1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * Copyright (c) 1994 John S. Dyson 6 * Copyright (c) 1994 David Greenman 7 * Copyright (c) 2003 Peter Wemm 8 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu> 9 * Copyright (c) 2008, 2009 The DragonFly Project. 10 * Copyright (c) 2008, 2009 Jordan Gordeev. 11 * All rights reserved. 12 * 13 * This code is derived from software contributed to Berkeley by 14 * the Systems Programming Group of the University of Utah Computer 15 * Science Department and William Jolitz of UUNET Technologies Inc. 16 * 17 * Redistribution and use in source and binary forms, with or without 18 * modification, are permitted provided that the following conditions 19 * are met: 20 * 1. Redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer. 22 * 2. Redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution. 25 * 3. All advertising materials mentioning features or use of this software 26 * must display the following acknowledgement: 27 * This product includes software developed by the University of 28 * California, Berkeley and its contributors. 29 * 4. Neither the name of the University nor the names of its contributors 30 * may be used to endorse or promote products derived from this software 31 * without specific prior written permission. 32 * 33 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 34 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 35 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 36 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 37 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 38 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 39 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 40 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 41 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 42 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 43 * SUCH DAMAGE. 44 * 45 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 46 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $ 47 */ 48 49 /* 50 * Manages physical address maps. 51 * 52 * In addition to hardware address maps, this 53 * module is called upon to provide software-use-only 54 * maps which may or may not be stored in the same 55 * form as hardware maps. These pseudo-maps are 56 * used to store intermediate results from copy 57 * operations to and from address spaces. 58 * 59 * Since the information managed by this module is 60 * also stored by the logical address mapping module, 61 * this module may throw away valid virtual-to-physical 62 * mappings at almost any time. However, invalidations 63 * of virtual-to-physical mappings must be done as 64 * requested. 65 * 66 * In order to cope with hardware architectures which 67 * make virtual-to-physical map invalidates expensive, 68 * this module may delay invalidate or reduced protection 69 * operations until such time as they are actually 70 * necessary. This module is given full information as 71 * to which processors are currently using which maps, 72 * and to when physical maps must be made correct. 73 */ 74 75 #if JG 76 #include "opt_disable_pse.h" 77 #include "opt_pmap.h" 78 #endif 79 #include "opt_msgbuf.h" 80 81 #include <sys/param.h> 82 #include <sys/systm.h> 83 #include <sys/kernel.h> 84 #include <sys/proc.h> 85 #include <sys/msgbuf.h> 86 #include <sys/vmmeter.h> 87 #include <sys/mman.h> 88 89 #include <vm/vm.h> 90 #include <vm/vm_param.h> 91 #include <sys/sysctl.h> 92 #include <sys/lock.h> 93 #include <vm/vm_kern.h> 94 #include <vm/vm_page.h> 95 #include <vm/vm_map.h> 96 #include <vm/vm_object.h> 97 #include <vm/vm_extern.h> 98 #include <vm/vm_pageout.h> 99 #include <vm/vm_pager.h> 100 #include <vm/vm_zone.h> 101 102 #include <sys/user.h> 103 #include <sys/thread2.h> 104 #include <sys/sysref2.h> 105 106 #include <machine/cputypes.h> 107 #include <machine/md_var.h> 108 #include <machine/specialreg.h> 109 #include <machine/smp.h> 110 #include <machine_base/apic/apicreg.h> 111 #include <machine/globaldata.h> 112 #include <machine/pmap.h> 113 #include <machine/pmap_inval.h> 114 115 #include <ddb/ddb.h> 116 117 #define PMAP_KEEP_PDIRS 118 #ifndef PMAP_SHPGPERPROC 119 #define PMAP_SHPGPERPROC 200 120 #endif 121 122 #if defined(DIAGNOSTIC) 123 #define PMAP_DIAGNOSTIC 124 #endif 125 126 #define MINPV 2048 127 128 /* 129 * Get PDEs and PTEs for user/kernel address space 130 */ 131 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va); 132 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 133 134 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0) 135 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0) 136 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0) 137 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0) 138 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0) 139 140 141 /* 142 * Given a map and a machine independent protection code, 143 * convert to a vax protection code. 144 */ 145 #define pte_prot(m, p) \ 146 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)]) 147 static int protection_codes[8]; 148 149 struct pmap kernel_pmap; 150 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list); 151 152 vm_paddr_t avail_start; /* PA of first available physical page */ 153 vm_paddr_t avail_end; /* PA of last available physical page */ 154 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */ 155 vm_offset_t virtual2_end; 156 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */ 157 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 158 vm_offset_t KvaStart; /* VA start of KVA space */ 159 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */ 160 vm_offset_t KvaSize; /* max size of kernel virtual address space */ 161 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 162 static int pgeflag; /* PG_G or-in */ 163 static int pseflag; /* PG_PS or-in */ 164 165 static vm_object_t kptobj; 166 167 static int ndmpdp; 168 static vm_paddr_t dmaplimit; 169 static int nkpt; 170 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS; 171 172 static uint64_t KPTbase; 173 static uint64_t KPTphys; 174 static uint64_t KPDphys; /* phys addr of kernel level 2 */ 175 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */ 176 uint64_t KPDPphys; /* phys addr of kernel level 3 */ 177 uint64_t KPML4phys; /* phys addr of kernel level 4 */ 178 179 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */ 180 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */ 181 182 /* 183 * Data for the pv entry allocation mechanism 184 */ 185 static vm_zone_t pvzone; 186 static struct vm_zone pvzone_store; 187 static struct vm_object pvzone_obj; 188 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0; 189 static int pmap_pagedaemon_waken = 0; 190 static struct pv_entry *pvinit; 191 192 /* 193 * All those kernel PT submaps that BSD is so fond of 194 */ 195 pt_entry_t *CMAP1 = 0, *ptmmap; 196 caddr_t CADDR1 = 0, ptvmmap = 0; 197 static pt_entry_t *msgbufmap; 198 struct msgbuf *msgbufp=0; 199 200 /* 201 * Crashdump maps. 202 */ 203 static pt_entry_t *pt_crashdumpmap; 204 static caddr_t crashdumpmap; 205 206 #define DISABLE_PSE 207 208 static pv_entry_t get_pv_entry (void); 209 static void i386_protection_init (void); 210 static void create_pagetables(vm_paddr_t *firstaddr); 211 static void pmap_remove_all (vm_page_t m); 212 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq, 213 vm_offset_t sva, pmap_inval_info_t info); 214 static void pmap_remove_page (struct pmap *pmap, 215 vm_offset_t va, pmap_inval_info_t info); 216 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m, 217 vm_offset_t va, pmap_inval_info_t info); 218 static boolean_t pmap_testbit (vm_page_t m, int bit); 219 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va, 220 vm_page_t mpte, vm_page_t m); 221 222 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va); 223 224 static int pmap_release_free_page (pmap_t pmap, vm_page_t p); 225 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex); 226 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va); 227 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex); 228 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m, 229 pmap_inval_info_t info); 230 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t); 231 static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 232 233 static unsigned pdir4mb; 234 235 /* 236 * Move the kernel virtual free pointer to the next 237 * 2MB. This is used to help improve performance 238 * by using a large (2MB) page for much of the kernel 239 * (.text, .data, .bss) 240 */ 241 static 242 vm_offset_t 243 pmap_kmem_choose(vm_offset_t addr) 244 { 245 vm_offset_t newaddr = addr; 246 247 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); 248 return newaddr; 249 } 250 251 /* 252 * pmap_pte_quick: 253 * 254 * Super fast pmap_pte routine best used when scanning the pv lists. 255 * This eliminates many course-grained invltlb calls. Note that many of 256 * the pv list scans are across different pmaps and it is very wasteful 257 * to do an entire invltlb when checking a single mapping. 258 * 259 * Should only be called while in a critical section. 260 */ 261 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va); 262 263 static 264 pt_entry_t * 265 pmap_pte_quick(pmap_t pmap, vm_offset_t va) 266 { 267 return pmap_pte(pmap, va); 268 } 269 270 /* Return a non-clipped PD index for a given VA */ 271 static __inline 272 vm_pindex_t 273 pmap_pde_pindex(vm_offset_t va) 274 { 275 return va >> PDRSHIFT; 276 } 277 278 /* Return various clipped indexes for a given VA */ 279 static __inline 280 vm_pindex_t 281 pmap_pte_index(vm_offset_t va) 282 { 283 284 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1)); 285 } 286 287 static __inline 288 vm_pindex_t 289 pmap_pde_index(vm_offset_t va) 290 { 291 292 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1)); 293 } 294 295 static __inline 296 vm_pindex_t 297 pmap_pdpe_index(vm_offset_t va) 298 { 299 300 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1)); 301 } 302 303 static __inline 304 vm_pindex_t 305 pmap_pml4e_index(vm_offset_t va) 306 { 307 308 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1)); 309 } 310 311 /* Return a pointer to the PML4 slot that corresponds to a VA */ 312 static __inline 313 pml4_entry_t * 314 pmap_pml4e(pmap_t pmap, vm_offset_t va) 315 { 316 317 return (&pmap->pm_pml4[pmap_pml4e_index(va)]); 318 } 319 320 /* Return a pointer to the PDP slot that corresponds to a VA */ 321 static __inline 322 pdp_entry_t * 323 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va) 324 { 325 pdp_entry_t *pdpe; 326 327 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME); 328 return (&pdpe[pmap_pdpe_index(va)]); 329 } 330 331 /* Return a pointer to the PDP slot that corresponds to a VA */ 332 static __inline 333 pdp_entry_t * 334 pmap_pdpe(pmap_t pmap, vm_offset_t va) 335 { 336 pml4_entry_t *pml4e; 337 338 pml4e = pmap_pml4e(pmap, va); 339 if ((*pml4e & PG_V) == 0) 340 return NULL; 341 return (pmap_pml4e_to_pdpe(pml4e, va)); 342 } 343 344 /* Return a pointer to the PD slot that corresponds to a VA */ 345 static __inline 346 pd_entry_t * 347 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va) 348 { 349 pd_entry_t *pde; 350 351 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME); 352 return (&pde[pmap_pde_index(va)]); 353 } 354 355 /* Return a pointer to the PD slot that corresponds to a VA */ 356 static __inline 357 pd_entry_t * 358 pmap_pde(pmap_t pmap, vm_offset_t va) 359 { 360 pdp_entry_t *pdpe; 361 362 pdpe = pmap_pdpe(pmap, va); 363 if (pdpe == NULL || (*pdpe & PG_V) == 0) 364 return NULL; 365 return (pmap_pdpe_to_pde(pdpe, va)); 366 } 367 368 /* Return a pointer to the PT slot that corresponds to a VA */ 369 static __inline 370 pt_entry_t * 371 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va) 372 { 373 pt_entry_t *pte; 374 375 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME); 376 return (&pte[pmap_pte_index(va)]); 377 } 378 379 /* Return a pointer to the PT slot that corresponds to a VA */ 380 static __inline 381 pt_entry_t * 382 pmap_pte(pmap_t pmap, vm_offset_t va) 383 { 384 pd_entry_t *pde; 385 386 pde = pmap_pde(pmap, va); 387 if (pde == NULL || (*pde & PG_V) == 0) 388 return NULL; 389 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */ 390 return ((pt_entry_t *)pde); 391 return (pmap_pde_to_pte(pde, va)); 392 } 393 394 static __inline 395 pt_entry_t * 396 vtopte(vm_offset_t va) 397 { 398 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); 399 400 return (PTmap + ((va >> PAGE_SHIFT) & mask)); 401 } 402 403 static __inline 404 pd_entry_t * 405 vtopde(vm_offset_t va) 406 { 407 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); 408 409 return (PDmap + ((va >> PDRSHIFT) & mask)); 410 } 411 412 static uint64_t 413 allocpages(vm_paddr_t *firstaddr, long n) 414 { 415 uint64_t ret; 416 417 ret = *firstaddr; 418 bzero((void *)ret, n * PAGE_SIZE); 419 *firstaddr += n * PAGE_SIZE; 420 return (ret); 421 } 422 423 static 424 void 425 create_pagetables(vm_paddr_t *firstaddr) 426 { 427 long i; /* must be 64 bits */ 428 long nkpt_base; 429 long nkpt_phys; 430 431 /* 432 * We are running (mostly) V=P at this point 433 * 434 * Calculate NKPT - number of kernel page tables. We have to 435 * accomodoate prealloction of the vm_page_array, dump bitmap, 436 * MSGBUF_SIZE, and other stuff. Be generous. 437 * 438 * Maxmem is in pages. 439 */ 440 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT; 441 if (ndmpdp < 4) /* Minimum 4GB of dirmap */ 442 ndmpdp = 4; 443 444 /* 445 * Starting at the beginning of kvm (not KERNBASE). 446 */ 447 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR; 448 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR; 449 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E + ndmpdp) + 450 511) / 512; 451 nkpt_phys += 128; 452 453 /* 454 * Starting at KERNBASE - map 2G worth of page table pages. 455 * KERNBASE is offset -2G from the end of kvm. 456 */ 457 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */ 458 459 /* 460 * Allocate pages 461 */ 462 KPTbase = allocpages(firstaddr, nkpt_base); 463 KPTphys = allocpages(firstaddr, nkpt_phys); 464 KPML4phys = allocpages(firstaddr, 1); 465 KPDPphys = allocpages(firstaddr, NKPML4E); 466 KPDphys = allocpages(firstaddr, NKPDPE); 467 468 /* 469 * Calculate the page directory base for KERNBASE, 470 * that is where we start populating the page table pages. 471 * Basically this is the end - 2. 472 */ 473 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT); 474 475 DMPDPphys = allocpages(firstaddr, NDMPML4E); 476 if ((amd_feature & AMDID_PAGE1GB) == 0) 477 DMPDphys = allocpages(firstaddr, ndmpdp); 478 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT; 479 480 /* 481 * Fill in the underlying page table pages for the area around 482 * KERNBASE. This remaps low physical memory to KERNBASE. 483 * 484 * Read-only from zero to physfree 485 * XXX not fully used, underneath 2M pages 486 */ 487 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) { 488 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT; 489 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G; 490 } 491 492 /* 493 * Now map the initial kernel page tables. One block of page 494 * tables is placed at the beginning of kernel virtual memory, 495 * and another block is placed at KERNBASE to map the kernel binary, 496 * data, bss, and initial pre-allocations. 497 */ 498 for (i = 0; i < nkpt_base; i++) { 499 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT); 500 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V; 501 } 502 for (i = 0; i < nkpt_phys; i++) { 503 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT); 504 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V; 505 } 506 507 /* 508 * Map from zero to end of allocations using 2M pages as an 509 * optimization. This will bypass some of the KPTBase pages 510 * above in the KERNBASE area. 511 */ 512 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) { 513 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT; 514 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G; 515 } 516 517 /* 518 * And connect up the PD to the PDP. The kernel pmap is expected 519 * to pre-populate all of its PDs. See NKPDPE in vmparam.h. 520 */ 521 for (i = 0; i < NKPDPE; i++) { 522 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] = 523 KPDphys + (i << PAGE_SHIFT); 524 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |= 525 PG_RW | PG_V | PG_U; 526 } 527 528 /* Now set up the direct map space using either 2MB or 1GB pages */ 529 /* Preset PG_M and PG_A because demotion expects it */ 530 if ((amd_feature & AMDID_PAGE1GB) == 0) { 531 for (i = 0; i < NPDEPG * ndmpdp; i++) { 532 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT; 533 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS | 534 PG_G | PG_M | PG_A; 535 } 536 /* And the direct map space's PDP */ 537 for (i = 0; i < ndmpdp; i++) { 538 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys + 539 (i << PAGE_SHIFT); 540 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U; 541 } 542 } else { 543 for (i = 0; i < ndmpdp; i++) { 544 ((pdp_entry_t *)DMPDPphys)[i] = 545 (vm_paddr_t)i << PDPSHIFT; 546 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS | 547 PG_G | PG_M | PG_A; 548 } 549 } 550 551 /* And recursively map PML4 to itself in order to get PTmap */ 552 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys; 553 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U; 554 555 /* Connect the Direct Map slot up to the PML4 */ 556 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys; 557 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U; 558 559 /* Connect the KVA slot up to the PML4 */ 560 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys; 561 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U; 562 } 563 564 /* 565 * Bootstrap the system enough to run with virtual memory. 566 * 567 * On the i386 this is called after mapping has already been enabled 568 * and just syncs the pmap module with what has already been done. 569 * [We can't call it easily with mapping off since the kernel is not 570 * mapped with PA == VA, hence we would have to relocate every address 571 * from the linked base (virtual) address "KERNBASE" to the actual 572 * (physical) address starting relative to 0] 573 */ 574 void 575 pmap_bootstrap(vm_paddr_t *firstaddr) 576 { 577 vm_offset_t va; 578 pt_entry_t *pte; 579 struct mdglobaldata *gd; 580 int pg; 581 582 KvaStart = VM_MIN_KERNEL_ADDRESS; 583 KvaEnd = VM_MAX_KERNEL_ADDRESS; 584 KvaSize = KvaEnd - KvaStart; 585 586 avail_start = *firstaddr; 587 588 /* 589 * Create an initial set of page tables to run the kernel in. 590 */ 591 create_pagetables(firstaddr); 592 593 virtual2_start = KvaStart; 594 virtual2_end = PTOV_OFFSET; 595 596 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr; 597 virtual_start = pmap_kmem_choose(virtual_start); 598 599 virtual_end = VM_MAX_KERNEL_ADDRESS; 600 601 /* XXX do %cr0 as well */ 602 load_cr4(rcr4() | CR4_PGE | CR4_PSE); 603 load_cr3(KPML4phys); 604 605 /* 606 * Initialize protection array. 607 */ 608 i386_protection_init(); 609 610 /* 611 * The kernel's pmap is statically allocated so we don't have to use 612 * pmap_create, which is unlikely to work correctly at this part of 613 * the boot sequence (XXX and which no longer exists). 614 */ 615 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys); 616 kernel_pmap.pm_count = 1; 617 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK; 618 TAILQ_INIT(&kernel_pmap.pm_pvlist); 619 620 /* 621 * Reserve some special page table entries/VA space for temporary 622 * mapping of pages. 623 */ 624 #define SYSMAP(c, p, v, n) \ 625 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 626 627 va = virtual_start; 628 pte = vtopte(va); 629 630 /* 631 * CMAP1/CMAP2 are used for zeroing and copying pages. 632 */ 633 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 634 635 /* 636 * Crashdump maps. 637 */ 638 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 639 640 /* 641 * ptvmmap is used for reading arbitrary physical pages via 642 * /dev/mem. 643 */ 644 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 645 646 /* 647 * msgbufp is used to map the system message buffer. 648 * XXX msgbufmap is not used. 649 */ 650 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 651 atop(round_page(MSGBUF_SIZE))) 652 653 virtual_start = va; 654 655 *CMAP1 = 0; 656 657 /* 658 * PG_G is terribly broken on SMP because we IPI invltlb's in some 659 * cases rather then invl1pg. Actually, I don't even know why it 660 * works under UP because self-referential page table mappings 661 */ 662 #ifdef SMP 663 pgeflag = 0; 664 #else 665 if (cpu_feature & CPUID_PGE) 666 pgeflag = PG_G; 667 #endif 668 669 /* 670 * Initialize the 4MB page size flag 671 */ 672 pseflag = 0; 673 /* 674 * The 4MB page version of the initial 675 * kernel page mapping. 676 */ 677 pdir4mb = 0; 678 679 #if !defined(DISABLE_PSE) 680 if (cpu_feature & CPUID_PSE) { 681 pt_entry_t ptditmp; 682 /* 683 * Note that we have enabled PSE mode 684 */ 685 pseflag = PG_PS; 686 ptditmp = *(PTmap + x86_64_btop(KERNBASE)); 687 ptditmp &= ~(NBPDR - 1); 688 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag; 689 pdir4mb = ptditmp; 690 691 #ifndef SMP 692 /* 693 * Enable the PSE mode. If we are SMP we can't do this 694 * now because the APs will not be able to use it when 695 * they boot up. 696 */ 697 load_cr4(rcr4() | CR4_PSE); 698 699 /* 700 * We can do the mapping here for the single processor 701 * case. We simply ignore the old page table page from 702 * now on. 703 */ 704 /* 705 * For SMP, we still need 4K pages to bootstrap APs, 706 * PSE will be enabled as soon as all APs are up. 707 */ 708 PTD[KPTDI] = (pd_entry_t)ptditmp; 709 cpu_invltlb(); 710 #endif 711 } 712 #endif 713 714 /* 715 * We need to finish setting up the globaldata page for the BSP. 716 * locore has already populated the page table for the mdglobaldata 717 * portion. 718 */ 719 pg = MDGLOBALDATA_BASEALLOC_PAGES; 720 gd = &CPU_prvspace[0].mdglobaldata; 721 722 cpu_invltlb(); 723 } 724 725 #ifdef SMP 726 /* 727 * Set 4mb pdir for mp startup 728 */ 729 void 730 pmap_set_opt(void) 731 { 732 if (pseflag && (cpu_feature & CPUID_PSE)) { 733 load_cr4(rcr4() | CR4_PSE); 734 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */ 735 cpu_invltlb(); 736 } 737 } 738 } 739 #endif 740 741 /* 742 * XXX: Hack. Required by pmap_init() 743 */ 744 extern vm_offset_t cpu_apic_addr; 745 746 /* 747 * Initialize the pmap module. 748 * Called by vm_init, to initialize any structures that the pmap 749 * system needs to map virtual memory. 750 * pmap_init has been enhanced to support in a fairly consistant 751 * way, discontiguous physical memory. 752 */ 753 void 754 pmap_init(void) 755 { 756 int i; 757 int initial_pvs; 758 759 /* 760 * object for kernel page table pages 761 */ 762 /* JG I think the number can be arbitrary */ 763 kptobj = vm_object_allocate(OBJT_DEFAULT, 5); 764 765 /* 766 * Allocate memory for random pmap data structures. Includes the 767 * pv_head_table. 768 */ 769 770 for(i = 0; i < vm_page_array_size; i++) { 771 vm_page_t m; 772 773 m = &vm_page_array[i]; 774 TAILQ_INIT(&m->md.pv_list); 775 m->md.pv_list_count = 0; 776 } 777 778 /* 779 * init the pv free list 780 */ 781 initial_pvs = vm_page_array_size; 782 if (initial_pvs < MINPV) 783 initial_pvs = MINPV; 784 pvzone = &pvzone_store; 785 pvinit = (void *)kmem_alloc(&kernel_map, 786 initial_pvs * sizeof (struct pv_entry)); 787 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), 788 pvinit, initial_pvs); 789 790 /* 791 * Now it is safe to enable pv_table recording. 792 */ 793 pmap_initialized = TRUE; 794 #ifdef SMP 795 /* 796 * XXX: Hack 797 */ 798 lapic = pmap_mapdev_uncacheable(cpu_apic_addr, sizeof(struct LAPIC)); 799 #endif 800 } 801 802 /* 803 * Initialize the address space (zone) for the pv_entries. Set a 804 * high water mark so that the system can recover from excessive 805 * numbers of pv entries. 806 */ 807 void 808 pmap_init2(void) 809 { 810 int shpgperproc = PMAP_SHPGPERPROC; 811 int entry_max; 812 813 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 814 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 815 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 816 pv_entry_high_water = 9 * (pv_entry_max / 10); 817 818 /* 819 * Subtract out pages already installed in the zone (hack) 820 */ 821 entry_max = pv_entry_max - vm_page_array_size; 822 if (entry_max <= 0) 823 entry_max = 1; 824 825 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1); 826 } 827 828 829 /*************************************************** 830 * Low level helper routines..... 831 ***************************************************/ 832 833 #if defined(PMAP_DIAGNOSTIC) 834 835 /* 836 * This code checks for non-writeable/modified pages. 837 * This should be an invalid condition. 838 */ 839 static 840 int 841 pmap_nw_modified(pt_entry_t pte) 842 { 843 if ((pte & (PG_M|PG_RW)) == PG_M) 844 return 1; 845 else 846 return 0; 847 } 848 #endif 849 850 851 /* 852 * this routine defines the region(s) of memory that should 853 * not be tested for the modified bit. 854 */ 855 static __inline 856 int 857 pmap_track_modified(vm_offset_t va) 858 { 859 if ((va < clean_sva) || (va >= clean_eva)) 860 return 1; 861 else 862 return 0; 863 } 864 865 /* 866 * Extract the physical page address associated with the map/VA pair. 867 * 868 * The caller must hold vm_token if non-blocking operation is desired. 869 */ 870 vm_paddr_t 871 pmap_extract(pmap_t pmap, vm_offset_t va) 872 { 873 vm_paddr_t rtval; 874 pt_entry_t *pte; 875 pd_entry_t pde, *pdep; 876 877 lwkt_gettoken(&vm_token); 878 rtval = 0; 879 pdep = pmap_pde(pmap, va); 880 if (pdep != NULL) { 881 pde = *pdep; 882 if (pde) { 883 if ((pde & PG_PS) != 0) { 884 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK); 885 } else { 886 pte = pmap_pde_to_pte(pdep, va); 887 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK); 888 } 889 } 890 } 891 lwkt_reltoken(&vm_token); 892 return rtval; 893 } 894 895 /* 896 * Extract the physical page address associated kernel virtual address. 897 */ 898 vm_paddr_t 899 pmap_kextract(vm_offset_t va) 900 { 901 pd_entry_t pde; 902 vm_paddr_t pa; 903 904 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) { 905 pa = DMAP_TO_PHYS(va); 906 } else { 907 pde = *vtopde(va); 908 if (pde & PG_PS) { 909 pa = (pde & PG_PS_FRAME) | (va & PDRMASK); 910 } else { 911 /* 912 * Beware of a concurrent promotion that changes the 913 * PDE at this point! For example, vtopte() must not 914 * be used to access the PTE because it would use the 915 * new PDE. It is, however, safe to use the old PDE 916 * because the page table page is preserved by the 917 * promotion. 918 */ 919 pa = *pmap_pde_to_pte(&pde, va); 920 pa = (pa & PG_FRAME) | (va & PAGE_MASK); 921 } 922 } 923 return pa; 924 } 925 926 /*************************************************** 927 * Low level mapping routines..... 928 ***************************************************/ 929 930 /* 931 * Routine: pmap_kenter 932 * Function: 933 * Add a wired page to the KVA 934 * NOTE! note that in order for the mapping to take effect -- you 935 * should do an invltlb after doing the pmap_kenter(). 936 */ 937 void 938 pmap_kenter(vm_offset_t va, vm_paddr_t pa) 939 { 940 pt_entry_t *pte; 941 pt_entry_t npte; 942 pmap_inval_info info; 943 944 pmap_inval_init(&info); 945 npte = pa | PG_RW | PG_V | pgeflag; 946 pte = vtopte(va); 947 pmap_inval_interlock(&info, &kernel_pmap, va); 948 *pte = npte; 949 pmap_inval_deinterlock(&info, &kernel_pmap); 950 pmap_inval_done(&info); 951 } 952 953 /* 954 * Routine: pmap_kenter_quick 955 * Function: 956 * Similar to pmap_kenter(), except we only invalidate the 957 * mapping on the current CPU. 958 */ 959 void 960 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa) 961 { 962 pt_entry_t *pte; 963 pt_entry_t npte; 964 965 npte = pa | PG_RW | PG_V | pgeflag; 966 pte = vtopte(va); 967 *pte = npte; 968 cpu_invlpg((void *)va); 969 } 970 971 void 972 pmap_kenter_sync(vm_offset_t va) 973 { 974 pmap_inval_info info; 975 976 pmap_inval_init(&info); 977 pmap_inval_interlock(&info, &kernel_pmap, va); 978 pmap_inval_deinterlock(&info, &kernel_pmap); 979 pmap_inval_done(&info); 980 } 981 982 void 983 pmap_kenter_sync_quick(vm_offset_t va) 984 { 985 cpu_invlpg((void *)va); 986 } 987 988 /* 989 * remove a page from the kernel pagetables 990 */ 991 void 992 pmap_kremove(vm_offset_t va) 993 { 994 pt_entry_t *pte; 995 pmap_inval_info info; 996 997 pmap_inval_init(&info); 998 pte = vtopte(va); 999 pmap_inval_interlock(&info, &kernel_pmap, va); 1000 *pte = 0; 1001 pmap_inval_deinterlock(&info, &kernel_pmap); 1002 pmap_inval_done(&info); 1003 } 1004 1005 void 1006 pmap_kremove_quick(vm_offset_t va) 1007 { 1008 pt_entry_t *pte; 1009 pte = vtopte(va); 1010 *pte = 0; 1011 cpu_invlpg((void *)va); 1012 } 1013 1014 /* 1015 * XXX these need to be recoded. They are not used in any critical path. 1016 */ 1017 void 1018 pmap_kmodify_rw(vm_offset_t va) 1019 { 1020 *vtopte(va) |= PG_RW; 1021 cpu_invlpg((void *)va); 1022 } 1023 1024 void 1025 pmap_kmodify_nc(vm_offset_t va) 1026 { 1027 *vtopte(va) |= PG_N; 1028 cpu_invlpg((void *)va); 1029 } 1030 1031 /* 1032 * Used to map a range of physical addresses into kernel virtual 1033 * address space during the low level boot, typically to map the 1034 * dump bitmap, message buffer, and vm_page_array. 1035 * 1036 * These mappings are typically made at some pointer after the end of the 1037 * kernel text+data. 1038 * 1039 * We could return PHYS_TO_DMAP(start) here and not allocate any 1040 * via (*virtp), but then kmem from userland and kernel dumps won't 1041 * have access to the related pointers. 1042 */ 1043 vm_offset_t 1044 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot) 1045 { 1046 vm_offset_t va; 1047 vm_offset_t va_start; 1048 1049 /*return PHYS_TO_DMAP(start);*/ 1050 1051 va_start = *virtp; 1052 va = va_start; 1053 1054 while (start < end) { 1055 pmap_kenter_quick(va, start); 1056 va += PAGE_SIZE; 1057 start += PAGE_SIZE; 1058 } 1059 *virtp = va; 1060 return va_start; 1061 } 1062 1063 1064 /* 1065 * Add a list of wired pages to the kva 1066 * this routine is only used for temporary 1067 * kernel mappings that do not need to have 1068 * page modification or references recorded. 1069 * Note that old mappings are simply written 1070 * over. The page *must* be wired. 1071 */ 1072 void 1073 pmap_qenter(vm_offset_t va, vm_page_t *m, int count) 1074 { 1075 vm_offset_t end_va; 1076 1077 end_va = va + count * PAGE_SIZE; 1078 1079 while (va < end_va) { 1080 pt_entry_t *pte; 1081 1082 pte = vtopte(va); 1083 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag; 1084 cpu_invlpg((void *)va); 1085 va += PAGE_SIZE; 1086 m++; 1087 } 1088 smp_invltlb(); 1089 } 1090 1091 /* 1092 * This routine jerks page mappings from the 1093 * kernel -- it is meant only for temporary mappings. 1094 * 1095 * MPSAFE, INTERRUPT SAFE (cluster callback) 1096 */ 1097 void 1098 pmap_qremove(vm_offset_t va, int count) 1099 { 1100 vm_offset_t end_va; 1101 1102 end_va = va + count * PAGE_SIZE; 1103 1104 while (va < end_va) { 1105 pt_entry_t *pte; 1106 1107 pte = vtopte(va); 1108 *pte = 0; 1109 cpu_invlpg((void *)va); 1110 va += PAGE_SIZE; 1111 } 1112 smp_invltlb(); 1113 } 1114 1115 /* 1116 * This routine works like vm_page_lookup() but also blocks as long as the 1117 * page is busy. This routine does not busy the page it returns. 1118 * 1119 * Unless the caller is managing objects whos pages are in a known state, 1120 * the call should be made with a critical section held so the page's object 1121 * association remains valid on return. 1122 */ 1123 static 1124 vm_page_t 1125 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex) 1126 { 1127 vm_page_t m; 1128 1129 do { 1130 m = vm_page_lookup(object, pindex); 1131 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp")); 1132 1133 return(m); 1134 } 1135 1136 /* 1137 * Create a new thread and optionally associate it with a (new) process. 1138 * NOTE! the new thread's cpu may not equal the current cpu. 1139 */ 1140 void 1141 pmap_init_thread(thread_t td) 1142 { 1143 /* enforce pcb placement & alignment */ 1144 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1; 1145 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF); 1146 td->td_savefpu = &td->td_pcb->pcb_save; 1147 td->td_sp = (char *)td->td_pcb; /* no -16 */ 1148 } 1149 1150 /* 1151 * This routine directly affects the fork perf for a process. 1152 */ 1153 void 1154 pmap_init_proc(struct proc *p) 1155 { 1156 } 1157 1158 /* 1159 * Dispose the UPAGES for a process that has exited. 1160 * This routine directly impacts the exit perf of a process. 1161 */ 1162 void 1163 pmap_dispose_proc(struct proc *p) 1164 { 1165 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p)); 1166 } 1167 1168 /*************************************************** 1169 * Page table page management routines..... 1170 ***************************************************/ 1171 1172 /* 1173 * This routine unholds page table pages, and if the hold count 1174 * drops to zero, then it decrements the wire count. 1175 */ 1176 static __inline 1177 int 1178 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m, 1179 pmap_inval_info_t info) 1180 { 1181 KKASSERT(m->hold_count > 0); 1182 if (m->hold_count > 1) { 1183 vm_page_unhold(m); 1184 return 0; 1185 } else { 1186 return _pmap_unwire_pte_hold(pmap, va, m, info); 1187 } 1188 } 1189 1190 static 1191 int 1192 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m, 1193 pmap_inval_info_t info) 1194 { 1195 /* 1196 * Wait until we can busy the page ourselves. We cannot have 1197 * any active flushes if we block. We own one hold count on the 1198 * page so it cannot be freed out from under us. 1199 */ 1200 if (m->flags & PG_BUSY) { 1201 while (vm_page_sleep_busy(m, FALSE, "pmuwpt")) 1202 ; 1203 } 1204 KASSERT(m->queue == PQ_NONE, 1205 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m)); 1206 1207 /* 1208 * This case can occur if new references were acquired while 1209 * we were blocked. 1210 */ 1211 if (m->hold_count > 1) { 1212 KKASSERT(m->hold_count > 1); 1213 vm_page_unhold(m); 1214 return 0; 1215 } 1216 1217 /* 1218 * Unmap the page table page 1219 */ 1220 KKASSERT(m->hold_count == 1); 1221 vm_page_busy(m); 1222 pmap_inval_interlock(info, pmap, -1); 1223 1224 if (m->pindex >= (NUPDE + NUPDPE)) { 1225 /* PDP page */ 1226 pml4_entry_t *pml4; 1227 pml4 = pmap_pml4e(pmap, va); 1228 *pml4 = 0; 1229 } else if (m->pindex >= NUPDE) { 1230 /* PD page */ 1231 pdp_entry_t *pdp; 1232 pdp = pmap_pdpe(pmap, va); 1233 *pdp = 0; 1234 } else { 1235 /* PT page */ 1236 pd_entry_t *pd; 1237 pd = pmap_pde(pmap, va); 1238 *pd = 0; 1239 } 1240 1241 KKASSERT(pmap->pm_stats.resident_count > 0); 1242 --pmap->pm_stats.resident_count; 1243 1244 if (pmap->pm_ptphint == m) 1245 pmap->pm_ptphint = NULL; 1246 pmap_inval_deinterlock(info, pmap); 1247 1248 if (m->pindex < NUPDE) { 1249 /* We just released a PT, unhold the matching PD */ 1250 vm_page_t pdpg; 1251 1252 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME); 1253 pmap_unwire_pte_hold(pmap, va, pdpg, info); 1254 } 1255 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) { 1256 /* We just released a PD, unhold the matching PDP */ 1257 vm_page_t pdppg; 1258 1259 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME); 1260 pmap_unwire_pte_hold(pmap, va, pdppg, info); 1261 } 1262 1263 /* 1264 * This was our last hold, the page had better be unwired 1265 * after we decrement wire_count. 1266 * 1267 * FUTURE NOTE: shared page directory page could result in 1268 * multiple wire counts. 1269 */ 1270 vm_page_unhold(m); 1271 --m->wire_count; 1272 KKASSERT(m->wire_count == 0); 1273 --vmstats.v_wire_count; 1274 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 1275 vm_page_flash(m); 1276 vm_page_free_zero(m); 1277 1278 return 1; 1279 } 1280 1281 /* 1282 * After removing a page table entry, this routine is used to 1283 * conditionally free the page, and manage the hold/wire counts. 1284 */ 1285 static 1286 int 1287 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte, 1288 pmap_inval_info_t info) 1289 { 1290 vm_pindex_t ptepindex; 1291 1292 if (va >= VM_MAX_USER_ADDRESS) 1293 return 0; 1294 1295 if (mpte == NULL) { 1296 ptepindex = pmap_pde_pindex(va); 1297 #if JGHINT 1298 if (pmap->pm_ptphint && 1299 (pmap->pm_ptphint->pindex == ptepindex)) { 1300 mpte = pmap->pm_ptphint; 1301 } else { 1302 #endif 1303 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 1304 pmap->pm_ptphint = mpte; 1305 #if JGHINT 1306 } 1307 #endif 1308 } 1309 return pmap_unwire_pte_hold(pmap, va, mpte, info); 1310 } 1311 1312 /* 1313 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because 1314 * it, and IdlePTD, represents the template used to update all other pmaps. 1315 * 1316 * On architectures where the kernel pmap is not integrated into the user 1317 * process pmap, this pmap represents the process pmap, not the kernel pmap. 1318 * kernel_pmap should be used to directly access the kernel_pmap. 1319 */ 1320 void 1321 pmap_pinit0(struct pmap *pmap) 1322 { 1323 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys); 1324 pmap->pm_count = 1; 1325 pmap->pm_active = 0; 1326 pmap->pm_ptphint = NULL; 1327 TAILQ_INIT(&pmap->pm_pvlist); 1328 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1329 } 1330 1331 /* 1332 * Initialize a preallocated and zeroed pmap structure, 1333 * such as one in a vmspace structure. 1334 */ 1335 void 1336 pmap_pinit(struct pmap *pmap) 1337 { 1338 vm_page_t ptdpg; 1339 1340 /* 1341 * No need to allocate page table space yet but we do need a valid 1342 * page directory table. 1343 */ 1344 if (pmap->pm_pml4 == NULL) { 1345 pmap->pm_pml4 = 1346 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE); 1347 } 1348 1349 /* 1350 * Allocate an object for the ptes 1351 */ 1352 if (pmap->pm_pteobj == NULL) 1353 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1); 1354 1355 /* 1356 * Allocate the page directory page, unless we already have 1357 * one cached. If we used the cached page the wire_count will 1358 * already be set appropriately. 1359 */ 1360 if ((ptdpg = pmap->pm_pdirm) == NULL) { 1361 ptdpg = vm_page_grab(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I, 1362 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1363 pmap->pm_pdirm = ptdpg; 1364 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY); 1365 ptdpg->valid = VM_PAGE_BITS_ALL; 1366 if (ptdpg->wire_count == 0) 1367 ++vmstats.v_wire_count; 1368 ptdpg->wire_count = 1; 1369 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg)); 1370 } 1371 if ((ptdpg->flags & PG_ZERO) == 0) 1372 bzero(pmap->pm_pml4, PAGE_SIZE); 1373 #ifdef PMAP_DEBUG 1374 else 1375 pmap_page_assertzero(VM_PAGE_TO_PHYS(ptdpg)); 1376 #endif 1377 1378 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U; 1379 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U; 1380 1381 /* install self-referential address mapping entry */ 1382 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M; 1383 1384 pmap->pm_count = 1; 1385 pmap->pm_active = 0; 1386 pmap->pm_ptphint = NULL; 1387 TAILQ_INIT(&pmap->pm_pvlist); 1388 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1389 pmap->pm_stats.resident_count = 1; 1390 } 1391 1392 /* 1393 * Clean up a pmap structure so it can be physically freed. This routine 1394 * is called by the vmspace dtor function. A great deal of pmap data is 1395 * left passively mapped to improve vmspace management so we have a bit 1396 * of cleanup work to do here. 1397 */ 1398 void 1399 pmap_puninit(pmap_t pmap) 1400 { 1401 vm_page_t p; 1402 1403 KKASSERT(pmap->pm_active == 0); 1404 lwkt_gettoken(&vm_token); 1405 if ((p = pmap->pm_pdirm) != NULL) { 1406 KKASSERT(pmap->pm_pml4 != NULL); 1407 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys)); 1408 pmap_kremove((vm_offset_t)pmap->pm_pml4); 1409 p->wire_count--; 1410 vmstats.v_wire_count--; 1411 KKASSERT((p->flags & PG_BUSY) == 0); 1412 vm_page_busy(p); 1413 vm_page_free_zero(p); 1414 pmap->pm_pdirm = NULL; 1415 } 1416 if (pmap->pm_pml4) { 1417 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys)); 1418 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE); 1419 pmap->pm_pml4 = NULL; 1420 } 1421 if (pmap->pm_pteobj) { 1422 vm_object_deallocate(pmap->pm_pteobj); 1423 pmap->pm_pteobj = NULL; 1424 } 1425 lwkt_reltoken(&vm_token); 1426 } 1427 1428 /* 1429 * Wire in kernel global address entries. To avoid a race condition 1430 * between pmap initialization and pmap_growkernel, this procedure 1431 * adds the pmap to the master list (which growkernel scans to update), 1432 * then copies the template. 1433 */ 1434 void 1435 pmap_pinit2(struct pmap *pmap) 1436 { 1437 crit_enter(); 1438 lwkt_gettoken(&vm_token); 1439 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode); 1440 /* XXX copies current process, does not fill in MPPTDI */ 1441 lwkt_reltoken(&vm_token); 1442 crit_exit(); 1443 } 1444 1445 /* 1446 * Attempt to release and free a vm_page in a pmap. Returns 1 on success, 1447 * 0 on failure (if the procedure had to sleep). 1448 * 1449 * When asked to remove the page directory page itself, we actually just 1450 * leave it cached so we do not have to incur the SMP inval overhead of 1451 * removing the kernel mapping. pmap_puninit() will take care of it. 1452 */ 1453 static 1454 int 1455 pmap_release_free_page(struct pmap *pmap, vm_page_t p) 1456 { 1457 /* 1458 * This code optimizes the case of freeing non-busy 1459 * page-table pages. Those pages are zero now, and 1460 * might as well be placed directly into the zero queue. 1461 */ 1462 if (vm_page_sleep_busy(p, FALSE, "pmaprl")) 1463 return 0; 1464 1465 vm_page_busy(p); 1466 1467 /* 1468 * Remove the page table page from the processes address space. 1469 */ 1470 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) { 1471 /* 1472 * We are the pml4 table itself. 1473 */ 1474 /* XXX anything to do here? */ 1475 } else if (p->pindex >= (NUPDE + NUPDPE)) { 1476 /* 1477 * Remove a PDP page from the PML4. We do not maintain 1478 * hold counts on the PML4 page. 1479 */ 1480 pml4_entry_t *pml4; 1481 vm_page_t m4; 1482 int idx; 1483 1484 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I); 1485 KKASSERT(m4 != NULL); 1486 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4)); 1487 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG; 1488 KKASSERT(pml4[idx] != 0); 1489 pml4[idx] = 0; 1490 } else if (p->pindex >= NUPDE) { 1491 /* 1492 * Remove a PD page from the PDP and drop the hold count 1493 * on the PDP. The PDP is left cached in the pmap if 1494 * the hold count drops to 0 so the wire count remains 1495 * intact. 1496 */ 1497 vm_page_t m3; 1498 pdp_entry_t *pdp; 1499 int idx; 1500 1501 m3 = vm_page_lookup(pmap->pm_pteobj, 1502 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG); 1503 KKASSERT(m3 != NULL); 1504 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3)); 1505 idx = (p->pindex - NUPDE) % NPDPEPG; 1506 KKASSERT(pdp[idx] != 0); 1507 pdp[idx] = 0; 1508 m3->hold_count--; 1509 } else { 1510 /* 1511 * Remove a PT page from the PD and drop the hold count 1512 * on the PD. The PD is left cached in the pmap if 1513 * the hold count drops to 0 so the wire count remains 1514 * intact. 1515 */ 1516 vm_page_t m2; 1517 pd_entry_t *pd; 1518 int idx; 1519 1520 m2 = vm_page_lookup(pmap->pm_pteobj, 1521 NUPDE + p->pindex / NPDEPG); 1522 KKASSERT(m2 != NULL); 1523 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2)); 1524 idx = p->pindex % NPDEPG; 1525 pd[idx] = 0; 1526 m2->hold_count--; 1527 } 1528 1529 /* 1530 * One fewer mappings in the pmap. p's hold count had better 1531 * be zero. 1532 */ 1533 KKASSERT(pmap->pm_stats.resident_count > 0); 1534 --pmap->pm_stats.resident_count; 1535 if (p->hold_count) 1536 panic("pmap_release: freeing held page table page"); 1537 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex)) 1538 pmap->pm_ptphint = NULL; 1539 1540 /* 1541 * We leave the top-level page table page cached, wired, and mapped in 1542 * the pmap until the dtor function (pmap_puninit()) gets called. 1543 * However, still clean it up so we can set PG_ZERO. 1544 */ 1545 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) { 1546 bzero(pmap->pm_pml4, PAGE_SIZE); 1547 vm_page_flag_set(p, PG_ZERO); 1548 vm_page_wakeup(p); 1549 } else { 1550 p->wire_count--; 1551 KKASSERT(p->wire_count == 0); 1552 vmstats.v_wire_count--; 1553 /* JG eventually revert to using vm_page_free_zero() */ 1554 vm_page_free(p); 1555 } 1556 return 1; 1557 } 1558 1559 /* 1560 * This routine is called when various levels in the page table need to 1561 * be populated. This routine cannot fail. 1562 */ 1563 static 1564 vm_page_t 1565 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex) 1566 { 1567 vm_page_t m; 1568 1569 /* 1570 * Find or fabricate a new pagetable page. This will busy the page. 1571 */ 1572 m = vm_page_grab(pmap->pm_pteobj, ptepindex, 1573 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY); 1574 if ((m->flags & PG_ZERO) == 0) { 1575 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 1576 } 1577 #ifdef PMAP_DEBUG 1578 else { 1579 pmap_page_assertzero(VM_PAGE_TO_PHYS(m)); 1580 } 1581 #endif 1582 1583 KASSERT(m->queue == PQ_NONE, 1584 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1585 1586 /* 1587 * Increment the hold count for the page we will be returning to 1588 * the caller. 1589 */ 1590 m->hold_count++; 1591 if (m->wire_count++ == 0) 1592 vmstats.v_wire_count++; 1593 m->valid = VM_PAGE_BITS_ALL; 1594 vm_page_flag_clear(m, PG_ZERO); 1595 1596 /* 1597 * Map the pagetable page into the process address space, if 1598 * it isn't already there. 1599 * 1600 * It is possible that someone else got in and mapped the page 1601 * directory page while we were blocked, if so just unbusy and 1602 * return the held page. 1603 */ 1604 if (ptepindex >= (NUPDE + NUPDPE)) { 1605 /* 1606 * Wire up a new PDP page in the PML4 1607 */ 1608 vm_pindex_t pml4index; 1609 pml4_entry_t *pml4; 1610 1611 pml4index = ptepindex - (NUPDE + NUPDPE); 1612 pml4 = &pmap->pm_pml4[pml4index]; 1613 if (*pml4 & PG_V) { 1614 if (--m->wire_count == 0) 1615 --vmstats.v_wire_count; 1616 vm_page_wakeup(m); 1617 return(m); 1618 } 1619 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; 1620 } else if (ptepindex >= NUPDE) { 1621 /* 1622 * Wire up a new PD page in the PDP 1623 */ 1624 vm_pindex_t pml4index; 1625 vm_pindex_t pdpindex; 1626 vm_page_t pdppg; 1627 pml4_entry_t *pml4; 1628 pdp_entry_t *pdp; 1629 1630 pdpindex = ptepindex - NUPDE; 1631 pml4index = pdpindex >> NPML4EPGSHIFT; 1632 1633 pml4 = &pmap->pm_pml4[pml4index]; 1634 if ((*pml4 & PG_V) == 0) { 1635 /* 1636 * Have to allocate a new PDP page, recurse. 1637 * This always succeeds. Returned page will 1638 * be held. 1639 */ 1640 pdppg = _pmap_allocpte(pmap, 1641 NUPDE + NUPDPE + pml4index); 1642 } else { 1643 /* 1644 * Add a held reference to the PDP page. 1645 */ 1646 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME); 1647 pdppg->hold_count++; 1648 } 1649 1650 /* 1651 * Now find the pdp_entry and map the PDP. If the PDP 1652 * has already been mapped unwind and return the 1653 * already-mapped PDP held. 1654 * 1655 * pdppg is left held (hold_count is incremented for 1656 * each PD in the PDP). 1657 */ 1658 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME); 1659 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; 1660 if (*pdp & PG_V) { 1661 vm_page_unhold(pdppg); 1662 if (--m->wire_count == 0) 1663 --vmstats.v_wire_count; 1664 vm_page_wakeup(m); 1665 return(m); 1666 } 1667 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; 1668 } else { 1669 /* 1670 * Wire up the new PT page in the PD 1671 */ 1672 vm_pindex_t pml4index; 1673 vm_pindex_t pdpindex; 1674 pml4_entry_t *pml4; 1675 pdp_entry_t *pdp; 1676 pd_entry_t *pd; 1677 vm_page_t pdpg; 1678 1679 pdpindex = ptepindex >> NPDPEPGSHIFT; 1680 pml4index = pdpindex >> NPML4EPGSHIFT; 1681 1682 /* 1683 * Locate the PDP page in the PML4, then the PD page in 1684 * the PDP. If either does not exist we simply recurse 1685 * to allocate them. 1686 * 1687 * We can just recurse on the PD page as it will recurse 1688 * on the PDP if necessary. 1689 */ 1690 pml4 = &pmap->pm_pml4[pml4index]; 1691 if ((*pml4 & PG_V) == 0) { 1692 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex); 1693 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME); 1694 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; 1695 } else { 1696 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME); 1697 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; 1698 if ((*pdp & PG_V) == 0) { 1699 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex); 1700 } else { 1701 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME); 1702 pdpg->hold_count++; 1703 } 1704 } 1705 1706 /* 1707 * Now fill in the pte in the PD. If the pte already exists 1708 * (again, if we raced the grab), unhold pdpg and unwire 1709 * m, returning a held m. 1710 * 1711 * pdpg is left held (hold_count is incremented for 1712 * each PT in the PD). 1713 */ 1714 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME); 1715 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)]; 1716 if (*pd != 0) { 1717 vm_page_unhold(pdpg); 1718 if (--m->wire_count == 0) 1719 --vmstats.v_wire_count; 1720 vm_page_wakeup(m); 1721 return(m); 1722 } 1723 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; 1724 } 1725 1726 /* 1727 * We successfully loaded a PDP, PD, or PTE. Set the page table hint, 1728 * valid bits, mapped flag, unbusy, and we're done. 1729 */ 1730 pmap->pm_ptphint = m; 1731 ++pmap->pm_stats.resident_count; 1732 1733 #if 0 1734 m->valid = VM_PAGE_BITS_ALL; 1735 vm_page_flag_clear(m, PG_ZERO); 1736 #endif 1737 vm_page_flag_set(m, PG_MAPPED); 1738 vm_page_wakeup(m); 1739 1740 return (m); 1741 } 1742 1743 static 1744 vm_page_t 1745 pmap_allocpte(pmap_t pmap, vm_offset_t va) 1746 { 1747 vm_pindex_t ptepindex; 1748 pd_entry_t *pd; 1749 vm_page_t m; 1750 1751 /* 1752 * Calculate pagetable page index 1753 */ 1754 ptepindex = pmap_pde_pindex(va); 1755 1756 /* 1757 * Get the page directory entry 1758 */ 1759 pd = pmap_pde(pmap, va); 1760 1761 /* 1762 * This supports switching from a 2MB page to a 1763 * normal 4K page. 1764 */ 1765 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) { 1766 panic("no promotion/demotion yet"); 1767 *pd = 0; 1768 pd = NULL; 1769 cpu_invltlb(); 1770 smp_invltlb(); 1771 } 1772 1773 /* 1774 * If the page table page is mapped, we just increment the 1775 * hold count, and activate it. 1776 */ 1777 if (pd != NULL && (*pd & PG_V) != 0) { 1778 /* YYY hint is used here on i386 */ 1779 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex); 1780 pmap->pm_ptphint = m; 1781 m->hold_count++; 1782 return m; 1783 } 1784 /* 1785 * Here if the pte page isn't mapped, or if it has been deallocated. 1786 */ 1787 return _pmap_allocpte(pmap, ptepindex); 1788 } 1789 1790 1791 /*************************************************** 1792 * Pmap allocation/deallocation routines. 1793 ***************************************************/ 1794 1795 /* 1796 * Release any resources held by the given physical map. 1797 * Called when a pmap initialized by pmap_pinit is being released. 1798 * Should only be called if the map contains no valid mappings. 1799 */ 1800 static int pmap_release_callback(struct vm_page *p, void *data); 1801 1802 void 1803 pmap_release(struct pmap *pmap) 1804 { 1805 vm_object_t object = pmap->pm_pteobj; 1806 struct rb_vm_page_scan_info info; 1807 1808 KASSERT(pmap->pm_active == 0, 1809 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active)); 1810 #if defined(DIAGNOSTIC) 1811 if (object->ref_count != 1) 1812 panic("pmap_release: pteobj reference count != 1"); 1813 #endif 1814 1815 info.pmap = pmap; 1816 info.object = object; 1817 crit_enter(); 1818 lwkt_gettoken(&vm_token); 1819 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode); 1820 crit_exit(); 1821 1822 do { 1823 crit_enter(); 1824 info.error = 0; 1825 info.mpte = NULL; 1826 info.limit = object->generation; 1827 1828 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 1829 pmap_release_callback, &info); 1830 if (info.error == 0 && info.mpte) { 1831 if (!pmap_release_free_page(pmap, info.mpte)) 1832 info.error = 1; 1833 } 1834 crit_exit(); 1835 } while (info.error); 1836 lwkt_reltoken(&vm_token); 1837 } 1838 1839 static 1840 int 1841 pmap_release_callback(struct vm_page *p, void *data) 1842 { 1843 struct rb_vm_page_scan_info *info = data; 1844 1845 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) { 1846 info->mpte = p; 1847 return(0); 1848 } 1849 if (!pmap_release_free_page(info->pmap, p)) { 1850 info->error = 1; 1851 return(-1); 1852 } 1853 if (info->object->generation != info->limit) { 1854 info->error = 1; 1855 return(-1); 1856 } 1857 return(0); 1858 } 1859 1860 /* 1861 * Grow the number of kernel page table entries, if needed. 1862 * 1863 * This routine is always called to validate any address space 1864 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address 1865 * space below KERNBASE. 1866 */ 1867 void 1868 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend) 1869 { 1870 vm_paddr_t paddr; 1871 vm_offset_t ptppaddr; 1872 vm_page_t nkpg; 1873 pd_entry_t *pde, newpdir; 1874 pdp_entry_t newpdp; 1875 int update_kernel_vm_end; 1876 1877 crit_enter(); 1878 lwkt_gettoken(&vm_token); 1879 1880 /* 1881 * bootstrap kernel_vm_end on first real VM use 1882 */ 1883 if (kernel_vm_end == 0) { 1884 kernel_vm_end = VM_MIN_KERNEL_ADDRESS; 1885 nkpt = 0; 1886 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) { 1887 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & 1888 ~(PAGE_SIZE * NPTEPG - 1); 1889 nkpt++; 1890 if (kernel_vm_end - 1 >= kernel_map.max_offset) { 1891 kernel_vm_end = kernel_map.max_offset; 1892 break; 1893 } 1894 } 1895 } 1896 1897 /* 1898 * Fill in the gaps. kernel_vm_end is only adjusted for ranges 1899 * below KERNBASE. Ranges above KERNBASE are kldloaded and we 1900 * do not want to force-fill 128G worth of page tables. 1901 */ 1902 if (kstart < KERNBASE) { 1903 if (kstart > kernel_vm_end) 1904 kstart = kernel_vm_end; 1905 KKASSERT(kend <= KERNBASE); 1906 update_kernel_vm_end = 1; 1907 } else { 1908 update_kernel_vm_end = 0; 1909 } 1910 1911 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG); 1912 kend = roundup2(kend, PAGE_SIZE * NPTEPG); 1913 1914 if (kend - 1 >= kernel_map.max_offset) 1915 kend = kernel_map.max_offset; 1916 1917 while (kstart < kend) { 1918 pde = pmap_pde(&kernel_pmap, kstart); 1919 if (pde == NULL) { 1920 /* We need a new PDP entry */ 1921 nkpg = vm_page_alloc(kptobj, nkpt, 1922 VM_ALLOC_NORMAL | 1923 VM_ALLOC_SYSTEM | 1924 VM_ALLOC_INTERRUPT); 1925 if (nkpg == NULL) { 1926 panic("pmap_growkernel: no memory to grow " 1927 "kernel"); 1928 } 1929 paddr = VM_PAGE_TO_PHYS(nkpg); 1930 if ((nkpg->flags & PG_ZERO) == 0) 1931 pmap_zero_page(paddr); 1932 vm_page_flag_clear(nkpg, PG_ZERO); 1933 newpdp = (pdp_entry_t) 1934 (paddr | PG_V | PG_RW | PG_A | PG_M); 1935 *pmap_pdpe(&kernel_pmap, kstart) = newpdp; 1936 nkpt++; 1937 continue; /* try again */ 1938 } 1939 if ((*pde & PG_V) != 0) { 1940 kstart = (kstart + PAGE_SIZE * NPTEPG) & 1941 ~(PAGE_SIZE * NPTEPG - 1); 1942 if (kstart - 1 >= kernel_map.max_offset) { 1943 kstart = kernel_map.max_offset; 1944 break; 1945 } 1946 continue; 1947 } 1948 1949 /* 1950 * This index is bogus, but out of the way 1951 */ 1952 nkpg = vm_page_alloc(kptobj, nkpt, 1953 VM_ALLOC_NORMAL | 1954 VM_ALLOC_SYSTEM | 1955 VM_ALLOC_INTERRUPT); 1956 if (nkpg == NULL) 1957 panic("pmap_growkernel: no memory to grow kernel"); 1958 1959 vm_page_wire(nkpg); 1960 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1961 pmap_zero_page(ptppaddr); 1962 vm_page_flag_clear(nkpg, PG_ZERO); 1963 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1964 *pmap_pde(&kernel_pmap, kstart) = newpdir; 1965 nkpt++; 1966 1967 kstart = (kstart + PAGE_SIZE * NPTEPG) & 1968 ~(PAGE_SIZE * NPTEPG - 1); 1969 1970 if (kstart - 1 >= kernel_map.max_offset) { 1971 kstart = kernel_map.max_offset; 1972 break; 1973 } 1974 } 1975 1976 /* 1977 * Only update kernel_vm_end for areas below KERNBASE. 1978 */ 1979 if (update_kernel_vm_end && kernel_vm_end < kstart) 1980 kernel_vm_end = kstart; 1981 1982 lwkt_reltoken(&vm_token); 1983 crit_exit(); 1984 } 1985 1986 /* 1987 * Retire the given physical map from service. 1988 * Should only be called if the map contains 1989 * no valid mappings. 1990 */ 1991 void 1992 pmap_destroy(pmap_t pmap) 1993 { 1994 int count; 1995 1996 if (pmap == NULL) 1997 return; 1998 1999 lwkt_gettoken(&vm_token); 2000 count = --pmap->pm_count; 2001 if (count == 0) { 2002 pmap_release(pmap); 2003 panic("destroying a pmap is not yet implemented"); 2004 } 2005 lwkt_reltoken(&vm_token); 2006 } 2007 2008 /* 2009 * Add a reference to the specified pmap. 2010 */ 2011 void 2012 pmap_reference(pmap_t pmap) 2013 { 2014 if (pmap != NULL) { 2015 lwkt_gettoken(&vm_token); 2016 pmap->pm_count++; 2017 lwkt_reltoken(&vm_token); 2018 } 2019 } 2020 2021 /*************************************************** 2022 * page management routines. 2023 ***************************************************/ 2024 2025 /* 2026 * free the pv_entry back to the free list. This function may be 2027 * called from an interrupt. 2028 */ 2029 static __inline 2030 void 2031 free_pv_entry(pv_entry_t pv) 2032 { 2033 pv_entry_count--; 2034 KKASSERT(pv_entry_count >= 0); 2035 zfree(pvzone, pv); 2036 } 2037 2038 /* 2039 * get a new pv_entry, allocating a block from the system 2040 * when needed. This function may be called from an interrupt. 2041 */ 2042 static 2043 pv_entry_t 2044 get_pv_entry(void) 2045 { 2046 pv_entry_count++; 2047 if (pv_entry_high_water && 2048 (pv_entry_count > pv_entry_high_water) && 2049 (pmap_pagedaemon_waken == 0)) { 2050 pmap_pagedaemon_waken = 1; 2051 wakeup(&vm_pages_needed); 2052 } 2053 return zalloc(pvzone); 2054 } 2055 2056 /* 2057 * This routine is very drastic, but can save the system 2058 * in a pinch. 2059 */ 2060 void 2061 pmap_collect(void) 2062 { 2063 int i; 2064 vm_page_t m; 2065 static int warningdone=0; 2066 2067 if (pmap_pagedaemon_waken == 0) 2068 return; 2069 lwkt_gettoken(&vm_token); 2070 if (warningdone < 5) { 2071 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n"); 2072 warningdone++; 2073 } 2074 2075 for(i = 0; i < vm_page_array_size; i++) { 2076 m = &vm_page_array[i]; 2077 if (m->wire_count || m->hold_count || m->busy || 2078 (m->flags & PG_BUSY)) 2079 continue; 2080 pmap_remove_all(m); 2081 } 2082 pmap_pagedaemon_waken = 0; 2083 lwkt_reltoken(&vm_token); 2084 } 2085 2086 2087 /* 2088 * If it is the first entry on the list, it is actually 2089 * in the header and we must copy the following entry up 2090 * to the header. Otherwise we must search the list for 2091 * the entry. In either case we free the now unused entry. 2092 */ 2093 static 2094 int 2095 pmap_remove_entry(struct pmap *pmap, vm_page_t m, 2096 vm_offset_t va, pmap_inval_info_t info) 2097 { 2098 pv_entry_t pv; 2099 int rtval; 2100 2101 crit_enter(); 2102 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 2103 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2104 if (pmap == pv->pv_pmap && va == pv->pv_va) 2105 break; 2106 } 2107 } else { 2108 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 2109 if (va == pv->pv_va) 2110 break; 2111 } 2112 } 2113 2114 rtval = 0; 2115 KKASSERT(pv); 2116 2117 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2118 m->md.pv_list_count--; 2119 m->object->agg_pv_list_count--; 2120 KKASSERT(m->md.pv_list_count >= 0); 2121 if (TAILQ_EMPTY(&m->md.pv_list)) 2122 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 2123 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 2124 ++pmap->pm_generation; 2125 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info); 2126 free_pv_entry(pv); 2127 2128 crit_exit(); 2129 return rtval; 2130 } 2131 2132 /* 2133 * Create a pv entry for page at pa for 2134 * (pmap, va). 2135 */ 2136 static 2137 void 2138 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 2139 { 2140 pv_entry_t pv; 2141 2142 crit_enter(); 2143 pv = get_pv_entry(); 2144 pv->pv_va = va; 2145 pv->pv_pmap = pmap; 2146 pv->pv_ptem = mpte; 2147 2148 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 2149 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2150 ++pmap->pm_generation; 2151 m->md.pv_list_count++; 2152 m->object->agg_pv_list_count++; 2153 2154 crit_exit(); 2155 } 2156 2157 /* 2158 * pmap_remove_pte: do the things to unmap a page in a process 2159 */ 2160 static 2161 int 2162 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va, 2163 pmap_inval_info_t info) 2164 { 2165 pt_entry_t oldpte; 2166 vm_page_t m; 2167 2168 pmap_inval_interlock(info, pmap, va); 2169 oldpte = pte_load_clear(ptq); 2170 pmap_inval_deinterlock(info, pmap); 2171 if (oldpte & PG_W) 2172 pmap->pm_stats.wired_count -= 1; 2173 /* 2174 * Machines that don't support invlpg, also don't support 2175 * PG_G. XXX PG_G is disabled for SMP so don't worry about 2176 * the SMP case. 2177 */ 2178 if (oldpte & PG_G) 2179 cpu_invlpg((void *)va); 2180 KKASSERT(pmap->pm_stats.resident_count > 0); 2181 --pmap->pm_stats.resident_count; 2182 if (oldpte & PG_MANAGED) { 2183 m = PHYS_TO_VM_PAGE(oldpte); 2184 if (oldpte & PG_M) { 2185 #if defined(PMAP_DIAGNOSTIC) 2186 if (pmap_nw_modified((pt_entry_t) oldpte)) { 2187 kprintf( 2188 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n", 2189 va, oldpte); 2190 } 2191 #endif 2192 if (pmap_track_modified(va)) 2193 vm_page_dirty(m); 2194 } 2195 if (oldpte & PG_A) 2196 vm_page_flag_set(m, PG_REFERENCED); 2197 return pmap_remove_entry(pmap, m, va, info); 2198 } else { 2199 return pmap_unuse_pt(pmap, va, NULL, info); 2200 } 2201 2202 return 0; 2203 } 2204 2205 /* 2206 * pmap_remove_page: 2207 * 2208 * Remove a single page from a process address space. 2209 * 2210 * This function may not be called from an interrupt if the pmap is 2211 * not kernel_pmap. 2212 */ 2213 static 2214 void 2215 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info) 2216 { 2217 pt_entry_t *pte; 2218 2219 pte = pmap_pte(pmap, va); 2220 if (pte == NULL) 2221 return; 2222 if ((*pte & PG_V) == 0) 2223 return; 2224 pmap_remove_pte(pmap, pte, va, info); 2225 } 2226 2227 /* 2228 * pmap_remove: 2229 * 2230 * Remove the given range of addresses from the specified map. 2231 * 2232 * It is assumed that the start and end are properly 2233 * rounded to the page size. 2234 * 2235 * This function may not be called from an interrupt if the pmap is 2236 * not kernel_pmap. 2237 */ 2238 void 2239 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva) 2240 { 2241 vm_offset_t va_next; 2242 pml4_entry_t *pml4e; 2243 pdp_entry_t *pdpe; 2244 pd_entry_t ptpaddr, *pde; 2245 pt_entry_t *pte; 2246 struct pmap_inval_info info; 2247 2248 if (pmap == NULL) 2249 return; 2250 2251 lwkt_gettoken(&vm_token); 2252 if (pmap->pm_stats.resident_count == 0) { 2253 lwkt_reltoken(&vm_token); 2254 return; 2255 } 2256 2257 pmap_inval_init(&info); 2258 2259 /* 2260 * special handling of removing one page. a very 2261 * common operation and easy to short circuit some 2262 * code. 2263 */ 2264 if (sva + PAGE_SIZE == eva) { 2265 pde = pmap_pde(pmap, sva); 2266 if (pde && (*pde & PG_PS) == 0) { 2267 pmap_remove_page(pmap, sva, &info); 2268 pmap_inval_done(&info); 2269 lwkt_reltoken(&vm_token); 2270 return; 2271 } 2272 } 2273 2274 for (; sva < eva; sva = va_next) { 2275 pml4e = pmap_pml4e(pmap, sva); 2276 if ((*pml4e & PG_V) == 0) { 2277 va_next = (sva + NBPML4) & ~PML4MASK; 2278 if (va_next < sva) 2279 va_next = eva; 2280 continue; 2281 } 2282 2283 pdpe = pmap_pml4e_to_pdpe(pml4e, sva); 2284 if ((*pdpe & PG_V) == 0) { 2285 va_next = (sva + NBPDP) & ~PDPMASK; 2286 if (va_next < sva) 2287 va_next = eva; 2288 continue; 2289 } 2290 2291 /* 2292 * Calculate index for next page table. 2293 */ 2294 va_next = (sva + NBPDR) & ~PDRMASK; 2295 if (va_next < sva) 2296 va_next = eva; 2297 2298 pde = pmap_pdpe_to_pde(pdpe, sva); 2299 ptpaddr = *pde; 2300 2301 /* 2302 * Weed out invalid mappings. 2303 */ 2304 if (ptpaddr == 0) 2305 continue; 2306 2307 /* 2308 * Check for large page. 2309 */ 2310 if ((ptpaddr & PG_PS) != 0) { 2311 /* JG FreeBSD has more complex treatment here */ 2312 pmap_inval_interlock(&info, pmap, -1); 2313 *pde = 0; 2314 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 2315 pmap_inval_deinterlock(&info, pmap); 2316 continue; 2317 } 2318 2319 /* 2320 * Limit our scan to either the end of the va represented 2321 * by the current page table page, or to the end of the 2322 * range being removed. 2323 */ 2324 if (va_next > eva) 2325 va_next = eva; 2326 2327 /* 2328 * NOTE: pmap_remove_pte() can block. 2329 */ 2330 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, 2331 sva += PAGE_SIZE) { 2332 if (*pte == 0) 2333 continue; 2334 if (pmap_remove_pte(pmap, pte, sva, &info)) 2335 break; 2336 } 2337 } 2338 pmap_inval_done(&info); 2339 lwkt_reltoken(&vm_token); 2340 } 2341 2342 /* 2343 * pmap_remove_all: 2344 * 2345 * Removes this physical page from all physical maps in which it resides. 2346 * Reflects back modify bits to the pager. 2347 * 2348 * This routine may not be called from an interrupt. 2349 */ 2350 2351 static 2352 void 2353 pmap_remove_all(vm_page_t m) 2354 { 2355 struct pmap_inval_info info; 2356 pt_entry_t *pte, tpte; 2357 pv_entry_t pv; 2358 2359 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2360 return; 2361 2362 lwkt_gettoken(&vm_token); 2363 pmap_inval_init(&info); 2364 crit_enter(); 2365 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2366 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0); 2367 --pv->pv_pmap->pm_stats.resident_count; 2368 2369 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2370 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va); 2371 tpte = pte_load_clear(pte); 2372 if (tpte & PG_W) 2373 pv->pv_pmap->pm_stats.wired_count--; 2374 pmap_inval_deinterlock(&info, pv->pv_pmap); 2375 if (tpte & PG_A) 2376 vm_page_flag_set(m, PG_REFERENCED); 2377 2378 /* 2379 * Update the vm_page_t clean and reference bits. 2380 */ 2381 if (tpte & PG_M) { 2382 #if defined(PMAP_DIAGNOSTIC) 2383 if (pmap_nw_modified(tpte)) { 2384 kprintf( 2385 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n", 2386 pv->pv_va, tpte); 2387 } 2388 #endif 2389 if (pmap_track_modified(pv->pv_va)) 2390 vm_page_dirty(m); 2391 } 2392 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2393 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2394 ++pv->pv_pmap->pm_generation; 2395 m->md.pv_list_count--; 2396 m->object->agg_pv_list_count--; 2397 KKASSERT(m->md.pv_list_count >= 0); 2398 if (TAILQ_EMPTY(&m->md.pv_list)) 2399 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 2400 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info); 2401 free_pv_entry(pv); 2402 } 2403 crit_exit(); 2404 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0); 2405 pmap_inval_done(&info); 2406 lwkt_reltoken(&vm_token); 2407 } 2408 2409 /* 2410 * pmap_protect: 2411 * 2412 * Set the physical protection on the specified range of this map 2413 * as requested. 2414 * 2415 * This function may not be called from an interrupt if the map is 2416 * not the kernel_pmap. 2417 */ 2418 void 2419 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 2420 { 2421 vm_offset_t va_next; 2422 pml4_entry_t *pml4e; 2423 pdp_entry_t *pdpe; 2424 pd_entry_t ptpaddr, *pde; 2425 pt_entry_t *pte; 2426 pmap_inval_info info; 2427 2428 /* JG review for NX */ 2429 2430 if (pmap == NULL) 2431 return; 2432 2433 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 2434 pmap_remove(pmap, sva, eva); 2435 return; 2436 } 2437 2438 if (prot & VM_PROT_WRITE) 2439 return; 2440 2441 lwkt_gettoken(&vm_token); 2442 pmap_inval_init(&info); 2443 2444 for (; sva < eva; sva = va_next) { 2445 2446 pml4e = pmap_pml4e(pmap, sva); 2447 if ((*pml4e & PG_V) == 0) { 2448 va_next = (sva + NBPML4) & ~PML4MASK; 2449 if (va_next < sva) 2450 va_next = eva; 2451 continue; 2452 } 2453 2454 pdpe = pmap_pml4e_to_pdpe(pml4e, sva); 2455 if ((*pdpe & PG_V) == 0) { 2456 va_next = (sva + NBPDP) & ~PDPMASK; 2457 if (va_next < sva) 2458 va_next = eva; 2459 continue; 2460 } 2461 2462 va_next = (sva + NBPDR) & ~PDRMASK; 2463 if (va_next < sva) 2464 va_next = eva; 2465 2466 pde = pmap_pdpe_to_pde(pdpe, sva); 2467 ptpaddr = *pde; 2468 2469 /* 2470 * Check for large page. 2471 */ 2472 if ((ptpaddr & PG_PS) != 0) { 2473 pmap_inval_interlock(&info, pmap, -1); 2474 *pde &= ~(PG_M|PG_RW); 2475 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 2476 pmap_inval_deinterlock(&info, pmap); 2477 continue; 2478 } 2479 2480 /* 2481 * Weed out invalid mappings. Note: we assume that the page 2482 * directory table is always allocated, and in kernel virtual. 2483 */ 2484 if (ptpaddr == 0) 2485 continue; 2486 2487 if (va_next > eva) 2488 va_next = eva; 2489 2490 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, 2491 sva += PAGE_SIZE) { 2492 pt_entry_t pbits; 2493 pt_entry_t cbits; 2494 vm_page_t m; 2495 2496 /* 2497 * XXX non-optimal. 2498 */ 2499 pmap_inval_interlock(&info, pmap, sva); 2500 again: 2501 pbits = *pte; 2502 cbits = pbits; 2503 if ((pbits & PG_V) == 0) { 2504 pmap_inval_deinterlock(&info, pmap); 2505 continue; 2506 } 2507 if (pbits & PG_MANAGED) { 2508 m = NULL; 2509 if (pbits & PG_A) { 2510 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME); 2511 vm_page_flag_set(m, PG_REFERENCED); 2512 cbits &= ~PG_A; 2513 } 2514 if (pbits & PG_M) { 2515 if (pmap_track_modified(sva)) { 2516 if (m == NULL) 2517 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME); 2518 vm_page_dirty(m); 2519 cbits &= ~PG_M; 2520 } 2521 } 2522 } 2523 cbits &= ~PG_RW; 2524 if (pbits != cbits && 2525 !atomic_cmpset_long(pte, pbits, cbits)) { 2526 goto again; 2527 } 2528 pmap_inval_deinterlock(&info, pmap); 2529 } 2530 } 2531 pmap_inval_done(&info); 2532 lwkt_reltoken(&vm_token); 2533 } 2534 2535 /* 2536 * Insert the given physical page (p) at 2537 * the specified virtual address (v) in the 2538 * target physical map with the protection requested. 2539 * 2540 * If specified, the page will be wired down, meaning 2541 * that the related pte can not be reclaimed. 2542 * 2543 * NB: This is the only routine which MAY NOT lazy-evaluate 2544 * or lose information. That is, this routine must actually 2545 * insert this page into the given map NOW. 2546 */ 2547 void 2548 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 2549 boolean_t wired) 2550 { 2551 vm_paddr_t pa; 2552 pd_entry_t *pde; 2553 pt_entry_t *pte; 2554 vm_paddr_t opa; 2555 pt_entry_t origpte, newpte; 2556 vm_page_t mpte; 2557 pmap_inval_info info; 2558 2559 if (pmap == NULL) 2560 return; 2561 2562 va = trunc_page(va); 2563 #ifdef PMAP_DIAGNOSTIC 2564 if (va >= KvaEnd) 2565 panic("pmap_enter: toobig"); 2566 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 2567 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va); 2568 #endif 2569 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) { 2570 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n"); 2571 #ifdef DDB 2572 db_print_backtrace(); 2573 #endif 2574 } 2575 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) { 2576 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n"); 2577 #ifdef DDB 2578 db_print_backtrace(); 2579 #endif 2580 } 2581 2582 lwkt_gettoken(&vm_token); 2583 2584 /* 2585 * In the case that a page table page is not 2586 * resident, we are creating it here. 2587 */ 2588 if (va < VM_MAX_USER_ADDRESS) 2589 mpte = pmap_allocpte(pmap, va); 2590 else 2591 mpte = NULL; 2592 2593 pmap_inval_init(&info); 2594 pde = pmap_pde(pmap, va); 2595 if (pde != NULL && (*pde & PG_V) != 0) { 2596 if ((*pde & PG_PS) != 0) 2597 panic("pmap_enter: attempted pmap_enter on 2MB page"); 2598 pte = pmap_pde_to_pte(pde, va); 2599 } else 2600 panic("pmap_enter: invalid page directory va=%#lx", va); 2601 2602 KKASSERT(pte != NULL); 2603 pa = VM_PAGE_TO_PHYS(m); 2604 origpte = *pte; 2605 opa = origpte & PG_FRAME; 2606 2607 /* 2608 * Mapping has not changed, must be protection or wiring change. 2609 */ 2610 if (origpte && (opa == pa)) { 2611 /* 2612 * Wiring change, just update stats. We don't worry about 2613 * wiring PT pages as they remain resident as long as there 2614 * are valid mappings in them. Hence, if a user page is wired, 2615 * the PT page will be also. 2616 */ 2617 if (wired && ((origpte & PG_W) == 0)) 2618 pmap->pm_stats.wired_count++; 2619 else if (!wired && (origpte & PG_W)) 2620 pmap->pm_stats.wired_count--; 2621 2622 #if defined(PMAP_DIAGNOSTIC) 2623 if (pmap_nw_modified(origpte)) { 2624 kprintf( 2625 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n", 2626 va, origpte); 2627 } 2628 #endif 2629 2630 /* 2631 * Remove the extra pte reference. Note that we cannot 2632 * optimize the RO->RW case because we have adjusted the 2633 * wiring count above and may need to adjust the wiring 2634 * bits below. 2635 */ 2636 if (mpte) 2637 mpte->hold_count--; 2638 2639 /* 2640 * We might be turning off write access to the page, 2641 * so we go ahead and sense modify status. 2642 */ 2643 if (origpte & PG_MANAGED) { 2644 if ((origpte & PG_M) && pmap_track_modified(va)) { 2645 vm_page_t om; 2646 om = PHYS_TO_VM_PAGE(opa); 2647 vm_page_dirty(om); 2648 } 2649 pa |= PG_MANAGED; 2650 KKASSERT(m->flags & PG_MAPPED); 2651 } 2652 goto validate; 2653 } 2654 /* 2655 * Mapping has changed, invalidate old range and fall through to 2656 * handle validating new mapping. 2657 */ 2658 while (opa) { 2659 int err; 2660 err = pmap_remove_pte(pmap, pte, va, &info); 2661 if (err) 2662 panic("pmap_enter: pte vanished, va: 0x%lx", va); 2663 origpte = *pte; 2664 opa = origpte & PG_FRAME; 2665 if (opa) { 2666 kprintf("pmap_enter: Warning, raced pmap %p va %p\n", 2667 pmap, (void *)va); 2668 } 2669 } 2670 2671 /* 2672 * Enter on the PV list if part of our managed memory. Note that we 2673 * raise IPL while manipulating pv_table since pmap_enter can be 2674 * called at interrupt time. 2675 */ 2676 if (pmap_initialized && 2677 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2678 pmap_insert_entry(pmap, va, mpte, m); 2679 pa |= PG_MANAGED; 2680 vm_page_flag_set(m, PG_MAPPED); 2681 } 2682 2683 /* 2684 * Increment counters 2685 */ 2686 ++pmap->pm_stats.resident_count; 2687 if (wired) 2688 pmap->pm_stats.wired_count++; 2689 2690 validate: 2691 /* 2692 * Now validate mapping with desired protection/wiring. 2693 */ 2694 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V); 2695 2696 if (wired) 2697 newpte |= PG_W; 2698 if (va < VM_MAX_USER_ADDRESS) 2699 newpte |= PG_U; 2700 if (pmap == &kernel_pmap) 2701 newpte |= pgeflag; 2702 2703 /* 2704 * if the mapping or permission bits are different, we need 2705 * to update the pte. 2706 */ 2707 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2708 pmap_inval_interlock(&info, pmap, va); 2709 *pte = newpte | PG_A; 2710 pmap_inval_deinterlock(&info, pmap); 2711 if (newpte & PG_RW) 2712 vm_page_flag_set(m, PG_WRITEABLE); 2713 } 2714 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED)); 2715 pmap_inval_done(&info); 2716 lwkt_reltoken(&vm_token); 2717 } 2718 2719 /* 2720 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired. 2721 * This code also assumes that the pmap has no pre-existing entry for this 2722 * VA. 2723 * 2724 * This code currently may only be used on user pmaps, not kernel_pmap. 2725 */ 2726 void 2727 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m) 2728 { 2729 pt_entry_t *pte; 2730 vm_paddr_t pa; 2731 vm_page_t mpte; 2732 vm_pindex_t ptepindex; 2733 pd_entry_t *ptepa; 2734 pmap_inval_info info; 2735 2736 lwkt_gettoken(&vm_token); 2737 pmap_inval_init(&info); 2738 2739 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) { 2740 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n"); 2741 #ifdef DDB 2742 db_print_backtrace(); 2743 #endif 2744 } 2745 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) { 2746 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n"); 2747 #ifdef DDB 2748 db_print_backtrace(); 2749 #endif 2750 } 2751 2752 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */ 2753 2754 /* 2755 * Calculate the page table page (mpte), allocating it if necessary. 2756 * 2757 * A held page table page (mpte), or NULL, is passed onto the 2758 * section following. 2759 */ 2760 if (va < VM_MAX_USER_ADDRESS) { 2761 /* 2762 * Calculate pagetable page index 2763 */ 2764 ptepindex = pmap_pde_pindex(va); 2765 2766 do { 2767 /* 2768 * Get the page directory entry 2769 */ 2770 ptepa = pmap_pde(pmap, va); 2771 2772 /* 2773 * If the page table page is mapped, we just increment 2774 * the hold count, and activate it. 2775 */ 2776 if (ptepa && (*ptepa & PG_V) != 0) { 2777 if (*ptepa & PG_PS) 2778 panic("pmap_enter_quick: unexpected mapping into 2MB page"); 2779 // if (pmap->pm_ptphint && 2780 // (pmap->pm_ptphint->pindex == ptepindex)) { 2781 // mpte = pmap->pm_ptphint; 2782 // } else { 2783 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex); 2784 pmap->pm_ptphint = mpte; 2785 // } 2786 if (mpte) 2787 mpte->hold_count++; 2788 } else { 2789 mpte = _pmap_allocpte(pmap, ptepindex); 2790 } 2791 } while (mpte == NULL); 2792 } else { 2793 mpte = NULL; 2794 /* this code path is not yet used */ 2795 } 2796 2797 /* 2798 * With a valid (and held) page directory page, we can just use 2799 * vtopte() to get to the pte. If the pte is already present 2800 * we do not disturb it. 2801 */ 2802 pte = vtopte(va); 2803 if (*pte & PG_V) { 2804 if (mpte) 2805 pmap_unwire_pte_hold(pmap, va, mpte, &info); 2806 pa = VM_PAGE_TO_PHYS(m); 2807 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0); 2808 pmap_inval_done(&info); 2809 lwkt_reltoken(&vm_token); 2810 return; 2811 } 2812 2813 /* 2814 * Enter on the PV list if part of our managed memory 2815 */ 2816 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2817 pmap_insert_entry(pmap, va, mpte, m); 2818 vm_page_flag_set(m, PG_MAPPED); 2819 } 2820 2821 /* 2822 * Increment counters 2823 */ 2824 ++pmap->pm_stats.resident_count; 2825 2826 pa = VM_PAGE_TO_PHYS(m); 2827 2828 /* 2829 * Now validate mapping with RO protection 2830 */ 2831 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2832 *pte = pa | PG_V | PG_U; 2833 else 2834 *pte = pa | PG_V | PG_U | PG_MANAGED; 2835 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */ 2836 pmap_inval_done(&info); 2837 lwkt_reltoken(&vm_token); 2838 } 2839 2840 /* 2841 * Make a temporary mapping for a physical address. This is only intended 2842 * to be used for panic dumps. 2843 */ 2844 /* JG Needed on x86_64? */ 2845 void * 2846 pmap_kenter_temporary(vm_paddr_t pa, long i) 2847 { 2848 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa); 2849 return ((void *)crashdumpmap); 2850 } 2851 2852 #define MAX_INIT_PT (96) 2853 2854 /* 2855 * This routine preloads the ptes for a given object into the specified pmap. 2856 * This eliminates the blast of soft faults on process startup and 2857 * immediately after an mmap. 2858 */ 2859 static int pmap_object_init_pt_callback(vm_page_t p, void *data); 2860 2861 void 2862 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot, 2863 vm_object_t object, vm_pindex_t pindex, 2864 vm_size_t size, int limit) 2865 { 2866 struct rb_vm_page_scan_info info; 2867 struct lwp *lp; 2868 vm_size_t psize; 2869 2870 /* 2871 * We can't preinit if read access isn't set or there is no pmap 2872 * or object. 2873 */ 2874 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL) 2875 return; 2876 2877 /* 2878 * We can't preinit if the pmap is not the current pmap 2879 */ 2880 lp = curthread->td_lwp; 2881 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace)) 2882 return; 2883 2884 psize = x86_64_btop(size); 2885 2886 if ((object->type != OBJT_VNODE) || 2887 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) && 2888 (object->resident_page_count > MAX_INIT_PT))) { 2889 return; 2890 } 2891 2892 if (psize + pindex > object->size) { 2893 if (object->size < pindex) 2894 return; 2895 psize = object->size - pindex; 2896 } 2897 2898 if (psize == 0) 2899 return; 2900 2901 /* 2902 * Use a red-black scan to traverse the requested range and load 2903 * any valid pages found into the pmap. 2904 * 2905 * We cannot safely scan the object's memq unless we are in a 2906 * critical section since interrupts can remove pages from objects. 2907 */ 2908 info.start_pindex = pindex; 2909 info.end_pindex = pindex + psize - 1; 2910 info.limit = limit; 2911 info.mpte = NULL; 2912 info.addr = addr; 2913 info.pmap = pmap; 2914 2915 crit_enter(); 2916 lwkt_gettoken(&vm_token); 2917 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 2918 pmap_object_init_pt_callback, &info); 2919 lwkt_reltoken(&vm_token); 2920 crit_exit(); 2921 } 2922 2923 static 2924 int 2925 pmap_object_init_pt_callback(vm_page_t p, void *data) 2926 { 2927 struct rb_vm_page_scan_info *info = data; 2928 vm_pindex_t rel_index; 2929 /* 2930 * don't allow an madvise to blow away our really 2931 * free pages allocating pv entries. 2932 */ 2933 if ((info->limit & MAP_PREFAULT_MADVISE) && 2934 vmstats.v_free_count < vmstats.v_free_reserved) { 2935 return(-1); 2936 } 2937 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2938 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2939 vm_page_busy(p); 2940 if ((p->queue - p->pc) == PQ_CACHE) 2941 vm_page_deactivate(p); 2942 rel_index = p->pindex - info->start_pindex; 2943 pmap_enter_quick(info->pmap, 2944 info->addr + x86_64_ptob(rel_index), p); 2945 vm_page_wakeup(p); 2946 } 2947 return(0); 2948 } 2949 2950 /* 2951 * Return TRUE if the pmap is in shape to trivially 2952 * pre-fault the specified address. 2953 * 2954 * Returns FALSE if it would be non-trivial or if a 2955 * pte is already loaded into the slot. 2956 */ 2957 int 2958 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr) 2959 { 2960 pt_entry_t *pte; 2961 pd_entry_t *pde; 2962 int ret; 2963 2964 lwkt_gettoken(&vm_token); 2965 pde = pmap_pde(pmap, addr); 2966 if (pde == NULL || *pde == 0) { 2967 ret = 0; 2968 } else { 2969 pte = vtopte(addr); 2970 ret = (*pte) ? 0 : 1; 2971 } 2972 lwkt_reltoken(&vm_token); 2973 return(ret); 2974 } 2975 2976 /* 2977 * Routine: pmap_change_wiring 2978 * Function: Change the wiring attribute for a map/virtual-address 2979 * pair. 2980 * In/out conditions: 2981 * The mapping must already exist in the pmap. 2982 */ 2983 void 2984 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired) 2985 { 2986 pt_entry_t *pte; 2987 2988 if (pmap == NULL) 2989 return; 2990 2991 lwkt_gettoken(&vm_token); 2992 pte = pmap_pte(pmap, va); 2993 2994 if (wired && !pmap_pte_w(pte)) 2995 pmap->pm_stats.wired_count++; 2996 else if (!wired && pmap_pte_w(pte)) 2997 pmap->pm_stats.wired_count--; 2998 2999 /* 3000 * Wiring is not a hardware characteristic so there is no need to 3001 * invalidate TLB. However, in an SMP environment we must use 3002 * a locked bus cycle to update the pte (if we are not using 3003 * the pmap_inval_*() API that is)... it's ok to do this for simple 3004 * wiring changes. 3005 */ 3006 #ifdef SMP 3007 if (wired) 3008 atomic_set_long(pte, PG_W); 3009 else 3010 atomic_clear_long(pte, PG_W); 3011 #else 3012 if (wired) 3013 atomic_set_long_nonlocked(pte, PG_W); 3014 else 3015 atomic_clear_long_nonlocked(pte, PG_W); 3016 #endif 3017 lwkt_reltoken(&vm_token); 3018 } 3019 3020 3021 3022 /* 3023 * Copy the range specified by src_addr/len 3024 * from the source map to the range dst_addr/len 3025 * in the destination map. 3026 * 3027 * This routine is only advisory and need not do anything. 3028 */ 3029 void 3030 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, 3031 vm_size_t len, vm_offset_t src_addr) 3032 { 3033 return; 3034 #if 0 3035 pmap_inval_info info; 3036 vm_offset_t addr; 3037 vm_offset_t end_addr = src_addr + len; 3038 vm_offset_t pdnxt; 3039 pd_entry_t src_frame, dst_frame; 3040 vm_page_t m; 3041 3042 if (dst_addr != src_addr) 3043 return; 3044 #if JGPMAP32 3045 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME; 3046 if (src_frame != (PTDpde & PG_FRAME)) { 3047 return; 3048 } 3049 3050 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME; 3051 if (dst_frame != (APTDpde & PG_FRAME)) { 3052 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V); 3053 /* The page directory is not shared between CPUs */ 3054 cpu_invltlb(); 3055 } 3056 #endif 3057 pmap_inval_init(&info); 3058 pmap_inval_add(&info, dst_pmap, -1); 3059 pmap_inval_add(&info, src_pmap, -1); 3060 3061 /* 3062 * critical section protection is required to maintain the page/object 3063 * association, interrupts can free pages and remove them from 3064 * their objects. 3065 */ 3066 crit_enter(); 3067 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 3068 pt_entry_t *src_pte, *dst_pte; 3069 vm_page_t dstmpte, srcmpte; 3070 vm_offset_t srcptepaddr; 3071 vm_pindex_t ptepindex; 3072 3073 if (addr >= UPT_MIN_ADDRESS) 3074 panic("pmap_copy: invalid to pmap_copy page tables\n"); 3075 3076 /* 3077 * Don't let optional prefaulting of pages make us go 3078 * way below the low water mark of free pages or way 3079 * above high water mark of used pv entries. 3080 */ 3081 if (vmstats.v_free_count < vmstats.v_free_reserved || 3082 pv_entry_count > pv_entry_high_water) 3083 break; 3084 3085 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1)); 3086 ptepindex = addr >> PDRSHIFT; 3087 3088 #if JGPMAP32 3089 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex]; 3090 #endif 3091 if (srcptepaddr == 0) 3092 continue; 3093 3094 if (srcptepaddr & PG_PS) { 3095 #if JGPMAP32 3096 if (dst_pmap->pm_pdir[ptepindex] == 0) { 3097 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr; 3098 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; 3099 } 3100 #endif 3101 continue; 3102 } 3103 3104 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex); 3105 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) || 3106 (srcmpte->flags & PG_BUSY)) { 3107 continue; 3108 } 3109 3110 if (pdnxt > end_addr) 3111 pdnxt = end_addr; 3112 3113 src_pte = vtopte(addr); 3114 #if JGPMAP32 3115 dst_pte = avtopte(addr); 3116 #endif 3117 while (addr < pdnxt) { 3118 pt_entry_t ptetemp; 3119 3120 ptetemp = *src_pte; 3121 /* 3122 * we only virtual copy managed pages 3123 */ 3124 if ((ptetemp & PG_MANAGED) != 0) { 3125 /* 3126 * We have to check after allocpte for the 3127 * pte still being around... allocpte can 3128 * block. 3129 * 3130 * pmap_allocpte() can block. If we lose 3131 * our page directory mappings we stop. 3132 */ 3133 dstmpte = pmap_allocpte(dst_pmap, addr); 3134 3135 #if JGPMAP32 3136 if (src_frame != (PTDpde & PG_FRAME) || 3137 dst_frame != (APTDpde & PG_FRAME) 3138 ) { 3139 kprintf("WARNING: pmap_copy: detected and corrected race\n"); 3140 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info); 3141 goto failed; 3142 } else if ((*dst_pte == 0) && 3143 (ptetemp = *src_pte) != 0 && 3144 (ptetemp & PG_MANAGED)) { 3145 /* 3146 * Clear the modified and 3147 * accessed (referenced) bits 3148 * during the copy. 3149 */ 3150 m = PHYS_TO_VM_PAGE(ptetemp); 3151 *dst_pte = ptetemp & ~(PG_M | PG_A); 3152 ++dst_pmap->pm_stats.resident_count; 3153 pmap_insert_entry(dst_pmap, addr, 3154 dstmpte, m); 3155 KKASSERT(m->flags & PG_MAPPED); 3156 } else { 3157 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n"); 3158 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info); 3159 goto failed; 3160 } 3161 #endif 3162 if (dstmpte->hold_count >= srcmpte->hold_count) 3163 break; 3164 } 3165 addr += PAGE_SIZE; 3166 src_pte++; 3167 dst_pte++; 3168 } 3169 } 3170 failed: 3171 crit_exit(); 3172 pmap_inval_done(&info); 3173 #endif 3174 } 3175 3176 /* 3177 * pmap_zero_page: 3178 * 3179 * Zero the specified physical page. 3180 * 3181 * This function may be called from an interrupt and no locking is 3182 * required. 3183 */ 3184 void 3185 pmap_zero_page(vm_paddr_t phys) 3186 { 3187 vm_offset_t va = PHYS_TO_DMAP(phys); 3188 3189 pagezero((void *)va); 3190 } 3191 3192 /* 3193 * pmap_page_assertzero: 3194 * 3195 * Assert that a page is empty, panic if it isn't. 3196 */ 3197 void 3198 pmap_page_assertzero(vm_paddr_t phys) 3199 { 3200 vm_offset_t va = PHYS_TO_DMAP(phys); 3201 size_t i; 3202 3203 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) { 3204 if (*(long *)((char *)va + i) != 0) { 3205 panic("pmap_page_assertzero() @ %p not zero!\n", 3206 (void *)(intptr_t)va); 3207 } 3208 } 3209 } 3210 3211 /* 3212 * pmap_zero_page: 3213 * 3214 * Zero part of a physical page by mapping it into memory and clearing 3215 * its contents with bzero. 3216 * 3217 * off and size may not cover an area beyond a single hardware page. 3218 */ 3219 void 3220 pmap_zero_page_area(vm_paddr_t phys, int off, int size) 3221 { 3222 vm_offset_t virt = PHYS_TO_DMAP(phys); 3223 3224 bzero((char *)virt + off, size); 3225 } 3226 3227 /* 3228 * pmap_copy_page: 3229 * 3230 * Copy the physical page from the source PA to the target PA. 3231 * This function may be called from an interrupt. No locking 3232 * is required. 3233 */ 3234 void 3235 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst) 3236 { 3237 vm_offset_t src_virt, dst_virt; 3238 3239 src_virt = PHYS_TO_DMAP(src); 3240 dst_virt = PHYS_TO_DMAP(dst); 3241 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE); 3242 } 3243 3244 /* 3245 * pmap_copy_page_frag: 3246 * 3247 * Copy the physical page from the source PA to the target PA. 3248 * This function may be called from an interrupt. No locking 3249 * is required. 3250 */ 3251 void 3252 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes) 3253 { 3254 vm_offset_t src_virt, dst_virt; 3255 3256 src_virt = PHYS_TO_DMAP(src); 3257 dst_virt = PHYS_TO_DMAP(dst); 3258 3259 bcopy((char *)src_virt + (src & PAGE_MASK), 3260 (char *)dst_virt + (dst & PAGE_MASK), 3261 bytes); 3262 } 3263 3264 /* 3265 * Returns true if the pmap's pv is one of the first 3266 * 16 pvs linked to from this page. This count may 3267 * be changed upwards or downwards in the future; it 3268 * is only necessary that true be returned for a small 3269 * subset of pmaps for proper page aging. 3270 */ 3271 boolean_t 3272 pmap_page_exists_quick(pmap_t pmap, vm_page_t m) 3273 { 3274 pv_entry_t pv; 3275 int loops = 0; 3276 3277 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3278 return FALSE; 3279 3280 crit_enter(); 3281 lwkt_gettoken(&vm_token); 3282 3283 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3284 if (pv->pv_pmap == pmap) { 3285 lwkt_reltoken(&vm_token); 3286 crit_exit(); 3287 return TRUE; 3288 } 3289 loops++; 3290 if (loops >= 16) 3291 break; 3292 } 3293 lwkt_reltoken(&vm_token); 3294 crit_exit(); 3295 return (FALSE); 3296 } 3297 3298 /* 3299 * Remove all pages from specified address space 3300 * this aids process exit speeds. Also, this code 3301 * is special cased for current process only, but 3302 * can have the more generic (and slightly slower) 3303 * mode enabled. This is much faster than pmap_remove 3304 * in the case of running down an entire address space. 3305 */ 3306 void 3307 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 3308 { 3309 struct lwp *lp; 3310 pt_entry_t *pte, tpte; 3311 pv_entry_t pv, npv; 3312 vm_page_t m; 3313 pmap_inval_info info; 3314 int iscurrentpmap; 3315 int save_generation; 3316 3317 lp = curthread->td_lwp; 3318 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace)) 3319 iscurrentpmap = 1; 3320 else 3321 iscurrentpmap = 0; 3322 3323 lwkt_gettoken(&vm_token); 3324 pmap_inval_init(&info); 3325 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 3326 if (pv->pv_va >= eva || pv->pv_va < sva) { 3327 npv = TAILQ_NEXT(pv, pv_plist); 3328 continue; 3329 } 3330 3331 KKASSERT(pmap == pv->pv_pmap); 3332 3333 if (iscurrentpmap) 3334 pte = vtopte(pv->pv_va); 3335 else 3336 pte = pmap_pte_quick(pmap, pv->pv_va); 3337 pmap_inval_interlock(&info, pmap, pv->pv_va); 3338 3339 /* 3340 * We cannot remove wired pages from a process' mapping 3341 * at this time 3342 */ 3343 if (*pte & PG_W) { 3344 pmap_inval_deinterlock(&info, pmap); 3345 npv = TAILQ_NEXT(pv, pv_plist); 3346 continue; 3347 } 3348 tpte = pte_load_clear(pte); 3349 3350 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME); 3351 3352 KASSERT(m < &vm_page_array[vm_page_array_size], 3353 ("pmap_remove_pages: bad tpte %lx", tpte)); 3354 3355 KKASSERT(pmap->pm_stats.resident_count > 0); 3356 --pmap->pm_stats.resident_count; 3357 pmap_inval_deinterlock(&info, pmap); 3358 3359 /* 3360 * Update the vm_page_t clean and reference bits. 3361 */ 3362 if (tpte & PG_M) { 3363 vm_page_dirty(m); 3364 } 3365 3366 npv = TAILQ_NEXT(pv, pv_plist); 3367 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 3368 save_generation = ++pmap->pm_generation; 3369 3370 m->md.pv_list_count--; 3371 m->object->agg_pv_list_count--; 3372 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 3373 if (TAILQ_EMPTY(&m->md.pv_list)) 3374 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 3375 3376 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info); 3377 free_pv_entry(pv); 3378 3379 /* 3380 * Restart the scan if we blocked during the unuse or free 3381 * calls and other removals were made. 3382 */ 3383 if (save_generation != pmap->pm_generation) { 3384 kprintf("Warning: pmap_remove_pages race-A avoided\n"); 3385 npv = TAILQ_FIRST(&pmap->pm_pvlist); 3386 } 3387 } 3388 pmap_inval_done(&info); 3389 lwkt_reltoken(&vm_token); 3390 } 3391 3392 /* 3393 * pmap_testbit tests bits in pte's 3394 * note that the testbit/clearbit routines are inline, 3395 * and a lot of things compile-time evaluate. 3396 */ 3397 static 3398 boolean_t 3399 pmap_testbit(vm_page_t m, int bit) 3400 { 3401 pv_entry_t pv; 3402 pt_entry_t *pte; 3403 3404 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3405 return FALSE; 3406 3407 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 3408 return FALSE; 3409 3410 crit_enter(); 3411 3412 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3413 /* 3414 * if the bit being tested is the modified bit, then 3415 * mark clean_map and ptes as never 3416 * modified. 3417 */ 3418 if (bit & (PG_A|PG_M)) { 3419 if (!pmap_track_modified(pv->pv_va)) 3420 continue; 3421 } 3422 3423 #if defined(PMAP_DIAGNOSTIC) 3424 if (pv->pv_pmap == NULL) { 3425 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va); 3426 continue; 3427 } 3428 #endif 3429 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3430 if (*pte & bit) { 3431 crit_exit(); 3432 return TRUE; 3433 } 3434 } 3435 crit_exit(); 3436 return (FALSE); 3437 } 3438 3439 /* 3440 * this routine is used to modify bits in ptes 3441 */ 3442 static __inline 3443 void 3444 pmap_clearbit(vm_page_t m, int bit) 3445 { 3446 struct pmap_inval_info info; 3447 pv_entry_t pv; 3448 pt_entry_t *pte; 3449 pt_entry_t pbits; 3450 3451 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3452 return; 3453 3454 pmap_inval_init(&info); 3455 3456 /* 3457 * Loop over all current mappings setting/clearing as appropos If 3458 * setting RO do we need to clear the VAC? 3459 */ 3460 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3461 /* 3462 * don't write protect pager mappings 3463 */ 3464 if (bit == PG_RW) { 3465 if (!pmap_track_modified(pv->pv_va)) 3466 continue; 3467 } 3468 3469 #if defined(PMAP_DIAGNOSTIC) 3470 if (pv->pv_pmap == NULL) { 3471 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va); 3472 continue; 3473 } 3474 #endif 3475 3476 /* 3477 * Careful here. We can use a locked bus instruction to 3478 * clear PG_A or PG_M safely but we need to synchronize 3479 * with the target cpus when we mess with PG_RW. 3480 * 3481 * We do not have to force synchronization when clearing 3482 * PG_M even for PTEs generated via virtual memory maps, 3483 * because the virtual kernel will invalidate the pmap 3484 * entry when/if it needs to resynchronize the Modify bit. 3485 */ 3486 if (bit & PG_RW) 3487 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va); 3488 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3489 again: 3490 pbits = *pte; 3491 if (pbits & bit) { 3492 if (bit == PG_RW) { 3493 if (pbits & PG_M) { 3494 vm_page_dirty(m); 3495 atomic_clear_long(pte, PG_M|PG_RW); 3496 } else { 3497 /* 3498 * The cpu may be trying to set PG_M 3499 * simultaniously with our clearing 3500 * of PG_RW. 3501 */ 3502 if (!atomic_cmpset_long(pte, pbits, 3503 pbits & ~PG_RW)) 3504 goto again; 3505 } 3506 } else if (bit == PG_M) { 3507 /* 3508 * We could also clear PG_RW here to force 3509 * a fault on write to redetect PG_M for 3510 * virtual kernels, but it isn't necessary 3511 * since virtual kernels invalidate the pte 3512 * when they clear the VPTE_M bit in their 3513 * virtual page tables. 3514 */ 3515 atomic_clear_long(pte, PG_M); 3516 } else { 3517 atomic_clear_long(pte, bit); 3518 } 3519 } 3520 if (bit & PG_RW) 3521 pmap_inval_deinterlock(&info, pv->pv_pmap); 3522 } 3523 pmap_inval_done(&info); 3524 } 3525 3526 /* 3527 * pmap_page_protect: 3528 * 3529 * Lower the permission for all mappings to a given page. 3530 */ 3531 void 3532 pmap_page_protect(vm_page_t m, vm_prot_t prot) 3533 { 3534 /* JG NX support? */ 3535 if ((prot & VM_PROT_WRITE) == 0) { 3536 lwkt_gettoken(&vm_token); 3537 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 3538 pmap_clearbit(m, PG_RW); 3539 vm_page_flag_clear(m, PG_WRITEABLE); 3540 } else { 3541 pmap_remove_all(m); 3542 } 3543 lwkt_reltoken(&vm_token); 3544 } 3545 } 3546 3547 vm_paddr_t 3548 pmap_phys_address(vm_pindex_t ppn) 3549 { 3550 return (x86_64_ptob(ppn)); 3551 } 3552 3553 /* 3554 * pmap_ts_referenced: 3555 * 3556 * Return a count of reference bits for a page, clearing those bits. 3557 * It is not necessary for every reference bit to be cleared, but it 3558 * is necessary that 0 only be returned when there are truly no 3559 * reference bits set. 3560 * 3561 * XXX: The exact number of bits to check and clear is a matter that 3562 * should be tested and standardized at some point in the future for 3563 * optimal aging of shared pages. 3564 */ 3565 int 3566 pmap_ts_referenced(vm_page_t m) 3567 { 3568 pv_entry_t pv, pvf, pvn; 3569 pt_entry_t *pte; 3570 int rtval = 0; 3571 3572 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3573 return (rtval); 3574 3575 crit_enter(); 3576 lwkt_gettoken(&vm_token); 3577 3578 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 3579 3580 pvf = pv; 3581 3582 do { 3583 pvn = TAILQ_NEXT(pv, pv_list); 3584 3585 crit_enter(); 3586 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 3587 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 3588 crit_exit(); 3589 3590 if (!pmap_track_modified(pv->pv_va)) 3591 continue; 3592 3593 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3594 3595 if (pte && (*pte & PG_A)) { 3596 #ifdef SMP 3597 atomic_clear_long(pte, PG_A); 3598 #else 3599 atomic_clear_long_nonlocked(pte, PG_A); 3600 #endif 3601 rtval++; 3602 if (rtval > 4) { 3603 break; 3604 } 3605 } 3606 } while ((pv = pvn) != NULL && pv != pvf); 3607 } 3608 lwkt_reltoken(&vm_token); 3609 crit_exit(); 3610 3611 return (rtval); 3612 } 3613 3614 /* 3615 * pmap_is_modified: 3616 * 3617 * Return whether or not the specified physical page was modified 3618 * in any physical maps. 3619 */ 3620 boolean_t 3621 pmap_is_modified(vm_page_t m) 3622 { 3623 boolean_t res; 3624 3625 lwkt_gettoken(&vm_token); 3626 res = pmap_testbit(m, PG_M); 3627 lwkt_reltoken(&vm_token); 3628 return (res); 3629 } 3630 3631 /* 3632 * Clear the modify bits on the specified physical page. 3633 */ 3634 void 3635 pmap_clear_modify(vm_page_t m) 3636 { 3637 lwkt_gettoken(&vm_token); 3638 pmap_clearbit(m, PG_M); 3639 lwkt_reltoken(&vm_token); 3640 } 3641 3642 /* 3643 * pmap_clear_reference: 3644 * 3645 * Clear the reference bit on the specified physical page. 3646 */ 3647 void 3648 pmap_clear_reference(vm_page_t m) 3649 { 3650 lwkt_gettoken(&vm_token); 3651 pmap_clearbit(m, PG_A); 3652 lwkt_reltoken(&vm_token); 3653 } 3654 3655 /* 3656 * Miscellaneous support routines follow 3657 */ 3658 3659 static 3660 void 3661 i386_protection_init(void) 3662 { 3663 int *kp, prot; 3664 3665 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */ 3666 kp = protection_codes; 3667 for (prot = 0; prot < 8; prot++) { 3668 switch (prot) { 3669 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE: 3670 /* 3671 * Read access is also 0. There isn't any execute bit, 3672 * so just make it readable. 3673 */ 3674 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE: 3675 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE: 3676 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE: 3677 *kp++ = 0; 3678 break; 3679 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE: 3680 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE: 3681 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE: 3682 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE: 3683 *kp++ = PG_RW; 3684 break; 3685 } 3686 } 3687 } 3688 3689 /* 3690 * Map a set of physical memory pages into the kernel virtual 3691 * address space. Return a pointer to where it is mapped. This 3692 * routine is intended to be used for mapping device memory, 3693 * NOT real memory. 3694 * 3695 * NOTE: we can't use pgeflag unless we invalidate the pages one at 3696 * a time. 3697 */ 3698 void * 3699 pmap_mapdev(vm_paddr_t pa, vm_size_t size) 3700 { 3701 vm_offset_t va, tmpva, offset; 3702 pt_entry_t *pte; 3703 3704 offset = pa & PAGE_MASK; 3705 size = roundup(offset + size, PAGE_SIZE); 3706 3707 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE); 3708 if (va == 0) 3709 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 3710 3711 pa = pa & ~PAGE_MASK; 3712 for (tmpva = va; size > 0;) { 3713 pte = vtopte(tmpva); 3714 *pte = pa | PG_RW | PG_V; /* | pgeflag; */ 3715 size -= PAGE_SIZE; 3716 tmpva += PAGE_SIZE; 3717 pa += PAGE_SIZE; 3718 } 3719 cpu_invltlb(); 3720 smp_invltlb(); 3721 3722 return ((void *)(va + offset)); 3723 } 3724 3725 void * 3726 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size) 3727 { 3728 vm_offset_t va, tmpva, offset; 3729 pt_entry_t *pte; 3730 3731 offset = pa & PAGE_MASK; 3732 size = roundup(offset + size, PAGE_SIZE); 3733 3734 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE); 3735 if (va == 0) 3736 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 3737 3738 pa = pa & ~PAGE_MASK; 3739 for (tmpva = va; size > 0;) { 3740 pte = vtopte(tmpva); 3741 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */ 3742 size -= PAGE_SIZE; 3743 tmpva += PAGE_SIZE; 3744 pa += PAGE_SIZE; 3745 } 3746 cpu_invltlb(); 3747 smp_invltlb(); 3748 3749 return ((void *)(va + offset)); 3750 } 3751 3752 void 3753 pmap_unmapdev(vm_offset_t va, vm_size_t size) 3754 { 3755 vm_offset_t base, offset; 3756 3757 base = va & ~PAGE_MASK; 3758 offset = va & PAGE_MASK; 3759 size = roundup(offset + size, PAGE_SIZE); 3760 pmap_qremove(va, size >> PAGE_SHIFT); 3761 kmem_free(&kernel_map, base, size); 3762 } 3763 3764 /* 3765 * perform the pmap work for mincore 3766 */ 3767 int 3768 pmap_mincore(pmap_t pmap, vm_offset_t addr) 3769 { 3770 pt_entry_t *ptep, pte; 3771 vm_page_t m; 3772 int val = 0; 3773 3774 lwkt_gettoken(&vm_token); 3775 ptep = pmap_pte(pmap, addr); 3776 3777 if (ptep && (pte = *ptep) != 0) { 3778 vm_offset_t pa; 3779 3780 val = MINCORE_INCORE; 3781 if ((pte & PG_MANAGED) == 0) 3782 goto done; 3783 3784 pa = pte & PG_FRAME; 3785 3786 m = PHYS_TO_VM_PAGE(pa); 3787 3788 /* 3789 * Modified by us 3790 */ 3791 if (pte & PG_M) 3792 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 3793 /* 3794 * Modified by someone 3795 */ 3796 else if (m->dirty || pmap_is_modified(m)) 3797 val |= MINCORE_MODIFIED_OTHER; 3798 /* 3799 * Referenced by us 3800 */ 3801 if (pte & PG_A) 3802 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 3803 3804 /* 3805 * Referenced by someone 3806 */ 3807 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) { 3808 val |= MINCORE_REFERENCED_OTHER; 3809 vm_page_flag_set(m, PG_REFERENCED); 3810 } 3811 } 3812 done: 3813 lwkt_reltoken(&vm_token); 3814 return val; 3815 } 3816 3817 /* 3818 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new 3819 * vmspace will be ref'd and the old one will be deref'd. 3820 * 3821 * The vmspace for all lwps associated with the process will be adjusted 3822 * and cr3 will be reloaded if any lwp is the current lwp. 3823 */ 3824 void 3825 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs) 3826 { 3827 struct vmspace *oldvm; 3828 struct lwp *lp; 3829 3830 crit_enter(); 3831 oldvm = p->p_vmspace; 3832 if (oldvm != newvm) { 3833 p->p_vmspace = newvm; 3834 KKASSERT(p->p_nthreads == 1); 3835 lp = RB_ROOT(&p->p_lwp_tree); 3836 pmap_setlwpvm(lp, newvm); 3837 if (adjrefs) { 3838 sysref_get(&newvm->vm_sysref); 3839 sysref_put(&oldvm->vm_sysref); 3840 } 3841 } 3842 crit_exit(); 3843 } 3844 3845 /* 3846 * Set the vmspace for a LWP. The vmspace is almost universally set the 3847 * same as the process vmspace, but virtual kernels need to swap out contexts 3848 * on a per-lwp basis. 3849 */ 3850 void 3851 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm) 3852 { 3853 struct vmspace *oldvm; 3854 struct pmap *pmap; 3855 3856 crit_enter(); 3857 oldvm = lp->lwp_vmspace; 3858 3859 if (oldvm != newvm) { 3860 lp->lwp_vmspace = newvm; 3861 if (curthread->td_lwp == lp) { 3862 pmap = vmspace_pmap(newvm); 3863 #if defined(SMP) 3864 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask); 3865 if (pmap->pm_active & CPUMASK_LOCK) 3866 pmap_interlock_wait(newvm); 3867 #else 3868 pmap->pm_active |= 1; 3869 #endif 3870 #if defined(SWTCH_OPTIM_STATS) 3871 tlb_flush_count++; 3872 #endif 3873 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4); 3874 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V; 3875 load_cr3(curthread->td_pcb->pcb_cr3); 3876 pmap = vmspace_pmap(oldvm); 3877 #if defined(SMP) 3878 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask); 3879 #else 3880 pmap->pm_active &= ~(cpumask_t)1; 3881 #endif 3882 } 3883 } 3884 crit_exit(); 3885 } 3886 3887 #ifdef SMP 3888 3889 /* 3890 * Called when switching to a locked pmap 3891 */ 3892 void 3893 pmap_interlock_wait(struct vmspace *vm) 3894 { 3895 struct pmap *pmap = &vm->vm_pmap; 3896 3897 if (pmap->pm_active & CPUMASK_LOCK) { 3898 DEBUG_PUSH_INFO("pmap_interlock_wait"); 3899 while (pmap->pm_active & CPUMASK_LOCK) { 3900 cpu_pause(); 3901 cpu_ccfence(); 3902 lwkt_process_ipiq(); 3903 } 3904 DEBUG_POP_INFO(); 3905 } 3906 } 3907 3908 #endif 3909 3910 vm_offset_t 3911 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3912 { 3913 3914 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3915 return addr; 3916 } 3917 3918 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); 3919 return addr; 3920 } 3921 3922 /* 3923 * Used by kmalloc/kfree, page already exists at va 3924 */ 3925 vm_page_t 3926 pmap_kvtom(vm_offset_t va) 3927 { 3928 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME)); 3929 } 3930