sys/platform/pc64/x86_64/pmap.c

   1 /*
   2  * Copyright (c) 1991 Regents of the University of California.
   3  * Copyright (c) 1994 John S. Dyson
   4  * Copyright (c) 1994 David Greenman
   5  * Copyright (c) 2003 Peter Wemm
   6  * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
   7  * Copyright (c) 2008, 2009 The DragonFly Project.
   8  * Copyright (c) 2008, 2009 Jordan Gordeev.
   9  * Copyright (c) 2011-2012 Matthew Dillon
  10  * All rights reserved.
  11  *
  12  * This code is derived from software contributed to Berkeley by
  13  * the Systems Programming Group of the University of Utah Computer
  14  * Science Department and William Jolitz of UUNET Technologies Inc.
  15  *
  16  * Redistribution and use in source and binary forms, with or without
  17  * modification, are permitted provided that the following conditions
  18  * are met:
  19  * 1. Redistributions of source code must retain the above copyright
  20  *    notice, this list of conditions and the following disclaimer.
  21  * 2. Redistributions in binary form must reproduce the above copyright
  22  *    notice, this list of conditions and the following disclaimer in the
  23  *    documentation and/or other materials provided with the distribution.
  24  * 3. All advertising materials mentioning features or use of this software
  25  *    must display the following acknowledgement:
  26  *      This product includes software developed by the University of
  27  *      California, Berkeley and its contributors.
  28  * 4. Neither the name of the University nor the names of its contributors
  29  *    may be used to endorse or promote products derived from this software
  30  *    without specific prior written permission.
  31  *
  32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  42  * SUCH DAMAGE.
  43  */
  44 /*
  45  * Manage physical address maps for x86-64 systems.
  46  */
  47
  48 #if 0 /* JG */
  49 #include "opt_disable_pse.h"
  50 #include "opt_pmap.h"
  51 #endif
  52 #include "opt_msgbuf.h"
  53
  54 #include <sys/param.h>
  55 #include <sys/kernel.h>
  56 #include <sys/proc.h>
  57 #include <sys/msgbuf.h>
  58 #include <sys/vmmeter.h>
  59 #include <sys/mman.h>
  60 #include <sys/systm.h>
  61
  62 #include <vm/vm.h>
  63 #include <vm/vm_param.h>
  64 #include <sys/sysctl.h>
  65 #include <sys/lock.h>
  66 #include <vm/vm_kern.h>
  67 #include <vm/vm_page.h>
  68 #include <vm/vm_map.h>
  69 #include <vm/vm_object.h>
  70 #include <vm/vm_extern.h>
  71 #include <vm/vm_pageout.h>
  72 #include <vm/vm_pager.h>
  73 #include <vm/vm_zone.h>
  74
  75 #include <sys/user.h>
  76 #include <sys/thread2.h>
  77 #include <sys/sysref2.h>
  78 #include <sys/spinlock2.h>
  79 #include <vm/vm_page2.h>
  80
  81 #include <machine/cputypes.h>
  82 #include <machine/md_var.h>
  83 #include <machine/specialreg.h>
  84 #include <machine/smp.h>
  85 #include <machine_base/apic/apicreg.h>
  86 #include <machine/globaldata.h>
  87 #include <machine/pmap.h>
  88 #include <machine/pmap_inval.h>
  89 #include <machine/inttypes.h>
  90
  91 #include <ddb/ddb.h>
  92
  93 #define PMAP_KEEP_PDIRS
  94 #ifndef PMAP_SHPGPERPROC
  95 #define PMAP_SHPGPERPROC 2000
  96 #endif
  97
  98 #if defined(DIAGNOSTIC)
  99 #define PMAP_DIAGNOSTIC
 100 #endif
 101
 102 #define MINPV 2048
 103
 104 /*
 105  * pmap debugging will report who owns a pv lock when blocking.
 106  */
 107 #ifdef PMAP_DEBUG
 108
 109 #define PMAP_DEBUG_DECL         ,const char *func, int lineno
 110 #define PMAP_DEBUG_ARGS         , __func__, __LINE__
 111 #define PMAP_DEBUG_COPY         , func, lineno
 112
 113 #define pv_get(pmap, pindex, pmarkp)    _pv_get(pmap, pindex, pmarkp    \
 114                                                         PMAP_DEBUG_ARGS)
 115 #define pv_lock(pv)                     _pv_lock(pv                     \
 116                                                         PMAP_DEBUG_ARGS)
 117 #define pv_hold_try(pv)                 _pv_hold_try(pv                 \
 118                                                         PMAP_DEBUG_ARGS)
 119 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp  \
 120                                                         PMAP_DEBUG_ARGS)
 121
 122 #define pv_free(pv, pvp)                _pv_free(pv, pvp PMAP_DEBUG_ARGS)
 123
 124 #else
 125
 126 #define PMAP_DEBUG_DECL
 127 #define PMAP_DEBUG_ARGS
 128 #define PMAP_DEBUG_COPY
 129
 130 #define pv_get(pmap, pindex, pmarkp)            _pv_get(pmap, pindex, pmarkp)
 131 #define pv_lock(pv)                     _pv_lock(pv)
 132 #define pv_hold_try(pv)                 _pv_hold_try(pv)
 133 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp)
 134 #define pv_free(pv, pvp)                _pv_free(pv, pvp)
 135
 136 #endif
 137
 138 /*
 139  * Get PDEs and PTEs for user/kernel address space
 140  */
 141 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
 142
 143 #define pmap_pde_v(pmap, pte)           ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 144 #define pmap_pte_w(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
 145 #define pmap_pte_m(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
 146 #define pmap_pte_u(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
 147 #define pmap_pte_v(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 148
 149 /*
 150  * Given a map and a machine independent protection code,
 151  * convert to a vax protection code.
 152  */
 153 #define pte_prot(m, p)          \
 154         (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
 155 static int protection_codes[PROTECTION_CODES_SIZE];
 156
 157 struct pmap kernel_pmap;
 158
 159 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
 160
 161 vm_paddr_t avail_start;         /* PA of first available physical page */
 162 vm_paddr_t avail_end;           /* PA of last available physical page */
 163 vm_offset_t virtual2_start;     /* cutout free area prior to kernel start */
 164 vm_offset_t virtual2_end;
 165 vm_offset_t virtual_start;      /* VA of first avail page (after kernel bss) */
 166 vm_offset_t virtual_end;        /* VA of last avail page (end of kernel AS) */
 167 vm_offset_t KvaStart;           /* VA start of KVA space */
 168 vm_offset_t KvaEnd;             /* VA end of KVA space (non-inclusive) */
 169 vm_offset_t KvaSize;            /* max size of kernel virtual address space */
 170 static boolean_t pmap_initialized = FALSE;      /* Has pmap_init completed? */
 171 //static int pgeflag;           /* PG_G or-in */
 172 //static int pseflag;           /* PG_PS or-in */
 173 uint64_t PatMsr;
 174
 175 static int ndmpdp;
 176 static vm_paddr_t dmaplimit;
 177 static int nkpt;
 178 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
 179
 180 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE];        /* PAT -> PG_ bits */
 181 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/    /* PAT -> PG_ bits */
 182
 183 static uint64_t KPTbase;
 184 static uint64_t KPTphys;
 185 static uint64_t KPDphys;        /* phys addr of kernel level 2 */
 186 static uint64_t KPDbase;        /* phys addr of kernel level 2 @ KERNBASE */
 187 uint64_t KPDPphys;      /* phys addr of kernel level 3 */
 188 uint64_t KPML4phys;     /* phys addr of kernel level 4 */
 189
 190 static uint64_t DMPDphys;       /* phys addr of direct mapped level 2 */
 191 static uint64_t DMPDPphys;      /* phys addr of direct mapped level 3 */
 192
 193 /*
 194  * Data for the pv entry allocation mechanism
 195  */
 196 static vm_zone_t pvzone;
 197 static struct vm_zone pvzone_store;
 198 static struct vm_object pvzone_obj;
 199 static int pv_entry_max=0, pv_entry_high_water=0;
 200 static int pmap_pagedaemon_waken = 0;
 201 static struct pv_entry *pvinit;
 202
 203 /*
 204  * All those kernel PT submaps that BSD is so fond of
 205  */
 206 pt_entry_t *CMAP1 = NULL, *ptmmap;
 207 caddr_t CADDR1 = NULL, ptvmmap = NULL;
 208 static pt_entry_t *msgbufmap;
 209 struct msgbuf *msgbufp=NULL;
 210
 211 /*
 212  * PMAP default PG_* bits. Needed to be able to add
 213  * EPT/NPT pagetable pmap_bits for the VMM module
 214  */
 215 uint64_t pmap_bits_default[] = {
 216                 REGULAR_PMAP,                                   /* TYPE_IDX             0 */
 217                 X86_PG_V,                                       /* PG_V_IDX             1 */
 218                 X86_PG_RW,                                      /* PG_RW_IDX            2 */
 219                 X86_PG_U,                                       /* PG_U_IDX             3 */
 220                 X86_PG_A,                                       /* PG_A_IDX             4 */
 221                 X86_PG_M,                                       /* PG_M_IDX             5 */
 222                 X86_PG_PS,                                      /* PG_PS_IDX3           6 */
 223                 X86_PG_G,                                       /* PG_G_IDX             7 */
 224                 X86_PG_AVAIL1,                                  /* PG_AVAIL1_IDX        8 */
 225                 X86_PG_AVAIL2,                                  /* PG_AVAIL2_IDX        9 */
 226                 X86_PG_AVAIL3,                                  /* PG_AVAIL3_IDX        10 */
 227                 X86_PG_NC_PWT | X86_PG_NC_PCD,                  /* PG_N_IDX     11 */
 228 };
 229 /*
 230  * Crashdump maps.
 231  */
 232 static pt_entry_t *pt_crashdumpmap;
 233 static caddr_t crashdumpmap;
 234
 235 static int pmap_debug = 0;
 236 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
 237     &pmap_debug, 0, "Debug pmap's");
 238 #ifdef PMAP_DEBUG2
 239 static int pmap_enter_debug = 0;
 240 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
 241     &pmap_enter_debug, 0, "Debug pmap_enter's");
 242 #endif
 243 static int pmap_yield_count = 64;
 244 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
 245     &pmap_yield_count, 0, "Yield during init_pt/release");
 246 static int pmap_mmu_optimize = 0;
 247 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
 248     &pmap_mmu_optimize, 0, "Share page table pages when possible");
 249 int pmap_fast_kernel_cpusync = 0;
 250 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
 251     &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
 252 int pmap_dynamic_delete = -1;
 253 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
 254     &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
 255
 256 #define DISABLE_PSE
 257
 258 /* Standard user access funtions */
 259 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
 260     size_t *lencopied);
 261 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
 262 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
 263 extern int std_fubyte (const uint8_t *base);
 264 extern int std_subyte (uint8_t *base, uint8_t byte);
 265 extern int32_t std_fuword32 (const uint32_t *base);
 266 extern int64_t std_fuword64 (const uint64_t *base);
 267 extern int std_suword64 (uint64_t *base, uint64_t word);
 268 extern int std_suword32 (uint32_t *base, int word);
 269 extern uint32_t std_swapu32 (volatile uint32_t *base, uint32_t v);
 270 extern uint64_t std_swapu64 (volatile uint64_t *base, uint64_t v);
 271
 272 static void pv_hold(pv_entry_t pv);
 273 static int _pv_hold_try(pv_entry_t pv
 274                                 PMAP_DEBUG_DECL);
 275 static void pv_drop(pv_entry_t pv);
 276 static void _pv_lock(pv_entry_t pv
 277                                 PMAP_DEBUG_DECL);
 278 static void pv_unlock(pv_entry_t pv);
 279 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
 280                                 PMAP_DEBUG_DECL);
 281 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp
 282                                 PMAP_DEBUG_DECL);
 283 static void _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL);
 284 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex,
 285                                 vm_pindex_t **pmarkp, int *errorp);
 286 static void pv_put(pv_entry_t pv);
 287 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
 288 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
 289                       pv_entry_t *pvpp);
 290 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
 291                       pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
 292 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
 293                         pmap_inval_bulk_t *bulk, int destroy);
 294 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
 295 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
 296                         pmap_inval_bulk_t *bulk);
 297
 298 struct pmap_scan_info;
 299 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
 300                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
 301                       pv_entry_t pt_pv, int sharept,
 302                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 303 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
 304                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
 305                       pv_entry_t pt_pv, int sharept,
 306                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 307
 308 static void i386_protection_init (void);
 309 static void create_pagetables(vm_paddr_t *firstaddr);
 310 static void pmap_remove_all (vm_page_t m);
 311 static boolean_t pmap_testbit (vm_page_t m, int bit);
 312
 313 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
 314 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
 315
 316 static void pmap_pinit_defaults(struct pmap *pmap);
 317 static void pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark);
 318 static void pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark);
 319
 320 static unsigned pdir4mb;
 321
 322 static int
 323 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
 324 {
 325         if (pv1->pv_pindex < pv2->pv_pindex)
 326                 return(-1);
 327         if (pv1->pv_pindex > pv2->pv_pindex)
 328                 return(1);
 329         return(0);
 330 }
 331
 332 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
 333              pv_entry_compare, vm_pindex_t, pv_pindex);
 334
 335 static __inline
 336 void
 337 pmap_page_stats_adding(vm_page_t m)
 338 {
 339         globaldata_t gd = mycpu;
 340
 341         if (TAILQ_EMPTY(&m->md.pv_list)) {
 342                 ++gd->gd_vmtotal.t_arm;
 343         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 344                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 345                 ++gd->gd_vmtotal.t_armshr;
 346                 ++gd->gd_vmtotal.t_avmshr;
 347         } else {
 348                 ++gd->gd_vmtotal.t_avmshr;
 349         }
 350 }
 351
 352 static __inline
 353 void
 354 pmap_page_stats_deleting(vm_page_t m)
 355 {
 356         globaldata_t gd = mycpu;
 357
 358         if (TAILQ_EMPTY(&m->md.pv_list)) {
 359                 --gd->gd_vmtotal.t_arm;
 360         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 361                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 362                 --gd->gd_vmtotal.t_armshr;
 363                 --gd->gd_vmtotal.t_avmshr;
 364         } else {
 365                 --gd->gd_vmtotal.t_avmshr;
 366         }
 367 }
 368
 369 /*
 370  * Move the kernel virtual free pointer to the next
 371  * 2MB.  This is used to help improve performance
 372  * by using a large (2MB) page for much of the kernel
 373  * (.text, .data, .bss)
 374  */
 375 static
 376 vm_offset_t
 377 pmap_kmem_choose(vm_offset_t addr)
 378 {
 379         vm_offset_t newaddr = addr;
 380
 381         newaddr = roundup2(addr, NBPDR);
 382         return newaddr;
 383 }
 384
 385 /*
 386  * pmap_pte_quick:
 387  *
 388  *      Super fast pmap_pte routine best used when scanning the pv lists.
 389  *      This eliminates many course-grained invltlb calls.  Note that many of
 390  *      the pv list scans are across different pmaps and it is very wasteful
 391  *      to do an entire invltlb when checking a single mapping.
 392  */
 393 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
 394
 395 static
 396 pt_entry_t *
 397 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
 398 {
 399         return pmap_pte(pmap, va);
 400 }
 401
 402 /*
 403  * Returns the pindex of a page table entry (representing a terminal page).
 404  * There are NUPTE_TOTAL page table entries possible (a huge number)
 405  *
 406  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
 407  * We want to properly translate negative KVAs.
 408  */
 409 static __inline
 410 vm_pindex_t
 411 pmap_pte_pindex(vm_offset_t va)
 412 {
 413         return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
 414 }
 415
 416 /*
 417  * Returns the pindex of a page table.
 418  */
 419 static __inline
 420 vm_pindex_t
 421 pmap_pt_pindex(vm_offset_t va)
 422 {
 423         return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
 424 }
 425
 426 /*
 427  * Returns the pindex of a page directory.
 428  */
 429 static __inline
 430 vm_pindex_t
 431 pmap_pd_pindex(vm_offset_t va)
 432 {
 433         return (NUPTE_TOTAL + NUPT_TOTAL +
 434                 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
 435 }
 436
 437 static __inline
 438 vm_pindex_t
 439 pmap_pdp_pindex(vm_offset_t va)
 440 {
 441         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
 442                 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
 443 }
 444
 445 static __inline
 446 vm_pindex_t
 447 pmap_pml4_pindex(void)
 448 {
 449         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
 450 }
 451
 452 /*
 453  * Return various clipped indexes for a given VA
 454  *
 455  * Returns the index of a pt in a page directory, representing a page
 456  * table.
 457  */
 458 static __inline
 459 vm_pindex_t
 460 pmap_pt_index(vm_offset_t va)
 461 {
 462         return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
 463 }
 464
 465 /*
 466  * Returns the index of a pd in a page directory page, representing a page
 467  * directory.
 468  */
 469 static __inline
 470 vm_pindex_t
 471 pmap_pd_index(vm_offset_t va)
 472 {
 473         return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
 474 }
 475
 476 /*
 477  * Returns the index of a pdp in the pml4 table, representing a page
 478  * directory page.
 479  */
 480 static __inline
 481 vm_pindex_t
 482 pmap_pdp_index(vm_offset_t va)
 483 {
 484         return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
 485 }
 486
 487 /*
 488  * The placemarker hash must be broken up into four zones so lock
 489  * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
 490  *
 491  * Placemarkers are used to 'lock' page table indices that do not have
 492  * a pv_entry.  This allows the pmap to support managed and unmanaged
 493  * pages and shared page tables.
 494  */
 495 #define PM_PLACE_BASE   (PM_PLACEMARKS >> 2)
 496
 497 static __inline
 498 vm_pindex_t *
 499 pmap_placemarker_hash(pmap_t pmap, vm_pindex_t pindex)
 500 {
 501         int hi;
 502
 503         if (pindex < pmap_pt_pindex(0))         /* zone 0 - PTE */
 504                 hi = 0;
 505         else if (pindex < pmap_pd_pindex(0))    /* zone 1 - PT */
 506                 hi = PM_PLACE_BASE;
 507         else if (pindex < pmap_pdp_pindex(0))   /* zone 2 - PD */
 508                 hi = PM_PLACE_BASE << 1;
 509         else                                    /* zone 3 - PDP (and PML4E) */
 510                 hi = PM_PLACE_BASE | (PM_PLACE_BASE << 1);
 511         hi += pindex & (PM_PLACE_BASE - 1);
 512
 513         return (&pmap->pm_placemarks[hi]);
 514 }
 515
 516
 517 /*
 518  * Generic procedure to index a pte from a pt, pd, or pdp.
 519  *
 520  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
 521  *       a page table page index but is instead of PV lookup index.
 522  */
 523 static
 524 void *
 525 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
 526 {
 527         pt_entry_t *pte;
 528
 529         pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
 530         return(&pte[pindex]);
 531 }
 532
 533 /*
 534  * Return pointer to PDP slot in the PML4
 535  */
 536 static __inline
 537 pml4_entry_t *
 538 pmap_pdp(pmap_t pmap, vm_offset_t va)
 539 {
 540         return (&pmap->pm_pml4[pmap_pdp_index(va)]);
 541 }
 542
 543 /*
 544  * Return pointer to PD slot in the PDP given a pointer to the PDP
 545  */
 546 static __inline
 547 pdp_entry_t *
 548 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
 549 {
 550         pdp_entry_t *pd;
 551
 552         pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
 553         return (&pd[pmap_pd_index(va)]);
 554 }
 555
 556 /*
 557  * Return pointer to PD slot in the PDP.
 558  */
 559 static __inline
 560 pdp_entry_t *
 561 pmap_pd(pmap_t pmap, vm_offset_t va)
 562 {
 563         pml4_entry_t *pdp;
 564
 565         pdp = pmap_pdp(pmap, va);
 566         if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
 567                 return NULL;
 568         return (pmap_pdp_to_pd(*pdp, va));
 569 }
 570
 571 /*
 572  * Return pointer to PT slot in the PD given a pointer to the PD
 573  */
 574 static __inline
 575 pd_entry_t *
 576 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
 577 {
 578         pd_entry_t *pt;
 579
 580         pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
 581         return (&pt[pmap_pt_index(va)]);
 582 }
 583
 584 /*
 585  * Return pointer to PT slot in the PD
 586  *
 587  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
 588  *                   so we cannot lookup the PD via the PDP.  Instead we
 589  *                   must look it up via the pmap.
 590  */
 591 static __inline
 592 pd_entry_t *
 593 pmap_pt(pmap_t pmap, vm_offset_t va)
 594 {
 595         pdp_entry_t *pd;
 596         pv_entry_t pv;
 597         vm_pindex_t pd_pindex;
 598
 599         if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
 600                 pd_pindex = pmap_pd_pindex(va);
 601                 spin_lock(&pmap->pm_spin);
 602                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
 603                 spin_unlock(&pmap->pm_spin);
 604                 if (pv == NULL || pv->pv_m == NULL)
 605                         return NULL;
 606                 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
 607         } else {
 608                 pd = pmap_pd(pmap, va);
 609                 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
 610                          return NULL;
 611                 return (pmap_pd_to_pt(*pd, va));
 612         }
 613 }
 614
 615 /*
 616  * Return pointer to PTE slot in the PT given a pointer to the PT
 617  */
 618 static __inline
 619 pt_entry_t *
 620 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
 621 {
 622         pt_entry_t *pte;
 623
 624         pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
 625         return (&pte[pmap_pte_index(va)]);
 626 }
 627
 628 /*
 629  * Return pointer to PTE slot in the PT
 630  */
 631 static __inline
 632 pt_entry_t *
 633 pmap_pte(pmap_t pmap, vm_offset_t va)
 634 {
 635         pd_entry_t *pt;
 636
 637         pt = pmap_pt(pmap, va);
 638         if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
 639                  return NULL;
 640         if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
 641                 return ((pt_entry_t *)pt);
 642         return (pmap_pt_to_pte(*pt, va));
 643 }
 644
 645 /*
 646  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
 647  * the PT layer.  This will speed up core pmap operations considerably.
 648  *
 649  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
 650  *       must be in a known associated state (typically by being locked when
 651  *       the pmap spinlock isn't held).  We allow the race for that case.
 652  *
 653  * NOTE: pm_pvhint is only accessed (read) with the spin-lock held, using
 654  *       cpu_ccfence() to prevent compiler optimizations from reloading the
 655  *       field.
 656  */
 657 static __inline
 658 void
 659 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
 660 {
 661         if (pindex >= pmap_pt_pindex(0) && pindex < pmap_pd_pindex(0)) {
 662                 if (pv->pv_pmap)
 663                         pv->pv_pmap->pm_pvhint = pv;
 664         }
 665 }
 666
 667
 668 /*
 669  * Return address of PT slot in PD (KVM only)
 670  *
 671  * Cannot be used for user page tables because it might interfere with
 672  * the shared page-table-page optimization (pmap_mmu_optimize).
 673  */
 674 static __inline
 675 pd_entry_t *
 676 vtopt(vm_offset_t va)
 677 {
 678         uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
 679                                   NPML4EPGSHIFT)) - 1);
 680
 681         return (PDmap + ((va >> PDRSHIFT) & mask));
 682 }
 683
 684 /*
 685  * KVM - return address of PTE slot in PT
 686  */
 687 static __inline
 688 pt_entry_t *
 689 vtopte(vm_offset_t va)
 690 {
 691         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 692                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 693
 694         return (PTmap + ((va >> PAGE_SHIFT) & mask));
 695 }
 696
 697 static uint64_t
 698 allocpages(vm_paddr_t *firstaddr, long n)
 699 {
 700         uint64_t ret;
 701
 702         ret = *firstaddr;
 703         bzero((void *)ret, n * PAGE_SIZE);
 704         *firstaddr += n * PAGE_SIZE;
 705         return (ret);
 706 }
 707
 708 static
 709 void
 710 create_pagetables(vm_paddr_t *firstaddr)
 711 {
 712         long i;         /* must be 64 bits */
 713         long nkpt_base;
 714         long nkpt_phys;
 715         int j;
 716
 717         /*
 718          * We are running (mostly) V=P at this point
 719          *
 720          * Calculate NKPT - number of kernel page tables.  We have to
 721          * accomodoate prealloction of the vm_page_array, dump bitmap,
 722          * MSGBUF_SIZE, and other stuff.  Be generous.
 723          *
 724          * Maxmem is in pages.
 725          *
 726          * ndmpdp is the number of 1GB pages we wish to map.
 727          */
 728         ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
 729         if (ndmpdp < 4)         /* Minimum 4GB of dirmap */
 730                 ndmpdp = 4;
 731         KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
 732
 733         /*
 734          * Starting at the beginning of kvm (not KERNBASE).
 735          */
 736         nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
 737         nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
 738         nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
 739                        ndmpdp) + 511) / 512;
 740         nkpt_phys += 128;
 741
 742         /*
 743          * Starting at KERNBASE - map 2G worth of page table pages.
 744          * KERNBASE is offset -2G from the end of kvm.
 745          */
 746         nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
 747
 748         /*
 749          * Allocate pages
 750          */
 751         KPTbase = allocpages(firstaddr, nkpt_base);
 752         KPTphys = allocpages(firstaddr, nkpt_phys);
 753         KPML4phys = allocpages(firstaddr, 1);
 754         KPDPphys = allocpages(firstaddr, NKPML4E);
 755         KPDphys = allocpages(firstaddr, NKPDPE);
 756
 757         /*
 758          * Calculate the page directory base for KERNBASE,
 759          * that is where we start populating the page table pages.
 760          * Basically this is the end - 2.
 761          */
 762         KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
 763
 764         DMPDPphys = allocpages(firstaddr, NDMPML4E);
 765         if ((amd_feature & AMDID_PAGE1GB) == 0)
 766                 DMPDphys = allocpages(firstaddr, ndmpdp);
 767         dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
 768
 769         /*
 770          * Fill in the underlying page table pages for the area around
 771          * KERNBASE.  This remaps low physical memory to KERNBASE.
 772          *
 773          * Read-only from zero to physfree
 774          * XXX not fully used, underneath 2M pages
 775          */
 776         for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
 777                 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
 778                 ((pt_entry_t *)KPTbase)[i] |=
 779                     pmap_bits_default[PG_RW_IDX] |
 780                     pmap_bits_default[PG_V_IDX] |
 781                     pmap_bits_default[PG_G_IDX];
 782         }
 783
 784         /*
 785          * Now map the initial kernel page tables.  One block of page
 786          * tables is placed at the beginning of kernel virtual memory,
 787          * and another block is placed at KERNBASE to map the kernel binary,
 788          * data, bss, and initial pre-allocations.
 789          */
 790         for (i = 0; i < nkpt_base; i++) {
 791                 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
 792                 ((pd_entry_t *)KPDbase)[i] |=
 793                     pmap_bits_default[PG_RW_IDX] |
 794                     pmap_bits_default[PG_V_IDX];
 795         }
 796         for (i = 0; i < nkpt_phys; i++) {
 797                 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
 798                 ((pd_entry_t *)KPDphys)[i] |=
 799                     pmap_bits_default[PG_RW_IDX] |
 800                     pmap_bits_default[PG_V_IDX];
 801         }
 802
 803         /*
 804          * Map from zero to end of allocations using 2M pages as an
 805          * optimization.  This will bypass some of the KPTBase pages
 806          * above in the KERNBASE area.
 807          */
 808         for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
 809                 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
 810                 ((pd_entry_t *)KPDbase)[i] |=
 811                     pmap_bits_default[PG_RW_IDX] |
 812                     pmap_bits_default[PG_V_IDX] |
 813                     pmap_bits_default[PG_PS_IDX] |
 814                     pmap_bits_default[PG_G_IDX];
 815         }
 816
 817         /*
 818          * And connect up the PD to the PDP.  The kernel pmap is expected
 819          * to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
 820          */
 821         for (i = 0; i < NKPDPE; i++) {
 822                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
 823                                 KPDphys + (i << PAGE_SHIFT);
 824                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
 825                     pmap_bits_default[PG_RW_IDX] |
 826                     pmap_bits_default[PG_V_IDX] |
 827                     pmap_bits_default[PG_U_IDX];
 828         }
 829
 830         /*
 831          * Now set up the direct map space using either 2MB or 1GB pages
 832          * Preset PG_M and PG_A because demotion expects it.
 833          *
 834          * When filling in entries in the PD pages make sure any excess
 835          * entries are set to zero as we allocated enough PD pages
 836          */
 837         if ((amd_feature & AMDID_PAGE1GB) == 0) {
 838                 for (i = 0; i < NPDEPG * ndmpdp; i++) {
 839                         ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
 840                         ((pd_entry_t *)DMPDphys)[i] |=
 841                             pmap_bits_default[PG_RW_IDX] |
 842                             pmap_bits_default[PG_V_IDX] |
 843                             pmap_bits_default[PG_PS_IDX] |
 844                             pmap_bits_default[PG_G_IDX] |
 845                             pmap_bits_default[PG_M_IDX] |
 846                             pmap_bits_default[PG_A_IDX];
 847                 }
 848
 849                 /*
 850                  * And the direct map space's PDP
 851                  */
 852                 for (i = 0; i < ndmpdp; i++) {
 853                         ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
 854                                                         (i << PAGE_SHIFT);
 855                         ((pdp_entry_t *)DMPDPphys)[i] |=
 856                             pmap_bits_default[PG_RW_IDX] |
 857                             pmap_bits_default[PG_V_IDX] |
 858                             pmap_bits_default[PG_U_IDX];
 859                 }
 860         } else {
 861                 for (i = 0; i < ndmpdp; i++) {
 862                         ((pdp_entry_t *)DMPDPphys)[i] =
 863                                                 (vm_paddr_t)i << PDPSHIFT;
 864                         ((pdp_entry_t *)DMPDPphys)[i] |=
 865                             pmap_bits_default[PG_RW_IDX] |
 866                             pmap_bits_default[PG_V_IDX] |
 867                             pmap_bits_default[PG_PS_IDX] |
 868                             pmap_bits_default[PG_G_IDX] |
 869                             pmap_bits_default[PG_M_IDX] |
 870                             pmap_bits_default[PG_A_IDX];
 871                 }
 872         }
 873
 874         /* And recursively map PML4 to itself in order to get PTmap */
 875         ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
 876         ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
 877             pmap_bits_default[PG_RW_IDX] |
 878             pmap_bits_default[PG_V_IDX] |
 879             pmap_bits_default[PG_U_IDX];
 880
 881         /*
 882          * Connect the Direct Map slots up to the PML4
 883          */
 884         for (j = 0; j < NDMPML4E; ++j) {
 885                 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
 886                     (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
 887                     pmap_bits_default[PG_RW_IDX] |
 888                     pmap_bits_default[PG_V_IDX] |
 889                     pmap_bits_default[PG_U_IDX];
 890         }
 891
 892         /*
 893          * Connect the KVA slot up to the PML4
 894          */
 895         ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
 896         ((pdp_entry_t *)KPML4phys)[KPML4I] |=
 897             pmap_bits_default[PG_RW_IDX] |
 898             pmap_bits_default[PG_V_IDX] |
 899             pmap_bits_default[PG_U_IDX];
 900 }
 901
 902 /*
 903  *      Bootstrap the system enough to run with virtual memory.
 904  *
 905  *      On the i386 this is called after mapping has already been enabled
 906  *      and just syncs the pmap module with what has already been done.
 907  *      [We can't call it easily with mapping off since the kernel is not
 908  *      mapped with PA == VA, hence we would have to relocate every address
 909  *      from the linked base (virtual) address "KERNBASE" to the actual
 910  *      (physical) address starting relative to 0]
 911  */
 912 void
 913 pmap_bootstrap(vm_paddr_t *firstaddr)
 914 {
 915         vm_offset_t va;
 916         pt_entry_t *pte;
 917         int i;
 918
 919         KvaStart = VM_MIN_KERNEL_ADDRESS;
 920         KvaEnd = VM_MAX_KERNEL_ADDRESS;
 921         KvaSize = KvaEnd - KvaStart;
 922
 923         avail_start = *firstaddr;
 924
 925         /*
 926          * Create an initial set of page tables to run the kernel in.
 927          */
 928         create_pagetables(firstaddr);
 929
 930         virtual2_start = KvaStart;
 931         virtual2_end = PTOV_OFFSET;
 932
 933         virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
 934         virtual_start = pmap_kmem_choose(virtual_start);
 935
 936         virtual_end = VM_MAX_KERNEL_ADDRESS;
 937
 938         /* XXX do %cr0 as well */
 939         load_cr4(rcr4() | CR4_PGE | CR4_PSE);
 940         load_cr3(KPML4phys);
 941
 942         /*
 943          * Initialize protection array.
 944          */
 945         i386_protection_init();
 946
 947         /*
 948          * The kernel's pmap is statically allocated so we don't have to use
 949          * pmap_create, which is unlikely to work correctly at this part of
 950          * the boot sequence (XXX and which no longer exists).
 951          */
 952         kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
 953         kernel_pmap.pm_count = 1;
 954         CPUMASK_ASSALLONES(kernel_pmap.pm_active);
 955         RB_INIT(&kernel_pmap.pm_pvroot);
 956         spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
 957         for (i = 0; i < PM_PLACEMARKS; ++i)
 958                 kernel_pmap.pm_placemarks[i] = PM_NOPLACEMARK;
 959
 960         /*
 961          * Reserve some special page table entries/VA space for temporary
 962          * mapping of pages.
 963          */
 964 #define SYSMAP(c, p, v, n)      \
 965         v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
 966
 967         va = virtual_start;
 968         pte = vtopte(va);
 969
 970         /*
 971          * CMAP1/CMAP2 are used for zeroing and copying pages.
 972          */
 973         SYSMAP(caddr_t, CMAP1, CADDR1, 1)
 974
 975         /*
 976          * Crashdump maps.
 977          */
 978         SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
 979
 980         /*
 981          * ptvmmap is used for reading arbitrary physical pages via
 982          * /dev/mem.
 983          */
 984         SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
 985
 986         /*
 987          * msgbufp is used to map the system message buffer.
 988          * XXX msgbufmap is not used.
 989          */
 990         SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
 991                atop(round_page(MSGBUF_SIZE)))
 992
 993         virtual_start = va;
 994         virtual_start = pmap_kmem_choose(virtual_start);
 995
 996         *CMAP1 = 0;
 997
 998         /*
 999          * PG_G is terribly broken on SMP because we IPI invltlb's in some
1000          * cases rather then invl1pg.  Actually, I don't even know why it
1001          * works under UP because self-referential page table mappings
1002          */
1003 //      pgeflag = 0;
1004
1005 /*
1006  * Initialize the 4MB page size flag
1007  */
1008 //      pseflag = 0;
1009 /*
1010  * The 4MB page version of the initial
1011  * kernel page mapping.
1012  */
1013         pdir4mb = 0;
1014
1015 #if !defined(DISABLE_PSE)
1016         if (cpu_feature & CPUID_PSE) {
1017                 pt_entry_t ptditmp;
1018                 /*
1019                  * Note that we have enabled PSE mode
1020                  */
1021 //              pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
1022                 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
1023                 ptditmp &= ~(NBPDR - 1);
1024                 ptditmp |= pmap_bits_default[PG_V_IDX] |
1025                     pmap_bits_default[PG_RW_IDX] |
1026                     pmap_bits_default[PG_PS_IDX] |
1027                     pmap_bits_default[PG_U_IDX];
1028 //                  pgeflag;
1029                 pdir4mb = ptditmp;
1030         }
1031 #endif
1032         cpu_invltlb();
1033
1034         /* Initialize the PAT MSR */
1035         pmap_init_pat();
1036         pmap_pinit_defaults(&kernel_pmap);
1037
1038         TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1039                           &pmap_fast_kernel_cpusync);
1040
1041 }
1042
1043 /*
1044  * Setup the PAT MSR.
1045  */
1046 void
1047 pmap_init_pat(void)
1048 {
1049         uint64_t pat_msr;
1050         u_long cr0, cr4;
1051
1052         /*
1053          * Default values mapping PATi,PCD,PWT bits at system reset.
1054          * The default values effectively ignore the PATi bit by
1055          * repeating the encodings for 0-3 in 4-7, and map the PCD
1056          * and PWT bit combinations to the expected PAT types.
1057          */
1058         pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |        /* 000 */
1059                   PAT_VALUE(1, PAT_WRITE_THROUGH) |     /* 001 */
1060                   PAT_VALUE(2, PAT_UNCACHED) |          /* 010 */
1061                   PAT_VALUE(3, PAT_UNCACHEABLE) |       /* 011 */
1062                   PAT_VALUE(4, PAT_WRITE_BACK) |        /* 100 */
1063                   PAT_VALUE(5, PAT_WRITE_THROUGH) |     /* 101 */
1064                   PAT_VALUE(6, PAT_UNCACHED) |          /* 110 */
1065                   PAT_VALUE(7, PAT_UNCACHEABLE);        /* 111 */
1066         pat_pte_index[PAT_WRITE_BACK]   = 0;
1067         pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
1068         pat_pte_index[PAT_UNCACHED]     = X86_PG_NC_PCD;
1069         pat_pte_index[PAT_UNCACHEABLE]  = X86_PG_NC_PCD | X86_PG_NC_PWT;
1070         pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1071         pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1072
1073         if (cpu_feature & CPUID_PAT) {
1074                 /*
1075                  * If we support the PAT then set-up entries for
1076                  * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1077                  * 4 and 5.
1078                  */
1079                 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1080                           PAT_VALUE(4, PAT_WRITE_PROTECTED);
1081                 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1082                           PAT_VALUE(5, PAT_WRITE_COMBINING);
1083                 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1084                 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1085
1086                 /*
1087                  * Then enable the PAT
1088                  */
1089
1090                 /* Disable PGE. */
1091                 cr4 = rcr4();
1092                 load_cr4(cr4 & ~CR4_PGE);
1093
1094                 /* Disable caches (CD = 1, NW = 0). */
1095                 cr0 = rcr0();
1096                 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1097
1098                 /* Flushes caches and TLBs. */
1099                 wbinvd();
1100                 cpu_invltlb();
1101
1102                 /* Update PAT and index table. */
1103                 wrmsr(MSR_PAT, pat_msr);
1104
1105                 /* Flush caches and TLBs again. */
1106                 wbinvd();
1107                 cpu_invltlb();
1108
1109                 /* Restore caches and PGE. */
1110                 load_cr0(cr0);
1111                 load_cr4(cr4);
1112                 PatMsr = pat_msr;
1113         }
1114 }
1115
1116 /*
1117  * Set 4mb pdir for mp startup
1118  */
1119 void
1120 pmap_set_opt(void)
1121 {
1122         if (cpu_feature & CPUID_PSE) {
1123                 load_cr4(rcr4() | CR4_PSE);
1124                 if (pdir4mb && mycpu->gd_cpuid == 0) {  /* only on BSP */
1125                         cpu_invltlb();
1126                 }
1127         }
1128 }
1129
1130 /*
1131  *      Initialize the pmap module.
1132  *      Called by vm_init, to initialize any structures that the pmap
1133  *      system needs to map virtual memory.
1134  *      pmap_init has been enhanced to support in a fairly consistant
1135  *      way, discontiguous physical memory.
1136  */
1137 void
1138 pmap_init(void)
1139 {
1140         int i;
1141         int initial_pvs;
1142
1143         /*
1144          * Allocate memory for random pmap data structures.  Includes the
1145          * pv_head_table.
1146          */
1147
1148         for (i = 0; i < vm_page_array_size; i++) {
1149                 vm_page_t m;
1150
1151                 m = &vm_page_array[i];
1152                 TAILQ_INIT(&m->md.pv_list);
1153         }
1154
1155         /*
1156          * init the pv free list
1157          */
1158         initial_pvs = vm_page_array_size;
1159         if (initial_pvs < MINPV)
1160                 initial_pvs = MINPV;
1161         pvzone = &pvzone_store;
1162         pvinit = (void *)kmem_alloc(&kernel_map,
1163                                     initial_pvs * sizeof (struct pv_entry),
1164                                     VM_SUBSYS_PVENTRY);
1165         zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1166                   pvinit, initial_pvs);
1167
1168         /*
1169          * Now it is safe to enable pv_table recording.
1170          */
1171         pmap_initialized = TRUE;
1172 }
1173
1174 /*
1175  * Initialize the address space (zone) for the pv_entries.  Set a
1176  * high water mark so that the system can recover from excessive
1177  * numbers of pv entries.
1178  */
1179 void
1180 pmap_init2(void)
1181 {
1182         int shpgperproc = PMAP_SHPGPERPROC;
1183         int entry_max;
1184
1185         TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1186         pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1187         TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1188         pv_entry_high_water = 9 * (pv_entry_max / 10);
1189
1190         /*
1191          * Subtract out pages already installed in the zone (hack)
1192          */
1193         entry_max = pv_entry_max - vm_page_array_size;
1194         if (entry_max <= 0)
1195                 entry_max = 1;
1196
1197         zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT);
1198
1199         /*
1200          * Enable dynamic deletion of empty higher-level page table pages
1201          * by default only if system memory is < 8GB (use 7GB for slop).
1202          * This can save a little memory, but imposes significant
1203          * performance overhead for things like bulk builds, and for programs
1204          * which do a lot of memory mapping and memory unmapping.
1205          */
1206         if (pmap_dynamic_delete < 0) {
1207                 if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1208                         pmap_dynamic_delete = 1;
1209                 else
1210                         pmap_dynamic_delete = 0;
1211         }
1212 }
1213
1214 /*
1215  * Typically used to initialize a fictitious page by vm/device_pager.c
1216  */
1217 void
1218 pmap_page_init(struct vm_page *m)
1219 {
1220         vm_page_init(m);
1221         TAILQ_INIT(&m->md.pv_list);
1222 }
1223
1224 /***************************************************
1225  * Low level helper routines.....
1226  ***************************************************/
1227
1228 /*
1229  * this routine defines the region(s) of memory that should
1230  * not be tested for the modified bit.
1231  */
1232 static __inline
1233 int
1234 pmap_track_modified(vm_pindex_t pindex)
1235 {
1236         vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1237         if ((va < clean_sva) || (va >= clean_eva))
1238                 return 1;
1239         else
1240                 return 0;
1241 }
1242
1243 /*
1244  * Extract the physical page address associated with the map/VA pair.
1245  * The page must be wired for this to work reliably.
1246  */
1247 vm_paddr_t
1248 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
1249 {
1250         vm_paddr_t rtval;
1251         pv_entry_t pt_pv;
1252         pt_entry_t *ptep;
1253
1254         rtval = 0;
1255         if (va >= VM_MAX_USER_ADDRESS) {
1256                 /*
1257                  * Kernel page directories might be direct-mapped and
1258                  * there is typically no PV tracking of pte's
1259                  */
1260                 pd_entry_t *pt;
1261
1262                 pt = pmap_pt(pmap, va);
1263                 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1264                         if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1265                                 rtval = *pt & PG_PS_FRAME;
1266                                 rtval |= va & PDRMASK;
1267                         } else {
1268                                 ptep = pmap_pt_to_pte(*pt, va);
1269                                 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1270                                         rtval = *ptep & PG_FRAME;
1271                                         rtval |= va & PAGE_MASK;
1272                                 }
1273                         }
1274                 }
1275                 if (handlep)
1276                         *handlep = NULL;
1277         } else {
1278                 /*
1279                  * User pages currently do not direct-map the page directory
1280                  * and some pages might not used managed PVs.  But all PT's
1281                  * will have a PV.
1282                  */
1283                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1284                 if (pt_pv) {
1285                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1286                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1287                                 rtval = *ptep & PG_FRAME;
1288                                 rtval |= va & PAGE_MASK;
1289                         }
1290                         if (handlep)
1291                                 *handlep = pt_pv;       /* locked until done */
1292                         else
1293                                 pv_put (pt_pv);
1294                 } else if (handlep) {
1295                         *handlep = NULL;
1296                 }
1297         }
1298         return rtval;
1299 }
1300
1301 void
1302 pmap_extract_done(void *handle)
1303 {
1304         if (handle)
1305                 pv_put((pv_entry_t)handle);
1306 }
1307
1308 /*
1309  * Similar to extract but checks protections, SMP-friendly short-cut for
1310  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1311  * fall-through to the real fault code.  Does not work with HVM page
1312  * tables.
1313  *
1314  * The returned page, if not NULL, is held (and not busied).
1315  *
1316  * WARNING! THE RETURNED PAGE IS ONLY HELD AND NOT SUITABLE FOR READING
1317  *          OR WRITING AS-IS.
1318  */
1319 vm_page_t
1320 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot, int *busyp)
1321 {
1322         if (pmap &&
1323             va < VM_MAX_USER_ADDRESS &&
1324             (pmap->pm_flags & PMAP_HVM) == 0) {
1325                 pv_entry_t pt_pv;
1326                 pv_entry_t pte_pv;
1327                 pt_entry_t *ptep;
1328                 pt_entry_t req;
1329                 vm_page_t m;
1330                 int error;
1331
1332                 req = pmap->pmap_bits[PG_V_IDX] |
1333                       pmap->pmap_bits[PG_U_IDX];
1334                 if (prot & VM_PROT_WRITE)
1335                         req |= pmap->pmap_bits[PG_RW_IDX];
1336
1337                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1338                 if (pt_pv == NULL)
1339                         return (NULL);
1340                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1341                 if ((*ptep & req) != req) {
1342                         pv_put(pt_pv);
1343                         return (NULL);
1344                 }
1345                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), NULL, &error);
1346                 if (pte_pv && error == 0) {
1347                         m = pte_pv->pv_m;
1348                         if (prot & VM_PROT_WRITE) {
1349                                 /* interlocked by presence of pv_entry */
1350                                 vm_page_dirty(m);
1351                         }
1352                         if (busyp) {
1353                                 if (prot & VM_PROT_WRITE) {
1354                                         if (vm_page_busy_try(m, TRUE))
1355                                                 m = NULL;
1356                                         *busyp = 1;
1357                                 } else {
1358                                         vm_page_hold(m);
1359                                         *busyp = 0;
1360                                 }
1361                         } else {
1362                                 vm_page_hold(m);
1363                         }
1364                         pv_put(pte_pv);
1365                 } else if (pte_pv) {
1366                         pv_drop(pte_pv);
1367                         m = NULL;
1368                 } else {
1369                         /* error, since we didn't request a placemarker */
1370                         m = NULL;
1371                 }
1372                 pv_put(pt_pv);
1373                 return(m);
1374         } else {
1375                 return(NULL);
1376         }
1377 }
1378
1379 /*
1380  * Extract the physical page address associated kernel virtual address.
1381  */
1382 vm_paddr_t
1383 pmap_kextract(vm_offset_t va)
1384 {
1385         pd_entry_t pt;          /* pt entry in pd */
1386         vm_paddr_t pa;
1387
1388         if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1389                 pa = DMAP_TO_PHYS(va);
1390         } else {
1391                 pt = *vtopt(va);
1392                 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1393                         pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1394                 } else {
1395                         /*
1396                          * Beware of a concurrent promotion that changes the
1397                          * PDE at this point!  For example, vtopte() must not
1398                          * be used to access the PTE because it would use the
1399                          * new PDE.  It is, however, safe to use the old PDE
1400                          * because the page table page is preserved by the
1401                          * promotion.
1402                          */
1403                         pa = *pmap_pt_to_pte(pt, va);
1404                         pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1405                 }
1406         }
1407         return pa;
1408 }
1409
1410 /***************************************************
1411  * Low level mapping routines.....
1412  ***************************************************/
1413
1414 /*
1415  * Routine: pmap_kenter
1416  * Function:
1417  *      Add a wired page to the KVA
1418  *      NOTE! note that in order for the mapping to take effect -- you
1419  *      should do an invltlb after doing the pmap_kenter().
1420  */
1421 void
1422 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1423 {
1424         pt_entry_t *ptep;
1425         pt_entry_t npte;
1426
1427         npte = pa |
1428                kernel_pmap.pmap_bits[PG_RW_IDX] |
1429                kernel_pmap.pmap_bits[PG_V_IDX];
1430 //             pgeflag;
1431         ptep = vtopte(va);
1432 #if 1
1433         pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1434 #else
1435         /* FUTURE */
1436         if (*ptep)
1437                 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1438         else
1439                 *ptep = npte;
1440 #endif
1441 }
1442
1443 /*
1444  * Similar to pmap_kenter(), except we only invalidate the mapping on the
1445  * current CPU.  Returns 0 if the previous pte was 0, 1 if it wasn't
1446  * (caller can conditionalize calling smp_invltlb()).
1447  */
1448 int
1449 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1450 {
1451         pt_entry_t *ptep;
1452         pt_entry_t npte;
1453         int res;
1454
1455         npte = pa | kernel_pmap.pmap_bits[PG_RW_IDX] |
1456                     kernel_pmap.pmap_bits[PG_V_IDX];
1457         // npte |= pgeflag;
1458         ptep = vtopte(va);
1459 #if 1
1460         res = 1;
1461 #else
1462         /* FUTURE */
1463         res = (*ptep != 0);
1464 #endif
1465         atomic_swap_long(ptep, npte);
1466         cpu_invlpg((void *)va);
1467
1468         return res;
1469 }
1470
1471 /*
1472  * Enter addresses into the kernel pmap but don't bother
1473  * doing any tlb invalidations.  Caller will do a rollup
1474  * invalidation via pmap_rollup_inval().
1475  */
1476 int
1477 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1478 {
1479         pt_entry_t *ptep;
1480         pt_entry_t npte;
1481         int res;
1482
1483         npte = pa |
1484             kernel_pmap.pmap_bits[PG_RW_IDX] |
1485             kernel_pmap.pmap_bits[PG_V_IDX];
1486 //          pgeflag;
1487         ptep = vtopte(va);
1488 #if 1
1489         res = 1;
1490 #else
1491         /* FUTURE */
1492         res = (*ptep != 0);
1493 #endif
1494         atomic_swap_long(ptep, npte);
1495         cpu_invlpg((void *)va);
1496
1497         return res;
1498 }
1499
1500 /*
1501  * remove a page from the kernel pagetables
1502  */
1503 void
1504 pmap_kremove(vm_offset_t va)
1505 {
1506         pt_entry_t *ptep;
1507
1508         ptep = vtopte(va);
1509         pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1510 }
1511
1512 void
1513 pmap_kremove_quick(vm_offset_t va)
1514 {
1515         pt_entry_t *ptep;
1516
1517         ptep = vtopte(va);
1518         (void)pte_load_clear(ptep);
1519         cpu_invlpg((void *)va);
1520 }
1521
1522 /*
1523  * Remove addresses from the kernel pmap but don't bother
1524  * doing any tlb invalidations.  Caller will do a rollup
1525  * invalidation via pmap_rollup_inval().
1526  */
1527 void
1528 pmap_kremove_noinval(vm_offset_t va)
1529 {
1530         pt_entry_t *ptep;
1531
1532         ptep = vtopte(va);
1533         (void)pte_load_clear(ptep);
1534 }
1535
1536 /*
1537  * XXX these need to be recoded.  They are not used in any critical path.
1538  */
1539 void
1540 pmap_kmodify_rw(vm_offset_t va)
1541 {
1542         atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1543         cpu_invlpg((void *)va);
1544 }
1545
1546 /* NOT USED
1547 void
1548 pmap_kmodify_nc(vm_offset_t va)
1549 {
1550         atomic_set_long(vtopte(va), PG_N);
1551         cpu_invlpg((void *)va);
1552 }
1553 */
1554
1555 /*
1556  * Used to map a range of physical addresses into kernel virtual
1557  * address space during the low level boot, typically to map the
1558  * dump bitmap, message buffer, and vm_page_array.
1559  *
1560  * These mappings are typically made at some pointer after the end of the
1561  * kernel text+data.
1562  *
1563  * We could return PHYS_TO_DMAP(start) here and not allocate any
1564  * via (*virtp), but then kmem from userland and kernel dumps won't
1565  * have access to the related pointers.
1566  */
1567 vm_offset_t
1568 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1569 {
1570         vm_offset_t va;
1571         vm_offset_t va_start;
1572
1573         /*return PHYS_TO_DMAP(start);*/
1574
1575         va_start = *virtp;
1576         va = va_start;
1577
1578         while (start < end) {
1579                 pmap_kenter_quick(va, start);
1580                 va += PAGE_SIZE;
1581                 start += PAGE_SIZE;
1582         }
1583         *virtp = va;
1584         return va_start;
1585 }
1586
1587 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
1588
1589 /*
1590  * Remove the specified set of pages from the data and instruction caches.
1591  *
1592  * In contrast to pmap_invalidate_cache_range(), this function does not
1593  * rely on the CPU's self-snoop feature, because it is intended for use
1594  * when moving pages into a different cache domain.
1595  */
1596 void
1597 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1598 {
1599         vm_offset_t daddr, eva;
1600         int i;
1601
1602         if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1603             (cpu_feature & CPUID_CLFSH) == 0)
1604                 wbinvd();
1605         else {
1606                 cpu_mfence();
1607                 for (i = 0; i < count; i++) {
1608                         daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1609                         eva = daddr + PAGE_SIZE;
1610                         for (; daddr < eva; daddr += cpu_clflush_line_size)
1611                                 clflush(daddr);
1612                 }
1613                 cpu_mfence();
1614         }
1615 }
1616
1617 void
1618 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1619 {
1620         KASSERT((sva & PAGE_MASK) == 0,
1621             ("pmap_invalidate_cache_range: sva not page-aligned"));
1622         KASSERT((eva & PAGE_MASK) == 0,
1623             ("pmap_invalidate_cache_range: eva not page-aligned"));
1624
1625         if (cpu_feature & CPUID_SS) {
1626                 ; /* If "Self Snoop" is supported, do nothing. */
1627         } else {
1628                 /* Globally invalidate caches */
1629                 cpu_wbinvd_on_all_cpus();
1630         }
1631 }
1632
1633 /*
1634  * Invalidate the specified range of virtual memory on all cpus associated
1635  * with the pmap.
1636  */
1637 void
1638 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1639 {
1640         pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1641 }
1642
1643 /*
1644  * Add a list of wired pages to the kva.  This routine is used for temporary
1645  * kernel mappings such as those found in buffer cache buffer.  Page
1646  * modifications and accesses are not tracked or recorded.
1647  *
1648  * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1649  *       semantics as previous mappings may have been zerod without any
1650  *       invalidation.
1651  *
1652  * The page *must* be wired.
1653  */
1654 void
1655 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1656 {
1657         vm_offset_t end_va;
1658         vm_offset_t va;
1659
1660         end_va = beg_va + count * PAGE_SIZE;
1661
1662         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1663                 pt_entry_t pte;
1664                 pt_entry_t *ptep;
1665
1666                 ptep = vtopte(va);
1667                 pte = VM_PAGE_TO_PHYS(*m) |
1668                         kernel_pmap.pmap_bits[PG_RW_IDX] |
1669                         kernel_pmap.pmap_bits[PG_V_IDX] |
1670                         kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1671 //              pgeflag;
1672                 atomic_swap_long(ptep, pte);
1673                 m++;
1674         }
1675         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1676 }
1677
1678 /*
1679  * This routine jerks page mappings from the kernel -- it is meant only
1680  * for temporary mappings such as those found in buffer cache buffers.
1681  * No recording modified or access status occurs.
1682  *
1683  * MPSAFE, INTERRUPT SAFE (cluster callback)
1684  */
1685 void
1686 pmap_qremove(vm_offset_t beg_va, int count)
1687 {
1688         vm_offset_t end_va;
1689         vm_offset_t va;
1690
1691         end_va = beg_va + count * PAGE_SIZE;
1692
1693         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1694                 pt_entry_t *pte;
1695
1696                 pte = vtopte(va);
1697                 (void)pte_load_clear(pte);
1698                 cpu_invlpg((void *)va);
1699         }
1700         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1701 }
1702
1703 /*
1704  * This routine removes temporary kernel mappings, only invalidating them
1705  * on the current cpu.  It should only be used under carefully controlled
1706  * conditions.
1707  */
1708 void
1709 pmap_qremove_quick(vm_offset_t beg_va, int count)
1710 {
1711         vm_offset_t end_va;
1712         vm_offset_t va;
1713
1714         end_va = beg_va + count * PAGE_SIZE;
1715
1716         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1717                 pt_entry_t *pte;
1718
1719                 pte = vtopte(va);
1720                 (void)pte_load_clear(pte);
1721                 cpu_invlpg((void *)va);
1722         }
1723 }
1724
1725 /*
1726  * This routine removes temporary kernel mappings *without* invalidating
1727  * the TLB.  It can only be used on permanent kva reservations such as those
1728  * found in buffer cache buffers, under carefully controlled circumstances.
1729  *
1730  * NOTE: Repopulating these KVAs requires unconditional invalidation.
1731  *       (pmap_qenter() does unconditional invalidation).
1732  */
1733 void
1734 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1735 {
1736         vm_offset_t end_va;
1737         vm_offset_t va;
1738
1739         end_va = beg_va + count * PAGE_SIZE;
1740
1741         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1742                 pt_entry_t *pte;
1743
1744                 pte = vtopte(va);
1745                 (void)pte_load_clear(pte);
1746         }
1747 }
1748
1749 /*
1750  * Create a new thread and optionally associate it with a (new) process.
1751  * NOTE! the new thread's cpu may not equal the current cpu.
1752  */
1753 void
1754 pmap_init_thread(thread_t td)
1755 {
1756         /* enforce pcb placement & alignment */
1757         td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1758         td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1759         td->td_savefpu = &td->td_pcb->pcb_save;
1760         td->td_sp = (char *)td->td_pcb; /* no -16 */
1761 }
1762
1763 /*
1764  * This routine directly affects the fork perf for a process.
1765  */
1766 void
1767 pmap_init_proc(struct proc *p)
1768 {
1769 }
1770
1771 static void
1772 pmap_pinit_defaults(struct pmap *pmap)
1773 {
1774         bcopy(pmap_bits_default, pmap->pmap_bits,
1775               sizeof(pmap_bits_default));
1776         bcopy(protection_codes, pmap->protection_codes,
1777               sizeof(protection_codes));
1778         bcopy(pat_pte_index, pmap->pmap_cache_bits,
1779               sizeof(pat_pte_index));
1780         pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1781         pmap->copyinstr = std_copyinstr;
1782         pmap->copyin = std_copyin;
1783         pmap->copyout = std_copyout;
1784         pmap->fubyte = std_fubyte;
1785         pmap->subyte = std_subyte;
1786         pmap->fuword32 = std_fuword32;
1787         pmap->fuword64 = std_fuword64;
1788         pmap->suword32 = std_suword32;
1789         pmap->suword64 = std_suword64;
1790         pmap->swapu32 = std_swapu32;
1791         pmap->swapu64 = std_swapu64;
1792 }
1793 /*
1794  * Initialize pmap0/vmspace0.
1795  *
1796  * On architectures where the kernel pmap is not integrated into the user
1797  * process pmap, this pmap represents the process pmap, not the kernel pmap.
1798  * kernel_pmap should be used to directly access the kernel_pmap.
1799  */
1800 void
1801 pmap_pinit0(struct pmap *pmap)
1802 {
1803         int i;
1804
1805         pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1806         pmap->pm_count = 1;
1807         CPUMASK_ASSZERO(pmap->pm_active);
1808         pmap->pm_pvhint = NULL;
1809         RB_INIT(&pmap->pm_pvroot);
1810         spin_init(&pmap->pm_spin, "pmapinit0");
1811         for (i = 0; i < PM_PLACEMARKS; ++i)
1812                 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1813         bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1814         pmap_pinit_defaults(pmap);
1815 }
1816
1817 /*
1818  * Initialize a preallocated and zeroed pmap structure,
1819  * such as one in a vmspace structure.
1820  */
1821 static void
1822 pmap_pinit_simple(struct pmap *pmap)
1823 {
1824         int i;
1825
1826         /*
1827          * Misc initialization
1828          */
1829         pmap->pm_count = 1;
1830         CPUMASK_ASSZERO(pmap->pm_active);
1831         pmap->pm_pvhint = NULL;
1832         pmap->pm_flags = PMAP_FLAG_SIMPLE;
1833
1834         pmap_pinit_defaults(pmap);
1835
1836         /*
1837          * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1838          * for this).
1839          */
1840         if (pmap->pm_pmlpv == NULL) {
1841                 RB_INIT(&pmap->pm_pvroot);
1842                 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1843                 spin_init(&pmap->pm_spin, "pmapinitsimple");
1844                 for (i = 0; i < PM_PLACEMARKS; ++i)
1845                         pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1846         }
1847 }
1848
1849 void
1850 pmap_pinit(struct pmap *pmap)
1851 {
1852         pv_entry_t pv;
1853         int j;
1854
1855         if (pmap->pm_pmlpv) {
1856                 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1857                         pmap_puninit(pmap);
1858                 }
1859         }
1860
1861         pmap_pinit_simple(pmap);
1862         pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1863
1864         /*
1865          * No need to allocate page table space yet but we do need a valid
1866          * page directory table.
1867          */
1868         if (pmap->pm_pml4 == NULL) {
1869                 pmap->pm_pml4 =
1870                     (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1871                                                         PAGE_SIZE,
1872                                                         VM_SUBSYS_PML4);
1873         }
1874
1875         /*
1876          * Allocate the page directory page, which wires it even though
1877          * it isn't being entered into some higher level page table (it
1878          * being the highest level).  If one is already cached we don't
1879          * have to do anything.
1880          */
1881         if ((pv = pmap->pm_pmlpv) == NULL) {
1882                 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1883                 pmap->pm_pmlpv = pv;
1884                 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1885                             VM_PAGE_TO_PHYS(pv->pv_m));
1886                 pv_put(pv);
1887
1888                 /*
1889                  * Install DMAP and KMAP.
1890                  */
1891                 for (j = 0; j < NDMPML4E; ++j) {
1892                         pmap->pm_pml4[DMPML4I + j] =
1893                             (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1894                             pmap->pmap_bits[PG_RW_IDX] |
1895                             pmap->pmap_bits[PG_V_IDX] |
1896                             pmap->pmap_bits[PG_U_IDX];
1897                 }
1898                 pmap->pm_pml4[KPML4I] = KPDPphys |
1899                     pmap->pmap_bits[PG_RW_IDX] |
1900                     pmap->pmap_bits[PG_V_IDX] |
1901                     pmap->pmap_bits[PG_U_IDX];
1902
1903                 /*
1904                  * install self-referential address mapping entry
1905                  */
1906                 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1907                     pmap->pmap_bits[PG_V_IDX] |
1908                     pmap->pmap_bits[PG_RW_IDX] |
1909                     pmap->pmap_bits[PG_A_IDX] |
1910                     pmap->pmap_bits[PG_M_IDX];
1911         } else {
1912                 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1913                 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1914         }
1915         KKASSERT(pmap->pm_pml4[255] == 0);
1916         KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1917         KKASSERT(pv->pv_entry.rbe_left == NULL);
1918         KKASSERT(pv->pv_entry.rbe_right == NULL);
1919 }
1920
1921 /*
1922  * Clean up a pmap structure so it can be physically freed.  This routine
1923  * is called by the vmspace dtor function.  A great deal of pmap data is
1924  * left passively mapped to improve vmspace management so we have a bit
1925  * of cleanup work to do here.
1926  */
1927 void
1928 pmap_puninit(pmap_t pmap)
1929 {
1930         pv_entry_t pv;
1931         vm_page_t p;
1932
1933         KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1934         if ((pv = pmap->pm_pmlpv) != NULL) {
1935                 if (pv_hold_try(pv) == 0)
1936                         pv_lock(pv);
1937                 KKASSERT(pv == pmap->pm_pmlpv);
1938                 p = pmap_remove_pv_page(pv);
1939                 pv_free(pv, NULL);
1940                 pv = NULL;      /* safety */
1941                 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1942                 vm_page_busy_wait(p, FALSE, "pgpun");
1943                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1944                 vm_page_unwire(p, 0);
1945                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1946
1947                 /*
1948                  * XXX eventually clean out PML4 static entries and
1949                  * use vm_page_free_zero()
1950                  */
1951                 vm_page_free(p);
1952                 pmap->pm_pmlpv = NULL;
1953         }
1954         if (pmap->pm_pml4) {
1955                 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1956                 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1957                 pmap->pm_pml4 = NULL;
1958         }
1959         KKASSERT(pmap->pm_stats.resident_count == 0);
1960         KKASSERT(pmap->pm_stats.wired_count == 0);
1961 }
1962
1963 /*
1964  * This function is now unused (used to add the pmap to the pmap_list)
1965  */
1966 void
1967 pmap_pinit2(struct pmap *pmap)
1968 {
1969 }
1970
1971 /*
1972  * This routine is called when various levels in the page table need to
1973  * be populated.  This routine cannot fail.
1974  *
1975  * This function returns two locked pv_entry's, one representing the
1976  * requested pv and one representing the requested pv's parent pv.  If
1977  * an intermediate page table does not exist it will be created, mapped,
1978  * wired, and the parent page table will be given an additional hold
1979  * count representing the presence of the child pv_entry.
1980  */
1981 static
1982 pv_entry_t
1983 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1984 {
1985         pt_entry_t *ptep;
1986         pv_entry_t pv;
1987         pv_entry_t pvp;
1988         pt_entry_t v;
1989         vm_pindex_t pt_pindex;
1990         vm_page_t m;
1991         int isnew;
1992         int ispt;
1993
1994         /*
1995          * If the pv already exists and we aren't being asked for the
1996          * parent page table page we can just return it.  A locked+held pv
1997          * is returned.  The pv will also have a second hold related to the
1998          * pmap association that we don't have to worry about.
1999          */
2000         ispt = 0;
2001         pv = pv_alloc(pmap, ptepindex, &isnew);
2002         if (isnew == 0 && pvpp == NULL)
2003                 return(pv);
2004
2005         /*
2006          * Special case terminal PVs.  These are not page table pages so
2007          * no vm_page is allocated (the caller supplied the vm_page).  If
2008          * pvpp is non-NULL we are being asked to also removed the pt_pv
2009          * for this pv.
2010          *
2011          * Note that pt_pv's are only returned for user VAs. We assert that
2012          * a pt_pv is not being requested for kernel VAs.  The kernel
2013          * pre-wires all higher-level page tables so don't overload managed
2014          * higher-level page tables on top of it!
2015          */
2016         if (ptepindex < pmap_pt_pindex(0)) {
2017                 if (ptepindex >= NUPTE_USER) {
2018                         /* kernel manages this manually for KVM */
2019                         KKASSERT(pvpp == NULL);
2020                 } else {
2021                         KKASSERT(pvpp != NULL);
2022                         pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
2023                         pvp = pmap_allocpte(pmap, pt_pindex, NULL);
2024                         if (isnew)
2025                                 vm_page_wire_quick(pvp->pv_m);
2026                         *pvpp = pvp;
2027                 }
2028                 return(pv);
2029         }
2030
2031         /*
2032          * The kernel never uses managed PT/PD/PDP pages.
2033          */
2034         KKASSERT(pmap != &kernel_pmap);
2035
2036         /*
2037          * Non-terminal PVs allocate a VM page to represent the page table,
2038          * so we have to resolve pvp and calculate ptepindex for the pvp
2039          * and then for the page table entry index in the pvp for
2040          * fall-through.
2041          */
2042         if (ptepindex < pmap_pd_pindex(0)) {
2043                 /*
2044                  * pv is PT, pvp is PD
2045                  */
2046                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
2047                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
2048                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2049
2050                 /*
2051                  * PT index in PD
2052                  */
2053                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
2054                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
2055                 ispt = 1;
2056         } else if (ptepindex < pmap_pdp_pindex(0)) {
2057                 /*
2058                  * pv is PD, pvp is PDP
2059                  *
2060                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2061                  *                   the PD.
2062                  */
2063                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
2064                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2065
2066                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
2067                         KKASSERT(pvpp == NULL);
2068                         pvp = NULL;
2069                 } else {
2070                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
2071                 }
2072
2073                 /*
2074                  * PD index in PDP
2075                  */
2076                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
2077                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
2078         } else if (ptepindex < pmap_pml4_pindex()) {
2079                 /*
2080                  * pv is PDP, pvp is the root pml4 table
2081                  */
2082                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2083
2084                 /*
2085                  * PDP index in PML4
2086                  */
2087                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2088                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2089         } else {
2090                 /*
2091                  * pv represents the top-level PML4, there is no parent.
2092                  */
2093                 pvp = NULL;
2094         }
2095
2096         if (isnew == 0)
2097                 goto notnew;
2098
2099         /*
2100          * (isnew) is TRUE, pv is not terminal.
2101          *
2102          * (1) Add a wire count to the parent page table (pvp).
2103          * (2) Allocate a VM page for the page table.
2104          * (3) Enter the VM page into the parent page table.
2105          *
2106          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2107          */
2108         if (pvp)
2109                 vm_page_wire_quick(pvp->pv_m);
2110
2111         for (;;) {
2112                 m = vm_page_alloc(NULL, pv->pv_pindex,
2113                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2114                                   VM_ALLOC_INTERRUPT);
2115                 if (m)
2116                         break;
2117                 vm_wait(0);
2118         }
2119         vm_page_wire(m);        /* wire for mapping in parent */
2120         vm_page_unmanage(m);    /* m must be spinunlocked */
2121         pmap_zero_page(VM_PAGE_TO_PHYS(m));
2122         m->valid = VM_PAGE_BITS_ALL;
2123
2124         vm_page_spin_lock(m);
2125         pmap_page_stats_adding(m);
2126         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2127         pv->pv_m = m;
2128         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2129         vm_page_spin_unlock(m);
2130
2131         /*
2132          * (isnew) is TRUE, pv is not terminal.
2133          *
2134          * Wire the page into pvp.  Bump the resident_count for the pmap.
2135          * There is no pvp for the top level, address the pm_pml4[] array
2136          * directly.
2137          *
2138          * If the caller wants the parent we return it, otherwise
2139          * we just put it away.
2140          *
2141          * No interlock is needed for pte 0 -> non-zero.
2142          *
2143          * In the situation where *ptep is valid we might have an unmanaged
2144          * page table page shared from another page table which we need to
2145          * unshare before installing our private page table page.
2146          */
2147         if (pvp) {
2148                 v = VM_PAGE_TO_PHYS(m) |
2149                     (pmap->pmap_bits[PG_U_IDX] |
2150                      pmap->pmap_bits[PG_RW_IDX] |
2151                      pmap->pmap_bits[PG_V_IDX] |
2152                      pmap->pmap_bits[PG_A_IDX] |
2153                      pmap->pmap_bits[PG_M_IDX]);
2154                 ptep = pv_pte_lookup(pvp, ptepindex);
2155                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2156                         pt_entry_t pte;
2157
2158                         if (ispt == 0) {
2159                                 panic("pmap_allocpte: unexpected pte %p/%d",
2160                                       pvp, (int)ptepindex);
2161                         }
2162                         pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, v);
2163                         if (vm_page_unwire_quick(
2164                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2165                                 panic("pmap_allocpte: shared pgtable "
2166                                       "pg bad wirecount");
2167                         }
2168                 } else {
2169                         pt_entry_t pte;
2170
2171                         pte = atomic_swap_long(ptep, v);
2172                         if (pte != 0) {
2173                                 kprintf("install pgtbl mixup 0x%016jx "
2174                                         "old/new 0x%016jx/0x%016jx\n",
2175                                         (intmax_t)ptepindex, pte, v);
2176                         }
2177                 }
2178         }
2179         vm_page_wakeup(m);
2180
2181         /*
2182          * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2183          */
2184 notnew:
2185         if (pvp) {
2186                 KKASSERT(pvp->pv_m != NULL);
2187                 ptep = pv_pte_lookup(pvp, ptepindex);
2188                 v = VM_PAGE_TO_PHYS(pv->pv_m) |
2189                     (pmap->pmap_bits[PG_U_IDX] |
2190                      pmap->pmap_bits[PG_RW_IDX] |
2191                      pmap->pmap_bits[PG_V_IDX] |
2192                      pmap->pmap_bits[PG_A_IDX] |
2193                      pmap->pmap_bits[PG_M_IDX]);
2194                 if (*ptep != v) {
2195                         kprintf("mismatched upper level pt %016jx/%016jx\n",
2196                                 *ptep, v);
2197                 }
2198         }
2199         if (pvpp)
2200                 *pvpp = pvp;
2201         else if (pvp)
2202                 pv_put(pvp);
2203         return (pv);
2204 }
2205
2206 /*
2207  * This version of pmap_allocpte() checks for possible segment optimizations
2208  * that would allow page-table sharing.  It can be called for terminal
2209  * page or page table page ptepindex's.
2210  *
2211  * The function is called with page table page ptepindex's for fictitious
2212  * and unmanaged terminal pages.  That is, we don't want to allocate a
2213  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
2214  * for this case.
2215  *
2216  * This function can return a pv and *pvpp associated with the passed in pmap
2217  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
2218  * an unmanaged page table page will be entered into the pass in pmap.
2219  */
2220 static
2221 pv_entry_t
2222 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2223                   vm_map_entry_t entry, vm_offset_t va)
2224 {
2225         vm_object_t object;
2226         pmap_t obpmap;
2227         pmap_t *obpmapp;
2228         vm_offset_t b;
2229         pv_entry_t pte_pv;      /* in original or shared pmap */
2230         pv_entry_t pt_pv;       /* in original or shared pmap */
2231         pv_entry_t proc_pd_pv;  /* in original pmap */
2232         pv_entry_t proc_pt_pv;  /* in original pmap */
2233         pv_entry_t xpv;         /* PT in shared pmap */
2234         pd_entry_t *pt;         /* PT entry in PD of original pmap */
2235         pd_entry_t opte;        /* contents of *pt */
2236         pd_entry_t npte;        /* contents of *pt */
2237         vm_page_t m;
2238
2239 retry:
2240         /*
2241          * Basic tests, require a non-NULL vm_map_entry, require proper
2242          * alignment and type for the vm_map_entry, require that the
2243          * underlying object already be allocated.
2244          *
2245          * We allow almost any type of object to use this optimization.
2246          * The object itself does NOT have to be sized to a multiple of the
2247          * segment size, but the memory mapping does.
2248          *
2249          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2250          *     won't work as expected.
2251          */
2252         if (entry == NULL ||
2253             pmap_mmu_optimize == 0 ||                   /* not enabled */
2254             (pmap->pm_flags & PMAP_HVM) ||              /* special pmap */
2255             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal or pt */
2256             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2257             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2258             entry->object.vm_object == NULL ||          /* needs VM object */
2259             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2260             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2261             (entry->offset & SEG_MASK) ||               /* must be aligned */
2262             (entry->start & SEG_MASK)) {
2263                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2264         }
2265
2266         /*
2267          * Make sure the full segment can be represented.
2268          */
2269         b = va & ~(vm_offset_t)SEG_MASK;
2270         if (b < entry->start || b + SEG_SIZE > entry->end)
2271                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2272
2273         /*
2274          * If the full segment can be represented dive the VM object's
2275          * shared pmap, allocating as required.
2276          */
2277         object = entry->object.vm_object;
2278
2279         if (entry->protection & VM_PROT_WRITE)
2280                 obpmapp = &object->md.pmap_rw;
2281         else
2282                 obpmapp = &object->md.pmap_ro;
2283
2284 #ifdef PMAP_DEBUG2
2285         if (pmap_enter_debug > 0) {
2286                 --pmap_enter_debug;
2287                 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2288                         "obpmapp %p %p\n",
2289                         va, entry->protection, object,
2290                         obpmapp, *obpmapp);
2291                 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2292                         entry, entry->start, entry->end);
2293         }
2294 #endif
2295
2296         /*
2297          * We allocate what appears to be a normal pmap but because portions
2298          * of this pmap are shared with other unrelated pmaps we have to
2299          * set pm_active to point to all cpus.
2300          *
2301          * XXX Currently using pmap_spin to interlock the update, can't use
2302          *     vm_object_hold/drop because the token might already be held
2303          *     shared OR exclusive and we don't know.
2304          */
2305         while ((obpmap = *obpmapp) == NULL) {
2306                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2307                 pmap_pinit_simple(obpmap);
2308                 pmap_pinit2(obpmap);
2309                 spin_lock(&pmap_spin);
2310                 if (*obpmapp != NULL) {
2311                         /*
2312                          * Handle race
2313                          */
2314                         spin_unlock(&pmap_spin);
2315                         pmap_release(obpmap);
2316                         pmap_puninit(obpmap);
2317                         kfree(obpmap, M_OBJPMAP);
2318                         obpmap = *obpmapp; /* safety */
2319                 } else {
2320                         obpmap->pm_active = smp_active_mask;
2321                         obpmap->pm_flags |= PMAP_SEGSHARED;
2322                         *obpmapp = obpmap;
2323                         spin_unlock(&pmap_spin);
2324                 }
2325         }
2326
2327         /*
2328          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2329          * pte/pt using the shared pmap from the object but also adjust
2330          * the process pmap's page table page as a side effect.
2331          */
2332
2333         /*
2334          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2335          * we will return.  This is true if ptepindex represents a terminal
2336          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2337          * the PD.
2338          */
2339         pt_pv = NULL;
2340         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2341         if (ptepindex >= pmap_pt_pindex(0))
2342                 xpv = pte_pv;
2343         else
2344                 xpv = pt_pv;
2345
2346         /*
2347          * Resolve the PD in the process pmap so we can properly share the
2348          * page table page.  Lock order is bottom-up (leaf first)!
2349          *
2350          * NOTE: proc_pt_pv can be NULL.
2351          */
2352         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b), NULL);
2353         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2354 #ifdef PMAP_DEBUG2
2355         if (pmap_enter_debug > 0) {
2356                 --pmap_enter_debug;
2357                 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2358                         proc_pt_pv,
2359                         (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2360                         proc_pd_pv,
2361                         va);
2362         }
2363 #endif
2364
2365         /*
2366          * xpv is the page table page pv from the shared object
2367          * (for convenience), from above.
2368          *
2369          * Calculate the pte value for the PT to load into the process PD.
2370          * If we have to change it we must properly dispose of the previous
2371          * entry.
2372          */
2373         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2374         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2375                (pmap->pmap_bits[PG_U_IDX] |
2376                 pmap->pmap_bits[PG_RW_IDX] |
2377                 pmap->pmap_bits[PG_V_IDX] |
2378                 pmap->pmap_bits[PG_A_IDX] |
2379                 pmap->pmap_bits[PG_M_IDX]);
2380
2381         /*
2382          * Dispose of previous page table page if it was local to the
2383          * process pmap.  If the old pt is not empty we cannot dispose of it
2384          * until we clean it out.  This case should not arise very often so
2385          * it is not optimized.
2386          */
2387         if (proc_pt_pv) {
2388                 pmap_inval_bulk_t bulk;
2389
2390                 if (proc_pt_pv->pv_m->wire_count != 1) {
2391                         pv_put(proc_pd_pv);
2392                         pv_put(proc_pt_pv);
2393                         pv_put(pt_pv);
2394                         pv_put(pte_pv);
2395                         pmap_remove(pmap,
2396                                     va & ~(vm_offset_t)SEG_MASK,
2397                                     (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2398                         goto retry;
2399                 }
2400
2401                 /*
2402                  * The release call will indirectly clean out *pt
2403                  */
2404                 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2405                 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2406                 pmap_inval_bulk_flush(&bulk);
2407                 proc_pt_pv = NULL;
2408                 /* relookup */
2409                 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2410         }
2411
2412         /*
2413          * Handle remaining cases.
2414          */
2415         if (*pt == 0) {
2416                 atomic_swap_long(pt, npte);
2417                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2418                 vm_page_wire_quick(proc_pd_pv->pv_m);   /* proc pd for sh pt */
2419                 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2420         } else if (*pt != npte) {
2421                 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2422
2423 #if 0
2424                 opte = pte_load_clear(pt);
2425                 KKASSERT(opte && opte != npte);
2426
2427                 *pt = npte;
2428 #endif
2429                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2430
2431                 /*
2432                  * Clean up opte, bump the wire_count for the process
2433                  * PD page representing the new entry if it was
2434                  * previously empty.
2435                  *
2436                  * If the entry was not previously empty and we have
2437                  * a PT in the proc pmap then opte must match that
2438                  * pt.  The proc pt must be retired (this is done
2439                  * later on in this procedure).
2440                  *
2441                  * NOTE: replacing valid pte, wire_count on proc_pd_pv
2442                  * stays the same.
2443                  */
2444                 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2445                 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2446                 if (vm_page_unwire_quick(m)) {
2447                         panic("pmap_allocpte_seg: "
2448                               "bad wire count %p",
2449                               m);
2450                 }
2451         }
2452
2453         /*
2454          * The existing process page table was replaced and must be destroyed
2455          * here.
2456          */
2457         if (proc_pd_pv)
2458                 pv_put(proc_pd_pv);
2459         if (pvpp)
2460                 *pvpp = pt_pv;
2461         else
2462                 pv_put(pt_pv);
2463
2464         return (pte_pv);
2465 }
2466
2467 /*
2468  * Release any resources held by the given physical map.
2469  *
2470  * Called when a pmap initialized by pmap_pinit is being released.  Should
2471  * only be called if the map contains no valid mappings.
2472  */
2473 struct pmap_release_info {
2474         pmap_t  pmap;
2475         int     retry;
2476         pv_entry_t pvp;
2477 };
2478
2479 static int pmap_release_callback(pv_entry_t pv, void *data);
2480
2481 void
2482 pmap_release(struct pmap *pmap)
2483 {
2484         struct pmap_release_info info;
2485
2486         KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2487                 ("pmap still active! %016jx",
2488                 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2489
2490         /*
2491          * There is no longer a pmap_list, if there were we would remove the
2492          * pmap from it here.
2493          */
2494
2495         /*
2496          * Pull pv's off the RB tree in order from low to high and release
2497          * each page.
2498          */
2499         info.pmap = pmap;
2500         do {
2501                 info.retry = 0;
2502                 info.pvp = NULL;
2503
2504                 spin_lock(&pmap->pm_spin);
2505                 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2506                         pmap_release_callback, &info);
2507                 spin_unlock(&pmap->pm_spin);
2508
2509                 if (info.pvp)
2510                         pv_put(info.pvp);
2511         } while (info.retry);
2512
2513
2514         /*
2515          * One resident page (the pml4 page) should remain.
2516          * No wired pages should remain.
2517          */
2518 #if 1
2519         if (pmap->pm_stats.resident_count !=
2520             ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1) ||
2521             pmap->pm_stats.wired_count != 0) {
2522                 kprintf("fatal pmap problem - pmap %p flags %08x "
2523                         "rescnt=%jd wirecnt=%jd\n",
2524                         pmap,
2525                         pmap->pm_flags,
2526                         pmap->pm_stats.resident_count,
2527                         pmap->pm_stats.wired_count);
2528                 tsleep(pmap, 0, "DEAD", 0);
2529         }
2530 #else
2531         KKASSERT(pmap->pm_stats.resident_count ==
2532                  ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2533         KKASSERT(pmap->pm_stats.wired_count == 0);
2534 #endif
2535 }
2536
2537 /*
2538  * Called from low to high.  We must cache the proper parent pv so we
2539  * can adjust its wired count.
2540  */
2541 static int
2542 pmap_release_callback(pv_entry_t pv, void *data)
2543 {
2544         struct pmap_release_info *info = data;
2545         pmap_t pmap = info->pmap;
2546         vm_pindex_t pindex;
2547         int r;
2548
2549         /*
2550          * Acquire a held and locked pv, check for release race
2551          */
2552         pindex = pv->pv_pindex;
2553         if (info->pvp == pv) {
2554                 spin_unlock(&pmap->pm_spin);
2555                 info->pvp = NULL;
2556         } else if (pv_hold_try(pv)) {
2557                 spin_unlock(&pmap->pm_spin);
2558         } else {
2559                 spin_unlock(&pmap->pm_spin);
2560                 pv_lock(pv);
2561                 pv_unlock(pv);
2562                 pv_drop(pv);
2563                 info->retry = 1;
2564                 return -1;
2565         }
2566         KKASSERT(pv->pv_pmap == pmap && pindex == pv->pv_pindex);
2567
2568         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2569                 /*
2570                  * I am PTE, parent is PT
2571                  */
2572                 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2573                 pindex += NUPTE_TOTAL;
2574         } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2575                 /*
2576                  * I am PT, parent is PD
2577                  */
2578                 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2579                 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2580         } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2581                 /*
2582                  * I am PD, parent is PDP
2583                  */
2584                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2585                          NPDPEPGSHIFT;
2586                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2587         } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2588                 /*
2589                  * I am PDP, parent is PML4 (there's only one)
2590                  */
2591 #if 0
2592                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2593                            NUPD_TOTAL) >> NPML4EPGSHIFT;
2594                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2595 #endif
2596                 pindex = pmap_pml4_pindex();
2597         } else {
2598                 /*
2599                  * parent is NULL
2600                  */
2601                 if (info->pvp) {
2602                         pv_put(info->pvp);
2603                         info->pvp = NULL;
2604                 }
2605                 pindex = 0;
2606         }
2607         if (pindex) {
2608                 if (info->pvp && info->pvp->pv_pindex != pindex) {
2609                         pv_put(info->pvp);
2610                         info->pvp = NULL;
2611                 }
2612                 if (info->pvp == NULL)
2613                         info->pvp = pv_get(pmap, pindex, NULL);
2614         } else {
2615                 if (info->pvp) {
2616                         pv_put(info->pvp);
2617                         info->pvp = NULL;
2618                 }
2619         }
2620         r = pmap_release_pv(pv, info->pvp, NULL);
2621         spin_lock(&pmap->pm_spin);
2622
2623         return(r);
2624 }
2625
2626 /*
2627  * Called with held (i.e. also locked) pv.  This function will dispose of
2628  * the lock along with the pv.
2629  *
2630  * If the caller already holds the locked parent page table for pv it
2631  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2632  * pass NULL for pvp.
2633  */
2634 static int
2635 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2636 {
2637         vm_page_t p;
2638
2639         /*
2640          * The pmap is currently not spinlocked, pv is held+locked.
2641          * Remove the pv's page from its parent's page table.  The
2642          * parent's page table page's wire_count will be decremented.
2643          *
2644          * This will clean out the pte at any level of the page table.
2645          * If smp != 0 all cpus are affected.
2646          *
2647          * Do not tear-down recursively, its faster to just let the
2648          * release run its course.
2649          */
2650         pmap_remove_pv_pte(pv, pvp, bulk, 0);
2651
2652         /*
2653          * Terminal pvs are unhooked from their vm_pages.  Because
2654          * terminal pages aren't page table pages they aren't wired
2655          * by us, so we have to be sure not to unwire them either.
2656          */
2657         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2658                 pmap_remove_pv_page(pv);
2659                 goto skip;
2660         }
2661
2662         /*
2663          * We leave the top-level page table page cached, wired, and
2664          * mapped in the pmap until the dtor function (pmap_puninit())
2665          * gets called.
2666          *
2667          * Since we are leaving the top-level pv intact we need
2668          * to break out of what would otherwise be an infinite loop.
2669          */
2670         if (pv->pv_pindex == pmap_pml4_pindex()) {
2671                 pv_put(pv);
2672                 return(-1);
2673         }
2674
2675         /*
2676          * For page table pages (other than the top-level page),
2677          * remove and free the vm_page.  The representitive mapping
2678          * removed above by pmap_remove_pv_pte() did not undo the
2679          * last wire_count so we have to do that as well.
2680          */
2681         p = pmap_remove_pv_page(pv);
2682         vm_page_busy_wait(p, FALSE, "pmaprl");
2683         if (p->wire_count != 1) {
2684                 kprintf("p->wire_count was %016lx %d\n",
2685                         pv->pv_pindex, p->wire_count);
2686         }
2687         KKASSERT(p->wire_count == 1);
2688         KKASSERT(p->flags & PG_UNMANAGED);
2689
2690         vm_page_unwire(p, 0);
2691         KKASSERT(p->wire_count == 0);
2692
2693         vm_page_free(p);
2694 skip:
2695         pv_free(pv, pvp);
2696
2697         return 0;
2698 }
2699
2700 /*
2701  * This function will remove the pte associated with a pv from its parent.
2702  * Terminal pv's are supported.  All cpus specified by (bulk) are properly
2703  * invalidated.
2704  *
2705  * The wire count will be dropped on the parent page table.  The wire
2706  * count on the page being removed (pv->pv_m) from the parent page table
2707  * is NOT touched.  Note that terminal pages will not have any additional
2708  * wire counts while page table pages will have at least one representing
2709  * the mapping, plus others representing sub-mappings.
2710  *
2711  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2712  *       pages and user page table and terminal pages.
2713  *
2714  * NOTE: The pte being removed might be unmanaged, and the pv supplied might
2715  *       be freshly allocated and not imply that the pte is managed.  In this
2716  *       case pv->pv_m should be NULL.
2717  *
2718  * The pv must be locked.  The pvp, if supplied, must be locked.  All
2719  * supplied pv's will remain locked on return.
2720  *
2721  * XXX must lock parent pv's if they exist to remove pte XXX
2722  */
2723 static
2724 void
2725 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2726                    int destroy)
2727 {
2728         vm_pindex_t ptepindex = pv->pv_pindex;
2729         pmap_t pmap = pv->pv_pmap;
2730         vm_page_t p;
2731         int gotpvp = 0;
2732
2733         KKASSERT(pmap);
2734
2735         if (ptepindex == pmap_pml4_pindex()) {
2736                 /*
2737                  * We are the top level PML4E table, there is no parent.
2738                  */
2739                 p = pmap->pm_pmlpv->pv_m;
2740                 KKASSERT(pv->pv_m == p);        /* debugging */
2741         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2742                 /*
2743                  * Remove a PDP page from the PML4E.  This can only occur
2744                  * with user page tables.  We do not have to lock the
2745                  * pml4 PV so just ignore pvp.
2746                  */
2747                 vm_pindex_t pml4_pindex;
2748                 vm_pindex_t pdp_index;
2749                 pml4_entry_t *pdp;
2750
2751                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2752                 if (pvp == NULL) {
2753                         pml4_pindex = pmap_pml4_pindex();
2754                         pvp = pv_get(pv->pv_pmap, pml4_pindex, NULL);
2755                         KKASSERT(pvp);
2756                         gotpvp = 1;
2757                 }
2758
2759                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2760                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2761                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2762                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2763                 KKASSERT(pv->pv_m == p);        /* debugging */
2764         } else if (ptepindex >= pmap_pd_pindex(0)) {
2765                 /*
2766                  * Remove a PD page from the PDP
2767                  *
2768                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2769                  *                   of a simple pmap because it stops at
2770                  *                   the PD page.
2771                  */
2772                 vm_pindex_t pdp_pindex;
2773                 vm_pindex_t pd_index;
2774                 pdp_entry_t *pd;
2775
2776                 pd_index = ptepindex - pmap_pd_pindex(0);
2777
2778                 if (pvp == NULL) {
2779                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2780                                      (pd_index >> NPML4EPGSHIFT);
2781                         pvp = pv_get(pv->pv_pmap, pdp_pindex, NULL);
2782                         gotpvp = 1;
2783                 }
2784
2785                 if (pvp) {
2786                         pd = pv_pte_lookup(pvp, pd_index &
2787                                                 ((1ul << NPDPEPGSHIFT) - 1));
2788                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2789                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2790                         pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2791                 } else {
2792                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2793                         p = pv->pv_m;           /* degenerate test later */
2794                 }
2795                 KKASSERT(pv->pv_m == p);        /* debugging */
2796         } else if (ptepindex >= pmap_pt_pindex(0)) {
2797                 /*
2798                  *  Remove a PT page from the PD
2799                  */
2800                 vm_pindex_t pd_pindex;
2801                 vm_pindex_t pt_index;
2802                 pd_entry_t *pt;
2803
2804                 pt_index = ptepindex - pmap_pt_pindex(0);
2805
2806                 if (pvp == NULL) {
2807                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2808                                     (pt_index >> NPDPEPGSHIFT);
2809                         pvp = pv_get(pv->pv_pmap, pd_pindex, NULL);
2810                         KKASSERT(pvp);
2811                         gotpvp = 1;
2812                 }
2813
2814                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2815 #if 0
2816                 KASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0,
2817                         ("*pt unexpectedly invalid %016jx "
2818                          "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
2819                         *pt, gotpvp, ptepindex, pt_index, pv, pvp));
2820                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2821 #else
2822                 if ((*pt & pmap->pmap_bits[PG_V_IDX]) == 0) {
2823                         kprintf("*pt unexpectedly invalid %016jx "
2824                                 "gotpvp=%d ptepindex=%ld ptindex=%ld "
2825                                 "pv=%p pvp=%p\n",
2826                                 *pt, gotpvp, ptepindex, pt_index, pv, pvp);
2827                         tsleep(pt, 0, "DEAD", 0);
2828                         p = pv->pv_m;
2829                 } else {
2830                         p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2831                 }
2832 #endif
2833                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2834                 KKASSERT(pv->pv_m == p);        /* debugging */
2835         } else {
2836                 /*
2837                  * Remove a PTE from the PT page.  The PV might exist even if
2838                  * the PTE is not managed, in whichcase pv->pv_m should be
2839                  * NULL.
2840                  *
2841                  * NOTE: Userland pmaps manage the parent PT/PD/PDP page
2842                  *       table pages but the kernel_pmap does not.
2843                  *
2844                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2845                  *       pv is a pte_pv so we can safely lock pt_pv.
2846                  *
2847                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2848                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2849                  *       terminal ptes.
2850                  */
2851                 vm_pindex_t pt_pindex;
2852                 pt_entry_t *ptep;
2853                 pt_entry_t pte;
2854                 vm_offset_t va;
2855
2856                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2857                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2858
2859                 if (ptepindex >= NUPTE_USER) {
2860                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2861                         KKASSERT(pvp == NULL);
2862                         /* pvp remains NULL */
2863                 } else {
2864                         if (pvp == NULL) {
2865                                 pt_pindex = NUPTE_TOTAL +
2866                                             (ptepindex >> NPDPEPGSHIFT);
2867                                 pvp = pv_get(pv->pv_pmap, pt_pindex, NULL);
2868                                 KKASSERT(pvp);
2869                                 gotpvp = 1;
2870                         }
2871                         ptep = pv_pte_lookup(pvp, ptepindex &
2872                                                   ((1ul << NPDPEPGSHIFT) - 1));
2873                 }
2874                 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2875                 if (bulk == NULL)               /* XXX */
2876                         cpu_invlpg((void *)va); /* XXX */
2877
2878                 /*
2879                  * Now update the vm_page_t
2880                  */
2881                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
2882                     (pte & pmap->pmap_bits[PG_V_IDX])) {
2883                         /*
2884                          * Valid managed page, adjust (p).
2885                          */
2886                         if (pte & pmap->pmap_bits[PG_DEVICE_IDX]) {
2887                                 p = pv->pv_m;
2888                         } else {
2889                                 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2890                                 KKASSERT(pv->pv_m == p);
2891                         }
2892                         if (pte & pmap->pmap_bits[PG_M_IDX]) {
2893                                 if (pmap_track_modified(ptepindex))
2894                                         vm_page_dirty(p);
2895                         }
2896                         if (pte & pmap->pmap_bits[PG_A_IDX]) {
2897                                 vm_page_flag_set(p, PG_REFERENCED);
2898                         }
2899                 } else {
2900                         /*
2901                          * Unmanaged page, do not try to adjust the vm_page_t.
2902                          * pv could be freshly allocated for a pmap_enter(),
2903                          * replacing an unmanaged page with a managed one.
2904                          *
2905                          * pv->pv_m might reflect the new page and not the
2906                          * existing page.
2907                          *
2908                          * We could extract p from the physical address and
2909                          * adjust it but we explicitly do not for unmanaged
2910                          * pages.
2911                          */
2912                         p = NULL;
2913                 }
2914                 if (pte & pmap->pmap_bits[PG_W_IDX])
2915                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
2916                 if (pte & pmap->pmap_bits[PG_G_IDX])
2917                         cpu_invlpg((void *)va);
2918         }
2919
2920         /*
2921          * If requested, scrap the underlying pv->pv_m and the underlying
2922          * pv.  If this is a page-table-page we must also free the page.
2923          *
2924          * pvp must be returned locked.
2925          */
2926         if (destroy == 1) {
2927                 /*
2928                  * page table page (PT, PD, PDP, PML4), caller was responsible
2929                  * for testing wired_count.
2930                  */
2931                 KKASSERT(pv->pv_m->wire_count == 1);
2932                 p = pmap_remove_pv_page(pv);
2933                 pv_free(pv, pvp);
2934                 pv = NULL;
2935
2936                 vm_page_busy_wait(p, FALSE, "pgpun");
2937                 vm_page_unwire(p, 0);
2938                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2939                 vm_page_free(p);
2940         } else if (destroy == 2) {
2941                 /*
2942                  * Normal page, remove from pmap and leave the underlying
2943                  * page untouched.
2944                  */
2945                 pmap_remove_pv_page(pv);
2946                 pv_free(pv, pvp);
2947                 pv = NULL;              /* safety */
2948         }
2949
2950         /*
2951          * If we acquired pvp ourselves then we are responsible for
2952          * recursively deleting it.
2953          */
2954         if (pvp && gotpvp) {
2955                 /*
2956                  * Recursively destroy higher-level page tables.
2957                  *
2958                  * This is optional.  If we do not, they will still
2959                  * be destroyed when the process exits.
2960                  *
2961                  * NOTE: Do not destroy pv_entry's with extra hold refs,
2962                  *       a caller may have unlocked it and intends to
2963                  *       continue to use it.
2964                  */
2965                 if (pmap_dynamic_delete &&
2966                     pvp->pv_m &&
2967                     pvp->pv_m->wire_count == 1 &&
2968                     (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
2969                     pvp->pv_pindex != pmap_pml4_pindex()) {
2970                         if (pmap_dynamic_delete == 2)
2971                                 kprintf("A %jd %08x\n", pvp->pv_pindex, pvp->pv_hold);
2972                         if (pmap != &kernel_pmap) {
2973                                 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
2974                                 pvp = NULL;     /* safety */
2975                         } else {
2976                                 kprintf("Attempt to remove kernel_pmap pindex "
2977                                         "%jd\n", pvp->pv_pindex);
2978                                 pv_put(pvp);
2979                         }
2980                 } else {
2981                         pv_put(pvp);
2982                 }
2983         }
2984 }
2985
2986 /*
2987  * Remove the vm_page association to a pv.  The pv must be locked.
2988  */
2989 static
2990 vm_page_t
2991 pmap_remove_pv_page(pv_entry_t pv)
2992 {
2993         vm_page_t m;
2994
2995         m = pv->pv_m;
2996         vm_page_spin_lock(m);
2997         KKASSERT(m && m == pv->pv_m);
2998         pv->pv_m = NULL;
2999         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3000         pmap_page_stats_deleting(m);
3001         if (TAILQ_EMPTY(&m->md.pv_list))
3002                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3003         vm_page_spin_unlock(m);
3004
3005         return(m);
3006 }
3007
3008 /*
3009  * Grow the number of kernel page table entries, if needed.
3010  *
3011  * This routine is always called to validate any address space
3012  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
3013  * space below KERNBASE.
3014  *
3015  * kernel_map must be locked exclusively by the caller.
3016  */
3017 void
3018 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
3019 {
3020         vm_paddr_t paddr;
3021         vm_offset_t ptppaddr;
3022         vm_page_t nkpg;
3023         pd_entry_t *pt, newpt;
3024         pdp_entry_t newpd;
3025         int update_kernel_vm_end;
3026
3027         /*
3028          * bootstrap kernel_vm_end on first real VM use
3029          */
3030         if (kernel_vm_end == 0) {
3031                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
3032                 nkpt = 0;
3033                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3034                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
3035                                         ~(PAGE_SIZE * NPTEPG - 1);
3036                         nkpt++;
3037                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
3038                                 kernel_vm_end = kernel_map.max_offset;
3039                                 break;
3040                         }
3041                 }
3042         }
3043
3044         /*
3045          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
3046          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
3047          * do not want to force-fill 128G worth of page tables.
3048          */
3049         if (kstart < KERNBASE) {
3050                 if (kstart > kernel_vm_end)
3051                         kstart = kernel_vm_end;
3052                 KKASSERT(kend <= KERNBASE);
3053                 update_kernel_vm_end = 1;
3054         } else {
3055                 update_kernel_vm_end = 0;
3056         }
3057
3058         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
3059         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
3060
3061         if (kend - 1 >= kernel_map.max_offset)
3062                 kend = kernel_map.max_offset;
3063
3064         while (kstart < kend) {
3065                 pt = pmap_pt(&kernel_pmap, kstart);
3066                 if (pt == NULL) {
3067                         /* We need a new PD entry */
3068                         nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3069                                              VM_ALLOC_NORMAL |
3070                                              VM_ALLOC_SYSTEM |
3071                                              VM_ALLOC_INTERRUPT);
3072                         if (nkpg == NULL) {
3073                                 panic("pmap_growkernel: no memory to grow "
3074                                       "kernel");
3075                         }
3076                         paddr = VM_PAGE_TO_PHYS(nkpg);
3077                         pmap_zero_page(paddr);
3078                         newpd = (pdp_entry_t)
3079                             (paddr |
3080                             kernel_pmap.pmap_bits[PG_V_IDX] |
3081                             kernel_pmap.pmap_bits[PG_RW_IDX] |
3082                             kernel_pmap.pmap_bits[PG_A_IDX] |
3083                             kernel_pmap.pmap_bits[PG_M_IDX]);
3084                         *pmap_pd(&kernel_pmap, kstart) = newpd;
3085                         continue; /* try again */
3086                 }
3087                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3088                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
3089                                  ~(PAGE_SIZE * NPTEPG - 1);
3090                         if (kstart - 1 >= kernel_map.max_offset) {
3091                                 kstart = kernel_map.max_offset;
3092                                 break;
3093                         }
3094                         continue;
3095                 }
3096
3097                 /*
3098                  * We need a new PT
3099                  *
3100                  * This index is bogus, but out of the way
3101                  */
3102                 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3103                                      VM_ALLOC_NORMAL |
3104                                      VM_ALLOC_SYSTEM |
3105                                      VM_ALLOC_INTERRUPT);
3106                 if (nkpg == NULL)
3107                         panic("pmap_growkernel: no memory to grow kernel");
3108
3109                 vm_page_wire(nkpg);
3110                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
3111                 pmap_zero_page(ptppaddr);
3112                 newpt = (pd_entry_t)(ptppaddr |
3113                                      kernel_pmap.pmap_bits[PG_V_IDX] |
3114                                      kernel_pmap.pmap_bits[PG_RW_IDX] |
3115                                      kernel_pmap.pmap_bits[PG_A_IDX] |
3116                                      kernel_pmap.pmap_bits[PG_M_IDX]);
3117                 atomic_swap_long(pmap_pt(&kernel_pmap, kstart), newpt);
3118
3119                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3120                           ~(PAGE_SIZE * NPTEPG - 1);
3121
3122                 if (kstart - 1 >= kernel_map.max_offset) {
3123                         kstart = kernel_map.max_offset;
3124                         break;
3125                 }
3126         }
3127
3128         /*
3129          * Only update kernel_vm_end for areas below KERNBASE.
3130          */
3131         if (update_kernel_vm_end && kernel_vm_end < kstart)
3132                 kernel_vm_end = kstart;
3133 }
3134
3135 /*
3136  *      Add a reference to the specified pmap.
3137  */
3138 void
3139 pmap_reference(pmap_t pmap)
3140 {
3141         if (pmap != NULL)
3142                 atomic_add_int(&pmap->pm_count, 1);
3143 }
3144
3145 /***************************************************
3146  * page management routines.
3147  ***************************************************/
3148
3149 /*
3150  * Hold a pv without locking it
3151  */
3152 static void
3153 pv_hold(pv_entry_t pv)
3154 {
3155         atomic_add_int(&pv->pv_hold, 1);
3156 }
3157
3158 /*
3159  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
3160  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
3161  * the pv properly.
3162  *
3163  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3164  * pv list via its page) must be held by the caller in order to stabilize
3165  * the pv.
3166  */
3167 static int
3168 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3169 {
3170         u_int count;
3171
3172         /*
3173          * Critical path shortcut expects pv to already have one ref
3174          * (for the pv->pv_pmap).
3175          */
3176         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3177 #ifdef PMAP_DEBUG
3178                 pv->pv_func = func;
3179                 pv->pv_line = lineno;
3180 #endif
3181                 return TRUE;
3182         }
3183
3184         for (;;) {
3185                 count = pv->pv_hold;
3186                 cpu_ccfence();
3187                 if ((count & PV_HOLD_LOCKED) == 0) {
3188                         if (atomic_cmpset_int(&pv->pv_hold, count,
3189                                               (count + 1) | PV_HOLD_LOCKED)) {
3190 #ifdef PMAP_DEBUG
3191                                 pv->pv_func = func;
3192                                 pv->pv_line = lineno;
3193 #endif
3194                                 return TRUE;
3195                         }
3196                 } else {
3197                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3198                                 return FALSE;
3199                 }
3200                 /* retry */
3201         }
3202 }
3203
3204 /*
3205  * Drop a previously held pv_entry which could not be locked, allowing its
3206  * destruction.
3207  *
3208  * Must not be called with a spinlock held as we might zfree() the pv if it
3209  * is no longer associated with a pmap and this was the last hold count.
3210  */
3211 static void
3212 pv_drop(pv_entry_t pv)
3213 {
3214         u_int count;
3215
3216         for (;;) {
3217                 count = pv->pv_hold;
3218                 cpu_ccfence();
3219                 KKASSERT((count & PV_HOLD_MASK) > 0);
3220                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3221                          (PV_HOLD_LOCKED | 1));
3222                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3223                         if ((count & PV_HOLD_MASK) == 1) {
3224 #ifdef PMAP_DEBUG2
3225                                 if (pmap_enter_debug > 0) {
3226                                         --pmap_enter_debug;
3227                                         kprintf("pv_drop: free pv %p\n", pv);
3228                                 }
3229 #endif
3230                                 KKASSERT(count == 1);
3231                                 KKASSERT(pv->pv_pmap == NULL);
3232                                 zfree(pvzone, pv);
3233                         }
3234                         return;
3235                 }
3236                 /* retry */
3237         }
3238 }
3239
3240 /*
3241  * Find or allocate the requested PV entry, returning a locked, held pv.
3242  *
3243  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3244  * for the caller and one representing the pmap and vm_page association.
3245  *
3246  * If (*isnew) is zero, the returned pv will have only one hold count.
3247  *
3248  * Since both associations can only be adjusted while the pv is locked,
3249  * together they represent just one additional hold.
3250  */
3251 static
3252 pv_entry_t
3253 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3254 {
3255         pv_entry_t pv;
3256         pv_entry_t pnew = NULL;
3257
3258         spin_lock(&pmap->pm_spin);
3259         for (;;) {
3260                 /*
3261                  * Shortcut cache
3262                  */
3263                 pv = pmap->pm_pvhint;
3264                 cpu_ccfence();
3265                 if (pv == NULL ||
3266                     pv->pv_pmap != pmap ||
3267                     pv->pv_pindex != pindex) {
3268                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3269                                                         pindex);
3270                 }
3271                 if (pv == NULL) {
3272                         vm_pindex_t *pmark;
3273
3274                         /*
3275                          * We need to stage a new pv entry
3276                          */
3277                         if (pnew == NULL) {
3278                                 spin_unlock(&pmap->pm_spin);
3279                                 pnew = zalloc(pvzone);
3280                                 spin_lock(&pmap->pm_spin);
3281                                 continue;
3282                         }
3283
3284                         /*
3285                          * We need to block if someone is holding a
3286                          * placemarker.  The exclusive spinlock is a
3287                          * sufficient interlock, as long as we determine
3288                          * the placemarker has not been aquired we do not
3289                          * need to get it.
3290                          */
3291                         pmark = pmap_placemarker_hash(pmap, pindex);
3292
3293                         if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3294                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3295                                 ssleep(pmark, &pmap->pm_spin, 0, "pvplc", 0);
3296                                 continue;
3297                         }
3298
3299                         /*
3300                          * Setup the new entry
3301                          */
3302                         pnew->pv_pmap = pmap;
3303                         pnew->pv_pindex = pindex;
3304                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
3305 #ifdef PMAP_DEBUG
3306                         pnew->pv_func = func;
3307                         pnew->pv_line = lineno;
3308                         if (pnew->pv_line_lastfree > 0) {
3309                                 pnew->pv_line_lastfree =
3310                                                 -pnew->pv_line_lastfree;
3311                         }
3312 #endif
3313                         pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3314                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
3315                         spin_unlock(&pmap->pm_spin);
3316                         *isnew = 1;
3317
3318                         KKASSERT(pv == NULL);
3319                         return(pnew);
3320                 }
3321
3322                 /*
3323                  * We have an entry, clean up any staged pv we had allocated,
3324                  * then block until we can lock the entry.
3325                  */
3326                 if (pnew) {
3327                         spin_unlock(&pmap->pm_spin);
3328                         zfree(pvzone, pnew);
3329                         pnew = NULL;
3330                         spin_lock(&pmap->pm_spin);
3331                         continue;
3332                 }
3333                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3334                         spin_unlock(&pmap->pm_spin);
3335                         KKASSERT(pv->pv_pmap == pmap &&
3336                                  pv->pv_pindex == pindex);
3337                         *isnew = 0;
3338                         return(pv);
3339                 }
3340                 spin_unlock(&pmap->pm_spin);
3341                 _pv_lock(pv PMAP_DEBUG_COPY);
3342                 pv_unlock(pv);
3343                 pv_drop(pv);
3344                 spin_lock(&pmap->pm_spin);
3345         }
3346 }
3347
3348 /*
3349  * Find the requested PV entry, returning a locked+held pv or NULL
3350  */
3351 static
3352 pv_entry_t
3353 _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp PMAP_DEBUG_DECL)
3354 {
3355         pv_entry_t pv;
3356
3357         spin_lock(&pmap->pm_spin);
3358         for (;;) {
3359                 /*
3360                  * Shortcut cache
3361                  */
3362                 pv = pmap->pm_pvhint;
3363                 cpu_ccfence();
3364                 if (pv == NULL ||
3365                     pv->pv_pmap != pmap ||
3366                     pv->pv_pindex != pindex) {
3367                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3368                                                         pindex);
3369                 }
3370                 if (pv == NULL) {
3371                         /*
3372                          * Block if there is a placemarker.  If we are to
3373                          * return it, we must also aquire the spot.
3374                          */
3375                         vm_pindex_t *pmark;
3376
3377                         pmark = pmap_placemarker_hash(pmap, pindex);
3378
3379                         if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3380                             ((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3381                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3382                                 ssleep(pmark, &pmap->pm_spin, 0, "pvpld", 0);
3383                                 continue;
3384                         }
3385                         if (pmarkp) {
3386                                 if (atomic_swap_long(pmark, pindex) !=
3387                                     PM_NOPLACEMARK) {
3388                                         panic("_pv_get: pmark race");
3389                                 }
3390                                 *pmarkp = pmark;
3391                         }
3392                         spin_unlock(&pmap->pm_spin);
3393                         return NULL;
3394                 }
3395                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3396                         pv_cache(pv, pindex);
3397                         spin_unlock(&pmap->pm_spin);
3398                         KKASSERT(pv->pv_pmap == pmap &&
3399                                  pv->pv_pindex == pindex);
3400                         return(pv);
3401                 }
3402                 spin_unlock(&pmap->pm_spin);
3403                 _pv_lock(pv PMAP_DEBUG_COPY);
3404                 pv_unlock(pv);
3405                 pv_drop(pv);
3406                 spin_lock(&pmap->pm_spin);
3407         }
3408 }
3409
3410 /*
3411  * Lookup, hold, and attempt to lock (pmap,pindex).
3412  *
3413  * If the entry does not exist NULL is returned and *errorp is set to 0
3414  *
3415  * If the entry exists and could be successfully locked it is returned and
3416  * errorp is set to 0.
3417  *
3418  * If the entry exists but could NOT be successfully locked it is returned
3419  * held and *errorp is set to 1.
3420  *
3421  * If the entry is placemarked by someone else NULL is returned and *errorp
3422  * is set to 1.
3423  */
3424 static
3425 pv_entry_t
3426 pv_get_try(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp, int *errorp)
3427 {
3428         pv_entry_t pv;
3429
3430         spin_lock_shared(&pmap->pm_spin);
3431
3432         pv = pmap->pm_pvhint;
3433         cpu_ccfence();
3434         if (pv == NULL ||
3435             pv->pv_pmap != pmap ||
3436             pv->pv_pindex != pindex) {
3437                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3438         }
3439
3440         if (pv == NULL) {
3441                 vm_pindex_t *pmark;
3442
3443                 pmark = pmap_placemarker_hash(pmap, pindex);
3444
3445                 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3446                         *errorp = 1;
3447                 } else if (pmarkp &&
3448                            atomic_cmpset_long(pmark, PM_NOPLACEMARK, pindex)) {
3449                         *errorp = 0;
3450                 } else {
3451                         /*
3452                          * Can't set a placemark with a NULL pmarkp, or if
3453                          * pmarkp is non-NULL but we failed to set our
3454                          * placemark.
3455                          */
3456                         *errorp = 1;
3457                 }
3458                 if (pmarkp)
3459                         *pmarkp = pmark;
3460                 spin_unlock_shared(&pmap->pm_spin);
3461
3462                 return NULL;
3463         }
3464
3465         /*
3466          * XXX This has problems if the lock is shared, why?
3467          */
3468         if (pv_hold_try(pv)) {
3469                 pv_cache(pv, pindex);   /* overwrite ok (shared lock) */
3470                 spin_unlock_shared(&pmap->pm_spin);
3471                 *errorp = 0;
3472                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3473                 return(pv);     /* lock succeeded */
3474         }
3475         spin_unlock_shared(&pmap->pm_spin);
3476         *errorp = 1;
3477
3478         return (pv);            /* lock failed */
3479 }
3480
3481 /*
3482  * Lock a held pv, keeping the hold count
3483  */
3484 static
3485 void
3486 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3487 {
3488         u_int count;
3489
3490         for (;;) {
3491                 count = pv->pv_hold;
3492                 cpu_ccfence();
3493                 if ((count & PV_HOLD_LOCKED) == 0) {
3494                         if (atomic_cmpset_int(&pv->pv_hold, count,
3495                                               count | PV_HOLD_LOCKED)) {
3496 #ifdef PMAP_DEBUG
3497                                 pv->pv_func = func;
3498                                 pv->pv_line = lineno;
3499 #endif
3500                                 return;
3501                         }
3502                         continue;
3503                 }
3504                 tsleep_interlock(pv, 0);
3505                 if (atomic_cmpset_int(&pv->pv_hold, count,
3506                                       count | PV_HOLD_WAITING)) {
3507 #ifdef PMAP_DEBUG2
3508                         if (pmap_enter_debug > 0) {
3509                                 --pmap_enter_debug;
3510                                 kprintf("pv waiting on %s:%d\n",
3511                                         pv->pv_func, pv->pv_line);
3512                         }
3513 #endif
3514                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
3515                 }
3516                 /* retry */
3517         }
3518 }
3519
3520 /*
3521  * Unlock a held and locked pv, keeping the hold count.
3522  */
3523 static
3524 void
3525 pv_unlock(pv_entry_t pv)
3526 {
3527         u_int count;
3528
3529         for (;;) {
3530                 count = pv->pv_hold;
3531                 cpu_ccfence();
3532                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3533                          (PV_HOLD_LOCKED | 1));
3534                 if (atomic_cmpset_int(&pv->pv_hold, count,
3535                                       count &
3536                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3537                         if (count & PV_HOLD_WAITING)
3538                                 wakeup(pv);
3539                         break;
3540                 }
3541         }
3542 }
3543
3544 /*
3545  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3546  * and the hold count drops to zero we will free it.
3547  *
3548  * Caller should not hold any spin locks.  We are protected from hold races
3549  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3550  * lock held.  A pv cannot be located otherwise.
3551  */
3552 static
3553 void
3554 pv_put(pv_entry_t pv)
3555 {
3556 #ifdef PMAP_DEBUG2
3557         if (pmap_enter_debug > 0) {
3558                 --pmap_enter_debug;
3559                 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3560         }
3561 #endif
3562
3563         /*
3564          * Normal put-aways must have a pv_m associated with the pv.
3565          */
3566         KKASSERT(pv->pv_m != NULL);
3567
3568         /*
3569          * Fast - shortcut most common condition
3570          */
3571         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3572                 return;
3573
3574         /*
3575          * Slow
3576          */
3577         pv_unlock(pv);
3578         pv_drop(pv);
3579 }
3580
3581 /*
3582  * Remove the pmap association from a pv, require that pv_m already be removed,
3583  * then unlock and drop the pv.  Any pte operations must have already been
3584  * completed.  This call may result in a last-drop which will physically free
3585  * the pv.
3586  *
3587  * Removing the pmap association entails an additional drop.
3588  *
3589  * pv must be exclusively locked on call and will be disposed of on return.
3590  */
3591 static
3592 void
3593 _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL)
3594 {
3595         pmap_t pmap;
3596
3597 #ifdef PMAP_DEBUG
3598         pv->pv_func_lastfree = func;
3599         pv->pv_line_lastfree = lineno;
3600 #endif
3601         KKASSERT(pv->pv_m == NULL);
3602         KKASSERT((pv->pv_hold & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
3603                   (PV_HOLD_LOCKED|1));
3604         if ((pmap = pv->pv_pmap) != NULL) {
3605                 spin_lock(&pmap->pm_spin);
3606                 KKASSERT(pv->pv_pmap == pmap);
3607                 if (pmap->pm_pvhint == pv)
3608                         pmap->pm_pvhint = NULL;
3609                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3610                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3611                 pv->pv_pmap = NULL;
3612                 pv->pv_pindex = 0;
3613                 spin_unlock(&pmap->pm_spin);
3614
3615                 /*
3616                  * Try to shortcut three atomic ops, otherwise fall through
3617                  * and do it normally.  Drop two refs and the lock all in
3618                  * one go.
3619                  */
3620                 if (pvp)
3621                         vm_page_unwire_quick(pvp->pv_m);
3622                 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3623 #ifdef PMAP_DEBUG2
3624                         if (pmap_enter_debug > 0) {
3625                                 --pmap_enter_debug;
3626                                 kprintf("pv_free: free pv %p\n", pv);
3627                         }
3628 #endif
3629                         zfree(pvzone, pv);
3630                         return;
3631                 }
3632                 pv_drop(pv);    /* ref for pv_pmap */
3633         }
3634         pv_unlock(pv);
3635         pv_drop(pv);
3636 }
3637
3638 /*
3639  * This routine is very drastic, but can save the system
3640  * in a pinch.
3641  */
3642 void
3643 pmap_collect(void)
3644 {
3645         int i;
3646         vm_page_t m;
3647         static int warningdone=0;
3648
3649         if (pmap_pagedaemon_waken == 0)
3650                 return;
3651         pmap_pagedaemon_waken = 0;
3652         if (warningdone < 5) {
3653                 kprintf("pmap_collect: collecting pv entries -- "
3654                         "suggest increasing PMAP_SHPGPERPROC\n");
3655                 warningdone++;
3656         }
3657
3658         for (i = 0; i < vm_page_array_size; i++) {
3659                 m = &vm_page_array[i];
3660                 if (m->wire_count || m->hold_count)
3661                         continue;
3662                 if (vm_page_busy_try(m, TRUE) == 0) {
3663                         if (m->wire_count == 0 && m->hold_count == 0) {
3664                                 pmap_remove_all(m);
3665                         }
3666                         vm_page_wakeup(m);
3667                 }
3668         }
3669 }
3670
3671 /*
3672  * Scan the pmap for active page table entries and issue a callback.
3673  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3674  * its parent page table.
3675  *
3676  * pte_pv will be NULL if the page or page table is unmanaged.
3677  * pt_pv will point to the page table page containing the pte for the page.
3678  *
3679  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3680  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3681  *       process pmap's PD and page to the callback function.  This can be
3682  *       confusing because the pt_pv is really a pd_pv, and the target page
3683  *       table page is simply aliased by the pmap and not owned by it.
3684  *
3685  * It is assumed that the start and end are properly rounded to the page size.
3686  *
3687  * It is assumed that PD pages and above are managed and thus in the RB tree,
3688  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3689  */
3690 struct pmap_scan_info {
3691         struct pmap *pmap;
3692         vm_offset_t sva;
3693         vm_offset_t eva;
3694         vm_pindex_t sva_pd_pindex;
3695         vm_pindex_t eva_pd_pindex;
3696         void (*func)(pmap_t, struct pmap_scan_info *,
3697                      pv_entry_t, vm_pindex_t *, pv_entry_t,
3698                      int, vm_offset_t,
3699                      pt_entry_t *, void *);
3700         void *arg;
3701         pmap_inval_bulk_t bulk_core;
3702         pmap_inval_bulk_t *bulk;
3703         int count;
3704         int stop;
3705 };
3706
3707 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3708 static int pmap_scan_callback(pv_entry_t pv, void *data);
3709
3710 static void
3711 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3712 {
3713         struct pmap *pmap = info->pmap;
3714         pv_entry_t pd_pv;       /* A page directory PV */
3715         pv_entry_t pt_pv;       /* A page table PV */
3716         pv_entry_t pte_pv;      /* A page table entry PV */
3717         vm_pindex_t *pte_placemark;
3718         vm_pindex_t *pt_placemark;
3719         pt_entry_t *ptep;
3720         pt_entry_t oldpte;
3721         struct pv_entry dummy_pv;
3722
3723         info->stop = 0;
3724         if (pmap == NULL)
3725                 return;
3726         if (smp_inval) {
3727                 info->bulk = &info->bulk_core;
3728                 pmap_inval_bulk_init(&info->bulk_core, pmap);
3729         } else {
3730                 info->bulk = NULL;
3731         }
3732
3733         /*
3734          * Hold the token for stability; if the pmap is empty we have nothing
3735          * to do.
3736          */
3737 #if 0
3738         if (pmap->pm_stats.resident_count == 0) {
3739                 return;
3740         }
3741 #endif
3742
3743         info->count = 0;
3744
3745         /*
3746          * Special handling for scanning one page, which is a very common
3747          * operation (it is?).
3748          *
3749          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3750          */
3751         if (info->sva + PAGE_SIZE == info->eva) {
3752                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3753                         /*
3754                          * Kernel mappings do not track wire counts on
3755                          * page table pages and only maintain pd_pv and
3756                          * pte_pv levels so pmap_scan() works.
3757                          */
3758                         pt_pv = NULL;
3759                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3760                                         &pte_placemark);
3761                         ptep = vtopte(info->sva);
3762                 } else {
3763                         /*
3764                          * User pages which are unmanaged will not have a
3765                          * pte_pv.  User page table pages which are unmanaged
3766                          * (shared from elsewhere) will also not have a pt_pv.
3767                          * The func() callback will pass both pte_pv and pt_pv
3768                          * as NULL in that case.
3769                          *
3770                          * We hold pte_placemark across the operation for
3771                          * unmanaged pages.
3772                          *
3773                          * WARNING!  We must hold pt_placemark across the
3774                          *           *ptep test to prevent misintepreting
3775                          *           a non-zero *ptep as a shared page
3776                          *           table page.  Hold it across the function
3777                          *           callback as well for SMP safety.
3778                          */
3779                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3780                                         &pte_placemark);
3781                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva),
3782                                         &pt_placemark);
3783                         if (pt_pv == NULL) {
3784                                 KKASSERT(pte_pv == NULL);
3785                                 pd_pv = pv_get(pmap,
3786                                                pmap_pd_pindex(info->sva),
3787                                                NULL);
3788                                 if (pd_pv) {
3789                                         ptep = pv_pte_lookup(pd_pv,
3790                                                     pmap_pt_index(info->sva));
3791                                         if (*ptep) {
3792                                                 info->func(pmap, info,
3793                                                      NULL, pt_placemark,
3794                                                      pd_pv, 1,
3795                                                      info->sva, ptep,
3796                                                      info->arg);
3797                                         } else {
3798                                                 pv_placemarker_wakeup(pmap,
3799                                                                   pt_placemark);
3800                                         }
3801                                         pv_put(pd_pv);
3802                                 } else {
3803                                         pv_placemarker_wakeup(pmap,
3804                                                               pt_placemark);
3805                                 }
3806                                 pv_placemarker_wakeup(pmap, pte_placemark);
3807                                 goto fast_skip;
3808                         }
3809                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3810                 }
3811
3812                 /*
3813                  * NOTE: *ptep can't be ripped out from under us if we hold
3814                  *       pte_pv (or pte_placemark) locked, but bits can
3815                  *       change.
3816                  */
3817                 oldpte = *ptep;
3818                 cpu_ccfence();
3819                 if (oldpte == 0) {
3820                         KKASSERT(pte_pv == NULL);
3821                         pv_placemarker_wakeup(pmap, pte_placemark);
3822                 } else if (pte_pv) {
3823                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3824                                            pmap->pmap_bits[PG_V_IDX])) ==
3825                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
3826                                  pmap->pmap_bits[PG_V_IDX]),
3827                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p",
3828                             *ptep, oldpte, info->sva, pte_pv));
3829                         info->func(pmap, info, pte_pv, NULL, pt_pv, 0,
3830                                    info->sva, ptep, info->arg);
3831                 } else {
3832                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3833                                            pmap->pmap_bits[PG_V_IDX])) ==
3834                             pmap->pmap_bits[PG_V_IDX],
3835                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
3836                             *ptep, oldpte, info->sva));
3837                         info->func(pmap, info, NULL, pte_placemark, pt_pv, 0,
3838                                    info->sva, ptep, info->arg);
3839                 }
3840                 if (pt_pv)
3841                         pv_put(pt_pv);
3842 fast_skip:
3843                 pmap_inval_bulk_flush(info->bulk);
3844                 return;
3845         }
3846
3847         /*
3848          * Nominal scan case, RB_SCAN() for PD pages and iterate from
3849          * there.
3850          */
3851         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3852         info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3853
3854         if (info->sva >= VM_MAX_USER_ADDRESS) {
3855                 /*
3856                  * The kernel does not currently maintain any pv_entry's for
3857                  * higher-level page tables.
3858                  */
3859                 bzero(&dummy_pv, sizeof(dummy_pv));
3860                 dummy_pv.pv_pindex = info->sva_pd_pindex;
3861                 spin_lock(&pmap->pm_spin);
3862                 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3863                         pmap_scan_callback(&dummy_pv, info);
3864                         ++dummy_pv.pv_pindex;
3865                 }
3866                 spin_unlock(&pmap->pm_spin);
3867         } else {
3868                 /*
3869                  * User page tables maintain local PML4, PDP, and PD
3870                  * pv_entry's at the very least.  PT pv's might be
3871                  * unmanaged and thus not exist.  PTE pv's might be
3872                  * unmanaged and thus not exist.
3873                  */
3874                 spin_lock(&pmap->pm_spin);
3875                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot, pmap_scan_cmp,
3876                                          pmap_scan_callback, info);
3877                 spin_unlock(&pmap->pm_spin);
3878         }
3879         pmap_inval_bulk_flush(info->bulk);
3880 }
3881
3882 /*
3883  * WARNING! pmap->pm_spin held
3884  */
3885 static int
3886 pmap_scan_cmp(pv_entry_t pv, void *data)
3887 {
3888         struct pmap_scan_info *info = data;
3889         if (pv->pv_pindex < info->sva_pd_pindex)
3890                 return(-1);
3891         if (pv->pv_pindex >= info->eva_pd_pindex)
3892                 return(1);
3893         return(0);
3894 }
3895
3896 /*
3897  * pmap_scan() by PDs
3898  *
3899  * WARNING! pmap->pm_spin held
3900  */
3901 static int
3902 pmap_scan_callback(pv_entry_t pv, void *data)
3903 {
3904         struct pmap_scan_info *info = data;
3905         struct pmap *pmap = info->pmap;
3906         pv_entry_t pd_pv;       /* A page directory PV */
3907         pv_entry_t pt_pv;       /* A page table PV */
3908         vm_pindex_t *pt_placemark;
3909         pt_entry_t *ptep;
3910         pt_entry_t oldpte;
3911         vm_offset_t sva;
3912         vm_offset_t eva;
3913         vm_offset_t va_next;
3914         vm_pindex_t pd_pindex;
3915         int error;
3916
3917         /*
3918          * Stop if requested
3919          */
3920         if (info->stop)
3921                 return -1;
3922
3923         /*
3924          * Pull the PD pindex from the pv before releasing the spinlock.
3925          *
3926          * WARNING: pv is faked for kernel pmap scans.
3927          */
3928         pd_pindex = pv->pv_pindex;
3929         spin_unlock(&pmap->pm_spin);
3930         pv = NULL;      /* invalid after spinlock unlocked */
3931
3932         /*
3933          * Calculate the page range within the PD.  SIMPLE pmaps are
3934          * direct-mapped for the entire 2^64 address space.  Normal pmaps
3935          * reflect the user and kernel address space which requires
3936          * cannonicalization w/regards to converting pd_pindex's back
3937          * into addresses.
3938          */
3939         sva = (pd_pindex - pmap_pd_pindex(0)) << PDPSHIFT;
3940         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3941             (sva & PML4_SIGNMASK)) {
3942                 sva |= PML4_SIGNMASK;
3943         }
3944         eva = sva + NBPDP;      /* can overflow */
3945         if (sva < info->sva)
3946                 sva = info->sva;
3947         if (eva < info->sva || eva > info->eva)
3948                 eva = info->eva;
3949
3950         /*
3951          * NOTE: kernel mappings do not track page table pages, only
3952          *       terminal pages.
3953          *
3954          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3955          *       However, for the scan to be efficient we try to
3956          *       cache items top-down.
3957          */
3958         pd_pv = NULL;
3959         pt_pv = NULL;
3960
3961         for (; sva < eva; sva = va_next) {
3962                 if (info->stop)
3963                         break;
3964                 if (sva >= VM_MAX_USER_ADDRESS) {
3965                         if (pt_pv) {
3966                                 pv_put(pt_pv);
3967                                 pt_pv = NULL;
3968                         }
3969                         goto kernel_skip;
3970                 }
3971
3972                 /*
3973                  * PD cache, scan shortcut if it doesn't exist.
3974                  */
3975                 if (pd_pv == NULL) {
3976                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
3977                 } else if (pd_pv->pv_pmap != pmap ||
3978                            pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3979                         pv_put(pd_pv);
3980                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
3981                 }
3982                 if (pd_pv == NULL) {
3983                         va_next = (sva + NBPDP) & ~PDPMASK;
3984                         if (va_next < sva)
3985                                 va_next = eva;
3986                         continue;
3987                 }
3988
3989                 /*
3990                  * PT cache
3991                  */
3992                 if (pt_pv && (pt_pv->pv_pmap != pmap ||
3993                               pt_pv->pv_pindex != pmap_pt_pindex(sva))) {
3994                         pv_put(pt_pv);
3995                         pt_pv = NULL;
3996                 }
3997                 if (pt_pv == NULL) {
3998                         pt_pv = pv_get_try(pmap, pmap_pt_pindex(sva),
3999                                            &pt_placemark, &error);
4000                         if (error) {
4001                                 pv_put(pd_pv);  /* lock order */
4002                                 pd_pv = NULL;
4003                                 if (pt_pv) {
4004                                         pv_lock(pt_pv);
4005                                         pv_unlock(pt_pv);
4006                                         pv_drop(pt_pv);
4007                                         pt_pv = NULL;
4008                                 } else {
4009                                         pv_placemarker_wait(pmap, pt_placemark);
4010                                 }
4011                                 va_next = sva;
4012                                 continue;
4013                         }
4014                         /* may have to re-check later if pt_pv is NULL here */
4015                 }
4016
4017                 /*
4018                  * If pt_pv is NULL we either have an shared page table
4019                  * page and must issue a callback specific to that case,
4020                  * or there is no page table page.
4021                  *
4022                  * Either way we can skip the page table page.
4023                  *
4024                  * WARNING! pt_pv can also be NULL due to a pv creation
4025                  *          race where we find it to be NULL and then
4026                  *          later see a pte_pv.  But its possible the pt_pv
4027                  *          got created inbetween the two operations, so
4028                  *          we must check.
4029                  */
4030                 if (pt_pv == NULL) {
4031                         /*
4032                          * Possible unmanaged (shared from another pmap)
4033                          * page table page.
4034                          *
4035                          * WARNING!  We must hold pt_placemark across the
4036                          *           *ptep test to prevent misintepreting
4037                          *           a non-zero *ptep as a shared page
4038                          *           table page.  Hold it across the function
4039                          *           callback as well for SMP safety.
4040                          */
4041                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
4042                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
4043                                 info->func(pmap, info, NULL, pt_placemark,
4044                                            pd_pv, 1,
4045                                            sva, ptep, info->arg);
4046                         } else {
4047                                 pv_placemarker_wakeup(pmap, pt_placemark);
4048                         }
4049
4050                         /*
4051                          * Done, move to next page table page.
4052                          */
4053                         va_next = (sva + NBPDR) & ~PDRMASK;
4054                         if (va_next < sva)
4055                                 va_next = eva;
4056                         continue;
4057                 }
4058
4059                 /*
4060                  * From this point in the loop testing pt_pv for non-NULL
4061                  * means we are in UVM, else if it is NULL we are in KVM.
4062                  *
4063                  * Limit our scan to either the end of the va represented
4064                  * by the current page table page, or to the end of the
4065                  * range being removed.
4066                  */
4067 kernel_skip:
4068                 va_next = (sva + NBPDR) & ~PDRMASK;
4069                 if (va_next < sva)
4070                         va_next = eva;
4071                 if (va_next > eva)
4072                         va_next = eva;
4073
4074                 /*
4075                  * Scan the page table for pages.  Some pages may not be
4076                  * managed (might not have a pv_entry).
4077                  *
4078                  * There is no page table management for kernel pages so
4079                  * pt_pv will be NULL in that case, but otherwise pt_pv
4080                  * is non-NULL, locked, and referenced.
4081                  */
4082
4083                 /*
4084                  * At this point a non-NULL pt_pv means a UVA, and a NULL
4085                  * pt_pv means a KVA.
4086                  */
4087                 if (pt_pv)
4088                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
4089                 else
4090                         ptep = vtopte(sva);
4091
4092                 while (sva < va_next) {
4093                         pv_entry_t pte_pv;
4094                         vm_pindex_t *pte_placemark;
4095
4096                         /*
4097                          * Yield every 64 pages, stop if requested.
4098                          */
4099                         if ((++info->count & 63) == 0)
4100                                 lwkt_user_yield();
4101                         if (info->stop)
4102                                 break;
4103
4104                         /*
4105                          * We can shortcut our scan if *ptep == 0.  This is
4106                          * an unlocked check.
4107                          */
4108                         if (*ptep == 0) {
4109                                 sva += PAGE_SIZE;
4110                                 ++ptep;
4111                                 continue;
4112                         }
4113                         cpu_ccfence();
4114
4115                         /*
4116                          * Acquire the related pte_pv, if any.  If *ptep == 0
4117                          * the related pte_pv should not exist, but if *ptep
4118                          * is not zero the pte_pv may or may not exist (e.g.
4119                          * will not exist for an unmanaged page).
4120                          *
4121                          * However a multitude of races are possible here
4122                          * so if we cannot lock definite state we clean out
4123                          * our cache and break the inner while() loop to
4124                          * force a loop up to the top of the for().
4125                          *
4126                          * XXX unlock/relock pd_pv, pt_pv, and re-test their
4127                          *     validity instead of looping up?
4128                          */
4129                         pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
4130                                             &pte_placemark, &error);
4131                         if (error) {
4132                                 pv_put(pd_pv);          /* lock order */
4133                                 pd_pv = NULL;
4134                                 if (pt_pv) {
4135                                         pv_put(pt_pv);  /* lock order */
4136                                         pt_pv = NULL;
4137                                 }
4138                                 if (pte_pv) {           /* block */
4139                                         pv_lock(pte_pv);
4140                                         pv_unlock(pte_pv);
4141                                         pv_drop(pte_pv);
4142                                         pte_pv = NULL;
4143                                 } else {
4144                                         pv_placemarker_wait(pmap,
4145                                                         pte_placemark);
4146                                 }
4147                                 va_next = sva;          /* retry */
4148                                 break;
4149                         }
4150
4151                         /*
4152                          * Reload *ptep after successfully locking the
4153                          * pindex.  If *ptep == 0 we had better NOT have a
4154                          * pte_pv.
4155                          */
4156                         cpu_ccfence();
4157                         oldpte = *ptep;
4158                         if (oldpte == 0) {
4159                                 if (pte_pv) {
4160                                         kprintf("Unexpected non-NULL pte_pv "
4161                                                 "%p pt_pv %p "
4162                                                 "*ptep = %016lx/%016lx\n",
4163                                                 pte_pv, pt_pv, *ptep, oldpte);
4164                                         panic("Unexpected non-NULL pte_pv");
4165                                 } else {
4166                                         pv_placemarker_wakeup(pmap, pte_placemark);
4167                                 }
4168                                 sva += PAGE_SIZE;
4169                                 ++ptep;
4170                                 continue;
4171                         }
4172
4173                         /*
4174                          * We can't hold pd_pv across the callback (because
4175                          * we don't pass it to the callback and the callback
4176                          * might deadlock)
4177                          */
4178                         if (pd_pv) {
4179                                 vm_page_wire_quick(pd_pv->pv_m);
4180                                 pv_unlock(pd_pv);
4181                         }
4182
4183                         /*
4184                          * Ready for the callback.  The locked pte_pv (if any)
4185                          * is consumed by the callback.  pte_pv will exist if
4186                          * the page is managed, and will not exist if it
4187                          * isn't.
4188                          */
4189                         if (oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4190                                 /*
4191                                  * Managed pte
4192                                  */
4193                                 KASSERT(pte_pv &&
4194                                          (oldpte & pmap->pmap_bits[PG_V_IDX]),
4195                                     ("badC *ptep %016lx/%016lx sva %016lx "
4196                                     "pte_pv %p",
4197                                     *ptep, oldpte, sva, pte_pv));
4198                                 /*
4199                                  * We must unlock pd_pv across the callback
4200                                  * to avoid deadlocks on any recursive
4201                                  * disposal.  Re-check that it still exists
4202                                  * after re-locking.
4203                                  *
4204                                  * Call target disposes of pte_pv and may
4205                                  * destroy but will not dispose of pt_pv.
4206                                  */
4207                                 info->func(pmap, info, pte_pv, NULL,
4208                                            pt_pv, 0,
4209                                            sva, ptep, info->arg);
4210                         } else {
4211                                 /*
4212                                  * Unmanaged pte
4213                                  *
4214                                  * We must unlock pd_pv across the callback
4215                                  * to avoid deadlocks on any recursive
4216                                  * disposal.  Re-check that it still exists
4217                                  * after re-locking.
4218                                  *
4219                                  * Call target disposes of pte_pv or
4220                                  * pte_placemark and may destroy but will
4221                                  * not dispose of pt_pv.
4222                                  */
4223                                 KASSERT(pte_pv == NULL &&
4224                                         (oldpte & pmap->pmap_bits[PG_V_IDX]),
4225                                     ("badD *ptep %016lx/%016lx sva %016lx "
4226                                     "pte_pv %p pte_pv->pv_m %p ",
4227                                      *ptep, oldpte, sva,
4228                                      pte_pv, (pte_pv ? pte_pv->pv_m : NULL)));
4229                                 if (pte_pv)
4230                                         kprintf("RaceD\n");
4231                                 if (pte_pv) {
4232                                         info->func(pmap, info,
4233                                                    pte_pv, NULL,
4234                                                    pt_pv, 0,
4235                                                    sva, ptep, info->arg);
4236                                 } else {
4237                                         info->func(pmap, info,
4238                                                    NULL, pte_placemark,
4239                                                    pt_pv, 0,
4240                                                    sva, ptep, info->arg);
4241                                 }
4242                         }
4243                         if (pd_pv) {
4244                                 pv_lock(pd_pv);
4245                                 vm_page_unwire_quick(pd_pv->pv_m);
4246                                 if (pd_pv->pv_pmap == NULL) {
4247                                         va_next = sva;          /* retry */
4248                                         break;
4249                                 }
4250                         }
4251                         pte_pv = NULL;
4252                         sva += PAGE_SIZE;
4253                         ++ptep;
4254                 }
4255         }
4256         if (pd_pv) {
4257                 pv_put(pd_pv);
4258                 pd_pv = NULL;
4259         }
4260         if (pt_pv) {
4261                 pv_put(pt_pv);
4262                 pt_pv = NULL;
4263         }
4264         if ((++info->count & 7) == 0)
4265                 lwkt_user_yield();
4266
4267         /*
4268          * Relock before returning.
4269          */
4270         spin_lock(&pmap->pm_spin);
4271         return (0);
4272 }
4273
4274 void
4275 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4276 {
4277         struct pmap_scan_info info;
4278
4279         info.pmap = pmap;
4280         info.sva = sva;
4281         info.eva = eva;
4282         info.func = pmap_remove_callback;
4283         info.arg = NULL;
4284         pmap_scan(&info, 1);
4285 }
4286
4287 static void
4288 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4289 {
4290         struct pmap_scan_info info;
4291
4292         info.pmap = pmap;
4293         info.sva = sva;
4294         info.eva = eva;
4295         info.func = pmap_remove_callback;
4296         info.arg = NULL;
4297         pmap_scan(&info, 0);
4298 }
4299
4300 static void
4301 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4302                      pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4303                      pv_entry_t pt_pv, int sharept,
4304                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4305 {
4306         pt_entry_t pte;
4307
4308         if (pte_pv) {
4309                 /*
4310                  * Managed entry
4311                  *
4312                  * This will also drop pt_pv's wire_count. Note that
4313                  * terminal pages are not wired based on mmu presence.
4314                  *
4315                  * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4316                  */
4317                 KKASSERT(pte_pv->pv_m != NULL);
4318                 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4319                 pte_pv = NULL;  /* safety */
4320
4321                 /*
4322                  * Recursively destroy higher-level page tables.
4323                  *
4324                  * This is optional.  If we do not, they will still
4325                  * be destroyed when the process exits.
4326                  *
4327                  * NOTE: Do not destroy pv_entry's with extra hold refs,
4328                  *       a caller may have unlocked it and intends to
4329                  *       continue to use it.
4330                  */
4331                 if (pmap_dynamic_delete &&
4332                     pt_pv &&
4333                     pt_pv->pv_m &&
4334                     pt_pv->pv_m->wire_count == 1 &&
4335                     (pt_pv->pv_hold & PV_HOLD_MASK) == 2 &&
4336                     pt_pv->pv_pindex != pmap_pml4_pindex()) {
4337                         if (pmap_dynamic_delete == 2)
4338                                 kprintf("B %jd %08x\n", pt_pv->pv_pindex, pt_pv->pv_hold);
4339                         pv_hold(pt_pv); /* extra hold */
4340                         pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4341                         pv_lock(pt_pv); /* prior extra hold + relock */
4342                 }
4343         } else if (sharept == 0) {
4344                 /*
4345                  * Unmanaged page table (pt, pd, or pdp. Not pte).
4346                  *
4347                  * pt_pv's wire_count is still bumped by unmanaged pages
4348                  * so we must decrement it manually.
4349                  *
4350                  * We have to unwire the target page table page.
4351                  *
4352                  * It is unclear how we can invalidate a segment so we
4353                  * invalidate -1 which invlidates the tlb.
4354                  */
4355                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4356                 if (pte & pmap->pmap_bits[PG_W_IDX])
4357                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4358                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4359                 if (vm_page_unwire_quick(pt_pv->pv_m))
4360                         panic("pmap_remove: insufficient wirecount");
4361                 pv_placemarker_wakeup(pmap, pte_placemark);
4362         } else {
4363                 /*
4364                  * Unmanaged page table (pt, pd, or pdp. Not pte) for
4365                  * a shared page table.
4366                  *
4367                  * pt_pv is actually the pd_pv for our pmap (not the shared
4368                  * object pmap).
4369                  *
4370                  * We have to unwire the target page table page and we
4371                  * have to unwire our page directory page.
4372                  *
4373                  * It is unclear how we can invalidate a segment so we
4374                  * invalidate -1 which invlidates the tlb.
4375                  */
4376                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4377                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4378                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4379                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4380                         panic("pmap_remove: shared pgtable1 bad wirecount");
4381                 if (vm_page_unwire_quick(pt_pv->pv_m))
4382                         panic("pmap_remove: shared pgtable2 bad wirecount");
4383                 pv_placemarker_wakeup(pmap, pte_placemark);
4384         }
4385 }
4386
4387 /*
4388  * Removes this physical page from all physical maps in which it resides.
4389  * Reflects back modify bits to the pager.
4390  *
4391  * This routine may not be called from an interrupt.
4392  */
4393 static
4394 void
4395 pmap_remove_all(vm_page_t m)
4396 {
4397         pv_entry_t pv;
4398         pmap_inval_bulk_t bulk;
4399
4400         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4401                 return;
4402
4403         vm_page_spin_lock(m);
4404         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4405                 KKASSERT(pv->pv_m == m);
4406                 if (pv_hold_try(pv)) {
4407                         vm_page_spin_unlock(m);
4408                 } else {
4409                         vm_page_spin_unlock(m);
4410                         pv_lock(pv);
4411                         pv_unlock(pv);
4412                         pv_drop(pv);
4413                         vm_page_spin_lock(m);
4414                         continue;
4415                 }
4416                 KKASSERT(pv->pv_pmap && pv->pv_m == m);
4417
4418                 /*
4419                  * Holding no spinlocks, pv is locked.  Once we scrap
4420                  * pv we can no longer use it as a list iterator (but
4421                  * we are doing a TAILQ_FIRST() so we are ok).
4422                  */
4423                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4424                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4425                 pv = NULL;      /* safety */
4426                 pmap_inval_bulk_flush(&bulk);
4427                 vm_page_spin_lock(m);
4428         }
4429         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4430         vm_page_spin_unlock(m);
4431 }
4432
4433 /*
4434  * Removes the page from a particular pmap
4435  */
4436 void
4437 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4438 {
4439         pv_entry_t pv;
4440         pmap_inval_bulk_t bulk;
4441
4442         if (!pmap_initialized)
4443                 return;
4444
4445 again:
4446         vm_page_spin_lock(m);
4447         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4448                 if (pv->pv_pmap != pmap)
4449                         continue;
4450                 KKASSERT(pv->pv_m == m);
4451                 if (pv_hold_try(pv)) {
4452                         vm_page_spin_unlock(m);
4453                 } else {
4454                         vm_page_spin_unlock(m);
4455                         pv_lock(pv);
4456                         pv_unlock(pv);
4457                         pv_drop(pv);
4458                         goto again;
4459                 }
4460                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
4461
4462                 /*
4463                  * Holding no spinlocks, pv is locked.  Once gone it can't
4464                  * be used as an iterator.  In fact, because we couldn't
4465                  * necessarily lock it atomically it may have moved within
4466                  * the list and ALSO cannot be used as an iterator.
4467                  */
4468                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4469                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4470                 pv = NULL;      /* safety */
4471                 pmap_inval_bulk_flush(&bulk);
4472                 goto again;
4473         }
4474         vm_page_spin_unlock(m);
4475 }
4476
4477 /*
4478  * Set the physical protection on the specified range of this map
4479  * as requested.  This function is typically only used for debug watchpoints
4480  * and COW pages.
4481  *
4482  * This function may not be called from an interrupt if the map is
4483  * not the kernel_pmap.
4484  *
4485  * NOTE!  For shared page table pages we just unmap the page.
4486  */
4487 void
4488 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4489 {
4490         struct pmap_scan_info info;
4491         /* JG review for NX */
4492
4493         if (pmap == NULL)
4494                 return;
4495         if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
4496                 pmap_remove(pmap, sva, eva);
4497                 return;
4498         }
4499         if (prot & VM_PROT_WRITE)
4500                 return;
4501         info.pmap = pmap;
4502         info.sva = sva;
4503         info.eva = eva;
4504         info.func = pmap_protect_callback;
4505         info.arg = &prot;
4506         pmap_scan(&info, 1);
4507 }
4508
4509 static
4510 void
4511 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4512                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4513                       pv_entry_t pt_pv, int sharept,
4514                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4515 {
4516         pt_entry_t pbits;
4517         pt_entry_t cbits;
4518         pt_entry_t pte;
4519         vm_page_t m;
4520
4521 again:
4522         pbits = *ptep;
4523         cbits = pbits;
4524         if (pte_pv) {
4525                 KKASSERT(pte_pv->pv_m != NULL);
4526                 m = NULL;
4527                 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4528                         if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4529                                 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4530                                 KKASSERT(m == pte_pv->pv_m);
4531                                 vm_page_flag_set(m, PG_REFERENCED);
4532                         }
4533                         cbits &= ~pmap->pmap_bits[PG_A_IDX];
4534                 }
4535                 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4536                         if (pmap_track_modified(pte_pv->pv_pindex)) {
4537                                 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4538                                         if (m == NULL) {
4539                                                 m = PHYS_TO_VM_PAGE(pbits &
4540                                                                     PG_FRAME);
4541                                         }
4542                                         vm_page_dirty(m);
4543                                 }
4544                                 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4545                         }
4546                 }
4547         } else if (sharept) {
4548                 /*
4549                  * Unmanaged page table, pt_pv is actually the pd_pv
4550                  * for our pmap (not the object's shared pmap).
4551                  *
4552                  * When asked to protect something in a shared page table
4553                  * page we just unmap the page table page.  We have to
4554                  * invalidate the tlb in this situation.
4555                  *
4556                  * XXX Warning, shared page tables will not be used for
4557                  * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4558                  * so PHYS_TO_VM_PAGE() should be safe here.
4559                  */
4560                 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4561                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4562                         panic("pmap_protect: pgtable1 pg bad wirecount");
4563                 if (vm_page_unwire_quick(pt_pv->pv_m))
4564                         panic("pmap_protect: pgtable2 pg bad wirecount");
4565                 ptep = NULL;
4566         }
4567         /* else unmanaged page, adjust bits, no wire changes */
4568
4569         if (ptep) {
4570                 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4571 #ifdef PMAP_DEBUG2
4572                 if (pmap_enter_debug > 0) {
4573                         --pmap_enter_debug;
4574                         kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4575                                 "pt_pv=%p cbits=%08lx\n",
4576                                 va, ptep, pte_pv,
4577                                 pt_pv, cbits
4578                         );
4579                 }
4580 #endif
4581                 if (pbits != cbits) {
4582                         if (!pmap_inval_smp_cmpset(pmap, (vm_offset_t)-1,
4583                                                    ptep, pbits, cbits)) {
4584                                 goto again;
4585                         }
4586                 }
4587         }
4588         if (pte_pv)
4589                 pv_put(pte_pv);
4590         else
4591                 pv_placemarker_wakeup(pmap, pte_placemark);
4592 }
4593
4594 /*
4595  * Insert the vm_page (m) at the virtual address (va), replacing any prior
4596  * mapping at that address.  Set protection and wiring as requested.
4597  *
4598  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4599  * possible.  If it is we enter the page into the appropriate shared pmap
4600  * hanging off the related VM object instead of the passed pmap, then we
4601  * share the page table page from the VM object's pmap into the current pmap.
4602  *
4603  * NOTE: This routine MUST insert the page into the pmap now, it cannot
4604  *       lazy-evaluate.
4605  *
4606  * NOTE: If (m) is PG_UNMANAGED it may also be a temporary fake vm_page_t.
4607  *       never record it.
4608  */
4609 void
4610 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4611            boolean_t wired, vm_map_entry_t entry)
4612 {
4613         pv_entry_t pt_pv;       /* page table */
4614         pv_entry_t pte_pv;      /* page table entry */
4615         vm_pindex_t *pte_placemark;
4616         pt_entry_t *ptep;
4617         vm_paddr_t opa;
4618         pt_entry_t origpte, newpte;
4619         vm_paddr_t pa;
4620
4621         if (pmap == NULL)
4622                 return;
4623         va = trunc_page(va);
4624 #ifdef PMAP_DIAGNOSTIC
4625         if (va >= KvaEnd)
4626                 panic("pmap_enter: toobig");
4627         if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4628                 panic("pmap_enter: invalid to pmap_enter page table "
4629                       "pages (va: 0x%lx)", va);
4630 #endif
4631         if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4632                 kprintf("Warning: pmap_enter called on UVA with "
4633                         "kernel_pmap\n");
4634 #ifdef DDB
4635                 db_print_backtrace();
4636 #endif
4637         }
4638         if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4639                 kprintf("Warning: pmap_enter called on KVA without"
4640                         "kernel_pmap\n");
4641 #ifdef DDB
4642                 db_print_backtrace();
4643 #endif
4644         }
4645
4646         /*
4647          * Get locked PV entries for our new page table entry (pte_pv or
4648          * pte_placemark) and for its parent page table (pt_pv).  We need
4649          * the parent so we can resolve the location of the ptep.
4650          *
4651          * Only hardware MMU actions can modify the ptep out from
4652          * under us.
4653          *
4654          * if (m) is fictitious or unmanaged we do not create a managing
4655          * pte_pv for it.  Any pre-existing page's management state must
4656          * match (avoiding code complexity).
4657          *
4658          * If the pmap is still being initialized we assume existing
4659          * page tables.
4660          *
4661          * Kernel mapppings do not track page table pages (i.e. pt_pv).
4662          *
4663          * WARNING! If replacing a managed mapping with an unmanaged mapping
4664          *          pte_pv will wind up being non-NULL and must be handled
4665          *          below.
4666          */
4667         if (pmap_initialized == FALSE) {
4668                 pte_pv = NULL;
4669                 pt_pv = NULL;
4670                 pte_placemark = NULL;
4671                 ptep = vtopte(va);
4672                 origpte = *ptep;
4673         } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4674                 pte_pv = pv_get(pmap, pmap_pte_pindex(va), &pte_placemark);
4675                 KKASSERT(pte_pv == NULL);
4676                 if (va >= VM_MAX_USER_ADDRESS) {
4677                         pt_pv = NULL;
4678                         ptep = vtopte(va);
4679                 } else {
4680                         pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4681                                                   NULL, entry, va);
4682                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4683                 }
4684                 origpte = *ptep;
4685                 cpu_ccfence();
4686                 KASSERT(origpte == 0 ||
4687                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4688                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4689         } else {
4690                 if (va >= VM_MAX_USER_ADDRESS) {
4691                         /*
4692                          * Kernel map, pv_entry-tracked.
4693                          */
4694                         pt_pv = NULL;
4695                         pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4696                         ptep = vtopte(va);
4697                 } else {
4698                         /*
4699                          * User map
4700                          */
4701                         pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4702                                                    &pt_pv, entry, va);
4703                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4704                 }
4705                 pte_placemark = NULL;   /* safety */
4706                 origpte = *ptep;
4707                 cpu_ccfence();
4708                 KASSERT(origpte == 0 ||
4709                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4710                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4711         }
4712
4713         pa = VM_PAGE_TO_PHYS(m);
4714         opa = origpte & PG_FRAME;
4715
4716         /*
4717          * Calculate the new PTE.  Note that pte_pv alone does not mean
4718          * the new pte_pv is managed, it could exist because the old pte
4719          * was managed even if the new one is not.
4720          */
4721         newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4722                  pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4723         if (wired)
4724                 newpte |= pmap->pmap_bits[PG_W_IDX];
4725         if (va < VM_MAX_USER_ADDRESS)
4726                 newpte |= pmap->pmap_bits[PG_U_IDX];
4727         if (pte_pv && (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) == 0)
4728                 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4729 //      if (pmap == &kernel_pmap)
4730 //              newpte |= pgeflag;
4731         newpte |= pmap->pmap_cache_bits[m->pat_mode];
4732         if (m->flags & PG_FICTITIOUS)
4733                 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4734
4735         /*
4736          * It is possible for multiple faults to occur in threaded
4737          * environments, the existing pte might be correct.
4738          */
4739         if (((origpte ^ newpte) &
4740             ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4741                           pmap->pmap_bits[PG_A_IDX])) == 0) {
4742                 goto done;
4743         }
4744
4745         /*
4746          * Ok, either the address changed or the protection or wiring
4747          * changed.
4748          *
4749          * Clear the current entry, interlocking the removal.  For managed
4750          * pte's this will also flush the modified state to the vm_page.
4751          * Atomic ops are mandatory in order to ensure that PG_M events are
4752          * not lost during any transition.
4753          *
4754          * WARNING: The caller has busied the new page but not the original
4755          *          vm_page which we are trying to replace.  Because we hold
4756          *          the pte_pv lock, but have not busied the page, PG bits
4757          *          can be cleared out from under us.
4758          */
4759         if (opa) {
4760                 if (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4761                         /*
4762                          * Old page was managed.  Expect pte_pv to exist.
4763                          * (it might also exist if the old page was unmanaged).
4764                          *
4765                          * NOTE: pt_pv won't exist for a kernel page
4766                          *       (managed or otherwise).
4767                          *
4768                          * NOTE: We may be reusing the pte_pv so we do not
4769                          *       destroy it in pmap_remove_pv_pte().
4770                          */
4771                         KKASSERT(pte_pv && pte_pv->pv_m);
4772                         if (prot & VM_PROT_NOSYNC) {
4773                                 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4774                         } else {
4775                                 pmap_inval_bulk_t bulk;
4776
4777                                 pmap_inval_bulk_init(&bulk, pmap);
4778                                 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4779                                 pmap_inval_bulk_flush(&bulk);
4780                         }
4781                         pmap_remove_pv_page(pte_pv);
4782                         /* will either set pte_pv->pv_m or pv_free() later */
4783                 } else {
4784                         /*
4785                          * Old page was not managed.  If we have a pte_pv
4786                          * it better not have a pv_m assigned to it.  If the
4787                          * new page is managed the pte_pv will be destroyed
4788                          * near the end (we need its interlock).
4789                          *
4790                          * NOTE: We leave the wire count on the PT page
4791                          *       intact for the followup enter, but adjust
4792                          *       the wired-pages count on the pmap.
4793                          */
4794                         KKASSERT(pte_pv == NULL);
4795                         if (prot & VM_PROT_NOSYNC) {
4796                                 /*
4797                                  * NOSYNC (no mmu sync) requested.
4798                                  */
4799                                 (void)pte_load_clear(ptep);
4800                                 cpu_invlpg((void *)va);
4801                         } else {
4802                                 /*
4803                                  * Nominal SYNC
4804                                  */
4805                                 pmap_inval_smp(pmap, va, 1, ptep, 0);
4806                         }
4807
4808                         /*
4809                          * We must adjust pm_stats manually for unmanaged
4810                          * pages.
4811                          */
4812                         if (pt_pv) {
4813                                 atomic_add_long(&pmap->pm_stats.
4814                                                 resident_count, -1);
4815                         }
4816                         if (origpte & pmap->pmap_bits[PG_W_IDX]) {
4817                                 atomic_add_long(&pmap->pm_stats.
4818                                                 wired_count, -1);
4819                         }
4820                 }
4821                 KKASSERT(*ptep == 0);
4822         }
4823
4824 #ifdef PMAP_DEBUG2
4825         if (pmap_enter_debug > 0) {
4826                 --pmap_enter_debug;
4827                 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4828                         " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4829                         va, m,
4830                         origpte, newpte, ptep,
4831                         pte_pv, pt_pv, opa, prot);
4832         }
4833 #endif
4834
4835         if ((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
4836                 /*
4837                  * Entering an unmanaged page.  We must wire the pt_pv unless
4838                  * we retained the wiring from an unmanaged page we had
4839                  * removed (if we retained it via pte_pv that will go away
4840                  * soon).
4841                  */
4842                 if (pt_pv && (opa == 0 ||
4843                               (origpte & pmap->pmap_bits[PG_MANAGED_IDX]))) {
4844                         vm_page_wire_quick(pt_pv->pv_m);
4845                 }
4846                 if (wired)
4847                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
4848
4849                 /*
4850                  * Unmanaged pages need manual resident_count tracking.
4851                  */
4852                 if (pt_pv) {
4853                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.
4854                                         resident_count, 1);
4855                 }
4856         } else {
4857                 /*
4858                  * Entering a managed page.  Our pte_pv takes care of the
4859                  * PT wiring, so if we had removed an unmanaged page before
4860                  * we must adjust.
4861                  *
4862                  * We have to take care of the pmap wired count ourselves.
4863                  *
4864                  * Enter on the PV list if part of our managed memory.
4865                  */
4866                 KKASSERT(pte_pv && (pte_pv->pv_m == NULL || pte_pv->pv_m == m));
4867                 vm_page_spin_lock(m);
4868                 pte_pv->pv_m = m;
4869                 pmap_page_stats_adding(m);
4870                 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4871                 vm_page_flag_set(m, PG_MAPPED);
4872                 vm_page_spin_unlock(m);
4873
4874                 if (pt_pv && opa &&
4875                     (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
4876                         vm_page_unwire_quick(pt_pv->pv_m);
4877                 }
4878
4879                 /*
4880                  * Adjust pmap wired pages count for new entry.
4881                  */
4882                 if (wired) {
4883                         atomic_add_long(&pte_pv->pv_pmap->pm_stats.
4884                                         wired_count, 1);
4885                 }
4886         }
4887
4888         /*
4889          * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4890          *
4891          * User VMAs do not because those will be zero->non-zero, so no
4892          * stale entries to worry about at this point.
4893          *
4894          * For KVM there appear to still be issues.  Theoretically we
4895          * should be able to scrap the interlocks entirely but we
4896          * get crashes.
4897          */
4898         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4899                 pmap_inval_smp(pmap, va, 1, ptep, newpte);
4900         } else {
4901                 atomic_swap_long(ptep, newpte);
4902                 if (pt_pv == NULL)
4903                         cpu_invlpg((void *)va);
4904         }
4905
4906         if (newpte & pmap->pmap_bits[PG_RW_IDX])
4907                 vm_page_flag_set(m, PG_WRITEABLE);
4908
4909         /*
4910          * Cleanup
4911          */
4912 done:
4913         KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4914                  (m->flags & PG_MAPPED));
4915
4916         /*
4917          * Cleanup the pv entry, allowing other accessors.  If the new page
4918          * is not managed but we have a pte_pv (which was locking our
4919          * operation), we can free it now.  pte_pv->pv_m should be NULL.
4920          */
4921         if (pte_pv && (newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
4922                 pv_free(pte_pv, pt_pv);
4923         } else if (pte_pv) {
4924                 pv_put(pte_pv);
4925         } else if (pte_placemark) {
4926                 pv_placemarker_wakeup(pmap, pte_placemark);
4927         }
4928         if (pt_pv)
4929                 pv_put(pt_pv);
4930 }
4931
4932 /*
4933  * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4934  * This code also assumes that the pmap has no pre-existing entry for this
4935  * VA.
4936  *
4937  * This code currently may only be used on user pmaps, not kernel_pmap.
4938  */
4939 void
4940 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4941 {
4942         pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4943 }
4944
4945 /*
4946  * Make a temporary mapping for a physical address.  This is only intended
4947  * to be used for panic dumps.
4948  *
4949  * The caller is responsible for calling smp_invltlb().
4950  */
4951 void *
4952 pmap_kenter_temporary(vm_paddr_t pa, long i)
4953 {
4954         pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4955         return ((void *)crashdumpmap);
4956 }
4957
4958 #define MAX_INIT_PT (96)
4959
4960 /*
4961  * This routine preloads the ptes for a given object into the specified pmap.
4962  * This eliminates the blast of soft faults on process startup and
4963  * immediately after an mmap.
4964  */
4965 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4966
4967 void
4968 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4969                     vm_object_t object, vm_pindex_t pindex,
4970                     vm_size_t size, int limit)
4971 {
4972         struct rb_vm_page_scan_info info;
4973         struct lwp *lp;
4974         vm_size_t psize;
4975
4976         /*
4977          * We can't preinit if read access isn't set or there is no pmap
4978          * or object.
4979          */
4980         if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4981                 return;
4982
4983         /*
4984          * We can't preinit if the pmap is not the current pmap
4985          */
4986         lp = curthread->td_lwp;
4987         if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4988                 return;
4989
4990         /*
4991          * Misc additional checks
4992          */
4993         psize = x86_64_btop(size);
4994
4995         if ((object->type != OBJT_VNODE) ||
4996                 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4997                         (object->resident_page_count > MAX_INIT_PT))) {
4998                 return;
4999         }
5000
5001         if (pindex + psize > object->size) {
5002                 if (object->size < pindex)
5003                         return;
5004                 psize = object->size - pindex;
5005         }
5006
5007         if (psize == 0)
5008                 return;
5009
5010         /*
5011          * If everything is segment-aligned do not pre-init here.  Instead
5012          * allow the normal vm_fault path to pass a segment hint to
5013          * pmap_enter() which will then use an object-referenced shared
5014          * page table page.
5015          */
5016         if ((addr & SEG_MASK) == 0 &&
5017             (ctob(psize) & SEG_MASK) == 0 &&
5018             (ctob(pindex) & SEG_MASK) == 0) {
5019                 return;
5020         }
5021
5022         /*
5023          * Use a red-black scan to traverse the requested range and load
5024          * any valid pages found into the pmap.
5025          *
5026          * We cannot safely scan the object's memq without holding the
5027          * object token.
5028          */
5029         info.start_pindex = pindex;
5030         info.end_pindex = pindex + psize - 1;
5031         info.limit = limit;
5032         info.mpte = NULL;
5033         info.addr = addr;
5034         info.pmap = pmap;
5035
5036         vm_object_hold_shared(object);
5037         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
5038                                 pmap_object_init_pt_callback, &info);
5039         vm_object_drop(object);
5040 }
5041
5042 static
5043 int
5044 pmap_object_init_pt_callback(vm_page_t p, void *data)
5045 {
5046         struct rb_vm_page_scan_info *info = data;
5047         vm_pindex_t rel_index;
5048
5049         /*
5050          * don't allow an madvise to blow away our really
5051          * free pages allocating pv entries.
5052          */
5053         if ((info->limit & MAP_PREFAULT_MADVISE) &&
5054                 vmstats.v_free_count < vmstats.v_free_reserved) {
5055                     return(-1);
5056         }
5057
5058         /*
5059          * Ignore list markers and ignore pages we cannot instantly
5060          * busy (while holding the object token).
5061          */
5062         if (p->flags & PG_MARKER)
5063                 return 0;
5064         if (vm_page_busy_try(p, TRUE))
5065                 return 0;
5066         if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
5067             (p->flags & PG_FICTITIOUS) == 0) {
5068                 if ((p->queue - p->pc) == PQ_CACHE)
5069                         vm_page_deactivate(p);
5070                 rel_index = p->pindex - info->start_pindex;
5071                 pmap_enter_quick(info->pmap,
5072                                  info->addr + x86_64_ptob(rel_index), p);
5073         }
5074         vm_page_wakeup(p);
5075         lwkt_yield();
5076         return(0);
5077 }
5078
5079 /*
5080  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
5081  * address.
5082  *
5083  * Returns FALSE if it would be non-trivial or if a pte is already loaded
5084  * into the slot.
5085  *
5086  * XXX This is safe only because page table pages are not freed.
5087  */
5088 int
5089 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
5090 {
5091         pt_entry_t *pte;
5092
5093         /*spin_lock(&pmap->pm_spin);*/
5094         if ((pte = pmap_pte(pmap, addr)) != NULL) {
5095                 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
5096                         /*spin_unlock(&pmap->pm_spin);*/
5097                         return FALSE;
5098                 }
5099         }
5100         /*spin_unlock(&pmap->pm_spin);*/
5101         return TRUE;
5102 }
5103
5104 /*
5105  * Change the wiring attribute for a pmap/va pair.  The mapping must already
5106  * exist in the pmap.  The mapping may or may not be managed.  The wiring in
5107  * the page is not changed, the page is returned so the caller can adjust
5108  * its wiring (the page is not locked in any way).
5109  *
5110  * Wiring is not a hardware characteristic so there is no need to invalidate
5111  * TLB.  However, in an SMP environment we must use a locked bus cycle to
5112  * update the pte (if we are not using the pmap_inval_*() API that is)...
5113  * it's ok to do this for simple wiring changes.
5114  */
5115 vm_page_t
5116 pmap_unwire(pmap_t pmap, vm_offset_t va)
5117 {
5118         pt_entry_t *ptep;
5119         pv_entry_t pt_pv;
5120         vm_paddr_t pa;
5121         vm_page_t m;
5122
5123         if (pmap == NULL)
5124                 return NULL;
5125
5126         /*
5127          * Assume elements in the kernel pmap are stable
5128          */
5129         if (pmap == &kernel_pmap) {
5130                 if (pmap_pt(pmap, va) == 0)
5131                         return NULL;
5132                 ptep = pmap_pte_quick(pmap, va);
5133
5134                 if (pmap_pte_w(pmap, ptep))
5135                         atomic_add_long(&pmap->pm_stats.wired_count,-1);
5136                 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5137                 pa = *ptep & PG_FRAME;
5138                 m = PHYS_TO_VM_PAGE(pa);
5139         } else {
5140                 /*
5141                  * We can only [un]wire pmap-local pages (we cannot wire
5142                  * shared pages)
5143                  */
5144                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
5145                 if (pt_pv == NULL)
5146                         return NULL;
5147
5148                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
5149                 if ((*ptep & pmap->pmap_bits[PG_V_IDX]) == 0) {
5150                         pv_put(pt_pv);
5151                         return NULL;
5152                 }
5153
5154                 if (pmap_pte_w(pmap, ptep)) {
5155                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.wired_count,
5156                                         -1);
5157                 }
5158                 /* XXX else return NULL so caller doesn't unwire m ? */
5159
5160                 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5161
5162                 pa = *ptep & PG_FRAME;
5163                 m = PHYS_TO_VM_PAGE(pa);        /* held by wired count */
5164                 pv_put(pt_pv);
5165         }
5166         return m;
5167 }
5168
5169 /*
5170  * Copy the range specified by src_addr/len from the source map to
5171  * the range dst_addr/len in the destination map.
5172  *
5173  * This routine is only advisory and need not do anything.
5174  */
5175 void
5176 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
5177           vm_size_t len, vm_offset_t src_addr)
5178 {
5179 }
5180
5181 /*
5182  * pmap_zero_page:
5183  *
5184  *      Zero the specified physical page.
5185  *
5186  *      This function may be called from an interrupt and no locking is
5187  *      required.
5188  */
5189 void
5190 pmap_zero_page(vm_paddr_t phys)
5191 {
5192         vm_offset_t va = PHYS_TO_DMAP(phys);
5193
5194         pagezero((void *)va);
5195 }
5196
5197 /*
5198  * pmap_zero_page:
5199  *
5200  *      Zero part of a physical page by mapping it into memory and clearing
5201  *      its contents with bzero.
5202  *
5203  *      off and size may not cover an area beyond a single hardware page.
5204  */
5205 void
5206 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
5207 {
5208         vm_offset_t virt = PHYS_TO_DMAP(phys);
5209
5210         bzero((char *)virt + off, size);
5211 }
5212
5213 /*
5214  * pmap_copy_page:
5215  *
5216  *      Copy the physical page from the source PA to the target PA.
5217  *      This function may be called from an interrupt.  No locking
5218  *      is required.
5219  */
5220 void
5221 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
5222 {
5223         vm_offset_t src_virt, dst_virt;
5224
5225         src_virt = PHYS_TO_DMAP(src);
5226         dst_virt = PHYS_TO_DMAP(dst);
5227         bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
5228 }
5229
5230 /*
5231  * pmap_copy_page_frag:
5232  *
5233  *      Copy the physical page from the source PA to the target PA.
5234  *      This function may be called from an interrupt.  No locking
5235  *      is required.
5236  */
5237 void
5238 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
5239 {
5240         vm_offset_t src_virt, dst_virt;
5241
5242         src_virt = PHYS_TO_DMAP(src);
5243         dst_virt = PHYS_TO_DMAP(dst);
5244
5245         bcopy((char *)src_virt + (src & PAGE_MASK),
5246               (char *)dst_virt + (dst & PAGE_MASK),
5247               bytes);
5248 }
5249
5250 /*
5251  * Returns true if the pmap's pv is one of the first 16 pvs linked to from
5252  * this page.  This count may be changed upwards or downwards in the future;
5253  * it is only necessary that true be returned for a small subset of pmaps
5254  * for proper page aging.
5255  */
5256 boolean_t
5257 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
5258 {
5259         pv_entry_t pv;
5260         int loops = 0;
5261
5262         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5263                 return FALSE;
5264
5265         vm_page_spin_lock(m);
5266         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5267                 if (pv->pv_pmap == pmap) {
5268                         vm_page_spin_unlock(m);
5269                         return TRUE;
5270                 }
5271                 loops++;
5272                 if (loops >= 16)
5273                         break;
5274         }
5275         vm_page_spin_unlock(m);
5276         return (FALSE);
5277 }
5278
5279 /*
5280  * Remove all pages from specified address space this aids process exit
5281  * speeds.  Also, this code may be special cased for the current process
5282  * only.
5283  */
5284 void
5285 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5286 {
5287         pmap_remove_noinval(pmap, sva, eva);
5288         cpu_invltlb();
5289 }
5290
5291 /*
5292  * pmap_testbit tests bits in pte's note that the testbit/clearbit
5293  * routines are inline, and a lot of things compile-time evaluate.
5294  */
5295 static
5296 boolean_t
5297 pmap_testbit(vm_page_t m, int bit)
5298 {
5299         pv_entry_t pv;
5300         pt_entry_t *pte;
5301         pmap_t pmap;
5302
5303         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5304                 return FALSE;
5305
5306         if (TAILQ_FIRST(&m->md.pv_list) == NULL)
5307                 return FALSE;
5308         vm_page_spin_lock(m);
5309         if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
5310                 vm_page_spin_unlock(m);
5311                 return FALSE;
5312         }
5313
5314         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5315
5316 #if defined(PMAP_DIAGNOSTIC)
5317                 if (pv->pv_pmap == NULL) {
5318                         kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
5319                             pv->pv_pindex);
5320                         continue;
5321                 }
5322 #endif
5323                 pmap = pv->pv_pmap;
5324
5325                 /*
5326                  * If the bit being tested is the modified bit, then
5327                  * mark clean_map and ptes as never
5328                  * modified.
5329                  *
5330                  * WARNING!  Because we do not lock the pv, *pte can be in a
5331                  *           state of flux.  Despite this the value of *pte
5332                  *           will still be related to the vm_page in some way
5333                  *           because the pv cannot be destroyed as long as we
5334                  *           hold the vm_page spin lock.
5335                  */
5336                 if (bit == PG_A_IDX || bit == PG_M_IDX) {
5337                                 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5338                         if (!pmap_track_modified(pv->pv_pindex))
5339                                 continue;
5340                 }
5341
5342                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5343                 if (*pte & pmap->pmap_bits[bit]) {
5344                         vm_page_spin_unlock(m);
5345                         return TRUE;
5346                 }
5347         }
5348         vm_page_spin_unlock(m);
5349         return (FALSE);
5350 }
5351
5352 /*
5353  * This routine is used to modify bits in ptes.  Only one bit should be
5354  * specified.  PG_RW requires special handling.
5355  *
5356  * Caller must NOT hold any spin locks
5357  */
5358 static __inline
5359 void
5360 pmap_clearbit(vm_page_t m, int bit_index)
5361 {
5362         pv_entry_t pv;
5363         pt_entry_t *pte;
5364         pt_entry_t pbits;
5365         pmap_t pmap;
5366
5367         if (bit_index == PG_RW_IDX)
5368                 vm_page_flag_clear(m, PG_WRITEABLE);
5369         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5370                 return;
5371
5372         /*
5373          * PG_M or PG_A case
5374          *
5375          * Loop over all current mappings setting/clearing as appropos If
5376          * setting RO do we need to clear the VAC?
5377          *
5378          * NOTE: When clearing PG_M we could also (not implemented) drop
5379          *       through to the PG_RW code and clear PG_RW too, forcing
5380          *       a fault on write to redetect PG_M for virtual kernels, but
5381          *       it isn't necessary since virtual kernels invalidate the
5382          *       pte when they clear the VPTE_M bit in their virtual page
5383          *       tables.
5384          *
5385          * NOTE: Does not re-dirty the page when clearing only PG_M.
5386          *
5387          * NOTE: Because we do not lock the pv, *pte can be in a state of
5388          *       flux.  Despite this the value of *pte is still somewhat
5389          *       related while we hold the vm_page spin lock.
5390          *
5391          *       *pte can be zero due to this race.  Since we are clearing
5392          *       bits we basically do no harm when this race  ccurs.
5393          */
5394         if (bit_index != PG_RW_IDX) {
5395                 vm_page_spin_lock(m);
5396                 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5397 #if defined(PMAP_DIAGNOSTIC)
5398                         if (pv->pv_pmap == NULL) {
5399                                 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5400                                     pv->pv_pindex);
5401                                 continue;
5402                         }
5403 #endif
5404                         pmap = pv->pv_pmap;
5405                         pte = pmap_pte_quick(pv->pv_pmap,
5406                                              pv->pv_pindex << PAGE_SHIFT);
5407                         pbits = *pte;
5408                         if (pbits & pmap->pmap_bits[bit_index])
5409                                 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5410                 }
5411                 vm_page_spin_unlock(m);
5412                 return;
5413         }
5414
5415         /*
5416          * Clear PG_RW.  Also clears PG_M and marks the page dirty if PG_M
5417          * was set.
5418          */
5419 restart:
5420         vm_page_spin_lock(m);
5421         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5422                 /*
5423                  * don't write protect pager mappings
5424                  */
5425                 if (!pmap_track_modified(pv->pv_pindex))
5426                         continue;
5427
5428 #if defined(PMAP_DIAGNOSTIC)
5429                 if (pv->pv_pmap == NULL) {
5430                         kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5431                                 pv->pv_pindex);
5432                         continue;
5433                 }
5434 #endif
5435                 pmap = pv->pv_pmap;
5436
5437                 /*
5438                  * Skip pages which do not have PG_RW set.
5439                  */
5440                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5441                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5442                         continue;
5443
5444                 /*
5445                  * Lock the PV
5446                  */
5447                 if (pv_hold_try(pv)) {
5448                         vm_page_spin_unlock(m);
5449                 } else {
5450                         vm_page_spin_unlock(m);
5451                         pv_lock(pv);    /* held, now do a blocking lock */
5452                         pv_unlock(pv);
5453                         pv_drop(pv);
5454                         goto restart;
5455                 }
5456                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
5457                 for (;;) {
5458                         pt_entry_t nbits;
5459
5460                         pbits = *pte;
5461                         cpu_ccfence();
5462                         nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5463                                           pmap->pmap_bits[PG_M_IDX]);
5464                         if (pmap_inval_smp_cmpset(pmap,
5465                                      ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5466                                      pte, pbits, nbits)) {
5467                                 break;
5468                         }
5469                         cpu_pause();
5470                 }
5471
5472                 /*
5473                  * If PG_M was found to be set while we were clearing PG_RW
5474                  * we also clear PG_M (done above) and mark the page dirty.
5475                  * Callers expect this behavior.
5476                  *
5477                  * we lost pv so it cannot be used as an iterator.  In fact,
5478                  * because we couldn't necessarily lock it atomically it may
5479                  * have moved within the list and ALSO cannot be used as an
5480                  * iterator.
5481                  */
5482                 vm_page_spin_lock(m);
5483                 if (pbits & pmap->pmap_bits[PG_M_IDX])
5484                         vm_page_dirty(m);
5485                 vm_page_spin_unlock(m);
5486                 pv_put(pv);
5487                 goto restart;
5488         }
5489         vm_page_spin_unlock(m);
5490 }
5491
5492 /*
5493  * Lower the permission for all mappings to a given page.
5494  *
5495  * Page must be busied by caller.  Because page is busied by caller this
5496  * should not be able to race a pmap_enter().
5497  */
5498 void
5499 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5500 {
5501         /* JG NX support? */
5502         if ((prot & VM_PROT_WRITE) == 0) {
5503                 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5504                         /*
5505                          * NOTE: pmap_clearbit(.. PG_RW) also clears
5506                          *       the PG_WRITEABLE flag in (m).
5507                          */
5508                         pmap_clearbit(m, PG_RW_IDX);
5509                 } else {
5510                         pmap_remove_all(m);
5511                 }
5512         }
5513 }
5514
5515 vm_paddr_t
5516 pmap_phys_address(vm_pindex_t ppn)
5517 {
5518         return (x86_64_ptob(ppn));
5519 }
5520
5521 /*
5522  * Return a count of reference bits for a page, clearing those bits.
5523  * It is not necessary for every reference bit to be cleared, but it
5524  * is necessary that 0 only be returned when there are truly no
5525  * reference bits set.
5526  *
5527  * XXX: The exact number of bits to check and clear is a matter that
5528  * should be tested and standardized at some point in the future for
5529  * optimal aging of shared pages.
5530  *
5531  * This routine may not block.
5532  */
5533 int
5534 pmap_ts_referenced(vm_page_t m)
5535 {
5536         pv_entry_t pv;
5537         pt_entry_t *pte;
5538         pmap_t pmap;
5539         int rtval = 0;
5540
5541         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5542                 return (rtval);
5543
5544         vm_page_spin_lock(m);
5545         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5546                 if (!pmap_track_modified(pv->pv_pindex))
5547                         continue;
5548                 pmap = pv->pv_pmap;
5549                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5550                 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5551                         atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5552                         rtval++;
5553                         if (rtval > 4)
5554                                 break;
5555                 }
5556         }
5557         vm_page_spin_unlock(m);
5558         return (rtval);
5559 }
5560
5561 /*
5562  *      pmap_is_modified:
5563  *
5564  *      Return whether or not the specified physical page was modified
5565  *      in any physical maps.
5566  */
5567 boolean_t
5568 pmap_is_modified(vm_page_t m)
5569 {
5570         boolean_t res;
5571
5572         res = pmap_testbit(m, PG_M_IDX);
5573         return (res);
5574 }
5575
5576 /*
5577  *      Clear the modify bits on the specified physical page.
5578  */
5579 void
5580 pmap_clear_modify(vm_page_t m)
5581 {
5582         pmap_clearbit(m, PG_M_IDX);
5583 }
5584
5585 /*
5586  *      pmap_clear_reference:
5587  *
5588  *      Clear the reference bit on the specified physical page.
5589  */
5590 void
5591 pmap_clear_reference(vm_page_t m)
5592 {
5593         pmap_clearbit(m, PG_A_IDX);
5594 }
5595
5596 /*
5597  * Miscellaneous support routines follow
5598  */
5599
5600 static
5601 void
5602 i386_protection_init(void)
5603 {
5604         int *kp, prot;
5605
5606         /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit  */
5607         kp = protection_codes;
5608         for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5609                 switch (prot) {
5610                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5611                         /*
5612                          * Read access is also 0. There isn't any execute bit,
5613                          * so just make it readable.
5614                          */
5615                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5616                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5617                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5618                         *kp++ = 0;
5619                         break;
5620                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5621                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5622                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5623                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5624                         *kp++ = pmap_bits_default[PG_RW_IDX];
5625                         break;
5626                 }
5627         }
5628 }
5629
5630 /*
5631  * Map a set of physical memory pages into the kernel virtual
5632  * address space. Return a pointer to where it is mapped. This
5633  * routine is intended to be used for mapping device memory,
5634  * NOT real memory.
5635  *
5636  * NOTE: We can't use pgeflag unless we invalidate the pages one at
5637  *       a time.
5638  *
5639  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5640  *       work whether the cpu supports PAT or not.  The remaining PAT
5641  *       attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5642  *       supports PAT.
5643  */
5644 void *
5645 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5646 {
5647         return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5648 }
5649
5650 void *
5651 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5652 {
5653         return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5654 }
5655
5656 void *
5657 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5658 {
5659         return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5660 }
5661
5662 /*
5663  * Map a set of physical memory pages into the kernel virtual
5664  * address space. Return a pointer to where it is mapped. This
5665  * routine is intended to be used for mapping device memory,
5666  * NOT real memory.
5667  */
5668 void *
5669 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5670 {
5671         vm_offset_t va, tmpva, offset;
5672         pt_entry_t *pte;
5673         vm_size_t tmpsize;
5674
5675         offset = pa & PAGE_MASK;
5676         size = roundup(offset + size, PAGE_SIZE);
5677
5678         va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5679         if (va == 0)
5680                 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5681
5682         pa = pa & ~PAGE_MASK;
5683         for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5684                 pte = vtopte(tmpva);
5685                 *pte = pa |
5686                     kernel_pmap.pmap_bits[PG_RW_IDX] |
5687                     kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5688                     kernel_pmap.pmap_cache_bits[mode];
5689                 tmpsize -= PAGE_SIZE;
5690                 tmpva += PAGE_SIZE;
5691                 pa += PAGE_SIZE;
5692         }
5693         pmap_invalidate_range(&kernel_pmap, va, va + size);
5694         pmap_invalidate_cache_range(va, va + size);
5695
5696         return ((void *)(va + offset));
5697 }
5698
5699 void
5700 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5701 {
5702         vm_offset_t base, offset;
5703
5704         base = va & ~PAGE_MASK;
5705         offset = va & PAGE_MASK;
5706         size = roundup(offset + size, PAGE_SIZE);
5707         pmap_qremove(va, size >> PAGE_SHIFT);
5708         kmem_free(&kernel_map, base, size);
5709 }
5710
5711 /*
5712  * Sets the memory attribute for the specified page.
5713  */
5714 void
5715 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5716 {
5717
5718     m->pat_mode = ma;
5719
5720     /*
5721      * If "m" is a normal page, update its direct mapping.  This update
5722      * can be relied upon to perform any cache operations that are
5723      * required for data coherence.
5724      */
5725     if ((m->flags & PG_FICTITIOUS) == 0)
5726         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5727 }
5728
5729 /*
5730  * Change the PAT attribute on an existing kernel memory map.  Caller
5731  * must ensure that the virtual memory in question is not accessed
5732  * during the adjustment.
5733  */
5734 void
5735 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5736 {
5737         pt_entry_t *pte;
5738         vm_offset_t base;
5739         int changed = 0;
5740
5741         if (va == 0)
5742                 panic("pmap_change_attr: va is NULL");
5743         base = trunc_page(va);
5744
5745         while (count) {
5746                 pte = vtopte(va);
5747                 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5748                        kernel_pmap.pmap_cache_bits[mode];
5749                 --count;
5750                 va += PAGE_SIZE;
5751         }
5752
5753         changed = 1;    /* XXX: not optimal */
5754
5755         /*
5756          * Flush CPU caches if required to make sure any data isn't cached that
5757          * shouldn't be, etc.
5758          */
5759         if (changed) {
5760                 pmap_invalidate_range(&kernel_pmap, base, va);
5761                 pmap_invalidate_cache_range(base, va);
5762         }
5763 }
5764
5765 /*
5766  * perform the pmap work for mincore
5767  */
5768 int
5769 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5770 {
5771         pt_entry_t *ptep, pte;
5772         vm_page_t m;
5773         int val = 0;
5774
5775         ptep = pmap_pte(pmap, addr);
5776
5777         if (ptep && (pte = *ptep) != 0) {
5778                 vm_offset_t pa;
5779
5780                 val = MINCORE_INCORE;
5781                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5782                         goto done;
5783
5784                 pa = pte & PG_FRAME;
5785
5786                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5787                         m = NULL;
5788                 else
5789                         m = PHYS_TO_VM_PAGE(pa);
5790
5791                 /*
5792                  * Modified by us
5793                  */
5794                 if (pte & pmap->pmap_bits[PG_M_IDX])
5795                         val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5796                 /*
5797                  * Modified by someone
5798                  */
5799                 else if (m && (m->dirty || pmap_is_modified(m)))
5800                         val |= MINCORE_MODIFIED_OTHER;
5801                 /*
5802                  * Referenced by us
5803                  */
5804                 if (pte & pmap->pmap_bits[PG_A_IDX])
5805                         val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5806
5807                 /*
5808                  * Referenced by someone
5809                  */
5810                 else if (m && ((m->flags & PG_REFERENCED) ||
5811                                 pmap_ts_referenced(m))) {
5812                         val |= MINCORE_REFERENCED_OTHER;
5813                         vm_page_flag_set(m, PG_REFERENCED);
5814                 }
5815         }
5816 done:
5817
5818         return val;
5819 }
5820
5821 /*
5822  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
5823  * vmspace will be ref'd and the old one will be deref'd.
5824  *
5825  * The vmspace for all lwps associated with the process will be adjusted
5826  * and cr3 will be reloaded if any lwp is the current lwp.
5827  *
5828  * The process must hold the vmspace->vm_map.token for oldvm and newvm
5829  */
5830 void
5831 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5832 {
5833         struct vmspace *oldvm;
5834         struct lwp *lp;
5835
5836         oldvm = p->p_vmspace;
5837         if (oldvm != newvm) {
5838                 if (adjrefs)
5839                         vmspace_ref(newvm);
5840                 p->p_vmspace = newvm;
5841                 KKASSERT(p->p_nthreads == 1);
5842                 lp = RB_ROOT(&p->p_lwp_tree);
5843                 pmap_setlwpvm(lp, newvm);
5844                 if (adjrefs)
5845                         vmspace_rel(oldvm);
5846         }
5847 }
5848
5849 /*
5850  * Set the vmspace for a LWP.  The vmspace is almost universally set the
5851  * same as the process vmspace, but virtual kernels need to swap out contexts
5852  * on a per-lwp basis.
5853  *
5854  * Caller does not necessarily hold any vmspace tokens.  Caller must control
5855  * the lwp (typically be in the context of the lwp).  We use a critical
5856  * section to protect against statclock and hardclock (statistics collection).
5857  */
5858 void
5859 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5860 {
5861         struct vmspace *oldvm;
5862         struct pmap *pmap;
5863
5864         oldvm = lp->lwp_vmspace;
5865
5866         if (oldvm != newvm) {
5867                 crit_enter();
5868                 lp->lwp_vmspace = newvm;
5869                 if (curthread->td_lwp == lp) {
5870                         pmap = vmspace_pmap(newvm);
5871                         ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5872                         if (pmap->pm_active_lock & CPULOCK_EXCL)
5873                                 pmap_interlock_wait(newvm);
5874 #if defined(SWTCH_OPTIM_STATS)
5875                         tlb_flush_count++;
5876 #endif
5877                         if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5878                                 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5879                         } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5880                                 curthread->td_pcb->pcb_cr3 = KPML4phys;
5881                         } else {
5882                                 panic("pmap_setlwpvm: unknown pmap type\n");
5883                         }
5884                         load_cr3(curthread->td_pcb->pcb_cr3);
5885                         pmap = vmspace_pmap(oldvm);
5886                         ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5887                                                mycpu->gd_cpuid);
5888                 }
5889                 crit_exit();
5890         }
5891 }
5892
5893 /*
5894  * Called when switching to a locked pmap, used to interlock against pmaps
5895  * undergoing modifications to prevent us from activating the MMU for the
5896  * target pmap until all such modifications have completed.  We have to do
5897  * this because the thread making the modifications has already set up its
5898  * SMP synchronization mask.
5899  *
5900  * This function cannot sleep!
5901  *
5902  * No requirements.
5903  */
5904 void
5905 pmap_interlock_wait(struct vmspace *vm)
5906 {
5907         struct pmap *pmap = &vm->vm_pmap;
5908
5909         if (pmap->pm_active_lock & CPULOCK_EXCL) {
5910                 crit_enter();
5911                 KKASSERT(curthread->td_critcount >= 2);
5912                 DEBUG_PUSH_INFO("pmap_interlock_wait");
5913                 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5914                         cpu_ccfence();
5915                         lwkt_process_ipiq();
5916                 }
5917                 DEBUG_POP_INFO();
5918                 crit_exit();
5919         }
5920 }
5921
5922 vm_offset_t
5923 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5924 {
5925
5926         if ((obj == NULL) || (size < NBPDR) ||
5927             ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5928                 return addr;
5929         }
5930
5931         addr = roundup2(addr, NBPDR);
5932         return addr;
5933 }
5934
5935 /*
5936  * Used by kmalloc/kfree, page already exists at va
5937  */
5938 vm_page_t
5939 pmap_kvtom(vm_offset_t va)
5940 {
5941         pt_entry_t *ptep = vtopte(va);
5942
5943         KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5944         return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5945 }
5946
5947 /*
5948  * Initialize machine-specific shared page directory support.  This
5949  * is executed when a VM object is created.
5950  */
5951 void
5952 pmap_object_init(vm_object_t object)
5953 {
5954         object->md.pmap_rw = NULL;
5955         object->md.pmap_ro = NULL;
5956 }
5957
5958 /*
5959  * Clean up machine-specific shared page directory support.  This
5960  * is executed when a VM object is destroyed.
5961  */
5962 void
5963 pmap_object_free(vm_object_t object)
5964 {
5965         pmap_t pmap;
5966
5967         if ((pmap = object->md.pmap_rw) != NULL) {
5968                 object->md.pmap_rw = NULL;
5969                 pmap_remove_noinval(pmap,
5970                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5971                 CPUMASK_ASSZERO(pmap->pm_active);
5972                 pmap_release(pmap);
5973                 pmap_puninit(pmap);
5974                 kfree(pmap, M_OBJPMAP);
5975         }
5976         if ((pmap = object->md.pmap_ro) != NULL) {
5977                 object->md.pmap_ro = NULL;
5978                 pmap_remove_noinval(pmap,
5979                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5980                 CPUMASK_ASSZERO(pmap->pm_active);
5981                 pmap_release(pmap);
5982                 pmap_puninit(pmap);
5983                 kfree(pmap, M_OBJPMAP);
5984         }
5985 }
5986
5987 /*
5988  * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
5989  * VM page and issue a pginfo->callback.
5990  *
5991  * We are expected to dispose of any non-NULL pte_pv.
5992  */
5993 static
5994 void
5995 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
5996                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
5997                       pv_entry_t pt_pv, int sharept,
5998                       vm_offset_t va, pt_entry_t *ptep, void *arg)
5999 {
6000         struct pmap_pgscan_info *pginfo = arg;
6001         vm_page_t m;
6002
6003         if (pte_pv) {
6004                 /*
6005                  * Try to busy the page while we hold the pte_pv locked.
6006                  */
6007                 KKASSERT(pte_pv->pv_m);
6008                 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
6009                 if (vm_page_busy_try(m, TRUE) == 0) {
6010                         if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
6011                                 /*
6012                                  * The callback is issued with the pte_pv
6013                                  * unlocked and put away, and the pt_pv
6014                                  * unlocked.
6015                                  */
6016                                 pv_put(pte_pv);
6017                                 if (pt_pv) {
6018                                         vm_page_wire_quick(pt_pv->pv_m);
6019                                         pv_unlock(pt_pv);
6020                                 }
6021                                 if (pginfo->callback(pginfo, va, m) < 0)
6022                                         info->stop = 1;
6023                                 if (pt_pv) {
6024                                         pv_lock(pt_pv);
6025                                         vm_page_unwire_quick(pt_pv->pv_m);
6026                                 }
6027                         } else {
6028                                 vm_page_wakeup(m);
6029                                 pv_put(pte_pv);
6030                         }
6031                 } else {
6032                         ++pginfo->busycount;
6033                         pv_put(pte_pv);
6034                 }
6035         } else {
6036                 /*
6037                  * Shared page table or unmanaged page (sharept or !sharept)
6038                  */
6039                 pv_placemarker_wakeup(pmap, pte_placemark);
6040         }
6041 }
6042
6043 void
6044 pmap_pgscan(struct pmap_pgscan_info *pginfo)
6045 {
6046         struct pmap_scan_info info;
6047
6048         pginfo->offset = pginfo->beg_addr;
6049         info.pmap = pginfo->pmap;
6050         info.sva = pginfo->beg_addr;
6051         info.eva = pginfo->end_addr;
6052         info.func = pmap_pgscan_callback;
6053         info.arg = pginfo;
6054         pmap_scan(&info, 0);
6055         if (info.stop == 0)
6056                 pginfo->offset = pginfo->end_addr;
6057 }
6058
6059 /*
6060  * Wait for a placemarker that we do not own to clear.  The placemarker
6061  * in question is not necessary set to the pindex we want, we may have
6062  * to wait on the element because we want to reserve it ourselves.
6063  */
6064 static
6065 void
6066 pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark)
6067 {
6068         spin_lock(&pmap->pm_spin);
6069         if (*pmark != PM_NOPLACEMARK) {
6070                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
6071                 ssleep(pmark, &pmap->pm_spin, 0, "pvplw", 0);
6072         }
6073         spin_unlock(&pmap->pm_spin);
6074 }
6075
6076 /*
6077  * Wakeup a placemarker that we own.  Replace the entry with
6078  * PM_NOPLACEMARK and issue a wakeup() if necessary.
6079  */
6080 static
6081 void
6082 pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark)
6083 {
6084         vm_pindex_t pindex;
6085
6086         spin_lock(&pmap->pm_spin);
6087         pindex = atomic_swap_long(pmark, PM_NOPLACEMARK);
6088         spin_unlock(&pmap->pm_spin);
6089         KKASSERT(pindex != PM_NOPLACEMARK);
6090         if (pindex & PM_PLACEMARK_WAKEUP)
6091                 wakeup(pmark);
6092 }