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
 156 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
 157
 158 vm_paddr_t avail_start;         /* PA of first available physical page */
 159 vm_paddr_t avail_end;           /* PA of last available physical page */
 160 vm_offset_t virtual2_start;     /* cutout free area prior to kernel start */
 161 vm_offset_t virtual2_end;
 162 vm_offset_t virtual_start;      /* VA of first avail page (after kernel bss) */
 163 vm_offset_t virtual_end;        /* VA of last avail page (end of kernel AS) */
 164 vm_offset_t KvaStart;           /* VA start of KVA space */
 165 vm_offset_t KvaEnd;             /* VA end of KVA space (non-inclusive) */
 166 vm_offset_t KvaSize;            /* max size of kernel virtual address space */
 167 static boolean_t pmap_initialized = FALSE;      /* Has pmap_init completed? */
 168 //static int pgeflag;           /* PG_G or-in */
 169 //static int pseflag;           /* PG_PS or-in */
 170 uint64_t PatMsr;
 171
 172 static int ndmpdp;
 173 static vm_paddr_t dmaplimit;
 174 static int nkpt;
 175 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
 176
 177 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE];        /* PAT -> PG_ bits */
 178 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/    /* PAT -> PG_ bits */
 179
 180 static uint64_t KPTbase;
 181 static uint64_t KPTphys;
 182 static uint64_t KPDphys;        /* phys addr of kernel level 2 */
 183 static uint64_t KPDbase;        /* phys addr of kernel level 2 @ KERNBASE */
 184 uint64_t KPDPphys;      /* phys addr of kernel level 3 */
 185 uint64_t KPML4phys;     /* phys addr of kernel level 4 */
 186
 187 static uint64_t DMPDphys;       /* phys addr of direct mapped level 2 */
 188 static uint64_t DMPDPphys;      /* phys addr of direct mapped level 3 */
 189
 190 /*
 191  * Data for the pv entry allocation mechanism
 192  */
 193 static vm_zone_t pvzone;
 194 static struct vm_zone pvzone_store;
 195 static struct vm_object pvzone_obj;
 196 static int pv_entry_max=0, pv_entry_high_water=0;
 197 static int pmap_pagedaemon_waken = 0;
 198 static struct pv_entry *pvinit;
 199
 200 /*
 201  * All those kernel PT submaps that BSD is so fond of
 202  */
 203 pt_entry_t *CMAP1 = NULL, *ptmmap;
 204 caddr_t CADDR1 = NULL, ptvmmap = NULL;
 205 static pt_entry_t *msgbufmap;
 206 struct msgbuf *msgbufp=NULL;
 207
 208 /*
 209  * PMAP default PG_* bits. Needed to be able to add
 210  * EPT/NPT pagetable pmap_bits for the VMM module
 211  */
 212 uint64_t pmap_bits_default[] = {
 213                 REGULAR_PMAP,                                   /* TYPE_IDX             0 */
 214                 X86_PG_V,                                       /* PG_V_IDX             1 */
 215                 X86_PG_RW,                                      /* PG_RW_IDX            2 */
 216                 X86_PG_U,                                       /* PG_U_IDX             3 */
 217                 X86_PG_A,                                       /* PG_A_IDX             4 */
 218                 X86_PG_M,                                       /* PG_M_IDX             5 */
 219                 X86_PG_PS,                                      /* PG_PS_IDX3           6 */
 220                 X86_PG_G,                                       /* PG_G_IDX             7 */
 221                 X86_PG_AVAIL1,                                  /* PG_AVAIL1_IDX        8 */
 222                 X86_PG_AVAIL2,                                  /* PG_AVAIL2_IDX        9 */
 223                 X86_PG_AVAIL3,                                  /* PG_AVAIL3_IDX        10 */
 224                 X86_PG_NC_PWT | X86_PG_NC_PCD,                  /* PG_N_IDX     11 */
 225 };
 226 /*
 227  * Crashdump maps.
 228  */
 229 static pt_entry_t *pt_crashdumpmap;
 230 static caddr_t crashdumpmap;
 231
 232 static int pmap_debug = 0;
 233 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
 234     &pmap_debug, 0, "Debug pmap's");
 235 #ifdef PMAP_DEBUG2
 236 static int pmap_enter_debug = 0;
 237 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
 238     &pmap_enter_debug, 0, "Debug pmap_enter's");
 239 #endif
 240 static int pmap_yield_count = 64;
 241 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
 242     &pmap_yield_count, 0, "Yield during init_pt/release");
 243 static int pmap_mmu_optimize = 0;
 244 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
 245     &pmap_mmu_optimize, 0, "Share page table pages when possible");
 246 int pmap_fast_kernel_cpusync = 0;
 247 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
 248     &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
 249 int pmap_dynamic_delete = -1;
 250 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
 251     &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
 252
 253 #define DISABLE_PSE
 254
 255 /* Standard user access funtions */
 256 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
 257     size_t *lencopied);
 258 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
 259 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
 260 extern int std_fubyte (const void *base);
 261 extern int std_subyte (void *base, int byte);
 262 extern long std_fuword (const void *base);
 263 extern int std_suword (void *base, long word);
 264 extern int std_suword32 (void *base, int word);
 265
 266 static void pv_hold(pv_entry_t pv);
 267 static int _pv_hold_try(pv_entry_t pv
 268                                 PMAP_DEBUG_DECL);
 269 static void pv_drop(pv_entry_t pv);
 270 static void _pv_lock(pv_entry_t pv
 271                                 PMAP_DEBUG_DECL);
 272 static void pv_unlock(pv_entry_t pv);
 273 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
 274                                 PMAP_DEBUG_DECL);
 275 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
 276                                 PMAP_DEBUG_DECL);
 277 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
 278 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
 279 static void pv_put(pv_entry_t pv);
 280 static void pv_free(pv_entry_t pv, pv_entry_t pvp, int putaway);
 281 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
 282 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
 283                       pv_entry_t *pvpp);
 284 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
 285                       pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
 286 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
 287                         pmap_inval_bulk_t *bulk, int destroy);
 288 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
 289 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
 290                         pmap_inval_bulk_t *bulk);
 291
 292 struct pmap_scan_info;
 293 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
 294                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
 295                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 296 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
 297                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
 298                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 299
 300 static void i386_protection_init (void);
 301 static void create_pagetables(vm_paddr_t *firstaddr);
 302 static void pmap_remove_all (vm_page_t m);
 303 static boolean_t pmap_testbit (vm_page_t m, int bit);
 304
 305 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
 306 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
 307
 308 static void pmap_pinit_defaults(struct pmap *pmap);
 309
 310 static unsigned pdir4mb;
 311
 312 static int
 313 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
 314 {
 315         if (pv1->pv_pindex < pv2->pv_pindex)
 316                 return(-1);
 317         if (pv1->pv_pindex > pv2->pv_pindex)
 318                 return(1);
 319         return(0);
 320 }
 321
 322 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
 323              pv_entry_compare, vm_pindex_t, pv_pindex);
 324
 325 static __inline
 326 void
 327 pmap_page_stats_adding(vm_page_t m)
 328 {
 329         globaldata_t gd = mycpu;
 330
 331         if (TAILQ_EMPTY(&m->md.pv_list)) {
 332                 ++gd->gd_vmtotal.t_arm;
 333         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 334                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 335                 ++gd->gd_vmtotal.t_armshr;
 336                 ++gd->gd_vmtotal.t_avmshr;
 337         } else {
 338                 ++gd->gd_vmtotal.t_avmshr;
 339         }
 340 }
 341
 342 static __inline
 343 void
 344 pmap_page_stats_deleting(vm_page_t m)
 345 {
 346         globaldata_t gd = mycpu;
 347
 348         if (TAILQ_EMPTY(&m->md.pv_list)) {
 349                 --gd->gd_vmtotal.t_arm;
 350         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 351                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 352                 --gd->gd_vmtotal.t_armshr;
 353                 --gd->gd_vmtotal.t_avmshr;
 354         } else {
 355                 --gd->gd_vmtotal.t_avmshr;
 356         }
 357 }
 358
 359 /*
 360  * Move the kernel virtual free pointer to the next
 361  * 2MB.  This is used to help improve performance
 362  * by using a large (2MB) page for much of the kernel
 363  * (.text, .data, .bss)
 364  */
 365 static
 366 vm_offset_t
 367 pmap_kmem_choose(vm_offset_t addr)
 368 {
 369         vm_offset_t newaddr = addr;
 370
 371         newaddr = roundup2(addr, NBPDR);
 372         return newaddr;
 373 }
 374
 375 /*
 376  * pmap_pte_quick:
 377  *
 378  *      Super fast pmap_pte routine best used when scanning the pv lists.
 379  *      This eliminates many course-grained invltlb calls.  Note that many of
 380  *      the pv list scans are across different pmaps and it is very wasteful
 381  *      to do an entire invltlb when checking a single mapping.
 382  */
 383 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
 384
 385 static
 386 pt_entry_t *
 387 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
 388 {
 389         return pmap_pte(pmap, va);
 390 }
 391
 392 /*
 393  * Returns the pindex of a page table entry (representing a terminal page).
 394  * There are NUPTE_TOTAL page table entries possible (a huge number)
 395  *
 396  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
 397  * We want to properly translate negative KVAs.
 398  */
 399 static __inline
 400 vm_pindex_t
 401 pmap_pte_pindex(vm_offset_t va)
 402 {
 403         return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
 404 }
 405
 406 /*
 407  * Returns the pindex of a page table.
 408  */
 409 static __inline
 410 vm_pindex_t
 411 pmap_pt_pindex(vm_offset_t va)
 412 {
 413         return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
 414 }
 415
 416 /*
 417  * Returns the pindex of a page directory.
 418  */
 419 static __inline
 420 vm_pindex_t
 421 pmap_pd_pindex(vm_offset_t va)
 422 {
 423         return (NUPTE_TOTAL + NUPT_TOTAL +
 424                 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
 425 }
 426
 427 static __inline
 428 vm_pindex_t
 429 pmap_pdp_pindex(vm_offset_t va)
 430 {
 431         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
 432                 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
 433 }
 434
 435 static __inline
 436 vm_pindex_t
 437 pmap_pml4_pindex(void)
 438 {
 439         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
 440 }
 441
 442 /*
 443  * Return various clipped indexes for a given VA
 444  *
 445  * Returns the index of a pt in a page directory, representing a page
 446  * table.
 447  */
 448 static __inline
 449 vm_pindex_t
 450 pmap_pt_index(vm_offset_t va)
 451 {
 452         return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
 453 }
 454
 455 /*
 456  * Returns the index of a pd in a page directory page, representing a page
 457  * directory.
 458  */
 459 static __inline
 460 vm_pindex_t
 461 pmap_pd_index(vm_offset_t va)
 462 {
 463         return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
 464 }
 465
 466 /*
 467  * Returns the index of a pdp in the pml4 table, representing a page
 468  * directory page.
 469  */
 470 static __inline
 471 vm_pindex_t
 472 pmap_pdp_index(vm_offset_t va)
 473 {
 474         return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
 475 }
 476
 477 /*
 478  * Generic procedure to index a pte from a pt, pd, or pdp.
 479  *
 480  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
 481  *       a page table page index but is instead of PV lookup index.
 482  */
 483 static
 484 void *
 485 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
 486 {
 487         pt_entry_t *pte;
 488
 489         pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
 490         return(&pte[pindex]);
 491 }
 492
 493 /*
 494  * Return pointer to PDP slot in the PML4
 495  */
 496 static __inline
 497 pml4_entry_t *
 498 pmap_pdp(pmap_t pmap, vm_offset_t va)
 499 {
 500         return (&pmap->pm_pml4[pmap_pdp_index(va)]);
 501 }
 502
 503 /*
 504  * Return pointer to PD slot in the PDP given a pointer to the PDP
 505  */
 506 static __inline
 507 pdp_entry_t *
 508 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
 509 {
 510         pdp_entry_t *pd;
 511
 512         pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
 513         return (&pd[pmap_pd_index(va)]);
 514 }
 515
 516 /*
 517  * Return pointer to PD slot in the PDP.
 518  */
 519 static __inline
 520 pdp_entry_t *
 521 pmap_pd(pmap_t pmap, vm_offset_t va)
 522 {
 523         pml4_entry_t *pdp;
 524
 525         pdp = pmap_pdp(pmap, va);
 526         if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
 527                 return NULL;
 528         return (pmap_pdp_to_pd(*pdp, va));
 529 }
 530
 531 /*
 532  * Return pointer to PT slot in the PD given a pointer to the PD
 533  */
 534 static __inline
 535 pd_entry_t *
 536 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
 537 {
 538         pd_entry_t *pt;
 539
 540         pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
 541         return (&pt[pmap_pt_index(va)]);
 542 }
 543
 544 /*
 545  * Return pointer to PT slot in the PD
 546  *
 547  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
 548  *                   so we cannot lookup the PD via the PDP.  Instead we
 549  *                   must look it up via the pmap.
 550  */
 551 static __inline
 552 pd_entry_t *
 553 pmap_pt(pmap_t pmap, vm_offset_t va)
 554 {
 555         pdp_entry_t *pd;
 556         pv_entry_t pv;
 557         vm_pindex_t pd_pindex;
 558
 559         if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
 560                 pd_pindex = pmap_pd_pindex(va);
 561                 spin_lock(&pmap->pm_spin);
 562                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
 563                 spin_unlock(&pmap->pm_spin);
 564                 if (pv == NULL || pv->pv_m == NULL)
 565                         return NULL;
 566                 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
 567         } else {
 568                 pd = pmap_pd(pmap, va);
 569                 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
 570                          return NULL;
 571                 return (pmap_pd_to_pt(*pd, va));
 572         }
 573 }
 574
 575 /*
 576  * Return pointer to PTE slot in the PT given a pointer to the PT
 577  */
 578 static __inline
 579 pt_entry_t *
 580 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
 581 {
 582         pt_entry_t *pte;
 583
 584         pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
 585         return (&pte[pmap_pte_index(va)]);
 586 }
 587
 588 /*
 589  * Return pointer to PTE slot in the PT
 590  */
 591 static __inline
 592 pt_entry_t *
 593 pmap_pte(pmap_t pmap, vm_offset_t va)
 594 {
 595         pd_entry_t *pt;
 596
 597         pt = pmap_pt(pmap, va);
 598         if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
 599                  return NULL;
 600         if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
 601                 return ((pt_entry_t *)pt);
 602         return (pmap_pt_to_pte(*pt, va));
 603 }
 604
 605 /*
 606  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
 607  * the PT layer.  This will speed up core pmap operations considerably.
 608  *
 609  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
 610  *       must be in a known associated state (typically by being locked when
 611  *       the pmap spinlock isn't held).  We allow the race for that case.
 612  */
 613 static __inline
 614 void
 615 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
 616 {
 617         if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
 618                 pv->pv_pmap->pm_pvhint = pv;
 619 }
 620
 621
 622 /*
 623  * Return address of PT slot in PD (KVM only)
 624  *
 625  * Cannot be used for user page tables because it might interfere with
 626  * the shared page-table-page optimization (pmap_mmu_optimize).
 627  */
 628 static __inline
 629 pd_entry_t *
 630 vtopt(vm_offset_t va)
 631 {
 632         uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
 633                                   NPML4EPGSHIFT)) - 1);
 634
 635         return (PDmap + ((va >> PDRSHIFT) & mask));
 636 }
 637
 638 /*
 639  * KVM - return address of PTE slot in PT
 640  */
 641 static __inline
 642 pt_entry_t *
 643 vtopte(vm_offset_t va)
 644 {
 645         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 646                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 647
 648         return (PTmap + ((va >> PAGE_SHIFT) & mask));
 649 }
 650
 651 static uint64_t
 652 allocpages(vm_paddr_t *firstaddr, long n)
 653 {
 654         uint64_t ret;
 655
 656         ret = *firstaddr;
 657         bzero((void *)ret, n * PAGE_SIZE);
 658         *firstaddr += n * PAGE_SIZE;
 659         return (ret);
 660 }
 661
 662 static
 663 void
 664 create_pagetables(vm_paddr_t *firstaddr)
 665 {
 666         long i;         /* must be 64 bits */
 667         long nkpt_base;
 668         long nkpt_phys;
 669         int j;
 670
 671         /*
 672          * We are running (mostly) V=P at this point
 673          *
 674          * Calculate NKPT - number of kernel page tables.  We have to
 675          * accomodoate prealloction of the vm_page_array, dump bitmap,
 676          * MSGBUF_SIZE, and other stuff.  Be generous.
 677          *
 678          * Maxmem is in pages.
 679          *
 680          * ndmpdp is the number of 1GB pages we wish to map.
 681          */
 682         ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
 683         if (ndmpdp < 4)         /* Minimum 4GB of dirmap */
 684                 ndmpdp = 4;
 685         KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
 686
 687         /*
 688          * Starting at the beginning of kvm (not KERNBASE).
 689          */
 690         nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
 691         nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
 692         nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
 693                        ndmpdp) + 511) / 512;
 694         nkpt_phys += 128;
 695
 696         /*
 697          * Starting at KERNBASE - map 2G worth of page table pages.
 698          * KERNBASE is offset -2G from the end of kvm.
 699          */
 700         nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
 701
 702         /*
 703          * Allocate pages
 704          */
 705         KPTbase = allocpages(firstaddr, nkpt_base);
 706         KPTphys = allocpages(firstaddr, nkpt_phys);
 707         KPML4phys = allocpages(firstaddr, 1);
 708         KPDPphys = allocpages(firstaddr, NKPML4E);
 709         KPDphys = allocpages(firstaddr, NKPDPE);
 710
 711         /*
 712          * Calculate the page directory base for KERNBASE,
 713          * that is where we start populating the page table pages.
 714          * Basically this is the end - 2.
 715          */
 716         KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
 717
 718         DMPDPphys = allocpages(firstaddr, NDMPML4E);
 719         if ((amd_feature & AMDID_PAGE1GB) == 0)
 720                 DMPDphys = allocpages(firstaddr, ndmpdp);
 721         dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
 722
 723         /*
 724          * Fill in the underlying page table pages for the area around
 725          * KERNBASE.  This remaps low physical memory to KERNBASE.
 726          *
 727          * Read-only from zero to physfree
 728          * XXX not fully used, underneath 2M pages
 729          */
 730         for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
 731                 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
 732                 ((pt_entry_t *)KPTbase)[i] |=
 733                     pmap_bits_default[PG_RW_IDX] |
 734                     pmap_bits_default[PG_V_IDX] |
 735                     pmap_bits_default[PG_G_IDX];
 736         }
 737
 738         /*
 739          * Now map the initial kernel page tables.  One block of page
 740          * tables is placed at the beginning of kernel virtual memory,
 741          * and another block is placed at KERNBASE to map the kernel binary,
 742          * data, bss, and initial pre-allocations.
 743          */
 744         for (i = 0; i < nkpt_base; i++) {
 745                 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
 746                 ((pd_entry_t *)KPDbase)[i] |=
 747                     pmap_bits_default[PG_RW_IDX] |
 748                     pmap_bits_default[PG_V_IDX];
 749         }
 750         for (i = 0; i < nkpt_phys; i++) {
 751                 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
 752                 ((pd_entry_t *)KPDphys)[i] |=
 753                     pmap_bits_default[PG_RW_IDX] |
 754                     pmap_bits_default[PG_V_IDX];
 755         }
 756
 757         /*
 758          * Map from zero to end of allocations using 2M pages as an
 759          * optimization.  This will bypass some of the KPTBase pages
 760          * above in the KERNBASE area.
 761          */
 762         for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
 763                 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
 764                 ((pd_entry_t *)KPDbase)[i] |=
 765                     pmap_bits_default[PG_RW_IDX] |
 766                     pmap_bits_default[PG_V_IDX] |
 767                     pmap_bits_default[PG_PS_IDX] |
 768                     pmap_bits_default[PG_G_IDX];
 769         }
 770
 771         /*
 772          * And connect up the PD to the PDP.  The kernel pmap is expected
 773          * to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
 774          */
 775         for (i = 0; i < NKPDPE; i++) {
 776                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
 777                                 KPDphys + (i << PAGE_SHIFT);
 778                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
 779                     pmap_bits_default[PG_RW_IDX] |
 780                     pmap_bits_default[PG_V_IDX] |
 781                     pmap_bits_default[PG_U_IDX];
 782         }
 783
 784         /*
 785          * Now set up the direct map space using either 2MB or 1GB pages
 786          * Preset PG_M and PG_A because demotion expects it.
 787          *
 788          * When filling in entries in the PD pages make sure any excess
 789          * entries are set to zero as we allocated enough PD pages
 790          */
 791         if ((amd_feature & AMDID_PAGE1GB) == 0) {
 792                 for (i = 0; i < NPDEPG * ndmpdp; i++) {
 793                         ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
 794                         ((pd_entry_t *)DMPDphys)[i] |=
 795                             pmap_bits_default[PG_RW_IDX] |
 796                             pmap_bits_default[PG_V_IDX] |
 797                             pmap_bits_default[PG_PS_IDX] |
 798                             pmap_bits_default[PG_G_IDX] |
 799                             pmap_bits_default[PG_M_IDX] |
 800                             pmap_bits_default[PG_A_IDX];
 801                 }
 802
 803                 /*
 804                  * And the direct map space's PDP
 805                  */
 806                 for (i = 0; i < ndmpdp; i++) {
 807                         ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
 808                                                         (i << PAGE_SHIFT);
 809                         ((pdp_entry_t *)DMPDPphys)[i] |=
 810                             pmap_bits_default[PG_RW_IDX] |
 811                             pmap_bits_default[PG_V_IDX] |
 812                             pmap_bits_default[PG_U_IDX];
 813                 }
 814         } else {
 815                 for (i = 0; i < ndmpdp; i++) {
 816                         ((pdp_entry_t *)DMPDPphys)[i] =
 817                                                 (vm_paddr_t)i << PDPSHIFT;
 818                         ((pdp_entry_t *)DMPDPphys)[i] |=
 819                             pmap_bits_default[PG_RW_IDX] |
 820                             pmap_bits_default[PG_V_IDX] |
 821                             pmap_bits_default[PG_PS_IDX] |
 822                             pmap_bits_default[PG_G_IDX] |
 823                             pmap_bits_default[PG_M_IDX] |
 824                             pmap_bits_default[PG_A_IDX];
 825                 }
 826         }
 827
 828         /* And recursively map PML4 to itself in order to get PTmap */
 829         ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
 830         ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
 831             pmap_bits_default[PG_RW_IDX] |
 832             pmap_bits_default[PG_V_IDX] |
 833             pmap_bits_default[PG_U_IDX];
 834
 835         /*
 836          * Connect the Direct Map slots up to the PML4
 837          */
 838         for (j = 0; j < NDMPML4E; ++j) {
 839                 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
 840                     (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
 841                     pmap_bits_default[PG_RW_IDX] |
 842                     pmap_bits_default[PG_V_IDX] |
 843                     pmap_bits_default[PG_U_IDX];
 844         }
 845
 846         /*
 847          * Connect the KVA slot up to the PML4
 848          */
 849         ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
 850         ((pdp_entry_t *)KPML4phys)[KPML4I] |=
 851             pmap_bits_default[PG_RW_IDX] |
 852             pmap_bits_default[PG_V_IDX] |
 853             pmap_bits_default[PG_U_IDX];
 854 }
 855
 856 /*
 857  *      Bootstrap the system enough to run with virtual memory.
 858  *
 859  *      On the i386 this is called after mapping has already been enabled
 860  *      and just syncs the pmap module with what has already been done.
 861  *      [We can't call it easily with mapping off since the kernel is not
 862  *      mapped with PA == VA, hence we would have to relocate every address
 863  *      from the linked base (virtual) address "KERNBASE" to the actual
 864  *      (physical) address starting relative to 0]
 865  */
 866 void
 867 pmap_bootstrap(vm_paddr_t *firstaddr)
 868 {
 869         vm_offset_t va;
 870         pt_entry_t *pte;
 871
 872         KvaStart = VM_MIN_KERNEL_ADDRESS;
 873         KvaEnd = VM_MAX_KERNEL_ADDRESS;
 874         KvaSize = KvaEnd - KvaStart;
 875
 876         avail_start = *firstaddr;
 877
 878         /*
 879          * Create an initial set of page tables to run the kernel in.
 880          */
 881         create_pagetables(firstaddr);
 882
 883         virtual2_start = KvaStart;
 884         virtual2_end = PTOV_OFFSET;
 885
 886         virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
 887         virtual_start = pmap_kmem_choose(virtual_start);
 888
 889         virtual_end = VM_MAX_KERNEL_ADDRESS;
 890
 891         /* XXX do %cr0 as well */
 892         load_cr4(rcr4() | CR4_PGE | CR4_PSE);
 893         load_cr3(KPML4phys);
 894
 895         /*
 896          * Initialize protection array.
 897          */
 898         i386_protection_init();
 899
 900         /*
 901          * The kernel's pmap is statically allocated so we don't have to use
 902          * pmap_create, which is unlikely to work correctly at this part of
 903          * the boot sequence (XXX and which no longer exists).
 904          */
 905         kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
 906         kernel_pmap.pm_count = 1;
 907         CPUMASK_ASSALLONES(kernel_pmap.pm_active);
 908         RB_INIT(&kernel_pmap.pm_pvroot);
 909         spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
 910         lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
 911
 912         /*
 913          * Reserve some special page table entries/VA space for temporary
 914          * mapping of pages.
 915          */
 916 #define SYSMAP(c, p, v, n)      \
 917         v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
 918
 919         va = virtual_start;
 920         pte = vtopte(va);
 921
 922         /*
 923          * CMAP1/CMAP2 are used for zeroing and copying pages.
 924          */
 925         SYSMAP(caddr_t, CMAP1, CADDR1, 1)
 926
 927         /*
 928          * Crashdump maps.
 929          */
 930         SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
 931
 932         /*
 933          * ptvmmap is used for reading arbitrary physical pages via
 934          * /dev/mem.
 935          */
 936         SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
 937
 938         /*
 939          * msgbufp is used to map the system message buffer.
 940          * XXX msgbufmap is not used.
 941          */
 942         SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
 943                atop(round_page(MSGBUF_SIZE)))
 944
 945         virtual_start = va;
 946         virtual_start = pmap_kmem_choose(virtual_start);
 947
 948         *CMAP1 = 0;
 949
 950         /*
 951          * PG_G is terribly broken on SMP because we IPI invltlb's in some
 952          * cases rather then invl1pg.  Actually, I don't even know why it
 953          * works under UP because self-referential page table mappings
 954          */
 955 //      pgeflag = 0;
 956
 957 /*
 958  * Initialize the 4MB page size flag
 959  */
 960 //      pseflag = 0;
 961 /*
 962  * The 4MB page version of the initial
 963  * kernel page mapping.
 964  */
 965         pdir4mb = 0;
 966
 967 #if !defined(DISABLE_PSE)
 968         if (cpu_feature & CPUID_PSE) {
 969                 pt_entry_t ptditmp;
 970                 /*
 971                  * Note that we have enabled PSE mode
 972                  */
 973 //              pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
 974                 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
 975                 ptditmp &= ~(NBPDR - 1);
 976                 ptditmp |= pmap_bits_default[PG_V_IDX] |
 977                     pmap_bits_default[PG_RW_IDX] |
 978                     pmap_bits_default[PG_PS_IDX] |
 979                     pmap_bits_default[PG_U_IDX];
 980 //                  pgeflag;
 981                 pdir4mb = ptditmp;
 982         }
 983 #endif
 984         cpu_invltlb();
 985
 986         /* Initialize the PAT MSR */
 987         pmap_init_pat();
 988         pmap_pinit_defaults(&kernel_pmap);
 989
 990         TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
 991                           &pmap_fast_kernel_cpusync);
 992
 993 }
 994
 995 /*
 996  * Setup the PAT MSR.
 997  */
 998 void
 999 pmap_init_pat(void)
1000 {
1001         uint64_t pat_msr;
1002         u_long cr0, cr4;
1003
1004         /*
1005          * Default values mapping PATi,PCD,PWT bits at system reset.
1006          * The default values effectively ignore the PATi bit by
1007          * repeating the encodings for 0-3 in 4-7, and map the PCD
1008          * and PWT bit combinations to the expected PAT types.
1009          */
1010         pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |        /* 000 */
1011                   PAT_VALUE(1, PAT_WRITE_THROUGH) |     /* 001 */
1012                   PAT_VALUE(2, PAT_UNCACHED) |          /* 010 */
1013                   PAT_VALUE(3, PAT_UNCACHEABLE) |       /* 011 */
1014                   PAT_VALUE(4, PAT_WRITE_BACK) |        /* 100 */
1015                   PAT_VALUE(5, PAT_WRITE_THROUGH) |     /* 101 */
1016                   PAT_VALUE(6, PAT_UNCACHED) |          /* 110 */
1017                   PAT_VALUE(7, PAT_UNCACHEABLE);        /* 111 */
1018         pat_pte_index[PAT_WRITE_BACK]   = 0;
1019         pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
1020         pat_pte_index[PAT_UNCACHED]     = X86_PG_NC_PCD;
1021         pat_pte_index[PAT_UNCACHEABLE]  = X86_PG_NC_PCD | X86_PG_NC_PWT;
1022         pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1023         pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1024
1025         if (cpu_feature & CPUID_PAT) {
1026                 /*
1027                  * If we support the PAT then set-up entries for
1028                  * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1029                  * 4 and 5.
1030                  */
1031                 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1032                           PAT_VALUE(4, PAT_WRITE_PROTECTED);
1033                 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1034                           PAT_VALUE(5, PAT_WRITE_COMBINING);
1035                 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1036                 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1037
1038                 /*
1039                  * Then enable the PAT
1040                  */
1041
1042                 /* Disable PGE. */
1043                 cr4 = rcr4();
1044                 load_cr4(cr4 & ~CR4_PGE);
1045
1046                 /* Disable caches (CD = 1, NW = 0). */
1047                 cr0 = rcr0();
1048                 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1049
1050                 /* Flushes caches and TLBs. */
1051                 wbinvd();
1052                 cpu_invltlb();
1053
1054                 /* Update PAT and index table. */
1055                 wrmsr(MSR_PAT, pat_msr);
1056
1057                 /* Flush caches and TLBs again. */
1058                 wbinvd();
1059                 cpu_invltlb();
1060
1061                 /* Restore caches and PGE. */
1062                 load_cr0(cr0);
1063                 load_cr4(cr4);
1064                 PatMsr = pat_msr;
1065         }
1066 }
1067
1068 /*
1069  * Set 4mb pdir for mp startup
1070  */
1071 void
1072 pmap_set_opt(void)
1073 {
1074         if (cpu_feature & CPUID_PSE) {
1075                 load_cr4(rcr4() | CR4_PSE);
1076                 if (pdir4mb && mycpu->gd_cpuid == 0) {  /* only on BSP */
1077                         cpu_invltlb();
1078                 }
1079         }
1080 }
1081
1082 /*
1083  *      Initialize the pmap module.
1084  *      Called by vm_init, to initialize any structures that the pmap
1085  *      system needs to map virtual memory.
1086  *      pmap_init has been enhanced to support in a fairly consistant
1087  *      way, discontiguous physical memory.
1088  */
1089 void
1090 pmap_init(void)
1091 {
1092         int i;
1093         int initial_pvs;
1094
1095         /*
1096          * Allocate memory for random pmap data structures.  Includes the
1097          * pv_head_table.
1098          */
1099
1100         for (i = 0; i < vm_page_array_size; i++) {
1101                 vm_page_t m;
1102
1103                 m = &vm_page_array[i];
1104                 TAILQ_INIT(&m->md.pv_list);
1105         }
1106
1107         /*
1108          * init the pv free list
1109          */
1110         initial_pvs = vm_page_array_size;
1111         if (initial_pvs < MINPV)
1112                 initial_pvs = MINPV;
1113         pvzone = &pvzone_store;
1114         pvinit = (void *)kmem_alloc(&kernel_map,
1115                                     initial_pvs * sizeof (struct pv_entry),
1116                                     VM_SUBSYS_PVENTRY);
1117         zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1118                   pvinit, initial_pvs);
1119
1120         /*
1121          * Now it is safe to enable pv_table recording.
1122          */
1123         pmap_initialized = TRUE;
1124 }
1125
1126 /*
1127  * Initialize the address space (zone) for the pv_entries.  Set a
1128  * high water mark so that the system can recover from excessive
1129  * numbers of pv entries.
1130  */
1131 void
1132 pmap_init2(void)
1133 {
1134         int shpgperproc = PMAP_SHPGPERPROC;
1135         int entry_max;
1136
1137         TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1138         pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1139         TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1140         pv_entry_high_water = 9 * (pv_entry_max / 10);
1141
1142         /*
1143          * Subtract out pages already installed in the zone (hack)
1144          */
1145         entry_max = pv_entry_max - vm_page_array_size;
1146         if (entry_max <= 0)
1147                 entry_max = 1;
1148
1149         zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT);
1150
1151         /*
1152          * Enable dynamic deletion of empty higher-level page table pages
1153          * by default only if system memory is < 8GB (use 7GB for slop).
1154          * This can save a little memory, but imposes significant
1155          * performance overhead for things like bulk builds, and for programs
1156          * which do a lot of memory mapping and memory unmapping.
1157          */
1158         if (pmap_dynamic_delete < 0) {
1159                 if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1160                         pmap_dynamic_delete = 1;
1161                 else
1162                         pmap_dynamic_delete = 0;
1163         }
1164 }
1165
1166 /*
1167  * Typically used to initialize a fictitious page by vm/device_pager.c
1168  */
1169 void
1170 pmap_page_init(struct vm_page *m)
1171 {
1172         vm_page_init(m);
1173         TAILQ_INIT(&m->md.pv_list);
1174 }
1175
1176 /***************************************************
1177  * Low level helper routines.....
1178  ***************************************************/
1179
1180 /*
1181  * this routine defines the region(s) of memory that should
1182  * not be tested for the modified bit.
1183  */
1184 static __inline
1185 int
1186 pmap_track_modified(vm_pindex_t pindex)
1187 {
1188         vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1189         if ((va < clean_sva) || (va >= clean_eva))
1190                 return 1;
1191         else
1192                 return 0;
1193 }
1194
1195 /*
1196  * Extract the physical page address associated with the map/VA pair.
1197  * The page must be wired for this to work reliably.
1198  *
1199  * XXX for the moment we're using pv_find() instead of pv_get(), as
1200  *     callers might be expecting non-blocking operation.
1201  */
1202 vm_paddr_t
1203 pmap_extract(pmap_t pmap, vm_offset_t va)
1204 {
1205         vm_paddr_t rtval;
1206         pv_entry_t pt_pv;
1207         pt_entry_t *ptep;
1208
1209         rtval = 0;
1210         if (va >= VM_MAX_USER_ADDRESS) {
1211                 /*
1212                  * Kernel page directories might be direct-mapped and
1213                  * there is typically no PV tracking of pte's
1214                  */
1215                 pd_entry_t *pt;
1216
1217                 pt = pmap_pt(pmap, va);
1218                 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1219                         if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1220                                 rtval = *pt & PG_PS_FRAME;
1221                                 rtval |= va & PDRMASK;
1222                         } else {
1223                                 ptep = pmap_pt_to_pte(*pt, va);
1224                                 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1225                                         rtval = *ptep & PG_FRAME;
1226                                         rtval |= va & PAGE_MASK;
1227                                 }
1228                         }
1229                 }
1230         } else {
1231                 /*
1232                  * User pages currently do not direct-map the page directory
1233                  * and some pages might not used managed PVs.  But all PT's
1234                  * will have a PV.
1235                  */
1236                 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1237                 if (pt_pv) {
1238                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1239                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1240                                 rtval = *ptep & PG_FRAME;
1241                                 rtval |= va & PAGE_MASK;
1242                         }
1243                         pv_drop(pt_pv);
1244                 }
1245         }
1246         return rtval;
1247 }
1248
1249 /*
1250  * Similar to extract but checks protections, SMP-friendly short-cut for
1251  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1252  * fall-through to the real fault code.
1253  *
1254  * The returned page, if not NULL, is held (and not busied).
1255  */
1256 vm_page_t
1257 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1258 {
1259         if (pmap && va < VM_MAX_USER_ADDRESS) {
1260                 pv_entry_t pt_pv;
1261                 pv_entry_t pte_pv;
1262                 pt_entry_t *ptep;
1263                 pt_entry_t req;
1264                 vm_page_t m;
1265                 int error;
1266
1267                 req = pmap->pmap_bits[PG_V_IDX] |
1268                       pmap->pmap_bits[PG_U_IDX];
1269                 if (prot & VM_PROT_WRITE)
1270                         req |= pmap->pmap_bits[PG_RW_IDX];
1271
1272                 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1273                 if (pt_pv == NULL)
1274                         return (NULL);
1275                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1276                 if ((*ptep & req) != req) {
1277                         pv_drop(pt_pv);
1278                         return (NULL);
1279                 }
1280                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), &error);
1281                 if (pte_pv && error == 0) {
1282                         m = pte_pv->pv_m;
1283                         vm_page_hold(m);
1284                         if (prot & VM_PROT_WRITE)
1285                                 vm_page_dirty(m);
1286                         pv_put(pte_pv);
1287                 } else if (pte_pv) {
1288                         pv_drop(pte_pv);
1289                         m = NULL;
1290                 } else {
1291                         m = NULL;
1292                 }
1293                 pv_drop(pt_pv);
1294                 return(m);
1295         } else {
1296                 return(NULL);
1297         }
1298 }
1299
1300 /*
1301  * Extract the physical page address associated kernel virtual address.
1302  */
1303 vm_paddr_t
1304 pmap_kextract(vm_offset_t va)
1305 {
1306         pd_entry_t pt;          /* pt entry in pd */
1307         vm_paddr_t pa;
1308
1309         if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1310                 pa = DMAP_TO_PHYS(va);
1311         } else {
1312                 pt = *vtopt(va);
1313                 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1314                         pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1315                 } else {
1316                         /*
1317                          * Beware of a concurrent promotion that changes the
1318                          * PDE at this point!  For example, vtopte() must not
1319                          * be used to access the PTE because it would use the
1320                          * new PDE.  It is, however, safe to use the old PDE
1321                          * because the page table page is preserved by the
1322                          * promotion.
1323                          */
1324                         pa = *pmap_pt_to_pte(pt, va);
1325                         pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1326                 }
1327         }
1328         return pa;
1329 }
1330
1331 /***************************************************
1332  * Low level mapping routines.....
1333  ***************************************************/
1334
1335 /*
1336  * Routine: pmap_kenter
1337  * Function:
1338  *      Add a wired page to the KVA
1339  *      NOTE! note that in order for the mapping to take effect -- you
1340  *      should do an invltlb after doing the pmap_kenter().
1341  */
1342 void
1343 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1344 {
1345         pt_entry_t *ptep;
1346         pt_entry_t npte;
1347
1348         npte = pa |
1349             kernel_pmap.pmap_bits[PG_RW_IDX] |
1350             kernel_pmap.pmap_bits[PG_V_IDX];
1351 //          pgeflag;
1352         ptep = vtopte(va);
1353 #if 1
1354         pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1355 #else
1356         /* FUTURE */
1357         if (*ptep)
1358                 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1359         else
1360                 *ptep = npte;
1361 #endif
1362 }
1363
1364 /*
1365  * Similar to pmap_kenter(), except we only invalidate the mapping on the
1366  * current CPU.  Returns 0 if the previous pte was 0, 1 if it wasn't
1367  * (caller can conditionalize calling smp_invltlb()).
1368  */
1369 int
1370 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1371 {
1372         pt_entry_t *ptep;
1373         pt_entry_t npte;
1374         int res;
1375
1376         npte = pa |
1377             kernel_pmap.pmap_bits[PG_RW_IDX] |
1378             kernel_pmap.pmap_bits[PG_V_IDX];
1379 //          pgeflag;
1380         ptep = vtopte(va);
1381 #if 1
1382         res = 1;
1383 #else
1384         /* FUTURE */
1385         res = (*ptep != 0);
1386 #endif
1387         *ptep = npte;
1388         cpu_invlpg((void *)va);
1389
1390         return res;
1391 }
1392
1393 /*
1394  * Enter addresses into the kernel pmap but don't bother
1395  * doing any tlb invalidations.  Caller will do a rollup
1396  * invalidation via pmap_rollup_inval().
1397  */
1398 int
1399 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1400 {
1401         pt_entry_t *ptep;
1402         pt_entry_t npte;
1403         int res;
1404
1405         npte = pa |
1406             kernel_pmap.pmap_bits[PG_RW_IDX] |
1407             kernel_pmap.pmap_bits[PG_V_IDX];
1408 //          pgeflag;
1409         ptep = vtopte(va);
1410 #if 1
1411         res = 1;
1412 #else
1413         /* FUTURE */
1414         res = (*ptep != 0);
1415 #endif
1416         *ptep = npte;
1417         cpu_invlpg((void *)va);
1418
1419         return res;
1420 }
1421
1422 /*
1423  * remove a page from the kernel pagetables
1424  */
1425 void
1426 pmap_kremove(vm_offset_t va)
1427 {
1428         pt_entry_t *ptep;
1429
1430         ptep = vtopte(va);
1431         pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1432 }
1433
1434 void
1435 pmap_kremove_quick(vm_offset_t va)
1436 {
1437         pt_entry_t *ptep;
1438
1439         ptep = vtopte(va);
1440         (void)pte_load_clear(ptep);
1441         cpu_invlpg((void *)va);
1442 }
1443
1444 /*
1445  * Remove addresses from the kernel pmap but don't bother
1446  * doing any tlb invalidations.  Caller will do a rollup
1447  * invalidation via pmap_rollup_inval().
1448  */
1449 void
1450 pmap_kremove_noinval(vm_offset_t va)
1451 {
1452         pt_entry_t *ptep;
1453
1454         ptep = vtopte(va);
1455         (void)pte_load_clear(ptep);
1456 }
1457
1458 /*
1459  * XXX these need to be recoded.  They are not used in any critical path.
1460  */
1461 void
1462 pmap_kmodify_rw(vm_offset_t va)
1463 {
1464         atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1465         cpu_invlpg((void *)va);
1466 }
1467
1468 /* NOT USED
1469 void
1470 pmap_kmodify_nc(vm_offset_t va)
1471 {
1472         atomic_set_long(vtopte(va), PG_N);
1473         cpu_invlpg((void *)va);
1474 }
1475 */
1476
1477 /*
1478  * Used to map a range of physical addresses into kernel virtual
1479  * address space during the low level boot, typically to map the
1480  * dump bitmap, message buffer, and vm_page_array.
1481  *
1482  * These mappings are typically made at some pointer after the end of the
1483  * kernel text+data.
1484  *
1485  * We could return PHYS_TO_DMAP(start) here and not allocate any
1486  * via (*virtp), but then kmem from userland and kernel dumps won't
1487  * have access to the related pointers.
1488  */
1489 vm_offset_t
1490 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1491 {
1492         vm_offset_t va;
1493         vm_offset_t va_start;
1494
1495         /*return PHYS_TO_DMAP(start);*/
1496
1497         va_start = *virtp;
1498         va = va_start;
1499
1500         while (start < end) {
1501                 pmap_kenter_quick(va, start);
1502                 va += PAGE_SIZE;
1503                 start += PAGE_SIZE;
1504         }
1505         *virtp = va;
1506         return va_start;
1507 }
1508
1509 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
1510
1511 /*
1512  * Remove the specified set of pages from the data and instruction caches.
1513  *
1514  * In contrast to pmap_invalidate_cache_range(), this function does not
1515  * rely on the CPU's self-snoop feature, because it is intended for use
1516  * when moving pages into a different cache domain.
1517  */
1518 void
1519 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1520 {
1521         vm_offset_t daddr, eva;
1522         int i;
1523
1524         if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1525             (cpu_feature & CPUID_CLFSH) == 0)
1526                 wbinvd();
1527         else {
1528                 cpu_mfence();
1529                 for (i = 0; i < count; i++) {
1530                         daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1531                         eva = daddr + PAGE_SIZE;
1532                         for (; daddr < eva; daddr += cpu_clflush_line_size)
1533                                 clflush(daddr);
1534                 }
1535                 cpu_mfence();
1536         }
1537 }
1538
1539 void
1540 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1541 {
1542         KASSERT((sva & PAGE_MASK) == 0,
1543             ("pmap_invalidate_cache_range: sva not page-aligned"));
1544         KASSERT((eva & PAGE_MASK) == 0,
1545             ("pmap_invalidate_cache_range: eva not page-aligned"));
1546
1547         if (cpu_feature & CPUID_SS) {
1548                 ; /* If "Self Snoop" is supported, do nothing. */
1549         } else {
1550                 /* Globally invalidate caches */
1551                 cpu_wbinvd_on_all_cpus();
1552         }
1553 }
1554
1555 /*
1556  * Invalidate the specified range of virtual memory on all cpus associated
1557  * with the pmap.
1558  */
1559 void
1560 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1561 {
1562         pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1563 }
1564
1565 /*
1566  * Add a list of wired pages to the kva.  This routine is used for temporary
1567  * kernel mappings such as those found in buffer cache buffer.  Page
1568  * modifications and accesses are not tracked or recorded.
1569  *
1570  * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1571  *       semantics as previous mappings may have been zerod without any
1572  *       invalidation.
1573  *
1574  * The page *must* be wired.
1575  */
1576 void
1577 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1578 {
1579         vm_offset_t end_va;
1580         vm_offset_t va;
1581
1582         end_va = beg_va + count * PAGE_SIZE;
1583
1584         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1585                 pt_entry_t *pte;
1586
1587                 pte = vtopte(va);
1588                 *pte = VM_PAGE_TO_PHYS(*m) |
1589                     kernel_pmap.pmap_bits[PG_RW_IDX] |
1590                     kernel_pmap.pmap_bits[PG_V_IDX] |
1591                     kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1592 //              pgeflag;
1593                 m++;
1594         }
1595         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1596 }
1597
1598 /*
1599  * This routine jerks page mappings from the kernel -- it is meant only
1600  * for temporary mappings such as those found in buffer cache buffers.
1601  * No recording modified or access status occurs.
1602  *
1603  * MPSAFE, INTERRUPT SAFE (cluster callback)
1604  */
1605 void
1606 pmap_qremove(vm_offset_t beg_va, int count)
1607 {
1608         vm_offset_t end_va;
1609         vm_offset_t va;
1610
1611         end_va = beg_va + count * PAGE_SIZE;
1612
1613         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1614                 pt_entry_t *pte;
1615
1616                 pte = vtopte(va);
1617                 (void)pte_load_clear(pte);
1618                 cpu_invlpg((void *)va);
1619         }
1620         pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1621 }
1622
1623 /*
1624  * This routine removes temporary kernel mappings, only invalidating them
1625  * on the current cpu.  It should only be used under carefully controlled
1626  * conditions.
1627  */
1628 void
1629 pmap_qremove_quick(vm_offset_t beg_va, int count)
1630 {
1631         vm_offset_t end_va;
1632         vm_offset_t va;
1633
1634         end_va = beg_va + count * PAGE_SIZE;
1635
1636         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1637                 pt_entry_t *pte;
1638
1639                 pte = vtopte(va);
1640                 (void)pte_load_clear(pte);
1641                 cpu_invlpg((void *)va);
1642         }
1643 }
1644
1645 /*
1646  * This routine removes temporary kernel mappings *without* invalidating
1647  * the TLB.  It can only be used on permanent kva reservations such as those
1648  * found in buffer cache buffers, under carefully controlled circumstances.
1649  *
1650  * NOTE: Repopulating these KVAs requires unconditional invalidation.
1651  *       (pmap_qenter() does unconditional invalidation).
1652  */
1653 void
1654 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1655 {
1656         vm_offset_t end_va;
1657         vm_offset_t va;
1658
1659         end_va = beg_va + count * PAGE_SIZE;
1660
1661         for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1662                 pt_entry_t *pte;
1663
1664                 pte = vtopte(va);
1665                 (void)pte_load_clear(pte);
1666         }
1667 }
1668
1669 /*
1670  * Create a new thread and optionally associate it with a (new) process.
1671  * NOTE! the new thread's cpu may not equal the current cpu.
1672  */
1673 void
1674 pmap_init_thread(thread_t td)
1675 {
1676         /* enforce pcb placement & alignment */
1677         td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1678         td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1679         td->td_savefpu = &td->td_pcb->pcb_save;
1680         td->td_sp = (char *)td->td_pcb; /* no -16 */
1681 }
1682
1683 /*
1684  * This routine directly affects the fork perf for a process.
1685  */
1686 void
1687 pmap_init_proc(struct proc *p)
1688 {
1689 }
1690
1691 static void
1692 pmap_pinit_defaults(struct pmap *pmap)
1693 {
1694         bcopy(pmap_bits_default, pmap->pmap_bits,
1695               sizeof(pmap_bits_default));
1696         bcopy(protection_codes, pmap->protection_codes,
1697               sizeof(protection_codes));
1698         bcopy(pat_pte_index, pmap->pmap_cache_bits,
1699               sizeof(pat_pte_index));
1700         pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1701         pmap->copyinstr = std_copyinstr;
1702         pmap->copyin = std_copyin;
1703         pmap->copyout = std_copyout;
1704         pmap->fubyte = std_fubyte;
1705         pmap->subyte = std_subyte;
1706         pmap->fuword = std_fuword;
1707         pmap->suword = std_suword;
1708         pmap->suword32 = std_suword32;
1709 }
1710 /*
1711  * Initialize pmap0/vmspace0.
1712  *
1713  * On architectures where the kernel pmap is not integrated into the user
1714  * process pmap, this pmap represents the process pmap, not the kernel pmap.
1715  * kernel_pmap should be used to directly access the kernel_pmap.
1716  */
1717 void
1718 pmap_pinit0(struct pmap *pmap)
1719 {
1720         pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1721         pmap->pm_count = 1;
1722         CPUMASK_ASSZERO(pmap->pm_active);
1723         pmap->pm_pvhint = NULL;
1724         RB_INIT(&pmap->pm_pvroot);
1725         spin_init(&pmap->pm_spin, "pmapinit0");
1726         lwkt_token_init(&pmap->pm_token, "pmap_tok");
1727         bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1728         pmap_pinit_defaults(pmap);
1729 }
1730
1731 /*
1732  * Initialize a preallocated and zeroed pmap structure,
1733  * such as one in a vmspace structure.
1734  */
1735 static void
1736 pmap_pinit_simple(struct pmap *pmap)
1737 {
1738         /*
1739          * Misc initialization
1740          */
1741         pmap->pm_count = 1;
1742         CPUMASK_ASSZERO(pmap->pm_active);
1743         pmap->pm_pvhint = NULL;
1744         pmap->pm_flags = PMAP_FLAG_SIMPLE;
1745
1746         pmap_pinit_defaults(pmap);
1747
1748         /*
1749          * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1750          * for this).
1751          */
1752         if (pmap->pm_pmlpv == NULL) {
1753                 RB_INIT(&pmap->pm_pvroot);
1754                 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1755                 spin_init(&pmap->pm_spin, "pmapinitsimple");
1756                 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1757         }
1758 }
1759
1760 void
1761 pmap_pinit(struct pmap *pmap)
1762 {
1763         pv_entry_t pv;
1764         int j;
1765
1766         if (pmap->pm_pmlpv) {
1767                 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1768                         pmap_puninit(pmap);
1769                 }
1770         }
1771
1772         pmap_pinit_simple(pmap);
1773         pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1774
1775         /*
1776          * No need to allocate page table space yet but we do need a valid
1777          * page directory table.
1778          */
1779         if (pmap->pm_pml4 == NULL) {
1780                 pmap->pm_pml4 =
1781                     (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1782                                                         PAGE_SIZE,
1783                                                         VM_SUBSYS_PML4);
1784         }
1785
1786         /*
1787          * Allocate the page directory page, which wires it even though
1788          * it isn't being entered into some higher level page table (it
1789          * being the highest level).  If one is already cached we don't
1790          * have to do anything.
1791          */
1792         if ((pv = pmap->pm_pmlpv) == NULL) {
1793                 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1794                 pmap->pm_pmlpv = pv;
1795                 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1796                             VM_PAGE_TO_PHYS(pv->pv_m));
1797                 pv_put(pv);
1798
1799                 /*
1800                  * Install DMAP and KMAP.
1801                  */
1802                 for (j = 0; j < NDMPML4E; ++j) {
1803                         pmap->pm_pml4[DMPML4I + j] =
1804                             (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1805                             pmap->pmap_bits[PG_RW_IDX] |
1806                             pmap->pmap_bits[PG_V_IDX] |
1807                             pmap->pmap_bits[PG_U_IDX];
1808                 }
1809                 pmap->pm_pml4[KPML4I] = KPDPphys |
1810                     pmap->pmap_bits[PG_RW_IDX] |
1811                     pmap->pmap_bits[PG_V_IDX] |
1812                     pmap->pmap_bits[PG_U_IDX];
1813
1814                 /*
1815                  * install self-referential address mapping entry
1816                  */
1817                 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1818                     pmap->pmap_bits[PG_V_IDX] |
1819                     pmap->pmap_bits[PG_RW_IDX] |
1820                     pmap->pmap_bits[PG_A_IDX] |
1821                     pmap->pmap_bits[PG_M_IDX];
1822         } else {
1823                 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1824                 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1825         }
1826         KKASSERT(pmap->pm_pml4[255] == 0);
1827         KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1828         KKASSERT(pv->pv_entry.rbe_left == NULL);
1829         KKASSERT(pv->pv_entry.rbe_right == NULL);
1830 }
1831
1832 /*
1833  * Clean up a pmap structure so it can be physically freed.  This routine
1834  * is called by the vmspace dtor function.  A great deal of pmap data is
1835  * left passively mapped to improve vmspace management so we have a bit
1836  * of cleanup work to do here.
1837  */
1838 void
1839 pmap_puninit(pmap_t pmap)
1840 {
1841         pv_entry_t pv;
1842         vm_page_t p;
1843
1844         KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1845         if ((pv = pmap->pm_pmlpv) != NULL) {
1846                 if (pv_hold_try(pv) == 0)
1847                         pv_lock(pv);
1848                 KKASSERT(pv == pmap->pm_pmlpv);
1849                 p = pmap_remove_pv_page(pv);
1850                 pv_free(pv, NULL, 1);
1851                 pv = NULL;      /* safety */
1852                 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1853                 vm_page_busy_wait(p, FALSE, "pgpun");
1854                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1855                 vm_page_unwire(p, 0);
1856                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1857
1858                 /*
1859                  * XXX eventually clean out PML4 static entries and
1860                  * use vm_page_free_zero()
1861                  */
1862                 vm_page_free(p);
1863                 pmap->pm_pmlpv = NULL;
1864         }
1865         if (pmap->pm_pml4) {
1866                 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1867                 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1868                 pmap->pm_pml4 = NULL;
1869         }
1870         KKASSERT(pmap->pm_stats.resident_count == 0);
1871         KKASSERT(pmap->pm_stats.wired_count == 0);
1872 }
1873
1874 /*
1875  * This function is now unused (used to add the pmap to the pmap_list)
1876  */
1877 void
1878 pmap_pinit2(struct pmap *pmap)
1879 {
1880 }
1881
1882 /*
1883  * This routine is called when various levels in the page table need to
1884  * be populated.  This routine cannot fail.
1885  *
1886  * This function returns two locked pv_entry's, one representing the
1887  * requested pv and one representing the requested pv's parent pv.  If
1888  * an intermediate page table does not exist it will be created, mapped,
1889  * wired, and the parent page table will be given an additional hold
1890  * count representing the presence of the child pv_entry.
1891  */
1892 static
1893 pv_entry_t
1894 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1895 {
1896         pt_entry_t *ptep;
1897         pv_entry_t pv;
1898         pv_entry_t pvp;
1899         vm_pindex_t pt_pindex;
1900         vm_page_t m;
1901         int isnew;
1902         int ispt;
1903
1904         /*
1905          * If the pv already exists and we aren't being asked for the
1906          * parent page table page we can just return it.  A locked+held pv
1907          * is returned.  The pv will also have a second hold related to the
1908          * pmap association that we don't have to worry about.
1909          */
1910         ispt = 0;
1911         pv = pv_alloc(pmap, ptepindex, &isnew);
1912         if (isnew == 0 && pvpp == NULL)
1913                 return(pv);
1914
1915         /*
1916          * Special case terminal PVs.  These are not page table pages so
1917          * no vm_page is allocated (the caller supplied the vm_page).  If
1918          * pvpp is non-NULL we are being asked to also removed the pt_pv
1919          * for this pv.
1920          *
1921          * Note that pt_pv's are only returned for user VAs. We assert that
1922          * a pt_pv is not being requested for kernel VAs.  The kernel
1923          * pre-wires all higher-level page tables so don't overload managed
1924          * higher-level page tables on top of it!
1925          */
1926         if (ptepindex < pmap_pt_pindex(0)) {
1927                 if (ptepindex >= NUPTE_USER) {
1928                         /* kernel manages this manually for KVM */
1929                         KKASSERT(pvpp == NULL);
1930                 } else {
1931                         KKASSERT(pvpp != NULL);
1932                         pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1933                         pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1934                         if (isnew) {
1935                                 vm_page_wire_quick(pvp->pv_m);
1936                                 if (pvpp)
1937                                         *pvpp = pvp;
1938                                 else
1939                                         pv_put(pvp);
1940                         } else {
1941                                 *pvpp = pvp;
1942                         }
1943                 }
1944                 return(pv);
1945         }
1946
1947         /*
1948          * The kernel never uses managed PT/PD/PDP pages.
1949          */
1950         KKASSERT(pmap != &kernel_pmap);
1951
1952         /*
1953          * Non-terminal PVs allocate a VM page to represent the page table,
1954          * so we have to resolve pvp and calculate ptepindex for the pvp
1955          * and then for the page table entry index in the pvp for
1956          * fall-through.
1957          */
1958         if (ptepindex < pmap_pd_pindex(0)) {
1959                 /*
1960                  * pv is PT, pvp is PD
1961                  */
1962                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1963                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1964                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1965                 if (!isnew)
1966                         goto notnew;
1967
1968                 /*
1969                  * PT index in PD
1970                  */
1971                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1972                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1973                 ispt = 1;
1974         } else if (ptepindex < pmap_pdp_pindex(0)) {
1975                 /*
1976                  * pv is PD, pvp is PDP
1977                  *
1978                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1979                  *                   the PD.
1980                  */
1981                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1982                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1983
1984                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1985                         KKASSERT(pvpp == NULL);
1986                         pvp = NULL;
1987                 } else {
1988                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
1989                 }
1990                 if (!isnew)
1991                         goto notnew;
1992
1993                 /*
1994                  * PD index in PDP
1995                  */
1996                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1997                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1998         } else if (ptepindex < pmap_pml4_pindex()) {
1999                 /*
2000                  * pv is PDP, pvp is the root pml4 table
2001                  */
2002                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2003                 if (!isnew)
2004                         goto notnew;
2005
2006                 /*
2007                  * PDP index in PML4
2008                  */
2009                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2010                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2011         } else {
2012                 /*
2013                  * pv represents the top-level PML4, there is no parent.
2014                  */
2015                 pvp = NULL;
2016                 if (!isnew)
2017                         goto notnew;
2018         }
2019
2020         /*
2021          * (isnew) is TRUE, pv is not terminal.
2022          *
2023          * (1) Add a wire count to the parent page table (pvp).
2024          * (2) Allocate a VM page for the page table.
2025          * (3) Enter the VM page into the parent page table.
2026          *
2027          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2028          */
2029         if (pvp)
2030                 vm_page_wire_quick(pvp->pv_m);
2031
2032         for (;;) {
2033                 m = vm_page_alloc(NULL, pv->pv_pindex,
2034                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2035                                   VM_ALLOC_INTERRUPT);
2036                 if (m)
2037                         break;
2038                 vm_wait(0);
2039         }
2040         vm_page_spin_lock(m);
2041         pmap_page_stats_adding(m);
2042         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2043         pv->pv_m = m;
2044         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2045         vm_page_spin_unlock(m);
2046         vm_page_unmanage(m);    /* m must be spinunlocked */
2047
2048         pmap_zero_page(VM_PAGE_TO_PHYS(m));
2049         m->valid = VM_PAGE_BITS_ALL;
2050         vm_page_wire(m);        /* wire for mapping in parent */
2051
2052         /*
2053          * Wire the page into pvp.  Bump the resident_count for the pmap.
2054          * There is no pvp for the top level, address the pm_pml4[] array
2055          * directly.
2056          *
2057          * If the caller wants the parent we return it, otherwise
2058          * we just put it away.
2059          *
2060          * No interlock is needed for pte 0 -> non-zero.
2061          *
2062          * In the situation where *ptep is valid we might have an unmanaged
2063          * page table page shared from another page table which we need to
2064          * unshare before installing our private page table page.
2065          */
2066         if (pvp) {
2067                 ptep = pv_pte_lookup(pvp, ptepindex);
2068                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2069                         pt_entry_t pte;
2070
2071                         if (ispt == 0) {
2072                                 panic("pmap_allocpte: unexpected pte %p/%d",
2073                                       pvp, (int)ptepindex);
2074                         }
2075                         pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
2076                         if (vm_page_unwire_quick(
2077                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2078                                 panic("pmap_allocpte: shared pgtable "
2079                                       "pg bad wirecount");
2080                         }
2081                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
2082                 }
2083                 *ptep = VM_PAGE_TO_PHYS(m) |
2084                         (pmap->pmap_bits[PG_U_IDX] |
2085                          pmap->pmap_bits[PG_RW_IDX] |
2086                          pmap->pmap_bits[PG_V_IDX] |
2087                          pmap->pmap_bits[PG_A_IDX] |
2088                          pmap->pmap_bits[PG_M_IDX]);
2089         }
2090         vm_page_wakeup(m);
2091 notnew:
2092         if (pvpp)
2093                 *pvpp = pvp;
2094         else if (pvp)
2095                 pv_put(pvp);
2096         return (pv);
2097 }
2098
2099 /*
2100  * This version of pmap_allocpte() checks for possible segment optimizations
2101  * that would allow page-table sharing.  It can be called for terminal
2102  * page or page table page ptepindex's.
2103  *
2104  * The function is called with page table page ptepindex's for fictitious
2105  * and unmanaged terminal pages.  That is, we don't want to allocate a
2106  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
2107  * for this case.
2108  *
2109  * This function can return a pv and *pvpp associated with the passed in pmap
2110  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
2111  * an unmanaged page table page will be entered into the pass in pmap.
2112  */
2113 static
2114 pv_entry_t
2115 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2116                   vm_map_entry_t entry, vm_offset_t va)
2117 {
2118         vm_object_t object;
2119         pmap_t obpmap;
2120         pmap_t *obpmapp;
2121         vm_offset_t b;
2122         pv_entry_t pte_pv;      /* in original or shared pmap */
2123         pv_entry_t pt_pv;       /* in original or shared pmap */
2124         pv_entry_t proc_pd_pv;  /* in original pmap */
2125         pv_entry_t proc_pt_pv;  /* in original pmap */
2126         pv_entry_t xpv;         /* PT in shared pmap */
2127         pd_entry_t *pt;         /* PT entry in PD of original pmap */
2128         pd_entry_t opte;        /* contents of *pt */
2129         pd_entry_t npte;        /* contents of *pt */
2130         vm_page_t m;
2131
2132 retry:
2133         /*
2134          * Basic tests, require a non-NULL vm_map_entry, require proper
2135          * alignment and type for the vm_map_entry, require that the
2136          * underlying object already be allocated.
2137          *
2138          * We allow almost any type of object to use this optimization.
2139          * The object itself does NOT have to be sized to a multiple of the
2140          * segment size, but the memory mapping does.
2141          *
2142          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2143          *     won't work as expected.
2144          */
2145         if (entry == NULL ||
2146             pmap_mmu_optimize == 0 ||                   /* not enabled */
2147             (pmap->pm_flags & PMAP_HVM) ||              /* special pmap */
2148             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal or pt */
2149             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2150             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2151             entry->object.vm_object == NULL ||          /* needs VM object */
2152             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2153             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2154             (entry->offset & SEG_MASK) ||               /* must be aligned */
2155             (entry->start & SEG_MASK)) {
2156                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2157         }
2158
2159         /*
2160          * Make sure the full segment can be represented.
2161          */
2162         b = va & ~(vm_offset_t)SEG_MASK;
2163         if (b < entry->start || b + SEG_SIZE > entry->end)
2164                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2165
2166         /*
2167          * If the full segment can be represented dive the VM object's
2168          * shared pmap, allocating as required.
2169          */
2170         object = entry->object.vm_object;
2171
2172         if (entry->protection & VM_PROT_WRITE)
2173                 obpmapp = &object->md.pmap_rw;
2174         else
2175                 obpmapp = &object->md.pmap_ro;
2176
2177 #ifdef PMAP_DEBUG2
2178         if (pmap_enter_debug > 0) {
2179                 --pmap_enter_debug;
2180                 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2181                         "obpmapp %p %p\n",
2182                         va, entry->protection, object,
2183                         obpmapp, *obpmapp);
2184                 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2185                         entry, entry->start, entry->end);
2186         }
2187 #endif
2188
2189         /*
2190          * We allocate what appears to be a normal pmap but because portions
2191          * of this pmap are shared with other unrelated pmaps we have to
2192          * set pm_active to point to all cpus.
2193          *
2194          * XXX Currently using pmap_spin to interlock the update, can't use
2195          *     vm_object_hold/drop because the token might already be held
2196          *     shared OR exclusive and we don't know.
2197          */
2198         while ((obpmap = *obpmapp) == NULL) {
2199                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2200                 pmap_pinit_simple(obpmap);
2201                 pmap_pinit2(obpmap);
2202                 spin_lock(&pmap_spin);
2203                 if (*obpmapp != NULL) {
2204                         /*
2205                          * Handle race
2206                          */
2207                         spin_unlock(&pmap_spin);
2208                         pmap_release(obpmap);
2209                         pmap_puninit(obpmap);
2210                         kfree(obpmap, M_OBJPMAP);
2211                         obpmap = *obpmapp; /* safety */
2212                 } else {
2213                         obpmap->pm_active = smp_active_mask;
2214                         obpmap->pm_flags |= PMAP_SEGSHARED;
2215                         *obpmapp = obpmap;
2216                         spin_unlock(&pmap_spin);
2217                 }
2218         }
2219
2220         /*
2221          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2222          * pte/pt using the shared pmap from the object but also adjust
2223          * the process pmap's page table page as a side effect.
2224          */
2225
2226         /*
2227          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2228          * we will return.  This is true if ptepindex represents a terminal
2229          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2230          * the PD.
2231          */
2232         pt_pv = NULL;
2233         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2234         if (ptepindex >= pmap_pt_pindex(0))
2235                 xpv = pte_pv;
2236         else
2237                 xpv = pt_pv;
2238
2239         /*
2240          * Resolve the PD in the process pmap so we can properly share the
2241          * page table page.  Lock order is bottom-up (leaf first)!
2242          *
2243          * NOTE: proc_pt_pv can be NULL.
2244          */
2245         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2246         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2247 #ifdef PMAP_DEBUG2
2248         if (pmap_enter_debug > 0) {
2249                 --pmap_enter_debug;
2250                 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2251                         proc_pt_pv,
2252                         (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2253                         proc_pd_pv,
2254                         va);
2255         }
2256 #endif
2257
2258         /*
2259          * xpv is the page table page pv from the shared object
2260          * (for convenience), from above.
2261          *
2262          * Calculate the pte value for the PT to load into the process PD.
2263          * If we have to change it we must properly dispose of the previous
2264          * entry.
2265          */
2266         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2267         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2268                (pmap->pmap_bits[PG_U_IDX] |
2269                 pmap->pmap_bits[PG_RW_IDX] |
2270                 pmap->pmap_bits[PG_V_IDX] |
2271                 pmap->pmap_bits[PG_A_IDX] |
2272                 pmap->pmap_bits[PG_M_IDX]);
2273
2274         /*
2275          * Dispose of previous page table page if it was local to the
2276          * process pmap.  If the old pt is not empty we cannot dispose of it
2277          * until we clean it out.  This case should not arise very often so
2278          * it is not optimized.
2279          */
2280         if (proc_pt_pv) {
2281                 pmap_inval_bulk_t bulk;
2282
2283                 if (proc_pt_pv->pv_m->wire_count != 1) {
2284                         pv_put(proc_pd_pv);
2285                         pv_put(proc_pt_pv);
2286                         pv_put(pt_pv);
2287                         pv_put(pte_pv);
2288                         pmap_remove(pmap,
2289                                     va & ~(vm_offset_t)SEG_MASK,
2290                                     (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2291                         goto retry;
2292                 }
2293
2294                 /*
2295                  * The release call will indirectly clean out *pt
2296                  */
2297                 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2298                 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2299                 pmap_inval_bulk_flush(&bulk);
2300                 proc_pt_pv = NULL;
2301                 /* relookup */
2302                 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2303         }
2304
2305         /*
2306          * Handle remaining cases.
2307          */
2308         if (*pt == 0) {
2309                 *pt = npte;
2310                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2311                 vm_page_wire_quick(proc_pd_pv->pv_m);   /* proc pd for sh pt */
2312                 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2313         } else if (*pt != npte) {
2314                 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2315
2316 #if 0
2317                 opte = pte_load_clear(pt);
2318                 KKASSERT(opte && opte != npte);
2319
2320                 *pt = npte;
2321 #endif
2322                 vm_page_wire_quick(xpv->pv_m);          /* shared pt -> proc */
2323
2324                 /*
2325                  * Clean up opte, bump the wire_count for the process
2326                  * PD page representing the new entry if it was
2327                  * previously empty.
2328                  *
2329                  * If the entry was not previously empty and we have
2330                  * a PT in the proc pmap then opte must match that
2331                  * pt.  The proc pt must be retired (this is done
2332                  * later on in this procedure).
2333                  *
2334                  * NOTE: replacing valid pte, wire_count on proc_pd_pv
2335                  * stays the same.
2336                  */
2337                 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2338                 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2339                 if (vm_page_unwire_quick(m)) {
2340                         panic("pmap_allocpte_seg: "
2341                               "bad wire count %p",
2342                               m);
2343                 }
2344         }
2345
2346         /*
2347          * The existing process page table was replaced and must be destroyed
2348          * here.
2349          */
2350         if (proc_pd_pv)
2351                 pv_put(proc_pd_pv);
2352         if (pvpp)
2353                 *pvpp = pt_pv;
2354         else
2355                 pv_put(pt_pv);
2356
2357         return (pte_pv);
2358 }
2359
2360 /*
2361  * Release any resources held by the given physical map.
2362  *
2363  * Called when a pmap initialized by pmap_pinit is being released.  Should
2364  * only be called if the map contains no valid mappings.
2365  *
2366  * Caller must hold pmap->pm_token
2367  */
2368 struct pmap_release_info {
2369         pmap_t  pmap;
2370         int     retry;
2371         pv_entry_t pvp;
2372 };
2373
2374 static int pmap_release_callback(pv_entry_t pv, void *data);
2375
2376 void
2377 pmap_release(struct pmap *pmap)
2378 {
2379         struct pmap_release_info info;
2380
2381         KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2382                 ("pmap still active! %016jx",
2383                 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2384
2385         /*
2386          * There is no longer a pmap_list, if there were we would remove the
2387          * pmap from it here.
2388          */
2389
2390         /*
2391          * Pull pv's off the RB tree in order from low to high and release
2392          * each page.
2393          */
2394         info.pmap = pmap;
2395         do {
2396                 info.retry = 0;
2397                 info.pvp = NULL;
2398
2399                 spin_lock(&pmap->pm_spin);
2400                 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2401                         pmap_release_callback, &info);
2402                 spin_unlock(&pmap->pm_spin);
2403
2404                 if (info.pvp)
2405                         pv_put(info.pvp);
2406         } while (info.retry);
2407
2408
2409         /*
2410          * One resident page (the pml4 page) should remain.
2411          * No wired pages should remain.
2412          */
2413         KKASSERT(pmap->pm_stats.resident_count ==
2414                  ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2415
2416         KKASSERT(pmap->pm_stats.wired_count == 0);
2417 }
2418
2419 /*
2420  * Called from low to high.  We must cache the proper parent pv so we
2421  * can adjust its wired count.
2422  */
2423 static int
2424 pmap_release_callback(pv_entry_t pv, void *data)
2425 {
2426         struct pmap_release_info *info = data;
2427         pmap_t pmap = info->pmap;
2428         vm_pindex_t pindex;
2429         int r;
2430
2431         if (info->pvp == pv) {
2432                 spin_unlock(&pmap->pm_spin);
2433                 info->pvp = NULL;
2434         } else if (pv_hold_try(pv)) {
2435                 spin_unlock(&pmap->pm_spin);
2436         } else {
2437                 spin_unlock(&pmap->pm_spin);
2438                 pv_lock(pv);
2439         }
2440         if (pv->pv_pmap != pmap) {
2441                 pv_put(pv);
2442                 spin_lock(&pmap->pm_spin);
2443                 info->retry = 1;
2444                 return(-1);
2445         }
2446
2447         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2448                 /*
2449                  * parent is PT
2450                  */
2451                 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2452                 pindex += NUPTE_TOTAL;
2453         } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2454                 /*
2455                  * parent is PD
2456                  */
2457                 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2458                 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2459         } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2460                 /*
2461                  * parent is PDP
2462                  */
2463                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2464                          NPDPEPGSHIFT;
2465                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2466         } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2467                 /*
2468                  * parent is PML4 (there's only one)
2469                  */
2470 #if 0
2471                 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2472                            NUPD_TOTAL) >> NPML4EPGSHIFT;
2473                 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2474 #endif
2475                 pindex = pmap_pml4_pindex();
2476         } else {
2477                 /*
2478                  * parent is NULL
2479                  */
2480                 if (info->pvp) {
2481                         pv_put(info->pvp);
2482                         info->pvp = NULL;
2483                 }
2484                 pindex = 0;
2485         }
2486         if (pindex) {
2487                 if (info->pvp && info->pvp->pv_pindex != pindex) {
2488                         pv_put(info->pvp);
2489                         info->pvp = NULL;
2490                 }
2491                 if (info->pvp == NULL)
2492                         info->pvp = pv_get(pmap, pindex);
2493         } else {
2494                 if (info->pvp) {
2495                         pv_put(info->pvp);
2496                         info->pvp = NULL;
2497                 }
2498         }
2499         r = pmap_release_pv(pv, info->pvp, NULL);
2500         spin_lock(&pmap->pm_spin);
2501         return(r);
2502 }
2503
2504 /*
2505  * Called with held (i.e. also locked) pv.  This function will dispose of
2506  * the lock along with the pv.
2507  *
2508  * If the caller already holds the locked parent page table for pv it
2509  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2510  * pass NULL for pvp.
2511  */
2512 static int
2513 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2514 {
2515         vm_page_t p;
2516
2517         /*
2518          * The pmap is currently not spinlocked, pv is held+locked.
2519          * Remove the pv's page from its parent's page table.  The
2520          * parent's page table page's wire_count will be decremented.
2521          *
2522          * This will clean out the pte at any level of the page table.
2523          * If smp != 0 all cpus are affected.
2524          *
2525          * Do not tear-down recursively, its faster to just let the
2526          * release run its course.
2527          */
2528         pmap_remove_pv_pte(pv, pvp, bulk, 0);
2529
2530         /*
2531          * Terminal pvs are unhooked from their vm_pages.  Because
2532          * terminal pages aren't page table pages they aren't wired
2533          * by us, so we have to be sure not to unwire them either.
2534          */
2535         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2536                 pmap_remove_pv_page(pv);
2537                 goto skip;
2538         }
2539
2540         /*
2541          * We leave the top-level page table page cached, wired, and
2542          * mapped in the pmap until the dtor function (pmap_puninit())
2543          * gets called.
2544          *
2545          * Since we are leaving the top-level pv intact we need
2546          * to break out of what would otherwise be an infinite loop.
2547          */
2548         if (pv->pv_pindex == pmap_pml4_pindex()) {
2549                 pv_put(pv);
2550                 return(-1);
2551         }
2552
2553         /*
2554          * For page table pages (other than the top-level page),
2555          * remove and free the vm_page.  The representitive mapping
2556          * removed above by pmap_remove_pv_pte() did not undo the
2557          * last wire_count so we have to do that as well.
2558          */
2559         p = pmap_remove_pv_page(pv);
2560         vm_page_busy_wait(p, FALSE, "pmaprl");
2561         if (p->wire_count != 1) {
2562                 kprintf("p->wire_count was %016lx %d\n",
2563                         pv->pv_pindex, p->wire_count);
2564         }
2565         KKASSERT(p->wire_count == 1);
2566         KKASSERT(p->flags & PG_UNMANAGED);
2567
2568         vm_page_unwire(p, 0);
2569         KKASSERT(p->wire_count == 0);
2570
2571         vm_page_free(p);
2572 skip:
2573         pv_free(pv, pvp, 1);
2574
2575         return 0;
2576 }
2577
2578 /*
2579  * This function will remove the pte associated with a pv from its parent.
2580  * Terminal pv's are supported.  All cpus specified by (bulk) are properly
2581  * invalidated.
2582  *
2583  * The wire count will be dropped on the parent page table.  The wire
2584  * count on the page being removed (pv->pv_m) from the parent page table
2585  * is NOT touched.  Note that terminal pages will not have any additional
2586  * wire counts while page table pages will have at least one representing
2587  * the mapping, plus others representing sub-mappings.
2588  *
2589  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2590  *       pages and user page table and terminal pages.
2591  *
2592  * The pv must be locked.  The pvp, if supplied, must be locked.  All
2593  * supplied pv's will remain locked on return.
2594  *
2595  * XXX must lock parent pv's if they exist to remove pte XXX
2596  */
2597 static
2598 void
2599 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2600                    int destroy)
2601 {
2602         vm_pindex_t ptepindex = pv->pv_pindex;
2603         pmap_t pmap = pv->pv_pmap;
2604         vm_page_t p;
2605         int gotpvp = 0;
2606
2607         KKASSERT(pmap);
2608
2609         if (ptepindex == pmap_pml4_pindex()) {
2610                 /*
2611                  * We are the top level pml4 table, there is no parent.
2612                  */
2613                 p = pmap->pm_pmlpv->pv_m;
2614         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2615                 /*
2616                  * Remove a PDP page from the pml4e.  This can only occur
2617                  * with user page tables.  We do not have to lock the
2618                  * pml4 PV so just ignore pvp.
2619                  */
2620                 vm_pindex_t pml4_pindex;
2621                 vm_pindex_t pdp_index;
2622                 pml4_entry_t *pdp;
2623
2624                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2625                 if (pvp == NULL) {
2626                         pml4_pindex = pmap_pml4_pindex();
2627                         pvp = pv_get(pv->pv_pmap, pml4_pindex);
2628                         KKASSERT(pvp);
2629                         gotpvp = 1;
2630                 }
2631                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2632                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2633                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2634                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2635         } else if (ptepindex >= pmap_pd_pindex(0)) {
2636                 /*
2637                  * Remove a PD page from the pdp
2638                  *
2639                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2640                  *                   of a simple pmap because it stops at
2641                  *                   the PD page.
2642                  */
2643                 vm_pindex_t pdp_pindex;
2644                 vm_pindex_t pd_index;
2645                 pdp_entry_t *pd;
2646
2647                 pd_index = ptepindex - pmap_pd_pindex(0);
2648
2649                 if (pvp == NULL) {
2650                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2651                                      (pd_index >> NPML4EPGSHIFT);
2652                         pvp = pv_get(pv->pv_pmap, pdp_pindex);
2653                         gotpvp = 1;
2654                 }
2655                 if (pvp) {
2656                         pd = pv_pte_lookup(pvp, pd_index &
2657                                                 ((1ul << NPDPEPGSHIFT) - 1));
2658                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2659                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2660                         pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2661                 } else {
2662                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2663                         p = pv->pv_m;           /* degenerate test later */
2664                 }
2665         } else if (ptepindex >= pmap_pt_pindex(0)) {
2666                 /*
2667                  *  Remove a PT page from the pd
2668                  */
2669                 vm_pindex_t pd_pindex;
2670                 vm_pindex_t pt_index;
2671                 pd_entry_t *pt;
2672
2673                 pt_index = ptepindex - pmap_pt_pindex(0);
2674
2675                 if (pvp == NULL) {
2676                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2677                                     (pt_index >> NPDPEPGSHIFT);
2678                         pvp = pv_get(pv->pv_pmap, pd_pindex);
2679                         KKASSERT(pvp);
2680                         gotpvp = 1;
2681                 }
2682                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2683                 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2684                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2685                 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2686         } else {
2687                 /*
2688                  * Remove a managed PTE from the PT page.  Userland pmaps
2689                  * manage PT/PD/PDP page tables pages but the kernel_pmap
2690                  * does not.
2691                  *
2692                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2693                  *       pv is a pte_pv so we can safely lock pt_pv.
2694                  *
2695                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2696                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2697                  *       terminal ptes.
2698                  */
2699                 vm_pindex_t pt_pindex;
2700                 pt_entry_t *ptep;
2701                 pt_entry_t pte;
2702                 vm_offset_t va;
2703
2704                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2705                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2706
2707                 if (ptepindex >= NUPTE_USER) {
2708                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2709                         KKASSERT(pvp == NULL);
2710                         /* pvp remains NULL */
2711                 } else {
2712                         if (pvp == NULL) {
2713                                 pt_pindex = NUPTE_TOTAL +
2714                                             (ptepindex >> NPDPEPGSHIFT);
2715                                 pvp = pv_get(pv->pv_pmap, pt_pindex);
2716                                 KKASSERT(pvp);
2717                                 gotpvp = 1;
2718                         }
2719                         ptep = pv_pte_lookup(pvp, ptepindex &
2720                                                   ((1ul << NPDPEPGSHIFT) - 1));
2721                 }
2722                 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2723                 if (bulk == NULL)               /* XXX */
2724                         cpu_invlpg((void *)va); /* XXX */
2725
2726                 /*
2727                  * Now update the vm_page_t
2728                  */
2729                 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] |
2730                             pmap->pmap_bits[PG_V_IDX])) !=
2731                     (pmap->pmap_bits[PG_MANAGED_IDX] |
2732                      pmap->pmap_bits[PG_V_IDX])) {
2733                         kprintf("remove_pte badpte %016lx %016lx %d\n",
2734                                 pte, pv->pv_pindex,
2735                                 pv->pv_pindex < pmap_pt_pindex(0));
2736                 }
2737
2738                 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2739                 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2740                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2741                         p = pv->pv_m;
2742                 else
2743                         p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2744                 /* p = pv->pv_m; */
2745
2746                 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2747                         if (pmap_track_modified(ptepindex))
2748                                 vm_page_dirty(p);
2749                 }
2750                 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2751                         vm_page_flag_set(p, PG_REFERENCED);
2752                 }
2753                 if (pte & pmap->pmap_bits[PG_W_IDX])
2754                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
2755                 if (pte & pmap->pmap_bits[PG_G_IDX])
2756                         cpu_invlpg((void *)va);
2757         }
2758         KKASSERT(pv->pv_m == p);        /* XXX remove me later */
2759
2760         /*
2761          * If requested, scrap the underlying pv->pv_m and the underlying
2762          * pv.  If this is a page-table-page we must also free the page.
2763          *
2764          * pvp must be returned locked.
2765          */
2766         if (destroy == 1) {
2767                 /*
2768                  * page table page (PT, PD, PDP, PML4), caller was responsible
2769                  * for testing wired_count.
2770                  */
2771                 vm_page_t p;
2772
2773                 KKASSERT(pv->pv_m->wire_count == 1);
2774                 p = pmap_remove_pv_page(pv);
2775                 pv_free(pv, pvp, 1);
2776                 pv = NULL;
2777
2778                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
2779                 vm_page_busy_wait(p, FALSE, "pgpun");
2780                 vm_page_unwire(p, 0);
2781                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2782                 vm_page_free(p);
2783         } else if (destroy == 2) {
2784                 /*
2785                  * Normal page, remove from pmap and leave the underlying
2786                  * page untouched.
2787                  */
2788                 pmap_remove_pv_page(pv);
2789                 pv_free(pv, pvp, 1);
2790                 pv = NULL;              /* safety */
2791         }
2792
2793         /*
2794          * If we acquired pvp ourselves then we are responsible for
2795          * recursively deleting it.
2796          */
2797         if (pvp && gotpvp) {
2798                 /*
2799                  * Recursively destroy higher-level page tables.
2800                  *
2801                  * This is optional.  If we do not, they will still
2802                  * be destroyed when the process exits.
2803                  *
2804                  * NOTE: Do not destroy pv_entry's with extra hold refs,
2805                  *       a caller may have unlocked it and intends to
2806                  *       continue to use it.
2807                  */
2808                 if (pmap_dynamic_delete &&
2809                     pvp->pv_m &&
2810                     pvp->pv_m->wire_count == 1 &&
2811                     (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
2812                     pvp->pv_pindex != pmap_pml4_pindex()) {
2813                         if (pmap_dynamic_delete == 2)
2814                                 kprintf("A %jd %08x\n", pvp->pv_pindex, pvp->pv_hold);
2815                         if (pmap != &kernel_pmap) {
2816                                 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
2817                                 pvp = NULL;     /* safety */
2818                         } else {
2819                                 kprintf("Attempt to remove kernel_pmap pindex "
2820                                         "%jd\n", pvp->pv_pindex);
2821                                 pv_put(pvp);
2822                         }
2823                 } else {
2824                         pv_put(pvp);
2825                 }
2826         }
2827 }
2828
2829 /*
2830  * Remove the vm_page association to a pv.  The pv must be locked.
2831  */
2832 static
2833 vm_page_t
2834 pmap_remove_pv_page(pv_entry_t pv)
2835 {
2836         vm_page_t m;
2837
2838         m = pv->pv_m;
2839         KKASSERT(m);
2840         vm_page_spin_lock(m);
2841         pv->pv_m = NULL;
2842         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2843         pmap_page_stats_deleting(m);
2844         /*
2845         if (m->object)
2846                 atomic_add_int(&m->object->agg_pv_list_count, -1);
2847         */
2848         if (TAILQ_EMPTY(&m->md.pv_list))
2849                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2850         vm_page_spin_unlock(m);
2851
2852         return(m);
2853 }
2854
2855 /*
2856  * Grow the number of kernel page table entries, if needed.
2857  *
2858  * This routine is always called to validate any address space
2859  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
2860  * space below KERNBASE.
2861  */
2862 void
2863 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2864 {
2865         vm_paddr_t paddr;
2866         vm_offset_t ptppaddr;
2867         vm_page_t nkpg;
2868         pd_entry_t *pt, newpt;
2869         pdp_entry_t newpd;
2870         int update_kernel_vm_end;
2871
2872         /*
2873          * bootstrap kernel_vm_end on first real VM use
2874          */
2875         if (kernel_vm_end == 0) {
2876                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2877                 nkpt = 0;
2878                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2879                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2880                                         ~(PAGE_SIZE * NPTEPG - 1);
2881                         nkpt++;
2882                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2883                                 kernel_vm_end = kernel_map.max_offset;
2884                                 break;
2885                         }
2886                 }
2887         }
2888
2889         /*
2890          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
2891          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
2892          * do not want to force-fill 128G worth of page tables.
2893          */
2894         if (kstart < KERNBASE) {
2895                 if (kstart > kernel_vm_end)
2896                         kstart = kernel_vm_end;
2897                 KKASSERT(kend <= KERNBASE);
2898                 update_kernel_vm_end = 1;
2899         } else {
2900                 update_kernel_vm_end = 0;
2901         }
2902
2903         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2904         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2905
2906         if (kend - 1 >= kernel_map.max_offset)
2907                 kend = kernel_map.max_offset;
2908
2909         while (kstart < kend) {
2910                 pt = pmap_pt(&kernel_pmap, kstart);
2911                 if (pt == NULL) {
2912                         /* We need a new PD entry */
2913                         nkpg = vm_page_alloc(NULL, nkpt,
2914                                              VM_ALLOC_NORMAL |
2915                                              VM_ALLOC_SYSTEM |
2916                                              VM_ALLOC_INTERRUPT);
2917                         if (nkpg == NULL) {
2918                                 panic("pmap_growkernel: no memory to grow "
2919                                       "kernel");
2920                         }
2921                         paddr = VM_PAGE_TO_PHYS(nkpg);
2922                         pmap_zero_page(paddr);
2923                         newpd = (pdp_entry_t)
2924                             (paddr |
2925                             kernel_pmap.pmap_bits[PG_V_IDX] |
2926                             kernel_pmap.pmap_bits[PG_RW_IDX] |
2927                             kernel_pmap.pmap_bits[PG_A_IDX] |
2928                             kernel_pmap.pmap_bits[PG_M_IDX]);
2929                         *pmap_pd(&kernel_pmap, kstart) = newpd;
2930                         nkpt++;
2931                         continue; /* try again */
2932                 }
2933                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2934                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
2935                                  ~(PAGE_SIZE * NPTEPG - 1);
2936                         if (kstart - 1 >= kernel_map.max_offset) {
2937                                 kstart = kernel_map.max_offset;
2938                                 break;
2939                         }
2940                         continue;
2941                 }
2942
2943                 /*
2944                  * We need a new PT
2945                  *
2946                  * This index is bogus, but out of the way
2947                  */
2948                 nkpg = vm_page_alloc(NULL, nkpt,
2949                                      VM_ALLOC_NORMAL |
2950                                      VM_ALLOC_SYSTEM |
2951                                      VM_ALLOC_INTERRUPT);
2952                 if (nkpg == NULL)
2953                         panic("pmap_growkernel: no memory to grow kernel");
2954
2955                 vm_page_wire(nkpg);
2956                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2957                 pmap_zero_page(ptppaddr);
2958                 newpt = (pd_entry_t) (ptppaddr |
2959                     kernel_pmap.pmap_bits[PG_V_IDX] |
2960                     kernel_pmap.pmap_bits[PG_RW_IDX] |
2961                     kernel_pmap.pmap_bits[PG_A_IDX] |
2962                     kernel_pmap.pmap_bits[PG_M_IDX]);
2963                 *pmap_pt(&kernel_pmap, kstart) = newpt;
2964                 nkpt++;
2965
2966                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2967                           ~(PAGE_SIZE * NPTEPG - 1);
2968
2969                 if (kstart - 1 >= kernel_map.max_offset) {
2970                         kstart = kernel_map.max_offset;
2971                         break;
2972                 }
2973         }
2974
2975         /*
2976          * Only update kernel_vm_end for areas below KERNBASE.
2977          */
2978         if (update_kernel_vm_end && kernel_vm_end < kstart)
2979                 kernel_vm_end = kstart;
2980 }
2981
2982 /*
2983  *      Add a reference to the specified pmap.
2984  */
2985 void
2986 pmap_reference(pmap_t pmap)
2987 {
2988         if (pmap != NULL) {
2989                 lwkt_gettoken(&pmap->pm_token);
2990                 ++pmap->pm_count;
2991                 lwkt_reltoken(&pmap->pm_token);
2992         }
2993 }
2994
2995 /***************************************************
2996  * page management routines.
2997  ***************************************************/
2998
2999 /*
3000  * Hold a pv without locking it
3001  */
3002 static void
3003 pv_hold(pv_entry_t pv)
3004 {
3005         atomic_add_int(&pv->pv_hold, 1);
3006 }
3007
3008 /*
3009  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
3010  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
3011  * the pv properly.
3012  *
3013  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3014  * pv list via its page) must be held by the caller.
3015  */
3016 static int
3017 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3018 {
3019         u_int count;
3020
3021         /*
3022          * Critical path shortcut expects pv to already have one ref
3023          * (for the pv->pv_pmap).
3024          */
3025         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3026 #ifdef PMAP_DEBUG
3027                 pv->pv_func = func;
3028                 pv->pv_line = lineno;
3029 #endif
3030                 return TRUE;
3031         }
3032
3033         for (;;) {
3034                 count = pv->pv_hold;
3035                 cpu_ccfence();
3036                 if ((count & PV_HOLD_LOCKED) == 0) {
3037                         if (atomic_cmpset_int(&pv->pv_hold, count,
3038                                               (count + 1) | PV_HOLD_LOCKED)) {
3039 #ifdef PMAP_DEBUG
3040                                 pv->pv_func = func;
3041                                 pv->pv_line = lineno;
3042 #endif
3043                                 return TRUE;
3044                         }
3045                 } else {
3046                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3047                                 return FALSE;
3048                 }
3049                 /* retry */
3050         }
3051 }
3052
3053 /*
3054  * Drop a previously held pv_entry which could not be locked, allowing its
3055  * destruction.
3056  *
3057  * Must not be called with a spinlock held as we might zfree() the pv if it
3058  * is no longer associated with a pmap and this was the last hold count.
3059  */
3060 static void
3061 pv_drop(pv_entry_t pv)
3062 {
3063         u_int count;
3064
3065         for (;;) {
3066                 count = pv->pv_hold;
3067                 cpu_ccfence();
3068                 KKASSERT((count & PV_HOLD_MASK) > 0);
3069                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3070                          (PV_HOLD_LOCKED | 1));
3071                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3072                         if ((count & PV_HOLD_MASK) == 1) {
3073 #ifdef PMAP_DEBUG2
3074                                 if (pmap_enter_debug > 0) {
3075                                         --pmap_enter_debug;
3076                                         kprintf("pv_drop: free pv %p\n", pv);
3077                                 }
3078 #endif
3079                                 KKASSERT(count == 1);
3080                                 KKASSERT(pv->pv_pmap == NULL);
3081                                 zfree(pvzone, pv);
3082                         }
3083                         return;
3084                 }
3085                 /* retry */
3086         }
3087 }
3088
3089 /*
3090  * Find or allocate the requested PV entry, returning a locked, held pv.
3091  *
3092  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3093  * for the caller and one representing the pmap and vm_page association.
3094  *
3095  * If (*isnew) is zero, the returned pv will have only one hold count.
3096  *
3097  * Since both associations can only be adjusted while the pv is locked,
3098  * together they represent just one additional hold.
3099  */
3100 static
3101 pv_entry_t
3102 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3103 {
3104         pv_entry_t pv;
3105         pv_entry_t pnew = NULL;
3106
3107         spin_lock(&pmap->pm_spin);
3108         for (;;) {
3109                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3110                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3111                                                         pindex);
3112                 }
3113                 if (pv == NULL) {
3114                         if (pnew == NULL) {
3115                                 spin_unlock(&pmap->pm_spin);
3116                                 pnew = zalloc(pvzone);
3117                                 spin_lock(&pmap->pm_spin);
3118                                 continue;
3119                         }
3120                         pnew->pv_pmap = pmap;
3121                         pnew->pv_pindex = pindex;
3122                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
3123 #ifdef PMAP_DEBUG
3124                         pnew->pv_func = func;
3125                         pnew->pv_line = lineno;
3126 #endif
3127                         pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3128                         ++pmap->pm_generation;
3129                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
3130                         spin_unlock(&pmap->pm_spin);
3131                         *isnew = 1;
3132                         return(pnew);
3133                 }
3134                 if (pnew) {
3135                         spin_unlock(&pmap->pm_spin);
3136                         zfree(pvzone, pnew);
3137                         pnew = NULL;
3138                         spin_lock(&pmap->pm_spin);
3139                         continue;
3140                 }
3141                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3142                         spin_unlock(&pmap->pm_spin);
3143                 } else {
3144                         spin_unlock(&pmap->pm_spin);
3145                         _pv_lock(pv PMAP_DEBUG_COPY);
3146                 }
3147                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3148                         *isnew = 0;
3149                         return(pv);
3150                 }
3151                 pv_put(pv);
3152                 spin_lock(&pmap->pm_spin);
3153         }
3154 }
3155
3156 /*
3157  * Find the requested PV entry, returning a locked+held pv or NULL
3158  */
3159 static
3160 pv_entry_t
3161 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
3162 {
3163         pv_entry_t pv;
3164
3165         spin_lock(&pmap->pm_spin);
3166         for (;;) {
3167                 /*
3168                  * Shortcut cache
3169                  */
3170                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3171                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3172                                                         pindex);
3173                 }
3174                 if (pv == NULL) {
3175                         spin_unlock(&pmap->pm_spin);
3176                         return NULL;
3177                 }
3178                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3179                         spin_unlock(&pmap->pm_spin);
3180                 } else {
3181                         spin_unlock(&pmap->pm_spin);
3182                         _pv_lock(pv PMAP_DEBUG_COPY);
3183                 }
3184                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3185                         pv_cache(pv, pindex);
3186                         return(pv);
3187                 }
3188                 pv_put(pv);
3189                 spin_lock(&pmap->pm_spin);
3190         }
3191 }
3192
3193 /*
3194  * Lookup, hold, and attempt to lock (pmap,pindex).
3195  *
3196  * If the entry does not exist NULL is returned and *errorp is set to 0
3197  *
3198  * If the entry exists and could be successfully locked it is returned and
3199  * errorp is set to 0.
3200  *
3201  * If the entry exists but could NOT be successfully locked it is returned
3202  * held and *errorp is set to 1.
3203  */
3204 static
3205 pv_entry_t
3206 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
3207 {
3208         pv_entry_t pv;
3209
3210         spin_lock_shared(&pmap->pm_spin);
3211         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3212                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3213         if (pv == NULL) {
3214                 spin_unlock_shared(&pmap->pm_spin);
3215                 *errorp = 0;
3216                 return NULL;
3217         }
3218         if (pv_hold_try(pv)) {
3219                 pv_cache(pv, pindex);
3220                 spin_unlock_shared(&pmap->pm_spin);
3221                 *errorp = 0;
3222                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3223                 return(pv);     /* lock succeeded */
3224         }
3225         spin_unlock_shared(&pmap->pm_spin);
3226         *errorp = 1;
3227         return (pv);            /* lock failed */
3228 }
3229
3230 /*
3231  * Find the requested PV entry, returning a held pv or NULL
3232  */
3233 static
3234 pv_entry_t
3235 pv_find(pmap_t pmap, vm_pindex_t pindex)
3236 {
3237         pv_entry_t pv;
3238
3239         spin_lock_shared(&pmap->pm_spin);
3240
3241         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3242                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3243         if (pv == NULL) {
3244                 spin_unlock_shared(&pmap->pm_spin);
3245                 return NULL;
3246         }
3247         pv_hold(pv);
3248         pv_cache(pv, pindex);
3249         spin_unlock_shared(&pmap->pm_spin);
3250         return(pv);
3251 }
3252
3253 /*
3254  * Lock a held pv, keeping the hold count
3255  */
3256 static
3257 void
3258 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3259 {
3260         u_int count;
3261
3262         for (;;) {
3263                 count = pv->pv_hold;
3264                 cpu_ccfence();
3265                 if ((count & PV_HOLD_LOCKED) == 0) {
3266                         if (atomic_cmpset_int(&pv->pv_hold, count,
3267                                               count | PV_HOLD_LOCKED)) {
3268 #ifdef PMAP_DEBUG
3269                                 pv->pv_func = func;
3270                                 pv->pv_line = lineno;
3271 #endif
3272                                 return;
3273                         }
3274                         continue;
3275                 }
3276                 tsleep_interlock(pv, 0);
3277                 if (atomic_cmpset_int(&pv->pv_hold, count,
3278                                       count | PV_HOLD_WAITING)) {
3279 #ifdef PMAP_DEBUG
3280                         kprintf("pv waiting on %s:%d\n",
3281                                         pv->pv_func, pv->pv_line);
3282 #endif
3283                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
3284                 }
3285                 /* retry */
3286         }
3287 }
3288
3289 /*
3290  * Unlock a held and locked pv, keeping the hold count.
3291  */
3292 static
3293 void
3294 pv_unlock(pv_entry_t pv)
3295 {
3296         u_int count;
3297
3298         for (;;) {
3299                 count = pv->pv_hold;
3300                 cpu_ccfence();
3301                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3302                          (PV_HOLD_LOCKED | 1));
3303                 if (atomic_cmpset_int(&pv->pv_hold, count,
3304                                       count &
3305                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3306                         if (count & PV_HOLD_WAITING)
3307                                 wakeup(pv);
3308                         break;
3309                 }
3310         }
3311 }
3312
3313 /*
3314  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3315  * and the hold count drops to zero we will free it.
3316  *
3317  * Caller should not hold any spin locks.  We are protected from hold races
3318  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3319  * lock held.  A pv cannot be located otherwise.
3320  */
3321 static
3322 void
3323 pv_put(pv_entry_t pv)
3324 {
3325 #ifdef PMAP_DEBUG2
3326         if (pmap_enter_debug > 0) {
3327                 --pmap_enter_debug;
3328                 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3329         }
3330 #endif
3331
3332         /*
3333          * Fast - shortcut most common condition
3334          */
3335         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3336                 return;
3337
3338         /*
3339          * Slow
3340          */
3341         pv_unlock(pv);
3342         pv_drop(pv);
3343 }
3344
3345 /*
3346  * Remove the pmap association from a pv, require that pv_m already be removed,
3347  * then unlock and drop the pv.  Any pte operations must have already been
3348  * completed.  This call may result in a last-drop which will physically free
3349  * the pv.
3350  *
3351  * Removing the pmap association entails an additional drop.
3352  *
3353  * pv must be exclusively locked on call and will be disposed of on return.
3354  */
3355 static
3356 void
3357 pv_free(pv_entry_t pv, pv_entry_t pvp, int putaway)
3358 {
3359         pmap_t pmap;
3360
3361         KKASSERT(pv->pv_m == NULL);
3362         KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3363         if ((pmap = pv->pv_pmap) != NULL) {
3364                 spin_lock(&pmap->pm_spin);
3365                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3366                 ++pmap->pm_generation;
3367                 if (pmap->pm_pvhint == pv)
3368                         pmap->pm_pvhint = NULL;
3369                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3370                 pv->pv_pmap = NULL;
3371                 pv->pv_pindex = 0;
3372                 spin_unlock(&pmap->pm_spin);
3373
3374                 /*
3375                  * Try to shortcut three atomic ops, otherwise fall through
3376                  * and do it normally.  Drop two refs and the lock all in
3377                  * one go.
3378                  */
3379                 if (putaway &&
3380                     atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3381 #ifdef PMAP_DEBUG2
3382                         if (pmap_enter_debug > 0) {
3383                                 --pmap_enter_debug;
3384                                 kprintf("pv_free: free pv %p\n", pv);
3385                         }
3386 #endif
3387                         zfree(pvzone, pv);
3388                         if (pvp)
3389                                 vm_page_unwire_quick(pvp->pv_m);
3390                         return;
3391                 }
3392                 pv_drop(pv);    /* ref for pv_pmap */
3393         }
3394         if (putaway)
3395                 pv_put(pv);
3396         if (pvp)
3397                 vm_page_unwire_quick(pvp->pv_m);
3398 }
3399
3400 /*
3401  * This routine is very drastic, but can save the system
3402  * in a pinch.
3403  */
3404 void
3405 pmap_collect(void)
3406 {
3407         int i;
3408         vm_page_t m;
3409         static int warningdone=0;
3410
3411         if (pmap_pagedaemon_waken == 0)
3412                 return;
3413         pmap_pagedaemon_waken = 0;
3414         if (warningdone < 5) {
3415                 kprintf("pmap_collect: collecting pv entries -- "
3416                         "suggest increasing PMAP_SHPGPERPROC\n");
3417                 warningdone++;
3418         }
3419
3420         for (i = 0; i < vm_page_array_size; i++) {
3421                 m = &vm_page_array[i];
3422                 if (m->wire_count || m->hold_count)
3423                         continue;
3424                 if (vm_page_busy_try(m, TRUE) == 0) {
3425                         if (m->wire_count == 0 && m->hold_count == 0) {
3426                                 pmap_remove_all(m);
3427                         }
3428                         vm_page_wakeup(m);
3429                 }
3430         }
3431 }
3432
3433 /*
3434  * Scan the pmap for active page table entries and issue a callback.
3435  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3436  * its parent page table.
3437  *
3438  * pte_pv will be NULL if the page or page table is unmanaged.
3439  * pt_pv will point to the page table page containing the pte for the page.
3440  *
3441  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3442  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3443  *       process pmap's PD and page to the callback function.  This can be
3444  *       confusing because the pt_pv is really a pd_pv, and the target page
3445  *       table page is simply aliased by the pmap and not owned by it.
3446  *
3447  * It is assumed that the start and end are properly rounded to the page size.
3448  *
3449  * It is assumed that PD pages and above are managed and thus in the RB tree,
3450  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3451  */
3452 struct pmap_scan_info {
3453         struct pmap *pmap;
3454         vm_offset_t sva;
3455         vm_offset_t eva;
3456         vm_pindex_t sva_pd_pindex;
3457         vm_pindex_t eva_pd_pindex;
3458         void (*func)(pmap_t, struct pmap_scan_info *,
3459                      pv_entry_t, pv_entry_t, int, vm_offset_t,
3460                      pt_entry_t *, void *);
3461         void *arg;
3462         pmap_inval_bulk_t bulk_core;
3463         pmap_inval_bulk_t *bulk;
3464         int count;
3465         int stop;
3466 };
3467
3468 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3469 static int pmap_scan_callback(pv_entry_t pv, void *data);
3470
3471 static void
3472 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3473 {
3474         struct pmap *pmap = info->pmap;
3475         pv_entry_t pd_pv;       /* A page directory PV */
3476         pv_entry_t pt_pv;       /* A page table PV */
3477         pv_entry_t pte_pv;      /* A page table entry PV */
3478         pt_entry_t *ptep;
3479         pt_entry_t oldpte;
3480         struct pv_entry dummy_pv;
3481         int generation;
3482
3483         info->stop = 0;
3484         if (pmap == NULL)
3485                 return;
3486         if (smp_inval) {
3487                 info->bulk = &info->bulk_core;
3488                 pmap_inval_bulk_init(&info->bulk_core, pmap);
3489         } else {
3490                 info->bulk = NULL;
3491         }
3492
3493         /*
3494          * Hold the token for stability; if the pmap is empty we have nothing
3495          * to do.
3496          */
3497         lwkt_gettoken(&pmap->pm_token);
3498 #if 0
3499         if (pmap->pm_stats.resident_count == 0) {
3500                 lwkt_reltoken(&pmap->pm_token);
3501                 return;
3502         }
3503 #endif
3504
3505         info->count = 0;
3506
3507 again:
3508         /*
3509          * Special handling for scanning one page, which is a very common
3510          * operation (it is?).
3511          *
3512          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3513          */
3514         if (info->sva + PAGE_SIZE == info->eva) {
3515                 generation = pmap->pm_generation;
3516                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3517                         /*
3518                          * Kernel mappings do not track wire counts on
3519                          * page table pages and only maintain pd_pv and
3520                          * pte_pv levels so pmap_scan() works.
3521                          */
3522                         pt_pv = NULL;
3523                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3524                         ptep = vtopte(info->sva);
3525                 } else {
3526                         /*
3527                          * User pages which are unmanaged will not have a
3528                          * pte_pv.  User page table pages which are unmanaged
3529                          * (shared from elsewhere) will also not have a pt_pv.
3530                          * The func() callback will pass both pte_pv and pt_pv
3531                          * as NULL in that case.
3532                          */
3533                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3534                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3535                         if (pt_pv == NULL) {
3536                                 KKASSERT(pte_pv == NULL);
3537                                 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3538                                 if (pd_pv) {
3539                                         ptep = pv_pte_lookup(pd_pv,
3540                                                     pmap_pt_index(info->sva));
3541                                         if (*ptep) {
3542                                                 info->func(pmap, info,
3543                                                      NULL, pd_pv, 1,
3544                                                      info->sva, ptep,
3545                                                      info->arg);
3546                                         }
3547                                         pv_put(pd_pv);
3548                                 }
3549                                 goto fast_skip;
3550                         }
3551                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3552                 }
3553
3554                 /*
3555                  * NOTE: *ptep can't be ripped out from under us if we hold
3556                  *       pte_pv locked, but bits can change.  However, there is
3557                  *       a race where another thread may be inserting pte_pv
3558                  *       and setting *ptep just after our pte_pv lookup fails.
3559                  *
3560                  *       In this situation we can end up with a NULL pte_pv
3561                  *       but find that we have a managed *ptep.  We explicitly
3562                  *       check for this race.
3563                  */
3564                 oldpte = *ptep;
3565                 cpu_ccfence();
3566                 if (oldpte == 0) {
3567                         /*
3568                          * Unlike the pv_find() case below we actually
3569                          * acquired a locked pv in this case so any
3570                          * race should have been resolved.  It is expected
3571                          * to not exist.
3572                          */
3573                         KKASSERT(pte_pv == NULL);
3574                 } else if (pte_pv) {
3575                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3576                                            pmap->pmap_bits[PG_V_IDX])) ==
3577                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
3578                                  pmap->pmap_bits[PG_V_IDX]),
3579                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3580                              "generation %d/%d",
3581                             *ptep, oldpte, info->sva, pte_pv,
3582                             generation, pmap->pm_generation));
3583                         info->func(pmap, info, pte_pv, pt_pv, 0,
3584                                    info->sva, ptep, info->arg);
3585                 } else {
3586                         /*
3587                          * Check for insertion race
3588                          */
3589                         if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3590                             pt_pv) {
3591                                 pte_pv = pv_find(pmap,
3592                                                  pmap_pte_pindex(info->sva));
3593                                 if (pte_pv) {
3594                                         pv_drop(pte_pv);
3595                                         pv_put(pt_pv);
3596                                         kprintf("pmap_scan: RACE1 "
3597                                                 "%016jx, %016lx\n",
3598                                                 info->sva, oldpte);
3599                                         goto again;
3600                                 }
3601                         }
3602
3603                         /*
3604                          * Didn't race
3605                          */
3606                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3607                                            pmap->pmap_bits[PG_V_IDX])) ==
3608                             pmap->pmap_bits[PG_V_IDX],
3609                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3610                              "generation %d/%d",
3611                             *ptep, oldpte, info->sva,
3612                             generation, pmap->pm_generation));
3613                         info->func(pmap, info, NULL, pt_pv, 0,
3614                             info->sva, ptep, info->arg);
3615                 }
3616                 if (pt_pv)
3617                         pv_put(pt_pv);
3618 fast_skip:
3619                 pmap_inval_bulk_flush(info->bulk);
3620                 lwkt_reltoken(&pmap->pm_token);
3621                 return;
3622         }
3623
3624         /*
3625          * Nominal scan case, RB_SCAN() for PD pages and iterate from
3626          * there.
3627          */
3628         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3629         info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3630
3631         if (info->sva >= VM_MAX_USER_ADDRESS) {
3632                 /*
3633                  * The kernel does not currently maintain any pv_entry's for
3634                  * higher-level page tables.
3635                  */
3636                 bzero(&dummy_pv, sizeof(dummy_pv));
3637                 dummy_pv.pv_pindex = info->sva_pd_pindex;
3638                 spin_lock(&pmap->pm_spin);
3639                 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3640                         pmap_scan_callback(&dummy_pv, info);
3641                         ++dummy_pv.pv_pindex;
3642                 }
3643                 spin_unlock(&pmap->pm_spin);
3644         } else {
3645                 /*
3646                  * User page tables maintain local PML4, PDP, and PD
3647                  * pv_entry's at the very least.  PT pv's might be
3648                  * unmanaged and thus not exist.  PTE pv's might be
3649                  * unmanaged and thus not exist.
3650                  */
3651                 spin_lock(&pmap->pm_spin);
3652                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3653                         pmap_scan_cmp, pmap_scan_callback, info);
3654                 spin_unlock(&pmap->pm_spin);
3655         }
3656         pmap_inval_bulk_flush(info->bulk);
3657         lwkt_reltoken(&pmap->pm_token);
3658 }
3659
3660 /*
3661  * WARNING! pmap->pm_spin held
3662  */
3663 static int
3664 pmap_scan_cmp(pv_entry_t pv, void *data)
3665 {
3666         struct pmap_scan_info *info = data;
3667         if (pv->pv_pindex < info->sva_pd_pindex)
3668                 return(-1);
3669         if (pv->pv_pindex >= info->eva_pd_pindex)
3670                 return(1);
3671         return(0);
3672 }
3673
3674 /*
3675  * WARNING! pmap->pm_spin held
3676  */
3677 static int
3678 pmap_scan_callback(pv_entry_t pv, void *data)
3679 {
3680         struct pmap_scan_info *info = data;
3681         struct pmap *pmap = info->pmap;
3682         pv_entry_t pd_pv;       /* A page directory PV */
3683         pv_entry_t pt_pv;       /* A page table PV */
3684         pv_entry_t pte_pv;      /* A page table entry PV */
3685         pt_entry_t *ptep;
3686         pt_entry_t oldpte;
3687         vm_offset_t sva;
3688         vm_offset_t eva;
3689         vm_offset_t va_next;
3690         vm_pindex_t pd_pindex;
3691         int error;
3692         int generation;
3693
3694         /*
3695          * Stop if requested
3696          */
3697         if (info->stop)
3698                 return -1;
3699
3700         /*
3701          * Pull the PD pindex from the pv before releasing the spinlock.
3702          *
3703          * WARNING: pv is faked for kernel pmap scans.
3704          */
3705         pd_pindex = pv->pv_pindex;
3706         spin_unlock(&pmap->pm_spin);
3707         pv = NULL;      /* invalid after spinlock unlocked */
3708
3709         /*
3710          * Calculate the page range within the PD.  SIMPLE pmaps are
3711          * direct-mapped for the entire 2^64 address space.  Normal pmaps
3712          * reflect the user and kernel address space which requires
3713          * cannonicalization w/regards to converting pd_pindex's back
3714          * into addresses.
3715          */
3716         sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3717         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3718             (sva & PML4_SIGNMASK)) {
3719                 sva |= PML4_SIGNMASK;
3720         }
3721         eva = sva + NBPDP;      /* can overflow */
3722         if (sva < info->sva)
3723                 sva = info->sva;
3724         if (eva < info->sva || eva > info->eva)
3725                 eva = info->eva;
3726
3727         /*
3728          * NOTE: kernel mappings do not track page table pages, only
3729          *       terminal pages.
3730          *
3731          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3732          *       However, for the scan to be efficient we try to
3733          *       cache items top-down.
3734          */
3735         pd_pv = NULL;
3736         pt_pv = NULL;
3737
3738         for (; sva < eva; sva = va_next) {
3739                 if (info->stop)
3740                         break;
3741                 if (sva >= VM_MAX_USER_ADDRESS) {
3742                         if (pt_pv) {
3743                                 pv_put(pt_pv);
3744                                 pt_pv = NULL;
3745                         }
3746                         goto kernel_skip;
3747                 }
3748
3749                 /*
3750                  * PD cache (degenerate case if we skip).  It is possible
3751                  * for the PD to not exist due to races.  This is ok.
3752                  */
3753                 if (pd_pv == NULL) {
3754                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3755                 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3756                         pv_put(pd_pv);
3757                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3758                 }
3759                 if (pd_pv == NULL) {
3760                         va_next = (sva + NBPDP) & ~PDPMASK;
3761                         if (va_next < sva)
3762                                 va_next = eva;
3763                         continue;
3764                 }
3765
3766                 /*
3767                  * PT cache
3768                  */
3769                 if (pt_pv == NULL) {
3770                         vm_page_wire_quick(pd_pv->pv_m);
3771                         pv_unlock(pd_pv);
3772                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3773                         pv_lock(pd_pv);
3774                         vm_page_unwire_quick(pd_pv->pv_m);
3775                 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3776                         vm_page_wire_quick(pd_pv->pv_m);
3777                         pv_unlock(pd_pv);
3778                         pv_put(pt_pv);
3779                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3780                         pv_lock(pd_pv);
3781                         vm_page_unwire_quick(pd_pv->pv_m);
3782                 }
3783
3784                 /*
3785                  * If pt_pv is NULL we either have an shared page table
3786                  * page and must issue a callback specific to that case,
3787                  * or there is no page table page.
3788                  *
3789                  * Either way we can skip the page table page.
3790                  */
3791                 if (pt_pv == NULL) {
3792                         /*
3793                          * Possible unmanaged (shared from another pmap)
3794                          * page table page.
3795                          */
3796                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3797                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3798                                 info->func(pmap, info, NULL, pd_pv, 1,
3799                                            sva, ptep, info->arg);
3800                         }
3801
3802                         /*
3803                          * Done, move to next page table page.
3804                          */
3805                         va_next = (sva + NBPDR) & ~PDRMASK;
3806                         if (va_next < sva)
3807                                 va_next = eva;
3808                         continue;
3809                 }
3810
3811                 /*
3812                  * From this point in the loop testing pt_pv for non-NULL
3813                  * means we are in UVM, else if it is NULL we are in KVM.
3814                  *
3815                  * Limit our scan to either the end of the va represented
3816                  * by the current page table page, or to the end of the
3817                  * range being removed.
3818                  */
3819 kernel_skip:
3820                 va_next = (sva + NBPDR) & ~PDRMASK;
3821                 if (va_next < sva)
3822                         va_next = eva;
3823                 if (va_next > eva)
3824                         va_next = eva;
3825
3826                 /*
3827                  * Scan the page table for pages.  Some pages may not be
3828                  * managed (might not have a pv_entry).
3829                  *
3830                  * There is no page table management for kernel pages so
3831                  * pt_pv will be NULL in that case, but otherwise pt_pv
3832                  * is non-NULL, locked, and referenced.
3833                  */
3834
3835                 /*
3836                  * At this point a non-NULL pt_pv means a UVA, and a NULL
3837                  * pt_pv means a KVA.
3838                  */
3839                 if (pt_pv)
3840                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3841                 else
3842                         ptep = vtopte(sva);
3843
3844                 while (sva < va_next) {
3845                         /*
3846                          * Yield every 64 pages, stop if requested.
3847                          */
3848                         if ((++info->count & 63) == 0)
3849                                 lwkt_user_yield();
3850                         if (info->stop)
3851                                 break;
3852
3853                         /*
3854                          * Check if pt_pv has been lost (probably due to
3855                          * a remove of the underlying pages).
3856                          */
3857                         if (pt_pv && pt_pv->pv_pmap == NULL)
3858                                 break;
3859
3860                         /*
3861                          * Acquire the related pte_pv, if any.  If *ptep == 0
3862                          * the related pte_pv should not exist, but if *ptep
3863                          * is not zero the pte_pv may or may not exist (e.g.
3864                          * will not exist for an unmanaged page).
3865                          *
3866                          * However a multitude of races are possible here.
3867                          *
3868                          * In addition, the (pt_pv, pte_pv) lock order is
3869                          * backwards, so we have to be careful in aquiring
3870                          * a properly locked pte_pv.
3871                          */
3872                         generation = pmap->pm_generation;
3873                         if (pt_pv) {
3874                                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3875                                                     &error);
3876                                 if (error) {
3877                                         if (pd_pv) {
3878                                                 vm_page_wire_quick(pd_pv->pv_m);
3879                                                 pv_unlock(pd_pv);
3880                                         }
3881                                         vm_page_wire_quick(pt_pv->pv_m);
3882                                         pv_unlock(pt_pv);/* must be non-NULL */
3883                                         pv_lock(pte_pv); /* safe to block now */
3884                                         pv_put(pte_pv);
3885                                         pte_pv = NULL;
3886                                         pv_lock(pt_pv);
3887                                         vm_page_unwire_quick(pt_pv->pv_m);
3888
3889                                         /*
3890                                          * pt_pv reloaded, need new ptep
3891                                          */
3892                                         KKASSERT(pt_pv != NULL);
3893                                         ptep = pv_pte_lookup(pt_pv,
3894                                                         pmap_pte_index(sva));
3895                                         if (pd_pv) {
3896                                                 pv_lock(pd_pv);
3897                                                 vm_page_unwire_quick(pd_pv->pv_m);
3898                                         }
3899                                         continue;
3900                                 }
3901                         } else {
3902                                 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3903                         }
3904
3905                         /*
3906                          * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3907                          */
3908                         oldpte = *ptep;
3909                         if (oldpte == 0) {
3910                                 if (pte_pv) {
3911                                         kprintf("Unexpected non-NULL pte_pv "
3912                                                 "%p pt_pv %p "
3913                                                 "*ptep = %016lx/%016lx\n",
3914                                                 pte_pv, pt_pv, *ptep, oldpte);
3915                                         panic("Unexpected non-NULL pte_pv");
3916                                 }
3917                                 sva += PAGE_SIZE;
3918                                 ++ptep;
3919                                 continue;
3920                         }
3921
3922                         /*
3923                          * Ready for the callback.  The locked pte_pv (if any)
3924                          * is consumed by the callback.  pte_pv will exist if
3925                          * the page is managed, and will not exist if it
3926                          * isn't.
3927                          */
3928                         if (pte_pv) {
3929                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3930                                     (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3931                                     ("badC *ptep %016lx/%016lx sva %016lx "
3932                                     "pte_pv %p pm_generation %d/%d",
3933                                     *ptep, oldpte, sva, pte_pv,
3934                                     generation, pmap->pm_generation));
3935                                 /*
3936                                  * We must unlock pd_pv across the callback
3937                                  * to avoid deadlocks on any recursive
3938                                  * disposal.  Re-check that it still exists
3939                                  * after re-locking.
3940                                  */
3941                                 if (pd_pv)
3942                                         pv_unlock(pd_pv);
3943                                 info->func(pmap, info, pte_pv, pt_pv, 0,
3944                                            sva, ptep, info->arg);
3945                                 if (pd_pv) {
3946                                         pv_lock(pd_pv);
3947                                         if (pd_pv->pv_pmap == NULL) {
3948                                                 pv_put(pd_pv);
3949                                                 pd_pv = NULL;
3950                                         }
3951                                 }
3952                         } else {
3953                                 /*
3954                                  * Check for insertion race.  Since there is no
3955                                  * pte_pv to guard us it is possible for us
3956                                  * to race another thread doing an insertion.
3957                                  * Our lookup misses the pte_pv but our *ptep
3958                                  * check sees the inserted pte.
3959                                  *
3960                                  * XXX panic case seems to occur within a
3961                                  * vm_fork() of /bin/sh, which frankly
3962                                  * shouldn't happen since no other threads
3963                                  * should be inserting to our pmap in that
3964                                  * situation.  Removing, possibly.  Inserting,
3965                                  * shouldn't happen.
3966                                  */
3967                                 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3968                                     pt_pv) {
3969                                         pte_pv = pv_find(pmap,
3970                                                          pmap_pte_pindex(sva));
3971                                         if (pte_pv) {
3972                                                 pv_drop(pte_pv);
3973                                                 kprintf("pmap_scan: RACE2 "
3974                                                         "%016jx, %016lx\n",
3975                                                         sva, oldpte);
3976                                                 continue;
3977                                         }
3978                                 }
3979
3980                                 /*
3981                                  * Didn't race
3982                                  *
3983                                  * We must unlock pd_pv across the callback
3984                                  * to avoid deadlocks on any recursive
3985                                  * disposal.  Re-check that it still exists
3986                                  * after re-locking.
3987                                  */
3988                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3989                                     pmap->pmap_bits[PG_V_IDX],
3990                                     ("badD *ptep %016lx/%016lx sva %016lx "
3991                                     "pte_pv NULL pm_generation %d/%d",
3992                                      *ptep, oldpte, sva,
3993                                      generation, pmap->pm_generation));
3994                                 if (pd_pv)
3995                                         pv_unlock(pd_pv);
3996                                 info->func(pmap, info, NULL, pt_pv, 0,
3997                                            sva, ptep, info->arg);
3998                                 if (pd_pv) {
3999                                         pv_lock(pd_pv);
4000                                         if (pd_pv->pv_pmap == NULL) {
4001                                                 pv_put(pd_pv);
4002                                                 pd_pv = NULL;
4003                                         }
4004                                 }
4005                         }
4006                         pte_pv = NULL;
4007                         sva += PAGE_SIZE;
4008                         ++ptep;
4009                 }
4010         }
4011         if (pd_pv) {
4012                 pv_put(pd_pv);
4013                 pd_pv = NULL;
4014         }
4015         if (pt_pv) {
4016                 pv_put(pt_pv);
4017                 pt_pv = NULL;
4018         }
4019         if ((++info->count & 7) == 0)
4020                 lwkt_user_yield();
4021
4022         /*
4023          * Relock before returning.
4024          */
4025         spin_lock(&pmap->pm_spin);
4026         return (0);
4027 }
4028
4029 void
4030 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4031 {
4032         struct pmap_scan_info info;
4033
4034         info.pmap = pmap;
4035         info.sva = sva;
4036         info.eva = eva;
4037         info.func = pmap_remove_callback;
4038         info.arg = NULL;
4039         pmap_scan(&info, 1);
4040 }
4041
4042 static void
4043 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4044 {
4045         struct pmap_scan_info info;
4046
4047         info.pmap = pmap;
4048         info.sva = sva;
4049         info.eva = eva;
4050         info.func = pmap_remove_callback;
4051         info.arg = NULL;
4052         pmap_scan(&info, 0);
4053 }
4054
4055 static void
4056 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4057                      pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4058                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4059 {
4060         pt_entry_t pte;
4061
4062         if (pte_pv) {
4063                 /*
4064                  * This will also drop pt_pv's wire_count. Note that
4065                  * terminal pages are not wired based on mmu presence.
4066                  *
4067                  * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4068                  */
4069                 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4070                 pte_pv = NULL;  /* safety */
4071
4072                 /*
4073                  * Recursively destroy higher-level page tables.
4074                  *
4075                  * This is optional.  If we do not, they will still
4076                  * be destroyed when the process exits.
4077                  *
4078                  * NOTE: Do not destroy pv_entry's with extra hold refs,
4079                  *       a caller may have unlocked it and intends to
4080                  *       continue to use it.
4081                  */
4082                 if (pmap_dynamic_delete &&
4083                     pt_pv &&
4084                     pt_pv->pv_m &&
4085                     pt_pv->pv_m->wire_count == 1 &&
4086                     (pt_pv->pv_hold & PV_HOLD_MASK) == 2 &&
4087                     pt_pv->pv_pindex != pmap_pml4_pindex()) {
4088                         if (pmap_dynamic_delete == 2)
4089                                 kprintf("B %jd %08x\n", pt_pv->pv_pindex, pt_pv->pv_hold);
4090                         pv_hold(pt_pv);
4091                         pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4092                         pv_lock(pt_pv);
4093                 }
4094         } else if (sharept == 0) {
4095                 /*
4096                  * Unmanaged page table (pt, pd, or pdp. Not pte).
4097                  *
4098                  * pt_pv's wire_count is still bumped by unmanaged pages
4099                  * so we must decrement it manually.
4100                  *
4101                  * We have to unwire the target page table page.
4102                  *
4103                  * It is unclear how we can invalidate a segment so we
4104                  * invalidate -1 which invlidates the tlb.
4105                  */
4106                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4107                 if (pte & pmap->pmap_bits[PG_W_IDX])
4108                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4109                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4110                 if (vm_page_unwire_quick(pt_pv->pv_m))
4111                         panic("pmap_remove: insufficient wirecount");
4112         } else {
4113                 /*
4114                  * Unmanaged page table (pt, pd, or pdp. Not pte) for
4115                  * a shared page table.
4116                  *
4117                  * pt_pv is actually the pd_pv for our pmap (not the shared
4118                  * object pmap).
4119                  *
4120                  * We have to unwire the target page table page and we
4121                  * have to unwire our page directory page.
4122                  *
4123                  * It is unclear how we can invalidate a segment so we
4124                  * invalidate -1 which invlidates the tlb.
4125                  */
4126                 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4127                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4128                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4129                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4130                         panic("pmap_remove: shared pgtable1 bad wirecount");
4131                 if (vm_page_unwire_quick(pt_pv->pv_m))
4132                         panic("pmap_remove: shared pgtable2 bad wirecount");
4133         }
4134 }
4135
4136 /*
4137  * Removes this physical page from all physical maps in which it resides.
4138  * Reflects back modify bits to the pager.
4139  *
4140  * This routine may not be called from an interrupt.
4141  */
4142 static
4143 void
4144 pmap_remove_all(vm_page_t m)
4145 {
4146         pv_entry_t pv;
4147         pmap_inval_bulk_t bulk;
4148
4149         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4150                 return;
4151
4152         vm_page_spin_lock(m);
4153         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4154                 KKASSERT(pv->pv_m == m);
4155                 if (pv_hold_try(pv)) {
4156                         vm_page_spin_unlock(m);
4157                 } else {
4158                         vm_page_spin_unlock(m);
4159                         pv_lock(pv);
4160                 }
4161                 if (pv->pv_m != m) {
4162                         pv_put(pv);
4163                         vm_page_spin_lock(m);
4164                         continue;
4165                 }
4166
4167                 /*
4168                  * Holding no spinlocks, pv is locked.
4169                  */
4170                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4171                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4172                 pv = NULL;      /* safety */
4173                 pmap_inval_bulk_flush(&bulk);
4174 #if 0
4175                 pmap_remove_pv_page(pv);
4176                 pv_free(pv, 1);
4177 #endif
4178                 vm_page_spin_lock(m);
4179         }
4180         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4181         vm_page_spin_unlock(m);
4182 }
4183
4184 /*
4185  * Removes the page from a particular pmap
4186  */
4187 void
4188 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4189 {
4190         pv_entry_t pv;
4191         pmap_inval_bulk_t bulk;
4192
4193         if (!pmap_initialized)
4194                 return;
4195
4196 again:
4197         vm_page_spin_lock(m);
4198         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4199                 if (pv->pv_pmap != pmap)
4200                         continue;
4201                 KKASSERT(pv->pv_m == m);
4202                 if (pv_hold_try(pv)) {
4203                         vm_page_spin_unlock(m);
4204                 } else {
4205                         vm_page_spin_unlock(m);
4206                         pv_lock(pv);
4207                 }
4208                 if (pv->pv_m != m) {
4209                         pv_put(pv);
4210                         goto again;
4211                 }
4212
4213                 /*
4214                  * Holding no spinlocks, pv is locked.
4215                  */
4216                 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4217                 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4218                 pv = NULL;      /* safety */
4219                 pmap_inval_bulk_flush(&bulk);
4220 #if 0
4221                 pmap_remove_pv_page(pv);
4222                 pv_free(pv, 1);
4223 #endif
4224                 goto again;
4225         }
4226         vm_page_spin_unlock(m);
4227 }
4228
4229 /*
4230  * Set the physical protection on the specified range of this map
4231  * as requested.  This function is typically only used for debug watchpoints
4232  * and COW pages.
4233  *
4234  * This function may not be called from an interrupt if the map is
4235  * not the kernel_pmap.
4236  *
4237  * NOTE!  For shared page table pages we just unmap the page.
4238  */
4239 void
4240 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4241 {
4242         struct pmap_scan_info info;
4243         /* JG review for NX */
4244
4245         if (pmap == NULL)
4246                 return;
4247         if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
4248                 pmap_remove(pmap, sva, eva);
4249                 return;
4250         }
4251         if (prot & VM_PROT_WRITE)
4252                 return;
4253         info.pmap = pmap;
4254         info.sva = sva;
4255         info.eva = eva;
4256         info.func = pmap_protect_callback;
4257         info.arg = &prot;
4258         pmap_scan(&info, 1);
4259 }
4260
4261 static
4262 void
4263 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4264                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4265                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4266 {
4267         pt_entry_t pbits;
4268         pt_entry_t cbits;
4269         pt_entry_t pte;
4270         vm_page_t m;
4271
4272 again:
4273         pbits = *ptep;
4274         cbits = pbits;
4275         if (pte_pv) {
4276                 m = NULL;
4277                 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4278                         if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4279                                 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4280                                 KKASSERT(m == pte_pv->pv_m);
4281                                 vm_page_flag_set(m, PG_REFERENCED);
4282                         }
4283                         cbits &= ~pmap->pmap_bits[PG_A_IDX];
4284                 }
4285                 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4286                         if (pmap_track_modified(pte_pv->pv_pindex)) {
4287                                 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4288                                         if (m == NULL) {
4289                                                 m = PHYS_TO_VM_PAGE(pbits &
4290                                                                     PG_FRAME);
4291                                         }
4292                                         vm_page_dirty(m);
4293                                 }
4294                                 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4295                         }
4296                 }
4297         } else if (sharept) {
4298                 /*
4299                  * Unmanaged page table, pt_pv is actually the pd_pv
4300                  * for our pmap (not the object's shared pmap).
4301                  *
4302                  * When asked to protect something in a shared page table
4303                  * page we just unmap the page table page.  We have to
4304                  * invalidate the tlb in this situation.
4305                  *
4306                  * XXX Warning, shared page tables will not be used for
4307                  * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4308                  * so PHYS_TO_VM_PAGE() should be safe here.
4309                  */
4310                 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4311                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4312                         panic("pmap_protect: pgtable1 pg bad wirecount");
4313                 if (vm_page_unwire_quick(pt_pv->pv_m))
4314                         panic("pmap_protect: pgtable2 pg bad wirecount");
4315                 ptep = NULL;
4316         }
4317         /* else unmanaged page, adjust bits, no wire changes */
4318
4319         if (ptep) {
4320                 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4321 #ifdef PMAP_DEBUG2
4322                 if (pmap_enter_debug > 0) {
4323                         --pmap_enter_debug;
4324                         kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4325                                 "pt_pv=%p cbits=%08lx\n",
4326                                 va, ptep, pte_pv,
4327                                 pt_pv, cbits
4328                         );
4329                 }
4330 #endif
4331                 if (pbits != cbits) {
4332                         if (!pmap_inval_smp_cmpset(pmap, (vm_offset_t)-1,
4333                                                    ptep, pbits, cbits)) {
4334                                 goto again;
4335                         }
4336                 }
4337         }
4338         if (pte_pv)
4339                 pv_put(pte_pv);
4340 }
4341
4342 /*
4343  * Insert the vm_page (m) at the virtual address (va), replacing any prior
4344  * mapping at that address.  Set protection and wiring as requested.
4345  *
4346  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4347  * possible.  If it is we enter the page into the appropriate shared pmap
4348  * hanging off the related VM object instead of the passed pmap, then we
4349  * share the page table page from the VM object's pmap into the current pmap.
4350  *
4351  * NOTE: This routine MUST insert the page into the pmap now, it cannot
4352  *       lazy-evaluate.
4353  */
4354 void
4355 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4356            boolean_t wired, vm_map_entry_t entry)
4357 {
4358         pv_entry_t pt_pv;       /* page table */
4359         pv_entry_t pte_pv;      /* page table entry */
4360         pt_entry_t *ptep;
4361         vm_paddr_t opa;
4362         pt_entry_t origpte, newpte;
4363         vm_paddr_t pa;
4364
4365         if (pmap == NULL)
4366                 return;
4367         va = trunc_page(va);
4368 #ifdef PMAP_DIAGNOSTIC
4369         if (va >= KvaEnd)
4370                 panic("pmap_enter: toobig");
4371         if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4372                 panic("pmap_enter: invalid to pmap_enter page table "
4373                       "pages (va: 0x%lx)", va);
4374 #endif
4375         if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4376                 kprintf("Warning: pmap_enter called on UVA with "
4377                         "kernel_pmap\n");
4378 #ifdef DDB
4379                 db_print_backtrace();
4380 #endif
4381         }
4382         if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4383                 kprintf("Warning: pmap_enter called on KVA without"
4384                         "kernel_pmap\n");
4385 #ifdef DDB
4386                 db_print_backtrace();
4387 #endif
4388         }
4389
4390         /*
4391          * Get locked PV entries for our new page table entry (pte_pv)
4392          * and for its parent page table (pt_pv).  We need the parent
4393          * so we can resolve the location of the ptep.
4394          *
4395          * Only hardware MMU actions can modify the ptep out from
4396          * under us.
4397          *
4398          * if (m) is fictitious or unmanaged we do not create a managing
4399          * pte_pv for it.  Any pre-existing page's management state must
4400          * match (avoiding code complexity).
4401          *
4402          * If the pmap is still being initialized we assume existing
4403          * page tables.
4404          *
4405          * Kernel mapppings do not track page table pages (i.e. pt_pv).
4406          */
4407         if (pmap_initialized == FALSE) {
4408                 pte_pv = NULL;
4409                 pt_pv = NULL;
4410                 ptep = vtopte(va);
4411                 origpte = *ptep;
4412         } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4413                 pte_pv = NULL;
4414                 if (va >= VM_MAX_USER_ADDRESS) {
4415                         pt_pv = NULL;
4416                         ptep = vtopte(va);
4417                 } else {
4418                         pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4419                                                   NULL, entry, va);
4420                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4421                 }
4422                 origpte = *ptep;
4423                 cpu_ccfence();
4424                 KASSERT(origpte == 0 ||
4425                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4426                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4427         } else {
4428                 if (va >= VM_MAX_USER_ADDRESS) {
4429                         /*
4430                          * Kernel map, pv_entry-tracked.
4431                          */
4432                         pt_pv = NULL;
4433                         pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4434                         ptep = vtopte(va);
4435                 } else {
4436                         /*
4437                          * User map
4438                          */
4439                         pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4440                                                    &pt_pv, entry, va);
4441                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4442                 }
4443                 origpte = *ptep;
4444                 cpu_ccfence();
4445                 KASSERT(origpte == 0 ||
4446                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4447                          ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4448         }
4449
4450         pa = VM_PAGE_TO_PHYS(m);
4451         opa = origpte & PG_FRAME;
4452
4453         newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4454                  pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4455         if (wired)
4456                 newpte |= pmap->pmap_bits[PG_W_IDX];
4457         if (va < VM_MAX_USER_ADDRESS)
4458                 newpte |= pmap->pmap_bits[PG_U_IDX];
4459         if (pte_pv)
4460                 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4461 //      if (pmap == &kernel_pmap)
4462 //              newpte |= pgeflag;
4463         newpte |= pmap->pmap_cache_bits[m->pat_mode];
4464         if (m->flags & PG_FICTITIOUS)
4465                 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4466
4467         /*
4468          * It is possible for multiple faults to occur in threaded
4469          * environments, the existing pte might be correct.
4470          */
4471         if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4472             pmap->pmap_bits[PG_A_IDX])) == 0)
4473                 goto done;
4474
4475         /*
4476          * Ok, either the address changed or the protection or wiring
4477          * changed.
4478          *
4479          * Clear the current entry, interlocking the removal.  For managed
4480          * pte's this will also flush the modified state to the vm_page.
4481          * Atomic ops are mandatory in order to ensure that PG_M events are
4482          * not lost during any transition.
4483          *
4484          * WARNING: The caller has busied the new page but not the original
4485          *          vm_page which we are trying to replace.  Because we hold
4486          *          the pte_pv lock, but have not busied the page, PG bits
4487          *          can be cleared out from under us.
4488          */
4489         if (opa) {
4490                 if (pte_pv) {
4491                         /*
4492                          * pt_pv won't exist for a kernel page (managed or
4493                          * otherwise).
4494                          */
4495                         if (prot & VM_PROT_NOSYNC) {
4496                                 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4497                         } else {
4498                                 pmap_inval_bulk_t bulk;
4499
4500                                 pmap_inval_bulk_init(&bulk, pmap);
4501                                 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4502                                 pmap_inval_bulk_flush(&bulk);
4503                         }
4504                         if (pte_pv->pv_m)
4505                                 pmap_remove_pv_page(pte_pv);
4506                 } else if (prot & VM_PROT_NOSYNC) {
4507                         /*
4508                          * Unmanaged page, NOSYNC (no mmu sync) requested.
4509                          *
4510                          * Leave wire count on PT page intact.
4511                          */
4512                         (void)pte_load_clear(ptep);
4513                         cpu_invlpg((void *)va);
4514                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
4515                 } else {
4516                         /*
4517                          * Unmanaged page, normal enter.
4518                          *
4519                          * Leave wire count on PT page intact.
4520                          */
4521                         pmap_inval_smp(pmap, va, 1, ptep, 0);
4522                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
4523                 }
4524                 KKASSERT(*ptep == 0);
4525         }
4526
4527 #ifdef PMAP_DEBUG2
4528         if (pmap_enter_debug > 0) {
4529                 --pmap_enter_debug;
4530                 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4531                         " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4532                         va, m,
4533                         origpte, newpte, ptep,
4534                         pte_pv, pt_pv, opa, prot);
4535         }
4536 #endif
4537
4538         if (pte_pv) {
4539                 /*
4540                  * Enter on the PV list if part of our managed memory.
4541                  * Wiring of the PT page is already handled.
4542                  */
4543                 KKASSERT(pte_pv->pv_m == NULL);
4544                 vm_page_spin_lock(m);
4545                 pte_pv->pv_m = m;
4546                 pmap_page_stats_adding(m);
4547                 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4548                 vm_page_flag_set(m, PG_MAPPED);
4549                 vm_page_spin_unlock(m);
4550         } else if (pt_pv && opa == 0) {
4551                 /*
4552                  * We have to adjust the wire count on the PT page ourselves
4553                  * for unmanaged entries.  If opa was non-zero we retained
4554                  * the existing wire count from the removal.
4555                  */
4556                 vm_page_wire_quick(pt_pv->pv_m);
4557         }
4558
4559         /*
4560          * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4561          *
4562          * User VMAs do not because those will be zero->non-zero, so no
4563          * stale entries to worry about at this point.
4564          *
4565          * For KVM there appear to still be issues.  Theoretically we
4566          * should be able to scrap the interlocks entirely but we
4567          * get crashes.
4568          */
4569         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4570                 pmap_inval_smp(pmap, va, 1, ptep, newpte);
4571         } else {
4572                 *(volatile pt_entry_t *)ptep = newpte;
4573                 if (pt_pv == NULL)
4574                         cpu_invlpg((void *)va);
4575         }
4576
4577         if (wired) {
4578                 if (pte_pv) {
4579                         atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4580                                         1);
4581                 } else {
4582                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
4583                 }
4584         }
4585         if (newpte & pmap->pmap_bits[PG_RW_IDX])
4586                 vm_page_flag_set(m, PG_WRITEABLE);
4587
4588         /*
4589          * Unmanaged pages need manual resident_count tracking.
4590          */
4591         if (pte_pv == NULL && pt_pv)
4592                 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4593
4594         /*
4595          * Cleanup
4596          */
4597 done:
4598         KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4599                  (m->flags & PG_MAPPED));
4600
4601         /*
4602          * Cleanup the pv entry, allowing other accessors.
4603          */
4604         if (pte_pv)
4605                 pv_put(pte_pv);
4606         if (pt_pv)
4607                 pv_put(pt_pv);
4608 }
4609
4610 /*
4611  * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4612  * This code also assumes that the pmap has no pre-existing entry for this
4613  * VA.
4614  *
4615  * This code currently may only be used on user pmaps, not kernel_pmap.
4616  */
4617 void
4618 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4619 {
4620         pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4621 }
4622
4623 /*
4624  * Make a temporary mapping for a physical address.  This is only intended
4625  * to be used for panic dumps.
4626  *
4627  * The caller is responsible for calling smp_invltlb().
4628  */
4629 void *
4630 pmap_kenter_temporary(vm_paddr_t pa, long i)
4631 {
4632         pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4633         return ((void *)crashdumpmap);
4634 }
4635
4636 #define MAX_INIT_PT (96)
4637
4638 /*
4639  * This routine preloads the ptes for a given object into the specified pmap.
4640  * This eliminates the blast of soft faults on process startup and
4641  * immediately after an mmap.
4642  */
4643 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4644
4645 void
4646 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4647                     vm_object_t object, vm_pindex_t pindex,
4648                     vm_size_t size, int limit)
4649 {
4650         struct rb_vm_page_scan_info info;
4651         struct lwp *lp;
4652         vm_size_t psize;
4653
4654         /*
4655          * We can't preinit if read access isn't set or there is no pmap
4656          * or object.
4657          */
4658         if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4659                 return;
4660
4661         /*
4662          * We can't preinit if the pmap is not the current pmap
4663          */
4664         lp = curthread->td_lwp;
4665         if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4666                 return;
4667
4668         /*
4669          * Misc additional checks
4670          */
4671         psize = x86_64_btop(size);
4672
4673         if ((object->type != OBJT_VNODE) ||
4674                 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4675                         (object->resident_page_count > MAX_INIT_PT))) {
4676                 return;
4677         }
4678
4679         if (pindex + psize > object->size) {
4680                 if (object->size < pindex)
4681                         return;
4682                 psize = object->size - pindex;
4683         }
4684
4685         if (psize == 0)
4686                 return;
4687
4688         /*
4689          * If everything is segment-aligned do not pre-init here.  Instead
4690          * allow the normal vm_fault path to pass a segment hint to
4691          * pmap_enter() which will then use an object-referenced shared
4692          * page table page.
4693          */
4694         if ((addr & SEG_MASK) == 0 &&
4695             (ctob(psize) & SEG_MASK) == 0 &&
4696             (ctob(pindex) & SEG_MASK) == 0) {
4697                 return;
4698         }
4699
4700         /*
4701          * Use a red-black scan to traverse the requested range and load
4702          * any valid pages found into the pmap.
4703          *
4704          * We cannot safely scan the object's memq without holding the
4705          * object token.
4706          */
4707         info.start_pindex = pindex;
4708         info.end_pindex = pindex + psize - 1;
4709         info.limit = limit;
4710         info.mpte = NULL;
4711         info.addr = addr;
4712         info.pmap = pmap;
4713
4714         vm_object_hold_shared(object);
4715         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4716                                 pmap_object_init_pt_callback, &info);
4717         vm_object_drop(object);
4718 }
4719
4720 static
4721 int
4722 pmap_object_init_pt_callback(vm_page_t p, void *data)
4723 {
4724         struct rb_vm_page_scan_info *info = data;
4725         vm_pindex_t rel_index;
4726
4727         /*
4728          * don't allow an madvise to blow away our really
4729          * free pages allocating pv entries.
4730          */
4731         if ((info->limit & MAP_PREFAULT_MADVISE) &&
4732                 vmstats.v_free_count < vmstats.v_free_reserved) {
4733                     return(-1);
4734         }
4735
4736         /*
4737          * Ignore list markers and ignore pages we cannot instantly
4738          * busy (while holding the object token).
4739          */
4740         if (p->flags & PG_MARKER)
4741                 return 0;
4742         if (vm_page_busy_try(p, TRUE))
4743                 return 0;
4744         if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4745             (p->flags & PG_FICTITIOUS) == 0) {
4746                 if ((p->queue - p->pc) == PQ_CACHE)
4747                         vm_page_deactivate(p);
4748                 rel_index = p->pindex - info->start_pindex;
4749                 pmap_enter_quick(info->pmap,
4750                                  info->addr + x86_64_ptob(rel_index), p);
4751         }
4752         vm_page_wakeup(p);
4753         lwkt_yield();
4754         return(0);
4755 }
4756
4757 /*
4758  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4759  * address.
4760  *
4761  * Returns FALSE if it would be non-trivial or if a pte is already loaded
4762  * into the slot.
4763  *
4764  * XXX This is safe only because page table pages are not freed.
4765  */
4766 int
4767 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4768 {
4769         pt_entry_t *pte;
4770
4771         /*spin_lock(&pmap->pm_spin);*/
4772         if ((pte = pmap_pte(pmap, addr)) != NULL) {
4773                 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4774                         /*spin_unlock(&pmap->pm_spin);*/
4775                         return FALSE;
4776                 }
4777         }
4778         /*spin_unlock(&pmap->pm_spin);*/
4779         return TRUE;
4780 }
4781
4782 /*
4783  * Change the wiring attribute for a pmap/va pair.  The mapping must already
4784  * exist in the pmap.  The mapping may or may not be managed.
4785  *
4786  * Wiring is not a hardware characteristic so there is no need to invalidate
4787  * TLB.  However, in an SMP environment we must use a locked bus cycle to
4788  * update the pte (if we are not using the pmap_inval_*() API that is)...
4789  * it's ok to do this for simple wiring changes.
4790  */
4791 void
4792 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
4793                    vm_map_entry_t entry)
4794 {
4795         pt_entry_t *ptep;
4796         pv_entry_t pv;
4797
4798         if (pmap == NULL)
4799                 return;
4800
4801         lwkt_gettoken(&pmap->pm_token);
4802         if (pmap == &kernel_pmap) {
4803                 /*
4804                  * The kernel may have managed pages, but not managed
4805                  * page tables.
4806                  */
4807                 ptep = pmap_pte_quick(pmap, va);
4808
4809                 if (wired && !pmap_pte_w(pmap, ptep))
4810                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
4811                 else if (!wired && pmap_pte_w(pmap, ptep))
4812                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
4813
4814                 if (wired)
4815                         atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4816                 else
4817                         atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4818         } else {
4819                 /*
4820                  * Userland, the pmap of the possibly shared segment might
4821                  * not be (pmap).
4822                  */
4823                 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL,
4824                                        entry, va);
4825                 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
4826
4827                 if (wired && !pmap_pte_w(pmap, ptep))
4828                         atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4829                 else if (!wired && pmap_pte_w(pmap, ptep))
4830                         atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4831
4832                 if (wired)
4833                         atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4834                 else
4835                         atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4836                 pv_put(pv);
4837         }
4838         lwkt_reltoken(&pmap->pm_token);
4839 }
4840
4841
4842
4843 /*
4844  * Copy the range specified by src_addr/len from the source map to
4845  * the range dst_addr/len in the destination map.
4846  *
4847  * This routine is only advisory and need not do anything.
4848  */
4849 void
4850 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4851           vm_size_t len, vm_offset_t src_addr)
4852 {
4853 }
4854
4855 /*
4856  * pmap_zero_page:
4857  *
4858  *      Zero the specified physical page.
4859  *
4860  *      This function may be called from an interrupt and no locking is
4861  *      required.
4862  */
4863 void
4864 pmap_zero_page(vm_paddr_t phys)
4865 {
4866         vm_offset_t va = PHYS_TO_DMAP(phys);
4867
4868         pagezero((void *)va);
4869 }
4870
4871 /*
4872  * pmap_zero_page:
4873  *
4874  *      Zero part of a physical page by mapping it into memory and clearing
4875  *      its contents with bzero.
4876  *
4877  *      off and size may not cover an area beyond a single hardware page.
4878  */
4879 void
4880 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4881 {
4882         vm_offset_t virt = PHYS_TO_DMAP(phys);
4883
4884         bzero((char *)virt + off, size);
4885 }
4886
4887 /*
4888  * pmap_copy_page:
4889  *
4890  *      Copy the physical page from the source PA to the target PA.
4891  *      This function may be called from an interrupt.  No locking
4892  *      is required.
4893  */
4894 void
4895 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4896 {
4897         vm_offset_t src_virt, dst_virt;
4898
4899         src_virt = PHYS_TO_DMAP(src);
4900         dst_virt = PHYS_TO_DMAP(dst);
4901         bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4902 }
4903
4904 /*
4905  * pmap_copy_page_frag:
4906  *
4907  *      Copy the physical page from the source PA to the target PA.
4908  *      This function may be called from an interrupt.  No locking
4909  *      is required.
4910  */
4911 void
4912 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4913 {
4914         vm_offset_t src_virt, dst_virt;
4915
4916         src_virt = PHYS_TO_DMAP(src);
4917         dst_virt = PHYS_TO_DMAP(dst);
4918
4919         bcopy((char *)src_virt + (src & PAGE_MASK),
4920               (char *)dst_virt + (dst & PAGE_MASK),
4921               bytes);
4922 }
4923
4924 /*
4925  * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4926  * this page.  This count may be changed upwards or downwards in the future;
4927  * it is only necessary that true be returned for a small subset of pmaps
4928  * for proper page aging.
4929  */
4930 boolean_t
4931 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4932 {
4933         pv_entry_t pv;
4934         int loops = 0;
4935
4936         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4937                 return FALSE;
4938
4939         vm_page_spin_lock(m);
4940         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4941                 if (pv->pv_pmap == pmap) {
4942                         vm_page_spin_unlock(m);
4943                         return TRUE;
4944                 }
4945                 loops++;
4946                 if (loops >= 16)
4947                         break;
4948         }
4949         vm_page_spin_unlock(m);
4950         return (FALSE);
4951 }
4952
4953 /*
4954  * Remove all pages from specified address space this aids process exit
4955  * speeds.  Also, this code may be special cased for the current process
4956  * only.
4957  */
4958 void
4959 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4960 {
4961         pmap_remove_noinval(pmap, sva, eva);
4962         cpu_invltlb();
4963 }
4964
4965 /*
4966  * pmap_testbit tests bits in pte's note that the testbit/clearbit
4967  * routines are inline, and a lot of things compile-time evaluate.
4968  */
4969 static
4970 boolean_t
4971 pmap_testbit(vm_page_t m, int bit)
4972 {
4973         pv_entry_t pv;
4974         pt_entry_t *pte;
4975         pmap_t pmap;
4976
4977         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4978                 return FALSE;
4979
4980         if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4981                 return FALSE;
4982         vm_page_spin_lock(m);
4983         if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4984                 vm_page_spin_unlock(m);
4985                 return FALSE;
4986         }
4987
4988         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4989
4990 #if defined(PMAP_DIAGNOSTIC)
4991                 if (pv->pv_pmap == NULL) {
4992                         kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4993                             pv->pv_pindex);
4994                         continue;
4995                 }
4996 #endif
4997                 pmap = pv->pv_pmap;
4998
4999                 /*
5000                  * If the bit being tested is the modified bit, then
5001                  * mark clean_map and ptes as never
5002                  * modified.
5003                  *
5004                  * WARNING!  Because we do not lock the pv, *pte can be in a
5005                  *           state of flux.  Despite this the value of *pte
5006                  *           will still be related to the vm_page in some way
5007                  *           because the pv cannot be destroyed as long as we
5008                  *           hold the vm_page spin lock.
5009                  */
5010                 if (bit == PG_A_IDX || bit == PG_M_IDX) {
5011                                 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5012                         if (!pmap_track_modified(pv->pv_pindex))
5013                                 continue;
5014                 }
5015
5016                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5017                 if (*pte & pmap->pmap_bits[bit]) {
5018                         vm_page_spin_unlock(m);
5019                         return TRUE;
5020                 }
5021         }
5022         vm_page_spin_unlock(m);
5023         return (FALSE);
5024 }
5025
5026 /*
5027  * This routine is used to modify bits in ptes.  Only one bit should be
5028  * specified.  PG_RW requires special handling.
5029  *
5030  * Caller must NOT hold any spin locks
5031  */
5032 static __inline
5033 void
5034 pmap_clearbit(vm_page_t m, int bit_index)
5035 {
5036         pv_entry_t pv;
5037         pt_entry_t *pte;
5038         pt_entry_t pbits;
5039         pmap_t pmap;
5040
5041         if (bit_index == PG_RW_IDX)
5042                 vm_page_flag_clear(m, PG_WRITEABLE);
5043         if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5044                 return;
5045         }
5046
5047         /*
5048          * PG_M or PG_A case
5049          *
5050          * Loop over all current mappings setting/clearing as appropos If
5051          * setting RO do we need to clear the VAC?
5052          *
5053          * NOTE: When clearing PG_M we could also (not implemented) drop
5054          *       through to the PG_RW code and clear PG_RW too, forcing
5055          *       a fault on write to redetect PG_M for virtual kernels, but
5056          *       it isn't necessary since virtual kernels invalidate the
5057          *       pte when they clear the VPTE_M bit in their virtual page
5058          *       tables.
5059          *
5060          * NOTE: Does not re-dirty the page when clearing only PG_M.
5061          *
5062          * NOTE: Because we do not lock the pv, *pte can be in a state of
5063          *       flux.  Despite this the value of *pte is still somewhat
5064          *       related while we hold the vm_page spin lock.
5065          *
5066          *       *pte can be zero due to this race.  Since we are clearing
5067          *       bits we basically do no harm when this race  ccurs.
5068          */
5069         if (bit_index != PG_RW_IDX) {
5070                 vm_page_spin_lock(m);
5071                 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5072 #if defined(PMAP_DIAGNOSTIC)
5073                         if (pv->pv_pmap == NULL) {
5074                                 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5075                                     pv->pv_pindex);
5076                                 continue;
5077                         }
5078 #endif
5079                         pmap = pv->pv_pmap;
5080                         pte = pmap_pte_quick(pv->pv_pmap,
5081                                              pv->pv_pindex << PAGE_SHIFT);
5082                         pbits = *pte;
5083                         if (pbits & pmap->pmap_bits[bit_index])
5084                                 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5085                 }
5086                 vm_page_spin_unlock(m);
5087                 return;
5088         }
5089
5090         /*
5091          * Clear PG_RW.  Also clears PG_M and marks the page dirty if PG_M
5092          * was set.
5093          */
5094 restart:
5095         vm_page_spin_lock(m);
5096         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5097                 /*
5098                  * don't write protect pager mappings
5099                  */
5100                 if (!pmap_track_modified(pv->pv_pindex))
5101                         continue;
5102
5103 #if defined(PMAP_DIAGNOSTIC)
5104                 if (pv->pv_pmap == NULL) {
5105                         kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5106                             pv->pv_pindex);
5107                         continue;
5108                 }
5109 #endif
5110                 pmap = pv->pv_pmap;
5111                 /*
5112                  * Skip pages which do not have PG_RW set.
5113                  */
5114                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5115                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5116                         continue;
5117
5118                 /*
5119                  * Lock the PV
5120                  */
5121                 if (pv_hold_try(pv)) {
5122                         vm_page_spin_unlock(m);
5123                 } else {
5124                         vm_page_spin_unlock(m);
5125                         pv_lock(pv);    /* held, now do a blocking lock */
5126                 }
5127                 if (pv->pv_pmap != pmap || pv->pv_m != m) {
5128                         pv_put(pv);     /* and release */
5129                         goto restart;   /* anything could have happened */
5130                 }
5131                 KKASSERT(pv->pv_pmap == pmap);
5132                 for (;;) {
5133                         pt_entry_t nbits;
5134
5135                         pbits = *pte;
5136                         cpu_ccfence();
5137                         nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5138                                           pmap->pmap_bits[PG_M_IDX]);
5139                         if (pmap_inval_smp_cmpset(pmap,
5140                                      ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5141                                      pte, pbits, nbits)) {
5142                                 break;
5143                         }
5144                         cpu_pause();
5145                 }
5146                 vm_page_spin_lock(m);
5147
5148                 /*
5149                  * If PG_M was found to be set while we were clearing PG_RW
5150                  * we also clear PG_M (done above) and mark the page dirty.
5151                  * Callers expect this behavior.
5152                  */
5153                 if (pbits & pmap->pmap_bits[PG_M_IDX])
5154                         vm_page_dirty(m);
5155                 pv_put(pv);
5156         }
5157         vm_page_spin_unlock(m);
5158 }
5159
5160 /*
5161  * Lower the permission for all mappings to a given page.
5162  *
5163  * Page must be busied by caller.  Because page is busied by caller this
5164  * should not be able to race a pmap_enter().
5165  */
5166 void
5167 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5168 {
5169         /* JG NX support? */
5170         if ((prot & VM_PROT_WRITE) == 0) {
5171                 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5172                         /*
5173                          * NOTE: pmap_clearbit(.. PG_RW) also clears
5174                          *       the PG_WRITEABLE flag in (m).
5175                          */
5176                         pmap_clearbit(m, PG_RW_IDX);
5177                 } else {
5178                         pmap_remove_all(m);
5179                 }
5180         }
5181 }
5182
5183 vm_paddr_t
5184 pmap_phys_address(vm_pindex_t ppn)
5185 {
5186         return (x86_64_ptob(ppn));
5187 }
5188
5189 /*
5190  * Return a count of reference bits for a page, clearing those bits.
5191  * It is not necessary for every reference bit to be cleared, but it
5192  * is necessary that 0 only be returned when there are truly no
5193  * reference bits set.
5194  *
5195  * XXX: The exact number of bits to check and clear is a matter that
5196  * should be tested and standardized at some point in the future for
5197  * optimal aging of shared pages.
5198  *
5199  * This routine may not block.
5200  */
5201 int
5202 pmap_ts_referenced(vm_page_t m)
5203 {
5204         pv_entry_t pv;
5205         pt_entry_t *pte;
5206         pmap_t pmap;
5207         int rtval = 0;
5208
5209         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5210                 return (rtval);
5211
5212         vm_page_spin_lock(m);
5213         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5214                 if (!pmap_track_modified(pv->pv_pindex))
5215                         continue;
5216                 pmap = pv->pv_pmap;
5217                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5218                 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5219                         atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5220                         rtval++;
5221                         if (rtval > 4)
5222                                 break;
5223                 }
5224         }
5225         vm_page_spin_unlock(m);
5226         return (rtval);
5227 }
5228
5229 /*
5230  *      pmap_is_modified:
5231  *
5232  *      Return whether or not the specified physical page was modified
5233  *      in any physical maps.
5234  */
5235 boolean_t
5236 pmap_is_modified(vm_page_t m)
5237 {
5238         boolean_t res;
5239
5240         res = pmap_testbit(m, PG_M_IDX);
5241         return (res);
5242 }
5243
5244 /*
5245  *      Clear the modify bits on the specified physical page.
5246  */
5247 void
5248 pmap_clear_modify(vm_page_t m)
5249 {
5250         pmap_clearbit(m, PG_M_IDX);
5251 }
5252
5253 /*
5254  *      pmap_clear_reference:
5255  *
5256  *      Clear the reference bit on the specified physical page.
5257  */
5258 void
5259 pmap_clear_reference(vm_page_t m)
5260 {
5261         pmap_clearbit(m, PG_A_IDX);
5262 }
5263
5264 /*
5265  * Miscellaneous support routines follow
5266  */
5267
5268 static
5269 void
5270 i386_protection_init(void)
5271 {
5272         int *kp, prot;
5273
5274         /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit  */
5275         kp = protection_codes;
5276         for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5277                 switch (prot) {
5278                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5279                         /*
5280                          * Read access is also 0. There isn't any execute bit,
5281                          * so just make it readable.
5282                          */
5283                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5284                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5285                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5286                         *kp++ = 0;
5287                         break;
5288                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5289                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5290                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5291                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5292                         *kp++ = pmap_bits_default[PG_RW_IDX];
5293                         break;
5294                 }
5295         }
5296 }
5297
5298 /*
5299  * Map a set of physical memory pages into the kernel virtual
5300  * address space. Return a pointer to where it is mapped. This
5301  * routine is intended to be used for mapping device memory,
5302  * NOT real memory.
5303  *
5304  * NOTE: We can't use pgeflag unless we invalidate the pages one at
5305  *       a time.
5306  *
5307  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5308  *       work whether the cpu supports PAT or not.  The remaining PAT
5309  *       attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5310  *       supports PAT.
5311  */
5312 void *
5313 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5314 {
5315         return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5316 }
5317
5318 void *
5319 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5320 {
5321         return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5322 }
5323
5324 void *
5325 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5326 {
5327         return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5328 }
5329
5330 /*
5331  * Map a set of physical memory pages into the kernel virtual
5332  * address space. Return a pointer to where it is mapped. This
5333  * routine is intended to be used for mapping device memory,
5334  * NOT real memory.
5335  */
5336 void *
5337 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5338 {
5339         vm_offset_t va, tmpva, offset;
5340         pt_entry_t *pte;
5341         vm_size_t tmpsize;
5342
5343         offset = pa & PAGE_MASK;
5344         size = roundup(offset + size, PAGE_SIZE);
5345
5346         va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5347         if (va == 0)
5348                 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5349
5350         pa = pa & ~PAGE_MASK;
5351         for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5352                 pte = vtopte(tmpva);
5353                 *pte = pa |
5354                     kernel_pmap.pmap_bits[PG_RW_IDX] |
5355                     kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5356                     kernel_pmap.pmap_cache_bits[mode];
5357                 tmpsize -= PAGE_SIZE;
5358                 tmpva += PAGE_SIZE;
5359                 pa += PAGE_SIZE;
5360         }
5361         pmap_invalidate_range(&kernel_pmap, va, va + size);
5362         pmap_invalidate_cache_range(va, va + size);
5363
5364         return ((void *)(va + offset));
5365 }
5366
5367 void
5368 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5369 {
5370         vm_offset_t base, offset;
5371
5372         base = va & ~PAGE_MASK;
5373         offset = va & PAGE_MASK;
5374         size = roundup(offset + size, PAGE_SIZE);
5375         pmap_qremove(va, size >> PAGE_SHIFT);
5376         kmem_free(&kernel_map, base, size);
5377 }
5378
5379 /*
5380  * Sets the memory attribute for the specified page.
5381  */
5382 void
5383 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5384 {
5385
5386     m->pat_mode = ma;
5387
5388     /*
5389      * If "m" is a normal page, update its direct mapping.  This update
5390      * can be relied upon to perform any cache operations that are
5391      * required for data coherence.
5392      */
5393     if ((m->flags & PG_FICTITIOUS) == 0)
5394         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5395 }
5396
5397 /*
5398  * Change the PAT attribute on an existing kernel memory map.  Caller
5399  * must ensure that the virtual memory in question is not accessed
5400  * during the adjustment.
5401  */
5402 void
5403 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5404 {
5405         pt_entry_t *pte;
5406         vm_offset_t base;
5407         int changed = 0;
5408
5409         if (va == 0)
5410                 panic("pmap_change_attr: va is NULL");
5411         base = trunc_page(va);
5412
5413         while (count) {
5414                 pte = vtopte(va);
5415                 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5416                        kernel_pmap.pmap_cache_bits[mode];
5417                 --count;
5418                 va += PAGE_SIZE;
5419         }
5420
5421         changed = 1;    /* XXX: not optimal */
5422
5423         /*
5424          * Flush CPU caches if required to make sure any data isn't cached that
5425          * shouldn't be, etc.
5426          */
5427         if (changed) {
5428                 pmap_invalidate_range(&kernel_pmap, base, va);
5429                 pmap_invalidate_cache_range(base, va);
5430         }
5431 }
5432
5433 /*
5434  * perform the pmap work for mincore
5435  */
5436 int
5437 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5438 {
5439         pt_entry_t *ptep, pte;
5440         vm_page_t m;
5441         int val = 0;
5442
5443         lwkt_gettoken(&pmap->pm_token);
5444         ptep = pmap_pte(pmap, addr);
5445
5446         if (ptep && (pte = *ptep) != 0) {
5447                 vm_offset_t pa;
5448
5449                 val = MINCORE_INCORE;
5450                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5451                         goto done;
5452
5453                 pa = pte & PG_FRAME;
5454
5455                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5456                         m = NULL;
5457                 else
5458                         m = PHYS_TO_VM_PAGE(pa);
5459
5460                 /*
5461                  * Modified by us
5462                  */
5463                 if (pte & pmap->pmap_bits[PG_M_IDX])
5464                         val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5465                 /*
5466                  * Modified by someone
5467                  */
5468                 else if (m && (m->dirty || pmap_is_modified(m)))
5469                         val |= MINCORE_MODIFIED_OTHER;
5470                 /*
5471                  * Referenced by us
5472                  */
5473                 if (pte & pmap->pmap_bits[PG_A_IDX])
5474                         val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5475
5476                 /*
5477                  * Referenced by someone
5478                  */
5479                 else if (m && ((m->flags & PG_REFERENCED) ||
5480                                 pmap_ts_referenced(m))) {
5481                         val |= MINCORE_REFERENCED_OTHER;
5482                         vm_page_flag_set(m, PG_REFERENCED);
5483                 }
5484         }
5485 done:
5486         lwkt_reltoken(&pmap->pm_token);
5487
5488         return val;
5489 }
5490
5491 /*
5492  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
5493  * vmspace will be ref'd and the old one will be deref'd.
5494  *
5495  * The vmspace for all lwps associated with the process will be adjusted
5496  * and cr3 will be reloaded if any lwp is the current lwp.
5497  *
5498  * The process must hold the vmspace->vm_map.token for oldvm and newvm
5499  */
5500 void
5501 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5502 {
5503         struct vmspace *oldvm;
5504         struct lwp *lp;
5505
5506         oldvm = p->p_vmspace;
5507         if (oldvm != newvm) {
5508                 if (adjrefs)
5509                         vmspace_ref(newvm);
5510                 p->p_vmspace = newvm;
5511                 KKASSERT(p->p_nthreads == 1);
5512                 lp = RB_ROOT(&p->p_lwp_tree);
5513                 pmap_setlwpvm(lp, newvm);
5514                 if (adjrefs)
5515                         vmspace_rel(oldvm);
5516         }
5517 }
5518
5519 /*
5520  * Set the vmspace for a LWP.  The vmspace is almost universally set the
5521  * same as the process vmspace, but virtual kernels need to swap out contexts
5522  * on a per-lwp basis.
5523  *
5524  * Caller does not necessarily hold any vmspace tokens.  Caller must control
5525  * the lwp (typically be in the context of the lwp).  We use a critical
5526  * section to protect against statclock and hardclock (statistics collection).
5527  */
5528 void
5529 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5530 {
5531         struct vmspace *oldvm;
5532         struct pmap *pmap;
5533
5534         oldvm = lp->lwp_vmspace;
5535
5536         if (oldvm != newvm) {
5537                 crit_enter();
5538                 lp->lwp_vmspace = newvm;
5539                 if (curthread->td_lwp == lp) {
5540                         pmap = vmspace_pmap(newvm);
5541                         ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5542                         if (pmap->pm_active_lock & CPULOCK_EXCL)
5543                                 pmap_interlock_wait(newvm);
5544 #if defined(SWTCH_OPTIM_STATS)
5545                         tlb_flush_count++;
5546 #endif
5547                         if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5548                                 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5549                         } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5550                                 curthread->td_pcb->pcb_cr3 = KPML4phys;
5551                         } else {
5552                                 panic("pmap_setlwpvm: unknown pmap type\n");
5553                         }
5554                         load_cr3(curthread->td_pcb->pcb_cr3);
5555                         pmap = vmspace_pmap(oldvm);
5556                         ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5557                                                mycpu->gd_cpuid);
5558                 }
5559                 crit_exit();
5560         }
5561 }
5562
5563 /*
5564  * Called when switching to a locked pmap, used to interlock against pmaps
5565  * undergoing modifications to prevent us from activating the MMU for the
5566  * target pmap until all such modifications have completed.  We have to do
5567  * this because the thread making the modifications has already set up its
5568  * SMP synchronization mask.
5569  *
5570  * This function cannot sleep!
5571  *
5572  * No requirements.
5573  */
5574 void
5575 pmap_interlock_wait(struct vmspace *vm)
5576 {
5577         struct pmap *pmap = &vm->vm_pmap;
5578
5579         if (pmap->pm_active_lock & CPULOCK_EXCL) {
5580                 crit_enter();
5581                 KKASSERT(curthread->td_critcount >= 2);
5582                 DEBUG_PUSH_INFO("pmap_interlock_wait");
5583                 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5584                         cpu_ccfence();
5585                         lwkt_process_ipiq();
5586                 }
5587                 DEBUG_POP_INFO();
5588                 crit_exit();
5589         }
5590 }
5591
5592 vm_offset_t
5593 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5594 {
5595
5596         if ((obj == NULL) || (size < NBPDR) ||
5597             ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5598                 return addr;
5599         }
5600
5601         addr = roundup2(addr, NBPDR);
5602         return addr;
5603 }
5604
5605 /*
5606  * Used by kmalloc/kfree, page already exists at va
5607  */
5608 vm_page_t
5609 pmap_kvtom(vm_offset_t va)
5610 {
5611         pt_entry_t *ptep = vtopte(va);
5612
5613         KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5614         return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5615 }
5616
5617 /*
5618  * Initialize machine-specific shared page directory support.  This
5619  * is executed when a VM object is created.
5620  */
5621 void
5622 pmap_object_init(vm_object_t object)
5623 {
5624         object->md.pmap_rw = NULL;
5625         object->md.pmap_ro = NULL;
5626 }
5627
5628 /*
5629  * Clean up machine-specific shared page directory support.  This
5630  * is executed when a VM object is destroyed.
5631  */
5632 void
5633 pmap_object_free(vm_object_t object)
5634 {
5635         pmap_t pmap;
5636
5637         if ((pmap = object->md.pmap_rw) != NULL) {
5638                 object->md.pmap_rw = NULL;
5639                 pmap_remove_noinval(pmap,
5640                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5641                 CPUMASK_ASSZERO(pmap->pm_active);
5642                 pmap_release(pmap);
5643                 pmap_puninit(pmap);
5644                 kfree(pmap, M_OBJPMAP);
5645         }
5646         if ((pmap = object->md.pmap_ro) != NULL) {
5647                 object->md.pmap_ro = NULL;
5648                 pmap_remove_noinval(pmap,
5649                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5650                 CPUMASK_ASSZERO(pmap->pm_active);
5651                 pmap_release(pmap);
5652                 pmap_puninit(pmap);
5653                 kfree(pmap, M_OBJPMAP);
5654         }
5655 }
5656
5657 /*
5658  * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
5659  * VM page and issue a pginfo->callback.
5660  *
5661  * We are expected to dispose of any non-NULL pte_pv.
5662  */
5663 static
5664 void
5665 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
5666                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
5667                       vm_offset_t va, pt_entry_t *ptep, void *arg)
5668 {
5669         struct pmap_pgscan_info *pginfo = arg;
5670         vm_page_t m;
5671
5672         if (pte_pv) {
5673                 /*
5674                  * Try to busy the page while we hold the pte_pv locked.
5675                  */
5676                 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
5677                 if (vm_page_busy_try(m, TRUE) == 0) {
5678                         if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
5679                                 /*
5680                                  * The callback is issued with the pte_pv
5681                                  * unlocked and put away, and the pt_pv
5682                                  * unlocked.
5683                                  */
5684                                 pv_put(pte_pv);
5685                                 if (pt_pv)
5686                                         pv_unlock(pt_pv);
5687                                 if (pginfo->callback(pginfo, va, m) < 0)
5688                                         info->stop = 1;
5689                                 if (pt_pv)
5690                                         pv_lock(pt_pv);
5691                         } else {
5692                                 vm_page_wakeup(m);
5693                                 pv_put(pte_pv);
5694                         }
5695                 } else {
5696                         ++pginfo->busycount;
5697                         pv_put(pte_pv);
5698                 }
5699         } else if (sharept) {
5700                 /* shared page table */
5701         } else {
5702                 /* else unmanaged page */
5703         }
5704 }
5705
5706 void
5707 pmap_pgscan(struct pmap_pgscan_info *pginfo)
5708 {
5709         struct pmap_scan_info info;
5710
5711         pginfo->offset = pginfo->beg_addr;
5712         info.pmap = pginfo->pmap;
5713         info.sva = pginfo->beg_addr;
5714         info.eva = pginfo->end_addr;
5715         info.func = pmap_pgscan_callback;
5716         info.arg = pginfo;
5717         pmap_scan(&info, 0);
5718         if (info.stop == 0)
5719                 pginfo->offset = pginfo->end_addr;
5720 }