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