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 53 #if JG 54 #include "opt_pmap.h" 55 #endif 56 #include "opt_msgbuf.h" 57 58 #include <sys/param.h> 59 #include <sys/systm.h> 60 #include <sys/kernel.h> 61 #include <sys/proc.h> 62 #include <sys/msgbuf.h> 63 #include <sys/vmmeter.h> 64 #include <sys/mman.h> 65 #include <sys/vmspace.h> 66 67 #include <vm/vm.h> 68 #include <vm/vm_param.h> 69 #include <sys/sysctl.h> 70 #include <sys/lock.h> 71 #include <vm/vm_kern.h> 72 #include <vm/vm_page.h> 73 #include <vm/vm_map.h> 74 #include <vm/vm_object.h> 75 #include <vm/vm_extern.h> 76 #include <vm/vm_pageout.h> 77 #include <vm/vm_pager.h> 78 #include <vm/vm_zone.h> 79 80 #include <sys/user.h> 81 #include <sys/thread2.h> 82 #include <sys/sysref2.h> 83 #include <sys/spinlock2.h> 84 #include <vm/vm_page2.h> 85 86 #include <machine/cputypes.h> 87 #include <machine/md_var.h> 88 #include <machine/specialreg.h> 89 #include <machine/smp.h> 90 #include <machine/globaldata.h> 91 #include <machine/pmap.h> 92 #include <machine/pmap_inval.h> 93 94 #include <ddb/ddb.h> 95 96 #include <stdio.h> 97 #include <assert.h> 98 #include <stdlib.h> 99 #include <pthread.h> 100 101 #define PMAP_KEEP_PDIRS 102 #ifndef PMAP_SHPGPERPROC 103 #define PMAP_SHPGPERPROC 1000 104 #endif 105 106 #if defined(DIAGNOSTIC) 107 #define PMAP_DIAGNOSTIC 108 #endif 109 110 #define MINPV 2048 111 112 #if !defined(PMAP_DIAGNOSTIC) 113 #define PMAP_INLINE __inline 114 #else 115 #define PMAP_INLINE 116 #endif 117 118 /* 119 * Get PDEs and PTEs for user/kernel address space 120 */ 121 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va); 122 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 123 124 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & VPTE_V) != 0) 125 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & VPTE_WIRED) != 0) 126 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & VPTE_M) != 0) 127 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & VPTE_A) != 0) 128 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & VPTE_V) != 0) 129 130 /* 131 * Given a map and a machine independent protection code, 132 * convert to a vax protection code. 133 */ 134 #define pte_prot(m, p) \ 135 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)]) 136 static int protection_codes[8]; 137 138 struct pmap kernel_pmap; 139 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list); 140 141 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 142 143 static vm_object_t kptobj; 144 145 static int nkpt; 146 147 static uint64_t KPDphys; /* phys addr of kernel level 2 */ 148 uint64_t KPDPphys; /* phys addr of kernel level 3 */ 149 uint64_t KPML4phys; /* phys addr of kernel level 4 */ 150 151 extern int vmm_enabled; 152 extern void *vkernel_stack; 153 154 /* 155 * Data for the pv entry allocation mechanism 156 */ 157 static vm_zone_t pvzone; 158 static struct vm_zone pvzone_store; 159 static struct vm_object pvzone_obj; 160 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0; 161 static int pmap_pagedaemon_waken = 0; 162 static struct pv_entry *pvinit; 163 164 /* 165 * All those kernel PT submaps that BSD is so fond of 166 */ 167 pt_entry_t *CMAP1 = NULL, *ptmmap; 168 caddr_t CADDR1 = NULL; 169 static pt_entry_t *msgbufmap; 170 171 uint64_t KPTphys; 172 173 static PMAP_INLINE void free_pv_entry (pv_entry_t pv); 174 static pv_entry_t get_pv_entry (void); 175 static void i386_protection_init (void); 176 static __inline void pmap_clearbit (vm_page_t m, int bit); 177 178 static void pmap_remove_all (vm_page_t m); 179 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq, 180 vm_offset_t sva); 181 static void pmap_remove_page (struct pmap *pmap, vm_offset_t va); 182 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m, 183 vm_offset_t va); 184 static boolean_t pmap_testbit (vm_page_t m, int bit); 185 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va, 186 vm_page_t mpte, vm_page_t m); 187 188 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va); 189 190 static int pmap_release_free_page (pmap_t pmap, vm_page_t p); 191 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex); 192 #if JGPMAP32 193 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va); 194 #endif 195 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex); 196 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t); 197 198 /* 199 * pmap_pte_quick: 200 * 201 * Super fast pmap_pte routine best used when scanning the pv lists. 202 * This eliminates many course-grained invltlb calls. Note that many of 203 * the pv list scans are across different pmaps and it is very wasteful 204 * to do an entire invltlb when checking a single mapping. 205 * 206 * Should only be called while in a critical section. 207 */ 208 #if JGPMAP32 209 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va); 210 211 static pt_entry_t * 212 pmap_pte_quick(pmap_t pmap, vm_offset_t va) 213 { 214 return pmap_pte(pmap, va); 215 } 216 #endif 217 218 /* Return a non-clipped PD index for a given VA */ 219 static __inline vm_pindex_t 220 pmap_pde_pindex(vm_offset_t va) 221 { 222 return va >> PDRSHIFT; 223 } 224 225 /* Return various clipped indexes for a given VA */ 226 static __inline vm_pindex_t 227 pmap_pte_index(vm_offset_t va) 228 { 229 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1)); 230 } 231 232 static __inline vm_pindex_t 233 pmap_pde_index(vm_offset_t va) 234 { 235 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1)); 236 } 237 238 static __inline vm_pindex_t 239 pmap_pdpe_index(vm_offset_t va) 240 { 241 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1)); 242 } 243 244 static __inline vm_pindex_t 245 pmap_pml4e_index(vm_offset_t va) 246 { 247 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1)); 248 } 249 250 /* Return a pointer to the PML4 slot that corresponds to a VA */ 251 static __inline pml4_entry_t * 252 pmap_pml4e(pmap_t pmap, vm_offset_t va) 253 { 254 return (&pmap->pm_pml4[pmap_pml4e_index(va)]); 255 } 256 257 /* Return a pointer to the PDP slot that corresponds to a VA */ 258 static __inline pdp_entry_t * 259 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va) 260 { 261 pdp_entry_t *pdpe; 262 263 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & VPTE_FRAME); 264 return (&pdpe[pmap_pdpe_index(va)]); 265 } 266 267 /* Return a pointer to the PDP slot that corresponds to a VA */ 268 static __inline pdp_entry_t * 269 pmap_pdpe(pmap_t pmap, vm_offset_t va) 270 { 271 pml4_entry_t *pml4e; 272 273 pml4e = pmap_pml4e(pmap, va); 274 if ((*pml4e & VPTE_V) == 0) 275 return NULL; 276 return (pmap_pml4e_to_pdpe(pml4e, va)); 277 } 278 279 /* Return a pointer to the PD slot that corresponds to a VA */ 280 static __inline pd_entry_t * 281 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va) 282 { 283 pd_entry_t *pde; 284 285 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & VPTE_FRAME); 286 return (&pde[pmap_pde_index(va)]); 287 } 288 289 /* Return a pointer to the PD slot that corresponds to a VA */ 290 static __inline pd_entry_t * 291 pmap_pde(pmap_t pmap, vm_offset_t va) 292 { 293 pdp_entry_t *pdpe; 294 295 pdpe = pmap_pdpe(pmap, va); 296 if (pdpe == NULL || (*pdpe & VPTE_V) == 0) 297 return NULL; 298 return (pmap_pdpe_to_pde(pdpe, va)); 299 } 300 301 /* Return a pointer to the PT slot that corresponds to a VA */ 302 static __inline pt_entry_t * 303 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va) 304 { 305 pt_entry_t *pte; 306 307 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & VPTE_FRAME); 308 return (&pte[pmap_pte_index(va)]); 309 } 310 311 /* Return a pointer to the PT slot that corresponds to a VA */ 312 static __inline pt_entry_t * 313 pmap_pte(pmap_t pmap, vm_offset_t va) 314 { 315 pd_entry_t *pde; 316 317 pde = pmap_pde(pmap, va); 318 if (pde == NULL || (*pde & VPTE_V) == 0) 319 return NULL; 320 if ((*pde & VPTE_PS) != 0) /* compat with i386 pmap_pte() */ 321 return ((pt_entry_t *)pde); 322 return (pmap_pde_to_pte(pde, va)); 323 } 324 325 326 #if JGV 327 PMAP_INLINE pt_entry_t * 328 vtopte(vm_offset_t va) 329 { 330 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + 331 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); 332 333 return (PTmap + ((va >> PAGE_SHIFT) & mask)); 334 } 335 336 static __inline pd_entry_t * 337 vtopde(vm_offset_t va) 338 { 339 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + 340 NPML4EPGSHIFT)) - 1); 341 342 return (PDmap + ((va >> PDRSHIFT) & mask)); 343 } 344 #else 345 static PMAP_INLINE pt_entry_t * 346 vtopte(vm_offset_t va) 347 { 348 pt_entry_t *x; 349 x = pmap_pte(&kernel_pmap, va); 350 assert(x != NULL); 351 return x; 352 } 353 354 static __inline pd_entry_t * 355 vtopde(vm_offset_t va) 356 { 357 pd_entry_t *x; 358 x = pmap_pde(&kernel_pmap, va); 359 assert(x != NULL); 360 return x; 361 } 362 #endif 363 364 static uint64_t 365 allocpages(vm_paddr_t *firstaddr, int n) 366 { 367 uint64_t ret; 368 369 ret = *firstaddr; 370 #if JGV 371 bzero((void *)ret, n * PAGE_SIZE); 372 #endif 373 *firstaddr += n * PAGE_SIZE; 374 return (ret); 375 } 376 377 static void 378 create_dmap_vmm(vm_paddr_t *firstaddr) 379 { 380 void *stack_addr; 381 int pml4_stack_index; 382 int pdp_stack_index; 383 int pd_stack_index; 384 long i,j; 385 int regs[4]; 386 int amd_feature; 387 388 uint64_t KPDP_DMAP_phys = allocpages(firstaddr, NDMPML4E); 389 uint64_t KPDP_VSTACK_phys = allocpages(firstaddr, 1); 390 uint64_t KPD_VSTACK_phys = allocpages(firstaddr, 1); 391 392 pml4_entry_t *KPML4virt = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys); 393 pdp_entry_t *KPDP_DMAP_virt = (pdp_entry_t *)PHYS_TO_DMAP(KPDP_DMAP_phys); 394 pdp_entry_t *KPDP_VSTACK_virt = (pdp_entry_t *)PHYS_TO_DMAP(KPDP_VSTACK_phys); 395 pd_entry_t *KPD_VSTACK_virt = (pd_entry_t *)PHYS_TO_DMAP(KPD_VSTACK_phys); 396 397 bzero(KPDP_DMAP_virt, NDMPML4E * PAGE_SIZE); 398 bzero(KPDP_VSTACK_virt, 1 * PAGE_SIZE); 399 bzero(KPD_VSTACK_virt, 1 * PAGE_SIZE); 400 401 do_cpuid(0x80000001, regs); 402 amd_feature = regs[3]; 403 404 /* Build the mappings for the first 512GB */ 405 if (amd_feature & AMDID_PAGE1GB) { 406 /* In pages of 1 GB, if supported */ 407 for (i = 0; i < NPDPEPG; i++) { 408 KPDP_DMAP_virt[i] = ((uint64_t)i << PDPSHIFT); 409 KPDP_DMAP_virt[i] |= VPTE_RW | VPTE_V | VPTE_PS | VPTE_U; 410 } 411 } else { 412 /* In page of 2MB, otherwise */ 413 for (i = 0; i < NPDPEPG; i++) { 414 uint64_t KPD_DMAP_phys = allocpages(firstaddr, 1); 415 pd_entry_t *KPD_DMAP_virt = (pd_entry_t *)PHYS_TO_DMAP(KPD_DMAP_phys); 416 417 bzero(KPD_DMAP_virt, PAGE_SIZE); 418 419 KPDP_DMAP_virt[i] = KPD_DMAP_phys; 420 KPDP_DMAP_virt[i] |= VPTE_RW | VPTE_V | VPTE_U; 421 422 /* For each PD, we have to allocate NPTEPG PT */ 423 for (j = 0; j < NPTEPG; j++) { 424 KPD_DMAP_virt[j] = (i << PDPSHIFT) | (j << PDRSHIFT); 425 KPD_DMAP_virt[j] |= VPTE_RW | VPTE_V | VPTE_PS | VPTE_U; 426 } 427 } 428 } 429 430 /* DMAP for the first 512G */ 431 KPML4virt[0] = KPDP_DMAP_phys; 432 KPML4virt[0] |= VPTE_RW | VPTE_V | VPTE_U; 433 434 /* create a 2 MB map of the new stack */ 435 pml4_stack_index = (uint64_t)&stack_addr >> PML4SHIFT; 436 KPML4virt[pml4_stack_index] = KPDP_VSTACK_phys; 437 KPML4virt[pml4_stack_index] |= VPTE_RW | VPTE_V | VPTE_U; 438 439 pdp_stack_index = ((uint64_t)&stack_addr & PML4MASK) >> PDPSHIFT; 440 KPDP_VSTACK_virt[pdp_stack_index] = KPD_VSTACK_phys; 441 KPDP_VSTACK_virt[pdp_stack_index] |= VPTE_RW | VPTE_V | VPTE_U; 442 443 pd_stack_index = ((uint64_t)&stack_addr & PDPMASK) >> PDRSHIFT; 444 KPD_VSTACK_virt[pd_stack_index] = (uint64_t) vkernel_stack; 445 KPD_VSTACK_virt[pd_stack_index] |= VPTE_RW | VPTE_V | VPTE_U | VPTE_PS; 446 } 447 448 static void 449 create_pagetables(vm_paddr_t *firstaddr, int64_t ptov_offset) 450 { 451 int i; 452 pml4_entry_t *KPML4virt; 453 pdp_entry_t *KPDPvirt; 454 pd_entry_t *KPDvirt; 455 pt_entry_t *KPTvirt; 456 int kpml4i = pmap_pml4e_index(ptov_offset); 457 int kpdpi = pmap_pdpe_index(ptov_offset); 458 int kpdi = pmap_pde_index(ptov_offset); 459 460 /* 461 * Calculate NKPT - number of kernel page tables. We have to 462 * accomodoate prealloction of the vm_page_array, dump bitmap, 463 * MSGBUF_SIZE, and other stuff. Be generous. 464 * 465 * Maxmem is in pages. 466 */ 467 nkpt = (Maxmem * (sizeof(struct vm_page) * 2) + MSGBUF_SIZE) / NBPDR; 468 /* 469 * Allocate pages 470 */ 471 KPML4phys = allocpages(firstaddr, 1); 472 KPDPphys = allocpages(firstaddr, NKPML4E); 473 KPDphys = allocpages(firstaddr, NKPDPE); 474 KPTphys = allocpages(firstaddr, nkpt); 475 476 KPML4virt = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys); 477 KPDPvirt = (pdp_entry_t *)PHYS_TO_DMAP(KPDPphys); 478 KPDvirt = (pd_entry_t *)PHYS_TO_DMAP(KPDphys); 479 KPTvirt = (pt_entry_t *)PHYS_TO_DMAP(KPTphys); 480 481 bzero(KPML4virt, 1 * PAGE_SIZE); 482 bzero(KPDPvirt, NKPML4E * PAGE_SIZE); 483 bzero(KPDvirt, NKPDPE * PAGE_SIZE); 484 bzero(KPTvirt, nkpt * PAGE_SIZE); 485 486 /* Now map the page tables at their location within PTmap */ 487 for (i = 0; i < nkpt; i++) { 488 KPDvirt[i + kpdi] = KPTphys + (i << PAGE_SHIFT); 489 KPDvirt[i + kpdi] |= VPTE_RW | VPTE_V | VPTE_U; 490 } 491 492 /* And connect up the PD to the PDP */ 493 for (i = 0; i < NKPDPE; i++) { 494 KPDPvirt[i + kpdpi] = KPDphys + (i << PAGE_SHIFT); 495 KPDPvirt[i + kpdpi] |= VPTE_RW | VPTE_V | VPTE_U; 496 } 497 498 /* And recursively map PML4 to itself in order to get PTmap */ 499 KPML4virt[PML4PML4I] = KPML4phys; 500 KPML4virt[PML4PML4I] |= VPTE_RW | VPTE_V | VPTE_U; 501 502 /* Connect the KVA slot up to the PML4 */ 503 KPML4virt[kpml4i] = KPDPphys; 504 KPML4virt[kpml4i] |= VPTE_RW | VPTE_V | VPTE_U; 505 } 506 507 /* 508 * Typically used to initialize a fictitious page by vm/device_pager.c 509 */ 510 void 511 pmap_page_init(struct vm_page *m) 512 { 513 vm_page_init(m); 514 TAILQ_INIT(&m->md.pv_list); 515 } 516 517 /* 518 * Bootstrap the system enough to run with virtual memory. 519 * 520 * On the i386 this is called after mapping has already been enabled 521 * and just syncs the pmap module with what has already been done. 522 * [We can't call it easily with mapping off since the kernel is not 523 * mapped with PA == VA, hence we would have to relocate every address 524 * from the linked base (virtual) address "KERNBASE" to the actual 525 * (physical) address starting relative to 0] 526 */ 527 void 528 pmap_bootstrap(vm_paddr_t *firstaddr, int64_t ptov_offset) 529 { 530 vm_offset_t va; 531 pt_entry_t *pte; 532 533 /* 534 * Create an initial set of page tables to run the kernel in. 535 */ 536 create_pagetables(firstaddr, ptov_offset); 537 538 /* Create the DMAP for the VMM */ 539 if (vmm_enabled) { 540 create_dmap_vmm(firstaddr); 541 } 542 543 virtual_start = KvaStart; 544 virtual_end = KvaEnd; 545 546 /* 547 * Initialize protection array. 548 */ 549 i386_protection_init(); 550 551 /* 552 * The kernel's pmap is statically allocated so we don't have to use 553 * pmap_create, which is unlikely to work correctly at this part of 554 * the boot sequence (XXX and which no longer exists). 555 * 556 * The kernel_pmap's pm_pteobj is used only for locking and not 557 * for mmu pages. 558 */ 559 kernel_pmap.pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys); 560 kernel_pmap.pm_count = 1; 561 /* don't allow deactivation */ 562 CPUMASK_ASSALLONES(kernel_pmap.pm_active); 563 kernel_pmap.pm_pteobj = NULL; /* see pmap_init */ 564 TAILQ_INIT(&kernel_pmap.pm_pvlist); 565 TAILQ_INIT(&kernel_pmap.pm_pvlist_free); 566 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok"); 567 spin_init(&kernel_pmap.pm_spin, "pmapbootstrap"); 568 569 /* 570 * Reserve some special page table entries/VA space for temporary 571 * mapping of pages. 572 */ 573 #define SYSMAP(c, p, v, n) \ 574 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 575 576 va = virtual_start; 577 pte = pmap_pte(&kernel_pmap, va); 578 /* 579 * CMAP1/CMAP2 are used for zeroing and copying pages. 580 */ 581 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 582 583 #if JGV 584 /* 585 * Crashdump maps. 586 */ 587 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 588 #endif 589 590 /* 591 * ptvmmap is used for reading arbitrary physical pages via 592 * /dev/mem. 593 */ 594 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 595 596 /* 597 * msgbufp is used to map the system message buffer. 598 * XXX msgbufmap is not used. 599 */ 600 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 601 atop(round_page(MSGBUF_SIZE))) 602 603 virtual_start = va; 604 605 *CMAP1 = 0; 606 /* Not ready to do an invltlb yet for VMM*/ 607 if (!vmm_enabled) 608 cpu_invltlb(); 609 610 } 611 612 /* 613 * Initialize the pmap module. 614 * Called by vm_init, to initialize any structures that the pmap 615 * system needs to map virtual memory. 616 * pmap_init has been enhanced to support in a fairly consistant 617 * way, discontiguous physical memory. 618 */ 619 void 620 pmap_init(void) 621 { 622 int i; 623 int initial_pvs; 624 625 /* 626 * object for kernel page table pages 627 */ 628 /* JG I think the number can be arbitrary */ 629 kptobj = vm_object_allocate(OBJT_DEFAULT, 5); 630 kernel_pmap.pm_pteobj = kptobj; 631 632 /* 633 * Allocate memory for random pmap data structures. Includes the 634 * pv_head_table. 635 */ 636 for(i = 0; i < vm_page_array_size; i++) { 637 vm_page_t m; 638 639 m = &vm_page_array[i]; 640 TAILQ_INIT(&m->md.pv_list); 641 m->md.pv_list_count = 0; 642 } 643 644 /* 645 * init the pv free list 646 */ 647 initial_pvs = vm_page_array_size; 648 if (initial_pvs < MINPV) 649 initial_pvs = MINPV; 650 pvzone = &pvzone_store; 651 pvinit = (struct pv_entry *) 652 kmem_alloc(&kernel_map, 653 initial_pvs * sizeof (struct pv_entry), 654 VM_SUBSYS_PVENTRY); 655 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit, 656 initial_pvs); 657 658 /* 659 * Now it is safe to enable pv_table recording. 660 */ 661 pmap_initialized = TRUE; 662 } 663 664 /* 665 * Initialize the address space (zone) for the pv_entries. Set a 666 * high water mark so that the system can recover from excessive 667 * numbers of pv entries. 668 */ 669 void 670 pmap_init2(void) 671 { 672 int shpgperproc = PMAP_SHPGPERPROC; 673 674 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 675 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 676 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 677 pv_entry_high_water = 9 * (pv_entry_max / 10); 678 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT); 679 } 680 681 682 /*************************************************** 683 * Low level helper routines..... 684 ***************************************************/ 685 686 /* 687 * The modification bit is not tracked for any pages in this range. XXX 688 * such pages in this maps should always use pmap_k*() functions and not 689 * be managed anyhow. 690 * 691 * XXX User and kernel address spaces are independant for virtual kernels, 692 * this function only applies to the kernel pmap. 693 */ 694 static int 695 pmap_track_modified(pmap_t pmap, vm_offset_t va) 696 { 697 if (pmap != &kernel_pmap) 698 return 1; 699 if ((va < clean_sva) || (va >= clean_eva)) 700 return 1; 701 else 702 return 0; 703 } 704 705 /* 706 * Extract the physical page address associated with the map/VA pair. 707 * 708 * No requirements. 709 */ 710 vm_paddr_t 711 pmap_extract(pmap_t pmap, vm_offset_t va) 712 { 713 vm_paddr_t rtval; 714 pt_entry_t *pte; 715 pd_entry_t pde, *pdep; 716 717 lwkt_gettoken(&vm_token); 718 rtval = 0; 719 pdep = pmap_pde(pmap, va); 720 if (pdep != NULL) { 721 pde = *pdep; 722 if (pde) { 723 if ((pde & VPTE_PS) != 0) { 724 /* JGV */ 725 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK); 726 } else { 727 pte = pmap_pde_to_pte(pdep, va); 728 rtval = (*pte & VPTE_FRAME) | (va & PAGE_MASK); 729 } 730 } 731 } 732 lwkt_reltoken(&vm_token); 733 return rtval; 734 } 735 736 /* 737 * Similar to extract but checks protections, SMP-friendly short-cut for 738 * vm_fault_page[_quick](). 739 */ 740 vm_page_t 741 pmap_fault_page_quick(pmap_t pmap __unused, vm_offset_t vaddr __unused, 742 vm_prot_t prot __unused) 743 { 744 return(NULL); 745 } 746 747 /* 748 * Routine: pmap_kextract 749 * Function: 750 * Extract the physical page address associated 751 * kernel virtual address. 752 */ 753 vm_paddr_t 754 pmap_kextract(vm_offset_t va) 755 { 756 pd_entry_t pde; 757 vm_paddr_t pa; 758 759 KKASSERT(va >= KvaStart && va < KvaEnd); 760 761 /* 762 * The DMAP region is not included in [KvaStart, KvaEnd) 763 */ 764 #if 0 765 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) { 766 pa = DMAP_TO_PHYS(va); 767 } else { 768 #endif 769 pde = *vtopde(va); 770 if (pde & VPTE_PS) { 771 /* JGV */ 772 pa = (pde & PG_PS_FRAME) | (va & PDRMASK); 773 } else { 774 /* 775 * Beware of a concurrent promotion that changes the 776 * PDE at this point! For example, vtopte() must not 777 * be used to access the PTE because it would use the 778 * new PDE. It is, however, safe to use the old PDE 779 * because the page table page is preserved by the 780 * promotion. 781 */ 782 pa = *pmap_pde_to_pte(&pde, va); 783 pa = (pa & VPTE_FRAME) | (va & PAGE_MASK); 784 } 785 #if 0 786 } 787 #endif 788 return pa; 789 } 790 791 /*************************************************** 792 * Low level mapping routines..... 793 ***************************************************/ 794 795 /* 796 * Enter a mapping into kernel_pmap. Mappings created in this fashion 797 * are not managed. Mappings must be immediately accessible on all cpus. 798 * 799 * Call pmap_inval_pte() to invalidate the virtual pte and clean out the 800 * real pmap and handle related races before storing the new vpte. The 801 * new semantics for kenter require use to do an UNCONDITIONAL invalidation, 802 * because the entry may have previously been cleared without an invalidation. 803 */ 804 void 805 pmap_kenter(vm_offset_t va, vm_paddr_t pa) 806 { 807 pt_entry_t *pte; 808 pt_entry_t npte; 809 810 KKASSERT(va >= KvaStart && va < KvaEnd); 811 npte = pa | VPTE_RW | VPTE_V | VPTE_U; 812 pte = vtopte(va); 813 814 #if 1 815 *pte = 0; 816 pmap_inval_pte(pte, &kernel_pmap, va); 817 #else 818 if (*pte & VPTE_V) 819 pmap_inval_pte(pte, &kernel_pmap, va); 820 #endif 821 *pte = npte; 822 } 823 824 /* 825 * Enter an unmanaged KVA mapping for the private use of the current 826 * cpu only. 827 * 828 * It is illegal for the mapping to be accessed by other cpus without 829 * proper invalidation. 830 */ 831 int 832 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa) 833 { 834 pt_entry_t *ptep; 835 pt_entry_t npte; 836 int res; 837 838 KKASSERT(va >= KvaStart && va < KvaEnd); 839 840 npte = (vpte_t)pa | VPTE_RW | VPTE_V | VPTE_U; 841 ptep = vtopte(va); 842 #if 1 843 res = 1; 844 #else 845 /* FUTURE */ 846 res = (*ptep != 0); 847 #endif 848 849 if (*ptep & VPTE_V) 850 pmap_inval_pte_quick(ptep, &kernel_pmap, va); 851 *ptep = npte; 852 853 return res; 854 } 855 856 int 857 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa) 858 { 859 pt_entry_t *ptep; 860 pt_entry_t npte; 861 int res; 862 863 KKASSERT(va >= KvaStart && va < KvaEnd); 864 865 npte = (vpte_t)pa | VPTE_RW | VPTE_V | VPTE_U; 866 ptep = vtopte(va); 867 #if 1 868 res = 1; 869 #else 870 /* FUTURE */ 871 res = (*ptep != 0); 872 #endif 873 874 *ptep = npte; 875 876 return res; 877 } 878 879 /* 880 * Remove an unmanaged mapping created with pmap_kenter*(). 881 */ 882 void 883 pmap_kremove(vm_offset_t va) 884 { 885 pt_entry_t *pte; 886 887 KKASSERT(va >= KvaStart && va < KvaEnd); 888 889 pte = vtopte(va); 890 *pte = 0; 891 pmap_inval_pte(pte, &kernel_pmap, va); 892 } 893 894 /* 895 * Remove an unmanaged mapping created with pmap_kenter*() but synchronize 896 * only with this cpu. 897 * 898 * Unfortunately because we optimize new entries by testing VPTE_V later 899 * on, we actually still have to synchronize with all the cpus. XXX maybe 900 * store a junk value and test against 0 in the other places instead? 901 */ 902 void 903 pmap_kremove_quick(vm_offset_t va) 904 { 905 pt_entry_t *pte; 906 907 KKASSERT(va >= KvaStart && va < KvaEnd); 908 909 pte = vtopte(va); 910 *pte = 0; 911 pmap_inval_pte(pte, &kernel_pmap, va); /* NOT _quick */ 912 } 913 914 void 915 pmap_kremove_noinval(vm_offset_t va) 916 { 917 pt_entry_t *pte; 918 919 KKASSERT(va >= KvaStart && va < KvaEnd); 920 921 pte = vtopte(va); 922 *pte = 0; 923 } 924 925 /* 926 * Used to map a range of physical addresses into kernel 927 * virtual address space. 928 * 929 * For now, VM is already on, we only need to map the 930 * specified memory. 931 */ 932 vm_offset_t 933 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot) 934 { 935 return PHYS_TO_DMAP(start); 936 } 937 938 /* 939 * Map a set of unmanaged VM pages into KVM. 940 */ 941 void 942 pmap_qenter(vm_offset_t va, vm_page_t *m, int count) 943 { 944 vm_offset_t end_va; 945 946 end_va = va + count * PAGE_SIZE; 947 KKASSERT(va >= KvaStart && end_va < KvaEnd); 948 949 while (va < end_va) { 950 pt_entry_t *pte; 951 952 pte = vtopte(va); 953 *pte = 0; 954 pmap_inval_pte(pte, &kernel_pmap, va); 955 *pte = VM_PAGE_TO_PHYS(*m) | VPTE_RW | VPTE_V | VPTE_U; 956 va += PAGE_SIZE; 957 m++; 958 } 959 } 960 961 /* 962 * Undo the effects of pmap_qenter*(). 963 */ 964 void 965 pmap_qremove(vm_offset_t va, int count) 966 { 967 vm_offset_t end_va; 968 969 end_va = va + count * PAGE_SIZE; 970 KKASSERT(va >= KvaStart && end_va < KvaEnd); 971 972 while (va < end_va) { 973 pt_entry_t *pte; 974 975 pte = vtopte(va); 976 atomic_swap_long(pte, 0); 977 pmap_inval_pte(pte, &kernel_pmap, va); 978 va += PAGE_SIZE; 979 } 980 } 981 982 void 983 pmap_qremove_quick(vm_offset_t va, int count) 984 { 985 vm_offset_t end_va; 986 987 end_va = va + count * PAGE_SIZE; 988 KKASSERT(va >= KvaStart && end_va < KvaEnd); 989 990 while (va < end_va) { 991 pt_entry_t *pte; 992 993 pte = vtopte(va); 994 atomic_swap_long(pte, 0); 995 cpu_invlpg((void *)va); 996 va += PAGE_SIZE; 997 } 998 } 999 1000 void 1001 pmap_qremove_noinval(vm_offset_t va, int count) 1002 { 1003 vm_offset_t end_va; 1004 1005 end_va = va + count * PAGE_SIZE; 1006 KKASSERT(va >= KvaStart && end_va < KvaEnd); 1007 1008 while (va < end_va) { 1009 pt_entry_t *pte; 1010 1011 pte = vtopte(va); 1012 atomic_swap_long(pte, 0); 1013 va += PAGE_SIZE; 1014 } 1015 } 1016 1017 /* 1018 * This routine works like vm_page_lookup() but also blocks as long as the 1019 * page is busy. This routine does not busy the page it returns. 1020 * 1021 * Unless the caller is managing objects whos pages are in a known state, 1022 * the call should be made with a critical section held so the page's object 1023 * association remains valid on return. 1024 */ 1025 static vm_page_t 1026 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex) 1027 { 1028 vm_page_t m; 1029 1030 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1031 m = vm_page_lookup_busy_wait(object, pindex, FALSE, "pplookp"); 1032 1033 return(m); 1034 } 1035 1036 /* 1037 * Create a new thread and optionally associate it with a (new) process. 1038 * NOTE! the new thread's cpu may not equal the current cpu. 1039 */ 1040 void 1041 pmap_init_thread(thread_t td) 1042 { 1043 /* enforce pcb placement */ 1044 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1; 1045 td->td_savefpu = &td->td_pcb->pcb_save; 1046 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */ 1047 } 1048 1049 /* 1050 * This routine directly affects the fork perf for a process. 1051 */ 1052 void 1053 pmap_init_proc(struct proc *p) 1054 { 1055 } 1056 1057 /*************************************************** 1058 * Page table page management routines..... 1059 ***************************************************/ 1060 1061 static __inline int pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, 1062 vm_page_t m); 1063 1064 /* 1065 * This routine unholds page table pages, and if the hold count 1066 * drops to zero, then it decrements the wire count. 1067 * 1068 * We must recheck that this is the last hold reference after busy-sleeping 1069 * on the page. 1070 */ 1071 static int 1072 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m) 1073 { 1074 vm_page_busy_wait(m, FALSE, "pmuwpt"); 1075 KASSERT(m->queue == PQ_NONE, 1076 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m)); 1077 1078 if (m->hold_count == 1) { 1079 /* 1080 * Unmap the page table page. 1081 */ 1082 //abort(); /* JG */ 1083 /* pmap_inval_add(info, pmap, -1); */ 1084 1085 if (m->pindex >= (NUPDE + NUPDPE)) { 1086 /* PDP page */ 1087 pml4_entry_t *pml4; 1088 pml4 = pmap_pml4e(pmap, va); 1089 *pml4 = 0; 1090 } else if (m->pindex >= NUPDE) { 1091 /* PD page */ 1092 pdp_entry_t *pdp; 1093 pdp = pmap_pdpe(pmap, va); 1094 *pdp = 0; 1095 } else { 1096 /* PT page */ 1097 pd_entry_t *pd; 1098 pd = pmap_pde(pmap, va); 1099 *pd = 0; 1100 } 1101 1102 KKASSERT(pmap->pm_stats.resident_count > 0); 1103 --pmap->pm_stats.resident_count; 1104 1105 if (pmap->pm_ptphint == m) 1106 pmap->pm_ptphint = NULL; 1107 1108 if (m->pindex < NUPDE) { 1109 /* We just released a PT, unhold the matching PD */ 1110 vm_page_t pdpg; 1111 1112 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & VPTE_FRAME); 1113 pmap_unwire_pte_hold(pmap, va, pdpg); 1114 } 1115 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) { 1116 /* We just released a PD, unhold the matching PDP */ 1117 vm_page_t pdppg; 1118 1119 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & VPTE_FRAME); 1120 pmap_unwire_pte_hold(pmap, va, pdppg); 1121 } 1122 1123 /* 1124 * This was our last hold, the page had better be unwired 1125 * after we decrement wire_count. 1126 * 1127 * FUTURE NOTE: shared page directory page could result in 1128 * multiple wire counts. 1129 */ 1130 vm_page_unhold(m); 1131 --m->wire_count; 1132 KKASSERT(m->wire_count == 0); 1133 atomic_add_int(&vmstats.v_wire_count, -1); 1134 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 1135 vm_page_flash(m); 1136 vm_page_free_zero(m); 1137 return 1; 1138 } else { 1139 KKASSERT(m->hold_count > 1); 1140 vm_page_unhold(m); 1141 vm_page_wakeup(m); 1142 return 0; 1143 } 1144 } 1145 1146 static __inline int 1147 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m) 1148 { 1149 KKASSERT(m->hold_count > 0); 1150 if (m->hold_count > 1) { 1151 vm_page_unhold(m); 1152 return 0; 1153 } else { 1154 return _pmap_unwire_pte_hold(pmap, va, m); 1155 } 1156 } 1157 1158 /* 1159 * After removing a page table entry, this routine is used to 1160 * conditionally free the page, and manage the hold/wire counts. 1161 */ 1162 static int 1163 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 1164 { 1165 /* JG Use FreeBSD/amd64 or FreeBSD/i386 ptepde approaches? */ 1166 vm_pindex_t ptepindex; 1167 1168 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj)); 1169 1170 if (mpte == NULL) { 1171 /* 1172 * page table pages in the kernel_pmap are not managed. 1173 */ 1174 if (pmap == &kernel_pmap) 1175 return(0); 1176 ptepindex = pmap_pde_pindex(va); 1177 if (pmap->pm_ptphint && 1178 (pmap->pm_ptphint->pindex == ptepindex)) { 1179 mpte = pmap->pm_ptphint; 1180 } else { 1181 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 1182 pmap->pm_ptphint = mpte; 1183 vm_page_wakeup(mpte); 1184 } 1185 } 1186 1187 return pmap_unwire_pte_hold(pmap, va, mpte); 1188 } 1189 1190 /* 1191 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we 1192 * just dummy it up so it works well enough for fork(). 1193 * 1194 * In DragonFly, process pmaps may only be used to manipulate user address 1195 * space, never kernel address space. 1196 */ 1197 void 1198 pmap_pinit0(struct pmap *pmap) 1199 { 1200 pmap_pinit(pmap); 1201 } 1202 1203 /* 1204 * Initialize a preallocated and zeroed pmap structure, 1205 * such as one in a vmspace structure. 1206 */ 1207 void 1208 pmap_pinit(struct pmap *pmap) 1209 { 1210 vm_page_t ptdpg; 1211 1212 /* 1213 * No need to allocate page table space yet but we do need a valid 1214 * page directory table. 1215 */ 1216 if (pmap->pm_pml4 == NULL) { 1217 pmap->pm_pml4 = (pml4_entry_t *) 1218 kmem_alloc_pageable(&kernel_map, PAGE_SIZE, 1219 VM_SUBSYS_PML4); 1220 } 1221 1222 /* 1223 * Allocate an object for the ptes 1224 */ 1225 if (pmap->pm_pteobj == NULL) 1226 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1); 1227 1228 /* 1229 * Allocate the page directory page, unless we already have 1230 * one cached. If we used the cached page the wire_count will 1231 * already be set appropriately. 1232 */ 1233 if ((ptdpg = pmap->pm_pdirm) == NULL) { 1234 ptdpg = vm_page_grab(pmap->pm_pteobj, 1235 NUPDE + NUPDPE + PML4PML4I, 1236 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | 1237 VM_ALLOC_ZERO); 1238 pmap->pm_pdirm = ptdpg; 1239 vm_page_flag_clear(ptdpg, PG_MAPPED); 1240 vm_page_wire(ptdpg); 1241 vm_page_wakeup(ptdpg); 1242 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg)); 1243 } 1244 pmap->pm_count = 1; 1245 CPUMASK_ASSZERO(pmap->pm_active); 1246 pmap->pm_ptphint = NULL; 1247 TAILQ_INIT(&pmap->pm_pvlist); 1248 TAILQ_INIT(&pmap->pm_pvlist_free); 1249 spin_init(&pmap->pm_spin, "pmapinit"); 1250 lwkt_token_init(&pmap->pm_token, "pmap_tok"); 1251 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1252 pmap->pm_stats.resident_count = 1; 1253 } 1254 1255 /* 1256 * Clean up a pmap structure so it can be physically freed. This routine 1257 * is called by the vmspace dtor function. A great deal of pmap data is 1258 * left passively mapped to improve vmspace management so we have a bit 1259 * of cleanup work to do here. 1260 * 1261 * No requirements. 1262 */ 1263 void 1264 pmap_puninit(pmap_t pmap) 1265 { 1266 vm_page_t p; 1267 1268 KKASSERT(CPUMASK_TESTZERO(pmap->pm_active)); 1269 if ((p = pmap->pm_pdirm) != NULL) { 1270 KKASSERT(pmap->pm_pml4 != NULL); 1271 pmap_kremove((vm_offset_t)pmap->pm_pml4); 1272 vm_page_busy_wait(p, FALSE, "pgpun"); 1273 p->wire_count--; 1274 atomic_add_int(&vmstats.v_wire_count, -1); 1275 vm_page_free_zero(p); 1276 pmap->pm_pdirm = NULL; 1277 } 1278 if (pmap->pm_pml4) { 1279 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE); 1280 pmap->pm_pml4 = NULL; 1281 } 1282 if (pmap->pm_pteobj) { 1283 vm_object_deallocate(pmap->pm_pteobj); 1284 pmap->pm_pteobj = NULL; 1285 } 1286 } 1287 1288 /* 1289 * Wire in kernel global address entries. To avoid a race condition 1290 * between pmap initialization and pmap_growkernel, this procedure 1291 * adds the pmap to the master list (which growkernel scans to update), 1292 * then copies the template. 1293 * 1294 * In a virtual kernel there are no kernel global address entries. 1295 * 1296 * No requirements. 1297 */ 1298 void 1299 pmap_pinit2(struct pmap *pmap) 1300 { 1301 spin_lock(&pmap_spin); 1302 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode); 1303 spin_unlock(&pmap_spin); 1304 } 1305 1306 /* 1307 * Attempt to release and free a vm_page in a pmap. Returns 1 on success, 1308 * 0 on failure (if the procedure had to sleep). 1309 * 1310 * When asked to remove the page directory page itself, we actually just 1311 * leave it cached so we do not have to incur the SMP inval overhead of 1312 * removing the kernel mapping. pmap_puninit() will take care of it. 1313 */ 1314 static int 1315 pmap_release_free_page(struct pmap *pmap, vm_page_t p) 1316 { 1317 /* 1318 * This code optimizes the case of freeing non-busy 1319 * page-table pages. Those pages are zero now, and 1320 * might as well be placed directly into the zero queue. 1321 */ 1322 if (vm_page_busy_try(p, FALSE)) { 1323 vm_page_sleep_busy(p, FALSE, "pmaprl"); 1324 return 0; 1325 } 1326 1327 /* 1328 * Remove the page table page from the processes address space. 1329 */ 1330 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) { 1331 /* 1332 * We are the pml4 table itself. 1333 */ 1334 /* XXX anything to do here? */ 1335 } else if (p->pindex >= (NUPDE + NUPDPE)) { 1336 /* 1337 * We are a PDP page. 1338 * We look for the PML4 entry that points to us. 1339 */ 1340 vm_page_t m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I); 1341 KKASSERT(m4 != NULL); 1342 pml4_entry_t *pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4)); 1343 int idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG; 1344 KKASSERT(pml4[idx] != 0); 1345 pml4[idx] = 0; 1346 m4->hold_count--; 1347 /* JG What about wire_count? */ 1348 } else if (p->pindex >= NUPDE) { 1349 /* 1350 * We are a PD page. 1351 * We look for the PDP entry that points to us. 1352 */ 1353 vm_page_t m3 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG); 1354 KKASSERT(m3 != NULL); 1355 pdp_entry_t *pdp = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3)); 1356 int idx = (p->pindex - NUPDE) % NPDPEPG; 1357 KKASSERT(pdp[idx] != 0); 1358 pdp[idx] = 0; 1359 m3->hold_count--; 1360 /* JG What about wire_count? */ 1361 } else { 1362 /* We are a PT page. 1363 * We look for the PD entry that points to us. 1364 */ 1365 vm_page_t m2 = vm_page_lookup(pmap->pm_pteobj, NUPDE + p->pindex / NPDEPG); 1366 KKASSERT(m2 != NULL); 1367 pd_entry_t *pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2)); 1368 int idx = p->pindex % NPDEPG; 1369 pd[idx] = 0; 1370 m2->hold_count--; 1371 /* JG What about wire_count? */ 1372 } 1373 KKASSERT(pmap->pm_stats.resident_count > 0); 1374 --pmap->pm_stats.resident_count; 1375 1376 if (p->hold_count) { 1377 panic("pmap_release: freeing held pt page " 1378 "pmap=%p pg=%p dmap=%p pi=%ld {%ld,%ld,%ld}", 1379 pmap, p, (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(p)), 1380 p->pindex, NUPDE, NUPDPE, PML4PML4I); 1381 } 1382 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex)) 1383 pmap->pm_ptphint = NULL; 1384 1385 /* 1386 * We leave the top-level page table page cached, wired, and mapped in 1387 * the pmap until the dtor function (pmap_puninit()) gets called. 1388 * However, still clean it up. 1389 */ 1390 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) { 1391 bzero(pmap->pm_pml4, PAGE_SIZE); 1392 vm_page_wakeup(p); 1393 } else { 1394 abort(); 1395 p->wire_count--; 1396 atomic_add_int(&vmstats.v_wire_count, -1); 1397 /* JG eventually revert to using vm_page_free_zero() */ 1398 vm_page_free(p); 1399 } 1400 return 1; 1401 } 1402 1403 /* 1404 * this routine is called if the page table page is not 1405 * mapped correctly. 1406 */ 1407 static vm_page_t 1408 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex) 1409 { 1410 vm_page_t m, pdppg, pdpg; 1411 1412 /* 1413 * Find or fabricate a new pagetable page. Handle allocation 1414 * races by checking m->valid. 1415 */ 1416 m = vm_page_grab(pmap->pm_pteobj, ptepindex, 1417 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY); 1418 1419 KASSERT(m->queue == PQ_NONE, 1420 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1421 1422 /* 1423 * Increment the hold count for the page we will be returning to 1424 * the caller. 1425 */ 1426 m->hold_count++; 1427 vm_page_wire(m); 1428 1429 /* 1430 * Map the pagetable page into the process address space, if 1431 * it isn't already there. 1432 */ 1433 ++pmap->pm_stats.resident_count; 1434 1435 if (ptepindex >= (NUPDE + NUPDPE)) { 1436 pml4_entry_t *pml4; 1437 vm_pindex_t pml4index; 1438 1439 /* Wire up a new PDP page */ 1440 pml4index = ptepindex - (NUPDE + NUPDPE); 1441 pml4 = &pmap->pm_pml4[pml4index]; 1442 *pml4 = VM_PAGE_TO_PHYS(m) | 1443 VPTE_RW | VPTE_V | VPTE_U | 1444 VPTE_A | VPTE_M; 1445 } else if (ptepindex >= NUPDE) { 1446 vm_pindex_t pml4index; 1447 vm_pindex_t pdpindex; 1448 pml4_entry_t *pml4; 1449 pdp_entry_t *pdp; 1450 1451 /* Wire up a new PD page */ 1452 pdpindex = ptepindex - NUPDE; 1453 pml4index = pdpindex >> NPML4EPGSHIFT; 1454 1455 pml4 = &pmap->pm_pml4[pml4index]; 1456 if ((*pml4 & VPTE_V) == 0) { 1457 /* Have to allocate a new PDP page, recurse */ 1458 if (_pmap_allocpte(pmap, NUPDE + NUPDPE + pml4index) 1459 == NULL) { 1460 --m->wire_count; 1461 vm_page_free(m); 1462 return (NULL); 1463 } 1464 } else { 1465 /* Add reference to the PDP page */ 1466 pdppg = PHYS_TO_VM_PAGE(*pml4 & VPTE_FRAME); 1467 pdppg->hold_count++; 1468 } 1469 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME); 1470 1471 /* Now find the pdp page */ 1472 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; 1473 KKASSERT(*pdp == 0); /* JG DEBUG64 */ 1474 *pdp = VM_PAGE_TO_PHYS(m) | VPTE_RW | VPTE_V | VPTE_U | 1475 VPTE_A | VPTE_M; 1476 } else { 1477 vm_pindex_t pml4index; 1478 vm_pindex_t pdpindex; 1479 pml4_entry_t *pml4; 1480 pdp_entry_t *pdp; 1481 pd_entry_t *pd; 1482 1483 /* Wire up a new PT page */ 1484 pdpindex = ptepindex >> NPDPEPGSHIFT; 1485 pml4index = pdpindex >> NPML4EPGSHIFT; 1486 1487 /* First, find the pdp and check that its valid. */ 1488 pml4 = &pmap->pm_pml4[pml4index]; 1489 if ((*pml4 & VPTE_V) == 0) { 1490 /* We miss a PDP page. We ultimately need a PD page. 1491 * Recursively allocating a PD page will allocate 1492 * the missing PDP page and will also allocate 1493 * the PD page we need. 1494 */ 1495 /* Have to allocate a new PD page, recurse */ 1496 if (_pmap_allocpte(pmap, NUPDE + pdpindex) 1497 == NULL) { 1498 --m->wire_count; 1499 vm_page_free(m); 1500 return (NULL); 1501 } 1502 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME); 1503 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; 1504 } else { 1505 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME); 1506 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; 1507 if ((*pdp & VPTE_V) == 0) { 1508 /* Have to allocate a new PD page, recurse */ 1509 if (_pmap_allocpte(pmap, NUPDE + pdpindex) 1510 == NULL) { 1511 --m->wire_count; 1512 vm_page_free(m); 1513 return (NULL); 1514 } 1515 } else { 1516 /* Add reference to the PD page */ 1517 pdpg = PHYS_TO_VM_PAGE(*pdp & VPTE_FRAME); 1518 pdpg->hold_count++; 1519 } 1520 } 1521 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & VPTE_FRAME); 1522 1523 /* Now we know where the page directory page is */ 1524 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)]; 1525 KKASSERT(*pd == 0); /* JG DEBUG64 */ 1526 *pd = VM_PAGE_TO_PHYS(m) | VPTE_RW | VPTE_V | VPTE_U | 1527 VPTE_A | VPTE_M; 1528 } 1529 1530 /* 1531 * Set the page table hint 1532 */ 1533 pmap->pm_ptphint = m; 1534 vm_page_flag_set(m, PG_MAPPED); 1535 vm_page_wakeup(m); 1536 1537 return m; 1538 } 1539 1540 /* 1541 * Determine the page table page required to access the VA in the pmap 1542 * and allocate it if necessary. Return a held vm_page_t for the page. 1543 * 1544 * Only used with user pmaps. 1545 */ 1546 static vm_page_t 1547 pmap_allocpte(pmap_t pmap, vm_offset_t va) 1548 { 1549 vm_pindex_t ptepindex; 1550 pd_entry_t *pd; 1551 vm_page_t m; 1552 1553 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj)); 1554 1555 /* 1556 * Calculate pagetable page index 1557 */ 1558 ptepindex = pmap_pde_pindex(va); 1559 1560 /* 1561 * Get the page directory entry 1562 */ 1563 pd = pmap_pde(pmap, va); 1564 1565 /* 1566 * This supports switching from a 2MB page to a 1567 * normal 4K page. 1568 */ 1569 if (pd != NULL && (*pd & (VPTE_PS | VPTE_V)) == (VPTE_PS | VPTE_V)) { 1570 panic("no promotion/demotion yet"); 1571 *pd = 0; 1572 pd = NULL; 1573 /*cpu_invltlb();*/ 1574 /*smp_invltlb();*/ 1575 } 1576 1577 /* 1578 * If the page table page is mapped, we just increment the 1579 * hold count, and activate it. 1580 */ 1581 if (pd != NULL && (*pd & VPTE_V) != 0) { 1582 /* YYY hint is used here on i386 */ 1583 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 1584 pmap->pm_ptphint = m; 1585 vm_page_hold(m); 1586 vm_page_wakeup(m); 1587 return m; 1588 } 1589 /* 1590 * Here if the pte page isn't mapped, or if it has been deallocated. 1591 */ 1592 return _pmap_allocpte(pmap, ptepindex); 1593 } 1594 1595 1596 /*************************************************** 1597 * Pmap allocation/deallocation routines. 1598 ***************************************************/ 1599 1600 /* 1601 * Release any resources held by the given physical map. 1602 * Called when a pmap initialized by pmap_pinit is being released. 1603 * Should only be called if the map contains no valid mappings. 1604 * 1605 * Caller must hold pmap->pm_token 1606 */ 1607 static int pmap_release_callback(struct vm_page *p, void *data); 1608 1609 void 1610 pmap_release(struct pmap *pmap) 1611 { 1612 vm_object_t object = pmap->pm_pteobj; 1613 struct rb_vm_page_scan_info info; 1614 1615 KKASSERT(pmap != &kernel_pmap); 1616 1617 lwkt_gettoken(&vm_token); 1618 #if defined(DIAGNOSTIC) 1619 if (object->ref_count != 1) 1620 panic("pmap_release: pteobj reference count != 1"); 1621 #endif 1622 1623 info.pmap = pmap; 1624 info.object = object; 1625 1626 KASSERT(CPUMASK_TESTZERO(pmap->pm_active), 1627 ("pmap %p still active! %016jx", 1628 pmap, 1629 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active))); 1630 1631 spin_lock(&pmap_spin); 1632 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode); 1633 spin_unlock(&pmap_spin); 1634 1635 vm_object_hold(object); 1636 do { 1637 info.error = 0; 1638 info.mpte = NULL; 1639 info.limit = object->generation; 1640 1641 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 1642 pmap_release_callback, &info); 1643 if (info.error == 0 && info.mpte) { 1644 if (!pmap_release_free_page(pmap, info.mpte)) 1645 info.error = 1; 1646 } 1647 } while (info.error); 1648 vm_object_drop(object); 1649 lwkt_reltoken(&vm_token); 1650 } 1651 1652 static int 1653 pmap_release_callback(struct vm_page *p, void *data) 1654 { 1655 struct rb_vm_page_scan_info *info = data; 1656 1657 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) { 1658 info->mpte = p; 1659 return(0); 1660 } 1661 if (!pmap_release_free_page(info->pmap, p)) { 1662 info->error = 1; 1663 return(-1); 1664 } 1665 if (info->object->generation != info->limit) { 1666 info->error = 1; 1667 return(-1); 1668 } 1669 return(0); 1670 } 1671 1672 /* 1673 * Grow the number of kernel page table entries, if needed. 1674 * 1675 * No requirements. 1676 */ 1677 void 1678 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend) 1679 { 1680 vm_offset_t addr; 1681 vm_paddr_t paddr; 1682 vm_offset_t ptppaddr; 1683 vm_page_t nkpg; 1684 pd_entry_t *pde, newpdir; 1685 pdp_entry_t newpdp; 1686 1687 addr = kend; 1688 1689 vm_object_hold(kptobj); 1690 if (kernel_vm_end == 0) { 1691 kernel_vm_end = KvaStart; 1692 nkpt = 0; 1693 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & VPTE_V) != 0) { 1694 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1695 nkpt++; 1696 if (kernel_vm_end - 1 >= kernel_map.max_offset) { 1697 kernel_vm_end = kernel_map.max_offset; 1698 break; 1699 } 1700 } 1701 } 1702 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1703 if (addr - 1 >= kernel_map.max_offset) 1704 addr = kernel_map.max_offset; 1705 while (kernel_vm_end < addr) { 1706 pde = pmap_pde(&kernel_pmap, kernel_vm_end); 1707 if (pde == NULL) { 1708 /* We need a new PDP entry */ 1709 nkpg = vm_page_alloc(kptobj, nkpt, 1710 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM 1711 | VM_ALLOC_INTERRUPT); 1712 if (nkpg == NULL) { 1713 panic("pmap_growkernel: no memory to " 1714 "grow kernel"); 1715 } 1716 paddr = VM_PAGE_TO_PHYS(nkpg); 1717 pmap_zero_page(paddr); 1718 newpdp = (pdp_entry_t)(paddr | 1719 VPTE_V | VPTE_RW | VPTE_U | 1720 VPTE_A | VPTE_M); 1721 *pmap_pdpe(&kernel_pmap, kernel_vm_end) = newpdp; 1722 nkpt++; 1723 continue; /* try again */ 1724 } 1725 if ((*pde & VPTE_V) != 0) { 1726 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & 1727 ~(PAGE_SIZE * NPTEPG - 1); 1728 if (kernel_vm_end - 1 >= kernel_map.max_offset) { 1729 kernel_vm_end = kernel_map.max_offset; 1730 break; 1731 } 1732 continue; 1733 } 1734 1735 /* 1736 * This index is bogus, but out of the way 1737 */ 1738 nkpg = vm_page_alloc(kptobj, nkpt, 1739 VM_ALLOC_NORMAL | 1740 VM_ALLOC_SYSTEM | 1741 VM_ALLOC_INTERRUPT); 1742 if (nkpg == NULL) 1743 panic("pmap_growkernel: no memory to grow kernel"); 1744 1745 vm_page_wire(nkpg); 1746 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1747 pmap_zero_page(ptppaddr); 1748 newpdir = (pd_entry_t)(ptppaddr | 1749 VPTE_V | VPTE_RW | VPTE_U | 1750 VPTE_A | VPTE_M); 1751 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir; 1752 nkpt++; 1753 1754 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & 1755 ~(PAGE_SIZE * NPTEPG - 1); 1756 if (kernel_vm_end - 1 >= kernel_map.max_offset) { 1757 kernel_vm_end = kernel_map.max_offset; 1758 break; 1759 } 1760 } 1761 vm_object_drop(kptobj); 1762 } 1763 1764 /* 1765 * Add a reference to the specified pmap. 1766 * 1767 * No requirements. 1768 */ 1769 void 1770 pmap_reference(pmap_t pmap) 1771 { 1772 if (pmap) { 1773 lwkt_gettoken(&vm_token); 1774 ++pmap->pm_count; 1775 lwkt_reltoken(&vm_token); 1776 } 1777 } 1778 1779 /************************************************************************ 1780 * VMSPACE MANAGEMENT * 1781 ************************************************************************ 1782 * 1783 * The VMSPACE management we do in our virtual kernel must be reflected 1784 * in the real kernel. This is accomplished by making vmspace system 1785 * calls to the real kernel. 1786 */ 1787 void 1788 cpu_vmspace_alloc(struct vmspace *vm) 1789 { 1790 int r; 1791 void *rp; 1792 vpte_t vpte; 1793 1794 /* 1795 * If VMM enable, don't do nothing, we 1796 * are able to use real page tables 1797 */ 1798 if (vmm_enabled) 1799 return; 1800 1801 #define USER_SIZE (VM_MAX_USER_ADDRESS - VM_MIN_USER_ADDRESS) 1802 1803 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0) 1804 panic("vmspace_create() failed"); 1805 1806 rp = vmspace_mmap(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE, 1807 PROT_READ|PROT_WRITE, 1808 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED, 1809 MemImageFd, 0); 1810 if (rp == MAP_FAILED) 1811 panic("vmspace_mmap: failed"); 1812 vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE, 1813 MADV_NOSYNC, 0); 1814 vpte = VM_PAGE_TO_PHYS(vmspace_pmap(vm)->pm_pdirm) | VPTE_RW | VPTE_V | VPTE_U; 1815 r = vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE, 1816 MADV_SETMAP, vpte); 1817 if (r < 0) 1818 panic("vmspace_mcontrol: failed"); 1819 } 1820 1821 void 1822 cpu_vmspace_free(struct vmspace *vm) 1823 { 1824 /* 1825 * If VMM enable, don't do nothing, we 1826 * are able to use real page tables 1827 */ 1828 if (vmm_enabled) 1829 return; 1830 1831 if (vmspace_destroy(&vm->vm_pmap) < 0) 1832 panic("vmspace_destroy() failed"); 1833 } 1834 1835 /*************************************************** 1836 * page management routines. 1837 ***************************************************/ 1838 1839 /* 1840 * free the pv_entry back to the free list. This function may be 1841 * called from an interrupt. 1842 */ 1843 static __inline void 1844 free_pv_entry(pv_entry_t pv) 1845 { 1846 pv_entry_count--; 1847 KKASSERT(pv_entry_count >= 0); 1848 zfree(pvzone, pv); 1849 } 1850 1851 /* 1852 * get a new pv_entry, allocating a block from the system 1853 * when needed. This function may be called from an interrupt. 1854 */ 1855 static pv_entry_t 1856 get_pv_entry(void) 1857 { 1858 pv_entry_count++; 1859 if (pv_entry_high_water && 1860 (pv_entry_count > pv_entry_high_water) && 1861 (pmap_pagedaemon_waken == 0)) { 1862 pmap_pagedaemon_waken = 1; 1863 wakeup(&vm_pages_needed); 1864 } 1865 return zalloc(pvzone); 1866 } 1867 1868 /* 1869 * This routine is very drastic, but can save the system 1870 * in a pinch. 1871 * 1872 * No requirements. 1873 */ 1874 void 1875 pmap_collect(void) 1876 { 1877 int i; 1878 vm_page_t m; 1879 static int warningdone=0; 1880 1881 if (pmap_pagedaemon_waken == 0) 1882 return; 1883 lwkt_gettoken(&vm_token); 1884 pmap_pagedaemon_waken = 0; 1885 1886 if (warningdone < 5) { 1887 kprintf("pmap_collect: collecting pv entries -- " 1888 "suggest increasing PMAP_SHPGPERPROC\n"); 1889 warningdone++; 1890 } 1891 1892 for (i = 0; i < vm_page_array_size; i++) { 1893 m = &vm_page_array[i]; 1894 if (m->wire_count || m->hold_count) 1895 continue; 1896 if (vm_page_busy_try(m, TRUE) == 0) { 1897 if (m->wire_count == 0 && m->hold_count == 0) { 1898 pmap_remove_all(m); 1899 } 1900 vm_page_wakeup(m); 1901 } 1902 } 1903 lwkt_reltoken(&vm_token); 1904 } 1905 1906 1907 /* 1908 * If it is the first entry on the list, it is actually 1909 * in the header and we must copy the following entry up 1910 * to the header. Otherwise we must search the list for 1911 * the entry. In either case we free the now unused entry. 1912 * 1913 * caller must hold vm_token. 1914 */ 1915 static int 1916 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va) 1917 { 1918 pv_entry_t pv; 1919 int rtval; 1920 1921 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1922 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1923 if (pmap == pv->pv_pmap && va == pv->pv_va) 1924 break; 1925 } 1926 } else { 1927 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1928 if (va == pv->pv_va) 1929 break; 1930 } 1931 } 1932 1933 /* 1934 * Note that pv_ptem is NULL if the page table page itself is not 1935 * managed, even if the page being removed IS managed. 1936 */ 1937 rtval = 0; 1938 /* JGXXX When can 'pv' be NULL? */ 1939 if (pv) { 1940 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1941 m->md.pv_list_count--; 1942 atomic_add_int(&m->object->agg_pv_list_count, -1); 1943 KKASSERT(m->md.pv_list_count >= 0); 1944 if (TAILQ_EMPTY(&m->md.pv_list)) 1945 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 1946 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1947 ++pmap->pm_generation; 1948 KKASSERT(pmap->pm_pteobj != NULL); 1949 vm_object_hold(pmap->pm_pteobj); 1950 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1951 vm_object_drop(pmap->pm_pteobj); 1952 free_pv_entry(pv); 1953 } 1954 return rtval; 1955 } 1956 1957 /* 1958 * Create a pv entry for page at pa for (pmap, va). If the page table page 1959 * holding the VA is managed, mpte will be non-NULL. 1960 */ 1961 static void 1962 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1963 { 1964 pv_entry_t pv; 1965 1966 crit_enter(); 1967 pv = get_pv_entry(); 1968 pv->pv_va = va; 1969 pv->pv_pmap = pmap; 1970 pv->pv_ptem = mpte; 1971 1972 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1973 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1974 m->md.pv_list_count++; 1975 atomic_add_int(&m->object->agg_pv_list_count, 1); 1976 1977 crit_exit(); 1978 } 1979 1980 /* 1981 * pmap_remove_pte: do the things to unmap a page in a process 1982 */ 1983 static int 1984 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va) 1985 { 1986 pt_entry_t oldpte; 1987 vm_page_t m; 1988 1989 oldpte = pmap_inval_loadandclear(ptq, pmap, va); 1990 if (oldpte & VPTE_WIRED) 1991 --pmap->pm_stats.wired_count; 1992 KKASSERT(pmap->pm_stats.wired_count >= 0); 1993 1994 #if 0 1995 /* 1996 * Machines that don't support invlpg, also don't support 1997 * PG_G. XXX PG_G is disabled for SMP so don't worry about 1998 * the SMP case. 1999 */ 2000 if (oldpte & PG_G) 2001 cpu_invlpg((void *)va); 2002 #endif 2003 KKASSERT(pmap->pm_stats.resident_count > 0); 2004 --pmap->pm_stats.resident_count; 2005 if (oldpte & VPTE_MANAGED) { 2006 m = PHYS_TO_VM_PAGE(oldpte); 2007 if (oldpte & VPTE_M) { 2008 #if defined(PMAP_DIAGNOSTIC) 2009 if (pmap_nw_modified(oldpte)) { 2010 kprintf("pmap_remove: modified page not " 2011 "writable: va: 0x%lx, pte: 0x%lx\n", 2012 va, oldpte); 2013 } 2014 #endif 2015 if (pmap_track_modified(pmap, va)) 2016 vm_page_dirty(m); 2017 } 2018 if (oldpte & VPTE_A) 2019 vm_page_flag_set(m, PG_REFERENCED); 2020 return pmap_remove_entry(pmap, m, va); 2021 } else { 2022 return pmap_unuse_pt(pmap, va, NULL); 2023 } 2024 2025 return 0; 2026 } 2027 2028 /* 2029 * pmap_remove_page: 2030 * 2031 * Remove a single page from a process address space. 2032 * 2033 * This function may not be called from an interrupt if the pmap is 2034 * not kernel_pmap. 2035 */ 2036 static void 2037 pmap_remove_page(struct pmap *pmap, vm_offset_t va) 2038 { 2039 pt_entry_t *pte; 2040 2041 pte = pmap_pte(pmap, va); 2042 if (pte == NULL) 2043 return; 2044 if ((*pte & VPTE_V) == 0) 2045 return; 2046 pmap_remove_pte(pmap, pte, va); 2047 } 2048 2049 /* 2050 * Remove the given range of addresses from the specified map. 2051 * 2052 * It is assumed that the start and end are properly rounded to 2053 * the page size. 2054 * 2055 * This function may not be called from an interrupt if the pmap is 2056 * not kernel_pmap. 2057 * 2058 * No requirements. 2059 */ 2060 void 2061 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva) 2062 { 2063 vm_offset_t va_next; 2064 pml4_entry_t *pml4e; 2065 pdp_entry_t *pdpe; 2066 pd_entry_t ptpaddr, *pde; 2067 pt_entry_t *pte; 2068 2069 if (pmap == NULL) 2070 return; 2071 2072 vm_object_hold(pmap->pm_pteobj); 2073 lwkt_gettoken(&vm_token); 2074 KKASSERT(pmap->pm_stats.resident_count >= 0); 2075 if (pmap->pm_stats.resident_count == 0) { 2076 lwkt_reltoken(&vm_token); 2077 vm_object_drop(pmap->pm_pteobj); 2078 return; 2079 } 2080 2081 /* 2082 * special handling of removing one page. a very 2083 * common operation and easy to short circuit some 2084 * code. 2085 */ 2086 if (sva + PAGE_SIZE == eva) { 2087 pde = pmap_pde(pmap, sva); 2088 if (pde && (*pde & VPTE_PS) == 0) { 2089 pmap_remove_page(pmap, sva); 2090 lwkt_reltoken(&vm_token); 2091 vm_object_drop(pmap->pm_pteobj); 2092 return; 2093 } 2094 } 2095 2096 for (; sva < eva; sva = va_next) { 2097 pml4e = pmap_pml4e(pmap, sva); 2098 if ((*pml4e & VPTE_V) == 0) { 2099 va_next = (sva + NBPML4) & ~PML4MASK; 2100 if (va_next < sva) 2101 va_next = eva; 2102 continue; 2103 } 2104 2105 pdpe = pmap_pml4e_to_pdpe(pml4e, sva); 2106 if ((*pdpe & VPTE_V) == 0) { 2107 va_next = (sva + NBPDP) & ~PDPMASK; 2108 if (va_next < sva) 2109 va_next = eva; 2110 continue; 2111 } 2112 2113 /* 2114 * Calculate index for next page table. 2115 */ 2116 va_next = (sva + NBPDR) & ~PDRMASK; 2117 if (va_next < sva) 2118 va_next = eva; 2119 2120 pde = pmap_pdpe_to_pde(pdpe, sva); 2121 ptpaddr = *pde; 2122 2123 /* 2124 * Weed out invalid mappings. 2125 */ 2126 if (ptpaddr == 0) 2127 continue; 2128 2129 /* 2130 * Check for large page. 2131 */ 2132 if ((ptpaddr & VPTE_PS) != 0) { 2133 /* JG FreeBSD has more complex treatment here */ 2134 KKASSERT(*pde != 0); 2135 pmap_inval_pde(pde, pmap, sva); 2136 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 2137 continue; 2138 } 2139 2140 /* 2141 * Limit our scan to either the end of the va represented 2142 * by the current page table page, or to the end of the 2143 * range being removed. 2144 */ 2145 if (va_next > eva) 2146 va_next = eva; 2147 2148 /* 2149 * NOTE: pmap_remove_pte() can block. 2150 */ 2151 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, 2152 sva += PAGE_SIZE) { 2153 if (*pte == 0) 2154 continue; 2155 if (pmap_remove_pte(pmap, pte, sva)) 2156 break; 2157 } 2158 } 2159 lwkt_reltoken(&vm_token); 2160 vm_object_drop(pmap->pm_pteobj); 2161 } 2162 2163 /* 2164 * Removes this physical page from all physical maps in which it resides. 2165 * Reflects back modify bits to the pager. 2166 * 2167 * This routine may not be called from an interrupt. 2168 * 2169 * No requirements. 2170 */ 2171 static void 2172 pmap_remove_all(vm_page_t m) 2173 { 2174 pt_entry_t *pte, tpte; 2175 pv_entry_t pv; 2176 2177 #if defined(PMAP_DIAGNOSTIC) 2178 /* 2179 * XXX this makes pmap_page_protect(NONE) illegal for non-managed 2180 * pages! 2181 */ 2182 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 2183 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m)); 2184 } 2185 #endif 2186 2187 lwkt_gettoken(&vm_token); 2188 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2189 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0); 2190 --pv->pv_pmap->pm_stats.resident_count; 2191 2192 pte = pmap_pte(pv->pv_pmap, pv->pv_va); 2193 KKASSERT(pte != NULL); 2194 2195 tpte = pmap_inval_loadandclear(pte, pv->pv_pmap, pv->pv_va); 2196 if (tpte & VPTE_WIRED) 2197 pv->pv_pmap->pm_stats.wired_count--; 2198 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0); 2199 2200 if (tpte & VPTE_A) 2201 vm_page_flag_set(m, PG_REFERENCED); 2202 2203 /* 2204 * Update the vm_page_t clean and reference bits. 2205 */ 2206 if (tpte & VPTE_M) { 2207 #if defined(PMAP_DIAGNOSTIC) 2208 if (pmap_nw_modified(tpte)) { 2209 kprintf( 2210 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n", 2211 pv->pv_va, tpte); 2212 } 2213 #endif 2214 if (pmap_track_modified(pv->pv_pmap, pv->pv_va)) 2215 vm_page_dirty(m); 2216 } 2217 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2218 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2219 ++pv->pv_pmap->pm_generation; 2220 m->md.pv_list_count--; 2221 atomic_add_int(&m->object->agg_pv_list_count, -1); 2222 KKASSERT(m->md.pv_list_count >= 0); 2223 if (TAILQ_EMPTY(&m->md.pv_list)) 2224 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 2225 vm_object_hold(pv->pv_pmap->pm_pteobj); 2226 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2227 vm_object_drop(pv->pv_pmap->pm_pteobj); 2228 free_pv_entry(pv); 2229 } 2230 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0); 2231 lwkt_reltoken(&vm_token); 2232 } 2233 2234 /* 2235 * Set the physical protection on the specified range of this map 2236 * as requested. 2237 * 2238 * This function may not be called from an interrupt if the map is 2239 * not the kernel_pmap. 2240 * 2241 * No requirements. 2242 */ 2243 void 2244 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 2245 { 2246 vm_offset_t va_next; 2247 pml4_entry_t *pml4e; 2248 pdp_entry_t *pdpe; 2249 pd_entry_t ptpaddr, *pde; 2250 pt_entry_t *pte; 2251 2252 /* JG review for NX */ 2253 2254 if (pmap == NULL) 2255 return; 2256 2257 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 2258 pmap_remove(pmap, sva, eva); 2259 return; 2260 } 2261 2262 if (prot & VM_PROT_WRITE) 2263 return; 2264 2265 lwkt_gettoken(&vm_token); 2266 2267 for (; sva < eva; sva = va_next) { 2268 2269 pml4e = pmap_pml4e(pmap, sva); 2270 if ((*pml4e & VPTE_V) == 0) { 2271 va_next = (sva + NBPML4) & ~PML4MASK; 2272 if (va_next < sva) 2273 va_next = eva; 2274 continue; 2275 } 2276 2277 pdpe = pmap_pml4e_to_pdpe(pml4e, sva); 2278 if ((*pdpe & VPTE_V) == 0) { 2279 va_next = (sva + NBPDP) & ~PDPMASK; 2280 if (va_next < sva) 2281 va_next = eva; 2282 continue; 2283 } 2284 2285 va_next = (sva + NBPDR) & ~PDRMASK; 2286 if (va_next < sva) 2287 va_next = eva; 2288 2289 pde = pmap_pdpe_to_pde(pdpe, sva); 2290 ptpaddr = *pde; 2291 2292 /* 2293 * Check for large page. 2294 */ 2295 if ((ptpaddr & VPTE_PS) != 0) { 2296 /* JG correct? */ 2297 pmap_clean_pde(pde, pmap, sva); 2298 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 2299 continue; 2300 } 2301 2302 /* 2303 * Weed out invalid mappings. Note: we assume that the page 2304 * directory table is always allocated, and in kernel virtual. 2305 */ 2306 if (ptpaddr == 0) 2307 continue; 2308 2309 if (va_next > eva) 2310 va_next = eva; 2311 2312 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, 2313 sva += PAGE_SIZE) { 2314 pt_entry_t pbits; 2315 vm_page_t m; 2316 2317 /* 2318 * Clean managed pages and also check the accessed 2319 * bit. Just remove write perms for unmanaged 2320 * pages. Be careful of races, turning off write 2321 * access will force a fault rather then setting 2322 * the modified bit at an unexpected time. 2323 */ 2324 if (*pte & VPTE_MANAGED) { 2325 pbits = pmap_clean_pte(pte, pmap, sva); 2326 m = NULL; 2327 if (pbits & VPTE_A) { 2328 m = PHYS_TO_VM_PAGE(pbits & VPTE_FRAME); 2329 vm_page_flag_set(m, PG_REFERENCED); 2330 atomic_clear_long(pte, VPTE_A); 2331 } 2332 if (pbits & VPTE_M) { 2333 if (pmap_track_modified(pmap, sva)) { 2334 if (m == NULL) 2335 m = PHYS_TO_VM_PAGE(pbits & VPTE_FRAME); 2336 vm_page_dirty(m); 2337 } 2338 } 2339 } else { 2340 pbits = pmap_setro_pte(pte, pmap, sva); 2341 } 2342 } 2343 } 2344 lwkt_reltoken(&vm_token); 2345 } 2346 2347 /* 2348 * Enter a managed page into a pmap. If the page is not wired related pmap 2349 * data can be destroyed at any time for later demand-operation. 2350 * 2351 * Insert the vm_page (m) at virtual address (v) in (pmap), with the 2352 * specified protection, and wire the mapping if requested. 2353 * 2354 * NOTE: This routine may not lazy-evaluate or lose information. The 2355 * page must actually be inserted into the given map NOW. 2356 * 2357 * NOTE: When entering a page at a KVA address, the pmap must be the 2358 * kernel_pmap. 2359 * 2360 * No requirements. 2361 */ 2362 void 2363 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 2364 boolean_t wired, vm_map_entry_t entry __unused) 2365 { 2366 vm_paddr_t pa; 2367 pd_entry_t *pde; 2368 pt_entry_t *pte; 2369 vm_paddr_t opa; 2370 pt_entry_t origpte, newpte; 2371 vm_page_t mpte; 2372 2373 if (pmap == NULL) 2374 return; 2375 2376 va = trunc_page(va); 2377 2378 vm_object_hold(pmap->pm_pteobj); 2379 lwkt_gettoken(&vm_token); 2380 2381 /* 2382 * Get the page table page. The kernel_pmap's page table pages 2383 * are preallocated and have no associated vm_page_t. 2384 */ 2385 if (pmap == &kernel_pmap) 2386 mpte = NULL; 2387 else 2388 mpte = pmap_allocpte(pmap, va); 2389 2390 pde = pmap_pde(pmap, va); 2391 if (pde != NULL && (*pde & VPTE_V) != 0) { 2392 if ((*pde & VPTE_PS) != 0) 2393 panic("pmap_enter: attempted pmap_enter on 2MB page"); 2394 pte = pmap_pde_to_pte(pde, va); 2395 } else { 2396 panic("pmap_enter: invalid page directory va=%#lx", va); 2397 } 2398 2399 KKASSERT(pte != NULL); 2400 /* 2401 * Deal with races on the original mapping (though don't worry 2402 * about VPTE_A races) by cleaning it. This will force a fault 2403 * if an attempt is made to write to the page. 2404 */ 2405 pa = VM_PAGE_TO_PHYS(m); 2406 origpte = pmap_clean_pte(pte, pmap, va); 2407 opa = origpte & VPTE_FRAME; 2408 2409 if (origpte & VPTE_PS) 2410 panic("pmap_enter: attempted pmap_enter on 2MB page"); 2411 2412 /* 2413 * Mapping has not changed, must be protection or wiring change. 2414 */ 2415 if (origpte && (opa == pa)) { 2416 /* 2417 * Wiring change, just update stats. We don't worry about 2418 * wiring PT pages as they remain resident as long as there 2419 * are valid mappings in them. Hence, if a user page is wired, 2420 * the PT page will be also. 2421 */ 2422 if (wired && ((origpte & VPTE_WIRED) == 0)) 2423 ++pmap->pm_stats.wired_count; 2424 else if (!wired && (origpte & VPTE_WIRED)) 2425 --pmap->pm_stats.wired_count; 2426 2427 /* 2428 * Remove the extra pte reference. Note that we cannot 2429 * optimize the RO->RW case because we have adjusted the 2430 * wiring count above and may need to adjust the wiring 2431 * bits below. 2432 */ 2433 if (mpte) 2434 mpte->hold_count--; 2435 2436 /* 2437 * We might be turning off write access to the page, 2438 * so we go ahead and sense modify status. 2439 */ 2440 if (origpte & VPTE_MANAGED) { 2441 if ((origpte & VPTE_M) && 2442 pmap_track_modified(pmap, va)) { 2443 vm_page_t om; 2444 om = PHYS_TO_VM_PAGE(opa); 2445 vm_page_dirty(om); 2446 } 2447 pa |= VPTE_MANAGED; 2448 KKASSERT(m->flags & PG_MAPPED); 2449 } 2450 goto validate; 2451 } 2452 /* 2453 * Mapping has changed, invalidate old range and fall through to 2454 * handle validating new mapping. 2455 */ 2456 if (opa) { 2457 int err; 2458 err = pmap_remove_pte(pmap, pte, va); 2459 if (err) 2460 panic("pmap_enter: pte vanished, va: 0x%lx", va); 2461 } 2462 2463 /* 2464 * Enter on the PV list if part of our managed memory. Note that we 2465 * raise IPL while manipulating pv_table since pmap_enter can be 2466 * called at interrupt time. 2467 */ 2468 if (pmap_initialized && 2469 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2470 pmap_insert_entry(pmap, va, mpte, m); 2471 pa |= VPTE_MANAGED; 2472 vm_page_flag_set(m, PG_MAPPED); 2473 } 2474 2475 /* 2476 * Increment counters 2477 */ 2478 ++pmap->pm_stats.resident_count; 2479 if (wired) 2480 pmap->pm_stats.wired_count++; 2481 2482 validate: 2483 /* 2484 * Now validate mapping with desired protection/wiring. 2485 */ 2486 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | VPTE_V | VPTE_U); 2487 2488 if (wired) 2489 newpte |= VPTE_WIRED; 2490 // if (pmap != &kernel_pmap) 2491 newpte |= VPTE_U; 2492 2493 /* 2494 * If the mapping or permission bits are different from the 2495 * (now cleaned) original pte, an update is needed. We've 2496 * already downgraded or invalidated the page so all we have 2497 * to do now is update the bits. 2498 * 2499 * XXX should we synchronize RO->RW changes to avoid another 2500 * fault? 2501 */ 2502 if ((origpte & ~(VPTE_RW|VPTE_M|VPTE_A)) != newpte) { 2503 *pte = newpte | VPTE_A; 2504 if (newpte & VPTE_RW) 2505 vm_page_flag_set(m, PG_WRITEABLE); 2506 } 2507 KKASSERT((newpte & VPTE_MANAGED) == 0 || (m->flags & PG_MAPPED)); 2508 lwkt_reltoken(&vm_token); 2509 vm_object_drop(pmap->pm_pteobj); 2510 } 2511 2512 /* 2513 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired. 2514 * 2515 * Currently this routine may only be used on user pmaps, not kernel_pmap. 2516 * 2517 * No requirements. 2518 */ 2519 void 2520 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m) 2521 { 2522 pt_entry_t *pte; 2523 vm_paddr_t pa; 2524 vm_page_t mpte; 2525 vm_pindex_t ptepindex; 2526 pd_entry_t *ptepa; 2527 2528 KKASSERT(pmap != &kernel_pmap); 2529 2530 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS); 2531 2532 /* 2533 * Calculate pagetable page index 2534 */ 2535 ptepindex = pmap_pde_pindex(va); 2536 2537 vm_object_hold(pmap->pm_pteobj); 2538 lwkt_gettoken(&vm_token); 2539 2540 do { 2541 /* 2542 * Get the page directory entry 2543 */ 2544 ptepa = pmap_pde(pmap, va); 2545 2546 /* 2547 * If the page table page is mapped, we just increment 2548 * the hold count, and activate it. 2549 */ 2550 if (ptepa && (*ptepa & VPTE_V) != 0) { 2551 if (*ptepa & VPTE_PS) 2552 panic("pmap_enter_quick: unexpected mapping into 2MB page"); 2553 if (pmap->pm_ptphint && 2554 (pmap->pm_ptphint->pindex == ptepindex)) { 2555 mpte = pmap->pm_ptphint; 2556 } else { 2557 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex); 2558 pmap->pm_ptphint = mpte; 2559 vm_page_wakeup(mpte); 2560 } 2561 if (mpte) 2562 mpte->hold_count++; 2563 } else { 2564 mpte = _pmap_allocpte(pmap, ptepindex); 2565 } 2566 } while (mpte == NULL); 2567 2568 /* 2569 * Ok, now that the page table page has been validated, get the pte. 2570 * If the pte is already mapped undo mpte's hold_count and 2571 * just return. 2572 */ 2573 pte = pmap_pte(pmap, va); 2574 if (*pte & VPTE_V) { 2575 KKASSERT(mpte != NULL); 2576 pmap_unwire_pte_hold(pmap, va, mpte); 2577 pa = VM_PAGE_TO_PHYS(m); 2578 KKASSERT(((*pte ^ pa) & VPTE_FRAME) == 0); 2579 lwkt_reltoken(&vm_token); 2580 vm_object_drop(pmap->pm_pteobj); 2581 return; 2582 } 2583 2584 /* 2585 * Enter on the PV list if part of our managed memory 2586 */ 2587 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2588 pmap_insert_entry(pmap, va, mpte, m); 2589 vm_page_flag_set(m, PG_MAPPED); 2590 } 2591 2592 /* 2593 * Increment counters 2594 */ 2595 ++pmap->pm_stats.resident_count; 2596 2597 pa = VM_PAGE_TO_PHYS(m); 2598 2599 /* 2600 * Now validate mapping with RO protection 2601 */ 2602 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2603 *pte = (vpte_t)pa | VPTE_V | VPTE_U; 2604 else 2605 *pte = (vpte_t)pa | VPTE_V | VPTE_U | VPTE_MANAGED; 2606 /*pmap_inval_add(&info, pmap, va); shouldn't be needed 0->valid */ 2607 /*pmap_inval_flush(&info); don't need for vkernel */ 2608 lwkt_reltoken(&vm_token); 2609 vm_object_drop(pmap->pm_pteobj); 2610 } 2611 2612 /* 2613 * Make a temporary mapping for a physical address. This is only intended 2614 * to be used for panic dumps. 2615 * 2616 * The caller is responsible for calling smp_invltlb(). 2617 */ 2618 void * 2619 pmap_kenter_temporary(vm_paddr_t pa, long i) 2620 { 2621 pmap_kenter_quick(crashdumpmap + (i * PAGE_SIZE), pa); 2622 return ((void *)crashdumpmap); 2623 } 2624 2625 #define MAX_INIT_PT (96) 2626 2627 /* 2628 * This routine preloads the ptes for a given object into the specified pmap. 2629 * This eliminates the blast of soft faults on process startup and 2630 * immediately after an mmap. 2631 * 2632 * No requirements. 2633 */ 2634 static int pmap_object_init_pt_callback(vm_page_t p, void *data); 2635 2636 void 2637 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot, 2638 vm_object_t object, vm_pindex_t pindex, 2639 vm_size_t size, int limit) 2640 { 2641 struct rb_vm_page_scan_info info; 2642 struct lwp *lp; 2643 vm_size_t psize; 2644 2645 /* 2646 * We can't preinit if read access isn't set or there is no pmap 2647 * or object. 2648 */ 2649 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL) 2650 return; 2651 2652 /* 2653 * We can't preinit if the pmap is not the current pmap 2654 */ 2655 lp = curthread->td_lwp; 2656 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace)) 2657 return; 2658 2659 psize = x86_64_btop(size); 2660 2661 if ((object->type != OBJT_VNODE) || 2662 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) && 2663 (object->resident_page_count > MAX_INIT_PT))) { 2664 return; 2665 } 2666 2667 if (psize + pindex > object->size) { 2668 if (object->size < pindex) 2669 return; 2670 psize = object->size - pindex; 2671 } 2672 2673 if (psize == 0) 2674 return; 2675 2676 /* 2677 * Use a red-black scan to traverse the requested range and load 2678 * any valid pages found into the pmap. 2679 * 2680 * We cannot safely scan the object's memq unless we are in a 2681 * critical section since interrupts can remove pages from objects. 2682 */ 2683 info.start_pindex = pindex; 2684 info.end_pindex = pindex + psize - 1; 2685 info.limit = limit; 2686 info.mpte = NULL; 2687 info.addr = addr; 2688 info.pmap = pmap; 2689 2690 vm_object_hold_shared(object); 2691 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 2692 pmap_object_init_pt_callback, &info); 2693 vm_object_drop(object); 2694 } 2695 2696 static 2697 int 2698 pmap_object_init_pt_callback(vm_page_t p, void *data) 2699 { 2700 struct rb_vm_page_scan_info *info = data; 2701 vm_pindex_t rel_index; 2702 /* 2703 * don't allow an madvise to blow away our really 2704 * free pages allocating pv entries. 2705 */ 2706 if ((info->limit & MAP_PREFAULT_MADVISE) && 2707 vmstats.v_free_count < vmstats.v_free_reserved) { 2708 return(-1); 2709 } 2710 2711 /* 2712 * Ignore list markers and ignore pages we cannot instantly 2713 * busy (while holding the object token). 2714 */ 2715 if (p->flags & PG_MARKER) 2716 return 0; 2717 if (vm_page_busy_try(p, TRUE)) 2718 return 0; 2719 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2720 (p->flags & PG_FICTITIOUS) == 0) { 2721 if ((p->queue - p->pc) == PQ_CACHE) 2722 vm_page_deactivate(p); 2723 rel_index = p->pindex - info->start_pindex; 2724 pmap_enter_quick(info->pmap, 2725 info->addr + x86_64_ptob(rel_index), p); 2726 } 2727 vm_page_wakeup(p); 2728 return(0); 2729 } 2730 2731 /* 2732 * Return TRUE if the pmap is in shape to trivially 2733 * pre-fault the specified address. 2734 * 2735 * Returns FALSE if it would be non-trivial or if a 2736 * pte is already loaded into the slot. 2737 * 2738 * No requirements. 2739 */ 2740 int 2741 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr) 2742 { 2743 pt_entry_t *pte; 2744 pd_entry_t *pde; 2745 int ret; 2746 2747 lwkt_gettoken(&vm_token); 2748 pde = pmap_pde(pmap, addr); 2749 if (pde == NULL || *pde == 0) { 2750 ret = 0; 2751 } else { 2752 pte = pmap_pde_to_pte(pde, addr); 2753 ret = (*pte) ? 0 : 1; 2754 } 2755 lwkt_reltoken(&vm_token); 2756 return (ret); 2757 } 2758 2759 /* 2760 * Change the wiring attribute for a map/virtual-address pair. 2761 * 2762 * The mapping must already exist in the pmap. 2763 * No other requirements. 2764 */ 2765 void 2766 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired, 2767 vm_map_entry_t entry __unused) 2768 { 2769 pt_entry_t *pte; 2770 2771 if (pmap == NULL) 2772 return; 2773 2774 lwkt_gettoken(&vm_token); 2775 pte = pmap_pte(pmap, va); 2776 2777 if (wired && !pmap_pte_w(pte)) 2778 pmap->pm_stats.wired_count++; 2779 else if (!wired && pmap_pte_w(pte)) 2780 pmap->pm_stats.wired_count--; 2781 2782 /* 2783 * Wiring is not a hardware characteristic so there is no need to 2784 * invalidate TLB. However, in an SMP environment we must use 2785 * a locked bus cycle to update the pte (if we are not using 2786 * the pmap_inval_*() API that is)... it's ok to do this for simple 2787 * wiring changes. 2788 */ 2789 if (wired) 2790 atomic_set_long(pte, VPTE_WIRED); 2791 else 2792 atomic_clear_long(pte, VPTE_WIRED); 2793 lwkt_reltoken(&vm_token); 2794 } 2795 2796 /* 2797 * Copy the range specified by src_addr/len 2798 * from the source map to the range dst_addr/len 2799 * in the destination map. 2800 * 2801 * This routine is only advisory and need not do anything. 2802 */ 2803 void 2804 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, 2805 vm_size_t len, vm_offset_t src_addr) 2806 { 2807 /* 2808 * XXX BUGGY. Amoung other things srcmpte is assumed to remain 2809 * valid through blocking calls, and that's just not going to 2810 * be the case. 2811 * 2812 * FIXME! 2813 */ 2814 return; 2815 } 2816 2817 /* 2818 * pmap_zero_page: 2819 * 2820 * Zero the specified physical page. 2821 * 2822 * This function may be called from an interrupt and no locking is 2823 * required. 2824 */ 2825 void 2826 pmap_zero_page(vm_paddr_t phys) 2827 { 2828 vm_offset_t va = PHYS_TO_DMAP(phys); 2829 2830 bzero((void *)va, PAGE_SIZE); 2831 } 2832 2833 /* 2834 * pmap_zero_page: 2835 * 2836 * Zero part of a physical page by mapping it into memory and clearing 2837 * its contents with bzero. 2838 * 2839 * off and size may not cover an area beyond a single hardware page. 2840 */ 2841 void 2842 pmap_zero_page_area(vm_paddr_t phys, int off, int size) 2843 { 2844 crit_enter(); 2845 vm_offset_t virt = PHYS_TO_DMAP(phys); 2846 bzero((char *)virt + off, size); 2847 crit_exit(); 2848 } 2849 2850 /* 2851 * pmap_copy_page: 2852 * 2853 * Copy the physical page from the source PA to the target PA. 2854 * This function may be called from an interrupt. No locking 2855 * is required. 2856 */ 2857 void 2858 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst) 2859 { 2860 vm_offset_t src_virt, dst_virt; 2861 2862 crit_enter(); 2863 src_virt = PHYS_TO_DMAP(src); 2864 dst_virt = PHYS_TO_DMAP(dst); 2865 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE); 2866 crit_exit(); 2867 } 2868 2869 /* 2870 * pmap_copy_page_frag: 2871 * 2872 * Copy the physical page from the source PA to the target PA. 2873 * This function may be called from an interrupt. No locking 2874 * is required. 2875 */ 2876 void 2877 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes) 2878 { 2879 vm_offset_t src_virt, dst_virt; 2880 2881 crit_enter(); 2882 src_virt = PHYS_TO_DMAP(src); 2883 dst_virt = PHYS_TO_DMAP(dst); 2884 bcopy((char *)src_virt + (src & PAGE_MASK), 2885 (char *)dst_virt + (dst & PAGE_MASK), 2886 bytes); 2887 crit_exit(); 2888 } 2889 2890 /* 2891 * Returns true if the pmap's pv is one of the first 16 pvs linked to 2892 * from this page. This count may be changed upwards or downwards 2893 * in the future; it is only necessary that true be returned for a small 2894 * subset of pmaps for proper page aging. 2895 * 2896 * No other requirements. 2897 */ 2898 boolean_t 2899 pmap_page_exists_quick(pmap_t pmap, vm_page_t m) 2900 { 2901 pv_entry_t pv; 2902 int loops = 0; 2903 2904 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2905 return FALSE; 2906 2907 crit_enter(); 2908 lwkt_gettoken(&vm_token); 2909 2910 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2911 if (pv->pv_pmap == pmap) { 2912 lwkt_reltoken(&vm_token); 2913 crit_exit(); 2914 return TRUE; 2915 } 2916 loops++; 2917 if (loops >= 16) 2918 break; 2919 } 2920 lwkt_reltoken(&vm_token); 2921 crit_exit(); 2922 return (FALSE); 2923 } 2924 2925 /* 2926 * Remove all pages from specified address space this aids process 2927 * exit speeds. Also, this code is special cased for current 2928 * process only, but can have the more generic (and slightly slower) 2929 * mode enabled. This is much faster than pmap_remove in the case 2930 * of running down an entire address space. 2931 * 2932 * No other requirements. 2933 */ 2934 void 2935 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 2936 { 2937 pt_entry_t *pte, tpte; 2938 pv_entry_t pv, npv; 2939 vm_page_t m; 2940 int save_generation; 2941 2942 if (pmap->pm_pteobj) 2943 vm_object_hold(pmap->pm_pteobj); 2944 lwkt_gettoken(&vm_token); 2945 2946 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2947 if (pv->pv_va >= eva || pv->pv_va < sva) { 2948 npv = TAILQ_NEXT(pv, pv_plist); 2949 continue; 2950 } 2951 2952 KKASSERT(pmap == pv->pv_pmap); 2953 2954 pte = pmap_pte(pmap, pv->pv_va); 2955 2956 /* 2957 * We cannot remove wired pages from a process' mapping 2958 * at this time 2959 */ 2960 if (*pte & VPTE_WIRED) { 2961 npv = TAILQ_NEXT(pv, pv_plist); 2962 continue; 2963 } 2964 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va); 2965 2966 m = PHYS_TO_VM_PAGE(tpte & VPTE_FRAME); 2967 2968 KASSERT(m < &vm_page_array[vm_page_array_size], 2969 ("pmap_remove_pages: bad tpte %lx", tpte)); 2970 2971 KKASSERT(pmap->pm_stats.resident_count > 0); 2972 --pmap->pm_stats.resident_count; 2973 2974 /* 2975 * Update the vm_page_t clean and reference bits. 2976 */ 2977 if (tpte & VPTE_M) { 2978 vm_page_dirty(m); 2979 } 2980 2981 npv = TAILQ_NEXT(pv, pv_plist); 2982 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 2983 save_generation = ++pmap->pm_generation; 2984 2985 m->md.pv_list_count--; 2986 atomic_add_int(&m->object->agg_pv_list_count, -1); 2987 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2988 if (TAILQ_EMPTY(&m->md.pv_list)) 2989 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 2990 2991 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem); 2992 free_pv_entry(pv); 2993 2994 /* 2995 * Restart the scan if we blocked during the unuse or free 2996 * calls and other removals were made. 2997 */ 2998 if (save_generation != pmap->pm_generation) { 2999 kprintf("Warning: pmap_remove_pages race-A avoided\n"); 3000 npv = TAILQ_FIRST(&pmap->pm_pvlist); 3001 } 3002 } 3003 lwkt_reltoken(&vm_token); 3004 if (pmap->pm_pteobj) 3005 vm_object_drop(pmap->pm_pteobj); 3006 } 3007 3008 /* 3009 * pmap_testbit tests bits in active mappings of a VM page. 3010 */ 3011 static boolean_t 3012 pmap_testbit(vm_page_t m, int bit) 3013 { 3014 pv_entry_t pv; 3015 pt_entry_t *pte; 3016 3017 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3018 return FALSE; 3019 3020 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 3021 return FALSE; 3022 3023 crit_enter(); 3024 3025 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3026 /* 3027 * if the bit being tested is the modified bit, then 3028 * mark clean_map and ptes as never 3029 * modified. 3030 */ 3031 if (bit & (VPTE_A|VPTE_M)) { 3032 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va)) 3033 continue; 3034 } 3035 3036 #if defined(PMAP_DIAGNOSTIC) 3037 if (pv->pv_pmap == NULL) { 3038 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va); 3039 continue; 3040 } 3041 #endif 3042 pte = pmap_pte(pv->pv_pmap, pv->pv_va); 3043 if (*pte & bit) { 3044 crit_exit(); 3045 return TRUE; 3046 } 3047 } 3048 crit_exit(); 3049 return (FALSE); 3050 } 3051 3052 /* 3053 * This routine is used to clear bits in ptes. Certain bits require special 3054 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit. 3055 * 3056 * This routine is only called with certain VPTE_* bit combinations. 3057 */ 3058 static __inline void 3059 pmap_clearbit(vm_page_t m, int bit) 3060 { 3061 pv_entry_t pv; 3062 pt_entry_t *pte; 3063 pt_entry_t pbits; 3064 3065 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3066 return; 3067 3068 crit_enter(); 3069 3070 /* 3071 * Loop over all current mappings setting/clearing as appropos If 3072 * setting RO do we need to clear the VAC? 3073 */ 3074 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3075 /* 3076 * don't write protect pager mappings 3077 */ 3078 if (bit == VPTE_RW) { 3079 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va)) 3080 continue; 3081 } 3082 3083 #if defined(PMAP_DIAGNOSTIC) 3084 if (pv->pv_pmap == NULL) { 3085 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va); 3086 continue; 3087 } 3088 #endif 3089 3090 /* 3091 * Careful here. We can use a locked bus instruction to 3092 * clear VPTE_A or VPTE_M safely but we need to synchronize 3093 * with the target cpus when we mess with VPTE_RW. 3094 * 3095 * On virtual kernels we must force a new fault-on-write 3096 * in the real kernel if we clear the Modify bit ourselves, 3097 * otherwise the real kernel will not get a new fault and 3098 * will never set our Modify bit again. 3099 */ 3100 pte = pmap_pte(pv->pv_pmap, pv->pv_va); 3101 if (*pte & bit) { 3102 if (bit == VPTE_RW) { 3103 /* 3104 * We must also clear VPTE_M when clearing 3105 * VPTE_RW 3106 */ 3107 pbits = pmap_clean_pte(pte, pv->pv_pmap, 3108 pv->pv_va); 3109 if (pbits & VPTE_M) 3110 vm_page_dirty(m); 3111 } else if (bit == VPTE_M) { 3112 /* 3113 * We do not have to make the page read-only 3114 * when clearing the Modify bit. The real 3115 * kernel will make the real PTE read-only 3116 * or otherwise detect the write and set 3117 * our VPTE_M again simply by us invalidating 3118 * the real kernel VA for the pmap (as we did 3119 * above). This allows the real kernel to 3120 * handle the write fault without forwarding 3121 * the fault to us. 3122 */ 3123 atomic_clear_long(pte, VPTE_M); 3124 } else if ((bit & (VPTE_RW|VPTE_M)) == (VPTE_RW|VPTE_M)) { 3125 /* 3126 * We've been asked to clear W & M, I guess 3127 * the caller doesn't want us to update 3128 * the dirty status of the VM page. 3129 */ 3130 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va); 3131 } else { 3132 /* 3133 * We've been asked to clear bits that do 3134 * not interact with hardware. 3135 */ 3136 atomic_clear_long(pte, bit); 3137 } 3138 } 3139 } 3140 crit_exit(); 3141 } 3142 3143 /* 3144 * Lower the permission for all mappings to a given page. 3145 * 3146 * No other requirements. 3147 */ 3148 void 3149 pmap_page_protect(vm_page_t m, vm_prot_t prot) 3150 { 3151 /* JG NX support? */ 3152 if ((prot & VM_PROT_WRITE) == 0) { 3153 lwkt_gettoken(&vm_token); 3154 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 3155 pmap_clearbit(m, VPTE_RW); 3156 vm_page_flag_clear(m, PG_WRITEABLE); 3157 } else { 3158 pmap_remove_all(m); 3159 } 3160 lwkt_reltoken(&vm_token); 3161 } 3162 } 3163 3164 vm_paddr_t 3165 pmap_phys_address(vm_pindex_t ppn) 3166 { 3167 return (x86_64_ptob(ppn)); 3168 } 3169 3170 /* 3171 * Return a count of reference bits for a page, clearing those bits. 3172 * It is not necessary for every reference bit to be cleared, but it 3173 * is necessary that 0 only be returned when there are truly no 3174 * reference bits set. 3175 * 3176 * XXX: The exact number of bits to check and clear is a matter that 3177 * should be tested and standardized at some point in the future for 3178 * optimal aging of shared pages. 3179 * 3180 * No other requirements. 3181 */ 3182 int 3183 pmap_ts_referenced(vm_page_t m) 3184 { 3185 pv_entry_t pv, pvf, pvn; 3186 pt_entry_t *pte; 3187 int rtval = 0; 3188 3189 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3190 return (rtval); 3191 3192 crit_enter(); 3193 lwkt_gettoken(&vm_token); 3194 3195 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 3196 3197 pvf = pv; 3198 3199 do { 3200 pvn = TAILQ_NEXT(pv, pv_list); 3201 3202 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 3203 3204 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 3205 3206 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va)) 3207 continue; 3208 3209 pte = pmap_pte(pv->pv_pmap, pv->pv_va); 3210 3211 if (pte && (*pte & VPTE_A)) { 3212 atomic_clear_long(pte, VPTE_A); 3213 rtval++; 3214 if (rtval > 4) { 3215 break; 3216 } 3217 } 3218 } while ((pv = pvn) != NULL && pv != pvf); 3219 } 3220 lwkt_reltoken(&vm_token); 3221 crit_exit(); 3222 3223 return (rtval); 3224 } 3225 3226 /* 3227 * Return whether or not the specified physical page was modified 3228 * in any physical maps. 3229 * 3230 * No other requirements. 3231 */ 3232 boolean_t 3233 pmap_is_modified(vm_page_t m) 3234 { 3235 boolean_t res; 3236 3237 lwkt_gettoken(&vm_token); 3238 res = pmap_testbit(m, VPTE_M); 3239 lwkt_reltoken(&vm_token); 3240 return (res); 3241 } 3242 3243 /* 3244 * Clear the modify bits on the specified physical page. 3245 * 3246 * No other requirements. 3247 */ 3248 void 3249 pmap_clear_modify(vm_page_t m) 3250 { 3251 lwkt_gettoken(&vm_token); 3252 pmap_clearbit(m, VPTE_M); 3253 lwkt_reltoken(&vm_token); 3254 } 3255 3256 /* 3257 * Clear the reference bit on the specified physical page. 3258 * 3259 * No other requirements. 3260 */ 3261 void 3262 pmap_clear_reference(vm_page_t m) 3263 { 3264 lwkt_gettoken(&vm_token); 3265 pmap_clearbit(m, VPTE_A); 3266 lwkt_reltoken(&vm_token); 3267 } 3268 3269 /* 3270 * Miscellaneous support routines follow 3271 */ 3272 3273 static void 3274 i386_protection_init(void) 3275 { 3276 int *kp, prot; 3277 3278 kp = protection_codes; 3279 for (prot = 0; prot < 8; prot++) { 3280 if (prot & VM_PROT_READ) 3281 *kp |= 0; /* if it's VALID is readeable */ 3282 if (prot & VM_PROT_WRITE) 3283 *kp |= VPTE_RW; 3284 if (prot & VM_PROT_EXECUTE) 3285 *kp |= 0; /* if it's VALID is executable */ 3286 ++kp; 3287 } 3288 } 3289 3290 /* 3291 * Sets the memory attribute for the specified page. 3292 */ 3293 void 3294 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma) 3295 { 3296 /* This is a vkernel, do nothing */ 3297 } 3298 3299 /* 3300 * Change the PAT attribute on an existing kernel memory map. Caller 3301 * must ensure that the virtual memory in question is not accessed 3302 * during the adjustment. 3303 */ 3304 void 3305 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode) 3306 { 3307 /* This is a vkernel, do nothing */ 3308 } 3309 3310 /* 3311 * Perform the pmap work for mincore 3312 * 3313 * No other requirements. 3314 */ 3315 int 3316 pmap_mincore(pmap_t pmap, vm_offset_t addr) 3317 { 3318 pt_entry_t *ptep, pte; 3319 vm_page_t m; 3320 int val = 0; 3321 3322 lwkt_gettoken(&vm_token); 3323 ptep = pmap_pte(pmap, addr); 3324 3325 if (ptep && (pte = *ptep) != 0) { 3326 vm_paddr_t pa; 3327 3328 val = MINCORE_INCORE; 3329 if ((pte & VPTE_MANAGED) == 0) 3330 goto done; 3331 3332 pa = pte & VPTE_FRAME; 3333 3334 m = PHYS_TO_VM_PAGE(pa); 3335 3336 /* 3337 * Modified by us 3338 */ 3339 if (pte & VPTE_M) 3340 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 3341 /* 3342 * Modified by someone 3343 */ 3344 else if (m->dirty || pmap_is_modified(m)) 3345 val |= MINCORE_MODIFIED_OTHER; 3346 /* 3347 * Referenced by us 3348 */ 3349 if (pte & VPTE_A) 3350 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 3351 3352 /* 3353 * Referenced by someone 3354 */ 3355 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) { 3356 val |= MINCORE_REFERENCED_OTHER; 3357 vm_page_flag_set(m, PG_REFERENCED); 3358 } 3359 } 3360 done: 3361 lwkt_reltoken(&vm_token); 3362 return val; 3363 } 3364 3365 /* 3366 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new 3367 * vmspace will be ref'd and the old one will be deref'd. 3368 * 3369 * Caller must hold vmspace->vm_map.token for oldvm and newvm 3370 */ 3371 void 3372 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs) 3373 { 3374 struct vmspace *oldvm; 3375 struct lwp *lp; 3376 3377 crit_enter(); 3378 oldvm = p->p_vmspace; 3379 if (oldvm != newvm) { 3380 if (adjrefs) 3381 vmspace_ref(newvm); 3382 p->p_vmspace = newvm; 3383 KKASSERT(p->p_nthreads == 1); 3384 lp = RB_ROOT(&p->p_lwp_tree); 3385 pmap_setlwpvm(lp, newvm); 3386 if (adjrefs) 3387 vmspace_rel(oldvm); 3388 } 3389 crit_exit(); 3390 } 3391 3392 /* 3393 * Set the vmspace for a LWP. The vmspace is almost universally set the 3394 * same as the process vmspace, but virtual kernels need to swap out contexts 3395 * on a per-lwp basis. 3396 */ 3397 void 3398 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm) 3399 { 3400 struct vmspace *oldvm; 3401 struct pmap *pmap; 3402 3403 oldvm = lp->lwp_vmspace; 3404 if (oldvm != newvm) { 3405 crit_enter(); 3406 lp->lwp_vmspace = newvm; 3407 if (curthread->td_lwp == lp) { 3408 pmap = vmspace_pmap(newvm); 3409 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid); 3410 if (pmap->pm_active_lock & CPULOCK_EXCL) 3411 pmap_interlock_wait(newvm); 3412 #if defined(SWTCH_OPTIM_STATS) 3413 tlb_flush_count++; 3414 #endif 3415 pmap = vmspace_pmap(oldvm); 3416 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active, 3417 mycpu->gd_cpuid); 3418 } 3419 crit_exit(); 3420 } 3421 } 3422 3423 /* 3424 * The swtch code tried to switch in a heavy weight process whos pmap 3425 * is locked by another cpu. We have to wait for the lock to clear before 3426 * the pmap can be used. 3427 */ 3428 void 3429 pmap_interlock_wait (struct vmspace *vm) 3430 { 3431 pmap_t pmap = vmspace_pmap(vm); 3432 3433 if (pmap->pm_active_lock & CPULOCK_EXCL) { 3434 crit_enter(); 3435 while (pmap->pm_active_lock & CPULOCK_EXCL) { 3436 cpu_ccfence(); 3437 pthread_yield(); 3438 } 3439 crit_exit(); 3440 } 3441 } 3442 3443 vm_offset_t 3444 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3445 { 3446 3447 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3448 return addr; 3449 } 3450 3451 addr = roundup2(addr, NBPDR); 3452 return addr; 3453 } 3454 3455 /* 3456 * Used by kmalloc/kfree, page already exists at va 3457 */ 3458 vm_page_t 3459 pmap_kvtom(vm_offset_t va) 3460 { 3461 vpte_t *ptep; 3462 3463 KKASSERT(va >= KvaStart && va < KvaEnd); 3464 ptep = vtopte(va); 3465 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME)); 3466 } 3467 3468 void 3469 pmap_object_init(vm_object_t object) 3470 { 3471 /* empty */ 3472 } 3473 3474 void 3475 pmap_object_free(vm_object_t object) 3476 { 3477 /* empty */ 3478 } 3479