pmap.c revision 160525
1/*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1994 John S. Dyson 5 * All rights reserved. 6 * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * Copyright (c) 2005 Alan L. Cox <alc@cs.rice.edu> 9 * All rights reserved. 10 * 11 * This code is derived from software contributed to Berkeley by 12 * the Systems Programming Group of the University of Utah Computer 13 * Science Department and William Jolitz of UUNET Technologies Inc. 14 * 15 * Redistribution and use in source and binary forms, with or without 16 * modification, are permitted provided that the following conditions 17 * are met: 18 * 1. Redistributions of source code must retain the above copyright 19 * notice, this list of conditions and the following disclaimer. 20 * 2. Redistributions in binary form must reproduce the above copyright 21 * notice, this list of conditions and the following disclaimer in the 22 * documentation and/or other materials provided with the distribution. 23 * 3. All advertising materials mentioning features or use of this software 24 * must display the following acknowledgement: 25 * This product includes software developed by the University of 26 * California, Berkeley and its contributors. 27 * 4. Neither the name of the University nor the names of its contributors 28 * may be used to endorse or promote products derived from this software 29 * without specific prior written permission. 30 * 31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 41 * SUCH DAMAGE. 42 * 43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 44 */ 45/*- 46 * Copyright (c) 2003 Networks Associates Technology, Inc. 47 * All rights reserved. 48 * 49 * This software was developed for the FreeBSD Project by Jake Burkholder, 50 * Safeport Network Services, and Network Associates Laboratories, the 51 * Security Research Division of Network Associates, Inc. under 52 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA 53 * CHATS research program. 54 * 55 * Redistribution and use in source and binary forms, with or without 56 * modification, are permitted provided that the following conditions 57 * are met: 58 * 1. Redistributions of source code must retain the above copyright 59 * notice, this list of conditions and the following disclaimer. 60 * 2. Redistributions in binary form must reproduce the above copyright 61 * notice, this list of conditions and the following disclaimer in the 62 * documentation and/or other materials provided with the distribution. 63 * 64 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 65 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 66 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 67 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 68 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 69 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 70 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 71 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 72 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 73 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 74 * SUCH DAMAGE. 75 */ 76 77#include <sys/cdefs.h> 78__FBSDID("$FreeBSD: head/sys/i386/i386/pmap.c 160525 2006-07-20 17:48:41Z alc $"); 79 80/* 81 * Manages physical address maps. 82 * 83 * In addition to hardware address maps, this 84 * module is called upon to provide software-use-only 85 * maps which may or may not be stored in the same 86 * form as hardware maps. These pseudo-maps are 87 * used to store intermediate results from copy 88 * operations to and from address spaces. 89 * 90 * Since the information managed by this module is 91 * also stored by the logical address mapping module, 92 * this module may throw away valid virtual-to-physical 93 * mappings at almost any time. However, invalidations 94 * of virtual-to-physical mappings must be done as 95 * requested. 96 * 97 * In order to cope with hardware architectures which 98 * make virtual-to-physical map invalidates expensive, 99 * this module may delay invalidate or reduced protection 100 * operations until such time as they are actually 101 * necessary. This module is given full information as 102 * to which processors are currently using which maps, 103 * and to when physical maps must be made correct. 104 */ 105 106#include "opt_cpu.h" 107#include "opt_pmap.h" 108#include "opt_msgbuf.h" 109#include "opt_smp.h" 110#include "opt_xbox.h" 111 112#include <sys/param.h> 113#include <sys/systm.h> 114#include <sys/kernel.h> 115#include <sys/lock.h> 116#include <sys/malloc.h> 117#include <sys/mman.h> 118#include <sys/msgbuf.h> 119#include <sys/mutex.h> 120#include <sys/proc.h> 121#include <sys/sx.h> 122#include <sys/vmmeter.h> 123#include <sys/sched.h> 124#include <sys/sysctl.h> 125#ifdef SMP 126#include <sys/smp.h> 127#endif 128 129#include <vm/vm.h> 130#include <vm/vm_param.h> 131#include <vm/vm_kern.h> 132#include <vm/vm_page.h> 133#include <vm/vm_map.h> 134#include <vm/vm_object.h> 135#include <vm/vm_extern.h> 136#include <vm/vm_pageout.h> 137#include <vm/vm_pager.h> 138#include <vm/uma.h> 139 140#include <machine/cpu.h> 141#include <machine/cputypes.h> 142#include <machine/md_var.h> 143#include <machine/pcb.h> 144#include <machine/specialreg.h> 145#ifdef SMP 146#include <machine/smp.h> 147#endif 148 149#ifdef XBOX 150#include <machine/xbox.h> 151#endif 152 153#if !defined(CPU_DISABLE_SSE) && defined(I686_CPU) 154#define CPU_ENABLE_SSE 155#endif 156 157#ifndef PMAP_SHPGPERPROC 158#define PMAP_SHPGPERPROC 200 159#endif 160 161#if defined(DIAGNOSTIC) 162#define PMAP_DIAGNOSTIC 163#endif 164 165#if !defined(PMAP_DIAGNOSTIC) 166#define PMAP_INLINE __inline 167#else 168#define PMAP_INLINE 169#endif 170 171#define PV_STATS 172#ifdef PV_STATS 173#define PV_STAT(x) do { x ; } while (0) 174#else 175#define PV_STAT(x) do { } while (0) 176#endif 177 178/* 179 * Get PDEs and PTEs for user/kernel address space 180 */ 181#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 182#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 183 184#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 185#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 186#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 187#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 188#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 189 190#define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \ 191 atomic_clear_int((u_int *)(pte), PG_W)) 192#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 193 194struct pmap kernel_pmap_store; 195LIST_HEAD(pmaplist, pmap); 196static struct pmaplist allpmaps; 197static struct mtx allpmaps_lock; 198 199vm_paddr_t avail_end; /* PA of last available physical page */ 200vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 201vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 202int pgeflag = 0; /* PG_G or-in */ 203int pseflag = 0; /* PG_PS or-in */ 204 205static int nkpt; 206vm_offset_t kernel_vm_end; 207extern u_int32_t KERNend; 208 209#ifdef PAE 210static uma_zone_t pdptzone; 211#endif 212 213/* 214 * Data for the pv entry allocation mechanism 215 */ 216static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 217static int shpgperproc = PMAP_SHPGPERPROC; 218 219struct pv_chunk *pv_chunkbase; /* KVA block for pv_chunks */ 220int pv_maxchunks; /* How many chunks we have KVA for */ 221vm_offset_t pv_vafree; /* freelist stored in the PTE */ 222 223/* 224 * All those kernel PT submaps that BSD is so fond of 225 */ 226struct sysmaps { 227 struct mtx lock; 228 pt_entry_t *CMAP1; 229 pt_entry_t *CMAP2; 230 caddr_t CADDR1; 231 caddr_t CADDR2; 232}; 233static struct sysmaps sysmaps_pcpu[MAXCPU]; 234pt_entry_t *CMAP1 = 0; 235static pt_entry_t *CMAP3; 236caddr_t CADDR1 = 0, ptvmmap = 0; 237static caddr_t CADDR3; 238struct msgbuf *msgbufp = 0; 239 240/* 241 * Crashdump maps. 242 */ 243static caddr_t crashdumpmap; 244 245#ifdef SMP 246extern pt_entry_t *SMPpt; 247#endif 248static pt_entry_t *PMAP1 = 0, *PMAP2; 249static pt_entry_t *PADDR1 = 0, *PADDR2; 250#ifdef SMP 251static int PMAP1cpu; 252static int PMAP1changedcpu; 253SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD, 254 &PMAP1changedcpu, 0, 255 "Number of times pmap_pte_quick changed CPU with same PMAP1"); 256#endif 257static int PMAP1changed; 258SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD, 259 &PMAP1changed, 0, 260 "Number of times pmap_pte_quick changed PMAP1"); 261static int PMAP1unchanged; 262SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD, 263 &PMAP1unchanged, 0, 264 "Number of times pmap_pte_quick didn't change PMAP1"); 265static struct mtx PMAP2mutex; 266 267static void free_pv_entry(pmap_t pmap, pv_entry_t pv); 268static pv_entry_t get_pv_entry(pmap_t locked_pmap, int try); 269 270static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, 271 vm_page_t m, vm_prot_t prot, vm_page_t mpte); 272static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 273static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 274static void pmap_remove_entry(struct pmap *pmap, vm_page_t m, 275 vm_offset_t va); 276static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); 277static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, 278 vm_page_t m); 279 280static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags); 281 282static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags); 283static int _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m); 284static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 285static void pmap_pte_release(pt_entry_t *pte); 286static int pmap_unuse_pt(pmap_t, vm_offset_t); 287static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 288#ifdef PAE 289static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 290#endif 291 292CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 293CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 294 295/* 296 * Move the kernel virtual free pointer to the next 297 * 4MB. This is used to help improve performance 298 * by using a large (4MB) page for much of the kernel 299 * (.text, .data, .bss) 300 */ 301static vm_offset_t 302pmap_kmem_choose(vm_offset_t addr) 303{ 304 vm_offset_t newaddr = addr; 305 306#ifndef DISABLE_PSE 307 if (cpu_feature & CPUID_PSE) 308 newaddr = (addr + PDRMASK) & ~PDRMASK; 309#endif 310 return newaddr; 311} 312 313/* 314 * Bootstrap the system enough to run with virtual memory. 315 * 316 * On the i386 this is called after mapping has already been enabled 317 * and just syncs the pmap module with what has already been done. 318 * [We can't call it easily with mapping off since the kernel is not 319 * mapped with PA == VA, hence we would have to relocate every address 320 * from the linked base (virtual) address "KERNBASE" to the actual 321 * (physical) address starting relative to 0] 322 */ 323void 324pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr) 325{ 326 vm_offset_t va; 327 pt_entry_t *pte, *unused; 328 struct sysmaps *sysmaps; 329 int i; 330 331 /* 332 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 333 * large. It should instead be correctly calculated in locore.s and 334 * not based on 'first' (which is a physical address, not a virtual 335 * address, for the start of unused physical memory). The kernel 336 * page tables are NOT double mapped and thus should not be included 337 * in this calculation. 338 */ 339 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 340 virtual_avail = pmap_kmem_choose(virtual_avail); 341 342 virtual_end = VM_MAX_KERNEL_ADDRESS; 343 344 /* 345 * Initialize the kernel pmap (which is statically allocated). 346 */ 347 PMAP_LOCK_INIT(kernel_pmap); 348 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 349#ifdef PAE 350 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 351#endif 352 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 353 TAILQ_INIT(&kernel_pmap->pm_pvchunk); 354 LIST_INIT(&allpmaps); 355 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 356 mtx_lock_spin(&allpmaps_lock); 357 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 358 mtx_unlock_spin(&allpmaps_lock); 359 nkpt = NKPT; 360 361 /* 362 * Reserve some special page table entries/VA space for temporary 363 * mapping of pages. 364 */ 365#define SYSMAP(c, p, v, n) \ 366 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 367 368 va = virtual_avail; 369 pte = vtopte(va); 370 371 /* 372 * CMAP1/CMAP2 are used for zeroing and copying pages. 373 * CMAP3 is used for the idle process page zeroing. 374 */ 375 for (i = 0; i < MAXCPU; i++) { 376 sysmaps = &sysmaps_pcpu[i]; 377 mtx_init(&sysmaps->lock, "SYSMAPS", NULL, MTX_DEF); 378 SYSMAP(caddr_t, sysmaps->CMAP1, sysmaps->CADDR1, 1) 379 SYSMAP(caddr_t, sysmaps->CMAP2, sysmaps->CADDR2, 1) 380 } 381 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 382 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 383 *CMAP3 = 0; 384 385 /* 386 * Crashdump maps. 387 */ 388 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS) 389 390 /* 391 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 392 */ 393 SYSMAP(caddr_t, unused, ptvmmap, 1) 394 395 /* 396 * msgbufp is used to map the system message buffer. 397 */ 398 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE))) 399 400 /* 401 * ptemap is used for pmap_pte_quick 402 */ 403 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 404 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1); 405 406 mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF); 407 408 virtual_avail = va; 409 410 *CMAP1 = 0; 411 412#ifdef XBOX 413 /* FIXME: This is gross, but needed for the XBOX. Since we are in such 414 * an early stadium, we cannot yet neatly map video memory ... :-( 415 * Better fixes are very welcome! */ 416 if (!arch_i386_is_xbox) 417#endif 418 for (i = 0; i < NKPT; i++) 419 PTD[i] = 0; 420 421 /* Initialize the PAT MSR if present. */ 422 pmap_init_pat(); 423 424 /* Turn on PG_G on kernel page(s) */ 425 pmap_set_pg(); 426} 427 428/* 429 * Setup the PAT MSR. 430 */ 431void 432pmap_init_pat(void) 433{ 434 uint64_t pat_msr; 435 436 /* Bail if this CPU doesn't implement PAT. */ 437 if (!(cpu_feature & CPUID_PAT)) 438 return; 439 440#ifdef PAT_WORKS 441 /* 442 * Leave the indices 0-3 at the default of WB, WT, UC, and UC-. 443 * Program 4 and 5 as WP and WC. 444 * Leave 6 and 7 as UC and UC-. 445 */ 446 pat_msr = rdmsr(MSR_PAT); 447 pat_msr &= ~(PAT_MASK(4) | PAT_MASK(5)); 448 pat_msr |= PAT_VALUE(4, PAT_WRITE_PROTECTED) | 449 PAT_VALUE(5, PAT_WRITE_COMBINING); 450#else 451 /* 452 * Due to some Intel errata, we can only safely use the lower 4 453 * PAT entries. Thus, just replace PAT Index 2 with WC instead 454 * of UC-. 455 * 456 * Intel Pentium III Processor Specification Update 457 * Errata E.27 (Upper Four PAT Entries Not Usable With Mode B 458 * or Mode C Paging) 459 * 460 * Intel Pentium IV Processor Specification Update 461 * Errata N46 (PAT Index MSB May Be Calculated Incorrectly) 462 */ 463 pat_msr = rdmsr(MSR_PAT); 464 pat_msr &= ~PAT_MASK(2); 465 pat_msr |= PAT_VALUE(2, PAT_WRITE_COMBINING); 466#endif 467 wrmsr(MSR_PAT, pat_msr); 468} 469 470/* 471 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 472 */ 473void 474pmap_set_pg(void) 475{ 476 pd_entry_t pdir; 477 pt_entry_t *pte; 478 vm_offset_t va, endva; 479 int i; 480 481 if (pgeflag == 0) 482 return; 483 484 i = KERNLOAD/NBPDR; 485 endva = KERNBASE + KERNend; 486 487 if (pseflag) { 488 va = KERNBASE + KERNLOAD; 489 while (va < endva) { 490 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 491 pdir |= pgeflag; 492 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 493 invltlb(); /* Play it safe, invltlb() every time */ 494 i++; 495 va += NBPDR; 496 } 497 } else { 498 va = (vm_offset_t)btext; 499 while (va < endva) { 500 pte = vtopte(va); 501 if (*pte) 502 *pte |= pgeflag; 503 invltlb(); /* Play it safe, invltlb() every time */ 504 va += PAGE_SIZE; 505 } 506 } 507} 508 509/* 510 * Initialize a vm_page's machine-dependent fields. 511 */ 512void 513pmap_page_init(vm_page_t m) 514{ 515 516 TAILQ_INIT(&m->md.pv_list); 517 m->md.pv_list_count = 0; 518} 519 520#ifdef PAE 521 522static MALLOC_DEFINE(M_PMAPPDPT, "pmap", "pmap pdpt"); 523 524static void * 525pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 526{ 527 *flags = UMA_SLAB_PRIV; 528 return (contigmalloc(PAGE_SIZE, M_PMAPPDPT, 0, 0x0ULL, 0xffffffffULL, 529 1, 0)); 530} 531#endif 532 533/* 534 * ABuse the pte nodes for unmapped kva to thread a kva freelist through. 535 * Requirements: 536 * - Must deal with pages in order to ensure that none of the PG_* bits 537 * are ever set, PG_V in particular. 538 * - Assumes we can write to ptes without pte_store() atomic ops, even 539 * on PAE systems. This should be ok. 540 * - Assumes nothing will ever test these addresses for 0 to indicate 541 * no mapping instead of correctly checking PG_V. 542 * - Assumes a vm_offset_t will fit in a pte (true for i386). 543 * Because PG_V is never set, there can be no mappings to invalidate. 544 */ 545static vm_offset_t 546pmap_ptelist_alloc(vm_offset_t *head) 547{ 548 pt_entry_t *pte; 549 vm_offset_t va; 550 551 va = *head; 552 if (va == 0) 553 return (va); /* Out of memory */ 554 pte = vtopte(va); 555 *head = *pte; 556 if (*head & PG_V) 557 panic("pmap_ptelist_alloc: va with PG_V set!"); 558 *pte = 0; 559 return (va); 560} 561 562static void 563pmap_ptelist_free(vm_offset_t *head, vm_offset_t va) 564{ 565 pt_entry_t *pte; 566 567 if (va & PG_V) 568 panic("pmap_ptelist_free: freeing va with PG_V set!"); 569 pte = vtopte(va); 570 *pte = *head; /* virtual! PG_V is 0 though */ 571 *head = va; 572} 573 574static void 575pmap_ptelist_init(vm_offset_t *head, void *base, int npages) 576{ 577 int i; 578 vm_offset_t va; 579 580 *head = 0; 581 for (i = npages - 1; i >= 0; i--) { 582 va = (vm_offset_t)base + i * PAGE_SIZE; 583 pmap_ptelist_free(head, va); 584 } 585} 586 587 588/* 589 * Initialize the pmap module. 590 * Called by vm_init, to initialize any structures that the pmap 591 * system needs to map virtual memory. 592 */ 593void 594pmap_init(void) 595{ 596 597 /* 598 * Initialize the address space (zone) for the pv entries. Set a 599 * high water mark so that the system can recover from excessive 600 * numbers of pv entries. 601 */ 602 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 603 pv_entry_max = shpgperproc * maxproc + cnt.v_page_count; 604 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 605 pv_entry_max = roundup(pv_entry_max, _NPCPV); 606 pv_entry_high_water = 9 * (pv_entry_max / 10); 607 608 pv_maxchunks = pv_entry_max / _NPCPV; 609 pv_chunkbase = (struct pv_chunk *)kmem_alloc_nofault(kernel_map, 610 PAGE_SIZE * pv_maxchunks); 611 if (pv_chunkbase == NULL) 612 panic("pmap_init: not enough kvm for pv chunks"); 613 pmap_ptelist_init(&pv_vafree, pv_chunkbase, pv_maxchunks); 614#ifdef PAE 615 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 616 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 617 UMA_ZONE_VM | UMA_ZONE_NOFREE); 618 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 619#endif 620} 621 622 623SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters"); 624SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_max, CTLFLAG_RD, &pv_entry_max, 0, 625 "Max number of PV entries"); 626SYSCTL_INT(_vm_pmap, OID_AUTO, shpgperproc, CTLFLAG_RD, &shpgperproc, 0, 627 "Page share factor per proc"); 628 629/*************************************************** 630 * Low level helper routines..... 631 ***************************************************/ 632 633#ifdef SMP 634/* 635 * For SMP, these functions have to use the IPI mechanism for coherence. 636 */ 637void 638pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 639{ 640 u_int cpumask; 641 u_int other_cpus; 642 643 if (smp_started) { 644 if (!(read_eflags() & PSL_I)) 645 panic("%s: interrupts disabled", __func__); 646 mtx_lock_spin(&smp_ipi_mtx); 647 } else 648 critical_enter(); 649 /* 650 * We need to disable interrupt preemption but MUST NOT have 651 * interrupts disabled here. 652 * XXX we may need to hold schedlock to get a coherent pm_active 653 * XXX critical sections disable interrupts again 654 */ 655 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 656 invlpg(va); 657 smp_invlpg(va); 658 } else { 659 cpumask = PCPU_GET(cpumask); 660 other_cpus = PCPU_GET(other_cpus); 661 if (pmap->pm_active & cpumask) 662 invlpg(va); 663 if (pmap->pm_active & other_cpus) 664 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 665 } 666 if (smp_started) 667 mtx_unlock_spin(&smp_ipi_mtx); 668 else 669 critical_exit(); 670} 671 672void 673pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 674{ 675 u_int cpumask; 676 u_int other_cpus; 677 vm_offset_t addr; 678 679 if (smp_started) { 680 if (!(read_eflags() & PSL_I)) 681 panic("%s: interrupts disabled", __func__); 682 mtx_lock_spin(&smp_ipi_mtx); 683 } else 684 critical_enter(); 685 /* 686 * We need to disable interrupt preemption but MUST NOT have 687 * interrupts disabled here. 688 * XXX we may need to hold schedlock to get a coherent pm_active 689 * XXX critical sections disable interrupts again 690 */ 691 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 692 for (addr = sva; addr < eva; addr += PAGE_SIZE) 693 invlpg(addr); 694 smp_invlpg_range(sva, eva); 695 } else { 696 cpumask = PCPU_GET(cpumask); 697 other_cpus = PCPU_GET(other_cpus); 698 if (pmap->pm_active & cpumask) 699 for (addr = sva; addr < eva; addr += PAGE_SIZE) 700 invlpg(addr); 701 if (pmap->pm_active & other_cpus) 702 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 703 sva, eva); 704 } 705 if (smp_started) 706 mtx_unlock_spin(&smp_ipi_mtx); 707 else 708 critical_exit(); 709} 710 711void 712pmap_invalidate_all(pmap_t pmap) 713{ 714 u_int cpumask; 715 u_int other_cpus; 716 717 if (smp_started) { 718 if (!(read_eflags() & PSL_I)) 719 panic("%s: interrupts disabled", __func__); 720 mtx_lock_spin(&smp_ipi_mtx); 721 } else 722 critical_enter(); 723 /* 724 * We need to disable interrupt preemption but MUST NOT have 725 * interrupts disabled here. 726 * XXX we may need to hold schedlock to get a coherent pm_active 727 * XXX critical sections disable interrupts again 728 */ 729 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 730 invltlb(); 731 smp_invltlb(); 732 } else { 733 cpumask = PCPU_GET(cpumask); 734 other_cpus = PCPU_GET(other_cpus); 735 if (pmap->pm_active & cpumask) 736 invltlb(); 737 if (pmap->pm_active & other_cpus) 738 smp_masked_invltlb(pmap->pm_active & other_cpus); 739 } 740 if (smp_started) 741 mtx_unlock_spin(&smp_ipi_mtx); 742 else 743 critical_exit(); 744} 745 746void 747pmap_invalidate_cache(void) 748{ 749 750 if (smp_started) { 751 if (!(read_eflags() & PSL_I)) 752 panic("%s: interrupts disabled", __func__); 753 mtx_lock_spin(&smp_ipi_mtx); 754 } else 755 critical_enter(); 756 /* 757 * We need to disable interrupt preemption but MUST NOT have 758 * interrupts disabled here. 759 * XXX we may need to hold schedlock to get a coherent pm_active 760 * XXX critical sections disable interrupts again 761 */ 762 wbinvd(); 763 smp_cache_flush(); 764 if (smp_started) 765 mtx_unlock_spin(&smp_ipi_mtx); 766 else 767 critical_exit(); 768} 769#else /* !SMP */ 770/* 771 * Normal, non-SMP, 486+ invalidation functions. 772 * We inline these within pmap.c for speed. 773 */ 774PMAP_INLINE void 775pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 776{ 777 778 if (pmap == kernel_pmap || pmap->pm_active) 779 invlpg(va); 780} 781 782PMAP_INLINE void 783pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 784{ 785 vm_offset_t addr; 786 787 if (pmap == kernel_pmap || pmap->pm_active) 788 for (addr = sva; addr < eva; addr += PAGE_SIZE) 789 invlpg(addr); 790} 791 792PMAP_INLINE void 793pmap_invalidate_all(pmap_t pmap) 794{ 795 796 if (pmap == kernel_pmap || pmap->pm_active) 797 invltlb(); 798} 799 800PMAP_INLINE void 801pmap_invalidate_cache(void) 802{ 803 804 wbinvd(); 805} 806#endif /* !SMP */ 807 808/* 809 * Are we current address space or kernel? N.B. We return FALSE when 810 * a pmap's page table is in use because a kernel thread is borrowing 811 * it. The borrowed page table can change spontaneously, making any 812 * dependence on its continued use subject to a race condition. 813 */ 814static __inline int 815pmap_is_current(pmap_t pmap) 816{ 817 818 return (pmap == kernel_pmap || 819 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 820 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 821} 822 823/* 824 * If the given pmap is not the current or kernel pmap, the returned pte must 825 * be released by passing it to pmap_pte_release(). 826 */ 827pt_entry_t * 828pmap_pte(pmap_t pmap, vm_offset_t va) 829{ 830 pd_entry_t newpf; 831 pd_entry_t *pde; 832 833 pde = pmap_pde(pmap, va); 834 if (*pde & PG_PS) 835 return (pde); 836 if (*pde != 0) { 837 /* are we current address space or kernel? */ 838 if (pmap_is_current(pmap)) 839 return (vtopte(va)); 840 mtx_lock(&PMAP2mutex); 841 newpf = *pde & PG_FRAME; 842 if ((*PMAP2 & PG_FRAME) != newpf) { 843 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; 844 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 845 } 846 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); 847 } 848 return (0); 849} 850 851/* 852 * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte 853 * being NULL. 854 */ 855static __inline void 856pmap_pte_release(pt_entry_t *pte) 857{ 858 859 if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2) 860 mtx_unlock(&PMAP2mutex); 861} 862 863static __inline void 864invlcaddr(void *caddr) 865{ 866 867 invlpg((u_int)caddr); 868} 869 870/* 871 * Super fast pmap_pte routine best used when scanning 872 * the pv lists. This eliminates many coarse-grained 873 * invltlb calls. Note that many of the pv list 874 * scans are across different pmaps. It is very wasteful 875 * to do an entire invltlb for checking a single mapping. 876 * 877 * If the given pmap is not the current pmap, vm_page_queue_mtx 878 * must be held and curthread pinned to a CPU. 879 */ 880static pt_entry_t * 881pmap_pte_quick(pmap_t pmap, vm_offset_t va) 882{ 883 pd_entry_t newpf; 884 pd_entry_t *pde; 885 886 pde = pmap_pde(pmap, va); 887 if (*pde & PG_PS) 888 return (pde); 889 if (*pde != 0) { 890 /* are we current address space or kernel? */ 891 if (pmap_is_current(pmap)) 892 return (vtopte(va)); 893 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 894 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 895 newpf = *pde & PG_FRAME; 896 if ((*PMAP1 & PG_FRAME) != newpf) { 897 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; 898#ifdef SMP 899 PMAP1cpu = PCPU_GET(cpuid); 900#endif 901 invlcaddr(PADDR1); 902 PMAP1changed++; 903 } else 904#ifdef SMP 905 if (PMAP1cpu != PCPU_GET(cpuid)) { 906 PMAP1cpu = PCPU_GET(cpuid); 907 invlcaddr(PADDR1); 908 PMAP1changedcpu++; 909 } else 910#endif 911 PMAP1unchanged++; 912 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); 913 } 914 return (0); 915} 916 917/* 918 * Routine: pmap_extract 919 * Function: 920 * Extract the physical page address associated 921 * with the given map/virtual_address pair. 922 */ 923vm_paddr_t 924pmap_extract(pmap_t pmap, vm_offset_t va) 925{ 926 vm_paddr_t rtval; 927 pt_entry_t *pte; 928 pd_entry_t pde; 929 930 rtval = 0; 931 PMAP_LOCK(pmap); 932 pde = pmap->pm_pdir[va >> PDRSHIFT]; 933 if (pde != 0) { 934 if ((pde & PG_PS) != 0) { 935 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 936 PMAP_UNLOCK(pmap); 937 return rtval; 938 } 939 pte = pmap_pte(pmap, va); 940 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK); 941 pmap_pte_release(pte); 942 } 943 PMAP_UNLOCK(pmap); 944 return (rtval); 945} 946 947/* 948 * Routine: pmap_extract_and_hold 949 * Function: 950 * Atomically extract and hold the physical page 951 * with the given pmap and virtual address pair 952 * if that mapping permits the given protection. 953 */ 954vm_page_t 955pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 956{ 957 pd_entry_t pde; 958 pt_entry_t pte; 959 vm_page_t m; 960 961 m = NULL; 962 vm_page_lock_queues(); 963 PMAP_LOCK(pmap); 964 pde = *pmap_pde(pmap, va); 965 if (pde != 0) { 966 if (pde & PG_PS) { 967 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) { 968 m = PHYS_TO_VM_PAGE((pde & ~PDRMASK) | 969 (va & PDRMASK)); 970 vm_page_hold(m); 971 } 972 } else { 973 sched_pin(); 974 pte = *pmap_pte_quick(pmap, va); 975 if (pte != 0 && 976 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) { 977 m = PHYS_TO_VM_PAGE(pte & PG_FRAME); 978 vm_page_hold(m); 979 } 980 sched_unpin(); 981 } 982 } 983 vm_page_unlock_queues(); 984 PMAP_UNLOCK(pmap); 985 return (m); 986} 987 988/*************************************************** 989 * Low level mapping routines..... 990 ***************************************************/ 991 992/* 993 * Add a wired page to the kva. 994 * Note: not SMP coherent. 995 */ 996PMAP_INLINE void 997pmap_kenter(vm_offset_t va, vm_paddr_t pa) 998{ 999 pt_entry_t *pte; 1000 1001 pte = vtopte(va); 1002 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 1003} 1004 1005/* 1006 * Remove a page from the kernel pagetables. 1007 * Note: not SMP coherent. 1008 */ 1009PMAP_INLINE void 1010pmap_kremove(vm_offset_t va) 1011{ 1012 pt_entry_t *pte; 1013 1014 pte = vtopte(va); 1015 pte_clear(pte); 1016} 1017 1018/* 1019 * Used to map a range of physical addresses into kernel 1020 * virtual address space. 1021 * 1022 * The value passed in '*virt' is a suggested virtual address for 1023 * the mapping. Architectures which can support a direct-mapped 1024 * physical to virtual region can return the appropriate address 1025 * within that region, leaving '*virt' unchanged. Other 1026 * architectures should map the pages starting at '*virt' and 1027 * update '*virt' with the first usable address after the mapped 1028 * region. 1029 */ 1030vm_offset_t 1031pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 1032{ 1033 vm_offset_t va, sva; 1034 1035 va = sva = *virt; 1036 while (start < end) { 1037 pmap_kenter(va, start); 1038 va += PAGE_SIZE; 1039 start += PAGE_SIZE; 1040 } 1041 pmap_invalidate_range(kernel_pmap, sva, va); 1042 *virt = va; 1043 return (sva); 1044} 1045 1046 1047/* 1048 * Add a list of wired pages to the kva 1049 * this routine is only used for temporary 1050 * kernel mappings that do not need to have 1051 * page modification or references recorded. 1052 * Note that old mappings are simply written 1053 * over. The page *must* be wired. 1054 * Note: SMP coherent. Uses a ranged shootdown IPI. 1055 */ 1056void 1057pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count) 1058{ 1059 pt_entry_t *endpte, oldpte, *pte; 1060 1061 oldpte = 0; 1062 pte = vtopte(sva); 1063 endpte = pte + count; 1064 while (pte < endpte) { 1065 oldpte |= *pte; 1066 pte_store(pte, VM_PAGE_TO_PHYS(*ma) | pgeflag | PG_RW | PG_V); 1067 pte++; 1068 ma++; 1069 } 1070 if ((oldpte & PG_V) != 0) 1071 pmap_invalidate_range(kernel_pmap, sva, sva + count * 1072 PAGE_SIZE); 1073} 1074 1075/* 1076 * This routine tears out page mappings from the 1077 * kernel -- it is meant only for temporary mappings. 1078 * Note: SMP coherent. Uses a ranged shootdown IPI. 1079 */ 1080void 1081pmap_qremove(vm_offset_t sva, int count) 1082{ 1083 vm_offset_t va; 1084 1085 va = sva; 1086 while (count-- > 0) { 1087 pmap_kremove(va); 1088 va += PAGE_SIZE; 1089 } 1090 pmap_invalidate_range(kernel_pmap, sva, va); 1091} 1092 1093/*************************************************** 1094 * Page table page management routines..... 1095 ***************************************************/ 1096 1097/* 1098 * This routine unholds page table pages, and if the hold count 1099 * drops to zero, then it decrements the wire count. 1100 */ 1101static PMAP_INLINE int 1102pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1103{ 1104 1105 --m->wire_count; 1106 if (m->wire_count == 0) 1107 return _pmap_unwire_pte_hold(pmap, m); 1108 else 1109 return 0; 1110} 1111 1112static int 1113_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1114{ 1115 vm_offset_t pteva; 1116 1117 /* 1118 * unmap the page table page 1119 */ 1120 pmap->pm_pdir[m->pindex] = 0; 1121 --pmap->pm_stats.resident_count; 1122 1123 /* 1124 * Do an invltlb to make the invalidated mapping 1125 * take effect immediately. 1126 */ 1127 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1128 pmap_invalidate_page(pmap, pteva); 1129 1130 vm_page_free_zero(m); 1131 atomic_subtract_int(&cnt.v_wire_count, 1); 1132 return 1; 1133} 1134 1135/* 1136 * After removing a page table entry, this routine is used to 1137 * conditionally free the page, and manage the hold/wire counts. 1138 */ 1139static int 1140pmap_unuse_pt(pmap_t pmap, vm_offset_t va) 1141{ 1142 pd_entry_t ptepde; 1143 vm_page_t mpte; 1144 1145 if (va >= VM_MAXUSER_ADDRESS) 1146 return 0; 1147 ptepde = *pmap_pde(pmap, va); 1148 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); 1149 return pmap_unwire_pte_hold(pmap, mpte); 1150} 1151 1152void 1153pmap_pinit0(pmap_t pmap) 1154{ 1155 1156 PMAP_LOCK_INIT(pmap); 1157 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1158#ifdef PAE 1159 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1160#endif 1161 pmap->pm_active = 0; 1162 PCPU_SET(curpmap, pmap); 1163 TAILQ_INIT(&pmap->pm_pvchunk); 1164 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1165 mtx_lock_spin(&allpmaps_lock); 1166 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1167 mtx_unlock_spin(&allpmaps_lock); 1168} 1169 1170/* 1171 * Initialize a preallocated and zeroed pmap structure, 1172 * such as one in a vmspace structure. 1173 */ 1174void 1175pmap_pinit(pmap_t pmap) 1176{ 1177 vm_page_t m, ptdpg[NPGPTD]; 1178 vm_paddr_t pa; 1179 static int color; 1180 int i; 1181 1182 PMAP_LOCK_INIT(pmap); 1183 1184 /* 1185 * No need to allocate page table space yet but we do need a valid 1186 * page directory table. 1187 */ 1188 if (pmap->pm_pdir == NULL) { 1189 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1190 NBPTD); 1191#ifdef PAE 1192 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1193 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1194 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1195 ("pmap_pinit: pdpt misaligned")); 1196 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1197 ("pmap_pinit: pdpt above 4g")); 1198#endif 1199 } 1200 1201 /* 1202 * allocate the page directory page(s) 1203 */ 1204 for (i = 0; i < NPGPTD;) { 1205 m = vm_page_alloc(NULL, color++, 1206 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1207 VM_ALLOC_ZERO); 1208 if (m == NULL) 1209 VM_WAIT; 1210 else { 1211 ptdpg[i++] = m; 1212 } 1213 } 1214 1215 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1216 1217 for (i = 0; i < NPGPTD; i++) { 1218 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1219 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1220 } 1221 1222 mtx_lock_spin(&allpmaps_lock); 1223 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1224 mtx_unlock_spin(&allpmaps_lock); 1225 /* Wire in kernel global address entries. */ 1226 /* XXX copies current process, does not fill in MPPTDI */ 1227 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1228#ifdef SMP 1229 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1230#endif 1231 1232 /* install self-referential address mapping entry(s) */ 1233 for (i = 0; i < NPGPTD; i++) { 1234 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1235 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1236#ifdef PAE 1237 pmap->pm_pdpt[i] = pa | PG_V; 1238#endif 1239 } 1240 1241 pmap->pm_active = 0; 1242 TAILQ_INIT(&pmap->pm_pvchunk); 1243 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1244} 1245 1246/* 1247 * this routine is called if the page table page is not 1248 * mapped correctly. 1249 */ 1250static vm_page_t 1251_pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags) 1252{ 1253 vm_paddr_t ptepa; 1254 vm_page_t m; 1255 1256 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT || 1257 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK, 1258 ("_pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK")); 1259 1260 /* 1261 * Allocate a page table page. 1262 */ 1263 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1264 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1265 if (flags & M_WAITOK) { 1266 PMAP_UNLOCK(pmap); 1267 vm_page_unlock_queues(); 1268 VM_WAIT; 1269 vm_page_lock_queues(); 1270 PMAP_LOCK(pmap); 1271 } 1272 1273 /* 1274 * Indicate the need to retry. While waiting, the page table 1275 * page may have been allocated. 1276 */ 1277 return (NULL); 1278 } 1279 if ((m->flags & PG_ZERO) == 0) 1280 pmap_zero_page(m); 1281 1282 /* 1283 * Map the pagetable page into the process address space, if 1284 * it isn't already there. 1285 */ 1286 1287 pmap->pm_stats.resident_count++; 1288 1289 ptepa = VM_PAGE_TO_PHYS(m); 1290 pmap->pm_pdir[ptepindex] = 1291 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1292 1293 return m; 1294} 1295 1296static vm_page_t 1297pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags) 1298{ 1299 unsigned ptepindex; 1300 pd_entry_t ptepa; 1301 vm_page_t m; 1302 1303 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT || 1304 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK, 1305 ("pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK")); 1306 1307 /* 1308 * Calculate pagetable page index 1309 */ 1310 ptepindex = va >> PDRSHIFT; 1311retry: 1312 /* 1313 * Get the page directory entry 1314 */ 1315 ptepa = pmap->pm_pdir[ptepindex]; 1316 1317 /* 1318 * This supports switching from a 4MB page to a 1319 * normal 4K page. 1320 */ 1321 if (ptepa & PG_PS) { 1322 pmap->pm_pdir[ptepindex] = 0; 1323 ptepa = 0; 1324 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1325 pmap_invalidate_all(kernel_pmap); 1326 } 1327 1328 /* 1329 * If the page table page is mapped, we just increment the 1330 * hold count, and activate it. 1331 */ 1332 if (ptepa) { 1333 m = PHYS_TO_VM_PAGE(ptepa); 1334 m->wire_count++; 1335 } else { 1336 /* 1337 * Here if the pte page isn't mapped, or if it has 1338 * been deallocated. 1339 */ 1340 m = _pmap_allocpte(pmap, ptepindex, flags); 1341 if (m == NULL && (flags & M_WAITOK)) 1342 goto retry; 1343 } 1344 return (m); 1345} 1346 1347 1348/*************************************************** 1349* Pmap allocation/deallocation routines. 1350 ***************************************************/ 1351 1352#ifdef SMP 1353/* 1354 * Deal with a SMP shootdown of other users of the pmap that we are 1355 * trying to dispose of. This can be a bit hairy. 1356 */ 1357static u_int *lazymask; 1358static u_int lazyptd; 1359static volatile u_int lazywait; 1360 1361void pmap_lazyfix_action(void); 1362 1363void 1364pmap_lazyfix_action(void) 1365{ 1366 u_int mymask = PCPU_GET(cpumask); 1367 1368#ifdef COUNT_IPIS 1369 *ipi_lazypmap_counts[PCPU_GET(cpuid)]++; 1370#endif 1371 if (rcr3() == lazyptd) 1372 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1373 atomic_clear_int(lazymask, mymask); 1374 atomic_store_rel_int(&lazywait, 1); 1375} 1376 1377static void 1378pmap_lazyfix_self(u_int mymask) 1379{ 1380 1381 if (rcr3() == lazyptd) 1382 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1383 atomic_clear_int(lazymask, mymask); 1384} 1385 1386 1387static void 1388pmap_lazyfix(pmap_t pmap) 1389{ 1390 u_int mymask; 1391 u_int mask; 1392 u_int spins; 1393 1394 while ((mask = pmap->pm_active) != 0) { 1395 spins = 50000000; 1396 mask = mask & -mask; /* Find least significant set bit */ 1397 mtx_lock_spin(&smp_ipi_mtx); 1398#ifdef PAE 1399 lazyptd = vtophys(pmap->pm_pdpt); 1400#else 1401 lazyptd = vtophys(pmap->pm_pdir); 1402#endif 1403 mymask = PCPU_GET(cpumask); 1404 if (mask == mymask) { 1405 lazymask = &pmap->pm_active; 1406 pmap_lazyfix_self(mymask); 1407 } else { 1408 atomic_store_rel_int((u_int *)&lazymask, 1409 (u_int)&pmap->pm_active); 1410 atomic_store_rel_int(&lazywait, 0); 1411 ipi_selected(mask, IPI_LAZYPMAP); 1412 while (lazywait == 0) { 1413 ia32_pause(); 1414 if (--spins == 0) 1415 break; 1416 } 1417 } 1418 mtx_unlock_spin(&smp_ipi_mtx); 1419 if (spins == 0) 1420 printf("pmap_lazyfix: spun for 50000000\n"); 1421 } 1422} 1423 1424#else /* SMP */ 1425 1426/* 1427 * Cleaning up on uniprocessor is easy. For various reasons, we're 1428 * unlikely to have to even execute this code, including the fact 1429 * that the cleanup is deferred until the parent does a wait(2), which 1430 * means that another userland process has run. 1431 */ 1432static void 1433pmap_lazyfix(pmap_t pmap) 1434{ 1435 u_int cr3; 1436 1437 cr3 = vtophys(pmap->pm_pdir); 1438 if (cr3 == rcr3()) { 1439 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1440 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1441 } 1442} 1443#endif /* SMP */ 1444 1445/* 1446 * Release any resources held by the given physical map. 1447 * Called when a pmap initialized by pmap_pinit is being released. 1448 * Should only be called if the map contains no valid mappings. 1449 */ 1450void 1451pmap_release(pmap_t pmap) 1452{ 1453 vm_page_t m, ptdpg[NPGPTD]; 1454 int i; 1455 1456 KASSERT(pmap->pm_stats.resident_count == 0, 1457 ("pmap_release: pmap resident count %ld != 0", 1458 pmap->pm_stats.resident_count)); 1459 1460 pmap_lazyfix(pmap); 1461 mtx_lock_spin(&allpmaps_lock); 1462 LIST_REMOVE(pmap, pm_list); 1463 mtx_unlock_spin(&allpmaps_lock); 1464 1465 for (i = 0; i < NPGPTD; i++) 1466 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1467 1468 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1469 sizeof(*pmap->pm_pdir)); 1470#ifdef SMP 1471 pmap->pm_pdir[MPPTDI] = 0; 1472#endif 1473 1474 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1475 1476 vm_page_lock_queues(); 1477 for (i = 0; i < NPGPTD; i++) { 1478 m = ptdpg[i]; 1479#ifdef PAE 1480 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1481 ("pmap_release: got wrong ptd page")); 1482#endif 1483 m->wire_count--; 1484 atomic_subtract_int(&cnt.v_wire_count, 1); 1485 vm_page_free_zero(m); 1486 } 1487 vm_page_unlock_queues(); 1488 PMAP_LOCK_DESTROY(pmap); 1489} 1490 1491static int 1492kvm_size(SYSCTL_HANDLER_ARGS) 1493{ 1494 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1495 1496 return sysctl_handle_long(oidp, &ksize, 0, req); 1497} 1498SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1499 0, 0, kvm_size, "IU", "Size of KVM"); 1500 1501static int 1502kvm_free(SYSCTL_HANDLER_ARGS) 1503{ 1504 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1505 1506 return sysctl_handle_long(oidp, &kfree, 0, req); 1507} 1508SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1509 0, 0, kvm_free, "IU", "Amount of KVM free"); 1510 1511/* 1512 * grow the number of kernel page table entries, if needed 1513 */ 1514void 1515pmap_growkernel(vm_offset_t addr) 1516{ 1517 struct pmap *pmap; 1518 vm_paddr_t ptppaddr; 1519 vm_page_t nkpg; 1520 pd_entry_t newpdir; 1521 pt_entry_t *pde; 1522 1523 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1524 if (kernel_vm_end == 0) { 1525 kernel_vm_end = KERNBASE; 1526 nkpt = 0; 1527 while (pdir_pde(PTD, kernel_vm_end)) { 1528 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1529 nkpt++; 1530 if (kernel_vm_end - 1 >= kernel_map->max_offset) { 1531 kernel_vm_end = kernel_map->max_offset; 1532 break; 1533 } 1534 } 1535 } 1536 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1537 if (addr - 1 >= kernel_map->max_offset) 1538 addr = kernel_map->max_offset; 1539 while (kernel_vm_end < addr) { 1540 if (pdir_pde(PTD, kernel_vm_end)) { 1541 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1542 if (kernel_vm_end - 1 >= kernel_map->max_offset) { 1543 kernel_vm_end = kernel_map->max_offset; 1544 break; 1545 } 1546 continue; 1547 } 1548 1549 /* 1550 * This index is bogus, but out of the way 1551 */ 1552 nkpg = vm_page_alloc(NULL, nkpt, 1553 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1554 if (!nkpg) 1555 panic("pmap_growkernel: no memory to grow kernel"); 1556 1557 nkpt++; 1558 1559 pmap_zero_page(nkpg); 1560 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1561 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1562 pdir_pde(PTD, kernel_vm_end) = newpdir; 1563 1564 mtx_lock_spin(&allpmaps_lock); 1565 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1566 pde = pmap_pde(pmap, kernel_vm_end); 1567 pde_store(pde, newpdir); 1568 } 1569 mtx_unlock_spin(&allpmaps_lock); 1570 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1571 if (kernel_vm_end - 1 >= kernel_map->max_offset) { 1572 kernel_vm_end = kernel_map->max_offset; 1573 break; 1574 } 1575 } 1576} 1577 1578 1579/*************************************************** 1580 * page management routines. 1581 ***************************************************/ 1582 1583CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE); 1584CTASSERT(_NPCM == 11); 1585 1586static __inline struct pv_chunk * 1587pv_to_chunk(pv_entry_t pv) 1588{ 1589 1590 return (struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK); 1591} 1592 1593#define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap) 1594 1595#define PC_FREE0_9 0xfffffffful /* Free values for index 0 through 9 */ 1596#define PC_FREE10 0x0000fffful /* Free values for index 10 */ 1597 1598static uint32_t pc_freemask[11] = { 1599 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, 1600 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, 1601 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, 1602 PC_FREE0_9, PC_FREE10 1603}; 1604 1605SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0, 1606 "Current number of pv entries"); 1607 1608#ifdef PV_STATS 1609static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail; 1610 1611SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0, 1612 "Current number of pv entry chunks"); 1613SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0, 1614 "Current number of pv entry chunks allocated"); 1615SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0, 1616 "Current number of pv entry chunks frees"); 1617SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0, 1618 "Number of times tried to get a chunk page but failed."); 1619 1620static long pv_entry_frees, pv_entry_allocs; 1621static int pv_entry_spare; 1622 1623SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0, 1624 "Current number of pv entry frees"); 1625SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0, 1626 "Current number of pv entry allocs"); 1627SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0, 1628 "Current number of spare pv entries"); 1629 1630static int pmap_collect_inactive, pmap_collect_active; 1631 1632SYSCTL_INT(_vm_pmap, OID_AUTO, pmap_collect_inactive, CTLFLAG_RD, &pmap_collect_inactive, 0, 1633 "Current number times pmap_collect called on inactive queue"); 1634SYSCTL_INT(_vm_pmap, OID_AUTO, pmap_collect_active, CTLFLAG_RD, &pmap_collect_active, 0, 1635 "Current number times pmap_collect called on active queue"); 1636#endif 1637 1638/* 1639 * We are in a serious low memory condition. Resort to 1640 * drastic measures to free some pages so we can allocate 1641 * another pv entry chunk. This is normally called to 1642 * unmap inactive pages, and if necessary, active pages. 1643 */ 1644static void 1645pmap_collect(pmap_t locked_pmap, struct vpgqueues *vpq) 1646{ 1647 pmap_t pmap; 1648 pt_entry_t *pte, tpte; 1649 pv_entry_t next_pv, pv; 1650 vm_offset_t va; 1651 vm_page_t m; 1652 1653 sched_pin(); 1654 TAILQ_FOREACH(m, &vpq->pl, pageq) { 1655 if (m->hold_count || m->busy || (m->flags & PG_BUSY)) 1656 continue; 1657 TAILQ_FOREACH_SAFE(pv, &m->md.pv_list, pv_list, next_pv) { 1658 va = pv->pv_va; 1659 pmap = PV_PMAP(pv); 1660 /* Avoid deadlock and lock recursion. */ 1661 if (pmap > locked_pmap) 1662 PMAP_LOCK(pmap); 1663 else if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap)) 1664 continue; 1665 pmap->pm_stats.resident_count--; 1666 pte = pmap_pte_quick(pmap, va); 1667 tpte = pte_load_clear(pte); 1668 KASSERT((tpte & PG_W) == 0, 1669 ("pmap_collect: wired pte %#jx", (uintmax_t)tpte)); 1670 if (tpte & PG_A) 1671 vm_page_flag_set(m, PG_REFERENCED); 1672 if (tpte & PG_M) { 1673 KASSERT((tpte & PG_RW), 1674 ("pmap_collect: modified page not writable: va: %#x, pte: %#jx", 1675 va, (uintmax_t)tpte)); 1676 vm_page_dirty(m); 1677 } 1678 pmap_invalidate_page(pmap, va); 1679 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1680 if (TAILQ_EMPTY(&m->md.pv_list)) 1681 vm_page_flag_clear(m, PG_WRITEABLE); 1682 m->md.pv_list_count--; 1683 pmap_unuse_pt(pmap, va); 1684 free_pv_entry(pmap, pv); 1685 if (pmap != locked_pmap) 1686 PMAP_UNLOCK(pmap); 1687 } 1688 } 1689 sched_unpin(); 1690} 1691 1692 1693/* 1694 * free the pv_entry back to the free list 1695 */ 1696static void 1697free_pv_entry(pmap_t pmap, pv_entry_t pv) 1698{ 1699 vm_page_t m; 1700 struct pv_chunk *pc; 1701 int idx, field, bit; 1702 1703 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1704 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1705 PV_STAT(pv_entry_frees++); 1706 PV_STAT(pv_entry_spare++); 1707 pv_entry_count--; 1708 pc = pv_to_chunk(pv); 1709 idx = pv - &pc->pc_pventry[0]; 1710 field = idx / 32; 1711 bit = idx % 32; 1712 pc->pc_map[field] |= 1ul << bit; 1713 /* move to head of list */ 1714 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); 1715 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); 1716 for (idx = 0; idx < _NPCM; idx++) 1717 if (pc->pc_map[idx] != pc_freemask[idx]) 1718 return; 1719 PV_STAT(pv_entry_spare -= _NPCPV); 1720 PV_STAT(pc_chunk_count--); 1721 PV_STAT(pc_chunk_frees++); 1722 /* entire chunk is free, return it */ 1723 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); 1724 m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc)); 1725 pmap_qremove((vm_offset_t)pc, 1); 1726 vm_page_unwire(m, 0); 1727 vm_page_free(m); 1728 pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc); 1729} 1730 1731/* 1732 * get a new pv_entry, allocating a block from the system 1733 * when needed. 1734 */ 1735static pv_entry_t 1736get_pv_entry(pmap_t pmap, int try) 1737{ 1738 static const struct timeval printinterval = { 60, 0 }; 1739 static struct timeval lastprint; 1740 static vm_pindex_t colour; 1741 int bit, field, page_req; 1742 pv_entry_t pv; 1743 struct pv_chunk *pc; 1744 vm_page_t m; 1745 1746 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1747 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1748 PV_STAT(pv_entry_allocs++); 1749 pv_entry_count++; 1750 if (pv_entry_count > pv_entry_high_water) 1751 pagedaemon_wakeup(); 1752 pc = TAILQ_FIRST(&pmap->pm_pvchunk); 1753 if (pc != NULL) { 1754 for (field = 0; field < _NPCM; field++) { 1755 if (pc->pc_map[field]) { 1756 bit = bsfl(pc->pc_map[field]); 1757 break; 1758 } 1759 } 1760 if (field < _NPCM) { 1761 pv = &pc->pc_pventry[field * 32 + bit]; 1762 pc->pc_map[field] &= ~(1ul << bit); 1763 /* If this was the last item, move it to tail */ 1764 for (field = 0; field < _NPCM; field++) 1765 if (pc->pc_map[field] != 0) { 1766 PV_STAT(pv_entry_spare--); 1767 return (pv); /* not full, return */ 1768 } 1769 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); 1770 TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); 1771 PV_STAT(pv_entry_spare--); 1772 return (pv); 1773 } 1774 } 1775 pc = (struct pv_chunk *)pmap_ptelist_alloc(&pv_vafree); 1776 page_req = try ? VM_ALLOC_NORMAL : VM_ALLOC_SYSTEM; 1777 m = vm_page_alloc(NULL, colour, page_req | 1778 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED); 1779 if (m == NULL || pc == NULL) { 1780 if (try) { 1781 pv_entry_count--; 1782 PV_STAT(pc_chunk_tryfail++); 1783 if (m) { 1784 vm_page_lock_queues(); 1785 vm_page_unwire(m, 0); 1786 vm_page_free(m); 1787 vm_page_unlock_queues(); 1788 } 1789 if (pc) 1790 pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc); 1791 return (NULL); 1792 } 1793 /* 1794 * Reclaim pv entries: At first, destroy mappings to 1795 * inactive pages. After that, if a pv chunk entry 1796 * is still needed, destroy mappings to active pages. 1797 */ 1798 if (ratecheck(&lastprint, &printinterval)) 1799 printf("Approaching the limit on PV entries, " 1800 "consider increasing tunables " 1801 "vm.pmap.shpgperproc or " 1802 "vm.pmap.pv_entry_max\n"); 1803 PV_STAT(pmap_collect_inactive++); 1804 pmap_collect(pmap, &vm_page_queues[PQ_INACTIVE]); 1805 if (m == NULL) 1806 m = vm_page_alloc(NULL, colour, VM_ALLOC_SYSTEM | 1807 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED); 1808 if (pc == NULL) 1809 pc = (struct pv_chunk *)pmap_ptelist_alloc(&pv_vafree); 1810 if (m == NULL || pc == NULL) { 1811 PV_STAT(pmap_collect_active++); 1812 pmap_collect(pmap, &vm_page_queues[PQ_ACTIVE]); 1813 if (m == NULL) 1814 m = vm_page_alloc(NULL, colour, 1815 VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 1816 VM_ALLOC_WIRED); 1817 if (pc == NULL) 1818 pc = (struct pv_chunk *) 1819 pmap_ptelist_alloc(&pv_vafree); 1820 if (m == NULL || pc == NULL) 1821 panic("get_pv_entry: increase vm.pmap.shpgperproc"); 1822 } 1823 } 1824 PV_STAT(pc_chunk_count++); 1825 PV_STAT(pc_chunk_allocs++); 1826 colour++; 1827 pmap_qenter((vm_offset_t)pc, &m, 1); 1828 pc->pc_pmap = pmap; 1829 pc->pc_map[0] = pc_freemask[0] & ~1ul; /* preallocated bit 0 */ 1830 for (field = 1; field < _NPCM; field++) 1831 pc->pc_map[field] = pc_freemask[field]; 1832 pv = &pc->pc_pventry[0]; 1833 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); 1834 PV_STAT(pv_entry_spare += _NPCPV - 1); 1835 return (pv); 1836} 1837 1838static void 1839pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1840{ 1841 pv_entry_t pv; 1842 1843 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1844 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1845 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1846 if (pmap == PV_PMAP(pv) && va == pv->pv_va) 1847 break; 1848 } 1849 KASSERT(pv != NULL, ("pmap_remove_entry: pv not found")); 1850 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1851 m->md.pv_list_count--; 1852 if (TAILQ_EMPTY(&m->md.pv_list)) 1853 vm_page_flag_clear(m, PG_WRITEABLE); 1854 free_pv_entry(pmap, pv); 1855} 1856 1857/* 1858 * Create a pv entry for page at pa for 1859 * (pmap, va). 1860 */ 1861static void 1862pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) 1863{ 1864 pv_entry_t pv; 1865 1866 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1867 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1868 pv = get_pv_entry(pmap, FALSE); 1869 pv->pv_va = va; 1870 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1871 m->md.pv_list_count++; 1872} 1873 1874/* 1875 * Conditionally create a pv entry. 1876 */ 1877static boolean_t 1878pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) 1879{ 1880 pv_entry_t pv; 1881 1882 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1883 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1884 if (pv_entry_count < pv_entry_high_water && 1885 (pv = get_pv_entry(pmap, TRUE)) != NULL) { 1886 pv->pv_va = va; 1887 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1888 m->md.pv_list_count++; 1889 return (TRUE); 1890 } else 1891 return (FALSE); 1892} 1893 1894/* 1895 * pmap_remove_pte: do the things to unmap a page in a process 1896 */ 1897static int 1898pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1899{ 1900 pt_entry_t oldpte; 1901 vm_page_t m; 1902 1903 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1904 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1905 oldpte = pte_load_clear(ptq); 1906 if (oldpte & PG_W) 1907 pmap->pm_stats.wired_count -= 1; 1908 /* 1909 * Machines that don't support invlpg, also don't support 1910 * PG_G. 1911 */ 1912 if (oldpte & PG_G) 1913 pmap_invalidate_page(kernel_pmap, va); 1914 pmap->pm_stats.resident_count -= 1; 1915 if (oldpte & PG_MANAGED) { 1916 m = PHYS_TO_VM_PAGE(oldpte); 1917 if (oldpte & PG_M) { 1918 KASSERT((oldpte & PG_RW), 1919 ("pmap_remove_pte: modified page not writable: va: %#x, pte: %#jx", 1920 va, (uintmax_t)oldpte)); 1921 vm_page_dirty(m); 1922 } 1923 if (oldpte & PG_A) 1924 vm_page_flag_set(m, PG_REFERENCED); 1925 pmap_remove_entry(pmap, m, va); 1926 } 1927 return (pmap_unuse_pt(pmap, va)); 1928} 1929 1930/* 1931 * Remove a single page from a process address space 1932 */ 1933static void 1934pmap_remove_page(pmap_t pmap, vm_offset_t va) 1935{ 1936 pt_entry_t *pte; 1937 1938 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1939 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 1940 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1941 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) 1942 return; 1943 pmap_remove_pte(pmap, pte, va); 1944 pmap_invalidate_page(pmap, va); 1945} 1946 1947/* 1948 * Remove the given range of addresses from the specified map. 1949 * 1950 * It is assumed that the start and end are properly 1951 * rounded to the page size. 1952 */ 1953void 1954pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1955{ 1956 vm_offset_t pdnxt; 1957 pd_entry_t ptpaddr; 1958 pt_entry_t *pte; 1959 int anyvalid; 1960 1961 /* 1962 * Perform an unsynchronized read. This is, however, safe. 1963 */ 1964 if (pmap->pm_stats.resident_count == 0) 1965 return; 1966 1967 anyvalid = 0; 1968 1969 vm_page_lock_queues(); 1970 sched_pin(); 1971 PMAP_LOCK(pmap); 1972 1973 /* 1974 * special handling of removing one page. a very 1975 * common operation and easy to short circuit some 1976 * code. 1977 */ 1978 if ((sva + PAGE_SIZE == eva) && 1979 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1980 pmap_remove_page(pmap, sva); 1981 goto out; 1982 } 1983 1984 for (; sva < eva; sva = pdnxt) { 1985 unsigned pdirindex; 1986 1987 /* 1988 * Calculate index for next page table. 1989 */ 1990 pdnxt = (sva + NBPDR) & ~PDRMASK; 1991 if (pmap->pm_stats.resident_count == 0) 1992 break; 1993 1994 pdirindex = sva >> PDRSHIFT; 1995 ptpaddr = pmap->pm_pdir[pdirindex]; 1996 1997 /* 1998 * Weed out invalid mappings. Note: we assume that the page 1999 * directory table is always allocated, and in kernel virtual. 2000 */ 2001 if (ptpaddr == 0) 2002 continue; 2003 2004 /* 2005 * Check for large page. 2006 */ 2007 if ((ptpaddr & PG_PS) != 0) { 2008 pmap->pm_pdir[pdirindex] = 0; 2009 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 2010 anyvalid = 1; 2011 continue; 2012 } 2013 2014 /* 2015 * Limit our scan to either the end of the va represented 2016 * by the current page table page, or to the end of the 2017 * range being removed. 2018 */ 2019 if (pdnxt > eva) 2020 pdnxt = eva; 2021 2022 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, 2023 sva += PAGE_SIZE) { 2024 if (*pte == 0) 2025 continue; 2026 2027 /* 2028 * The TLB entry for a PG_G mapping is invalidated 2029 * by pmap_remove_pte(). 2030 */ 2031 if ((*pte & PG_G) == 0) 2032 anyvalid = 1; 2033 if (pmap_remove_pte(pmap, pte, sva)) 2034 break; 2035 } 2036 } 2037out: 2038 sched_unpin(); 2039 vm_page_unlock_queues(); 2040 if (anyvalid) 2041 pmap_invalidate_all(pmap); 2042 PMAP_UNLOCK(pmap); 2043} 2044 2045/* 2046 * Routine: pmap_remove_all 2047 * Function: 2048 * Removes this physical page from 2049 * all physical maps in which it resides. 2050 * Reflects back modify bits to the pager. 2051 * 2052 * Notes: 2053 * Original versions of this routine were very 2054 * inefficient because they iteratively called 2055 * pmap_remove (slow...) 2056 */ 2057 2058void 2059pmap_remove_all(vm_page_t m) 2060{ 2061 pv_entry_t pv; 2062 pmap_t pmap; 2063 pt_entry_t *pte, tpte; 2064 2065#if defined(PMAP_DIAGNOSTIC) 2066 /* 2067 * XXX This makes pmap_remove_all() illegal for non-managed pages! 2068 */ 2069 if (m->flags & PG_FICTITIOUS) { 2070 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 2071 VM_PAGE_TO_PHYS(m)); 2072 } 2073#endif 2074 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2075 sched_pin(); 2076 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2077 pmap = PV_PMAP(pv); 2078 PMAP_LOCK(pmap); 2079 pmap->pm_stats.resident_count--; 2080 pte = pmap_pte_quick(pmap, pv->pv_va); 2081 tpte = pte_load_clear(pte); 2082 if (tpte & PG_W) 2083 pmap->pm_stats.wired_count--; 2084 if (tpte & PG_A) 2085 vm_page_flag_set(m, PG_REFERENCED); 2086 2087 /* 2088 * Update the vm_page_t clean and reference bits. 2089 */ 2090 if (tpte & PG_M) { 2091 KASSERT((tpte & PG_RW), 2092 ("pmap_remove_all: modified page not writable: va: %#x, pte: %#jx", 2093 pv->pv_va, (uintmax_t)tpte)); 2094 vm_page_dirty(m); 2095 } 2096 pmap_invalidate_page(pmap, pv->pv_va); 2097 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2098 m->md.pv_list_count--; 2099 pmap_unuse_pt(pmap, pv->pv_va); 2100 free_pv_entry(pmap, pv); 2101 PMAP_UNLOCK(pmap); 2102 } 2103 vm_page_flag_clear(m, PG_WRITEABLE); 2104 sched_unpin(); 2105} 2106 2107/* 2108 * Set the physical protection on the 2109 * specified range of this map as requested. 2110 */ 2111void 2112pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 2113{ 2114 vm_offset_t pdnxt; 2115 pd_entry_t ptpaddr; 2116 pt_entry_t *pte; 2117 int anychanged; 2118 2119 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 2120 pmap_remove(pmap, sva, eva); 2121 return; 2122 } 2123 2124 if (prot & VM_PROT_WRITE) 2125 return; 2126 2127 anychanged = 0; 2128 2129 vm_page_lock_queues(); 2130 sched_pin(); 2131 PMAP_LOCK(pmap); 2132 for (; sva < eva; sva = pdnxt) { 2133 unsigned obits, pbits, pdirindex; 2134 2135 pdnxt = (sva + NBPDR) & ~PDRMASK; 2136 2137 pdirindex = sva >> PDRSHIFT; 2138 ptpaddr = pmap->pm_pdir[pdirindex]; 2139 2140 /* 2141 * Weed out invalid mappings. Note: we assume that the page 2142 * directory table is always allocated, and in kernel virtual. 2143 */ 2144 if (ptpaddr == 0) 2145 continue; 2146 2147 /* 2148 * Check for large page. 2149 */ 2150 if ((ptpaddr & PG_PS) != 0) { 2151 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 2152 anychanged = 1; 2153 continue; 2154 } 2155 2156 if (pdnxt > eva) 2157 pdnxt = eva; 2158 2159 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, 2160 sva += PAGE_SIZE) { 2161 vm_page_t m; 2162 2163retry: 2164 /* 2165 * Regardless of whether a pte is 32 or 64 bits in 2166 * size, PG_RW, PG_A, and PG_M are among the least 2167 * significant 32 bits. 2168 */ 2169 obits = pbits = *(u_int *)pte; 2170 if (pbits & PG_MANAGED) { 2171 m = NULL; 2172 if (pbits & PG_A) { 2173 m = PHYS_TO_VM_PAGE(*pte); 2174 vm_page_flag_set(m, PG_REFERENCED); 2175 pbits &= ~PG_A; 2176 } 2177 if ((pbits & PG_M) != 0) { 2178 if (m == NULL) 2179 m = PHYS_TO_VM_PAGE(*pte); 2180 vm_page_dirty(m); 2181 } 2182 } 2183 2184 pbits &= ~(PG_RW | PG_M); 2185 2186 if (pbits != obits) { 2187 if (!atomic_cmpset_int((u_int *)pte, obits, 2188 pbits)) 2189 goto retry; 2190 if (obits & PG_G) 2191 pmap_invalidate_page(pmap, sva); 2192 else 2193 anychanged = 1; 2194 } 2195 } 2196 } 2197 sched_unpin(); 2198 vm_page_unlock_queues(); 2199 if (anychanged) 2200 pmap_invalidate_all(pmap); 2201 PMAP_UNLOCK(pmap); 2202} 2203 2204/* 2205 * Insert the given physical page (p) at 2206 * the specified virtual address (v) in the 2207 * target physical map with the protection requested. 2208 * 2209 * If specified, the page will be wired down, meaning 2210 * that the related pte can not be reclaimed. 2211 * 2212 * NB: This is the only routine which MAY NOT lazy-evaluate 2213 * or lose information. That is, this routine must actually 2214 * insert this page into the given map NOW. 2215 */ 2216void 2217pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 2218 boolean_t wired) 2219{ 2220 vm_paddr_t pa; 2221 pd_entry_t *pde; 2222 pt_entry_t *pte; 2223 vm_paddr_t opa; 2224 pt_entry_t origpte, newpte; 2225 vm_page_t mpte, om; 2226 boolean_t invlva; 2227 2228 va &= PG_FRAME; 2229#ifdef PMAP_DIAGNOSTIC 2230 if (va > VM_MAX_KERNEL_ADDRESS) 2231 panic("pmap_enter: toobig"); 2232 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 2233 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 2234#endif 2235 2236 mpte = NULL; 2237 2238 vm_page_lock_queues(); 2239 PMAP_LOCK(pmap); 2240 sched_pin(); 2241 2242 /* 2243 * In the case that a page table page is not 2244 * resident, we are creating it here. 2245 */ 2246 if (va < VM_MAXUSER_ADDRESS) { 2247 mpte = pmap_allocpte(pmap, va, M_WAITOK); 2248 } 2249#if 0 && defined(PMAP_DIAGNOSTIC) 2250 else { 2251 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 2252 origpte = *pdeaddr; 2253 if ((origpte & PG_V) == 0) { 2254 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 2255 pmap->pm_pdir[PTDPTDI], origpte, va); 2256 } 2257 } 2258#endif 2259 2260 pde = pmap_pde(pmap, va); 2261 if ((*pde & PG_PS) != 0) 2262 panic("pmap_enter: attempted pmap_enter on 4MB page"); 2263 pte = pmap_pte_quick(pmap, va); 2264 2265 /* 2266 * Page Directory table entry not valid, we need a new PT page 2267 */ 2268 if (pte == NULL) { 2269 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 2270 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 2271 } 2272 2273 pa = VM_PAGE_TO_PHYS(m); 2274 om = NULL; 2275 origpte = *pte; 2276 opa = origpte & PG_FRAME; 2277 2278 /* 2279 * Mapping has not changed, must be protection or wiring change. 2280 */ 2281 if (origpte && (opa == pa)) { 2282 /* 2283 * Wiring change, just update stats. We don't worry about 2284 * wiring PT pages as they remain resident as long as there 2285 * are valid mappings in them. Hence, if a user page is wired, 2286 * the PT page will be also. 2287 */ 2288 if (wired && ((origpte & PG_W) == 0)) 2289 pmap->pm_stats.wired_count++; 2290 else if (!wired && (origpte & PG_W)) 2291 pmap->pm_stats.wired_count--; 2292 2293 /* 2294 * Remove extra pte reference 2295 */ 2296 if (mpte) 2297 mpte->wire_count--; 2298 2299 /* 2300 * We might be turning off write access to the page, 2301 * so we go ahead and sense modify status. 2302 */ 2303 if (origpte & PG_MANAGED) { 2304 om = m; 2305 pa |= PG_MANAGED; 2306 } 2307 goto validate; 2308 } 2309 /* 2310 * Mapping has changed, invalidate old range and fall through to 2311 * handle validating new mapping. 2312 */ 2313 if (opa) { 2314 if (origpte & PG_W) 2315 pmap->pm_stats.wired_count--; 2316 if (origpte & PG_MANAGED) { 2317 om = PHYS_TO_VM_PAGE(opa); 2318 pmap_remove_entry(pmap, om, va); 2319 } 2320 if (mpte != NULL) { 2321 mpte->wire_count--; 2322 KASSERT(mpte->wire_count > 0, 2323 ("pmap_enter: missing reference to page table page," 2324 " va: 0x%x", va)); 2325 } 2326 } else 2327 pmap->pm_stats.resident_count++; 2328 2329 /* 2330 * Enter on the PV list if part of our managed memory. 2331 */ 2332 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0) { 2333 KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva, 2334 ("pmap_enter: managed mapping within the clean submap")); 2335 pmap_insert_entry(pmap, va, m); 2336 pa |= PG_MANAGED; 2337 } 2338 2339 /* 2340 * Increment counters 2341 */ 2342 if (wired) 2343 pmap->pm_stats.wired_count++; 2344 2345validate: 2346 /* 2347 * Now validate mapping with desired protection/wiring. 2348 */ 2349 newpte = (pt_entry_t)(pa | PG_V); 2350 if ((prot & VM_PROT_WRITE) != 0) 2351 newpte |= PG_RW; 2352 if (wired) 2353 newpte |= PG_W; 2354 if (va < VM_MAXUSER_ADDRESS) 2355 newpte |= PG_U; 2356 if (pmap == kernel_pmap) 2357 newpte |= pgeflag; 2358 2359 /* 2360 * if the mapping or permission bits are different, we need 2361 * to update the pte. 2362 */ 2363 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2364 if (origpte & PG_V) { 2365 invlva = FALSE; 2366 origpte = pte_load_store(pte, newpte | PG_A); 2367 if (origpte & PG_A) { 2368 if (origpte & PG_MANAGED) 2369 vm_page_flag_set(om, PG_REFERENCED); 2370 if (opa != VM_PAGE_TO_PHYS(m)) 2371 invlva = TRUE; 2372 } 2373 if (origpte & PG_M) { 2374 KASSERT((origpte & PG_RW), 2375 ("pmap_enter: modified page not writable: va: %#x, pte: %#jx", 2376 va, (uintmax_t)origpte)); 2377 if ((origpte & PG_MANAGED) != 0) 2378 vm_page_dirty(om); 2379 if ((prot & VM_PROT_WRITE) == 0) 2380 invlva = TRUE; 2381 } 2382 if (invlva) 2383 pmap_invalidate_page(pmap, va); 2384 } else 2385 pte_store(pte, newpte | PG_A); 2386 } 2387 sched_unpin(); 2388 vm_page_unlock_queues(); 2389 PMAP_UNLOCK(pmap); 2390} 2391 2392/* 2393 * Maps a sequence of resident pages belonging to the same object. 2394 * The sequence begins with the given page m_start. This page is 2395 * mapped at the given virtual address start. Each subsequent page is 2396 * mapped at a virtual address that is offset from start by the same 2397 * amount as the page is offset from m_start within the object. The 2398 * last page in the sequence is the page with the largest offset from 2399 * m_start that can be mapped at a virtual address less than the given 2400 * virtual address end. Not every virtual page between start and end 2401 * is mapped; only those for which a resident page exists with the 2402 * corresponding offset from m_start are mapped. 2403 */ 2404void 2405pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, 2406 vm_page_t m_start, vm_prot_t prot) 2407{ 2408 vm_page_t m, mpte; 2409 vm_pindex_t diff, psize; 2410 2411 VM_OBJECT_LOCK_ASSERT(m_start->object, MA_OWNED); 2412 psize = atop(end - start); 2413 mpte = NULL; 2414 m = m_start; 2415 PMAP_LOCK(pmap); 2416 while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { 2417 mpte = pmap_enter_quick_locked(pmap, start + ptoa(diff), m, 2418 prot, mpte); 2419 m = TAILQ_NEXT(m, listq); 2420 } 2421 PMAP_UNLOCK(pmap); 2422} 2423 2424/* 2425 * this code makes some *MAJOR* assumptions: 2426 * 1. Current pmap & pmap exists. 2427 * 2. Not wired. 2428 * 3. Read access. 2429 * 4. No page table pages. 2430 * but is *MUCH* faster than pmap_enter... 2431 */ 2432 2433void 2434pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) 2435{ 2436 2437 PMAP_LOCK(pmap); 2438 (void) pmap_enter_quick_locked(pmap, va, m, prot, NULL); 2439 PMAP_UNLOCK(pmap); 2440} 2441 2442static vm_page_t 2443pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, 2444 vm_prot_t prot, vm_page_t mpte) 2445{ 2446 pt_entry_t *pte; 2447 vm_paddr_t pa; 2448 2449 KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva || 2450 (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0, 2451 ("pmap_enter_quick_locked: managed mapping within the clean submap")); 2452 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2453 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 2454 2455 /* 2456 * In the case that a page table page is not 2457 * resident, we are creating it here. 2458 */ 2459 if (va < VM_MAXUSER_ADDRESS) { 2460 unsigned ptepindex; 2461 pd_entry_t ptepa; 2462 2463 /* 2464 * Calculate pagetable page index 2465 */ 2466 ptepindex = va >> PDRSHIFT; 2467 if (mpte && (mpte->pindex == ptepindex)) { 2468 mpte->wire_count++; 2469 } else { 2470 /* 2471 * Get the page directory entry 2472 */ 2473 ptepa = pmap->pm_pdir[ptepindex]; 2474 2475 /* 2476 * If the page table page is mapped, we just increment 2477 * the hold count, and activate it. 2478 */ 2479 if (ptepa) { 2480 if (ptepa & PG_PS) 2481 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2482 mpte = PHYS_TO_VM_PAGE(ptepa); 2483 mpte->wire_count++; 2484 } else { 2485 mpte = _pmap_allocpte(pmap, ptepindex, 2486 M_NOWAIT); 2487 if (mpte == NULL) 2488 return (mpte); 2489 } 2490 } 2491 } else { 2492 mpte = NULL; 2493 } 2494 2495 /* 2496 * This call to vtopte makes the assumption that we are 2497 * entering the page into the current pmap. In order to support 2498 * quick entry into any pmap, one would likely use pmap_pte_quick. 2499 * But that isn't as quick as vtopte. 2500 */ 2501 pte = vtopte(va); 2502 if (*pte) { 2503 if (mpte != NULL) { 2504 pmap_unwire_pte_hold(pmap, mpte); 2505 mpte = NULL; 2506 } 2507 return (mpte); 2508 } 2509 2510 /* 2511 * Enter on the PV list if part of our managed memory. 2512 */ 2513 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0 && 2514 !pmap_try_insert_pv_entry(pmap, va, m)) { 2515 if (mpte != NULL) { 2516 pmap_unwire_pte_hold(pmap, mpte); 2517 mpte = NULL; 2518 } 2519 return (mpte); 2520 } 2521 2522 /* 2523 * Increment counters 2524 */ 2525 pmap->pm_stats.resident_count++; 2526 2527 pa = VM_PAGE_TO_PHYS(m); 2528 2529 /* 2530 * Now validate mapping with RO protection 2531 */ 2532 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2533 pte_store(pte, pa | PG_V | PG_U); 2534 else 2535 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2536 return mpte; 2537} 2538 2539/* 2540 * Make a temporary mapping for a physical address. This is only intended 2541 * to be used for panic dumps. 2542 */ 2543void * 2544pmap_kenter_temporary(vm_paddr_t pa, int i) 2545{ 2546 vm_offset_t va; 2547 2548 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2549 pmap_kenter(va, pa); 2550 invlpg(va); 2551 return ((void *)crashdumpmap); 2552} 2553 2554/* 2555 * This code maps large physical mmap regions into the 2556 * processor address space. Note that some shortcuts 2557 * are taken, but the code works. 2558 */ 2559void 2560pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2561 vm_object_t object, vm_pindex_t pindex, 2562 vm_size_t size) 2563{ 2564 vm_page_t p; 2565 2566 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2567 KASSERT(object->type == OBJT_DEVICE, 2568 ("pmap_object_init_pt: non-device object")); 2569 if (pseflag && 2570 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2571 int i; 2572 vm_page_t m[1]; 2573 unsigned int ptepindex; 2574 int npdes; 2575 pd_entry_t ptepa; 2576 2577 PMAP_LOCK(pmap); 2578 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2579 goto out; 2580 PMAP_UNLOCK(pmap); 2581retry: 2582 p = vm_page_lookup(object, pindex); 2583 if (p != NULL) { 2584 vm_page_lock_queues(); 2585 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2586 goto retry; 2587 } else { 2588 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2589 if (p == NULL) 2590 return; 2591 m[0] = p; 2592 2593 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2594 vm_page_lock_queues(); 2595 vm_page_free(p); 2596 vm_page_unlock_queues(); 2597 return; 2598 } 2599 2600 p = vm_page_lookup(object, pindex); 2601 vm_page_lock_queues(); 2602 vm_page_wakeup(p); 2603 } 2604 vm_page_unlock_queues(); 2605 2606 ptepa = VM_PAGE_TO_PHYS(p); 2607 if (ptepa & (NBPDR - 1)) 2608 return; 2609 2610 p->valid = VM_PAGE_BITS_ALL; 2611 2612 PMAP_LOCK(pmap); 2613 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2614 npdes = size >> PDRSHIFT; 2615 for(i = 0; i < npdes; i++) { 2616 pde_store(&pmap->pm_pdir[ptepindex], 2617 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2618 ptepa += NBPDR; 2619 ptepindex += 1; 2620 } 2621 pmap_invalidate_all(pmap); 2622out: 2623 PMAP_UNLOCK(pmap); 2624 } 2625} 2626 2627/* 2628 * Routine: pmap_change_wiring 2629 * Function: Change the wiring attribute for a map/virtual-address 2630 * pair. 2631 * In/out conditions: 2632 * The mapping must already exist in the pmap. 2633 */ 2634void 2635pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired) 2636{ 2637 pt_entry_t *pte; 2638 2639 PMAP_LOCK(pmap); 2640 pte = pmap_pte(pmap, va); 2641 2642 if (wired && !pmap_pte_w(pte)) 2643 pmap->pm_stats.wired_count++; 2644 else if (!wired && pmap_pte_w(pte)) 2645 pmap->pm_stats.wired_count--; 2646 2647 /* 2648 * Wiring is not a hardware characteristic so there is no need to 2649 * invalidate TLB. 2650 */ 2651 pmap_pte_set_w(pte, wired); 2652 pmap_pte_release(pte); 2653 PMAP_UNLOCK(pmap); 2654} 2655 2656 2657 2658/* 2659 * Copy the range specified by src_addr/len 2660 * from the source map to the range dst_addr/len 2661 * in the destination map. 2662 * 2663 * This routine is only advisory and need not do anything. 2664 */ 2665 2666void 2667pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2668 vm_offset_t src_addr) 2669{ 2670 vm_offset_t addr; 2671 vm_offset_t end_addr = src_addr + len; 2672 vm_offset_t pdnxt; 2673 2674 if (dst_addr != src_addr) 2675 return; 2676 2677 if (!pmap_is_current(src_pmap)) 2678 return; 2679 2680 vm_page_lock_queues(); 2681 if (dst_pmap < src_pmap) { 2682 PMAP_LOCK(dst_pmap); 2683 PMAP_LOCK(src_pmap); 2684 } else { 2685 PMAP_LOCK(src_pmap); 2686 PMAP_LOCK(dst_pmap); 2687 } 2688 sched_pin(); 2689 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2690 pt_entry_t *src_pte, *dst_pte; 2691 vm_page_t dstmpte, srcmpte; 2692 pd_entry_t srcptepaddr; 2693 unsigned ptepindex; 2694 2695 if (addr >= UPT_MIN_ADDRESS) 2696 panic("pmap_copy: invalid to pmap_copy page tables"); 2697 2698 pdnxt = (addr + NBPDR) & ~PDRMASK; 2699 ptepindex = addr >> PDRSHIFT; 2700 2701 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2702 if (srcptepaddr == 0) 2703 continue; 2704 2705 if (srcptepaddr & PG_PS) { 2706 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2707 dst_pmap->pm_pdir[ptepindex] = srcptepaddr & 2708 ~PG_W; 2709 dst_pmap->pm_stats.resident_count += 2710 NBPDR / PAGE_SIZE; 2711 } 2712 continue; 2713 } 2714 2715 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2716 if (srcmpte->wire_count == 0) 2717 panic("pmap_copy: source page table page is unused"); 2718 2719 if (pdnxt > end_addr) 2720 pdnxt = end_addr; 2721 2722 src_pte = vtopte(addr); 2723 while (addr < pdnxt) { 2724 pt_entry_t ptetemp; 2725 ptetemp = *src_pte; 2726 /* 2727 * we only virtual copy managed pages 2728 */ 2729 if ((ptetemp & PG_MANAGED) != 0) { 2730 dstmpte = pmap_allocpte(dst_pmap, addr, 2731 M_NOWAIT); 2732 if (dstmpte == NULL) 2733 break; 2734 dst_pte = pmap_pte_quick(dst_pmap, addr); 2735 if (*dst_pte == 0 && 2736 pmap_try_insert_pv_entry(dst_pmap, addr, 2737 PHYS_TO_VM_PAGE(ptetemp & PG_FRAME))) { 2738 /* 2739 * Clear the wired, modified, and 2740 * accessed (referenced) bits 2741 * during the copy. 2742 */ 2743 *dst_pte = ptetemp & ~(PG_W | PG_M | 2744 PG_A); 2745 dst_pmap->pm_stats.resident_count++; 2746 } else 2747 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2748 if (dstmpte->wire_count >= srcmpte->wire_count) 2749 break; 2750 } 2751 addr += PAGE_SIZE; 2752 src_pte++; 2753 } 2754 } 2755 sched_unpin(); 2756 vm_page_unlock_queues(); 2757 PMAP_UNLOCK(src_pmap); 2758 PMAP_UNLOCK(dst_pmap); 2759} 2760 2761static __inline void 2762pagezero(void *page) 2763{ 2764#if defined(I686_CPU) 2765 if (cpu_class == CPUCLASS_686) { 2766#if defined(CPU_ENABLE_SSE) 2767 if (cpu_feature & CPUID_SSE2) 2768 sse2_pagezero(page); 2769 else 2770#endif 2771 i686_pagezero(page); 2772 } else 2773#endif 2774 bzero(page, PAGE_SIZE); 2775} 2776 2777/* 2778 * pmap_zero_page zeros the specified hardware page by mapping 2779 * the page into KVM and using bzero to clear its contents. 2780 */ 2781void 2782pmap_zero_page(vm_page_t m) 2783{ 2784 struct sysmaps *sysmaps; 2785 2786 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2787 mtx_lock(&sysmaps->lock); 2788 if (*sysmaps->CMAP2) 2789 panic("pmap_zero_page: CMAP2 busy"); 2790 sched_pin(); 2791 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2792 invlcaddr(sysmaps->CADDR2); 2793 pagezero(sysmaps->CADDR2); 2794 *sysmaps->CMAP2 = 0; 2795 sched_unpin(); 2796 mtx_unlock(&sysmaps->lock); 2797} 2798 2799/* 2800 * pmap_zero_page_area zeros the specified hardware page by mapping 2801 * the page into KVM and using bzero to clear its contents. 2802 * 2803 * off and size may not cover an area beyond a single hardware page. 2804 */ 2805void 2806pmap_zero_page_area(vm_page_t m, int off, int size) 2807{ 2808 struct sysmaps *sysmaps; 2809 2810 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2811 mtx_lock(&sysmaps->lock); 2812 if (*sysmaps->CMAP2) 2813 panic("pmap_zero_page: CMAP2 busy"); 2814 sched_pin(); 2815 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2816 invlcaddr(sysmaps->CADDR2); 2817 if (off == 0 && size == PAGE_SIZE) 2818 pagezero(sysmaps->CADDR2); 2819 else 2820 bzero((char *)sysmaps->CADDR2 + off, size); 2821 *sysmaps->CMAP2 = 0; 2822 sched_unpin(); 2823 mtx_unlock(&sysmaps->lock); 2824} 2825 2826/* 2827 * pmap_zero_page_idle zeros the specified hardware page by mapping 2828 * the page into KVM and using bzero to clear its contents. This 2829 * is intended to be called from the vm_pagezero process only and 2830 * outside of Giant. 2831 */ 2832void 2833pmap_zero_page_idle(vm_page_t m) 2834{ 2835 2836 if (*CMAP3) 2837 panic("pmap_zero_page: CMAP3 busy"); 2838 sched_pin(); 2839 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2840 invlcaddr(CADDR3); 2841 pagezero(CADDR3); 2842 *CMAP3 = 0; 2843 sched_unpin(); 2844} 2845 2846/* 2847 * pmap_copy_page copies the specified (machine independent) 2848 * page by mapping the page into virtual memory and using 2849 * bcopy to copy the page, one machine dependent page at a 2850 * time. 2851 */ 2852void 2853pmap_copy_page(vm_page_t src, vm_page_t dst) 2854{ 2855 struct sysmaps *sysmaps; 2856 2857 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2858 mtx_lock(&sysmaps->lock); 2859 if (*sysmaps->CMAP1) 2860 panic("pmap_copy_page: CMAP1 busy"); 2861 if (*sysmaps->CMAP2) 2862 panic("pmap_copy_page: CMAP2 busy"); 2863 sched_pin(); 2864 invlpg((u_int)sysmaps->CADDR1); 2865 invlpg((u_int)sysmaps->CADDR2); 2866 *sysmaps->CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2867 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2868 bcopy(sysmaps->CADDR1, sysmaps->CADDR2, PAGE_SIZE); 2869 *sysmaps->CMAP1 = 0; 2870 *sysmaps->CMAP2 = 0; 2871 sched_unpin(); 2872 mtx_unlock(&sysmaps->lock); 2873} 2874 2875/* 2876 * Returns true if the pmap's pv is one of the first 2877 * 16 pvs linked to from this page. This count may 2878 * be changed upwards or downwards in the future; it 2879 * is only necessary that true be returned for a small 2880 * subset of pmaps for proper page aging. 2881 */ 2882boolean_t 2883pmap_page_exists_quick(pmap_t pmap, vm_page_t m) 2884{ 2885 pv_entry_t pv; 2886 int loops = 0; 2887 2888 if (m->flags & PG_FICTITIOUS) 2889 return FALSE; 2890 2891 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2892 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2893 if (PV_PMAP(pv) == pmap) { 2894 return TRUE; 2895 } 2896 loops++; 2897 if (loops >= 16) 2898 break; 2899 } 2900 return (FALSE); 2901} 2902 2903/* 2904 * Remove all pages from specified address space 2905 * this aids process exit speeds. Also, this code 2906 * is special cased for current process only, but 2907 * can have the more generic (and slightly slower) 2908 * mode enabled. This is much faster than pmap_remove 2909 * in the case of running down an entire address space. 2910 */ 2911void 2912pmap_remove_pages(pmap_t pmap) 2913{ 2914 pt_entry_t *pte, tpte; 2915 vm_page_t m; 2916 pv_entry_t pv; 2917 struct pv_chunk *pc, *npc; 2918 int field, idx; 2919 int32_t bit; 2920 uint32_t inuse, bitmask; 2921 int allfree; 2922 2923 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) { 2924 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2925 return; 2926 } 2927 vm_page_lock_queues(); 2928 PMAP_LOCK(pmap); 2929 sched_pin(); 2930 TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) { 2931 allfree = 1; 2932 for (field = 0; field < _NPCM; field++) { 2933 inuse = (~(pc->pc_map[field])) & pc_freemask[field]; 2934 while (inuse != 0) { 2935 bit = bsfl(inuse); 2936 bitmask = 1UL << bit; 2937 idx = field * 32 + bit; 2938 pv = &pc->pc_pventry[idx]; 2939 inuse &= ~bitmask; 2940 2941 pte = vtopte(pv->pv_va); 2942 tpte = *pte; 2943 2944 if (tpte == 0) { 2945 printf( 2946 "TPTE at %p IS ZERO @ VA %08x\n", 2947 pte, pv->pv_va); 2948 panic("bad pte"); 2949 } 2950 2951/* 2952 * We cannot remove wired pages from a process' mapping at this time 2953 */ 2954 if (tpte & PG_W) { 2955 allfree = 0; 2956 continue; 2957 } 2958 2959 m = PHYS_TO_VM_PAGE(tpte); 2960 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2961 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2962 m, (uintmax_t)m->phys_addr, 2963 (uintmax_t)tpte)); 2964 2965 KASSERT(m < &vm_page_array[vm_page_array_size], 2966 ("pmap_remove_pages: bad tpte %#jx", 2967 (uintmax_t)tpte)); 2968 2969 pmap->pm_stats.resident_count--; 2970 2971 pte_clear(pte); 2972 2973 /* 2974 * Update the vm_page_t clean/reference bits. 2975 */ 2976 if (tpte & PG_M) 2977 vm_page_dirty(m); 2978 2979 /* Mark free */ 2980 PV_STAT(pv_entry_frees++); 2981 PV_STAT(pv_entry_spare++); 2982 pv_entry_count--; 2983 pc->pc_map[field] |= bitmask; 2984 m->md.pv_list_count--; 2985 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2986 if (TAILQ_EMPTY(&m->md.pv_list)) 2987 vm_page_flag_clear(m, PG_WRITEABLE); 2988 2989 pmap_unuse_pt(pmap, pv->pv_va); 2990 } 2991 } 2992 if (allfree) { 2993 PV_STAT(pv_entry_spare -= _NPCPV); 2994 PV_STAT(pc_chunk_count--); 2995 PV_STAT(pc_chunk_frees++); 2996 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); 2997 m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc)); 2998 pmap_qremove((vm_offset_t)pc, 1); 2999 vm_page_unwire(m, 0); 3000 vm_page_free(m); 3001 pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc); 3002 } 3003 } 3004 sched_unpin(); 3005 vm_page_unlock_queues(); 3006 pmap_invalidate_all(pmap); 3007 PMAP_UNLOCK(pmap); 3008} 3009 3010/* 3011 * pmap_is_modified: 3012 * 3013 * Return whether or not the specified physical page was modified 3014 * in any physical maps. 3015 */ 3016boolean_t 3017pmap_is_modified(vm_page_t m) 3018{ 3019 pv_entry_t pv; 3020 pt_entry_t *pte; 3021 pmap_t pmap; 3022 boolean_t rv; 3023 3024 rv = FALSE; 3025 if (m->flags & PG_FICTITIOUS) 3026 return (rv); 3027 3028 sched_pin(); 3029 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3030 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3031 pmap = PV_PMAP(pv); 3032 PMAP_LOCK(pmap); 3033 pte = pmap_pte_quick(pmap, pv->pv_va); 3034 rv = (*pte & PG_M) != 0; 3035 PMAP_UNLOCK(pmap); 3036 if (rv) 3037 break; 3038 } 3039 sched_unpin(); 3040 return (rv); 3041} 3042 3043/* 3044 * pmap_is_prefaultable: 3045 * 3046 * Return whether or not the specified virtual address is elgible 3047 * for prefault. 3048 */ 3049boolean_t 3050pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 3051{ 3052 pt_entry_t *pte; 3053 boolean_t rv; 3054 3055 rv = FALSE; 3056 PMAP_LOCK(pmap); 3057 if (*pmap_pde(pmap, addr)) { 3058 pte = vtopte(addr); 3059 rv = *pte == 0; 3060 } 3061 PMAP_UNLOCK(pmap); 3062 return (rv); 3063} 3064 3065/* 3066 * Clear the write and modified bits in each of the given page's mappings. 3067 */ 3068void 3069pmap_clear_write(vm_page_t m) 3070{ 3071 pv_entry_t pv; 3072 pmap_t pmap; 3073 pt_entry_t oldpte, *pte; 3074 3075 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3076 if ((m->flags & PG_FICTITIOUS) != 0 || 3077 (m->flags & PG_WRITEABLE) == 0) 3078 return; 3079 sched_pin(); 3080 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3081 pmap = PV_PMAP(pv); 3082 PMAP_LOCK(pmap); 3083 pte = pmap_pte_quick(pmap, pv->pv_va); 3084retry: 3085 oldpte = *pte; 3086 if ((oldpte & PG_RW) != 0) { 3087 /* 3088 * Regardless of whether a pte is 32 or 64 bits 3089 * in size, PG_RW and PG_M are among the least 3090 * significant 32 bits. 3091 */ 3092 if (!atomic_cmpset_int((u_int *)pte, oldpte, 3093 oldpte & ~(PG_RW | PG_M))) 3094 goto retry; 3095 if ((oldpte & PG_M) != 0) 3096 vm_page_dirty(m); 3097 pmap_invalidate_page(pmap, pv->pv_va); 3098 } 3099 PMAP_UNLOCK(pmap); 3100 } 3101 vm_page_flag_clear(m, PG_WRITEABLE); 3102 sched_unpin(); 3103} 3104 3105/* 3106 * pmap_page_protect: 3107 * 3108 * Lower the permission for all mappings to a given page. 3109 */ 3110void 3111pmap_page_protect(vm_page_t m, vm_prot_t prot) 3112{ 3113 if ((prot & VM_PROT_WRITE) == 0) { 3114 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 3115 pmap_clear_write(m); 3116 } else { 3117 pmap_remove_all(m); 3118 } 3119 } 3120} 3121 3122/* 3123 * pmap_ts_referenced: 3124 * 3125 * Return a count of reference bits for a page, clearing those bits. 3126 * It is not necessary for every reference bit to be cleared, but it 3127 * is necessary that 0 only be returned when there are truly no 3128 * reference bits set. 3129 * 3130 * XXX: The exact number of bits to check and clear is a matter that 3131 * should be tested and standardized at some point in the future for 3132 * optimal aging of shared pages. 3133 */ 3134int 3135pmap_ts_referenced(vm_page_t m) 3136{ 3137 pv_entry_t pv, pvf, pvn; 3138 pmap_t pmap; 3139 pt_entry_t *pte; 3140 int rtval = 0; 3141 3142 if (m->flags & PG_FICTITIOUS) 3143 return (rtval); 3144 sched_pin(); 3145 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3146 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 3147 pvf = pv; 3148 do { 3149 pvn = TAILQ_NEXT(pv, pv_list); 3150 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 3151 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 3152 pmap = PV_PMAP(pv); 3153 PMAP_LOCK(pmap); 3154 pte = pmap_pte_quick(pmap, pv->pv_va); 3155 if ((*pte & PG_A) != 0) { 3156 atomic_clear_int((u_int *)pte, PG_A); 3157 pmap_invalidate_page(pmap, pv->pv_va); 3158 rtval++; 3159 if (rtval > 4) 3160 pvn = NULL; 3161 } 3162 PMAP_UNLOCK(pmap); 3163 } while ((pv = pvn) != NULL && pv != pvf); 3164 } 3165 sched_unpin(); 3166 return (rtval); 3167} 3168 3169/* 3170 * Clear the modify bits on the specified physical page. 3171 */ 3172void 3173pmap_clear_modify(vm_page_t m) 3174{ 3175 pv_entry_t pv; 3176 pmap_t pmap; 3177 pt_entry_t *pte; 3178 3179 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3180 if ((m->flags & PG_FICTITIOUS) != 0) 3181 return; 3182 sched_pin(); 3183 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3184 pmap = PV_PMAP(pv); 3185 PMAP_LOCK(pmap); 3186 pte = pmap_pte_quick(pmap, pv->pv_va); 3187 if ((*pte & PG_M) != 0) { 3188 /* 3189 * Regardless of whether a pte is 32 or 64 bits 3190 * in size, PG_M is among the least significant 3191 * 32 bits. 3192 */ 3193 atomic_clear_int((u_int *)pte, PG_M); 3194 pmap_invalidate_page(pmap, pv->pv_va); 3195 } 3196 PMAP_UNLOCK(pmap); 3197 } 3198 sched_unpin(); 3199} 3200 3201/* 3202 * pmap_clear_reference: 3203 * 3204 * Clear the reference bit on the specified physical page. 3205 */ 3206void 3207pmap_clear_reference(vm_page_t m) 3208{ 3209 pv_entry_t pv; 3210 pmap_t pmap; 3211 pt_entry_t *pte; 3212 3213 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3214 if ((m->flags & PG_FICTITIOUS) != 0) 3215 return; 3216 sched_pin(); 3217 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3218 pmap = PV_PMAP(pv); 3219 PMAP_LOCK(pmap); 3220 pte = pmap_pte_quick(pmap, pv->pv_va); 3221 if ((*pte & PG_A) != 0) { 3222 /* 3223 * Regardless of whether a pte is 32 or 64 bits 3224 * in size, PG_A is among the least significant 3225 * 32 bits. 3226 */ 3227 atomic_clear_int((u_int *)pte, PG_A); 3228 pmap_invalidate_page(pmap, pv->pv_va); 3229 } 3230 PMAP_UNLOCK(pmap); 3231 } 3232 sched_unpin(); 3233} 3234 3235/* 3236 * Miscellaneous support routines follow 3237 */ 3238 3239/* 3240 * Map a set of physical memory pages into the kernel virtual 3241 * address space. Return a pointer to where it is mapped. This 3242 * routine is intended to be used for mapping device memory, 3243 * NOT real memory. 3244 */ 3245void * 3246pmap_mapdev(vm_paddr_t pa, vm_size_t size) 3247{ 3248 vm_offset_t va, tmpva, offset; 3249 3250 offset = pa & PAGE_MASK; 3251 size = roundup(offset + size, PAGE_SIZE); 3252 pa = pa & PG_FRAME; 3253 3254 if (pa < KERNLOAD && pa + size <= KERNLOAD) 3255 va = KERNBASE + pa; 3256 else 3257 va = kmem_alloc_nofault(kernel_map, size); 3258 if (!va) 3259 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 3260 3261 for (tmpva = va; size > 0; ) { 3262 pmap_kenter(tmpva, pa); 3263 size -= PAGE_SIZE; 3264 tmpva += PAGE_SIZE; 3265 pa += PAGE_SIZE; 3266 } 3267 pmap_invalidate_range(kernel_pmap, va, tmpva); 3268 return ((void *)(va + offset)); 3269} 3270 3271void 3272pmap_unmapdev(vm_offset_t va, vm_size_t size) 3273{ 3274 vm_offset_t base, offset, tmpva; 3275 3276 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 3277 return; 3278 base = va & PG_FRAME; 3279 offset = va & PAGE_MASK; 3280 size = roundup(offset + size, PAGE_SIZE); 3281 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 3282 pmap_kremove(tmpva); 3283 pmap_invalidate_range(kernel_pmap, va, tmpva); 3284 kmem_free(kernel_map, base, size); 3285} 3286 3287/* 3288 * perform the pmap work for mincore 3289 */ 3290int 3291pmap_mincore(pmap_t pmap, vm_offset_t addr) 3292{ 3293 pt_entry_t *ptep, pte; 3294 vm_page_t m; 3295 int val = 0; 3296 3297 PMAP_LOCK(pmap); 3298 ptep = pmap_pte(pmap, addr); 3299 pte = (ptep != NULL) ? *ptep : 0; 3300 pmap_pte_release(ptep); 3301 PMAP_UNLOCK(pmap); 3302 3303 if (pte != 0) { 3304 vm_paddr_t pa; 3305 3306 val = MINCORE_INCORE; 3307 if ((pte & PG_MANAGED) == 0) 3308 return val; 3309 3310 pa = pte & PG_FRAME; 3311 3312 m = PHYS_TO_VM_PAGE(pa); 3313 3314 /* 3315 * Modified by us 3316 */ 3317 if (pte & PG_M) 3318 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 3319 else { 3320 /* 3321 * Modified by someone else 3322 */ 3323 vm_page_lock_queues(); 3324 if (m->dirty || pmap_is_modified(m)) 3325 val |= MINCORE_MODIFIED_OTHER; 3326 vm_page_unlock_queues(); 3327 } 3328 /* 3329 * Referenced by us 3330 */ 3331 if (pte & PG_A) 3332 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 3333 else { 3334 /* 3335 * Referenced by someone else 3336 */ 3337 vm_page_lock_queues(); 3338 if ((m->flags & PG_REFERENCED) || 3339 pmap_ts_referenced(m)) { 3340 val |= MINCORE_REFERENCED_OTHER; 3341 vm_page_flag_set(m, PG_REFERENCED); 3342 } 3343 vm_page_unlock_queues(); 3344 } 3345 } 3346 return val; 3347} 3348 3349void 3350pmap_activate(struct thread *td) 3351{ 3352 pmap_t pmap, oldpmap; 3353 u_int32_t cr3; 3354 3355 critical_enter(); 3356 pmap = vmspace_pmap(td->td_proc->p_vmspace); 3357 oldpmap = PCPU_GET(curpmap); 3358#if defined(SMP) 3359 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 3360 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 3361#else 3362 oldpmap->pm_active &= ~1; 3363 pmap->pm_active |= 1; 3364#endif 3365#ifdef PAE 3366 cr3 = vtophys(pmap->pm_pdpt); 3367#else 3368 cr3 = vtophys(pmap->pm_pdir); 3369#endif 3370 /* 3371 * pmap_activate is for the current thread on the current cpu 3372 */ 3373 td->td_pcb->pcb_cr3 = cr3; 3374 load_cr3(cr3); 3375 PCPU_SET(curpmap, pmap); 3376 critical_exit(); 3377} 3378 3379vm_offset_t 3380pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3381{ 3382 3383 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3384 return addr; 3385 } 3386 3387 addr = (addr + PDRMASK) & ~PDRMASK; 3388 return addr; 3389} 3390 3391 3392#if defined(PMAP_DEBUG) 3393pmap_pid_dump(int pid) 3394{ 3395 pmap_t pmap; 3396 struct proc *p; 3397 int npte = 0; 3398 int index; 3399 3400 sx_slock(&allproc_lock); 3401 LIST_FOREACH(p, &allproc, p_list) { 3402 if (p->p_pid != pid) 3403 continue; 3404 3405 if (p->p_vmspace) { 3406 int i,j; 3407 index = 0; 3408 pmap = vmspace_pmap(p->p_vmspace); 3409 for (i = 0; i < NPDEPTD; i++) { 3410 pd_entry_t *pde; 3411 pt_entry_t *pte; 3412 vm_offset_t base = i << PDRSHIFT; 3413 3414 pde = &pmap->pm_pdir[i]; 3415 if (pde && pmap_pde_v(pde)) { 3416 for (j = 0; j < NPTEPG; j++) { 3417 vm_offset_t va = base + (j << PAGE_SHIFT); 3418 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3419 if (index) { 3420 index = 0; 3421 printf("\n"); 3422 } 3423 sx_sunlock(&allproc_lock); 3424 return npte; 3425 } 3426 pte = pmap_pte(pmap, va); 3427 if (pte && pmap_pte_v(pte)) { 3428 pt_entry_t pa; 3429 vm_page_t m; 3430 pa = *pte; 3431 m = PHYS_TO_VM_PAGE(pa); 3432 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3433 va, pa, m->hold_count, m->wire_count, m->flags); 3434 npte++; 3435 index++; 3436 if (index >= 2) { 3437 index = 0; 3438 printf("\n"); 3439 } else { 3440 printf(" "); 3441 } 3442 } 3443 } 3444 } 3445 } 3446 } 3447 } 3448 sx_sunlock(&allproc_lock); 3449 return npte; 3450} 3451#endif 3452 3453#if defined(DEBUG) 3454 3455static void pads(pmap_t pm); 3456void pmap_pvdump(vm_offset_t pa); 3457 3458/* print address space of pmap*/ 3459static void 3460pads(pmap_t pm) 3461{ 3462 int i, j; 3463 vm_paddr_t va; 3464 pt_entry_t *ptep; 3465 3466 if (pm == kernel_pmap) 3467 return; 3468 for (i = 0; i < NPDEPTD; i++) 3469 if (pm->pm_pdir[i]) 3470 for (j = 0; j < NPTEPG; j++) { 3471 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3472 if (pm == kernel_pmap && va < KERNBASE) 3473 continue; 3474 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3475 continue; 3476 ptep = pmap_pte(pm, va); 3477 if (pmap_pte_v(ptep)) 3478 printf("%x:%x ", va, *ptep); 3479 }; 3480 3481} 3482 3483void 3484pmap_pvdump(vm_paddr_t pa) 3485{ 3486 pv_entry_t pv; 3487 pmap_t pmap; 3488 vm_page_t m; 3489 3490 printf("pa %x", pa); 3491 m = PHYS_TO_VM_PAGE(pa); 3492 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3493 pmap = PV_PMAP(pv); 3494 printf(" -> pmap %p, va %x", (void *)pmap, pv->pv_va); 3495 pads(pmap); 3496 } 3497 printf(" "); 3498} 3499#endif 3500