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 int pv_entry_max=0, pv_entry_high_water=0;
 199 static int pmap_pagedaemon_waken = 0;
 200 static struct pv_entry *pvinit;
 201
 202 /*
 203  * All those kernel PT submaps that BSD is so fond of
 204  */
 205 pt_entry_t *CMAP1 = NULL, *ptmmap;
 206 caddr_t CADDR1 = NULL, ptvmmap = NULL;
 207 static pt_entry_t *msgbufmap;
 208 struct msgbuf *msgbufp=NULL;
 209
 210 /*
 211  * PMAP default PG_* bits. Needed to be able to add
 212  * EPT/NPT pagetable pmap_bits for the VMM module
 213  */
 214 uint64_t pmap_bits_default[] = {
 215                 REGULAR_PMAP,                   /* TYPE_IDX             0 */
 216                 X86_PG_V,                       /* PG_V_IDX             1 */
 217                 X86_PG_RW,                      /* PG_RW_IDX            2 */
 218                 X86_PG_U,                       /* PG_U_IDX             3 */
 219                 X86_PG_A,                       /* PG_A_IDX             4 */
 220                 X86_PG_M,                       /* PG_M_IDX             5 */
 221                 X86_PG_PS,                      /* PG_PS_IDX3           6 */
 222                 X86_PG_G,                       /* PG_G_IDX             7 */
 223                 X86_PG_AVAIL1,                  /* PG_AVAIL1_IDX        8 */
 224                 X86_PG_AVAIL2,                  /* PG_AVAIL2_IDX        9 */
 225                 X86_PG_AVAIL3,                  /* PG_AVAIL3_IDX        10 */
 226                 X86_PG_NC_PWT | X86_PG_NC_PCD,  /* PG_N_IDX             11 */
 227 };
 228 /*
 229  * Crashdump maps.
 230  */
 231 static pt_entry_t *pt_crashdumpmap;
 232 static caddr_t crashdumpmap;
 233
 234 static int pmap_debug = 0;
 235 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
 236     &pmap_debug, 0, "Debug pmap's");
 237 #ifdef PMAP_DEBUG2
 238 static int pmap_enter_debug = 0;
 239 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
 240     &pmap_enter_debug, 0, "Debug pmap_enter's");
 241 #endif
 242 static int pmap_yield_count = 64;
 243 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
 244     &pmap_yield_count, 0, "Yield during init_pt/release");
 245 static int pmap_mmu_optimize = 0;
 246 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
 247     &pmap_mmu_optimize, 0, "Share page table pages when possible");
 248 int pmap_fast_kernel_cpusync = 0;
 249 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
 250     &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
 251 int pmap_dynamic_delete = 0;
 252 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
 253     &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
 254
 255 #define DISABLE_PSE
 256
 257 /* Standard user access funtions */
 258 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
 259     size_t *lencopied);
 260 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
 261 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
 262 extern int std_fubyte (const uint8_t *base);
 263 extern int std_subyte (uint8_t *base, uint8_t byte);
 264 extern int32_t std_fuword32 (const uint32_t *base);
 265 extern int64_t std_fuword64 (const uint64_t *base);
 266 extern int std_suword64 (uint64_t *base, uint64_t word);
 267 extern int std_suword32 (uint32_t *base, int word);
 268 extern uint32_t std_swapu32 (volatile uint32_t *base, uint32_t v);
 269 extern uint64_t std_swapu64 (volatile uint64_t *base, uint64_t v);
 270
 271 static void pv_hold(pv_entry_t pv);
 272 static int _pv_hold_try(pv_entry_t pv
 273                                 PMAP_DEBUG_DECL);
 274 static void pv_drop(pv_entry_t pv);
 275 static void _pv_lock(pv_entry_t pv
 276                                 PMAP_DEBUG_DECL);
 277 static void pv_unlock(pv_entry_t pv);
 278 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
 279                                 PMAP_DEBUG_DECL);
 280 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp
 281                                 PMAP_DEBUG_DECL);
 282 static void _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL);
 283 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex,
 284                                 vm_pindex_t **pmarkp, int *errorp);
 285 static void pv_put(pv_entry_t pv);
 286 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
 287 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
 288                       pv_entry_t *pvpp);
 289 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
 290                       pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
 291 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
 292                         pmap_inval_bulk_t *bulk, int destroy);
 293 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
 294 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
 295                         pmap_inval_bulk_t *bulk);
 296
 297 struct pmap_scan_info;
 298 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
 299                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
 300                       pv_entry_t pt_pv, int sharept,
 301                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 302 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
 303                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
 304                       pv_entry_t pt_pv, int sharept,
 305                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 306
 307 static void i386_protection_init (void);
 308 static void create_pagetables(vm_paddr_t *firstaddr);
 309 static void pmap_remove_all (vm_page_t m);
 310 static boolean_t pmap_testbit (vm_page_t m, int bit);
 311
 312 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
 313 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
 314
 315 static void pmap_pinit_defaults(struct pmap *pmap);
 316 static void pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark);
 317 static void pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark);
 318
 319 static unsigned pdir4mb;
 320
 321 static int
 322 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
 323 {
 324         if (pv1->pv_pindex < pv2->pv_pindex)
 325                 return(-1);
 326         if (pv1->pv_pindex > pv2->pv_pindex)
 327                 return(1);
 328         return(0);
 329 }
 330
 331 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
 332              pv_entry_compare, vm_pindex_t, pv_pindex);
 333
 334 static __inline
 335 void
 336 pmap_page_stats_adding(vm_page_t m)
 337 {
 338         globaldata_t gd = mycpu;
 339
 340         if (TAILQ_EMPTY(&m->md.pv_list)) {
 341                 ++gd->gd_vmtotal.t_arm;
 342         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 343                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 344                 ++gd->gd_vmtotal.t_armshr;
 345                 ++gd->gd_vmtotal.t_avmshr;
 346         } else {
 347                 ++gd->gd_vmtotal.t_avmshr;
 348         }
 349 }
 350
 351 static __inline
 352 void
 353 pmap_page_stats_deleting(vm_page_t m)
 354 {
 355         globaldata_t gd = mycpu;
 356
 357         if (TAILQ_EMPTY(&m->md.pv_list)) {
 358                 --gd->gd_vmtotal.t_arm;
 359         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 360                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 361                 --gd->gd_vmtotal.t_armshr;
 362                 --gd->gd_vmtotal.t_avmshr;
 363         } else {
 364                 --gd->gd_vmtotal.t_avmshr;
 365         }
 366 }
 367
 368 /*
 369  * Move the kernel virtual free pointer to the next
 370  * 2MB.  This is used to help improve performance
 371  * by using a large (2MB) page for much of the kernel
 372  * (.text, .data, .bss)
 373  */
 374 static
 375 vm_offset_t
 376 pmap_kmem_choose(vm_offset_t addr)
 377 {
 378         vm_offset_t newaddr = addr;
 379
 380         newaddr = roundup2(addr, NBPDR);
 381         return newaddr;
 382 }
 383
 384 /*
 385  * pmap_pte_quick:
 386  *
 387  *      Super fast pmap_pte routine best used when scanning the pv lists.
 388  *      This eliminates many course-grained invltlb calls.  Note that many of
 389  *      the pv list scans are across different pmaps and it is very wasteful
 390  *      to do an entire invltlb when checking a single mapping.
 391  */
 392 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
 393
 394 static
 395 pt_entry_t *
 396 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
 397 {
 398         return pmap_pte(pmap, va);
 399 }
 400
 401 /*
 402  * Returns the pindex of a page table entry (representing a terminal page).
 403  * There are NUPTE_TOTAL page table entries possible (a huge number)
 404  *
 405  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
 406  * We want to properly translate negative KVAs.
 407  */
 408 static __inline
 409 vm_pindex_t
 410 pmap_pte_pindex(vm_offset_t va)
 411 {
 412         return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
 413 }
 414
 415 /*
 416  * Returns the pindex of a page table.
 417  */
 418 static __inline
 419 vm_pindex_t
 420 pmap_pt_pindex(vm_offset_t va)
 421 {
 422         return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
 423 }
 424
 425 /*
 426  * Returns the pindex of a page directory.
 427  */
 428 static __inline
 429 vm_pindex_t
 430 pmap_pd_pindex(vm_offset_t va)
 431 {
 432         return (NUPTE_TOTAL + NUPT_TOTAL +
 433                 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
 434 }
 435
 436 static __inline
 437 vm_pindex_t
 438 pmap_pdp_pindex(vm_offset_t va)
 439 {
 440         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
 441                 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
 442 }
 443
 444 static __inline
 445 vm_pindex_t
 446 pmap_pml4_pindex(void)
 447 {
 448         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
 449 }
 450
 451 /*
 452  * Return various clipped indexes for a given VA
 453  *
 454  * Returns the index of a pt in a page directory, representing a page
 455  * table.
 456  */
 457 static __inline
 458 vm_pindex_t
 459 pmap_pt_index(vm_offset_t va)
 460 {
 461         return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
 462 }
 463
 464 /*
 465  * Returns the index of a pd in a page directory page, representing a page
 466  * directory.
 467  */
 468 static __inline
 469 vm_pindex_t
 470 pmap_pd_index(vm_offset_t va)
 471 {
 472         return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
 473 }
 474
 475 /*
 476  * Returns the index of a pdp in the pml4 table, representing a page
 477  * directory page.
 478  */
 479 static __inline
 480 vm_pindex_t
 481 pmap_pdp_index(vm_offset_t va)
 482 {
 483         return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
 484 }
 485
 486 /*
 487  * The placemarker hash must be broken up into four zones so lock
 488  * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
 489  *
 490  * Placemarkers are used to 'lock' page table indices that do not have
 491  * a pv_entry.  This allows the pmap to support managed and unmanaged
 492  * pages and shared page tables.
 493  */
 494 #define PM_PLACE_BASE   (PM_PLACEMARKS >> 2)
 495
 496 static __inline
 497 vm_pindex_t *
 498 pmap_placemarker_hash(pmap_t pmap, vm_pindex_t pindex)
 499 {
 500         int hi;
 501
 502         if (pindex < pmap_pt_pindex(0))         /* zone 0 - PTE */
 503                 hi = 0;
 504         else if (pindex < pmap_pd_pindex(0))    /* zone 1 - PT */
 505                 hi = PM_PLACE_BASE;
 506         else if (pindex < pmap_pdp_pindex(0))   /* zone 2 - PD */
 507                 hi = PM_PLACE_BASE << 1;
 508         else                                    /* zone 3 - PDP (and PML4E) */
 509                 hi = PM_PLACE_BASE | (PM_PLACE_BASE << 1);
 510         hi += pindex & (PM_PLACE_BASE - 1);
 511
 512         return (&pmap->pm_placemarks[hi]);
 513 }
 514
 515
 516 /*
 517  * Generic procedure to index a pte from a pt, pd, or pdp.
 518  *
 519  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
 520  *       a page table page index but is instead of PV lookup index.
 521  */
 522 static
 523 void *
 524 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
 525 {
 526         pt_entry_t *pte;
 527
 528         pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
 529         return(&pte[pindex]);
 530 }
 531
 532 /*
 533  * Return pointer to PDP slot in the PML4
 534  */
 535 static __inline
 536 pml4_entry_t *
 537 pmap_pdp(pmap_t pmap, vm_offset_t va)
 538 {
 539         return (&pmap->pm_pml4[pmap_pdp_index(va)]);
 540 }
 541
 542 /*
 543  * Return pointer to PD slot in the PDP given a pointer to the PDP
 544  */
 545 static __inline
 546 pdp_entry_t *
 547 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
 548 {
 549         pdp_entry_t *pd;
 550
 551         pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
 552         return (&pd[pmap_pd_index(va)]);
 553 }
 554
 555 /*
 556  * Return pointer to PD slot in the PDP.
 557  */
 558 static __inline
 559 pdp_entry_t *
 560 pmap_pd(pmap_t pmap, vm_offset_t va)
 561 {
 562         pml4_entry_t *pdp;
 563
 564         pdp = pmap_pdp(pmap, va);
 565         if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
 566                 return NULL;
 567         return (pmap_pdp_to_pd(*pdp, va));
 568 }
 569
 570 /*
 571  * Return pointer to PT slot in the PD given a pointer to the PD
 572  */
 573 static __inline
 574 pd_entry_t *
 575 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
 576 {
 577         pd_entry_t *pt;
 578
 579         pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
 580         return (&pt[pmap_pt_index(va)]);
 581 }
 582
 583 /*
 584  * Return pointer to PT slot in the PD
 585  *
 586  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
 587  *                   so we cannot lookup the PD via the PDP.  Instead we
 588  *                   must look it up via the pmap.
 589  */
 590 static __inline
 591 pd_entry_t *
 592 pmap_pt(pmap_t pmap, vm_offset_t va)
 593 {
 594         pdp_entry_t *pd;
 595         pv_entry_t pv;
 596         vm_pindex_t pd_pindex;
 597
 598         if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
 599                 pd_pindex = pmap_pd_pindex(va);
 600                 spin_lock(&pmap->pm_spin);
 601                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
 602                 spin_unlock(&pmap->pm_spin);
 603                 if (pv == NULL || pv->pv_m == NULL)
 604                         return NULL;
 605                 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
 606         } else {
 607                 pd = pmap_pd(pmap, va);
 608                 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
 609                          return NULL;
 610                 return (pmap_pd_to_pt(*pd, va));
 611         }
 612 }
 613
 614 /*
 615  * Return pointer to PTE slot in the PT given a pointer to the PT
 616  */
 617 static __inline
 618 pt_entry_t *
 619 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
 620 {
 621         pt_entry_t *pte;
 622
 623         pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
 624         return (&pte[pmap_pte_index(va)]);
 625 }
 626
 627 /*
 628  * Return pointer to PTE slot in the PT
 629  */
 630 static __inline
 631 pt_entry_t *
 632 pmap_pte(pmap_t pmap, vm_offset_t va)
 633 {
 634         pd_entry_t *pt;
 635
 636         pt = pmap_pt(pmap, va);
 637         if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
 638                  return NULL;
 639         if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
 640                 return ((pt_entry_t *)pt);
 641         return (pmap_pt_to_pte(*pt, va));
 642 }
 643
 644 /*
 645  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
 646  * the PT layer.  This will speed up core pmap operations considerably.
 647  *
 648  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
 649  *       must be in a known associated state (typically by being locked when
 650  *       the pmap spinlock isn't held).  We allow the race for that case.
 651  *
 652  * NOTE: pm_pvhint is only accessed (read) with the spin-lock held, using
 653  *       cpu_ccfence() to prevent compiler optimizations from reloading the
 654  *       field.
 655  */
 656 static __inline
 657 void
 658 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
 659 {
 660         if (pindex >= pmap_pt_pindex(0) && pindex < pmap_pd_pindex(0)) {
 661                 if (pv->pv_pmap)
 662                         pv->pv_pmap->pm_pvhint = pv;
 663         }
 664 }
 665
 666
 667 /*
 668  * Return address of PT slot in PD (KVM only)
 669  *
 670  * Cannot be used for user page tables because it might interfere with
 671  * the shared page-table-page optimization (pmap_mmu_optimize).
 672  */
 673 static __inline
 674 pd_entry_t *
 675 vtopt(vm_offset_t va)
 676 {
 677         uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
 678                                   NPML4EPGSHIFT)) - 1);
 679
 680         return (PDmap + ((va >> PDRSHIFT) & mask));
 681 }
 682
 683 /*
 684  * KVM - return address of PTE slot in PT
 685  */
 686 static __inline
 687 pt_entry_t *
 688 vtopte(vm_offset_t va)
 689 {
 690         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 691                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 692
 693         return (PTmap + ((va >> PAGE_SHIFT) & mask));
 694 }
 695
 696 static uint64_t
 697 allocpages(vm_paddr_t *firstaddr, long n)
 698 {
 699         uint64_t ret;
 700
 701         ret = *firstaddr;
 702         bzero((void *)ret, n * PAGE_SIZE);
 703         *firstaddr += n * PAGE_SIZE;
 704         return (ret);
 705 }
 706
 707 static
 708 void
 709 create_pagetables(vm_paddr_t *firstaddr)
 710 {
 711         long i;         /* must be 64 bits */
 712         long nkpt_base;
 713         long nkpt_phys;
 714         int j;
 715
 716         /*
 717          * We are running (mostly) V=P at this point
 718          *
 719          * Calculate NKPT - number of kernel page tables.  We have to
 720          * accomodoate prealloction of the vm_page_array, dump bitmap,
 721          * MSGBUF_SIZE, and other stuff.  Be generous.
 722          *
 723          * Maxmem is in pages.
 724          *
 725          * ndmpdp is the number of 1GB pages we wish to map.
 726          */
 727         ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
 728         if (ndmpdp < 4)         /* Minimum 4GB of dirmap */
 729                 ndmpdp = 4;
 730         KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
 731
 732         /*
 733          * Starting at the beginning of kvm (not KERNBASE).
 734          */
 735         nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
 736         nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
 737         nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
 738                        ndmpdp) + 511) / 512;
 739         nkpt_phys += 128;
 740
 741         /*
 742          * Starting at KERNBASE - map 2G worth of page table pages.
 743          * KERNBASE is offset -2G from the end of kvm.
 744          */
 745         nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
 746
 747         /*
 748          * Allocate pages
 749          */
 750         KPTbase = allocpages(firstaddr, nkpt_base);
 751         KPTphys = allocpages(firstaddr, nkpt_phys);
 752         KPML4phys = allocpages(firstaddr, 1);
 753         KPDPphys = allocpages(firstaddr, NKPML4E);
 754         KPDphys = allocpages(firstaddr, NKPDPE);
 755
 756         /*
 757          * Calculate the page directory base for KERNBASE,
 758          * that is where we start populating the page table pages.
 759          * Basically this is the end - 2.
 760          */
 761         KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
 762
 763         DMPDPphys = allocpages(firstaddr, NDMPML4E);
 764         if ((amd_feature & AMDID_PAGE1GB) == 0)
 765                 DMPDphys = allocpages(firstaddr, ndmpdp);
 766         dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
 767
 768         /*
 769          * Fill in the underlying page table pages for the area around
 770          * KERNBASE.  This remaps low physical memory to KERNBASE.
 771          *
 772          * Read-only from zero to physfree
 773          * XXX not fully used, underneath 2M pages
 774          */
 775         for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
 776                 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
 777                 ((pt_entry_t *)KPTbase)[i] |=
 778                     pmap_bits_default[PG_RW_IDX] |
 779                     pmap_bits_default[PG_V_IDX] |
 780                     pmap_bits_default[PG_G_IDX];
 781         }
 782
 783         /*
 784          * Now map the initial kernel page tables.  One block of page
 785          * tables is placed at the beginning of kernel virtual memory,
 786          * and another block is placed at KERNBASE to map the kernel binary,
 787          * data, bss, and initial pre-allocations.
 788          */
 789         for (i = 0; i < nkpt_base; i++) {
 790                 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
 791                 ((pd_entry_t *)KPDbase)[i] |=
 792                     pmap_bits_default[PG_RW_IDX] |
 793                     pmap_bits_default[PG_V_IDX];
 794         }
 795         for (i = 0; i < nkpt_phys; i++) {
 796                 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
 797                 ((pd_entry_t *)KPDphys)[i] |=
 798                     pmap_bits_default[PG_RW_IDX] |
 799                     pmap_bits_default[PG_V_IDX];
 800         }
 801
 802         /*
 803          * Map from zero to end of allocations using 2M pages as an
 804          * optimization.  This will bypass some of the KPTBase pages
 805          * above in the KERNBASE area.
 806          */
 807         for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
 808                 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
 809                 ((pd_entry_t *)KPDbase)[i] |=
 810                     pmap_bits_default[PG_RW_IDX] |
 811                     pmap_bits_default[PG_V_IDX] |
 812                     pmap_bits_default[PG_PS_IDX] |
 813                     pmap_bits_default[PG_G_IDX];
 814         }
 815
 816         /*
 817          * And connect up the PD to the PDP.  The kernel pmap is expected
 818          * to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
 819          */
 820         for (i = 0; i < NKPDPE; i++) {
 821                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
 822                                 KPDphys + (i << PAGE_SHIFT);
 823                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
 824                     pmap_bits_default[PG_RW_IDX] |
 825                     pmap_bits_default[PG_V_IDX] |
 826                     pmap_bits_default[PG_U_IDX];
 827         }
 828
 829         /*
 830          * Now set up the direct map space using either 2MB or 1GB pages
 831          * Preset PG_M and PG_A because demotion expects it.
 832          *
 833          * When filling in entries in the PD pages make sure any excess
 834          * entries are set to zero as we allocated enough PD pages
 835          */
 836         if ((amd_feature & AMDID_PAGE1GB) == 0) {
 837                 for (i = 0; i < NPDEPG * ndmpdp; i++) {
 838                         ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
 839                         ((pd_entry_t *)DMPDphys)[i] |=
 840                             pmap_bits_default[PG_RW_IDX] |
 841                             pmap_bits_default[PG_V_IDX] |
 842                             pmap_bits_default[PG_PS_IDX] |
 843                             pmap_bits_default[PG_G_IDX] |
 844                             pmap_bits_default[PG_M_IDX] |
 845                             pmap_bits_default[PG_A_IDX];
 846                 }
 847
 848                 /*
 849                  * And the direct map space's PDP
 850                  */
 851                 for (i = 0; i < ndmpdp; i++) {
 852                         ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
 853                                                         (i << PAGE_SHIFT);
 854                         ((pdp_entry_t *)DMPDPphys)[i] |=
 855                             pmap_bits_default[PG_RW_IDX] |
 856                             pmap_bits_default[PG_V_IDX] |
 857                             pmap_bits_default[PG_U_IDX];
 858                 }
 859         } else {
 860                 for (i = 0; i < ndmpdp; i++) {
 861                         ((pdp_entry_t *)DMPDPphys)[i] =
 862                                                 (vm_paddr_t)i << PDPSHIFT;
 863                         ((pdp_entry_t *)DMPDPphys)[i] |=
 864                             pmap_bits_default[PG_RW_IDX] |
 865                             pmap_bits_default[PG_V_IDX] |
 866                             pmap_bits_default[PG_PS_IDX] |
 867                             pmap_bits_default[PG_G_IDX] |
 868                             pmap_bits_default[PG_M_IDX] |
 869                             pmap_bits_default[PG_A_IDX];
 870                 }
 871         }
 872
 873         /* And recursively map PML4 to itself in order to get PTmap */
 874         ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
 875         ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
 876             pmap_bits_default[PG_RW_IDX] |
 877             pmap_bits_default[PG_V_IDX] |
 878             pmap_bits_default[PG_U_IDX];
 879
 880         /*
 881          * Connect the Direct Map slots up to the PML4
 882          */
 883         for (j = 0; j < NDMPML4E; ++j) {
 884                 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
 885                     (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
 886                     pmap_bits_default[PG_RW_IDX] |
 887                     pmap_bits_default[PG_V_IDX] |
 888                     pmap_bits_default[PG_U_IDX];
 889         }
 890
 891         /*
 892          * Connect the KVA slot up to the PML4
 893          */
 894         ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
 895         ((pdp_entry_t *)KPML4phys)[KPML4I] |=
 896             pmap_bits_default[PG_RW_IDX] |
 897             pmap_bits_default[PG_V_IDX] |
 898             pmap_bits_default[PG_U_IDX];
 899 }
 900
 901 /*
 902  *      Bootstrap the system enough to run with virtual memory.
 903  *
 904  *      On the i386 this is called after mapping has already been enabled
 905  *      and just syncs the pmap module with what has already been done.
 906  *      [We can't call it easily with mapping off since the kernel is not
 907  *      mapped with PA == VA, hence we would have to relocate every address
 908  *      from the linked base (virtual) address "KERNBASE" to the actual
 909  *      (physical) address starting relative to 0]
 910  */
 911 void
 912 pmap_bootstrap(vm_paddr_t *firstaddr)
 913 {
 914         vm_offset_t va;
 915         pt_entry_t *pte;
 916         int i;
 917
 918         KvaStart = VM_MIN_KERNEL_ADDRESS;
 919         KvaEnd = VM_MAX_KERNEL_ADDRESS;
 920         KvaSize = KvaEnd - KvaStart;
 921
 922         avail_start = *firstaddr;
 923
 924         /*
 925          * Create an initial set of page tables to run the kernel in.
 926          */
 927         create_pagetables(firstaddr);
 928
 929         virtual2_start = KvaStart;
 930         virtual2_end = PTOV_OFFSET;
 931
 932         virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
 933         virtual_start = pmap_kmem_choose(virtual_start);
 934
 935         virtual_end = VM_MAX_KERNEL_ADDRESS;
 936
 937         /* XXX do %cr0 as well */
 938         load_cr4(rcr4() | CR4_PGE | CR4_PSE);
 939         load_cr3(KPML4phys);
 940
 941         /*
 942          * Initialize protection array.
 943          */
 944         i386_protection_init();
 945
 946         /*
 947          * The kernel's pmap is statically allocated so we don't have to use
 948          * pmap_create, which is unlikely to work correctly at this part of
 949          * the boot sequence (XXX and which no longer exists).
 950          */
 951         kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
 952         kernel_pmap.pm_count = 1;
 953         CPUMASK_ASSALLONES(kernel_pmap.pm_active);
 954         RB_INIT(&kernel_pmap.pm_pvroot);
 955         spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
 956         for (i = 0; i < PM_PLACEMARKS; ++i)
 957                 kernel_pmap.pm_placemarks[i] = PM_NOPLACEMARK;
 958
 959         /*
 960          * Reserve some special page table entries/VA space for temporary
 961          * mapping of pages.
 962          */
 963 #define SYSMAP(c, p, v, n)      \
 964         v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
 965
 966         va = virtual_start;
 967         pte = vtopte(va);
 968
 969         /*
 970          * CMAP1/CMAP2 are used for zeroing and copying pages.
 971          */
 972         SYSMAP(caddr_t, CMAP1, CADDR1, 1)
 973
 974         /*
 975          * Crashdump maps.
 976          */
 977         SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
 978
 979         /*
 980          * ptvmmap is used for reading arbitrary physical pages via
 981          * /dev/mem.
 982          */
 983         SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
 984
 985         /*
 986          * msgbufp is used to map the system message buffer.
 987          * XXX msgbufmap is not used.
 988          */
 989         SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
 990                atop(round_page(MSGBUF_SIZE)))
 991
 992         virtual_start = va;
 993         virtual_start = pmap_kmem_choose(virtual_start);
 994
 995         *CMAP1 = 0;
 996
 997         /*
 998          * PG_G is terribly broken on SMP because we IPI invltlb's in some
 999          * cases rather then invl1pg.  Actually, I don't even know why it
1000          * works under UP because self-referential page table mappings
1001          */
1002 //      pgeflag = 0;
1003
1004 /*
1005  * Initialize the 4MB page size flag
1006  */
1007 //      pseflag = 0;
1008 /*
1009  * The 4MB page version of the initial
1010  * kernel page mapping.
1011  */
1012         pdir4mb = 0;
1013
1014 #if !defined(DISABLE_PSE)
1015         if (cpu_feature & CPUID_PSE) {
1016                 pt_entry_t ptditmp;
1017                 /*
1018                  * Note that we have enabled PSE mode
1019                  */
1020 //              pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
1021                 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
1022                 ptditmp &= ~(NBPDR - 1);
1023                 ptditmp |= pmap_bits_default[PG_V_IDX] |
1024                     pmap_bits_default[PG_RW_IDX] |
1025                     pmap_bits_default[PG_PS_IDX] |
1026                     pmap_bits_default[PG_U_IDX];
1027 //                  pgeflag;
1028                 pdir4mb = ptditmp;
1029         }
1030 #endif
1031         cpu_invltlb();
1032
1033         /* Initialize the PAT MSR */
1034         pmap_init_pat();
1035         pmap_pinit_defaults(&kernel_pmap);
1036
1037         TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1038                           &pmap_fast_kernel_cpusync);
1039
1040 }
1041
1042 /*
1043  * Setup the PAT MSR.
1044  */
1045 void
1046 pmap_init_pat(void)
1047 {
1048         uint64_t pat_msr;
1049         u_long cr0, cr4;
1050
1051         /*
1052          * Default values mapping PATi,PCD,PWT bits at system reset.
1053          * The default values effectively ignore the PATi bit by
1054          * repeating the encodings for 0-3 in 4-7, and map the PCD
1055          * and PWT bit combinations to the expected PAT types.
1056          */
1057         pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |        /* 000 */
1058                   PAT_VALUE(1, PAT_WRITE_THROUGH) |     /* 001 */
1059                   PAT_VALUE(2, PAT_UNCACHED) |          /* 010 */
1060                   PAT_VALUE(3, PAT_UNCACHEABLE) |       /* 011 */
1061                   PAT_VALUE(4, PAT_WRITE_BACK) |        /* 100 */
1062                   PAT_VALUE(5, PAT_WRITE_THROUGH) |     /* 101 */
1063                   PAT_VALUE(6, PAT_UNCACHED) |          /* 110 */
1064                   PAT_VALUE(7, PAT_UNCACHEABLE);        /* 111 */
1065         pat_pte_index[PAT_WRITE_BACK]   = 0;
1066         pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
1067         pat_pte_index[PAT_UNCACHED]     = X86_PG_NC_PCD;
1068         pat_pte_index[PAT_UNCACHEABLE]  = X86_PG_NC_PCD | X86_PG_NC_PWT;
1069         pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1070         pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1071
1072         if (cpu_feature & CPUID_PAT) {
1073                 /*
1074                  * If we support the PAT then set-up entries for
1075                  * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1076                  * 4 and 5.
1077                  */
1078                 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1079                           PAT_VALUE(4, PAT_WRITE_PROTECTED);
1080                 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1081                           PAT_VALUE(5, PAT_WRITE_COMBINING);
1082                 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1083                 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1084
1085                 /*
1086                  * Then enable the PAT
1087                  */
1088
1089                 /* Disable PGE. */
1090                 cr4 = rcr4();
1091                 load_cr4(cr4 & ~CR4_PGE);
1092
1093                 /* Disable caches (CD = 1, NW = 0). */
1094                 cr0 = rcr0();
1095                 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1096
1097                 /* Flushes caches and TLBs. */
1098                 wbinvd();
1099                 cpu_invltlb();
1100
1101                 /* Update PAT and index table. */
1102                 wrmsr(MSR_PAT, pat_msr);
1103
1104                 /* Flush caches and TLBs again. */
1105                 wbinvd();
1106                 cpu_invltlb();
1107
1108                 /* Restore caches and PGE. */
1109                 load_cr0(cr0);
1110                 load_cr4(cr4);
1111                 PatMsr = pat_msr;
1112         }
1113 }
1114
1115 /*
1116  * Set 4mb pdir for mp startup
1117  */
1118 void
1119 pmap_set_opt(void)
1120 {
1121         if (cpu_feature & CPUID_PSE) {
1122                 load_cr4(rcr4() | CR4_PSE);
1123                 if (pdir4mb && mycpu->gd_cpuid == 0) {  /* only on BSP */
1124                         cpu_invltlb();
1125                 }
1126         }
1127 }
1128
1129 /*
1130  *      Initialize the pmap module.
1131  *      Called by vm_init, to initialize any structures that the pmap
1132  *      system needs to map virtual memory.
1133  *      pmap_init has been enhanced to support in a fairly consistant
1134  *      way, discontiguous physical memory.
1135  */
1136 void
1137 pmap_init(void)
1138 {
1139         int i;
1140         int initial_pvs;
1141
1142         /*
1143          * Allocate memory for random pmap data structures.  Includes the
1144          * pv_head_table.
1145          */
1146
1147         for (i = 0; i < vm_page_array_size; i++) {
1148                 vm_page_t m;
1149
1150                 m = &vm_page_array[i];
1151                 TAILQ_INIT(&m->md.pv_list);
1152         }
1153
1154         /*
1155          * init the pv free list
1156          */
1157         initial_pvs = vm_page_array_size;
1158         if (initial_pvs < MINPV)
1159                 initial_pvs = MINPV;
1160         pvzone = &pvzone_store;
1161         pvinit = (void *)kmem_alloc(&kernel_map,
1162                                     initial_pvs * sizeof (struct pv_entry),
1163                                     VM_SUBSYS_PVENTRY);
1164         zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1165                   pvinit, initial_pvs);
1166
1167         /*
1168          * Now it is safe to enable pv_table recording.
1169          */
1170         pmap_initialized = TRUE;
1171 }
1172
1173 /*
1174  * Initialize the address space (zone) for the pv_entries.  Set a
1175  * high water mark so that the system can recover from excessive
1176  * numbers of pv entries.
1177  */
1178 void
1179 pmap_init2(void)
1180 {
1181         int shpgperproc = PMAP_SHPGPERPROC;
1182         int entry_max;
1183
1184         TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1185         pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1186         TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1187         pv_entry_high_water = 9 * (pv_entry_max / 10);
1188
1189         /*
1190          * Subtract out pages already installed in the zone (hack)
1191          */
1192         entry_max = pv_entry_max - vm_page_array_size;
1193         if (entry_max <= 0)
1194                 entry_max = 1;
1195
1196         zinitna(pvzone, NULL, 0, entry_max, ZONE_INTERRUPT);
1197
1198         /*
1199          * Enable dynamic deletion of empty higher-level page table pages
1200          * by default only if system memory is < 8GB (use 7GB for slop).
1201          * This can save a little memory, but imposes significant
1202          * performance overhead for things like bulk builds, and for programs
1203          * which do a lot of memory mapping and memory unmapping.
1204          */
1205         if (pmap_dynamic_delete < 0) {
1206                 if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1207                         pmap_dynamic_delete = 1;
1208                 else
1209                         pmap_dynamic_delete = 0;
1210         }
1211 }
1212
1213 /*
1214  * Typically used to initialize a fictitious page by vm/device_pager.c
1215  */
1216 void
1217 pmap_page_init(struct vm_page *m)
1218 {
1219         vm_page_init(m);
1220         TAILQ_INIT(&m->md.pv_list);
1221 }
1222
1223 /***************************************************
1224  * Low level helper routines.....
1225  ***************************************************/
1226
1227 /*
1228  * this routine defines the region(s) of memory that should
1229  * not be tested for the modified bit.
1230  */
1231 static __inline
1232 int
1233 pmap_track_modified(vm_pindex_t pindex)
1234 {
1235         vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1236         if ((va < clean_sva) || (va >= clean_eva))
1237                 return 1;
1238         else
1239                 return 0;
1240 }
1241
1242 /*
1243  * Extract the physical page address associated with the map/VA pair.
1244  * The page must be wired for this to work reliably.
1245  */
1246 vm_paddr_t
1247 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
1248 {
1249         vm_paddr_t rtval;
1250         pv_entry_t pt_pv;
1251         pt_entry_t *ptep;
1252
1253         rtval = 0;
1254         if (va >= VM_MAX_USER_ADDRESS) {
1255                 /*
1256                  * Kernel page directories might be direct-mapped and
1257                  * there is typically no PV tracking of pte's
1258                  */
1259                 pd_entry_t *pt;
1260
1261                 pt = pmap_pt(pmap, va);
1262                 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1263                         if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1264                                 rtval = *pt & PG_PS_FRAME;
1265                                 rtval |= va & PDRMASK;
1266                         } else {
1267                                 ptep = pmap_pt_to_pte(*pt, va);
1268                                 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1269                                         rtval = *ptep & PG_FRAME;
1270                                         rtval |= va & PAGE_MASK;
1271                                 }
1272                         }
1273                 }
1274                 if (handlep)
1275                         *handlep = NULL;
1276         } else {
1277                 /*
1278                  * User pages currently do not direct-map the page directory
1279                  * and some pages might not used managed PVs.  But all PT's
1280                  * will have a PV.
1281                  */
1282                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1283                 if (pt_pv) {
1284                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1285                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1286                                 rtval = *ptep & PG_FRAME;
1287                                 rtval |= va & PAGE_MASK;
1288                         }
1289                         if (handlep)
1290                                 *handlep = pt_pv;       /* locked until done */
1291                         else
1292                                 pv_put (pt_pv);
1293                 } else if (handlep) {
1294                         *handlep = NULL;
1295                 }
1296         }
1297         return rtval;
1298 }
1299
1300 void
1301 pmap_extract_done(void *handle)
1302 {
1303         if (handle)
1304                 pv_put((pv_entry_t)handle);
1305 }
1306
1307 /*
1308  * Similar to extract but checks protections, SMP-friendly short-cut for
1309  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1310  * fall-through to the real fault code.  Does not work with HVM page
1311  * tables.
1312  *
1313  * The returned page, if not NULL, is held (and not busied).
1314  *
1315  * WARNING! THE RETURNED PAGE IS ONLY HELD AND NOT SUITABLE FOR READING
1316  *          OR WRITING AS-IS.
1317  */
1318 vm_page_t
1319 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot, int *busyp)
1320 {
1321         if (pmap &&
1322             va < VM_MAX_USER_ADDRESS &&
1323             (pmap->pm_flags & PMAP_HVM) == 0) {
1324                 pv_entry_t pt_pv;
1325                 pv_entry_t pte_pv;
1326                 pt_entry_t *ptep;
1327                 pt_entry_t req;
1328                 vm_page_t m;
1329                 int error;
1330
1331                 req = pmap->pmap_bits[PG_V_IDX] |
1332                       pmap->pmap_bits[PG_U_IDX];
1333                 if (prot & VM_PROT_WRITE)
1334                         req |= pmap->pmap_bits[PG_RW_IDX];
1335
1336                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1337                 if (pt_pv == NULL)
1338                         return (NULL);
1339                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1340                 if ((*ptep & req) != req) {
1341                         pv_put(pt_pv);
1342                         return (NULL);
1343                 }
1344                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), NULL, &error);
1345                 if (pte_pv && error == 0) {
1346                         m = pte_pv->pv_m;
1347                         if (prot & VM_PROT_WRITE) {
1348                                 /* interlocked by presence of pv_entry */
1349                                 vm_page_dirty(m);
1350                         }
1351                         if (busyp) {
1352                                 if (prot & VM_PROT_WRITE) {
1353                                         if (vm_page_busy_try(m, TRUE))
1354                                                 m = NULL;
1355                                         *busyp = 1;
1356                                 } else {
1357                                         vm_page_hold(m);
1358                                         *busyp = 0;
1359                                 }
1360                         } else {
1361                                 vm_page_hold(m);
1362                         }
1363                         pv_put(pte_pv);
1364                 } else if (pte_pv) {
1365                         pv_drop(pte_pv);
1366                         m = NULL;
1367                 } else {
1368                         /* error, since we didn't request a placemarker */
1369                         m = NULL;
1370                 }
1371                 pv_put(pt_pv);
1372                 return(m);
1373         } else {
1374                 return(NULL);
1375         }
1376 }
1377
1378 /*
1379  * Extract the physical page address associated kernel virtual address.
1380  */
1381 vm_paddr_t
1382 pmap_kextract(vm_offset_t va)
1383 {
1384         pd_entry_t pt;          /* pt entry in pd */
1385         vm_paddr_t pa;
1386
1387         if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1388                 pa = DMAP_TO_PHYS(va);
1389         } else {
1390                 pt = *vtopt(va);
1391                 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1392                         pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1393                 } else {
1394                         /*
1395                          * Beware of a concurrent promotion that changes the
1396                          * PDE at this point!  For example, vtopte() must not
1397                          * be used to access the PTE because it would use the
1398                          * new PDE.  It is, however, safe to use the old PDE
1399                          * because the page table page is preserved by the
1400                          * promotion.
1401                          */
1402                         pa = *pmap_pt_to_pte(pt, va);
1403                         pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1404                 }
1405         }
1406         return pa;
1407 }
1408
1409 /***************************************************
1410  * Low level mapping routines.....
1411  ***************************************************/
1412
1413 /*
1414  * Routine: pmap_kenter
1415  * Function:
1416  *      Add a wired page to the KVA
1417  *      NOTE! note that in order for the mapping to take effect -- you
1418  *      should do an invltlb after doing the pmap_kenter().
1419  */
1420 void
1421 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1422 {
1423         pt_entry_t *ptep;
1424         pt_entry_t npte;
1425
1426         npte = pa |
1427                kernel_pmap.pmap_bits[PG_RW_IDX] |
1428                kernel_pmap.pmap_bits[PG_V_IDX];
1429 //             pgeflag;
1430         ptep = vtopte(va);
1431 #if 1
1432         pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1433 #else
1434         /* FUTURE */
1435         if (*ptep)
1436                 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1437         else
1438                 *ptep = npte;
1439 #endif
1440 }
1441
1442 /*
1443  * Similar to pmap_kenter(), except we only invalidate the mapping on the
1444  * current CPU.  Returns 0 if the previous pte was 0, 1 if it wasn't
1445  * (caller can conditionalize calling smp_invltlb()).
1446  */
1447 int
1448 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1449 {
1450         pt_entry_t *ptep;
1451         pt_entry_t npte;
1452         int res;
1453
1454         npte = pa | kernel_pmap.pmap_bits[PG_RW_IDX] |
1455                     kernel_pmap.pmap_bits[PG_V_IDX];
1456         // npte |= pgeflag;
1457         ptep = vtopte(va);
1458 #if 1
1459         res = 1;
1460 #else
1461         /* FUTURE */
1462         res = (*ptep != 0);
1463 #endif
1464         atomic_swap_long(ptep, npte);
1465         cpu_invlpg((void *)va);
1466
1467         return res;
1468 }
1469
1470 /*
1471  * Enter addresses into the kernel pmap but don't bother
1472  * doing any tlb invalidations.  Caller will do a rollup
1473  * invalidation via pmap_rollup_inval().
1474  */
1475 int
1476 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1477 {
1478         pt_entry_t *ptep;
1479         pt_entry_t npte;
1480         int res;
1481
1482         npte = pa |
1483             kernel_pmap.pmap_bits[PG_RW_IDX] |
1484             kernel_pmap.pmap_bits[PG_V_IDX];
1485 //          pgeflag;
1486         ptep = vtopte(va);
1487 #if 1
1488         res = 1;
1489 #else
1490         /* FUTURE */
1491         res = (*ptep != 0);
1492 #endif
1493         atomic_swap_long(ptep, npte);
1494         cpu_invlpg((void *)va);
1495
1496         return res;
1497 }
1498
1499 /*
1500  * remove a page from the kernel pagetables
1501  */
1502 void
1503 pmap_kremove(vm_offset_t va)
1504 {
1505         pt_entry_t *ptep;
1506
1507         ptep = vtopte(va);
1508         pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1509 }
1510
1511 void
1512 pmap_kremove_quick(vm_offset_t va)
1513 {
1514         pt_entry_t *ptep;
1515
1516         ptep = vtopte(va);
1517         (void)pte_load_clear(ptep);
1518         cpu_invlpg((void *)va);
1519 }
1520
1521 /*
1522  * Remove addresses from the kernel pmap but don't bother
1523  * doing any tlb invalidations.  Caller will do a rollup
1524  * invalidation via pmap_rollup_inval().
1525  */
1526 void
1527 pmap_kremove_noinval(vm_offset_t va)
1528 {
1529         pt_entry_t *ptep;
1530
1531         ptep = vtopte(va);
1532         (void)pte_load_clear(ptep);
1533 }
1534
1535 /*
1536  * XXX these need to be recoded.  They are not used in any critical path.
1537  */
1538 void
1539 pmap_kmodify_rw(vm_offset_t va)
1540 {
1541         atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1542         cpu_invlpg((void *)va);
1543 }
1544
1545 /* NOT USED
1546 void
1547 pmap_kmodify_nc(vm_offset_t va)
1548 {
1549         atomic_set_long(vtopte(va), PG_N);
1550         cpu_invlpg((void *)va);
1551 }
1552 */
1553
1554 /*
1555  * Used to map a range of physical addresses into kernel virtual
1556  * address space during the low level boot, typically to map the
1557  * dump bitmap, message buffer, and vm_page_array.
1558  *
1559  * These mappings are typically made at some pointer after the end of the
1560  * kernel text+data.
1561  *
1562  * We could return PHYS_TO_DMAP(start) here and not allocate any
1563  * via (*virtp), but then kmem from userland and kernel dumps won't
1564  * have access to the related pointers.
1565  */
1566 vm_offset_t
1567 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1568 {
1569         vm_offset_t va;
1570         vm_offset_t va_start;
1571
1572         /*return PHYS_TO_DMAP(start);*/
1573
1574         va_start = *virtp;
1575         va = va_start;
1576
1577         while (start < end) {
1578                 pmap_kenter_quick(va, start);
1579                 va += PAGE_SIZE;
1580                 start += PAGE_SIZE;
1581         }
1582         *virtp = va;
1583         return va_start;
1584 }
1585
1586 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
1587
1588 /*
1589  * Remove the specified set of pages from the data and instruction caches.
1590  *
1591  * In contrast to pmap_invalidate_cache_range(), this function does not
1592  * rely on the CPU's self-snoop feature, because it is intended for use
1593  * when moving pages into a different cache domain.
1594  */
1595 void
1596 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1597 {
1598         vm_offset_t daddr, eva;
1599         int i;
1600
1601         if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1602             (cpu_feature & CPUID_CLFSH) == 0)
1603                 wbinvd();
1604         else {
1605                 cpu_mfence();
1606                 for (i = 0; i < count; i++) {
1607                         daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1608                         eva = daddr + PAGE_SIZE;
1609                         for (; daddr < eva; daddr += cpu_clflush_line_size)
1610                                 clflush(daddr);
1611                 }
1612                 cpu_mfence();
1613         }
1614 }
1615
1616 void
1617 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1618 {
1619         KASSERT((sva & PAGE_MASK) == 0,
1620             ("pmap_invalidate_cache_range: sva not page-aligned"));
1621         KASSERT((eva & PAGE_MASK) == 0,
1622             ("pmap_invalidate_cache_range: eva not page-aligned"));
1623
1624         if (cpu_feature & CPUID_SS) {
1625                 ; /* If "Self Snoop" is supported, do nothing. */
1626         } else {
1627                 /* Globally invalidate caches */
1628                 cpu_wbinvd_on_all_cpus();
1629         }
1630 }
1631
1632 /*
1633  * Invalidate the specified range of virtual memory on all cpus associated
1634  * with the pmap.
1635  */
1636 void
1637 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1638 {
1639         pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1640 }
1641
1642 /*
1643  * Add a list of wired pages to the kva.  This routine is used for temporary
1644  * kernel mappings such as those found in buffer cache buffer.  Page
1645  * modifications and accesses are not tracked or recorded.
1646  *
1647  * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1648  *       semantics as previous mappings may have been zerod without any
1649  *       invalidation.
1650  *
1651  * The page *must* be wired.
1652  */
1653 void
1654 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1655 {
1656         vm_offset_t end_va;
1657         vm_offset_t va;
1658
1659         end_va = beg_va + count * PAGE_SIZE;
1660
1661         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1662                 pt_entry_t pte;
1663                 pt_entry_t *ptep;
1664
1665                 ptep = vtopte(va);
1666                 pte = VM_PAGE_TO_PHYS(*m) |
1667                         kernel_pmap.pmap_bits[PG_RW_IDX] |
1668                         kernel_pmap.pmap_bits[PG_V_IDX] |
1669                         kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1670 //              pgeflag;
1671                 atomic_swap_long(ptep, pte);
1672                 m++;
1673         }
1674         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1675 }
1676
1677 /*
1678  * This routine jerks page mappings from the kernel -- it is meant only
1679  * for temporary mappings such as those found in buffer cache buffers.
1680  * No recording modified or access status occurs.
1681  *
1682  * MPSAFE, INTERRUPT SAFE (cluster callback)
1683  */
1684 void
1685 pmap_qremove(vm_offset_t beg_va, int count)
1686 {
1687         vm_offset_t end_va;
1688         vm_offset_t va;
1689
1690         end_va = beg_va + count * PAGE_SIZE;
1691
1692         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1693                 pt_entry_t *pte;
1694
1695                 pte = vtopte(va);
1696                 (void)pte_load_clear(pte);
1697                 cpu_invlpg((void *)va);
1698         }
1699         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1700 }
1701
1702 /*
1703  * This routine removes temporary kernel mappings, only invalidating them
1704  * on the current cpu.  It should only be used under carefully controlled
1705  * conditions.
1706  */
1707 void
1708 pmap_qremove_quick(vm_offset_t beg_va, int count)
1709 {
1710         vm_offset_t end_va;
1711         vm_offset_t va;
1712
1713         end_va = beg_va + count * PAGE_SIZE;
1714
1715         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1716                 pt_entry_t *pte;
1717
1718                 pte = vtopte(va);
1719                 (void)pte_load_clear(pte);
1720                 cpu_invlpg((void *)va);
1721         }
1722 }
1723
1724 /*
1725  * This routine removes temporary kernel mappings *without* invalidating
1726  * the TLB.  It can only be used on permanent kva reservations such as those
1727  * found in buffer cache buffers, under carefully controlled circumstances.
1728  *
1729  * NOTE: Repopulating these KVAs requires unconditional invalidation.
1730  *       (pmap_qenter() does unconditional invalidation).
1731  */
1732 void
1733 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1734 {
1735         vm_offset_t end_va;
1736         vm_offset_t va;
1737
1738         end_va = beg_va + count * PAGE_SIZE;
1739
1740         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1741                 pt_entry_t *pte;
1742
1743                 pte = vtopte(va);
1744                 (void)pte_load_clear(pte);
1745         }
1746 }
1747
1748 /*
1749  * Create a new thread and optionally associate it with a (new) process.
1750  * NOTE! the new thread's cpu may not equal the current cpu.
1751  */
1752 void
1753 pmap_init_thread(thread_t td)
1754 {
1755         /* enforce pcb placement & alignment */
1756         td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1757         td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1758         td->td_savefpu = &td->td_pcb->pcb_save;
1759         td->td_sp = (char *)td->td_pcb; /* no -16 */
1760 }
1761
1762 /*
1763  * This routine directly affects the fork perf for a process.
1764  */
1765 void
1766 pmap_init_proc(struct proc *p)
1767 {
1768 }
1769
1770 static void
1771 pmap_pinit_defaults(struct pmap *pmap)
1772 {
1773         bcopy(pmap_bits_default, pmap->pmap_bits,
1774               sizeof(pmap_bits_default));
1775         bcopy(protection_codes, pmap->protection_codes,
1776               sizeof(protection_codes));
1777         bcopy(pat_pte_index, pmap->pmap_cache_bits,
1778               sizeof(pat_pte_index));
1779         pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1780         pmap->copyinstr = std_copyinstr;
1781         pmap->copyin = std_copyin;
1782         pmap->copyout = std_copyout;
1783         pmap->fubyte = std_fubyte;
1784         pmap->subyte = std_subyte;
1785         pmap->fuword32 = std_fuword32;
1786         pmap->fuword64 = std_fuword64;
1787         pmap->suword32 = std_suword32;
1788         pmap->suword64 = std_suword64;
1789         pmap->swapu32 = std_swapu32;
1790         pmap->swapu64 = std_swapu64;
1791 }
1792 /*
1793  * Initialize pmap0/vmspace0.
1794  *
1795  * On architectures where the kernel pmap is not integrated into the user
1796  * process pmap, this pmap represents the process pmap, not the kernel pmap.
1797  * kernel_pmap should be used to directly access the kernel_pmap.
1798  */
1799 void
1800 pmap_pinit0(struct pmap *pmap)
1801 {
1802         int i;
1803
1804         pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1805         pmap->pm_count = 1;
1806         CPUMASK_ASSZERO(pmap->pm_active);
1807         pmap->pm_pvhint = NULL;
1808         RB_INIT(&pmap->pm_pvroot);
1809         spin_init(&pmap->pm_spin, "pmapinit0");
1810         for (i = 0; i < PM_PLACEMARKS; ++i)
1811                 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1812         bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1813         pmap_pinit_defaults(pmap);
1814 }
1815
1816 /*
1817  * Initialize a preallocated and zeroed pmap structure,
1818  * such as one in a vmspace structure.
1819  */
1820 static void
1821 pmap_pinit_simple(struct pmap *pmap)
1822 {
1823         int i;
1824
1825         /*
1826          * Misc initialization
1827          */
1828         pmap->pm_count = 1;
1829         CPUMASK_ASSZERO(pmap->pm_active);
1830         pmap->pm_pvhint = NULL;
1831         pmap->pm_flags = PMAP_FLAG_SIMPLE;
1832
1833         pmap_pinit_defaults(pmap);
1834
1835         /*
1836          * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1837          * for this).
1838          */
1839         if (pmap->pm_pmlpv == NULL) {
1840                 RB_INIT(&pmap->pm_pvroot);
1841                 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1842                 spin_init(&pmap->pm_spin, "pmapinitsimple");
1843                 for (i = 0; i < PM_PLACEMARKS; ++i)
1844                         pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1845         }
1846 }
1847
1848 void
1849 pmap_pinit(struct pmap *pmap)
1850 {
1851         pv_entry_t pv;
1852         int j;
1853
1854         if (pmap->pm_pmlpv) {
1855                 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1856                         pmap_puninit(pmap);
1857                 }
1858         }
1859
1860         pmap_pinit_simple(pmap);
1861         pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1862
1863         /*
1864          * No need to allocate page table space yet but we do need a valid
1865          * page directory table.
1866          */
1867         if (pmap->pm_pml4 == NULL) {
1868                 pmap->pm_pml4 =
1869                     (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1870                                                         PAGE_SIZE,
1871                                                         VM_SUBSYS_PML4);
1872         }
1873
1874         /*
1875          * Allocate the page directory page, which wires it even though
1876          * it isn't being entered into some higher level page table (it
1877          * being the highest level).  If one is already cached we don't
1878          * have to do anything.
1879          */
1880         if ((pv = pmap->pm_pmlpv) == NULL) {
1881                 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1882                 pmap->pm_pmlpv = pv;
1883                 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1884                             VM_PAGE_TO_PHYS(pv->pv_m));
1885                 pv_put(pv);
1886
1887                 /*
1888                  * Install DMAP and KMAP.
1889                  */
1890                 for (j = 0; j < NDMPML4E; ++j) {
1891                         pmap->pm_pml4[DMPML4I + j] =
1892                             (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1893                             pmap->pmap_bits[PG_RW_IDX] |
1894                             pmap->pmap_bits[PG_V_IDX] |
1895                             pmap->pmap_bits[PG_U_IDX];
1896                 }
1897                 pmap->pm_pml4[KPML4I] = KPDPphys |
1898                     pmap->pmap_bits[PG_RW_IDX] |
1899                     pmap->pmap_bits[PG_V_IDX] |
1900                     pmap->pmap_bits[PG_U_IDX];
1901
1902                 /*
1903                  * install self-referential address mapping entry
1904                  */
1905                 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1906                     pmap->pmap_bits[PG_V_IDX] |
1907                     pmap->pmap_bits[PG_RW_IDX] |
1908                     pmap->pmap_bits[PG_A_IDX] |
1909                     pmap->pmap_bits[PG_M_IDX];
1910         } else {
1911                 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1912                 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1913         }
1914         KKASSERT(pmap->pm_pml4[255] == 0);
1915         KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1916         KKASSERT(pv->pv_entry.rbe_left == NULL);
1917         KKASSERT(pv->pv_entry.rbe_right == NULL);
1918 }
1919
1920 /*
1921  * Clean up a pmap structure so it can be physically freed.  This routine
1922  * is called by the vmspace dtor function.  A great deal of pmap data is
1923  * left passively mapped to improve vmspace management so we have a bit
1924  * of cleanup work to do here.
1925  */
1926 void
1927 pmap_puninit(pmap_t pmap)
1928 {
1929         pv_entry_t pv;
1930         vm_page_t p;
1931
1932         KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1933         if ((pv = pmap->pm_pmlpv) != NULL) {
1934                 if (pv_hold_try(pv) == 0)
1935                         pv_lock(pv);
1936                 KKASSERT(pv == pmap->pm_pmlpv);
1937                 p = pmap_remove_pv_page(pv);
1938                 pv_free(pv, NULL);
1939                 pv = NULL;      /* safety */
1940                 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1941                 vm_page_busy_wait(p, FALSE, "pgpun");
1942                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1943                 vm_page_unwire(p, 0);
1944                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1945
1946                 /*
1947                  * XXX eventually clean out PML4 static entries and
1948                  * use vm_page_free_zero()
1949                  */
1950                 vm_page_free(p);
1951                 pmap->pm_pmlpv = NULL;
1952         }
1953         if (pmap->pm_pml4) {
1954                 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1955                 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1956                 pmap->pm_pml4 = NULL;
1957         }
1958         KKASSERT(pmap->pm_stats.resident_count == 0);
1959         KKASSERT(pmap->pm_stats.wired_count == 0);
1960 }
1961
1962 /*
1963  * This function is now unused (used to add the pmap to the pmap_list)
1964  */
1965 void
1966 pmap_pinit2(struct pmap *pmap)
1967 {
1968 }
1969
1970 /*
1971  * This routine is called when various levels in the page table need to
1972  * be populated.  This routine cannot fail.
1973  *
1974  * This function returns two locked pv_entry's, one representing the
1975  * requested pv and one representing the requested pv's parent pv.  If
1976  * an intermediate page table does not exist it will be created, mapped,
1977  * wired, and the parent page table will be given an additional hold
1978  * count representing the presence of the child pv_entry.
1979  */
1980 static
1981 pv_entry_t
1982 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1983 {
1984         pt_entry_t *ptep;
1985         pv_entry_t pv;
1986         pv_entry_t pvp;
1987         pt_entry_t v;
1988         vm_pindex_t pt_pindex;
1989         vm_page_t m;
1990         int isnew;
1991         int ispt;
1992
1993         /*
1994          * If the pv already exists and we aren't being asked for the
1995          * parent page table page we can just return it.  A locked+held pv
1996          * is returned.  The pv will also have a second hold related to the
1997          * pmap association that we don't have to worry about.
1998          */
1999         ispt = 0;
2000         pv = pv_alloc(pmap, ptepindex, &isnew);
2001         if (isnew == 0 && pvpp == NULL)
2002                 return(pv);
2003
2004         /*
2005          * Special case terminal PVs.  These are not page table pages so
2006          * no vm_page is allocated (the caller supplied the vm_page).  If
2007          * pvpp is non-NULL we are being asked to also removed the pt_pv
2008          * for this pv.
2009          *
2010          * Note that pt_pv's are only returned for user VAs. We assert that
2011          * a pt_pv is not being requested for kernel VAs.  The kernel
2012          * pre-wires all higher-level page tables so don't overload managed
2013          * higher-level page tables on top of it!
2014          */
2015         if (ptepindex < pmap_pt_pindex(0)) {
2016                 if (ptepindex >= NUPTE_USER) {
2017                         /* kernel manages this manually for KVM */
2018                         KKASSERT(pvpp == NULL);
2019                 } else {
2020                         KKASSERT(pvpp != NULL);
2021                         pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
2022                         pvp = pmap_allocpte(pmap, pt_pindex, NULL);
2023                         if (isnew)
2024                                 vm_page_wire_quick(pvp->pv_m);
2025                         *pvpp = pvp;
2026                 }
2027                 return(pv);
2028         }
2029
2030         /*
2031          * The kernel never uses managed PT/PD/PDP pages.
2032          */
2033         KKASSERT(pmap != &kernel_pmap);
2034
2035         /*
2036          * Non-terminal PVs allocate a VM page to represent the page table,
2037          * so we have to resolve pvp and calculate ptepindex for the pvp
2038          * and then for the page table entry index in the pvp for
2039          * fall-through.
2040          */
2041         if (ptepindex < pmap_pd_pindex(0)) {
2042                 /*
2043                  * pv is PT, pvp is PD
2044                  */
2045                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
2046                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
2047                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2048
2049                 /*
2050                  * PT index in PD
2051                  */
2052                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
2053                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
2054                 ispt = 1;
2055         } else if (ptepindex < pmap_pdp_pindex(0)) {
2056                 /*
2057                  * pv is PD, pvp is PDP
2058                  *
2059                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2060                  *                   the PD.
2061                  */
2062                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
2063                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2064
2065                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
2066                         KKASSERT(pvpp == NULL);
2067                         pvp = NULL;
2068                 } else {
2069                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
2070                 }
2071
2072                 /*
2073                  * PD index in PDP
2074                  */
2075                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
2076                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
2077         } else if (ptepindex < pmap_pml4_pindex()) {
2078                 /*
2079                  * pv is PDP, pvp is the root pml4 table
2080                  */
2081                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2082
2083                 /*
2084                  * PDP index in PML4
2085                  */
2086                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2087                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2088         } else {
2089                 /*
2090                  * pv represents the top-level PML4, there is no parent.
2091                  */
2092                 pvp = NULL;
2093         }
2094
2095         if (isnew == 0)
2096                 goto notnew;
2097
2098         /*
2099          * (isnew) is TRUE, pv is not terminal.
2100          *
2101          * (1) Add a wire count to the parent page table (pvp).
2102          * (2) Allocate a VM page for the page table.
2103          * (3) Enter the VM page into the parent page table.
2104          *
2105          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2106          */
2107         if (pvp)
2108                 vm_page_wire_quick(pvp->pv_m);
2109
2110         for (;;) {
2111                 m = vm_page_alloc(NULL, pv->pv_pindex,
2112                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2113                                   VM_ALLOC_INTERRUPT);
2114                 if (m)
2115                         break;
2116                 vm_wait(0);
2117         }
2118         vm_page_wire(m);        /* wire for mapping in parent */
2119         vm_page_unmanage(m);    /* m must be spinunlocked */
2120         pmap_zero_page(VM_PAGE_TO_PHYS(m));
2121         m->valid = VM_PAGE_BITS_ALL;
2122
2123         vm_page_spin_lock(m);
2124         pmap_page_stats_adding(m);
2125         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2126         pv->pv_m = m;
2127         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2128         vm_page_spin_unlock(m);
2129
2130         /*
2131          * (isnew) is TRUE, pv is not terminal.
2132          *
2133          * Wire the page into pvp.  Bump the resident_count for the pmap.
2134          * There is no pvp for the top level, address the pm_pml4[] array
2135          * directly.
2136          *
2137          * If the caller wants the parent we return it, otherwise
2138          * we just put it away.
2139          *
2140          * No interlock is needed for pte 0 -> non-zero.
2141          *
2142          * In the situation where *ptep is valid we might have an unmanaged
2143          * page table page shared from another page table which we need to
2144          * unshare before installing our private page table page.
2145          */
2146         if (pvp) {
2147                 v = VM_PAGE_TO_PHYS(m) |
2148                     (pmap->pmap_bits[PG_U_IDX] |
2149                      pmap->pmap_bits[PG_RW_IDX] |
2150                      pmap->pmap_bits[PG_V_IDX] |
2151                      pmap->pmap_bits[PG_A_IDX] |
2152                      pmap->pmap_bits[PG_M_IDX]);
2153                 ptep = pv_pte_lookup(pvp, ptepindex);
2154                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2155                         pt_entry_t pte;
2156
2157                         if (ispt == 0) {
2158                                 panic("pmap_allocpte: unexpected pte %p/%d",
2159                                       pvp, (int)ptepindex);
2160                         }
2161                         pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, v);
2162                         if (vm_page_unwire_quick(
2163                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2164                                 panic("pmap_allocpte: shared pgtable "
2165                                       "pg bad wirecount");
2166                         }
2167                 } else {
2168                         pt_entry_t pte;
2169
2170                         pte = atomic_swap_long(ptep, v);
2171                         if (pte != 0) {
2172                                 kprintf("install pgtbl mixup 0x%016jx "
2173                                         "old/new 0x%016jx/0x%016jx\n",
2174                                         (intmax_t)ptepindex, pte, v);
2175                         }
2176                 }
2177         }
2178         vm_page_wakeup(m);
2179
2180         /*
2181          * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2182          */
2183 notnew:
2184         if (pvp) {
2185                 KKASSERT(pvp->pv_m != NULL);
2186                 ptep = pv_pte_lookup(pvp, ptepindex);
2187                 v = VM_PAGE_TO_PHYS(pv->pv_m) |
2188                     (pmap->pmap_bits[PG_U_IDX] |
2189                      pmap->pmap_bits[PG_RW_IDX] |
2190                      pmap->pmap_bits[PG_V_IDX] |
2191                      pmap->pmap_bits[PG_A_IDX] |
2192                      pmap->pmap_bits[PG_M_IDX]);
2193                 if (*ptep != v) {
2194                         kprintf("mismatched upper level pt %016jx/%016jx\n",
2195                                 *ptep, v);
2196                 }
2197         }
2198         if (pvpp)
2199                 *pvpp = pvp;
2200         else if (pvp)
2201                 pv_put(pvp);
2202         return (pv);
2203 }
2204
2205 /*
2206  * This version of pmap_allocpte() checks for possible segment optimizations
2207  * that would allow page-table sharing.  It can be called for terminal
2208  * page or page table page ptepindex's.
2209  *
2210  * The function is called with page table page ptepindex's for fictitious
2211  * and unmanaged terminal pages.  That is, we don't want to allocate a
2212  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
2213  * for this case.
2214  *
2215  * This function can return a pv and *pvpp associated with the passed in pmap
2216  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
2217  * an unmanaged page table page will be entered into the pass in pmap.
2218  */
2219 static
2220 pv_entry_t
2221 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2222                   vm_map_entry_t entry, vm_offset_t va)
2223 {
2224         vm_object_t object;
2225         pmap_t obpmap;
2226         pmap_t *obpmapp;
2227         vm_offset_t b;
2228         pv_entry_t pte_pv;      /* in original or shared pmap */
2229         pv_entry_t pt_pv;       /* in original or shared pmap */
2230         pv_entry_t proc_pd_pv;  /* in original pmap */
2231         pv_entry_t proc_pt_pv;  /* in original pmap */
2232         pv_entry_t xpv;         /* PT in shared pmap */
2233         pd_entry_t *pt;         /* PT entry in PD of original pmap */
2234         pd_entry_t opte;        /* contents of *pt */
2235         pd_entry_t npte;        /* contents of *pt */
2236         vm_page_t m;
2237
2238         /*
2239          * Basic tests, require a non-NULL vm_map_entry, require proper
2240          * alignment and type for the vm_map_entry, require that the
2241          * underlying object already be allocated.
2242          *
2243          * We allow almost any type of object to use this optimization.
2244          * The object itself does NOT have to be sized to a multiple of the
2245          * segment size, but the memory mapping does.
2246          *
2247          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2248          *     won't work as expected.
2249          */
2250         if (entry == NULL ||
2251             pmap_mmu_optimize == 0 ||                   /* not enabled */
2252             (pmap->pm_flags & PMAP_HVM) ||              /* special pmap */
2253             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal or pt */
2254             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2255             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2256             entry->object.vm_object == NULL ||          /* needs VM object */
2257             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2258             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2259             (entry->offset & SEG_MASK) ||               /* must be aligned */
2260             (entry->start & SEG_MASK)) {
2261                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2262         }
2263
2264         /*
2265          * Make sure the full segment can be represented.
2266          */
2267         b = va & ~(vm_offset_t)SEG_MASK;
2268         if (b < entry->start || b + SEG_SIZE > entry->end)
2269                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2270
2271         /*
2272          * If the full segment can be represented dive the VM object's
2273          * shared pmap, allocating as required.
2274          */
2275         object = entry->object.vm_object;
2276
2277         if (entry->protection & VM_PROT_WRITE)
2278                 obpmapp = &object->md.pmap_rw;
2279         else
2280                 obpmapp = &object->md.pmap_ro;
2281
2282 #ifdef PMAP_DEBUG2
2283         if (pmap_enter_debug > 0) {
2284                 --pmap_enter_debug;
2285                 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2286                         "obpmapp %p %p\n",
2287                         va, entry->protection, object,
2288                         obpmapp, *obpmapp);
2289                 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2290                         entry, entry->start, entry->end);
2291         }
2292 #endif
2293
2294         /*
2295          * We allocate what appears to be a normal pmap but because portions
2296          * of this pmap are shared with other unrelated pmaps we have to
2297          * set pm_active to point to all cpus.
2298          *
2299          * XXX Currently using pmap_spin to interlock the update, can't use
2300          *     vm_object_hold/drop because the token might already be held
2301          *     shared OR exclusive and we don't know.
2302          */
2303         while ((obpmap = *obpmapp) == NULL) {
2304                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2305                 pmap_pinit_simple(obpmap);
2306                 pmap_pinit2(obpmap);
2307                 spin_lock(&pmap_spin);
2308                 if (*obpmapp != NULL) {
2309                         /*
2310                          * Handle race
2311                          */
2312                         spin_unlock(&pmap_spin);
2313                         pmap_release(obpmap);
2314                         pmap_puninit(obpmap);
2315                         kfree(obpmap, M_OBJPMAP);
2316                         obpmap = *obpmapp; /* safety */
2317                 } else {
2318                         obpmap->pm_active = smp_active_mask;
2319                         obpmap->pm_flags |= PMAP_SEGSHARED;
2320                         *obpmapp = obpmap;
2321                         spin_unlock(&pmap_spin);
2322                 }
2323         }
2324
2325         /*
2326          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2327          * pte/pt using the shared pmap from the object but also adjust
2328          * the process pmap's page table page as a side effect.
2329          */
2330
2331         /*
2332          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2333          * we will return.  This is true if ptepindex represents a terminal
2334          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2335          * the PD.
2336          */
2337         pt_pv = NULL;
2338         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2339 retry:
2340         if (ptepindex >= pmap_pt_pindex(0))
2341                 xpv = pte_pv;
2342         else
2343                 xpv = pt_pv;
2344
2345         /*
2346          * Resolve the PD in the process pmap so we can properly share the
2347          * page table page.  Lock order is bottom-up (leaf first)!
2348          *
2349          * NOTE: proc_pt_pv can be NULL.
2350          */
2351         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b), NULL);
2352         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2353 #ifdef PMAP_DEBUG2
2354         if (pmap_enter_debug > 0) {
2355                 --pmap_enter_debug;
2356                 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2357                         proc_pt_pv,
2358                         (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2359                         proc_pd_pv,
2360                         va);
2361         }
2362 #endif
2363
2364         /*
2365          * xpv is the page table page pv from the shared object
2366          * (for convenience), from above.
2367          *
2368          * Calculate the pte value for the PT to load into the process PD.
2369          * If we have to change it we must properly dispose of the previous
2370          * entry.
2371          */
2372         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2373         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2374                (pmap->pmap_bits[PG_U_IDX] |
2375                 pmap->pmap_bits[PG_RW_IDX] |
2376                 pmap->pmap_bits[PG_V_IDX] |
2377                 pmap->pmap_bits[PG_A_IDX] |
2378                 pmap->pmap_bits[PG_M_IDX]);
2379
2380         /*
2381          * Dispose of previous page table page if it was local to the
2382          * process pmap.  If the old pt is not empty we cannot dispose of it
2383          * until we clean it out.  This case should not arise very often so
2384          * it is not optimized.
2385          *
2386          * Leave pt_pv and pte_pv (in our object pmap) locked and intact
2387          * for the retry.
2388          */
2389         if (proc_pt_pv) {
2390                 pmap_inval_bulk_t bulk;
2391
2392                 if (proc_pt_pv->pv_m->wire_count != 1) {
2393                         pv_put(proc_pd_pv);
2394                         pv_put(proc_pt_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_put(pv);
2562                 info->retry = 1;
2563                 spin_lock(&pmap->pm_spin);
2564
2565                 return -1;
2566         }
2567         KKASSERT(pv->pv_pmap == pmap && pindex == pv->pv_pindex);
2568
2569         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2570                 /*
2571                  * I am PTE, parent is PT
2572                  */
2573                 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2574                 pindex += NUPTE_TOTAL;
2575         } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2576                 /*
2577                  * I am PT, parent is PD
2578                  */
2579                 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2580                 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2581         } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2582                 /*
2583                  * I am PD, parent is PDP
2584                  */
2585                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2586                          NPDPEPGSHIFT;
2587                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2588         } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2589                 /*
2590                  * I am PDP, parent is PML4 (there's only one)
2591                  */
2592 #if 0
2593                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2594                            NUPD_TOTAL) >> NPML4EPGSHIFT;
2595                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2596 #endif
2597                 pindex = pmap_pml4_pindex();
2598         } else {
2599                 /*
2600                  * parent is NULL
2601                  */
2602                 if (info->pvp) {
2603                         pv_put(info->pvp);
2604                         info->pvp = NULL;
2605                 }
2606                 pindex = 0;
2607         }
2608         if (pindex) {
2609                 if (info->pvp && info->pvp->pv_pindex != pindex) {
2610                         pv_put(info->pvp);
2611                         info->pvp = NULL;
2612                 }
2613                 if (info->pvp == NULL)
2614                         info->pvp = pv_get(pmap, pindex, NULL);
2615         } else {
2616                 if (info->pvp) {
2617                         pv_put(info->pvp);
2618                         info->pvp = NULL;
2619                 }
2620         }
2621         r = pmap_release_pv(pv, info->pvp, NULL);
2622         spin_lock(&pmap->pm_spin);
2623
2624         return(r);
2625 }
2626
2627 /*
2628  * Called with held (i.e. also locked) pv.  This function will dispose of
2629  * the lock along with the pv.
2630  *
2631  * If the caller already holds the locked parent page table for pv it
2632  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2633  * pass NULL for pvp.
2634  */
2635 static int
2636 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2637 {
2638         vm_page_t p;
2639
2640         /*
2641          * The pmap is currently not spinlocked, pv is held+locked.
2642          * Remove the pv's page from its parent's page table.  The
2643          * parent's page table page's wire_count will be decremented.
2644          *
2645          * This will clean out the pte at any level of the page table.
2646          * If smp != 0 all cpus are affected.
2647          *
2648          * Do not tear-down recursively, its faster to just let the
2649          * release run its course.
2650          */
2651         pmap_remove_pv_pte(pv, pvp, bulk, 0);
2652
2653         /*
2654          * Terminal pvs are unhooked from their vm_pages.  Because
2655          * terminal pages aren't page table pages they aren't wired
2656          * by us, so we have to be sure not to unwire them either.
2657          */
2658         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2659                 pmap_remove_pv_page(pv);
2660                 goto skip;
2661         }
2662
2663         /*
2664          * We leave the top-level page table page cached, wired, and
2665          * mapped in the pmap until the dtor function (pmap_puninit())
2666          * gets called.
2667          *
2668          * Since we are leaving the top-level pv intact we need
2669          * to break out of what would otherwise be an infinite loop.
2670          */
2671         if (pv->pv_pindex == pmap_pml4_pindex()) {
2672                 pv_put(pv);
2673                 return(-1);
2674         }
2675
2676         /*
2677          * For page table pages (other than the top-level page),
2678          * remove and free the vm_page.  The representitive mapping
2679          * removed above by pmap_remove_pv_pte() did not undo the
2680          * last wire_count so we have to do that as well.
2681          */
2682         p = pmap_remove_pv_page(pv);
2683         vm_page_busy_wait(p, FALSE, "pmaprl");
2684         if (p->wire_count != 1) {
2685                 kprintf("p->wire_count was %016lx %d\n",
2686                         pv->pv_pindex, p->wire_count);
2687         }
2688         KKASSERT(p->wire_count == 1);
2689         KKASSERT(p->flags & PG_UNMANAGED);
2690
2691         vm_page_unwire(p, 0);
2692         KKASSERT(p->wire_count == 0);
2693
2694         vm_page_free(p);
2695 skip:
2696         pv_free(pv, pvp);
2697
2698         return 0;
2699 }
2700
2701 /*
2702  * This function will remove the pte associated with a pv from its parent.
2703  * Terminal pv's are supported.  All cpus specified by (bulk) are properly
2704  * invalidated.
2705  *
2706  * The wire count will be dropped on the parent page table.  The wire
2707  * count on the page being removed (pv->pv_m) from the parent page table
2708  * is NOT touched.  Note that terminal pages will not have any additional
2709  * wire counts while page table pages will have at least one representing
2710  * the mapping, plus others representing sub-mappings.
2711  *
2712  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2713  *       pages and user page table and terminal pages.
2714  *
2715  * NOTE: The pte being removed might be unmanaged, and the pv supplied might
2716  *       be freshly allocated and not imply that the pte is managed.  In this
2717  *       case pv->pv_m should be NULL.
2718  *
2719  * The pv must be locked.  The pvp, if supplied, must be locked.  All
2720  * supplied pv's will remain locked on return.
2721  *
2722  * XXX must lock parent pv's if they exist to remove pte XXX
2723  */
2724 static
2725 void
2726 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2727                    int destroy)
2728 {
2729         vm_pindex_t ptepindex = pv->pv_pindex;
2730         pmap_t pmap = pv->pv_pmap;
2731         vm_page_t p;
2732         int gotpvp = 0;
2733
2734         KKASSERT(pmap);
2735
2736         if (ptepindex == pmap_pml4_pindex()) {
2737                 /*
2738                  * We are the top level PML4E table, there is no parent.
2739                  */
2740                 p = pmap->pm_pmlpv->pv_m;
2741                 KKASSERT(pv->pv_m == p);        /* debugging */
2742         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2743                 /*
2744                  * Remove a PDP page from the PML4E.  This can only occur
2745                  * with user page tables.  We do not have to lock the
2746                  * pml4 PV so just ignore pvp.
2747                  */
2748                 vm_pindex_t pml4_pindex;
2749                 vm_pindex_t pdp_index;
2750                 pml4_entry_t *pdp;
2751
2752                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2753                 if (pvp == NULL) {
2754                         pml4_pindex = pmap_pml4_pindex();
2755                         pvp = pv_get(pv->pv_pmap, pml4_pindex, NULL);
2756                         KKASSERT(pvp);
2757                         gotpvp = 1;
2758                 }
2759
2760                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2761                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2762                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2763                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2764                 KKASSERT(pv->pv_m == p);        /* debugging */
2765         } else if (ptepindex >= pmap_pd_pindex(0)) {
2766                 /*
2767                  * Remove a PD page from the PDP
2768                  *
2769                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2770                  *                   of a simple pmap because it stops at
2771                  *                   the PD page.
2772                  */
2773                 vm_pindex_t pdp_pindex;
2774                 vm_pindex_t pd_index;
2775                 pdp_entry_t *pd;
2776
2777                 pd_index = ptepindex - pmap_pd_pindex(0);
2778
2779                 if (pvp == NULL) {
2780                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2781                                      (pd_index >> NPML4EPGSHIFT);
2782                         pvp = pv_get(pv->pv_pmap, pdp_pindex, NULL);
2783                         gotpvp = 1;
2784                 }
2785
2786                 if (pvp) {
2787                         pd = pv_pte_lookup(pvp, pd_index &
2788                                                 ((1ul << NPDPEPGSHIFT) - 1));
2789                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2790                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2791                         pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2792                 } else {
2793                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2794                         p = pv->pv_m;           /* degenerate test later */
2795                 }
2796                 KKASSERT(pv->pv_m == p);        /* debugging */
2797         } else if (ptepindex >= pmap_pt_pindex(0)) {
2798                 /*
2799                  *  Remove a PT page from the PD
2800                  */
2801                 vm_pindex_t pd_pindex;
2802                 vm_pindex_t pt_index;
2803                 pd_entry_t *pt;
2804
2805                 pt_index = ptepindex - pmap_pt_pindex(0);
2806
2807                 if (pvp == NULL) {
2808                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2809                                     (pt_index >> NPDPEPGSHIFT);
2810                         pvp = pv_get(pv->pv_pmap, pd_pindex, NULL);
2811                         KKASSERT(pvp);
2812                         gotpvp = 1;
2813                 }
2814
2815                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2816 #if 0
2817                 KASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0,
2818                         ("*pt unexpectedly invalid %016jx "
2819                          "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
2820                         *pt, gotpvp, ptepindex, pt_index, pv, pvp));
2821                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2822 #else
2823                 if ((*pt & pmap->pmap_bits[PG_V_IDX]) == 0) {
2824                         kprintf("*pt unexpectedly invalid %016jx "
2825                                 "gotpvp=%d ptepindex=%ld ptindex=%ld "
2826                                 "pv=%p pvp=%p\n",
2827                                 *pt, gotpvp, ptepindex, pt_index, pv, pvp);
2828                         tsleep(pt, 0, "DEAD", 0);
2829                         p = pv->pv_m;
2830                 } else {
2831                         p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2832                 }
2833 #endif
2834                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2835                 KKASSERT(pv->pv_m == p);        /* debugging */
2836         } else {
2837                 /*
2838                  * Remove a PTE from the PT page.  The PV might exist even if
2839                  * the PTE is not managed, in whichcase pv->pv_m should be
2840                  * NULL.
2841                  *
2842                  * NOTE: Userland pmaps manage the parent PT/PD/PDP page
2843                  *       table pages but the kernel_pmap does not.
2844                  *
2845                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2846                  *       pv is a pte_pv so we can safely lock pt_pv.
2847                  *
2848                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2849                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2850                  *       terminal ptes.
2851                  */
2852                 vm_pindex_t pt_pindex;
2853                 pt_entry_t *ptep;
2854                 pt_entry_t pte;
2855                 vm_offset_t va;
2856
2857                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2858                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2859
2860                 if (ptepindex >= NUPTE_USER) {
2861                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2862                         KKASSERT(pvp == NULL);
2863                         /* pvp remains NULL */
2864                 } else {
2865                         if (pvp == NULL) {
2866                                 pt_pindex = NUPTE_TOTAL +
2867                                             (ptepindex >> NPDPEPGSHIFT);
2868                                 pvp = pv_get(pv->pv_pmap, pt_pindex, NULL);
2869                                 KKASSERT(pvp);
2870                                 gotpvp = 1;
2871                         }
2872                         ptep = pv_pte_lookup(pvp, ptepindex &
2873                                                   ((1ul << NPDPEPGSHIFT) - 1));
2874                 }
2875                 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2876                 if (bulk == NULL)               /* XXX */
2877                         cpu_invlpg((void *)va); /* XXX */
2878
2879                 /*
2880                  * Now update the vm_page_t
2881                  */
2882                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
2883                     (pte & pmap->pmap_bits[PG_V_IDX])) {
2884                         /*
2885                          * Valid managed page, adjust (p).
2886                          */
2887                         if (pte & pmap->pmap_bits[PG_DEVICE_IDX]) {
2888                                 p = pv->pv_m;
2889                         } else {
2890                                 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2891                                 KKASSERT(pv->pv_m == p);
2892                         }
2893                         if (pte & pmap->pmap_bits[PG_M_IDX]) {
2894                                 if (pmap_track_modified(ptepindex))
2895                                         vm_page_dirty(p);
2896                         }
2897                         if (pte & pmap->pmap_bits[PG_A_IDX]) {
2898                                 vm_page_flag_set(p, PG_REFERENCED);
2899                         }
2900                 } else {
2901                         /*
2902                          * Unmanaged page, do not try to adjust the vm_page_t.
2903                          * pv could be freshly allocated for a pmap_enter(),
2904                          * replacing an unmanaged page with a managed one.
2905                          *
2906                          * pv->pv_m might reflect the new page and not the
2907                          * existing page.
2908                          *
2909                          * We could extract p from the physical address and
2910                          * adjust it but we explicitly do not for unmanaged
2911                          * pages.
2912                          */
2913                         p = NULL;
2914                 }
2915                 if (pte & pmap->pmap_bits[PG_W_IDX])
2916                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
2917                 if (pte & pmap->pmap_bits[PG_G_IDX])
2918                         cpu_invlpg((void *)va);
2919         }
2920
2921         /*
2922          * If requested, scrap the underlying pv->pv_m and the underlying
2923          * pv.  If this is a page-table-page we must also free the page.
2924          *
2925          * pvp must be returned locked.
2926          */
2927         if (destroy == 1) {
2928                 /*
2929                  * page table page (PT, PD, PDP, PML4), caller was responsible
2930                  * for testing wired_count.
2931                  */
2932                 KKASSERT(pv->pv_m->wire_count == 1);
2933                 p = pmap_remove_pv_page(pv);
2934                 pv_free(pv, pvp);
2935                 pv = NULL;
2936
2937                 vm_page_busy_wait(p, FALSE, "pgpun");
2938                 vm_page_unwire(p, 0);
2939                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2940                 vm_page_free(p);
2941         } else if (destroy == 2) {
2942                 /*
2943                  * Normal page, remove from pmap and leave the underlying
2944                  * page untouched.
2945                  */
2946                 pmap_remove_pv_page(pv);
2947                 pv_free(pv, pvp);
2948                 pv = NULL;              /* safety */
2949         }
2950
2951         /*
2952          * If we acquired pvp ourselves then we are responsible for
2953          * recursively deleting it.
2954          */
2955         if (pvp && gotpvp) {
2956                 /*
2957                  * Recursively destroy higher-level page tables.
2958                  *
2959                  * This is optional.  If we do not, they will still
2960                  * be destroyed when the process exits.
2961                  *
2962                  * NOTE: Do not destroy pv_entry's with extra hold refs,
2963                  *       a caller may have unlocked it and intends to
2964                  *       continue to use it.
2965                  */
2966                 if (pmap_dynamic_delete &&
2967                     pvp->pv_m &&
2968                     pvp->pv_m->wire_count == 1 &&
2969                     (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
2970                     pvp->pv_pindex != pmap_pml4_pindex()) {
2971                         if (pmap_dynamic_delete == 2)
2972                                 kprintf("A %jd %08x\n", pvp->pv_pindex, pvp->pv_hold);
2973                         if (pmap != &kernel_pmap) {
2974                                 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
2975                                 pvp = NULL;     /* safety */
2976                         } else {
2977                                 kprintf("Attempt to remove kernel_pmap pindex "
2978                                         "%jd\n", pvp->pv_pindex);
2979                                 pv_put(pvp);
2980                         }
2981                 } else {
2982                         pv_put(pvp);
2983                 }
2984         }
2985 }
2986
2987 /*
2988  * Remove the vm_page association to a pv.  The pv must be locked.
2989  */
2990 static
2991 vm_page_t
2992 pmap_remove_pv_page(pv_entry_t pv)
2993 {
2994         vm_page_t m;
2995
2996         m = pv->pv_m;
2997         vm_page_spin_lock(m);
2998         KKASSERT(m && m == pv->pv_m);
2999         pv->pv_m = NULL;
3000         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3001         pmap_page_stats_deleting(m);
3002         if (TAILQ_EMPTY(&m->md.pv_list))
3003                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3004         vm_page_spin_unlock(m);
3005
3006         return(m);
3007 }
3008
3009 /*
3010  * Grow the number of kernel page table entries, if needed.
3011  *
3012  * This routine is always called to validate any address space
3013  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
3014  * space below KERNBASE.
3015  *
3016  * kernel_map must be locked exclusively by the caller.
3017  */
3018 void
3019 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
3020 {
3021         vm_paddr_t paddr;
3022         vm_offset_t ptppaddr;
3023         vm_page_t nkpg;
3024         pd_entry_t *pt, newpt;
3025         pdp_entry_t newpd;
3026         int update_kernel_vm_end;
3027
3028         /*
3029          * bootstrap kernel_vm_end on first real VM use
3030          */
3031         if (kernel_vm_end == 0) {
3032                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
3033                 nkpt = 0;
3034                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3035                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
3036                                         ~(PAGE_SIZE * NPTEPG - 1);
3037                         nkpt++;
3038                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
3039                                 kernel_vm_end = kernel_map.max_offset;
3040                                 break;
3041                         }
3042                 }
3043         }
3044
3045         /*
3046          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
3047          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
3048          * do not want to force-fill 128G worth of page tables.
3049          */
3050         if (kstart < KERNBASE) {
3051                 if (kstart > kernel_vm_end)
3052                         kstart = kernel_vm_end;
3053                 KKASSERT(kend <= KERNBASE);
3054                 update_kernel_vm_end = 1;
3055         } else {
3056                 update_kernel_vm_end = 0;
3057         }
3058
3059         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
3060         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
3061
3062         if (kend - 1 >= kernel_map.max_offset)
3063                 kend = kernel_map.max_offset;
3064
3065         while (kstart < kend) {
3066                 pt = pmap_pt(&kernel_pmap, kstart);
3067                 if (pt == NULL) {
3068                         /* We need a new PD entry */
3069                         nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3070                                              VM_ALLOC_NORMAL |
3071                                              VM_ALLOC_SYSTEM |
3072                                              VM_ALLOC_INTERRUPT);
3073                         if (nkpg == NULL) {
3074                                 panic("pmap_growkernel: no memory to grow "
3075                                       "kernel");
3076                         }
3077                         paddr = VM_PAGE_TO_PHYS(nkpg);
3078                         pmap_zero_page(paddr);
3079                         newpd = (pdp_entry_t)
3080                             (paddr |
3081                             kernel_pmap.pmap_bits[PG_V_IDX] |
3082                             kernel_pmap.pmap_bits[PG_RW_IDX] |
3083                             kernel_pmap.pmap_bits[PG_A_IDX] |
3084                             kernel_pmap.pmap_bits[PG_M_IDX]);
3085                         *pmap_pd(&kernel_pmap, kstart) = newpd;
3086                         continue; /* try again */
3087                 }
3088                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3089                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
3090                                  ~(PAGE_SIZE * NPTEPG - 1);
3091                         if (kstart - 1 >= kernel_map.max_offset) {
3092                                 kstart = kernel_map.max_offset;
3093                                 break;
3094                         }
3095                         continue;
3096                 }
3097
3098                 /*
3099                  * We need a new PT
3100                  *
3101                  * This index is bogus, but out of the way
3102                  */
3103                 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3104                                      VM_ALLOC_NORMAL |
3105                                      VM_ALLOC_SYSTEM |
3106                                      VM_ALLOC_INTERRUPT);
3107                 if (nkpg == NULL)
3108                         panic("pmap_growkernel: no memory to grow kernel");
3109
3110                 vm_page_wire(nkpg);
3111                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
3112                 pmap_zero_page(ptppaddr);
3113                 newpt = (pd_entry_t)(ptppaddr |
3114                                      kernel_pmap.pmap_bits[PG_V_IDX] |
3115                                      kernel_pmap.pmap_bits[PG_RW_IDX] |
3116                                      kernel_pmap.pmap_bits[PG_A_IDX] |
3117                                      kernel_pmap.pmap_bits[PG_M_IDX]);
3118                 atomic_swap_long(pmap_pt(&kernel_pmap, kstart), newpt);
3119
3120                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3121                           ~(PAGE_SIZE * NPTEPG - 1);
3122
3123                 if (kstart - 1 >= kernel_map.max_offset) {
3124                         kstart = kernel_map.max_offset;
3125                         break;
3126                 }
3127         }
3128
3129         /*
3130          * Only update kernel_vm_end for areas below KERNBASE.
3131          */
3132         if (update_kernel_vm_end && kernel_vm_end < kstart)
3133                 kernel_vm_end = kstart;
3134 }
3135
3136 /*
3137  *      Add a reference to the specified pmap.
3138  */
3139 void
3140 pmap_reference(pmap_t pmap)
3141 {
3142         if (pmap != NULL)
3143                 atomic_add_int(&pmap->pm_count, 1);
3144 }
3145
3146 /***************************************************
3147  * page management routines.
3148  ***************************************************/
3149
3150 /*
3151  * Hold a pv without locking it
3152  */
3153 static void
3154 pv_hold(pv_entry_t pv)
3155 {
3156         atomic_add_int(&pv->pv_hold, 1);
3157 }
3158
3159 /*
3160  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
3161  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
3162  * the pv properly.
3163  *
3164  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3165  * pv list via its page) must be held by the caller in order to stabilize
3166  * the pv.
3167  */
3168 static int
3169 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3170 {
3171         u_int count;
3172
3173         /*
3174          * Critical path shortcut expects pv to already have one ref
3175          * (for the pv->pv_pmap).
3176          */
3177         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3178 #ifdef PMAP_DEBUG
3179                 pv->pv_func = func;
3180                 pv->pv_line = lineno;
3181 #endif
3182                 return TRUE;
3183         }
3184
3185         for (;;) {
3186                 count = pv->pv_hold;
3187                 cpu_ccfence();
3188                 if ((count & PV_HOLD_LOCKED) == 0) {
3189                         if (atomic_cmpset_int(&pv->pv_hold, count,
3190                                               (count + 1) | PV_HOLD_LOCKED)) {
3191 #ifdef PMAP_DEBUG
3192                                 pv->pv_func = func;
3193                                 pv->pv_line = lineno;
3194 #endif
3195                                 return TRUE;
3196                         }
3197                 } else {
3198                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3199                                 return FALSE;
3200                 }
3201                 /* retry */
3202         }
3203 }
3204
3205 /*
3206  * Drop a previously held pv_entry which could not be locked, allowing its
3207  * destruction.
3208  *
3209  * Must not be called with a spinlock held as we might zfree() the pv if it
3210  * is no longer associated with a pmap and this was the last hold count.
3211  */
3212 static void
3213 pv_drop(pv_entry_t pv)
3214 {
3215         u_int count;
3216
3217         for (;;) {
3218                 count = pv->pv_hold;
3219                 cpu_ccfence();
3220                 KKASSERT((count & PV_HOLD_MASK) > 0);
3221                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3222                          (PV_HOLD_LOCKED | 1));
3223                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3224                         if ((count & PV_HOLD_MASK) == 1) {
3225 #ifdef PMAP_DEBUG2
3226                                 if (pmap_enter_debug > 0) {
3227                                         --pmap_enter_debug;
3228                                         kprintf("pv_drop: free pv %p\n", pv);
3229                                 }
3230 #endif
3231                                 KKASSERT(count == 1);
3232                                 KKASSERT(pv->pv_pmap == NULL);
3233                                 zfree(pvzone, pv);
3234                         }
3235                         return;
3236                 }
3237                 /* retry */
3238         }
3239 }
3240
3241 /*
3242  * Find or allocate the requested PV entry, returning a locked, held pv.
3243  *
3244  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3245  * for the caller and one representing the pmap and vm_page association.
3246  *
3247  * If (*isnew) is zero, the returned pv will have only one hold count.
3248  *
3249  * Since both associations can only be adjusted while the pv is locked,
3250  * together they represent just one additional hold.
3251  */
3252 static
3253 pv_entry_t
3254 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3255 {
3256         pv_entry_t pv;
3257         pv_entry_t pnew;
3258         struct mdglobaldata *md = mdcpu;
3259
3260         pnew = NULL;
3261         if (md->gd_newpv) {
3262 #if 0
3263                 pnew = atomic_swap_ptr((void *)&md->gd_newpv, NULL);
3264 #else
3265                 crit_enter();
3266                 pnew = md->gd_newpv;    /* might race NULL */
3267                 md->gd_newpv = NULL;
3268                 crit_exit();
3269 #endif
3270         }
3271         if (pnew == NULL)
3272                 pnew = zalloc(pvzone);
3273
3274         spin_lock(&pmap->pm_spin);
3275         for (;;) {
3276                 /*
3277                  * Shortcut cache
3278                  */
3279                 pv = pmap->pm_pvhint;
3280                 cpu_ccfence();
3281                 if (pv == NULL ||
3282                     pv->pv_pmap != pmap ||
3283                     pv->pv_pindex != pindex) {
3284                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3285                                                         pindex);
3286                 }
3287                 if (pv == NULL) {
3288                         vm_pindex_t *pmark;
3289
3290                         /*
3291                          * We need to block if someone is holding our
3292                          * placemarker.  As long as we determine the
3293                          * placemarker has not been aquired we do not
3294                          * need to get it as acquision also requires
3295                          * the pmap spin lock.
3296                          *
3297                          * However, we can race the wakeup.
3298                          */
3299                         pmark = pmap_placemarker_hash(pmap, pindex);
3300
3301                         if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3302                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3303                                 tsleep_interlock(pmark, 0);
3304                                 if (((*pmark ^ pindex) &
3305                                      ~PM_PLACEMARK_WAKEUP) == 0) {
3306                                         spin_unlock(&pmap->pm_spin);
3307                                         tsleep(pmark, PINTERLOCKED, "pvplc", 0);
3308                                         spin_lock(&pmap->pm_spin);
3309                                 }
3310                                 continue;
3311                         }
3312
3313                         /*
3314                          * Setup the new entry
3315                          */
3316                         pnew->pv_pmap = pmap;
3317                         pnew->pv_pindex = pindex;
3318                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
3319 #ifdef PMAP_DEBUG
3320                         pnew->pv_func = func;
3321                         pnew->pv_line = lineno;
3322                         if (pnew->pv_line_lastfree > 0) {
3323                                 pnew->pv_line_lastfree =
3324                                                 -pnew->pv_line_lastfree;
3325                         }
3326 #endif
3327                         pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3328                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
3329                         spin_unlock(&pmap->pm_spin);
3330                         *isnew = 1;
3331
3332                         KKASSERT(pv == NULL);
3333                         return(pnew);
3334                 }
3335
3336                 /*
3337                  * We already have an entry, cleanup the staged pnew if
3338                  * we can get the lock, otherwise block and retry.
3339                  */
3340                 if (__predict_true(_pv_hold_try(pv PMAP_DEBUG_COPY))) {
3341                         spin_unlock(&pmap->pm_spin);
3342 #if 0
3343                         pnew = atomic_swap_ptr((void *)&md->gd_newpv, pnew);
3344                         if (pnew)
3345                                 zfree(pvzone, pnew);
3346 #else
3347                         crit_enter();
3348                         if (md->gd_newpv == NULL)
3349                                 md->gd_newpv = pnew;
3350                         else
3351                                 zfree(pvzone, pnew);
3352                         crit_exit();
3353 #endif
3354                         KKASSERT(pv->pv_pmap == pmap &&
3355                                  pv->pv_pindex == pindex);
3356                         *isnew = 0;
3357                         return(pv);
3358                 }
3359                 spin_unlock(&pmap->pm_spin);
3360                 _pv_lock(pv PMAP_DEBUG_COPY);
3361                 pv_put(pv);
3362                 spin_lock(&pmap->pm_spin);
3363         }
3364         /* NOT REACHED */
3365 }
3366
3367 /*
3368  * Find the requested PV entry, returning a locked+held pv or NULL
3369  */
3370 static
3371 pv_entry_t
3372 _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp PMAP_DEBUG_DECL)
3373 {
3374         pv_entry_t pv;
3375
3376         spin_lock(&pmap->pm_spin);
3377         for (;;) {
3378                 /*
3379                  * Shortcut cache
3380                  */
3381                 pv = pmap->pm_pvhint;
3382                 cpu_ccfence();
3383                 if (pv == NULL ||
3384                     pv->pv_pmap != pmap ||
3385                     pv->pv_pindex != pindex) {
3386                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3387                                                         pindex);
3388                 }
3389                 if (pv == NULL) {
3390                         /*
3391                          * Block if there is ANY placemarker.  If we are to
3392                          * return it, we must also aquire the spot, so we
3393                          * have to block even if the placemarker is held on
3394                          * a different address.
3395                          *
3396                          * OPTIMIZATION: If pmarkp is passed as NULL the
3397                          * caller is just probing (or looking for a real
3398                          * pv_entry), and in this case we only need to check
3399                          * to see if the placemarker matches pindex.
3400                          */
3401                         vm_pindex_t *pmark;
3402
3403                         pmark = pmap_placemarker_hash(pmap, pindex);
3404
3405                         if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3406                             ((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3407                                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3408                                 tsleep_interlock(pmark, 0);
3409                                 if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3410                                     ((*pmark ^ pindex) &
3411                                      ~PM_PLACEMARK_WAKEUP) == 0) {
3412                                         spin_unlock(&pmap->pm_spin);
3413                                         tsleep(pmark, PINTERLOCKED, "pvpld", 0);
3414                                         spin_lock(&pmap->pm_spin);
3415                                 }
3416                                 continue;
3417                         }
3418                         if (pmarkp) {
3419                                 if (atomic_swap_long(pmark, pindex) !=
3420                                     PM_NOPLACEMARK) {
3421                                         panic("_pv_get: pmark race");
3422                                 }
3423                                 *pmarkp = pmark;
3424                         }
3425                         spin_unlock(&pmap->pm_spin);
3426                         return NULL;
3427                 }
3428                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3429                         pv_cache(pv, pindex);
3430                         spin_unlock(&pmap->pm_spin);
3431                         KKASSERT(pv->pv_pmap == pmap &&
3432                                  pv->pv_pindex == pindex);
3433                         return(pv);
3434                 }
3435                 spin_unlock(&pmap->pm_spin);
3436                 _pv_lock(pv PMAP_DEBUG_COPY);
3437                 pv_put(pv);
3438                 spin_lock(&pmap->pm_spin);
3439         }
3440 }
3441
3442 /*
3443  * Lookup, hold, and attempt to lock (pmap,pindex).
3444  *
3445  * If the entry does not exist NULL is returned and *errorp is set to 0
3446  *
3447  * If the entry exists and could be successfully locked it is returned and
3448  * errorp is set to 0.
3449  *
3450  * If the entry exists but could NOT be successfully locked it is returned
3451  * held and *errorp is set to 1.
3452  *
3453  * If the entry is placemarked by someone else NULL is returned and *errorp
3454  * is set to 1.
3455  */
3456 static
3457 pv_entry_t
3458 pv_get_try(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp, int *errorp)
3459 {
3460         pv_entry_t pv;
3461
3462         spin_lock_shared(&pmap->pm_spin);
3463
3464         pv = pmap->pm_pvhint;
3465         cpu_ccfence();
3466         if (pv == NULL ||
3467             pv->pv_pmap != pmap ||
3468             pv->pv_pindex != pindex) {
3469                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3470         }
3471
3472         if (pv == NULL) {
3473                 vm_pindex_t *pmark;
3474
3475                 pmark = pmap_placemarker_hash(pmap, pindex);
3476
3477                 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3478                         *errorp = 1;
3479                 } else if (pmarkp &&
3480                            atomic_cmpset_long(pmark, PM_NOPLACEMARK, pindex)) {
3481                         *errorp = 0;
3482                 } else {
3483                         /*
3484                          * Can't set a placemark with a NULL pmarkp, or if
3485                          * pmarkp is non-NULL but we failed to set our
3486                          * placemark.
3487                          */
3488                         *errorp = 1;
3489                 }
3490                 if (pmarkp)
3491                         *pmarkp = pmark;
3492                 spin_unlock_shared(&pmap->pm_spin);
3493
3494                 return NULL;
3495         }
3496
3497         /*
3498          * XXX This has problems if the lock is shared, why?
3499          */
3500         if (pv_hold_try(pv)) {
3501                 pv_cache(pv, pindex);   /* overwrite ok (shared lock) */
3502                 spin_unlock_shared(&pmap->pm_spin);
3503                 *errorp = 0;
3504                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3505                 return(pv);     /* lock succeeded */
3506         }
3507         spin_unlock_shared(&pmap->pm_spin);
3508         *errorp = 1;
3509
3510         return (pv);            /* lock failed */
3511 }
3512
3513 /*
3514  * Lock a held pv, keeping the hold count
3515  */
3516 static
3517 void
3518 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3519 {
3520         u_int count;
3521
3522         for (;;) {
3523                 count = pv->pv_hold;
3524                 cpu_ccfence();
3525                 if ((count & PV_HOLD_LOCKED) == 0) {
3526                         if (atomic_cmpset_int(&pv->pv_hold, count,
3527                                               count | PV_HOLD_LOCKED)) {
3528 #ifdef PMAP_DEBUG
3529                                 pv->pv_func = func;
3530                                 pv->pv_line = lineno;
3531 #endif
3532                                 return;
3533                         }
3534                         continue;
3535                 }
3536                 tsleep_interlock(pv, 0);
3537                 if (atomic_cmpset_int(&pv->pv_hold, count,
3538                                       count | PV_HOLD_WAITING)) {
3539 #ifdef PMAP_DEBUG2
3540                         if (pmap_enter_debug > 0) {
3541                                 --pmap_enter_debug;
3542                                 kprintf("pv waiting on %s:%d\n",
3543                                         pv->pv_func, pv->pv_line);
3544                         }
3545 #endif
3546                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
3547                 }
3548                 /* retry */
3549         }
3550 }
3551
3552 /*
3553  * Unlock a held and locked pv, keeping the hold count.
3554  */
3555 static
3556 void
3557 pv_unlock(pv_entry_t pv)
3558 {
3559         u_int count;
3560
3561         for (;;) {
3562                 count = pv->pv_hold;
3563                 cpu_ccfence();
3564                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3565                          (PV_HOLD_LOCKED | 1));
3566                 if (atomic_cmpset_int(&pv->pv_hold, count,
3567                                       count &
3568                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3569                         if (count & PV_HOLD_WAITING)
3570                                 wakeup(pv);
3571                         break;
3572                 }
3573         }
3574 }
3575
3576 /*
3577  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3578  * and the hold count drops to zero we will free it.
3579  *
3580  * Caller should not hold any spin locks.  We are protected from hold races
3581  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3582  * lock held.  A pv cannot be located otherwise.
3583  */
3584 static
3585 void
3586 pv_put(pv_entry_t pv)
3587 {
3588 #ifdef PMAP_DEBUG2
3589         if (pmap_enter_debug > 0) {
3590                 --pmap_enter_debug;
3591                 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3592         }
3593 #endif
3594
3595         /*
3596          * Normal put-aways must have a pv_m associated with the pv,
3597          * but allow the case where the pv has been destructed due
3598          * to pmap_dynamic_delete.
3599          */
3600         KKASSERT(pv->pv_pmap == NULL || pv->pv_m != NULL);
3601
3602         /*
3603          * Fast - shortcut most common condition
3604          */
3605         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3606                 return;
3607
3608         /*
3609          * Slow
3610          */
3611         pv_unlock(pv);
3612         pv_drop(pv);
3613 }
3614
3615 /*
3616  * Remove the pmap association from a pv, require that pv_m already be removed,
3617  * then unlock and drop the pv.  Any pte operations must have already been
3618  * completed.  This call may result in a last-drop which will physically free
3619  * the pv.
3620  *
3621  * Removing the pmap association entails an additional drop.
3622  *
3623  * pv must be exclusively locked on call and will be disposed of on return.
3624  */
3625 static
3626 void
3627 _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL)
3628 {
3629         pmap_t pmap;
3630
3631 #ifdef PMAP_DEBUG
3632         pv->pv_func_lastfree = func;
3633         pv->pv_line_lastfree = lineno;
3634 #endif
3635         KKASSERT(pv->pv_m == NULL);
3636         KKASSERT((pv->pv_hold & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
3637                   (PV_HOLD_LOCKED|1));
3638         if ((pmap = pv->pv_pmap) != NULL) {
3639                 spin_lock(&pmap->pm_spin);
3640                 KKASSERT(pv->pv_pmap == pmap);
3641                 if (pmap->pm_pvhint == pv)
3642                         pmap->pm_pvhint = NULL;
3643                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3644                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3645                 pv->pv_pmap = NULL;
3646                 pv->pv_pindex = 0;
3647                 spin_unlock(&pmap->pm_spin);
3648
3649                 /*
3650                  * Try to shortcut three atomic ops, otherwise fall through
3651                  * and do it normally.  Drop two refs and the lock all in
3652                  * one go.
3653                  */
3654                 if (pvp)
3655                         vm_page_unwire_quick(pvp->pv_m);
3656                 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3657 #ifdef PMAP_DEBUG2
3658                         if (pmap_enter_debug > 0) {
3659                                 --pmap_enter_debug;
3660                                 kprintf("pv_free: free pv %p\n", pv);
3661                         }
3662 #endif
3663                         zfree(pvzone, pv);
3664                         return;
3665                 }
3666                 pv_drop(pv);    /* ref for pv_pmap */
3667         }
3668         pv_unlock(pv);
3669         pv_drop(pv);
3670 }
3671
3672 /*
3673  * This routine is very drastic, but can save the system
3674  * in a pinch.
3675  */
3676 void
3677 pmap_collect(void)
3678 {
3679         int i;
3680         vm_page_t m;
3681         static int warningdone=0;
3682
3683         if (pmap_pagedaemon_waken == 0)
3684                 return;
3685         pmap_pagedaemon_waken = 0;
3686         if (warningdone < 5) {
3687                 kprintf("pmap_collect: collecting pv entries -- "
3688                         "suggest increasing PMAP_SHPGPERPROC\n");
3689                 warningdone++;
3690         }
3691
3692         for (i = 0; i < vm_page_array_size; i++) {
3693                 m = &vm_page_array[i];
3694                 if (m->wire_count || m->hold_count)
3695                         continue;
3696                 if (vm_page_busy_try(m, TRUE) == 0) {
3697                         if (m->wire_count == 0 && m->hold_count == 0) {
3698                                 pmap_remove_all(m);
3699                         }
3700                         vm_page_wakeup(m);
3701                 }
3702         }
3703 }
3704
3705 /*
3706  * Scan the pmap for active page table entries and issue a callback.
3707  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3708  * its parent page table.
3709  *
3710  * pte_pv will be NULL if the page or page table is unmanaged.
3711  * pt_pv will point to the page table page containing the pte for the page.
3712  *
3713  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3714  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3715  *       process pmap's PD and page to the callback function.  This can be
3716  *       confusing because the pt_pv is really a pd_pv, and the target page
3717  *       table page is simply aliased by the pmap and not owned by it.
3718  *
3719  * It is assumed that the start and end are properly rounded to the page size.
3720  *
3721  * It is assumed that PD pages and above are managed and thus in the RB tree,
3722  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3723  */
3724 struct pmap_scan_info {
3725         struct pmap *pmap;
3726         vm_offset_t sva;
3727         vm_offset_t eva;
3728         vm_pindex_t sva_pd_pindex;
3729         vm_pindex_t eva_pd_pindex;
3730         void (*func)(pmap_t, struct pmap_scan_info *,
3731                      pv_entry_t, vm_pindex_t *, pv_entry_t,
3732                      int, vm_offset_t,
3733                      pt_entry_t *, void *);
3734         void *arg;
3735         pmap_inval_bulk_t bulk_core;
3736         pmap_inval_bulk_t *bulk;
3737         int count;
3738         int stop;
3739 };
3740
3741 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3742 static int pmap_scan_callback(pv_entry_t pv, void *data);
3743
3744 static void
3745 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3746 {
3747         struct pmap *pmap = info->pmap;
3748         pv_entry_t pd_pv;       /* A page directory PV */
3749         pv_entry_t pt_pv;       /* A page table PV */
3750         pv_entry_t pte_pv;      /* A page table entry PV */
3751         vm_pindex_t *pte_placemark;
3752         vm_pindex_t *pt_placemark;
3753         pt_entry_t *ptep;
3754         pt_entry_t oldpte;
3755         struct pv_entry dummy_pv;
3756
3757         info->stop = 0;
3758         if (pmap == NULL)
3759                 return;
3760         if (info->sva == info->eva)
3761                 return;
3762         if (smp_inval) {
3763                 info->bulk = &info->bulk_core;
3764                 pmap_inval_bulk_init(&info->bulk_core, pmap);
3765         } else {
3766                 info->bulk = NULL;
3767         }
3768
3769         /*
3770          * Hold the token for stability; if the pmap is empty we have nothing
3771          * to do.
3772          */
3773 #if 0
3774         if (pmap->pm_stats.resident_count == 0) {
3775                 return;
3776         }
3777 #endif
3778
3779         info->count = 0;
3780
3781         /*
3782          * Special handling for scanning one page, which is a very common
3783          * operation (it is?).
3784          *
3785          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3786          */
3787         if (info->sva + PAGE_SIZE == info->eva) {
3788                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3789                         /*
3790                          * Kernel mappings do not track wire counts on
3791                          * page table pages and only maintain pd_pv and
3792                          * pte_pv levels so pmap_scan() works.
3793                          */
3794                         pt_pv = NULL;
3795                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3796                                         &pte_placemark);
3797                         ptep = vtopte(info->sva);
3798                 } else {
3799                         /*
3800                          * User pages which are unmanaged will not have a
3801                          * pte_pv.  User page table pages which are unmanaged
3802                          * (shared from elsewhere) will also not have a pt_pv.
3803                          * The func() callback will pass both pte_pv and pt_pv
3804                          * as NULL in that case.
3805                          *
3806                          * We hold pte_placemark across the operation for
3807                          * unmanaged pages.
3808                          *
3809                          * WARNING!  We must hold pt_placemark across the
3810                          *           *ptep test to prevent misintepreting
3811                          *           a non-zero *ptep as a shared page
3812                          *           table page.  Hold it across the function
3813                          *           callback as well for SMP safety.
3814                          */
3815                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3816                                         &pte_placemark);
3817                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva),
3818                                         &pt_placemark);
3819                         if (pt_pv == NULL) {
3820                                 KKASSERT(pte_pv == NULL);
3821                                 pd_pv = pv_get(pmap,
3822                                                pmap_pd_pindex(info->sva),
3823                                                NULL);
3824                                 if (pd_pv) {
3825                                         ptep = pv_pte_lookup(pd_pv,
3826                                                     pmap_pt_index(info->sva));
3827                                         if (*ptep) {
3828                                                 info->func(pmap, info,
3829                                                      NULL, pt_placemark,
3830                                                      pd_pv, 1,
3831                                                      info->sva, ptep,
3832                                                      info->arg);
3833                                         } else {
3834                                                 pv_placemarker_wakeup(pmap,
3835                                                                   pt_placemark);
3836                                         }
3837                                         pv_put(pd_pv);
3838                                 } else {
3839                                         pv_placemarker_wakeup(pmap,
3840                                                               pt_placemark);
3841                                 }
3842                                 pv_placemarker_wakeup(pmap, pte_placemark);
3843                                 goto fast_skip;
3844                         }
3845                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3846                 }
3847
3848                 /*
3849                  * NOTE: *ptep can't be ripped out from under us if we hold
3850                  *       pte_pv (or pte_placemark) locked, but bits can
3851                  *       change.
3852                  */
3853                 oldpte = *ptep;
3854                 cpu_ccfence();
3855                 if (oldpte == 0) {
3856                         KKASSERT(pte_pv == NULL);
3857                         pv_placemarker_wakeup(pmap, pte_placemark);
3858                 } else if (pte_pv) {
3859                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3860                                            pmap->pmap_bits[PG_V_IDX])) ==
3861                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
3862                                  pmap->pmap_bits[PG_V_IDX]),
3863                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p",
3864                             *ptep, oldpte, info->sva, pte_pv));
3865                         info->func(pmap, info, pte_pv, NULL, pt_pv, 0,
3866                                    info->sva, ptep, info->arg);
3867                 } else {
3868                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3869                                            pmap->pmap_bits[PG_V_IDX])) ==
3870                             pmap->pmap_bits[PG_V_IDX],
3871                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
3872                             *ptep, oldpte, info->sva));
3873                         info->func(pmap, info, NULL, pte_placemark, pt_pv, 0,
3874                                    info->sva, ptep, info->arg);
3875                 }
3876                 if (pt_pv)
3877                         pv_put(pt_pv);
3878 fast_skip:
3879                 pmap_inval_bulk_flush(info->bulk);
3880                 return;
3881         }
3882
3883         /*
3884          * Nominal scan case, RB_SCAN() for PD pages and iterate from
3885          * there.
3886          *
3887          * WARNING! eva can overflow our standard ((N + mask) >> bits)
3888          *          bounds, resulting in a pd_pindex of 0.  To solve the
3889          *          problem we use an inclusive range.
3890          */
3891         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3892         info->eva_pd_pindex = pmap_pd_pindex(info->eva - PAGE_SIZE);
3893
3894         if (info->sva >= VM_MAX_USER_ADDRESS) {
3895                 /*
3896                  * The kernel does not currently maintain any pv_entry's for
3897                  * higher-level page tables.
3898                  */
3899                 bzero(&dummy_pv, sizeof(dummy_pv));
3900                 dummy_pv.pv_pindex = info->sva_pd_pindex;
3901                 spin_lock(&pmap->pm_spin);
3902                 while (dummy_pv.pv_pindex <= info->eva_pd_pindex) {
3903                         pmap_scan_callback(&dummy_pv, info);
3904                         ++dummy_pv.pv_pindex;
3905                         if (dummy_pv.pv_pindex < info->sva_pd_pindex) /*wrap*/
3906                                 break;
3907                 }
3908                 spin_unlock(&pmap->pm_spin);
3909         } else {
3910                 /*
3911                  * User page tables maintain local PML4, PDP, and PD
3912                  * pv_entry's at the very least.  PT pv's might be
3913                  * unmanaged and thus not exist.  PTE pv's might be
3914                  * unmanaged and thus not exist.
3915                  */
3916                 spin_lock(&pmap->pm_spin);
3917                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot, pmap_scan_cmp,
3918                                          pmap_scan_callback, info);
3919                 spin_unlock(&pmap->pm_spin);
3920         }
3921         pmap_inval_bulk_flush(info->bulk);
3922 }
3923
3924 /*
3925  * WARNING! pmap->pm_spin held
3926  *
3927  * WARNING! eva can overflow our standard ((N + mask) >> bits)
3928  *          bounds, resulting in a pd_pindex of 0.  To solve the
3929  *          problem we use an inclusive range.
3930  */
3931 static int
3932 pmap_scan_cmp(pv_entry_t pv, void *data)
3933 {
3934         struct pmap_scan_info *info = data;
3935         if (pv->pv_pindex < info->sva_pd_pindex)
3936                 return(-1);
3937         if (pv->pv_pindex > info->eva_pd_pindex)
3938                 return(1);
3939         return(0);
3940 }
3941
3942 /*
3943  * pmap_scan() by PDs
3944  *
3945  * WARNING! pmap->pm_spin held
3946  */
3947 static int
3948 pmap_scan_callback(pv_entry_t pv, void *data)
3949 {
3950         struct pmap_scan_info *info = data;
3951         struct pmap *pmap = info->pmap;
3952         pv_entry_t pd_pv;       /* A page directory PV */
3953         pv_entry_t pt_pv;       /* A page table PV */
3954         vm_pindex_t *pt_placemark;
3955         pt_entry_t *ptep;
3956         pt_entry_t oldpte;
3957         vm_offset_t sva;
3958         vm_offset_t eva;
3959         vm_offset_t va_next;
3960         vm_pindex_t pd_pindex;
3961         int error;
3962
3963         /*
3964          * Stop if requested
3965          */
3966         if (info->stop)
3967                 return -1;
3968
3969         /*
3970          * Pull the PD pindex from the pv before releasing the spinlock.
3971          *
3972          * WARNING: pv is faked for kernel pmap scans.
3973          */
3974         pd_pindex = pv->pv_pindex;
3975         spin_unlock(&pmap->pm_spin);
3976         pv = NULL;      /* invalid after spinlock unlocked */
3977
3978         /*
3979          * Calculate the page range within the PD.  SIMPLE pmaps are
3980          * direct-mapped for the entire 2^64 address space.  Normal pmaps
3981          * reflect the user and kernel address space which requires
3982          * cannonicalization w/regards to converting pd_pindex's back
3983          * into addresses.
3984          */
3985         sva = (pd_pindex - pmap_pd_pindex(0)) << PDPSHIFT;
3986         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3987             (sva & PML4_SIGNMASK)) {
3988                 sva |= PML4_SIGNMASK;
3989         }
3990         eva = sva + NBPDP;      /* can overflow */
3991         if (sva < info->sva)
3992                 sva = info->sva;
3993         if (eva < info->sva || eva > info->eva)
3994                 eva = info->eva;
3995
3996         /*
3997          * NOTE: kernel mappings do not track page table pages, only
3998          *       terminal pages.
3999          *
4000          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
4001          *       However, for the scan to be efficient we try to
4002          *       cache items top-down.
4003          */
4004         pd_pv = NULL;
4005         pt_pv = NULL;
4006
4007         for (; sva < eva; sva = va_next) {
4008                 if (info->stop)
4009                         break;
4010                 if (sva >= VM_MAX_USER_ADDRESS) {
4011                         if (pt_pv) {
4012                                 pv_put(pt_pv);
4013                                 pt_pv = NULL;
4014                         }
4015                         goto kernel_skip;
4016                 }
4017
4018                 /*
4019                  * PD cache, scan shortcut if it doesn't exist.
4020                  */
4021                 if (pd_pv == NULL) {
4022                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4023                 } else if (pd_pv->pv_pmap != pmap ||
4024                            pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
4025                         pv_put(pd_pv);
4026                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4027                 }
4028                 if (pd_pv == NULL) {
4029                         va_next = (sva + NBPDP) & ~PDPMASK;
4030                         if (va_next < sva)
4031                                 va_next = eva;
4032                         continue;
4033                 }
4034
4035                 /*
4036                  * PT cache
4037                  *
4038                  * NOTE: The cached pt_pv can be removed from the pmap when
4039                  *       pmap_dynamic_delete is enabled.
4040                  */
4041                 if (pt_pv && (pt_pv->pv_pmap != pmap ||
4042                               pt_pv->pv_pindex != pmap_pt_pindex(sva))) {
4043                         pv_put(pt_pv);
4044                         pt_pv = NULL;
4045                 }
4046                 if (pt_pv == NULL) {
4047                         pt_pv = pv_get_try(pmap, pmap_pt_pindex(sva),
4048                                            &pt_placemark, &error);
4049                         if (error) {
4050                                 pv_put(pd_pv);  /* lock order */
4051                                 pd_pv = NULL;
4052                                 if (pt_pv) {
4053                                         pv_lock(pt_pv);
4054                                         pv_put(pt_pv);
4055                                         pt_pv = NULL;
4056                                 } else {
4057                                         pv_placemarker_wait(pmap, pt_placemark);
4058                                 }
4059                                 va_next = sva;
4060                                 continue;
4061                         }
4062                         /* may have to re-check later if pt_pv is NULL here */
4063                 }
4064
4065                 /*
4066                  * If pt_pv is NULL we either have an shared page table
4067                  * page and must issue a callback specific to that case,
4068                  * or there is no page table page.
4069                  *
4070                  * Either way we can skip the page table page.
4071                  *
4072                  * WARNING! pt_pv can also be NULL due to a pv creation
4073                  *          race where we find it to be NULL and then
4074                  *          later see a pte_pv.  But its possible the pt_pv
4075                  *          got created inbetween the two operations, so
4076                  *          we must check.
4077                  */
4078                 if (pt_pv == NULL) {
4079                         /*
4080                          * Possible unmanaged (shared from another pmap)
4081                          * page table page.
4082                          *
4083                          * WARNING!  We must hold pt_placemark across the
4084                          *           *ptep test to prevent misintepreting
4085                          *           a non-zero *ptep as a shared page
4086                          *           table page.  Hold it across the function
4087                          *           callback as well for SMP safety.
4088                          */
4089                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
4090                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
4091                                 info->func(pmap, info, NULL, pt_placemark,
4092                                            pd_pv, 1,
4093                                            sva, ptep, info->arg);
4094                         } else {
4095                                 pv_placemarker_wakeup(pmap, pt_placemark);
4096                         }
4097
4098                         /*
4099                          * Done, move to next page table page.
4100                          */
4101                         va_next = (sva + NBPDR) & ~PDRMASK;
4102                         if (va_next < sva)
4103                                 va_next = eva;
4104                         continue;
4105                 }
4106
4107                 /*
4108                  * From this point in the loop testing pt_pv for non-NULL
4109                  * means we are in UVM, else if it is NULL we are in KVM.
4110                  *
4111                  * Limit our scan to either the end of the va represented
4112                  * by the current page table page, or to the end of the
4113                  * range being removed.
4114                  */
4115 kernel_skip:
4116                 va_next = (sva + NBPDR) & ~PDRMASK;
4117                 if (va_next < sva)
4118                         va_next = eva;
4119                 if (va_next > eva)
4120                         va_next = eva;
4121
4122                 /*
4123                  * Scan the page table for pages.  Some pages may not be
4124                  * managed (might not have a pv_entry).
4125                  *
4126                  * There is no page table management for kernel pages so
4127                  * pt_pv will be NULL in that case, but otherwise pt_pv
4128                  * is non-NULL, locked, and referenced.
4129                  */
4130
4131                 /*
4132                  * At this point a non-NULL pt_pv means a UVA, and a NULL
4133                  * pt_pv means a KVA.
4134                  */
4135                 if (pt_pv)
4136                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
4137                 else
4138                         ptep = vtopte(sva);
4139
4140                 while (sva < va_next) {
4141                         pv_entry_t pte_pv;
4142                         vm_pindex_t *pte_placemark;
4143
4144                         /*
4145                          * Yield every 64 pages, stop if requested.
4146                          */
4147                         if ((++info->count & 63) == 0)
4148                                 lwkt_user_yield();
4149                         if (info->stop)
4150                                 break;
4151
4152                         /*
4153                          * We can shortcut our scan if *ptep == 0.  This is
4154                          * an unlocked check.
4155                          */
4156                         if (*ptep == 0) {
4157                                 sva += PAGE_SIZE;
4158                                 ++ptep;
4159                                 continue;
4160                         }
4161                         cpu_ccfence();
4162
4163                         /*
4164                          * Acquire the related pte_pv, if any.  If *ptep == 0
4165                          * the related pte_pv should not exist, but if *ptep
4166                          * is not zero the pte_pv may or may not exist (e.g.
4167                          * will not exist for an unmanaged page).
4168                          *
4169                          * However a multitude of races are possible here
4170                          * so if we cannot lock definite state we clean out
4171                          * our cache and break the inner while() loop to
4172                          * force a loop up to the top of the for().
4173                          *
4174                          * XXX unlock/relock pd_pv, pt_pv, and re-test their
4175                          *     validity instead of looping up?
4176                          */
4177                         pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
4178                                             &pte_placemark, &error);
4179                         if (error) {
4180                                 pv_put(pd_pv);          /* lock order */
4181                                 pd_pv = NULL;
4182                                 if (pt_pv) {
4183                                         pv_put(pt_pv);  /* lock order */
4184                                         pt_pv = NULL;
4185                                 }
4186                                 if (pte_pv) {           /* block */
4187                                         pv_lock(pte_pv);
4188                                         pv_put(pte_pv);
4189                                         pte_pv = NULL;
4190                                 } else {
4191                                         pv_placemarker_wait(pmap,
4192                                                         pte_placemark);
4193                                 }
4194                                 va_next = sva;          /* retry */
4195                                 break;
4196                         }
4197
4198                         /*
4199                          * Reload *ptep after successfully locking the
4200                          * pindex.  If *ptep == 0 we had better NOT have a
4201                          * pte_pv.
4202                          */
4203                         cpu_ccfence();
4204                         oldpte = *ptep;
4205                         if (oldpte == 0) {
4206                                 if (pte_pv) {
4207                                         kprintf("Unexpected non-NULL pte_pv "
4208                                                 "%p pt_pv %p "
4209                                                 "*ptep = %016lx/%016lx\n",
4210                                                 pte_pv, pt_pv, *ptep, oldpte);
4211                                         panic("Unexpected non-NULL pte_pv");
4212                                 } else {
4213                                         pv_placemarker_wakeup(pmap, pte_placemark);
4214                                 }
4215                                 sva += PAGE_SIZE;
4216                                 ++ptep;
4217                                 continue;
4218                         }
4219
4220                         /*
4221                          * We can't hold pd_pv across the callback (because
4222                          * we don't pass it to the callback and the callback
4223                          * might deadlock)
4224                          */
4225                         if (pd_pv) {
4226                                 vm_page_wire_quick(pd_pv->pv_m);
4227                                 pv_unlock(pd_pv);
4228                         }
4229
4230                         /*
4231                          * Ready for the callback.  The locked pte_pv (if any)
4232                          * is consumed by the callback.  pte_pv will exist if
4233                          * the page is managed, and will not exist if it
4234                          * isn't.
4235                          */
4236                         if (oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4237                                 /*
4238                                  * Managed pte
4239                                  */
4240                                 KASSERT(pte_pv &&
4241                                          (oldpte & pmap->pmap_bits[PG_V_IDX]),
4242                                     ("badC *ptep %016lx/%016lx sva %016lx "
4243                                     "pte_pv %p",
4244                                     *ptep, oldpte, sva, pte_pv));
4245                                 /*
4246                                  * We must unlock pd_pv across the callback
4247                                  * to avoid deadlocks on any recursive
4248                                  * disposal.  Re-check that it still exists
4249                                  * after re-locking.
4250                                  *
4251                                  * Call target disposes of pte_pv and may
4252                                  * destroy but will not dispose of pt_pv.
4253                                  */
4254                                 info->func(pmap, info, pte_pv, NULL,
4255                                            pt_pv, 0,
4256                                            sva, ptep, info->arg);
4257                         } else {
4258                                 /*
4259                                  * Unmanaged pte
4260                                  *
4261                                  * We must unlock pd_pv across the callback
4262                                  * to avoid deadlocks on any recursive
4263                                  * disposal.  Re-check that it still exists
4264                                  * after re-locking.
4265                                  *
4266                                  * Call target disposes of pte_pv or
4267                                  * pte_placemark and may destroy but will
4268                                  * not dispose of pt_pv.
4269                                  */
4270                                 KASSERT(pte_pv == NULL &&
4271                                         (oldpte & pmap->pmap_bits[PG_V_IDX]),
4272                                     ("badD *ptep %016lx/%016lx sva %016lx "
4273                                     "pte_pv %p pte_pv->pv_m %p ",
4274                                      *ptep, oldpte, sva,
4275                                      pte_pv, (pte_pv ? pte_pv->pv_m : NULL)));
4276                                 if (pte_pv)
4277                                         kprintf("RaceD\n");
4278                                 if (pte_pv) {
4279                                         info->func(pmap, info,
4280                                                    pte_pv, NULL,
4281                                                    pt_pv, 0,
4282                                                    sva, ptep, info->arg);
4283                                 } else {
4284                                         info->func(pmap, info,
4285                                                    NULL, pte_placemark,
4286                                                    pt_pv, 0,
4287                                                    sva, ptep, info->arg);
4288                                 }
4289                         }
4290                         if (pd_pv) {
4291                                 pv_lock(pd_pv);
4292                                 vm_page_unwire_quick(pd_pv->pv_m);
4293                                 if (pd_pv->pv_pmap == NULL) {
4294                                         va_next = sva;          /* retry */
4295                                         break;
4296                                 }
4297                         }
4298
4299                         /*
4300                          * NOTE: The cached pt_pv can be removed from the
4301                          *       pmap when pmap_dynamic_delete is enabled,
4302                          *       which will cause ptep to become stale.
4303                          *
4304                          *       This also means that no pages remain under
4305                          *       the PT, so we can just break out of the inner
4306                          *       loop and let the outer loop clean everything
4307                          *       up.
4308                          */
4309                         if (pt_pv && pt_pv->pv_pmap != pmap)
4310                                 break;
4311                         pte_pv = NULL;
4312                         sva += PAGE_SIZE;
4313                         ++ptep;
4314                 }
4315         }
4316         if (pd_pv) {
4317                 pv_put(pd_pv);
4318                 pd_pv = NULL;
4319         }
4320         if (pt_pv) {
4321                 pv_put(pt_pv);
4322                 pt_pv = NULL;
4323         }
4324         if ((++info->count & 7) == 0)
4325                 lwkt_user_yield();
4326
4327         /*
4328          * Relock before returning.
4329          */
4330         spin_lock(&pmap->pm_spin);
4331         return (0);
4332 }
4333
4334 void
4335 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4336 {
4337         struct pmap_scan_info info;
4338
4339         info.pmap = pmap;
4340         info.sva = sva;
4341         info.eva = eva;
4342         info.func = pmap_remove_callback;
4343         info.arg = NULL;
4344         pmap_scan(&info, 1);
4345 #if 0
4346         cpu_invltlb();
4347         if (eva - sva < 1024*1024) {
4348                 while (sva < eva) {
4349                         cpu_invlpg((void *)sva);
4350                         sva += PAGE_SIZE;
4351                 }
4352         }
4353 #endif
4354 }
4355
4356 static void
4357 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4358 {
4359         struct pmap_scan_info info;
4360
4361         info.pmap = pmap;
4362         info.sva = sva;
4363         info.eva = eva;
4364         info.func = pmap_remove_callback;
4365         info.arg = NULL;
4366         pmap_scan(&info, 0);
4367 }
4368
4369 static void
4370 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4371                      pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4372                      pv_entry_t pt_pv, int sharept,
4373                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4374 {
4375         pt_entry_t pte;
4376
4377         if (pte_pv) {
4378                 /*
4379                  * Managed entry
4380                  *
4381                  * This will also drop pt_pv's wire_count. Note that
4382                  * terminal pages are not wired based on mmu presence.
4383                  *
4384                  * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4385                  */
4386                 KKASSERT(pte_pv->pv_m != NULL);
4387                 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4388                 pte_pv = NULL;  /* safety */
4389
4390                 /*
4391                  * Recursively destroy higher-level page tables.
4392                  *
4393                  * This is optional.  If we do not, they will still
4394                  * be destroyed when the process exits.
4395                  *
4396                  * NOTE: Do not destroy pv_entry's with extra hold refs,
4397                  *       a caller may have unlocked it and intends to
4398                  *       continue to use it.
4399                  */
4400                 if (pmap_dynamic_delete &&
4401                     pt_pv &&
4402                     pt_pv->pv_m &&
4403                     pt_pv->pv_m->wire_count == 1 &&
4404                     (pt_pv->pv_hold & PV_HOLD_MASK) == 2 &&
4405                     pt_pv->pv_pindex != pmap_pml4_pindex()) {
4406                         if (pmap_dynamic_delete == 2)
4407                                 kprintf("B %jd %08x\n", pt_pv->pv_pindex, pt_pv->pv_hold);
4408                         pv_hold(pt_pv); /* extra hold */
4409                         pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4410                         pv_lock(pt_pv); /* prior extra hold + relock */
4411                 }
4412         } else if (sharept == 0) {
4413                 /*
4414                  * Unmanaged pte (pte_placemark is non-NULL)
4415                  *
4416                  * pt_pv's wire_count is still bumped by unmanaged pages
4417                  * so we must decrement it manually.
4418                  *
4419                  * We have to unwire the target page table page.
4420                  */
4421                 pte = pmap_inval_bulk(info->bulk, va, ptep, 0);
4422                 if (pte & pmap->pmap_bits[PG_W_IDX])
4423                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4424                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4425                 if (vm_page_unwire_quick(pt_pv->pv_m))
4426                         panic("pmap_remove: insufficient wirecount");
4427                 pv_placemarker_wakeup(pmap, pte_placemark);
4428         } else {
4429                 /*
4430                  * Unmanaged page table (pt, pd, or pdp. Not pte) for
4431                  * a shared page table.
4432                  *
4433                  * pt_pv is actually the pd_pv for our pmap (not the shared
4434                  * object pmap).
4435                  *
4436                  * We have to unwire the target page table page and we
4437                  * have to unwire our page directory page.
4438                  *
4439                  * It is unclear how we can invalidate a segment so we
4440                  * invalidate -1 which invlidates the tlb.
4441                  */
4442                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4443                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4444                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4445                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4446                         panic("pmap_remove: shared pgtable1 bad wirecount");
4447                 if (vm_page_unwire_quick(pt_pv->pv_m))
4448                         panic("pmap_remove: shared pgtable2 bad wirecount");
4449                 pv_placemarker_wakeup(pmap, pte_placemark);
4450         }
4451 }
4452
4453 /*
4454  * Removes this physical page from all physical maps in which it resides.
4455  * Reflects back modify bits to the pager.
4456  *
4457  * This routine may not be called from an interrupt.
4458  */
4459 static
4460 void
4461 pmap_remove_all(vm_page_t m)
4462 {
4463         pv_entry_t pv;
4464         pmap_inval_bulk_t bulk;
4465
4466         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4467                 return;
4468
4469         vm_page_spin_lock(m);
4470         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4471                 KKASSERT(pv->pv_m == m);
4472                 if (pv_hold_try(pv)) {
4473                         vm_page_spin_unlock(m);
4474                 } else {
4475                         vm_page_spin_unlock(m);
4476                         pv_lock(pv);
4477                         pv_put(pv);
4478                         vm_page_spin_lock(m);
4479                         continue;
4480                 }
4481                 KKASSERT(pv->pv_pmap && pv->pv_m == m);
4482
4483                 /*
4484                  * Holding no spinlocks, pv is locked.  Once we scrap
4485                  * pv we can no longer use it as a list iterator (but
4486                  * we are doing a TAILQ_FIRST() so we are ok).
4487                  */
4488                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4489                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4490                 pv = NULL;      /* safety */
4491                 pmap_inval_bulk_flush(&bulk);
4492                 vm_page_spin_lock(m);
4493         }
4494         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4495         vm_page_spin_unlock(m);
4496 }
4497
4498 /*
4499  * Removes the page from a particular pmap
4500  */
4501 void
4502 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4503 {
4504         pv_entry_t pv;
4505         pmap_inval_bulk_t bulk;
4506
4507         if (!pmap_initialized)
4508                 return;
4509
4510 again:
4511         vm_page_spin_lock(m);
4512         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4513                 if (pv->pv_pmap != pmap)
4514                         continue;
4515                 KKASSERT(pv->pv_m == m);
4516                 if (pv_hold_try(pv)) {
4517                         vm_page_spin_unlock(m);
4518                 } else {
4519                         vm_page_spin_unlock(m);
4520                         pv_lock(pv);
4521                         pv_put(pv);
4522                         goto again;
4523                 }
4524                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
4525
4526                 /*
4527                  * Holding no spinlocks, pv is locked.  Once gone it can't
4528                  * be used as an iterator.  In fact, because we couldn't
4529                  * necessarily lock it atomically it may have moved within
4530                  * the list and ALSO cannot be used as an iterator.
4531                  */
4532                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4533                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4534                 pv = NULL;      /* safety */
4535                 pmap_inval_bulk_flush(&bulk);
4536                 goto again;
4537         }
4538         vm_page_spin_unlock(m);
4539 }
4540
4541 /*
4542  * Set the physical protection on the specified range of this map
4543  * as requested.  This function is typically only used for debug watchpoints
4544  * and COW pages.
4545  *
4546  * This function may not be called from an interrupt if the map is
4547  * not the kernel_pmap.
4548  *
4549  * NOTE!  For shared page table pages we just unmap the page.
4550  */
4551 void
4552 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4553 {
4554         struct pmap_scan_info info;
4555         /* JG review for NX */
4556
4557         if (pmap == NULL)
4558                 return;
4559         if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
4560                 pmap_remove(pmap, sva, eva);
4561                 return;
4562         }
4563         if (prot & VM_PROT_WRITE)
4564                 return;
4565         info.pmap = pmap;
4566         info.sva = sva;
4567         info.eva = eva;
4568         info.func = pmap_protect_callback;
4569         info.arg = &prot;
4570         pmap_scan(&info, 1);
4571 }
4572
4573 static
4574 void
4575 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4576                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4577                       pv_entry_t pt_pv, int sharept,
4578                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4579 {
4580         pt_entry_t pbits;
4581         pt_entry_t cbits;
4582         pt_entry_t pte;
4583         vm_page_t m;
4584
4585 again:
4586         pbits = *ptep;
4587         cbits = pbits;
4588         if (pte_pv) {
4589                 KKASSERT(pte_pv->pv_m != NULL);
4590                 m = NULL;
4591                 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4592                         if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4593                                 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4594                                 KKASSERT(m == pte_pv->pv_m);
4595                                 vm_page_flag_set(m, PG_REFERENCED);
4596                         }
4597                         cbits &= ~pmap->pmap_bits[PG_A_IDX];
4598                 }
4599                 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4600                         if (pmap_track_modified(pte_pv->pv_pindex)) {
4601                                 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4602                                         if (m == NULL) {
4603                                                 m = PHYS_TO_VM_PAGE(pbits &
4604                                                                     PG_FRAME);
4605                                         }
4606                                         vm_page_dirty(m);
4607                                 }
4608                                 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4609                         }
4610                 }
4611         } else if (sharept) {
4612                 /*
4613                  * Unmanaged page table, pt_pv is actually the pd_pv
4614                  * for our pmap (not the object's shared pmap).
4615                  *
4616                  * When asked to protect something in a shared page table
4617                  * page we just unmap the page table page.  We have to
4618                  * invalidate the tlb in this situation.
4619                  *
4620                  * XXX Warning, shared page tables will not be used for
4621                  * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4622                  * so PHYS_TO_VM_PAGE() should be safe here.
4623                  */
4624                 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4625                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4626                         panic("pmap_protect: pgtable1 pg bad wirecount");
4627                 if (vm_page_unwire_quick(pt_pv->pv_m))
4628                         panic("pmap_protect: pgtable2 pg bad wirecount");
4629                 ptep = NULL;
4630         }
4631         /* else unmanaged page, adjust bits, no wire changes */
4632
4633         if (ptep) {
4634                 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4635 #ifdef PMAP_DEBUG2
4636                 if (pmap_enter_debug > 0) {
4637                         --pmap_enter_debug;
4638                         kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4639                                 "pt_pv=%p cbits=%08lx\n",
4640                                 va, ptep, pte_pv,
4641                                 pt_pv, cbits
4642                         );
4643                 }
4644 #endif
4645                 if (pbits != cbits) {
4646                         vm_offset_t xva;
4647
4648                         xva = (sharept) ? (vm_offset_t)-1 : va;
4649                         if (!pmap_inval_smp_cmpset(pmap, xva,
4650                                                    ptep, pbits, cbits)) {
4651                                 goto again;
4652                         }
4653                 }
4654         }
4655         if (pte_pv)
4656                 pv_put(pte_pv);
4657         else
4658                 pv_placemarker_wakeup(pmap, pte_placemark);
4659 }
4660
4661 /*
4662  * Insert the vm_page (m) at the virtual address (va), replacing any prior
4663  * mapping at that address.  Set protection and wiring as requested.
4664  *
4665  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4666  * possible.  If it is we enter the page into the appropriate shared pmap
4667  * hanging off the related VM object instead of the passed pmap, then we
4668  * share the page table page from the VM object's pmap into the current pmap.
4669  *
4670  * NOTE: This routine MUST insert the page into the pmap now, it cannot
4671  *       lazy-evaluate.
4672  *
4673  * NOTE: If (m) is PG_UNMANAGED it may also be a temporary fake vm_page_t.
4674  *       never record it.
4675  */
4676 void
4677 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4678            boolean_t wired, vm_map_entry_t entry)
4679 {
4680         pv_entry_t pt_pv;       /* page table */
4681         pv_entry_t pte_pv;      /* page table entry */
4682         vm_pindex_t *pte_placemark;
4683         pt_entry_t *ptep;
4684         vm_paddr_t opa;
4685         pt_entry_t origpte, newpte;
4686         vm_paddr_t pa;
4687
4688         if (pmap == NULL)
4689                 return;
4690         va = trunc_page(va);
4691 #ifdef PMAP_DIAGNOSTIC
4692         if (va >= KvaEnd)
4693                 panic("pmap_enter: toobig");
4694         if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4695                 panic("pmap_enter: invalid to pmap_enter page table "
4696                       "pages (va: 0x%lx)", va);
4697 #endif
4698         if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4699                 kprintf("Warning: pmap_enter called on UVA with "
4700                         "kernel_pmap\n");
4701 #ifdef DDB
4702                 db_print_backtrace();
4703 #endif
4704         }
4705         if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4706                 kprintf("Warning: pmap_enter called on KVA without"
4707                         "kernel_pmap\n");
4708 #ifdef DDB
4709                 db_print_backtrace();
4710 #endif
4711         }
4712
4713         /*
4714          * Get locked PV entries for our new page table entry (pte_pv or
4715          * pte_placemark) and for its parent page table (pt_pv).  We need
4716          * the parent so we can resolve the location of the ptep.
4717          *
4718          * Only hardware MMU actions can modify the ptep out from
4719          * under us.
4720          *
4721          * if (m) is fictitious or unmanaged we do not create a managing
4722          * pte_pv for it.  Any pre-existing page's management state must
4723          * match (avoiding code complexity).
4724          *
4725          * If the pmap is still being initialized we assume existing
4726          * page tables.
4727          *
4728          * Kernel mapppings do not track page table pages (i.e. pt_pv).
4729          *
4730          * WARNING! If replacing a managed mapping with an unmanaged mapping
4731          *          pte_pv will wind up being non-NULL and must be handled
4732          *          below.
4733          */
4734         if (pmap_initialized == FALSE) {
4735                 pte_pv = NULL;
4736                 pt_pv = NULL;
4737                 pte_placemark = NULL;
4738                 ptep = vtopte(va);
4739                 origpte = *ptep;
4740         } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4741                 pte_pv = pv_get(pmap, pmap_pte_pindex(va), &pte_placemark);
4742                 KKASSERT(pte_pv == NULL);
4743                 if (va >= VM_MAX_USER_ADDRESS) {
4744                         pt_pv = NULL;
4745                         ptep = vtopte(va);
4746                 } else {
4747                         pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4748                                                   NULL, entry, va);
4749                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4750                 }
4751                 origpte = *ptep;
4752                 cpu_ccfence();
4753                 KASSERT(origpte == 0 ||
4754                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4755                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4756         } else {
4757                 if (va >= VM_MAX_USER_ADDRESS) {
4758                         /*
4759                          * Kernel map, pv_entry-tracked.
4760                          */
4761                         pt_pv = NULL;
4762                         pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4763                         ptep = vtopte(va);
4764                 } else {
4765                         /*
4766                          * User map
4767                          */
4768                         pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4769                                                    &pt_pv, entry, va);
4770                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4771                 }
4772                 pte_placemark = NULL;   /* safety */
4773                 origpte = *ptep;
4774                 cpu_ccfence();
4775                 KASSERT(origpte == 0 ||
4776                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4777                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4778         }
4779
4780         pa = VM_PAGE_TO_PHYS(m);
4781         opa = origpte & PG_FRAME;
4782
4783         /*
4784          * Calculate the new PTE.  Note that pte_pv alone does not mean
4785          * the new pte_pv is managed, it could exist because the old pte
4786          * was managed even if the new one is not.
4787          */
4788         newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4789                  pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4790         if (wired)
4791                 newpte |= pmap->pmap_bits[PG_W_IDX];
4792         if (va < VM_MAX_USER_ADDRESS)
4793                 newpte |= pmap->pmap_bits[PG_U_IDX];
4794         if (pte_pv && (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) == 0)
4795                 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4796 //      if (pmap == &kernel_pmap)
4797 //              newpte |= pgeflag;
4798         newpte |= pmap->pmap_cache_bits[m->pat_mode];
4799         if (m->flags & PG_FICTITIOUS)
4800                 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4801
4802         /*
4803          * It is possible for multiple faults to occur in threaded
4804          * environments, the existing pte might be correct.
4805          */
4806         if (((origpte ^ newpte) &
4807             ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4808                           pmap->pmap_bits[PG_A_IDX])) == 0) {
4809                 goto done;
4810         }
4811
4812         /*
4813          * Ok, either the address changed or the protection or wiring
4814          * changed.
4815          *
4816          * Clear the current entry, interlocking the removal.  For managed
4817          * pte's this will also flush the modified state to the vm_page.
4818          * Atomic ops are mandatory in order to ensure that PG_M events are
4819          * not lost during any transition.
4820          *
4821          * WARNING: The caller has busied the new page but not the original
4822          *          vm_page which we are trying to replace.  Because we hold
4823          *          the pte_pv lock, but have not busied the page, PG bits
4824          *          can be cleared out from under us.
4825          */
4826         if (opa) {
4827                 if (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4828                         /*
4829                          * Old page was managed.  Expect pte_pv to exist.
4830                          * (it might also exist if the old page was unmanaged).
4831                          *
4832                          * NOTE: pt_pv won't exist for a kernel page
4833                          *       (managed or otherwise).
4834                          *
4835                          * NOTE: We may be reusing the pte_pv so we do not
4836                          *       destroy it in pmap_remove_pv_pte().
4837                          */
4838                         KKASSERT(pte_pv && pte_pv->pv_m);
4839                         if (prot & VM_PROT_NOSYNC) {
4840                                 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4841                         } else {
4842                                 pmap_inval_bulk_t bulk;
4843
4844                                 pmap_inval_bulk_init(&bulk, pmap);
4845                                 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4846                                 pmap_inval_bulk_flush(&bulk);
4847                         }
4848                         pmap_remove_pv_page(pte_pv);
4849                         /* will either set pte_pv->pv_m or pv_free() later */
4850                 } else {
4851                         /*
4852                          * Old page was not managed.  If we have a pte_pv
4853                          * it better not have a pv_m assigned to it.  If the
4854                          * new page is managed the pte_pv will be destroyed
4855                          * near the end (we need its interlock).
4856                          *
4857                          * NOTE: We leave the wire count on the PT page
4858                          *       intact for the followup enter, but adjust
4859                          *       the wired-pages count on the pmap.
4860                          */
4861                         KKASSERT(pte_pv == NULL);
4862                         if (prot & VM_PROT_NOSYNC) {
4863                                 /*
4864                                  * NOSYNC (no mmu sync) requested.
4865                                  */
4866                                 (void)pte_load_clear(ptep);
4867                                 cpu_invlpg((void *)va);
4868                         } else {
4869                                 /*
4870                                  * Nominal SYNC
4871                                  */
4872                                 pmap_inval_smp(pmap, va, 1, ptep, 0);
4873                         }
4874
4875                         /*
4876                          * We must adjust pm_stats manually for unmanaged
4877                          * pages.
4878                          */
4879                         if (pt_pv) {
4880                                 atomic_add_long(&pmap->pm_stats.
4881                                                 resident_count, -1);
4882                         }
4883                         if (origpte & pmap->pmap_bits[PG_W_IDX]) {
4884                                 atomic_add_long(&pmap->pm_stats.
4885                                                 wired_count, -1);
4886                         }
4887                 }
4888                 KKASSERT(*ptep == 0);
4889         }
4890
4891 #ifdef PMAP_DEBUG2
4892         if (pmap_enter_debug > 0) {
4893                 --pmap_enter_debug;
4894                 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4895                         " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4896                         va, m,
4897                         origpte, newpte, ptep,
4898                         pte_pv, pt_pv, opa, prot);
4899         }
4900 #endif
4901
4902         if ((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
4903                 /*
4904                  * Entering an unmanaged page.  We must wire the pt_pv unless
4905                  * we retained the wiring from an unmanaged page we had
4906                  * removed (if we retained it via pte_pv that will go away
4907                  * soon).
4908                  */
4909                 if (pt_pv && (opa == 0 ||
4910                               (origpte & pmap->pmap_bits[PG_MANAGED_IDX]))) {
4911                         vm_page_wire_quick(pt_pv->pv_m);
4912                 }
4913                 if (wired)
4914                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
4915
4916                 /*
4917                  * Unmanaged pages need manual resident_count tracking.
4918                  */
4919                 if (pt_pv) {
4920                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.
4921                                         resident_count, 1);
4922                 }
4923                 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4924                         vm_page_flag_set(m, PG_WRITEABLE);
4925         } else {
4926                 /*
4927                  * Entering a managed page.  Our pte_pv takes care of the
4928                  * PT wiring, so if we had removed an unmanaged page before
4929                  * we must adjust.
4930                  *
4931                  * We have to take care of the pmap wired count ourselves.
4932                  *
4933                  * Enter on the PV list if part of our managed memory.
4934                  */
4935                 KKASSERT(pte_pv && (pte_pv->pv_m == NULL || pte_pv->pv_m == m));
4936                 vm_page_spin_lock(m);
4937                 pte_pv->pv_m = m;
4938                 pmap_page_stats_adding(m);
4939                 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4940                 vm_page_flag_set(m, PG_MAPPED);
4941                 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4942                         vm_page_flag_set(m, PG_WRITEABLE);
4943                 vm_page_spin_unlock(m);
4944
4945                 if (pt_pv && opa &&
4946                     (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
4947                         vm_page_unwire_quick(pt_pv->pv_m);
4948                 }
4949
4950                 /*
4951                  * Adjust pmap wired pages count for new entry.
4952                  */
4953                 if (wired) {
4954                         atomic_add_long(&pte_pv->pv_pmap->pm_stats.
4955                                         wired_count, 1);
4956                 }
4957         }
4958
4959         /*
4960          * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4961          *
4962          * User VMAs do not because those will be zero->non-zero, so no
4963          * stale entries to worry about at this point.
4964          *
4965          * For KVM there appear to still be issues.  Theoretically we
4966          * should be able to scrap the interlocks entirely but we
4967          * get crashes.
4968          */
4969         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4970                 pmap_inval_smp(pmap, va, 1, ptep, newpte);
4971         } else {
4972                 origpte = atomic_swap_long(ptep, newpte);
4973                 if (origpte & pmap->pmap_bits[PG_M_IDX]) {
4974                         kprintf("pmap [M] race @ %016jx\n", va);
4975                         atomic_set_long(ptep, pmap->pmap_bits[PG_M_IDX]);
4976                 }
4977                 if (pt_pv == NULL)
4978                         cpu_invlpg((void *)va);
4979         }
4980
4981         /*
4982          * Cleanup
4983          */
4984 done:
4985         KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4986                  (m->flags & PG_MAPPED));
4987
4988         /*
4989          * Cleanup the pv entry, allowing other accessors.  If the new page
4990          * is not managed but we have a pte_pv (which was locking our
4991          * operation), we can free it now.  pte_pv->pv_m should be NULL.
4992          */
4993         if (pte_pv && (newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
4994                 pv_free(pte_pv, pt_pv);
4995         } else if (pte_pv) {
4996                 pv_put(pte_pv);
4997         } else if (pte_placemark) {
4998                 pv_placemarker_wakeup(pmap, pte_placemark);
4999         }
5000         if (pt_pv)
5001                 pv_put(pt_pv);
5002 }
5003
5004 /*
5005  * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
5006  * This code also assumes that the pmap has no pre-existing entry for this
5007  * VA.
5008  *
5009  * This code currently may only be used on user pmaps, not kernel_pmap.
5010  */
5011 void
5012 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
5013 {
5014         pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
5015 }
5016
5017 /*
5018  * Make a temporary mapping for a physical address.  This is only intended
5019  * to be used for panic dumps.
5020  *
5021  * The caller is responsible for calling smp_invltlb().
5022  */
5023 void *
5024 pmap_kenter_temporary(vm_paddr_t pa, long i)
5025 {
5026         pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
5027         return ((void *)crashdumpmap);
5028 }
5029
5030 #define MAX_INIT_PT (96)
5031
5032 /*
5033  * This routine preloads the ptes for a given object into the specified pmap.
5034  * This eliminates the blast of soft faults on process startup and
5035  * immediately after an mmap.
5036  */
5037 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
5038
5039 void
5040 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
5041                     vm_object_t object, vm_pindex_t pindex,
5042                     vm_size_t size, int limit)
5043 {
5044         struct rb_vm_page_scan_info info;
5045         struct lwp *lp;
5046         vm_size_t psize;
5047
5048         /*
5049          * We can't preinit if read access isn't set or there is no pmap
5050          * or object.
5051          */
5052         if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
5053                 return;
5054
5055         /*
5056          * We can't preinit if the pmap is not the current pmap
5057          */
5058         lp = curthread->td_lwp;
5059         if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
5060                 return;
5061
5062         /*
5063          * Misc additional checks
5064          */
5065         psize = x86_64_btop(size);
5066
5067         if ((object->type != OBJT_VNODE) ||
5068                 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
5069                         (object->resident_page_count > MAX_INIT_PT))) {
5070                 return;
5071         }
5072
5073         if (pindex + psize > object->size) {
5074                 if (object->size < pindex)
5075                         return;
5076                 psize = object->size - pindex;
5077         }
5078
5079         if (psize == 0)
5080                 return;
5081
5082         /*
5083          * If everything is segment-aligned do not pre-init here.  Instead
5084          * allow the normal vm_fault path to pass a segment hint to
5085          * pmap_enter() which will then use an object-referenced shared
5086          * page table page.
5087          */
5088         if ((addr & SEG_MASK) == 0 &&
5089             (ctob(psize) & SEG_MASK) == 0 &&
5090             (ctob(pindex) & SEG_MASK) == 0) {
5091                 return;
5092         }
5093
5094         /*
5095          * Use a red-black scan to traverse the requested range and load
5096          * any valid pages found into the pmap.
5097          *
5098          * We cannot safely scan the object's memq without holding the
5099          * object token.
5100          */
5101         info.start_pindex = pindex;
5102         info.end_pindex = pindex + psize - 1;
5103         info.limit = limit;
5104         info.mpte = NULL;
5105         info.addr = addr;
5106         info.pmap = pmap;
5107
5108         vm_object_hold_shared(object);
5109         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
5110                                 pmap_object_init_pt_callback, &info);
5111         vm_object_drop(object);
5112 }
5113
5114 static
5115 int
5116 pmap_object_init_pt_callback(vm_page_t p, void *data)
5117 {
5118         struct rb_vm_page_scan_info *info = data;
5119         vm_pindex_t rel_index;
5120
5121         /*
5122          * don't allow an madvise to blow away our really
5123          * free pages allocating pv entries.
5124          */
5125         if ((info->limit & MAP_PREFAULT_MADVISE) &&
5126                 vmstats.v_free_count < vmstats.v_free_reserved) {
5127                     return(-1);
5128         }
5129
5130         /*
5131          * Ignore list markers and ignore pages we cannot instantly
5132          * busy (while holding the object token).
5133          */
5134         if (p->flags & PG_MARKER)
5135                 return 0;
5136         if (vm_page_busy_try(p, TRUE))
5137                 return 0;
5138         if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
5139             (p->flags & PG_FICTITIOUS) == 0) {
5140                 if ((p->queue - p->pc) == PQ_CACHE)
5141                         vm_page_deactivate(p);
5142                 rel_index = p->pindex - info->start_pindex;
5143                 pmap_enter_quick(info->pmap,
5144                                  info->addr + x86_64_ptob(rel_index), p);
5145         }
5146         vm_page_wakeup(p);
5147         lwkt_yield();
5148         return(0);
5149 }
5150
5151 /*
5152  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
5153  * address.
5154  *
5155  * Returns FALSE if it would be non-trivial or if a pte is already loaded
5156  * into the slot.
5157  *
5158  * XXX This is safe only because page table pages are not freed.
5159  */
5160 int
5161 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
5162 {
5163         pt_entry_t *pte;
5164
5165         /*spin_lock(&pmap->pm_spin);*/
5166         if ((pte = pmap_pte(pmap, addr)) != NULL) {
5167                 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
5168                         /*spin_unlock(&pmap->pm_spin);*/
5169                         return FALSE;
5170                 }
5171         }
5172         /*spin_unlock(&pmap->pm_spin);*/
5173         return TRUE;
5174 }
5175
5176 /*
5177  * Change the wiring attribute for a pmap/va pair.  The mapping must already
5178  * exist in the pmap.  The mapping may or may not be managed.  The wiring in
5179  * the page is not changed, the page is returned so the caller can adjust
5180  * its wiring (the page is not locked in any way).
5181  *
5182  * Wiring is not a hardware characteristic so there is no need to invalidate
5183  * TLB.  However, in an SMP environment we must use a locked bus cycle to
5184  * update the pte (if we are not using the pmap_inval_*() API that is)...
5185  * it's ok to do this for simple wiring changes.
5186  */
5187 vm_page_t
5188 pmap_unwire(pmap_t pmap, vm_offset_t va)
5189 {
5190         pt_entry_t *ptep;
5191         pv_entry_t pt_pv;
5192         vm_paddr_t pa;
5193         vm_page_t m;
5194
5195         if (pmap == NULL)
5196                 return NULL;
5197
5198         /*
5199          * Assume elements in the kernel pmap are stable
5200          */
5201         if (pmap == &kernel_pmap) {
5202                 if (pmap_pt(pmap, va) == 0)
5203                         return NULL;
5204                 ptep = pmap_pte_quick(pmap, va);
5205                 if (pmap_pte_v(pmap, ptep)) {
5206                         if (pmap_pte_w(pmap, ptep))
5207                                 atomic_add_long(&pmap->pm_stats.wired_count,-1);
5208                         atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5209                         pa = *ptep & PG_FRAME;
5210                         m = PHYS_TO_VM_PAGE(pa);
5211                 } else {
5212                         m = NULL;
5213                 }
5214         } else {
5215                 /*
5216                  * We can only [un]wire pmap-local pages (we cannot wire
5217                  * shared pages)
5218                  */
5219                 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
5220                 if (pt_pv == NULL)
5221                         return NULL;
5222
5223                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
5224                 if ((*ptep & pmap->pmap_bits[PG_V_IDX]) == 0) {
5225                         pv_put(pt_pv);
5226                         return NULL;
5227                 }
5228
5229                 if (pmap_pte_w(pmap, ptep)) {
5230                         atomic_add_long(&pt_pv->pv_pmap->pm_stats.wired_count,
5231                                         -1);
5232                 }
5233                 /* XXX else return NULL so caller doesn't unwire m ? */
5234
5235                 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5236
5237                 pa = *ptep & PG_FRAME;
5238                 m = PHYS_TO_VM_PAGE(pa);        /* held by wired count */
5239                 pv_put(pt_pv);
5240         }
5241         return m;
5242 }
5243
5244 /*
5245  * Copy the range specified by src_addr/len from the source map to
5246  * the range dst_addr/len in the destination map.
5247  *
5248  * This routine is only advisory and need not do anything.
5249  */
5250 void
5251 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
5252           vm_size_t len, vm_offset_t src_addr)
5253 {
5254 }
5255
5256 /*
5257  * pmap_zero_page:
5258  *
5259  *      Zero the specified physical page.
5260  *
5261  *      This function may be called from an interrupt and no locking is
5262  *      required.
5263  */
5264 void
5265 pmap_zero_page(vm_paddr_t phys)
5266 {
5267         vm_offset_t va = PHYS_TO_DMAP(phys);
5268
5269         pagezero((void *)va);
5270 }
5271
5272 /*
5273  * pmap_zero_page:
5274  *
5275  *      Zero part of a physical page by mapping it into memory and clearing
5276  *      its contents with bzero.
5277  *
5278  *      off and size may not cover an area beyond a single hardware page.
5279  */
5280 void
5281 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
5282 {
5283         vm_offset_t virt = PHYS_TO_DMAP(phys);
5284
5285         bzero((char *)virt + off, size);
5286 }
5287
5288 /*
5289  * pmap_copy_page:
5290  *
5291  *      Copy the physical page from the source PA to the target PA.
5292  *      This function may be called from an interrupt.  No locking
5293  *      is required.
5294  */
5295 void
5296 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
5297 {
5298         vm_offset_t src_virt, dst_virt;
5299
5300         src_virt = PHYS_TO_DMAP(src);
5301         dst_virt = PHYS_TO_DMAP(dst);
5302         bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
5303 }
5304
5305 /*
5306  * pmap_copy_page_frag:
5307  *
5308  *      Copy the physical page from the source PA to the target PA.
5309  *      This function may be called from an interrupt.  No locking
5310  *      is required.
5311  */
5312 void
5313 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
5314 {
5315         vm_offset_t src_virt, dst_virt;
5316
5317         src_virt = PHYS_TO_DMAP(src);
5318         dst_virt = PHYS_TO_DMAP(dst);
5319
5320         bcopy((char *)src_virt + (src & PAGE_MASK),
5321               (char *)dst_virt + (dst & PAGE_MASK),
5322               bytes);
5323 }
5324
5325 /*
5326  * Returns true if the pmap's pv is one of the first 16 pvs linked to from
5327  * this page.  This count may be changed upwards or downwards in the future;
5328  * it is only necessary that true be returned for a small subset of pmaps
5329  * for proper page aging.
5330  */
5331 boolean_t
5332 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
5333 {
5334         pv_entry_t pv;
5335         int loops = 0;
5336
5337         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5338                 return FALSE;
5339
5340         vm_page_spin_lock(m);
5341         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5342                 if (pv->pv_pmap == pmap) {
5343                         vm_page_spin_unlock(m);
5344                         return TRUE;
5345                 }
5346                 loops++;
5347                 if (loops >= 16)
5348                         break;
5349         }
5350         vm_page_spin_unlock(m);
5351         return (FALSE);
5352 }
5353
5354 /*
5355  * Remove all pages from specified address space this aids process exit
5356  * speeds.  Also, this code may be special cased for the current process
5357  * only.
5358  */
5359 void
5360 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5361 {
5362         pmap_remove_noinval(pmap, sva, eva);
5363         cpu_invltlb();
5364 }
5365
5366 /*
5367  * pmap_testbit tests bits in pte's note that the testbit/clearbit
5368  * routines are inline, and a lot of things compile-time evaluate.
5369  */
5370 static
5371 boolean_t
5372 pmap_testbit(vm_page_t m, int bit)
5373 {
5374         pv_entry_t pv;
5375         pt_entry_t *pte;
5376         pmap_t pmap;
5377
5378         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5379                 return FALSE;
5380
5381         if (TAILQ_FIRST(&m->md.pv_list) == NULL)
5382                 return FALSE;
5383         vm_page_spin_lock(m);
5384         if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
5385                 vm_page_spin_unlock(m);
5386                 return FALSE;
5387         }
5388
5389         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5390
5391 #if defined(PMAP_DIAGNOSTIC)
5392                 if (pv->pv_pmap == NULL) {
5393                         kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
5394                             pv->pv_pindex);
5395                         continue;
5396                 }
5397 #endif
5398                 pmap = pv->pv_pmap;
5399
5400                 /*
5401                  * If the bit being tested is the modified bit, then
5402                  * mark clean_map and ptes as never
5403                  * modified.
5404                  *
5405                  * WARNING!  Because we do not lock the pv, *pte can be in a
5406                  *           state of flux.  Despite this the value of *pte
5407                  *           will still be related to the vm_page in some way
5408                  *           because the pv cannot be destroyed as long as we
5409                  *           hold the vm_page spin lock.
5410                  */
5411                 if (bit == PG_A_IDX || bit == PG_M_IDX) {
5412                                 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5413                         if (!pmap_track_modified(pv->pv_pindex))
5414                                 continue;
5415                 }
5416
5417                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5418                 if (*pte & pmap->pmap_bits[bit]) {
5419                         vm_page_spin_unlock(m);
5420                         return TRUE;
5421                 }
5422         }
5423         vm_page_spin_unlock(m);
5424         return (FALSE);
5425 }
5426
5427 /*
5428  * This routine is used to modify bits in ptes.  Only one bit should be
5429  * specified.  PG_RW requires special handling.
5430  *
5431  * Caller must NOT hold any spin locks
5432  */
5433 static __inline
5434 void
5435 pmap_clearbit(vm_page_t m, int bit_index)
5436 {
5437         pv_entry_t pv;
5438         pt_entry_t *pte;
5439         pt_entry_t pbits;
5440         pmap_t pmap;
5441
5442         if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5443                 if (bit_index == PG_RW_IDX)
5444                         vm_page_flag_clear(m, PG_WRITEABLE);
5445                 return;
5446         }
5447
5448         /*
5449          * PG_M or PG_A case
5450          *
5451          * Loop over all current mappings setting/clearing as appropos If
5452          * setting RO do we need to clear the VAC?
5453          *
5454          * NOTE: When clearing PG_M we could also (not implemented) drop
5455          *       through to the PG_RW code and clear PG_RW too, forcing
5456          *       a fault on write to redetect PG_M for virtual kernels, but
5457          *       it isn't necessary since virtual kernels invalidate the
5458          *       pte when they clear the VPTE_M bit in their virtual page
5459          *       tables.
5460          *
5461          * NOTE: Does not re-dirty the page when clearing only PG_M.
5462          *
5463          * NOTE: Because we do not lock the pv, *pte can be in a state of
5464          *       flux.  Despite this the value of *pte is still somewhat
5465          *       related while we hold the vm_page spin lock.
5466          *
5467          *       *pte can be zero due to this race.  Since we are clearing
5468          *       bits we basically do no harm when this race occurs.
5469          */
5470         if (bit_index != PG_RW_IDX) {
5471                 vm_page_spin_lock(m);
5472                 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5473 #if defined(PMAP_DIAGNOSTIC)
5474                         if (pv->pv_pmap == NULL) {
5475                                 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5476                                     pv->pv_pindex);
5477                                 continue;
5478                         }
5479 #endif
5480                         pmap = pv->pv_pmap;
5481                         pte = pmap_pte_quick(pv->pv_pmap,
5482                                              pv->pv_pindex << PAGE_SHIFT);
5483                         pbits = *pte;
5484                         if (pbits & pmap->pmap_bits[bit_index])
5485                                 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5486                 }
5487                 vm_page_spin_unlock(m);
5488                 return;
5489         }
5490
5491         /*
5492          * Clear PG_RW.  Also clears PG_M and marks the page dirty if PG_M
5493          * was set.
5494          */
5495 restart:
5496         vm_page_spin_lock(m);
5497         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5498                 /*
5499                  * don't write protect pager mappings
5500                  */
5501                 if (!pmap_track_modified(pv->pv_pindex))
5502                         continue;
5503
5504 #if defined(PMAP_DIAGNOSTIC)
5505                 if (pv->pv_pmap == NULL) {
5506                         kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5507                                 pv->pv_pindex);
5508                         continue;
5509                 }
5510 #endif
5511                 pmap = pv->pv_pmap;
5512
5513                 /*
5514                  * Skip pages which do not have PG_RW set.
5515                  */
5516                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5517                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5518                         continue;
5519
5520                 /*
5521                  * We must lock the PV to be able to safely test the pte.
5522                  */
5523                 if (pv_hold_try(pv)) {
5524                         vm_page_spin_unlock(m);
5525                 } else {
5526                         vm_page_spin_unlock(m);
5527                         pv_lock(pv);    /* held, now do a blocking lock */
5528                         pv_put(pv);
5529                         goto restart;
5530                 }
5531
5532                 /*
5533                  * Reload pte after acquiring pv.
5534                  */
5535                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5536 #if 0
5537                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0) {
5538                         pv_put(pv);
5539                         goto restart;
5540                 }
5541 #endif
5542
5543                 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
5544                 for (;;) {
5545                         pt_entry_t nbits;
5546
5547                         pbits = *pte;
5548                         cpu_ccfence();
5549                         nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5550                                           pmap->pmap_bits[PG_M_IDX]);
5551                         if (pmap_inval_smp_cmpset(pmap,
5552                                      ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5553                                      pte, pbits, nbits)) {
5554                                 break;
5555                         }
5556                         cpu_pause();
5557                 }
5558
5559                 /*
5560                  * If PG_M was found to be set while we were clearing PG_RW
5561                  * we also clear PG_M (done above) and mark the page dirty.
5562                  * Callers expect this behavior.
5563                  *
5564                  * we lost pv so it cannot be used as an iterator.  In fact,
5565                  * because we couldn't necessarily lock it atomically it may
5566                  * have moved within the list and ALSO cannot be used as an
5567                  * iterator.
5568                  */
5569                 vm_page_spin_lock(m);
5570                 if (pbits & pmap->pmap_bits[PG_M_IDX])
5571                         vm_page_dirty(m);
5572                 vm_page_spin_unlock(m);
5573                 pv_put(pv);
5574                 goto restart;
5575         }
5576         if (bit_index == PG_RW_IDX)
5577                 vm_page_flag_clear(m, PG_WRITEABLE);
5578         vm_page_spin_unlock(m);
5579 }
5580
5581 /*
5582  * Lower the permission for all mappings to a given page.
5583  *
5584  * Page must be busied by caller.  Because page is busied by caller this
5585  * should not be able to race a pmap_enter().
5586  */
5587 void
5588 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5589 {
5590         /* JG NX support? */
5591         if ((prot & VM_PROT_WRITE) == 0) {
5592                 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5593                         /*
5594                          * NOTE: pmap_clearbit(.. PG_RW) also clears
5595                          *       the PG_WRITEABLE flag in (m).
5596                          */
5597                         pmap_clearbit(m, PG_RW_IDX);
5598                 } else {
5599                         pmap_remove_all(m);
5600                 }
5601         }
5602 }
5603
5604 vm_paddr_t
5605 pmap_phys_address(vm_pindex_t ppn)
5606 {
5607         return (x86_64_ptob(ppn));
5608 }
5609
5610 /*
5611  * Return a count of reference bits for a page, clearing those bits.
5612  * It is not necessary for every reference bit to be cleared, but it
5613  * is necessary that 0 only be returned when there are truly no
5614  * reference bits set.
5615  *
5616  * XXX: The exact number of bits to check and clear is a matter that
5617  * should be tested and standardized at some point in the future for
5618  * optimal aging of shared pages.
5619  *
5620  * This routine may not block.
5621  */
5622 int
5623 pmap_ts_referenced(vm_page_t m)
5624 {
5625         pv_entry_t pv;
5626         pt_entry_t *pte;
5627         pmap_t pmap;
5628         int rtval = 0;
5629
5630         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5631                 return (rtval);
5632
5633         vm_page_spin_lock(m);
5634         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5635                 if (!pmap_track_modified(pv->pv_pindex))
5636                         continue;
5637                 pmap = pv->pv_pmap;
5638                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5639                 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5640                         atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5641                         rtval++;
5642                         if (rtval > 4)
5643                                 break;
5644                 }
5645         }
5646         vm_page_spin_unlock(m);
5647         return (rtval);
5648 }
5649
5650 /*
5651  *      pmap_is_modified:
5652  *
5653  *      Return whether or not the specified physical page was modified
5654  *      in any physical maps.
5655  */
5656 boolean_t
5657 pmap_is_modified(vm_page_t m)
5658 {
5659         boolean_t res;
5660
5661         res = pmap_testbit(m, PG_M_IDX);
5662         return (res);
5663 }
5664
5665 /*
5666  *      Clear the modify bits on the specified physical page.
5667  */
5668 void
5669 pmap_clear_modify(vm_page_t m)
5670 {
5671         pmap_clearbit(m, PG_M_IDX);
5672 }
5673
5674 /*
5675  *      pmap_clear_reference:
5676  *
5677  *      Clear the reference bit on the specified physical page.
5678  */
5679 void
5680 pmap_clear_reference(vm_page_t m)
5681 {
5682         pmap_clearbit(m, PG_A_IDX);
5683 }
5684
5685 /*
5686  * Miscellaneous support routines follow
5687  */
5688
5689 static
5690 void
5691 i386_protection_init(void)
5692 {
5693         int *kp, prot;
5694
5695         /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit  */
5696         kp = protection_codes;
5697         for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5698                 switch (prot) {
5699                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5700                         /*
5701                          * Read access is also 0. There isn't any execute bit,
5702                          * so just make it readable.
5703                          */
5704                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5705                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5706                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5707                         *kp++ = 0;
5708                         break;
5709                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5710                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5711                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5712                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5713                         *kp++ = pmap_bits_default[PG_RW_IDX];
5714                         break;
5715                 }
5716         }
5717 }
5718
5719 /*
5720  * Map a set of physical memory pages into the kernel virtual
5721  * address space. Return a pointer to where it is mapped. This
5722  * routine is intended to be used for mapping device memory,
5723  * NOT real memory.
5724  *
5725  * NOTE: We can't use pgeflag unless we invalidate the pages one at
5726  *       a time.
5727  *
5728  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5729  *       work whether the cpu supports PAT or not.  The remaining PAT
5730  *       attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5731  *       supports PAT.
5732  */
5733 void *
5734 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5735 {
5736         return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5737 }
5738
5739 void *
5740 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5741 {
5742         return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5743 }
5744
5745 void *
5746 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5747 {
5748         return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5749 }
5750
5751 /*
5752  * Map a set of physical memory pages into the kernel virtual
5753  * address space. Return a pointer to where it is mapped. This
5754  * routine is intended to be used for mapping device memory,
5755  * NOT real memory.
5756  */
5757 void *
5758 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5759 {
5760         vm_offset_t va, tmpva, offset;
5761         pt_entry_t *pte;
5762         vm_size_t tmpsize;
5763
5764         offset = pa & PAGE_MASK;
5765         size = roundup(offset + size, PAGE_SIZE);
5766
5767         va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5768         if (va == 0)
5769                 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5770
5771         pa = pa & ~PAGE_MASK;
5772         for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5773                 pte = vtopte(tmpva);
5774                 *pte = pa |
5775                     kernel_pmap.pmap_bits[PG_RW_IDX] |
5776                     kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5777                     kernel_pmap.pmap_cache_bits[mode];
5778                 tmpsize -= PAGE_SIZE;
5779                 tmpva += PAGE_SIZE;
5780                 pa += PAGE_SIZE;
5781         }
5782         pmap_invalidate_range(&kernel_pmap, va, va + size);
5783         pmap_invalidate_cache_range(va, va + size);
5784
5785         return ((void *)(va + offset));
5786 }
5787
5788 void
5789 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5790 {
5791         vm_offset_t base, offset;
5792
5793         base = va & ~PAGE_MASK;
5794         offset = va & PAGE_MASK;
5795         size = roundup(offset + size, PAGE_SIZE);
5796         pmap_qremove(va, size >> PAGE_SHIFT);
5797         kmem_free(&kernel_map, base, size);
5798 }
5799
5800 /*
5801  * Sets the memory attribute for the specified page.
5802  */
5803 void
5804 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5805 {
5806
5807     m->pat_mode = ma;
5808
5809     /*
5810      * If "m" is a normal page, update its direct mapping.  This update
5811      * can be relied upon to perform any cache operations that are
5812      * required for data coherence.
5813      */
5814     if ((m->flags & PG_FICTITIOUS) == 0)
5815         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5816 }
5817
5818 /*
5819  * Change the PAT attribute on an existing kernel memory map.  Caller
5820  * must ensure that the virtual memory in question is not accessed
5821  * during the adjustment.
5822  */
5823 void
5824 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5825 {
5826         pt_entry_t *pte;
5827         vm_offset_t base;
5828         int changed = 0;
5829
5830         if (va == 0)
5831                 panic("pmap_change_attr: va is NULL");
5832         base = trunc_page(va);
5833
5834         while (count) {
5835                 pte = vtopte(va);
5836                 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5837                        kernel_pmap.pmap_cache_bits[mode];
5838                 --count;
5839                 va += PAGE_SIZE;
5840         }
5841
5842         changed = 1;    /* XXX: not optimal */
5843
5844         /*
5845          * Flush CPU caches if required to make sure any data isn't cached that
5846          * shouldn't be, etc.
5847          */
5848         if (changed) {
5849                 pmap_invalidate_range(&kernel_pmap, base, va);
5850                 pmap_invalidate_cache_range(base, va);
5851         }
5852 }
5853
5854 /*
5855  * perform the pmap work for mincore
5856  */
5857 int
5858 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5859 {
5860         pt_entry_t *ptep, pte;
5861         vm_page_t m;
5862         int val = 0;
5863
5864         ptep = pmap_pte(pmap, addr);
5865
5866         if (ptep && (pte = *ptep) != 0) {
5867                 vm_offset_t pa;
5868
5869                 val = MINCORE_INCORE;
5870                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5871                         goto done;
5872
5873                 pa = pte & PG_FRAME;
5874
5875                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5876                         m = NULL;
5877                 else
5878                         m = PHYS_TO_VM_PAGE(pa);
5879
5880                 /*
5881                  * Modified by us
5882                  */
5883                 if (pte & pmap->pmap_bits[PG_M_IDX])
5884                         val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5885                 /*
5886                  * Modified by someone
5887                  */
5888                 else if (m && (m->dirty || pmap_is_modified(m)))
5889                         val |= MINCORE_MODIFIED_OTHER;
5890                 /*
5891                  * Referenced by us
5892                  */
5893                 if (pte & pmap->pmap_bits[PG_A_IDX])
5894                         val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5895
5896                 /*
5897                  * Referenced by someone
5898                  */
5899                 else if (m && ((m->flags & PG_REFERENCED) ||
5900                                 pmap_ts_referenced(m))) {
5901                         val |= MINCORE_REFERENCED_OTHER;
5902                         vm_page_flag_set(m, PG_REFERENCED);
5903                 }
5904         }
5905 done:
5906
5907         return val;
5908 }
5909
5910 /*
5911  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
5912  * vmspace will be ref'd and the old one will be deref'd.
5913  *
5914  * The vmspace for all lwps associated with the process will be adjusted
5915  * and cr3 will be reloaded if any lwp is the current lwp.
5916  *
5917  * The process must hold the vmspace->vm_map.token for oldvm and newvm
5918  */
5919 void
5920 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5921 {
5922         struct vmspace *oldvm;
5923         struct lwp *lp;
5924
5925         oldvm = p->p_vmspace;
5926         if (oldvm != newvm) {
5927                 if (adjrefs)
5928                         vmspace_ref(newvm);
5929                 p->p_vmspace = newvm;
5930                 KKASSERT(p->p_nthreads == 1);
5931                 lp = RB_ROOT(&p->p_lwp_tree);
5932                 pmap_setlwpvm(lp, newvm);
5933                 if (adjrefs)
5934                         vmspace_rel(oldvm);
5935         }
5936 }
5937
5938 /*
5939  * Set the vmspace for a LWP.  The vmspace is almost universally set the
5940  * same as the process vmspace, but virtual kernels need to swap out contexts
5941  * on a per-lwp basis.
5942  *
5943  * Caller does not necessarily hold any vmspace tokens.  Caller must control
5944  * the lwp (typically be in the context of the lwp).  We use a critical
5945  * section to protect against statclock and hardclock (statistics collection).
5946  */
5947 void
5948 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5949 {
5950         struct vmspace *oldvm;
5951         struct pmap *pmap;
5952
5953         oldvm = lp->lwp_vmspace;
5954
5955         if (oldvm != newvm) {
5956                 crit_enter();
5957                 KKASSERT((newvm->vm_refcnt & VM_REF_DELETED) == 0);
5958                 lp->lwp_vmspace = newvm;
5959                 if (curthread->td_lwp == lp) {
5960                         pmap = vmspace_pmap(newvm);
5961                         ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5962                         if (pmap->pm_active_lock & CPULOCK_EXCL)
5963                                 pmap_interlock_wait(newvm);
5964 #if defined(SWTCH_OPTIM_STATS)
5965                         tlb_flush_count++;
5966 #endif
5967                         if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5968                                 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5969                         } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5970                                 curthread->td_pcb->pcb_cr3 = KPML4phys;
5971                         } else {
5972                                 panic("pmap_setlwpvm: unknown pmap type\n");
5973                         }
5974                         load_cr3(curthread->td_pcb->pcb_cr3);
5975                         pmap = vmspace_pmap(oldvm);
5976                         ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5977                                                mycpu->gd_cpuid);
5978                 }
5979                 crit_exit();
5980         }
5981 }
5982
5983 /*
5984  * Called when switching to a locked pmap, used to interlock against pmaps
5985  * undergoing modifications to prevent us from activating the MMU for the
5986  * target pmap until all such modifications have completed.  We have to do
5987  * this because the thread making the modifications has already set up its
5988  * SMP synchronization mask.
5989  *
5990  * This function cannot sleep!
5991  *
5992  * No requirements.
5993  */
5994 void
5995 pmap_interlock_wait(struct vmspace *vm)
5996 {
5997         struct pmap *pmap = &vm->vm_pmap;
5998
5999         if (pmap->pm_active_lock & CPULOCK_EXCL) {
6000                 crit_enter();
6001                 KKASSERT(curthread->td_critcount >= 2);
6002                 DEBUG_PUSH_INFO("pmap_interlock_wait");
6003                 while (pmap->pm_active_lock & CPULOCK_EXCL) {
6004                         cpu_ccfence();
6005                         lwkt_process_ipiq();
6006                 }
6007                 DEBUG_POP_INFO();
6008                 crit_exit();
6009         }
6010 }
6011
6012 vm_offset_t
6013 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
6014 {
6015
6016         if ((obj == NULL) || (size < NBPDR) ||
6017             ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
6018                 return addr;
6019         }
6020
6021         addr = roundup2(addr, NBPDR);
6022         return addr;
6023 }
6024
6025 /*
6026  * Used by kmalloc/kfree, page already exists at va
6027  */
6028 vm_page_t
6029 pmap_kvtom(vm_offset_t va)
6030 {
6031         pt_entry_t *ptep = vtopte(va);
6032
6033         KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
6034         return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
6035 }
6036
6037 /*
6038  * Initialize machine-specific shared page directory support.  This
6039  * is executed when a VM object is created.
6040  */
6041 void
6042 pmap_object_init(vm_object_t object)
6043 {
6044         object->md.pmap_rw = NULL;
6045         object->md.pmap_ro = NULL;
6046 }
6047
6048 /*
6049  * Clean up machine-specific shared page directory support.  This
6050  * is executed when a VM object is destroyed.
6051  */
6052 void
6053 pmap_object_free(vm_object_t object)
6054 {
6055         pmap_t pmap;
6056
6057         if ((pmap = object->md.pmap_rw) != NULL) {
6058                 object->md.pmap_rw = NULL;
6059                 pmap_remove_noinval(pmap,
6060                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6061                 CPUMASK_ASSZERO(pmap->pm_active);
6062                 pmap_release(pmap);
6063                 pmap_puninit(pmap);
6064                 kfree(pmap, M_OBJPMAP);
6065         }
6066         if ((pmap = object->md.pmap_ro) != NULL) {
6067                 object->md.pmap_ro = NULL;
6068                 pmap_remove_noinval(pmap,
6069                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6070                 CPUMASK_ASSZERO(pmap->pm_active);
6071                 pmap_release(pmap);
6072                 pmap_puninit(pmap);
6073                 kfree(pmap, M_OBJPMAP);
6074         }
6075 }
6076
6077 /*
6078  * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
6079  * VM page and issue a pginfo->callback.
6080  *
6081  * We are expected to dispose of any non-NULL pte_pv.
6082  */
6083 static
6084 void
6085 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
6086                       pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
6087                       pv_entry_t pt_pv, int sharept,
6088                       vm_offset_t va, pt_entry_t *ptep, void *arg)
6089 {
6090         struct pmap_pgscan_info *pginfo = arg;
6091         vm_page_t m;
6092
6093         if (pte_pv) {
6094                 /*
6095                  * Try to busy the page while we hold the pte_pv locked.
6096                  */
6097                 KKASSERT(pte_pv->pv_m);
6098                 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
6099                 if (vm_page_busy_try(m, TRUE) == 0) {
6100                         if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
6101                                 /*
6102                                  * The callback is issued with the pte_pv
6103                                  * unlocked and put away, and the pt_pv
6104                                  * unlocked.
6105                                  */
6106                                 pv_put(pte_pv);
6107                                 if (pt_pv) {
6108                                         vm_page_wire_quick(pt_pv->pv_m);
6109                                         pv_unlock(pt_pv);
6110                                 }
6111                                 if (pginfo->callback(pginfo, va, m) < 0)
6112                                         info->stop = 1;
6113                                 if (pt_pv) {
6114                                         pv_lock(pt_pv);
6115                                         vm_page_unwire_quick(pt_pv->pv_m);
6116                                 }
6117                         } else {
6118                                 vm_page_wakeup(m);
6119                                 pv_put(pte_pv);
6120                         }
6121                 } else {
6122                         ++pginfo->busycount;
6123                         pv_put(pte_pv);
6124                 }
6125         } else {
6126                 /*
6127                  * Shared page table or unmanaged page (sharept or !sharept)
6128                  */
6129                 pv_placemarker_wakeup(pmap, pte_placemark);
6130         }
6131 }
6132
6133 void
6134 pmap_pgscan(struct pmap_pgscan_info *pginfo)
6135 {
6136         struct pmap_scan_info info;
6137
6138         pginfo->offset = pginfo->beg_addr;
6139         info.pmap = pginfo->pmap;
6140         info.sva = pginfo->beg_addr;
6141         info.eva = pginfo->end_addr;
6142         info.func = pmap_pgscan_callback;
6143         info.arg = pginfo;
6144         pmap_scan(&info, 0);
6145         if (info.stop == 0)
6146                 pginfo->offset = pginfo->end_addr;
6147 }
6148
6149 /*
6150  * Wait for a placemarker that we do not own to clear.  The placemarker
6151  * in question is not necessarily set to the pindex we want, we may have
6152  * to wait on the element because we want to reserve it ourselves.
6153  *
6154  * NOTE: PM_PLACEMARK_WAKEUP sets a bit which is already set in
6155  *       PM_NOPLACEMARK, so it does not interfere with placemarks
6156  *       which have already been woken up.
6157  */
6158 static
6159 void
6160 pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark)
6161 {
6162         if (*pmark != PM_NOPLACEMARK) {
6163                 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
6164                 tsleep_interlock(pmark, 0);
6165                 if (*pmark != PM_NOPLACEMARK)
6166                         tsleep(pmark, PINTERLOCKED, "pvplw", 0);
6167         }
6168 }
6169
6170 /*
6171  * Wakeup a placemarker that we own.  Replace the entry with
6172  * PM_NOPLACEMARK and issue a wakeup() if necessary.
6173  */
6174 static
6175 void
6176 pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark)
6177 {
6178         vm_pindex_t pindex;
6179
6180         pindex = atomic_swap_long(pmark, PM_NOPLACEMARK);
6181         KKASSERT(pindex != PM_NOPLACEMARK);
6182         if (pindex & PM_PLACEMARK_WAKEUP)
6183                 wakeup(pmark);
6184 }