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