_ts_referenced(vm_page_t m)
5129 {
5130         pv_entry_t pv;
5131         pt_entry_t *pte;
5132         pmap_t pmap;
5133         int rtval = 0;
5134
5135         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5136                 return (rtval);
5137
5138         vm_page_spin_lock(m);
5139         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5140                 if (!pmap_track_modified(pv->pv_pindex))
5141                         continue;
5142                 pmap = pv->pv_pmap;
5143                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5144                 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5145                         atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5146                         rtval++;
5147                         if (rtval > 4)
5148                                 break;
5149                 }
5150         }
5151         vm_page_spin_unlock(m);
5152         return (rtval);
5153 }
5154
5155 /*
5156  *      pmap_is_modified:
5157  *
5158  *      Return whether or not the specified physical page was modified
5159  *      in any physical maps.
5160  */
5161 boolean_t
5162 pmap_is_modified(vm_page_t m)
5163 {
5164         boolean_t res;
5165
5166         res = pmap_testbit(m, PG_M_IDX);
5167         return (res);
5168 }
5169
5170 /*
5171  *      Clear the modify bits on the specified physical page.
5172  */
5173 void
5174 pmap_clear_modify(vm_page_t m)
5175 {
5176         pmap_clearbit(m, PG_M_IDX);
5177 }
5178
5179 /*
5180  *      pmap_clear_reference:
5181  *
5182  *      Clear the reference bit on the specified physical page.
5183  */
5184 void
5185 pmap_clear_reference(vm_page_t m)
5186 {
5187         pmap_clearbit(m, PG_A_IDX);
5188 }
5189
5190 /*
5191  * Miscellaneous support routines follow
5192  */
5193
5194 static
5195 void
5196 i386_protection_init(void)
5197 {
5198         int *kp, prot;
5199
5200         /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit  */
5201         kp = protection_codes;
5202         for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5203                 switch (prot) {
5204                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5205                         /*
5206                          * Read access is also 0. There isn't any execute bit,
5207                          * so just make it readable.
5208                          */
5209                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5210                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5211                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5212                         *kp++ = 0;
5213                         break;
5214                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5215                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5216                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5217                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5218                         *kp++ = pmap_bits_default[PG_RW_IDX];
5219                         break;
5220                 }
5221         }
5222 }
5223
5224 /*
5225  * Map a set of physical memory pages into the kernel virtual
5226  * address space. Return a pointer to where it is mapped. This
5227  * routine is intended to be used for mapping device memory,
5228  * NOT real memory.
5229  *
5230  * NOTE: We can't use pgeflag unless we invalidate the pages one at
5231  *       a time.
5232  *
5233  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5234  *       work whether the cpu supports PAT or not.  The remaining PAT
5235  *       attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5236  *       supports PAT.
5237  */
5238 void *
5239 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5240 {
5241         return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5242 }
5243
5244 void *
5245 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5246 {
5247         return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5248 }
5249
5250 void *
5251 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5252 {
5253         return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5254 }
5255
5256 /*
5257  * Map a set of physical memory pages into the kernel virtual
5258  * address space. Return a pointer to where it is mapped. This
5259  * routine is intended to be used for mapping device memory,
5260  * NOT real memory.
5261  */
5262 void *
5263 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5264 {
5265         vm_offset_t va, tmpva, offset;
5266         pt_entry_t *pte;
5267         vm_size_t tmpsize;
5268
5269         offset = pa & PAGE_MASK;
5270         size = roundup(offset + size, PAGE_SIZE);
5271
5272         va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5273         if (va == 0)
5274                 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5275
5276         pa = pa & ~PAGE_MASK;
5277         for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5278                 pte = vtopte(tmpva);
5279                 *pte = pa |
5280                     kernel_pmap.pmap_bits[PG_RW_IDX] |
5281                     kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5282                     kernel_pmap.pmap_cache_bits[mode];
5283                 tmpsize -= PAGE_SIZE;
5284                 tmpva += PAGE_SIZE;
5285                 pa += PAGE_SIZE;
5286         }
5287         pmap_invalidate_range(&kernel_pmap, va, va + size);
5288         pmap_invalidate_cache_range(va, va + size);
5289
5290         return ((void *)(va + offset));
5291 }
5292
5293 void
5294 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5295 {
5296         vm_offset_t base, offset;
5297
5298         base = va & ~PAGE_MASK;
5299         offset = va & PAGE_MASK;
5300         size = roundup(offset + size, PAGE_SIZE);
5301         pmap_qremove(va, size >> PAGE_SHIFT);
5302         kmem_free(&kernel_map, base, size);
5303 }
5304
5305 /*
5306  * Sets the memory attribute for the specified page.
5307  */
5308 void
5309 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5310 {
5311
5312     m->pat_mode = ma;
5313
5314     /*
5315      * If "m" is a normal page, update its direct mapping.  This update
5316      * can be relied upon to perform any cache operations that are
5317      * required for data coherence.
5318      */
5319     if ((m->flags & PG_FICTITIOUS) == 0)
5320         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5321 }
5322
5323 /*
5324  * Change the PAT attribute on an existing kernel memory map.  Caller
5325  * must ensure that the virtual memory in question is not accessed
5326  * during the adjustment.
5327  */
5328 void
5329 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5330 {
5331         pt_entry_t *pte;
5332         vm_offset_t base;
5333         int changed = 0;
5334
5335         if (va == 0)
5336                 panic("pmap_change_attr: va is NULL");
5337         base = trunc_page(va);
5338
5339         while (count) {
5340                 pte = vtopte(va);
5341                 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5342                        kernel_pmap.pmap_cache_bits[mode];
5343                 --count;
5344                 va += PAGE_SIZE;
5345         }
5346
5347         changed = 1;    /* XXX: not optimal */
5348
5349         /*
5350          * Flush CPU caches if required to make sure any data isn't cached that
5351          * shouldn't be, etc.
5352          */
5353         if (changed) {
5354                 pmap_invalidate_range(&kernel_pmap, base, va);
5355                 pmap_invalidate_cache_range(base, va);
5356         }
5357 }
5358
5359 /*
5360  * perform the pmap work for mincore
5361  */
5362 int
5363 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5364 {
5365         pt_entry_t *ptep, pte;
5366         vm_page_t m;
5367         int val = 0;
5368
5369         lwkt_gettoken(&pmap->pm_token);
5370         ptep = pmap_pte(pmap, addr);
5371
5372         if (ptep && (pte = *ptep) != 0) {
5373                 vm_offset_t pa;
5374
5375                 val = MINCORE_INCORE;
5376                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5377                         goto done;
5378
5379                 pa = pte & PG_FRAME;
5380
5381                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5382                         m = NULL;
5383                 else
5384                         m = PHYS_TO_VM_PAGE(pa);
5385
5386                 /*
5387                  * Modified by us
5388                  */
5389                 if (pte & pmap->pmap_bits[PG_M_IDX])
5390                         val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5391                 /*
5392                  * Modified by someone
5393                  */
5394                 else if (m && (m->dirty || pmap_is_modified(m)))
5395                         val |= MINCORE_MODIFIED_OTHER;
5396                 /*
5397                  * Referenced by us
5398                  */
5399                 if (pte & pmap->pmap_bits[PG_A_IDX])
5400                         val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5401
5402                 /*
5403                  * Referenced by someone
5404                  */
5405                 else if (m && ((m->flags & PG_REFERENCED) ||
5406                                 pmap_ts_referenced(m))) {
5407                         val |= MINCORE_REFERENCED_OTHER;
5408                         vm_page_flag_set(m, PG_REFERENCED);
5409                 }
5410         }
5411 done:
5412         lwkt_reltoken(&pmap->pm_token);
5413
5414         return val;
5415 }
5416
5417 /*
5418  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
5419  * vmspace will be ref'd and the old one will be deref'd.
5420  *
5421  * The vmspace for all lwps associated with the process will be adjusted
5422  * and cr3 will be reloaded if any lwp is the current lwp.
5423  *
5424  * The process must hold the vmspace->vm_map.token for oldvm and newvm
5425  */
5426 void
5427 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5428 {
5429         struct vmspace *oldvm;
5430         struct lwp *lp;
5431
5432         oldvm = p->p_vmspace;
5433         if (oldvm != newvm) {
5434                 if (adjrefs)
5435                         vmspace_ref(newvm);
5436                 p->p_vmspace = newvm;
5437                 KKASSERT(p->p_nthreads == 1);
5438                 lp = RB_ROOT(&p->p_lwp_tree);
5439                 pmap_setlwpvm(lp, newvm);
5440                 if (adjrefs)
5441                         vmspace_rel(oldvm);
5442         }
5443 }
5444
5445 /*
5446  * Set the vmspace for a LWP.  The vmspace is almost universally set the
5447  * same as the process vmspace, but virtual kernels need to swap out contexts
5448  * on a per-lwp basis.
5449  *
5450  * Caller does not necessarily hold any vmspace tokens.  Caller must control
5451  * the lwp (typically be in the context of the lwp).  We use a critical
5452  * section to protect against statclock and hardclock (statistics collection).
5453  */
5454 void
5455 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5456 {
5457         struct vmspace *oldvm;
5458         struct pmap *pmap;
5459
5460         oldvm = lp->lwp_vmspace;
5461
5462         if (oldvm != newvm) {
5463                 crit_enter();
5464                 lp->lwp_vmspace = newvm;
5465                 if (curthread->td_lwp == lp) {
5466                         pmap = vmspace_pmap(newvm);
5467                         ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5468                         if (pmap->pm_active_lock & CPULOCK_EXCL)
5469                                 pmap_interlock_wait(newvm);
5470 #if defined(SWTCH_OPTIM_STATS)
5471                         tlb_flush_count++;
5472 #endif
5473                         if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5474                                 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5475                         } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5476                                 curthread->td_pcb->pcb_cr3 = KPML4phys;
5477                         } else {
5478                                 panic("pmap_setlwpvm: unknown pmap type\n");
5479                         }
5480                         load_cr3(curthread->td_pcb->pcb_cr3);
5481                         pmap = vmspace_pmap(oldvm);
5482                         ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5483                                                mycpu->gd_cpuid);
5484                 }
5485                 crit_exit();
5486         }
5487 }
5488
5489 /*
5490  * Called when switching to a locked pmap, used to interlock against pmaps
5491  * undergoing modifications to prevent us from activating the MMU for the
5492  * target pmap until all such modifications have completed.  We have to do
5493  * this because the thread making the modifications has already set up its
5494  * SMP synchronization mask.
5495  *
5496  * This function cannot sleep!
5497  *
5498  * No requirements.
5499  */
5500 void
5501 pmap_interlock_wait(struct vmspace *vm)
5502 {
5503         struct pmap *pmap = &vm->vm_pmap;
5504
5505         if (pmap->pm_active_lock & CPULOCK_EXCL) {
5506                 crit_enter();
5507                 KKASSERT(curthread->td_critcount >= 2);
5508                 DEBUG_PUSH_INFO("pmap_interlock_wait");
5509                 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5510                         cpu_ccfence();
5511                         lwkt_process_ipiq();
5512                 }
5513                 DEBUG_POP_INFO();
5514                 crit_exit();
5515         }
5516 }
5517
5518 vm_offset_t
5519 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5520 {
5521
5522         if ((obj == NULL) || (size < NBPDR) ||
5523             ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5524                 return addr;
5525         }
5526
5527         addr = roundup2(addr, NBPDR);
5528         return addr;
5529 }
5530
5531 /*
5532  * Used by kmalloc/kfree, page already exists at va
5533  */
5534 vm_page_t
5535 pmap_kvtom(vm_offset_t va)
5536 {
5537         pt_entry_t *ptep = vtopte(va);
5538
5539         KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5540         return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5541 }
5542
5543 /*
5544  * Initialize machine-specific shared page directory support.  This
5545  * is executed when a VM object is created.
5546  */
5547 void
5548 pmap_object_init(vm_object_t object)
5549 {
5550         object->md.pmap_rw = NULL;
5551         object->md.pmap_ro = NULL;
5552 }
5553
5554 /*
5555  * Clean up machine-specific shared page directory support.  This
5556  * is executed when a VM object is destroyed.
5557  */
5558 void
5559 pmap_object_free(vm_object_t object)
5560 {
5561         pmap_t pmap;
5562
5563         if ((pmap = object->md.pmap_rw) != NULL) {
5564                 object->md.pmap_rw = NULL;
5565                 pmap_remove_noinval(pmap,
5566                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5567                 CPUMASK_ASSZERO(pmap->pm_active);
5568                 pmap_release(pmap);
5569                 pmap_puninit(pmap);
5570                 kfree(pmap, M_OBJPMAP);
5571         }
5572         if ((pmap = object->md.pmap_ro) != NULL) {
5573                 object->md.pmap_ro = NULL;
5574                 pmap_remove_noinval(pmap,
5575                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5576                 CPUMASK_ASSZERO(pmap->pm_active);
5577                 pmap_release(pmap);
5578                 pmap_puninit(pmap);
5579                 kfree(pmap, M_OBJPMAP);
5580         }
5581 }
5582
5583 /*
5584  * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
5585  * VM page and issue a pginfo->callback.
5586  *
5587  * We are expected to dispose of any non-NULL pte_pv.
5588  */
5589 static
5590 void
5591 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
5592                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
5593                       vm_offset_t va, pt_entry_t *ptep, void *arg)
5594 {
5595         struct pmap_pgscan_info *pginfo = arg;
5596         vm_page_t m;
5597
5598         if (pte_pv) {
5599                 /*
5600                  * Try to busy the page while we hold the pte_pv locked.
5601                  */
5602                 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
5603                 if (vm_page_busy_try(m, TRUE) == 0) {
5604                         if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
5605                                 /*
5606                                  * The callback is issued with the pte_pv
5607                                  * unlocked and put away, and the pt_pv
5608                                  * unlocked.
5609                                  */
5610                                 pv_put(pte_pv);
5611                                 if (pt_pv)
5612                                         pv_unlock(pt_pv);
5613                                 if (pginfo->callback(pginfo, va, m) < 0)
5614                                         info->stop = 1;
5615                                 if (pt_pv)
5616                                         pv_lock(pt_pv);
5617                         } else {
5618                                 pv_put(pte_pv);
5619                         }
5620                 } else {
5621                         ++pginfo->busycount;
5622                         pv_put(pte_pv);
5623                 }
5624         } else if (sharept) {
5625                 /* shared page table */
5626         } else {
5627                 /* else unmanaged page */
5628         }
5629 }
5630
5631 void
5632 pmap_pgscan(struct pmap_pgscan_info *pginfo)
5633 {
5634         struct pmap_scan_info info;
5635
5636         pginfo->offset = pginfo->beg_addr;
5637         info.pmap = pginfo->pmap;
5638         info.sva = pginfo->beg_addr;
5639         info.eva = pginfo->end_addr;
5640         info.func = pmap_pgscan_callback;
5641         info.arg = pginfo;
5642         pmap_scan(&info, 0);
5643         if (info.stop == 0)
5644                 pginfo->offset = pginfo->end_addr;
5645 }