sys/platform/pc64/x86_64/pmap.c

   1 /*
   2  * Copyright (c) 1991 Regents of the University of California.
   3  * Copyright (c) 1994 John S. Dyson
   4  * Copyright (c) 1994 David Greenman
   5  * Copyright (c) 2003 Peter Wemm
   6  * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
   7  * Copyright (c) 2008, 2009 The DragonFly Project.
   8  * Copyright (c) 2008, 2009 Jordan Gordeev.
   9  * Copyright (c) 2011-2012 Matthew Dillon
  10  * All rights reserved.
  11  *
  12  * This code is derived from software contributed to Berkeley by
  13  * the Systems Programming Group of the University of Utah Computer
  14  * Science Department and William Jolitz of UUNET Technologies Inc.
  15  *
  16  * Redistribution and use in source and binary forms, with or without
  17  * modification, are permitted provided that the following conditions
  18  * are met:
  19  * 1. Redistributions of source code must retain the above copyright
  20  *    notice, this list of conditions and the following disclaimer.
  21  * 2. Redistributions in binary form must reproduce the above copyright
  22  *    notice, this list of conditions and the following disclaimer in the
  23  *    documentation and/or other materials provided with the distribution.
  24  * 3. All advertising materials mentioning features or use of this software
  25  *    must display the following acknowledgement:
  26  *      This product includes software developed by the University of
  27  *      California, Berkeley and its contributors.
  28  * 4. Neither the name of the University nor the names of its contributors
  29  *    may be used to endorse or promote products derived from this software
  30  *    without specific prior written permission.
  31  *
  32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  42  * SUCH DAMAGE.
  43  */
  44 /*
  45  * Manage physical address maps for x86-64 systems.
  46  */
  47
  48 #if 0 /* JG */
  49 #include "opt_disable_pse.h"
  50 #include "opt_pmap.h"
  51 #endif
  52 #include "opt_msgbuf.h"
  53
  54 #include <sys/param.h>
  55 #include <sys/kernel.h>
  56 #include <sys/proc.h>
  57 #include <sys/msgbuf.h>
  58 #include <sys/vmmeter.h>
  59 #include <sys/mman.h>
  60 #include <sys/systm.h>
  61
  62 #include <vm/vm.h>
  63 #include <vm/vm_param.h>
  64 #include <sys/sysctl.h>
  65 #include <sys/lock.h>
  66 #include <vm/vm_kern.h>
  67 #include <vm/vm_page.h>
  68 #include <vm/vm_map.h>
  69 #include <vm/vm_object.h>
  70 #include <vm/vm_extern.h>
  71 #include <vm/vm_pageout.h>
  72 #include <vm/vm_pager.h>
  73 #include <vm/vm_zone.h>
  74
  75 #include <sys/user.h>
  76 #include <sys/thread2.h>
  77 #include <sys/sysref2.h>
  78 #include <sys/spinlock2.h>
  79 #include <vm/vm_page2.h>
  80
  81 #include <machine/cputypes.h>
  82 #include <machine/md_var.h>
  83 #include <machine/specialreg.h>
  84 #include <machine/smp.h>
  85 #include <machine_base/apic/apicreg.h>
  86 #include <machine/globaldata.h>
  87 #include <machine/pmap.h>
  88 #include <machine/pmap_inval.h>
  89 #include <machine/inttypes.h>
  90
  91 #include <ddb/ddb.h>
  92
  93 #define PMAP_KEEP_PDIRS
  94 #ifndef PMAP_SHPGPERPROC
  95 #define PMAP_SHPGPERPROC 2000
  96 #endif
  97
  98 #if defined(DIAGNOSTIC)
  99 #define PMAP_DIAGNOSTIC
 100 #endif
 101
 102 #define MINPV 2048
 103
 104 /*
 105  * pmap debugging will report who owns a pv lock when blocking.
 106  */
 107 #ifdef PMAP_DEBUG
 108
 109 #define PMAP_DEBUG_DECL         ,const char *func, int lineno
 110 #define PMAP_DEBUG_ARGS         , __func__, __LINE__
 111 #define PMAP_DEBUG_COPY         , func, lineno
 112
 113 #define pv_get(pmap, pindex)            _pv_get(pmap, pindex            \
 114                                                         PMAP_DEBUG_ARGS)
 115 #define pv_lock(pv)                     _pv_lock(pv                     \
 116                                                         PMAP_DEBUG_ARGS)
 117 #define pv_hold_try(pv)                 _pv_hold_try(pv                 \
 118                                                         PMAP_DEBUG_ARGS)
 119 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp  \
 120                                                         PMAP_DEBUG_ARGS)
 121
 122 #else
 123
 124 #define PMAP_DEBUG_DECL
 125 #define PMAP_DEBUG_ARGS
 126 #define PMAP_DEBUG_COPY
 127
 128 #define pv_get(pmap, pindex)            _pv_get(pmap, pindex)
 129 #define pv_lock(pv)                     _pv_lock(pv)
 130 #define pv_hold_try(pv)                 _pv_hold_try(pv)
 131 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp)
 132
 133 #endif
 134
 135 /*
 136  * Get PDEs and PTEs for user/kernel address space
 137  */
 138 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
 139
 140 #define pmap_pde_v(pmap, pte)           ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 141 #define pmap_pte_w(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
 142 #define pmap_pte_m(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
 143 #define pmap_pte_u(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
 144 #define pmap_pte_v(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 145
 146 /*
 147  * Given a map and a machine independent protection code,
 148  * convert to a vax protection code.
 149  */
 150 #define pte_prot(m, p)          \
 151         (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
 152 static int protection_codes[PROTECTION_CODES_SIZE];
 153
 154 struct pmap kernel_pmap;
 155 static TAILQ_HEAD(,pmap)        pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
 156
 157 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
 158
 159 vm_paddr_t avail_start;         /* PA of first available physical page */
 160 vm_paddr_t avail_end;           /* PA of last available physical page */
 161 vm_offset_t virtual2_start;     /* cutout free area prior to kernel start */
 162 vm_offset_t virtual2_end;
 163 vm_offset_t virtual_start;      /* VA of first avail page (after kernel bss) */
 164 vm_offset_t virtual_end;        /* VA of last avail page (end of kernel AS) */
 165 vm_offset_t KvaStart;           /* VA start of KVA space */
 166 vm_offset_t KvaEnd;             /* VA end of KVA space (non-inclusive) */
 167 vm_offset_t KvaSize;            /* max size of kernel virtual address space */
 168 static boolean_t pmap_initialized = FALSE;      /* Has pmap_init completed? */
 169 //static int pgeflag;           /* PG_G or-in */
 170 //static int pseflag;           /* PG_PS or-in */
 171 uint64_t PatMsr;
 172
 173 static int ndmpdp;
 174 static vm_paddr_t dmaplimit;
 175 static int nkpt;
 176 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
 177
 178 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE];        /* PAT -> PG_ bits */
 179 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/    /* PAT -> PG_ bits */
 180
 181 static uint64_t KPTbase;
 182 static uint64_t KPTphys;
 183 static uint64_t KPDphys;        /* phys addr of kernel level 2 */
 184 static uint64_t KPDbase;        /* phys addr of kernel level 2 @ KERNBASE */
 185 uint64_t KPDPphys;      /* phys addr of kernel level 3 */
 186 uint64_t KPML4phys;     /* phys addr of kernel level 4 */
 187
 188 static uint64_t DMPDphys;       /* phys addr of direct mapped level 2 */
 189 static uint64_t DMPDPphys;      /* phys addr of direct mapped level 3 */
 190
 191 /*
 192  * Data for the pv entry allocation mechanism
 193  */
 194 static vm_zone_t pvzone;
 195 static struct vm_zone pvzone_store;
 196 static struct vm_object pvzone_obj;
 197 static int pv_entry_max=0, pv_entry_high_water=0;
 198 static int pmap_pagedaemon_waken = 0;
 199 static struct pv_entry *pvinit;
 200
 201 /*
 202  * All those kernel PT submaps that BSD is so fond of
 203  */
 204 pt_entry_t *CMAP1 = NULL, *ptmmap;
 205 caddr_t CADDR1 = NULL, ptvmmap = NULL;
 206 static pt_entry_t *msgbufmap;
 207 struct msgbuf *msgbufp=NULL;
 208
 209 /*
 210  * PMAP default PG_* bits. Needed to be able to add
 211  * EPT/NPT pagetable pmap_bits for the VMM module
 212  */
 213 uint64_t pmap_bits_default[] = {
 214                 REGULAR_PMAP,                                   /* TYPE_IDX             0 */
 215                 X86_PG_V,                                       /* PG_V_IDX             1 */
 216                 X86_PG_RW,                                      /* PG_RW_IDX            2 */
 217                 X86_PG_U,                                       /* PG_U_IDX             3 */
 218                 X86_PG_A,                                       /* PG_A_IDX             4 */
 219                 X86_PG_M,                                       /* PG_M_IDX             5 */
 220                 X86_PG_PS,                                      /* PG_PS_IDX3           6 */
 221                 X86_PG_G,                                       /* PG_G_IDX             7 */
 222                 X86_PG_AVAIL1,                                  /* PG_AVAIL1_IDX        8 */
 223                 X86_PG_AVAIL2,                                  /* PG_AVAIL2_IDX        9 */
 224                 X86_PG_AVAIL3,                                  /* PG_AVAIL3_IDX        10 */
 225                 X86_PG_NC_PWT | X86_PG_NC_PCD,                  /* PG_N_IDX     11 */
 226 };
 227 /*
 228  * Crashdump maps.
 229  */
 230 static pt_entry_t *pt_crashdumpmap;
 231 static caddr_t crashdumpmap;
 232
 233 static int pmap_debug = 0;
 234 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
 235     &pmap_debug, 0, "Debug pmap's");
 236 #ifdef PMAP_DEBUG2
 237 static int pmap_enter_debug = 0;
 238 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
 239     &pmap_enter_debug, 0, "Debug pmap_enter's");
 240 #endif
 241 static int pmap_yield_count = 64;
 242 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
 243     &pmap_yield_count, 0, "Yield during init_pt/release");
 244 static int pmap_mmu_optimize = 0;
 245 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
 246     &pmap_mmu_optimize, 0, "Share page table pages when possible");
 247 int pmap_fast_kernel_cpusync = 0;
 248 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
 249     &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
 250 int pmap_dynamic_delete = 1;
 251 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
 252     &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
 253
 254 #define DISABLE_PSE
 255
 256 /* Standard user access funtions */
 257 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
 258     size_t *lencopied);
 259 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
 260 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
 261 extern int std_fubyte (const void *base);
 262 extern int std_subyte (void *base, int byte);
 263 extern long std_fuword (const void *base);
 264 extern int std_suword (void *base, long word);
 265 extern int std_suword32 (void *base, int word);
 266
 267 static void pv_hold(pv_entry_t pv);
 268 static int _pv_hold_try(pv_entry_t pv
 269                                 PMAP_DEBUG_DECL);
 270 static void pv_drop(pv_entry_t pv);
 271 static void _pv_lock(pv_entry_t pv
 272                                 PMAP_DEBUG_DECL);
 273 static void pv_unlock(pv_entry_t pv);
 274 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
 275                                 PMAP_DEBUG_DECL);
 276 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
 277                                 PMAP_DEBUG_DECL);
 278 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
 279 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
 280 static void pv_put(pv_entry_t pv);
 281 static void pv_free(pv_entry_t pv, pv_entry_t pvp, int putaway);
 282 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
 283 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
 284                       pv_entry_t *pvpp);
 285 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
 286                       pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
 287 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
 288                         pmap_inval_bulk_t *bulk, int destroy);
 289 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
 290 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
 291                         pmap_inval_bulk_t *bulk);
 292
 293 struct pmap_scan_info;
 294 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
 295                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
 296                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 297 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
 298                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
 299                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 300
 301 static void i386_protection_init (void);
 302 static void create_pagetables(vm_paddr_t *firstaddr);
 303 static void pmap_remove_all (vm_page_t m);
 304 static boolean_t pmap_testbit (vm_page_t m, int bit);
 305
 306 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
 307 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
 308
 309 static void pmap_pinit_defaults(struct pmap *pmap);
 310
 311 static unsigned pdir4mb;
 312
 313 static int
 314 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
 315 {
 316         if (pv1->pv_pindex < pv2->pv_pindex)
 317                 return(-1);
 318         if (pv1->pv_pindex > pv2->pv_pindex)
 319                 return(1);
 320         return(0);
 321 }
 322
 323 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
 324              pv_entry_compare, vm_pindex_t, pv_pindex);
 325
 326 static __inline
 327 void
 328 pmap_page_stats_adding(vm_page_t m)
 329 {
 330         globaldata_t gd = mycpu;
 331
 332         if (TAILQ_EMPTY(&m->md.pv_list)) {
 333                 ++gd->gd_vmtotal.t_arm;
 334         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 335                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 336                 ++gd->gd_vmtotal.t_armshr;
 337                 ++gd->gd_vmtotal.t_avmshr;
 338         } else {
 339                 ++gd->gd_vmtotal.t_avmshr;
 340         }
 341 }
 342
 343 static __inline
 344 void
 345 pmap_page_stats_deleting(vm_page_t m)
 346 {
 347         globaldata_t gd = mycpu;
 348
 349         if (TAILQ_EMPTY(&m->md.pv_list)) {
 350                 --gd->gd_vmtotal.t_arm;
 351         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 352                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 353                 --gd->gd_vmtotal.t_armshr;
 354                 --gd->gd_vmtotal.t_avmshr;
 355         } else {
 356                 --gd->gd_vmtotal.t_avmshr;
 357         }
 358 }
 359
 360 /*
 361  * Move the kernel virtual free pointer to the next
 362  * 2MB.  This is used to help improve performance
 363  * by using a large (2MB) page for much of the kernel
 364  * (.text, .data, .bss)
 365  */
 366 static
 367 vm_offset_t
 368 pmap_kmem_choose(vm_offset_t addr)
 369 {
 370         vm_offset_t newaddr = addr;
 371
 372         newaddr = roundup2(addr, NBPDR);
 373         return newaddr;
 374 }
 375
 376 /*
 377  * pmap_pte_quick:
 378  *
 379  *      Super fast pmap_pte routine best used when scanning the pv lists.
 380  *      This eliminates many course-grained invltlb calls.  Note that many of
 381  *      the pv list scans are across different pmaps and it is very wasteful
 382  *      to do an entire invltlb when checking a single mapping.
 383  */
 384 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
 385
 386 static
 387 pt_entry_t *
 388 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
 389 {
 390         return pmap_pte(pmap, va);
 391 }
 392
 393 /*
 394  * Returns the pindex of a page table entry (representing a terminal page).
 395  * There are NUPTE_TOTAL page table entries possible (a huge number)
 396  *
 397  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
 398  * We want to properly translate negative KVAs.
 399  */
 400 static __inline
 401 vm_pindex_t
 402 pmap_pte_pindex(vm_offset_t va)
 403 {
 404         return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
 405 }
 406
 407 /*
 408  * Returns the pindex of a page table.
 409  */
 410 static __inline
 411 vm_pindex_t
 412 pmap_pt_pindex(vm_offset_t va)
 413 {
 414         return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
 415 }
 416
 417 /*
 418  * Returns the pindex of a page directory.
 419  */
 420 static __inline
 421 vm_pindex_t
 422 pmap_pd_pindex(vm_offset_t va)
 423 {
 424         return (NUPTE_TOTAL + NUPT_TOTAL +
 425                 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
 426 }
 427
 428 static __inline
 429 vm_pindex_t
 430 pmap_pdp_pindex(vm_offset_t va)
 431 {
 432         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
 433                 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
 434 }
 435
 436 static __inline
 437 vm_pindex_t
 438 pmap_pml4_pindex(void)
 439 {
 440         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
 441 }
 442
 443 /*
 444  * Return various clipped indexes for a given VA
 445  *
 446  * Returns the index of a pt in a page directory, representing a page
 447  * table.
 448  */
 449 static __inline
 450 vm_pindex_t
 451 pmap_pt_index(vm_offset_t va)
 452 {
 453         return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
 454 }
 455
 456 /*
 457  * Returns the index of a pd in a page directory page, representing a page
 458  * directory.
 459  */
 460 static __inline
 461 vm_pindex_t
 462 pmap_pd_index(vm_offset_t va)
 463 {
 464         return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
 465 }
 466
 467 /*
 468  * Returns the index of a pdp in the pml4 table, representing a page
 469  * directory page.
 470  */
 471 static __inline
 472 vm_pindex_t
 473 pmap_pdp_index(vm_offset_t va)
 474 {
 475         return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
 476 }
 477
 478 /*
 479  * Generic procedure to index a pte from a pt, pd, or pdp.
 480  *
 481  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
 482  *       a page table page index but is instead of PV lookup index.
 483  */
 484 static
 485 void *
 486 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
 487 {
 488         pt_entry_t *pte;
 489
 490         pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
 491         return(&pte[pindex]);
 492 }
 493
 494 /*
 495  * Return pointer to PDP slot in the PML4
 496  */
 497 static __inline
 498 pml4_entry_t *
 499 pmap_pdp(pmap_t pmap, vm_offset_t va)
 500 {
 501         return (&pmap->pm_pml4[pmap_pdp_index(va)]);
 502 }
 503
 504 /*
 505  * Return pointer to PD slot in the PDP given a pointer to the PDP
 506  */
 507 static __inline
 508 pdp_entry_t *
 509 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
 510 {
 511         pdp_entry_t *pd;
 512
 513         pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
 514         return (&pd[pmap_pd_index(va)]);
 515 }
 516
 517 /*
 518  * Return pointer to PD slot in the PDP.
 519  */
 520 static __inline
 521 pdp_entry_t *
 522 pmap_pd(pmap_t pmap, vm_offset_t va)
 523 {
 524         pml4_entry_t *pdp;
 525
 526         pdp = pmap_pdp(pmap, va);
 527         if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
 528                 return NULL;
 529         return (pmap_pdp_to_pd(*pdp, va));
 530 }
 531
 532 /*
 533  * Return pointer to PT slot in the PD given a pointer to the PD
 534  */
 535 static __inline
 536 pd_entry_t *
 537 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
 538 {
 539         pd_entry_t *pt;
 540
 541         pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
 542         return (&pt[pmap_pt_index(va)]);
 543 }
 544
 545 /*
 546  * Return pointer to PT slot in the PD
 547  *
 548  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
 549  *                   so we cannot lookup the PD via the PDP.  Instead we
 550  *                   must look it up via the pmap.
 551  */
 552 static __inline
 553 pd_entry_t *
 554 pmap_pt(pmap_t pmap, vm_offset_t va)
 555 {
 556         pdp_entry_t *pd;
 557         pv_entry_t pv;
 558         vm_pindex_t pd_pindex;
 559
 560         if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
 561                 pd_pindex = pmap_pd_pindex(va);
 562                 spin_lock(&pmap->pm_spin);
 563                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
 564                 spin_unlock(&pmap->pm_spin);
 565                 if (pv == NULL || pv->pv_m == NULL)
 566                         return NULL;
 567                 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
 568         } else {
 569                 pd = pmap_pd(pmap, va);
 570                 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
 571                          return NULL;
 572                 return (pmap_pd_to_pt(*pd, va));
 573         }
 574 }
 575
 576 /*
 577  * Return pointer to PTE slot in the PT given a pointer to the PT
 578  */
 579 static __inline
 580 pt_entry_t *
 581 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
 582 {
 583         pt_entry_t *pte;
 584
 585         pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
 586         return (&pte[pmap_pte_index(va)]);
 587 }
 588
 589 /*
 590  * Return pointer to PTE slot in the PT
 591  */
 592 static __inline
 593 pt_entry_t *
 594 pmap_pte(pmap_t pmap, vm_offset_t va)
 595 {
 596         pd_entry_t *pt;
 597
 598         pt = pmap_pt(pmap, va);
 599         if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
 600                  return NULL;
 601         if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
 602                 return ((pt_entry_t *)pt);
 603         return (pmap_pt_to_pte(*pt, va));
 604 }
 605
 606 /*
 607  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
 608  * the PT layer.  This will speed up core pmap operations considerably.
 609  *
 610  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
 611  *       must be in a known associated state (typically by being locked when
 612  *       the pmap spinlock isn't held).  We allow the race for that case.
 613  */
 614 static __inline
 615 void
 616 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
 617 {
 618         if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
 619                 pv->pv_pmap->pm_pvhint = pv;
 620 }
 621
 622
 623 /*
 624  * Return address of PT slot in PD (KVM only)
 625  *
 626  * Cannot be used for user page tables because it might interfere with
 627  * the shared page-table-page optimization (pmap_mmu_optimize).
 628  */
 629 static __inline
 630 pd_entry_t *
 631 vtopt(vm_offset_t va)
 632 {
 633         uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
 634                                   NPML4EPGSHIFT)) - 1);
 635
 636         return (PDmap + ((va >> PDRSHIFT) & mask));
 637 }
 638
 639 /*
 640  * KVM - return address of PTE slot in PT
 641  */
 642 static __inline
 643 pt_entry_t *
 644 vtopte(vm_offset_t va)
 645 {
 646         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 647                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 648
 649         return (PTmap + ((va >> PAGE_SHIFT) & mask));
 650 }
 651
 652 static uint64_t
 653 allocpages(vm_paddr_t *firstaddr, long n)
 654 {
 655         uint64_t ret;
 656
 657         ret = *firstaddr;
 658         bzero((void *)ret, n * PAGE_SIZE);
 659         *firstaddr += n * PAGE_SIZE;
 660         return (ret);
 661 }
 662
 663 static
 664 void
 665 create_pagetables(vm_paddr_t *firstaddr)
 666 {
 667         long i;         /* must be 64 bits */
 668         long nkpt_base;
 669         long nkpt_phys;
 670         int j;
 671
 672         /*
 673          * We are running (mostly) V=P at this point
 674          *
 675          * Calculate NKPT - number of kernel page tables.  We have to
 676          * accomodoate prealloction of the vm_page_array, dump bitmap,
 677          * MSGBUF_SIZE, and other stuff.  Be generous.
 678          *
 679          * Maxmem is in pages.
 680          *
 681          * ndmpdp is the number of 1GB pages we wish to map.
 682          */
 683         ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
 684         if (ndmpdp < 4)         /* Minimum 4GB of dirmap */
 685                 ndmpdp = 4;
 686         KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
 687
 688         /*
 689          * Starting at the beginning of kvm (not KERNBASE).
 690          */
 691         nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
 692         nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
 693         nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
 694                        ndmpdp) + 511) / 512;
 695         nkpt_phys += 128;
 696
 697         /*
 698          * Starting at KERNBASE - map 2G worth of page table pages.
 699          * KERNBASE is offset -2G from the end of kvm.
 700          */
 701         nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
 702
 703         /*
 704          * Allocate pages
 705          */
 706         KPTbase = allocpages(firstaddr, nkpt_base);
 707         KPTphys = allocpages(firstaddr, nkpt_phys);
 708         KPML4phys = allocpages(firstaddr, 1);
 709         KPDPphys = allocpages(firstaddr, NKPML4E);
 710         KPDphys = allocpages(firstaddr, NKPDPE);
 711
 712         /*
 713          * Calculate the page directory base for KERNBASE,
 714          * that is where we start populating the page table pages.
 715          * Basically this is the end - 2.
 716          */
 717         KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
 718
 719         DMPDPphys = allocpages(firstaddr, NDMPML4E);
 720         if ((amd_feature & AMDID_PAGE1GB) == 0)
 721                 DMPDphys = allocpages(firstaddr, ndmpdp);
 722         dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
 723
 724         /*
 725          * Fill in the underlying page table pages for the area around
 726          * KERNBASE.  This remaps low physical memory to KERNBASE.
 727          *
 728          * Read-only from zero to physfree
 729          * XXX not fully used, underneath 2M pages
 730          */
 731         for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
 732                 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
 733                 ((pt_entry_t *)KPTbase)[i] |=
 734                     pmap_bits_default[PG_RW_IDX] |
 735                     pmap_bits_default[PG_V_IDX] |
 736                     pmap_bits_default[PG_G_IDX];
 737         }
 738
 739         /*
 740          * Now map the initial kernel page tables.  One block of page
 741          * tables is placed at the beginning of kernel virtual memory,
 742          * and another block is placed at KERNBASE to map the kernel binary,
 743          * data, bss, and initial pre-allocations.
 744          */
 745         for (i = 0; i < nkpt_base; i++) {
 746                 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
 747                 ((pd_entry_t *)KPDbase)[i] |=
 748                     pmap_bits_default[PG_RW_IDX] |
 749                     pmap_bits_default[PG_V_IDX];
 750         }
 751         for (i = 0; i < nkpt_phys; i++) {
 752                 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
 753                 ((pd_entry_t *)KPDphys)[i] |=
 754                     pmap_bits_default[PG_RW_IDX] |
 755                     pmap_bits_default[PG_V_IDX];
 756         }
 757
 758         /*
 759          * Map from zero to end of allocations using 2M pages as an
 760          * optimization.  This will bypass some of the KPTBase pages
 761          * above in the KERNBASE area.
 762          */
 763         for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
 764                 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
 765                 ((pd_entry_t *)KPDbase)[i] |=
 766                     pmap_bits_default[PG_RW_IDX] |
 767                     pmap_bits_default[PG_V_IDX] |
 768                     pmap_bits_default[PG_PS_IDX] |
 769                     pmap_bits_default[PG_G_IDX];
 770         }
 771
 772         /*
 773          * And connect up the PD to the PDP.  The kernel pmap is expected
 774          * to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
 775          */
 776         for (i = 0; i < NKPDPE; i++) {
 777                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
 778                                 KPDphys + (i << PAGE_SHIFT);
 779                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
 780                     pmap_bits_default[PG_RW_IDX] |
 781                     pmap_bits_default[PG_V_IDX] |
 782                     pmap_bits_default[PG_U_IDX];
 783         }
 784
 785         /*
 786          * Now set up the direct map space using either 2MB or 1GB pages
 787          * Preset PG_M and PG_A because demotion expects it.
 788          *
 789          * When filling in entries in the PD pages make sure any excess
 790          * entries are set to zero as we allocated enough PD pages
 791          */
 792         if ((amd_feature & AMDID_PAGE1GB) == 0) {
 793                 for (i = 0; i < NPDEPG * ndmpdp; i++) {
 794                         ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
 795                         ((pd_entry_t *)DMPDphys)[i] |=
 796                             pmap_bits_default[PG_RW_IDX] |
 797                             pmap_bits_default[PG_V_IDX] |
 798                             pmap_bits_default[PG_PS_IDX] |
 799                             pmap_bits_default[PG_G_IDX] |
 800                             pmap_bits_default[PG_M_IDX] |
 801                             pmap_bits_default[PG_A_IDX];
 802                 }
 803
 804                 /*
 805                  * And the direct map space's PDP
 806                  */
 807                 for (i = 0; i < ndmpdp; i++) {
 808                         ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
 809                                                         (i << PAGE_SHIFT);
 810                         ((pdp_entry_t *)DMPDPphys)[i] |=
 811                             pmap_bits_default[PG_RW_IDX] |
 812                             pmap_bits_default[PG_V_IDX] |
 813                             pmap_bits_default[PG_U_IDX];
 814                 }
 815         } else {
 816                 for (i = 0; i < ndmpdp; i++) {
 817                         ((pdp_entry_t *)DMPDPphys)[i] =
 818                                                 (vm_paddr_t)i << PDPSHIFT;
 819                         ((pdp_entry_t *)DMPDPphys)[i] |=
 820                             pmap_bits_default[PG_RW_IDX] |
 821                             pmap_bits_default[PG_V_IDX] |
 822                             pmap_bits_default[PG_PS_IDX] |
 823                             pmap_bits_default[PG_G_IDX] |
 824                             pmap_bits_default[PG_M_IDX] |
 825                             pmap_bits_default[PG_A_IDX];
 826                 }
 827         }
 828
 829         /* And recursively map PML4 to itself in order to get PTmap */
 830         ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
 831         ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
 832             pmap_bits_default[PG_RW_IDX] |
 833             pmap_bits_default[PG_V_IDX] |
 834             pmap_bits_default[PG_U_IDX];
 835
 836         /*
 837          * Connect the Direct Map slots up to the PML4
 838          */
 839         for (j = 0; j < NDMPML4E; ++j) {
 840                 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
 841                     (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
 842                     pmap_bits_default[PG_RW_IDX] |
 843                     pmap_bits_default[PG_V_IDX] |
 844                     pmap_bits_default[PG_U_IDX];
 845         }
 846
 847         /*
 848          * Connect the KVA slot up to the PML4
 849          */
 850         ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
 851         ((pdp_entry_t *)KPML4phys)[KPML4I] |=
 852             pmap_bits_default[PG_RW_IDX] |
 853             pmap_bits_default[PG_V_IDX] |
 854             pmap_bits_default[PG_U_IDX];
 855 }
 856
 857 /*
 858  *      Bootstrap the system enough to run with virtual memory.
 859  *
 860  *      On the i386 this is called after mapping has already been enabled
 861  *      and just syncs the pmap module with what has already been done.
 862  *      [We can't call it easily with mapping off since the kernel is not
 863  *      mapped with PA == VA, hence we would have to relocate every address
 864  *      from the linked base (virtual) address "KERNBASE" to the actual
 865  *      (physical) address starting relative to 0]
 866  */
 867 void
 868 pmap_bootstrap(vm_paddr_t *firstaddr)
 869 {
 870         vm_offset_t va;
 871         pt_entry_t *pte;
 872
 873         KvaStart = VM_MIN_KERNEL_ADDRESS;
 874         KvaEnd = VM_MAX_KERNEL_ADDRESS;
 875         KvaSize = KvaEnd - KvaStart;
 876
 877         avail_start = *firstaddr;
 878
 879         /*
 880          * Create an initial set of page tables to run the kernel in.
 881          */
 882         create_pagetables(firstaddr);
 883
 884         virtual2_start = KvaStart;
 885         virtual2_end = PTOV_OFFSET;
 886
 887         virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
 888         virtual_start = pmap_kmem_choose(virtual_start);
 889
 890         virtual_end = VM_MAX_KERNEL_ADDRESS;
 891
 892         /* XXX do %cr0 as well */
 893         load_cr4(rcr4() | CR4_PGE | CR4_PSE);
 894         load_cr3(KPML4phys);
 895
 896         /*
 897          * Initialize protection array.
 898          */
 899         i386_protection_init();
 900
 901         /*
 902          * The kernel's pmap is statically allocated so we don't have to use
 903          * pmap_create, which is unlikely to work correctly at this part of
 904          * the boot sequence (XXX and which no longer exists).
 905          */
 906         kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
 907         kernel_pmap.pm_count = 1;
 908         CPUMASK_ASSALLONES(kernel_pmap.pm_active);
 909         RB_INIT(&kernel_pmap.pm_pvroot);
 910         spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
 911         lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
 912
 913         /*
 914          * Reserve some special page table entries/VA space for temporary
 915          * mapping of pages.
 916          */
 917 #define SYSMAP(c, p, v, n)      \
 918         v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
 919
 920         va = virtual_start;
 921         pte = vtopte(va);
 922
 923         /*
 924          * CMAP1/CMAP2 are used for zeroing and copying pages.
 925          */
 926         SYSMAP(caddr_t, CMAP1, CADDR1, 1)
 927
 928         /*
 929          * Crashdump maps.
 930          */
 931         SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
 932
 933         /*
 934          * ptvmmap is used for reading arbitrary physical pages via
 935          * /dev/mem.
 936          */
 937         SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
 938
 939         /*
 940          * msgbufp is used to map the system message buffer.
 941          * XXX msgbufmap is not used.
 942          */
 943         SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
 944                atop(round_page(MSGBUF_SIZE)))
 945
 946         virtual_start = va;
 947         virtual_start = pmap_kmem_choose(virtual_start);
 948
 949         *CMAP1 = 0;
 950
 951         /*
 952          * PG_G is terribly broken on SMP because we IPI invltlb's in some
 953          * cases rather then invl1pg.  Actually, I don't even know why it
 954          * works under UP because self-referential page table mappings
 955          */
 956 //      pgeflag = 0;
 957
 958 /*
 959  * Initialize the 4MB page size flag
 960  */
 961 //      pseflag = 0;
 962 /*
 963  * The 4MB page version of the initial
 964  * kernel page mapping.
 965  */
 966         pdir4mb = 0;
 967
 968 #if !defined(DISABLE_PSE)
 969         if (cpu_feature & CPUID_PSE) {
 970                 pt_entry_t ptditmp;
 971                 /*
 972                  * Note that we have enabled PSE mode
 973                  */
 974 //              pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
 975                 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
 976                 ptditmp &= ~(NBPDR - 1);
 977                 ptditmp |= pmap_bits_default[PG_V_IDX] |
 978                     pmap_bits_default[PG_RW_IDX] |
 979                     pmap_bits_default[PG_PS_IDX] |
 980                     pmap_bits_default[PG_U_IDX];
 981 //                  pgeflag;
 982                 pdir4mb = ptditmp;
 983         }
 984 #endif
 985         cpu_invltlb();
 986
 987         /* Initialize the PAT MSR */
 988         pmap_init_pat();
 989         pmap_pinit_defaults(&kernel_pmap);
 990
 991         TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
 992                           &pmap_fast_kernel_cpusync);
 993
 994 }
 995
 996 /*
 997  * Setup the PAT MSR.
 998  */
 999 void
1000 pmap_init_pat(void)
1001 {
1002         uint64_t pat_msr;
1003         u_long cr0, cr4;
1004
1005         /*
1006          * Default values mapping PATi,PCD,PWT bits at system reset.
1007          * The default values effectively ignore the PATi bit by
1008          * repeating the encodings for 0-3 in 4-7, and map the PCD
1009          * and PWT bit combinations to the expected PAT types.
1010          */
1011         pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |        /* 000 */
1012                   PAT_VALUE(1, PAT_WRITE_THROUGH) |     /* 001 */
1013                   PAT_VALUE(2, PAT_UNCACHED) |          /* 010 */
1014                   PAT_VALUE(3, PAT_UNCACHEABLE) |       /* 011 */
1015                   PAT_VALUE(4, PAT_WRITE_BACK) |        /* 100 */
1016                   PAT_VALUE(5, PAT_WRITE_THROUGH) |     /* 101 */
1017                   PAT_VALUE(6, PAT_UNCACHED) |          /* 110 */
1018                   PAT_VALUE(7, PAT_UNCACHEABLE);        /* 111 */
1019         pat_pte_index[PAT_WRITE_BACK]   = 0;
1020         pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
1021         pat_pte_index[PAT_UNCACHED]     = X86_PG_NC_PCD;
1022         pat_pte_index[PAT_UNCACHEABLE]  = X86_PG_NC_PCD | X86_PG_NC_PWT;
1023         pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1024         pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1025
1026         if (cpu_feature & CPUID_PAT) {
1027                 /*
1028                  * If we support the PAT then set-up entries for
1029                  * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1030                  * 4 and 5.
1031                  */
1032                 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1033                           PAT_VALUE(4, PAT_WRITE_PROTECTED);
1034                 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1035                           PAT_VALUE(5, PAT_WRITE_COMBINING);
1036                 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1037                 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1038
1039                 /*
1040                  * Then enable the PAT
1041                  */
1042
1043                 /* Disable PGE. */
1044                 cr4 = rcr4();
1045                 load_cr4(cr4 & ~CR4_PGE);
1046
1047                 /* Disable caches (CD = 1, NW = 0). */
1048                 cr0 = rcr0();
1049                 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1050
1051                 /* Flushes caches and TLBs. */
1052                 wbinvd();
1053                 cpu_invltlb();
1054
1055                 /* Update PAT and index table. */
1056                 wrmsr(MSR_PAT, pat_msr);
1057
1058                 /* Flush caches and TLBs again. */
1059                 wbinvd();
1060                 cpu_invltlb();
1061
1062                 /* Restore caches and PGE. */
1063                 load_cr0(cr0);
1064                 load_cr4(cr4);
1065                 PatMsr = pat_msr;
1066         }
1067 }
1068
1069 /*
1070  * Set 4mb pdir for mp startup
1071  */
1072 void
1073 pmap_set_opt(void)
1074 {
1075         if (cpu_feature & CPUID_PSE) {
1076                 load_cr4(rcr4() | CR4_PSE);
1077                 if (pdir4mb && mycpu->gd_cpuid == 0) {  /* only on BSP */
1078                         cpu_invltlb();
1079                 }
1080         }
1081 }
1082
1083 /*
1084  *      Initialize the pmap module.
1085  *      Called by vm_init, to initialize any structures that the pmap
1086  *      system needs to map virtual memory.
1087  *      pmap_init has been enhanced to support in a fairly consistant
1088  *      way, discontiguous physical memory.
1089  */
1090 void
1091 pmap_init(void)
1092 {
1093         int i;
1094         int initial_pvs;
1095
1096         /*
1097          * Allocate memory for random pmap data structures.  Includes the
1098          * pv_head_table.
1099          */
1100
1101         for (i = 0; i < vm_page_array_size; i++) {
1102                 vm_page_t m;
1103
1104                 m = &vm_page_array[i];
1105                 TAILQ_INIT(&m->md.pv_list);
1106         }
1107
1108         /*
1109          * init the pv free list
1110          */
1111         initial_pvs = vm_page_array_size;
1112         if (initial_pvs < MINPV)
1113                 initial_pvs = MINPV;
1114         pvzone = &pvzone_store;
1115         pvinit = (void *)kmem_alloc(&kernel_map,
1116                                     initial_pvs * sizeof (struct pv_entry),
1117                                     VM_SUBSYS_PVENTRY);
1118         zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1119                   pvinit, initial_pvs);
1120
1121         /*
1122          * Now it is safe to enable pv_table recording.
1123          */
1124         pmap_initialized = TRUE;
1125 }
1126
1127 /*
1128  * Initialize the address space (zone) for the pv_entries.  Set a
1129  * high water mark so that the system can recover from excessive
1130  * numbers of pv entries.
1131  */
1132 void
1133 pmap_init2(void)
1134 {
1135         int shpgperproc = PMAP_SHPGPERPROC;
1136         int entry_max;
1137
1138         TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1139         pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1140         TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1141         pv_entry_high_water = 9 * (pv_entry_max / 10);
1142
1143         /*
1144          * Subtract out pages already installed in the zone (hack)
1145          */
1146         entry_max = pv_entry_max - vm_page_array_size;
1147         if (entry_max <= 0)
1148                 entry_max = 1;
1149
1150         zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT);
1151 }
1152
1153 /*
1154  * Typically used to initialize a fictitious page by vm/device_pager.c
1155  */
1156 void
1157 pmap_page_init(struct vm_page *m)
1158 {
1159         vm_page_init(m);
1160         TAILQ_INIT(&m->md.pv_list);
1161 }
1162
1163 /***************************************************
1164  * Low level helper routines.....
1165  ***************************************************/
1166
1167 /*
1168  * this routine defines the region(s) of memory that should
1169  * not be tested for the modified bit.
1170  */
1171 static __inline
1172 int
1173 pmap_track_modified(vm_pindex_t pindex)
1174 {
1175         vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1176         if ((va < clean_sva) || (va >= clean_eva))
1177                 return 1;
1178         else
1179                 return 0;
1180 }
1181
1182 /*
1183  * Extract the physical page address associated with the map/VA pair.
1184  * The page must be wired for this to work reliably.
1185  *
1186  * XXX for the moment we're using pv_find() instead of pv_get(), as
1187  *     callers might be expecting non-blocking operation.
1188  */
1189 vm_paddr_t
1190 pmap_extract(pmap_t pmap, vm_offset_t va)
1191 {
1192         vm_paddr_t rtval;
1193         pv_entry_t pt_pv;
1194         pt_entry_t *ptep;
1195
1196         rtval = 0;
1197         if (va >= VM_MAX_USER_ADDRESS) {
1198                 /*
1199                  * Kernel page directories might be direct-mapped and
1200                  * there is typically no PV tracking of pte's
1201                  */
1202                 pd_entry_t *pt;
1203
1204                 pt = pmap_pt(pmap, va);
1205                 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1206                         if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1207                                 rtval = *pt & PG_PS_FRAME;
1208                                 rtval |= va & PDRMASK;
1209                         } else {
1210                                 ptep = pmap_pt_to_pte(*pt, va);
1211                                 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1212                                         rtval = *ptep & PG_FRAME;
1213                                         rtval |= va & PAGE_MASK;
1214                                 }
1215                         }
1216                 }
1217         } else {
1218                 /*
1219                  * User pages currently do not direct-map the page directory
1220                  * and some pages might not used managed PVs.  But all PT's
1221                  * will have a PV.
1222                  */
1223                 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1224                 if (pt_pv) {
1225                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1226                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1227                                 rtval = *ptep & PG_FRAME;
1228                                 rtval |= va & PAGE_MASK;
1229                         }
1230                         pv_drop(pt_pv);
1231                 }
1232         }
1233         return rtval;
1234 }
1235
1236 /*
1237  * Similar to extract but checks protections, SMP-friendly short-cut for
1238  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1239  * fall-through to the real fault code.
1240  *
1241  * The returned page, if not NULL, is held (and not busied).
1242  */
1243 vm_page_t
1244 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1245 {
1246         if (pmap && va < VM_MAX_USER_ADDRESS) {
1247                 pv_entry_t pt_pv;
1248                 pv_entry_t pte_pv;
1249                 pt_entry_t *ptep;
1250                 pt_entry_t req;
1251                 vm_page_t m;
1252                 int error;
1253
1254                 req = pmap->pmap_bits[PG_V_IDX] |
1255                       pmap->pmap_bits[PG_U_IDX];
1256                 if (prot & VM_PROT_WRITE)
1257                         req |= pmap->pmap_bits[PG_RW_IDX];
1258
1259                 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1260                 if (pt_pv == NULL)
1261                         return (NULL);
1262                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1263                 if ((*ptep & req) != req) {
1264                         pv_drop(pt_pv);
1265                         return (NULL);
1266                 }
1267                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), &error);
1268                 if (pte_pv && error == 0) {
1269                         m = pte_pv->pv_m;
1270                         vm_page_hold(m);
1271                         if (prot & VM_PROT_WRITE)
1272                                 vm_page_dirty(m);
1273                         pv_put(pte_pv);
1274                 } else if (pte_pv) {
1275                         pv_drop(pte_pv);
1276                         m = NULL;
1277                 } else {
1278                         m = NULL;
1279                 }
1280                 pv_drop(pt_pv);
1281                 return(m);
1282         } else {
1283                 return(NULL);
1284         }
1285 }
1286
1287 /*
1288  * Extract the physical page address associated kernel virtual address.
1289  */
1290 vm_paddr_t
1291 pmap_kextract(vm_offset_t va)
1292 {
1293         pd_entry_t pt;          /* pt entry in pd */
1294         vm_paddr_t pa;
1295
1296         if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1297                 pa = DMAP_TO_PHYS(va);
1298         } else {
1299                 pt = *vtopt(va);
1300                 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1301                         pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1302                 } else {
1303                         /*
1304                          * Beware of a concurrent promotion that changes the
1305                          * PDE at this point!  For example, vtopte() must not
1306                          * be used to access the PTE because it would use the
1307                          * new PDE.  It is, however, safe to use the old PDE
1308                          * because the page table page is preserved by the
1309                          * promotion.
1310                          */
1311                         pa = *pmap_pt_to_pte(pt, va);
1312                         pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1313                 }
1314         }
1315         return pa;
1316 }
1317
1318 /***************************************************
1319  * Low level mapping routines.....
1320  ***************************************************/
1321
1322 /*
1323  * Routine: pmap_kenter
1324  * Function:
1325  *      Add a wired page to the KVA
1326  *      NOTE! note that in order for the mapping to take effect -- you
1327  *      should do an invltlb after doing the pmap_kenter().
1328  */
1329 void
1330 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1331 {
1332         pt_entry_t *ptep;
1333         pt_entry_t npte;
1334
1335         npte = pa |
1336             kernel_pmap.pmap_bits[PG_RW_IDX] |
1337             kernel_pmap.pmap_bits[PG_V_IDX];
1338 //          pgeflag;
1339         ptep = vtopte(va);
1340 #if 1
1341         pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1342 #else
1343         /* FUTURE */
1344         if (*ptep)
1345                 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1346         else
1347                 *ptep = npte;
1348 #endif
1349 }
1350
1351 /*
1352  * Similar to pmap_kenter(), except we only invalidate the mapping on the
1353  * current CPU.  Returns 0 if the previous pte was 0, 1 if it wasn't
1354  * (caller can conditionalize calling smp_invltlb()).
1355  */
1356 int
1357 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1358 {
1359         pt_entry_t *ptep;
1360         pt_entry_t npte;
1361         int res;
1362
1363         npte = pa |
1364             kernel_pmap.pmap_bits[PG_RW_IDX] |
1365             kernel_pmap.pmap_bits[PG_V_IDX];
1366 //          pgeflag;
1367         ptep = vtopte(va);
1368 #if 1
1369         res = 1;
1370 #else
1371         /* FUTURE */
1372         res = (*ptep != 0);
1373 #endif
1374         *ptep = npte;
1375         cpu_invlpg((void *)va);
1376
1377         return res;
1378 }
1379
1380 /*
1381  * Enter addresses into the kernel pmap but don't bother
1382  * doing any tlb invalidations.  Caller will do a rollup
1383  * invalidation via pmap_rollup_inval().
1384  */
1385 int
1386 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1387 {
1388         pt_entry_t *ptep;
1389         pt_entry_t npte;
1390         int res;
1391
1392         npte = pa |
1393             kernel_pmap.pmap_bits[PG_RW_IDX] |
1394             kernel_pmap.pmap_bits[PG_V_IDX];
1395 //          pgeflag;
1396         ptep = vtopte(va);
1397 #if 1
1398         res = 1;
1399 #else
1400         /* FUTURE */
1401         res = (*ptep != 0);
1402 #endif
1403         *ptep = npte;
1404         cpu_invlpg((void *)va);
1405
1406         return res;
1407 }
1408
1409 /*
1410  * remove a page from the kernel pagetables
1411  */
1412 void
1413 pmap_kremove(vm_offset_t va)
1414 {
1415         pt_entry_t *ptep;
1416
1417         ptep = vtopte(va);
1418         pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1419 }
1420
1421 void
1422 pmap_kremove_quick(vm_offset_t va)
1423 {
1424         pt_entry_t *ptep;
1425
1426         ptep = vtopte(va);
1427         (void)pte_load_clear(ptep);
1428         cpu_invlpg((void *)va);
1429 }
1430
1431 /*
1432  * Remove addresses from the kernel pmap but don't bother
1433  * doing any tlb invalidations.  Caller will do a rollup
1434  * invalidation via pmap_rollup_inval().
1435  */
1436 void
1437 pmap_kremove_noinval(vm_offset_t va)
1438 {
1439         pt_entry_t *ptep;
1440
1441         ptep = vtopte(va);
1442         (void)pte_load_clear(ptep);
1443 }
1444
1445 /*
1446  * XXX these need to be recoded.  They are not used in any critical path.
1447  */
1448 void
1449 pmap_kmodify_rw(vm_offset_t va)
1450 {
1451         atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1452         cpu_invlpg((void *)va);
1453 }
1454
1455 /* NOT USED
1456 void
1457 pmap_kmodify_nc(vm_offset_t va)
1458 {
1459         atomic_set_long(vtopte(va), PG_N);
1460         cpu_invlpg((void *)va);
1461 }
1462 */
1463
1464 /*
1465  * Used to map a range of physical addresses into kernel virtual
1466  * address space during the low level boot, typically to map the
1467  * dump bitmap, message buffer, and vm_page_array.
1468  *
1469  * These mappings are typically made at some pointer after the end of the
1470  * kernel text+data.
1471  *
1472  * We could return PHYS_TO_DMAP(start) here and not allocate any
1473  * via (*virtp), but then kmem from userland and kernel dumps won't
1474  * have access to the related pointers.
1475  */
1476 vm_offset_t
1477 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1478 {
1479         vm_offset_t va;
1480         vm_offset_t va_start;
1481
1482         /*return PHYS_TO_DMAP(start);*/
1483
1484         va_start = *virtp;
1485         va = va_start;
1486
1487         while (start < end) {
1488                 pmap_kenter_quick(va, start);
1489                 va += PAGE_SIZE;
1490                 start += PAGE_SIZE;
1491         }
1492         *virtp = va;
1493         return va_start;
1494 }
1495
1496 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
1497
1498 /*
1499  * Remove the specified set of pages from the data and instruction caches.
1500  *
1501  * In contrast to pmap_invalidate_cache_range(), this function does not
1502  * rely on the CPU's self-snoop feature, because it is intended for use
1503  * when moving pages into a different cache domain.
1504  */
1505 void
1506 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1507 {
1508         vm_offset_t daddr, eva;
1509         int i;
1510
1511         if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1512             (cpu_feature & CPUID_CLFSH) == 0)
1513                 wbinvd();
1514         else {
1515                 cpu_mfence();
1516                 for (i = 0; i < count; i++) {
1517                         daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1518                         eva = daddr + PAGE_SIZE;
1519                         for (; daddr < eva; daddr += cpu_clflush_line_size)
1520                                 clflush(daddr);
1521                 }
1522                 cpu_mfence();
1523         }
1524 }
1525
1526 void
1527 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1528 {
1529         KASSERT((sva & PAGE_MASK) == 0,
1530             ("pmap_invalidate_cache_range: sva not page-aligned"));
1531         KASSERT((eva & PAGE_MASK) == 0,
1532             ("pmap_invalidate_cache_range: eva not page-aligned"));
1533
1534         if (cpu_feature & CPUID_SS) {
1535                 ; /* If "Self Snoop" is supported, do nothing. */
1536         } else {
1537                 /* Globally invalidate caches */
1538                 cpu_wbinvd_on_all_cpus();
1539         }
1540 }
1541
1542 /*
1543  * Invalidate the specified range of virtual memory on all cpus associated
1544  * with the pmap.
1545  */
1546 void
1547 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1548 {
1549         pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1550 }
1551
1552 /*
1553  * Add a list of wired pages to the kva.  This routine is used for temporary
1554  * kernel mappings such as those found in buffer cache buffer.  Page
1555  * modifications and accesses are not tracked or recorded.
1556  *
1557  * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1558  *       semantics as previous mappings may have been zerod without any
1559  *       invalidation.
1560  *
1561  * The page *must* be wired.
1562  */
1563 void
1564 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1565 {
1566         vm_offset_t end_va;
1567         vm_offset_t va;
1568
1569         end_va = beg_va + count * PAGE_SIZE;
1570
1571         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1572                 pt_entry_t *pte;
1573
1574                 pte = vtopte(va);
1575                 *pte = VM_PAGE_TO_PHYS(*m) |
1576                     kernel_pmap.pmap_bits[PG_RW_IDX] |
1577                     kernel_pmap.pmap_bits[PG_V_IDX] |
1578                     kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1579 //              pgeflag;
1580                 m++;
1581         }
1582         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1583 }
1584
1585 /*
1586  * This routine jerks page mappings from the kernel -- it is meant only
1587  * for temporary mappings such as those found in buffer cache buffers.
1588  * No recording modified or access status occurs.
1589  *
1590  * MPSAFE, INTERRUPT SAFE (cluster callback)
1591  */
1592 void
1593 pmap_qremove(vm_offset_t beg_va, int count)
1594 {
1595         vm_offset_t end_va;
1596         vm_offset_t va;
1597
1598         end_va = beg_va + count * PAGE_SIZE;
1599
1600         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1601                 pt_entry_t *pte;
1602
1603                 pte = vtopte(va);
1604                 (void)pte_load_clear(pte);
1605                 cpu_invlpg((void *)va);
1606         }
1607         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1608 }
1609
1610 /*
1611  * This routine removes temporary kernel mappings, only invalidating them
1612  * on the current cpu.  It should only be used under carefully controlled
1613  * conditions.
1614  */
1615 void
1616 pmap_qremove_quick(vm_offset_t beg_va, int count)
1617 {
1618         vm_offset_t end_va;
1619         vm_offset_t va;
1620
1621         end_va = beg_va + count * PAGE_SIZE;
1622
1623         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1624                 pt_entry_t *pte;
1625
1626                 pte = vtopte(va);
1627                 (void)pte_load_clear(pte);
1628                 cpu_invlpg((void *)va);
1629         }
1630 }
1631
1632 /*
1633  * This routine removes temporary kernel mappings *without* invalidating
1634  * the TLB.  It can only be used on permanent kva reservations such as those
1635  * found in buffer cache buffers, under carefully controlled circumstances.
1636  *
1637  * NOTE: Repopulating these KVAs requires unconditional invalidation.
1638  *       (pmap_qenter() does unconditional invalidation).
1639  */
1640 void
1641 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1642 {
1643         vm_offset_t end_va;
1644         vm_offset_t va;
1645
1646         end_va = beg_va + count * PAGE_SIZE;
1647
1648         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1649                 pt_entry_t *pte;
1650
1651                 pte = vtopte(va);
1652                 (void)pte_load_clear(pte);
1653         }
1654 }
1655
1656 /*
1657  * Create a new thread and optionally associate it with a (new) process.
1658  * NOTE! the new thread's cpu may not equal the current cpu.
1659  */
1660 void
1661 pmap_init_thread(thread_t td)
1662 {
1663         /* enforce pcb placement & alignment */
1664         td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1665         td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1666         td->td_savefpu = &td->td_pcb->pcb_save;
1667         td->td_sp = (char *)td->td_pcb; /* no -16 */
1668 }
1669
1670 /*
1671  * This routine directly affects the fork perf for a process.
1672  */
1673 void
1674 pmap_init_proc(struct proc *p)
1675 {
1676 }
1677
1678 static void
1679 pmap_pinit_defaults(struct pmap *pmap)
1680 {
1681         bcopy(pmap_bits_default, pmap->pmap_bits,
1682               sizeof(pmap_bits_default));
1683         bcopy(protection_codes, pmap->protection_codes,
1684               sizeof(protection_codes));
1685         bcopy(pat_pte_index, pmap->pmap_cache_bits,
1686               sizeof(pat_pte_index));
1687         pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1688         pmap->copyinstr = std_copyinstr;
1689         pmap->copyin = std_copyin;
1690         pmap->copyout = std_copyout;
1691         pmap->fubyte = std_fubyte;
1692         pmap->subyte = std_subyte;
1693         pmap->fuword = std_fuword;
1694         pmap->suword = std_suword;
1695         pmap->suword32 = std_suword32;
1696 }
1697 /*
1698  * Initialize pmap0/vmspace0.  This pmap is not added to pmap_list because
1699  * it, and IdlePTD, represents the template used to update all other pmaps.
1700  *
1701  * On architectures where the kernel pmap is not integrated into the user
1702  * process pmap, this pmap represents the process pmap, not the kernel pmap.
1703  * kernel_pmap should be used to directly access the kernel_pmap.
1704  */
1705 void
1706 pmap_pinit0(struct pmap *pmap)
1707 {
1708         pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1709         pmap->pm_count = 1;
1710         CPUMASK_ASSZERO(pmap->pm_active);
1711         pmap->pm_pvhint = NULL;
1712         RB_INIT(&pmap->pm_pvroot);
1713         spin_init(&pmap->pm_spin, "pmapinit0");
1714         lwkt_token_init(&pmap->pm_token, "pmap_tok");
1715         bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1716         pmap_pinit_defaults(pmap);
1717 }
1718
1719 /*
1720  * Initialize a preallocated and zeroed pmap structure,
1721  * such as one in a vmspace structure.
1722  */
1723 static void
1724 pmap_pinit_simple(struct pmap *pmap)
1725 {
1726         /*
1727          * Misc initialization
1728          */
1729         pmap->pm_count = 1;
1730         CPUMASK_ASSZERO(pmap->pm_active);
1731         pmap->pm_pvhint = NULL;
1732         pmap->pm_flags = PMAP_FLAG_SIMPLE;
1733
1734         pmap_pinit_defaults(pmap);
1735
1736         /*
1737          * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1738          * for this).
1739          */
1740         if (pmap->pm_pmlpv == NULL) {
1741                 RB_INIT(&pmap->pm_pvroot);
1742                 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1743                 spin_init(&pmap->pm_spin, "pmapinitsimple");
1744                 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1745         }
1746 }
1747
1748 void
1749 pmap_pinit(struct pmap *pmap)
1750 {
1751         pv_entry_t pv;
1752         int j;
1753
1754         if (pmap->pm_pmlpv) {
1755                 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1756                         pmap_puninit(pmap);
1757                 }
1758         }
1759
1760         pmap_pinit_simple(pmap);
1761         pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1762
1763         /*
1764          * No need to allocate page table space yet but we do need a valid
1765          * page directory table.
1766          */
1767         if (pmap->pm_pml4 == NULL) {
1768                 pmap->pm_pml4 =
1769                     (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1770                                                         PAGE_SIZE,
1771                                                         VM_SUBSYS_PML4);
1772         }
1773
1774         /*
1775          * Allocate the page directory page, which wires it even though
1776          * it isn't being entered into some higher level page table (it
1777          * being the highest level).  If one is already cached we don't
1778          * have to do anything.
1779          */
1780         if ((pv = pmap->pm_pmlpv) == NULL) {
1781                 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1782                 pmap->pm_pmlpv = pv;
1783                 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1784                             VM_PAGE_TO_PHYS(pv->pv_m));
1785                 pv_put(pv);
1786
1787                 /*
1788                  * Install DMAP and KMAP.
1789                  */
1790                 for (j = 0; j < NDMPML4E; ++j) {
1791                         pmap->pm_pml4[DMPML4I + j] =
1792                             (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1793                             pmap->pmap_bits[PG_RW_IDX] |
1794                             pmap->pmap_bits[PG_V_IDX] |
1795                             pmap->pmap_bits[PG_U_IDX];
1796                 }
1797                 pmap->pm_pml4[KPML4I] = KPDPphys |
1798                     pmap->pmap_bits[PG_RW_IDX] |
1799                     pmap->pmap_bits[PG_V_IDX] |
1800                     pmap->pmap_bits[PG_U_IDX];
1801
1802                 /*
1803                  * install self-referential address mapping entry
1804                  */
1805                 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1806                     pmap->pmap_bits[PG_V_IDX] |
1807                     pmap->pmap_bits[PG_RW_IDX] |
1808                     pmap->pmap_bits[PG_A_IDX] |
1809                     pmap->pmap_bits[PG_M_IDX];
1810         } else {
1811                 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1812                 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1813         }
1814         KKASSERT(pmap->pm_pml4[255] == 0);
1815         KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1816         KKASSERT(pv->pv_entry.rbe_left == NULL);
1817         KKASSERT(pv->pv_entry.rbe_right == NULL);
1818 }
1819
1820 /*
1821  * Clean up a pmap structure so it can be physically freed.  This routine
1822  * is called by the vmspace dtor function.  A great deal of pmap data is
1823  * left passively mapped to improve vmspace management so we have a bit
1824  * of cleanup work to do here.
1825  */
1826 void
1827 pmap_puninit(pmap_t pmap)
1828 {
1829         pv_entry_t pv;
1830         vm_page_t p;
1831
1832         KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1833         if ((pv = pmap->pm_pmlpv) != NULL) {
1834                 if (pv_hold_try(pv) == 0)
1835                         pv_lock(pv);
1836                 KKASSERT(pv == pmap->pm_pmlpv);
1837                 p = pmap_remove_pv_page(pv);
1838                 pv_free(pv, NULL, 1);
1839                 pv = NULL;      /* safety */
1840                 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1841                 vm_page_busy_wait(p, FALSE, "pgpun");
1842                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1843                 vm_page_unwire(p, 0);
1844                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1845
1846                 /*
1847                  * XXX eventually clean out PML4 static entries and
1848                  * use vm_page_free_zero()
1849                  */
1850                 vm_page_free(p);
1851                 pmap->pm_pmlpv = NULL;
1852         }
1853         if (pmap->pm_pml4) {
1854                 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1855                 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1856                 pmap->pm_pml4 = NULL;
1857         }
1858         KKASSERT(pmap->pm_stats.resident_count == 0);
1859         KKASSERT(pmap->pm_stats.wired_count == 0);
1860 }
1861
1862 /*
1863  * Wire in kernel global address entries.  To avoid a race condition
1864  * between pmap initialization and pmap_growkernel, this procedure
1865  * adds the pmap to the master list (which growkernel scans to update),
1866  * then copies the template.
1867  */
1868 void
1869 pmap_pinit2(struct pmap *pmap)
1870 {
1871         spin_lock(&pmap_spin);
1872         TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1873         spin_unlock(&pmap_spin);
1874 }
1875
1876 /*
1877  * This routine is called when various levels in the page table need to
1878  * be populated.  This routine cannot fail.
1879  *
1880  * This function returns two locked pv_entry's, one representing the
1881  * requested pv and one representing the requested pv's parent pv.  If
1882  * an intermediate page table does not exist it will be created, mapped,
1883  * wired, and the parent page table will be given an additional hold
1884  * count representing the presence of the child pv_entry.
1885  */
1886 static
1887 pv_entry_t
1888 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1889 {
1890         pt_entry_t *ptep;
1891         pv_entry_t pv;
1892         pv_entry_t pvp;
1893         vm_pindex_t pt_pindex;
1894         vm_page_t m;
1895         int isnew;
1896         int ispt;
1897
1898         /*
1899          * If the pv already exists and we aren't being asked for the
1900          * parent page table page we can just return it.  A locked+held pv
1901          * is returned.  The pv will also have a second hold related to the
1902          * pmap association that we don't have to worry about.
1903          */
1904         ispt = 0;
1905         pv = pv_alloc(pmap, ptepindex, &isnew);
1906         if (isnew == 0 && pvpp == NULL)
1907                 return(pv);
1908
1909         /*
1910          * Special case terminal PVs.  These are not page table pages so
1911          * no vm_page is allocated (the caller supplied the vm_page).  If
1912          * pvpp is non-NULL we are being asked to also removed the pt_pv
1913          * for this pv.
1914          *
1915          * Note that pt_pv's are only returned for user VAs. We assert that
1916          * a pt_pv is not being requested for kernel VAs.  The kernel
1917          * pre-wires all higher-level page tables so don't overload managed
1918          * higher-level page tables on top of it!
1919          */
1920         if (ptepindex < pmap_pt_pindex(0)) {
1921                 if (ptepindex >= NUPTE_USER) {
1922                         /* kernel manages this manually for KVM */
1923                         KKASSERT(pvpp == NULL);
1924                 } else {
1925                         KKASSERT(pvpp != NULL);
1926                         pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1927                         pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1928                         if (isnew) {
1929                                 vm_page_wire_quick(pvp->pv_m);
1930                                 if (pvpp)
1931                                         *pvpp = pvp;
1932                                 else
1933                                         pv_put(pvp);
1934                         } else {
1935                                 *pvpp = pvp;
1936                         }
1937                 }
1938                 return(pv);
1939         }
1940
1941         /*
1942          * The kernel never uses managed PT/PD/PDP pages.
1943          */
1944         KKASSERT(pmap != &kernel_pmap);
1945
1946         /*
1947          * Non-terminal PVs allocate a VM page to represent the page table,
1948          * so we have to resolve pvp and calculate ptepindex for the pvp
1949          * and then for the page table entry index in the pvp for
1950          * fall-through.
1951          */
1952         if (ptepindex < pmap_pd_pindex(0)) {
1953                 /*
1954                  * pv is PT, pvp is PD
1955                  */
1956                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1957                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1958                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1959                 if (!isnew)
1960                         goto notnew;
1961
1962                 /*
1963                  * PT index in PD
1964                  */
1965                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1966                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1967                 ispt = 1;
1968         } else if (ptepindex < pmap_pdp_pindex(0)) {
1969                 /*
1970                  * pv is PD, pvp is PDP
1971                  *
1972                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1973                  *                   the PD.
1974                  */
1975                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1976                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1977
1978                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1979                         KKASSERT(pvpp == NULL);
1980                         pvp = NULL;
1981                 } else {
1982                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
1983                 }
1984                 if (!isnew)
1985                         goto notnew;
1986
1987                 /*
1988                  * PD index in PDP
1989                  */
1990                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1991                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1992         } else if (ptepindex < pmap_pml4_pindex()) {
1993                 /*
1994                  * pv is PDP, pvp is the root pml4 table
1995                  */
1996                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1997                 if (!isnew)
1998                         goto notnew;
1999
2000                 /*
2001                  * PDP index in PML4
2002                  */
2003                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2004                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2005         } else {
2006                 /*
2007                  * pv represents the top-level PML4, there is no parent.
2008                  */
2009                 pvp = NULL;
2010                 if (!isnew)
2011                         goto notnew;
2012         }
2013
2014         /*
2015          * (isnew) is TRUE, pv is not terminal.
2016          *
2017          * (1) Add a wire count to the parent page table (pvp).
2018          * (2) Allocate a VM page for the page table.
2019          * (3) Enter the VM page into the parent page table.
2020          *
2021          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2022          */
2023         if (pvp)
2024                 vm_page_wire_quick(pvp->pv_m);
2025
2026         for (;;) {
2027                 m = vm_page_alloc(NULL, pv->pv_pindex,
2028                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2029                                   VM_ALLOC_INTERRUPT);
2030                 if (m)
2031                         break;
2032                 vm_wait(0);
2033         }
2034         vm_page_spin_lock(m);
2035         pmap_page_stats_adding(m);
2036         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2037         pv->pv_m = m;
2038         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2039         vm_page_spin_unlock(m);
2040         vm_page_unmanage(m);    /* m must be spinunlocked */
2041
2042         pmap_zero_page(VM_PAGE_TO_PHYS(m));
2043         m->valid = VM_PAGE_BITS_ALL;
2044         vm_page_wire(m);        /* wire for mapping in parent */
2045
2046         /*
2047          * Wire the page into pvp.  Bump the resident_count for the pmap.
2048          * There is no pvp for the top level, address the pm_pml4[] array
2049          * directly.
2050          *
2051          * If the caller wants the parent we return it, otherwise
2052          * we just put it away.
2053          *
2054          * No interlock is needed for pte 0 -> non-zero.
2055          *
2056          * In the situation where *ptep is valid we might have an unmanaged
2057          * page table page shared from another page table which we need to
2058          * unshare before installing our private page table page.
2059          */
2060         if (pvp) {
2061                 ptep = pv_pte_lookup(pvp, ptepindex);
2062                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2063                         pt_entry_t pte;
2064
2065                         if (ispt == 0) {
2066                                 panic("pmap_allocpte: unexpected pte %p/%d",
2067                                       pvp, (int)ptepindex);
2068                         }
2069                         pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
2070                         if (vm_page_unwire_quick(
2071                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2072                                 panic("pmap_allocpte: shared pgtable "
2073                                       "pg bad wirecount");
2074                         }
2075                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
2076                 }
2077                 *ptep = VM_PAGE_TO_PHYS(m) |
2078                         (pmap->pmap_bits[PG_U_IDX] |
2079                          pmap->pmap_bits[PG_RW_IDX] |
2080                          pmap->pmap_bits[PG_V_IDX] |
2081                          pmap->pmap_bits[PG_A_IDX] |
2082                          pmap->pmap_bits[PG_M_IDX]);
2083         }
2084         vm_page_wakeup(m);
2085 notnew:
2086         if (pvpp)
2087                 *pvpp = pvp;
2088         else if (pvp)
2089                 pv_put(pvp);
2090         return (pv);
2091 }
2092
2093 /*
2094  * This version of pmap_allocpte() checks for possible segment optimizations
2095  * that would allow page-table sharing.  It can be called for terminal
2096  * page or page table page ptepindex's.
2097  *
2098  * The function is called with page table page ptepindex's for fictitious
2099  * and unmanaged terminal pages.  That is, we don't want to allocate a
2100  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
2101  * for this case.
2102  *
2103  * This function can return a pv and *pvpp associated with the passed in pmap
2104  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
2105  * an unmanaged page table page will be entered into the pass in pmap.
2106  */
2107 static
2108 pv_entry_t
2109 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2110                   vm_map_entry_t entry, vm_offset_t va)
2111 {
2112         vm_object_t object;
2113         pmap_t obpmap;
2114         pmap_t *obpmapp;
2115         vm_offset_t b;
2116         pv_entry_t pte_pv;      /* in original or shared pmap */
2117         pv_entry_t pt_pv;       /* in original or shared pmap */
2118         pv_entry_t proc_pd_pv;  /* in original pmap */
2119         pv_entry_t proc_pt_pv;  /* in original pmap */
2120         pv_entry_t xpv;         /* PT in shared pmap */
2121         pd_entry_t *pt;         /* PT entry in PD of original pmap */
2122         pd_entry_t opte;        /* contents of *pt */
2123         pd_entry_t npte;        /* contents of *pt */
2124         vm_page_t m;
2125
2126 retry:
2127         /*
2128          * Basic tests, require a non-NULL vm_map_entry, require proper
2129          * alignment and type for the vm_map_entry, require that the
2130          * underlying object already be allocated.
2131          *
2132          * We allow almost any type of object to use this optimization.
2133          * The object itself does NOT have to be sized to a multiple of the
2134          * segment size, but the memory mapping does.
2135          *
2136          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2137          *     won't work as expected.
2138          */
2139         if (entry == NULL ||
2140             pmap_mmu_optimize == 0 ||                   /* not enabled */
2141             (pmap->pm_flags & PMAP_HVM) ||              /* special pmap */
2142             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal or pt */
2143             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2144             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2145             entry->object.vm_object == NULL ||          /* needs VM object */
2146             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2147             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2148             (entry->offset & SEG_MASK) ||               /* must be aligned */
2149             (entry->start & SEG_MASK)) {
2150                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2151         }
2152
2153         /*
2154          * Make sure the full segment can be represented.
2155          */
2156         b = va & ~(vm_offset_t)SEG_MASK;
2157         if (b < entry->start || b + SEG_SIZE > entry->end)
2158                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2159
2160         /*
2161          * If the full segment can be represented dive the VM object's
2162          * shared pmap, allocating as required.
2163          */
2164         object = entry->object.vm_object;
2165
2166         if (entry->protection & VM_PROT_WRITE)
2167                 obpmapp = &object->md.pmap_rw;
2168         else
2169                 obpmapp = &object->md.pmap_ro;
2170
2171 #ifdef PMAP_DEBUG2
2172         if (pmap_enter_debug > 0) {
2173                 --pmap_enter_debug;
2174                 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2175                         "obpmapp %p %p\n",
2176                         va, entry->protection, object,
2177                         obpmapp, *obpmapp);
2178                 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2179                         entry, entry->start, entry->end);
2180         }
2181 #endif
2182
2183         /*
2184          * We allocate what appears to be a normal pmap but because portions
2185          * of this pmap are shared with other unrelated pmaps we have to
2186          * set pm_active to point to all cpus.
2187          *
2188          * XXX Currently using pmap_spin to interlock the update, can't use
2189          *     vm_object_hold/drop because the token might already be held
2190          *     shared OR exclusive and we don't know.
2191          */
2192         while ((obpmap = *obpmapp) == NULL) {
2193                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2194                 pmap_pinit_simple(obpmap);
2195                 pmap_pinit2(obpmap);
2196                 spin_lock(&pmap_spin);
2197                 if (*obpmapp != NULL) {
2198                         /*
2199                          * Handle race
2200                          */
2201                         spin_unlock(&pmap_spin);
2202                         pmap_release(obpmap);
2203                         pmap_puninit(obpmap);
2204                         kfree(obpmap, M_OBJPMAP);
2205                         obpmap = *obpmapp; /* safety */
2206                 } else {
2207                         obpmap->pm_active = smp_active_mask;
2208                         obpmap->pm_flags |= PMAP_SEGSHARED;
2209                         *obpmapp = obpmap;
2210                         spin_unlock(&pmap_spin);
2211                 }
2212         }
2213
2214         /*
2215          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2216          * pte/pt using the shared pmap from the object but also adjust
2217          * the process pmap's page table page as a side effect.
2218          */
2219
2220         /*
2221          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2222          * we will return.  This is true if ptepindex represents a terminal
2223          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2224          * the PD.
2225          */
2226         pt_pv = NULL;
2227         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2228         if (ptepindex >= pmap_pt_pindex(0))
2229                 xpv = pte_pv;
2230         else
2231                 xpv = pt_pv;
2232
2233         /*
2234          * Resolve the PD in the process pmap so we can properly share the
2235          * page table page.  Lock order is bottom-up (leaf first)!
2236          *
2237          * NOTE: proc_pt_pv can be NULL.
2238          */
2239         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2240         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2241 #ifdef PMAP_DEBUG2
2242         if (pmap_enter_debug > 0) {
2243                 --pmap_enter_debug;
2244                 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2245                         proc_pt_pv,
2246                         (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2247                         proc_pd_pv,
2248                         va);
2249         }
2250 #endif
2251
2252         /*
2253          * xpv is the page table page pv from the shared object
2254          * (for convenience), from above.
2255          *
2256          * Calculate the pte value for the PT to load into the process PD.
2257          * If we have to change it we must properly dispose of the previous
2258          * entry.
2259          */
2260         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2261         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2262                (pmap->pmap_bits[PG_U_IDX] |
2263                 pmap->pmap_bits[PG_RW_IDX] |
2264                 pmap->pmap_bits[PG_V_IDX] |
2265                 pmap->pmap_bits[PG_A_IDX] |
2266                 pmap->pmap_bits[PG_M_IDX]);
2267
2268         /*
2269          * Dispose of previous page table page if it was local to the
2270          * process pmap.  If the old pt is not empty we cannot dispose of it
2271          * until we clean it out.  This case should not arise very often so
2272          * it is not optimized.
2273          */
2274         if (proc_pt_pv) {
2275                 pmap_inval_bulk_t bulk;
2276
2277                 if (proc_pt_pv->pv_m->wire_count != 1) {
2278                         pv_put(proc_pd_pv);
2279                         pv_put(proc_pt_pv);
2280                         pv_put(pt_pv);
2281                         pv_put(pte_pv);
2282                         pmap_remove(pmap,
2283                                     va & ~(vm_offset_t)SEG_MASK,
2284                                     (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2285                         goto retry;
2286                 }
2287
2288                 /*
2289                  * The release call will indirectly clean out *pt
2290                  */
2291                 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2292                 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2293                 pmap_inval_bulk_flush(&bulk);
2294                 proc_pt_pv = NULL;
2295                 /* relookup */
2296                 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2297         }
2298
2299         /*
2300          * Handle remaining cases.
2301          */
2302         if (*pt == 0) {
2303                 *pt = npte;
2304                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2305                 vm_page_wire_quick(proc_pd_pv->pv_m);   /* proc pd for sh pt */
2306                 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2307         } else if (*pt != npte) {
2308                 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2309
2310 #if 0
2311                 opte = pte_load_clear(pt);
2312                 KKASSERT(opte && opte != npte);
2313
2314                 *pt = npte;
2315 #endif
2316                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2317
2318                 /*
2319                  * Clean up opte, bump the wire_count for the process
2320                  * PD page representing the new entry if it was
2321                  * previously empty.
2322                  *
2323                  * If the entry was not previously empty and we have
2324                  * a PT in the proc pmap then opte must match that
2325                  * pt.  The proc pt must be retired (this is done
2326                  * later on in this procedure).
2327                  *
2328                  * NOTE: replacing valid pte, wire_count on proc_pd_pv
2329                  * stays the same.
2330                  */
2331                 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2332                 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2333                 if (vm_page_unwire_quick(m)) {
2334                         panic("pmap_allocpte_seg: "
2335                               "bad wire count %p",
2336                               m);
2337                 }
2338         }
2339
2340         /*
2341          * The existing process page table was replaced and must be destroyed
2342          * here.
2343          */
2344         if (proc_pd_pv)
2345                 pv_put(proc_pd_pv);
2346         if (pvpp)
2347                 *pvpp = pt_pv;
2348         else
2349                 pv_put(pt_pv);
2350
2351         return (pte_pv);
2352 }
2353
2354 /*
2355  * Release any resources held by the given physical map.
2356  *
2357  * Called when a pmap initialized by pmap_pinit is being released.  Should
2358  * only be called if the map contains no valid mappings.
2359  *
2360  * Caller must hold pmap->pm_token
2361  */
2362 struct pmap_release_info {
2363         pmap_t  pmap;
2364         int     retry;
2365         pv_entry_t pvp;
2366 };
2367
2368 static int pmap_release_callback(pv_entry_t pv, void *data);
2369
2370 void
2371 pmap_release(struct pmap *pmap)
2372 {
2373         struct pmap_release_info info;
2374
2375         KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2376                 ("pmap still active! %016jx",
2377                 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2378
2379         spin_lock(&pmap_spin);
2380         TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
2381         spin_unlock(&pmap_spin);
2382
2383         /*
2384          * Pull pv's off the RB tree in order from low to high and release
2385          * each page.
2386          */
2387         info.pmap = pmap;
2388         do {
2389                 info.retry = 0;
2390                 info.pvp = NULL;
2391
2392                 spin_lock(&pmap->pm_spin);
2393                 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2394                         pmap_release_callback, &info);
2395                 spin_unlock(&pmap->pm_spin);
2396
2397                 if (info.pvp)
2398                         pv_put(info.pvp);
2399         } while (info.retry);
2400
2401
2402         /*
2403          * One resident page (the pml4 page) should remain.
2404          * No wired pages should remain.
2405          */
2406         KKASSERT(pmap->pm_stats.resident_count ==
2407                  ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2408
2409         KKASSERT(pmap->pm_stats.wired_count == 0);
2410 }
2411
2412 /*
2413  * Called from low to high.  We must cache the proper parent pv so we
2414  * can adjust its wired count.
2415  */
2416 static int
2417 pmap_release_callback(pv_entry_t pv, void *data)
2418 {
2419         struct pmap_release_info *info = data;
2420         pmap_t pmap = info->pmap;
2421         vm_pindex_t pindex;
2422         int r;
2423
2424         if (info->pvp == pv) {
2425                 spin_unlock(&pmap->pm_spin);
2426                 info->pvp = NULL;
2427         } else if (pv_hold_try(pv)) {
2428                 spin_unlock(&pmap->pm_spin);
2429         } else {
2430                 spin_unlock(&pmap->pm_spin);
2431                 pv_lock(pv);
2432         }
2433         if (pv->pv_pmap != pmap) {
2434                 pv_put(pv);
2435                 spin_lock(&pmap->pm_spin);
2436                 info->retry = 1;
2437                 return(-1);
2438         }
2439
2440         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2441                 /*
2442                  * parent is PT
2443                  */
2444                 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2445                 pindex += NUPTE_TOTAL;
2446         } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2447                 /*
2448                  * parent is PD
2449                  */
2450                 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2451                 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2452         } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2453                 /*
2454                  * parent is PDP
2455                  */
2456                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2457                          NPDPEPGSHIFT;
2458                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2459         } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2460                 /*
2461                  * parent is PML4 (there's only one)
2462                  */
2463 #if 0
2464                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2465                            NUPD_TOTAL) >> NPML4EPGSHIFT;
2466                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2467 #endif
2468                 pindex = pmap_pml4_pindex();
2469         } else {
2470                 /*
2471                  * parent is NULL
2472                  */
2473                 if (info->pvp) {
2474                         pv_put(info->pvp);
2475                         info->pvp = NULL;
2476                 }
2477                 pindex = 0;
2478         }
2479         if (pindex) {
2480                 if (info->pvp && info->pvp->pv_pindex != pindex) {
2481                         pv_put(info->pvp);
2482                         info->pvp = NULL;
2483                 }
2484                 if (info->pvp == NULL)
2485                         info->pvp = pv_get(pmap, pindex);
2486         } else {
2487                 if (info->pvp) {
2488                         pv_put(info->pvp);
2489                         info->pvp = NULL;
2490                 }
2491         }
2492         r = pmap_release_pv(pv, info->pvp, NULL);
2493         spin_lock(&pmap->pm_spin);
2494         return(r);
2495 }
2496
2497 /*
2498  * Called with held (i.e. also locked) pv.  This function will dispose of
2499  * the lock along with the pv.
2500  *
2501  * If the caller already holds the locked parent page table for pv it
2502  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2503  * pass NULL for pvp.
2504  */
2505 static int
2506 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2507 {
2508         vm_page_t p;
2509
2510         /*
2511          * The pmap is currently not spinlocked, pv is held+locked.
2512          * Remove the pv's page from its parent's page table.  The
2513          * parent's page table page's wire_count will be decremented.
2514          *
2515          * This will clean out the pte at any level of the page table.
2516          * If smp != 0 all cpus are affected.
2517          *
2518          * Do not tear-down recursively, its faster to just let the
2519          * release run its course.
2520          */
2521         pmap_remove_pv_pte(pv, pvp, bulk, 0);
2522
2523         /*
2524          * Terminal pvs are unhooked from their vm_pages.  Because
2525          * terminal pages aren't page table pages they aren't wired
2526          * by us, so we have to be sure not to unwire them either.
2527          */
2528         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2529                 pmap_remove_pv_page(pv);
2530                 goto skip;
2531         }
2532
2533         /*
2534          * We leave the top-level page table page cached, wired, and
2535          * mapped in the pmap until the dtor function (pmap_puninit())
2536          * gets called.
2537          *
2538          * Since we are leaving the top-level pv intact we need
2539          * to break out of what would otherwise be an infinite loop.
2540          */
2541         if (pv->pv_pindex == pmap_pml4_pindex()) {
2542                 pv_put(pv);
2543                 return(-1);
2544         }
2545
2546         /*
2547          * For page table pages (other than the top-level page),
2548          * remove and free the vm_page.  The representitive mapping
2549          * removed above by pmap_remove_pv_pte() did not undo the
2550          * last wire_count so we have to do that as well.
2551          */
2552         p = pmap_remove_pv_page(pv);
2553         vm_page_busy_wait(p, FALSE, "pmaprl");
2554         if (p->wire_count != 1) {
2555                 kprintf("p->wire_count was %016lx %d\n",
2556                         pv->pv_pindex, p->wire_count);
2557         }
2558         KKASSERT(p->wire_count == 1);
2559         KKASSERT(p->flags & PG_UNMANAGED);
2560
2561         vm_page_unwire(p, 0);
2562         KKASSERT(p->wire_count == 0);
2563
2564         vm_page_free(p);
2565 skip:
2566         pv_free(pv, pvp, 1);
2567
2568         return 0;
2569 }
2570
2571 /*
2572  * This function will remove the pte associated with a pv from its parent.
2573  * Terminal pv's are supported.  All cpus specified by (bulk) are properly
2574  * invalidated.
2575  *
2576  * The wire count will be dropped on the parent page table.  The wire
2577  * count on the page being removed (pv->pv_m) from the parent page table
2578  * is NOT touched.  Note that terminal pages will not have any additional
2579  * wire counts while page table pages will have at least one representing
2580  * the mapping, plus others representing sub-mappings.
2581  *
2582  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2583  *       pages and user page table and terminal pages.
2584  *
2585  * The pv must be locked.  The pvp, if supplied, must be locked.  All
2586  * supplied pv's will remain locked on return.
2587  *
2588  * XXX must lock parent pv's if they exist to remove pte XXX
2589  */
2590 static
2591 void
2592 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2593                    int destroy)
2594 {
2595         vm_pindex_t ptepindex = pv->pv_pindex;
2596         pmap_t pmap = pv->pv_pmap;
2597         vm_page_t p;
2598         int gotpvp = 0;
2599
2600         KKASSERT(pmap);
2601
2602         if (ptepindex == pmap_pml4_pindex()) {
2603                 /*
2604                  * We are the top level pml4 table, there is no parent.
2605                  */
2606                 p = pmap->pm_pmlpv->pv_m;
2607         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2608                 /*
2609                  * Remove a PDP page from the pml4e.  This can only occur
2610                  * with user page tables.  We do not have to lock the
2611                  * pml4 PV so just ignore pvp.
2612                  */
2613                 vm_pindex_t pml4_pindex;
2614                 vm_pindex_t pdp_index;
2615                 pml4_entry_t *pdp;
2616
2617                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2618                 if (pvp == NULL) {
2619                         pml4_pindex = pmap_pml4_pindex();
2620                         pvp = pv_get(pv->pv_pmap, pml4_pindex);
2621                         KKASSERT(pvp);
2622                         gotpvp = 1;
2623                 }
2624                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2625                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2626                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2627                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2628         } else if (ptepindex >= pmap_pd_pindex(0)) {
2629                 /*
2630                  * Remove a PD page from the pdp
2631                  *
2632                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2633                  *                   of a simple pmap because it stops at
2634                  *                   the PD page.
2635                  */
2636                 vm_pindex_t pdp_pindex;
2637                 vm_pindex_t pd_index;
2638                 pdp_entry_t *pd;
2639
2640                 pd_index = ptepindex - pmap_pd_pindex(0);
2641
2642                 if (pvp == NULL) {
2643                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2644                                      (pd_index >> NPML4EPGSHIFT);
2645                         pvp = pv_get(pv->pv_pmap, pdp_pindex);
2646                         gotpvp = 1;
2647                 }
2648                 if (pvp) {
2649                         pd = pv_pte_lookup(pvp, pd_index &
2650                                                 ((1ul << NPDPEPGSHIFT) - 1));
2651                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2652                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2653                         pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2654                 } else {
2655                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2656                         p = pv->pv_m;           /* degenerate test later */
2657                 }
2658         } else if (ptepindex >= pmap_pt_pindex(0)) {
2659                 /*
2660                  *  Remove a PT page from the pd
2661                  */
2662                 vm_pindex_t pd_pindex;
2663                 vm_pindex_t pt_index;
2664                 pd_entry_t *pt;
2665
2666                 pt_index = ptepindex - pmap_pt_pindex(0);
2667
2668                 if (pvp == NULL) {
2669                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2670                                     (pt_index >> NPDPEPGSHIFT);
2671                         pvp = pv_get(pv->pv_pmap, pd_pindex);
2672                         KKASSERT(pvp);
2673                         gotpvp = 1;
2674                 }
2675                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2676                 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2677                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2678                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2679         } else {
2680                 /*
2681                  * Remove a managed PTE from the PT page.  Userland pmaps
2682                  * manage PT/PD/PDP page tables pages but the kernel_pmap
2683                  * does not.
2684                  *
2685                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2686                  *       pv is a pte_pv so we can safely lock pt_pv.
2687                  *
2688                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2689                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2690                  *       terminal ptes.
2691                  */
2692                 vm_pindex_t pt_pindex;
2693                 pt_entry_t *ptep;
2694                 pt_entry_t pte;
2695                 vm_offset_t va;
2696
2697                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2698                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2699
2700                 if (ptepindex >= NUPTE_USER) {
2701                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2702                         KKASSERT(pvp == NULL);
2703                 } else {
2704                         if (pvp == NULL) {
2705                                 pt_pindex = NUPTE_TOTAL +
2706                                             (ptepindex >> NPDPEPGSHIFT);
2707                                 pvp = pv_get(pv->pv_pmap, pt_pindex);
2708                                 KKASSERT(pvp);
2709                                 gotpvp = 1;
2710                         }
2711                         ptep = pv_pte_lookup(pvp, ptepindex &
2712                                                   ((1ul << NPDPEPGSHIFT) - 1));
2713                 }
2714                 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2715                 if (bulk == NULL)               /* XXX */
2716                         cpu_invlpg((void *)va); /* XXX */
2717
2718                 /*
2719                  * Now update the vm_page_t
2720                  */
2721                 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] |
2722                             pmap->pmap_bits[PG_V_IDX])) !=
2723                     (pmap->pmap_bits[PG_MANAGED_IDX] |
2724                      pmap->pmap_bits[PG_V_IDX])) {
2725                         kprintf("remove_pte badpte %016lx %016lx %d\n",
2726                                 pte, pv->pv_pindex,
2727                                 pv->pv_pindex < pmap_pt_pindex(0));
2728                 }
2729
2730                 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2731                 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2732                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2733                         p = pv->pv_m;
2734                 else
2735                         p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2736                 /* p = pv->pv_m; */
2737
2738                 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2739                         if (pmap_track_modified(ptepindex))
2740                                 vm_page_dirty(p);
2741                 }
2742                 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2743                         vm_page_flag_set(p, PG_REFERENCED);
2744                 }
2745                 if (pte & pmap->pmap_bits[PG_W_IDX])
2746                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
2747                 if (pte & pmap->pmap_bits[PG_G_IDX])
2748                         cpu_invlpg((void *)va);
2749         }
2750         KKASSERT(pv->pv_m == p);        /* XXX remove me later */
2751
2752         /*
2753          * If requested, scrap the underlying pv->pv_m and the underlying
2754          * pv.  If this is a page-table-page we must also free the page.
2755          *
2756          * pvp must be returned locked.
2757          */
2758         if (destroy == 1) {
2759                 /*
2760                  * page table page (PT, PD, PDP, PML4), caller was responsible
2761                  * for testing wired_count.
2762                  */
2763                 vm_page_t p;
2764
2765                 KKASSERT(pv->pv_m->wire_count == 1);
2766                 p = pmap_remove_pv_page(pv);
2767                 pv_free(pv, pvp, 1);
2768                 pv = NULL;
2769
2770                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
2771                 vm_page_busy_wait(p, FALSE, "pgpun");
2772                 vm_page_unwire(p, 0);
2773                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2774                 vm_page_free(p);
2775         } else if (destroy == 2) {
2776                 /*
2777                  * Normal page (leave page untouched)
2778                  */
2779                 pmap_remove_pv_page(pv);
2780                 pv_free(pv, pvp, 1);
2781                 pv = NULL;              /* safety */
2782         }
2783
2784         /*
2785          * If we acquired pvp ourselves then we are responsible for
2786          * recursively deleting it.
2787          */
2788         if (pvp && gotpvp) {
2789                 /*
2790                  * Recursively destroy higher-level page tables.
2791                  *
2792                  * This is optional.  If we do not, they will still
2793                  * be destroyed when the process exits.
2794                  */
2795                 if (pmap_dynamic_delete &&
2796                     pvp->pv_m &&
2797                     pvp->pv_m->wire_count == 1 &&
2798                     pvp->pv_pindex != pmap_pml4_pindex()) {
2799                         if (pmap != &kernel_pmap) {
2800                                 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
2801                                 pvp = NULL;     /* safety */
2802                         } else {
2803                                 kprintf("Attempt to remove kernel_pmap pindex "
2804                                         "%jd\n", pvp->pv_pindex);
2805                                 pv_put(pvp);
2806                         }
2807                 } else {
2808                         pv_put(pvp);
2809                 }
2810         }
2811 }
2812
2813 /*
2814  * Remove the vm_page association to a pv.  The pv must be locked.
2815  */
2816 static
2817 vm_page_t
2818 pmap_remove_pv_page(pv_entry_t pv)
2819 {
2820         vm_page_t m;
2821
2822         m = pv->pv_m;
2823         KKASSERT(m);
2824         vm_page_spin_lock(m);
2825         pv->pv_m = NULL;
2826         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2827         pmap_page_stats_deleting(m);
2828         /*
2829         if (m->object)
2830                 atomic_add_int(&m->object->agg_pv_list_count, -1);
2831         */
2832         if (TAILQ_EMPTY(&m->md.pv_list))
2833                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2834         vm_page_spin_unlock(m);
2835
2836         return(m);
2837 }
2838
2839 /*
2840  * Grow the number of kernel page table entries, if needed.
2841  *
2842  * This routine is always called to validate any address space
2843  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
2844  * space below KERNBASE.
2845  */
2846 void
2847 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2848 {
2849         vm_paddr_t paddr;
2850         vm_offset_t ptppaddr;
2851         vm_page_t nkpg;
2852         pd_entry_t *pt, newpt;
2853         pdp_entry_t newpd;
2854         int update_kernel_vm_end;
2855
2856         /*
2857          * bootstrap kernel_vm_end on first real VM use
2858          */
2859         if (kernel_vm_end == 0) {
2860                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2861                 nkpt = 0;
2862                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2863                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2864                                         ~(PAGE_SIZE * NPTEPG - 1);
2865                         nkpt++;
2866                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2867                                 kernel_vm_end = kernel_map.max_offset;
2868                                 break;
2869                         }
2870                 }
2871         }
2872
2873         /*
2874          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
2875          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
2876          * do not want to force-fill 128G worth of page tables.
2877          */
2878         if (kstart < KERNBASE) {
2879                 if (kstart > kernel_vm_end)
2880                         kstart = kernel_vm_end;
2881                 KKASSERT(kend <= KERNBASE);
2882                 update_kernel_vm_end = 1;
2883         } else {
2884                 update_kernel_vm_end = 0;
2885         }
2886
2887         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2888         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2889
2890         if (kend - 1 >= kernel_map.max_offset)
2891                 kend = kernel_map.max_offset;
2892
2893         while (kstart < kend) {
2894                 pt = pmap_pt(&kernel_pmap, kstart);
2895                 if (pt == NULL) {
2896                         /* We need a new PD entry */
2897                         nkpg = vm_page_alloc(NULL, nkpt,
2898                                              VM_ALLOC_NORMAL |
2899                                              VM_ALLOC_SYSTEM |
2900                                              VM_ALLOC_INTERRUPT);
2901                         if (nkpg == NULL) {
2902                                 panic("pmap_growkernel: no memory to grow "
2903                                       "kernel");
2904                         }
2905                         paddr = VM_PAGE_TO_PHYS(nkpg);
2906                         pmap_zero_page(paddr);
2907                         newpd = (pdp_entry_t)
2908                             (paddr |
2909                             kernel_pmap.pmap_bits[PG_V_IDX] |
2910                             kernel_pmap.pmap_bits[PG_RW_IDX] |
2911                             kernel_pmap.pmap_bits[PG_A_IDX] |
2912                             kernel_pmap.pmap_bits[PG_M_IDX]);
2913                         *pmap_pd(&kernel_pmap, kstart) = newpd;
2914                         nkpt++;
2915                         continue; /* try again */
2916                 }
2917                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2918                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
2919                                  ~(PAGE_SIZE * NPTEPG - 1);
2920                         if (kstart - 1 >= kernel_map.max_offset) {
2921                                 kstart = kernel_map.max_offset;
2922                                 break;
2923                         }
2924                         continue;
2925                 }
2926
2927                 /*
2928                  * We need a new PT
2929                  *
2930                  * This index is bogus, but out of the way
2931                  */
2932                 nkpg = vm_page_alloc(NULL, nkpt,
2933                                      VM_ALLOC_NORMAL |
2934                                      VM_ALLOC_SYSTEM |
2935                                      VM_ALLOC_INTERRUPT);
2936                 if (nkpg == NULL)
2937                         panic("pmap_growkernel: no memory to grow kernel");
2938
2939                 vm_page_wire(nkpg);
2940                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2941                 pmap_zero_page(ptppaddr);
2942                 newpt = (pd_entry_t) (ptppaddr |
2943                     kernel_pmap.pmap_bits[PG_V_IDX] |
2944                     kernel_pmap.pmap_bits[PG_RW_IDX] |
2945                     kernel_pmap.pmap_bits[PG_A_IDX] |
2946                     kernel_pmap.pmap_bits[PG_M_IDX]);
2947                 *pmap_pt(&kernel_pmap, kstart) = newpt;
2948                 nkpt++;
2949
2950                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2951                           ~(PAGE_SIZE * NPTEPG - 1);
2952
2953                 if (kstart - 1 >= kernel_map.max_offset) {
2954                         kstart = kernel_map.max_offset;
2955                         break;
2956                 }
2957         }
2958
2959         /*
2960          * Only update kernel_vm_end for areas below KERNBASE.
2961          */
2962         if (update_kernel_vm_end && kernel_vm_end < kstart)
2963                 kernel_vm_end = kstart;
2964 }
2965
2966 /*
2967  *      Add a reference to the specified pmap.
2968  */
2969 void
2970 pmap_reference(pmap_t pmap)
2971 {
2972         if (pmap != NULL) {
2973                 lwkt_gettoken(&pmap->pm_token);
2974                 ++pmap->pm_count;
2975                 lwkt_reltoken(&pmap->pm_token);
2976         }
2977 }
2978
2979 /***************************************************
2980  * page management routines.
2981  ***************************************************/
2982
2983 /*
2984  * Hold a pv without locking it
2985  */
2986 static void
2987 pv_hold(pv_entry_t pv)
2988 {
2989         atomic_add_int(&pv->pv_hold, 1);
2990 }
2991
2992 /*
2993  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
2994  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
2995  * the pv properly.
2996  *
2997  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2998  * pv list via its page) must be held by the caller.
2999  */
3000 static int
3001 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3002 {
3003         u_int count;
3004
3005         /*
3006          * Critical path shortcut expects pv to already have one ref
3007          * (for the pv->pv_pmap).
3008          */
3009         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3010 #ifdef PMAP_DEBUG
3011                 pv->pv_func = func;
3012                 pv->pv_line = lineno;
3013 #endif
3014                 return TRUE;
3015         }
3016
3017         for (;;) {
3018                 count = pv->pv_hold;
3019                 cpu_ccfence();
3020                 if ((count & PV_HOLD_LOCKED) == 0) {
3021                         if (atomic_cmpset_int(&pv->pv_hold, count,
3022                                               (count + 1) | PV_HOLD_LOCKED)) {
3023 #ifdef PMAP_DEBUG
3024                                 pv->pv_func = func;
3025                                 pv->pv_line = lineno;
3026 #endif
3027                                 return TRUE;
3028                         }
3029                 } else {
3030                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3031                                 return FALSE;
3032                 }
3033                 /* retry */
3034         }
3035 }
3036
3037 /*
3038  * Drop a previously held pv_entry which could not be locked, allowing its
3039  * destruction.
3040  *
3041  * Must not be called with a spinlock held as we might zfree() the pv if it
3042  * is no longer associated with a pmap and this was the last hold count.
3043  */
3044 static void
3045 pv_drop(pv_entry_t pv)
3046 {
3047         u_int count;
3048
3049         for (;;) {
3050                 count = pv->pv_hold;
3051                 cpu_ccfence();
3052                 KKASSERT((count & PV_HOLD_MASK) > 0);
3053                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3054                          (PV_HOLD_LOCKED | 1));
3055                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3056                         if ((count & PV_HOLD_MASK) == 1) {
3057 #ifdef PMAP_DEBUG2
3058                                 if (pmap_enter_debug > 0) {
3059                                         --pmap_enter_debug;
3060                                         kprintf("pv_drop: free pv %p\n", pv);
3061                                 }
3062 #endif
3063                                 KKASSERT(count == 1);
3064                                 KKASSERT(pv->pv_pmap == NULL);
3065                                 zfree(pvzone, pv);
3066                         }
3067                         return;
3068                 }
3069                 /* retry */
3070         }
3071 }
3072
3073 /*
3074  * Find or allocate the requested PV entry, returning a locked, held pv.
3075  *
3076  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3077  * for the caller and one representing the pmap and vm_page association.
3078  *
3079  * If (*isnew) is zero, the returned pv will have only one hold count.
3080  *
3081  * Since both associations can only be adjusted while the pv is locked,
3082  * together they represent just one additional hold.
3083  */
3084 static
3085 pv_entry_t
3086 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3087 {
3088         pv_entry_t pv;
3089         pv_entry_t pnew = NULL;
3090
3091         spin_lock(&pmap->pm_spin);
3092         for (;;) {
3093                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3094                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3095                                                         pindex);
3096                 }
3097                 if (pv == NULL) {
3098                         if (pnew == NULL) {
3099                                 spin_unlock(&pmap->pm_spin);
3100                                 pnew = zalloc(pvzone);
3101                                 spin_lock(&pmap->pm_spin);
3102                                 continue;
3103                         }
3104                         pnew->pv_pmap = pmap;
3105                         pnew->pv_pindex = pindex;
3106                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
3107 #ifdef PMAP_DEBUG
3108                         pnew->pv_func = func;
3109                         pnew->pv_line = lineno;
3110 #endif
3111                         pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3112                         ++pmap->pm_generation;
3113                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
3114                         spin_unlock(&pmap->pm_spin);
3115                         *isnew = 1;
3116                         return(pnew);
3117                 }
3118                 if (pnew) {
3119                         spin_unlock(&pmap->pm_spin);
3120                         zfree(pvzone, pnew);
3121                         pnew = NULL;
3122                         spin_lock(&pmap->pm_spin);
3123                         continue;
3124                 }
3125                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3126                         spin_unlock(&pmap->pm_spin);
3127                 } else {
3128                         spin_unlock(&pmap->pm_spin);
3129                         _pv_lock(pv PMAP_DEBUG_COPY);
3130                 }
3131                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3132                         *isnew = 0;
3133                         return(pv);
3134                 }
3135                 pv_put(pv);
3136                 spin_lock(&pmap->pm_spin);
3137         }
3138 }
3139
3140 /*
3141  * Find the requested PV entry, returning a locked+held pv or NULL
3142  */
3143 static
3144 pv_entry_t
3145 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
3146 {
3147         pv_entry_t pv;
3148
3149         spin_lock(&pmap->pm_spin);
3150         for (;;) {
3151                 /*
3152                  * Shortcut cache
3153                  */
3154                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3155                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3156                                                         pindex);
3157                 }
3158                 if (pv == NULL) {
3159                         spin_unlock(&pmap->pm_spin);
3160                         return NULL;
3161                 }
3162                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3163                         spin_unlock(&pmap->pm_spin);
3164                 } else {
3165                         spin_unlock(&pmap->pm_spin);
3166                         _pv_lock(pv PMAP_DEBUG_COPY);
3167                 }
3168                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3169                         pv_cache(pv, pindex);
3170                         return(pv);
3171                 }
3172                 pv_put(pv);
3173                 spin_lock(&pmap->pm_spin);
3174         }
3175 }
3176
3177 /*
3178  * Lookup, hold, and attempt to lock (pmap,pindex).
3179  *
3180  * If the entry does not exist NULL is returned and *errorp is set to 0
3181  *
3182  * If the entry exists and could be successfully locked it is returned and
3183  * errorp is set to 0.
3184  *
3185  * If the entry exists but could NOT be successfully locked it is returned
3186  * held and *errorp is set to 1.
3187  */
3188 static
3189 pv_entry_t
3190 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
3191 {
3192         pv_entry_t pv;
3193
3194         spin_lock_shared(&pmap->pm_spin);
3195         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3196                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3197         if (pv == NULL) {
3198                 spin_unlock_shared(&pmap->pm_spin);
3199                 *errorp = 0;
3200                 return NULL;
3201         }
3202         if (pv_hold_try(pv)) {
3203                 pv_cache(pv, pindex);
3204                 spin_unlock_shared(&pmap->pm_spin);
3205                 *errorp = 0;
3206                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3207                 return(pv);     /* lock succeeded */
3208         }
3209         spin_unlock_shared(&pmap->pm_spin);
3210         *errorp = 1;
3211         return (pv);            /* lock failed */
3212 }
3213
3214 /*
3215  * Find the requested PV entry, returning a held pv or NULL
3216  */
3217 static
3218 pv_entry_t
3219 pv_find(pmap_t pmap, vm_pindex_t pindex)
3220 {
3221         pv_entry_t pv;
3222
3223         spin_lock_shared(&pmap->pm_spin);
3224
3225         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3226                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3227         if (pv == NULL) {
3228                 spin_unlock_shared(&pmap->pm_spin);
3229                 return NULL;
3230         }
3231         pv_hold(pv);
3232         pv_cache(pv, pindex);
3233         spin_unlock_shared(&pmap->pm_spin);
3234         return(pv);
3235 }
3236
3237 /*
3238  * Lock a held pv, keeping the hold count
3239  */
3240 static
3241 void
3242 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3243 {
3244         u_int count;
3245
3246         for (;;) {
3247                 count = pv->pv_hold;
3248                 cpu_ccfence();
3249                 if ((count & PV_HOLD_LOCKED) == 0) {
3250                         if (atomic_cmpset_int(&pv->pv_hold, count,
3251                                               count | PV_HOLD_LOCKED)) {
3252 #ifdef PMAP_DEBUG
3253                                 pv->pv_func = func;
3254                                 pv->pv_line = lineno;
3255 #endif
3256                                 return;
3257                         }
3258                         continue;
3259                 }
3260                 tsleep_interlock(pv, 0);
3261                 if (atomic_cmpset_int(&pv->pv_hold, count,
3262                                       count | PV_HOLD_WAITING)) {
3263 #ifdef PMAP_DEBUG
3264                         kprintf("pv waiting on %s:%d\n",
3265                                         pv->pv_func, pv->pv_line);
3266 #endif
3267                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
3268                 }
3269                 /* retry */
3270         }
3271 }
3272
3273 /*
3274  * Unlock a held and locked pv, keeping the hold count.
3275  */
3276 static
3277 void
3278 pv_unlock(pv_entry_t pv)
3279 {
3280         u_int count;
3281
3282         for (;;) {
3283                 count = pv->pv_hold;
3284                 cpu_ccfence();
3285                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3286                          (PV_HOLD_LOCKED | 1));
3287                 if (atomic_cmpset_int(&pv->pv_hold, count,
3288                                       count &
3289                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3290                         if (count & PV_HOLD_WAITING)
3291                                 wakeup(pv);
3292                         break;
3293                 }
3294         }
3295 }
3296
3297 /*
3298  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3299  * and the hold count drops to zero we will free it.
3300  *
3301  * Caller should not hold any spin locks.  We are protected from hold races
3302  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3303  * lock held.  A pv cannot be located otherwise.
3304  */
3305 static
3306 void
3307 pv_put(pv_entry_t pv)
3308 {
3309 #ifdef PMAP_DEBUG2
3310         if (pmap_enter_debug > 0) {
3311                 --pmap_enter_debug;
3312                 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3313         }
3314 #endif
3315
3316         /*
3317          * Fast - shortcut most common condition
3318          */
3319         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3320                 return;
3321
3322         /*
3323          * Slow
3324          */
3325         pv_unlock(pv);
3326         pv_drop(pv);
3327 }
3328
3329 /*
3330  * Remove the pmap association from a pv, require that pv_m already be removed,
3331  * then unlock and drop the pv.  Any pte operations must have already been
3332  * completed.  This call may result in a last-drop which will physically free
3333  * the pv.
3334  *
3335  * Removing the pmap association entails an additional drop.
3336  *
3337  * pv must be exclusively locked on call and will be disposed of on return.
3338  */
3339 static
3340 void
3341 pv_free(pv_entry_t pv, pv_entry_t pvp, int putaway)
3342 {
3343         pmap_t pmap;
3344
3345         KKASSERT(pv->pv_m == NULL);
3346         KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3347         if ((pmap = pv->pv_pmap) != NULL) {
3348                 spin_lock(&pmap->pm_spin);
3349                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3350                 ++pmap->pm_generation;
3351                 if (pmap->pm_pvhint == pv)
3352                         pmap->pm_pvhint = NULL;
3353                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3354                 pv->pv_pmap = NULL;
3355                 pv->pv_pindex = 0;
3356                 spin_unlock(&pmap->pm_spin);
3357
3358                 /*
3359                  * Try to shortcut three atomic ops, otherwise fall through
3360                  * and do it normally.  Drop two refs and the lock all in
3361                  * one go.
3362                  */
3363                 if (putaway &&
3364                     atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3365 #ifdef PMAP_DEBUG2
3366                         if (pmap_enter_debug > 0) {
3367                                 --pmap_enter_debug;
3368                                 kprintf("pv_free: free pv %p\n", pv);
3369                         }
3370 #endif
3371                         zfree(pvzone, pv);
3372                         if (pvp)
3373                                 vm_page_unwire_quick(pvp->pv_m);
3374                         return;
3375                 }
3376                 pv_drop(pv);    /* ref for pv_pmap */
3377         }
3378         if (putaway)
3379                 pv_put(pv);
3380         if (pvp)
3381                 vm_page_unwire_quick(pvp->pv_m);
3382 }
3383
3384 /*
3385  * This routine is very drastic, but can save the system
3386  * in a pinch.
3387  */
3388 void
3389 pmap_collect(void)
3390 {
3391         int i;
3392         vm_page_t m;
3393         static int warningdone=0;
3394
3395         if (pmap_pagedaemon_waken == 0)
3396                 return;
3397         pmap_pagedaemon_waken = 0;
3398         if (warningdone < 5) {
3399                 kprintf("pmap_collect: collecting pv entries -- "
3400                         "suggest increasing PMAP_SHPGPERPROC\n");
3401                 warningdone++;
3402         }
3403
3404         for (i = 0; i < vm_page_array_size; i++) {
3405                 m = &vm_page_array[i];
3406                 if (m->wire_count || m->hold_count)
3407                         continue;
3408                 if (vm_page_busy_try(m, TRUE) == 0) {
3409                         if (m->wire_count == 0 && m->hold_count == 0) {
3410                                 pmap_remove_all(m);
3411                         }
3412                         vm_page_wakeup(m);
3413                 }
3414         }
3415 }
3416
3417 /*
3418  * Scan the pmap for active page table entries and issue a callback.
3419  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3420  * its parent page table.
3421  *
3422  * pte_pv will be NULL if the page or page table is unmanaged.
3423  * pt_pv will point to the page table page containing the pte for the page.
3424  *
3425  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3426  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3427  *       process pmap's PD and page to the callback function.  This can be
3428  *       confusing because the pt_pv is really a pd_pv, and the target page
3429  *       table page is simply aliased by the pmap and not owned by it.
3430  *
3431  * It is assumed that the start and end are properly rounded to the page size.
3432  *
3433  * It is assumed that PD pages and above are managed and thus in the RB tree,
3434  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3435  */
3436 struct pmap_scan_info {
3437         struct pmap *pmap;
3438         vm_offset_t sva;
3439         vm_offset_t eva;
3440         vm_pindex_t sva_pd_pindex;
3441         vm_pindex_t eva_pd_pindex;
3442         void (*func)(pmap_t, struct pmap_scan_info *,
3443                      pv_entry_t, pv_entry_t, int, vm_offset_t,
3444                      pt_entry_t *, void *);
3445         void *arg;
3446         pmap_inval_bulk_t bulk_core;
3447         pmap_inval_bulk_t *bulk;
3448         int count;
3449         int stop;
3450 };
3451
3452 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3453 static int pmap_scan_callback(pv_entry_t pv, void *data);
3454
3455 static void
3456 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3457 {
3458         struct pmap *pmap = info->pmap;
3459         pv_entry_t pd_pv;       /* A page directory PV */
3460         pv_entry_t pt_pv;       /* A page table PV */
3461         pv_entry_t pte_pv;      /* A page table entry PV */
3462         pt_entry_t *ptep;
3463         pt_entry_t oldpte;
3464         struct pv_entry dummy_pv;
3465         int generation;
3466
3467         info->stop = 0;
3468         if (pmap == NULL)
3469                 return;
3470         if (smp_inval) {
3471                 info->bulk = &info->bulk_core;
3472                 pmap_inval_bulk_init(&info->bulk_core, pmap);
3473         } else {
3474                 info->bulk = NULL;
3475         }
3476
3477         /*
3478          * Hold the token for stability; if the pmap is empty we have nothing
3479          * to do.
3480          */
3481         lwkt_gettoken(&pmap->pm_token);
3482 #if 0
3483         if (pmap->pm_stats.resident_count == 0) {
3484                 lwkt_reltoken(&pmap->pm_token);
3485                 return;
3486         }
3487 #endif
3488
3489         info->count = 0;
3490
3491 again:
3492         /*
3493          * Special handling for scanning one page, which is a very common
3494          * operation (it is?).
3495          *
3496          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3497          */
3498         if (info->sva + PAGE_SIZE == info->eva) {
3499                 generation = pmap->pm_generation;
3500                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3501                         /*
3502                          * Kernel mappings do not track wire counts on
3503                          * page table pages and only maintain pd_pv and
3504                          * pte_pv levels so pmap_scan() works.
3505                          */
3506                         pt_pv = NULL;
3507                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3508                         ptep = vtopte(info->sva);
3509                 } else {
3510                         /*
3511                          * User pages which are unmanaged will not have a
3512                          * pte_pv.  User page table pages which are unmanaged
3513                          * (shared from elsewhere) will also not have a pt_pv.
3514                          * The func() callback will pass both pte_pv and pt_pv
3515                          * as NULL in that case.
3516                          */
3517                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3518                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3519                         if (pt_pv == NULL) {
3520                                 KKASSERT(pte_pv == NULL);
3521                                 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3522                                 if (pd_pv) {
3523                                         ptep = pv_pte_lookup(pd_pv,
3524                                                     pmap_pt_index(info->sva));
3525                                         if (*ptep) {
3526                                                 info->func(pmap, info,
3527                                                      NULL, pd_pv, 1,
3528                                                      info->sva, ptep,
3529                                                      info->arg);
3530                                         }
3531                                         pv_put(pd_pv);
3532                                 }
3533                                 goto fast_skip;
3534                         }
3535                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3536                 }
3537
3538                 /*
3539                  * NOTE: *ptep can't be ripped out from under us if we hold
3540                  *       pte_pv locked, but bits can change.  However, there is
3541                  *       a race where another thread may be inserting pte_pv
3542                  *       and setting *ptep just after our pte_pv lookup fails.
3543                  *
3544                  *       In this situation we can end up with a NULL pte_pv
3545                  *       but find that we have a managed *ptep.  We explicitly
3546                  *       check for this race.
3547                  */
3548                 oldpte = *ptep;
3549                 cpu_ccfence();
3550                 if (oldpte == 0) {
3551                         /*
3552                          * Unlike the pv_find() case below we actually
3553                          * acquired a locked pv in this case so any
3554                          * race should have been resolved.  It is expected
3555                          * to not exist.
3556                          */
3557                         KKASSERT(pte_pv == NULL);
3558                 } else if (pte_pv) {
3559                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3560                                            pmap->pmap_bits[PG_V_IDX])) ==
3561                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
3562                                  pmap->pmap_bits[PG_V_IDX]),
3563                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3564                              "generation %d/%d",
3565                             *ptep, oldpte, info->sva, pte_pv,
3566                             generation, pmap->pm_generation));
3567                         info->func(pmap, info, pte_pv, pt_pv, 0,
3568                                    info->sva, ptep, info->arg);
3569                 } else {
3570                         /*
3571                          * Check for insertion race
3572                          */
3573                         if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3574                             pt_pv) {
3575                                 pte_pv = pv_find(pmap,
3576                                                  pmap_pte_pindex(info->sva));
3577                                 if (pte_pv) {
3578                                         pv_drop(pte_pv);
3579                                         pv_put(pt_pv);
3580                                         kprintf("pmap_scan: RACE1 "
3581                                                 "%016jx, %016lx\n",
3582                                                 info->sva, oldpte);
3583                                         goto again;
3584                                 }
3585                         }
3586
3587                         /*
3588                          * Didn't race
3589                          */
3590                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3591                                            pmap->pmap_bits[PG_V_IDX])) ==
3592                             pmap->pmap_bits[PG_V_IDX],
3593                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3594                              "generation %d/%d",
3595                             *ptep, oldpte, info->sva,
3596                             generation, pmap->pm_generation));
3597                         info->func(pmap, info, NULL, pt_pv, 0,
3598                             info->sva, ptep, info->arg);
3599                 }
3600                 if (pt_pv)
3601                         pv_put(pt_pv);
3602 fast_skip:
3603                 pmap_inval_bulk_flush(info->bulk);
3604                 lwkt_reltoken(&pmap->pm_token);
3605                 return;
3606         }
3607
3608         /*
3609          * Nominal scan case, RB_SCAN() for PD pages and iterate from
3610          * there.
3611          */
3612         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3613         info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3614
3615         if (info->sva >= VM_MAX_USER_ADDRESS) {
3616                 /*
3617                  * The kernel does not currently maintain any pv_entry's for
3618                  * higher-level page tables.
3619                  */
3620                 bzero(&dummy_pv, sizeof(dummy_pv));
3621                 dummy_pv.pv_pindex = info->sva_pd_pindex;
3622                 spin_lock(&pmap->pm_spin);
3623                 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3624                         pmap_scan_callback(&dummy_pv, info);
3625                         ++dummy_pv.pv_pindex;
3626                 }
3627                 spin_unlock(&pmap->pm_spin);
3628         } else {
3629                 /*
3630                  * User page tables maintain local PML4, PDP, and PD
3631                  * pv_entry's at the very least.  PT pv's might be
3632                  * unmanaged and thus not exist.  PTE pv's might be
3633                  * unmanaged and thus not exist.
3634                  */
3635                 spin_lock(&pmap->pm_spin);
3636                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3637                         pmap_scan_cmp, pmap_scan_callback, info);
3638                 spin_unlock(&pmap->pm_spin);
3639         }
3640         pmap_inval_bulk_flush(info->bulk);
3641         lwkt_reltoken(&pmap->pm_token);
3642 }
3643
3644 /*
3645  * WARNING! pmap->pm_spin held
3646  */
3647 static int
3648 pmap_scan_cmp(pv_entry_t pv, void *data)
3649 {
3650         struct pmap_scan_info *info = data;
3651         if (pv->pv_pindex < info->sva_pd_pindex)
3652                 return(-1);
3653         if (pv->pv_pindex >= info->eva_pd_pindex)
3654                 return(1);
3655         return(0);
3656 }
3657
3658 /*
3659  * WARNING! pmap->pm_spin held
3660  */
3661 static int
3662 pmap_scan_callback(pv_entry_t pv, void *data)
3663 {
3664         struct pmap_scan_info *info = data;
3665         struct pmap *pmap = info->pmap;
3666         pv_entry_t pd_pv;       /* A page directory PV */
3667         pv_entry_t pt_pv;       /* A page table PV */
3668         pv_entry_t pte_pv;      /* A page table entry PV */
3669         pt_entry_t *ptep;
3670         pt_entry_t oldpte;
3671         vm_offset_t sva;
3672         vm_offset_t eva;
3673         vm_offset_t va_next;
3674         vm_pindex_t pd_pindex;
3675         int error;
3676         int generation;
3677
3678         /*
3679          * Stop if requested
3680          */
3681         if (info->stop)
3682                 return -1;
3683
3684         /*
3685          * Pull the PD pindex from the pv before releasing the spinlock.
3686          *
3687          * WARNING: pv is faked for kernel pmap scans.
3688          */
3689         pd_pindex = pv->pv_pindex;
3690         spin_unlock(&pmap->pm_spin);
3691         pv = NULL;      /* invalid after spinlock unlocked */
3692
3693         /*
3694          * Calculate the page range within the PD.  SIMPLE pmaps are
3695          * direct-mapped for the entire 2^64 address space.  Normal pmaps
3696          * reflect the user and kernel address space which requires
3697          * cannonicalization w/regards to converting pd_pindex's back
3698          * into addresses.
3699          */
3700         sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3701         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3702             (sva & PML4_SIGNMASK)) {
3703                 sva |= PML4_SIGNMASK;
3704         }
3705         eva = sva + NBPDP;      /* can overflow */
3706         if (sva < info->sva)
3707                 sva = info->sva;
3708         if (eva < info->sva || eva > info->eva)
3709                 eva = info->eva;
3710
3711         /*
3712          * NOTE: kernel mappings do not track page table pages, only
3713          *       terminal pages.
3714          *
3715          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3716          *       However, for the scan to be efficient we try to
3717          *       cache items top-down.
3718          */
3719         pd_pv = NULL;
3720         pt_pv = NULL;
3721
3722         for (; sva < eva; sva = va_next) {
3723                 if (info->stop)
3724                         break;
3725                 if (sva >= VM_MAX_USER_ADDRESS) {
3726                         if (pt_pv) {
3727                                 pv_put(pt_pv);
3728                                 pt_pv = NULL;
3729                         }
3730                         goto kernel_skip;
3731                 }
3732
3733                 /*
3734                  * PD cache (degenerate case if we skip).  It is possible
3735                  * for the PD to not exist due to races.  This is ok.
3736                  */
3737                 if (pd_pv == NULL) {
3738                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3739                 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3740                         pv_put(pd_pv);
3741                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3742                 }
3743                 if (pd_pv == NULL) {
3744                         va_next = (sva + NBPDP) & ~PDPMASK;
3745                         if (va_next < sva)
3746                                 va_next = eva;
3747                         continue;
3748                 }
3749
3750                 /*
3751                  * PT cache
3752                  */
3753                 if (pt_pv == NULL) {
3754                         vm_page_wire_quick(pd_pv->pv_m);
3755                         pv_unlock(pd_pv);
3756                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3757                         pv_lock(pd_pv);
3758                         vm_page_unwire_quick(pd_pv->pv_m);
3759                 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3760                         vm_page_wire_quick(pd_pv->pv_m);
3761                         pv_unlock(pd_pv);
3762                         pv_put(pt_pv);
3763                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3764                         pv_lock(pd_pv);
3765                         vm_page_unwire_quick(pd_pv->pv_m);
3766                 }
3767
3768                 /*
3769                  * If pt_pv is NULL we either have an shared page table
3770                  * page and must issue a callback specific to that case,
3771                  * or there is no page table page.
3772                  *
3773                  * Either way we can skip the page table page.
3774                  */
3775                 if (pt_pv == NULL) {
3776                         /*
3777                          * Possible unmanaged (shared from another pmap)
3778                          * page table page.
3779                          */
3780                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3781                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3782                                 info->func(pmap, info, NULL, pd_pv, 1,
3783                                            sva, ptep, info->arg);
3784                         }
3785
3786                         /*
3787                          * Done, move to next page table page.
3788                          */
3789                         va_next = (sva + NBPDR) & ~PDRMASK;
3790                         if (va_next < sva)
3791                                 va_next = eva;
3792                         continue;
3793                 }
3794
3795                 /*
3796                  * From this point in the loop testing pt_pv for non-NULL
3797                  * means we are in UVM, else if it is NULL we are in KVM.
3798                  *
3799                  * Limit our scan to either the end of the va represented
3800                  * by the current page table page, or to the end of the
3801                  * range being removed.
3802                  */
3803 kernel_skip:
3804                 va_next = (sva + NBPDR) & ~PDRMASK;
3805                 if (va_next < sva)
3806                         va_next = eva;
3807                 if (va_next > eva)
3808                         va_next = eva;
3809
3810                 /*
3811                  * Scan the page table for pages.  Some pages may not be
3812                  * managed (might not have a pv_entry).
3813                  *
3814                  * There is no page table management for kernel pages so
3815                  * pt_pv will be NULL in that case, but otherwise pt_pv
3816                  * is non-NULL, locked, and referenced.
3817                  */
3818
3819                 /*
3820                  * At this point a non-NULL pt_pv means a UVA, and a NULL
3821                  * pt_pv means a KVA.
3822                  */
3823                 if (pt_pv)
3824                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3825                 else
3826                         ptep = vtopte(sva);
3827
3828                 while (sva < va_next) {
3829                         /*
3830                          * Yield every 64 pages, stop if requested.
3831                          */
3832                         if ((++info->count & 63) == 0)
3833                                 lwkt_user_yield();
3834                         if (info->stop)
3835                                 break;
3836
3837                         /*
3838                          * Check if pt_pv has been lost (probably due to
3839                          * a remove of the underlying pages).
3840                          */
3841                         if (pt_pv && pt_pv->pv_pmap == NULL)
3842                                 break;
3843
3844                         /*
3845                          * Acquire the related pte_pv, if any.  If *ptep == 0
3846                          * the related pte_pv should not exist, but if *ptep
3847                          * is not zero the pte_pv may or may not exist (e.g.
3848                          * will not exist for an unmanaged page).
3849                          *
3850                          * However a multitude of races are possible here.
3851                          *
3852                          * In addition, the (pt_pv, pte_pv) lock order is
3853                          * backwards, so we have to be careful in aquiring
3854                          * a properly locked pte_pv.
3855                          */
3856                         generation = pmap->pm_generation;
3857                         if (pt_pv) {
3858                                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3859                                                     &error);
3860                                 if (error) {
3861                                         if (pd_pv) {
3862                                                 vm_page_wire_quick(pd_pv->pv_m);
3863                                                 pv_unlock(pd_pv);
3864                                         }
3865                                         vm_page_wire_quick(pt_pv->pv_m);
3866                                         pv_unlock(pt_pv);/* must be non-NULL */
3867                                         pv_lock(pte_pv); /* safe to block now */
3868                                         pv_put(pte_pv);
3869                                         pte_pv = NULL;
3870                                         pv_lock(pt_pv);
3871                                         vm_page_unwire_quick(pt_pv->pv_m);
3872
3873                                         /*
3874                                          * pt_pv reloaded, need new ptep
3875                                          */
3876                                         KKASSERT(pt_pv != NULL);
3877                                         ptep = pv_pte_lookup(pt_pv,
3878                                                         pmap_pte_index(sva));
3879                                         if (pd_pv) {
3880                                                 pv_lock(pd_pv);
3881                                                 vm_page_unwire_quick(pd_pv->pv_m);
3882                                         }
3883                                         continue;
3884                                 }
3885                         } else {
3886                                 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3887                         }
3888
3889                         /*
3890                          * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3891                          */
3892                         oldpte = *ptep;
3893                         if (oldpte == 0) {
3894                                 if (pte_pv) {
3895                                         kprintf("Unexpected non-NULL pte_pv "
3896                                                 "%p pt_pv %p "
3897                                                 "*ptep = %016lx/%016lx\n",
3898                                                 pte_pv, pt_pv, *ptep, oldpte);
3899                                         panic("Unexpected non-NULL pte_pv");
3900                                 }
3901                                 sva += PAGE_SIZE;
3902                                 ++ptep;
3903                                 continue;
3904                         }
3905
3906                         /*
3907                          * Ready for the callback.  The locked pte_pv (if any)
3908                          * is consumed by the callback.  pte_pv will exist if
3909                          * the page is managed, and will not exist if it
3910                          * isn't.
3911                          */
3912                         if (pte_pv) {
3913                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3914                                     (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3915                                     ("badC *ptep %016lx/%016lx sva %016lx "
3916                                     "pte_pv %p pm_generation %d/%d",
3917                                     *ptep, oldpte, sva, pte_pv,
3918                                     generation, pmap->pm_generation));
3919                                 /*
3920                                  * We must unlock pd_pv across the callback
3921                                  * to avoid deadlocks on any recursive
3922                                  * disposal.  Re-check that it still exists
3923                                  * after re-locking.
3924                                  */
3925                                 if (pd_pv)
3926                                         pv_unlock(pd_pv);
3927                                 info->func(pmap, info, pte_pv, pt_pv, 0,
3928                                            sva, ptep, info->arg);
3929                                 if (pd_pv) {
3930                                         pv_lock(pd_pv);
3931                                         if (pd_pv->pv_pmap == NULL) {
3932                                                 pv_put(pd_pv);
3933                                                 pd_pv = NULL;
3934                                         }
3935                                 }
3936                         } else {
3937                                 /*
3938                                  * Check for insertion race.  Since there is no
3939                                  * pte_pv to guard us it is possible for us
3940                                  * to race another thread doing an insertion.
3941                                  * Our lookup misses the pte_pv but our *ptep
3942                                  * check sees the inserted pte.
3943                                  *
3944                                  * XXX panic case seems to occur within a
3945                                  * vm_fork() of /bin/sh, which frankly
3946                                  * shouldn't happen since no other threads
3947                                  * should be inserting to our pmap in that
3948                                  * situation.  Removing, possibly.  Inserting,
3949                                  * shouldn't happen.
3950                                  */
3951                                 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3952                                     pt_pv) {
3953                                         pte_pv = pv_find(pmap,
3954                                                          pmap_pte_pindex(sva));
3955                                         if (pte_pv) {
3956                                                 pv_drop(pte_pv);
3957                                                 kprintf("pmap_scan: RACE2 "
3958                                                         "%016jx, %016lx\n",
3959                                                         sva, oldpte);
3960                                                 continue;
3961                                         }
3962                                 }
3963
3964                                 /*
3965                                  * Didn't race
3966                                  *
3967                                  * We must unlock pd_pv across the callback
3968                                  * to avoid deadlocks on any recursive
3969                                  * disposal.  Re-check that it still exists
3970                                  * after re-locking.
3971                                  */
3972                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3973                                     pmap->pmap_bits[PG_V_IDX],
3974                                     ("badD *ptep %016lx/%016lx sva %016lx "
3975                                     "pte_pv NULL pm_generation %d/%d",
3976                                      *ptep, oldpte, sva,
3977                                      generation, pmap->pm_generation));
3978                                 if (pd_pv)
3979                                         pv_unlock(pd_pv);
3980                                 info->func(pmap, info, NULL, pt_pv, 0,
3981                                            sva, ptep, info->arg);
3982                                 if (pd_pv) {
3983                                         pv_lock(pd_pv);
3984                                         if (pd_pv->pv_pmap == NULL) {
3985                                                 pv_put(pd_pv);
3986                                                 pd_pv = NULL;
3987                                         }
3988                                 }
3989                         }
3990                         pte_pv = NULL;
3991                         sva += PAGE_SIZE;
3992                         ++ptep;
3993                 }
3994         }
3995         if (pd_pv) {
3996                 pv_put(pd_pv);
3997                 pd_pv = NULL;
3998         }
3999         if (pt_pv) {
4000                 pv_put(pt_pv);
4001                 pt_pv = NULL;
4002         }
4003         if ((++info->count & 7) == 0)
4004                 lwkt_user_yield();
4005
4006         /*
4007          * Relock before returning.
4008          */
4009         spin_lock(&pmap->pm_spin);
4010         return (0);
4011 }
4012
4013 void
4014 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4015 {
4016         struct pmap_scan_info info;
4017
4018         info.pmap = pmap;
4019         info.sva = sva;
4020         info.eva = eva;
4021         info.func = pmap_remove_callback;
4022         info.arg = NULL;
4023         pmap_scan(&info, 1);
4024 }
4025
4026 static void
4027 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4028 {
4029         struct pmap_scan_info info;
4030
4031         info.pmap = pmap;
4032         info.sva = sva;
4033         info.eva = eva;
4034         info.func = pmap_remove_callback;
4035         info.arg = NULL;
4036         pmap_scan(&info, 0);
4037 }
4038
4039 static void
4040 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4041                      pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4042                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4043 {
4044         pt_entry_t pte;
4045
4046         if (pte_pv) {
4047                 /*
4048                  * This will also drop pt_pv's wire_count. Note that
4049                  * terminal pages are not wired based on mmu presence.
4050                  *
4051                  * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4052                  */
4053                 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4054                 pte_pv = NULL;  /* safety */
4055
4056                 /*
4057                  * Recursively destroy higher-level page tables.
4058                  *
4059                  * This is optional.  If we do not, they will still
4060                  * be destroyed when the process exits.
4061                  */
4062                 if (pmap_dynamic_delete &&
4063                     pt_pv &&
4064                     pt_pv->pv_m &&
4065                     pt_pv->pv_m->wire_count == 1 &&
4066                     pt_pv->pv_pindex != pmap_pml4_pindex()) {
4067                         pv_hold(pt_pv);
4068                         pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4069                         pv_lock(pt_pv);
4070                 }
4071         } else if (sharept == 0) {
4072                 /*
4073                  * Unmanaged page table (pt, pd, or pdp. Not pte).
4074                  *
4075                  * pt_pv's wire_count is still bumped by unmanaged pages
4076                  * so we must decrement it manually.
4077                  *
4078                  * We have to unwire the target page table page.
4079                  *
4080                  * It is unclear how we can invalidate a segment so we
4081                  * invalidate -1 which invlidates the tlb.
4082                  */
4083                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4084                 if (pte & pmap->pmap_bits[PG_W_IDX])
4085                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4086                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4087                 if (vm_page_unwire_quick(pt_pv->pv_m))
4088                         panic("pmap_remove: insufficient wirecount");
4089         } else {
4090                 /*
4091                  * Unmanaged page table (pt, pd, or pdp. Not pte) for
4092                  * a shared page table.
4093                  *
4094                  * pt_pv is actually the pd_pv for our pmap (not the shared
4095                  * object pmap).
4096                  *
4097                  * We have to unwire the target page table page and we
4098                  * have to unwire our page directory page.
4099                  *
4100                  * It is unclear how we can invalidate a segment so we
4101                  * invalidate -1 which invlidates the tlb.
4102                  */
4103                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4104                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4105                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4106                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4107                         panic("pmap_remove: shared pgtable1 bad wirecount");
4108                 if (vm_page_unwire_quick(pt_pv->pv_m))
4109                         panic("pmap_remove: shared pgtable2 bad wirecount");
4110         }
4111 }
4112
4113 /*
4114  * Removes this physical page from all physical maps in which it resides.
4115  * Reflects back modify bits to the pager.
4116  *
4117  * This routine may not be called from an interrupt.
4118  */
4119 static
4120 void
4121 pmap_remove_all(vm_page_t m)
4122 {
4123         pv_entry_t pv;
4124         pmap_inval_bulk_t bulk;
4125
4126         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4127                 return;
4128
4129         vm_page_spin_lock(m);
4130         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4131                 KKASSERT(pv->pv_m == m);
4132                 if (pv_hold_try(pv)) {
4133                         vm_page_spin_unlock(m);
4134                 } else {
4135                         vm_page_spin_unlock(m);
4136                         pv_lock(pv);
4137                 }
4138                 if (pv->pv_m != m) {
4139                         pv_put(pv);
4140                         vm_page_spin_lock(m);
4141                         continue;
4142                 }
4143
4144                 /*
4145                  * Holding no spinlocks, pv is locked.
4146                  */
4147                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4148                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4149                 pv = NULL;      /* safety */
4150                 pmap_inval_bulk_flush(&bulk);
4151 #if 0
4152                 pmap_remove_pv_page(pv);
4153                 pv_free(pv, 1);
4154 #endif
4155                 vm_page_spin_lock(m);
4156         }
4157         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4158         vm_page_spin_unlock(m);
4159 }
4160
4161 /*
4162  * Removes the page from a particular pmap
4163  */
4164 void
4165 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4166 {
4167         pv_entry_t pv;
4168         pmap_inval_bulk_t bulk;
4169
4170         if (!pmap_initialized)
4171                 return;
4172
4173 again:
4174         vm_page_spin_lock(m);
4175         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4176                 if (pv->pv_pmap != pmap)
4177                         continue;
4178                 KKASSERT(pv->pv_m == m);
4179                 if (pv_hold_try(pv)) {
4180                         vm_page_spin_unlock(m);
4181                 } else {
4182                         vm_page_spin_unlock(m);
4183                         pv_lock(pv);
4184                 }
4185                 if (pv->pv_m != m) {
4186                         pv_put(pv);
4187                         goto again;
4188                 }
4189
4190                 /*
4191                  * Holding no spinlocks, pv is locked.
4192                  */
4193                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4194                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4195                 pv = NULL;      /* safety */
4196                 pmap_inval_bulk_flush(&bulk);
4197 #if 0
4198                 pmap_remove_pv_page(pv);
4199                 pv_free(pv, 1);
4200 #endif
4201                 goto again;
4202         }
4203         vm_page_spin_unlock(m);
4204 }
4205
4206 /*
4207  * Set the physical protection on the specified range of this map
4208  * as requested.  This function is typically only used for debug watchpoints
4209  * and COW pages.
4210  *
4211  * This function may not be called from an interrupt if the map is
4212  * not the kernel_pmap.
4213  *
4214  * NOTE!  For shared page table pages we just unmap the page.
4215  */
4216 void
4217 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4218 {
4219         struct pmap_scan_info info;
4220         /* JG review for NX */
4221
4222         if (pmap == NULL)
4223                 return;
4224         if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
4225                 pmap_remove(pmap, sva, eva);
4226                 return;
4227         }
4228         if (prot & VM_PROT_WRITE)
4229                 return;
4230         info.pmap = pmap;
4231         info.sva = sva;
4232         info.eva = eva;
4233         info.func = pmap_protect_callback;
4234         info.arg = &prot;
4235         pmap_scan(&info, 1);
4236 }
4237
4238 static
4239 void
4240 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4241                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4242                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4243 {
4244         pt_entry_t pbits;
4245         pt_entry_t cbits;
4246         pt_entry_t pte;
4247         vm_page_t m;
4248
4249 again:
4250         pbits = *ptep;
4251         cbits = pbits;
4252         if (pte_pv) {
4253                 m = NULL;
4254                 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4255                         if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4256                                 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4257                                 KKASSERT(m == pte_pv->pv_m);
4258                                 vm_page_flag_set(m, PG_REFERENCED);
4259                         }
4260                         cbits &= ~pmap->pmap_bits[PG_A_IDX];
4261                 }
4262                 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4263                         if (pmap_track_modified(pte_pv->pv_pindex)) {
4264                                 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4265                                         if (m == NULL) {
4266                                                 m = PHYS_TO_VM_PAGE(pbits &
4267                                                                     PG_FRAME);
4268                                         }
4269                                         vm_page_dirty(m);
4270                                 }
4271                                 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4272                         }
4273                 }
4274         } else if (sharept) {
4275                 /*
4276                  * Unmanaged page table, pt_pv is actually the pd_pv
4277                  * for our pmap (not the object's shared pmap).
4278                  *
4279                  * When asked to protect something in a shared page table
4280                  * page we just unmap the page table page.  We have to
4281                  * invalidate the tlb in this situation.
4282                  *
4283                  * XXX Warning, shared page tables will not be used for
4284                  * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4285                  * so PHYS_TO_VM_PAGE() should be safe here.
4286                  */
4287                 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4288                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4289                         panic("pmap_protect: pgtable1 pg bad wirecount");
4290                 if (vm_page_unwire_quick(pt_pv->pv_m))
4291                         panic("pmap_protect: pgtable2 pg bad wirecount");
4292                 ptep = NULL;
4293         }
4294         /* else unmanaged page, adjust bits, no wire changes */
4295
4296         if (ptep) {
4297                 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4298 #ifdef PMAP_DEBUG2
4299                 if (pmap_enter_debug > 0) {
4300                         --pmap_enter_debug;
4301                         kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4302                                 "pt_pv=%p cbits=%08lx\n",
4303                                 va, ptep, pte_pv,
4304                                 pt_pv, cbits
4305                         );
4306                 }
4307 #endif
4308                 if (pbits != cbits) {
4309                         if (!pmap_inval_smp_cmpset(pmap, (vm_offset_t)-1,
4310                                                    ptep, pbits, cbits)) {
4311                                 goto again;
4312                         }
4313                 }
4314         }
4315         if (pte_pv)
4316                 pv_put(pte_pv);
4317 }
4318
4319 /*
4320  * Insert the vm_page (m) at the virtual address (va), replacing any prior
4321  * mapping at that address.  Set protection and wiring as requested.
4322  *
4323  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4324  * possible.  If it is we enter the page into the appropriate shared pmap
4325  * hanging off the related VM object instead of the passed pmap, then we
4326  * share the page table page from the VM object's pmap into the current pmap.
4327  *
4328  * NOTE: This routine MUST insert the page into the pmap now, it cannot
4329  *       lazy-evaluate.
4330  */
4331 void
4332 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4333            boolean_t wired, vm_map_entry_t entry)
4334 {
4335         pv_entry_t pt_pv;       /* page table */
4336         pv_entry_t pte_pv;      /* page table entry */
4337         pt_entry_t *ptep;
4338         vm_paddr_t opa;
4339         pt_entry_t origpte, newpte;
4340         vm_paddr_t pa;
4341
4342         if (pmap == NULL)
4343                 return;
4344         va = trunc_page(va);
4345 #ifdef PMAP_DIAGNOSTIC
4346         if (va >= KvaEnd)
4347                 panic("pmap_enter: toobig");
4348         if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4349                 panic("pmap_enter: invalid to pmap_enter page table "
4350                       "pages (va: 0x%lx)", va);
4351 #endif
4352         if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4353                 kprintf("Warning: pmap_enter called on UVA with "
4354                         "kernel_pmap\n");
4355 #ifdef DDB
4356                 db_print_backtrace();
4357 #endif
4358         }
4359         if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4360                 kprintf("Warning: pmap_enter called on KVA without"
4361                         "kernel_pmap\n");
4362 #ifdef DDB
4363                 db_print_backtrace();
4364 #endif
4365         }
4366
4367         /*
4368          * Get locked PV entries for our new page table entry (pte_pv)
4369          * and for its parent page table (pt_pv).  We need the parent
4370          * so we can resolve the location of the ptep.
4371          *
4372          * Only hardware MMU actions can modify the ptep out from
4373          * under us.
4374          *
4375          * if (m) is fictitious or unmanaged we do not create a managing
4376          * pte_pv for it.  Any pre-existing page's management state must
4377          * match (avoiding code complexity).
4378          *
4379          * If the pmap is still being initialized we assume existing
4380          * page tables.
4381          *
4382          * Kernel mapppings do not track page table pages (i.e. pt_pv).
4383          */
4384         if (pmap_initialized == FALSE) {
4385                 pte_pv = NULL;
4386                 pt_pv = NULL;
4387                 ptep = vtopte(va);
4388                 origpte = *ptep;
4389         } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4390                 pte_pv = NULL;
4391                 if (va >= VM_MAX_USER_ADDRESS) {
4392                         pt_pv = NULL;
4393                         ptep = vtopte(va);
4394                 } else {
4395                         pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4396                                                   NULL, entry, va);
4397                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4398                 }
4399                 origpte = *ptep;
4400                 cpu_ccfence();
4401                 KASSERT(origpte == 0 ||
4402                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4403                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4404         } else {
4405                 if (va >= VM_MAX_USER_ADDRESS) {
4406                         /*
4407                          * Kernel map, pv_entry-tracked.
4408                          */
4409                         pt_pv = NULL;
4410                         pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4411                         ptep = vtopte(va);
4412                 } else {
4413                         /*
4414                          * User map
4415                          */
4416                         pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4417                                                    &pt_pv, entry, va);
4418                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4419                 }
4420                 origpte = *ptep;
4421                 cpu_ccfence();
4422                 KASSERT(origpte == 0 ||
4423                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4424                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4425         }
4426
4427         pa = VM_PAGE_TO_PHYS(m);
4428         opa = origpte & PG_FRAME;
4429
4430         newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4431                  pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4432         if (wired)
4433                 newpte |= pmap->pmap_bits[PG_W_IDX];
4434         if (va < VM_MAX_USER_ADDRESS)
4435                 newpte |= pmap->pmap_bits[PG_U_IDX];
4436         if (pte_pv)
4437                 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4438 //      if (pmap == &kernel_pmap)
4439 //              newpte |= pgeflag;
4440         newpte |= pmap->pmap_cache_bits[m->pat_mode];
4441         if (m->flags & PG_FICTITIOUS)
4442                 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4443
4444         /*
4445          * It is possible for multiple faults to occur in threaded
4446          * environments, the existing pte might be correct.
4447          */
4448         if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4449             pmap->pmap_bits[PG_A_IDX])) == 0)
4450                 goto done;
4451
4452         /*
4453          * Ok, either the address changed or the protection or wiring
4454          * changed.
4455          *
4456          * Clear the current entry, interlocking the removal.  For managed
4457          * pte's this will also flush the modified state to the vm_page.
4458          * Atomic ops are mandatory in order to ensure that PG_M events are
4459          * not lost during any transition.
4460          *
4461          * WARNING: The caller has busied the new page but not the original
4462          *          vm_page which we are trying to replace.  Because we hold
4463          *          the pte_pv lock, but have not busied the page, PG bits
4464          *          can be cleared out from under us.
4465          */
4466         if (opa) {
4467                 if (pte_pv) {
4468                         /*
4469                          * pt_pv won't exist for a kernel page (managed or
4470                          * otherwise).
4471                          */
4472                         if (prot & VM_PROT_NOSYNC) {
4473                                 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4474                         } else {
4475                                 pmap_inval_bulk_t bulk;
4476
4477                                 pmap_inval_bulk_init(&bulk, pmap);
4478                                 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4479                                 pmap_inval_bulk_flush(&bulk);
4480                         }
4481                         if (pte_pv->pv_m)
4482                                 pmap_remove_pv_page(pte_pv);
4483                 } else if (prot & VM_PROT_NOSYNC) {
4484                         /*
4485                          * Unmanaged page, NOSYNC (no mmu sync) requested.
4486                          *
4487                          * Leave wire count on PT page intact.
4488                          */
4489                         (void)pte_load_clear(ptep);
4490                         cpu_invlpg((void *)va);
4491                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
4492                 } else {
4493                         /*
4494                          * Unmanaged page, normal enter.
4495                          *
4496                          * Leave wire count on PT page intact.
4497                          */
4498                         pmap_inval_smp(pmap, va, 1, ptep, 0);
4499                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
4500                 }
4501                 KKASSERT(*ptep == 0);
4502         }
4503
4504 #ifdef PMAP_DEBUG2
4505         if (pmap_enter_debug > 0) {
4506                 --pmap_enter_debug;
4507                 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4508                         " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4509                         va, m,
4510                         origpte, newpte, ptep,
4511                         pte_pv, pt_pv, opa, prot);
4512         }
4513 #endif
4514
4515         if (pte_pv) {
4516                 /*
4517                  * Enter on the PV list if part of our managed memory.
4518                  * Wiring of the PT page is already handled.
4519                  */
4520                 KKASSERT(pte_pv->pv_m == NULL);
4521                 vm_page_spin_lock(m);
4522                 pte_pv->pv_m = m;
4523                 pmap_page_stats_adding(m);
4524                 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4525                 vm_page_flag_set(m, PG_MAPPED);
4526                 vm_page_spin_unlock(m);
4527         } else if (pt_pv && opa == 0) {
4528                 /*
4529                  * We have to adjust the wire count on the PT page ourselves
4530                  * for unmanaged entries.  If opa was non-zero we retained
4531                  * the existing wire count from the removal.
4532                  */
4533                 vm_page_wire_quick(pt_pv->pv_m);
4534         }
4535
4536         /*
4537          * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4538          *
4539          * User VMAs do not because those will be zero->non-zero, so no
4540          * stale entries to worry about at this point.
4541          *
4542          * For KVM there appear to still be issues.  Theoretically we
4543          * should be able to scrap the interlocks entirely but we
4544          * get crashes.
4545          */
4546         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4547                 pmap_inval_smp(pmap, va, 1, ptep, newpte);
4548         } else {
4549                 *(volatile pt_entry_t *)ptep = newpte;
4550                 if (pt_pv == NULL)
4551                         cpu_invlpg((void *)va);
4552         }
4553
4554         if (wired) {
4555                 if (pte_pv) {
4556                         atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4557                                         1);
4558                 } else {
4559                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
4560                 }
4561         }
4562         if (newpte & pmap->pmap_bits[PG_RW_IDX])
4563                 vm_page_flag_set(m, PG_WRITEABLE);
4564
4565         /*
4566          * Unmanaged pages need manual resident_count tracking.
4567          */
4568         if (pte_pv == NULL && pt_pv)
4569                 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4570
4571         /*
4572          * Cleanup
4573          */
4574 done:
4575         KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4576                  (m->flags & PG_MAPPED));
4577
4578         /*
4579          * Cleanup the pv entry, allowing other accessors.
4580          */
4581         if (pte_pv)
4582                 pv_put(pte_pv);
4583         if (pt_pv)
4584                 pv_put(pt_pv);
4585 }
4586
4587 /*
4588  * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4589  * This code also assumes that the pmap has no pre-existing entry for this
4590  * VA.
4591  *
4592  * This code currently may only be used on user pmaps, not kernel_pmap.
4593  */
4594 void
4595 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4596 {
4597         pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4598 }
4599
4600 /*
4601  * Make a temporary mapping for a physical address.  This is only intended
4602  * to be used for panic dumps.
4603  *
4604  * The caller is responsible for calling smp_invltlb().
4605  */
4606 void *
4607 pmap_kenter_temporary(vm_paddr_t pa, long i)
4608 {
4609         pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4610         return ((void *)crashdumpmap);
4611 }
4612
4613 #define MAX_INIT_PT (96)
4614
4615 /*
4616  * This routine preloads the ptes for a given object into the specified pmap.
4617  * This eliminates the blast of soft faults on process startup and
4618  * immediately after an mmap.
4619  */
4620 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4621
4622 void
4623 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4624                     vm_object_t object, vm_pindex_t pindex,
4625                     vm_size_t size, int limit)
4626 {
4627         struct rb_vm_page_scan_info info;
4628         struct lwp *lp;
4629         vm_size_t psize;
4630
4631         /*
4632          * We can't preinit if read access isn't set or there is no pmap
4633          * or object.
4634          */
4635         if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4636                 return;
4637
4638         /*
4639          * We can't preinit if the pmap is not the current pmap
4640          */
4641         lp = curthread->td_lwp;
4642         if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4643                 return;
4644
4645         /*
4646          * Misc additional checks
4647          */
4648         psize = x86_64_btop(size);
4649
4650         if ((object->type != OBJT_VNODE) ||
4651                 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4652                         (object->resident_page_count > MAX_INIT_PT))) {
4653                 return;
4654         }
4655
4656         if (pindex + psize > object->size) {
4657                 if (object->size < pindex)
4658                         return;
4659                 psize = object->size - pindex;
4660         }
4661
4662         if (psize == 0)
4663                 return;
4664
4665         /*
4666          * If everything is segment-aligned do not pre-init here.  Instead
4667          * allow the normal vm_fault path to pass a segment hint to
4668          * pmap_enter() which will then use an object-referenced shared
4669          * page table page.
4670          */
4671         if ((addr & SEG_MASK) == 0 &&
4672             (ctob(psize) & SEG_MASK) == 0 &&
4673             (ctob(pindex) & SEG_MASK) == 0) {
4674                 return;
4675         }
4676
4677         /*
4678          * Use a red-black scan to traverse the requested range and load
4679          * any valid pages found into the pmap.
4680          *
4681          * We cannot safely scan the object's memq without holding the
4682          * object token.
4683          */
4684         info.start_pindex = pindex;
4685         info.end_pindex = pindex + psize - 1;
4686         info.limit = limit;
4687         info.mpte = NULL;
4688         info.addr = addr;
4689         info.pmap = pmap;
4690
4691         vm_object_hold_shared(object);
4692         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4693                                 pmap_object_init_pt_callback, &info);
4694         vm_object_drop(object);
4695 }
4696
4697 static
4698 int
4699 pmap_object_init_pt_callback(vm_page_t p, void *data)
4700 {
4701         struct rb_vm_page_scan_info *info = data;
4702         vm_pindex_t rel_index;
4703
4704         /*
4705          * don't allow an madvise to blow away our really
4706          * free pages allocating pv entries.
4707          */
4708         if ((info->limit & MAP_PREFAULT_MADVISE) &&
4709                 vmstats.v_free_count < vmstats.v_free_reserved) {
4710                     return(-1);
4711         }
4712
4713         /*
4714          * Ignore list markers and ignore pages we cannot instantly
4715          * busy (while holding the object token).
4716          */
4717         if (p->flags & PG_MARKER)
4718                 return 0;
4719         if (vm_page_busy_try(p, TRUE))
4720                 return 0;
4721         if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4722             (p->flags & PG_FICTITIOUS) == 0) {
4723                 if ((p->queue - p->pc) == PQ_CACHE)
4724                         vm_page_deactivate(p);
4725                 rel_index = p->pindex - info->start_pindex;
4726                 pmap_enter_quick(info->pmap,
4727                                  info->addr + x86_64_ptob(rel_index), p);
4728         }
4729         vm_page_wakeup(p);
4730         lwkt_yield();
4731         return(0);
4732 }
4733
4734 /*
4735  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4736  * address.
4737  *
4738  * Returns FALSE if it would be non-trivial or if a pte is already loaded
4739  * into the slot.
4740  *
4741  * XXX This is safe only because page table pages are not freed.
4742  */
4743 int
4744 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4745 {
4746         pt_entry_t *pte;
4747
4748         /*spin_lock(&pmap->pm_spin);*/
4749         if ((pte = pmap_pte(pmap, addr)) != NULL) {
4750                 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4751                         /*spin_unlock(&pmap->pm_spin);*/
4752                         return FALSE;
4753                 }
4754         }
4755         /*spin_unlock(&pmap->pm_spin);*/
4756         return TRUE;
4757 }
4758
4759 /*
4760  * Change the wiring attribute for a pmap/va pair.  The mapping must already
4761  * exist in the pmap.  The mapping may or may not be managed.
4762  *
4763  * Wiring is not a hardware characteristic so there is no need to invalidate
4764  * TLB.  However, in an SMP environment we must use a locked bus cycle to
4765  * update the pte (if we are not using the pmap_inval_*() API that is)...
4766  * it's ok to do this for simple wiring changes.
4767  */
4768 void
4769 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
4770                    vm_map_entry_t entry)
4771 {
4772         pt_entry_t *ptep;
4773         pv_entry_t pv;
4774
4775         if (pmap == NULL)
4776                 return;
4777
4778         lwkt_gettoken(&pmap->pm_token);
4779         if (pmap == &kernel_pmap) {
4780                 /*
4781                  * The kernel may have managed pages, but not managed
4782                  * page tables.
4783                  */
4784                 ptep = pmap_pte_quick(pmap, va);
4785
4786                 if (wired && !pmap_pte_w(pmap, ptep))
4787                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
4788                 else if (!wired && pmap_pte_w(pmap, ptep))
4789                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4790
4791                 if (wired)
4792                         atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4793                 else
4794                         atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4795         } else {
4796                 /*
4797                  * Userland, the pmap of the possibly shared segment might
4798                  * not be (pmap).
4799                  */
4800                 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL,
4801                                        entry, va);
4802                 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
4803
4804                 if (wired && !pmap_pte_w(pmap, ptep))
4805                         atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4806                 else if (!wired && pmap_pte_w(pmap, ptep))
4807                         atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4808
4809                 if (wired)
4810                         atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4811                 else
4812                         atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4813                 pv_put(pv);
4814         }
4815         lwkt_reltoken(&pmap->pm_token);
4816 }
4817
4818
4819
4820 /*
4821  * Copy the range specified by src_addr/len from the source map to
4822  * the range dst_addr/len in the destination map.
4823  *
4824  * This routine is only advisory and need not do anything.
4825  */
4826 void
4827 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4828           vm_size_t len, vm_offset_t src_addr)
4829 {
4830 }
4831
4832 /*
4833  * pmap_zero_page:
4834  *
4835  *      Zero the specified physical page.
4836  *
4837  *      This function may be called from an interrupt and no locking is
4838  *      required.
4839  */
4840 void
4841 pmap_zero_page(vm_paddr_t phys)
4842 {
4843         vm_offset_t va = PHYS_TO_DMAP(phys);
4844
4845         pagezero((void *)va);
4846 }
4847
4848 /*
4849  * pmap_zero_page:
4850  *
4851  *      Zero part of a physical page by mapping it into memory and clearing
4852  *      its contents with bzero.
4853  *
4854  *      off and size may not cover an area beyond a single hardware page.
4855  */
4856 void
4857 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4858 {
4859         vm_offset_t virt = PHYS_TO_DMAP(phys);
4860
4861         bzero((char *)virt + off, size);
4862 }
4863
4864 /*
4865  * pmap_copy_page:
4866  *
4867  *      Copy the physical page from the source PA to the target PA.
4868  *      This function may be called from an interrupt.  No locking
4869  *      is required.
4870  */
4871 void
4872 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4873 {
4874         vm_offset_t src_virt, dst_virt;
4875
4876         src_virt = PHYS_TO_DMAP(src);
4877         dst_virt = PHYS_TO_DMAP(dst);
4878         bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4879 }
4880
4881 /*
4882  * pmap_copy_page_frag:
4883  *
4884  *      Copy the physical page from the source PA to the target PA.
4885  *      This function may be called from an interrupt.  No locking
4886  *      is required.
4887  */
4888 void
4889 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4890 {
4891         vm_offset_t src_virt, dst_virt;
4892
4893         src_virt = PHYS_TO_DMAP(src);
4894         dst_virt = PHYS_TO_DMAP(dst);
4895
4896         bcopy((char *)src_virt + (src & PAGE_MASK),
4897               (char *)dst_virt + (dst & PAGE_MASK),
4898               bytes);
4899 }
4900
4901 /*
4902  * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4903  * this page.  This count may be changed upwards or downwards in the future;
4904  * it is only necessary that true be returned for a small subset of pmaps
4905  * for proper page aging.
4906  */
4907 boolean_t
4908 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4909 {
4910         pv_entry_t pv;
4911         int loops = 0;
4912
4913         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4914                 return FALSE;
4915
4916         vm_page_spin_lock(m);
4917         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4918                 if (pv->pv_pmap == pmap) {
4919                         vm_page_spin_unlock(m);
4920                         return TRUE;
4921                 }
4922                 loops++;
4923                 if (loops >= 16)
4924                         break;
4925         }
4926         vm_page_spin_unlock(m);
4927         return (FALSE);
4928 }
4929
4930 /*
4931  * Remove all pages from specified address space this aids process exit
4932  * speeds.  Also, this code may be special cased for the current process
4933  * only.
4934  */
4935 void
4936 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4937 {
4938         pmap_remove_noinval(pmap, sva, eva);
4939         cpu_invltlb();
4940 }
4941
4942 /*
4943  * pmap_testbit tests bits in pte's note that the testbit/clearbit
4944  * routines are inline, and a lot of things compile-time evaluate.
4945  */
4946 static
4947 boolean_t
4948 pmap_testbit(vm_page_t m, int bit)
4949 {
4950         pv_entry_t pv;
4951         pt_entry_t *pte;
4952         pmap_t pmap;
4953
4954         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4955                 return FALSE;
4956
4957         if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4958                 return FALSE;
4959         vm_page_spin_lock(m);
4960         if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4961                 vm_page_spin_unlock(m);
4962                 return FALSE;
4963         }
4964
4965         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4966
4967 #if defined(PMAP_DIAGNOSTIC)
4968                 if (pv->pv_pmap == NULL) {
4969                         kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4970                             pv->pv_pindex);
4971                         continue;
4972                 }
4973 #endif
4974                 pmap = pv->pv_pmap;
4975
4976                 /*
4977                  * If the bit being tested is the modified bit, then
4978                  * mark clean_map and ptes as never
4979                  * modified.
4980                  *
4981                  * WARNING!  Because we do not lock the pv, *pte can be in a
4982                  *           state of flux.  Despite this the value of *pte
4983                  *           will still be related to the vm_page in some way
4984                  *           because the pv cannot be destroyed as long as we
4985                  *           hold the vm_page spin lock.
4986                  */
4987                 if (bit == PG_A_IDX || bit == PG_M_IDX) {
4988                                 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
4989                         if (!pmap_track_modified(pv->pv_pindex))
4990                                 continue;
4991                 }
4992
4993                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4994                 if (*pte & pmap->pmap_bits[bit]) {
4995                         vm_page_spin_unlock(m);
4996                         return TRUE;
4997                 }
4998         }
4999         vm_page_spin_unlock(m);
5000         return (FALSE);
5001 }
5002
5003 /*
5004  * This routine is used to modify bits in ptes.  Only one bit should be
5005  * specified.  PG_RW requires special handling.
5006  *
5007  * Caller must NOT hold any spin locks
5008  */
5009 static __inline
5010 void
5011 pmap_clearbit(vm_page_t m, int bit_index)
5012 {
5013         pv_entry_t pv;
5014         pt_entry_t *pte;
5015         pt_entry_t pbits;
5016         pmap_t pmap;
5017
5018         if (bit_index == PG_RW_IDX)
5019                 vm_page_flag_clear(m, PG_WRITEABLE);
5020         if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5021                 return;
5022         }
5023
5024         /*
5025          * PG_M or PG_A case
5026          *
5027          * Loop over all current mappings setting/clearing as appropos If
5028          * setting RO do we need to clear the VAC?
5029          *
5030          * NOTE: When clearing PG_M we could also (not implemented) drop
5031          *       through to the PG_RW code and clear PG_RW too, forcing
5032          *       a fault on write to redetect PG_M for virtual kernels, but
5033          *       it isn't necessary since virtual kernels invalidate the
5034          *       pte when they clear the VPTE_M bit in their virtual page
5035          *       tables.
5036          *
5037          * NOTE: Does not re-dirty the page when clearing only PG_M.
5038          *
5039          * NOTE: Because we do not lock the pv, *pte can be in a state of
5040          *       flux.  Despite this the value of *pte is still somewhat
5041          *       related while we hold the vm_page spin lock.
5042          *
5043          *       *pte can be zero due to this race.  Since we are clearing
5044          *       bits we basically do no harm when this race  ccurs.
5045          */
5046         if (bit_index != PG_RW_IDX) {
5047                 vm_page_spin_lock(m);
5048                 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5049 #if defined(PMAP_DIAGNOSTIC)
5050                         if (pv->pv_pmap == NULL) {
5051                                 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5052                                     pv->pv_pindex);
5053                                 continue;
5054                         }
5055 #endif
5056                         pmap = pv->pv_pmap;
5057                         pte = pmap_pte_quick(pv->pv_pmap,
5058                                              pv->pv_pindex << PAGE_SHIFT);
5059                         pbits = *pte;
5060                         if (pbits & pmap->pmap_bits[bit_index])
5061                                 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5062                 }
5063                 vm_page_spin_unlock(m);
5064                 return;
5065         }
5066
5067         /*
5068          * Clear PG_RW.  Also clears PG_M and marks the page dirty if PG_M
5069          * was set.
5070          */
5071 restart:
5072         vm_page_spin_lock(m);
5073         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5074                 /*
5075                  * don't write protect pager mappings
5076                  */
5077                 if (!pmap_track_modified(pv->pv_pindex))
5078                         continue;
5079
5080 #if defined(PMAP_DIAGNOSTIC)
5081                 if (pv->pv_pmap == NULL) {
5082                         kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5083                             pv->pv_pindex);
5084                         continue;
5085                 }
5086 #endif
5087                 pmap = pv->pv_pmap;
5088                 /*
5089                  * Skip pages which do not have PG_RW set.
5090                  */
5091                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5092                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5093                         continue;
5094
5095                 /*
5096                  * Lock the PV
5097                  */
5098                 if (pv_hold_try(pv)) {
5099                         vm_page_spin_unlock(m);
5100                 } else {
5101                         vm_page_spin_unlock(m);
5102                         pv_lock(pv);    /* held, now do a blocking lock */
5103                 }
5104                 if (pv->pv_pmap != pmap || pv->pv_m != m) {
5105                         pv_put(pv);     /* and release */
5106                         goto restart;   /* anything could have happened */
5107                 }
5108                 KKASSERT(pv->pv_pmap == pmap);
5109                 for (;;) {
5110                         pt_entry_t nbits;
5111
5112                         pbits = *pte;
5113                         cpu_ccfence();
5114                         nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5115                                           pmap->pmap_bits[PG_M_IDX]);
5116                         if (pmap_inval_smp_cmpset(pmap,
5117                                      ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5118                                      pte, pbits, nbits)) {
5119                                 break;
5120                         }
5121                         cpu_pause();
5122                 }
5123                 vm_page_spin_lock(m);
5124
5125                 /*
5126                  * If PG_M was found to be set while we were clearing PG_RW
5127                  * we also clear PG_M (done above) and mark the page dirty.
5128                  * Callers expect this behavior.
5129                  */
5130                 if (pbits & pmap->pmap_bits[PG_M_IDX])
5131                         vm_page_dirty(m);
5132                 pv_put(pv);
5133         }
5134         vm_page_spin_unlock(m);
5135 }
5136
5137 /*
5138  * Lower the permission for all mappings to a given page.
5139  *
5140  * Page must be busied by caller.  Because page is busied by caller this
5141  * should not be able to race a pmap_enter().
5142  */
5143 void
5144 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5145 {
5146         /* JG NX support? */
5147         if ((prot & VM_PROT_WRITE) == 0) {
5148                 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5149                         /*
5150                          * NOTE: pmap_clearbit(.. PG_RW) also clears
5151                          *       the PG_WRITEABLE flag in (m).
5152                          */
5153                         pmap_clearbit(m, PG_RW_IDX);
5154                 } else {
5155                         pmap_remove_all(m);
5156                 }
5157         }
5158 }
5159
5160 vm_paddr_t
5161 pmap_phys_address(vm_pindex_t ppn)
5162 {
5163         return (x86_64_ptob(ppn));
5164 }
5165
5166 /*
5167  * Return a count of reference bits for a page, clearing those bits.
5168  * It is not necessary for every reference bit to be cleared, but it
5169  * is necessary that 0 only be returned when there are truly no
5170  * reference bits set.
5171  *
5172  * XXX: The exact number of bits to check and clear is a matter that
5173  * should be tested and standardized at some point in the future for
5174  * optimal aging of shared pages.
5175  *
5176  * This routine may not block.
5177  */
5178 int
5179 pmap_ts_referenced(vm_page_t m)
5180 {
5181         pv_entry_t pv;
5182         pt_entry_t *pte;
5183         pmap_t pmap;
5184         int rtval = 0;
5185
5186         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5187                 return (rtval);
5188
5189         vm_page_spin_lock(m);
5190         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5191                 if (!pmap_track_modified(pv->pv_pindex))
5192                         continue;
5193                 pmap = pv->pv_pmap;
5194                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5195                 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5196                         atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5197                         rtval++;
5198                         if (rtval > 4)
5199                                 break;
5200                 }
5201         }
5202         vm_page_spin_unlock(m);
5203         return (rtval);
5204 }
5205
5206 /*
5207  *      pmap_is_modified:
5208  *
5209  *      Return whether or not the specified physical page was modified
5210  *      in any physical maps.
5211  */
5212 boolean_t
5213 pmap_is_modified(vm_page_t m)
5214 {
5215         boolean_t res;
5216
5217         res = pmap_testbit(m, PG_M_IDX);
5218         return (res);
5219 }
5220
5221 /*
5222  *      Clear the modify bits on the specified physical page.
5223  */
5224 void
5225 pmap_clear_modify(vm_page_t m)
5226 {
5227         pmap_clearbit(m, PG_M_IDX);
5228 }
5229
5230 /*
5231  *      pmap_clear_reference:
5232  *
5233  *      Clear the reference bit on the specified physical page.
5234  */
5235 void
5236 pmap_clear_reference(vm_page_t m)
5237 {
5238         pmap_clearbit(m, PG_A_IDX);
5239 }
5240
5241 /*
5242  * Miscellaneous support routines follow
5243  */
5244
5245 static
5246 void
5247 i386_protection_init(void)
5248 {
5249         int *kp, prot;
5250
5251         /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit  */
5252         kp = protection_codes;
5253         for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5254                 switch (prot) {
5255                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5256                         /*
5257                          * Read access is also 0. There isn't any execute bit,
5258                          * so just make it readable.
5259                          */
5260                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5261                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5262                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5263                         *kp++ = 0;
5264                         break;
5265                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5266                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5267                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5268                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5269                         *kp++ = pmap_bits_default[PG_RW_IDX];
5270                         break;
5271                 }
5272         }
5273 }
5274
5275 /*
5276  * Map a set of physical memory pages into the kernel virtual
5277  * address space. Return a pointer to where it is mapped. This
5278  * routine is intended to be used for mapping device memory,
5279  * NOT real memory.
5280  *
5281  * NOTE: We can't use pgeflag unless we invalidate the pages one at
5282  *       a time.
5283  *
5284  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5285  *       work whether the cpu supports PAT or not.  The remaining PAT
5286  *       attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5287  *       supports PAT.
5288  */
5289 void *
5290 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5291 {
5292         return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5293 }
5294
5295 void *
5296 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5297 {
5298         return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5299 }
5300
5301 void *
5302 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5303 {
5304         return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5305 }
5306
5307 /*
5308  * Map a set of physical memory pages into the kernel virtual
5309  * address space. Return a pointer to where it is mapped. This
5310  * routine is intended to be used for mapping device memory,
5311  * NOT real memory.
5312  */
5313 void *
5314 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5315 {
5316         vm_offset_t va, tmpva, offset;
5317         pt_entry_t *pte;
5318         vm_size_t tmpsize;
5319
5320         offset = pa & PAGE_MASK;
5321         size = roundup(offset + size, PAGE_SIZE);
5322
5323         va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5324         if (va == 0)
5325                 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5326
5327         pa = pa & ~PAGE_MASK;
5328         for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5329                 pte = vtopte(tmpva);
5330                 *pte = pa |
5331                     kernel_pmap.pmap_bits[PG_RW_IDX] |
5332                     kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5333                     kernel_pmap.pmap_cache_bits[mode];
5334                 tmpsize -= PAGE_SIZE;
5335                 tmpva += PAGE_SIZE;
5336                 pa += PAGE_SIZE;
5337         }
5338         pmap_invalidate_range(&kernel_pmap, va, va + size);
5339         pmap_invalidate_cache_range(va, va + size);
5340
5341         return ((void *)(va + offset));
5342 }
5343
5344 void
5345 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5346 {
5347         vm_offset_t base, offset;
5348
5349         base = va & ~PAGE_MASK;
5350         offset = va & PAGE_MASK;
5351         size = roundup(offset + size, PAGE_SIZE);
5352         pmap_qremove(va, size >> PAGE_SHIFT);
5353         kmem_free(&kernel_map, base, size);
5354 }
5355
5356 /*
5357  * Sets the memory attribute for the specified page.
5358  */
5359 void
5360 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5361 {
5362
5363     m->pat_mode = ma;
5364
5365     /*
5366      * If "m" is a normal page, update its direct mapping.  This update
5367      * can be relied upon to perform any cache operations that are
5368      * required for data coherence.
5369      */
5370     if ((m->flags & PG_FICTITIOUS) == 0)
5371         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5372 }
5373
5374 /*
5375  * Change the PAT attribute on an existing kernel memory map.  Caller
5376  * must ensure that the virtual memory in question is not accessed
5377  * during the adjustment.
5378  */
5379 void
5380 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5381 {
5382         pt_entry_t *pte;
5383         vm_offset_t base;
5384         int changed = 0;
5385
5386         if (va == 0)
5387                 panic("pmap_change_attr: va is NULL");
5388         base = trunc_page(va);
5389
5390         while (count) {
5391                 pte = vtopte(va);
5392                 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5393                        kernel_pmap.pmap_cache_bits[mode];
5394                 --count;
5395                 va += PAGE_SIZE;
5396         }
5397
5398         changed = 1;    /* XXX: not optimal */
5399
5400         /*
5401          * Flush CPU caches if required to make sure any data isn't cached that
5402          * shouldn't be, etc.
5403          */
5404         if (changed) {
5405                 pmap_invalidate_range(&kernel_pmap, base, va);
5406                 pmap_invalidate_cache_range(base, va);
5407         }
5408 }
5409
5410 /*
5411  * perform the pmap work for mincore
5412  */
5413 int
5414 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5415 {
5416         pt_entry_t *ptep, pte;
5417         vm_page_t m;
5418         int val = 0;
5419
5420         lwkt_gettoken(&pmap->pm_token);
5421         ptep = pmap_pte(pmap, addr);
5422
5423         if (ptep && (pte = *ptep) != 0) {
5424                 vm_offset_t pa;
5425
5426                 val = MINCORE_INCORE;
5427                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5428                         goto done;
5429
5430                 pa = pte & PG_FRAME;
5431
5432                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5433                         m = NULL;
5434                 else
5435                         m = PHYS_TO_VM_PAGE(pa);
5436
5437                 /*
5438                  * Modified by us
5439                  */
5440                 if (pte & pmap->pmap_bits[PG_M_IDX])
5441                         val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5442                 /*
5443                  * Modified by someone
5444                  */
5445                 else if (m && (m->dirty || pmap_is_modified(m)))
5446                         val |= MINCORE_MODIFIED_OTHER;
5447                 /*
5448                  * Referenced by us
5449                  */
5450                 if (pte & pmap->pmap_bits[PG_A_IDX])
5451                         val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5452
5453                 /*
5454                  * Referenced by someone
5455                  */
5456                 else if (m && ((m->flags & PG_REFERENCED) ||
5457                                 pmap_ts_referenced(m))) {
5458                         val |= MINCORE_REFERENCED_OTHER;
5459                         vm_page_flag_set(m, PG_REFERENCED);
5460                 }
5461         }
5462 done:
5463         lwkt_reltoken(&pmap->pm_token);
5464
5465         return val;
5466 }
5467
5468 /*
5469  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
5470  * vmspace will be ref'd and the old one will be deref'd.
5471  *
5472  * The vmspace for all lwps associated with the process will be adjusted
5473  * and cr3 will be reloaded if any lwp is the current lwp.
5474  *
5475  * The process must hold the vmspace->vm_map.token for oldvm and newvm
5476  */
5477 void
5478 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5479 {
5480         struct vmspace *oldvm;
5481         struct lwp *lp;
5482
5483         oldvm = p->p_vmspace;
5484         if (oldvm != newvm) {
5485                 if (adjrefs)
5486                         vmspace_ref(newvm);
5487                 p->p_vmspace = newvm;
5488                 KKASSERT(p->p_nthreads == 1);
5489                 lp = RB_ROOT(&p->p_lwp_tree);
5490                 pmap_setlwpvm(lp, newvm);
5491                 if (adjrefs)
5492                         vmspace_rel(oldvm);
5493         }
5494 }
5495
5496 /*
5497  * Set the vmspace for a LWP.  The vmspace is almost universally set the
5498  * same as the process vmspace, but virtual kernels need to swap out contexts
5499  * on a per-lwp basis.
5500  *
5501  * Caller does not necessarily hold any vmspace tokens.  Caller must control
5502  * the lwp (typically be in the context of the lwp).  We use a critical
5503  * section to protect against statclock and hardclock (statistics collection).
5504  */
5505 void
5506 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5507 {
5508         struct vmspace *oldvm;
5509         struct pmap *pmap;
5510
5511         oldvm = lp->lwp_vmspace;
5512
5513         if (oldvm != newvm) {
5514                 crit_enter();
5515                 lp->lwp_vmspace = newvm;
5516                 if (curthread->td_lwp == lp) {
5517                         pmap = vmspace_pmap(newvm);
5518                         ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5519                         if (pmap->pm_active_lock & CPULOCK_EXCL)
5520                                 pmap_interlock_wait(newvm);
5521 #if defined(SWTCH_OPTIM_STATS)
5522                         tlb_flush_count++;
5523 #endif
5524                         if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5525                                 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5526                         } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5527                                 curthread->td_pcb->pcb_cr3 = KPML4phys;
5528                         } else {
5529                                 panic("pmap_setlwpvm: unknown pmap type\n");
5530                         }
5531                         load_cr3(curthread->td_pcb->pcb_cr3);
5532                         pmap = vmspace_pmap(oldvm);
5533                         ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5534                                                mycpu->gd_cpuid);
5535                 }
5536                 crit_exit();
5537         }
5538 }
5539
5540 /*
5541  * Called when switching to a locked pmap, used to interlock against pmaps
5542  * undergoing modifications to prevent us from activating the MMU for the
5543  * target pmap until all such modifications have completed.  We have to do
5544  * this because the thread making the modifications has already set up its
5545  * SMP synchronization mask.
5546  *
5547  * This function cannot sleep!
5548  *
5549  * No requirements.
5550  */
5551 void
5552 pmap_interlock_wait(struct vmspace *vm)
5553 {
5554         struct pmap *pmap = &vm->vm_pmap;
5555
5556         if (pmap->pm_active_lock & CPULOCK_EXCL) {
5557                 crit_enter();
5558                 KKASSERT(curthread->td_critcount >= 2);
5559                 DEBUG_PUSH_INFO("pmap_interlock_wait");
5560                 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5561                         cpu_ccfence();
5562                         lwkt_process_ipiq();
5563                 }
5564                 DEBUG_POP_INFO();
5565                 crit_exit();
5566         }
5567 }
5568
5569 vm_offset_t
5570 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5571 {
5572
5573         if ((obj == NULL) || (size < NBPDR) ||
5574             ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5575                 return addr;
5576         }
5577
5578         addr = roundup2(addr, NBPDR);
5579         return addr;
5580 }
5581
5582 /*
5583  * Used by kmalloc/kfree, page already exists at va
5584  */
5585 vm_page_t
5586 pmap_kvtom(vm_offset_t va)
5587 {
5588         pt_entry_t *ptep = vtopte(va);
5589
5590         KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5591         return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5592 }
5593
5594 /*
5595  * Initialize machine-specific shared page directory support.  This
5596  * is executed when a VM object is created.
5597  */
5598 void
5599 pmap_object_init(vm_object_t object)
5600 {
5601         object->md.pmap_rw = NULL;
5602         object->md.pmap_ro = NULL;
5603 }
5604
5605 /*
5606  * Clean up machine-specific shared page directory support.  This
5607  * is executed when a VM object is destroyed.
5608  */
5609 void
5610 pmap_object_free(vm_object_t object)
5611 {
5612         pmap_t pmap;
5613
5614         if ((pmap = object->md.pmap_rw) != NULL) {
5615                 object->md.pmap_rw = NULL;
5616                 pmap_remove_noinval(pmap,
5617                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5618                 CPUMASK_ASSZERO(pmap->pm_active);
5619                 pmap_release(pmap);
5620                 pmap_puninit(pmap);
5621                 kfree(pmap, M_OBJPMAP);
5622         }
5623         if ((pmap = object->md.pmap_ro) != NULL) {
5624                 object->md.pmap_ro = NULL;
5625                 pmap_remove_noinval(pmap,
5626                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5627                 CPUMASK_ASSZERO(pmap->pm_active);
5628                 pmap_release(pmap);
5629                 pmap_puninit(pmap);
5630                 kfree(pmap, M_OBJPMAP);
5631         }
5632 }
5633
5634 /*
5635  * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
5636  * VM page and issue a pginfo->callback.
5637  *
5638  * We are expected to dispose of any non-NULL pte_pv.
5639  */
5640 static
5641 void
5642 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
5643                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
5644                       vm_offset_t va, pt_entry_t *ptep, void *arg)
5645 {
5646         struct pmap_pgscan_info *pginfo = arg;
5647         vm_page_t m;
5648
5649         if (pte_pv) {
5650                 /*
5651                  * Try to busy the page while we hold the pte_pv locked.
5652                  */
5653                 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
5654                 if (vm_page_busy_try(m, TRUE) == 0) {
5655                         if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
5656                                 /*
5657                                  * The callback is issued with the pte_pv
5658                                  * unlocked and put away, and the pt_pv
5659                                  * unlocked.
5660                                  */
5661                                 pv_put(pte_pv);
5662                                 if (pt_pv)
5663                                         pv_unlock(pt_pv);
5664                                 if (pginfo->callback(pginfo, va, m) < 0)
5665                                         info->stop = 1;
5666                                 if (pt_pv)
5667                                         pv_lock(pt_pv);
5668                         } else {
5669                                 vm_page_wakeup(m);
5670                                 pv_put(pte_pv);
5671                         }
5672                 } else {
5673                         ++pginfo->busycount;
5674                         pv_put(pte_pv);
5675                 }
5676         } else if (sharept) {
5677                 /* shared page table */
5678         } else {
5679                 /* else unmanaged page */
5680         }
5681 }
5682
5683 void
5684 pmap_pgscan(struct pmap_pgscan_info *pginfo)
5685 {
5686         struct pmap_scan_info info;
5687
5688         pginfo->offset = pginfo->beg_addr;
5689         info.pmap = pginfo->pmap;
5690         info.sva = pginfo->beg_addr;
5691         info.eva = pginfo->end_addr;
5692         info.func = pmap_pgscan_callback;
5693         info.arg = pginfo;
5694         pmap_scan(&info, 0);
5695         if (info.stop == 0)
5696                 pginfo->offset = pginfo->end_addr;
5697 }