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 JG
  49 #include "opt_disable_pse.h"
  50 #include "opt_pmap.h"
  51 #endif
  52 #include "opt_msgbuf.h"
  53
  54 #include <sys/param.h>
  55 #include <sys/kernel.h>
  56 #include <sys/proc.h>
  57 #include <sys/msgbuf.h>
  58 #include <sys/vmmeter.h>
  59 #include <sys/mman.h>
  60 #include <sys/systm.h>
  61
  62 #include <vm/vm.h>
  63 #include <vm/vm_param.h>
  64 #include <sys/sysctl.h>
  65 #include <sys/lock.h>
  66 #include <vm/vm_kern.h>
  67 #include <vm/vm_page.h>
  68 #include <vm/vm_map.h>
  69 #include <vm/vm_object.h>
  70 #include <vm/vm_extern.h>
  71 #include <vm/vm_pageout.h>
  72 #include <vm/vm_pager.h>
  73 #include <vm/vm_zone.h>
  74
  75 #include <sys/user.h>
  76 #include <sys/thread2.h>
  77 #include <sys/sysref2.h>
  78 #include <sys/spinlock2.h>
  79 #include <vm/vm_page2.h>
  80
  81 #include <machine/cputypes.h>
  82 #include <machine/md_var.h>
  83 #include <machine/specialreg.h>
  84 #include <machine/smp.h>
  85 #include <machine_base/apic/apicreg.h>
  86 #include <machine/globaldata.h>
  87 #include <machine/pmap.h>
  88 #include <machine/pmap_inval.h>
  89 #include <machine/inttypes.h>
  90
  91 #include <ddb/ddb.h>
  92
  93 #define PMAP_KEEP_PDIRS
  94 #ifndef PMAP_SHPGPERPROC
  95 #define PMAP_SHPGPERPROC 2000
  96 #endif
  97
  98 #if defined(DIAGNOSTIC)
  99 #define PMAP_DIAGNOSTIC
 100 #endif
 101
 102 #define MINPV 2048
 103
 104 /*
 105  * pmap debugging will report who owns a pv lock when blocking.
 106  */
 107 #ifdef PMAP_DEBUG
 108
 109 #define PMAP_DEBUG_DECL         ,const char *func, int lineno
 110 #define PMAP_DEBUG_ARGS         , __func__, __LINE__
 111 #define PMAP_DEBUG_COPY         , func, lineno
 112
 113 #define pv_get(pmap, pindex)            _pv_get(pmap, pindex            \
 114                                                         PMAP_DEBUG_ARGS)
 115 #define pv_lock(pv)                     _pv_lock(pv                     \
 116                                                         PMAP_DEBUG_ARGS)
 117 #define pv_hold_try(pv)                 _pv_hold_try(pv                 \
 118                                                         PMAP_DEBUG_ARGS)
 119 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp  \
 120                                                         PMAP_DEBUG_ARGS)
 121
 122 #else
 123
 124 #define PMAP_DEBUG_DECL
 125 #define PMAP_DEBUG_ARGS
 126 #define PMAP_DEBUG_COPY
 127
 128 #define pv_get(pmap, pindex)            _pv_get(pmap, pindex)
 129 #define pv_lock(pv)                     _pv_lock(pv)
 130 #define pv_hold_try(pv)                 _pv_hold_try(pv)
 131 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp)
 132
 133 #endif
 134
 135 /*
 136  * Get PDEs and PTEs for user/kernel address space
 137  */
 138 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
 139
 140 #define pmap_pde_v(pmap, pte)           ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 141 #define pmap_pte_w(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
 142 #define pmap_pte_m(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
 143 #define pmap_pte_u(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
 144 #define pmap_pte_v(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
 145
 146 /*
 147  * Given a map and a machine independent protection code,
 148  * convert to a vax protection code.
 149  */
 150 #define pte_prot(m, p)          \
 151         (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
 152 static int protection_codes[PROTECTION_CODES_SIZE];
 153
 154 struct pmap kernel_pmap;
 155 static TAILQ_HEAD(,pmap)        pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
 156
 157 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
 158
 159 vm_paddr_t avail_start;         /* PA of first available physical page */
 160 vm_paddr_t avail_end;           /* PA of last available physical page */
 161 vm_offset_t virtual2_start;     /* cutout free area prior to kernel start */
 162 vm_offset_t virtual2_end;
 163 vm_offset_t virtual_start;      /* VA of first avail page (after kernel bss) */
 164 vm_offset_t virtual_end;        /* VA of last avail page (end of kernel AS) */
 165 vm_offset_t KvaStart;           /* VA start of KVA space */
 166 vm_offset_t KvaEnd;             /* VA end of KVA space (non-inclusive) */
 167 vm_offset_t KvaSize;            /* max size of kernel virtual address space */
 168 static boolean_t pmap_initialized = FALSE;      /* Has pmap_init completed? */
 169 //static int pgeflag;           /* PG_G or-in */
 170 //static int pseflag;           /* PG_PS or-in */
 171 uint64_t PatMsr;
 172
 173 static int ndmpdp;
 174 static vm_paddr_t dmaplimit;
 175 static int nkpt;
 176 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
 177
 178 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE];        /* PAT -> PG_ bits */
 179 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/    /* PAT -> PG_ bits */
 180
 181 static uint64_t KPTbase;
 182 static uint64_t KPTphys;
 183 static uint64_t KPDphys;        /* phys addr of kernel level 2 */
 184 static uint64_t KPDbase;        /* phys addr of kernel level 2 @ KERNBASE */
 185 uint64_t KPDPphys;      /* phys addr of kernel level 3 */
 186 uint64_t KPML4phys;     /* phys addr of kernel level 4 */
 187
 188 static uint64_t DMPDphys;       /* phys addr of direct mapped level 2 */
 189 static uint64_t DMPDPphys;      /* phys addr of direct mapped level 3 */
 190
 191 /*
 192  * Data for the pv entry allocation mechanism
 193  */
 194 static vm_zone_t pvzone;
 195 static struct vm_zone pvzone_store;
 196 static struct vm_object pvzone_obj;
 197 static int pv_entry_max=0, pv_entry_high_water=0;
 198 static int pmap_pagedaemon_waken = 0;
 199 static struct pv_entry *pvinit;
 200
 201 /*
 202  * All those kernel PT submaps that BSD is so fond of
 203  */
 204 pt_entry_t *CMAP1 = NULL, *ptmmap;
 205 caddr_t CADDR1 = NULL, ptvmmap = NULL;
 206 static pt_entry_t *msgbufmap;
 207 struct msgbuf *msgbufp=NULL;
 208
 209 /*
 210  * PMAP default PG_* bits. Needed to be able to add
 211  * EPT/NPT pagetable pmap_bits for the VMM module
 212  */
 213 uint64_t pmap_bits_default[] = {
 214                 REGULAR_PMAP,                                   /* TYPE_IDX             0 */
 215                 X86_PG_V,                                       /* PG_V_IDX             1 */
 216                 X86_PG_RW,                                      /* PG_RW_IDX            2 */
 217                 X86_PG_U,                                       /* PG_U_IDX             3 */
 218                 X86_PG_A,                                       /* PG_A_IDX             4 */
 219                 X86_PG_M,                                       /* PG_M_IDX             5 */
 220                 X86_PG_PS,                                      /* PG_PS_IDX3           6 */
 221                 X86_PG_G,                                       /* PG_G_IDX             7 */
 222                 X86_PG_AVAIL1,                                  /* PG_AVAIL1_IDX        8 */
 223                 X86_PG_AVAIL2,                                  /* PG_AVAIL2_IDX        9 */
 224                 X86_PG_AVAIL3,                                  /* PG_AVAIL3_IDX        10 */
 225                 X86_PG_NC_PWT | X86_PG_NC_PCD,                  /* PG_N_IDX     11 */
 226 };
 227 /*
 228  * Crashdump maps.
 229  */
 230 static pt_entry_t *pt_crashdumpmap;
 231 static caddr_t crashdumpmap;
 232
 233 #ifdef PMAP_DEBUG2
 234 static int pmap_enter_debug = 0;
 235 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
 236     &pmap_enter_debug, 0, "Debug pmap_enter's");
 237 #endif
 238 static int pmap_yield_count = 64;
 239 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
 240     &pmap_yield_count, 0, "Yield during init_pt/release");
 241 static int pmap_mmu_optimize = 0;
 242 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
 243     &pmap_mmu_optimize, 0, "Share page table pages when possible");
 244
 245 #define DISABLE_PSE
 246
 247 /* Standard user access funtions */
 248 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
 249     size_t *lencopied);
 250 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
 251 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
 252 extern int std_fubyte (const void *base);
 253 extern int std_subyte (void *base, int byte);
 254 extern long std_fuword (const void *base);
 255 extern int std_suword (void *base, long word);
 256 extern int std_suword32 (void *base, int word);
 257
 258 static void pv_hold(pv_entry_t pv);
 259 static int _pv_hold_try(pv_entry_t pv
 260                                 PMAP_DEBUG_DECL);
 261 static void pv_drop(pv_entry_t pv);
 262 static void _pv_lock(pv_entry_t pv
 263                                 PMAP_DEBUG_DECL);
 264 static void pv_unlock(pv_entry_t pv);
 265 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
 266                                 PMAP_DEBUG_DECL);
 267 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
 268                                 PMAP_DEBUG_DECL);
 269 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
 270 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
 271 static void pv_put(pv_entry_t pv);
 272 static void pv_free(pv_entry_t pv);
 273 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
 274 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
 275                       pv_entry_t *pvpp);
 276 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
 277                       pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
 278 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
 279                       struct pmap_inval_info *info);
 280 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
 281 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp);
 282
 283 struct pmap_scan_info;
 284 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
 285                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
 286                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 287 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
 288                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
 289                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
 290
 291 static void i386_protection_init (void);
 292 static void create_pagetables(vm_paddr_t *firstaddr);
 293 static void pmap_remove_all (vm_page_t m);
 294 static boolean_t pmap_testbit (vm_page_t m, int bit);
 295
 296 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
 297 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
 298
 299 static void pmap_pinit_defaults(struct pmap *pmap);
 300
 301 static unsigned pdir4mb;
 302
 303 static int
 304 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
 305 {
 306         if (pv1->pv_pindex < pv2->pv_pindex)
 307                 return(-1);
 308         if (pv1->pv_pindex > pv2->pv_pindex)
 309                 return(1);
 310         return(0);
 311 }
 312
 313 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
 314              pv_entry_compare, vm_pindex_t, pv_pindex);
 315
 316 static __inline
 317 void
 318 pmap_page_stats_adding(vm_page_t m)
 319 {
 320         globaldata_t gd = mycpu;
 321
 322         if (TAILQ_EMPTY(&m->md.pv_list)) {
 323                 ++gd->gd_vmtotal.t_arm;
 324         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 325                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 326                 ++gd->gd_vmtotal.t_armshr;
 327                 ++gd->gd_vmtotal.t_avmshr;
 328         } else {
 329                 ++gd->gd_vmtotal.t_avmshr;
 330         }
 331 }
 332
 333 static __inline
 334 void
 335 pmap_page_stats_deleting(vm_page_t m)
 336 {
 337         globaldata_t gd = mycpu;
 338
 339         if (TAILQ_EMPTY(&m->md.pv_list)) {
 340                 --gd->gd_vmtotal.t_arm;
 341         } else if (TAILQ_FIRST(&m->md.pv_list) ==
 342                    TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
 343                 --gd->gd_vmtotal.t_armshr;
 344                 --gd->gd_vmtotal.t_avmshr;
 345         } else {
 346                 --gd->gd_vmtotal.t_avmshr;
 347         }
 348 }
 349
 350 /*
 351  * Move the kernel virtual free pointer to the next
 352  * 2MB.  This is used to help improve performance
 353  * by using a large (2MB) page for much of the kernel
 354  * (.text, .data, .bss)
 355  */
 356 static
 357 vm_offset_t
 358 pmap_kmem_choose(vm_offset_t addr)
 359 {
 360         vm_offset_t newaddr = addr;
 361
 362         newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
 363         return newaddr;
 364 }
 365
 366 /*
 367  * pmap_pte_quick:
 368  *
 369  *      Super fast pmap_pte routine best used when scanning the pv lists.
 370  *      This eliminates many course-grained invltlb calls.  Note that many of
 371  *      the pv list scans are across different pmaps and it is very wasteful
 372  *      to do an entire invltlb when checking a single mapping.
 373  */
 374 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
 375
 376 static
 377 pt_entry_t *
 378 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
 379 {
 380         return pmap_pte(pmap, va);
 381 }
 382
 383 /*
 384  * Returns the pindex of a page table entry (representing a terminal page).
 385  * There are NUPTE_TOTAL page table entries possible (a huge number)
 386  *
 387  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
 388  * We want to properly translate negative KVAs.
 389  */
 390 static __inline
 391 vm_pindex_t
 392 pmap_pte_pindex(vm_offset_t va)
 393 {
 394         return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
 395 }
 396
 397 /*
 398  * Returns the pindex of a page table.
 399  */
 400 static __inline
 401 vm_pindex_t
 402 pmap_pt_pindex(vm_offset_t va)
 403 {
 404         return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
 405 }
 406
 407 /*
 408  * Returns the pindex of a page directory.
 409  */
 410 static __inline
 411 vm_pindex_t
 412 pmap_pd_pindex(vm_offset_t va)
 413 {
 414         return (NUPTE_TOTAL + NUPT_TOTAL +
 415                 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
 416 }
 417
 418 static __inline
 419 vm_pindex_t
 420 pmap_pdp_pindex(vm_offset_t va)
 421 {
 422         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
 423                 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
 424 }
 425
 426 static __inline
 427 vm_pindex_t
 428 pmap_pml4_pindex(void)
 429 {
 430         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
 431 }
 432
 433 /*
 434  * Return various clipped indexes for a given VA
 435  *
 436  * Returns the index of a pte in a page table, representing a terminal
 437  * page.
 438  */
 439 static __inline
 440 vm_pindex_t
 441 pmap_pte_index(vm_offset_t va)
 442 {
 443         return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
 444 }
 445
 446 /*
 447  * Returns the index of a pt in a page directory, representing a page
 448  * table.
 449  */
 450 static __inline
 451 vm_pindex_t
 452 pmap_pt_index(vm_offset_t va)
 453 {
 454         return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
 455 }
 456
 457 /*
 458  * Returns the index of a pd in a page directory page, representing a page
 459  * directory.
 460  */
 461 static __inline
 462 vm_pindex_t
 463 pmap_pd_index(vm_offset_t va)
 464 {
 465         return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
 466 }
 467
 468 /*
 469  * Returns the index of a pdp in the pml4 table, representing a page
 470  * directory page.
 471  */
 472 static __inline
 473 vm_pindex_t
 474 pmap_pdp_index(vm_offset_t va)
 475 {
 476         return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
 477 }
 478
 479 /*
 480  * Generic procedure to index a pte from a pt, pd, or pdp.
 481  *
 482  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
 483  *       a page table page index but is instead of PV lookup index.
 484  */
 485 static
 486 void *
 487 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
 488 {
 489         pt_entry_t *pte;
 490
 491         pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
 492         return(&pte[pindex]);
 493 }
 494
 495 /*
 496  * Return pointer to PDP slot in the PML4
 497  */
 498 static __inline
 499 pml4_entry_t *
 500 pmap_pdp(pmap_t pmap, vm_offset_t va)
 501 {
 502         return (&pmap->pm_pml4[pmap_pdp_index(va)]);
 503 }
 504
 505 /*
 506  * Return pointer to PD slot in the PDP given a pointer to the PDP
 507  */
 508 static __inline
 509 pdp_entry_t *
 510 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
 511 {
 512         pdp_entry_t *pd;
 513
 514         pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
 515         return (&pd[pmap_pd_index(va)]);
 516 }
 517
 518 /*
 519  * Return pointer to PD slot in the PDP.
 520  */
 521 static __inline
 522 pdp_entry_t *
 523 pmap_pd(pmap_t pmap, vm_offset_t va)
 524 {
 525         pml4_entry_t *pdp;
 526
 527         pdp = pmap_pdp(pmap, va);
 528         if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
 529                 return NULL;
 530         return (pmap_pdp_to_pd(*pdp, va));
 531 }
 532
 533 /*
 534  * Return pointer to PT slot in the PD given a pointer to the PD
 535  */
 536 static __inline
 537 pd_entry_t *
 538 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
 539 {
 540         pd_entry_t *pt;
 541
 542         pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
 543         return (&pt[pmap_pt_index(va)]);
 544 }
 545
 546 /*
 547  * Return pointer to PT slot in the PD
 548  *
 549  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
 550  *                   so we cannot lookup the PD via the PDP.  Instead we
 551  *                   must look it up via the pmap.
 552  */
 553 static __inline
 554 pd_entry_t *
 555 pmap_pt(pmap_t pmap, vm_offset_t va)
 556 {
 557         pdp_entry_t *pd;
 558         pv_entry_t pv;
 559         vm_pindex_t pd_pindex;
 560
 561         if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
 562                 pd_pindex = pmap_pd_pindex(va);
 563                 spin_lock(&pmap->pm_spin);
 564                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
 565                 spin_unlock(&pmap->pm_spin);
 566                 if (pv == NULL || pv->pv_m == NULL)
 567                         return NULL;
 568                 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
 569         } else {
 570                 pd = pmap_pd(pmap, va);
 571                 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
 572                          return NULL;
 573                 return (pmap_pd_to_pt(*pd, va));
 574         }
 575 }
 576
 577 /*
 578  * Return pointer to PTE slot in the PT given a pointer to the PT
 579  */
 580 static __inline
 581 pt_entry_t *
 582 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
 583 {
 584         pt_entry_t *pte;
 585
 586         pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
 587         return (&pte[pmap_pte_index(va)]);
 588 }
 589
 590 /*
 591  * Return pointer to PTE slot in the PT
 592  */
 593 static __inline
 594 pt_entry_t *
 595 pmap_pte(pmap_t pmap, vm_offset_t va)
 596 {
 597         pd_entry_t *pt;
 598
 599         pt = pmap_pt(pmap, va);
 600         if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
 601                  return NULL;
 602         if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
 603                 return ((pt_entry_t *)pt);
 604         return (pmap_pt_to_pte(*pt, va));
 605 }
 606
 607 /*
 608  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
 609  * the PT layer.  This will speed up core pmap operations considerably.
 610  *
 611  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
 612  *       must be in a known associated state (typically by being locked when
 613  *       the pmap spinlock isn't held).  We allow the race for that case.
 614  */
 615 static __inline
 616 void
 617 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
 618 {
 619         if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
 620                 pv->pv_pmap->pm_pvhint = pv;
 621 }
 622
 623
 624 /*
 625  * Return address of PT slot in PD (KVM only)
 626  *
 627  * Cannot be used for user page tables because it might interfere with
 628  * the shared page-table-page optimization (pmap_mmu_optimize).
 629  */
 630 static __inline
 631 pd_entry_t *
 632 vtopt(vm_offset_t va)
 633 {
 634         uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
 635                                   NPML4EPGSHIFT)) - 1);
 636
 637         return (PDmap + ((va >> PDRSHIFT) & mask));
 638 }
 639
 640 /*
 641  * KVM - return address of PTE slot in PT
 642  */
 643 static __inline
 644 pt_entry_t *
 645 vtopte(vm_offset_t va)
 646 {
 647         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
 648                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
 649
 650         return (PTmap + ((va >> PAGE_SHIFT) & mask));
 651 }
 652
 653 static uint64_t
 654 allocpages(vm_paddr_t *firstaddr, long n)
 655 {
 656         uint64_t ret;
 657
 658         ret = *firstaddr;
 659         bzero((void *)ret, n * PAGE_SIZE);
 660         *firstaddr += n * PAGE_SIZE;
 661         return (ret);
 662 }
 663
 664 static
 665 void
 666 create_pagetables(vm_paddr_t *firstaddr)
 667 {
 668         long i;         /* must be 64 bits */
 669         long nkpt_base;
 670         long nkpt_phys;
 671         int j;
 672
 673         /*
 674          * We are running (mostly) V=P at this point
 675          *
 676          * Calculate NKPT - number of kernel page tables.  We have to
 677          * accomodoate prealloction of the vm_page_array, dump bitmap,
 678          * MSGBUF_SIZE, and other stuff.  Be generous.
 679          *
 680          * Maxmem is in pages.
 681          *
 682          * ndmpdp is the number of 1GB pages we wish to map.
 683          */
 684         ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
 685         if (ndmpdp < 4)         /* Minimum 4GB of dirmap */
 686                 ndmpdp = 4;
 687         KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
 688
 689         /*
 690          * Starting at the beginning of kvm (not KERNBASE).
 691          */
 692         nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
 693         nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
 694         nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
 695                        ndmpdp) + 511) / 512;
 696         nkpt_phys += 128;
 697
 698         /*
 699          * Starting at KERNBASE - map 2G worth of page table pages.
 700          * KERNBASE is offset -2G from the end of kvm.
 701          */
 702         nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
 703
 704         /*
 705          * Allocate pages
 706          */
 707         KPTbase = allocpages(firstaddr, nkpt_base);
 708         KPTphys = allocpages(firstaddr, nkpt_phys);
 709         KPML4phys = allocpages(firstaddr, 1);
 710         KPDPphys = allocpages(firstaddr, NKPML4E);
 711         KPDphys = allocpages(firstaddr, NKPDPE);
 712
 713         /*
 714          * Calculate the page directory base for KERNBASE,
 715          * that is where we start populating the page table pages.
 716          * Basically this is the end - 2.
 717          */
 718         KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
 719
 720         DMPDPphys = allocpages(firstaddr, NDMPML4E);
 721         if ((amd_feature & AMDID_PAGE1GB) == 0)
 722                 DMPDphys = allocpages(firstaddr, ndmpdp);
 723         dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
 724
 725         /*
 726          * Fill in the underlying page table pages for the area around
 727          * KERNBASE.  This remaps low physical memory to KERNBASE.
 728          *
 729          * Read-only from zero to physfree
 730          * XXX not fully used, underneath 2M pages
 731          */
 732         for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
 733                 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
 734                 ((pt_entry_t *)KPTbase)[i] |=
 735                     pmap_bits_default[PG_RW_IDX] |
 736                     pmap_bits_default[PG_V_IDX] |
 737                     pmap_bits_default[PG_G_IDX];
 738         }
 739
 740         /*
 741          * Now map the initial kernel page tables.  One block of page
 742          * tables is placed at the beginning of kernel virtual memory,
 743          * and another block is placed at KERNBASE to map the kernel binary,
 744          * data, bss, and initial pre-allocations.
 745          */
 746         for (i = 0; i < nkpt_base; i++) {
 747                 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
 748                 ((pd_entry_t *)KPDbase)[i] |=
 749                     pmap_bits_default[PG_RW_IDX] |
 750                     pmap_bits_default[PG_V_IDX];
 751         }
 752         for (i = 0; i < nkpt_phys; i++) {
 753                 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
 754                 ((pd_entry_t *)KPDphys)[i] |=
 755                     pmap_bits_default[PG_RW_IDX] |
 756                     pmap_bits_default[PG_V_IDX];
 757         }
 758
 759         /*
 760          * Map from zero to end of allocations using 2M pages as an
 761          * optimization.  This will bypass some of the KPTBase pages
 762          * above in the KERNBASE area.
 763          */
 764         for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
 765                 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
 766                 ((pd_entry_t *)KPDbase)[i] |=
 767                     pmap_bits_default[PG_RW_IDX] |
 768                     pmap_bits_default[PG_V_IDX] |
 769                     pmap_bits_default[PG_PS_IDX] |
 770                     pmap_bits_default[PG_G_IDX];
 771         }
 772
 773         /*
 774          * And connect up the PD to the PDP.  The kernel pmap is expected
 775          * to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
 776          */
 777         for (i = 0; i < NKPDPE; i++) {
 778                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
 779                                 KPDphys + (i << PAGE_SHIFT);
 780                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
 781                     pmap_bits_default[PG_RW_IDX] |
 782                     pmap_bits_default[PG_V_IDX] |
 783                     pmap_bits_default[PG_U_IDX];
 784         }
 785
 786         /*
 787          * Now set up the direct map space using either 2MB or 1GB pages
 788          * Preset PG_M and PG_A because demotion expects it.
 789          *
 790          * When filling in entries in the PD pages make sure any excess
 791          * entries are set to zero as we allocated enough PD pages
 792          */
 793         if ((amd_feature & AMDID_PAGE1GB) == 0) {
 794                 for (i = 0; i < NPDEPG * ndmpdp; i++) {
 795                         ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
 796                         ((pd_entry_t *)DMPDphys)[i] |=
 797                             pmap_bits_default[PG_RW_IDX] |
 798                             pmap_bits_default[PG_V_IDX] |
 799                             pmap_bits_default[PG_PS_IDX] |
 800                             pmap_bits_default[PG_G_IDX] |
 801                             pmap_bits_default[PG_M_IDX] |
 802                             pmap_bits_default[PG_A_IDX];
 803                 }
 804
 805                 /*
 806                  * And the direct map space's PDP
 807                  */
 808                 for (i = 0; i < ndmpdp; i++) {
 809                         ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
 810                                                         (i << PAGE_SHIFT);
 811                         ((pdp_entry_t *)DMPDPphys)[i] |=
 812                             pmap_bits_default[PG_RW_IDX] |
 813                             pmap_bits_default[PG_V_IDX] |
 814                             pmap_bits_default[PG_U_IDX];
 815                 }
 816         } else {
 817                 for (i = 0; i < ndmpdp; i++) {
 818                         ((pdp_entry_t *)DMPDPphys)[i] =
 819                                                 (vm_paddr_t)i << PDPSHIFT;
 820                         ((pdp_entry_t *)DMPDPphys)[i] |=
 821                             pmap_bits_default[PG_RW_IDX] |
 822                             pmap_bits_default[PG_V_IDX] |
 823                             pmap_bits_default[PG_PS_IDX] |
 824                             pmap_bits_default[PG_G_IDX] |
 825                             pmap_bits_default[PG_M_IDX] |
 826                             pmap_bits_default[PG_A_IDX];
 827                 }
 828         }
 829
 830         /* And recursively map PML4 to itself in order to get PTmap */
 831         ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
 832         ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
 833             pmap_bits_default[PG_RW_IDX] |
 834             pmap_bits_default[PG_V_IDX] |
 835             pmap_bits_default[PG_U_IDX];
 836
 837         /*
 838          * Connect the Direct Map slots up to the PML4
 839          */
 840         for (j = 0; j < NDMPML4E; ++j) {
 841                 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
 842                     (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
 843                     pmap_bits_default[PG_RW_IDX] |
 844                     pmap_bits_default[PG_V_IDX] |
 845                     pmap_bits_default[PG_U_IDX];
 846         }
 847
 848         /*
 849          * Connect the KVA slot up to the PML4
 850          */
 851         ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
 852         ((pdp_entry_t *)KPML4phys)[KPML4I] |=
 853             pmap_bits_default[PG_RW_IDX] |
 854             pmap_bits_default[PG_V_IDX] |
 855             pmap_bits_default[PG_U_IDX];
 856 }
 857
 858 /*
 859  *      Bootstrap the system enough to run with virtual memory.
 860  *
 861  *      On the i386 this is called after mapping has already been enabled
 862  *      and just syncs the pmap module with what has already been done.
 863  *      [We can't call it easily with mapping off since the kernel is not
 864  *      mapped with PA == VA, hence we would have to relocate every address
 865  *      from the linked base (virtual) address "KERNBASE" to the actual
 866  *      (physical) address starting relative to 0]
 867  */
 868 void
 869 pmap_bootstrap(vm_paddr_t *firstaddr)
 870 {
 871         vm_offset_t va;
 872         pt_entry_t *pte;
 873
 874         KvaStart = VM_MIN_KERNEL_ADDRESS;
 875         KvaEnd = VM_MAX_KERNEL_ADDRESS;
 876         KvaSize = KvaEnd - KvaStart;
 877
 878         avail_start = *firstaddr;
 879
 880         /*
 881          * Create an initial set of page tables to run the kernel in.
 882          */
 883         create_pagetables(firstaddr);
 884
 885         virtual2_start = KvaStart;
 886         virtual2_end = PTOV_OFFSET;
 887
 888         virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
 889         virtual_start = pmap_kmem_choose(virtual_start);
 890
 891         virtual_end = VM_MAX_KERNEL_ADDRESS;
 892
 893         /* XXX do %cr0 as well */
 894         load_cr4(rcr4() | CR4_PGE | CR4_PSE);
 895         load_cr3(KPML4phys);
 896
 897         /*
 898          * Initialize protection array.
 899          */
 900         i386_protection_init();
 901
 902         /*
 903          * The kernel's pmap is statically allocated so we don't have to use
 904          * pmap_create, which is unlikely to work correctly at this part of
 905          * the boot sequence (XXX and which no longer exists).
 906          */
 907         kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
 908         kernel_pmap.pm_count = 1;
 909         kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
 910         RB_INIT(&kernel_pmap.pm_pvroot);
 911         spin_init(&kernel_pmap.pm_spin);
 912         lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
 913
 914         /*
 915          * Reserve some special page table entries/VA space for temporary
 916          * mapping of pages.
 917          */
 918 #define SYSMAP(c, p, v, n)      \
 919         v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
 920
 921         va = virtual_start;
 922         pte = vtopte(va);
 923
 924         /*
 925          * CMAP1/CMAP2 are used for zeroing and copying pages.
 926          */
 927         SYSMAP(caddr_t, CMAP1, CADDR1, 1)
 928
 929         /*
 930          * Crashdump maps.
 931          */
 932         SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
 933
 934         /*
 935          * ptvmmap is used for reading arbitrary physical pages via
 936          * /dev/mem.
 937          */
 938         SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
 939
 940         /*
 941          * msgbufp is used to map the system message buffer.
 942          * XXX msgbufmap is not used.
 943          */
 944         SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
 945                atop(round_page(MSGBUF_SIZE)))
 946
 947         virtual_start = va;
 948
 949         *CMAP1 = 0;
 950
 951         /*
 952          * PG_G is terribly broken on SMP because we IPI invltlb's in some
 953          * cases rather then invl1pg.  Actually, I don't even know why it
 954          * works under UP because self-referential page table mappings
 955          */
 956 //      pgeflag = 0;
 957
 958 /*
 959  * Initialize the 4MB page size flag
 960  */
 961 //      pseflag = 0;
 962 /*
 963  * The 4MB page version of the initial
 964  * kernel page mapping.
 965  */
 966         pdir4mb = 0;
 967
 968 #if !defined(DISABLE_PSE)
 969         if (cpu_feature & CPUID_PSE) {
 970                 pt_entry_t ptditmp;
 971                 /*
 972                  * Note that we have enabled PSE mode
 973                  */
 974 //              pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
 975                 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
 976                 ptditmp &= ~(NBPDR - 1);
 977                 ptditmp |= pmap_bits_default[PG_V_IDX] |
 978                     pmap_bits_default[PG_RW_IDX] |
 979                     pmap_bits_default[PG_PS_IDX] |
 980                     pmap_bits_default[PG_U_IDX];
 981 //                  pgeflag;
 982                 pdir4mb = ptditmp;
 983         }
 984 #endif
 985         cpu_invltlb();
 986
 987         /* Initialize the PAT MSR */
 988         pmap_init_pat();
 989
 990         pmap_pinit_defaults(&kernel_pmap);
 991 }
 992
 993 /*
 994  * Setup the PAT MSR.
 995  */
 996 void
 997 pmap_init_pat(void)
 998 {
 999         uint64_t pat_msr;
1000         u_long cr0, cr4;
1001
1002         /*
1003          * Default values mapping PATi,PCD,PWT bits at system reset.
1004          * The default values effectively ignore the PATi bit by
1005          * repeating the encodings for 0-3 in 4-7, and map the PCD
1006          * and PWT bit combinations to the expected PAT types.
1007          */
1008         pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |        /* 000 */
1009                   PAT_VALUE(1, PAT_WRITE_THROUGH) |     /* 001 */
1010                   PAT_VALUE(2, PAT_UNCACHED) |          /* 010 */
1011                   PAT_VALUE(3, PAT_UNCACHEABLE) |       /* 011 */
1012                   PAT_VALUE(4, PAT_WRITE_BACK) |        /* 100 */
1013                   PAT_VALUE(5, PAT_WRITE_THROUGH) |     /* 101 */
1014                   PAT_VALUE(6, PAT_UNCACHED) |          /* 110 */
1015                   PAT_VALUE(7, PAT_UNCACHEABLE);        /* 111 */
1016         pat_pte_index[PAT_WRITE_BACK]   = 0;
1017         pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
1018         pat_pte_index[PAT_UNCACHED]     = X86_PG_NC_PCD;
1019         pat_pte_index[PAT_UNCACHEABLE]  = X86_PG_NC_PCD | X86_PG_NC_PWT;
1020         pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1021         pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1022
1023         if (cpu_feature & CPUID_PAT) {
1024                 /*
1025                  * If we support the PAT then set-up entries for
1026                  * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1027                  * 4 and 5.
1028                  */
1029                 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1030                           PAT_VALUE(4, PAT_WRITE_PROTECTED);
1031                 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1032                           PAT_VALUE(5, PAT_WRITE_COMBINING);
1033                 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1034                 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1035
1036                 /*
1037                  * Then enable the PAT
1038                  */
1039
1040                 /* Disable PGE. */
1041                 cr4 = rcr4();
1042                 load_cr4(cr4 & ~CR4_PGE);
1043
1044                 /* Disable caches (CD = 1, NW = 0). */
1045                 cr0 = rcr0();
1046                 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1047
1048                 /* Flushes caches and TLBs. */
1049                 wbinvd();
1050                 cpu_invltlb();
1051
1052                 /* Update PAT and index table. */
1053                 wrmsr(MSR_PAT, pat_msr);
1054
1055                 /* Flush caches and TLBs again. */
1056                 wbinvd();
1057                 cpu_invltlb();
1058
1059                 /* Restore caches and PGE. */
1060                 load_cr0(cr0);
1061                 load_cr4(cr4);
1062                 PatMsr = pat_msr;
1063         }
1064 }
1065
1066 /*
1067  * Set 4mb pdir for mp startup
1068  */
1069 void
1070 pmap_set_opt(void)
1071 {
1072         if (cpu_feature & CPUID_PSE) {
1073                 load_cr4(rcr4() | CR4_PSE);
1074                 if (pdir4mb && mycpu->gd_cpuid == 0) {  /* only on BSP */
1075                         cpu_invltlb();
1076                 }
1077         }
1078 }
1079
1080 /*
1081  *      Initialize the pmap module.
1082  *      Called by vm_init, to initialize any structures that the pmap
1083  *      system needs to map virtual memory.
1084  *      pmap_init has been enhanced to support in a fairly consistant
1085  *      way, discontiguous physical memory.
1086  */
1087 void
1088 pmap_init(void)
1089 {
1090         int i;
1091         int initial_pvs;
1092
1093         /*
1094          * Allocate memory for random pmap data structures.  Includes the
1095          * pv_head_table.
1096          */
1097
1098         for (i = 0; i < vm_page_array_size; i++) {
1099                 vm_page_t m;
1100
1101                 m = &vm_page_array[i];
1102                 TAILQ_INIT(&m->md.pv_list);
1103         }
1104
1105         /*
1106          * init the pv free list
1107          */
1108         initial_pvs = vm_page_array_size;
1109         if (initial_pvs < MINPV)
1110                 initial_pvs = MINPV;
1111         pvzone = &pvzone_store;
1112         pvinit = (void *)kmem_alloc(&kernel_map,
1113                                     initial_pvs * sizeof (struct pv_entry));
1114         zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1115                   pvinit, initial_pvs);
1116
1117         /*
1118          * Now it is safe to enable pv_table recording.
1119          */
1120         pmap_initialized = TRUE;
1121 }
1122
1123 /*
1124  * Initialize the address space (zone) for the pv_entries.  Set a
1125  * high water mark so that the system can recover from excessive
1126  * numbers of pv entries.
1127  */
1128 void
1129 pmap_init2(void)
1130 {
1131         int shpgperproc = PMAP_SHPGPERPROC;
1132         int entry_max;
1133
1134         TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1135         pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1136         TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1137         pv_entry_high_water = 9 * (pv_entry_max / 10);
1138
1139         /*
1140          * Subtract out pages already installed in the zone (hack)
1141          */
1142         entry_max = pv_entry_max - vm_page_array_size;
1143         if (entry_max <= 0)
1144                 entry_max = 1;
1145
1146         zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
1147 }
1148
1149 /*
1150  * Typically used to initialize a fictitious page by vm/device_pager.c
1151  */
1152 void
1153 pmap_page_init(struct vm_page *m)
1154 {
1155         vm_page_init(m);
1156         TAILQ_INIT(&m->md.pv_list);
1157 }
1158
1159 /***************************************************
1160  * Low level helper routines.....
1161  ***************************************************/
1162
1163 /*
1164  * this routine defines the region(s) of memory that should
1165  * not be tested for the modified bit.
1166  */
1167 static __inline
1168 int
1169 pmap_track_modified(vm_pindex_t pindex)
1170 {
1171         vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1172         if ((va < clean_sva) || (va >= clean_eva))
1173                 return 1;
1174         else
1175                 return 0;
1176 }
1177
1178 /*
1179  * Extract the physical page address associated with the map/VA pair.
1180  * The page must be wired for this to work reliably.
1181  *
1182  * XXX for the moment we're using pv_find() instead of pv_get(), as
1183  *     callers might be expecting non-blocking operation.
1184  */
1185 vm_paddr_t
1186 pmap_extract(pmap_t pmap, vm_offset_t va)
1187 {
1188         vm_paddr_t rtval;
1189         pv_entry_t pt_pv;
1190         pt_entry_t *ptep;
1191
1192         rtval = 0;
1193         if (va >= VM_MAX_USER_ADDRESS) {
1194                 /*
1195                  * Kernel page directories might be direct-mapped and
1196                  * there is typically no PV tracking of pte's
1197                  */
1198                 pd_entry_t *pt;
1199
1200                 pt = pmap_pt(pmap, va);
1201                 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1202                         if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1203                                 rtval = *pt & PG_PS_FRAME;
1204                                 rtval |= va & PDRMASK;
1205                         } else {
1206                                 ptep = pmap_pt_to_pte(*pt, va);
1207                                 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1208                                         rtval = *ptep & PG_FRAME;
1209                                         rtval |= va & PAGE_MASK;
1210                                 }
1211                         }
1212                 }
1213         } else {
1214                 /*
1215                  * User pages currently do not direct-map the page directory
1216                  * and some pages might not used managed PVs.  But all PT's
1217                  * will have a PV.
1218                  */
1219                 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1220                 if (pt_pv) {
1221                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1222                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1223                                 rtval = *ptep & PG_FRAME;
1224                                 rtval |= va & PAGE_MASK;
1225                         }
1226                         pv_drop(pt_pv);
1227                 }
1228         }
1229         return rtval;
1230 }
1231
1232 /*
1233  * Similar to extract but checks protections, SMP-friendly short-cut for
1234  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1235  * fall-through to the real fault code.
1236  *
1237  * The returned page, if not NULL, is held (and not busied).
1238  */
1239 vm_page_t
1240 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1241 {
1242         if (pmap && va < VM_MAX_USER_ADDRESS) {
1243                 pv_entry_t pt_pv;
1244                 pv_entry_t pte_pv;
1245                 pt_entry_t *ptep;
1246                 pt_entry_t req;
1247                 vm_page_t m;
1248                 int error;
1249
1250                 req = pmap->pmap_bits[PG_V_IDX] |
1251                       pmap->pmap_bits[PG_U_IDX];
1252                 if (prot & VM_PROT_WRITE)
1253                         req |= pmap->pmap_bits[PG_RW_IDX];
1254
1255                 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1256                 if (pt_pv == NULL)
1257                         return (NULL);
1258                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1259                 if ((*ptep & req) != req) {
1260                         pv_drop(pt_pv);
1261                         return (NULL);
1262                 }
1263                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), &error);
1264                 if (pte_pv && error == 0) {
1265                         m = pte_pv->pv_m;
1266                         vm_page_hold(m);
1267                         if (prot & VM_PROT_WRITE)
1268                                 vm_page_dirty(m);
1269                         pv_put(pte_pv);
1270                 } else if (pte_pv) {
1271                         pv_drop(pte_pv);
1272                         m = NULL;
1273                 } else {
1274                         m = NULL;
1275                 }
1276                 pv_drop(pt_pv);
1277                 return(m);
1278         } else {
1279                 return(NULL);
1280         }
1281 }
1282
1283 /*
1284  * Extract the physical page address associated kernel virtual address.
1285  */
1286 vm_paddr_t
1287 pmap_kextract(vm_offset_t va)
1288 {
1289         pd_entry_t pt;          /* pt entry in pd */
1290         vm_paddr_t pa;
1291
1292         if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1293                 pa = DMAP_TO_PHYS(va);
1294         } else {
1295                 pt = *vtopt(va);
1296                 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1297                         pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1298                 } else {
1299                         /*
1300                          * Beware of a concurrent promotion that changes the
1301                          * PDE at this point!  For example, vtopte() must not
1302                          * be used to access the PTE because it would use the
1303                          * new PDE.  It is, however, safe to use the old PDE
1304                          * because the page table page is preserved by the
1305                          * promotion.
1306                          */
1307                         pa = *pmap_pt_to_pte(pt, va);
1308                         pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1309                 }
1310         }
1311         return pa;
1312 }
1313
1314 /***************************************************
1315  * Low level mapping routines.....
1316  ***************************************************/
1317
1318 /*
1319  * Routine: pmap_kenter
1320  * Function:
1321  *      Add a wired page to the KVA
1322  *      NOTE! note that in order for the mapping to take effect -- you
1323  *      should do an invltlb after doing the pmap_kenter().
1324  */
1325 void
1326 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1327 {
1328         pt_entry_t *pte;
1329         pt_entry_t npte;
1330         pmap_inval_info info;
1331
1332         pmap_inval_init(&info);                         /* XXX remove */
1333         npte = pa |
1334             kernel_pmap.pmap_bits[PG_RW_IDX] |
1335             kernel_pmap.pmap_bits[PG_V_IDX];
1336 //          pgeflag;
1337         pte = vtopte(va);
1338         pmap_inval_interlock(&info, &kernel_pmap, va);  /* XXX remove */
1339         *pte = npte;
1340         pmap_inval_deinterlock(&info, &kernel_pmap);    /* XXX remove */
1341         pmap_inval_done(&info);                         /* XXX remove */
1342 }
1343
1344 /*
1345  * Routine: pmap_kenter_quick
1346  * Function:
1347  *      Similar to pmap_kenter(), except we only invalidate the
1348  *      mapping on the current CPU.
1349  */
1350 void
1351 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1352 {
1353         pt_entry_t *pte;
1354         pt_entry_t npte;
1355
1356         npte = pa |
1357             kernel_pmap.pmap_bits[PG_RW_IDX] |
1358             kernel_pmap.pmap_bits[PG_V_IDX];
1359 //          pgeflag;
1360         pte = vtopte(va);
1361         *pte = npte;
1362         cpu_invlpg((void *)va);
1363 }
1364
1365 void
1366 pmap_kenter_sync(vm_offset_t va)
1367 {
1368         pmap_inval_info info;
1369
1370         pmap_inval_init(&info);
1371         pmap_inval_interlock(&info, &kernel_pmap, va);
1372         pmap_inval_deinterlock(&info, &kernel_pmap);
1373         pmap_inval_done(&info);
1374 }
1375
1376 void
1377 pmap_kenter_sync_quick(vm_offset_t va)
1378 {
1379         cpu_invlpg((void *)va);
1380 }
1381
1382 /*
1383  * remove a page from the kernel pagetables
1384  */
1385 void
1386 pmap_kremove(vm_offset_t va)
1387 {
1388         pt_entry_t *pte;
1389         pmap_inval_info info;
1390
1391         pmap_inval_init(&info);
1392         pte = vtopte(va);
1393         pmap_inval_interlock(&info, &kernel_pmap, va);
1394         (void)pte_load_clear(pte);
1395         pmap_inval_deinterlock(&info, &kernel_pmap);
1396         pmap_inval_done(&info);
1397 }
1398
1399 void
1400 pmap_kremove_quick(vm_offset_t va)
1401 {
1402         pt_entry_t *pte;
1403         pte = vtopte(va);
1404         (void)pte_load_clear(pte);
1405         cpu_invlpg((void *)va);
1406 }
1407
1408 /*
1409  * XXX these need to be recoded.  They are not used in any critical path.
1410  */
1411 void
1412 pmap_kmodify_rw(vm_offset_t va)
1413 {
1414         atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1415         cpu_invlpg((void *)va);
1416 }
1417
1418 /* NOT USED
1419 void
1420 pmap_kmodify_nc(vm_offset_t va)
1421 {
1422         atomic_set_long(vtopte(va), PG_N);
1423         cpu_invlpg((void *)va);
1424 }
1425 */
1426
1427 /*
1428  * Used to map a range of physical addresses into kernel virtual
1429  * address space during the low level boot, typically to map the
1430  * dump bitmap, message buffer, and vm_page_array.
1431  *
1432  * These mappings are typically made at some pointer after the end of the
1433  * kernel text+data.
1434  *
1435  * We could return PHYS_TO_DMAP(start) here and not allocate any
1436  * via (*virtp), but then kmem from userland and kernel dumps won't
1437  * have access to the related pointers.
1438  */
1439 vm_offset_t
1440 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1441 {
1442         vm_offset_t va;
1443         vm_offset_t va_start;
1444
1445         /*return PHYS_TO_DMAP(start);*/
1446
1447         va_start = *virtp;
1448         va = va_start;
1449
1450         while (start < end) {
1451                 pmap_kenter_quick(va, start);
1452                 va += PAGE_SIZE;
1453                 start += PAGE_SIZE;
1454         }
1455         *virtp = va;
1456         return va_start;
1457 }
1458
1459 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
1460
1461 /*
1462  * Remove the specified set of pages from the data and instruction caches.
1463  *
1464  * In contrast to pmap_invalidate_cache_range(), this function does not
1465  * rely on the CPU's self-snoop feature, because it is intended for use
1466  * when moving pages into a different cache domain.
1467  */
1468 void
1469 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1470 {
1471         vm_offset_t daddr, eva;
1472         int i;
1473
1474         if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1475             (cpu_feature & CPUID_CLFSH) == 0)
1476                 wbinvd();
1477         else {
1478                 cpu_mfence();
1479                 for (i = 0; i < count; i++) {
1480                         daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1481                         eva = daddr + PAGE_SIZE;
1482                         for (; daddr < eva; daddr += cpu_clflush_line_size)
1483                                 clflush(daddr);
1484                 }
1485                 cpu_mfence();
1486         }
1487 }
1488
1489 void
1490 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1491 {
1492         KASSERT((sva & PAGE_MASK) == 0,
1493             ("pmap_invalidate_cache_range: sva not page-aligned"));
1494         KASSERT((eva & PAGE_MASK) == 0,
1495             ("pmap_invalidate_cache_range: eva not page-aligned"));
1496
1497         if (cpu_feature & CPUID_SS) {
1498                 ; /* If "Self Snoop" is supported, do nothing. */
1499         } else {
1500                 /* Globally invalidate caches */
1501                 cpu_wbinvd_on_all_cpus();
1502         }
1503 }
1504 void
1505 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1506 {
1507         smp_invlpg_range(pmap->pm_active, sva, eva);
1508 }
1509
1510 /*
1511  * Add a list of wired pages to the kva
1512  * this routine is only used for temporary
1513  * kernel mappings that do not need to have
1514  * page modification or references recorded.
1515  * Note that old mappings are simply written
1516  * over.  The page *must* be wired.
1517  */
1518 void
1519 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1520 {
1521         vm_offset_t end_va;
1522
1523         end_va = va + count * PAGE_SIZE;
1524
1525         while (va < end_va) {
1526                 pt_entry_t *pte;
1527
1528                 pte = vtopte(va);
1529                 *pte = VM_PAGE_TO_PHYS(*m) |
1530                     kernel_pmap.pmap_bits[PG_RW_IDX] |
1531                     kernel_pmap.pmap_bits[PG_V_IDX] |
1532                     kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1533 //              pgeflag;
1534                 cpu_invlpg((void *)va);
1535                 va += PAGE_SIZE;
1536                 m++;
1537         }
1538         smp_invltlb();
1539 }
1540
1541 /*
1542  * This routine jerks page mappings from the
1543  * kernel -- it is meant only for temporary mappings.
1544  *
1545  * MPSAFE, INTERRUPT SAFE (cluster callback)
1546  */
1547 void
1548 pmap_qremove(vm_offset_t va, int count)
1549 {
1550         vm_offset_t end_va;
1551
1552         end_va = va + count * PAGE_SIZE;
1553
1554         while (va < end_va) {
1555                 pt_entry_t *pte;
1556
1557                 pte = vtopte(va);
1558                 (void)pte_load_clear(pte);
1559                 cpu_invlpg((void *)va);
1560                 va += PAGE_SIZE;
1561         }
1562         smp_invltlb();
1563 }
1564
1565 /*
1566  * Create a new thread and optionally associate it with a (new) process.
1567  * NOTE! the new thread's cpu may not equal the current cpu.
1568  */
1569 void
1570 pmap_init_thread(thread_t td)
1571 {
1572         /* enforce pcb placement & alignment */
1573         td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1574         td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1575         td->td_savefpu = &td->td_pcb->pcb_save;
1576         td->td_sp = (char *)td->td_pcb; /* no -16 */
1577 }
1578
1579 /*
1580  * This routine directly affects the fork perf for a process.
1581  */
1582 void
1583 pmap_init_proc(struct proc *p)
1584 {
1585 }
1586
1587 static void
1588 pmap_pinit_defaults(struct pmap *pmap)
1589 {
1590         bcopy(pmap_bits_default, pmap->pmap_bits,
1591               sizeof(pmap_bits_default));
1592         bcopy(protection_codes, pmap->protection_codes,
1593               sizeof(protection_codes));
1594         bcopy(pat_pte_index, pmap->pmap_cache_bits,
1595               sizeof(pat_pte_index));
1596         pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1597         pmap->copyinstr = std_copyinstr;
1598         pmap->copyin = std_copyin;
1599         pmap->copyout = std_copyout;
1600         pmap->fubyte = std_fubyte;
1601         pmap->subyte = std_subyte;
1602         pmap->fuword = std_fuword;
1603         pmap->suword = std_suword;
1604         pmap->suword32 = std_suword32;
1605 }
1606 /*
1607  * Initialize pmap0/vmspace0.  This pmap is not added to pmap_list because
1608  * it, and IdlePTD, represents the template used to update all other pmaps.
1609  *
1610  * On architectures where the kernel pmap is not integrated into the user
1611  * process pmap, this pmap represents the process pmap, not the kernel pmap.
1612  * kernel_pmap should be used to directly access the kernel_pmap.
1613  */
1614 void
1615 pmap_pinit0(struct pmap *pmap)
1616 {
1617         pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1618         pmap->pm_count = 1;
1619         pmap->pm_active = 0;
1620         pmap->pm_pvhint = NULL;
1621         RB_INIT(&pmap->pm_pvroot);
1622         spin_init(&pmap->pm_spin);
1623         lwkt_token_init(&pmap->pm_token, "pmap_tok");
1624         bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1625         pmap_pinit_defaults(pmap);
1626 }
1627
1628 /*
1629  * Initialize a preallocated and zeroed pmap structure,
1630  * such as one in a vmspace structure.
1631  */
1632 static void
1633 pmap_pinit_simple(struct pmap *pmap)
1634 {
1635         /*
1636          * Misc initialization
1637          */
1638         pmap->pm_count = 1;
1639         pmap->pm_active = 0;
1640         pmap->pm_pvhint = NULL;
1641         pmap->pm_flags = PMAP_FLAG_SIMPLE;
1642
1643         pmap_pinit_defaults(pmap);
1644
1645         /*
1646          * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1647          * for this).
1648          */
1649         if (pmap->pm_pmlpv == NULL) {
1650                 RB_INIT(&pmap->pm_pvroot);
1651                 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1652                 spin_init(&pmap->pm_spin);
1653                 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1654         }
1655 }
1656
1657 void
1658 pmap_pinit(struct pmap *pmap)
1659 {
1660         pv_entry_t pv;
1661         int j;
1662
1663         if (pmap->pm_pmlpv) {
1664                 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1665                         pmap_puninit(pmap);
1666                 }
1667         }
1668
1669         pmap_pinit_simple(pmap);
1670         pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1671
1672         /*
1673          * No need to allocate page table space yet but we do need a valid
1674          * page directory table.
1675          */
1676         if (pmap->pm_pml4 == NULL) {
1677                 pmap->pm_pml4 =
1678                     (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1679         }
1680
1681         /*
1682          * Allocate the page directory page, which wires it even though
1683          * it isn't being entered into some higher level page table (it
1684          * being the highest level).  If one is already cached we don't
1685          * have to do anything.
1686          */
1687         if ((pv = pmap->pm_pmlpv) == NULL) {
1688                 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1689                 pmap->pm_pmlpv = pv;
1690                 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1691                             VM_PAGE_TO_PHYS(pv->pv_m));
1692                 pv_put(pv);
1693
1694                 /*
1695                  * Install DMAP and KMAP.
1696                  */
1697                 for (j = 0; j < NDMPML4E; ++j) {
1698                         pmap->pm_pml4[DMPML4I + j] =
1699                             (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1700                             pmap->pmap_bits[PG_RW_IDX] |
1701                             pmap->pmap_bits[PG_V_IDX] |
1702                             pmap->pmap_bits[PG_U_IDX];
1703                 }
1704                 pmap->pm_pml4[KPML4I] = KPDPphys |
1705                     pmap->pmap_bits[PG_RW_IDX] |
1706                     pmap->pmap_bits[PG_V_IDX] |
1707                     pmap->pmap_bits[PG_U_IDX];
1708
1709                 /*
1710                  * install self-referential address mapping entry
1711                  */
1712                 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1713                     pmap->pmap_bits[PG_V_IDX] |
1714                     pmap->pmap_bits[PG_RW_IDX] |
1715                     pmap->pmap_bits[PG_A_IDX] |
1716                     pmap->pmap_bits[PG_M_IDX];
1717         } else {
1718                 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1719                 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1720         }
1721         KKASSERT(pmap->pm_pml4[255] == 0);
1722         KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1723         KKASSERT(pv->pv_entry.rbe_left == NULL);
1724         KKASSERT(pv->pv_entry.rbe_right == NULL);
1725 }
1726
1727 /*
1728  * Clean up a pmap structure so it can be physically freed.  This routine
1729  * is called by the vmspace dtor function.  A great deal of pmap data is
1730  * left passively mapped to improve vmspace management so we have a bit
1731  * of cleanup work to do here.
1732  */
1733 void
1734 pmap_puninit(pmap_t pmap)
1735 {
1736         pv_entry_t pv;
1737         vm_page_t p;
1738
1739         KKASSERT(pmap->pm_active == 0);
1740         if ((pv = pmap->pm_pmlpv) != NULL) {
1741                 if (pv_hold_try(pv) == 0)
1742                         pv_lock(pv);
1743                 KKASSERT(pv == pmap->pm_pmlpv);
1744                 p = pmap_remove_pv_page(pv);
1745                 pv_free(pv);
1746                 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1747                 vm_page_busy_wait(p, FALSE, "pgpun");
1748                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1749                 vm_page_unwire(p, 0);
1750                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1751
1752                 /*
1753                  * XXX eventually clean out PML4 static entries and
1754                  * use vm_page_free_zero()
1755                  */
1756                 vm_page_free(p);
1757                 pmap->pm_pmlpv = NULL;
1758         }
1759         if (pmap->pm_pml4) {
1760                 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1761                 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1762                 pmap->pm_pml4 = NULL;
1763         }
1764         KKASSERT(pmap->pm_stats.resident_count == 0);
1765         KKASSERT(pmap->pm_stats.wired_count == 0);
1766 }
1767
1768 /*
1769  * Wire in kernel global address entries.  To avoid a race condition
1770  * between pmap initialization and pmap_growkernel, this procedure
1771  * adds the pmap to the master list (which growkernel scans to update),
1772  * then copies the template.
1773  */
1774 void
1775 pmap_pinit2(struct pmap *pmap)
1776 {
1777         spin_lock(&pmap_spin);
1778         TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1779         spin_unlock(&pmap_spin);
1780 }
1781
1782 /*
1783  * This routine is called when various levels in the page table need to
1784  * be populated.  This routine cannot fail.
1785  *
1786  * This function returns two locked pv_entry's, one representing the
1787  * requested pv and one representing the requested pv's parent pv.  If
1788  * the pv did not previously exist it will be mapped into its parent
1789  * and wired, otherwise no additional wire count will be added.
1790  */
1791 static
1792 pv_entry_t
1793 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1794 {
1795         pt_entry_t *ptep;
1796         pv_entry_t pv;
1797         pv_entry_t pvp;
1798         vm_pindex_t pt_pindex;
1799         vm_page_t m;
1800         int isnew;
1801         int ispt;
1802
1803         /*
1804          * If the pv already exists and we aren't being asked for the
1805          * parent page table page we can just return it.  A locked+held pv
1806          * is returned.  The pv will also have a second hold related to the
1807          * pmap association that we don't have to worry about.
1808          */
1809         ispt = 0;
1810         pv = pv_alloc(pmap, ptepindex, &isnew);
1811         if (isnew == 0 && pvpp == NULL)
1812                 return(pv);
1813
1814         /*
1815          * Special case terminal PVs.  These are not page table pages so
1816          * no vm_page is allocated (the caller supplied the vm_page).  If
1817          * pvpp is non-NULL we are being asked to also removed the pt_pv
1818          * for this pv.
1819          *
1820          * Note that pt_pv's are only returned for user VAs. We assert that
1821          * a pt_pv is not being requested for kernel VAs.
1822          */
1823         if (ptepindex < pmap_pt_pindex(0)) {
1824                 if (ptepindex >= NUPTE_USER)
1825                         KKASSERT(pvpp == NULL);
1826                 else
1827                         KKASSERT(pvpp != NULL);
1828                 if (pvpp) {
1829                         pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1830                         pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1831                         if (isnew)
1832                                 vm_page_wire_quick(pvp->pv_m);
1833                         *pvpp = pvp;
1834                 } else {
1835                         pvp = NULL;
1836                 }
1837                 return(pv);
1838         }
1839
1840         /*
1841          * Non-terminal PVs allocate a VM page to represent the page table,
1842          * so we have to resolve pvp and calculate ptepindex for the pvp
1843          * and then for the page table entry index in the pvp for
1844          * fall-through.
1845          */
1846         if (ptepindex < pmap_pd_pindex(0)) {
1847                 /*
1848                  * pv is PT, pvp is PD
1849                  */
1850                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1851                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1852                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1853                 if (!isnew)
1854                         goto notnew;
1855
1856                 /*
1857                  * PT index in PD
1858                  */
1859                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1860                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1861                 ispt = 1;
1862         } else if (ptepindex < pmap_pdp_pindex(0)) {
1863                 /*
1864                  * pv is PD, pvp is PDP
1865                  *
1866                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1867                  *                   the PD.
1868                  */
1869                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1870                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1871
1872                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1873                         KKASSERT(pvpp == NULL);
1874                         pvp = NULL;
1875                 } else {
1876                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
1877                 }
1878                 if (!isnew)
1879                         goto notnew;
1880
1881                 /*
1882                  * PD index in PDP
1883                  */
1884                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1885                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1886         } else if (ptepindex < pmap_pml4_pindex()) {
1887                 /*
1888                  * pv is PDP, pvp is the root pml4 table
1889                  */
1890                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1891                 if (!isnew)
1892                         goto notnew;
1893
1894                 /*
1895                  * PDP index in PML4
1896                  */
1897                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1898                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1899         } else {
1900                 /*
1901                  * pv represents the top-level PML4, there is no parent.
1902                  */
1903                 pvp = NULL;
1904                 if (!isnew)
1905                         goto notnew;
1906         }
1907
1908         /*
1909          * This code is only reached if isnew is TRUE and this is not a
1910          * terminal PV.  We need to allocate a vm_page for the page table
1911          * at this level and enter it into the parent page table.
1912          *
1913          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1914          */
1915         for (;;) {
1916                 m = vm_page_alloc(NULL, pv->pv_pindex,
1917                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1918                                   VM_ALLOC_INTERRUPT);
1919                 if (m)
1920                         break;
1921                 vm_wait(0);
1922         }
1923         vm_page_spin_lock(m);
1924         pmap_page_stats_adding(m);
1925         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1926         pv->pv_m = m;
1927         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1928         vm_page_spin_unlock(m);
1929         vm_page_unmanage(m);    /* m must be spinunlocked */
1930
1931         if ((m->flags & PG_ZERO) == 0) {
1932                 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1933         }
1934 #ifdef PMAP_DEBUG
1935         else {
1936                 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1937         }
1938 #endif
1939         m->valid = VM_PAGE_BITS_ALL;
1940         vm_page_flag_clear(m, PG_ZERO);
1941         vm_page_wire(m);        /* wire for mapping in parent */
1942
1943         /*
1944          * Wire the page into pvp, bump the wire-count for pvp's page table
1945          * page.  Bump the resident_count for the pmap.  There is no pvp
1946          * for the top level, address the pm_pml4[] array directly.
1947          *
1948          * If the caller wants the parent we return it, otherwise
1949          * we just put it away.
1950          *
1951          * No interlock is needed for pte 0 -> non-zero.
1952          *
1953          * In the situation where *ptep is valid we might have an unmanaged
1954          * page table page shared from another page table which we need to
1955          * unshare before installing our private page table page.
1956          */
1957         if (pvp) {
1958                 ptep = pv_pte_lookup(pvp, ptepindex);
1959                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1960                         pt_entry_t pte;
1961                         pmap_inval_info info;
1962
1963                         if (ispt == 0) {
1964                                 panic("pmap_allocpte: unexpected pte %p/%d",
1965                                       pvp, (int)ptepindex);
1966                         }
1967                         pmap_inval_init(&info);
1968                         pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1969                         pte = pte_load_clear(ptep);
1970                         pmap_inval_deinterlock(&info, pmap);
1971                         pmap_inval_done(&info);
1972                         if (vm_page_unwire_quick(
1973                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
1974                                 panic("pmap_allocpte: shared pgtable "
1975                                       "pg bad wirecount");
1976                         }
1977                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
1978                 } else {
1979                         vm_page_wire_quick(pvp->pv_m);
1980                 }
1981                 *ptep = VM_PAGE_TO_PHYS(m) |
1982                     (pmap->pmap_bits[PG_U_IDX] |
1983                     pmap->pmap_bits[PG_RW_IDX] |
1984                     pmap->pmap_bits[PG_V_IDX] |
1985                     pmap->pmap_bits[PG_A_IDX] |
1986                     pmap->pmap_bits[PG_M_IDX]);
1987         }
1988         vm_page_wakeup(m);
1989 notnew:
1990         if (pvpp)
1991                 *pvpp = pvp;
1992         else if (pvp)
1993                 pv_put(pvp);
1994         return (pv);
1995 }
1996
1997 /*
1998  * This version of pmap_allocpte() checks for possible segment optimizations
1999  * that would allow page-table sharing.  It can be called for terminal
2000  * page or page table page ptepindex's.
2001  *
2002  * The function is called with page table page ptepindex's for fictitious
2003  * and unmanaged terminal pages.  That is, we don't want to allocate a
2004  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
2005  * for this case.
2006  *
2007  * This function can return a pv and *pvpp associated with the passed in pmap
2008  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
2009  * an unmanaged page table page will be entered into the pass in pmap.
2010  */
2011 static
2012 pv_entry_t
2013 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2014                   vm_map_entry_t entry, vm_offset_t va)
2015 {
2016         struct pmap_inval_info info;
2017         vm_object_t object;
2018         pmap_t obpmap;
2019         pmap_t *obpmapp;
2020         vm_offset_t b;
2021         pv_entry_t pte_pv;      /* in original or shared pmap */
2022         pv_entry_t pt_pv;       /* in original or shared pmap */
2023         pv_entry_t proc_pd_pv;  /* in original pmap */
2024         pv_entry_t proc_pt_pv;  /* in original pmap */
2025         pv_entry_t xpv;         /* PT in shared pmap */
2026         pd_entry_t *pt;         /* PT entry in PD of original pmap */
2027         pd_entry_t opte;        /* contents of *pt */
2028         pd_entry_t npte;        /* contents of *pt */
2029         vm_page_t m;
2030
2031 retry:
2032         /*
2033          * Basic tests, require a non-NULL vm_map_entry, require proper
2034          * alignment and type for the vm_map_entry, require that the
2035          * underlying object already be allocated.
2036          *
2037          * We allow almost any type of object to use this optimization.
2038          * The object itself does NOT have to be sized to a multiple of the
2039          * segment size, but the memory mapping does.
2040          *
2041          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2042          *     won't work as expected.
2043          */
2044         if (entry == NULL ||
2045             pmap_mmu_optimize == 0 ||                   /* not enabled */
2046             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal or pt */
2047             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2048             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2049             entry->object.vm_object == NULL ||          /* needs VM object */
2050             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2051             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2052             (entry->offset & SEG_MASK) ||               /* must be aligned */
2053             (entry->start & SEG_MASK)) {
2054                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2055         }
2056
2057         /*
2058          * Make sure the full segment can be represented.
2059          */
2060         b = va & ~(vm_offset_t)SEG_MASK;
2061         if (b < entry->start || b + SEG_SIZE > entry->end)
2062                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2063
2064         /*
2065          * If the full segment can be represented dive the VM object's
2066          * shared pmap, allocating as required.
2067          */
2068         object = entry->object.vm_object;
2069
2070         if (entry->protection & VM_PROT_WRITE)
2071                 obpmapp = &object->md.pmap_rw;
2072         else
2073                 obpmapp = &object->md.pmap_ro;
2074
2075 #ifdef PMAP_DEBUG2
2076         if (pmap_enter_debug > 0) {
2077                 --pmap_enter_debug;
2078                 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2079                         "obpmapp %p %p\n",
2080                         va, entry->protection, object,
2081                         obpmapp, *obpmapp);
2082                 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2083                         entry, entry->start, entry->end);
2084         }
2085 #endif
2086
2087         /*
2088          * We allocate what appears to be a normal pmap but because portions
2089          * of this pmap are shared with other unrelated pmaps we have to
2090          * set pm_active to point to all cpus.
2091          *
2092          * XXX Currently using pmap_spin to interlock the update, can't use
2093          *     vm_object_hold/drop because the token might already be held
2094          *     shared OR exclusive and we don't know.
2095          */
2096         while ((obpmap = *obpmapp) == NULL) {
2097                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2098                 pmap_pinit_simple(obpmap);
2099                 pmap_pinit2(obpmap);
2100                 spin_lock(&pmap_spin);
2101                 if (*obpmapp != NULL) {
2102                         /*
2103                          * Handle race
2104                          */
2105                         spin_unlock(&pmap_spin);
2106                         pmap_release(obpmap);
2107                         pmap_puninit(obpmap);
2108                         kfree(obpmap, M_OBJPMAP);
2109                         obpmap = *obpmapp; /* safety */
2110                 } else {
2111                         obpmap->pm_active = smp_active_mask;
2112                         *obpmapp = obpmap;
2113                         spin_unlock(&pmap_spin);
2114                 }
2115         }
2116
2117         /*
2118          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2119          * pte/pt using the shared pmap from the object but also adjust
2120          * the process pmap's page table page as a side effect.
2121          */
2122
2123         /*
2124          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2125          * we will return.  This is true if ptepindex represents a terminal
2126          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2127          * the PD.
2128          */
2129         pt_pv = NULL;
2130         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2131         if (ptepindex >= pmap_pt_pindex(0))
2132                 xpv = pte_pv;
2133         else
2134                 xpv = pt_pv;
2135
2136         /*
2137          * Resolve the PD in the process pmap so we can properly share the
2138          * page table page.  Lock order is bottom-up (leaf first)!
2139          *
2140          * NOTE: proc_pt_pv can be NULL.
2141          */
2142         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2143         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2144 #ifdef PMAP_DEBUG2
2145         if (pmap_enter_debug > 0) {
2146                 --pmap_enter_debug;
2147                 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2148                         proc_pt_pv,
2149                         (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2150                         proc_pd_pv,
2151                         va);
2152         }
2153 #endif
2154
2155         /*
2156          * xpv is the page table page pv from the shared object
2157          * (for convenience), from above.
2158          *
2159          * Calculate the pte value for the PT to load into the process PD.
2160          * If we have to change it we must properly dispose of the previous
2161          * entry.
2162          */
2163         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2164         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2165             (pmap->pmap_bits[PG_U_IDX] |
2166             pmap->pmap_bits[PG_RW_IDX] |
2167             pmap->pmap_bits[PG_V_IDX] |
2168             pmap->pmap_bits[PG_A_IDX] |
2169             pmap->pmap_bits[PG_M_IDX]);
2170
2171         /*
2172          * Dispose of previous page table page if it was local to the
2173          * process pmap.  If the old pt is not empty we cannot dispose of it
2174          * until we clean it out.  This case should not arise very often so
2175          * it is not optimized.
2176          */
2177         if (proc_pt_pv) {
2178                 if (proc_pt_pv->pv_m->wire_count != 1) {
2179                         pv_put(proc_pd_pv);
2180                         pv_put(proc_pt_pv);
2181                         pv_put(pt_pv);
2182                         pv_put(pte_pv);
2183                         pmap_remove(pmap,
2184                                     va & ~(vm_offset_t)SEG_MASK,
2185                                     (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2186                         goto retry;
2187                 }
2188                 pmap_release_pv(proc_pt_pv, proc_pd_pv);
2189                 proc_pt_pv = NULL;
2190                 /* relookup */
2191                 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2192         }
2193
2194         /*
2195          * Handle remaining cases.
2196          */
2197         if (*pt == 0) {
2198                 *pt = npte;
2199                 vm_page_wire_quick(xpv->pv_m);
2200                 vm_page_wire_quick(proc_pd_pv->pv_m);
2201                 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2202         } else if (*pt != npte) {
2203                 pmap_inval_init(&info);
2204                 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
2205
2206                 opte = pte_load_clear(pt);
2207                 KKASSERT(opte && opte != npte);
2208
2209                 *pt = npte;
2210                 vm_page_wire_quick(xpv->pv_m);  /* pgtable pg that is npte */
2211
2212                 /*
2213                  * Clean up opte, bump the wire_count for the process
2214                  * PD page representing the new entry if it was
2215                  * previously empty.
2216                  *
2217                  * If the entry was not previously empty and we have
2218                  * a PT in the proc pmap then opte must match that
2219                  * pt.  The proc pt must be retired (this is done
2220                  * later on in this procedure).
2221                  *
2222                  * NOTE: replacing valid pte, wire_count on proc_pd_pv
2223                  * stays the same.
2224                  */
2225                 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2226                 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2227                 if (vm_page_unwire_quick(m)) {
2228                         panic("pmap_allocpte_seg: "
2229                               "bad wire count %p",
2230                               m);
2231                 }
2232
2233                 pmap_inval_deinterlock(&info, pmap);
2234                 pmap_inval_done(&info);
2235         }
2236
2237         /*
2238          * The existing process page table was replaced and must be destroyed
2239          * here.
2240          */
2241         if (proc_pd_pv)
2242                 pv_put(proc_pd_pv);
2243         if (pvpp)
2244                 *pvpp = pt_pv;
2245         else
2246                 pv_put(pt_pv);
2247
2248         return (pte_pv);
2249 }
2250
2251 /*
2252  * Release any resources held by the given physical map.
2253  *
2254  * Called when a pmap initialized by pmap_pinit is being released.  Should
2255  * only be called if the map contains no valid mappings.
2256  *
2257  * Caller must hold pmap->pm_token
2258  */
2259 struct pmap_release_info {
2260         pmap_t  pmap;
2261         int     retry;
2262 };
2263
2264 static int pmap_release_callback(pv_entry_t pv, void *data);
2265
2266 void
2267 pmap_release(struct pmap *pmap)
2268 {
2269         struct pmap_release_info info;
2270
2271         KASSERT(pmap->pm_active == 0,
2272                 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
2273
2274         spin_lock(&pmap_spin);
2275         TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
2276         spin_unlock(&pmap_spin);
2277
2278         /*
2279          * Pull pv's off the RB tree in order from low to high and release
2280          * each page.
2281          */
2282         info.pmap = pmap;
2283         do {
2284                 info.retry = 0;
2285                 spin_lock(&pmap->pm_spin);
2286                 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2287                         pmap_release_callback, &info);
2288                 spin_unlock(&pmap->pm_spin);
2289         } while (info.retry);
2290
2291
2292         /*
2293          * One resident page (the pml4 page) should remain.
2294          * No wired pages should remain.
2295          */
2296         KKASSERT(pmap->pm_stats.resident_count ==
2297                  ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2298
2299         KKASSERT(pmap->pm_stats.wired_count == 0);
2300 }
2301
2302 static int
2303 pmap_release_callback(pv_entry_t pv, void *data)
2304 {
2305         struct pmap_release_info *info = data;
2306         pmap_t pmap = info->pmap;
2307         int r;
2308
2309         if (pv_hold_try(pv)) {
2310                 spin_unlock(&pmap->pm_spin);
2311         } else {
2312                 spin_unlock(&pmap->pm_spin);
2313                 pv_lock(pv);
2314         }
2315         if (pv->pv_pmap != pmap) {
2316                 pv_put(pv);
2317                 spin_lock(&pmap->pm_spin);
2318                 info->retry = 1;
2319                 return(-1);
2320         }
2321         r = pmap_release_pv(pv, NULL);
2322         spin_lock(&pmap->pm_spin);
2323         return(r);
2324 }
2325
2326 /*
2327  * Called with held (i.e. also locked) pv.  This function will dispose of
2328  * the lock along with the pv.
2329  *
2330  * If the caller already holds the locked parent page table for pv it
2331  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2332  * pass NULL for pvp.
2333  */
2334 static int
2335 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp)
2336 {
2337         vm_page_t p;
2338
2339         /*
2340          * The pmap is currently not spinlocked, pv is held+locked.
2341          * Remove the pv's page from its parent's page table.  The
2342          * parent's page table page's wire_count will be decremented.
2343          */
2344         pmap_remove_pv_pte(pv, pvp, NULL);
2345
2346         /*
2347          * Terminal pvs are unhooked from their vm_pages.  Because
2348          * terminal pages aren't page table pages they aren't wired
2349          * by us, so we have to be sure not to unwire them either.
2350          */
2351         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2352                 pmap_remove_pv_page(pv);
2353                 goto skip;
2354         }
2355
2356         /*
2357          * We leave the top-level page table page cached, wired, and
2358          * mapped in the pmap until the dtor function (pmap_puninit())
2359          * gets called.
2360          *
2361          * Since we are leaving the top-level pv intact we need
2362          * to break out of what would otherwise be an infinite loop.
2363          */
2364         if (pv->pv_pindex == pmap_pml4_pindex()) {
2365                 pv_put(pv);
2366                 return(-1);
2367         }
2368
2369         /*
2370          * For page table pages (other than the top-level page),
2371          * remove and free the vm_page.  The representitive mapping
2372          * removed above by pmap_remove_pv_pte() did not undo the
2373          * last wire_count so we have to do that as well.
2374          */
2375         p = pmap_remove_pv_page(pv);
2376         vm_page_busy_wait(p, FALSE, "pmaprl");
2377         if (p->wire_count != 1) {
2378                 kprintf("p->wire_count was %016lx %d\n",
2379                         pv->pv_pindex, p->wire_count);
2380         }
2381         KKASSERT(p->wire_count == 1);
2382         KKASSERT(p->flags & PG_UNMANAGED);
2383
2384         vm_page_unwire(p, 0);
2385         KKASSERT(p->wire_count == 0);
2386
2387         /*
2388          * Theoretically this page, if not the pml4 page, should contain
2389          * all-zeros.  But its just too dangerous to mark it PG_ZERO.  Free
2390          * normally.
2391          */
2392         vm_page_free(p);
2393 skip:
2394         pv_free(pv);
2395         return 0;
2396 }
2397
2398 /*
2399  * This function will remove the pte associated with a pv from its parent.
2400  * Terminal pv's are supported.  The removal will be interlocked if info
2401  * is non-NULL.  The caller must dispose of pv instead of just unlocking
2402  * it.
2403  *
2404  * The wire count will be dropped on the parent page table.  The wire
2405  * count on the page being removed (pv->pv_m) from the parent page table
2406  * is NOT touched.  Note that terminal pages will not have any additional
2407  * wire counts while page table pages will have at least one representing
2408  * the mapping, plus others representing sub-mappings.
2409  *
2410  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2411  *       pages and user page table and terminal pages.
2412  *
2413  * The pv must be locked.
2414  *
2415  * XXX must lock parent pv's if they exist to remove pte XXX
2416  */
2417 static
2418 void
2419 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
2420 {
2421         vm_pindex_t ptepindex = pv->pv_pindex;
2422         pmap_t pmap = pv->pv_pmap;
2423         vm_page_t p;
2424         int gotpvp = 0;
2425
2426         KKASSERT(pmap);
2427
2428         if (ptepindex == pmap_pml4_pindex()) {
2429                 /*
2430                  * We are the top level pml4 table, there is no parent.
2431                  */
2432                 p = pmap->pm_pmlpv->pv_m;
2433         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2434                 /*
2435                  * Remove a PDP page from the pml4e.  This can only occur
2436                  * with user page tables.  We do not have to lock the
2437                  * pml4 PV so just ignore pvp.
2438                  */
2439                 vm_pindex_t pml4_pindex;
2440                 vm_pindex_t pdp_index;
2441                 pml4_entry_t *pdp;
2442
2443                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2444                 if (pvp == NULL) {
2445                         pml4_pindex = pmap_pml4_pindex();
2446                         pvp = pv_get(pv->pv_pmap, pml4_pindex);
2447                         KKASSERT(pvp);
2448                         gotpvp = 1;
2449                 }
2450                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2451                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2452                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2453                 *pdp = 0;
2454                 KKASSERT(info == NULL);
2455         } else if (ptepindex >= pmap_pd_pindex(0)) {
2456                 /*
2457                  * Remove a PD page from the pdp
2458                  *
2459                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2460                  *                   of a simple pmap because it stops at
2461                  *                   the PD page.
2462                  */
2463                 vm_pindex_t pdp_pindex;
2464                 vm_pindex_t pd_index;
2465                 pdp_entry_t *pd;
2466
2467                 pd_index = ptepindex - pmap_pd_pindex(0);
2468
2469                 if (pvp == NULL) {
2470                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2471                                      (pd_index >> NPML4EPGSHIFT);
2472                         pvp = pv_get(pv->pv_pmap, pdp_pindex);
2473                         if (pvp)
2474                                 gotpvp = 1;
2475                 }
2476                 if (pvp) {
2477                         pd = pv_pte_lookup(pvp, pd_index &
2478                                                 ((1ul << NPDPEPGSHIFT) - 1));
2479                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2480                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2481                         *pd = 0;
2482                 } else {
2483                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2484                         p = pv->pv_m;           /* degenerate test later */
2485                 }
2486                 KKASSERT(info == NULL);
2487         } else if (ptepindex >= pmap_pt_pindex(0)) {
2488                 /*
2489                  *  Remove a PT page from the pd
2490                  */
2491                 vm_pindex_t pd_pindex;
2492                 vm_pindex_t pt_index;
2493                 pd_entry_t *pt;
2494
2495                 pt_index = ptepindex - pmap_pt_pindex(0);
2496
2497                 if (pvp == NULL) {
2498                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2499                                     (pt_index >> NPDPEPGSHIFT);
2500                         pvp = pv_get(pv->pv_pmap, pd_pindex);
2501                         KKASSERT(pvp);
2502                         gotpvp = 1;
2503                 }
2504                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2505                 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2506                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2507                 *pt = 0;
2508                 KKASSERT(info == NULL);
2509         } else {
2510                 /*
2511                  * Remove a PTE from the PT page
2512                  *
2513                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2514                  *       pv is a pte_pv so we can safely lock pt_pv.
2515                  *
2516                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2517                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2518                  *       terminal ptes.
2519                  */
2520                 vm_pindex_t pt_pindex;
2521                 pt_entry_t *ptep;
2522                 pt_entry_t pte;
2523                 vm_offset_t va;
2524
2525                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2526                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2527
2528                 if (ptepindex >= NUPTE_USER) {
2529                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2530                         KKASSERT(pvp == NULL);
2531                 } else {
2532                         if (pvp == NULL) {
2533                                 pt_pindex = NUPTE_TOTAL +
2534                                             (ptepindex >> NPDPEPGSHIFT);
2535                                 pvp = pv_get(pv->pv_pmap, pt_pindex);
2536                                 KKASSERT(pvp);
2537                                 gotpvp = 1;
2538                         }
2539                         ptep = pv_pte_lookup(pvp, ptepindex &
2540                                                   ((1ul << NPDPEPGSHIFT) - 1));
2541                 }
2542
2543                 if (info)
2544                         pmap_inval_interlock(info, pmap, va);
2545                 pte = pte_load_clear(ptep);
2546                 if (info)
2547                         pmap_inval_deinterlock(info, pmap);
2548                 else
2549                         cpu_invlpg((void *)va);
2550
2551                 /*
2552                  * Now update the vm_page_t
2553                  */
2554                 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) !=
2555                     (pmap->pmap_bits[PG_MANAGED_IDX]|pmap->pmap_bits[PG_V_IDX])) {
2556                         kprintf("remove_pte badpte %016lx %016lx %d\n",
2557                                 pte, pv->pv_pindex,
2558                                 pv->pv_pindex < pmap_pt_pindex(0));
2559                 }
2560                 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2561                 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2562                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2563                         p = pv->pv_m;
2564                 else
2565                         p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2566                 /* p = pv->pv_m; */
2567
2568                 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2569                         if (pmap_track_modified(ptepindex))
2570                                 vm_page_dirty(p);
2571                 }
2572                 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2573                         vm_page_flag_set(p, PG_REFERENCED);
2574                 }
2575                 if (pte & pmap->pmap_bits[PG_W_IDX])
2576                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
2577                 if (pte & pmap->pmap_bits[PG_G_IDX])
2578                         cpu_invlpg((void *)va);
2579         }
2580
2581         /*
2582          * Unwire the parent page table page.  The wire_count cannot go below
2583          * 1 here because the parent page table page is itself still mapped.
2584          *
2585          * XXX remove the assertions later.
2586          */
2587         KKASSERT(pv->pv_m == p);
2588         if (pvp && vm_page_unwire_quick(pvp->pv_m))
2589                 panic("pmap_remove_pv_pte: Insufficient wire_count");
2590
2591         if (gotpvp)
2592                 pv_put(pvp);
2593 }
2594
2595 /*
2596  * Remove the vm_page association to a pv.  The pv must be locked.
2597  */
2598 static
2599 vm_page_t
2600 pmap_remove_pv_page(pv_entry_t pv)
2601 {
2602         vm_page_t m;
2603
2604         m = pv->pv_m;
2605         KKASSERT(m);
2606         vm_page_spin_lock(m);
2607         pv->pv_m = NULL;
2608         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2609         pmap_page_stats_deleting(m);
2610         /*
2611         if (m->object)
2612                 atomic_add_int(&m->object->agg_pv_list_count, -1);
2613         */
2614         if (TAILQ_EMPTY(&m->md.pv_list))
2615                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2616         vm_page_spin_unlock(m);
2617         return(m);
2618 }
2619
2620 /*
2621  * Grow the number of kernel page table entries, if needed.
2622  *
2623  * This routine is always called to validate any address space
2624  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
2625  * space below KERNBASE.
2626  */
2627 void
2628 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2629 {
2630         vm_paddr_t paddr;
2631         vm_offset_t ptppaddr;
2632         vm_page_t nkpg;
2633         pd_entry_t *pt, newpt;
2634         pdp_entry_t newpd;
2635         int update_kernel_vm_end;
2636
2637         /*
2638          * bootstrap kernel_vm_end on first real VM use
2639          */
2640         if (kernel_vm_end == 0) {
2641                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2642                 nkpt = 0;
2643                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2644                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2645                                         ~(PAGE_SIZE * NPTEPG - 1);
2646                         nkpt++;
2647                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2648                                 kernel_vm_end = kernel_map.max_offset;
2649                                 break;
2650                         }
2651                 }
2652         }
2653
2654         /*
2655          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
2656          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
2657          * do not want to force-fill 128G worth of page tables.
2658          */
2659         if (kstart < KERNBASE) {
2660                 if (kstart > kernel_vm_end)
2661                         kstart = kernel_vm_end;
2662                 KKASSERT(kend <= KERNBASE);
2663                 update_kernel_vm_end = 1;
2664         } else {
2665                 update_kernel_vm_end = 0;
2666         }
2667
2668         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2669         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2670
2671         if (kend - 1 >= kernel_map.max_offset)
2672                 kend = kernel_map.max_offset;
2673
2674         while (kstart < kend) {
2675                 pt = pmap_pt(&kernel_pmap, kstart);
2676                 if (pt == NULL) {
2677                         /* We need a new PDP entry */
2678                         nkpg = vm_page_alloc(NULL, nkpt,
2679                                              VM_ALLOC_NORMAL |
2680                                              VM_ALLOC_SYSTEM |
2681                                              VM_ALLOC_INTERRUPT);
2682                         if (nkpg == NULL) {
2683                                 panic("pmap_growkernel: no memory to grow "
2684                                       "kernel");
2685                         }
2686                         paddr = VM_PAGE_TO_PHYS(nkpg);
2687                         if ((nkpg->flags & PG_ZERO) == 0)
2688                                 pmap_zero_page(paddr);
2689                         vm_page_flag_clear(nkpg, PG_ZERO);
2690                         newpd = (pdp_entry_t)
2691                             (paddr |
2692                             kernel_pmap.pmap_bits[PG_V_IDX] |
2693                             kernel_pmap.pmap_bits[PG_RW_IDX] |
2694                             kernel_pmap.pmap_bits[PG_A_IDX] |
2695                             kernel_pmap.pmap_bits[PG_M_IDX]);
2696                         *pmap_pd(&kernel_pmap, kstart) = newpd;
2697                         nkpt++;
2698                         continue; /* try again */
2699                 }
2700                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2701                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
2702                                  ~(PAGE_SIZE * NPTEPG - 1);
2703                         if (kstart - 1 >= kernel_map.max_offset) {
2704                                 kstart = kernel_map.max_offset;
2705                                 break;
2706                         }
2707                         continue;
2708                 }
2709
2710                 /*
2711                  * This index is bogus, but out of the way
2712                  */
2713                 nkpg = vm_page_alloc(NULL, nkpt,
2714                                      VM_ALLOC_NORMAL |
2715                                      VM_ALLOC_SYSTEM |
2716                                      VM_ALLOC_INTERRUPT);
2717                 if (nkpg == NULL)
2718                         panic("pmap_growkernel: no memory to grow kernel");
2719
2720                 vm_page_wire(nkpg);
2721                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2722                 pmap_zero_page(ptppaddr);
2723                 vm_page_flag_clear(nkpg, PG_ZERO);
2724                 newpt = (pd_entry_t) (ptppaddr |
2725                     kernel_pmap.pmap_bits[PG_V_IDX] |
2726                     kernel_pmap.pmap_bits[PG_RW_IDX] |
2727                     kernel_pmap.pmap_bits[PG_A_IDX] |
2728                     kernel_pmap.pmap_bits[PG_M_IDX]);
2729                 *pmap_pt(&kernel_pmap, kstart) = newpt;
2730                 nkpt++;
2731
2732                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2733                           ~(PAGE_SIZE * NPTEPG - 1);
2734
2735                 if (kstart - 1 >= kernel_map.max_offset) {
2736                         kstart = kernel_map.max_offset;
2737                         break;
2738                 }
2739         }
2740
2741         /*
2742          * Only update kernel_vm_end for areas below KERNBASE.
2743          */
2744         if (update_kernel_vm_end && kernel_vm_end < kstart)
2745                 kernel_vm_end = kstart;
2746 }
2747
2748 /*
2749  *      Add a reference to the specified pmap.
2750  */
2751 void
2752 pmap_reference(pmap_t pmap)
2753 {
2754         if (pmap != NULL) {
2755                 lwkt_gettoken(&pmap->pm_token);
2756                 ++pmap->pm_count;
2757                 lwkt_reltoken(&pmap->pm_token);
2758         }
2759 }
2760
2761 /***************************************************
2762  * page management routines.
2763  ***************************************************/
2764
2765 /*
2766  * Hold a pv without locking it
2767  */
2768 static void
2769 pv_hold(pv_entry_t pv)
2770 {
2771         atomic_add_int(&pv->pv_hold, 1);
2772 }
2773
2774 /*
2775  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
2776  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
2777  * the pv properly.
2778  *
2779  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2780  * pv list via its page) must be held by the caller.
2781  */
2782 static int
2783 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2784 {
2785         u_int count;
2786
2787         /*
2788          * Critical path shortcut expects pv to already have one ref
2789          * (for the pv->pv_pmap).
2790          */
2791         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
2792 #ifdef PMAP_DEBUG
2793                 pv->pv_func = func;
2794                 pv->pv_line = lineno;
2795 #endif
2796                 return TRUE;
2797         }
2798
2799         for (;;) {
2800                 count = pv->pv_hold;
2801                 cpu_ccfence();
2802                 if ((count & PV_HOLD_LOCKED) == 0) {
2803                         if (atomic_cmpset_int(&pv->pv_hold, count,
2804                                               (count + 1) | PV_HOLD_LOCKED)) {
2805 #ifdef PMAP_DEBUG
2806                                 pv->pv_func = func;
2807                                 pv->pv_line = lineno;
2808 #endif
2809                                 return TRUE;
2810                         }
2811                 } else {
2812                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2813                                 return FALSE;
2814                 }
2815                 /* retry */
2816         }
2817 }
2818
2819 /*
2820  * Drop a previously held pv_entry which could not be locked, allowing its
2821  * destruction.
2822  *
2823  * Must not be called with a spinlock held as we might zfree() the pv if it
2824  * is no longer associated with a pmap and this was the last hold count.
2825  */
2826 static void
2827 pv_drop(pv_entry_t pv)
2828 {
2829         u_int count;
2830
2831         for (;;) {
2832                 count = pv->pv_hold;
2833                 cpu_ccfence();
2834                 KKASSERT((count & PV_HOLD_MASK) > 0);
2835                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2836                          (PV_HOLD_LOCKED | 1));
2837                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2838                         if ((count & PV_HOLD_MASK) == 1) {
2839 #ifdef PMAP_DEBUG2
2840                                 if (pmap_enter_debug > 0) {
2841                                         --pmap_enter_debug;
2842                                         kprintf("pv_drop: free pv %p\n", pv);
2843                                 }
2844 #endif
2845                                 KKASSERT(count == 1);
2846                                 KKASSERT(pv->pv_pmap == NULL);
2847                                 zfree(pvzone, pv);
2848                         }
2849                         return;
2850                 }
2851                 /* retry */
2852         }
2853 }
2854
2855 /*
2856  * Find or allocate the requested PV entry, returning a locked, held pv.
2857  *
2858  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
2859  * for the caller and one representing the pmap and vm_page association.
2860  *
2861  * If (*isnew) is zero, the returned pv will have only one hold count.
2862  *
2863  * Since both associations can only be adjusted while the pv is locked,
2864  * together they represent just one additional hold.
2865  */
2866 static
2867 pv_entry_t
2868 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2869 {
2870         pv_entry_t pv;
2871         pv_entry_t pnew = NULL;
2872
2873         spin_lock(&pmap->pm_spin);
2874         for (;;) {
2875                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2876                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2877                                                         pindex);
2878                 }
2879                 if (pv == NULL) {
2880                         if (pnew == NULL) {
2881                                 spin_unlock(&pmap->pm_spin);
2882                                 pnew = zalloc(pvzone);
2883                                 spin_lock(&pmap->pm_spin);
2884                                 continue;
2885                         }
2886                         pnew->pv_pmap = pmap;
2887                         pnew->pv_pindex = pindex;
2888                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
2889 #ifdef PMAP_DEBUG
2890                         pnew->pv_func = func;
2891                         pnew->pv_line = lineno;
2892 #endif
2893                         pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2894                         ++pmap->pm_generation;
2895                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
2896                         spin_unlock(&pmap->pm_spin);
2897                         *isnew = 1;
2898                         return(pnew);
2899                 }
2900                 if (pnew) {
2901                         spin_unlock(&pmap->pm_spin);
2902                         zfree(pvzone, pnew);
2903                         pnew = NULL;
2904                         spin_lock(&pmap->pm_spin);
2905                         continue;
2906                 }
2907                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2908                         spin_unlock(&pmap->pm_spin);
2909                 } else {
2910                         spin_unlock(&pmap->pm_spin);
2911                         _pv_lock(pv PMAP_DEBUG_COPY);
2912                 }
2913                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2914                         *isnew = 0;
2915                         return(pv);
2916                 }
2917                 pv_put(pv);
2918                 spin_lock(&pmap->pm_spin);
2919         }
2920 }
2921
2922 /*
2923  * Find the requested PV entry, returning a locked+held pv or NULL
2924  */
2925 static
2926 pv_entry_t
2927 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2928 {
2929         pv_entry_t pv;
2930
2931         spin_lock(&pmap->pm_spin);
2932         for (;;) {
2933                 /*
2934                  * Shortcut cache
2935                  */
2936                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2937                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2938                                                         pindex);
2939                 }
2940                 if (pv == NULL) {
2941                         spin_unlock(&pmap->pm_spin);
2942                         return NULL;
2943                 }
2944                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2945                         spin_unlock(&pmap->pm_spin);
2946                 } else {
2947                         spin_unlock(&pmap->pm_spin);
2948                         _pv_lock(pv PMAP_DEBUG_COPY);
2949                 }
2950                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2951                         pv_cache(pv, pindex);
2952                         return(pv);
2953                 }
2954                 pv_put(pv);
2955                 spin_lock(&pmap->pm_spin);
2956         }
2957 }
2958
2959 /*
2960  * Lookup, hold, and attempt to lock (pmap,pindex).
2961  *
2962  * If the entry does not exist NULL is returned and *errorp is set to 0
2963  *
2964  * If the entry exists and could be successfully locked it is returned and
2965  * errorp is set to 0.
2966  *
2967  * If the entry exists but could NOT be successfully locked it is returned
2968  * held and *errorp is set to 1.
2969  */
2970 static
2971 pv_entry_t
2972 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2973 {
2974         pv_entry_t pv;
2975
2976         spin_lock_shared(&pmap->pm_spin);
2977         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2978                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2979         if (pv == NULL) {
2980                 spin_unlock_shared(&pmap->pm_spin);
2981                 *errorp = 0;
2982                 return NULL;
2983         }
2984         if (pv_hold_try(pv)) {
2985                 pv_cache(pv, pindex);
2986                 spin_unlock_shared(&pmap->pm_spin);
2987                 *errorp = 0;
2988                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
2989                 return(pv);     /* lock succeeded */
2990         }
2991         spin_unlock_shared(&pmap->pm_spin);
2992         *errorp = 1;
2993         return (pv);            /* lock failed */
2994 }
2995
2996 /*
2997  * Find the requested PV entry, returning a held pv or NULL
2998  */
2999 static
3000 pv_entry_t
3001 pv_find(pmap_t pmap, vm_pindex_t pindex)
3002 {
3003         pv_entry_t pv;
3004
3005         spin_lock_shared(&pmap->pm_spin);
3006
3007         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3008                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3009         if (pv == NULL) {
3010                 spin_unlock_shared(&pmap->pm_spin);
3011                 return NULL;
3012         }
3013         pv_hold(pv);
3014         pv_cache(pv, pindex);
3015         spin_unlock_shared(&pmap->pm_spin);
3016         return(pv);
3017 }
3018
3019 /*
3020  * Lock a held pv, keeping the hold count
3021  */
3022 static
3023 void
3024 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3025 {
3026         u_int count;
3027
3028         for (;;) {
3029                 count = pv->pv_hold;
3030                 cpu_ccfence();
3031                 if ((count & PV_HOLD_LOCKED) == 0) {
3032                         if (atomic_cmpset_int(&pv->pv_hold, count,
3033                                               count | PV_HOLD_LOCKED)) {
3034 #ifdef PMAP_DEBUG
3035                                 pv->pv_func = func;
3036                                 pv->pv_line = lineno;
3037 #endif
3038                                 return;
3039                         }
3040                         continue;
3041                 }
3042                 tsleep_interlock(pv, 0);
3043                 if (atomic_cmpset_int(&pv->pv_hold, count,
3044                                       count | PV_HOLD_WAITING)) {
3045 #ifdef PMAP_DEBUG
3046                         kprintf("pv waiting on %s:%d\n",
3047                                         pv->pv_func, pv->pv_line);
3048 #endif
3049                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
3050                 }
3051                 /* retry */
3052         }
3053 }
3054
3055 /*
3056  * Unlock a held and locked pv, keeping the hold count.
3057  */
3058 static
3059 void
3060 pv_unlock(pv_entry_t pv)
3061 {
3062         u_int count;
3063
3064         for (;;) {
3065                 count = pv->pv_hold;
3066                 cpu_ccfence();
3067                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3068                          (PV_HOLD_LOCKED | 1));
3069                 if (atomic_cmpset_int(&pv->pv_hold, count,
3070                                       count &
3071                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3072                         if (count & PV_HOLD_WAITING)
3073                                 wakeup(pv);
3074                         break;
3075                 }
3076         }
3077 }
3078
3079 /*
3080  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3081  * and the hold count drops to zero we will free it.
3082  *
3083  * Caller should not hold any spin locks.  We are protected from hold races
3084  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3085  * lock held.  A pv cannot be located otherwise.
3086  */
3087 static
3088 void
3089 pv_put(pv_entry_t pv)
3090 {
3091 #ifdef PMAP_DEBUG2
3092         if (pmap_enter_debug > 0) {
3093                 --pmap_enter_debug;
3094                 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3095         }
3096 #endif
3097
3098         /*
3099          * Fast - shortcut most common condition
3100          */
3101         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3102                 return;
3103
3104         /*
3105          * Slow
3106          */
3107         pv_unlock(pv);
3108         pv_drop(pv);
3109 }
3110
3111 /*
3112  * Remove the pmap association from a pv, require that pv_m already be removed,
3113  * then unlock and drop the pv.  Any pte operations must have already been
3114  * completed.  This call may result in a last-drop which will physically free
3115  * the pv.
3116  *
3117  * Removing the pmap association entails an additional drop.
3118  *
3119  * pv must be exclusively locked on call and will be disposed of on return.
3120  */
3121 static
3122 void
3123 pv_free(pv_entry_t pv)
3124 {
3125         pmap_t pmap;
3126
3127         KKASSERT(pv->pv_m == NULL);
3128         KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3129         if ((pmap = pv->pv_pmap) != NULL) {
3130                 spin_lock(&pmap->pm_spin);
3131                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3132                 ++pmap->pm_generation;
3133                 if (pmap->pm_pvhint == pv)
3134                         pmap->pm_pvhint = NULL;
3135                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3136                 pv->pv_pmap = NULL;
3137                 pv->pv_pindex = 0;
3138                 spin_unlock(&pmap->pm_spin);
3139
3140                 /*
3141                  * Try to shortcut three atomic ops, otherwise fall through
3142                  * and do it normally.  Drop two refs and the lock all in
3143                  * one go.
3144                  */
3145                 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3146 #ifdef PMAP_DEBUG2
3147                         if (pmap_enter_debug > 0) {
3148                                 --pmap_enter_debug;
3149                                 kprintf("pv_free: free pv %p\n", pv);
3150                         }
3151 #endif
3152                         zfree(pvzone, pv);
3153                         return;
3154                 }
3155                 pv_drop(pv);    /* ref for pv_pmap */
3156         }
3157         pv_put(pv);
3158 }
3159
3160 /*
3161  * This routine is very drastic, but can save the system
3162  * in a pinch.
3163  */
3164 void
3165 pmap_collect(void)
3166 {
3167         int i;
3168         vm_page_t m;
3169         static int warningdone=0;
3170
3171         if (pmap_pagedaemon_waken == 0)
3172                 return;
3173         pmap_pagedaemon_waken = 0;
3174         if (warningdone < 5) {
3175                 kprintf("pmap_collect: collecting pv entries -- "
3176                         "suggest increasing PMAP_SHPGPERPROC\n");
3177                 warningdone++;
3178         }
3179
3180         for (i = 0; i < vm_page_array_size; i++) {
3181                 m = &vm_page_array[i];
3182                 if (m->wire_count || m->hold_count)
3183                         continue;
3184                 if (vm_page_busy_try(m, TRUE) == 0) {
3185                         if (m->wire_count == 0 && m->hold_count == 0) {
3186                                 pmap_remove_all(m);
3187                         }
3188                         vm_page_wakeup(m);
3189                 }
3190         }
3191 }
3192
3193 /*
3194  * Scan the pmap for active page table entries and issue a callback.
3195  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3196  * its parent page table.
3197  *
3198  * pte_pv will be NULL if the page or page table is unmanaged.
3199  * pt_pv will point to the page table page containing the pte for the page.
3200  *
3201  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3202  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3203  *       process pmap's PD and page to the callback function.  This can be
3204  *       confusing because the pt_pv is really a pd_pv, and the target page
3205  *       table page is simply aliased by the pmap and not owned by it.
3206  *
3207  * It is assumed that the start and end are properly rounded to the page size.
3208  *
3209  * It is assumed that PD pages and above are managed and thus in the RB tree,
3210  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3211  */
3212 struct pmap_scan_info {
3213         struct pmap *pmap;
3214         vm_offset_t sva;
3215         vm_offset_t eva;
3216         vm_pindex_t sva_pd_pindex;
3217         vm_pindex_t eva_pd_pindex;
3218         void (*func)(pmap_t, struct pmap_scan_info *,
3219                      pv_entry_t, pv_entry_t, int, vm_offset_t,
3220                      pt_entry_t *, void *);
3221         void *arg;
3222         int doinval;
3223         struct pmap_inval_info inval;
3224 };
3225
3226 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3227 static int pmap_scan_callback(pv_entry_t pv, void *data);
3228
3229 static void
3230 pmap_scan(struct pmap_scan_info *info)
3231 {
3232         struct pmap *pmap = info->pmap;
3233         pv_entry_t pd_pv;       /* A page directory PV */
3234         pv_entry_t pt_pv;       /* A page table PV */
3235         pv_entry_t pte_pv;      /* A page table entry PV */
3236         pt_entry_t *ptep;
3237         pt_entry_t oldpte;
3238         struct pv_entry dummy_pv;
3239         int generation;
3240
3241         if (pmap == NULL)
3242                 return;
3243
3244         /*
3245          * Hold the token for stability; if the pmap is empty we have nothing
3246          * to do.
3247          */
3248         lwkt_gettoken(&pmap->pm_token);
3249 #if 0
3250         if (pmap->pm_stats.resident_count == 0) {
3251                 lwkt_reltoken(&pmap->pm_token);
3252                 return;
3253         }
3254 #endif
3255
3256         pmap_inval_init(&info->inval);
3257
3258 again:
3259         /*
3260          * Special handling for scanning one page, which is a very common
3261          * operation (it is?).
3262          *
3263          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3264          */
3265         if (info->sva + PAGE_SIZE == info->eva) {
3266                 generation = pmap->pm_generation;
3267                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3268                         /*
3269                          * Kernel mappings do not track wire counts on
3270                          * page table pages and only maintain pd_pv and
3271                          * pte_pv levels so pmap_scan() works.
3272                          */
3273                         pt_pv = NULL;
3274                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3275                         ptep = vtopte(info->sva);
3276                 } else {
3277                         /*
3278                          * User pages which are unmanaged will not have a
3279                          * pte_pv.  User page table pages which are unmanaged
3280                          * (shared from elsewhere) will also not have a pt_pv.
3281                          * The func() callback will pass both pte_pv and pt_pv
3282                          * as NULL in that case.
3283                          */
3284                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3285                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3286                         if (pt_pv == NULL) {
3287                                 KKASSERT(pte_pv == NULL);
3288                                 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3289                                 if (pd_pv) {
3290                                         ptep = pv_pte_lookup(pd_pv,
3291                                                     pmap_pt_index(info->sva));
3292                                         if (*ptep) {
3293                                                 info->func(pmap, info,
3294                                                      NULL, pd_pv, 1,
3295                                                      info->sva, ptep,
3296                                                      info->arg);
3297                                         }
3298                                         pv_put(pd_pv);
3299                                 }
3300                                 goto fast_skip;
3301                         }
3302                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3303                 }
3304
3305                 /*
3306                  * NOTE: *ptep can't be ripped out from under us if we hold
3307                  *       pte_pv locked, but bits can change.  However, there is
3308                  *       a race where another thread may be inserting pte_pv
3309                  *       and setting *ptep just after our pte_pv lookup fails.
3310                  *
3311                  *       In this situation we can end up with a NULL pte_pv
3312                  *       but find that we have a managed *ptep.  We explicitly
3313                  *       check for this race.
3314                  */
3315                 oldpte = *ptep;
3316                 cpu_ccfence();
3317                 if (oldpte == 0) {
3318                         /*
3319                          * Unlike the pv_find() case below we actually
3320                          * acquired a locked pv in this case so any
3321                          * race should have been resolved.  It is expected
3322                          * to not exist.
3323                          */
3324                         KKASSERT(pte_pv == NULL);
3325                 } else if (pte_pv) {
3326                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3327                                            pmap->pmap_bits[PG_V_IDX])) ==
3328                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
3329                                  pmap->pmap_bits[PG_V_IDX]),
3330                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3331                              "generation %d/%d",
3332                             *ptep, oldpte, info->sva, pte_pv,
3333                             generation, pmap->pm_generation));
3334                         info->func(pmap, info, pte_pv, pt_pv, 0,
3335                                    info->sva, ptep, info->arg);
3336                 } else {
3337                         /*
3338                          * Check for insertion race
3339                          */
3340                         if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3341                             pt_pv) {
3342                                 pte_pv = pv_find(pmap,
3343                                                  pmap_pte_pindex(info->sva));
3344                                 if (pte_pv) {
3345                                         pv_drop(pte_pv);
3346                                         pv_put(pt_pv);
3347                                         kprintf("pmap_scan: RACE1 "
3348                                                 "%016jx, %016lx\n",
3349                                                 info->sva, oldpte);
3350                                         goto again;
3351                                 }
3352                         }
3353
3354                         /*
3355                          * Didn't race
3356                          */
3357                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3358                                            pmap->pmap_bits[PG_V_IDX])) ==
3359                             pmap->pmap_bits[PG_V_IDX],
3360                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3361                              "generation %d/%d",
3362                             *ptep, oldpte, info->sva,
3363                             generation, pmap->pm_generation));
3364                         info->func(pmap, info, NULL, pt_pv, 0,
3365                             info->sva, ptep, info->arg);
3366                 }
3367                 if (pt_pv)
3368                         pv_put(pt_pv);
3369 fast_skip:
3370                 pmap_inval_done(&info->inval);
3371                 lwkt_reltoken(&pmap->pm_token);
3372                 return;
3373         }
3374
3375         /*
3376          * Nominal scan case, RB_SCAN() for PD pages and iterate from
3377          * there.
3378          */
3379         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3380         info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3381
3382         if (info->sva >= VM_MAX_USER_ADDRESS) {
3383                 /*
3384                  * The kernel does not currently maintain any pv_entry's for
3385                  * higher-level page tables.
3386                  */
3387                 bzero(&dummy_pv, sizeof(dummy_pv));
3388                 dummy_pv.pv_pindex = info->sva_pd_pindex;
3389                 spin_lock(&pmap->pm_spin);
3390                 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3391                         pmap_scan_callback(&dummy_pv, info);
3392                         ++dummy_pv.pv_pindex;
3393                 }
3394                 spin_unlock(&pmap->pm_spin);
3395         } else {
3396                 /*
3397                  * User page tables maintain local PML4, PDP, and PD
3398                  * pv_entry's at the very least.  PT pv's might be
3399                  * unmanaged and thus not exist.  PTE pv's might be
3400                  * unmanaged and thus not exist.
3401                  */
3402                 spin_lock(&pmap->pm_spin);
3403                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3404                         pmap_scan_cmp, pmap_scan_callback, info);
3405                 spin_unlock(&pmap->pm_spin);
3406         }
3407         pmap_inval_done(&info->inval);
3408         lwkt_reltoken(&pmap->pm_token);
3409 }
3410
3411 /*
3412  * WARNING! pmap->pm_spin held
3413  */
3414 static int
3415 pmap_scan_cmp(pv_entry_t pv, void *data)
3416 {
3417         struct pmap_scan_info *info = data;
3418         if (pv->pv_pindex < info->sva_pd_pindex)
3419                 return(-1);
3420         if (pv->pv_pindex >= info->eva_pd_pindex)
3421                 return(1);
3422         return(0);
3423 }
3424
3425 /*
3426  * WARNING! pmap->pm_spin held
3427  */
3428 static int
3429 pmap_scan_callback(pv_entry_t pv, void *data)
3430 {
3431         struct pmap_scan_info *info = data;
3432         struct pmap *pmap = info->pmap;
3433         pv_entry_t pd_pv;       /* A page directory PV */
3434         pv_entry_t pt_pv;       /* A page table PV */
3435         pv_entry_t pte_pv;      /* A page table entry PV */
3436         pt_entry_t *ptep;
3437         pt_entry_t oldpte;
3438         vm_offset_t sva;
3439         vm_offset_t eva;
3440         vm_offset_t va_next;
3441         vm_pindex_t pd_pindex;
3442         int error;
3443         int generation;
3444
3445         /*
3446          * Pull the PD pindex from the pv before releasing the spinlock.
3447          *
3448          * WARNING: pv is faked for kernel pmap scans.
3449          */
3450         pd_pindex = pv->pv_pindex;
3451         spin_unlock(&pmap->pm_spin);
3452         pv = NULL;      /* invalid after spinlock unlocked */
3453
3454         /*
3455          * Calculate the page range within the PD.  SIMPLE pmaps are
3456          * direct-mapped for the entire 2^64 address space.  Normal pmaps
3457          * reflect the user and kernel address space which requires
3458          * cannonicalization w/regards to converting pd_pindex's back
3459          * into addresses.
3460          */
3461         sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3462         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3463             (sva & PML4_SIGNMASK)) {
3464                 sva |= PML4_SIGNMASK;
3465         }
3466         eva = sva + NBPDP;      /* can overflow */
3467         if (sva < info->sva)
3468                 sva = info->sva;
3469         if (eva < info->sva || eva > info->eva)
3470                 eva = info->eva;
3471
3472         /*
3473          * NOTE: kernel mappings do not track page table pages, only
3474          *       terminal pages.
3475          *
3476          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3477          *       However, for the scan to be efficient we try to
3478          *       cache items top-down.
3479          */
3480         pd_pv = NULL;
3481         pt_pv = NULL;
3482
3483         for (; sva < eva; sva = va_next) {
3484                 if (sva >= VM_MAX_USER_ADDRESS) {
3485                         if (pt_pv) {
3486                                 pv_put(pt_pv);
3487                                 pt_pv = NULL;
3488                         }
3489                         goto kernel_skip;
3490                 }
3491
3492                 /*
3493                  * PD cache (degenerate case if we skip).  It is possible
3494                  * for the PD to not exist due to races.  This is ok.
3495                  */
3496                 if (pd_pv == NULL) {
3497                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3498                 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3499                         pv_put(pd_pv);
3500                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3501                 }
3502                 if (pd_pv == NULL) {
3503                         va_next = (sva + NBPDP) & ~PDPMASK;
3504                         if (va_next < sva)
3505                                 va_next = eva;
3506                         continue;
3507                 }
3508
3509                 /*
3510                  * PT cache
3511                  */
3512                 if (pt_pv == NULL) {
3513                         if (pd_pv) {
3514                                 pv_put(pd_pv);
3515                                 pd_pv = NULL;
3516                         }
3517                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3518                 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3519                         if (pd_pv) {
3520                                 pv_put(pd_pv);
3521                                 pd_pv = NULL;
3522                         }
3523                         pv_put(pt_pv);
3524                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3525                 }
3526
3527                 /*
3528                  * If pt_pv is NULL we either have an shared page table
3529                  * page and must issue a callback specific to that case,
3530                  * or there is no page table page.
3531                  *
3532                  * Either way we can skip the page table page.
3533                  */
3534                 if (pt_pv == NULL) {
3535                         /*
3536                          * Possible unmanaged (shared from another pmap)
3537                          * page table page.
3538                          */
3539                         if (pd_pv == NULL)
3540                                 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3541                         KKASSERT(pd_pv != NULL);
3542                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3543                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3544                                 info->func(pmap, info, NULL, pd_pv, 1,
3545                                            sva, ptep, info->arg);
3546                         }
3547
3548                         /*
3549                          * Done, move to next page table page.
3550                          */
3551                         va_next = (sva + NBPDR) & ~PDRMASK;
3552                         if (va_next < sva)
3553                                 va_next = eva;
3554                         continue;
3555                 }
3556
3557                 /*
3558                  * From this point in the loop testing pt_pv for non-NULL
3559                  * means we are in UVM, else if it is NULL we are in KVM.
3560                  *
3561                  * Limit our scan to either the end of the va represented
3562                  * by the current page table page, or to the end of the
3563                  * range being removed.
3564                  */
3565 kernel_skip:
3566                 va_next = (sva + NBPDR) & ~PDRMASK;
3567                 if (va_next < sva)
3568                         va_next = eva;
3569                 if (va_next > eva)
3570                         va_next = eva;
3571
3572                 /*
3573                  * Scan the page table for pages.  Some pages may not be
3574                  * managed (might not have a pv_entry).
3575                  *
3576                  * There is no page table management for kernel pages so
3577                  * pt_pv will be NULL in that case, but otherwise pt_pv
3578                  * is non-NULL, locked, and referenced.
3579                  */
3580
3581                 /*
3582                  * At this point a non-NULL pt_pv means a UVA, and a NULL
3583                  * pt_pv means a KVA.
3584                  */
3585                 if (pt_pv)
3586                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3587                 else
3588                         ptep = vtopte(sva);
3589
3590                 while (sva < va_next) {
3591                         /*
3592                          * Acquire the related pte_pv, if any.  If *ptep == 0
3593                          * the related pte_pv should not exist, but if *ptep
3594                          * is not zero the pte_pv may or may not exist (e.g.
3595                          * will not exist for an unmanaged page).
3596                          *
3597                          * However a multitude of races are possible here.
3598                          *
3599                          * In addition, the (pt_pv, pte_pv) lock order is
3600                          * backwards, so we have to be careful in aquiring
3601                          * a properly locked pte_pv.
3602                          */
3603                         generation = pmap->pm_generation;
3604                         if (pt_pv) {
3605                                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3606                                                     &error);
3607                                 if (error) {
3608                                         if (pd_pv) {
3609                                                 pv_put(pd_pv);
3610                                                 pd_pv = NULL;
3611                                         }
3612                                         pv_put(pt_pv);   /* must be non-NULL */
3613                                         pt_pv = NULL;
3614                                         pv_lock(pte_pv); /* safe to block now */
3615                                         pv_put(pte_pv);
3616                                         pte_pv = NULL;
3617                                         pt_pv = pv_get(pmap,
3618                                                        pmap_pt_pindex(sva));
3619                                         /*
3620                                          * pt_pv reloaded, need new ptep
3621                                          */
3622                                         KKASSERT(pt_pv != NULL);
3623                                         ptep = pv_pte_lookup(pt_pv,
3624                                                         pmap_pte_index(sva));
3625                                         continue;
3626                                 }
3627                         } else {
3628                                 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3629                         }
3630
3631                         /*
3632                          * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3633                          */
3634                         oldpte = *ptep;
3635                         if (oldpte == 0) {
3636                                 if (pte_pv) {
3637                                         kprintf("Unexpected non-NULL pte_pv "
3638                                                 "%p pt_pv %p "
3639                                                 "*ptep = %016lx/%016lx\n",
3640                                                 pte_pv, pt_pv, *ptep, oldpte);
3641                                         panic("Unexpected non-NULL pte_pv");
3642                                 }
3643                                 sva += PAGE_SIZE;
3644                                 ++ptep;
3645                                 continue;
3646                         }
3647
3648                         /*
3649                          * Ready for the callback.  The locked pte_pv (if any)
3650                          * is consumed by the callback.  pte_pv will exist if
3651                          *  the page is managed, and will not exist if it
3652                          * isn't.
3653                          */
3654                         if (pte_pv) {
3655                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3656                                     (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3657                                     ("badC *ptep %016lx/%016lx sva %016lx "
3658                                     "pte_pv %p pm_generation %d/%d",
3659                                     *ptep, oldpte, sva, pte_pv,
3660                                     generation, pmap->pm_generation));
3661                                 info->func(pmap, info, pte_pv, pt_pv, 0,
3662                                     sva, ptep, info->arg);
3663                         } else {
3664                                 /*
3665                                  * Check for insertion race.  Since there is no
3666                                  * pte_pv to guard us it is possible for us
3667                                  * to race another thread doing an insertion.
3668                                  * Our lookup misses the pte_pv but our *ptep
3669                                  * check sees the inserted pte.
3670                                  *
3671                                  * XXX panic case seems to occur within a
3672                                  * vm_fork() of /bin/sh, which frankly
3673                                  * shouldn't happen since no other threads
3674                                  * should be inserting to our pmap in that
3675                                  * situation.  Removing, possibly.  Inserting,
3676                                  * shouldn't happen.
3677                                  */
3678                                 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3679                                     pt_pv) {
3680                                         pte_pv = pv_find(pmap,
3681                                                          pmap_pte_pindex(sva));
3682                                         if (pte_pv) {
3683                                                 pv_drop(pte_pv);
3684                                                 kprintf("pmap_scan: RACE2 "
3685                                                         "%016jx, %016lx\n",
3686                                                         sva, oldpte);
3687                                                 continue;
3688                                         }
3689                                 }
3690
3691                                 /*
3692                                  * Didn't race
3693                                  */
3694                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3695                                     pmap->pmap_bits[PG_V_IDX],
3696                                     ("badD *ptep %016lx/%016lx sva %016lx "
3697                                     "pte_pv NULL pm_generation %d/%d",
3698                                      *ptep, oldpte, sva,
3699                                      generation, pmap->pm_generation));
3700                                 info->func(pmap, info, NULL, pt_pv, 0,
3701                                     sva, ptep, info->arg);
3702                         }
3703                         pte_pv = NULL;
3704                         sva += PAGE_SIZE;
3705                         ++ptep;
3706                 }
3707                 lwkt_yield();
3708         }
3709         if (pd_pv) {
3710                 pv_put(pd_pv);
3711                 pd_pv = NULL;
3712         }
3713         if (pt_pv) {
3714                 pv_put(pt_pv);
3715                 pt_pv = NULL;
3716         }
3717         lwkt_yield();
3718
3719         /*
3720          * Relock before returning.
3721          */
3722         spin_lock(&pmap->pm_spin);
3723         return (0);
3724 }
3725
3726 void
3727 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3728 {
3729         struct pmap_scan_info info;
3730
3731         info.pmap = pmap;
3732         info.sva = sva;
3733         info.eva = eva;
3734         info.func = pmap_remove_callback;
3735         info.arg = NULL;
3736         info.doinval = 1;       /* normal remove requires pmap inval */
3737         pmap_scan(&info);
3738 }
3739
3740 static void
3741 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3742 {
3743         struct pmap_scan_info info;
3744
3745         info.pmap = pmap;
3746         info.sva = sva;
3747         info.eva = eva;
3748         info.func = pmap_remove_callback;
3749         info.arg = NULL;
3750         info.doinval = 0;       /* normal remove requires pmap inval */
3751         pmap_scan(&info);
3752 }
3753
3754 static void
3755 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
3756                      pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3757                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3758 {
3759         pt_entry_t pte;
3760
3761         if (pte_pv) {
3762                 /*
3763                  * This will also drop pt_pv's wire_count. Note that
3764                  * terminal pages are not wired based on mmu presence.
3765                  */
3766                 if (info->doinval)
3767                         pmap_remove_pv_pte(pte_pv, pt_pv, &info->inval);
3768                 else
3769                         pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3770                 pmap_remove_pv_page(pte_pv);
3771                 pv_free(pte_pv);
3772         } else if (sharept == 0) {
3773                 /*
3774                  * Unmanaged page
3775                  *
3776                  * pt_pv's wire_count is still bumped by unmanaged pages
3777                  * so we must decrement it manually.
3778                  */
3779                 if (info->doinval)
3780                         pmap_inval_interlock(&info->inval, pmap, va);
3781                 pte = pte_load_clear(ptep);
3782                 if (info->doinval)
3783                         pmap_inval_deinterlock(&info->inval, pmap);
3784                 if (pte & pmap->pmap_bits[PG_W_IDX])
3785                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
3786                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3787                 if (vm_page_unwire_quick(pt_pv->pv_m))
3788                         panic("pmap_remove: insufficient wirecount");
3789         } else {
3790                 /*
3791                  * Unmanaged page table, pt_pv is actually the pd_pv
3792                  * for our pmap (not the share object pmap).
3793                  *
3794                  * We have to unwire the target page table page and we
3795                  * have to unwire our page directory page.
3796                  */
3797                 if (info->doinval)
3798                         pmap_inval_interlock(&info->inval, pmap, va);
3799                 pte = pte_load_clear(ptep);
3800                 if (info->doinval)
3801                         pmap_inval_deinterlock(&info->inval, pmap);
3802                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3803                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
3804                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3805                         panic("pmap_remove: shared pgtable1 bad wirecount");
3806                 if (vm_page_unwire_quick(pt_pv->pv_m))
3807                         panic("pmap_remove: shared pgtable2 bad wirecount");
3808         }
3809 }
3810
3811 /*
3812  * Removes this physical page from all physical maps in which it resides.
3813  * Reflects back modify bits to the pager.
3814  *
3815  * This routine may not be called from an interrupt.
3816  */
3817 static
3818 void
3819 pmap_remove_all(vm_page_t m)
3820 {
3821         struct pmap_inval_info info;
3822         pv_entry_t pv;
3823
3824         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
3825                 return;
3826
3827         pmap_inval_init(&info);
3828         vm_page_spin_lock(m);
3829         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3830                 KKASSERT(pv->pv_m == m);
3831                 if (pv_hold_try(pv)) {
3832                         vm_page_spin_unlock(m);
3833                 } else {
3834                         vm_page_spin_unlock(m);
3835                         pv_lock(pv);
3836                 }
3837                 if (pv->pv_m != m) {
3838                         pv_put(pv);
3839                         vm_page_spin_lock(m);
3840                         continue;
3841                 }
3842                 /*
3843                  * Holding no spinlocks, pv is locked.
3844                  */
3845                 pmap_remove_pv_pte(pv, NULL, &info);
3846                 pmap_remove_pv_page(pv);
3847                 pv_free(pv);
3848                 vm_page_spin_lock(m);
3849         }
3850         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
3851         vm_page_spin_unlock(m);
3852         pmap_inval_done(&info);
3853 }
3854
3855 /*
3856  * Set the physical protection on the specified range of this map
3857  * as requested.  This function is typically only used for debug watchpoints
3858  * and COW pages.
3859  *
3860  * This function may not be called from an interrupt if the map is
3861  * not the kernel_pmap.
3862  *
3863  * NOTE!  For shared page table pages we just unmap the page.
3864  */
3865 void</