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, pmarkp)    _pv_get(pmap, pindex, pmarkp    \
 114                                                         PMAP_DEBUG_ARGS)
 115 #define pv_lock(pv)                     _pv_lock(pv                     \
 116                                                         PMAP_DEBUG_ARGS)
 117 #define pv_hold_try(pv)                 _pv_hold_try(pv                 \
 118                                                         PMAP_DEBUG_ARGS)
 119 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp  \
 120                                                         PMAP_DEBUG_ARGS)
 121
 122 #define pv_free(pv, pvp)                _pv_free(pv, pvp PMAP_DEBUG_ARGS)
 123
 124 #else
 125
 126 #define PMAP_DEBUG_DECL
 127 #define PMAP_DEBUG_ARGS
 128 #define PMAP_DEBUG_COPY
 129
 130 #define pv_get(pmap, pindex, pmarkp)            _pv_get(pmap, pindex, pmarkp)
 131 #define pv_lock(pv)                     _pv_lock(pv)
 132 #define pv_hold_try(pv)                 _pv_hold_try(pv)
 133 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp)
 134 #define pv_free(pv, pvp)                _pv_free(pv, pvp)
 135
 136 #endif
 137
 138 /*
 139  * Get PDEs and PTEs for user/kernel address space
 140  */
 141 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
 142
 143 #define pmap_pde_v(pmap, pte)           ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 144 #define pmap_pte_w(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
 145 #define pmap_pte_m(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
 146 #define pmap_pte_u(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
 147 #define pmap_pte_v(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 148
 149 /*
 150  * Given a map and a machine independent protection code,
 151  * convert to a vax protection code.
 152  */
 153 #define pte_prot(m, p)          \
 154         (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
 155 static int protection_codes[PROTECTION_CODES_SIZE];
 156
 157 struct pmap kernel_pmap;
 158
 159 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
 160
 161 vm_paddr_t avail_start;         /* PA of first available physical page */
 162 vm_paddr_t avail_end;           /* PA of last available physical page */
 163 vm_offset_t virtual2_start;     /* cutout free area prior to kernel start */
 164 vm_offset_t virtual2_end;
 165 vm_offset_t virtual_start;      /* VA of first avail page (after kernel bss) */
 166 vm_offset_t virtual_end;        /* VA of last avail page (end of kernel AS) */
 167 vm_offset_t KvaStart;           /* VA start of KVA space */
 168 vm_offset_t KvaEnd;             /* VA end of KVA space (non-inclusive) */
 169 vm_offset_t KvaSize;            /* max size of kernel virtual address space */
 170 static boolean_t pmap_initialized = FALSE;      /* Has pmap_init completed? */
 171 //static int pgeflag;           /* PG_G or-in */
 172 //static int pseflag;           /* PG_PS or-in */
 173 uint64_t PatMsr;
 174
 175 static int ndmpdp;
 176 static vm_paddr_t dmaplimit;
 177 static int nkpt;
 178 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
 179
 180 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE];        /* PAT -> PG_ bits */
 181 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/    /* PAT -> PG_ bits */
 182
 183 static uint64_t KPTbase;
 184 static uint64_t KPTphys;
 185 static uint64_t KPDphys;        /* phys addr of kernel level 2 */
 186 static uint64_t KPDbase;        /* phys addr of kernel level 2 @ KERNBASE */
 187 uint64_t KPDPphys;      /* phys addr of kernel level 3 */
 188 uint64_t KPML4phys;     /* phys addr of kernel level 4 */
 189
 190 static uint64_t DMPDphys;       /* phys addr of direct mapped level 2 */
 191 static uint64_t DMPDPphys;      /* phys addr of direct mapped level 3 */
 192
 193 /*
 194  * Data for the pv entry allocation mechanism
 195  */
 196 static vm_zone_t pvzone;
 197 static struct vm_zone pvzone_store;
 198 static struct vm_object pvzone_obj;
 199 static int pv_entry_max=0, pv_entry_high_water=0;
 200 static int pmap_pagedaemon_waken = 0;
 201 static struct pv_entry *pvinit;
 202
 203 /*
 204  * All those kernel PT submaps that BSD is so fond of
 205  */
 206 pt_entry_t *CMAP1 = NULL, *ptmmap;
 207 caddr_t CADDR1 = NULL, ptvmmap = NULL;
 208 static pt_entry_t *msgbufmap;
 209 struct msgbuf *msgbufp=NULL;
 210
 211 /*
 212  * PMAP default PG_* bits. Needed to be able to add
 213  * EPT/NPT pagetable pmap_bits for the VMM module
 214  */
 215 uint64_t pmap_bits_default[] = {
 216                 REGULAR_PMAP,                   /* TYPE_IDX             0 */
 217                 X86_PG_V,                       /* PG_V_IDX             1 */
 218                 X86_PG_RW,                      /* PG_RW_IDX            2 */
 219                 X86_PG_U,                       /* PG_U_IDX             3 */
 220                 X86_PG_A,                       /* PG_A_IDX             4 */
 221                 X86_PG_M,                       /* PG_M_IDX             5 */
 222                 X86_PG_PS,                      /* PG_PS_IDX3           6 */
 223                 X86_PG_G,                       /* PG_G_IDX             7 */
 224                 X86_PG_AVAIL1,                  /* PG_AVAIL1_IDX        8 */
 225                 X86_PG_AVAIL2,                  /* PG_AVAIL2_IDX        9 */
 226                 X86_PG_AVAIL3,                  /* PG_AVAIL3_IDX        10 */
 227                 X86_PG_NC_PWT | X86_PG_NC_PCD,  /* PG_N_IDX             11 */
 228 };
 229 /*
 230  * Crashdump maps.
 231  */
 232 static pt_entry_t *pt_crashdumpmap;
 233 static caddr_t crashdumpmap;
 234
 235 static int pmap_debug = 0;
 236 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
 237     &pmap_debug, 0, "Debug pmap's");
 238 #ifdef PMAP_DEBUG2
 239 static int pmap_enter_debug = 0;
 240 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
 241     &pmap_enter_debug, 0, "Debug pmap_enter's");
 242 #endif
 243 static int pmap_yield_count = 64;
 244 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
 245     &pmap_yield_count, 0, "Yield during init_pt/release");
 246 static int pmap_mmu_optimize = 0;
 247 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
 248     &pmap_mmu_optimize, 0, "Share page table pages when possible");
 249 int pmap_fast_kernel_cpusync = 0;
 250 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
 251     &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
 252 int pmap_dynamic_delete = 0;
 253 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
 254     &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
 255
 256 #define DISABLE_PSE
 257
 258 /* Standard user access funtions */
 259 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
 260     size_t *lencopied);
 261 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
 262 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
 263 extern int std_fubyte (const uint8_t *base);
 264 extern int std_subyte (uint8_t *base, uint8_t byte);
 265 extern int32_t std_fuword32 (const uint32_t *base);
 266 extern int64_t std_fuword64 (const uint64_t *base);
 267 extern int std_suword64 (uint64_t *base, uint64_t word);
 268 extern int std_suword32 (uint32_t *base, int word);
 269 extern uint32_t std_swapu32 (volatile uint32_t *base, uint32_t v);
 270 extern uint64_t std_swapu64 (volatile uint64_t *base, uint64_t v);
 271
 272 static void pv_hold(pv_entry_t pv);
 273 static int _pv_hold_try(pv_entry_t pv
 274                                 PMAP_DEBUG_DECL);
 275 static void pv_drop(pv_entry_t pv);
 276 static void _pv_lock(pv_entry_t pv
 277                                 PMAP_DEBUG_DECL);
 278 static void pv_unlock(pv_entry_t pv);
 279 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
 280                                 PMAP_DEBUG_DECL);
 281 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp
 282                                 PMAP_DEBUG_DECL);
 283 static void _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL);
 284 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex,
 285                                 vm_pindex_t **pmarkp, int *errorp);
 286 static void pv_put(pv_entry_t pv);
 287 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
 288 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
 289                       pv_entry_t *pvpp);
 290 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
 291                       pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
 292 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
 293                         pmap_inval_bulk_t *bulk, int destroy);
 294 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
 295 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
 296                         pmap_inval_bulk_t *bulk);
 297
 298 struct pmap_scan_info;
 299 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
 300                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
 301                       pv_entry_t pt_pv, int sharept,
 302                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 303 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
 304                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
 305                       pv_entry_t pt_pv, int sharept,
 306                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 307
 308 static void i386_protection_init (void);
 309 static void create_pagetables(vm_paddr_t *firstaddr);
 310 static void pmap_remove_all (vm_page_t m);
 311 static boolean_t pmap_testbit (vm_page_t m, int bit);
 312
 313 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
 314 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
 315
 316 static void pmap_pinit_defaults(struct pmap *pmap);
 317 static void pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark);
 318 static void pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark);
 319
 320 static unsigned pdir4mb;
 321
 322 static int
 323 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
 324 {
 325         if (pv1->pv_pindex < pv2->pv_pindex)
 326                 return(-1);
 327         if (pv1->pv_pindex > pv2->pv_pindex)
 328                 return(1);
 329         return(0);
 330 }
 331
 332 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
 333              pv_entry_compare, vm_pindex_t, pv_pindex);
 334
 335 static __inline
 336 void
 337 pmap_page_stats_adding(vm_page_t m)
 338 {
 339         globaldata_t gd = mycpu;
 340
 341         if (TAILQ_EMPTY(&m->md.pv_list)) {
 342                 ++gd->gd_vmtotal.t_arm;
 343         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 344                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 345                 ++gd->gd_vmtotal.t_armshr;
 346                 ++gd->gd_vmtotal.t_avmshr;
 347         } else {
 348                 ++gd->gd_vmtotal.t_avmshr;
 349         }
 350 }
 351
 352 static __inline
 353 void
 354 pmap_page_stats_deleting(vm_page_t m)
 355 {
 356         globaldata_t gd = mycpu;
 357
 358         if (TAILQ_EMPTY(&m->md.pv_list)) {
 359                 --gd->gd_vmtotal.t_arm;
 360         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 361                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 362                 --gd->gd_vmtotal.t_armshr;
 363                 --gd->gd_vmtotal.t_avmshr;
 364         } else {
 365                 --gd->gd_vmtotal.t_avmshr;
 366         }
 367 }
 368
 369 /*
 370  * Move the kernel virtual free pointer to the next
 371  * 2MB.  This is used to help improve performance
 372  * by using a large (2MB) page for much of the kernel
 373  * (.text, .data, .bss)
 374  */
 375 static
 376 vm_offset_t
 377 pmap_kmem_choose(vm_offset_t addr)
 378 {
 379         vm_offset_t newaddr = addr;
 380
 381         newaddr = roundup2(addr, NBPDR);
 382         return newaddr;
 383 }
 384
 385 /*
 386  * pmap_pte_quick:
 387  *
 388  *      Super fast pmap_pte routine best used when scanning the pv lists.
 389  *      This eliminates many course-grained invltlb calls.  Note that many of
 390  *      the pv list scans are across different pmaps and it is very wasteful
 391  *      to do an entire invltlb when checking a single mapping.
 392  */
 393 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
 394
 395 static
 396 pt_entry_t *
 397 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
 398 {
 399         return pmap_pte(pmap, va);
 400 }
 401
 402 /*
 403  * Returns the pindex of a page table entry (representing a terminal page).
 404  * There are NUPTE_TOTAL page table entries possible (a huge number)
 405  *
 406  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
 407  * We want to properly translate negative KVAs.
 408  */
 409 static __inline
 410 vm_pindex_t
 411 pmap_pte_pindex(vm_offset_t va)
 412 {
 413         return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
 414 }
 415
 416 /*
 417  * Returns the pindex of a page table.
 418  */
 419 static __inline
 420 vm_pindex_t
 421 pmap_pt_pindex(vm_offset_t va)
 422 {
 423         return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
 424 }
 425
 426 /*
 427  * Returns the pindex of a page directory.
 428  */
 429 static __inline
 430 vm_pindex_t
 431 pmap_pd_pindex(vm_offset_t va)
 432 {
 433         return (NUPTE_TOTAL + NUPT_TOTAL +
 434                 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
 435 }
 436
 437 static __inline
 438 vm_pindex_t
 439 pmap_pdp_pindex(vm_offset_t va)
 440 {
 441         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
 442                 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
 443 }
 444
 445 static __inline
 446 vm_pindex_t
 447 pmap_pml4_pindex(void)
 448 {
 449         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
 450 }
 451
 452 /*
 453  * Return various clipped indexes for a given VA
 454  *
 455  * Returns the index of a pt in a page directory, representing a page
 456  * table.
 457  */
 458 static __inline
 459 vm_pindex_t
 460 pmap_pt_index(vm_offset_t va)
 461 {
 462         return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
 463 }
 464
 465 /*
 466  * Returns the index of a pd in a page directory page, representing a page
 467  * directory.
 468  */
 469 static __inline
 470 vm_pindex_t
 471 pmap_pd_index(vm_offset_t va)
 472 {
 473         return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
 474 }
 475
 476 /*
 477  * Returns the index of a pdp in the pml4 table, representing a page
 478  * directory page.
 479  */
 480 static __inline
 481 vm_pindex_t
 482 pmap_pdp_index(vm_offset_t va)
 483 {
 484         return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
 485 }
 486
 487 /*
 488  * The placemarker hash must be broken up into four zones so lock
 489  * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
 490  *
 491  * Placemarkers are used to 'lock' page table indices that do not have
 492  * a pv_entry.  This allows the pmap to support managed and unmanaged
 493  * pages and shared page tables.
 494  */
 495 #define PM_PLACE_BASE   (PM_PLACEMARKS >> 2)
 496
 497 static __inline
 498 vm_pindex_t *
 499 pmap_placemarker_hash(pmap_t pmap, vm_pindex_t pindex)
 500 {
 501         int hi;
 502
 503         if (pindex < pmap_pt_pindex(0))         /* zone 0 - PTE */
 504                 hi = 0;
 505         else if (pindex < pmap_pd_pindex(0))    /* zone 1 - PT */
 506                 hi = PM_PLACE_BASE;
 507         else if (pindex < pmap_pdp_pindex(0))   /* zone 2 - PD */
 508                 hi = PM_PLACE_BASE << 1;
 509         else                                    /* zone 3 - PDP (and PML4E) */
 510                 hi = PM_PLACE_BASE | (PM_PLACE_BASE << 1);
 511         hi += pindex & (PM_PLACE_BASE - 1);
 512
 513         return (&pmap->pm_placemarks[hi]);
 514 }
 515
 516
 517 /*
 518  * Generic procedure to index a pte from a pt, pd, or pdp.
 519  *
 520  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
 521  *       a page table page index but is instead of PV lookup index.
 522  */
 523 static
 524 void *
 525 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
 526 {
 527         pt_entry_t *pte;
 528
 529         pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
 530         return(&pte[pindex]);
 531 }
 532
 533 /*
 534  * Return pointer to PDP slot in the PML4
 535  */
 536 static __inline
 537 pml4_entry_t *
 538 pmap_pdp(pmap_t pmap, vm_offset_t va)
 539 {
 540         return (&pmap->pm_pml4[pmap_pdp_index(va)]);
 541 }
 542
 543 /*
 544  * Return pointer to PD slot in the PDP given a pointer to the PDP
 545  */
 546 static __inline
 547 pdp_entry_t *
 548 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
 549 {
 550         pdp_entry_t *pd;
 551
 552         pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
 553         return (&pd[pmap_pd_index(va)]);
 554 }
 555
 556 /*
 557  * Return pointer to PD slot in the PDP.
 558  */
 559 static __inline
 560 pdp_entry_t *
 561 pmap_pd(pmap_t pmap, vm_offset_t va)
 562 {
 563         pml4_entry_t *pdp;
 564
 565         pdp = pmap_pdp(pmap, va);
 566         if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
 567                 return NULL;
 568         return (pmap_pdp_to_pd(*pdp, va));
 569 }
 570
 571 /*
 572  * Return pointer to PT slot in the PD given a pointer to the PD
 573  */
 574 static __inline
 575 pd_entry_t *
 576 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
 577 {
 578         pd_entry_t *pt;
 579
 580         pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
 581         return (&pt[pmap_pt_index(va)]);
 582 }
 583
 584 /*
 585  * Return pointer to PT slot in the PD
 586  *
 587  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
 588  *                   so we cannot lookup the PD via the PDP.  Instead we
 589  *                   must look it up via the pmap.
 590  */
 591 static __inline
 592 pd_entry_t *
 593 pmap_pt(pmap_t pmap, vm_offset_t va)
 594 {
 595         pdp_entry_t *pd;
 596         pv_entry_t pv;
 597         vm_pindex_t pd_pindex;
 598
 599         if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
 600                 pd_pindex = pmap_pd_pindex(va);
 601                 spin_lock(&pmap->pm_spin);
 602                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
 603                 spin_unlock(&pmap->pm_spin);
 604                 if (pv == NULL || pv->pv_m == NULL)
 605                         return NULL;
 606                 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
 607         } else {
 608                 pd = pmap_pd(pmap, va);
 609                 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
 610                          return NULL;
 611                 return (pmap_pd_to_pt(*pd, va));
 612         }
 613 }
 614
 615 /*
 616  * Return pointer to PTE slot in the PT given a pointer to the PT
 617  */
 618 static __inline
 619 pt_entry_t *
 620 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
 621 {
 622         pt_entry_t *pte;
 623
 624         pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
 625         return (&pte[pmap_pte_index(va)]);
 626 }
 627
 628 /*
 629  * Return pointer to PTE slot in the PT
 630  */
 631 static __inline
 632 pt_entry_t *
 633 pmap_pte(pmap_t pmap, vm_offset_t va)
 634 {
 635         pd_entry_t *pt;
 636
 637         pt = pmap_pt(pmap, va);
 638         if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
 639                  return NULL;
 640         if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
 641                 return ((pt_entry_t *)pt);
 642         return (pmap_pt_to_pte(*pt, va));
 643 }
 644
 645 /*
 646  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
 647  * the PT layer.  This will speed up core pmap operations considerably.
 648  *
 649  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
 650  *       must be in a known associated state (typically by being locked when
 651  *       the pmap spinlock isn't held).  We allow the race for that case.
 652  *
 653  * NOTE: pm_pvhint is only accessed (read) with the spin-lock held, using
 654  *       cpu_ccfence() to prevent compiler optimizations from reloading the
 655  *       field.
 656  */
 657 static __inline
 658 void
 659 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
 660 {
 661         if (pindex >= pmap_pt_pindex(0) && pindex < pmap_pd_pindex(0)) {
 662                 if (pv->pv_pmap)
 663                         pv->pv_pmap->pm_pvhint = pv;
 664         }
 665 }
 666
 667
 668 /*
 669  * Return address of PT slot in PD (KVM only)
 670  *
 671  * Cannot be used for user page tables because it might interfere with
 672  * the shared page-table-page optimization (pmap_mmu_optimize).
 673  */
 674 static __inline
 675 pd_entry_t *
 676 vtopt(vm_offset_t va)
 677 {
 678         uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
 679                                   NPML4EPGSHIFT)) - 1);
 680
 681         return (PDmap + ((va >> PDRSHIFT) & mask));
 682 }
 683
 684 /*
 685  * KVM - return address of PTE slot in PT
 686  */
 687 static __inline
 688 pt_entry_t *
 689 vtopte(vm_offset_t va)
 690 {
 691         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 692                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 693
 694         return (PTmap + ((va >> PAGE_SHIFT) & mask));
 695 }
 696
 697 static uint64_t
 698 allocpages(vm_paddr_t *firstaddr, long n)
 699 {
 700         uint64_t ret;
 701
 702         ret = *firstaddr;
 703         bzero((void *)ret, n * PAGE_SIZE);
 704         *firstaddr += n * PAGE_SIZE;
 705         return (ret);
 706 }
 707
 708 static
 709 void
 710 create_pagetables(vm_paddr_t *firstaddr)
 711 {
 712         long i;         /* must be 64 bits */
 713         long nkpt_base;
 714         long nkpt_phys;
 715         int j;
 716
 717         /*
 718          * We are running (mostly) V=P at this point
 719          *
 720          * Calculate NKPT - number of kernel page tables.  We have to
 721          * accomodoate prealloction of the vm_page_array, dump bitmap,
 722          * MSGBUF_SIZE, and other stuff.  Be generous.
 723          *
 724          * Maxmem is in pages.
 725          *
 726          * ndmpdp is the number of 1GB pages we wish to map.
 727          */
 728         ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
 729         if (ndmpdp < 4)         /* Minimum 4GB of dirmap */
 730                 ndmpdp = 4;
 731         KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
 732
 733         /*
 734          * Starting at the beginning of kvm (not KERNBASE).
 735          */
 736         nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
 737         nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
 738         nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
 739                        ndmpdp) + 511) / 512;
 740         nkpt_phys += 128;
 741
 742         /*
 743          * Starting at KERNBASE - map 2G worth of page table pages.
 744          * KERNBASE is offset -2G from the end of kvm.
 745          */
 746         nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
 747
 748         /*
 749          * Allocate pages
 750          */
 751         KPTbase = allocpages(firstaddr, nkpt_base);
 752         KPTphys = allocpages(firstaddr, nkpt_phys);
 753         KPML4phys = allocpages(firstaddr, 1);
 754         KPDPphys = allocpages(firstaddr, NKPML4E);
 755         KPDphys = allocpages(firstaddr, NKPDPE);
 756
 757         /*
 758          * Calculate the page directory base for KERNBASE,
 759          * that is where we start populating the page table pages.
 760          * Basically this is the end - 2.
 761          */
 762         KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
 763
 764         DMPDPphys = allocpages(firstaddr, NDMPML4E);
 765         if ((amd_feature & AMDID_PAGE1GB) == 0)
 766                 DMPDphys = allocpages(firstaddr, ndmpdp);
 767         dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
 768
 769         /*
 770          * Fill in the underlying page table pages for the area around
 771          * KERNBASE.  This remaps low physical memory to KERNBASE.
 772          *
 773          * Read-only from zero to physfree
 774          * XXX not fully used, underneath 2M pages
 775          */
 776         for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
 777                 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
 778                 ((pt_entry_t *)KPTbase)[i] |=
 779                     pmap_bits_default[PG_RW_IDX] |
 780                     pmap_bits_default[PG_V_IDX] |
 781                     pmap_bits_default[PG_G_IDX];
 782         }
 783
 784         /*
 785          * Now map the initial kernel page tables.  One block of page
 786          * tables is placed at the beginning of kernel virtual memory,
 787          * and another block is placed at KERNBASE to map the kernel binary,
 788          * data, bss, and initial pre-allocations.
 789          */
 790         for (i = 0; i < nkpt_base; i++) {
 791                 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
 792                 ((pd_entry_t *)KPDbase)[i] |=
 793                     pmap_bits_default[PG_RW_IDX] |
 794                     pmap_bits_default[PG_V_IDX];
 795         }
 796         for (i = 0; i < nkpt_phys; i++) {
 797                 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
 798                 ((pd_entry_t *)KPDphys)[i] |=
 799                     pmap_bits_default[PG_RW_IDX] |
 800                     pmap_bits_default[PG_V_IDX];
 801         }
 802
 803         /*
 804          * Map from zero to end of allocations using 2M pages as an
 805          * optimization.  This will bypass some of the KPTBase pages
 806          * above in the KERNBASE area.
 807          */
 808         for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
 809                 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
 810                 ((pd_entry_t *)KPDbase)[i] |=
 811                     pmap_bits_default[PG_RW_IDX] |
 812                     pmap_bits_default[PG_V_IDX] |
 813                     pmap_bits_default[PG_PS_IDX] |
 814                     pmap_bits_default[PG_G_IDX];
 815         }
 816
 817         /*
 818          * And connect up the PD to the PDP.  The kernel pmap is expected
 819          * to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
 820          */
 821         for (i = 0; i < NKPDPE; i++) {
 822                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
 823                                 KPDphys + (i << PAGE_SHIFT);
 824                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
 825                     pmap_bits_default[PG_RW_IDX] |
 826                     pmap_bits_default[PG_V_IDX] |
 827                     pmap_bits_default[PG_U_IDX];
 828         }
 829
 830         /*
 831          * Now set up the direct map space using either 2MB or 1GB pages
 832          * Preset PG_M and PG_A because demotion expects it.
 833          *
 834          * When filling in entries in the PD pages make sure any excess
 835          * entries are set to zero as we allocated enough PD pages
 836          */
 837         if ((amd_feature & AMDID_PAGE1GB) == 0) {
 838                 for (i = 0; i < NPDEPG * ndmpdp; i++) {
 839                         ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
 840                         ((pd_entry_t *)DMPDphys)[i] |=
 841                             pmap_bits_default[PG_RW_IDX] |
 842                             pmap_bits_default[PG_V_IDX] |
 843                             pmap_bits_default[PG_PS_IDX] |
 844                             pmap_bits_default[PG_G_IDX] |
 845                             pmap_bits_default[PG_M_IDX] |
 846                             pmap_bits_default[PG_A_IDX];
 847                 }
 848
 849                 /*
 850                  * And the direct map space's PDP
 851                  */
 852                 for (i = 0; i < ndmpdp; i++) {
 853                         ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
 854                                                         (i << PAGE_SHIFT);
 855                         ((pdp_entry_t *)DMPDPphys)[i] |=
 856                             pmap_bits_default[PG_RW_IDX] |
 857                             pmap_bits_default[PG_V_IDX] |
 858                             pmap_bits_default[PG_U_IDX];
 859                 }
 860         } else {
 861                 for (i = 0; i < ndmpdp; i++) {
 862                         ((pdp_entry_t *)DMPDPphys)[i] =
 863                                                 (vm_paddr_t)i << PDPSHIFT;
 864                         ((pdp_entry_t *)DMPDPphys)[i] |=
 865                             pmap_bits_default[PG_RW_IDX] |
 866                             pmap_bits_default[PG_V_IDX] |
 867                             pmap_bits_default[PG_PS_IDX] |
 868                             pmap_bits_default[PG_G_IDX] |
 869                             pmap_bits_default[PG_M_IDX] |
 870                             pmap_bits_default[PG_A_IDX];
 871                 }
 872         }
 873
 874         /* And recursively map PML4 to itself in order to get PTmap */
 875         ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
 876         ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
 877             pmap_bits_default[PG_RW_IDX] |
 878             pmap_bits_default[PG_V_IDX] |
 879             pmap_bits_default[PG_U_IDX];
 880
 881         /*
 882          * Connect the Direct Map slots up to the PML4
 883          */
 884         for (j = 0; j < NDMPML4E; ++j) {
 885                 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
 886                     (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
 887                     pmap_bits_default[PG_RW_IDX] |
 888                     pmap_bits_default[PG_V_IDX] |
 889                     pmap_bits_default[PG_U_IDX];
 890         }
 891
 892         /*
 893          * Connect the KVA slot up to the PML4
 894          */
 895         ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
 896         ((pdp_entry_t *)KPML4phys)[KPML4I] |=
 897             pmap_bits_default[PG_RW_IDX] |
 898             pmap_bits_default[PG_V_IDX] |
 899             pmap_bits_default[PG_U_IDX];
 900 }
 901
 902 /*
 903  *      Bootstrap the system enough to run with virtual memory.
 904  *
 905  *      On the i386 this is called after mapping has already been enabled
 906  *      and just syncs the pmap module with what has already been done.
 907  *      [We can't call it easily with mapping off since the kernel is not
 908  *      mapped with PA == VA, hence we would have to relocate every address
 909  *      from the linked base (virtual) address "KERNBASE" to the actual
 910  *      (physical) address starting relative to 0]
 911  */
 912 void
 913 pmap_bootstrap(vm_paddr_t *firstaddr)
 914 {
 915         vm_offset_t va;
 916         pt_entry_t *pte;
 917         int i;
 918
 919         KvaStart = VM_MIN_KERNEL_ADDRESS;
 920         KvaEnd = VM_MAX_KERNEL_ADDRESS;
 921         KvaSize = KvaEnd - KvaStart;
 922
 923         avail_start = *firstaddr;
 924
 925         /*
 926          * Create an initial set of page tables to run the kernel in.
 927          */
 928         create_pagetables(firstaddr);
 929
 930         virtual2_start = KvaStart;
 931         virtual2_end = PTOV_OFFSET;
 932
 933         virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
 934         virtual_start = pmap_kmem_choose(virtual_start);
 935
 936         virtual_end = VM_MAX_KERNEL_ADDRESS;
 937
 938         /* XXX do %cr0 as well */
 939         load_cr4(rcr4() | CR4_PGE | CR4_PSE);
 940         load_cr3(KPML4phys);
 941
 942         /*
 943          * Initialize protection array.
 944          */
 945         i386_protection_init();
 946
 947         /*
 948          * The kernel's pmap is statically allocated so we don't have to use
 949          * pmap_create, which is unlikely to work correctly at this part of
 950          * the boot sequence (XXX and which no longer exists).
 951          */
 952         kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
 953         kernel_pmap.pm_count = 1;
 954         CPUMASK_ASSALLONES(kernel_pmap.pm_active);
 955         RB_INIT(&kernel_pmap.pm_pvroot);
 956         spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
 957         for (i = 0; i < PM_PLACEMARKS; ++i)
 958                 kernel_pmap.pm_placemarks[i] = PM_NOPLACEMARK;
 959
 960         /*
 961          * Reserve some special page table entries/VA space for temporary
 962          * mapping of pages.
 963          */
 964 #define SYSMAP(c, p, v, n)      \
 965         v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
 966
 967         va = virtual_start;
 968         pte = vtopte(va);
 969
 970         /*
 971          * CMAP1/CMAP2 are used for zeroing and copying pages.
 972          */
 973         SYSMAP(caddr_t, CMAP1, CADDR1, 1)
 974
 975         /*
 976          * Crashdump maps.
 977          */
 978         SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
 979
 980         /*
 981          * ptvmmap is used for reading arbitrary physical pages via
 982          * /dev/mem.
 983          */
 984         SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
 985
 986         /*
 987          * msgbufp is used to map the system message buffer.
 988          * XXX msgbufmap is not used.
 989          */
 990         SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
 991                atop(round_page(MSGBUF_SIZE)))
 992
 993         virtual_start = va;
 994         virtual_start = pmap_kmem_choose(virtual_start);
 995
 996         *CMAP1 = 0;
 997
 998         /*
 999          * PG_G is terribly broken on SMP because we IPI invltlb's in some
1000          * cases rather then invl1pg.  Actually, I don't even know why it
1001          * works under UP because self-referential page table mappings
1002          */
1003 //      pgeflag = 0;
1004
1005 /*
1006  * Initialize the 4MB page size flag
1007  */
1008 //      pseflag = 0;
1009 /*
1010  * The 4MB page version of the initial
1011  * kernel page mapping.
1012  */
1013         pdir4mb = 0;
1014
1015 #if !defined(DISABLE_PSE)
1016         if (cpu_feature & CPUID_PSE) {
1017                 pt_entry_t ptditmp;
1018                 /*
1019                  * Note that we have enabled PSE mode
1020                  */
1021 //              pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
1022                 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
1023                 ptditmp &= ~(NBPDR - 1);
1024                 ptditmp |= pmap_bits_default[PG_V_IDX] |
1025                     pmap_bits_default[PG_RW_IDX] |
1026                     pmap_bits_default[PG_PS_IDX] |
1027                     pmap_bits_default[PG_U_IDX];
1028 //                  pgeflag;
1029                 pdir4mb = ptditmp;
1030         }
1031 #endif
1032         cpu_invltlb();
1033
1034         /* Initialize the PAT MSR */
1035         pmap_init_pat();
1036         pmap_pinit_defaults(&kernel_pmap);
1037
1038         TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1039                           &pmap_fast_kernel_cpusync);
1040
1041 }
1042
1043 /*
1044  * Setup the PAT MSR.
1045  */
1046 void
1047 pmap_init_pat(void)
1048 {
1049         uint64_t pat_msr;
1050         u_long cr0, cr4;
1051
1052         /*
1053          * Default values mapping PATi,PCD,PWT bits at system reset.
1054          * The default values effectively ignore the PATi bit by
1055          * repeating the encodings for 0-3 in 4-7, and map the PCD
1056          * and PWT bit combinations to the expected PAT types.
1057          */
1058         pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |        /* 000 */
1059                   PAT_VALUE(1, PAT_WRITE_THROUGH) |     /* 001 */
1060                   PAT_VALUE(2, PAT_UNCACHED) |          /* 010 */
1061                   PAT_VALUE(3, PAT_UNCACHEABLE) |       /* 011 */
1062                   PAT_VALUE(4, PAT_WRITE_BACK) |        /* 100 */
1063                   PAT_VALUE(5, PAT_WRITE_THROUGH) |     /* 101 */
1064                   PAT_VALUE(6, PAT_UNCACHED) |          /* 110 */
1065                   PAT_VALUE(7, PAT_UNCACHEABLE);        /* 111 */
1066         pat_pte_index[PAT_WRITE_BACK]   = 0;
1067         pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
1068         pat_pte_index[PAT_UNCACHED]     = X86_PG_NC_PCD;
1069         pat_pte_index[PAT_UNCACHEABLE]  = X86_PG_NC_PCD | X86_PG_NC_PWT;
1070         pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1071         pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1072
1073         if (cpu_feature & CPUID_PAT) {
1074                 /*
1075                  * If we support the PAT then set-up entries for
1076                  * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1077                  * 4 and 5.
1078                  */
1079                 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1080                           PAT_VALUE(4, PAT_WRITE_PROTECTED);
1081                 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1082                           PAT_VALUE(5, PAT_WRITE_COMBINING);
1083                 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1084                 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1085
1086                 /*
1087                  * Then enable the PAT
1088                  */
1089
1090                 /* Disable PGE. */
1091                 cr4 = rcr4();
1092                 load_cr4(cr4 & ~CR4_PGE);
1093
1094                 /* Disable caches (CD = 1, NW = 0). */
1095                 cr0 = rcr0();
1096                 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1097
1098                 /* Flushes caches and TLBs. */
1099                 wbinvd();
1100                 cpu_invltlb();
1101
1102                 /* Update PAT and index table. */
1103                 wrmsr(MSR_PAT, pat_msr);
1104
1105                 /* Flush caches and TLBs again. */
1106                 wbinvd();
1107                 cpu_invltlb();
1108
1109                 /* Restore caches and PGE. */
1110                 load_cr0(cr0);
1111                 load_cr4(cr4);
1112                 PatMsr = pat_msr;
1113         }
1114 }
1115
1116 /*
1117  * Set 4mb pdir for mp startup
1118  */
1119 void
1120 pmap_set_opt(void)
1121 {
1122         if (cpu_feature & CPUID_PSE) {
1123                 load_cr4(rcr4() | CR4_PSE);
1124                 if (pdir4mb && mycpu->gd_cpuid == 0) {  /* only on BSP */
1125                         cpu_invltlb();
1126                 }
1127         }
1128 }
1129
1130 /*
1131  *      Initialize the pmap module.
1132  *      Called by vm_init, to initialize any structures that the pmap
1133  *      system needs to map virtual memory.
1134  *      pmap_init has been enhanced to support in a fairly consistant
1135  *      way, discontiguous physical memory.
1136  */
1137 void
1138 pmap_init(void)
1139 {
1140         int i;
1141         int initial_pvs;
1142
1143         /*
1144          * Allocate memory for random pmap data structures.  Includes the
1145          * pv_head_table.
1146          */
1147
1148         for (i = 0; i < vm_page_array_size; i++) {
1149                 vm_page_t m;
1150
1151                 m = &vm_page_array[i];
1152                 TAILQ_INIT(&m->md.pv_list);
1153         }
1154
1155         /*
1156          * init the pv free list
1157          */
1158         initial_pvs = vm_page_array_size;
1159         if (initial_pvs < MINPV)
1160                 initial_pvs = MINPV;
1161         pvzone = &pvzone_store;
1162         pvinit = (void *)kmem_alloc(&kernel_map,
1163                                     initial_pvs * sizeof (struct pv_entry),
1164                                     VM_SUBSYS_PVENTRY);
1165         zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1166                   pvinit, initial_pvs);
1167
1168         /*
1169          * Now it is safe to enable pv_table recording.
1170          */
1171         pmap_initialized = TRUE;
1172 }
1173
1174 /*
1175  * Initialize the address space (zone) for the pv_entries.  Set a
1176  * high water mark so that the system can recover from excessive
1177  * numbers of pv entries.
1178  */
1179 void
1180 pmap_init2(void)
1181 {
1182         int shpgperproc = PMAP_SHPGPERPROC;
1183         int entry_max;
1184
1185         TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1186         pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1187         TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1188         pv_entry_high_water = 9 * (pv_entry_max / 10);
1189
1190         /*
1191          * Subtract out pages already installed in the zone (hack)
1192          */
1193         entry_max = pv_entry_max - vm_page_array_size;
1194         if (entry_max <= 0)
1195                 entry_max = 1;
1196
1197         zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT);
1198
1199         /*
1200          * Enable dynamic deletion of empty higher-level page table pages
1201          * by default only if system memory is < 8GB (use 7GB for slop).
1202          * This can save a little memory, but imposes significant
1203          * performance overhead for things like bulk builds, and for programs
1204          * which do a lot of memory mapping and memory unmapping.
1205          */
1206         if (pmap_dynamic_delete < 0) {
1207                 if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1208                         pmap_dynamic_delete = 1;
1209                 else
1210                         pmap_dynamic_delete = 0;
1211         }
1212 }
1213
1214 /*
1215  * Typically used to initialize a fictitious page by vm/device_pager.c
1216  */
1217 void
1218 pmap_page_init(struct vm_page *m)
1219 {
1220         vm_page_init(m);
1221         TAILQ_INIT(&m->md.pv_list);
1222 }
1223
1224 /***************************************************
1225  * Low level helper routines.....
1226  ***************************************************/
1227
1228 /*
1229  * this routine defines the region(s) of memory that should
1230  * not be tested for the modified bit.
1231  */
1232 static __inline
1233 int
1234 pmap_track_modified(vm_pindex_t pindex)
1235 {
1236         vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1237         if ((va < clean_sva) || (va >= clean_eva))
1238                 return 1;
1239         else
1240                 return 0;
1241 }
1242
1243 /*
1244  * Extract the physical page address associated with the map/VA pair.
1245  * The page must be wired for this to work reliably.
1246  */
1247 vm_paddr_t
1248 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
1249 {
1250         vm_paddr_t rtval;
1251         pv_entry_t pt_pv;
1252         pt_entry_t *ptep;
1253
1254         rtval = 0;
1255         if (va >= VM_MAX_USER_ADDRESS) {
1256                 /*
1257                  * Kernel page directories might be direct-mapped and
1258                  * there is typically no PV tracking of pte's
1259                  */
1260                 pd_entry_t *pt;
1261
1262                 pt = pmap_pt(pmap, va);
1263                 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1264                         if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1265                                 rtval = *pt & PG_PS_FRAME;
1266                                 rtval |= va & PDRMASK;
1267                         } else {
1268                                 ptep = pmap_pt_to_pte(*pt, va);
1269                                 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1270                                         rtval = *ptep & PG_FRAME;
1271                                         rtval |= va & PAGE_MASK;
1272                                 }
1273                         }
1274                 }
1275                 if (handlep)
1276                         *handlep = NULL;
1277         } else {
1278                 /*
1279                  * User pages currently do not direct-map the page directory
1280                  * and some pages might not used managed PVs.  But all PT's
1281                  * will have a PV.
1282                  */
1283                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1284                 if (pt_pv) {
1285                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1286                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1287                                 rtval = *ptep & PG_FRAME;
1288                                 rtval |= va & PAGE_MASK;
1289                         }
1290                         if (handlep)
1291                                 *handlep = pt_pv;       /* locked until done */
1292                         else
1293                                 pv_put (pt_pv);
1294                 } else if (handlep) {
1295                         *handlep = NULL;
1296                 }
1297         }
1298         return rtval;
1299 }
1300
1301 void
1302 pmap_extract_done(void *handle)
1303 {
1304         if (handle)
1305                 pv_put((pv_entry_t)handle);
1306 }
1307
1308 /*
1309  * Similar to extract but checks protections, SMP-friendly short-cut for
1310  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1311  * fall-through to the real fault code.  Does not work with HVM page
1312  * tables.
1313  *
1314  * The returned page, if not NULL, is held (and not busied).
1315  *
1316  * WARNING! THE RETURNED PAGE IS ONLY HELD AND NOT SUITABLE FOR READING
1317  *          OR WRITING AS-IS.
1318  */
1319 vm_page_t
1320 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot, int *busyp)
1321 {
1322         if (pmap &&
1323             va < VM_MAX_USER_ADDRESS &&
1324             (pmap->pm_flags & PMAP_HVM) == 0) {
1325                 pv_entry_t pt_pv;
1326                 pv_entry_t pte_pv;
1327                 pt_entry_t *ptep;
1328                 pt_entry_t req;
1329                 vm_page_t m;
1330                 int error;
1331
1332                 req = pmap->pmap_bits[PG_V_IDX] |
1333                       pmap->pmap_bits[PG_U_IDX];
1334                 if (prot & VM_PROT_WRITE)
1335                         req |= pmap->pmap_bits[PG_RW_IDX];
1336
1337                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1338                 if (pt_pv == NULL)
1339                         return (NULL);
1340                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1341                 if ((*ptep & req) != req) {
1342                         pv_put(pt_pv);
1343                         return (NULL);
1344                 }
1345                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), NULL, &error);
1346                 if (pte_pv && error == 0) {
1347                         m = pte_pv->pv_m;
1348                         if (prot & VM_PROT_WRITE) {
1349                                 /* interlocked by presence of pv_entry */
1350                                 vm_page_dirty(m);
1351                         }
1352                         if (busyp) {
1353                                 if (prot & VM_PROT_WRITE) {
1354                                         if (vm_page_busy_try(m, TRUE))
1355                                                 m = NULL;
1356                                         *busyp = 1;
1357                                 } else {
1358                                         vm_page_hold(m);
1359                                         *busyp = 0;
1360                                 }
1361                         } else {
1362                                 vm_page_hold(m);
1363                         }
1364                         pv_put(pte_pv);
1365                 } else if (pte_pv) {
1366                         pv_drop(pte_pv);
1367                         m = NULL;
1368                 } else {
1369                         /* error, since we didn't request a placemarker */
1370                         m = NULL;
1371                 }
1372                 pv_put(pt_pv);
1373                 return(m);
1374         } else {
1375                 return(NULL);
1376         }
1377 }
1378
1379 /*
1380  * Extract the physical page address associated kernel virtual address.
1381  */
1382 vm_paddr_t
1383 pmap_kextract(vm_offset_t va)
1384 {
1385         pd_entry_t pt;          /* pt entry in pd */
1386         vm_paddr_t pa;
1387
1388         if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1389                 pa = DMAP_TO_PHYS(va);
1390         } else {
1391                 pt = *vtopt(va);
1392                 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1393                         pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1394                 } else {
1395                         /*
1396                          * Beware of a concurrent promotion that changes the
1397                          * PDE at this point!  For example, vtopte() must not
1398                          * be used to access the PTE because it would use the
1399                          * new PDE.  It is, however, safe to use the old PDE
1400                          * because the page table page is preserved by the
1401                          * promotion.
1402                          */
1403                         pa = *pmap_pt_to_pte(pt, va);
1404                         pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1405                 }
1406         }
1407         return pa;
1408 }
1409
1410 /***************************************************
1411  * Low level mapping routines.....
1412  ***************************************************/
1413
1414 /*
1415  * Routine: pmap_kenter
1416  * Function:
1417  *      Add a wired page to the KVA
1418  *      NOTE! note that in order for the mapping to take effect -- you
1419  *      should do an invltlb after doing the pmap_kenter().
1420  */
1421 void
1422 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1423 {
1424         pt_entry_t *ptep;
1425         pt_entry_t npte;
1426
1427         npte = pa |
1428                kernel_pmap.pmap_bits[PG_RW_IDX] |
1429                kernel_pmap.pmap_bits[PG_V_IDX];
1430 //             pgeflag;
1431         ptep = vtopte(va);
1432 #if 1
1433         pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1434 #else
1435         /* FUTURE */
1436         if (*ptep)
1437                 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1438         else
1439                 *ptep = npte;
1440 #endif
1441 }
1442
1443 /*
1444  * Similar to pmap_kenter(), except we only invalidate the mapping on the
1445  * current CPU.  Returns 0 if the previous pte was 0, 1 if it wasn't
1446  * (caller can conditionalize calling smp_invltlb()).
1447  */
1448 int
1449 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1450 {
1451         pt_entry_t *ptep;
1452         pt_entry_t npte;
1453         int res;
1454
1455         npte = pa | kernel_pmap.pmap_bits[PG_RW_IDX] |
1456                     kernel_pmap.pmap_bits[PG_V_IDX];
1457         // npte |= pgeflag;
1458         ptep = vtopte(va);
1459 #if 1
1460         res = 1;
1461 #else
1462         /* FUTURE */
1463         res = (*ptep != 0);
1464 #endif
1465         atomic_swap_long(ptep, npte);
1466         cpu_invlpg((void *)va);
1467
1468         return res;
1469 }
1470
1471 /*
1472  * Enter addresses into the kernel pmap but don't bother
1473  * doing any tlb invalidations.  Caller will do a rollup
1474  * invalidation via pmap_rollup_inval().
1475  */
1476 int
1477 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1478 {
1479         pt_entry_t *ptep;
1480         pt_entry_t npte;
1481         int res;
1482
1483         npte = pa |
1484             kernel_pmap.pmap_bits[PG_RW_IDX] |
1485             kernel_pmap.pmap_bits[PG_V_IDX];
1486 //          pgeflag;
1487         ptep = vtopte(va);
1488 #if 1
1489         res = 1;
1490 #else
1491         /* FUTURE */
1492         res = (*ptep != 0);
1493 #endif
1494         atomic_swap_long(ptep, npte);
1495         cpu_invlpg((void *)va);
1496
1497         return res;
1498 }
1499
1500 /*
1501  * remove a page from the kernel pagetables
1502  */
1503 void
1504 pmap_kremove(vm_offset_t va)
1505 {
1506         pt_entry_t *ptep;
1507
1508         ptep = vtopte(va);
1509         pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1510 }
1511
1512 void
1513 pmap_kremove_quick(vm_offset_t va)
1514 {
1515         pt_entry_t *ptep;
1516
1517         ptep = vtopte(va);
1518         (void)pte_load_clear(ptep);
1519         cpu_invlpg((void *)va);
1520 }
1521
1522 /*
1523  * Remove addresses from the kernel pmap but don't bother
1524  * doing any tlb invalidations.  Caller will do a rollup
1525  * invalidation via pmap_rollup_inval().
1526  */
1527 void
1528 pmap_kremove_noinval(vm_offset_t va)
1529 {
1530         pt_entry_t *ptep;
1531
1532         ptep = vtopte(va);
1533         (void)pte_load_clear(ptep);
1534 }
1535
1536 /*
1537  * XXX these need to be recoded.  They are not used in any critical path.
1538  */
1539 void
1540 pmap_kmodify_rw(vm_offset_t va)
1541 {
1542         atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1543         cpu_invlpg((void *)va);
1544 }
1545
1546 /* NOT USED
1547 void
1548 pmap_kmodify_nc(vm_offset_t va)
1549 {
1550         atomic_set_long(vtopte(va), PG_N);
1551         cpu_invlpg((void *)va);
1552 }
1553 */
1554
1555 /*
1556  * Used to map a range of physical addresses into kernel virtual
1557  * address space during the low level boot, typically to map the
1558  * dump bitmap, message buffer, and vm_page_array.
1559  *
1560  * These mappings are typically made at some pointer after the end of the
1561  * kernel text+data.
1562  *
1563  * We could return PHYS_TO_DMAP(start) here and not allocate any
1564  * via (*virtp), but then kmem from userland and kernel dumps won't
1565  * have access to the related pointers.
1566  */
1567 vm_offset_t
1568 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1569 {
1570         vm_offset_t va;
1571         vm_offset_t va_start;
1572
1573         /*return PHYS_TO_DMAP(start);*/
1574
1575         va_start = *virtp;
1576         va = va_start;
1577
1578         while (start < end) {
1579                 pmap_kenter_quick(va, start);
1580                 va += PAGE_SIZE;
1581                 start += PAGE_SIZE;
1582         }
1583         *virtp = va;
1584         return va_start;
1585 }
1586
1587 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
1588
1589 /*
1590  * Remove the specified set of pages from the data and instruction caches.
1591  *
1592  * In contrast to pmap_invalidate_cache_range(), this function does not
1593  * rely on the CPU's self-snoop feature, because it is intended for use
1594  * when moving pages into a different cache domain.
1595  */
1596 void
1597 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1598 {
1599         vm_offset_t daddr, eva;
1600         int i;
1601
1602         if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1603             (cpu_feature & CPUID_CLFSH) == 0)
1604                 wbinvd();
1605         else {
1606                 cpu_mfence();
1607                 for (i = 0; i < count; i++) {
1608                         daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1609                         eva = daddr + PAGE_SIZE;
1610                         for (; daddr < eva; daddr += cpu_clflush_line_size)
1611                                 clflush(daddr);
1612                 }
1613                 cpu_mfence();
1614         }
1615 }
1616
1617 void
1618 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1619 {
1620         KASSERT((sva & PAGE_MASK) == 0,
1621             ("pmap_invalidate_cache_range: sva not page-aligned"));
1622         KASSERT((eva & PAGE_MASK) == 0,
1623             ("pmap_invalidate_cache_range: eva not page-aligned"));
1624
1625         if (cpu_feature & CPUID_SS) {
1626                 ; /* If "Self Snoop" is supported, do nothing. */
1627         } else {
1628                 /* Globally invalidate caches */
1629                 cpu_wbinvd_on_all_cpus();
1630         }
1631 }
1632
1633 /*
1634  * Invalidate the specified range of virtual memory on all cpus associated
1635  * with the pmap.
1636  */
1637 void
1638 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1639 {
1640         pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1641 }
1642
1643 /*
1644  * Add a list of wired pages to the kva.  This routine is used for temporary
1645  * kernel mappings such as those found in buffer cache buffer.  Page
1646  * modifications and accesses are not tracked or recorded.
1647  *
1648  * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1649  *       semantics as previous mappings may have been zerod without any
1650  *       invalidation.
1651  *
1652  * The page *must* be wired.
1653  */
1654 void
1655 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1656 {
1657         vm_offset_t end_va;
1658         vm_offset_t va;
1659
1660         end_va = beg_va + count * PAGE_SIZE;
1661
1662         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1663                 pt_entry_t pte;
1664                 pt_entry_t *ptep;
1665
1666                 ptep = vtopte(va);
1667                 pte = VM_PAGE_TO_PHYS(*m) |
1668                         kernel_pmap.pmap_bits[PG_RW_IDX] |
1669                         kernel_pmap.pmap_bits[PG_V_IDX] |
1670                         kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1671 //              pgeflag;
1672                 atomic_swap_long(ptep, pte);
1673                 m++;
1674         }
1675         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1676 }
1677
1678 /*
1679  * This routine jerks page mappings from the kernel -- it is meant only
1680  * for temporary mappings such as those found in buffer cache buffers.
1681  * No recording modified or access status occurs.
1682  *
1683  * MPSAFE, INTERRUPT SAFE (cluster callback)
1684  */
1685 void
1686 pmap_qremove(vm_offset_t beg_va, int count)
1687 {
1688         vm_offset_t end_va;
1689         vm_offset_t va;
1690
1691         end_va = beg_va + count * PAGE_SIZE;
1692
1693         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1694                 pt_entry_t *pte;
1695
1696                 pte = vtopte(va);
1697                 (void)pte_load_clear(pte);
1698                 cpu_invlpg((void *)va);
1699         }
1700         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1701 }
1702
1703 /*
1704  * This routine removes temporary kernel mappings, only invalidating them
1705  * on the current cpu.  It should only be used under carefully controlled
1706  * conditions.
1707  */
1708 void
1709 pmap_qremove_quick(vm_offset_t beg_va, int count)
1710 {
1711         vm_offset_t end_va;
1712         vm_offset_t va;
1713
1714         end_va = beg_va + count * PAGE_SIZE;
1715
1716         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1717                 pt_entry_t *pte;
1718
1719                 pte = vtopte(va);
1720                 (void)pte_load_clear(pte);
1721                 cpu_invlpg((void *)va);
1722         }
1723 }
1724
1725 /*
1726  * This routine removes temporary kernel mappings *without* invalidating
1727  * the TLB.  It can only be used on permanent kva reservations such as those
1728  * found in buffer cache buffers, under carefully controlled circumstances.
1729  *
1730  * NOTE: Repopulating these KVAs requires unconditional invalidation.
1731  *       (pmap_qenter() does unconditional invalidation).
1732  */
1733 void
1734 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1735 {
1736         vm_offset_t end_va;
1737         vm_offset_t va;
1738
1739         end_va = beg_va + count * PAGE_SIZE;
1740
1741         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1742                 pt_entry_t *pte;
1743
1744                 pte = vtopte(va);
1745                 (void)pte_load_clear(pte);
1746         }
1747 }
1748
1749 /*
1750  * Create a new thread and optionally associate it with a (new) process.
1751  * NOTE! the new thread's cpu may not equal the current cpu.
1752  */
1753 void
1754 pmap_init_thread(thread_t td)
1755 {
1756         /* enforce pcb placement & alignment */
1757         td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1758         td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1759         td->td_savefpu = &td->td_pcb->pcb_save;
1760         td->td_sp = (char *)td->td_pcb; /* no -16 */
1761 }
1762
1763 /*
1764  * This routine directly affects the fork perf for a process.
1765  */
1766 void
1767 pmap_init_proc(struct proc *p)
1768 {
1769 }
1770
1771 static void
1772 pmap_pinit_defaults(struct pmap *pmap)
1773 {
1774         bcopy(pmap_bits_default, pmap->pmap_bits,
1775               sizeof(pmap_bits_default));
1776         bcopy(protection_codes, pmap->protection_codes,
1777               sizeof(protection_codes));
1778         bcopy(pat_pte_index, pmap->pmap_cache_bits,
1779               sizeof(pat_pte_index));
1780         pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1781         pmap->copyinstr = std_copyinstr;
1782         pmap->copyin = std_copyin;
1783         pmap->copyout = std_copyout;
1784         pmap->fubyte = std_fubyte;
1785         pmap->subyte = std_subyte;
1786         pmap->fuword32 = std_fuword32;
1787         pmap->fuword64 = std_fuword64;
1788         pmap->suword32 = std_suword32;
1789         pmap->suword64 = std_suword64;
1790         pmap->swapu32 = std_swapu32;
1791         pmap->swapu64 = std_swapu64;
1792 }
1793 /*
1794  * Initialize pmap0/vmspace0.
1795  *
1796  * On architectures where the kernel pmap is not integrated into the user
1797  * process pmap, this pmap represents the process pmap, not the kernel pmap.
1798  * kernel_pmap should be used to directly access the kernel_pmap.
1799  */
1800 void
1801 pmap_pinit0(struct pmap *pmap)
1802 {
1803         int i;
1804
1805         pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1806         pmap->pm_count = 1;
1807         CPUMASK_ASSZERO(pmap->pm_active);
1808         pmap->pm_pvhint = NULL;
1809         RB_INIT(&pmap->pm_pvroot);
1810         spin_init(&pmap->pm_spin, "pmapinit0");
1811         for (i = 0; i < PM_PLACEMARKS; ++i)
1812                 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1813         bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1814         pmap_pinit_defaults(pmap);
1815 }
1816
1817 /*
1818  * Initialize a preallocated and zeroed pmap structure,
1819  * such as one in a vmspace structure.
1820  */
1821 static void
1822 pmap_pinit_simple(struct pmap *pmap)
1823 {
1824         int i;
1825
1826         /*
1827          * Misc initialization
1828          */
1829         pmap->pm_count = 1;
1830         CPUMASK_ASSZERO(pmap->pm_active);
1831         pmap->pm_pvhint = NULL;
1832         pmap->pm_flags = PMAP_FLAG_SIMPLE;
1833
1834         pmap_pinit_defaults(pmap);
1835
1836         /*
1837          * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1838          * for this).
1839          */
1840         if (pmap->pm_pmlpv == NULL) {
1841                 RB_INIT(&pmap->pm_pvroot);
1842                 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1843                 spin_init(&pmap->pm_spin, "pmapinitsimple");
1844                 for (i = 0; i < PM_PLACEMARKS; ++i)
1845                         pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1846         }
1847 }
1848
1849 void
1850 pmap_pinit(struct pmap *pmap)
1851 {
1852         pv_entry_t pv;
1853         int j;
1854
1855         if (pmap->pm_pmlpv) {
1856                 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1857                         pmap_puninit(pmap);
1858                 }
1859         }
1860
1861         pmap_pinit_simple(pmap);
1862         pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1863
1864         /*
1865          * No need to allocate page table space yet but we do need a valid
1866          * page directory table.
1867          */
1868         if (pmap->pm_pml4 == NULL) {
1869                 pmap->pm_pml4 =
1870                     (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1871                                                         PAGE_SIZE,
1872                                                         VM_SUBSYS_PML4);
1873         }
1874
1875         /*
1876          * Allocate the page directory page, which wires it even though
1877          * it isn't being entered into some higher level page table (it
1878          * being the highest level).  If one is already cached we don't
1879          * have to do anything.
1880          */
1881         if ((pv = pmap->pm_pmlpv) == NULL) {
1882                 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1883                 pmap->pm_pmlpv = pv;
1884                 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1885                             VM_PAGE_TO_PHYS(pv->pv_m));
1886                 pv_put(pv);
1887
1888                 /*
1889                  * Install DMAP and KMAP.
1890                  */
1891                 for (j = 0; j < NDMPML4E; ++j) {
1892                         pmap->pm_pml4[DMPML4I + j] =
1893                             (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1894                             pmap->pmap_bits[PG_RW_IDX] |
1895                             pmap->pmap_bits[PG_V_IDX] |
1896                             pmap->pmap_bits[PG_U_IDX];
1897                 }
1898                 pmap->pm_pml4[KPML4I] = KPDPphys |
1899                     pmap->pmap_bits[PG_RW_IDX] |
1900                     pmap->pmap_bits[PG_V_IDX] |
1901                     pmap->pmap_bits[PG_U_IDX];
1902
1903                 /*
1904                  * install self-referential address mapping entry
1905                  */
1906                 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1907                     pmap->pmap_bits[PG_V_IDX] |
1908                     pmap->pmap_bits[PG_RW_IDX] |
1909                     pmap->pmap_bits[PG_A_IDX] |
1910                     pmap->pmap_bits[PG_M_IDX];
1911         } else {
1912                 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1913                 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1914         }
1915         KKASSERT(pmap->pm_pml4[255] == 0);
1916         KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1917         KKASSERT(pv->pv_entry.rbe_left == NULL);
1918         KKASSERT(pv->pv_entry.rbe_right == NULL);
1919 }
1920
1921 /*
1922  * Clean up a pmap structure so it can be physically freed.  This routine
1923  * is called by the vmspace dtor function.  A great deal of pmap data is
1924  * left passively mapped to improve vmspace management so we have a bit
1925  * of cleanup work to do here.
1926  */
1927 void
1928 pmap_puninit(pmap_t pmap)
1929 {
1930         pv_entry_t pv;
1931         vm_page_t p;
1932
1933         KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1934         if ((pv = pmap->pm_pmlpv) != NULL) {
1935                 if (pv_hold_try(pv) == 0)
1936                         pv_lock(pv);
1937                 KKASSERT(pv == pmap->pm_pmlpv);
1938                 p = pmap_remove_pv_page(pv);
1939                 pv_free(pv, NULL);
1940                 pv = NULL;      /* safety */
1941                 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1942                 vm_page_busy_wait(p, FALSE, "pgpun");
1943                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1944                 vm_page_unwire(p, 0);
1945                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1946
1947                 /*
1948                  * XXX eventually clean out PML4 static entries and
1949                  * use vm_page_free_zero()
1950                  */
1951                 vm_page_free(p);
1952                 pmap->pm_pmlpv = NULL;
1953         }
1954         if (pmap->pm_pml4) {
1955                 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1956                 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1957                 pmap->pm_pml4 = NULL;
1958         }
1959         KKASSERT(pmap->pm_stats.resident_count == 0);
1960         KKASSERT(pmap->pm_stats.wired_count == 0);
1961 }
1962
1963 /*
1964  * This function is now unused (used to add the pmap to the pmap_list)
1965  */
1966 void
1967 pmap_pinit2(struct pmap *pmap)
1968 {
1969 }
1970
1971 /*
1972  * This routine is called when various levels in the page table need to
1973  * be populated.  This routine cannot fail.
1974  *
1975  * This function returns two locked pv_entry's, one representing the
1976  * requested pv and one representing the requested pv's parent pv.  If
1977  * an intermediate page table does not exist it will be created, mapped,
1978  * wired, and the parent page table will be given an additional hold
1979  * count representing the presence of the child pv_entry.
1980  */
1981 static
1982 pv_entry_t
1983 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1984 {
1985         pt_entry_t *ptep;
1986         pv_entry_t pv;
1987         pv_entry_t pvp;
1988         pt_entry_t v;
1989         vm_pindex_t pt_pindex;
1990         vm_page_t m;
1991         int isnew;
1992         int ispt;
1993
1994         /*
1995          * If the pv already exists and we aren't being asked for the
1996          * parent page table page we can just return it.  A locked+held pv
1997          * is returned.  The pv will also have a second hold related to the
1998          * pmap association that we don't have to worry about.
1999          */
2000         ispt = 0;
2001         pv = pv_alloc(pmap, ptepindex, &isnew);
2002         if (isnew == 0 && pvpp == NULL)
2003                 return(pv);
2004
2005         /*
2006          * Special case terminal PVs.  These are not page table pages so
2007          * no vm_page is allocated (the caller supplied the vm_page).  If
2008          * pvpp is non-NULL we are being asked to also removed the pt_pv
2009          * for this pv.
2010          *
2011          * Note that pt_pv's are only returned for user VAs. We assert that
2012          * a pt_pv is not being requested for kernel VAs.  The kernel
2013          * pre-wires all higher-level page tables so don't overload managed
2014          * higher-level page tables on top of it!
2015          */
2016         if (ptepindex < pmap_pt_pindex(0)) {
2017                 if (ptepindex >= NUPTE_USER) {
2018                         /* kernel manages this manually for KVM */
2019                         KKASSERT(pvpp == NULL);
2020                 } else {
2021                         KKASSERT(pvpp != NULL);
2022                         pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
2023                         pvp = pmap_allocpte(pmap, pt_pindex, NULL);
2024                         if (isnew)
2025                                 vm_page_wire_quick(pvp->pv_m);
2026                         *pvpp = pvp;
2027                 }
2028                 return(pv);
2029         }
2030
2031         /*
2032          * The kernel never uses managed PT/PD/PDP pages.
2033          */
2034         KKASSERT(pmap != &kernel_pmap);
2035
2036         /*
2037          * Non-terminal PVs allocate a VM page to represent the page table,
2038          * so we have to resolve pvp and calculate ptepindex for the pvp
2039          * and then for the page table entry index in the pvp for
2040          * fall-through.
2041          */
2042         if (ptepindex < pmap_pd_pindex(0)) {
2043                 /*
2044                  * pv is PT, pvp is PD
2045                  */
2046                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
2047                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
2048                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2049
2050                 /*
2051                  * PT index in PD
2052                  */
2053                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
2054                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
2055                 ispt = 1;
2056         } else if (ptepindex < pmap_pdp_pindex(0)) {
2057                 /*
2058                  * pv is PD, pvp is PDP
2059                  *
2060                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2061                  *                   the PD.
2062                  */
2063                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
2064                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2065
2066                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
2067                         KKASSERT(pvpp == NULL);
2068                         pvp = NULL;
2069                 } else {
2070                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
2071                 }
2072
2073                 /*
2074                  * PD index in PDP
2075                  */
2076                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
2077                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
2078         } else if (ptepindex < pmap_pml4_pindex()) {
2079                 /*
2080                  * pv is PDP, pvp is the root pml4 table
2081                  */
2082                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2083
2084                 /*
2085                  * PDP index in PML4
2086                  */
2087                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2088                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2089         } else {
2090                 /*
2091                  * pv represents the top-level PML4, there is no parent.
2092                  */
2093                 pvp = NULL;
2094         }
2095
2096         if (isnew == 0)
2097                 goto notnew;
2098
2099         /*
2100          * (isnew) is TRUE, pv is not terminal.
2101          *
2102          * (1) Add a wire count to the parent page table (pvp).
2103          * (2) Allocate a VM page for the page table.
2104          * (3) Enter the VM page into the parent page table.
2105          *
2106          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2107          */
2108         if (pvp)
2109                 vm_page_wire_quick(pvp->pv_m);
2110
2111         for (;;) {
2112                 m = vm_page_alloc(NULL, pv->pv_pindex,
2113                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2114                                   VM_ALLOC_INTERRUPT);
2115                 if (m)
2116                         break;
2117                 vm_wait(0);
2118         }
2119         vm_page_wire(m);        /* wire for mapping in parent */
2120         vm_page_unmanage(m);    /* m must be spinunlocked */
2121         pmap_zero_page(VM_PAGE_TO_PHYS(m));
2122         m->valid = VM_PAGE_BITS_ALL;
2123
2124         vm_page_spin_lock(m);
2125         pmap_page_stats_adding(m);
2126         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2127         pv->pv_m = m;
2128         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2129         vm_page_spin_unlock(m);
2130
2131         /*
2132          * (isnew) is TRUE, pv is not terminal.
2133          *
2134          * Wire the page into pvp.  Bump the resident_count for the pmap.
2135          * There is no pvp for the top level, address the pm_pml4[] array
2136          * directly.
2137          *
2138          * If the caller wants the parent we return it, otherwise
2139          * we just put it away.
2140          *
2141          * No interlock is needed for pte 0 -> non-zero.
2142          *
2143          * In the situation where *ptep is valid we might have an unmanaged
2144          * page table page shared from another page table which we need to
2145          * unshare before installing our private page table page.
2146          */
2147         if (pvp) {
2148                 v = VM_PAGE_TO_PHYS(m) |
2149                     (pmap->pmap_bits[PG_U_IDX] |
2150                      pmap->pmap_bits[PG_RW_IDX] |
2151                      pmap->pmap_bits[PG_V_IDX] |
2152                      pmap->pmap_bits[PG_A_IDX] |
2153                      pmap->pmap_bits[PG_M_IDX]);
2154                 ptep = pv_pte_lookup(pvp, ptepindex);
2155                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2156                         pt_entry_t pte;
2157
2158                         if (ispt == 0) {
2159                                 panic("pmap_allocpte: unexpected pte %p/%d",
2160                                       pvp, (int)ptepindex);
2161                         }
2162                         pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, v);
2163                         if (vm_page_unwire_quick(
2164                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2165                                 panic("pmap_allocpte: shared pgtable "
2166                                       "pg bad wirecount");
2167                         }
2168                 } else {
2169                         pt_entry_t pte;
2170
2171                         pte = atomic_swap_long(ptep, v);
2172                         if (pte != 0) {
2173                                 kprintf("install pgtbl mixup 0x%016jx "
2174                                         "old/new 0x%016jx/0x%016jx\n",
2175                                         (intmax_t)ptepindex, pte, v);
2176                         }
2177                 }
2178         }
2179         vm_page_wakeup(m);
2180
2181         /*
2182          * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2183          */
2184 notnew:
2185         if (pvp) {
2186                 KKASSERT(pvp->pv_m != NULL);
2187                 ptep = pv_pte_lookup(pvp, ptepindex);
2188                 v = VM_PAGE_TO_PHYS(pv->pv_m) |
2189                     (pmap->pmap_bits[PG_U_IDX] |
2190                      pmap->pmap_bits[PG_RW_IDX] |
2191                      pmap->pmap_bits[PG_V_IDX] |
2192                      pmap->pmap_bits[PG_A_IDX] |
2193                      pmap->pmap_bits[PG_M_IDX]);
2194                 if (*ptep != v) {
2195                         kprintf("mismatched upper level pt %016jx/%016jx\n",
2196                                 *ptep, v);
2197                 }
2198         }
2199         if (pvpp)
2200                 *pvpp = pvp;
2201         else if (pvp)
2202                 pv_put(pvp);
2203         return (pv);
2204 }
2205
2206 /*
2207  * This version of pmap_allocpte() checks for possible segment optimizations
2208  * that would allow page-table sharing.  It can be called for terminal
2209  * page or page table page ptepindex's.
2210  *
2211  * The function is called with page table page ptepindex's for fictitious
2212  * and unmanaged terminal pages.  That is, we don't want to allocate a
2213  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
2214  * for this case.
2215  *
2216  * This function can return a pv and *pvpp associated with the passed in pmap
2217  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
2218  * an unmanaged page table page will be entered into the pass in pmap.
2219  */
2220 static
2221 pv_entry_t
2222 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2223                   vm_map_entry_t entry, vm_offset_t va)
2224 {
2225         vm_object_t object;
2226         pmap_t obpmap;
2227         pmap_t *obpmapp;
2228         vm_offset_t b;
2229         pv_entry_t pte_pv;      /* in original or shared pmap */
2230         pv_entry_t pt_pv;       /* in original or shared pmap */
2231         pv_entry_t proc_pd_pv;  /* in original pmap */
2232         pv_entry_t proc_pt_pv;  /* in original pmap */
2233         pv_entry_t xpv;         /* PT in shared pmap */
2234         pd_entry_t *pt;         /* PT entry in PD of original pmap */
2235         pd_entry_t opte;        /* contents of *pt */
2236         pd_entry_t npte;        /* contents of *pt */
2237         vm_page_t m;
2238
2239         /*
2240          * Basic tests, require a non-NULL vm_map_entry, require proper
2241          * alignment and type for the vm_map_entry, require that the
2242          * underlying object already be allocated.
2243          *
2244          * We allow almost any type of object to use this optimization.
2245          * The object itself does NOT have to be sized to a multiple of the
2246          * segment size, but the memory mapping does.
2247          *
2248          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2249          *     won't work as expected.
2250          */
2251         if (entry == NULL ||
2252             pmap_mmu_optimize == 0 ||                   /* not enabled */
2253             (pmap->pm_flags & PMAP_HVM) ||              /* special pmap */
2254             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal or pt */
2255             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2256             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2257             entry->object.vm_object == NULL ||          /* needs VM object */
2258             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2259             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2260             (entry->offset & SEG_MASK) ||               /* must be aligned */
2261             (entry->start & SEG_MASK)) {
2262                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2263         }
2264
2265         /*
2266          * Make sure the full segment can be represented.
2267          */
2268         b = va & ~(vm_offset_t)SEG_MASK;
2269         if (b < entry->start || b + SEG_SIZE > entry->end)
2270                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2271
2272         /*
2273          * If the full segment can be represented dive the VM object's
2274          * shared pmap, allocating as required.
2275          */
2276         object = entry->object.vm_object;
2277
2278         if (entry->protection & VM_PROT_WRITE)
2279                 obpmapp = &object->md.pmap_rw;
2280         else
2281                 obpmapp = &object->md.pmap_ro;
2282
2283 #ifdef PMAP_DEBUG2
2284         if (pmap_enter_debug > 0) {
2285                 --pmap_enter_debug;
2286                 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2287                         "obpmapp %p %p\n",
2288                         va, entry->protection, object,
2289                         obpmapp, *obpmapp);
2290                 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2291                         entry, entry->start, entry->end);
2292         }
2293 #endif
2294
2295         /*
2296          * We allocate what appears to be a normal pmap but because portions
2297          * of this pmap are shared with other unrelated pmaps we have to
2298          * set pm_active to point to all cpus.
2299          *
2300          * XXX Currently using pmap_spin to interlock the update, can't use
2301          *     vm_object_hold/drop because the token might already be held
2302          *     shared OR exclusive and we don't know.
2303          */
2304         while ((obpmap = *obpmapp) == NULL) {
2305                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2306                 pmap_pinit_simple(obpmap);
2307                 pmap_pinit2(obpmap);
2308                 spin_lock(&pmap_spin);
2309                 if (*obpmapp != NULL) {
2310                         /*
2311                          * Handle race
2312                          */
2313                         spin_unlock(&pmap_spin);
2314                         pmap_release(obpmap);
2315                         pmap_puninit(obpmap);
2316                         kfree(obpmap, M_OBJPMAP);
2317                         obpmap = *obpmapp; /* safety */
2318                 } else {
2319                         obpmap->pm_active = smp_active_mask;
2320                         obpmap->pm_flags |= PMAP_SEGSHARED;
2321                         *obpmapp = obpmap;
2322                         spin_unlock(&pmap_spin);
2323                 }
2324         }
2325
2326         /*
2327          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2328          * pte/pt using the shared pmap from the object but also adjust
2329          * the process pmap's page table page as a side effect.
2330          */
2331
2332         /*
2333          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2334          * we will return.  This is true if ptepindex represents a terminal
2335          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2336          * the PD.
2337          */
2338         pt_pv = NULL;
2339         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2340 retry:
2341         if (ptepindex >= pmap_pt_pindex(0))
2342                 xpv = pte_pv;
2343         else
2344                 xpv = pt_pv;
2345
2346         /*
2347          * Resolve the PD in the process pmap so we can properly share the
2348          * page table page.  Lock order is bottom-up (leaf first)!
2349          *
2350          * NOTE: proc_pt_pv can be NULL.
2351          */
2352         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b), NULL);
2353         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2354 #ifdef PMAP_DEBUG2
2355         if (pmap_enter_debug > 0) {
2356                 --pmap_enter_debug;
2357                 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2358                         proc_pt_pv,
2359                         (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2360                         proc_pd_pv,
2361                         va);
2362         }
2363 #endif
2364
2365         /*
2366          * xpv is the page table page pv from the shared object
2367          * (for convenience), from above.
2368          *
2369          * Calculate the pte value for the PT to load into the process PD.
2370          * If we have to change it we must properly dispose of the previous
2371          * entry.
2372          */
2373         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2374         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2375                (pmap->pmap_bits[PG_U_IDX] |
2376                 pmap->pmap_bits[PG_RW_IDX] |
2377                 pmap->pmap_bits[PG_V_IDX] |
2378                 pmap->pmap_bits[PG_A_IDX] |
2379                 pmap->pmap_bits[PG_M_IDX]);
2380
2381         /*
2382          * Dispose of previous page table page if it was local to the
2383          * process pmap.  If the old pt is not empty we cannot dispose of it
2384          * until we clean it out.  This case should not arise very often so
2385          * it is not optimized.
2386          *
2387          * Leave pt_pv and pte_pv (in our object pmap) locked and intact
2388          * for the retry.
2389          */
2390         if (proc_pt_pv) {
2391                 pmap_inval_bulk_t bulk;
2392
2393                 if (proc_pt_pv->pv_m->wire_count != 1) {
2394                         pv_put(proc_pd_pv);
2395                         pv_put(proc_pt_pv);
2396                         pmap_remove(pmap,
2397                                     va & ~(vm_offset_t)SEG_MASK,
2398                                     (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2399                         goto retry;
2400                 }
2401
2402                 /*
2403                  * The release call will indirectly clean out *pt
2404                  */
2405                 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2406                 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2407                 pmap_inval_bulk_flush(&bulk);
2408                 proc_pt_pv = NULL;
2409                 /* relookup */
2410                 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2411         }
2412
2413         /*
2414          * Handle remaining cases.
2415          */
2416         if (*pt == 0) {
2417                 atomic_swap_long(pt, npte);
2418                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2419                 vm_page_wire_quick(proc_pd_pv->pv_m);   /* proc pd for sh pt */
2420                 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2421         } else if (*pt != npte) {
2422                 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2423
2424 #if 0
2425                 opte = pte_load_clear(pt);
2426                 KKASSERT(opte && opte != npte);
2427
2428                 *pt = npte;
2429 #endif
2430                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2431
2432                 /*
2433                  * Clean up opte, bump the wire_count for the process
2434                  * PD page representing the new entry if it was
2435                  * previously empty.
2436                  *
2437                  * If the entry was not previously empty and we have
2438                  * a PT in the proc pmap then opte must match that
2439                  * pt.  The proc pt must be retired (this is done
2440                  * later on in this procedure).
2441                  *
2442                  * NOTE: replacing valid pte, wire_count on proc_pd_pv
2443                  * stays the same.
2444                  */
2445                 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2446                 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2447                 if (vm_page_unwire_quick(m)) {
2448                         panic("pmap_allocpte_seg: "
2449                               "bad wire count %p",
2450                               m);
2451                 }
2452         }
2453
2454         /*
2455          * The existing process page table was replaced and must be destroyed
2456          * here.
2457          */
2458         if (proc_pd_pv)
2459                 pv_put(proc_pd_pv);
2460         if (pvpp)
2461                 *pvpp = pt_pv;
2462         else
2463                 pv_put(pt_pv);
2464
2465         return (pte_pv);
2466 }
2467
2468 /*
2469  * Release any resources held by the given physical map.
2470  *
2471  * Called when a pmap initialized by pmap_pinit is being released.  Should
2472  * only be called if the map contains no valid mappings.
2473  */
2474 struct pmap_release_info {
2475         pmap_t  pmap;
2476         int     retry;
2477         pv_entry_t pvp;
2478 };
2479
2480 static int pmap_release_callback(pv_entry_t pv, void *data);
2481
2482 void
2483 pmap_release(struct pmap *pmap)
2484 {
2485         struct pmap_release_info info;
2486
2487         KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2488                 ("pmap still active! %016jx",
2489                 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2490
2491         /*
2492          * There is no longer a pmap_list, if there were we would remove the
2493          * pmap from it here.
2494          */
2495
2496         /*
2497          * Pull pv's off the RB tree in order from low to high and release
2498          * each page.
2499          */
2500         info.pmap = pmap;
2501         do {
2502                 info.retry = 0;
2503                 info.pvp = NULL;
2504
2505                 spin_lock(&pmap->pm_spin);
2506                 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2507                         pmap_release_callback, &info);
2508                 spin_unlock(&pmap->pm_spin);
2509
2510                 if (info.pvp)
2511                         pv_put(info.pvp);
2512         } while (info.retry);
2513
2514
2515         /*
2516          * One resident page (the pml4 page) should remain.
2517          * No wired pages should remain.
2518          */
2519 #if 1
2520         if (pmap->pm_stats.resident_count !=
2521             ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1) ||
2522             pmap->pm_stats.wired_count != 0) {
2523                 kprintf("fatal pmap problem - pmap %p flags %08x "
2524                         "rescnt=%jd wirecnt=%jd\n",
2525                         pmap,
2526                         pmap->pm_flags,
2527                         pmap->pm_stats.resident_count,
2528                         pmap->pm_stats.wired_count);
2529                 tsleep(pmap, 0, "DEAD", 0);
2530         }
2531 #else
2532         KKASSERT(pmap->pm_stats.resident_count ==
2533                  ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2534         KKASSERT(pmap->pm_stats.wired_count == 0);
2535 #endif
2536 }
2537
2538 /*
2539  * Called from low to high.  We must cache the proper parent pv so we
2540  * can adjust its wired count.
2541  */
2542 static int
2543 pmap_release_callback(pv_entry_t pv, void *data)
2544 {
2545         struct pmap_release_info *info = data;
2546         pmap_t pmap = info->pmap;
2547         vm_pindex_t pindex;
2548         int r;
2549
2550         /*
2551          * Acquire a held and locked pv, check for release race
2552          */
2553         pindex = pv->pv_pindex;
2554         if (info->pvp == pv) {
2555                 spin_unlock(&pmap->pm_spin);
2556                 info->pvp = NULL;
2557         } else if (pv_hold_try(pv)) {
2558                 spin_unlock(&pmap->pm_spin);
2559         } else {
2560                 spin_unlock(&pmap->pm_spin);
2561                 pv_lock(pv);
2562                 pv_put(pv);
2563                 info->retry = 1;
2564                 spin_lock(&pmap->pm_spin);
2565
2566                 return -1;
2567         }
2568         KKASSERT(pv->pv_pmap == pmap && pindex == pv->pv_pindex);
2569
2570         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2571                 /*
2572                  * I am PTE, parent is PT
2573                  */
2574                 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2575                 pindex += NUPTE_TOTAL;
2576         } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2577                 /*
2578                  * I am PT, parent is PD
2579                  */
2580                 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2581                 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2582         } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2583                 /*
2584                  * I am PD, parent is PDP
2585                  */
2586                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2587                          NPDPEPGSHIFT;
2588                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2589         } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2590                 /*
2591                  * I am PDP, parent is PML4 (there's only one)
2592                  */
2593 #if 0
2594                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2595                            NUPD_TOTAL) >> NPML4EPGSHIFT;
2596                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2597 #endif
2598                 pindex = pmap_pml4_pindex();
2599         } else {
2600                 /*
2601                  * parent is NULL
2602                  */
2603                 if (info->pvp) {
2604                         pv_put(info->pvp);
2605                         info->pvp = NULL;
2606                 }
2607                 pindex = 0;
2608         }
2609         if (pindex) {
2610                 if (info->pvp && info->pvp->pv_pindex != pindex) {
2611                         pv_put(info->pvp);
2612                         info->pvp = NULL;
2613                 }
2614                 if (info->pvp == NULL)
2615                         info->pvp = pv_get(pmap, pindex, NULL);
2616         } else {
2617                 if (info->pvp) {
2618                         pv_put(info->pvp);
2619                         info->pvp = NULL;
2620                 }
2621         }
2622         r = pmap_release_pv(pv, info->pvp, NULL);
2623         spin_lock(&pmap->pm_spin);
2624
2625         return(r);
2626 }
2627
2628 /*
2629  * Called with held (i.e. also locked) pv.  This function will dispose of
2630  * the lock along with the pv.
2631  *
2632  * If the caller already holds the locked parent page table for pv it
2633  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2634  * pass NULL for pvp.
2635  */
2636 static int
2637 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2638 {
2639         vm_page_t p;
2640
2641         /*
2642          * The pmap is currently not spinlocked, pv is held+locked.
2643          * Remove the pv's page from its parent's page table.  The
2644          * parent's page table page's wire_count will be decremented.
2645          *
2646          * This will clean out the pte at any level of the page table.
2647          * If smp != 0 all cpus are affected.
2648          *
2649          * Do not tear-down recursively, its faster to just let the
2650          * release run its course.
2651          */
2652         pmap_remove_pv_pte(pv, pvp, bulk, 0);
2653
2654         /*
2655          * Terminal pvs are unhooked from their vm_pages.  Because
2656          * terminal pages aren't page table pages they aren't wired
2657          * by us, so we have to be sure not to unwire them either.
2658          */
2659         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2660                 pmap_remove_pv_page(pv);
2661                 goto skip;
2662         }
2663
2664         /*
2665          * We leave the top-level page table page cached, wired, and
2666          * mapped in the pmap until the dtor function (pmap_puninit())
2667          * gets called.
2668          *
2669          * Since we are leaving the top-level pv intact we need
2670          * to break out of what would otherwise be an infinite loop.
2671          */
2672         if (pv->pv_pindex == pmap_pml4_pindex()) {
2673                 pv_put(pv);
2674                 return(-1);
2675         }
2676
2677         /*
2678          * For page table pages (other than the top-level page),
2679          * remove and free the vm_page.  The representitive mapping
2680          * removed above by pmap_remove_pv_pte() did not undo the
2681          * last wire_count so we have to do that as well.
2682          */
2683         p = pmap_remove_pv_page(pv);
2684         vm_page_busy_wait(p, FALSE, "pmaprl");
2685         if (p->wire_count != 1) {
2686                 kprintf("p->wire_count was %016lx %d\n",
2687                         pv->pv_pindex, p->wire_count);
2688         }
2689         KKASSERT(p->wire_count == 1);
2690         KKASSERT(p->flags & PG_UNMANAGED);
2691
2692         vm_page_unwire(p, 0);
2693         KKASSERT(p->wire_count == 0);
2694
2695         vm_page_free(p);
2696 skip:
2697         pv_free(pv, pvp);
2698
2699         return 0;
2700 }
2701
2702 /*
2703  * This function will remove the pte associated with a pv from its parent.
2704  * Terminal pv's are supported.  All cpus specified by (bulk) are properly
2705  * invalidated.
2706  *
2707  * The wire count will be dropped on the parent page table.  The wire
2708  * count on the page being removed (pv->pv_m) from the parent page table
2709  * is NOT touched.  Note that terminal pages will not have any additional
2710  * wire counts while page table pages will have at least one representing
2711  * the mapping, plus others representing sub-mappings.
2712  *
2713  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2714  *       pages and user page table and terminal pages.
2715  *
2716  * NOTE: The pte being removed might be unmanaged, and the pv supplied might
2717  *       be freshly allocated and not imply that the pte is managed.  In this
2718  *       case pv->pv_m should be NULL.
2719  *
2720  * The pv must be locked.  The pvp, if supplied, must be locked.  All
2721  * supplied pv's will remain locked on return.
2722  *
2723  * XXX must lock parent pv's if they exist to remove pte XXX
2724  */
2725 static
2726 void
2727 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2728                    int destroy)
2729 {
2730         vm_pindex_t ptepindex = pv->pv_pindex;
2731         pmap_t pmap = pv->pv_pmap;
2732         vm_page_t p;
2733         int gotpvp = 0;
2734
2735         KKASSERT(pmap);
2736
2737         if (ptepindex == pmap_pml4_pindex()) {
2738                 /*
2739                  * We are the top level PML4E table, there is no parent.
2740                  */
2741                 p = pmap->pm_pmlpv->pv_m;
2742                 KKASSERT(pv->pv_m == p);        /* debugging */
2743         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2744                 /*
2745                  * Remove a PDP page from the PML4E.  This can only occur
2746                  * with user page tables.  We do not have to lock the
2747                  * pml4 PV so just ignore pvp.
2748                  */
2749                 vm_pindex_t pml4_pindex;
2750                 vm_pindex_t pdp_index;
2751                 pml4_entry_t *pdp;
2752
2753                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2754                 if (pvp == NULL) {
2755                         pml4_pindex = pmap_pml4_pindex();
2756                         pvp = pv_get(pv->pv_pmap, pml4_pindex, NULL);
2757                         KKASSERT(pvp);
2758                         gotpvp = 1;
2759                 }
2760
2761                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2762                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2763                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2764                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2765                 KKASSERT(pv->pv_m == p);        /* debugging */
2766         } else if (ptepindex >= pmap_pd_pindex(0)) {
2767                 /*
2768                  * Remove a PD page from the PDP
2769                  *
2770                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2771                  *                   of a simple pmap because it stops at
2772                  *                   the PD page.
2773                  */
2774                 vm_pindex_t pdp_pindex;
2775                 vm_pindex_t pd_index;
2776                 pdp_entry_t *pd;
2777
2778                 pd_index = ptepindex - pmap_pd_pindex(0);
2779
2780                 if (pvp == NULL) {
2781                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2782                                      (pd_index >> NPML4EPGSHIFT);
2783                         pvp = pv_get(pv->pv_pmap, pdp_pindex, NULL);
2784                         gotpvp = 1;
2785                 }
2786
2787                 if (pvp) {
2788                         pd = pv_pte_lookup(pvp, pd_index &
2789                                                 ((1ul << NPDPEPGSHIFT) - 1));
2790                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2791                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2792                         pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2793                 } else {
2794                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2795                         p = pv->pv_m;           /* degenerate test later */
2796                 }
2797                 KKASSERT(pv->pv_m == p);        /* debugging */
2798         } else if (ptepindex >= pmap_pt_pindex(0)) {
2799                 /*
2800                  *  Remove a PT page from the PD
2801                  */
2802                 vm_pindex_t pd_pindex;
2803                 vm_pindex_t pt_index;
2804                 pd_entry_t *pt;
2805
2806                 pt_index = ptepindex - pmap_pt_pindex(0);
2807
2808                 if (pvp == NULL) {
2809                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2810                                     (pt_index >> NPDPEPGSHIFT);
2811                         pvp = pv_get(pv->pv_pmap, pd_pindex, NULL);
2812                         KKASSERT(pvp);
2813                         gotpvp = 1;
2814                 }
2815
2816                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2817 #if 0
2818                 KASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0,
2819                         ("*pt unexpectedly invalid %016jx "
2820                          "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
2821                         *pt, gotpvp, ptepindex, pt_index, pv, pvp));
2822                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2823 #else
2824                 if ((*pt & pmap->pmap_bits[PG_V_IDX]) == 0) {
2825                         kprintf("*pt unexpectedly invalid %016jx "
2826                                 "gotpvp=%d ptepindex=%ld ptindex=%ld "
2827                                 "pv=%p pvp=%p\n",
2828                                 *pt, gotpvp, ptepindex, pt_index, pv, pvp);
2829                         tsleep(pt, 0, "DEAD", 0);
2830                         p = pv->pv_m;
2831                 } else {
2832                         p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2833                 }
2834 #endif
2835                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2836                 KKASSERT(pv->pv_m == p);        /* debugging */
2837         } else {
2838                 /*
2839                  * Remove a PTE from the PT page.  The PV might exist even if
2840                  * the PTE is not managed, in whichcase pv->pv_m should be
2841                  * NULL.
2842                  *
2843                  * NOTE: Userland pmaps manage the parent PT/PD/PDP page
2844                  *       table pages but the kernel_pmap does not.
2845                  *
2846                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2847                  *       pv is a pte_pv so we can safely lock pt_pv.
2848                  *
2849                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2850                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2851                  *       terminal ptes.
2852                  */
2853                 vm_pindex_t pt_pindex;
2854                 pt_entry_t *ptep;
2855                 pt_entry_t pte;
2856                 vm_offset_t va;
2857
2858                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2859                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2860
2861                 if (ptepindex >= NUPTE_USER) {
2862                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2863                         KKASSERT(pvp == NULL);
2864                         /* pvp remains NULL */
2865                 } else {
2866                         if (pvp == NULL) {
2867                                 pt_pindex = NUPTE_TOTAL +
2868                                             (ptepindex >> NPDPEPGSHIFT);
2869                                 pvp = pv_get(pv->pv_pmap, pt_pindex, NULL);
2870                                 KKASSERT(pvp);
2871                                 gotpvp = 1;
2872                         }
2873                         ptep = pv_pte_lookup(pvp, ptepindex &
2874                                                   ((1ul << NPDPEPGSHIFT) - 1));
2875                 }
2876                 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2877                 if (bulk == NULL)               /* XXX */
2878                         cpu_invlpg((void *)va); /* XXX */
2879
2880                 /*
2881                  * Now update the vm_page_t
2882                  */
2883                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
2884                     (pte & pmap->pmap_bits[PG_V_IDX])) {
2885                         /*
2886                          * Valid managed page, adjust (p).
2887                          */
2888                         if (pte & pmap->pmap_bits[PG_DEVICE_IDX]) {
2889                                 p = pv->pv_m;
2890                         } else {
2891                                 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2892                                 KKASSERT(pv->pv_m == p);
2893                         }
2894                         if (pte & pmap->pmap_bits[PG_M_IDX]) {
2895                                 if (pmap_track_modified(ptepindex))
2896                                         vm_page_dirty(p);
2897                         }
2898                         if (pte & pmap->pmap_bits[PG_A_IDX]) {
2899                                 vm_page_flag_set(p, PG_REFERENCED);
2900                         }
2901                 } else {
2902                         /*
2903                          * Unmanaged page, do not try to adjust the vm_page_t.
2904                          * pv could be freshly allocated for a pmap_enter(),
2905                          * replacing an unmanaged page with a managed one.
2906                          *
2907                          * pv->pv_m might reflect the new page and not the
2908                          * existing page.
2909                          *
2910                          * We could extract p from the physical address and
2911                          * adjust it but we explicitly do not for unmanaged
2912                          * pages.
2913                          */
2914                         p = NULL;
2915                 }
2916                 if (pte & pmap->pmap_bits[PG_W_IDX])
2917                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
2918                 if (pte & pmap->pmap_bits[PG_G_IDX])
2919                         cpu_invlpg((void *)va);
2920         }
2921
2922         /*
2923          * If requested, scrap the underlying pv->pv_m and the underlying
2924          * pv.  If this is a page-table-page we must also free the page.
2925          *
2926          * pvp must be returned locked.
2927          */
2928         if (destroy == 1) {
2929                 /*
2930                  * page table page (PT, PD, PDP, PML4), caller was responsible
2931                  * for testing wired_count.
2932                  */
2933                 KKASSERT(pv->pv_m->wire_count == 1);
2934                 p = pmap_remove_pv_page(pv);
2935                 pv_free(pv, pvp);
2936                 pv = NULL;
2937
2938                 vm_page_busy_wait(p, FALSE, "pgpun");
2939                 vm_page_unwire(p, 0);
2940                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2941                 vm_page_free(p);
2942         } else if (destroy == 2) {
2943                 /*
2944                  * Normal page, remove from pmap and leave the underlying
2945                  * page untouched.
2946                  */
2947                 pmap_remove_pv_page(pv);
2948                 pv_free(pv, pvp);
2949                 pv = NULL;              /* safety */
2950         }
2951
2952         /*
2953          * If we acquired pvp ourselves then we are responsible for
2954          * recursively deleting it.
2955          */
2956         if (pvp && gotpvp) {
2957                 /*
2958                  * Recursively destroy higher-level page tables.
2959                  *
2960                  * This is optional.  If we do not, they will still
2961                  * be destroyed when the process exits.
2962                  *
2963                  * NOTE: Do not destroy pv_entry's with extra hold refs,
2964                  *       a caller may have unlocked it and intends to
2965                  *       continue to use it.
2966                  */
2967                 if (pmap_dynamic_delete &&
2968                     pvp->pv_m &&
2969                     pvp->pv_m->wire_count == 1 &&
2970                     (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
2971                     pvp->pv_pindex != pmap_pml4_pindex()) {
2972                         if (pmap_dynamic_delete == 2)
2973                                 kprintf("A %jd %08x\n", pvp->pv_pindex, pvp->pv_hold);
2974                         if (pmap != &kernel_pmap) {
2975                                 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
2976                                 pvp = NULL;     /* safety */
2977                         } else {
2978                                 kprintf("Attempt to remove kernel_pmap pindex "
2979                                         "%jd\n", pvp->pv_pindex);
2980                                 pv_put(pvp);
2981                         }
2982                 } else {
2983                         pv_put(pvp);
2984                 }
2985         }
2986 }
2987
2988 /*
2989  * Remove the vm_page association to a pv.  The pv must be locked.
2990  */
2991 static
2992 vm_page_t
2993 pmap_remove_pv_page(pv_entry_t pv)
2994 {
2995         vm_page_t m;
2996
2997         m = pv->pv_m;
2998         vm_page_spin_lock(m);
2999         KKASSERT(m && m == pv->pv_m);
3000         pv->pv_m = NULL;
3001         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3002         pmap_page_stats_deleting(m);
3003         if (TAILQ_EMPTY(&m->md.pv_list))
3004                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3005         vm_page_spin_unlock(m);
3006
3007         return(m);
3008 }
3009
3010 /*
3011  * Grow the number of kernel page table entries, if needed.
3012  *
3013  * This routine is always called to validate any address space
3014  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
3015  * space below KERNBASE.
3016  *
3017  * kernel_map must be locked exclusively by the caller.
3018  */
3019 void
3020 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
3021 {
3022         vm_paddr_t paddr;
3023         vm_offset_t ptppaddr;
3024         vm_page_t nkpg;
3025         pd_entry_t *pt, newpt;
3026         pdp_entry_t newpd;
3027         int update_kernel_vm_end;
3028
3029         /*
3030          * bootstrap kernel_vm_end on first real VM use
3031          */
3032         if (kernel_vm_end == 0) {
3033                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
3034                 nkpt = 0;
3035                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3036                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
3037                                         ~(PAGE_SIZE * NPTEPG - 1);
3038                         nkpt++;
3039                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
3040                                 kernel_vm_end = kernel_map.max_offset;
3041                                 break;
3042                         }
3043                 }
3044         }
3045
3046         /*
3047          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
3048          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
3049          * do not want to force-fill 128G worth of page tables.
3050          */
3051         if (kstart < KERNBASE) {
3052                 if (kstart > kernel_vm_end)
3053                         kstart = kernel_vm_end;
3054                 KKASSERT(kend <= KERNBASE);
3055                 update_kernel_vm_end = 1;
3056         } else {
3057                 update_kernel_vm_end = 0;
3058         }
3059
3060         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
3061         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
3062
3063         if (kend - 1 >= kernel_map.max_offset)
3064                 kend = kernel_map.max_offset;
3065
3066         while (kstart < kend) {
3067                 pt = pmap_pt(&kernel_pmap, kstart);
3068                 if (pt == NULL) {
3069                         /* We need a new PD entry */
3070                         nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3071                                              VM_ALLOC_NORMAL |
3072                                              VM_ALLOC_SYSTEM |
3073                                              VM_ALLOC_INTERRUPT);
3074                         if (nkpg == NULL) {
3075                                 panic("pmap_growkernel: no memory to grow "
3076                                       "kernel");
3077                         }
3078                         paddr = VM_PAGE_TO_PHYS(nkpg);
3079                         pmap_zero_page(paddr);
3080                         newpd = (pdp_entry_t)
3081                             (paddr |
3082                             kernel_pmap.pmap_bits[PG_V_IDX] |
3083                             kernel_pmap.pmap_bits[PG_RW_IDX] |
3084                             kernel_pmap.pmap_bits[PG_A_IDX] |
3085                             kernel_pmap.pmap_bits[PG_M_IDX]);
3086                         *pmap_pd(&kernel_pmap, kstart) = newpd;
3087                         continue; /* try again */
3088                 }
3089                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3090                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
3091                                  ~(PAGE_SIZE * NPTEPG - 1);
3092                         if (kstart - 1 >= kernel_map.max_offset) {
3093                                 kstart = kernel_map.max_offset;
3094                                 break;
3095                         }
3096                         continue;
3097                 }
3098
3099                 /*
3100                  * We need a new PT
3101                  *
3102                  * This index is bogus, but out of the way
3103                  */
3104                 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3105                                      VM_ALLOC_NORMAL |
3106                                      VM_ALLOC_SYSTEM |
3107                                      VM_ALLOC_INTERRUPT);
3108                 if (nkpg == NULL)
3109                         panic("pmap_growkernel: no memory to grow kernel");
3110
3111                 vm_page_wire(nkpg);
3112                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
3113                 pmap_zero_page(ptppaddr);
3114                 newpt = (pd_entry_t)(ptppaddr |
3115                                      kernel_pmap.pmap_bits[PG_V_IDX] |
3116                                      kernel_pmap.pmap_bits[PG_RW_IDX] |
3117                                      kernel_pmap.pmap_bits[PG_A_IDX] |
3118                                      kernel_pmap.pmap_bits[PG_M_IDX]);
3119                 atomic_swap_long(pmap_pt(&kernel_pmap, kstart), newpt);
3120
3121                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3122                           ~(PAGE_SIZE * NPTEPG - 1);
3123
3124                 if (kstart - 1 >= kernel_map.max_offset) {
3125                         kstart = kernel_map.max_offset;
3126                         break;
3127                 }
3128         }
3129
3130         /*
3131          * Only update kernel_vm_end for areas below KERNBASE.
3132          */
3133         if (update_kernel_vm_end && kernel_vm_end < kstart)
3134                 kernel_vm_end = kstart;
3135 }
3136
3137 /*
3138  *      Add a reference to the specified pmap.
3139  */
3140 void
3141 pmap_reference(pmap_t pmap)
3142 {
3143         if (pmap != NULL)
3144                 atomic_add_int(&pmap->pm_count, 1);
3145 }
3146
3147 /***************************************************
3148  * page management routines.
3149  ***************************************************/
3150
3151 /*
3152  * Hold a pv without locking it
3153  */
3154 static void
3155 pv_hold(pv_entry_t pv)
3156 {
3157         atomic_add_int(&pv->pv_hold, 1);
3158 }
3159
3160 /*
3161  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
3162  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
3163  * the pv properly.
3164  *
3165  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3166  * pv list via its page) must be held by the caller in order to stabilize
3167  * the pv.
3168  */
3169 static int
3170 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3171 {
3172         u_int count;
3173
3174         /*
3175          * Critical path shortcut expects pv to already have one ref
3176          * (for the pv->pv_pmap).
3177          */
3178         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3179 #ifdef PMAP_DEBUG
3180                 pv->pv_func = func;
3181                 pv->pv_line = lineno;
3182 #endif
3183                 return TRUE;
3184         }
3185
3186         for (;;) {
3187                 count = pv->pv_hold;
3188                 cpu_ccfence();
3189                 if ((count & PV_HOLD_LOCKED) == 0) {
3190                         if (atomic_cmpset_int(&pv->pv_hold, count,
3191                                               (count + 1) | PV_HOLD_LOCKED)) {
3192 #ifdef PMAP_DEBUG
3193                                 pv->pv_func = func;
3194                                 pv->pv_line = lineno;
3195 #endif
3196                                 return TRUE;
3197                         }
3198                 } else {
3199                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3200                                 return FALSE;
3201                 }
3202                 /* retry */
3203         }
3204 }
3205
3206 /*
3207  * Drop a previously held pv_entry which could not be locked, allowing its
3208  * destruction.
3209  *
3210  * Must not be called with a spinlock held as we might zfree() the pv if it
3211  * is no longer associated with a pmap and this was the last hold count.
3212  */
3213 static void
3214 pv_drop(pv_entry_t pv)
3215 {
3216         u_int count;
3217
3218         for (;;) {
3219                 count = pv->pv_hold;
3220                 cpu_ccfence();
3221                 KKASSERT((count & PV_HOLD_MASK) > 0);
3222                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3223                          (PV_HOLD_LOCKED | 1));
3224                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3225                         if ((count & PV_HOLD_MASK) == 1) {
3226 #ifdef PMAP_DEBUG2
3227                                 if (pmap_enter_debug > 0) {
3228                                         --pmap_enter_debug;
3229                                         kprintf("pv_drop: free pv %p\n", pv);
3230                                 }
3231 #endif
3232                                 KKASSERT(count == 1);
3233                                 KKASSERT(pv->pv_pmap == NULL);
3234                                 zfree(pvzone, pv);
3235                         }
3236                         return;
3237                 }
3238                 /* retry */
3239         }
3240 }
3241
3242 /*
3243  * Find or allocate the requested PV entry, returning a locked, held pv.
3244  *
3245  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3246  * for the caller and one representing the pmap and vm_page association.
3247  *
3248  * If (*isnew) is zero, the returned pv will have only one hold count.
3249  *
3250  * Since both associations can only be adjusted while the pv is locked,
3251  * together they represent just one additional hold.
3252  */
3253 static
3254 pv_entry_t
3255 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3256 {
3257         pv_entry_t pv;
3258         pv_entry_t pnew;
3259         struct mdglobaldata *md = mdcpu;
3260
3261         pnew = NULL;
3262         if (md->gd_newpv) {
3263 #if 0
3264                 pnew = atomic_swap_ptr((void *)&md->gd_newpv, NULL);
3265 #else
3266                 crit_enter();
3267                 pnew = md->gd_newpv;    /* might race NULL */
3268                 md->gd_newpv = NULL;
3269                 crit_exit();
3270 #endif
3271         }
3272         if (pnew == NULL)
3273                 pnew = zalloc(pvzone);
3274
3275         spin_lock(&pmap->pm_spin);
3276         for (;;) {
3277                 /*
3278                  * Shortcut cache
3279                  */
3280                 pv = pmap->pm_pvhint;
3281                 cpu_ccfence();
3282                 if (pv == NULL ||
3283                     pv->pv_pmap != pmap ||
3284                     pv->pv_pindex != pindex) {
3285                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3286                                                         pindex);
3287                 }
3288                 if (pv == NULL) {
3289                         vm_pindex_t *pmark;
3290
3291                         /*
3292                          * We need to block if someone is holding a
3293                          * placemarker.  The exclusive spinlock is a
3294                          * sufficient interlock, as long as we determine
3295                          * the placemarker has not been aquired we do not
3296                          * need to get it.
3297                          */
3298                         pmark = pmap_placemarker_hash(pmap, pindex);
3299
3300                         if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3301                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3302                                 ssleep(pmark, &pmap->pm_spin, 0, "pvplc", 0);
3303                                 continue;
3304                         }
3305
3306                         /*
3307                          * Setup the new entry
3308                          */
3309                         pnew->pv_pmap = pmap;
3310                         pnew->pv_pindex = pindex;
3311                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
3312 #ifdef PMAP_DEBUG
3313                         pnew->pv_func = func;
3314                         pnew->pv_line = lineno;
3315                         if (pnew->pv_line_lastfree > 0) {
3316                                 pnew->pv_line_lastfree =
3317                                                 -pnew->pv_line_lastfree;
3318                         }
3319 #endif
3320                         pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3321                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
3322                         spin_unlock(&pmap->pm_spin);
3323                         *isnew = 1;
3324
3325                         KKASSERT(pv == NULL);
3326                         return(pnew);
3327                 }
3328
3329                 /*
3330                  * We already have an entry, cleanup the staged pnew if
3331                  * we can get the lock, otherwise block and retry.
3332                  */
3333                 if (__predict_true(_pv_hold_try(pv PMAP_DEBUG_COPY))) {
3334                         spin_unlock(&pmap->pm_spin);
3335 #if 0
3336                         pnew = atomic_swap_ptr((void *)&md->gd_newpv, pnew);
3337                         if (pnew)
3338                                 zfree(pvzone, pnew);
3339 #else
3340                         crit_enter();
3341                         if (md->gd_newpv == NULL)
3342                                 md->gd_newpv = pnew;
3343                         else
3344                                 zfree(pvzone, pnew);
3345                         crit_exit();
3346 #endif
3347                         KKASSERT(pv->pv_pmap == pmap &&
3348                                  pv->pv_pindex == pindex);
3349                         *isnew = 0;
3350                         return(pv);
3351                 }
3352                 spin_unlock(&pmap->pm_spin);
3353                 _pv_lock(pv PMAP_DEBUG_COPY);
3354                 pv_put(pv);
3355                 spin_lock(&pmap->pm_spin);
3356         }
3357         /* NOT REACHED */
3358 }
3359
3360 /*
3361  * Find the requested PV entry, returning a locked+held pv or NULL
3362  */
3363 static
3364 pv_entry_t
3365 _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp PMAP_DEBUG_DECL)
3366 {
3367         pv_entry_t pv;
3368
3369         spin_lock(&pmap->pm_spin);
3370         for (;;) {
3371                 /*
3372                  * Shortcut cache
3373                  */
3374                 pv = pmap->pm_pvhint;
3375                 cpu_ccfence();
3376                 if (pv == NULL ||
3377                     pv->pv_pmap != pmap ||
3378                     pv->pv_pindex != pindex) {
3379                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3380                                                         pindex);
3381                 }
3382                 if (pv == NULL) {
3383                         /*
3384                          * Block if there is a placemarker.  If we are to
3385                          * return it, we must also aquire the spot.
3386                          */
3387                         vm_pindex_t *pmark;
3388
3389                         pmark = pmap_placemarker_hash(pmap, pindex);
3390
3391                         if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3392                             ((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3393                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3394                                 ssleep(pmark, &pmap->pm_spin, 0, "pvpld", 0);
3395                                 continue;
3396                         }
3397                         if (pmarkp) {
3398                                 if (atomic_swap_long(pmark, pindex) !=
3399                                     PM_NOPLACEMARK) {
3400                                         panic("_pv_get: pmark race");
3401                                 }
3402                                 *pmarkp = pmark;
3403                         }
3404                         spin_unlock(&pmap->pm_spin);
3405                         return NULL;
3406                 }
3407                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3408                         pv_cache(pv, pindex);
3409                         spin_unlock(&pmap->pm_spin);
3410                         KKASSERT(pv->pv_pmap == pmap &&
3411                                  pv->pv_pindex == pindex);
3412                         return(pv);
3413                 }
3414                 spin_unlock(&pmap->pm_spin);
3415                 _pv_lock(pv PMAP_DEBUG_COPY);
3416                 pv_put(pv);
3417                 spin_lock(&pmap->pm_spin);
3418         }
3419 }
3420
3421 /*
3422  * Lookup, hold, and attempt to lock (pmap,pindex).
3423  *
3424  * If the entry does not exist NULL is returned and *errorp is set to 0
3425  *
3426  * If the entry exists and could be successfully locked it is returned and
3427  * errorp is set to 0.
3428  *
3429  * If the entry exists but could NOT be successfully locked it is returned
3430  * held and *errorp is set to 1.
3431  *
3432  * If the entry is placemarked by someone else NULL is returned and *errorp
3433  * is set to 1.
3434  */
3435 static
3436 pv_entry_t
3437 pv_get_try(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp, int *errorp)
3438 {
3439         pv_entry_t pv;
3440
3441         spin_lock_shared(&pmap->pm_spin);
3442
3443         pv = pmap->pm_pvhint;
3444         cpu_ccfence();
3445         if (pv == NULL ||
3446             pv->pv_pmap != pmap ||
3447             pv->pv_pindex != pindex) {
3448                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3449         }
3450
3451         if (pv == NULL) {
3452                 vm_pindex_t *pmark;
3453
3454                 pmark = pmap_placemarker_hash(pmap, pindex);
3455
3456                 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3457                         *errorp = 1;
3458                 } else if (pmarkp &&
3459                            atomic_cmpset_long(pmark, PM_NOPLACEMARK, pindex)) {
3460                         *errorp = 0;
3461                 } else {
3462                         /*
3463                          * Can't set a placemark with a NULL pmarkp, or if
3464                          * pmarkp is non-NULL but we failed to set our
3465                          * placemark.
3466                          */
3467                         *errorp = 1;
3468                 }
3469                 if (pmarkp)
3470                         *pmarkp = pmark;
3471                 spin_unlock_shared(&pmap->pm_spin);
3472
3473                 return NULL;
3474         }
3475
3476         /*
3477          * XXX This has problems if the lock is shared, why?
3478          */
3479         if (pv_hold_try(pv)) {
3480                 pv_cache(pv, pindex);   /* overwrite ok (shared lock) */
3481                 spin_unlock_shared(&pmap->pm_spin);
3482                 *errorp = 0;
3483                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3484                 return(pv);     /* lock succeeded */
3485         }
3486         spin_unlock_shared(&pmap->pm_spin);
3487         *errorp = 1;
3488
3489         return (pv);            /* lock failed */
3490 }
3491
3492 /*
3493  * Lock a held pv, keeping the hold count
3494  */
3495 static
3496 void
3497 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3498 {
3499         u_int count;
3500
3501         for (;;) {
3502                 count = pv->pv_hold;
3503                 cpu_ccfence();
3504                 if ((count & PV_HOLD_LOCKED) == 0) {
3505                         if (atomic_cmpset_int(&pv->pv_hold, count,
3506                                               count | PV_HOLD_LOCKED)) {
3507 #ifdef PMAP_DEBUG
3508                                 pv->pv_func = func;
3509                                 pv->pv_line = lineno;
3510 #endif
3511                                 return;
3512                         }
3513                         continue;
3514                 }
3515                 tsleep_interlock(pv, 0);
3516                 if (atomic_cmpset_int(&pv->pv_hold, count,
3517                                       count | PV_HOLD_WAITING)) {
3518 #ifdef PMAP_DEBUG2
3519                         if (pmap_enter_debug > 0) {
3520                                 --pmap_enter_debug;
3521                                 kprintf("pv waiting on %s:%d\n",
3522                                         pv->pv_func, pv->pv_line);
3523                         }
3524 #endif
3525                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
3526                 }
3527                 /* retry */
3528         }
3529 }
3530
3531 /*
3532  * Unlock a held and locked pv, keeping the hold count.
3533  */
3534 static
3535 void
3536 pv_unlock(pv_entry_t pv)
3537 {
3538         u_int count;
3539
3540         for (;;) {
3541                 count = pv->pv_hold;
3542                 cpu_ccfence();
3543                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3544                          (PV_HOLD_LOCKED | 1));
3545                 if (atomic_cmpset_int(&pv->pv_hold, count,
3546                                       count &
3547                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3548                         if (count & PV_HOLD_WAITING)
3549                                 wakeup(pv);
3550                         break;
3551                 }
3552         }
3553 }
3554
3555 /*
3556  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3557  * and the hold count drops to zero we will free it.
3558  *
3559  * Caller should not hold any spin locks.  We are protected from hold races
3560  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3561  * lock held.  A pv cannot be located otherwise.
3562  */
3563 static
3564 void
3565 pv_put(pv_entry_t pv)
3566 {
3567 #ifdef PMAP_DEBUG2
3568         if (pmap_enter_debug > 0) {
3569                 --pmap_enter_debug;
3570                 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3571         }
3572 #endif
3573
3574         /*
3575          * Normal put-aways must have a pv_m associated with the pv,
3576          * but allow the case where the pv has been destructed due
3577          * to pmap_dynamic_delete.
3578          */
3579         KKASSERT(pv->pv_pmap == NULL || pv->pv_m != NULL);
3580
3581         /*
3582          * Fast - shortcut most common condition
3583          */
3584         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3585                 return;
3586
3587         /*
3588          * Slow
3589          */
3590         pv_unlock(pv);
3591         pv_drop(pv);
3592 }
3593
3594 /*
3595  * Remove the pmap association from a pv, require that pv_m already be removed,
3596  * then unlock and drop the pv.  Any pte operations must have already been
3597  * completed.  This call may result in a last-drop which will physically free
3598  * the pv.
3599  *
3600  * Removing the pmap association entails an additional drop.
3601  *
3602  * pv must be exclusively locked on call and will be disposed of on return.
3603  */
3604 static
3605 void
3606 _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL)
3607 {
3608         pmap_t pmap;
3609
3610 #ifdef PMAP_DEBUG
3611         pv->pv_func_lastfree = func;
3612         pv->pv_line_lastfree = lineno;
3613 #endif
3614         KKASSERT(pv->pv_m == NULL);
3615         KKASSERT((pv->pv_hold & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
3616                   (PV_HOLD_LOCKED|1));
3617         if ((pmap = pv->pv_pmap) != NULL) {
3618                 spin_lock(&pmap->pm_spin);
3619                 KKASSERT(pv->pv_pmap == pmap);
3620                 if (pmap->pm_pvhint == pv)
3621                         pmap->pm_pvhint = NULL;
3622                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3623                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3624                 pv->pv_pmap = NULL;
3625                 pv->pv_pindex = 0;
3626                 spin_unlock(&pmap->pm_spin);
3627
3628                 /*
3629                  * Try to shortcut three atomic ops, otherwise fall through
3630                  * and do it normally.  Drop two refs and the lock all in
3631                  * one go.
3632                  */
3633                 if (pvp)
3634                         vm_page_unwire_quick(pvp->pv_m);
3635                 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3636 #ifdef PMAP_DEBUG2
3637                         if (pmap_enter_debug > 0) {
3638                                 --pmap_enter_debug;
3639                                 kprintf("pv_free: free pv %p\n", pv);
3640                         }
3641 #endif
3642                         zfree(pvzone, pv);
3643                         return;
3644                 }
3645                 pv_drop(pv);    /* ref for pv_pmap */
3646         }
3647         pv_unlock(pv);
3648         pv_drop(pv);
3649 }
3650
3651 /*
3652  * This routine is very drastic, but can save the system
3653  * in a pinch.
3654  */
3655 void
3656 pmap_collect(void)
3657 {
3658         int i;
3659         vm_page_t m;
3660         static int warningdone=0;
3661
3662         if (pmap_pagedaemon_waken == 0)
3663                 return;
3664         pmap_pagedaemon_waken = 0;
3665         if (warningdone < 5) {
3666                 kprintf("pmap_collect: collecting pv entries -- "
3667                         "suggest increasing PMAP_SHPGPERPROC\n");
3668                 warningdone++;
3669         }
3670
3671         for (i = 0; i < vm_page_array_size; i++) {
3672                 m = &vm_page_array[i];
3673                 if (m->wire_count || m->hold_count)
3674                         continue;
3675                 if (vm_page_busy_try(m, TRUE) == 0) {
3676                         if (m->wire_count == 0 && m->hold_count == 0) {
3677                                 pmap_remove_all(m);
3678                         }
3679                         vm_page_wakeup(m);
3680                 }
3681         }
3682 }
3683
3684 /*
3685  * Scan the pmap for active page table entries and issue a callback.
3686  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3687  * its parent page table.
3688  *
3689  * pte_pv will be NULL if the page or page table is unmanaged.
3690  * pt_pv will point to the page table page containing the pte for the page.
3691  *
3692  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3693  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3694  *       process pmap's PD and page to the callback function.  This can be
3695  *       confusing because the pt_pv is really a pd_pv, and the target page
3696  *       table page is simply aliased by the pmap and not owned by it.
3697  *
3698  * It is assumed that the start and end are properly rounded to the page size.
3699  *
3700  * It is assumed that PD pages and above are managed and thus in the RB tree,
3701  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3702  */
3703 struct pmap_scan_info {
3704         struct pmap *pmap;
3705         vm_offset_t sva;
3706         vm_offset_t eva;
3707         vm_pindex_t sva_pd_pindex;
3708         vm_pindex_t eva_pd_pindex;
3709         void (*func)(pmap_t, struct pmap_scan_info *,
3710                      pv_entry_t, vm_pindex_t *, pv_entry_t,
3711                      int, vm_offset_t,
3712                      pt_entry_t *, void *);
3713         void *arg;
3714         pmap_inval_bulk_t bulk_core;
3715         pmap_inval_bulk_t *bulk;
3716         int count;
3717         int stop;
3718 };
3719
3720 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3721 static int pmap_scan_callback(pv_entry_t pv, void *data);
3722
3723 static void
3724 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3725 {
3726         struct pmap *pmap = info->pmap;
3727         pv_entry_t pd_pv;       /* A page directory PV */
3728         pv_entry_t pt_pv;       /* A page table PV */
3729         pv_entry_t pte_pv;      /* A page table entry PV */
3730         vm_pindex_t *pte_placemark;
3731         vm_pindex_t *pt_placemark;
3732         pt_entry_t *ptep;
3733         pt_entry_t oldpte;
3734         struct pv_entry dummy_pv;
3735
3736         info->stop = 0;
3737         if (pmap == NULL)
3738                 return;
3739         if (info->sva == info->eva)
3740                 return;
3741         if (smp_inval) {
3742                 info->bulk = &info->bulk_core;
3743                 pmap_inval_bulk_init(&info->bulk_core, pmap);
3744         } else {
3745                 info->bulk = NULL;
3746         }
3747
3748         /*
3749          * Hold the token for stability; if the pmap is empty we have nothing
3750          * to do.
3751          */
3752 #if 0
3753         if (pmap->pm_stats.resident_count == 0) {
3754                 return;
3755         }
3756 #endif
3757
3758         info->count = 0;
3759
3760         /*
3761          * Special handling for scanning one page, which is a very common
3762          * operation (it is?).
3763          *
3764          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3765          */
3766         if (info->sva + PAGE_SIZE == info->eva) {
3767                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3768                         /*
3769                          * Kernel mappings do not track wire counts on
3770                          * page table pages and only maintain pd_pv and
3771                          * pte_pv levels so pmap_scan() works.
3772                          */
3773                         pt_pv = NULL;
3774                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3775                                         &pte_placemark);
3776                         ptep = vtopte(info->sva);
3777                 } else {
3778                         /*
3779                          * User pages which are unmanaged will not have a
3780                          * pte_pv.  User page table pages which are unmanaged
3781                          * (shared from elsewhere) will also not have a pt_pv.
3782                          * The func() callback will pass both pte_pv and pt_pv
3783                          * as NULL in that case.
3784                          *
3785                          * We hold pte_placemark across the operation for
3786                          * unmanaged pages.
3787                          *
3788                          * WARNING!  We must hold pt_placemark across the
3789                          *           *ptep test to prevent misintepreting
3790                          *           a non-zero *ptep as a shared page
3791                          *           table page.  Hold it across the function
3792                          *           callback as well for SMP safety.
3793                          */
3794                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3795                                         &pte_placemark);
3796                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva),
3797                                         &pt_placemark);
3798                         if (pt_pv == NULL) {
3799                                 KKASSERT(pte_pv == NULL);
3800                                 pd_pv = pv_get(pmap,
3801                                                pmap_pd_pindex(info->sva),
3802                                                NULL);
3803                                 if (pd_pv) {
3804                                         ptep = pv_pte_lookup(pd_pv,
3805                                                     pmap_pt_index(info->sva));
3806                                         if (*ptep) {
3807                                                 info->func(pmap, info,
3808                                                      NULL, pt_placemark,
3809                                                      pd_pv, 1,
3810                                                      info->sva, ptep,
3811                                                      info->arg);
3812                                         } else {
3813                                                 pv_placemarker_wakeup(pmap,
3814                                                                   pt_placemark);
3815                                         }
3816                                         pv_put(pd_pv);
3817                                 } else {
3818                                         pv_placemarker_wakeup(pmap,
3819                                                               pt_placemark);
3820                                 }
3821                                 pv_placemarker_wakeup(pmap, pte_placemark);
3822                                 goto fast_skip;
3823                         }
3824                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3825                 }
3826
3827                 /*
3828                  * NOTE: *ptep can't be ripped out from under us if we hold
3829                  *       pte_pv (or pte_placemark) locked, but bits can
3830                  *       change.
3831                  */
3832                 oldpte = *ptep;
3833                 cpu_ccfence();
3834                 if (oldpte == 0) {
3835                         KKASSERT(pte_pv == NULL);
3836                         pv_placemarker_wakeup(pmap, pte_placemark);
3837                 } else if (pte_pv) {
3838                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3839                                            pmap->pmap_bits[PG_V_IDX])) ==
3840                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
3841                                  pmap->pmap_bits[PG_V_IDX]),
3842                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p",
3843                             *ptep, oldpte, info->sva, pte_pv));
3844                         info->func(pmap, info, pte_pv, NULL, pt_pv, 0,
3845                                    info->sva, ptep, info->arg);
3846                 } else {
3847                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3848                                            pmap->pmap_bits[PG_V_IDX])) ==
3849                             pmap->pmap_bits[PG_V_IDX],
3850                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
3851                             *ptep, oldpte, info->sva));
3852                         info->func(pmap, info, NULL, pte_placemark, pt_pv, 0,
3853                                    info->sva, ptep, info->arg);
3854                 }
3855                 if (pt_pv)
3856                         pv_put(pt_pv);
3857 fast_skip:
3858                 pmap_inval_bulk_flush(info->bulk);
3859                 return;
3860         }
3861
3862         /*
3863          * Nominal scan case, RB_SCAN() for PD pages and iterate from
3864          * there.
3865          *
3866          * WARNING! eva can overflow our standard ((N + mask) >> bits)
3867          *          bounds, resulting in a pd_pindex of 0.  To solve the
3868          *          problem we use an inclusive range.
3869          */
3870         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3871         info->eva_pd_pindex = pmap_pd_pindex(info->eva - PAGE_SIZE);
3872
3873         if (info->sva >= VM_MAX_USER_ADDRESS) {
3874                 /*
3875                  * The kernel does not currently maintain any pv_entry's for
3876                  * higher-level page tables.
3877                  */
3878                 bzero(&dummy_pv, sizeof(dummy_pv));
3879                 dummy_pv.pv_pindex = info->sva_pd_pindex;
3880                 spin_lock(&pmap->pm_spin);
3881                 while (dummy_pv.pv_pindex <= info->eva_pd_pindex) {
3882                         pmap_scan_callback(&dummy_pv, info);
3883                         ++dummy_pv.pv_pindex;
3884                         if (dummy_pv.pv_pindex < info->sva_pd_pindex) /*wrap*/
3885                                 break;
3886                 }
3887                 spin_unlock(&pmap->pm_spin);
3888         } else {
3889                 /*
3890                  * User page tables maintain local PML4, PDP, and PD
3891                  * pv_entry's at the very least.  PT pv's might be
3892                  * unmanaged and thus not exist.  PTE pv's might be
3893                  * unmanaged and thus not exist.
3894                  */
3895                 spin_lock(&pmap->pm_spin);
3896                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot, pmap_scan_cmp,
3897                                          pmap_scan_callback, info);
3898                 spin_unlock(&pmap->pm_spin);
3899         }
3900         pmap_inval_bulk_flush(info->bulk);
3901 }
3902
3903 /*
3904  * WARNING! pmap->pm_spin held
3905  *
3906  * WARNING! eva can overflow our standard ((N + mask) >> bits)
3907  *          bounds, resulting in a pd_pindex of 0.  To solve the
3908  *          problem we use an inclusive range.
3909  */
3910 static int
3911 pmap_scan_cmp(pv_entry_t pv, void *data)
3912 {
3913         struct pmap_scan_info *info = data;
3914         if (pv->pv_pindex < info->sva_pd_pindex)
3915                 return(-1);
3916         if (pv->pv_pindex > info->eva_pd_pindex)
3917                 return(1);
3918         return(0);
3919 }
3920
3921 /*
3922  * pmap_scan() by PDs
3923  *
3924  * WARNING! pmap->pm_spin held
3925  */
3926 static int
3927 pmap_scan_callback(pv_entry_t pv, void *data)
3928 {
3929         struct pmap_scan_info *info = data;
3930         struct pmap *pmap = info->pmap;
3931         pv_entry_t pd_pv;       /* A page directory PV */
3932         pv_entry_t pt_pv;       /* A page table PV */
3933         vm_pindex_t *pt_placemark;
3934         pt_entry_t *ptep;
3935         pt_entry_t oldpte;
3936         vm_offset_t sva;
3937         vm_offset_t eva;
3938         vm_offset_t va_next;
3939         vm_pindex_t pd_pindex;
3940         int error;
3941
3942         /*
3943          * Stop if requested
3944          */
3945         if (info->stop)
3946                 return -1;
3947
3948         /*
3949          * Pull the PD pindex from the pv before releasing the spinlock.
3950          *
3951          * WARNING: pv is faked for kernel pmap scans.
3952          */
3953         pd_pindex = pv->pv_pindex;
3954         spin_unlock(&pmap->pm_spin);
3955         pv = NULL;      /* invalid after spinlock unlocked */
3956
3957         /*
3958          * Calculate the page range within the PD.  SIMPLE pmaps are
3959          * direct-mapped for the entire 2^64 address space.  Normal pmaps
3960          * reflect the user and kernel address space which requires
3961          * cannonicalization w/regards to converting pd_pindex's back
3962          * into addresses.
3963          */
3964         sva = (pd_pindex - pmap_pd_pindex(0)) << PDPSHIFT;
3965         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3966             (sva & PML4_SIGNMASK)) {
3967                 sva |= PML4_SIGNMASK;
3968         }
3969         eva = sva + NBPDP;      /* can overflow */
3970         if (sva < info->sva)
3971                 sva = info->sva;
3972         if (eva < info->sva || eva > info->eva)
3973                 eva = info->eva;
3974
3975         /*
3976          * NOTE: kernel mappings do not track page table pages, only
3977          *       terminal pages.
3978          *
3979          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3980          *       However, for the scan to be efficient we try to
3981          *       cache items top-down.
3982          */
3983         pd_pv = NULL;
3984         pt_pv = NULL;
3985
3986         for (; sva < eva; sva = va_next) {
3987                 if (info->stop)
3988                         break;
3989                 if (sva >= VM_MAX_USER_ADDRESS) {
3990                         if (pt_pv) {
3991                                 pv_put(pt_pv);
3992                                 pt_pv = NULL;
3993                         }
3994                         goto kernel_skip;
3995                 }
3996
3997                 /*
3998                  * PD cache, scan shortcut if it doesn't exist.
3999                  */
4000                 if (pd_pv == NULL) {
4001                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4002                 } else if (pd_pv->pv_pmap != pmap ||
4003                            pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
4004                         pv_put(pd_pv);
4005                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4006                 }
4007                 if (pd_pv == NULL) {
4008                         va_next = (sva + NBPDP) & ~PDPMASK;
4009                         if (va_next < sva)
4010                                 va_next = eva;
4011                         continue;
4012                 }
4013
4014                 /*
4015                  * PT cache
4016                  *
4017                  * NOTE: The cached pt_pv can be removed from the pmap when
4018                  *       pmap_dynamic_delete is enabled.
4019                  */
4020                 if (pt_pv && (pt_pv->pv_pmap != pmap ||
4021                               pt_pv->pv_pindex != pmap_pt_pindex(sva))) {
4022                         pv_put(pt_pv);
4023                         pt_pv = NULL;
4024                 }
4025                 if (pt_pv == NULL) {
4026                         pt_pv = pv_get_try(pmap, pmap_pt_pindex(sva),
4027                                            &pt_placemark, &error);
4028                         if (error) {
4029                                 pv_put(pd_pv);  /* lock order */
4030                                 pd_pv = NULL;
4031                                 if (pt_pv) {
4032                                         pv_lock(pt_pv);
4033                                         pv_put(pt_pv);
4034                                         pt_pv = NULL;
4035                                 } else {
4036                                         pv_placemarker_wait(pmap, pt_placemark);
4037                                 }
4038                                 va_next = sva;
4039                                 continue;
4040                         }
4041                         /* may have to re-check later if pt_pv is NULL here */
4042                 }
4043
4044                 /*
4045                  * If pt_pv is NULL we either have an shared page table
4046                  * page and must issue a callback specific to that case,
4047                  * or there is no page table page.
4048                  *
4049                  * Either way we can skip the page table page.
4050                  *
4051                  * WARNING! pt_pv can also be NULL due to a pv creation
4052                  *          race where we find it to be NULL and then
4053                  *          later see a pte_pv.  But its possible the pt_pv
4054                  *          got created inbetween the two operations, so
4055                  *          we must check.
4056                  */
4057                 if (pt_pv == NULL) {
4058                         /*
4059                          * Possible unmanaged (shared from another pmap)
4060                          * page table page.
4061                          *
4062                          * WARNING!  We must hold pt_placemark across the
4063                          *           *ptep test to prevent misintepreting
4064                          *           a non-zero *ptep as a shared page
4065                          *           table page.  Hold it across the function
4066                          *           callback as well for SMP safety.
4067                          */
4068                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
4069                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
4070                                 info->func(pmap, info, NULL, pt_placemark,
4071                                            pd_pv, 1,
4072                                            sva, ptep, info->arg);
4073                         } else {
4074                                 pv_placemarker_wakeup(pmap, pt_placemark);
4075                         }
4076
4077                         /*
4078                          * Done, move to next page table page.
4079                          */
4080                         va_next = (sva + NBPDR) & ~PDRMASK;
4081                         if (va_next < sva)
4082                                 va_next = eva;
4083                         continue;
4084                 }
4085
4086                 /*
4087                  * From this point in the loop testing pt_pv for non-NULL
4088                  * means we are in UVM, else if it is NULL we are in KVM.
4089                  *
4090                  * Limit our scan to either the end of the va represented
4091                  * by the current page table page, or to the end of the
4092                  * range being removed.
4093                  */
4094 kernel_skip:
4095                 va_next = (sva + NBPDR) & ~PDRMASK;
4096                 if (va_next < sva)
4097                         va_next = eva;
4098                 if (va_next > eva)
4099                         va_next = eva;
4100
4101                 /*
4102                  * Scan the page table for pages.  Some pages may not be
4103                  * managed (might not have a pv_entry).
4104                  *
4105                  * There is no page table management for kernel pages so
4106                  * pt_pv will be NULL in that case, but otherwise pt_pv
4107                  * is non-NULL, locked, and referenced.
4108                  */
4109
4110                 /*
4111                  * At this point a non-NULL pt_pv means a UVA, and a NULL
4112                  * pt_pv means a KVA.
4113                  */
4114                 if (pt_pv)
4115                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
4116                 else
4117                         ptep = vtopte(sva);
4118
4119                 while (sva < va_next) {
4120                         pv_entry_t pte_pv;
4121                         vm_pindex_t *pte_placemark;
4122
4123                         /*
4124                          * Yield every 64 pages, stop if requested.
4125                          */
4126                         if ((++info->count & 63) == 0)
4127                                 lwkt_user_yield();
4128                         if (info->stop)
4129                                 break;
4130
4131                         /*
4132                          * We can shortcut our scan if *ptep == 0.  This is
4133                          * an unlocked check.
4134                          */
4135                         if (*ptep == 0) {
4136                                 sva += PAGE_SIZE;
4137                                 ++ptep;
4138                                 continue;
4139                         }
4140                         cpu_ccfence();
4141
4142                         /*
4143                          * Acquire the related pte_pv, if any.  If *ptep == 0
4144                          * the related pte_pv should not exist, but if *ptep
4145                          * is not zero the pte_pv may or may not exist (e.g.
4146                          * will not exist for an unmanaged page).
4147                          *
4148                          * However a multitude of races are possible here
4149                          * so if we cannot lock definite state we clean out
4150                          * our cache and break the inner while() loop to
4151                          * force a loop up to the top of the for().
4152                          *
4153                          * XXX unlock/relock pd_pv, pt_pv, and re-test their
4154                          *     validity instead of looping up?
4155                          */
4156                         pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
4157                                             &pte_placemark, &error);
4158                         if (error) {
4159                                 pv_put(pd_pv);          /* lock order */
4160                                 pd_pv = NULL;
4161                                 if (pt_pv) {
4162                                         pv_put(pt_pv);  /* lock order */
4163                                         pt_pv = NULL;
4164                                 }
4165                                 if (pte_pv) {           /* block */
4166                                         pv_lock(pte_pv);
4167                                         pv_put(pte_pv);
4168                                         pte_pv = NULL;
4169                                 } else {
4170                                         pv_placemarker_wait(pmap,
4171                                                         pte_placemark);
4172                                 }
4173                                 va_next = sva;          /* retry */
4174                                 break;
4175                         }
4176
4177                         /*
4178                          * Reload *ptep after successfully locking the
4179                          * pindex.  If *ptep == 0 we had better NOT have a
4180                          * pte_pv.
4181                          */
4182                         cpu_ccfence();
4183                         oldpte = *ptep;
4184                         if (oldpte == 0) {
4185                                 if (pte_pv) {
4186                                         kprintf("Unexpected non-NULL pte_pv "
4187                                                 "%p pt_pv %p "
4188                                                 "*ptep = %016lx/%016lx\n",
4189                                                 pte_pv, pt_pv, *ptep, oldpte);
4190                                         panic("Unexpected non-NULL pte_pv");
4191                                 } else {
4192                                         pv_placemarker_wakeup(pmap, pte_placemark);
4193                                 }
4194                                 sva += PAGE_SIZE;
4195                                 ++ptep;
4196                                 continue;
4197                         }
4198
4199                         /*
4200                          * We can't hold pd_pv across the callback (because
4201                          * we don't pass it to the callback and the callback
4202                          * might deadlock)
4203                          */
4204                         if (pd_pv) {
4205                                 vm_page_wire_quick(pd_pv->pv_m);
4206                                 pv_unlock(pd_pv);
4207                         }
4208
4209                         /*
4210                          * Ready for the callback.  The locked pte_pv (if any)
4211                          * is consumed by the callback.  pte_pv will exist if
4212                          * the page is managed, and will not exist if it
4213                          * isn't.
4214                          */
4215                         if (oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4216                                 /*
4217                                  * Managed pte
4218                                  */
4219                                 KASSERT(pte_pv &&
4220                                          (oldpte & pmap->pmap_bits[PG_V_IDX]),
4221                                     ("badC *ptep %016lx/%016lx sva %016lx "
4222                                     "pte_pv %p",
4223                                     *ptep, oldpte, sva, pte_pv));
4224                                 /*
4225                                  * We must unlock pd_pv across the callback
4226                                  * to avoid deadlocks on any recursive
4227                                  * disposal.  Re-check that it still exists
4228                                  * after re-locking.
4229                                  *
4230                                  * Call target disposes of pte_pv and may
4231                                  * destroy but will not dispose of pt_pv.
4232                                  */
4233                                 info->func(pmap, info, pte_pv, NULL,
4234                                            pt_pv, 0,
4235                                            sva, ptep, info->arg);
4236                         } else {
4237                                 /*
4238                                  * Unmanaged pte
4239                                  *
4240                                  * We must unlock pd_pv across the callback
4241                                  * to avoid deadlocks on any recursive
4242                                  * disposal.  Re-check that it still exists
4243                                  * after re-locking.
4244                                  *
4245                                  * Call target disposes of pte_pv or
4246                                  * pte_placemark and may destroy but will
4247                                  * not dispose of pt_pv.
4248                                  */
4249                                 KASSERT(pte_pv == NULL &&
4250                                         (oldpte & pmap->pmap_bits[PG_V_IDX]),
4251                                     ("badD *ptep %016lx/%016lx sva %016lx "
4252                                     "pte_pv %p pte_pv->pv_m %p ",
4253                                      *ptep, oldpte, sva,
4254                                      pte_pv, (pte_pv ? pte_pv->pv_m : NULL)));
4255                                 if (pte_pv)
4256                                         kprintf("RaceD\n");
4257                                 if (pte_pv) {
4258                                         info->func(pmap, info,
4259                                                    pte_pv, NULL,
4260                                                    pt_pv, 0,
4261                                                    sva, ptep, info->arg);
4262                                 } else {
4263                                         info->func(pmap, info,
4264                                                    NULL, pte_placemark,
4265                                                    pt_pv, 0,
4266                                                    sva, ptep, info->arg);
4267                                 }
4268                         }
4269                         if (pd_pv) {
4270                                 pv_lock(pd_pv);
4271                                 vm_page_unwire_quick(pd_pv->pv_m);
4272                                 if (pd_pv->pv_pmap == NULL) {
4273                                         va_next = sva;          /* retry */
4274                                         break;
4275                                 }
4276                         }
4277
4278                         /*
4279                          * NOTE: The cached pt_pv can be removed from the
4280                          *       pmap when pmap_dynamic_delete is enabled,
4281                          *       which will cause ptep to become stale.
4282                          *
4283                          *       This also means that no pages remain under
4284                          *       the PT, so we can just break out of the inner
4285                          *       loop and let the outer loop clean everything
4286                          *       up.
4287                          */
4288                         if (pt_pv && pt_pv->pv_pmap != pmap)
4289                                 break;
4290                         pte_pv = NULL;
4291                         sva += PAGE_SIZE;
4292                         ++ptep;
4293                 }
4294         }
4295         if (pd_pv) {
4296                 pv_put(pd_pv);
4297                 pd_pv = NULL;
4298         }
4299         if (pt_pv) {
4300                 pv_put(pt_pv);
4301                 pt_pv = NULL;
4302         }
4303         if ((++info->count & 7) == 0)
4304                 lwkt_user_yield();
4305
4306         /*
4307          * Relock before returning.
4308          */
4309         spin_lock(&pmap->pm_spin);
4310         return (0);
4311 }
4312
4313 void
4314 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4315 {
4316         struct pmap_scan_info info;
4317
4318         info.pmap = pmap;
4319         info.sva = sva;
4320         info.eva = eva;
4321         info.func = pmap_remove_callback;
4322         info.arg = NULL;
4323         pmap_scan(&info, 1);
4324 #if 0
4325         cpu_invltlb();
4326         if (eva - sva < 1024*1024) {
4327                 while (sva < eva) {
4328                         cpu_invlpg((void *)sva);
4329                         sva += PAGE_SIZE;
4330                 }
4331         }
4332 #endif
4333 }
4334
4335 static void
4336 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4337 {
4338         struct pmap_scan_info info;
4339
4340         info.pmap = pmap;
4341         info.sva = sva;
4342         info.eva = eva;
4343         info.func = pmap_remove_callback;
4344         info.arg = NULL;
4345         pmap_scan(&info, 0);
4346 }
4347
4348 static void
4349 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4350                      pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4351                      pv_entry_t pt_pv, int sharept,
4352                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4353 {
4354         pt_entry_t pte;
4355
4356         if (pte_pv) {
4357                 /*
4358                  * Managed entry
4359                  *
4360                  * This will also drop pt_pv's wire_count. Note that
4361                  * terminal pages are not wired based on mmu presence.
4362                  *
4363                  * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4364                  */
4365                 KKASSERT(pte_pv->pv_m != NULL);
4366                 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4367                 pte_pv = NULL;  /* safety */
4368
4369                 /*
4370                  * Recursively destroy higher-level page tables.
4371                  *
4372                  * This is optional.  If we do not, they will still
4373                  * be destroyed when the process exits.
4374                  *
4375                  * NOTE: Do not destroy pv_entry's with extra hold refs,
4376                  *       a caller may have unlocked it and intends to
4377                  *       continue to use it.
4378                  */
4379                 if (pmap_dynamic_delete &&
4380                     pt_pv &&
4381                     pt_pv->pv_m &&
4382                     pt_pv->pv_m->wire_count == 1 &&
4383                     (pt_pv->pv_hold & PV_HOLD_MASK) == 2 &&
4384                     pt_pv->pv_pindex != pmap_pml4_pindex()) {
4385                         if (pmap_dynamic_delete == 2)
4386                                 kprintf("B %jd %08x\n", pt_pv->pv_pindex, pt_pv->pv_hold);
4387                         pv_hold(pt_pv); /* extra hold */
4388                         pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4389                         pv_lock(pt_pv); /* prior extra hold + relock */
4390                 }
4391         } else if (sharept == 0) {
4392                 /*
4393                  * Unmanaged pte (pte_placemark is non-NULL)
4394                  *
4395                  * pt_pv's wire_count is still bumped by unmanaged pages
4396                  * so we must decrement it manually.
4397                  *
4398                  * We have to unwire the target page table page.
4399                  */
4400                 pte = pmap_inval_bulk(info->bulk, va, ptep, 0);
4401                 if (pte & pmap->pmap_bits[PG_W_IDX])
4402                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4403                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4404                 if (vm_page_unwire_quick(pt_pv->pv_m))
4405                         panic("pmap_remove: insufficient wirecount");
4406                 pv_placemarker_wakeup(pmap, pte_placemark);
4407         } else {
4408                 /*
4409                  * Unmanaged page table (pt, pd, or pdp. Not pte) for
4410                  * a shared page table.
4411                  *
4412                  * pt_pv is actually the pd_pv for our pmap (not the shared
4413                  * object pmap).
4414                  *
4415                  * We have to unwire the target page table page and we
4416                  * have to unwire our page directory page.
4417                  *
4418                  * It is unclear how we can invalidate a segment so we
4419                  * invalidate -1 which invlidates the tlb.
4420                  */
4421                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4422                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4423                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4424                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4425                         panic("pmap_remove: shared pgtable1 bad wirecount");
4426                 if (vm_page_unwire_quick(pt_pv->pv_m))
4427                         panic("pmap_remove: shared pgtable2 bad wirecount");
4428                 pv_placemarker_wakeup(pmap, pte_placemark);
4429         }
4430 }
4431
4432 /*
4433  * Removes this physical page from all physical maps in which it resides.
4434  * Reflects back modify bits to the pager.
4435  *
4436  * This routine may not be called from an interrupt.
4437  */
4438 static
4439 void
4440 pmap_remove_all(vm_page_t m)
4441 {
4442         pv_entry_t pv;
4443         pmap_inval_bulk_t bulk;
4444
4445         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4446                 return;
4447
4448         vm_page_spin_lock(m);
4449         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4450                 KKASSERT(pv->pv_m == m);
4451                 if (pv_hold_try(pv)) {
4452                         vm_page_spin_unlock(m);
4453                 } else {
4454                         vm_page_spin_unlock(m);
4455                         pv_lock(pv);
4456                         pv_put(pv);
4457                         vm_page_spin_lock(m);
4458                         continue;
4459                 }
4460                 KKASSERT(pv->pv_pmap && pv->pv_m == m);
4461
4462                 /*
4463                  * Holding no spinlocks, pv is locked.  Once we scrap
4464                  * pv we can no longer use it as a list iterator (but
4465                  * we are doing a TAILQ_FIRST() so we are ok).
4466                  */
4467                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4468                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4469                 pv = NULL;      /* safety */
4470                 pmap_inval_bulk_flush(&bulk);
4471                 vm_page_spin_lock(m);
4472         }
4473         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4474         vm_page_spin_unlock(m);
4475 }
4476
4477 /*
4478  * Removes the page from a particular pmap
4479  */
4480 void
4481 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4482 {
4483         pv_entry_t pv;
4484         pmap_inval_bulk_t bulk;
4485
4486         if (!pmap_initialized)
4487                 return;
4488
4489 again:
4490         vm_page_spin_lock(m);
4491         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4492                 if (pv->pv_pmap != pmap)
4493                         continue;
4494                 KKASSERT(pv->pv_m == m);
4495                 if (pv_hold_try(pv)) {
4496                         vm_page_spin_unlock(m);
4497                 } else {
4498                         vm_page_spin_unlock(m);
4499                         pv_lock(pv);
4500                         pv_put(pv);
4501                         goto again;
4502                 }
4503                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
4504
4505                 /*
4506                  * Holding no spinlocks, pv is locked.  Once gone it can't
4507                  * be used as an iterator.  In fact, because we couldn't
4508                  * necessarily lock it atomically it may have moved within
4509                  * the list and ALSO cannot be used as an iterator.
4510                  */
4511                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4512                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4513                 pv = NULL;      /* safety */
4514                 pmap_inval_bulk_flush(&bulk);
4515                 goto again;
4516         }
4517         vm_page_spin_unlock(m);
4518 }
4519
4520 /*
4521  * Set the physical protection on the specified range of this map
4522  * as requested.  This function is typically only used for debug watchpoints
4523  * and COW pages.
4524  *
4525  * This function may not be called from an interrupt if the map is
4526  * not the kernel_pmap.
4527  *
4528  * NOTE!  For shared page table pages we just unmap the page.
4529  */
4530 void
4531 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4532 {
4533         struct pmap_scan_info info;
4534         /* JG review for NX */
4535
4536         if (pmap == NULL)
4537                 return;
4538         if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
4539                 pmap_remove(pmap, sva, eva);
4540                 return;
4541         }
4542         if (prot & VM_PROT_WRITE)
4543                 return;
4544         info.pmap = pmap;
4545         info.sva = sva;
4546         info.eva = eva;
4547         info.func = pmap_protect_callback;
4548         info.arg = &prot;
4549         pmap_scan(&info, 1);
4550 }
4551
4552 static
4553 void
4554 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4555                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4556                       pv_entry_t pt_pv, int sharept,
4557                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4558 {
4559         pt_entry_t pbits;
4560         pt_entry_t cbits;
4561         pt_entry_t pte;
4562         vm_page_t m;
4563
4564 again:
4565         pbits = *ptep;
4566         cbits = pbits;
4567         if (pte_pv) {
4568                 KKASSERT(pte_pv->pv_m != NULL);
4569                 m = NULL;
4570                 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4571                         if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4572                                 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4573                                 KKASSERT(m == pte_pv->pv_m);
4574                                 vm_page_flag_set(m, PG_REFERENCED);
4575                         }
4576                         cbits &= ~pmap->pmap_bits[PG_A_IDX];
4577                 }
4578                 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4579                         if (pmap_track_modified(pte_pv->pv_pindex)) {
4580                                 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4581                                         if (m == NULL) {
4582                                                 m = PHYS_TO_VM_PAGE(pbits &
4583                                                                     PG_FRAME);
4584                                         }
4585                                         vm_page_dirty(m);
4586                                 }
4587                                 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4588                         }
4589                 }
4590         } else if (sharept) {
4591                 /*
4592                  * Unmanaged page table, pt_pv is actually the pd_pv
4593                  * for our pmap (not the object's shared pmap).
4594                  *
4595                  * When asked to protect something in a shared page table
4596                  * page we just unmap the page table page.  We have to
4597                  * invalidate the tlb in this situation.
4598                  *
4599                  * XXX Warning, shared page tables will not be used for
4600                  * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4601                  * so PHYS_TO_VM_PAGE() should be safe here.
4602                  */
4603                 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4604                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4605                         panic("pmap_protect: pgtable1 pg bad wirecount");
4606                 if (vm_page_unwire_quick(pt_pv->pv_m))
4607                         panic("pmap_protect: pgtable2 pg bad wirecount");
4608                 ptep = NULL;
4609         }
4610         /* else unmanaged page, adjust bits, no wire changes */
4611
4612         if (ptep) {
4613                 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4614 #ifdef PMAP_DEBUG2
4615                 if (pmap_enter_debug > 0) {
4616                         --pmap_enter_debug;
4617                         kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4618                                 "pt_pv=%p cbits=%08lx\n",
4619                                 va, ptep, pte_pv,
4620                                 pt_pv, cbits
4621                         );
4622                 }
4623 #endif
4624                 if (pbits != cbits) {
4625                         vm_offset_t xva;
4626
4627                         xva = (sharept) ? (vm_offset_t)-1 : va;
4628                         if (!pmap_inval_smp_cmpset(pmap, xva,
4629                                                    ptep, pbits, cbits)) {
4630                                 goto again;
4631                         }
4632                 }
4633         }
4634         if (pte_pv)
4635                 pv_put(pte_pv);
4636         else
4637                 pv_placemarker_wakeup(pmap, pte_placemark);
4638 }
4639
4640 /*
4641  * Insert the vm_page (m) at the virtual address (va), replacing any prior
4642  * mapping at that address.  Set protection and wiring as requested.
4643  *
4644  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4645  * possible.  If it is we enter the page into the appropriate shared pmap
4646  * hanging off the related VM object instead of the passed pmap, then we
4647  * share the page table page from the VM object's pmap into the current pmap.
4648  *
4649  * NOTE: This routine MUST insert the page into the pmap now, it cannot
4650  *       lazy-evaluate.
4651  *
4652  * NOTE: If (m) is PG_UNMANAGED it may also be a temporary fake vm_page_t.
4653  *       never record it.
4654  */
4655 void
4656 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4657            boolean_t wired, vm_map_entry_t entry)
4658 {
4659         pv_entry_t pt_pv;       /* page table */
4660         pv_entry_t pte_pv;      /* page table entry */
4661         vm_pindex_t *pte_placemark;
4662         pt_entry_t *ptep;
4663         vm_paddr_t opa;
4664         pt_entry_t origpte, newpte;
4665         vm_paddr_t pa;
4666
4667         if (pmap == NULL)
4668                 return;
4669         va = trunc_page(va);
4670 #ifdef PMAP_DIAGNOSTIC
4671         if (va >= KvaEnd)
4672                 panic("pmap_enter: toobig");
4673         if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4674                 panic("pmap_enter: invalid to pmap_enter page table "
4675                       "pages (va: 0x%lx)", va);
4676 #endif
4677         if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4678                 kprintf("Warning: pmap_enter called on UVA with "
4679                         "kernel_pmap\n");
4680 #ifdef DDB
4681                 db_print_backtrace();
4682 #endif
4683         }
4684         if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4685                 kprintf("Warning: pmap_enter called on KVA without"
4686                         "kernel_pmap\n");
4687 #ifdef DDB
4688                 db_print_backtrace();
4689 #endif
4690         }
4691
4692         /*
4693          * Get locked PV entries for our new page table entry (pte_pv or
4694          * pte_placemark) and for its parent page table (pt_pv).  We need
4695          * the parent so we can resolve the location of the ptep.
4696          *
4697          * Only hardware MMU actions can modify the ptep out from
4698          * under us.
4699          *
4700          * if (m) is fictitious or unmanaged we do not create a managing
4701          * pte_pv for it.  Any pre-existing page's management state must
4702          * match (avoiding code complexity).
4703          *
4704          * If the pmap is still being initialized we assume existing
4705          * page tables.
4706          *
4707          * Kernel mapppings do not track page table pages (i.e. pt_pv).
4708          *
4709          * WARNING! If replacing a managed mapping with an unmanaged mapping
4710          *          pte_pv will wind up being non-NULL and must be handled
4711          *          below.
4712          */
4713         if (pmap_initialized == FALSE) {
4714                 pte_pv = NULL;
4715                 pt_pv = NULL;
4716                 pte_placemark = NULL;
4717                 ptep = vtopte(va);
4718                 origpte = *ptep;
4719         } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4720                 pte_pv = pv_get(pmap, pmap_pte_pindex(va), &pte_placemark);
4721                 KKASSERT(pte_pv == NULL);
4722                 if (va >= VM_MAX_USER_ADDRESS) {
4723                         pt_pv = NULL;
4724                         ptep = vtopte(va);
4725                 } else {
4726                         pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4727                                                   NULL, entry, va);
4728                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4729                 }
4730                 origpte = *ptep;
4731                 cpu_ccfence();
4732                 KASSERT(origpte == 0 ||
4733                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4734                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4735         } else {
4736                 if (va >= VM_MAX_USER_ADDRESS) {
4737                         /*
4738                          * Kernel map, pv_entry-tracked.
4739                          */
4740                         pt_pv = NULL;
4741                         pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4742                         ptep = vtopte(va);
4743                 } else {
4744                         /*
4745                          * User map
4746                          */
4747                         pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4748                                                    &pt_pv, entry, va);
4749                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4750                 }
4751                 pte_placemark = NULL;   /* safety */
4752                 origpte = *ptep;
4753                 cpu_ccfence();
4754                 KASSERT(origpte == 0 ||
4755                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4756                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4757         }
4758
4759         pa = VM_PAGE_TO_PHYS(m);
4760         opa = origpte & PG_FRAME;
4761
4762         /*
4763          * Calculate the new PTE.  Note that pte_pv alone does not mean
4764          * the new pte_pv is managed, it could exist because the old pte
4765          * was managed even if the new one is not.
4766          */
4767         newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4768                  pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4769         if (wired)
4770                 newpte |= pmap->pmap_bits[PG_W_IDX];
4771         if (va < VM_MAX_USER_ADDRESS)
4772                 newpte |= pmap->pmap_bits[PG_U_IDX];
4773         if (pte_pv && (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) == 0)
4774                 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4775 //      if (pmap == &kernel_pmap)
4776 //              newpte |= pgeflag;
4777         newpte |= pmap->pmap_cache_bits[m->pat_mode];
4778         if (m->flags & PG_FICTITIOUS)
4779                 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4780
4781         /*
4782          * It is possible for multiple faults to occur in threaded
4783          * environments, the existing pte might be correct.
4784          */
4785         if (((origpte ^ newpte) &
4786             ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4787                           pmap->pmap_bits[PG_A_IDX])) == 0) {
4788                 goto done;
4789         }
4790
4791         /*
4792          * Ok, either the address changed or the protection or wiring
4793          * changed.
4794          *
4795          * Clear the current entry, interlocking the removal.  For managed
4796          * pte's this will also flush the modified state to the vm_page.
4797          * Atomic ops are mandatory in order to ensure that PG_M events are
4798          * not lost during any transition.
4799          *
4800          * WARNING: The caller has busied the new page but not the original
4801          *          vm_page which we are trying to replace.  Because we hold
4802          *          the pte_pv lock, but have not busied the page, PG bits
4803          *          can be cleared out from under us.
4804          */
4805         if (opa) {
4806                 if (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4807                         /*
4808                          * Old page was managed.  Expect pte_pv to exist.
4809                          * (it might also exist if the old page was unmanaged).
4810                          *
4811                          * NOTE: pt_pv won't exist for a kernel page
4812                          *       (managed or otherwise).
4813                          *
4814                          * NOTE: We may be reusing the pte_pv so we do not
4815                          *       destroy it in pmap_remove_pv_pte().
4816                          */
4817                         KKASSERT(pte_pv && pte_pv->pv_m);
4818                         if (prot & VM_PROT_NOSYNC) {
4819                                 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4820                         } else {
4821                                 pmap_inval_bulk_t bulk;
4822
4823                                 pmap_inval_bulk_init(&bulk, pmap);
4824                                 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4825                                 pmap_inval_bulk_flush(&bulk);
4826                         }
4827                         pmap_remove_pv_page(pte_pv);
4828                         /* will either set pte_pv->pv_m or pv_free() later */
4829                 } else {
4830                         /*
4831                          * Old page was not managed.  If we have a pte_pv
4832                          * it better not have a pv_m assigned to it.  If the
4833                          * new page is managed the pte_pv will be destroyed
4834                          * near the end (we need its interlock).
4835                          *
4836                          * NOTE: We leave the wire count on the PT page
4837                          *       intact for the followup enter, but adjust
4838                          *       the wired-pages count on the pmap.
4839                          */
4840                         KKASSERT(pte_pv == NULL);
4841                         if (prot & VM_PROT_NOSYNC) {
4842                                 /*
4843                                  * NOSYNC (no mmu sync) requested.
4844                                  */
4845                                 (void)pte_load_clear(ptep);
4846                                 cpu_invlpg((void *)va);
4847                         } else {
4848                                 /*
4849                                  * Nominal SYNC
4850                                  */
4851                                 pmap_inval_smp(pmap, va, 1, ptep, 0);
4852                         }
4853
4854                         /*
4855                          * We must adjust pm_stats manually for unmanaged
4856                          * pages.
4857                          */
4858                         if (pt_pv) {
4859                                 atomic_add_long(&pmap->pm_stats.
4860                                                 resident_count, -1);
4861                         }
4862                         if (origpte & pmap->pmap_bits[PG_W_IDX]) {
4863                                 atomic_add_long(&pmap->pm_stats.
4864                                                 wired_count, -1);
4865                         }
4866                 }
4867                 KKASSERT(*ptep == 0);
4868         }
4869
4870 #ifdef PMAP_DEBUG2
4871         if (pmap_enter_debug > 0) {
4872                 --pmap_enter_debug;
4873                 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4874                         " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4875                         va, m,
4876                         origpte, newpte, ptep,
4877                         pte_pv, pt_pv, opa, prot);
4878         }
4879 #endif
4880
4881         if ((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
4882                 /*
4883                  * Entering an unmanaged page.  We must wire the pt_pv unless
4884                  * we retained the wiring from an unmanaged page we had
4885                  * removed (if we retained it via pte_pv that will go away
4886                  * soon).
4887                  */
4888                 if (pt_pv && (opa == 0 ||
4889                               (origpte & pmap->pmap_bits[PG_MANAGED_IDX]))) {
4890                         vm_page_wire_quick(pt_pv->pv_m);
4891                 }
4892                 if (wired)
4893                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
4894
4895                 /*
4896                  * Unmanaged pages need manual resident_count tracking.
4897                  */
4898                 if (pt_pv) {
4899                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.
4900                                         resident_count, 1);
4901                 }
4902                 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4903                         vm_page_flag_set(m, PG_WRITEABLE);
4904         } else {
4905                 /*
4906                  * Entering a managed page.  Our pte_pv takes care of the
4907                  * PT wiring, so if we had removed an unmanaged page before
4908                  * we must adjust.
4909                  *
4910                  * We have to take care of the pmap wired count ourselves.
4911                  *
4912                  * Enter on the PV list if part of our managed memory.
4913                  */
4914                 KKASSERT(pte_pv && (pte_pv->pv_m == NULL || pte_pv->pv_m == m));
4915                 vm_page_spin_lock(m);
4916                 pte_pv->pv_m = m;
4917                 pmap_page_stats_adding(m);
4918                 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4919                 vm_page_flag_set(m, PG_MAPPED);
4920                 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4921                         vm_page_flag_set(m, PG_WRITEABLE);
4922                 vm_page_spin_unlock(m);
4923
4924                 if (pt_pv && opa &&
4925                     (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
4926                         vm_page_unwire_quick(pt_pv->pv_m);
4927                 }
4928
4929                 /*
4930                  * Adjust pmap wired pages count for new entry.
4931                  */
4932                 if (wired) {
4933                         atomic_add_long(&pte_pv->pv_pmap->pm_stats.
4934                                         wired_count, 1);
4935                 }
4936         }
4937
4938         /*
4939          * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4940          *
4941          * User VMAs do not because those will be zero->non-zero, so no
4942          * stale entries to worry about at this point.
4943          *
4944          * For KVM there appear to still be issues.  Theoretically we
4945          * should be able to scrap the interlocks entirely but we
4946          * get crashes.
4947          */
4948         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4949                 pmap_inval_smp(pmap, va, 1, ptep, newpte);
4950         } else {
4951                 origpte = atomic_swap_long(ptep, newpte);
4952                 if (origpte & pmap->pmap_bits[PG_M_IDX]) {
4953                         kprintf("pmap [M] race @ %016jx\n", va);
4954                         atomic_set_long(ptep, pmap->pmap_bits[PG_M_IDX]);
4955                 }
4956                 if (pt_pv == NULL)
4957                         cpu_invlpg((void *)va);
4958         }
4959
4960         /*
4961          * Cleanup
4962          */
4963 done:
4964         KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4965                  (m->flags & PG_MAPPED));
4966
4967         /*
4968          * Cleanup the pv entry, allowing other accessors.  If the new page
4969          * is not managed but we have a pte_pv (which was locking our
4970          * operation), we can free it now.  pte_pv->pv_m should be NULL.
4971          */
4972         if (pte_pv && (newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
4973                 pv_free(pte_pv, pt_pv);
4974         } else if (pte_pv) {
4975                 pv_put(pte_pv);
4976         } else if (pte_placemark) {
4977                 pv_placemarker_wakeup(pmap, pte_placemark);
4978         }
4979         if (pt_pv)
4980                 pv_put(pt_pv);
4981 }
4982
4983 /*
4984  * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4985  * This code also assumes that the pmap has no pre-existing entry for this
4986  * VA.
4987  *
4988  * This code currently may only be used on user pmaps, not kernel_pmap.
4989  */
4990 void
4991 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4992 {
4993         pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4994 }
4995
4996 /*
4997  * Make a temporary mapping for a physical address.  This is only intended
4998  * to be used for panic dumps.
4999  *
5000  * The caller is responsible for calling smp_invltlb().
5001  */
5002 void *
5003 pmap_kenter_temporary(vm_paddr_t pa, long i)
5004 {
5005         pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
5006         return ((void *)crashdumpmap);
5007 }
5008
5009 #define MAX_INIT_PT (96)
5010
5011 /*
5012  * This routine preloads the ptes for a given object into the specified pmap.
5013  * This eliminates the blast of soft faults on process startup and
5014  * immediately after an mmap.
5015  */
5016 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
5017
5018 void
5019 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
5020                     vm_object_t object, vm_pindex_t pindex,
5021                     vm_size_t size, int limit)
5022 {
5023         struct rb_vm_page_scan_info info;
5024         struct lwp *lp;
5025         vm_size_t psize;
5026
5027         /*
5028          * We can't preinit if read access isn't set or there is no pmap
5029          * or object.
5030          */
5031         if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
5032                 return;
5033
5034         /*
5035          * We can't preinit if the pmap is not the current pmap
5036          */
5037         lp = curthread->td_lwp;
5038         if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
5039                 return;
5040
5041         /*
5042          * Misc additional checks
5043          */
5044         psize = x86_64_btop(size);
5045
5046         if ((object->type != OBJT_VNODE) ||
5047                 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
5048                         (object->resident_page_count > MAX_INIT_PT))) {
5049                 return;
5050         }
5051
5052         if (pindex + psize > object->size) {
5053                 if (object->size < pindex)
5054                         return;
5055                 psize = object->size - pindex;
5056         }
5057
5058         if (psize == 0)
5059                 return;
5060
5061         /*
5062          * If everything is segment-aligned do not pre-init here.  Instead
5063          * allow the normal vm_fault path to pass a segment hint to
5064          * pmap_enter() which will then use an object-referenced shared
5065          * page table page.
5066          */
5067         if ((addr & SEG_MASK) == 0 &&
5068             (ctob(psize) & SEG_MASK) == 0 &&
5069             (ctob(pindex) & SEG_MASK) == 0) {
5070                 return;
5071         }
5072
5073         /*
5074          * Use a red-black scan to traverse the requested range and load
5075          * any valid pages found into the pmap.
5076          *
5077          * We cannot safely scan the object's memq without holding the
5078          * object token.
5079          */
5080         info.start_pindex = pindex;
5081         info.end_pindex = pindex + psize - 1;
5082         info.limit = limit;
5083         info.mpte = NULL;
5084         info.addr = addr;
5085         info.pmap = pmap;
5086
5087         vm_object_hold_shared(object);
5088         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
5089                                 pmap_object_init_pt_callback, &info);
5090         vm_object_drop(object);
5091 }
5092
5093 static
5094 int
5095 pmap_object_init_pt_callback(vm_page_t p, void *data)
5096 {
5097         struct rb_vm_page_scan_info *info = data;
5098         vm_pindex_t rel_index;
5099
5100         /*
5101          * don't allow an madvise to blow away our really
5102          * free pages allocating pv entries.
5103          */
5104         if ((info->limit & MAP_PREFAULT_MADVISE) &&
5105                 vmstats.v_free_count < vmstats.v_free_reserved) {
5106                     return(-1);
5107         }
5108
5109         /*
5110          * Ignore list markers and ignore pages we cannot instantly
5111          * busy (while holding the object token).
5112          */
5113         if (p->flags & PG_MARKER)
5114                 return 0;
5115         if (vm_page_busy_try(p, TRUE))
5116                 return 0;
5117         if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
5118             (p->flags & PG_FICTITIOUS) == 0) {
5119                 if ((p->queue - p->pc) == PQ_CACHE)
5120                         vm_page_deactivate(p);
5121                 rel_index = p->pindex - info->start_pindex;
5122                 pmap_enter_quick(info->pmap,
5123                                  info->addr + x86_64_ptob(rel_index), p);
5124         }
5125         vm_page_wakeup(p);
5126         lwkt_yield();
5127         return(0);
5128 }
5129
5130 /*
5131  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
5132  * address.
5133  *
5134  * Returns FALSE if it would be non-trivial or if a pte is already loaded
5135  * into the slot.
5136  *
5137  * XXX This is safe only because page table pages are not freed.
5138  */
5139 int
5140 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
5141 {
5142         pt_entry_t *pte;
5143
5144         /*spin_lock(&pmap->pm_spin);*/
5145         if ((pte = pmap_pte(pmap, addr)) != NULL) {
5146                 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
5147                         /*spin_unlock(&pmap->pm_spin);*/
5148                         return FALSE;
5149                 }
5150         }
5151         /*spin_unlock(&pmap->pm_spin);*/
5152         return TRUE;
5153 }
5154
5155 /*
5156  * Change the wiring attribute for a pmap/va pair.  The mapping must already
5157  * exist in the pmap.  The mapping may or may not be managed.  The wiring in
5158  * the page is not changed, the page is returned so the caller can adjust
5159  * its wiring (the page is not locked in any way).
5160  *
5161  * Wiring is not a hardware characteristic so there is no need to invalidate
5162  * TLB.  However, in an SMP environment we must use a locked bus cycle to
5163  * update the pte (if we are not using the pmap_inval_*() API that is)...
5164  * it's ok to do this for simple wiring changes.
5165  */
5166 vm_page_t
5167 pmap_unwire(pmap_t pmap, vm_offset_t va)
5168 {
5169         pt_entry_t *ptep;
5170         pv_entry_t pt_pv;
5171         vm_paddr_t pa;
5172         vm_page_t m;
5173
5174         if (pmap == NULL)
5175                 return NULL;
5176
5177         /*
5178          * Assume elements in the kernel pmap are stable
5179          */
5180         if (pmap == &kernel_pmap) {
5181                 if (pmap_pt(pmap, va) == 0)
5182                         return NULL;
5183                 ptep = pmap_pte_quick(pmap, va);
5184                 if (pmap_pte_v(pmap, ptep)) {
5185                         if (pmap_pte_w(pmap, ptep))
5186                                 atomic_add_long(&pmap->pm_stats.wired_count,-1);
5187                         atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5188                         pa = *ptep & PG_FRAME;
5189                         m = PHYS_TO_VM_PAGE(pa);
5190                 } else {
5191                         m = NULL;
5192                 }
5193         } else {
5194                 /*
5195                  * We can only [un]wire pmap-local pages (we cannot wire
5196                  * shared pages)
5197                  */
5198                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
5199                 if (pt_pv == NULL)
5200                         return NULL;
5201
5202                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
5203                 if ((*ptep & pmap->pmap_bits[PG_V_IDX]) == 0) {
5204                         pv_put(pt_pv);
5205                         return NULL;
5206                 }
5207
5208                 if (pmap_pte_w(pmap, ptep)) {
5209                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.wired_count,
5210                                         -1);
5211                 }
5212                 /* XXX else return NULL so caller doesn't unwire m ? */
5213
5214                 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5215
5216                 pa = *ptep & PG_FRAME;
5217                 m = PHYS_TO_VM_PAGE(pa);        /* held by wired count */
5218                 pv_put(pt_pv);
5219         }
5220         return m;
5221 }
5222
5223 /*
5224  * Copy the range specified by src_addr/len from the source map to
5225  * the range dst_addr/len in the destination map.
5226  *
5227  * This routine is only advisory and need not do anything.
5228  */
5229 void
5230 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
5231           vm_size_t len, vm_offset_t src_addr)
5232 {
5233 }
5234
5235 /*
5236  * pmap_zero_page:
5237  *
5238  *      Zero the specified physical page.
5239  *
5240  *      This function may be called from an interrupt and no locking is
5241  *      required.
5242  */
5243 void
5244 pmap_zero_page(vm_paddr_t phys)
5245 {
5246         vm_offset_t va = PHYS_TO_DMAP(phys);
5247
5248         pagezero((void *)va);
5249 }
5250
5251 /*
5252  * pmap_zero_page:
5253  *
5254  *      Zero part of a physical page by mapping it into memory and clearing
5255  *      its contents with bzero.
5256  *
5257  *      off and size may not cover an area beyond a single hardware page.
5258  */
5259 void
5260 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
5261 {
5262         vm_offset_t virt = PHYS_TO_DMAP(phys);
5263
5264         bzero((char *)virt + off, size);
5265 }
5266
5267 /*
5268  * pmap_copy_page:
5269  *
5270  *      Copy the physical page from the source PA to the target PA.
5271  *      This function may be called from an interrupt.  No locking
5272  *      is required.
5273  */
5274 void
5275 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
5276 {
5277         vm_offset_t src_virt, dst_virt;
5278
5279         src_virt = PHYS_TO_DMAP(src);
5280         dst_virt = PHYS_TO_DMAP(dst);
5281         bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
5282 }
5283
5284 /*
5285  * pmap_copy_page_frag:
5286  *
5287  *      Copy the physical page from the source PA to the target PA.
5288  *      This function may be called from an interrupt.  No locking
5289  *      is required.
5290  */
5291 void
5292 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
5293 {
5294         vm_offset_t src_virt, dst_virt;
5295
5296         src_virt = PHYS_TO_DMAP(src);
5297         dst_virt = PHYS_TO_DMAP(dst);
5298
5299         bcopy((char *)src_virt + (src & PAGE_MASK),
5300               (char *)dst_virt + (dst & PAGE_MASK),
5301               bytes);
5302 }
5303
5304 /*
5305  * Returns true if the pmap's pv is one of the first 16 pvs linked to from
5306  * this page.  This count may be changed upwards or downwards in the future;
5307  * it is only necessary that true be returned for a small subset of pmaps
5308  * for proper page aging.
5309  */
5310 boolean_t
5311 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
5312 {
5313         pv_entry_t pv;
5314         int loops = 0;
5315
5316         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5317                 return FALSE;
5318
5319         vm_page_spin_lock(m);
5320         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5321                 if (pv->pv_pmap == pmap) {
5322                         vm_page_spin_unlock(m);
5323                         return TRUE;
5324                 }
5325                 loops++;
5326                 if (loops >= 16)
5327                         break;
5328         }
5329         vm_page_spin_unlock(m);
5330         return (FALSE);
5331 }
5332
5333 /*
5334  * Remove all pages from specified address space this aids process exit
5335  * speeds.  Also, this code may be special cased for the current process
5336  * only.
5337  */
5338 void
5339 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5340 {
5341         pmap_remove_noinval(pmap, sva, eva);
5342         cpu_invltlb();
5343 }
5344
5345 /*
5346  * pmap_testbit tests bits in pte's note that the testbit/clearbit
5347  * routines are inline, and a lot of things compile-time evaluate.
5348  */
5349 static
5350 boolean_t
5351 pmap_testbit(vm_page_t m, int bit)
5352 {
5353         pv_entry_t pv;
5354         pt_entry_t *pte;
5355         pmap_t pmap;
5356
5357         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5358                 return FALSE;
5359
5360         if (TAILQ_FIRST(&m->md.pv_list) == NULL)
5361                 return FALSE;
5362         vm_page_spin_lock(m);
5363         if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
5364                 vm_page_spin_unlock(m);
5365                 return FALSE;
5366         }
5367
5368         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5369
5370 #if defined(PMAP_DIAGNOSTIC)
5371                 if (pv->pv_pmap == NULL) {
5372                         kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
5373                             pv->pv_pindex);
5374                         continue;
5375                 }
5376 #endif
5377                 pmap = pv->pv_pmap;
5378
5379                 /*
5380                  * If the bit being tested is the modified bit, then
5381                  * mark clean_map and ptes as never
5382                  * modified.
5383                  *
5384                  * WARNING!  Because we do not lock the pv, *pte can be in a
5385                  *           state of flux.  Despite this the value of *pte
5386                  *           will still be related to the vm_page in some way
5387                  *           because the pv cannot be destroyed as long as we
5388                  *           hold the vm_page spin lock.
5389                  */
5390                 if (bit == PG_A_IDX || bit == PG_M_IDX) {
5391                                 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5392                         if (!pmap_track_modified(pv->pv_pindex))
5393                                 continue;
5394                 }
5395
5396                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5397                 if (*pte & pmap->pmap_bits[bit]) {
5398                         vm_page_spin_unlock(m);
5399                         return TRUE;
5400                 }
5401         }
5402         vm_page_spin_unlock(m);
5403         return (FALSE);
5404 }
5405
5406 /*
5407  * This routine is used to modify bits in ptes.  Only one bit should be
5408  * specified.  PG_RW requires special handling.
5409  *
5410  * Caller must NOT hold any spin locks
5411  */
5412 static __inline
5413 void
5414 pmap_clearbit(vm_page_t m, int bit_index)
5415 {
5416         pv_entry_t pv;
5417         pt_entry_t *pte;
5418         pt_entry_t pbits;
5419         pmap_t pmap;
5420
5421         if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5422                 if (bit_index == PG_RW_IDX)
5423                         vm_page_flag_clear(m, PG_WRITEABLE);
5424                 return;
5425         }
5426
5427         /*
5428          * PG_M or PG_A case
5429          *
5430          * Loop over all current mappings setting/clearing as appropos If
5431          * setting RO do we need to clear the VAC?
5432          *
5433          * NOTE: When clearing PG_M we could also (not implemented) drop
5434          *       through to the PG_RW code and clear PG_RW too, forcing
5435          *       a fault on write to redetect PG_M for virtual kernels, but
5436          *       it isn't necessary since virtual kernels invalidate the
5437          *       pte when they clear the VPTE_M bit in their virtual page
5438          *       tables.
5439          *
5440          * NOTE: Does not re-dirty the page when clearing only PG_M.
5441          *
5442          * NOTE: Because we do not lock the pv, *pte can be in a state of
5443          *       flux.  Despite this the value of *pte is still somewhat
5444          *       related while we hold the vm_page spin lock.
5445          *
5446          *       *pte can be zero due to this race.  Since we are clearing
5447          *       bits we basically do no harm when this race occurs.
5448          */
5449         if (bit_index != PG_RW_IDX) {
5450                 vm_page_spin_lock(m);
5451                 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5452 #if defined(PMAP_DIAGNOSTIC)
5453                         if (pv->pv_pmap == NULL) {
5454                                 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5455                                     pv->pv_pindex);
5456                                 continue;
5457                         }
5458 #endif
5459                         pmap = pv->pv_pmap;
5460                         pte = pmap_pte_quick(pv->pv_pmap,
5461                                              pv->pv_pindex << PAGE_SHIFT);
5462                         pbits = *pte;
5463                         if (pbits & pmap->pmap_bits[bit_index])
5464                                 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5465                 }
5466                 vm_page_spin_unlock(m);
5467                 return;
5468         }
5469
5470         /*
5471          * Clear PG_RW.  Also clears PG_M and marks the page dirty if PG_M
5472          * was set.
5473          */
5474 restart:
5475         vm_page_spin_lock(m);
5476         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5477                 /*
5478                  * don't write protect pager mappings
5479                  */
5480                 if (!pmap_track_modified(pv->pv_pindex))
5481                         continue;
5482
5483 #if defined(PMAP_DIAGNOSTIC)
5484                 if (pv->pv_pmap == NULL) {
5485                         kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5486                                 pv->pv_pindex);
5487                         continue;
5488                 }
5489 #endif
5490                 pmap = pv->pv_pmap;
5491
5492                 /*
5493                  * Skip pages which do not have PG_RW set.
5494                  */
5495                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5496                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5497                         continue;
5498
5499                 /*
5500                  * We must lock the PV to be able to safely test the pte.
5501                  */
5502                 if (pv_hold_try(pv)) {
5503                         vm_page_spin_unlock(m);
5504                 } else {
5505                         vm_page_spin_unlock(m);
5506                         pv_lock(pv);    /* held, now do a blocking lock */
5507                         pv_put(pv);
5508                         goto restart;
5509                 }
5510
5511                 /*
5512                  * Reload pte after acquiring pv.
5513                  */
5514                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5515 #if 0
5516                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0) {
5517                         pv_put(pv);
5518                         goto restart;
5519                 }
5520 #endif
5521
5522                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
5523                 for (;;) {
5524                         pt_entry_t nbits;
5525
5526                         pbits = *pte;
5527                         cpu_ccfence();
5528                         nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5529                                           pmap->pmap_bits[PG_M_IDX]);
5530                         if (pmap_inval_smp_cmpset(pmap,
5531                                      ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5532                                      pte, pbits, nbits)) {
5533                                 break;
5534                         }
5535                         cpu_pause();
5536                 }
5537
5538                 /*
5539                  * If PG_M was found to be set while we were clearing PG_RW
5540                  * we also clear PG_M (done above) and mark the page dirty.
5541                  * Callers expect this behavior.
5542                  *
5543                  * we lost pv so it cannot be used as an iterator.  In fact,
5544                  * because we couldn't necessarily lock it atomically it may
5545                  * have moved within the list and ALSO cannot be used as an
5546                  * iterator.
5547                  */
5548                 vm_page_spin_lock(m);
5549                 if (pbits & pmap->pmap_bits[PG_M_IDX])
5550                         vm_page_dirty(m);
5551                 vm_page_spin_unlock(m);
5552                 pv_put(pv);
5553                 goto restart;
5554         }
5555         if (bit_index == PG_RW_IDX)
5556                 vm_page_flag_clear(m, PG_WRITEABLE);
5557         vm_page_spin_unlock(m);
5558 }
5559
5560 /*
5561  * Lower the permission for all mappings to a given page.
5562  *
5563  * Page must be busied by caller.  Because page is busied by caller this
5564  * should not be able to race a pmap_enter().
5565  */
5566 void
5567 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5568 {
5569         /* JG NX support? */
5570         if ((prot & VM_PROT_WRITE) == 0) {
5571                 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5572                         /*
5573                          * NOTE: pmap_clearbit(.. PG_RW) also clears
5574                          *       the PG_WRITEABLE flag in (m).
5575                          */
5576                         pmap_clearbit(m, PG_RW_IDX);
5577                 } else {
5578                         pmap_remove_all(m);
5579                 }
5580         }
5581 }
5582
5583 vm_paddr_t
5584 pmap_phys_address(vm_pindex_t ppn)
5585 {
5586         return (x86_64_ptob(ppn));
5587 }
5588
5589 /*
5590  * Return a count of reference bits for a page, clearing those bits.
5591  * It is not necessary for every reference bit to be cleared, but it
5592  * is necessary that 0 only be returned when there are truly no
5593  * reference bits set.
5594  *
5595  * XXX: The exact number of bits to check and clear is a matter that
5596  * should be tested and standardized at some point in the future for
5597  * optimal aging of shared pages.
5598  *
5599  * This routine may not block.
5600  */
5601 int
5602 pmap_ts_referenced(vm_page_t m)
5603 {
5604         pv_entry_t pv;
5605         pt_entry_t *pte;
5606         pmap_t pmap;
5607         int rtval = 0;
5608
5609         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5610                 return (rtval);
5611
5612         vm_page_spin_lock(m);
5613         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5614                 if (!pmap_track_modified(pv->pv_pindex))
5615                         continue;
5616                 pmap = pv->pv_pmap;
5617                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5618                 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5619                         atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5620                         rtval++;
5621                         if (rtval > 4)
5622                                 break;
5623                 }
5624         }
5625         vm_page_spin_unlock(m);
5626         return (rtval);
5627 }
5628
5629 /*
5630  *      pmap_is_modified:
5631  *
5632  *      Return whether or not the specified physical page was modified
5633  *      in any physical maps.
5634  */
5635 boolean_t
5636 pmap_is_modified(vm_page_t m)
5637 {
5638         boolean_t res;
5639
5640         res = pmap_testbit(m, PG_M_IDX);
5641         return (res);
5642 }
5643
5644 /*
5645  *      Clear the modify bits on the specified physical page.
5646  */
5647 void
5648 pmap_clear_modify(vm_page_t m)
5649 {
5650         pmap_clearbit(m, PG_M_IDX);
5651 }
5652
5653 /*
5654  *      pmap_clear_reference:
5655  *
5656  *      Clear the reference bit on the specified physical page.
5657  */
5658 void
5659 pmap_clear_reference(vm_page_t m)
5660 {
5661         pmap_clearbit(m, PG_A_IDX);
5662 }
5663
5664 /*
5665  * Miscellaneous support routines follow
5666  */
5667
5668 static
5669 void
5670 i386_protection_init(void)
5671 {
5672         int *kp, prot;
5673
5674         /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit  */
5675         kp = protection_codes;
5676         for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5677                 switch (prot) {
5678                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5679                         /*
5680                          * Read access is also 0. There isn't any execute bit,
5681                          * so just make it readable.
5682                          */
5683                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5684                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5685                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5686                         *kp++ = 0;
5687                         break;
5688                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5689                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5690                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5691                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5692                         *kp++ = pmap_bits_default[PG_RW_IDX];
5693                         break;
5694                 }
5695         }
5696 }
5697
5698 /*
5699  * Map a set of physical memory pages into the kernel virtual
5700  * address space. Return a pointer to where it is mapped. This
5701  * routine is intended to be used for mapping device memory,
5702  * NOT real memory.
5703  *
5704  * NOTE: We can't use pgeflag unless we invalidate the pages one at
5705  *       a time.
5706  *
5707  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5708  *       work whether the cpu supports PAT or not.  The remaining PAT
5709  *       attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5710  *       supports PAT.
5711  */
5712 void *
5713 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5714 {
5715         return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5716 }
5717
5718 void *
5719 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5720 {
5721         return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5722 }
5723
5724 void *
5725 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5726 {
5727         return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5728 }
5729
5730 /*
5731  * Map a set of physical memory pages into the kernel virtual
5732  * address space. Return a pointer to where it is mapped. This
5733  * routine is intended to be used for mapping device memory,
5734  * NOT real memory.
5735  */
5736 void *
5737 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5738 {
5739         vm_offset_t va, tmpva, offset;
5740         pt_entry_t *pte;
5741         vm_size_t tmpsize;
5742
5743         offset = pa & PAGE_MASK;
5744         size = roundup(offset + size, PAGE_SIZE);
5745
5746         va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5747         if (va == 0)
5748                 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5749
5750         pa = pa & ~PAGE_MASK;
5751         for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5752                 pte = vtopte(tmpva);
5753                 *pte = pa |
5754                     kernel_pmap.pmap_bits[PG_RW_IDX] |
5755                     kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5756                     kernel_pmap.pmap_cache_bits[mode];
5757                 tmpsize -= PAGE_SIZE;
5758                 tmpva += PAGE_SIZE;
5759                 pa += PAGE_SIZE;
5760         }
5761         pmap_invalidate_range(&kernel_pmap, va, va + size);
5762         pmap_invalidate_cache_range(va, va + size);
5763
5764         return ((void *)(va + offset));
5765 }
5766
5767 void
5768 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5769 {
5770         vm_offset_t base, offset;
5771
5772         base = va & ~PAGE_MASK;
5773         offset = va & PAGE_MASK;
5774         size = roundup(offset + size, PAGE_SIZE);
5775         pmap_qremove(va, size >> PAGE_SHIFT);
5776         kmem_free(&kernel_map, base, size);
5777 }
5778
5779 /*
5780  * Sets the memory attribute for the specified page.
5781  */
5782 void
5783 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5784 {
5785
5786     m->pat_mode = ma;
5787
5788     /*
5789      * If "m" is a normal page, update its direct mapping.  This update
5790      * can be relied upon to perform any cache operations that are
5791      * required for data coherence.
5792      */
5793     if ((m->flags & PG_FICTITIOUS) == 0)
5794         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5795 }
5796
5797 /*
5798  * Change the PAT attribute on an existing kernel memory map.  Caller
5799  * must ensure that the virtual memory in question is not accessed
5800  * during the adjustment.
5801  */
5802 void
5803 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5804 {
5805         pt_entry_t *pte;
5806         vm_offset_t base;
5807         int changed = 0;
5808
5809         if (va == 0)
5810                 panic("pmap_change_attr: va is NULL");
5811         base = trunc_page(va);
5812
5813         while (count) {
5814                 pte = vtopte(va);
5815                 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5816                        kernel_pmap.pmap_cache_bits[mode];
5817                 --count;
5818                 va += PAGE_SIZE;
5819         }
5820
5821         changed = 1;    /* XXX: not optimal */
5822
5823         /*
5824          * Flush CPU caches if required to make sure any data isn't cached that
5825          * shouldn't be, etc.
5826          */
5827         if (changed) {
5828                 pmap_invalidate_range(&kernel_pmap, base, va);
5829                 pmap_invalidate_cache_range(base, va);
5830         }
5831 }
5832
5833 /*
5834  * perform the pmap work for mincore
5835  */
5836 int
5837 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5838 {
5839         pt_entry_t *ptep, pte;
5840         vm_page_t m;
5841         int val = 0;
5842
5843         ptep = pmap_pte(pmap, addr);
5844
5845         if (ptep && (pte = *ptep) != 0) {
5846                 vm_offset_t pa;
5847
5848                 val = MINCORE_INCORE;
5849                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5850                         goto done;
5851
5852                 pa = pte & PG_FRAME;
5853
5854                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5855                         m = NULL;
5856                 else
5857                         m = PHYS_TO_VM_PAGE(pa);
5858
5859                 /*
5860                  * Modified by us
5861                  */
5862                 if (pte & pmap->pmap_bits[PG_M_IDX])
5863                         val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5864                 /*
5865                  * Modified by someone
5866                  */
5867                 else if (m && (m->dirty || pmap_is_modified(m)))
5868                         val |= MINCORE_MODIFIED_OTHER;
5869                 /*
5870                  * Referenced by us
5871                  */
5872                 if (pte & pmap->pmap_bits[PG_A_IDX])
5873                         val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5874
5875                 /*
5876                  * Referenced by someone
5877                  */
5878                 else if (m && ((m->flags & PG_REFERENCED) ||
5879                                 pmap_ts_referenced(m))) {
5880                         val |= MINCORE_REFERENCED_OTHER;
5881                         vm_page_flag_set(m, PG_REFERENCED);
5882                 }
5883         }
5884 done:
5885
5886         return val;
5887 }
5888
5889 /*
5890  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
5891  * vmspace will be ref'd and the old one will be deref'd.
5892  *
5893  * The vmspace for all lwps associated with the process will be adjusted
5894  * and cr3 will be reloaded if any lwp is the current lwp.
5895  *
5896  * The process must hold the vmspace->vm_map.token for oldvm and newvm
5897  */
5898 void
5899 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5900 {
5901         struct vmspace *oldvm;
5902         struct lwp *lp;
5903
5904         oldvm = p->p_vmspace;
5905         if (oldvm != newvm) {
5906                 if (adjrefs)
5907                         vmspace_ref(newvm);
5908                 p->p_vmspace = newvm;
5909                 KKASSERT(p->p_nthreads == 1);
5910                 lp = RB_ROOT(&p->p_lwp_tree);
5911                 pmap_setlwpvm(lp, newvm);
5912                 if (adjrefs)
5913                         vmspace_rel(oldvm);
5914         }
5915 }
5916
5917 /*
5918  * Set the vmspace for a LWP.  The vmspace is almost universally set the
5919  * same as the process vmspace, but virtual kernels need to swap out contexts
5920  * on a per-lwp basis.
5921  *
5922  * Caller does not necessarily hold any vmspace tokens.  Caller must control
5923  * the lwp (typically be in the context of the lwp).  We use a critical
5924  * section to protect against statclock and hardclock (statistics collection).
5925  */
5926 void
5927 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5928 {
5929         struct vmspace *oldvm;
5930         struct pmap *pmap;
5931
5932         oldvm = lp->lwp_vmspace;
5933
5934         if (oldvm != newvm) {
5935                 crit_enter();
5936                 KKASSERT((newvm->vm_refcnt & VM_REF_DELETED) == 0);
5937                 lp->lwp_vmspace = newvm;
5938                 if (curthread->td_lwp == lp) {
5939                         pmap = vmspace_pmap(newvm);
5940                         ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5941                         if (pmap->pm_active_lock & CPULOCK_EXCL)
5942                                 pmap_interlock_wait(newvm);
5943 #if defined(SWTCH_OPTIM_STATS)
5944                         tlb_flush_count++;
5945 #endif
5946                         if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5947                                 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5948                         } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5949                                 curthread->td_pcb->pcb_cr3 = KPML4phys;
5950                         } else {
5951                                 panic("pmap_setlwpvm: unknown pmap type\n");
5952                         }
5953                         load_cr3(curthread->td_pcb->pcb_cr3);
5954                         pmap = vmspace_pmap(oldvm);
5955                         ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5956                                                mycpu->gd_cpuid);
5957                 }
5958                 crit_exit();
5959         }
5960 }
5961
5962 /*
5963  * Called when switching to a locked pmap, used to interlock against pmaps
5964  * undergoing modifications to prevent us from activating the MMU for the
5965  * target pmap until all such modifications have completed.  We have to do
5966  * this because the thread making the modifications has already set up its
5967  * SMP synchronization mask.
5968  *
5969  * This function cannot sleep!
5970  *
5971  * No requirements.
5972  */
5973 void
5974 pmap_interlock_wait(struct vmspace *vm)
5975 {
5976         struct pmap *pmap = &vm->vm_pmap;
5977
5978         if (pmap->pm_active_lock & CPULOCK_EXCL) {
5979                 crit_enter();
5980                 KKASSERT(curthread->td_critcount >= 2);
5981                 DEBUG_PUSH_INFO("pmap_interlock_wait");
5982                 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5983                         cpu_ccfence();
5984                         lwkt_process_ipiq();
5985                 }
5986                 DEBUG_POP_INFO();
5987                 crit_exit();
5988         }
5989 }
5990
5991 vm_offset_t
5992 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5993 {
5994
5995         if ((obj == NULL) || (size < NBPDR) ||
5996             ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5997                 return addr;
5998         }
5999
6000         addr = roundup2(addr, NBPDR);
6001         return addr;
6002 }
6003
6004 /*
6005  * Used by kmalloc/kfree, page already exists at va
6006  */
6007 vm_page_t
6008 pmap_kvtom(vm_offset_t va)
6009 {
6010         pt_entry_t *ptep = vtopte(va);
6011
6012         KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
6013         return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
6014 }
6015
6016 /*
6017  * Initialize machine-specific shared page directory support.  This
6018  * is executed when a VM object is created.
6019  */
6020 void
6021 pmap_object_init(vm_object_t object)
6022 {
6023         object->md.pmap_rw = NULL;
6024         object->md.pmap_ro = NULL;
6025 }
6026
6027 /*
6028  * Clean up machine-specific shared page directory support.  This
6029  * is executed when a VM object is destroyed.
6030  */
6031 void
6032 pmap_object_free(vm_object_t object)
6033 {
6034         pmap_t pmap;
6035
6036         if ((pmap = object->md.pmap_rw) != NULL) {
6037                 object->md.pmap_rw = NULL;
6038                 pmap_remove_noinval(pmap,
6039                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6040                 CPUMASK_ASSZERO(pmap->pm_active);
6041                 pmap_release(pmap);
6042                 pmap_puninit(pmap);
6043                 kfree(pmap, M_OBJPMAP);
6044         }
6045         if ((pmap = object->md.pmap_ro) != NULL) {
6046                 object->md.pmap_ro = NULL;
6047                 pmap_remove_noinval(pmap,
6048                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6049                 CPUMASK_ASSZERO(pmap->pm_active);
6050                 pmap_release(pmap);
6051                 pmap_puninit(pmap);
6052                 kfree(pmap, M_OBJPMAP);
6053         }
6054 }
6055
6056 /*
6057  * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
6058  * VM page and issue a pginfo->callback.
6059  *
6060  * We are expected to dispose of any non-NULL pte_pv.
6061  */
6062 static
6063 void
6064 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
6065                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
6066                       pv_entry_t pt_pv, int sharept,
6067                       vm_offset_t va, pt_entry_t *ptep, void *arg)
6068 {
6069         struct pmap_pgscan_info *pginfo = arg;
6070         vm_page_t m;
6071
6072         if (pte_pv) {
6073                 /*
6074                  * Try to busy the page while we hold the pte_pv locked.
6075                  */
6076                 KKASSERT(pte_pv->pv_m);
6077                 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
6078                 if (vm_page_busy_try(m, TRUE) == 0) {
6079                         if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
6080                                 /*
6081                                  * The callback is issued with the pte_pv
6082                                  * unlocked and put away, and the pt_pv
6083                                  * unlocked.
6084                                  */
6085                                 pv_put(pte_pv);
6086                                 if (pt_pv) {
6087                                         vm_page_wire_quick(pt_pv->pv_m);
6088                                         pv_unlock(pt_pv);
6089                                 }
6090                                 if (pginfo->callback(pginfo, va, m) < 0)
6091                                         info->stop = 1;
6092                                 if (pt_pv) {
6093                                         pv_lock(pt_pv);
6094                                         vm_page_unwire_quick(pt_pv->pv_m);
6095                                 }
6096                         } else {
6097                                 vm_page_wakeup(m);
6098                                 pv_put(pte_pv);
6099                         }
6100                 } else {
6101                         ++pginfo->busycount;
6102                         pv_put(pte_pv);
6103                 }
6104         } else {
6105                 /*
6106                  * Shared page table or unmanaged page (sharept or !sharept)
6107                  */
6108                 pv_placemarker_wakeup(pmap, pte_placemark);
6109         }
6110 }
6111
6112 void
6113 pmap_pgscan(struct pmap_pgscan_info *pginfo)
6114 {
6115         struct pmap_scan_info info;
6116
6117         pginfo->offset = pginfo->beg_addr;
6118         info.pmap = pginfo->pmap;
6119         info.sva = pginfo->beg_addr;
6120         info.eva = pginfo->end_addr;
6121         info.func = pmap_pgscan_callback;
6122         info.arg = pginfo;
6123         pmap_scan(&info, 0);
6124         if (info.stop == 0)
6125                 pginfo->offset = pginfo->end_addr;
6126 }
6127
6128 /*
6129  * Wait for a placemarker that we do not own to clear.  The placemarker
6130  * in question is not necessary set to the pindex we want, we may have
6131  * to wait on the element because we want to reserve it ourselves.
6132  */
6133 static
6134 void
6135 pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark)
6136 {
6137         spin_lock(&pmap->pm_spin);
6138         if (*pmark != PM_NOPLACEMARK) {
6139                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
6140                 ssleep(pmark, &pmap->pm_spin, 0, "pvplw", 0);
6141         }
6142         spin_unlock(&pmap->pm_spin);
6143 }
6144
6145 /*
6146  * Wakeup a placemarker that we own.  Replace the entry with
6147  * PM_NOPLACEMARK and issue a wakeup() if necessary.
6148  */
6149 static
6150 void
6151 pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark)
6152 {
6153         vm_pindex_t pindex;
6154
6155         spin_lock(&pmap->pm_spin);
6156         pindex = atomic_swap_long(pmark, PM_NOPLACEMARK);
6157         spin_unlock(&pmap->pm_spin);
6158         KKASSERT(pindex != PM_NOPLACEMARK);
6159         if (pindex & PM_PLACEMARK_WAKEUP)
6160                 wakeup(pmark);
6161 }