1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * Copyright (c) 1994 John S. Dyson 6 * Copyright (c) 1994 David Greenman 7 * Copyright (c) 2003 Peter Wemm 8 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu> 9 * Copyright (c) 2008, 2009 The DragonFly Project. 10 * Copyright (c) 2008, 2009 Jordan Gordeev. 11 * All rights reserved. 12 * 13 * This code is derived from software contributed to Berkeley by 14 * the Systems Programming Group of the University of Utah Computer 15 * Science Department and William Jolitz of UUNET Technologies Inc. 16 * 17 * Redistribution and use in source and binary forms, with or without 18 * modification, are permitted provided that the following conditions 19 * are met: 20 * 1. Redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer. 22 * 2. Redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution. 25 * 3. All advertising materials mentioning features or use of this software 26 * must display the following acknowledgement: 27 * This product includes software developed by the University of 28 * California, Berkeley and its contributors. 29 * 4. Neither the name of the University nor the names of its contributors 30 * may be used to endorse or promote products derived from this software 31 * without specific prior written permission. 32 * 33 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 34 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 35 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 36 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 37 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 38 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 39 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 40 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 41 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 42 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 43 * SUCH DAMAGE. 44 * 45 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 46 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $ 47 */ 48 49 /* 50 * Manages physical address maps. 51 * 52 * In most cases the vm_token must be held when manipulating a user pmap 53 * or elements within a vm_page, and the kvm_token must be held when 54 * manipulating the kernel pmap. Operations on user pmaps may require 55 * additional synchronization. 56 * 57 * In some cases the caller may hold the required tokens to prevent pmap 58 * functions from blocking on those same tokens. This typically only works 59 * for lookup-style operations. 60 */ 61 62 #if JG 63 #include "opt_pmap.h" 64 #endif 65 #include "opt_msgbuf.h" 66 67 #include <sys/param.h> 68 #include <sys/systm.h> 69 #include <sys/kernel.h> 70 #include <sys/proc.h> 71 #include <sys/msgbuf.h> 72 #include <sys/vmmeter.h> 73 #include <sys/mman.h> 74 #include <sys/vmspace.h> 75 76 #include <vm/vm.h> 77 #include <vm/vm_param.h> 78 #include <sys/sysctl.h> 79 #include <sys/lock.h> 80 #include <vm/vm_kern.h> 81 #include <vm/vm_page.h> 82 #include <vm/vm_map.h> 83 #include <vm/vm_object.h> 84 #include <vm/vm_extern.h> 85 #include <vm/vm_pageout.h> 86 #include <vm/vm_pager.h> 87 #include <vm/vm_zone.h> 88 89 #include <sys/user.h> 90 #include <sys/thread2.h> 91 #include <sys/sysref2.h> 92 93 #include <machine/cputypes.h> 94 #include <machine/md_var.h> 95 #include <machine/specialreg.h> 96 #include <machine/smp.h> 97 #include <machine/globaldata.h> 98 #include <machine/pmap.h> 99 #include <machine/pmap_inval.h> 100 101 #include <ddb/ddb.h> 102 103 #include <stdio.h> 104 #include <assert.h> 105 #include <stdlib.h> 106 107 #define PMAP_KEEP_PDIRS 108 #ifndef PMAP_SHPGPERPROC 109 #define PMAP_SHPGPERPROC 200 110 #endif 111 112 #if defined(DIAGNOSTIC) 113 #define PMAP_DIAGNOSTIC 114 #endif 115 116 #define MINPV 2048 117 118 #if !defined(PMAP_DIAGNOSTIC) 119 #define PMAP_INLINE __inline 120 #else 121 #define PMAP_INLINE 122 #endif 123 124 /* 125 * Get PDEs and PTEs for user/kernel address space 126 */ 127 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va); 128 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 129 130 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & VPTE_V) != 0) 131 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & VPTE_WIRED) != 0) 132 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & VPTE_M) != 0) 133 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & VPTE_A) != 0) 134 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & VPTE_V) != 0) 135 136 /* 137 * Given a map and a machine independent protection code, 138 * convert to a vax protection code. 139 */ 140 #define pte_prot(m, p) \ 141 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)]) 142 static int protection_codes[8]; 143 144 struct pmap kernel_pmap; 145 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list); 146 147 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 148 149 static vm_object_t kptobj; 150 151 static int nkpt; 152 153 static uint64_t KPDphys; /* phys addr of kernel level 2 */ 154 uint64_t KPDPphys; /* phys addr of kernel level 3 */ 155 uint64_t KPML4phys; /* phys addr of kernel level 4 */ 156 157 158 /* 159 * Data for the pv entry allocation mechanism 160 */ 161 static vm_zone_t pvzone; 162 static struct vm_zone pvzone_store; 163 static struct vm_object pvzone_obj; 164 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0; 165 static int pmap_pagedaemon_waken = 0; 166 static struct pv_entry *pvinit; 167 168 /* 169 * All those kernel PT submaps that BSD is so fond of 170 */ 171 pt_entry_t *CMAP1 = 0, *ptmmap; 172 caddr_t CADDR1 = 0; 173 static pt_entry_t *msgbufmap; 174 175 uint64_t KPTphys; 176 177 static PMAP_INLINE void free_pv_entry (pv_entry_t pv); 178 static pv_entry_t get_pv_entry (void); 179 static void i386_protection_init (void); 180 static __inline void pmap_clearbit (vm_page_t m, int bit); 181 182 static void pmap_remove_all (vm_page_t m); 183 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq, 184 vm_offset_t sva); 185 static void pmap_remove_page (struct pmap *pmap, vm_offset_t va); 186 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m, 187 vm_offset_t va); 188 static boolean_t pmap_testbit (vm_page_t m, int bit); 189 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va, 190 vm_page_t mpte, vm_page_t m); 191 192 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va); 193 194 static int pmap_release_free_page (pmap_t pmap, vm_page_t p); 195 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex); 196 #if JGPMAP32 197 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va); 198 #endif 199 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex); 200 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t); 201 202 /* 203 * pmap_pte_quick: 204 * 205 * Super fast pmap_pte routine best used when scanning the pv lists. 206 * This eliminates many course-grained invltlb calls. Note that many of 207 * the pv list scans are across different pmaps and it is very wasteful 208 * to do an entire invltlb when checking a single mapping. 209 * 210 * Should only be called while in a critical section. 211 */ 212 #if JGPMAP32 213 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va); 214 215 static pt_entry_t * 216 pmap_pte_quick(pmap_t pmap, vm_offset_t va) 217 { 218 return pmap_pte(pmap, va); 219 } 220 #endif 221 222 /* Return a non-clipped PD index for a given VA */ 223 static __inline vm_pindex_t 224 pmap_pde_pindex(vm_offset_t va) 225 { 226 return va >> PDRSHIFT; 227 } 228 229 /* Return various clipped indexes for a given VA */ 230 static __inline vm_pindex_t 231 pmap_pte_index(vm_offset_t va) 232 { 233 234 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1)); 235 } 236 237 static __inline vm_pindex_t 238 pmap_pde_index(vm_offset_t va) 239 { 240 241 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1)); 242 } 243 244 static __inline vm_pindex_t 245 pmap_pdpe_index(vm_offset_t va) 246 { 247 248 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1)); 249 } 250 251 static __inline vm_pindex_t 252 pmap_pml4e_index(vm_offset_t va) 253 { 254 255 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1)); 256 } 257 258 /* Return a pointer to the PML4 slot that corresponds to a VA */ 259 static __inline pml4_entry_t * 260 pmap_pml4e(pmap_t pmap, vm_offset_t va) 261 { 262 263 return (&pmap->pm_pml4[pmap_pml4e_index(va)]); 264 } 265 266 /* Return a pointer to the PDP slot that corresponds to a VA */ 267 static __inline pdp_entry_t * 268 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va) 269 { 270 pdp_entry_t *pdpe; 271 272 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & VPTE_FRAME); 273 return (&pdpe[pmap_pdpe_index(va)]); 274 } 275 276 /* Return a pointer to the PDP slot that corresponds to a VA */ 277 static __inline pdp_entry_t * 278 pmap_pdpe(pmap_t pmap, vm_offset_t va) 279 { 280 pml4_entry_t *pml4e; 281 282 pml4e = pmap_pml4e(pmap, va); 283 if ((*pml4e & VPTE_V) == 0) 284 return NULL; 285 return (pmap_pml4e_to_pdpe(pml4e, va)); 286 } 287 288 /* Return a pointer to the PD slot that corresponds to a VA */ 289 static __inline pd_entry_t * 290 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va) 291 { 292 pd_entry_t *pde; 293 294 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & VPTE_FRAME); 295 return (&pde[pmap_pde_index(va)]); 296 } 297 298 /* Return a pointer to the PD slot that corresponds to a VA */ 299 static __inline pd_entry_t * 300 pmap_pde(pmap_t pmap, vm_offset_t va) 301 { 302 pdp_entry_t *pdpe; 303 304 pdpe = pmap_pdpe(pmap, va); 305 if (pdpe == NULL || (*pdpe & VPTE_V) == 0) 306 return NULL; 307 return (pmap_pdpe_to_pde(pdpe, va)); 308 } 309 310 /* Return a pointer to the PT slot that corresponds to a VA */ 311 static __inline pt_entry_t * 312 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va) 313 { 314 pt_entry_t *pte; 315 316 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & VPTE_FRAME); 317 return (&pte[pmap_pte_index(va)]); 318 } 319 320 /* Return a pointer to the PT slot that corresponds to a VA */ 321 static __inline pt_entry_t * 322 pmap_pte(pmap_t pmap, vm_offset_t va) 323 { 324 pd_entry_t *pde; 325 326 pde = pmap_pde(pmap, va); 327 if (pde == NULL || (*pde & VPTE_V) == 0) 328 return NULL; 329 if ((*pde & VPTE_PS) != 0) /* compat with i386 pmap_pte() */ 330 return ((pt_entry_t *)pde); 331 return (pmap_pde_to_pte(pde, va)); 332 } 333 334 335 #if JGV 336 PMAP_INLINE pt_entry_t * 337 vtopte(vm_offset_t va) 338 { 339 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); 340 341 return (PTmap + ((va >> PAGE_SHIFT) & mask)); 342 } 343 344 static __inline pd_entry_t * 345 vtopde(vm_offset_t va) 346 { 347 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); 348 349 return (PDmap + ((va >> PDRSHIFT) & mask)); 350 } 351 #else 352 static PMAP_INLINE pt_entry_t * 353 vtopte(vm_offset_t va) 354 { 355 pt_entry_t *x; 356 x = pmap_pte(&kernel_pmap, va); 357 assert(x != NULL); 358 return x; 359 } 360 361 static __inline pd_entry_t * 362 vtopde(vm_offset_t va) 363 { 364 pd_entry_t *x; 365 x = pmap_pde(&kernel_pmap, va); 366 assert(x != NULL); 367 return x; 368 } 369 #endif 370 371 static uint64_t 372 allocpages(vm_paddr_t *firstaddr, int n) 373 { 374 uint64_t ret; 375 376 ret = *firstaddr; 377 #if JGV 378 bzero((void *)ret, n * PAGE_SIZE); 379 #endif 380 *firstaddr += n * PAGE_SIZE; 381 return (ret); 382 } 383 384 static void 385 create_pagetables(vm_paddr_t *firstaddr, int64_t ptov_offset) 386 { 387 int i; 388 pml4_entry_t *KPML4virt; 389 pdp_entry_t *KPDPvirt; 390 pd_entry_t *KPDvirt; 391 pt_entry_t *KPTvirt; 392 int kpml4i = pmap_pml4e_index(ptov_offset); 393 int kpdpi = pmap_pdpe_index(ptov_offset); 394 395 /* 396 * Calculate NKPT - number of kernel page tables. We have to 397 * accomodoate prealloction of the vm_page_array, dump bitmap, 398 * MSGBUF_SIZE, and other stuff. Be generous. 399 * 400 * Maxmem is in pages. 401 */ 402 nkpt = (Maxmem * (sizeof(struct vm_page) * 2) + MSGBUF_SIZE) / NBPDR; 403 404 /* 405 * Allocate pages 406 */ 407 KPML4phys = allocpages(firstaddr, 1); 408 KPDPphys = allocpages(firstaddr, NKPML4E); 409 KPDphys = allocpages(firstaddr, NKPDPE); 410 KPTphys = allocpages(firstaddr, nkpt); 411 412 KPML4virt = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys); 413 KPDPvirt = (pdp_entry_t *)PHYS_TO_DMAP(KPDPphys); 414 KPDvirt = (pd_entry_t *)PHYS_TO_DMAP(KPDphys); 415 KPTvirt = (pt_entry_t *)PHYS_TO_DMAP(KPTphys); 416 417 bzero(KPML4virt, 1 * PAGE_SIZE); 418 bzero(KPDPvirt, NKPML4E * PAGE_SIZE); 419 bzero(KPDvirt, NKPDPE * PAGE_SIZE); 420 bzero(KPTvirt, nkpt * PAGE_SIZE); 421 422 /* Now map the page tables at their location within PTmap */ 423 for (i = 0; i < nkpt; i++) { 424 KPDvirt[i] = KPTphys + (i << PAGE_SHIFT); 425 KPDvirt[i] |= VPTE_R | VPTE_W | VPTE_V; 426 } 427 428 /* And connect up the PD to the PDP */ 429 for (i = 0; i < NKPDPE; i++) { 430 KPDPvirt[i + kpdpi] = KPDphys + (i << PAGE_SHIFT); 431 KPDPvirt[i + kpdpi] |= VPTE_R | VPTE_W | VPTE_V; 432 } 433 434 /* And recursively map PML4 to itself in order to get PTmap */ 435 KPML4virt[PML4PML4I] = KPML4phys; 436 KPML4virt[PML4PML4I] |= VPTE_R | VPTE_W | VPTE_V; 437 438 /* Connect the KVA slot up to the PML4 */ 439 KPML4virt[kpml4i] = KPDPphys; 440 KPML4virt[kpml4i] |= VPTE_R | VPTE_W | VPTE_V; 441 } 442 443 /* 444 * Bootstrap the system enough to run with virtual memory. 445 * 446 * On the i386 this is called after mapping has already been enabled 447 * and just syncs the pmap module with what has already been done. 448 * [We can't call it easily with mapping off since the kernel is not 449 * mapped with PA == VA, hence we would have to relocate every address 450 * from the linked base (virtual) address "KERNBASE" to the actual 451 * (physical) address starting relative to 0] 452 */ 453 void 454 pmap_bootstrap(vm_paddr_t *firstaddr, int64_t ptov_offset) 455 { 456 vm_offset_t va; 457 pt_entry_t *pte; 458 459 /* 460 * Create an initial set of page tables to run the kernel in. 461 */ 462 create_pagetables(firstaddr, ptov_offset); 463 464 virtual_start = KvaStart + *firstaddr; 465 virtual_end = KvaEnd; 466 467 /* 468 * Initialize protection array. 469 */ 470 i386_protection_init(); 471 472 /* 473 * The kernel's pmap is statically allocated so we don't have to use 474 * pmap_create, which is unlikely to work correctly at this part of 475 * the boot sequence (XXX and which no longer exists). 476 */ 477 kernel_pmap.pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys); 478 kernel_pmap.pm_count = 1; 479 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */ 480 TAILQ_INIT(&kernel_pmap.pm_pvlist); 481 482 /* 483 * Reserve some special page table entries/VA space for temporary 484 * mapping of pages. 485 */ 486 #define SYSMAP(c, p, v, n) \ 487 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 488 489 va = virtual_start; 490 pte = pmap_pte(&kernel_pmap, va); 491 492 /* 493 * CMAP1/CMAP2 are used for zeroing and copying pages. 494 */ 495 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 496 497 #if JGV 498 /* 499 * Crashdump maps. 500 */ 501 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 502 #endif 503 504 /* 505 * ptvmmap is used for reading arbitrary physical pages via 506 * /dev/mem. 507 */ 508 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 509 510 /* 511 * msgbufp is used to map the system message buffer. 512 * XXX msgbufmap is not used. 513 */ 514 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 515 atop(round_page(MSGBUF_SIZE))) 516 517 virtual_start = va; 518 519 *CMAP1 = 0; 520 521 cpu_invltlb(); 522 } 523 524 /* 525 * Initialize the pmap module. 526 * Called by vm_init, to initialize any structures that the pmap 527 * system needs to map virtual memory. 528 * pmap_init has been enhanced to support in a fairly consistant 529 * way, discontiguous physical memory. 530 */ 531 void 532 pmap_init(void) 533 { 534 int i; 535 int initial_pvs; 536 537 /* 538 * object for kernel page table pages 539 */ 540 /* JG I think the number can be arbitrary */ 541 kptobj = vm_object_allocate(OBJT_DEFAULT, 5); 542 543 /* 544 * Allocate memory for random pmap data structures. Includes the 545 * pv_head_table. 546 */ 547 548 for(i = 0; i < vm_page_array_size; i++) { 549 vm_page_t m; 550 551 m = &vm_page_array[i]; 552 TAILQ_INIT(&m->md.pv_list); 553 m->md.pv_list_count = 0; 554 } 555 556 /* 557 * init the pv free list 558 */ 559 initial_pvs = vm_page_array_size; 560 if (initial_pvs < MINPV) 561 initial_pvs = MINPV; 562 pvzone = &pvzone_store; 563 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map, 564 initial_pvs * sizeof (struct pv_entry)); 565 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit, 566 initial_pvs); 567 568 /* 569 * Now it is safe to enable pv_table recording. 570 */ 571 pmap_initialized = TRUE; 572 } 573 574 /* 575 * Initialize the address space (zone) for the pv_entries. Set a 576 * high water mark so that the system can recover from excessive 577 * numbers of pv entries. 578 */ 579 void 580 pmap_init2(void) 581 { 582 int shpgperproc = PMAP_SHPGPERPROC; 583 584 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 585 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 586 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 587 pv_entry_high_water = 9 * (pv_entry_max / 10); 588 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1); 589 } 590 591 592 /*************************************************** 593 * Low level helper routines..... 594 ***************************************************/ 595 596 /* 597 * The modification bit is not tracked for any pages in this range. XXX 598 * such pages in this maps should always use pmap_k*() functions and not 599 * be managed anyhow. 600 * 601 * XXX User and kernel address spaces are independant for virtual kernels, 602 * this function only applies to the kernel pmap. 603 */ 604 static int 605 pmap_track_modified(pmap_t pmap, vm_offset_t va) 606 { 607 if (pmap != &kernel_pmap) 608 return 1; 609 if ((va < clean_sva) || (va >= clean_eva)) 610 return 1; 611 else 612 return 0; 613 } 614 615 /* 616 * Extract the physical page address associated with the map/VA pair. 617 * 618 * No requirements. 619 */ 620 vm_paddr_t 621 pmap_extract(pmap_t pmap, vm_offset_t va) 622 { 623 vm_paddr_t rtval; 624 pt_entry_t *pte; 625 pd_entry_t pde, *pdep; 626 627 lwkt_gettoken(&vm_token); 628 rtval = 0; 629 pdep = pmap_pde(pmap, va); 630 if (pdep != NULL) { 631 pde = *pdep; 632 if (pde) { 633 if ((pde & VPTE_PS) != 0) { 634 /* JGV */ 635 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK); 636 } else { 637 pte = pmap_pde_to_pte(pdep, va); 638 rtval = (*pte & VPTE_FRAME) | (va & PAGE_MASK); 639 } 640 } 641 } 642 lwkt_reltoken(&vm_token); 643 return rtval; 644 } 645 646 /* 647 * Routine: pmap_kextract 648 * Function: 649 * Extract the physical page address associated 650 * kernel virtual address. 651 */ 652 vm_paddr_t 653 pmap_kextract(vm_offset_t va) 654 { 655 pd_entry_t pde; 656 vm_paddr_t pa; 657 658 KKASSERT(va >= KvaStart && va < KvaEnd); 659 660 /* 661 * The DMAP region is not included in [KvaStart, KvaEnd) 662 */ 663 #if 0 664 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) { 665 pa = DMAP_TO_PHYS(va); 666 } else { 667 #endif 668 pde = *vtopde(va); 669 if (pde & VPTE_PS) { 670 /* JGV */ 671 pa = (pde & PG_PS_FRAME) | (va & PDRMASK); 672 } else { 673 /* 674 * Beware of a concurrent promotion that changes the 675 * PDE at this point! For example, vtopte() must not 676 * be used to access the PTE because it would use the 677 * new PDE. It is, however, safe to use the old PDE 678 * because the page table page is preserved by the 679 * promotion. 680 */ 681 pa = *pmap_pde_to_pte(&pde, va); 682 pa = (pa & VPTE_FRAME) | (va & PAGE_MASK); 683 } 684 #if 0 685 } 686 #endif 687 return pa; 688 } 689 690 /*************************************************** 691 * Low level mapping routines..... 692 ***************************************************/ 693 694 /* 695 * Enter a mapping into kernel_pmap. Mappings created in this fashion 696 * are not managed. Mappings must be immediately accessible on all cpus. 697 * 698 * Call pmap_inval_pte() to invalidate the virtual pte and clean out the 699 * real pmap and handle related races before storing the new vpte. 700 */ 701 void 702 pmap_kenter(vm_offset_t va, vm_paddr_t pa) 703 { 704 pt_entry_t *pte; 705 pt_entry_t npte; 706 707 KKASSERT(va >= KvaStart && va < KvaEnd); 708 npte = pa | VPTE_R | VPTE_W | VPTE_V; 709 pte = vtopte(va); 710 if (*pte & VPTE_V) 711 pmap_inval_pte(pte, &kernel_pmap, va); 712 *pte = npte; 713 } 714 715 /* 716 * Enter an unmanaged KVA mapping for the private use of the current 717 * cpu only. pmap_kenter_sync() may be called to make the mapping usable 718 * by other cpus. 719 * 720 * It is illegal for the mapping to be accessed by other cpus unleess 721 * pmap_kenter_sync*() is called. 722 */ 723 void 724 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa) 725 { 726 pt_entry_t *pte; 727 pt_entry_t npte; 728 729 KKASSERT(va >= KvaStart && va < KvaEnd); 730 731 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V; 732 pte = vtopte(va); 733 if (*pte & VPTE_V) 734 pmap_inval_pte_quick(pte, &kernel_pmap, va); 735 *pte = npte; 736 //cpu_invlpg((void *)va); 737 } 738 739 /* 740 * Synchronize a kvm mapping originally made for the private use on 741 * some other cpu so it can be used on all cpus. 742 * 743 * XXX add MADV_RESYNC to improve performance. 744 */ 745 void 746 pmap_kenter_sync(vm_offset_t va) 747 { 748 madvise((void *)va, PAGE_SIZE, MADV_INVAL); 749 } 750 751 /* 752 * Synchronize a kvm mapping originally made for the private use on 753 * some other cpu so it can be used on our cpu. Turns out to be the 754 * same madvise() call, because we have to sync the real pmaps anyway. 755 * 756 * XXX add MADV_RESYNC to improve performance. 757 */ 758 void 759 pmap_kenter_sync_quick(vm_offset_t va) 760 { 761 madvise((void *)va, PAGE_SIZE, MADV_INVAL); 762 } 763 764 /* 765 * Remove an unmanaged mapping created with pmap_kenter*(). 766 */ 767 void 768 pmap_kremove(vm_offset_t va) 769 { 770 pt_entry_t *pte; 771 772 KKASSERT(va >= KvaStart && va < KvaEnd); 773 774 pte = vtopte(va); 775 if (*pte & VPTE_V) 776 pmap_inval_pte(pte, &kernel_pmap, va); 777 *pte = 0; 778 } 779 780 /* 781 * Remove an unmanaged mapping created with pmap_kenter*() but synchronize 782 * only with this cpu. 783 * 784 * Unfortunately because we optimize new entries by testing VPTE_V later 785 * on, we actually still have to synchronize with all the cpus. XXX maybe 786 * store a junk value and test against 0 in the other places instead? 787 */ 788 void 789 pmap_kremove_quick(vm_offset_t va) 790 { 791 pt_entry_t *pte; 792 793 KKASSERT(va >= KvaStart && va < KvaEnd); 794 795 pte = vtopte(va); 796 if (*pte & VPTE_V) 797 pmap_inval_pte(pte, &kernel_pmap, va); /* NOT _quick */ 798 *pte = 0; 799 } 800 801 /* 802 * Used to map a range of physical addresses into kernel 803 * virtual address space. 804 * 805 * For now, VM is already on, we only need to map the 806 * specified memory. 807 */ 808 vm_offset_t 809 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot) 810 { 811 return PHYS_TO_DMAP(start); 812 } 813 814 815 /* 816 * Map a set of unmanaged VM pages into KVM. 817 */ 818 void 819 pmap_qenter(vm_offset_t va, vm_page_t *m, int count) 820 { 821 vm_offset_t end_va; 822 823 end_va = va + count * PAGE_SIZE; 824 KKASSERT(va >= KvaStart && end_va < KvaEnd); 825 826 while (va < end_va) { 827 pt_entry_t *pte; 828 829 pte = vtopte(va); 830 if (*pte & VPTE_V) 831 pmap_inval_pte(pte, &kernel_pmap, va); 832 *pte = VM_PAGE_TO_PHYS(*m) | VPTE_R | VPTE_W | VPTE_V; 833 va += PAGE_SIZE; 834 m++; 835 } 836 } 837 838 /* 839 * Undo the effects of pmap_qenter*(). 840 */ 841 void 842 pmap_qremove(vm_offset_t va, int count) 843 { 844 vm_offset_t end_va; 845 846 end_va = va + count * PAGE_SIZE; 847 KKASSERT(va >= KvaStart && end_va < KvaEnd); 848 849 while (va < end_va) { 850 pt_entry_t *pte; 851 852 pte = vtopte(va); 853 if (*pte & VPTE_V) 854 pmap_inval_pte(pte, &kernel_pmap, va); 855 *pte = 0; 856 va += PAGE_SIZE; 857 } 858 } 859 860 /* 861 * This routine works like vm_page_lookup() but also blocks as long as the 862 * page is busy. This routine does not busy the page it returns. 863 * 864 * Unless the caller is managing objects whos pages are in a known state, 865 * the call should be made with a critical section held so the page's object 866 * association remains valid on return. 867 */ 868 static vm_page_t 869 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex) 870 { 871 vm_page_t m; 872 873 do { 874 m = vm_page_lookup(object, pindex); 875 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp")); 876 877 return(m); 878 } 879 880 /* 881 * Create a new thread and optionally associate it with a (new) process. 882 * NOTE! the new thread's cpu may not equal the current cpu. 883 */ 884 void 885 pmap_init_thread(thread_t td) 886 { 887 /* enforce pcb placement */ 888 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1; 889 td->td_savefpu = &td->td_pcb->pcb_save; 890 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */ 891 } 892 893 /* 894 * This routine directly affects the fork perf for a process. 895 */ 896 void 897 pmap_init_proc(struct proc *p) 898 { 899 } 900 901 /* 902 * Dispose the UPAGES for a process that has exited. 903 * This routine directly impacts the exit perf of a process. 904 */ 905 void 906 pmap_dispose_proc(struct proc *p) 907 { 908 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p)); 909 } 910 911 /*************************************************** 912 * Page table page management routines..... 913 ***************************************************/ 914 915 static __inline int pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, 916 vm_page_t m); 917 918 /* 919 * This routine unholds page table pages, and if the hold count 920 * drops to zero, then it decrements the wire count. 921 * 922 * We must recheck that this is the last hold reference after busy-sleeping 923 * on the page. 924 */ 925 static int 926 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m) 927 { 928 while (vm_page_sleep_busy(m, FALSE, "pmuwpt")) 929 ; 930 KASSERT(m->queue == PQ_NONE, 931 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m)); 932 933 if (m->hold_count == 1) { 934 /* 935 * Unmap the page table page. 936 */ 937 //abort(); /* JG */ 938 vm_page_busy(m); 939 /* pmap_inval_add(info, pmap, -1); */ 940 941 if (m->pindex >= (NUPDE + NUPDPE)) { 942 /* PDP page */ 943 pml4_entry_t *pml4; 944 pml4 = pmap_pml4e(pmap, va); 945 *pml4 = 0; 946 } else if (m->pindex >= NUPDE) { 947 /* PD page */ 948 pdp_entry_t *pdp; 949 pdp = pmap_pdpe(pmap, va); 950 *pdp = 0; 951 } else { 952 /* PT page */ 953 pd_entry_t *pd; 954 pd = pmap_pde(pmap, va); 955 *pd = 0; 956 } 957 958 KKASSERT(pmap->pm_stats.resident_count > 0); 959 --pmap->pm_stats.resident_count; 960 961 if (pmap->pm_ptphint == m) 962 pmap->pm_ptphint = NULL; 963 964 if (m->pindex < NUPDE) { 965 /* We just released a PT, unhold the matching PD */ 966 vm_page_t pdpg; 967 968 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & VPTE_FRAME); 969 pmap_unwire_pte_hold(pmap, va, pdpg); 970 } 971 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) { 972 /* We just released a PD, unhold the matching PDP */ 973 vm_page_t pdppg; 974 975 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & VPTE_FRAME); 976 pmap_unwire_pte_hold(pmap, va, pdppg); 977 } 978 979 /* 980 * This was our last hold, the page had better be unwired 981 * after we decrement wire_count. 982 * 983 * FUTURE NOTE: shared page directory page could result in 984 * multiple wire counts. 985 */ 986 vm_page_unhold(m); 987 --m->wire_count; 988 KKASSERT(m->wire_count == 0); 989 --vmstats.v_wire_count; 990 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 991 vm_page_flash(m); 992 vm_page_free_zero(m); 993 return 1; 994 } else { 995 KKASSERT(m->hold_count > 1); 996 vm_page_unhold(m); 997 return 0; 998 } 999 } 1000 1001 static __inline int 1002 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m) 1003 { 1004 KKASSERT(m->hold_count > 0); 1005 if (m->hold_count > 1) { 1006 vm_page_unhold(m); 1007 return 0; 1008 } else { 1009 return _pmap_unwire_pte_hold(pmap, va, m); 1010 } 1011 } 1012 1013 /* 1014 * After removing a page table entry, this routine is used to 1015 * conditionally free the page, and manage the hold/wire counts. 1016 */ 1017 static int 1018 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 1019 { 1020 /* JG Use FreeBSD/amd64 or FreeBSD/i386 ptepde approaches? */ 1021 vm_pindex_t ptepindex; 1022 1023 if (mpte == NULL) { 1024 /* 1025 * page table pages in the kernel_pmap are not managed. 1026 */ 1027 if (pmap == &kernel_pmap) 1028 return(0); 1029 ptepindex = pmap_pde_pindex(va); 1030 if (pmap->pm_ptphint && 1031 (pmap->pm_ptphint->pindex == ptepindex)) { 1032 mpte = pmap->pm_ptphint; 1033 } else { 1034 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 1035 pmap->pm_ptphint = mpte; 1036 } 1037 } 1038 1039 return pmap_unwire_pte_hold(pmap, va, mpte); 1040 } 1041 1042 /* 1043 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we 1044 * just dummy it up so it works well enough for fork(). 1045 * 1046 * In DragonFly, process pmaps may only be used to manipulate user address 1047 * space, never kernel address space. 1048 */ 1049 void 1050 pmap_pinit0(struct pmap *pmap) 1051 { 1052 pmap_pinit(pmap); 1053 } 1054 1055 /* 1056 * Initialize a preallocated and zeroed pmap structure, 1057 * such as one in a vmspace structure. 1058 */ 1059 void 1060 pmap_pinit(struct pmap *pmap) 1061 { 1062 vm_page_t ptdpg; 1063 1064 /* 1065 * No need to allocate page table space yet but we do need a valid 1066 * page directory table. 1067 */ 1068 if (pmap->pm_pml4 == NULL) { 1069 pmap->pm_pml4 = 1070 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE); 1071 } 1072 1073 /* 1074 * Allocate an object for the ptes 1075 */ 1076 if (pmap->pm_pteobj == NULL) 1077 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1); 1078 1079 /* 1080 * Allocate the page directory page, unless we already have 1081 * one cached. If we used the cached page the wire_count will 1082 * already be set appropriately. 1083 */ 1084 if ((ptdpg = pmap->pm_pdirm) == NULL) { 1085 ptdpg = vm_page_grab(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I, 1086 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1087 pmap->pm_pdirm = ptdpg; 1088 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY); 1089 ptdpg->valid = VM_PAGE_BITS_ALL; 1090 if (ptdpg->wire_count == 0) 1091 ++vmstats.v_wire_count; 1092 ptdpg->wire_count = 1; 1093 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg)); 1094 } 1095 if ((ptdpg->flags & PG_ZERO) == 0) 1096 bzero(pmap->pm_pml4, PAGE_SIZE); 1097 1098 pmap->pm_count = 1; 1099 pmap->pm_active = 0; 1100 pmap->pm_ptphint = NULL; 1101 TAILQ_INIT(&pmap->pm_pvlist); 1102 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1103 pmap->pm_stats.resident_count = 1; 1104 } 1105 1106 /* 1107 * Clean up a pmap structure so it can be physically freed. This routine 1108 * is called by the vmspace dtor function. A great deal of pmap data is 1109 * left passively mapped to improve vmspace management so we have a bit 1110 * of cleanup work to do here. 1111 * 1112 * No requirements. 1113 */ 1114 void 1115 pmap_puninit(pmap_t pmap) 1116 { 1117 vm_page_t p; 1118 1119 KKASSERT(pmap->pm_active == 0); 1120 lwkt_gettoken(&vm_token); 1121 if ((p = pmap->pm_pdirm) != NULL) { 1122 KKASSERT(pmap->pm_pml4 != NULL); 1123 pmap_kremove((vm_offset_t)pmap->pm_pml4); 1124 p->wire_count--; 1125 vmstats.v_wire_count--; 1126 KKASSERT((p->flags & PG_BUSY) == 0); 1127 vm_page_busy(p); 1128 vm_page_free_zero(p); 1129 pmap->pm_pdirm = NULL; 1130 } 1131 lwkt_reltoken(&vm_token); 1132 if (pmap->pm_pml4) { 1133 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE); 1134 pmap->pm_pml4 = NULL; 1135 } 1136 if (pmap->pm_pteobj) { 1137 vm_object_deallocate(pmap->pm_pteobj); 1138 pmap->pm_pteobj = NULL; 1139 } 1140 } 1141 1142 /* 1143 * Wire in kernel global address entries. To avoid a race condition 1144 * between pmap initialization and pmap_growkernel, this procedure 1145 * adds the pmap to the master list (which growkernel scans to update), 1146 * then copies the template. 1147 * 1148 * In a virtual kernel there are no kernel global address entries. 1149 * 1150 * No requirements. 1151 */ 1152 void 1153 pmap_pinit2(struct pmap *pmap) 1154 { 1155 crit_enter(); 1156 lwkt_gettoken(&vm_token); 1157 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode); 1158 lwkt_reltoken(&vm_token); 1159 crit_exit(); 1160 } 1161 1162 /* 1163 * Attempt to release and free a vm_page in a pmap. Returns 1 on success, 1164 * 0 on failure (if the procedure had to sleep). 1165 * 1166 * When asked to remove the page directory page itself, we actually just 1167 * leave it cached so we do not have to incur the SMP inval overhead of 1168 * removing the kernel mapping. pmap_puninit() will take care of it. 1169 */ 1170 static int 1171 pmap_release_free_page(struct pmap *pmap, vm_page_t p) 1172 { 1173 /* 1174 * This code optimizes the case of freeing non-busy 1175 * page-table pages. Those pages are zero now, and 1176 * might as well be placed directly into the zero queue. 1177 */ 1178 if (vm_page_sleep_busy(p, FALSE, "pmaprl")) 1179 return 0; 1180 1181 vm_page_busy(p); 1182 1183 /* 1184 * Remove the page table page from the processes address space. 1185 */ 1186 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) { 1187 /* 1188 * We are the pml4 table itself. 1189 */ 1190 /* XXX anything to do here? */ 1191 } else if (p->pindex >= (NUPDE + NUPDPE)) { 1192 /* 1193 * We are a PDP page. 1194 * We look for the PML4 entry that points to us. 1195 */ 1196 vm_page_t m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I); 1197 KKASSERT(m4 != NULL); 1198 pml4_entry_t *pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4)); 1199 int idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG; 1200 KKASSERT(pml4[idx] != 0); 1201 pml4[idx] = 0; 1202 m4->hold_count--; 1203 /* JG What about wire_count? */ 1204 } else if (p->pindex >= NUPDE) { 1205 /* 1206 * We are a PD page. 1207 * We look for the PDP entry that points to us. 1208 */ 1209 vm_page_t m3 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG); 1210 KKASSERT(m3 != NULL); 1211 pdp_entry_t *pdp = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3)); 1212 int idx = (p->pindex - NUPDE) % NPDPEPG; 1213 KKASSERT(pdp[idx] != 0); 1214 pdp[idx] = 0; 1215 m3->hold_count--; 1216 /* JG What about wire_count? */ 1217 } else { 1218 /* We are a PT page. 1219 * We look for the PD entry that points to us. 1220 */ 1221 vm_page_t m2 = vm_page_lookup(pmap->pm_pteobj, NUPDE + p->pindex / NPDEPG); 1222 KKASSERT(m2 != NULL); 1223 pd_entry_t *pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2)); 1224 int idx = p->pindex % NPDEPG; 1225 pd[idx] = 0; 1226 m2->hold_count--; 1227 /* JG What about wire_count? */ 1228 } 1229 KKASSERT(pmap->pm_stats.resident_count > 0); 1230 --pmap->pm_stats.resident_count; 1231 1232 if (p->hold_count) { 1233 panic("pmap_release: freeing held page table page"); 1234 } 1235 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex)) 1236 pmap->pm_ptphint = NULL; 1237 1238 /* 1239 * We leave the top-level page table page cached, wired, and mapped in 1240 * the pmap until the dtor function (pmap_puninit()) gets called. 1241 * However, still clean it up so we can set PG_ZERO. 1242 */ 1243 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) { 1244 bzero(pmap->pm_pml4, PAGE_SIZE); 1245 vm_page_flag_set(p, PG_ZERO); 1246 vm_page_wakeup(p); 1247 } else { 1248 abort(); 1249 p->wire_count--; 1250 vmstats.v_wire_count--; 1251 /* JG eventually revert to using vm_page_free_zero() */ 1252 vm_page_free(p); 1253 } 1254 return 1; 1255 } 1256 1257 /* 1258 * this routine is called if the page table page is not 1259 * mapped correctly. 1260 */ 1261 static vm_page_t 1262 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex) 1263 { 1264 vm_page_t m, pdppg, pdpg; 1265 1266 /* 1267 * Find or fabricate a new pagetable page 1268 */ 1269 m = vm_page_grab(pmap->pm_pteobj, ptepindex, 1270 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY); 1271 1272 if ((m->flags & PG_ZERO) == 0) { 1273 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 1274 } 1275 1276 KASSERT(m->queue == PQ_NONE, 1277 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1278 1279 /* 1280 * Increment the hold count for the page we will be returning to 1281 * the caller. 1282 */ 1283 m->hold_count++; 1284 1285 if (m->wire_count == 0) 1286 vmstats.v_wire_count++; 1287 m->wire_count++; 1288 1289 /* 1290 * Map the pagetable page into the process address space, if 1291 * it isn't already there. 1292 */ 1293 1294 ++pmap->pm_stats.resident_count; 1295 1296 if (ptepindex >= (NUPDE + NUPDPE)) { 1297 pml4_entry_t *pml4; 1298 vm_pindex_t pml4index; 1299 1300 /* Wire up a new PDP page */ 1301 pml4index = ptepindex - (NUPDE + NUPDPE); 1302 pml4 = &pmap->pm_pml4[pml4index]; 1303 *pml4 = VM_PAGE_TO_PHYS(m) | VPTE_R | VPTE_W | VPTE_V | 1304 VPTE_A | VPTE_M; 1305 } else if (ptepindex >= NUPDE) { 1306 vm_pindex_t pml4index; 1307 vm_pindex_t pdpindex; 1308 pml4_entry_t *pml4; 1309 pdp_entry_t *pdp; 1310 1311 /* Wire up a new PD page */ 1312 pdpindex = ptepindex - NUPDE; 1313 pml4index = pdpindex >> NPML4EPGSHIFT; 1314 1315 pml4 = &pmap->pm_pml4[pml4index]; 1316 if ((*pml4 & VPTE_V) == 0) { 1317 /* Have to allocate a new PDP page, recurse */ 1318 if (_pmap_allocpte(pmap, NUPDE + NUPDPE + pml4index) 1319 == NULL) { 1320 --m->wire_count; 1321 vm_page_free(m); 1322 return (NULL); 1323 } 1324 } else { 1325 /* Add reference to the PDP page */ 1326 pdppg = PHYS_TO_VM_PAGE(*pml4 & VPTE_FRAME); 1327 pdppg->hold_count++; 1328 } 1329 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME); 1330 1331 /* Now find the pdp page */ 1332 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; 1333 KKASSERT(*pdp == 0); /* JG DEBUG64 */ 1334 *pdp = VM_PAGE_TO_PHYS(m) | VPTE_R | VPTE_W | VPTE_V | 1335 VPTE_A | VPTE_M; 1336 } else { 1337 vm_pindex_t pml4index; 1338 vm_pindex_t pdpindex; 1339 pml4_entry_t *pml4; 1340 pdp_entry_t *pdp; 1341 pd_entry_t *pd; 1342 1343 /* Wire up a new PT page */ 1344 pdpindex = ptepindex >> NPDPEPGSHIFT; 1345 pml4index = pdpindex >> NPML4EPGSHIFT; 1346 1347 /* First, find the pdp and check that its valid. */ 1348 pml4 = &pmap->pm_pml4[pml4index]; 1349 if ((*pml4 & VPTE_V) == 0) { 1350 /* We miss a PDP page. We ultimately need a PD page. 1351 * Recursively allocating a PD page will allocate 1352 * the missing PDP page and will also allocate 1353 * the PD page we need. 1354 */ 1355 /* Have to allocate a new PD page, recurse */ 1356 if (_pmap_allocpte(pmap, NUPDE + pdpindex) 1357 == NULL) { 1358 --m->wire_count; 1359 vm_page_free(m); 1360 return (NULL); 1361 } 1362 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME); 1363 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; 1364 } else { 1365 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME); 1366 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; 1367 if ((*pdp & VPTE_V) == 0) { 1368 /* Have to allocate a new PD page, recurse */ 1369 if (_pmap_allocpte(pmap, NUPDE + pdpindex) 1370 == NULL) { 1371 --m->wire_count; 1372 vm_page_free(m); 1373 return (NULL); 1374 } 1375 } else { 1376 /* Add reference to the PD page */ 1377 pdpg = PHYS_TO_VM_PAGE(*pdp & VPTE_FRAME); 1378 pdpg->hold_count++; 1379 } 1380 } 1381 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & VPTE_FRAME); 1382 1383 /* Now we know where the page directory page is */ 1384 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)]; 1385 KKASSERT(*pd == 0); /* JG DEBUG64 */ 1386 *pd = VM_PAGE_TO_PHYS(m) | VPTE_R | VPTE_W | VPTE_V | 1387 VPTE_A | VPTE_M; 1388 } 1389 1390 /* 1391 * Set the page table hint 1392 */ 1393 pmap->pm_ptphint = m; 1394 1395 m->valid = VM_PAGE_BITS_ALL; 1396 vm_page_flag_clear(m, PG_ZERO); 1397 vm_page_flag_set(m, PG_MAPPED); 1398 vm_page_wakeup(m); 1399 1400 return m; 1401 } 1402 1403 /* 1404 * Determine the page table page required to access the VA in the pmap 1405 * and allocate it if necessary. Return a held vm_page_t for the page. 1406 * 1407 * Only used with user pmaps. 1408 */ 1409 static vm_page_t 1410 pmap_allocpte(pmap_t pmap, vm_offset_t va) 1411 { 1412 vm_pindex_t ptepindex; 1413 pd_entry_t *pd; 1414 vm_page_t m; 1415 1416 /* 1417 * Calculate pagetable page index 1418 */ 1419 ptepindex = pmap_pde_pindex(va); 1420 1421 /* 1422 * Get the page directory entry 1423 */ 1424 pd = pmap_pde(pmap, va); 1425 1426 /* 1427 * This supports switching from a 2MB page to a 1428 * normal 4K page. 1429 */ 1430 if (pd != NULL && (*pd & (VPTE_PS | VPTE_V)) == (VPTE_PS | VPTE_V)) { 1431 panic("no promotion/demotion yet"); 1432 *pd = 0; 1433 pd = NULL; 1434 /*cpu_invltlb();*/ 1435 /*smp_invltlb();*/ 1436 } 1437 1438 /* 1439 * If the page table page is mapped, we just increment the 1440 * hold count, and activate it. 1441 */ 1442 if (pd != NULL && (*pd & VPTE_V) != 0) { 1443 /* YYY hint is used here on i386 */ 1444 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex); 1445 pmap->pm_ptphint = m; 1446 m->hold_count++; 1447 return m; 1448 } 1449 /* 1450 * Here if the pte page isn't mapped, or if it has been deallocated. 1451 */ 1452 return _pmap_allocpte(pmap, ptepindex); 1453 } 1454 1455 1456 /*************************************************** 1457 * Pmap allocation/deallocation routines. 1458 ***************************************************/ 1459 1460 /* 1461 * Release any resources held by the given physical map. 1462 * Called when a pmap initialized by pmap_pinit is being released. 1463 * Should only be called if the map contains no valid mappings. 1464 * 1465 * No requirements. 1466 */ 1467 static int pmap_release_callback(struct vm_page *p, void *data); 1468 1469 void 1470 pmap_release(struct pmap *pmap) 1471 { 1472 vm_object_t object = pmap->pm_pteobj; 1473 struct rb_vm_page_scan_info info; 1474 1475 KKASSERT(pmap != &kernel_pmap); 1476 1477 #if defined(DIAGNOSTIC) 1478 if (object->ref_count != 1) 1479 panic("pmap_release: pteobj reference count != 1"); 1480 #endif 1481 1482 info.pmap = pmap; 1483 info.object = object; 1484 crit_enter(); 1485 lwkt_gettoken(&vm_token); 1486 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode); 1487 crit_exit(); 1488 1489 do { 1490 crit_enter(); 1491 info.error = 0; 1492 info.mpte = NULL; 1493 info.limit = object->generation; 1494 1495 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 1496 pmap_release_callback, &info); 1497 if (info.error == 0 && info.mpte) { 1498 if (!pmap_release_free_page(pmap, info.mpte)) 1499 info.error = 1; 1500 } 1501 crit_exit(); 1502 } while (info.error); 1503 lwkt_reltoken(&vm_token); 1504 } 1505 1506 static int 1507 pmap_release_callback(struct vm_page *p, void *data) 1508 { 1509 struct rb_vm_page_scan_info *info = data; 1510 1511 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) { 1512 info->mpte = p; 1513 return(0); 1514 } 1515 if (!pmap_release_free_page(info->pmap, p)) { 1516 info->error = 1; 1517 return(-1); 1518 } 1519 if (info->object->generation != info->limit) { 1520 info->error = 1; 1521 return(-1); 1522 } 1523 return(0); 1524 } 1525 1526 /* 1527 * Grow the number of kernel page table entries, if needed. 1528 * 1529 * No requirements. 1530 */ 1531 void 1532 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend) 1533 { 1534 vm_offset_t addr; 1535 vm_paddr_t paddr; 1536 vm_offset_t ptppaddr; 1537 vm_page_t nkpg; 1538 pd_entry_t *pde, newpdir; 1539 pdp_entry_t newpdp; 1540 1541 addr = kend; 1542 1543 crit_enter(); 1544 lwkt_gettoken(&vm_token); 1545 if (kernel_vm_end == 0) { 1546 kernel_vm_end = KvaStart; 1547 nkpt = 0; 1548 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & VPTE_V) != 0) { 1549 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1550 nkpt++; 1551 if (kernel_vm_end - 1 >= kernel_map.max_offset) { 1552 kernel_vm_end = kernel_map.max_offset; 1553 break; 1554 } 1555 } 1556 } 1557 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1558 if (addr - 1 >= kernel_map.max_offset) 1559 addr = kernel_map.max_offset; 1560 while (kernel_vm_end < addr) { 1561 pde = pmap_pde(&kernel_pmap, kernel_vm_end); 1562 if (pde == NULL) { 1563 /* We need a new PDP entry */ 1564 nkpg = vm_page_alloc(kptobj, nkpt, 1565 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM 1566 | VM_ALLOC_INTERRUPT); 1567 if (nkpg == NULL) { 1568 panic("pmap_growkernel: no memory to " 1569 "grow kernel"); 1570 } 1571 paddr = VM_PAGE_TO_PHYS(nkpg); 1572 if ((nkpg->flags & PG_ZERO) == 0) 1573 pmap_zero_page(paddr); 1574 vm_page_flag_clear(nkpg, PG_ZERO); 1575 newpdp = (pdp_entry_t)(paddr | VPTE_V | VPTE_R | 1576 VPTE_W | VPTE_A | VPTE_M); 1577 *pmap_pdpe(&kernel_pmap, kernel_vm_end) = newpdp; 1578 nkpt++; 1579 continue; /* try again */ 1580 } 1581 if ((*pde & VPTE_V) != 0) { 1582 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & 1583 ~(PAGE_SIZE * NPTEPG - 1); 1584 if (kernel_vm_end - 1 >= kernel_map.max_offset) { 1585 kernel_vm_end = kernel_map.max_offset; 1586 break; 1587 } 1588 continue; 1589 } 1590 1591 /* 1592 * This index is bogus, but out of the way 1593 */ 1594 nkpg = vm_page_alloc(kptobj, nkpt, 1595 VM_ALLOC_NORMAL | 1596 VM_ALLOC_SYSTEM | 1597 VM_ALLOC_INTERRUPT); 1598 if (nkpg == NULL) 1599 panic("pmap_growkernel: no memory to grow kernel"); 1600 1601 vm_page_wire(nkpg); 1602 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1603 pmap_zero_page(ptppaddr); 1604 vm_page_flag_clear(nkpg, PG_ZERO); 1605 newpdir = (pd_entry_t)(ptppaddr | VPTE_V | VPTE_R | 1606 VPTE_W | VPTE_A | VPTE_M); 1607 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir; 1608 nkpt++; 1609 1610 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & 1611 ~(PAGE_SIZE * NPTEPG - 1); 1612 if (kernel_vm_end - 1 >= kernel_map.max_offset) { 1613 kernel_vm_end = kernel_map.max_offset; 1614 break; 1615 } 1616 } 1617 lwkt_reltoken(&vm_token); 1618 crit_exit(); 1619 } 1620 1621 /* 1622 * Retire the given physical map from service. Should only be called 1623 * if the map contains no valid mappings. 1624 * 1625 * No requirements. 1626 */ 1627 void 1628 pmap_destroy(pmap_t pmap) 1629 { 1630 if (pmap == NULL) 1631 return; 1632 1633 lwkt_gettoken(&vm_token); 1634 if (--pmap->pm_count == 0) { 1635 pmap_release(pmap); 1636 panic("destroying a pmap is not yet implemented"); 1637 } 1638 lwkt_reltoken(&vm_token); 1639 } 1640 1641 /* 1642 * Add a reference to the specified pmap. 1643 * 1644 * No requirements. 1645 */ 1646 void 1647 pmap_reference(pmap_t pmap) 1648 { 1649 if (pmap) { 1650 lwkt_gettoken(&vm_token); 1651 ++pmap->pm_count; 1652 lwkt_reltoken(&vm_token); 1653 } 1654 } 1655 1656 /************************************************************************ 1657 * VMSPACE MANAGEMENT * 1658 ************************************************************************ 1659 * 1660 * The VMSPACE management we do in our virtual kernel must be reflected 1661 * in the real kernel. This is accomplished by making vmspace system 1662 * calls to the real kernel. 1663 */ 1664 void 1665 cpu_vmspace_alloc(struct vmspace *vm) 1666 { 1667 int r; 1668 void *rp; 1669 vpte_t vpte; 1670 1671 #define USER_SIZE (VM_MAX_USER_ADDRESS - VM_MIN_USER_ADDRESS) 1672 1673 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0) 1674 panic("vmspace_create() failed"); 1675 1676 rp = vmspace_mmap(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE, 1677 PROT_READ|PROT_WRITE, 1678 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED, 1679 MemImageFd, 0); 1680 if (rp == MAP_FAILED) 1681 panic("vmspace_mmap: failed"); 1682 vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE, 1683 MADV_NOSYNC, 0); 1684 vpte = VM_PAGE_TO_PHYS(vmspace_pmap(vm)->pm_pdirm) | VPTE_R | VPTE_W | VPTE_V; 1685 r = vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE, 1686 MADV_SETMAP, vpte); 1687 if (r < 0) 1688 panic("vmspace_mcontrol: failed"); 1689 } 1690 1691 void 1692 cpu_vmspace_free(struct vmspace *vm) 1693 { 1694 if (vmspace_destroy(&vm->vm_pmap) < 0) 1695 panic("vmspace_destroy() failed"); 1696 } 1697 1698 /*************************************************** 1699 * page management routines. 1700 ***************************************************/ 1701 1702 /* 1703 * free the pv_entry back to the free list. This function may be 1704 * called from an interrupt. 1705 */ 1706 static __inline void 1707 free_pv_entry(pv_entry_t pv) 1708 { 1709 pv_entry_count--; 1710 KKASSERT(pv_entry_count >= 0); 1711 zfree(pvzone, pv); 1712 } 1713 1714 /* 1715 * get a new pv_entry, allocating a block from the system 1716 * when needed. This function may be called from an interrupt. 1717 */ 1718 static pv_entry_t 1719 get_pv_entry(void) 1720 { 1721 pv_entry_count++; 1722 if (pv_entry_high_water && 1723 (pv_entry_count > pv_entry_high_water) && 1724 (pmap_pagedaemon_waken == 0)) { 1725 pmap_pagedaemon_waken = 1; 1726 wakeup(&vm_pages_needed); 1727 } 1728 return zalloc(pvzone); 1729 } 1730 1731 /* 1732 * This routine is very drastic, but can save the system 1733 * in a pinch. 1734 * 1735 * No requirements. 1736 */ 1737 void 1738 pmap_collect(void) 1739 { 1740 int i; 1741 vm_page_t m; 1742 static int warningdone=0; 1743 1744 if (pmap_pagedaemon_waken == 0) 1745 return; 1746 lwkt_gettoken(&vm_token); 1747 pmap_pagedaemon_waken = 0; 1748 1749 if (warningdone < 5) { 1750 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n"); 1751 warningdone++; 1752 } 1753 1754 for(i = 0; i < vm_page_array_size; i++) { 1755 m = &vm_page_array[i]; 1756 if (m->wire_count || m->hold_count || m->busy || 1757 (m->flags & PG_BUSY)) 1758 continue; 1759 pmap_remove_all(m); 1760 } 1761 lwkt_reltoken(&vm_token); 1762 } 1763 1764 1765 /* 1766 * If it is the first entry on the list, it is actually 1767 * in the header and we must copy the following entry up 1768 * to the header. Otherwise we must search the list for 1769 * the entry. In either case we free the now unused entry. 1770 */ 1771 static int 1772 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va) 1773 { 1774 pv_entry_t pv; 1775 int rtval; 1776 1777 crit_enter(); 1778 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1779 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1780 if (pmap == pv->pv_pmap && va == pv->pv_va) 1781 break; 1782 } 1783 } else { 1784 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1785 if (va == pv->pv_va) 1786 break; 1787 } 1788 } 1789 1790 /* 1791 * Note that pv_ptem is NULL if the page table page itself is not 1792 * managed, even if the page being removed IS managed. 1793 */ 1794 rtval = 0; 1795 /* JGXXX When can 'pv' be NULL? */ 1796 if (pv) { 1797 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1798 m->md.pv_list_count--; 1799 m->object->agg_pv_list_count--; 1800 KKASSERT(m->md.pv_list_count >= 0); 1801 if (TAILQ_EMPTY(&m->md.pv_list)) 1802 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 1803 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1804 ++pmap->pm_generation; 1805 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1806 free_pv_entry(pv); 1807 } 1808 crit_exit(); 1809 return rtval; 1810 } 1811 1812 /* 1813 * Create a pv entry for page at pa for (pmap, va). If the page table page 1814 * holding the VA is managed, mpte will be non-NULL. 1815 */ 1816 static void 1817 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1818 { 1819 pv_entry_t pv; 1820 1821 crit_enter(); 1822 pv = get_pv_entry(); 1823 pv->pv_va = va; 1824 pv->pv_pmap = pmap; 1825 pv->pv_ptem = mpte; 1826 1827 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1828 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1829 m->md.pv_list_count++; 1830 m->object->agg_pv_list_count++; 1831 1832 crit_exit(); 1833 } 1834 1835 /* 1836 * pmap_remove_pte: do the things to unmap a page in a process 1837 */ 1838 static int 1839 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va) 1840 { 1841 pt_entry_t oldpte; 1842 vm_page_t m; 1843 1844 oldpte = pmap_inval_loadandclear(ptq, pmap, va); 1845 if (oldpte & VPTE_WIRED) 1846 --pmap->pm_stats.wired_count; 1847 KKASSERT(pmap->pm_stats.wired_count >= 0); 1848 1849 #if 0 1850 /* 1851 * Machines that don't support invlpg, also don't support 1852 * PG_G. XXX PG_G is disabled for SMP so don't worry about 1853 * the SMP case. 1854 */ 1855 if (oldpte & PG_G) 1856 cpu_invlpg((void *)va); 1857 #endif 1858 KKASSERT(pmap->pm_stats.resident_count > 0); 1859 --pmap->pm_stats.resident_count; 1860 if (oldpte & VPTE_MANAGED) { 1861 m = PHYS_TO_VM_PAGE(oldpte); 1862 if (oldpte & VPTE_M) { 1863 #if defined(PMAP_DIAGNOSTIC) 1864 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1865 kprintf( 1866 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n", 1867 va, oldpte); 1868 } 1869 #endif 1870 if (pmap_track_modified(pmap, va)) 1871 vm_page_dirty(m); 1872 } 1873 if (oldpte & VPTE_A) 1874 vm_page_flag_set(m, PG_REFERENCED); 1875 return pmap_remove_entry(pmap, m, va); 1876 } else { 1877 return pmap_unuse_pt(pmap, va, NULL); 1878 } 1879 1880 return 0; 1881 } 1882 1883 /* 1884 * pmap_remove_page: 1885 * 1886 * Remove a single page from a process address space. 1887 * 1888 * This function may not be called from an interrupt if the pmap is 1889 * not kernel_pmap. 1890 */ 1891 static void 1892 pmap_remove_page(struct pmap *pmap, vm_offset_t va) 1893 { 1894 pt_entry_t *pte; 1895 1896 pte = pmap_pte(pmap, va); 1897 if (pte == NULL) 1898 return; 1899 if ((*pte & VPTE_V) == 0) 1900 return; 1901 pmap_remove_pte(pmap, pte, va); 1902 } 1903 1904 /* 1905 * Remove the given range of addresses from the specified map. 1906 * 1907 * It is assumed that the start and end are properly rounded to 1908 * the page size. 1909 * 1910 * This function may not be called from an interrupt if the pmap is 1911 * not kernel_pmap. 1912 * 1913 * No requirements. 1914 */ 1915 void 1916 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva) 1917 { 1918 vm_offset_t va_next; 1919 pml4_entry_t *pml4e; 1920 pdp_entry_t *pdpe; 1921 pd_entry_t ptpaddr, *pde; 1922 pt_entry_t *pte; 1923 1924 if (pmap == NULL) 1925 return; 1926 1927 lwkt_gettoken(&vm_token); 1928 KKASSERT(pmap->pm_stats.resident_count >= 0); 1929 if (pmap->pm_stats.resident_count == 0) { 1930 lwkt_reltoken(&vm_token); 1931 return; 1932 } 1933 1934 /* 1935 * special handling of removing one page. a very 1936 * common operation and easy to short circuit some 1937 * code. 1938 */ 1939 if (sva + PAGE_SIZE == eva) { 1940 pde = pmap_pde(pmap, sva); 1941 if (pde && (*pde & VPTE_PS) == 0) { 1942 pmap_remove_page(pmap, sva); 1943 lwkt_reltoken(&vm_token); 1944 return; 1945 } 1946 } 1947 1948 for (; sva < eva; sva = va_next) { 1949 pml4e = pmap_pml4e(pmap, sva); 1950 if ((*pml4e & VPTE_V) == 0) { 1951 va_next = (sva + NBPML4) & ~PML4MASK; 1952 if (va_next < sva) 1953 va_next = eva; 1954 continue; 1955 } 1956 1957 pdpe = pmap_pml4e_to_pdpe(pml4e, sva); 1958 if ((*pdpe & VPTE_V) == 0) { 1959 va_next = (sva + NBPDP) & ~PDPMASK; 1960 if (va_next < sva) 1961 va_next = eva; 1962 continue; 1963 } 1964 1965 /* 1966 * Calculate index for next page table. 1967 */ 1968 va_next = (sva + NBPDR) & ~PDRMASK; 1969 if (va_next < sva) 1970 va_next = eva; 1971 1972 pde = pmap_pdpe_to_pde(pdpe, sva); 1973 ptpaddr = *pde; 1974 1975 /* 1976 * Weed out invalid mappings. 1977 */ 1978 if (ptpaddr == 0) 1979 continue; 1980 1981 /* 1982 * Check for large page. 1983 */ 1984 if ((ptpaddr & VPTE_PS) != 0) { 1985 /* JG FreeBSD has more complex treatment here */ 1986 KKASSERT(*pde != 0); 1987 pmap_inval_pde(pde, pmap, sva); 1988 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1989 continue; 1990 } 1991 1992 /* 1993 * Limit our scan to either the end of the va represented 1994 * by the current page table page, or to the end of the 1995 * range being removed. 1996 */ 1997 if (va_next > eva) 1998 va_next = eva; 1999 2000 /* 2001 * NOTE: pmap_remove_pte() can block. 2002 */ 2003 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, 2004 sva += PAGE_SIZE) { 2005 if (*pte == 0) 2006 continue; 2007 if (pmap_remove_pte(pmap, pte, sva)) 2008 break; 2009 } 2010 } 2011 lwkt_reltoken(&vm_token); 2012 } 2013 2014 /* 2015 * Removes this physical page from all physical maps in which it resides. 2016 * Reflects back modify bits to the pager. 2017 * 2018 * This routine may not be called from an interrupt. 2019 * 2020 * No requirements. 2021 */ 2022 2023 static void 2024 pmap_remove_all(vm_page_t m) 2025 { 2026 pt_entry_t *pte, tpte; 2027 pv_entry_t pv; 2028 2029 #if defined(PMAP_DIAGNOSTIC) 2030 /* 2031 * XXX this makes pmap_page_protect(NONE) illegal for non-managed 2032 * pages! 2033 */ 2034 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 2035 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m)); 2036 } 2037 #endif 2038 2039 crit_enter(); 2040 lwkt_gettoken(&vm_token); 2041 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2042 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0); 2043 --pv->pv_pmap->pm_stats.resident_count; 2044 2045 pte = pmap_pte(pv->pv_pmap, pv->pv_va); 2046 KKASSERT(pte != NULL); 2047 2048 tpte = pmap_inval_loadandclear(pte, pv->pv_pmap, pv->pv_va); 2049 if (tpte & VPTE_WIRED) 2050 pv->pv_pmap->pm_stats.wired_count--; 2051 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0); 2052 2053 if (tpte & VPTE_A) 2054 vm_page_flag_set(m, PG_REFERENCED); 2055 2056 /* 2057 * Update the vm_page_t clean and reference bits. 2058 */ 2059 if (tpte & VPTE_M) { 2060 #if defined(PMAP_DIAGNOSTIC) 2061 if (pmap_nw_modified(tpte)) { 2062 kprintf( 2063 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n", 2064 pv->pv_va, tpte); 2065 } 2066 #endif 2067 if (pmap_track_modified(pv->pv_pmap, pv->pv_va)) 2068 vm_page_dirty(m); 2069 } 2070 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2071 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2072 ++pv->pv_pmap->pm_generation; 2073 m->md.pv_list_count--; 2074 m->object->agg_pv_list_count--; 2075 KKASSERT(m->md.pv_list_count >= 0); 2076 if (TAILQ_EMPTY(&m->md.pv_list)) 2077 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 2078 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2079 free_pv_entry(pv); 2080 } 2081 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0); 2082 lwkt_reltoken(&vm_token); 2083 crit_exit(); 2084 } 2085 2086 /* 2087 * Set the physical protection on the specified range of this map 2088 * as requested. 2089 * 2090 * This function may not be called from an interrupt if the map is 2091 * not the kernel_pmap. 2092 * 2093 * No requirements. 2094 */ 2095 void 2096 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 2097 { 2098 vm_offset_t va_next; 2099 pml4_entry_t *pml4e; 2100 pdp_entry_t *pdpe; 2101 pd_entry_t ptpaddr, *pde; 2102 pt_entry_t *pte; 2103 2104 /* JG review for NX */ 2105 2106 if (pmap == NULL) 2107 return; 2108 2109 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 2110 pmap_remove(pmap, sva, eva); 2111 return; 2112 } 2113 2114 if (prot & VM_PROT_WRITE) 2115 return; 2116 2117 lwkt_gettoken(&vm_token); 2118 2119 for (; sva < eva; sva = va_next) { 2120 2121 pml4e = pmap_pml4e(pmap, sva); 2122 if ((*pml4e & VPTE_V) == 0) { 2123 va_next = (sva + NBPML4) & ~PML4MASK; 2124 if (va_next < sva) 2125 va_next = eva; 2126 continue; 2127 } 2128 2129 pdpe = pmap_pml4e_to_pdpe(pml4e, sva); 2130 if ((*pdpe & VPTE_V) == 0) { 2131 va_next = (sva + NBPDP) & ~PDPMASK; 2132 if (va_next < sva) 2133 va_next = eva; 2134 continue; 2135 } 2136 2137 va_next = (sva + NBPDR) & ~PDRMASK; 2138 if (va_next < sva) 2139 va_next = eva; 2140 2141 pde = pmap_pdpe_to_pde(pdpe, sva); 2142 ptpaddr = *pde; 2143 2144 /* 2145 * Check for large page. 2146 */ 2147 if ((ptpaddr & VPTE_PS) != 0) { 2148 /* JG correct? */ 2149 pmap_clean_pde(pde, pmap, sva); 2150 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 2151 continue; 2152 } 2153 2154 /* 2155 * Weed out invalid mappings. Note: we assume that the page 2156 * directory table is always allocated, and in kernel virtual. 2157 */ 2158 if (ptpaddr == 0) 2159 continue; 2160 2161 if (va_next > eva) 2162 va_next = eva; 2163 2164 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, 2165 sva += PAGE_SIZE) { 2166 pt_entry_t pbits; 2167 vm_page_t m; 2168 2169 /* 2170 * Clean managed pages and also check the accessed 2171 * bit. Just remove write perms for unmanaged 2172 * pages. Be careful of races, turning off write 2173 * access will force a fault rather then setting 2174 * the modified bit at an unexpected time. 2175 */ 2176 if (*pte & VPTE_MANAGED) { 2177 pbits = pmap_clean_pte(pte, pmap, sva); 2178 m = NULL; 2179 if (pbits & VPTE_A) { 2180 m = PHYS_TO_VM_PAGE(pbits & VPTE_FRAME); 2181 vm_page_flag_set(m, PG_REFERENCED); 2182 atomic_clear_long(pte, VPTE_A); 2183 } 2184 if (pbits & VPTE_M) { 2185 if (pmap_track_modified(pmap, sva)) { 2186 if (m == NULL) 2187 m = PHYS_TO_VM_PAGE(pbits & VPTE_FRAME); 2188 vm_page_dirty(m); 2189 } 2190 } 2191 } else { 2192 pbits = pmap_setro_pte(pte, pmap, sva); 2193 } 2194 } 2195 } 2196 lwkt_reltoken(&vm_token); 2197 } 2198 2199 /* 2200 * Enter a managed page into a pmap. If the page is not wired related pmap 2201 * data can be destroyed at any time for later demand-operation. 2202 * 2203 * Insert the vm_page (m) at virtual address (v) in (pmap), with the 2204 * specified protection, and wire the mapping if requested. 2205 * 2206 * NOTE: This routine may not lazy-evaluate or lose information. The 2207 * page must actually be inserted into the given map NOW. 2208 * 2209 * NOTE: When entering a page at a KVA address, the pmap must be the 2210 * kernel_pmap. 2211 * 2212 * No requirements. 2213 */ 2214 void 2215 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 2216 boolean_t wired) 2217 { 2218 vm_paddr_t pa; 2219 pd_entry_t *pde; 2220 pt_entry_t *pte; 2221 vm_paddr_t opa; 2222 pt_entry_t origpte, newpte; 2223 vm_page_t mpte; 2224 2225 if (pmap == NULL) 2226 return; 2227 2228 va = trunc_page(va); 2229 2230 lwkt_gettoken(&vm_token); 2231 2232 /* 2233 * Get the page table page. The kernel_pmap's page table pages 2234 * are preallocated and have no associated vm_page_t. 2235 */ 2236 if (pmap == &kernel_pmap) 2237 mpte = NULL; 2238 else 2239 mpte = pmap_allocpte(pmap, va); 2240 2241 pde = pmap_pde(pmap, va); 2242 if (pde != NULL && (*pde & VPTE_V) != 0) { 2243 if ((*pde & VPTE_PS) != 0) 2244 panic("pmap_enter: attempted pmap_enter on 2MB page"); 2245 pte = pmap_pde_to_pte(pde, va); 2246 } else { 2247 panic("pmap_enter: invalid page directory va=%#lx", va); 2248 } 2249 2250 KKASSERT(pte != NULL); 2251 /* 2252 * Deal with races on the original mapping (though don't worry 2253 * about VPTE_A races) by cleaning it. This will force a fault 2254 * if an attempt is made to write to the page. 2255 */ 2256 pa = VM_PAGE_TO_PHYS(m); 2257 origpte = pmap_clean_pte(pte, pmap, va); 2258 opa = origpte & VPTE_FRAME; 2259 2260 if (origpte & VPTE_PS) 2261 panic("pmap_enter: attempted pmap_enter on 2MB page"); 2262 2263 /* 2264 * Mapping has not changed, must be protection or wiring change. 2265 */ 2266 if (origpte && (opa == pa)) { 2267 /* 2268 * Wiring change, just update stats. We don't worry about 2269 * wiring PT pages as they remain resident as long as there 2270 * are valid mappings in them. Hence, if a user page is wired, 2271 * the PT page will be also. 2272 */ 2273 if (wired && ((origpte & VPTE_WIRED) == 0)) 2274 ++pmap->pm_stats.wired_count; 2275 else if (!wired && (origpte & VPTE_WIRED)) 2276 --pmap->pm_stats.wired_count; 2277 2278 /* 2279 * Remove the extra pte reference. Note that we cannot 2280 * optimize the RO->RW case because we have adjusted the 2281 * wiring count above and may need to adjust the wiring 2282 * bits below. 2283 */ 2284 if (mpte) 2285 mpte->hold_count--; 2286 2287 /* 2288 * We might be turning off write access to the page, 2289 * so we go ahead and sense modify status. 2290 */ 2291 if (origpte & VPTE_MANAGED) { 2292 if ((origpte & VPTE_M) && 2293 pmap_track_modified(pmap, va)) { 2294 vm_page_t om; 2295 om = PHYS_TO_VM_PAGE(opa); 2296 vm_page_dirty(om); 2297 } 2298 pa |= VPTE_MANAGED; 2299 KKASSERT(m->flags & PG_MAPPED); 2300 } 2301 goto validate; 2302 } 2303 /* 2304 * Mapping has changed, invalidate old range and fall through to 2305 * handle validating new mapping. 2306 */ 2307 if (opa) { 2308 int err; 2309 err = pmap_remove_pte(pmap, pte, va); 2310 if (err) 2311 panic("pmap_enter: pte vanished, va: 0x%lx", va); 2312 } 2313 2314 /* 2315 * Enter on the PV list if part of our managed memory. Note that we 2316 * raise IPL while manipulating pv_table since pmap_enter can be 2317 * called at interrupt time. 2318 */ 2319 if (pmap_initialized && 2320 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2321 pmap_insert_entry(pmap, va, mpte, m); 2322 pa |= VPTE_MANAGED; 2323 vm_page_flag_set(m, PG_MAPPED); 2324 } 2325 2326 /* 2327 * Increment counters 2328 */ 2329 ++pmap->pm_stats.resident_count; 2330 if (wired) 2331 pmap->pm_stats.wired_count++; 2332 2333 validate: 2334 /* 2335 * Now validate mapping with desired protection/wiring. 2336 */ 2337 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | VPTE_V); 2338 2339 if (wired) 2340 newpte |= VPTE_WIRED; 2341 if (pmap != &kernel_pmap) 2342 newpte |= VPTE_U; 2343 2344 /* 2345 * If the mapping or permission bits are different from the 2346 * (now cleaned) original pte, an update is needed. We've 2347 * already downgraded or invalidated the page so all we have 2348 * to do now is update the bits. 2349 * 2350 * XXX should we synchronize RO->RW changes to avoid another 2351 * fault? 2352 */ 2353 if ((origpte & ~(VPTE_W|VPTE_M|VPTE_A)) != newpte) { 2354 *pte = newpte | VPTE_A; 2355 if (newpte & VPTE_W) 2356 vm_page_flag_set(m, PG_WRITEABLE); 2357 } 2358 KKASSERT((newpte & VPTE_MANAGED) == 0 || (m->flags & PG_MAPPED)); 2359 lwkt_reltoken(&vm_token); 2360 } 2361 2362 /* 2363 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired. 2364 * 2365 * Currently this routine may only be used on user pmaps, not kernel_pmap. 2366 * 2367 * No requirements. 2368 */ 2369 void 2370 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m) 2371 { 2372 pt_entry_t *pte; 2373 vm_paddr_t pa; 2374 vm_page_t mpte; 2375 vm_pindex_t ptepindex; 2376 pd_entry_t *ptepa; 2377 2378 KKASSERT(pmap != &kernel_pmap); 2379 2380 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS); 2381 2382 /* 2383 * Calculate pagetable page index 2384 */ 2385 ptepindex = pmap_pde_pindex(va); 2386 2387 lwkt_gettoken(&vm_token); 2388 2389 do { 2390 /* 2391 * Get the page directory entry 2392 */ 2393 ptepa = pmap_pde(pmap, va); 2394 2395 /* 2396 * If the page table page is mapped, we just increment 2397 * the hold count, and activate it. 2398 */ 2399 if (ptepa && (*ptepa & VPTE_V) != 0) { 2400 if (*ptepa & VPTE_PS) 2401 panic("pmap_enter_quick: unexpected mapping into 2MB page"); 2402 if (pmap->pm_ptphint && 2403 (pmap->pm_ptphint->pindex == ptepindex)) { 2404 mpte = pmap->pm_ptphint; 2405 } else { 2406 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex); 2407 pmap->pm_ptphint = mpte; 2408 } 2409 if (mpte) 2410 mpte->hold_count++; 2411 } else { 2412 mpte = _pmap_allocpte(pmap, ptepindex); 2413 } 2414 } while (mpte == NULL); 2415 2416 /* 2417 * Ok, now that the page table page has been validated, get the pte. 2418 * If the pte is already mapped undo mpte's hold_count and 2419 * just return. 2420 */ 2421 pte = pmap_pte(pmap, va); 2422 if (*pte & VPTE_V) { 2423 KKASSERT(mpte != NULL); 2424 pmap_unwire_pte_hold(pmap, va, mpte); 2425 pa = VM_PAGE_TO_PHYS(m); 2426 KKASSERT(((*pte ^ pa) & VPTE_FRAME) == 0); 2427 lwkt_reltoken(&vm_token); 2428 return; 2429 } 2430 2431 /* 2432 * Enter on the PV list if part of our managed memory 2433 */ 2434 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2435 pmap_insert_entry(pmap, va, mpte, m); 2436 vm_page_flag_set(m, PG_MAPPED); 2437 } 2438 2439 /* 2440 * Increment counters 2441 */ 2442 ++pmap->pm_stats.resident_count; 2443 2444 pa = VM_PAGE_TO_PHYS(m); 2445 2446 /* 2447 * Now validate mapping with RO protection 2448 */ 2449 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2450 *pte = (vpte_t)pa | VPTE_V | VPTE_U; 2451 else 2452 *pte = (vpte_t)pa | VPTE_V | VPTE_U | VPTE_MANAGED; 2453 /*pmap_inval_add(&info, pmap, va); shouldn't be needed 0->valid */ 2454 /*pmap_inval_flush(&info); don't need for vkernel */ 2455 lwkt_reltoken(&vm_token); 2456 } 2457 2458 /* 2459 * Make a temporary mapping for a physical address. This is only intended 2460 * to be used for panic dumps. 2461 */ 2462 void * 2463 pmap_kenter_temporary(vm_paddr_t pa, int i) 2464 { 2465 pmap_kenter(crashdumpmap + (i * PAGE_SIZE), pa); 2466 return ((void *)crashdumpmap); 2467 } 2468 2469 #define MAX_INIT_PT (96) 2470 2471 /* 2472 * This routine preloads the ptes for a given object into the specified pmap. 2473 * This eliminates the blast of soft faults on process startup and 2474 * immediately after an mmap. 2475 * 2476 * No requirements. 2477 */ 2478 static int pmap_object_init_pt_callback(vm_page_t p, void *data); 2479 2480 void 2481 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot, 2482 vm_object_t object, vm_pindex_t pindex, 2483 vm_size_t size, int limit) 2484 { 2485 struct rb_vm_page_scan_info info; 2486 struct lwp *lp; 2487 vm_size_t psize; 2488 2489 /* 2490 * We can't preinit if read access isn't set or there is no pmap 2491 * or object. 2492 */ 2493 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL) 2494 return; 2495 2496 /* 2497 * We can't preinit if the pmap is not the current pmap 2498 */ 2499 lp = curthread->td_lwp; 2500 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace)) 2501 return; 2502 2503 psize = x86_64_btop(size); 2504 2505 if ((object->type != OBJT_VNODE) || 2506 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) && 2507 (object->resident_page_count > MAX_INIT_PT))) { 2508 return; 2509 } 2510 2511 if (psize + pindex > object->size) { 2512 if (object->size < pindex) 2513 return; 2514 psize = object->size - pindex; 2515 } 2516 2517 if (psize == 0) 2518 return; 2519 2520 /* 2521 * Use a red-black scan to traverse the requested range and load 2522 * any valid pages found into the pmap. 2523 * 2524 * We cannot safely scan the object's memq unless we are in a 2525 * critical section since interrupts can remove pages from objects. 2526 */ 2527 info.start_pindex = pindex; 2528 info.end_pindex = pindex + psize - 1; 2529 info.limit = limit; 2530 info.mpte = NULL; 2531 info.addr = addr; 2532 info.pmap = pmap; 2533 2534 crit_enter(); 2535 lwkt_gettoken(&vm_token); 2536 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 2537 pmap_object_init_pt_callback, &info); 2538 lwkt_reltoken(&vm_token); 2539 crit_exit(); 2540 } 2541 2542 static 2543 int 2544 pmap_object_init_pt_callback(vm_page_t p, void *data) 2545 { 2546 struct rb_vm_page_scan_info *info = data; 2547 vm_pindex_t rel_index; 2548 /* 2549 * don't allow an madvise to blow away our really 2550 * free pages allocating pv entries. 2551 */ 2552 if ((info->limit & MAP_PREFAULT_MADVISE) && 2553 vmstats.v_free_count < vmstats.v_free_reserved) { 2554 return(-1); 2555 } 2556 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2557 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2558 if ((p->queue - p->pc) == PQ_CACHE) 2559 vm_page_deactivate(p); 2560 vm_page_busy(p); 2561 rel_index = p->pindex - info->start_pindex; 2562 pmap_enter_quick(info->pmap, 2563 info->addr + x86_64_ptob(rel_index), p); 2564 vm_page_wakeup(p); 2565 } 2566 return(0); 2567 } 2568 2569 /* 2570 * Return TRUE if the pmap is in shape to trivially 2571 * pre-fault the specified address. 2572 * 2573 * Returns FALSE if it would be non-trivial or if a 2574 * pte is already loaded into the slot. 2575 * 2576 * No requirements. 2577 */ 2578 int 2579 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr) 2580 { 2581 pt_entry_t *pte; 2582 pd_entry_t *pde; 2583 int ret; 2584 2585 lwkt_gettoken(&vm_token); 2586 pde = pmap_pde(pmap, addr); 2587 if (pde == NULL || *pde == 0) { 2588 ret = 0; 2589 } else { 2590 pte = pmap_pde_to_pte(pde, addr); 2591 ret = (*pte) ? 0 : 1; 2592 } 2593 lwkt_reltoken(&vm_token); 2594 return (ret); 2595 } 2596 2597 /* 2598 * Change the wiring attribute for a map/virtual-address pair. 2599 * 2600 * The mapping must already exist in the pmap. 2601 * No other requirements. 2602 */ 2603 void 2604 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired) 2605 { 2606 pt_entry_t *pte; 2607 2608 if (pmap == NULL) 2609 return; 2610 2611 lwkt_gettoken(&vm_token); 2612 pte = pmap_pte(pmap, va); 2613 2614 if (wired && !pmap_pte_w(pte)) 2615 pmap->pm_stats.wired_count++; 2616 else if (!wired && pmap_pte_w(pte)) 2617 pmap->pm_stats.wired_count--; 2618 2619 /* 2620 * Wiring is not a hardware characteristic so there is no need to 2621 * invalidate TLB. However, in an SMP environment we must use 2622 * a locked bus cycle to update the pte (if we are not using 2623 * the pmap_inval_*() API that is)... it's ok to do this for simple 2624 * wiring changes. 2625 */ 2626 if (wired) 2627 atomic_set_long(pte, VPTE_WIRED); 2628 else 2629 atomic_clear_long(pte, VPTE_WIRED); 2630 lwkt_reltoken(&vm_token); 2631 } 2632 2633 /* 2634 * Copy the range specified by src_addr/len 2635 * from the source map to the range dst_addr/len 2636 * in the destination map. 2637 * 2638 * This routine is only advisory and need not do anything. 2639 */ 2640 void 2641 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, 2642 vm_size_t len, vm_offset_t src_addr) 2643 { 2644 /* 2645 * XXX BUGGY. Amoung other things srcmpte is assumed to remain 2646 * valid through blocking calls, and that's just not going to 2647 * be the case. 2648 * 2649 * FIXME! 2650 */ 2651 return; 2652 } 2653 2654 /* 2655 * pmap_zero_page: 2656 * 2657 * Zero the specified physical page. 2658 * 2659 * This function may be called from an interrupt and no locking is 2660 * required. 2661 */ 2662 void 2663 pmap_zero_page(vm_paddr_t phys) 2664 { 2665 vm_offset_t va = PHYS_TO_DMAP(phys); 2666 2667 bzero((void *)va, PAGE_SIZE); 2668 } 2669 2670 /* 2671 * pmap_page_assertzero: 2672 * 2673 * Assert that a page is empty, panic if it isn't. 2674 */ 2675 void 2676 pmap_page_assertzero(vm_paddr_t phys) 2677 { 2678 int i; 2679 2680 crit_enter(); 2681 vm_offset_t virt = PHYS_TO_DMAP(phys); 2682 2683 for (i = 0; i < PAGE_SIZE; i += sizeof(int)) { 2684 if (*(int *)((char *)virt + i) != 0) { 2685 panic("pmap_page_assertzero() @ %p not zero!\n", 2686 (void *)virt); 2687 } 2688 } 2689 crit_exit(); 2690 } 2691 2692 /* 2693 * pmap_zero_page: 2694 * 2695 * Zero part of a physical page by mapping it into memory and clearing 2696 * its contents with bzero. 2697 * 2698 * off and size may not cover an area beyond a single hardware page. 2699 */ 2700 void 2701 pmap_zero_page_area(vm_paddr_t phys, int off, int size) 2702 { 2703 crit_enter(); 2704 vm_offset_t virt = PHYS_TO_DMAP(phys); 2705 bzero((char *)virt + off, size); 2706 crit_exit(); 2707 } 2708 2709 /* 2710 * pmap_copy_page: 2711 * 2712 * Copy the physical page from the source PA to the target PA. 2713 * This function may be called from an interrupt. No locking 2714 * is required. 2715 */ 2716 void 2717 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst) 2718 { 2719 vm_offset_t src_virt, dst_virt; 2720 2721 crit_enter(); 2722 src_virt = PHYS_TO_DMAP(src); 2723 dst_virt = PHYS_TO_DMAP(dst); 2724 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE); 2725 crit_exit(); 2726 } 2727 2728 /* 2729 * pmap_copy_page_frag: 2730 * 2731 * Copy the physical page from the source PA to the target PA. 2732 * This function may be called from an interrupt. No locking 2733 * is required. 2734 */ 2735 void 2736 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes) 2737 { 2738 vm_offset_t src_virt, dst_virt; 2739 2740 crit_enter(); 2741 src_virt = PHYS_TO_DMAP(src); 2742 dst_virt = PHYS_TO_DMAP(dst); 2743 bcopy((char *)src_virt + (src & PAGE_MASK), 2744 (char *)dst_virt + (dst & PAGE_MASK), 2745 bytes); 2746 crit_exit(); 2747 } 2748 2749 /* 2750 * Returns true if the pmap's pv is one of the first 16 pvs linked to 2751 * from this page. This count may be changed upwards or downwards 2752 * in the future; it is only necessary that true be returned for a small 2753 * subset of pmaps for proper page aging. 2754 * 2755 * No other requirements. 2756 */ 2757 boolean_t 2758 pmap_page_exists_quick(pmap_t pmap, vm_page_t m) 2759 { 2760 pv_entry_t pv; 2761 int loops = 0; 2762 2763 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2764 return FALSE; 2765 2766 crit_enter(); 2767 lwkt_gettoken(&vm_token); 2768 2769 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2770 if (pv->pv_pmap == pmap) { 2771 lwkt_reltoken(&vm_token); 2772 crit_exit(); 2773 return TRUE; 2774 } 2775 loops++; 2776 if (loops >= 16) 2777 break; 2778 } 2779 lwkt_reltoken(&vm_token); 2780 crit_exit(); 2781 return (FALSE); 2782 } 2783 2784 /* 2785 * Remove all pages from specified address space this aids process 2786 * exit speeds. Also, this code is special cased for current 2787 * process only, but can have the more generic (and slightly slower) 2788 * mode enabled. This is much faster than pmap_remove in the case 2789 * of running down an entire address space. 2790 * 2791 * No other requirements. 2792 */ 2793 void 2794 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 2795 { 2796 pt_entry_t *pte, tpte; 2797 pv_entry_t pv, npv; 2798 vm_page_t m; 2799 int save_generation; 2800 2801 crit_enter(); 2802 lwkt_gettoken(&vm_token); 2803 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2804 if (pv->pv_va >= eva || pv->pv_va < sva) { 2805 npv = TAILQ_NEXT(pv, pv_plist); 2806 continue; 2807 } 2808 2809 KKASSERT(pmap == pv->pv_pmap); 2810 2811 pte = pmap_pte(pmap, pv->pv_va); 2812 2813 /* 2814 * We cannot remove wired pages from a process' mapping 2815 * at this time 2816 */ 2817 if (*pte & VPTE_WIRED) { 2818 npv = TAILQ_NEXT(pv, pv_plist); 2819 continue; 2820 } 2821 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va); 2822 2823 m = PHYS_TO_VM_PAGE(tpte & VPTE_FRAME); 2824 2825 KASSERT(m < &vm_page_array[vm_page_array_size], 2826 ("pmap_remove_pages: bad tpte %lx", tpte)); 2827 2828 KKASSERT(pmap->pm_stats.resident_count > 0); 2829 --pmap->pm_stats.resident_count; 2830 2831 /* 2832 * Update the vm_page_t clean and reference bits. 2833 */ 2834 if (tpte & VPTE_M) { 2835 vm_page_dirty(m); 2836 } 2837 2838 npv = TAILQ_NEXT(pv, pv_plist); 2839 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 2840 save_generation = ++pmap->pm_generation; 2841 2842 m->md.pv_list_count--; 2843 m->object->agg_pv_list_count--; 2844 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2845 if (TAILQ_EMPTY(&m->md.pv_list)) 2846 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 2847 2848 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem); 2849 free_pv_entry(pv); 2850 2851 /* 2852 * Restart the scan if we blocked during the unuse or free 2853 * calls and other removals were made. 2854 */ 2855 if (save_generation != pmap->pm_generation) { 2856 kprintf("Warning: pmap_remove_pages race-A avoided\n"); 2857 npv = TAILQ_FIRST(&pmap->pm_pvlist); 2858 } 2859 } 2860 lwkt_reltoken(&vm_token); 2861 crit_exit(); 2862 } 2863 2864 /* 2865 * pmap_testbit tests bits in active mappings of a VM page. 2866 */ 2867 static boolean_t 2868 pmap_testbit(vm_page_t m, int bit) 2869 { 2870 pv_entry_t pv; 2871 pt_entry_t *pte; 2872 2873 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2874 return FALSE; 2875 2876 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 2877 return FALSE; 2878 2879 crit_enter(); 2880 2881 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2882 /* 2883 * if the bit being tested is the modified bit, then 2884 * mark clean_map and ptes as never 2885 * modified. 2886 */ 2887 if (bit & (VPTE_A|VPTE_M)) { 2888 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va)) 2889 continue; 2890 } 2891 2892 #if defined(PMAP_DIAGNOSTIC) 2893 if (pv->pv_pmap == NULL) { 2894 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va); 2895 continue; 2896 } 2897 #endif 2898 pte = pmap_pte(pv->pv_pmap, pv->pv_va); 2899 if (*pte & bit) { 2900 crit_exit(); 2901 return TRUE; 2902 } 2903 } 2904 crit_exit(); 2905 return (FALSE); 2906 } 2907 2908 /* 2909 * This routine is used to clear bits in ptes. Certain bits require special 2910 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit. 2911 * 2912 * This routine is only called with certain VPTE_* bit combinations. 2913 */ 2914 static __inline void 2915 pmap_clearbit(vm_page_t m, int bit) 2916 { 2917 pv_entry_t pv; 2918 pt_entry_t *pte; 2919 pt_entry_t pbits; 2920 2921 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2922 return; 2923 2924 crit_enter(); 2925 2926 /* 2927 * Loop over all current mappings setting/clearing as appropos If 2928 * setting RO do we need to clear the VAC? 2929 */ 2930 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2931 /* 2932 * don't write protect pager mappings 2933 */ 2934 if (bit == VPTE_W) { 2935 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va)) 2936 continue; 2937 } 2938 2939 #if defined(PMAP_DIAGNOSTIC) 2940 if (pv->pv_pmap == NULL) { 2941 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va); 2942 continue; 2943 } 2944 #endif 2945 2946 /* 2947 * Careful here. We can use a locked bus instruction to 2948 * clear VPTE_A or VPTE_M safely but we need to synchronize 2949 * with the target cpus when we mess with VPTE_W. 2950 * 2951 * On virtual kernels we must force a new fault-on-write 2952 * in the real kernel if we clear the Modify bit ourselves, 2953 * otherwise the real kernel will not get a new fault and 2954 * will never set our Modify bit again. 2955 */ 2956 pte = pmap_pte(pv->pv_pmap, pv->pv_va); 2957 if (*pte & bit) { 2958 if (bit == VPTE_W) { 2959 /* 2960 * We must also clear VPTE_M when clearing 2961 * VPTE_W 2962 */ 2963 pbits = pmap_clean_pte(pte, pv->pv_pmap, 2964 pv->pv_va); 2965 if (pbits & VPTE_M) 2966 vm_page_dirty(m); 2967 } else if (bit == VPTE_M) { 2968 /* 2969 * We do not have to make the page read-only 2970 * when clearing the Modify bit. The real 2971 * kernel will make the real PTE read-only 2972 * or otherwise detect the write and set 2973 * our VPTE_M again simply by us invalidating 2974 * the real kernel VA for the pmap (as we did 2975 * above). This allows the real kernel to 2976 * handle the write fault without forwarding 2977 * the fault to us. 2978 */ 2979 atomic_clear_long(pte, VPTE_M); 2980 } else if ((bit & (VPTE_W|VPTE_M)) == (VPTE_W|VPTE_M)) { 2981 /* 2982 * We've been asked to clear W & M, I guess 2983 * the caller doesn't want us to update 2984 * the dirty status of the VM page. 2985 */ 2986 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va); 2987 } else { 2988 /* 2989 * We've been asked to clear bits that do 2990 * not interact with hardware. 2991 */ 2992 atomic_clear_long(pte, bit); 2993 } 2994 } 2995 } 2996 crit_exit(); 2997 } 2998 2999 /* 3000 * Lower the permission for all mappings to a given page. 3001 * 3002 * No other requirements. 3003 */ 3004 void 3005 pmap_page_protect(vm_page_t m, vm_prot_t prot) 3006 { 3007 /* JG NX support? */ 3008 if ((prot & VM_PROT_WRITE) == 0) { 3009 lwkt_gettoken(&vm_token); 3010 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 3011 pmap_clearbit(m, VPTE_W); 3012 vm_page_flag_clear(m, PG_WRITEABLE); 3013 } else { 3014 pmap_remove_all(m); 3015 } 3016 lwkt_reltoken(&vm_token); 3017 } 3018 } 3019 3020 vm_paddr_t 3021 pmap_phys_address(vm_pindex_t ppn) 3022 { 3023 return (x86_64_ptob(ppn)); 3024 } 3025 3026 /* 3027 * Return a count of reference bits for a page, clearing those bits. 3028 * It is not necessary for every reference bit to be cleared, but it 3029 * is necessary that 0 only be returned when there are truly no 3030 * reference bits set. 3031 * 3032 * XXX: The exact number of bits to check and clear is a matter that 3033 * should be tested and standardized at some point in the future for 3034 * optimal aging of shared pages. 3035 * 3036 * No other requirements. 3037 */ 3038 int 3039 pmap_ts_referenced(vm_page_t m) 3040 { 3041 pv_entry_t pv, pvf, pvn; 3042 pt_entry_t *pte; 3043 int rtval = 0; 3044 3045 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3046 return (rtval); 3047 3048 crit_enter(); 3049 lwkt_gettoken(&vm_token); 3050 3051 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 3052 3053 pvf = pv; 3054 3055 do { 3056 pvn = TAILQ_NEXT(pv, pv_list); 3057 3058 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 3059 3060 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 3061 3062 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va)) 3063 continue; 3064 3065 pte = pmap_pte(pv->pv_pmap, pv->pv_va); 3066 3067 if (pte && (*pte & VPTE_A)) { 3068 #ifdef SMP 3069 atomic_clear_long(pte, VPTE_A); 3070 #else 3071 atomic_clear_long_nonlocked(pte, VPTE_A); 3072 #endif 3073 rtval++; 3074 if (rtval > 4) { 3075 break; 3076 } 3077 } 3078 } while ((pv = pvn) != NULL && pv != pvf); 3079 } 3080 lwkt_reltoken(&vm_token); 3081 crit_exit(); 3082 3083 return (rtval); 3084 } 3085 3086 /* 3087 * Return whether or not the specified physical page was modified 3088 * in any physical maps. 3089 * 3090 * No other requirements. 3091 */ 3092 boolean_t 3093 pmap_is_modified(vm_page_t m) 3094 { 3095 boolean_t res; 3096 3097 lwkt_gettoken(&vm_token); 3098 res = pmap_testbit(m, VPTE_M); 3099 lwkt_reltoken(&vm_token); 3100 return (res); 3101 } 3102 3103 /* 3104 * Clear the modify bits on the specified physical page. 3105 * 3106 * No other requirements. 3107 */ 3108 void 3109 pmap_clear_modify(vm_page_t m) 3110 { 3111 lwkt_gettoken(&vm_token); 3112 pmap_clearbit(m, VPTE_M); 3113 lwkt_reltoken(&vm_token); 3114 } 3115 3116 /* 3117 * Clear the reference bit on the specified physical page. 3118 * 3119 * No other requirements. 3120 */ 3121 void 3122 pmap_clear_reference(vm_page_t m) 3123 { 3124 lwkt_gettoken(&vm_token); 3125 pmap_clearbit(m, VPTE_A); 3126 lwkt_reltoken(&vm_token); 3127 } 3128 3129 /* 3130 * Miscellaneous support routines follow 3131 */ 3132 3133 static void 3134 i386_protection_init(void) 3135 { 3136 int *kp, prot; 3137 3138 kp = protection_codes; 3139 for (prot = 0; prot < 8; prot++) { 3140 if (prot & VM_PROT_READ) 3141 *kp |= VPTE_R; 3142 if (prot & VM_PROT_WRITE) 3143 *kp |= VPTE_W; 3144 if (prot & VM_PROT_EXECUTE) 3145 *kp |= VPTE_X; 3146 ++kp; 3147 } 3148 } 3149 3150 /* 3151 * Perform the pmap work for mincore 3152 * 3153 * No other requirements. 3154 */ 3155 int 3156 pmap_mincore(pmap_t pmap, vm_offset_t addr) 3157 { 3158 pt_entry_t *ptep, pte; 3159 vm_page_t m; 3160 int val = 0; 3161 3162 lwkt_gettoken(&vm_token); 3163 ptep = pmap_pte(pmap, addr); 3164 3165 if (ptep && (pte = *ptep) != 0) { 3166 vm_paddr_t pa; 3167 3168 val = MINCORE_INCORE; 3169 if ((pte & VPTE_MANAGED) == 0) 3170 goto done; 3171 3172 pa = pte & VPTE_FRAME; 3173 3174 m = PHYS_TO_VM_PAGE(pa); 3175 3176 /* 3177 * Modified by us 3178 */ 3179 if (pte & VPTE_M) 3180 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 3181 /* 3182 * Modified by someone 3183 */ 3184 else if (m->dirty || pmap_is_modified(m)) 3185 val |= MINCORE_MODIFIED_OTHER; 3186 /* 3187 * Referenced by us 3188 */ 3189 if (pte & VPTE_A) 3190 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 3191 3192 /* 3193 * Referenced by someone 3194 */ 3195 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) { 3196 val |= MINCORE_REFERENCED_OTHER; 3197 vm_page_flag_set(m, PG_REFERENCED); 3198 } 3199 } 3200 done: 3201 lwkt_reltoken(&vm_token); 3202 return val; 3203 } 3204 3205 /* 3206 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new 3207 * vmspace will be ref'd and the old one will be deref'd. 3208 */ 3209 void 3210 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs) 3211 { 3212 struct vmspace *oldvm; 3213 struct lwp *lp; 3214 3215 crit_enter(); 3216 oldvm = p->p_vmspace; 3217 if (oldvm != newvm) { 3218 p->p_vmspace = newvm; 3219 KKASSERT(p->p_nthreads == 1); 3220 lp = RB_ROOT(&p->p_lwp_tree); 3221 pmap_setlwpvm(lp, newvm); 3222 if (adjrefs) { 3223 sysref_get(&newvm->vm_sysref); 3224 sysref_put(&oldvm->vm_sysref); 3225 } 3226 } 3227 crit_exit(); 3228 } 3229 3230 /* 3231 * Set the vmspace for a LWP. The vmspace is almost universally set the 3232 * same as the process vmspace, but virtual kernels need to swap out contexts 3233 * on a per-lwp basis. 3234 */ 3235 void 3236 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm) 3237 { 3238 struct vmspace *oldvm; 3239 struct pmap *pmap; 3240 3241 crit_enter(); 3242 oldvm = lp->lwp_vmspace; 3243 3244 if (oldvm != newvm) { 3245 lp->lwp_vmspace = newvm; 3246 if (curthread->td_lwp == lp) { 3247 pmap = vmspace_pmap(newvm); 3248 #if defined(SMP) 3249 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid); 3250 #else 3251 pmap->pm_active |= 1; 3252 #endif 3253 #if defined(SWTCH_OPTIM_STATS) 3254 tlb_flush_count++; 3255 #endif 3256 pmap = vmspace_pmap(oldvm); 3257 #if defined(SMP) 3258 atomic_clear_int(&pmap->pm_active, 3259 1 << mycpu->gd_cpuid); 3260 #else 3261 pmap->pm_active &= ~1; 3262 #endif 3263 } 3264 } 3265 crit_exit(); 3266 } 3267 3268 vm_offset_t 3269 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3270 { 3271 3272 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3273 return addr; 3274 } 3275 3276 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); 3277 return addr; 3278 } 3279 3280 /* 3281 * Used by kmalloc/kfree, page already exists at va 3282 */ 3283 vm_page_t 3284 pmap_kvtom(vm_offset_t va) 3285 { 3286 vpte_t *ptep; 3287 3288 KKASSERT(va >= KvaStart && va < KvaEnd); 3289 ptep = vtopte(va); 3290 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME)); 3291 } 3292