pmap.c revision 135076
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 * 9 * This code is derived from software contributed to Berkeley by 10 * the Systems Programming Group of the University of Utah Computer 11 * Science Department and William Jolitz of UUNET Technologies Inc. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by the University of 24 * California, Berkeley and its contributors. 25 * 4. Neither the name of the University nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 39 * SUCH DAMAGE. 40 * 41 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 42 */ 43/*- 44 * Copyright (c) 2003 Networks Associates Technology, Inc. 45 * All rights reserved. 46 * 47 * This software was developed for the FreeBSD Project by Jake Burkholder, 48 * Safeport Network Services, and Network Associates Laboratories, the 49 * Security Research Division of Network Associates, Inc. under 50 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA 51 * CHATS research program. 52 * 53 * Redistribution and use in source and binary forms, with or without 54 * modification, are permitted provided that the following conditions 55 * are met: 56 * 1. Redistributions of source code must retain the above copyright 57 * notice, this list of conditions and the following disclaimer. 58 * 2. Redistributions in binary form must reproduce the above copyright 59 * notice, this list of conditions and the following disclaimer in the 60 * documentation and/or other materials provided with the distribution. 61 * 62 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 63 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 64 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 65 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 66 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 67 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 68 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 69 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 70 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 71 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 72 * SUCH DAMAGE. 73 */ 74 75#include <sys/cdefs.h> 76__FBSDID("$FreeBSD: head/sys/i386/i386/pmap.c 135076 2004-09-11 10:07:22Z scottl $"); 77 78/* 79 * Manages physical address maps. 80 * 81 * In addition to hardware address maps, this 82 * module is called upon to provide software-use-only 83 * maps which may or may not be stored in the same 84 * form as hardware maps. These pseudo-maps are 85 * used to store intermediate results from copy 86 * operations to and from address spaces. 87 * 88 * Since the information managed by this module is 89 * also stored by the logical address mapping module, 90 * this module may throw away valid virtual-to-physical 91 * mappings at almost any time. However, invalidations 92 * of virtual-to-physical mappings must be done as 93 * requested. 94 * 95 * In order to cope with hardware architectures which 96 * make virtual-to-physical map invalidates expensive, 97 * this module may delay invalidate or reduced protection 98 * operations until such time as they are actually 99 * necessary. This module is given full information as 100 * to which processors are currently using which maps, 101 * and to when physical maps must be made correct. 102 */ 103 104#include "opt_cpu.h" 105#include "opt_pmap.h" 106#include "opt_msgbuf.h" 107#include "opt_kstack_pages.h" 108 109#include <sys/param.h> 110#include <sys/systm.h> 111#include <sys/kernel.h> 112#include <sys/lock.h> 113#include <sys/mman.h> 114#include <sys/msgbuf.h> 115#include <sys/mutex.h> 116#include <sys/proc.h> 117#include <sys/sx.h> 118#include <sys/user.h> 119#include <sys/vmmeter.h> 120#include <sys/sched.h> 121#include <sys/sysctl.h> 122#ifdef SMP 123#include <sys/smp.h> 124#endif 125 126#include <vm/vm.h> 127#include <vm/vm_param.h> 128#include <vm/vm_kern.h> 129#include <vm/vm_page.h> 130#include <vm/vm_map.h> 131#include <vm/vm_object.h> 132#include <vm/vm_extern.h> 133#include <vm/vm_pageout.h> 134#include <vm/vm_pager.h> 135#include <vm/uma.h> 136 137#include <machine/cpu.h> 138#include <machine/cputypes.h> 139#include <machine/md_var.h> 140#include <machine/specialreg.h> 141#ifdef SMP 142#include <machine/smp.h> 143#endif 144 145#if !defined(CPU_ENABLE_SSE) && defined(I686_CPU) 146#define CPU_ENABLE_SSE 147#endif 148#if defined(CPU_DISABLE_SSE) 149#undef CPU_ENABLE_SSE 150#endif 151 152#ifndef PMAP_SHPGPERPROC 153#define PMAP_SHPGPERPROC 200 154#endif 155 156#if defined(DIAGNOSTIC) 157#define PMAP_DIAGNOSTIC 158#endif 159 160#define MINPV 2048 161 162#if !defined(PMAP_DIAGNOSTIC) 163#define PMAP_INLINE __inline 164#else 165#define PMAP_INLINE 166#endif 167 168/* 169 * Get PDEs and PTEs for user/kernel address space 170 */ 171#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 172#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 173 174#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 175#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 176#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 177#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 178#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 179 180#define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \ 181 atomic_clear_int((u_int *)(pte), PG_W)) 182#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 183 184struct pmap kernel_pmap_store; 185LIST_HEAD(pmaplist, pmap); 186static struct pmaplist allpmaps; 187static struct mtx allpmaps_lock; 188 189vm_paddr_t avail_end; /* PA of last available physical page */ 190vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 191vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 192static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 193int pgeflag = 0; /* PG_G or-in */ 194int pseflag = 0; /* PG_PS or-in */ 195 196static int nkpt; 197vm_offset_t kernel_vm_end; 198extern u_int32_t KERNend; 199 200#ifdef PAE 201static uma_zone_t pdptzone; 202#endif 203 204/* 205 * Data for the pv entry allocation mechanism 206 */ 207static uma_zone_t pvzone; 208static struct vm_object pvzone_obj; 209static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 210int pmap_pagedaemon_waken; 211 212/* 213 * All those kernel PT submaps that BSD is so fond of 214 */ 215pt_entry_t *CMAP1 = 0; 216static pt_entry_t *CMAP2, *CMAP3; 217caddr_t CADDR1 = 0, ptvmmap = 0; 218static caddr_t CADDR2, CADDR3; 219static struct mtx CMAPCADDR12_lock; 220struct msgbuf *msgbufp = 0; 221 222/* 223 * Crashdump maps. 224 */ 225static caddr_t crashdumpmap; 226 227#ifdef SMP 228extern pt_entry_t *SMPpt; 229#endif 230static pt_entry_t *PMAP1 = 0, *PMAP2; 231static pt_entry_t *PADDR1 = 0, *PADDR2; 232#ifdef SMP 233static int PMAP1cpu; 234static int PMAP1changedcpu; 235SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD, 236 &PMAP1changedcpu, 0, 237 "Number of times pmap_pte_quick changed CPU with same PMAP1"); 238#endif 239static int PMAP1changed; 240SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD, 241 &PMAP1changed, 0, 242 "Number of times pmap_pte_quick changed PMAP1"); 243static int PMAP1unchanged; 244SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD, 245 &PMAP1unchanged, 0, 246 "Number of times pmap_pte_quick didn't change PMAP1"); 247static struct mtx PMAP2mutex; 248 249static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 250static pv_entry_t get_pv_entry(void); 251static void pmap_clear_ptes(vm_page_t m, int bit); 252 253static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 254static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 255static int pmap_remove_entry(struct pmap *pmap, vm_page_t m, 256 vm_offset_t va); 257static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); 258 259static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va); 260 261static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex); 262static int _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m); 263static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 264static void pmap_pte_release(pt_entry_t *pte); 265static int pmap_unuse_pt(pmap_t, vm_offset_t); 266static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 267#ifdef PAE 268static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 269#endif 270 271CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 272CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 273 274/* 275 * Move the kernel virtual free pointer to the next 276 * 4MB. This is used to help improve performance 277 * by using a large (4MB) page for much of the kernel 278 * (.text, .data, .bss) 279 */ 280static vm_offset_t 281pmap_kmem_choose(vm_offset_t addr) 282{ 283 vm_offset_t newaddr = addr; 284 285#ifndef DISABLE_PSE 286 if (cpu_feature & CPUID_PSE) 287 newaddr = (addr + PDRMASK) & ~PDRMASK; 288#endif 289 return newaddr; 290} 291 292/* 293 * Bootstrap the system enough to run with virtual memory. 294 * 295 * On the i386 this is called after mapping has already been enabled 296 * and just syncs the pmap module with what has already been done. 297 * [We can't call it easily with mapping off since the kernel is not 298 * mapped with PA == VA, hence we would have to relocate every address 299 * from the linked base (virtual) address "KERNBASE" to the actual 300 * (physical) address starting relative to 0] 301 */ 302void 303pmap_bootstrap(firstaddr, loadaddr) 304 vm_paddr_t firstaddr; 305 vm_paddr_t loadaddr; 306{ 307 vm_offset_t va; 308 pt_entry_t *pte, *unused; 309 int i; 310 311 /* 312 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 313 * large. It should instead be correctly calculated in locore.s and 314 * not based on 'first' (which is a physical address, not a virtual 315 * address, for the start of unused physical memory). The kernel 316 * page tables are NOT double mapped and thus should not be included 317 * in this calculation. 318 */ 319 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 320 virtual_avail = pmap_kmem_choose(virtual_avail); 321 322 virtual_end = VM_MAX_KERNEL_ADDRESS; 323 324 /* 325 * Initialize the kernel pmap (which is statically allocated). 326 */ 327 PMAP_LOCK_INIT(kernel_pmap); 328 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 329#ifdef PAE 330 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 331#endif 332 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 333 TAILQ_INIT(&kernel_pmap->pm_pvlist); 334 LIST_INIT(&allpmaps); 335 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 336 mtx_lock_spin(&allpmaps_lock); 337 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 338 mtx_unlock_spin(&allpmaps_lock); 339 nkpt = NKPT; 340 341 /* 342 * Reserve some special page table entries/VA space for temporary 343 * mapping of pages. 344 */ 345#define SYSMAP(c, p, v, n) \ 346 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 347 348 va = virtual_avail; 349 pte = vtopte(va); 350 351 /* 352 * CMAP1/CMAP2 are used for zeroing and copying pages. 353 * CMAP3 is used for the idle process page zeroing. 354 */ 355 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 356 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 357 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 358 *CMAP3 = 0; 359 360 mtx_init(&CMAPCADDR12_lock, "CMAPCADDR12", NULL, MTX_DEF); 361 362 /* 363 * Crashdump maps. 364 */ 365 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS) 366 367 /* 368 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 369 */ 370 SYSMAP(caddr_t, unused, ptvmmap, 1) 371 372 /* 373 * msgbufp is used to map the system message buffer. 374 */ 375 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE))) 376 377 /* 378 * ptemap is used for pmap_pte_quick 379 */ 380 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 381 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1); 382 383 mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF); 384 385 virtual_avail = va; 386 387 *CMAP1 = *CMAP2 = 0; 388 for (i = 0; i < NKPT; i++) 389 PTD[i] = 0; 390 391 /* Turn on PG_G on kernel page(s) */ 392 pmap_set_pg(); 393} 394 395/* 396 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 397 */ 398void 399pmap_set_pg(void) 400{ 401 pd_entry_t pdir; 402 pt_entry_t *pte; 403 vm_offset_t va, endva; 404 int i; 405 406 if (pgeflag == 0) 407 return; 408 409 i = KERNLOAD/NBPDR; 410 endva = KERNBASE + KERNend; 411 412 if (pseflag) { 413 va = KERNBASE + KERNLOAD; 414 while (va < endva) { 415 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 416 pdir |= pgeflag; 417 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 418 invltlb(); /* Play it safe, invltlb() every time */ 419 i++; 420 va += NBPDR; 421 } 422 } else { 423 va = (vm_offset_t)btext; 424 while (va < endva) { 425 pte = vtopte(va); 426 if (*pte) 427 *pte |= pgeflag; 428 invltlb(); /* Play it safe, invltlb() every time */ 429 va += PAGE_SIZE; 430 } 431 } 432} 433 434#ifdef PAE 435 436static MALLOC_DEFINE(M_PMAPPDPT, "pmap", "pmap pdpt"); 437 438static void * 439pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 440{ 441 *flags = UMA_SLAB_PRIV; 442 return (contigmalloc(PAGE_SIZE, M_PMAPPDPT, 0, 0x0ULL, 0xffffffffULL, 443 1, 0)); 444} 445#endif 446 447/* 448 * Initialize the pmap module. 449 * Called by vm_init, to initialize any structures that the pmap 450 * system needs to map virtual memory. 451 * pmap_init has been enhanced to support in a fairly consistant 452 * way, discontiguous physical memory. 453 */ 454void 455pmap_init(void) 456{ 457 int i; 458 459 /* 460 * Allocate memory for random pmap data structures. Includes the 461 * pv_head_table. 462 */ 463 464 for(i = 0; i < vm_page_array_size; i++) { 465 vm_page_t m; 466 467 m = &vm_page_array[i]; 468 TAILQ_INIT(&m->md.pv_list); 469 m->md.pv_list_count = 0; 470 } 471 472 /* 473 * init the pv free list 474 */ 475 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL, 476 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 477 uma_prealloc(pvzone, MINPV); 478 479#ifdef PAE 480 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 481 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 482 UMA_ZONE_VM | UMA_ZONE_NOFREE); 483 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 484#endif 485 486 /* 487 * Now it is safe to enable pv_table recording. 488 */ 489 pmap_initialized = TRUE; 490} 491 492/* 493 * Initialize the address space (zone) for the pv_entries. Set a 494 * high water mark so that the system can recover from excessive 495 * numbers of pv entries. 496 */ 497void 498pmap_init2() 499{ 500 int shpgperproc = PMAP_SHPGPERPROC; 501 502 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 503 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 504 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 505 pv_entry_high_water = 9 * (pv_entry_max / 10); 506 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 507} 508 509 510/*************************************************** 511 * Low level helper routines..... 512 ***************************************************/ 513 514#if defined(PMAP_DIAGNOSTIC) 515 516/* 517 * This code checks for non-writeable/modified pages. 518 * This should be an invalid condition. 519 */ 520static int 521pmap_nw_modified(pt_entry_t ptea) 522{ 523 int pte; 524 525 pte = (int) ptea; 526 527 if ((pte & (PG_M|PG_RW)) == PG_M) 528 return 1; 529 else 530 return 0; 531} 532#endif 533 534 535/* 536 * this routine defines the region(s) of memory that should 537 * not be tested for the modified bit. 538 */ 539static PMAP_INLINE int 540pmap_track_modified(vm_offset_t va) 541{ 542 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 543 return 1; 544 else 545 return 0; 546} 547 548#ifdef I386_CPU 549/* 550 * i386 only has "invalidate everything" and no SMP to worry about. 551 */ 552PMAP_INLINE void 553pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 554{ 555 556 if (pmap == kernel_pmap || pmap->pm_active) 557 invltlb(); 558} 559 560PMAP_INLINE void 561pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 562{ 563 564 if (pmap == kernel_pmap || pmap->pm_active) 565 invltlb(); 566} 567 568PMAP_INLINE void 569pmap_invalidate_all(pmap_t pmap) 570{ 571 572 if (pmap == kernel_pmap || pmap->pm_active) 573 invltlb(); 574} 575#else /* !I386_CPU */ 576#ifdef SMP 577/* 578 * For SMP, these functions have to use the IPI mechanism for coherence. 579 */ 580void 581pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 582{ 583 u_int cpumask; 584 u_int other_cpus; 585 586 if (smp_started) { 587 if (!(read_eflags() & PSL_I)) 588 panic("%s: interrupts disabled", __func__); 589 mtx_lock_spin(&smp_ipi_mtx); 590 } else 591 critical_enter(); 592 /* 593 * We need to disable interrupt preemption but MUST NOT have 594 * interrupts disabled here. 595 * XXX we may need to hold schedlock to get a coherent pm_active 596 * XXX critical sections disable interrupts again 597 */ 598 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 599 invlpg(va); 600 smp_invlpg(va); 601 } else { 602 cpumask = PCPU_GET(cpumask); 603 other_cpus = PCPU_GET(other_cpus); 604 if (pmap->pm_active & cpumask) 605 invlpg(va); 606 if (pmap->pm_active & other_cpus) 607 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 608 } 609 if (smp_started) 610 mtx_unlock_spin(&smp_ipi_mtx); 611 else 612 critical_exit(); 613} 614 615void 616pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 617{ 618 u_int cpumask; 619 u_int other_cpus; 620 vm_offset_t addr; 621 622 if (smp_started) { 623 if (!(read_eflags() & PSL_I)) 624 panic("%s: interrupts disabled", __func__); 625 mtx_lock_spin(&smp_ipi_mtx); 626 } else 627 critical_enter(); 628 /* 629 * We need to disable interrupt preemption but MUST NOT have 630 * interrupts disabled here. 631 * XXX we may need to hold schedlock to get a coherent pm_active 632 * XXX critical sections disable interrupts again 633 */ 634 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 635 for (addr = sva; addr < eva; addr += PAGE_SIZE) 636 invlpg(addr); 637 smp_invlpg_range(sva, eva); 638 } else { 639 cpumask = PCPU_GET(cpumask); 640 other_cpus = PCPU_GET(other_cpus); 641 if (pmap->pm_active & cpumask) 642 for (addr = sva; addr < eva; addr += PAGE_SIZE) 643 invlpg(addr); 644 if (pmap->pm_active & other_cpus) 645 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 646 sva, eva); 647 } 648 if (smp_started) 649 mtx_unlock_spin(&smp_ipi_mtx); 650 else 651 critical_exit(); 652} 653 654void 655pmap_invalidate_all(pmap_t pmap) 656{ 657 u_int cpumask; 658 u_int other_cpus; 659 660 if (smp_started) { 661 if (!(read_eflags() & PSL_I)) 662 panic("%s: interrupts disabled", __func__); 663 mtx_lock_spin(&smp_ipi_mtx); 664 } else 665 critical_enter(); 666 /* 667 * We need to disable interrupt preemption but MUST NOT have 668 * interrupts disabled here. 669 * XXX we may need to hold schedlock to get a coherent pm_active 670 * XXX critical sections disable interrupts again 671 */ 672 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 673 invltlb(); 674 smp_invltlb(); 675 } else { 676 cpumask = PCPU_GET(cpumask); 677 other_cpus = PCPU_GET(other_cpus); 678 if (pmap->pm_active & cpumask) 679 invltlb(); 680 if (pmap->pm_active & other_cpus) 681 smp_masked_invltlb(pmap->pm_active & other_cpus); 682 } 683 if (smp_started) 684 mtx_unlock_spin(&smp_ipi_mtx); 685 else 686 critical_exit(); 687} 688#else /* !SMP */ 689/* 690 * Normal, non-SMP, 486+ invalidation functions. 691 * We inline these within pmap.c for speed. 692 */ 693PMAP_INLINE void 694pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 695{ 696 697 if (pmap == kernel_pmap || pmap->pm_active) 698 invlpg(va); 699} 700 701PMAP_INLINE void 702pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 703{ 704 vm_offset_t addr; 705 706 if (pmap == kernel_pmap || pmap->pm_active) 707 for (addr = sva; addr < eva; addr += PAGE_SIZE) 708 invlpg(addr); 709} 710 711PMAP_INLINE void 712pmap_invalidate_all(pmap_t pmap) 713{ 714 715 if (pmap == kernel_pmap || pmap->pm_active) 716 invltlb(); 717} 718#endif /* !SMP */ 719#endif /* !I386_CPU */ 720 721/* 722 * Are we current address space or kernel? N.B. We return FALSE when 723 * a pmap's page table is in use because a kernel thread is borrowing 724 * it. The borrowed page table can change spontaneously, making any 725 * dependence on its continued use subject to a race condition. 726 */ 727static __inline int 728pmap_is_current(pmap_t pmap) 729{ 730 731 return (pmap == kernel_pmap || 732 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 733 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 734} 735 736/* 737 * If the given pmap is not the current or kernel pmap, the returned pte must 738 * be released by passing it to pmap_pte_release(). 739 */ 740pt_entry_t * 741pmap_pte(pmap_t pmap, vm_offset_t va) 742{ 743 pd_entry_t newpf; 744 pd_entry_t *pde; 745 746 pde = pmap_pde(pmap, va); 747 if (*pde & PG_PS) 748 return (pde); 749 if (*pde != 0) { 750 /* are we current address space or kernel? */ 751 if (pmap_is_current(pmap)) 752 return (vtopte(va)); 753 mtx_lock(&PMAP2mutex); 754 newpf = *pde & PG_FRAME; 755 if ((*PMAP2 & PG_FRAME) != newpf) { 756 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; 757 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 758 } 759 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); 760 } 761 return (0); 762} 763 764/* 765 * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte 766 * being NULL. 767 */ 768static __inline void 769pmap_pte_release(pt_entry_t *pte) 770{ 771 772 if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2) 773 mtx_unlock(&PMAP2mutex); 774} 775 776static __inline void 777invlcaddr(void *caddr) 778{ 779#ifdef I386_CPU 780 invltlb(); 781#else 782 invlpg((u_int)caddr); 783#endif 784} 785 786/* 787 * Super fast pmap_pte routine best used when scanning 788 * the pv lists. This eliminates many coarse-grained 789 * invltlb calls. Note that many of the pv list 790 * scans are across different pmaps. It is very wasteful 791 * to do an entire invltlb for checking a single mapping. 792 * 793 * If the given pmap is not the current pmap, vm_page_queue_mtx 794 * must be held and curthread pinned to a CPU. 795 */ 796static pt_entry_t * 797pmap_pte_quick(pmap_t pmap, vm_offset_t va) 798{ 799 pd_entry_t newpf; 800 pd_entry_t *pde; 801 802 pde = pmap_pde(pmap, va); 803 if (*pde & PG_PS) 804 return (pde); 805 if (*pde != 0) { 806 /* are we current address space or kernel? */ 807 if (pmap_is_current(pmap)) 808 return (vtopte(va)); 809 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 810 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 811 newpf = *pde & PG_FRAME; 812 if ((*PMAP1 & PG_FRAME) != newpf) { 813 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; 814#ifdef SMP 815 PMAP1cpu = PCPU_GET(cpuid); 816#endif 817 invlcaddr(PADDR1); 818 PMAP1changed++; 819 } else 820#ifdef SMP 821 if (PMAP1cpu != PCPU_GET(cpuid)) { 822 PMAP1cpu = PCPU_GET(cpuid); 823 invlcaddr(PADDR1); 824 PMAP1changedcpu++; 825 } else 826#endif 827 PMAP1unchanged++; 828 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); 829 } 830 return (0); 831} 832 833/* 834 * Routine: pmap_extract 835 * Function: 836 * Extract the physical page address associated 837 * with the given map/virtual_address pair. 838 */ 839vm_paddr_t 840pmap_extract(pmap_t pmap, vm_offset_t va) 841{ 842 vm_paddr_t rtval; 843 pt_entry_t *pte; 844 pd_entry_t pde; 845 846 rtval = 0; 847 PMAP_LOCK(pmap); 848 pde = pmap->pm_pdir[va >> PDRSHIFT]; 849 if (pde != 0) { 850 if ((pde & PG_PS) != 0) { 851 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 852 PMAP_UNLOCK(pmap); 853 return rtval; 854 } 855 pte = pmap_pte(pmap, va); 856 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK); 857 pmap_pte_release(pte); 858 } 859 PMAP_UNLOCK(pmap); 860 return (rtval); 861} 862 863/* 864 * Routine: pmap_extract_and_hold 865 * Function: 866 * Atomically extract and hold the physical page 867 * with the given pmap and virtual address pair 868 * if that mapping permits the given protection. 869 */ 870vm_page_t 871pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 872{ 873 pd_entry_t pde; 874 pt_entry_t pte; 875 vm_page_t m; 876 877 m = NULL; 878 vm_page_lock_queues(); 879 PMAP_LOCK(pmap); 880 pde = *pmap_pde(pmap, va); 881 if (pde != 0) { 882 if (pde & PG_PS) { 883 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) { 884 m = PHYS_TO_VM_PAGE((pde & ~PDRMASK) | 885 (va & PDRMASK)); 886 vm_page_hold(m); 887 } 888 } else { 889 sched_pin(); 890 pte = *pmap_pte_quick(pmap, va); 891 if (pte != 0 && 892 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) { 893 m = PHYS_TO_VM_PAGE(pte & PG_FRAME); 894 vm_page_hold(m); 895 } 896 sched_unpin(); 897 } 898 } 899 vm_page_unlock_queues(); 900 PMAP_UNLOCK(pmap); 901 return (m); 902} 903 904/*************************************************** 905 * Low level mapping routines..... 906 ***************************************************/ 907 908/* 909 * Add a wired page to the kva. 910 * Note: not SMP coherent. 911 */ 912PMAP_INLINE void 913pmap_kenter(vm_offset_t va, vm_paddr_t pa) 914{ 915 pt_entry_t *pte; 916 917 pte = vtopte(va); 918 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 919} 920 921/* 922 * Remove a page from the kernel pagetables. 923 * Note: not SMP coherent. 924 */ 925PMAP_INLINE void 926pmap_kremove(vm_offset_t va) 927{ 928 pt_entry_t *pte; 929 930 pte = vtopte(va); 931 pte_clear(pte); 932} 933 934/* 935 * Used to map a range of physical addresses into kernel 936 * virtual address space. 937 * 938 * The value passed in '*virt' is a suggested virtual address for 939 * the mapping. Architectures which can support a direct-mapped 940 * physical to virtual region can return the appropriate address 941 * within that region, leaving '*virt' unchanged. Other 942 * architectures should map the pages starting at '*virt' and 943 * update '*virt' with the first usable address after the mapped 944 * region. 945 */ 946vm_offset_t 947pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 948{ 949 vm_offset_t va, sva; 950 951 va = sva = *virt; 952 while (start < end) { 953 pmap_kenter(va, start); 954 va += PAGE_SIZE; 955 start += PAGE_SIZE; 956 } 957 pmap_invalidate_range(kernel_pmap, sva, va); 958 *virt = va; 959 return (sva); 960} 961 962 963/* 964 * Add a list of wired pages to the kva 965 * this routine is only used for temporary 966 * kernel mappings that do not need to have 967 * page modification or references recorded. 968 * Note that old mappings are simply written 969 * over. The page *must* be wired. 970 * Note: SMP coherent. Uses a ranged shootdown IPI. 971 */ 972void 973pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 974{ 975 vm_offset_t va; 976 977 va = sva; 978 while (count-- > 0) { 979 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 980 va += PAGE_SIZE; 981 m++; 982 } 983 pmap_invalidate_range(kernel_pmap, sva, va); 984} 985 986/* 987 * This routine tears out page mappings from the 988 * kernel -- it is meant only for temporary mappings. 989 * Note: SMP coherent. Uses a ranged shootdown IPI. 990 */ 991void 992pmap_qremove(vm_offset_t sva, int count) 993{ 994 vm_offset_t va; 995 996 va = sva; 997 while (count-- > 0) { 998 pmap_kremove(va); 999 va += PAGE_SIZE; 1000 } 1001 pmap_invalidate_range(kernel_pmap, sva, va); 1002} 1003 1004/*************************************************** 1005 * Page table page management routines..... 1006 ***************************************************/ 1007 1008/* 1009 * This routine unholds page table pages, and if the hold count 1010 * drops to zero, then it decrements the wire count. 1011 */ 1012static PMAP_INLINE int 1013pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1014{ 1015 1016 vm_page_unhold(m); 1017 if (m->hold_count == 0) 1018 return _pmap_unwire_pte_hold(pmap, m); 1019 else 1020 return 0; 1021} 1022 1023static int 1024_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1025{ 1026 vm_offset_t pteva; 1027 1028 /* 1029 * unmap the page table page 1030 */ 1031 pmap->pm_pdir[m->pindex] = 0; 1032 --pmap->pm_stats.resident_count; 1033 /* 1034 * We never unwire a kernel page table page, making a 1035 * check for the kernel_pmap unnecessary. 1036 */ 1037 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)) { 1038 /* 1039 * Do an invltlb to make the invalidated mapping 1040 * take effect immediately. 1041 */ 1042 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1043 pmap_invalidate_page(pmap, pteva); 1044 } 1045 1046 /* 1047 * If the page is finally unwired, simply free it. 1048 */ 1049 --m->wire_count; 1050 if (m->wire_count == 0) { 1051 vm_page_free_zero(m); 1052 atomic_subtract_int(&cnt.v_wire_count, 1); 1053 } 1054 return 1; 1055} 1056 1057/* 1058 * After removing a page table entry, this routine is used to 1059 * conditionally free the page, and manage the hold/wire counts. 1060 */ 1061static int 1062pmap_unuse_pt(pmap_t pmap, vm_offset_t va) 1063{ 1064 pd_entry_t ptepde; 1065 vm_page_t mpte; 1066 1067 if (va >= VM_MAXUSER_ADDRESS) 1068 return 0; 1069 ptepde = *pmap_pde(pmap, va); 1070 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); 1071 return pmap_unwire_pte_hold(pmap, mpte); 1072} 1073 1074void 1075pmap_pinit0(pmap) 1076 struct pmap *pmap; 1077{ 1078 1079 PMAP_LOCK_INIT(pmap); 1080 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1081#ifdef PAE 1082 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1083#endif 1084 pmap->pm_active = 0; 1085 PCPU_SET(curpmap, pmap); 1086 TAILQ_INIT(&pmap->pm_pvlist); 1087 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1088 mtx_lock_spin(&allpmaps_lock); 1089 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1090 mtx_unlock_spin(&allpmaps_lock); 1091} 1092 1093/* 1094 * Initialize a preallocated and zeroed pmap structure, 1095 * such as one in a vmspace structure. 1096 */ 1097void 1098pmap_pinit(pmap) 1099 register struct pmap *pmap; 1100{ 1101 vm_page_t m, ptdpg[NPGPTD]; 1102 vm_paddr_t pa; 1103 static int color; 1104 int i; 1105 1106 PMAP_LOCK_INIT(pmap); 1107 1108 /* 1109 * No need to allocate page table space yet but we do need a valid 1110 * page directory table. 1111 */ 1112 if (pmap->pm_pdir == NULL) { 1113 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1114 NBPTD); 1115#ifdef PAE 1116 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1117 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1118 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1119 ("pmap_pinit: pdpt misaligned")); 1120 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1121 ("pmap_pinit: pdpt above 4g")); 1122#endif 1123 } 1124 1125 /* 1126 * allocate the page directory page(s) 1127 */ 1128 for (i = 0; i < NPGPTD;) { 1129 m = vm_page_alloc(NULL, color++, 1130 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1131 VM_ALLOC_ZERO); 1132 if (m == NULL) 1133 VM_WAIT; 1134 else { 1135 ptdpg[i++] = m; 1136 } 1137 } 1138 1139 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1140 1141 for (i = 0; i < NPGPTD; i++) { 1142 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1143 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1144 } 1145 1146 mtx_lock_spin(&allpmaps_lock); 1147 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1148 mtx_unlock_spin(&allpmaps_lock); 1149 /* Wire in kernel global address entries. */ 1150 /* XXX copies current process, does not fill in MPPTDI */ 1151 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1152#ifdef SMP 1153 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1154#endif 1155 1156 /* install self-referential address mapping entry(s) */ 1157 for (i = 0; i < NPGPTD; i++) { 1158 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1159 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1160#ifdef PAE 1161 pmap->pm_pdpt[i] = pa | PG_V; 1162#endif 1163 } 1164 1165 pmap->pm_active = 0; 1166 TAILQ_INIT(&pmap->pm_pvlist); 1167 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1168} 1169 1170/* 1171 * this routine is called if the page table page is not 1172 * mapped correctly. 1173 */ 1174static vm_page_t 1175_pmap_allocpte(pmap, ptepindex) 1176 pmap_t pmap; 1177 unsigned ptepindex; 1178{ 1179 vm_paddr_t ptepa; 1180 vm_page_t m; 1181 1182 /* 1183 * Allocate a page table page. 1184 */ 1185 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1186 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1187 PMAP_UNLOCK(pmap); 1188 vm_page_unlock_queues(); 1189 VM_WAIT; 1190 vm_page_lock_queues(); 1191 PMAP_LOCK(pmap); 1192 1193 /* 1194 * Indicate the need to retry. While waiting, the page table 1195 * page may have been allocated. 1196 */ 1197 return (NULL); 1198 } 1199 if ((m->flags & PG_ZERO) == 0) 1200 pmap_zero_page(m); 1201 1202 KASSERT(m->queue == PQ_NONE, 1203 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1204 1205 /* 1206 * Increment the hold count for the page table page 1207 * (denoting a new mapping.) 1208 */ 1209 m->hold_count++; 1210 1211 /* 1212 * Map the pagetable page into the process address space, if 1213 * it isn't already there. 1214 */ 1215 1216 pmap->pm_stats.resident_count++; 1217 1218 ptepa = VM_PAGE_TO_PHYS(m); 1219 pmap->pm_pdir[ptepindex] = 1220 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1221 1222 return m; 1223} 1224 1225static vm_page_t 1226pmap_allocpte(pmap_t pmap, vm_offset_t va) 1227{ 1228 unsigned ptepindex; 1229 pd_entry_t ptepa; 1230 vm_page_t m; 1231 1232 /* 1233 * Calculate pagetable page index 1234 */ 1235 ptepindex = va >> PDRSHIFT; 1236retry: 1237 /* 1238 * Get the page directory entry 1239 */ 1240 ptepa = pmap->pm_pdir[ptepindex]; 1241 1242 /* 1243 * This supports switching from a 4MB page to a 1244 * normal 4K page. 1245 */ 1246 if (ptepa & PG_PS) { 1247 pmap->pm_pdir[ptepindex] = 0; 1248 ptepa = 0; 1249 pmap_invalidate_all(kernel_pmap); 1250 } 1251 1252 /* 1253 * If the page table page is mapped, we just increment the 1254 * hold count, and activate it. 1255 */ 1256 if (ptepa) { 1257 m = PHYS_TO_VM_PAGE(ptepa); 1258 m->hold_count++; 1259 } else { 1260 /* 1261 * Here if the pte page isn't mapped, or if it has 1262 * been deallocated. 1263 */ 1264 m = _pmap_allocpte(pmap, ptepindex); 1265 if (m == NULL) 1266 goto retry; 1267 } 1268 return (m); 1269} 1270 1271 1272/*************************************************** 1273* Pmap allocation/deallocation routines. 1274 ***************************************************/ 1275 1276#ifdef SMP 1277/* 1278 * Deal with a SMP shootdown of other users of the pmap that we are 1279 * trying to dispose of. This can be a bit hairy. 1280 */ 1281static u_int *lazymask; 1282static u_int lazyptd; 1283static volatile u_int lazywait; 1284 1285void pmap_lazyfix_action(void); 1286 1287void 1288pmap_lazyfix_action(void) 1289{ 1290 u_int mymask = PCPU_GET(cpumask); 1291 1292 if (rcr3() == lazyptd) 1293 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1294 atomic_clear_int(lazymask, mymask); 1295 atomic_store_rel_int(&lazywait, 1); 1296} 1297 1298static void 1299pmap_lazyfix_self(u_int mymask) 1300{ 1301 1302 if (rcr3() == lazyptd) 1303 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1304 atomic_clear_int(lazymask, mymask); 1305} 1306 1307 1308static void 1309pmap_lazyfix(pmap_t pmap) 1310{ 1311 u_int mymask = PCPU_GET(cpumask); 1312 u_int mask; 1313 register u_int spins; 1314 1315 while ((mask = pmap->pm_active) != 0) { 1316 spins = 50000000; 1317 mask = mask & -mask; /* Find least significant set bit */ 1318 mtx_lock_spin(&smp_ipi_mtx); 1319#ifdef PAE 1320 lazyptd = vtophys(pmap->pm_pdpt); 1321#else 1322 lazyptd = vtophys(pmap->pm_pdir); 1323#endif 1324 if (mask == mymask) { 1325 lazymask = &pmap->pm_active; 1326 pmap_lazyfix_self(mymask); 1327 } else { 1328 atomic_store_rel_int((u_int *)&lazymask, 1329 (u_int)&pmap->pm_active); 1330 atomic_store_rel_int(&lazywait, 0); 1331 ipi_selected(mask, IPI_LAZYPMAP); 1332 while (lazywait == 0) { 1333 ia32_pause(); 1334 if (--spins == 0) 1335 break; 1336 } 1337 } 1338 mtx_unlock_spin(&smp_ipi_mtx); 1339 if (spins == 0) 1340 printf("pmap_lazyfix: spun for 50000000\n"); 1341 } 1342} 1343 1344#else /* SMP */ 1345 1346/* 1347 * Cleaning up on uniprocessor is easy. For various reasons, we're 1348 * unlikely to have to even execute this code, including the fact 1349 * that the cleanup is deferred until the parent does a wait(2), which 1350 * means that another userland process has run. 1351 */ 1352static void 1353pmap_lazyfix(pmap_t pmap) 1354{ 1355 u_int cr3; 1356 1357 cr3 = vtophys(pmap->pm_pdir); 1358 if (cr3 == rcr3()) { 1359 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1360 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1361 } 1362} 1363#endif /* SMP */ 1364 1365/* 1366 * Release any resources held by the given physical map. 1367 * Called when a pmap initialized by pmap_pinit is being released. 1368 * Should only be called if the map contains no valid mappings. 1369 */ 1370void 1371pmap_release(pmap_t pmap) 1372{ 1373 vm_page_t m, ptdpg[NPGPTD]; 1374 int i; 1375 1376 KASSERT(pmap->pm_stats.resident_count == 0, 1377 ("pmap_release: pmap resident count %ld != 0", 1378 pmap->pm_stats.resident_count)); 1379 1380 pmap_lazyfix(pmap); 1381 mtx_lock_spin(&allpmaps_lock); 1382 LIST_REMOVE(pmap, pm_list); 1383 mtx_unlock_spin(&allpmaps_lock); 1384 1385 for (i = 0; i < NPGPTD; i++) 1386 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1387 1388 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1389 sizeof(*pmap->pm_pdir)); 1390#ifdef SMP 1391 pmap->pm_pdir[MPPTDI] = 0; 1392#endif 1393 1394 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1395 1396 vm_page_lock_queues(); 1397 for (i = 0; i < NPGPTD; i++) { 1398 m = ptdpg[i]; 1399#ifdef PAE 1400 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1401 ("pmap_release: got wrong ptd page")); 1402#endif 1403 m->wire_count--; 1404 atomic_subtract_int(&cnt.v_wire_count, 1); 1405 vm_page_free_zero(m); 1406 } 1407 vm_page_unlock_queues(); 1408 PMAP_LOCK_DESTROY(pmap); 1409} 1410 1411static int 1412kvm_size(SYSCTL_HANDLER_ARGS) 1413{ 1414 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1415 1416 return sysctl_handle_long(oidp, &ksize, 0, req); 1417} 1418SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1419 0, 0, kvm_size, "IU", "Size of KVM"); 1420 1421static int 1422kvm_free(SYSCTL_HANDLER_ARGS) 1423{ 1424 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1425 1426 return sysctl_handle_long(oidp, &kfree, 0, req); 1427} 1428SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1429 0, 0, kvm_free, "IU", "Amount of KVM free"); 1430 1431/* 1432 * grow the number of kernel page table entries, if needed 1433 */ 1434void 1435pmap_growkernel(vm_offset_t addr) 1436{ 1437 struct pmap *pmap; 1438 vm_paddr_t ptppaddr; 1439 vm_page_t nkpg; 1440 pd_entry_t newpdir; 1441 pt_entry_t *pde; 1442 1443 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1444 if (kernel_vm_end == 0) { 1445 kernel_vm_end = KERNBASE; 1446 nkpt = 0; 1447 while (pdir_pde(PTD, kernel_vm_end)) { 1448 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1449 nkpt++; 1450 } 1451 } 1452 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1453 while (kernel_vm_end < addr) { 1454 if (pdir_pde(PTD, kernel_vm_end)) { 1455 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1456 continue; 1457 } 1458 1459 /* 1460 * This index is bogus, but out of the way 1461 */ 1462 nkpg = vm_page_alloc(NULL, nkpt, 1463 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1464 if (!nkpg) 1465 panic("pmap_growkernel: no memory to grow kernel"); 1466 1467 nkpt++; 1468 1469 pmap_zero_page(nkpg); 1470 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1471 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1472 pdir_pde(PTD, kernel_vm_end) = newpdir; 1473 1474 mtx_lock_spin(&allpmaps_lock); 1475 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1476 pde = pmap_pde(pmap, kernel_vm_end); 1477 pde_store(pde, newpdir); 1478 } 1479 mtx_unlock_spin(&allpmaps_lock); 1480 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1481 } 1482} 1483 1484 1485/*************************************************** 1486 * page management routines. 1487 ***************************************************/ 1488 1489/* 1490 * free the pv_entry back to the free list 1491 */ 1492static PMAP_INLINE void 1493free_pv_entry(pv_entry_t pv) 1494{ 1495 pv_entry_count--; 1496 uma_zfree(pvzone, pv); 1497} 1498 1499/* 1500 * get a new pv_entry, allocating a block from the system 1501 * when needed. 1502 * the memory allocation is performed bypassing the malloc code 1503 * because of the possibility of allocations at interrupt time. 1504 */ 1505static pv_entry_t 1506get_pv_entry(void) 1507{ 1508 pv_entry_count++; 1509 if (pv_entry_high_water && 1510 (pv_entry_count > pv_entry_high_water) && 1511 (pmap_pagedaemon_waken == 0)) { 1512 pmap_pagedaemon_waken = 1; 1513 wakeup (&vm_pages_needed); 1514 } 1515 return uma_zalloc(pvzone, M_NOWAIT); 1516} 1517 1518 1519static int 1520pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1521{ 1522 pv_entry_t pv; 1523 int rtval; 1524 1525 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1526 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1527 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1528 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1529 if (pmap == pv->pv_pmap && va == pv->pv_va) 1530 break; 1531 } 1532 } else { 1533 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1534 if (va == pv->pv_va) 1535 break; 1536 } 1537 } 1538 1539 rtval = 0; 1540 if (pv) { 1541 rtval = pmap_unuse_pt(pmap, va); 1542 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1543 m->md.pv_list_count--; 1544 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1545 vm_page_flag_clear(m, PG_WRITEABLE); 1546 1547 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1548 free_pv_entry(pv); 1549 } 1550 1551 return rtval; 1552} 1553 1554/* 1555 * Create a pv entry for page at pa for 1556 * (pmap, va). 1557 */ 1558static void 1559pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) 1560{ 1561 pv_entry_t pv; 1562 1563 pv = get_pv_entry(); 1564 pv->pv_va = va; 1565 pv->pv_pmap = pmap; 1566 1567 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1568 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1569 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1570 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1571 m->md.pv_list_count++; 1572} 1573 1574/* 1575 * pmap_remove_pte: do the things to unmap a page in a process 1576 */ 1577static int 1578pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1579{ 1580 pt_entry_t oldpte; 1581 vm_page_t m; 1582 1583 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1584 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1585 oldpte = pte_load_clear(ptq); 1586 if (oldpte & PG_W) 1587 pmap->pm_stats.wired_count -= 1; 1588 /* 1589 * Machines that don't support invlpg, also don't support 1590 * PG_G. 1591 */ 1592 if (oldpte & PG_G) 1593 pmap_invalidate_page(kernel_pmap, va); 1594 pmap->pm_stats.resident_count -= 1; 1595 if (oldpte & PG_MANAGED) { 1596 m = PHYS_TO_VM_PAGE(oldpte); 1597 if (oldpte & PG_M) { 1598#if defined(PMAP_DIAGNOSTIC) 1599 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1600 printf( 1601 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1602 va, oldpte); 1603 } 1604#endif 1605 if (pmap_track_modified(va)) 1606 vm_page_dirty(m); 1607 } 1608 if (oldpte & PG_A) 1609 vm_page_flag_set(m, PG_REFERENCED); 1610 return pmap_remove_entry(pmap, m, va); 1611 } else { 1612 return pmap_unuse_pt(pmap, va); 1613 } 1614} 1615 1616/* 1617 * Remove a single page from a process address space 1618 */ 1619static void 1620pmap_remove_page(pmap_t pmap, vm_offset_t va) 1621{ 1622 pt_entry_t *pte; 1623 1624 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1625 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 1626 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1627 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) 1628 return; 1629 pmap_remove_pte(pmap, pte, va); 1630 pmap_invalidate_page(pmap, va); 1631} 1632 1633/* 1634 * Remove the given range of addresses from the specified map. 1635 * 1636 * It is assumed that the start and end are properly 1637 * rounded to the page size. 1638 */ 1639void 1640pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1641{ 1642 vm_offset_t pdnxt; 1643 pd_entry_t ptpaddr; 1644 pt_entry_t *pte; 1645 int anyvalid; 1646 1647 /* 1648 * Perform an unsynchronized read. This is, however, safe. 1649 */ 1650 if (pmap->pm_stats.resident_count == 0) 1651 return; 1652 1653 vm_page_lock_queues(); 1654 sched_pin(); 1655 PMAP_LOCK(pmap); 1656 1657 /* 1658 * special handling of removing one page. a very 1659 * common operation and easy to short circuit some 1660 * code. 1661 */ 1662 if ((sva + PAGE_SIZE == eva) && 1663 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1664 pmap_remove_page(pmap, sva); 1665 goto out; 1666 } 1667 1668 anyvalid = 0; 1669 1670 for (; sva < eva; sva = pdnxt) { 1671 unsigned pdirindex; 1672 1673 /* 1674 * Calculate index for next page table. 1675 */ 1676 pdnxt = (sva + NBPDR) & ~PDRMASK; 1677 if (pmap->pm_stats.resident_count == 0) 1678 break; 1679 1680 pdirindex = sva >> PDRSHIFT; 1681 ptpaddr = pmap->pm_pdir[pdirindex]; 1682 1683 /* 1684 * Weed out invalid mappings. Note: we assume that the page 1685 * directory table is always allocated, and in kernel virtual. 1686 */ 1687 if (ptpaddr == 0) 1688 continue; 1689 1690 /* 1691 * Check for large page. 1692 */ 1693 if ((ptpaddr & PG_PS) != 0) { 1694 pmap->pm_pdir[pdirindex] = 0; 1695 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1696 anyvalid = 1; 1697 continue; 1698 } 1699 1700 /* 1701 * Limit our scan to either the end of the va represented 1702 * by the current page table page, or to the end of the 1703 * range being removed. 1704 */ 1705 if (pdnxt > eva) 1706 pdnxt = eva; 1707 1708 for (; sva != pdnxt; sva += PAGE_SIZE) { 1709 if ((pte = pmap_pte_quick(pmap, sva)) == NULL || 1710 *pte == 0) 1711 continue; 1712 anyvalid = 1; 1713 if (pmap_remove_pte(pmap, pte, sva)) 1714 break; 1715 } 1716 } 1717 1718 if (anyvalid) 1719 pmap_invalidate_all(pmap); 1720out: 1721 sched_unpin(); 1722 vm_page_unlock_queues(); 1723 PMAP_UNLOCK(pmap); 1724} 1725 1726/* 1727 * Routine: pmap_remove_all 1728 * Function: 1729 * Removes this physical page from 1730 * all physical maps in which it resides. 1731 * Reflects back modify bits to the pager. 1732 * 1733 * Notes: 1734 * Original versions of this routine were very 1735 * inefficient because they iteratively called 1736 * pmap_remove (slow...) 1737 */ 1738 1739void 1740pmap_remove_all(vm_page_t m) 1741{ 1742 register pv_entry_t pv; 1743 pt_entry_t *pte, tpte; 1744 1745#if defined(PMAP_DIAGNOSTIC) 1746 /* 1747 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1748 */ 1749 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1750 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1751 VM_PAGE_TO_PHYS(m)); 1752 } 1753#endif 1754 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1755 sched_pin(); 1756 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1757 PMAP_LOCK(pv->pv_pmap); 1758 pv->pv_pmap->pm_stats.resident_count--; 1759 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1760 tpte = pte_load_clear(pte); 1761 if (tpte & PG_W) 1762 pv->pv_pmap->pm_stats.wired_count--; 1763 if (tpte & PG_A) 1764 vm_page_flag_set(m, PG_REFERENCED); 1765 1766 /* 1767 * Update the vm_page_t clean and reference bits. 1768 */ 1769 if (tpte & PG_M) { 1770#if defined(PMAP_DIAGNOSTIC) 1771 if (pmap_nw_modified((pt_entry_t) tpte)) { 1772 printf( 1773 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1774 pv->pv_va, tpte); 1775 } 1776#endif 1777 if (pmap_track_modified(pv->pv_va)) 1778 vm_page_dirty(m); 1779 } 1780 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1781 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1782 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1783 m->md.pv_list_count--; 1784 pmap_unuse_pt(pv->pv_pmap, pv->pv_va); 1785 PMAP_UNLOCK(pv->pv_pmap); 1786 free_pv_entry(pv); 1787 } 1788 vm_page_flag_clear(m, PG_WRITEABLE); 1789 sched_unpin(); 1790} 1791 1792/* 1793 * Set the physical protection on the 1794 * specified range of this map as requested. 1795 */ 1796void 1797pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1798{ 1799 vm_offset_t pdnxt; 1800 pd_entry_t ptpaddr; 1801 int anychanged; 1802 1803 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1804 pmap_remove(pmap, sva, eva); 1805 return; 1806 } 1807 1808 if (prot & VM_PROT_WRITE) 1809 return; 1810 1811 anychanged = 0; 1812 1813 vm_page_lock_queues(); 1814 sched_pin(); 1815 PMAP_LOCK(pmap); 1816 for (; sva < eva; sva = pdnxt) { 1817 unsigned pdirindex; 1818 1819 pdnxt = (sva + NBPDR) & ~PDRMASK; 1820 1821 pdirindex = sva >> PDRSHIFT; 1822 ptpaddr = pmap->pm_pdir[pdirindex]; 1823 1824 /* 1825 * Weed out invalid mappings. Note: we assume that the page 1826 * directory table is always allocated, and in kernel virtual. 1827 */ 1828 if (ptpaddr == 0) 1829 continue; 1830 1831 /* 1832 * Check for large page. 1833 */ 1834 if ((ptpaddr & PG_PS) != 0) { 1835 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1836 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1837 anychanged = 1; 1838 continue; 1839 } 1840 1841 if (pdnxt > eva) 1842 pdnxt = eva; 1843 1844 for (; sva != pdnxt; sva += PAGE_SIZE) { 1845 pt_entry_t pbits; 1846 pt_entry_t *pte; 1847 vm_page_t m; 1848 1849 if ((pte = pmap_pte_quick(pmap, sva)) == NULL) 1850 continue; 1851 pbits = *pte; 1852 if (pbits & PG_MANAGED) { 1853 m = NULL; 1854 if (pbits & PG_A) { 1855 m = PHYS_TO_VM_PAGE(pbits); 1856 vm_page_flag_set(m, PG_REFERENCED); 1857 pbits &= ~PG_A; 1858 } 1859 if ((pbits & PG_M) != 0 && 1860 pmap_track_modified(sva)) { 1861 if (m == NULL) 1862 m = PHYS_TO_VM_PAGE(pbits); 1863 vm_page_dirty(m); 1864 pbits &= ~PG_M; 1865 } 1866 } 1867 1868 pbits &= ~PG_RW; 1869 1870 if (pbits != *pte) { 1871 pte_store(pte, pbits); 1872 anychanged = 1; 1873 } 1874 } 1875 } 1876 if (anychanged) 1877 pmap_invalidate_all(pmap); 1878 sched_unpin(); 1879 vm_page_unlock_queues(); 1880 PMAP_UNLOCK(pmap); 1881} 1882 1883/* 1884 * Insert the given physical page (p) at 1885 * the specified virtual address (v) in the 1886 * target physical map with the protection requested. 1887 * 1888 * If specified, the page will be wired down, meaning 1889 * that the related pte can not be reclaimed. 1890 * 1891 * NB: This is the only routine which MAY NOT lazy-evaluate 1892 * or lose information. That is, this routine must actually 1893 * insert this page into the given map NOW. 1894 */ 1895void 1896pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1897 boolean_t wired) 1898{ 1899 vm_paddr_t pa; 1900 register pt_entry_t *pte; 1901 vm_paddr_t opa; 1902 pt_entry_t origpte, newpte; 1903 vm_page_t mpte; 1904 1905 va &= PG_FRAME; 1906#ifdef PMAP_DIAGNOSTIC 1907 if (va > VM_MAX_KERNEL_ADDRESS) 1908 panic("pmap_enter: toobig"); 1909 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1910 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1911#endif 1912 1913 mpte = NULL; 1914 1915 vm_page_lock_queues(); 1916 PMAP_LOCK(pmap); 1917 sched_pin(); 1918 1919 /* 1920 * In the case that a page table page is not 1921 * resident, we are creating it here. 1922 */ 1923 if (va < VM_MAXUSER_ADDRESS) { 1924 mpte = pmap_allocpte(pmap, va); 1925 } 1926#if 0 && defined(PMAP_DIAGNOSTIC) 1927 else { 1928 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1929 origpte = *pdeaddr; 1930 if ((origpte & PG_V) == 0) { 1931 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1932 pmap->pm_pdir[PTDPTDI], origpte, va); 1933 } 1934 } 1935#endif 1936 1937 pte = pmap_pte_quick(pmap, va); 1938 1939 /* 1940 * Page Directory table entry not valid, we need a new PT page 1941 */ 1942 if (pte == NULL) { 1943 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1944 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1945 } 1946 1947 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 1948 origpte = *pte; 1949 opa = origpte & PG_FRAME; 1950 1951 if (origpte & PG_PS) { 1952 /* 1953 * Yes, I know this will truncate upper address bits for PAE, 1954 * but I'm actually more interested in the lower bits 1955 */ 1956 printf("pmap_enter: va %p, pte %p, origpte %p\n", 1957 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 1958 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1959 } 1960 1961 /* 1962 * Mapping has not changed, must be protection or wiring change. 1963 */ 1964 if (origpte && (opa == pa)) { 1965 /* 1966 * Wiring change, just update stats. We don't worry about 1967 * wiring PT pages as they remain resident as long as there 1968 * are valid mappings in them. Hence, if a user page is wired, 1969 * the PT page will be also. 1970 */ 1971 if (wired && ((origpte & PG_W) == 0)) 1972 pmap->pm_stats.wired_count++; 1973 else if (!wired && (origpte & PG_W)) 1974 pmap->pm_stats.wired_count--; 1975 1976#if defined(PMAP_DIAGNOSTIC) 1977 if (pmap_nw_modified((pt_entry_t) origpte)) { 1978 printf( 1979 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 1980 va, origpte); 1981 } 1982#endif 1983 1984 /* 1985 * Remove extra pte reference 1986 */ 1987 if (mpte) 1988 mpte->hold_count--; 1989 1990 /* 1991 * We might be turning off write access to the page, 1992 * so we go ahead and sense modify status. 1993 */ 1994 if (origpte & PG_MANAGED) { 1995 if ((origpte & PG_M) && pmap_track_modified(va)) { 1996 vm_page_t om; 1997 om = PHYS_TO_VM_PAGE(opa); 1998 vm_page_dirty(om); 1999 } 2000 pa |= PG_MANAGED; 2001 } 2002 goto validate; 2003 } 2004 /* 2005 * Mapping has changed, invalidate old range and fall through to 2006 * handle validating new mapping. 2007 */ 2008 if (opa) { 2009 int err; 2010 err = pmap_remove_pte(pmap, pte, va); 2011 if (err) 2012 panic("pmap_enter: pte vanished, va: 0x%x", va); 2013 } 2014 2015 /* 2016 * Enter on the PV list if part of our managed memory. Note that we 2017 * raise IPL while manipulating pv_table since pmap_enter can be 2018 * called at interrupt time. 2019 */ 2020 if (pmap_initialized && 2021 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2022 pmap_insert_entry(pmap, va, m); 2023 pa |= PG_MANAGED; 2024 } 2025 2026 /* 2027 * Increment counters 2028 */ 2029 pmap->pm_stats.resident_count++; 2030 if (wired) 2031 pmap->pm_stats.wired_count++; 2032 2033validate: 2034 /* 2035 * Now validate mapping with desired protection/wiring. 2036 */ 2037 newpte = (pt_entry_t)(pa | PG_V); 2038 if ((prot & VM_PROT_WRITE) != 0) 2039 newpte |= PG_RW; 2040 if (wired) 2041 newpte |= PG_W; 2042 if (va < VM_MAXUSER_ADDRESS) 2043 newpte |= PG_U; 2044 if (pmap == kernel_pmap) 2045 newpte |= pgeflag; 2046 2047 /* 2048 * if the mapping or permission bits are different, we need 2049 * to update the pte. 2050 */ 2051 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2052 pte_store(pte, newpte | PG_A); 2053 /*if (origpte)*/ { 2054 pmap_invalidate_page(pmap, va); 2055 } 2056 } 2057 sched_unpin(); 2058 vm_page_unlock_queues(); 2059 PMAP_UNLOCK(pmap); 2060} 2061 2062/* 2063 * this code makes some *MAJOR* assumptions: 2064 * 1. Current pmap & pmap exists. 2065 * 2. Not wired. 2066 * 3. Read access. 2067 * 4. No page table pages. 2068 * 5. Tlbflush is deferred to calling procedure. 2069 * 6. Page IS managed. 2070 * but is *MUCH* faster than pmap_enter... 2071 */ 2072 2073vm_page_t 2074pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2075{ 2076 pt_entry_t *pte; 2077 vm_paddr_t pa; 2078 2079 vm_page_lock_queues(); 2080 PMAP_LOCK(pmap); 2081 2082 /* 2083 * In the case that a page table page is not 2084 * resident, we are creating it here. 2085 */ 2086 if (va < VM_MAXUSER_ADDRESS) { 2087 unsigned ptepindex; 2088 pd_entry_t ptepa; 2089 2090 /* 2091 * Calculate pagetable page index 2092 */ 2093 ptepindex = va >> PDRSHIFT; 2094 if (mpte && (mpte->pindex == ptepindex)) { 2095 mpte->hold_count++; 2096 } else { 2097retry: 2098 /* 2099 * Get the page directory entry 2100 */ 2101 ptepa = pmap->pm_pdir[ptepindex]; 2102 2103 /* 2104 * If the page table page is mapped, we just increment 2105 * the hold count, and activate it. 2106 */ 2107 if (ptepa) { 2108 if (ptepa & PG_PS) 2109 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2110 mpte = PHYS_TO_VM_PAGE(ptepa); 2111 mpte->hold_count++; 2112 } else { 2113 mpte = _pmap_allocpte(pmap, ptepindex); 2114 if (mpte == NULL) 2115 goto retry; 2116 } 2117 } 2118 } else { 2119 mpte = NULL; 2120 } 2121 2122 /* 2123 * This call to vtopte makes the assumption that we are 2124 * entering the page into the current pmap. In order to support 2125 * quick entry into any pmap, one would likely use pmap_pte_quick. 2126 * But that isn't as quick as vtopte. 2127 */ 2128 pte = vtopte(va); 2129 if (*pte) { 2130 if (mpte != NULL) { 2131 pmap_unwire_pte_hold(pmap, mpte); 2132 mpte = NULL; 2133 } 2134 goto out; 2135 } 2136 2137 /* 2138 * Enter on the PV list if part of our managed memory. Note that we 2139 * raise IPL while manipulating pv_table since pmap_enter can be 2140 * called at interrupt time. 2141 */ 2142 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2143 pmap_insert_entry(pmap, va, m); 2144 2145 /* 2146 * Increment counters 2147 */ 2148 pmap->pm_stats.resident_count++; 2149 2150 pa = VM_PAGE_TO_PHYS(m); 2151 2152 /* 2153 * Now validate mapping with RO protection 2154 */ 2155 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2156 pte_store(pte, pa | PG_V | PG_U); 2157 else 2158 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2159out: 2160 vm_page_unlock_queues(); 2161 PMAP_UNLOCK(pmap); 2162 return mpte; 2163} 2164 2165/* 2166 * Make a temporary mapping for a physical address. This is only intended 2167 * to be used for panic dumps. 2168 */ 2169void * 2170pmap_kenter_temporary(vm_paddr_t pa, int i) 2171{ 2172 vm_offset_t va; 2173 2174 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2175 pmap_kenter(va, pa); 2176#ifndef I386_CPU 2177 invlpg(va); 2178#else 2179 invltlb(); 2180#endif 2181 return ((void *)crashdumpmap); 2182} 2183 2184/* 2185 * This code maps large physical mmap regions into the 2186 * processor address space. Note that some shortcuts 2187 * are taken, but the code works. 2188 */ 2189void 2190pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2191 vm_object_t object, vm_pindex_t pindex, 2192 vm_size_t size) 2193{ 2194 vm_page_t p; 2195 2196 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2197 KASSERT(object->type == OBJT_DEVICE, 2198 ("pmap_object_init_pt: non-device object")); 2199 if (pseflag && 2200 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2201 int i; 2202 vm_page_t m[1]; 2203 unsigned int ptepindex; 2204 int npdes; 2205 pd_entry_t ptepa; 2206 2207 PMAP_LOCK(pmap); 2208 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2209 goto out; 2210 PMAP_UNLOCK(pmap); 2211retry: 2212 p = vm_page_lookup(object, pindex); 2213 if (p != NULL) { 2214 vm_page_lock_queues(); 2215 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2216 goto retry; 2217 } else { 2218 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2219 if (p == NULL) 2220 return; 2221 m[0] = p; 2222 2223 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2224 vm_page_lock_queues(); 2225 vm_page_free(p); 2226 vm_page_unlock_queues(); 2227 return; 2228 } 2229 2230 p = vm_page_lookup(object, pindex); 2231 vm_page_lock_queues(); 2232 vm_page_wakeup(p); 2233 } 2234 vm_page_unlock_queues(); 2235 2236 ptepa = VM_PAGE_TO_PHYS(p); 2237 if (ptepa & (NBPDR - 1)) 2238 return; 2239 2240 p->valid = VM_PAGE_BITS_ALL; 2241 2242 PMAP_LOCK(pmap); 2243 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2244 npdes = size >> PDRSHIFT; 2245 for(i = 0; i < npdes; i++) { 2246 pde_store(&pmap->pm_pdir[ptepindex], 2247 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2248 ptepa += NBPDR; 2249 ptepindex += 1; 2250 } 2251 pmap_invalidate_all(pmap); 2252out: 2253 PMAP_UNLOCK(pmap); 2254 } 2255} 2256 2257/* 2258 * Routine: pmap_change_wiring 2259 * Function: Change the wiring attribute for a map/virtual-address 2260 * pair. 2261 * In/out conditions: 2262 * The mapping must already exist in the pmap. 2263 */ 2264void 2265pmap_change_wiring(pmap, va, wired) 2266 register pmap_t pmap; 2267 vm_offset_t va; 2268 boolean_t wired; 2269{ 2270 register pt_entry_t *pte; 2271 2272 PMAP_LOCK(pmap); 2273 pte = pmap_pte(pmap, va); 2274 2275 if (wired && !pmap_pte_w(pte)) 2276 pmap->pm_stats.wired_count++; 2277 else if (!wired && pmap_pte_w(pte)) 2278 pmap->pm_stats.wired_count--; 2279 2280 /* 2281 * Wiring is not a hardware characteristic so there is no need to 2282 * invalidate TLB. 2283 */ 2284 pmap_pte_set_w(pte, wired); 2285 pmap_pte_release(pte); 2286 PMAP_UNLOCK(pmap); 2287} 2288 2289 2290 2291/* 2292 * Copy the range specified by src_addr/len 2293 * from the source map to the range dst_addr/len 2294 * in the destination map. 2295 * 2296 * This routine is only advisory and need not do anything. 2297 */ 2298 2299void 2300pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2301 vm_offset_t src_addr) 2302{ 2303 vm_offset_t addr; 2304 vm_offset_t end_addr = src_addr + len; 2305 vm_offset_t pdnxt; 2306 vm_page_t m; 2307 2308 if (dst_addr != src_addr) 2309 return; 2310 2311 if (!pmap_is_current(src_pmap)) 2312 return; 2313 2314 vm_page_lock_queues(); 2315 PMAP_LOCK(dst_pmap); 2316 sched_pin(); 2317 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2318 pt_entry_t *src_pte, *dst_pte; 2319 vm_page_t dstmpte, srcmpte; 2320 pd_entry_t srcptepaddr; 2321 unsigned ptepindex; 2322 2323 if (addr >= UPT_MIN_ADDRESS) 2324 panic("pmap_copy: invalid to pmap_copy page tables"); 2325 2326 /* 2327 * Don't let optional prefaulting of pages make us go 2328 * way below the low water mark of free pages or way 2329 * above high water mark of used pv entries. 2330 */ 2331 if (cnt.v_free_count < cnt.v_free_reserved || 2332 pv_entry_count > pv_entry_high_water) 2333 break; 2334 2335 pdnxt = (addr + NBPDR) & ~PDRMASK; 2336 ptepindex = addr >> PDRSHIFT; 2337 2338 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2339 if (srcptepaddr == 0) 2340 continue; 2341 2342 if (srcptepaddr & PG_PS) { 2343 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2344 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2345 dst_pmap->pm_stats.resident_count += 2346 NBPDR / PAGE_SIZE; 2347 } 2348 continue; 2349 } 2350 2351 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2352 if (srcmpte->hold_count == 0) 2353 panic("pmap_copy: source page table page is unused"); 2354 2355 if (pdnxt > end_addr) 2356 pdnxt = end_addr; 2357 2358 src_pte = vtopte(addr); 2359 while (addr < pdnxt) { 2360 pt_entry_t ptetemp; 2361 ptetemp = *src_pte; 2362 /* 2363 * we only virtual copy managed pages 2364 */ 2365 if ((ptetemp & PG_MANAGED) != 0) { 2366 /* 2367 * We have to check after allocpte for the 2368 * pte still being around... allocpte can 2369 * block. 2370 */ 2371 dstmpte = pmap_allocpte(dst_pmap, addr); 2372 dst_pte = pmap_pte_quick(dst_pmap, addr); 2373 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2374 /* 2375 * Clear the modified and 2376 * accessed (referenced) bits 2377 * during the copy. 2378 */ 2379 m = PHYS_TO_VM_PAGE(ptetemp); 2380 *dst_pte = ptetemp & ~(PG_M | PG_A); 2381 dst_pmap->pm_stats.resident_count++; 2382 pmap_insert_entry(dst_pmap, addr, m); 2383 } else 2384 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2385 if (dstmpte->hold_count >= srcmpte->hold_count) 2386 break; 2387 } 2388 addr += PAGE_SIZE; 2389 src_pte++; 2390 } 2391 } 2392 sched_unpin(); 2393 vm_page_unlock_queues(); 2394 PMAP_UNLOCK(dst_pmap); 2395} 2396 2397static __inline void 2398pagezero(void *page) 2399{ 2400#if defined(I686_CPU) 2401 if (cpu_class == CPUCLASS_686) { 2402#if defined(CPU_ENABLE_SSE) 2403 if (cpu_feature & CPUID_SSE2) 2404 sse2_pagezero(page); 2405 else 2406#endif 2407 i686_pagezero(page); 2408 } else 2409#endif 2410 bzero(page, PAGE_SIZE); 2411} 2412 2413/* 2414 * pmap_zero_page zeros the specified hardware page by mapping 2415 * the page into KVM and using bzero to clear its contents. 2416 */ 2417void 2418pmap_zero_page(vm_page_t m) 2419{ 2420 2421 mtx_lock(&CMAPCADDR12_lock); 2422 if (*CMAP2) 2423 panic("pmap_zero_page: CMAP2 busy"); 2424 sched_pin(); 2425 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2426 invlcaddr(CADDR2); 2427 pagezero(CADDR2); 2428 *CMAP2 = 0; 2429 sched_unpin(); 2430 mtx_unlock(&CMAPCADDR12_lock); 2431} 2432 2433/* 2434 * pmap_zero_page_area zeros the specified hardware page by mapping 2435 * the page into KVM and using bzero to clear its contents. 2436 * 2437 * off and size may not cover an area beyond a single hardware page. 2438 */ 2439void 2440pmap_zero_page_area(vm_page_t m, int off, int size) 2441{ 2442 2443 mtx_lock(&CMAPCADDR12_lock); 2444 if (*CMAP2) 2445 panic("pmap_zero_page: CMAP2 busy"); 2446 sched_pin(); 2447 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2448 invlcaddr(CADDR2); 2449 if (off == 0 && size == PAGE_SIZE) 2450 pagezero(CADDR2); 2451 else 2452 bzero((char *)CADDR2 + off, size); 2453 *CMAP2 = 0; 2454 sched_unpin(); 2455 mtx_unlock(&CMAPCADDR12_lock); 2456} 2457 2458/* 2459 * pmap_zero_page_idle zeros the specified hardware page by mapping 2460 * the page into KVM and using bzero to clear its contents. This 2461 * is intended to be called from the vm_pagezero process only and 2462 * outside of Giant. 2463 */ 2464void 2465pmap_zero_page_idle(vm_page_t m) 2466{ 2467 2468 if (*CMAP3) 2469 panic("pmap_zero_page: CMAP3 busy"); 2470 sched_pin(); 2471 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2472 invlcaddr(CADDR3); 2473 pagezero(CADDR3); 2474 *CMAP3 = 0; 2475 sched_unpin(); 2476} 2477 2478/* 2479 * pmap_copy_page copies the specified (machine independent) 2480 * page by mapping the page into virtual memory and using 2481 * bcopy to copy the page, one machine dependent page at a 2482 * time. 2483 */ 2484void 2485pmap_copy_page(vm_page_t src, vm_page_t dst) 2486{ 2487 2488 mtx_lock(&CMAPCADDR12_lock); 2489 if (*CMAP1) 2490 panic("pmap_copy_page: CMAP1 busy"); 2491 if (*CMAP2) 2492 panic("pmap_copy_page: CMAP2 busy"); 2493 sched_pin(); 2494#ifdef I386_CPU 2495 invltlb(); 2496#else 2497 invlpg((u_int)CADDR1); 2498 invlpg((u_int)CADDR2); 2499#endif 2500 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2501 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2502 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2503 *CMAP1 = 0; 2504 *CMAP2 = 0; 2505 sched_unpin(); 2506 mtx_unlock(&CMAPCADDR12_lock); 2507} 2508 2509/* 2510 * Returns true if the pmap's pv is one of the first 2511 * 16 pvs linked to from this page. This count may 2512 * be changed upwards or downwards in the future; it 2513 * is only necessary that true be returned for a small 2514 * subset of pmaps for proper page aging. 2515 */ 2516boolean_t 2517pmap_page_exists_quick(pmap, m) 2518 pmap_t pmap; 2519 vm_page_t m; 2520{ 2521 pv_entry_t pv; 2522 int loops = 0; 2523 2524 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2525 return FALSE; 2526 2527 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2528 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2529 if (pv->pv_pmap == pmap) { 2530 return TRUE; 2531 } 2532 loops++; 2533 if (loops >= 16) 2534 break; 2535 } 2536 return (FALSE); 2537} 2538 2539#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2540/* 2541 * Remove all pages from specified address space 2542 * this aids process exit speeds. Also, this code 2543 * is special cased for current process only, but 2544 * can have the more generic (and slightly slower) 2545 * mode enabled. This is much faster than pmap_remove 2546 * in the case of running down an entire address space. 2547 */ 2548void 2549pmap_remove_pages(pmap, sva, eva) 2550 pmap_t pmap; 2551 vm_offset_t sva, eva; 2552{ 2553 pt_entry_t *pte, tpte; 2554 vm_page_t m; 2555 pv_entry_t pv, npv; 2556 2557#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2558 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) { 2559 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2560 return; 2561 } 2562#endif 2563 vm_page_lock_queues(); 2564 PMAP_LOCK(pmap); 2565 sched_pin(); 2566 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2567 2568 if (pv->pv_va >= eva || pv->pv_va < sva) { 2569 npv = TAILQ_NEXT(pv, pv_plist); 2570 continue; 2571 } 2572 2573#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2574 pte = vtopte(pv->pv_va); 2575#else 2576 pte = pmap_pte_quick(pmap, pv->pv_va); 2577#endif 2578 tpte = *pte; 2579 2580 if (tpte == 0) { 2581 printf("TPTE at %p IS ZERO @ VA %08x\n", 2582 pte, pv->pv_va); 2583 panic("bad pte"); 2584 } 2585 2586/* 2587 * We cannot remove wired pages from a process' mapping at this time 2588 */ 2589 if (tpte & PG_W) { 2590 npv = TAILQ_NEXT(pv, pv_plist); 2591 continue; 2592 } 2593 2594 m = PHYS_TO_VM_PAGE(tpte); 2595 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2596 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2597 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2598 2599 KASSERT(m < &vm_page_array[vm_page_array_size], 2600 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2601 2602 pmap->pm_stats.resident_count--; 2603 2604 pte_clear(pte); 2605 2606 /* 2607 * Update the vm_page_t clean and reference bits. 2608 */ 2609 if (tpte & PG_M) { 2610 vm_page_dirty(m); 2611 } 2612 2613 npv = TAILQ_NEXT(pv, pv_plist); 2614 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 2615 2616 m->md.pv_list_count--; 2617 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2618 if (TAILQ_EMPTY(&m->md.pv_list)) 2619 vm_page_flag_clear(m, PG_WRITEABLE); 2620 2621 pmap_unuse_pt(pmap, pv->pv_va); 2622 free_pv_entry(pv); 2623 } 2624 sched_unpin(); 2625 pmap_invalidate_all(pmap); 2626 PMAP_UNLOCK(pmap); 2627 vm_page_unlock_queues(); 2628} 2629 2630/* 2631 * pmap_is_modified: 2632 * 2633 * Return whether or not the specified physical page was modified 2634 * in any physical maps. 2635 */ 2636boolean_t 2637pmap_is_modified(vm_page_t m) 2638{ 2639 pv_entry_t pv; 2640 pt_entry_t *pte; 2641 boolean_t rv; 2642 2643 rv = FALSE; 2644 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2645 return (rv); 2646 2647 sched_pin(); 2648 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2649 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2650 /* 2651 * if the bit being tested is the modified bit, then 2652 * mark clean_map and ptes as never 2653 * modified. 2654 */ 2655 if (!pmap_track_modified(pv->pv_va)) 2656 continue; 2657#if defined(PMAP_DIAGNOSTIC) 2658 if (!pv->pv_pmap) { 2659 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 2660 continue; 2661 } 2662#endif 2663 PMAP_LOCK(pv->pv_pmap); 2664 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2665 rv = (*pte & PG_M) != 0; 2666 PMAP_UNLOCK(pv->pv_pmap); 2667 if (rv) 2668 break; 2669 } 2670 sched_unpin(); 2671 return (rv); 2672} 2673 2674/* 2675 * pmap_is_prefaultable: 2676 * 2677 * Return whether or not the specified virtual address is elgible 2678 * for prefault. 2679 */ 2680boolean_t 2681pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2682{ 2683 pt_entry_t *pte; 2684 boolean_t rv; 2685 2686 rv = FALSE; 2687 PMAP_LOCK(pmap); 2688 if (*pmap_pde(pmap, addr)) { 2689 pte = vtopte(addr); 2690 rv = *pte == 0; 2691 } 2692 PMAP_UNLOCK(pmap); 2693 return (rv); 2694} 2695 2696/* 2697 * Clear the given bit in each of the given page's ptes. The bit is 2698 * expressed as a 32-bit mask. Consequently, if the pte is 64 bits in 2699 * size, only a bit within the least significant 32 can be cleared. 2700 */ 2701static __inline void 2702pmap_clear_ptes(vm_page_t m, int bit) 2703{ 2704 register pv_entry_t pv; 2705 pt_entry_t pbits, *pte; 2706 2707 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) || 2708 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2709 return; 2710 2711 sched_pin(); 2712 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2713 /* 2714 * Loop over all current mappings setting/clearing as appropos If 2715 * setting RO do we need to clear the VAC? 2716 */ 2717 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2718 /* 2719 * don't write protect pager mappings 2720 */ 2721 if (bit == PG_RW) { 2722 if (!pmap_track_modified(pv->pv_va)) 2723 continue; 2724 } 2725 2726#if defined(PMAP_DIAGNOSTIC) 2727 if (!pv->pv_pmap) { 2728 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 2729 continue; 2730 } 2731#endif 2732 2733 PMAP_LOCK(pv->pv_pmap); 2734 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2735retry: 2736 pbits = *pte; 2737 if (pbits & bit) { 2738 if (bit == PG_RW) { 2739 /* 2740 * Regardless of whether a pte is 32 or 64 bits 2741 * in size, PG_RW and PG_M are among the least 2742 * significant 32 bits. 2743 */ 2744 if (!atomic_cmpset_int((u_int *)pte, pbits, 2745 pbits & ~(PG_RW | PG_M))) 2746 goto retry; 2747 if (pbits & PG_M) { 2748 vm_page_dirty(m); 2749 } 2750 } else { 2751 atomic_clear_int((u_int *)pte, bit); 2752 } 2753 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2754 } 2755 PMAP_UNLOCK(pv->pv_pmap); 2756 } 2757 if (bit == PG_RW) 2758 vm_page_flag_clear(m, PG_WRITEABLE); 2759 sched_unpin(); 2760} 2761 2762/* 2763 * pmap_page_protect: 2764 * 2765 * Lower the permission for all mappings to a given page. 2766 */ 2767void 2768pmap_page_protect(vm_page_t m, vm_prot_t prot) 2769{ 2770 if ((prot & VM_PROT_WRITE) == 0) { 2771 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2772 pmap_clear_ptes(m, PG_RW); 2773 } else { 2774 pmap_remove_all(m); 2775 } 2776 } 2777} 2778 2779/* 2780 * pmap_ts_referenced: 2781 * 2782 * Return a count of reference bits for a page, clearing those bits. 2783 * It is not necessary for every reference bit to be cleared, but it 2784 * is necessary that 0 only be returned when there are truly no 2785 * reference bits set. 2786 * 2787 * XXX: The exact number of bits to check and clear is a matter that 2788 * should be tested and standardized at some point in the future for 2789 * optimal aging of shared pages. 2790 */ 2791int 2792pmap_ts_referenced(vm_page_t m) 2793{ 2794 register pv_entry_t pv, pvf, pvn; 2795 pt_entry_t *pte; 2796 pt_entry_t v; 2797 int rtval = 0; 2798 2799 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2800 return (rtval); 2801 2802 sched_pin(); 2803 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2804 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2805 2806 pvf = pv; 2807 2808 do { 2809 pvn = TAILQ_NEXT(pv, pv_list); 2810 2811 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2812 2813 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2814 2815 if (!pmap_track_modified(pv->pv_va)) 2816 continue; 2817 2818 PMAP_LOCK(pv->pv_pmap); 2819 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2820 2821 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2822 atomic_clear_int((u_int *)pte, PG_A); 2823 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2824 2825 rtval++; 2826 if (rtval > 4) { 2827 PMAP_UNLOCK(pv->pv_pmap); 2828 break; 2829 } 2830 } 2831 PMAP_UNLOCK(pv->pv_pmap); 2832 } while ((pv = pvn) != NULL && pv != pvf); 2833 } 2834 sched_unpin(); 2835 2836 return (rtval); 2837} 2838 2839/* 2840 * Clear the modify bits on the specified physical page. 2841 */ 2842void 2843pmap_clear_modify(vm_page_t m) 2844{ 2845 pmap_clear_ptes(m, PG_M); 2846} 2847 2848/* 2849 * pmap_clear_reference: 2850 * 2851 * Clear the reference bit on the specified physical page. 2852 */ 2853void 2854pmap_clear_reference(vm_page_t m) 2855{ 2856 pmap_clear_ptes(m, PG_A); 2857} 2858 2859/* 2860 * Miscellaneous support routines follow 2861 */ 2862 2863/* 2864 * Map a set of physical memory pages into the kernel virtual 2865 * address space. Return a pointer to where it is mapped. This 2866 * routine is intended to be used for mapping device memory, 2867 * NOT real memory. 2868 */ 2869void * 2870pmap_mapdev(pa, size) 2871 vm_paddr_t pa; 2872 vm_size_t size; 2873{ 2874 vm_offset_t va, tmpva, offset; 2875 2876 offset = pa & PAGE_MASK; 2877 size = roundup(offset + size, PAGE_SIZE); 2878 pa = pa & PG_FRAME; 2879 2880 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2881 va = KERNBASE + pa; 2882 else 2883 va = kmem_alloc_nofault(kernel_map, size); 2884 if (!va) 2885 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2886 2887 for (tmpva = va; size > 0; ) { 2888 pmap_kenter(tmpva, pa); 2889 size -= PAGE_SIZE; 2890 tmpva += PAGE_SIZE; 2891 pa += PAGE_SIZE; 2892 } 2893 pmap_invalidate_range(kernel_pmap, va, tmpva); 2894 return ((void *)(va + offset)); 2895} 2896 2897void 2898pmap_unmapdev(va, size) 2899 vm_offset_t va; 2900 vm_size_t size; 2901{ 2902 vm_offset_t base, offset, tmpva; 2903 2904 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2905 return; 2906 base = va & PG_FRAME; 2907 offset = va & PAGE_MASK; 2908 size = roundup(offset + size, PAGE_SIZE); 2909 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2910 pmap_kremove(tmpva); 2911 pmap_invalidate_range(kernel_pmap, va, tmpva); 2912 kmem_free(kernel_map, base, size); 2913} 2914 2915/* 2916 * perform the pmap work for mincore 2917 */ 2918int 2919pmap_mincore(pmap, addr) 2920 pmap_t pmap; 2921 vm_offset_t addr; 2922{ 2923 pt_entry_t *ptep, pte; 2924 vm_page_t m; 2925 int val = 0; 2926 2927 PMAP_LOCK(pmap); 2928 ptep = pmap_pte(pmap, addr); 2929 pte = (ptep != NULL) ? *ptep : 0; 2930 pmap_pte_release(ptep); 2931 PMAP_UNLOCK(pmap); 2932 2933 if (pte != 0) { 2934 vm_paddr_t pa; 2935 2936 val = MINCORE_INCORE; 2937 if ((pte & PG_MANAGED) == 0) 2938 return val; 2939 2940 pa = pte & PG_FRAME; 2941 2942 m = PHYS_TO_VM_PAGE(pa); 2943 2944 /* 2945 * Modified by us 2946 */ 2947 if (pte & PG_M) 2948 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2949 else { 2950 /* 2951 * Modified by someone else 2952 */ 2953 vm_page_lock_queues(); 2954 if (m->dirty || pmap_is_modified(m)) 2955 val |= MINCORE_MODIFIED_OTHER; 2956 vm_page_unlock_queues(); 2957 } 2958 /* 2959 * Referenced by us 2960 */ 2961 if (pte & PG_A) 2962 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2963 else { 2964 /* 2965 * Referenced by someone else 2966 */ 2967 vm_page_lock_queues(); 2968 if ((m->flags & PG_REFERENCED) || 2969 pmap_ts_referenced(m)) { 2970 val |= MINCORE_REFERENCED_OTHER; 2971 vm_page_flag_set(m, PG_REFERENCED); 2972 } 2973 vm_page_unlock_queues(); 2974 } 2975 } 2976 return val; 2977} 2978 2979void 2980pmap_activate(struct thread *td) 2981{ 2982 struct proc *p = td->td_proc; 2983 pmap_t pmap, oldpmap; 2984 u_int32_t cr3; 2985 2986 critical_enter(); 2987 pmap = vmspace_pmap(td->td_proc->p_vmspace); 2988 oldpmap = PCPU_GET(curpmap); 2989#if defined(SMP) 2990 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 2991 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 2992#else 2993 oldpmap->pm_active &= ~1; 2994 pmap->pm_active |= 1; 2995#endif 2996#ifdef PAE 2997 cr3 = vtophys(pmap->pm_pdpt); 2998#else 2999 cr3 = vtophys(pmap->pm_pdir); 3000#endif 3001 /* XXXKSE this is wrong. 3002 * pmap_activate is for the current thread on the current cpu 3003 */ 3004 if (p->p_flag & P_SA) { 3005 /* Make sure all other cr3 entries are updated. */ 3006 /* what if they are running? XXXKSE (maybe abort them) */ 3007 FOREACH_THREAD_IN_PROC(p, td) { 3008 td->td_pcb->pcb_cr3 = cr3; 3009 } 3010 } else { 3011 td->td_pcb->pcb_cr3 = cr3; 3012 } 3013 load_cr3(cr3); 3014 PCPU_SET(curpmap, pmap); 3015 critical_exit(); 3016} 3017 3018vm_offset_t 3019pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3020{ 3021 3022 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3023 return addr; 3024 } 3025 3026 addr = (addr + PDRMASK) & ~PDRMASK; 3027 return addr; 3028} 3029 3030 3031#if defined(PMAP_DEBUG) 3032pmap_pid_dump(int pid) 3033{ 3034 pmap_t pmap; 3035 struct proc *p; 3036 int npte = 0; 3037 int index; 3038 3039 sx_slock(&allproc_lock); 3040 LIST_FOREACH(p, &allproc, p_list) { 3041 if (p->p_pid != pid) 3042 continue; 3043 3044 if (p->p_vmspace) { 3045 int i,j; 3046 index = 0; 3047 pmap = vmspace_pmap(p->p_vmspace); 3048 for (i = 0; i < NPDEPTD; i++) { 3049 pd_entry_t *pde; 3050 pt_entry_t *pte; 3051 vm_offset_t base = i << PDRSHIFT; 3052 3053 pde = &pmap->pm_pdir[i]; 3054 if (pde && pmap_pde_v(pde)) { 3055 for (j = 0; j < NPTEPG; j++) { 3056 vm_offset_t va = base + (j << PAGE_SHIFT); 3057 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3058 if (index) { 3059 index = 0; 3060 printf("\n"); 3061 } 3062 sx_sunlock(&allproc_lock); 3063 return npte; 3064 } 3065 pte = pmap_pte(pmap, va); 3066 if (pte && pmap_pte_v(pte)) { 3067 pt_entry_t pa; 3068 vm_page_t m; 3069 pa = *pte; 3070 m = PHYS_TO_VM_PAGE(pa); 3071 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3072 va, pa, m->hold_count, m->wire_count, m->flags); 3073 npte++; 3074 index++; 3075 if (index >= 2) { 3076 index = 0; 3077 printf("\n"); 3078 } else { 3079 printf(" "); 3080 } 3081 } 3082 } 3083 } 3084 } 3085 } 3086 } 3087 sx_sunlock(&allproc_lock); 3088 return npte; 3089} 3090#endif 3091 3092#if defined(DEBUG) 3093 3094static void pads(pmap_t pm); 3095void pmap_pvdump(vm_offset_t pa); 3096 3097/* print address space of pmap*/ 3098static void 3099pads(pm) 3100 pmap_t pm; 3101{ 3102 int i, j; 3103 vm_paddr_t va; 3104 pt_entry_t *ptep; 3105 3106 if (pm == kernel_pmap) 3107 return; 3108 for (i = 0; i < NPDEPTD; i++) 3109 if (pm->pm_pdir[i]) 3110 for (j = 0; j < NPTEPG; j++) { 3111 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3112 if (pm == kernel_pmap && va < KERNBASE) 3113 continue; 3114 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3115 continue; 3116 ptep = pmap_pte(pm, va); 3117 if (pmap_pte_v(ptep)) 3118 printf("%x:%x ", va, *ptep); 3119 }; 3120 3121} 3122 3123void 3124pmap_pvdump(pa) 3125 vm_paddr_t pa; 3126{ 3127 pv_entry_t pv; 3128 vm_page_t m; 3129 3130 printf("pa %x", pa); 3131 m = PHYS_TO_VM_PAGE(pa); 3132 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3133 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3134 pads(pv->pv_pmap); 3135 } 3136 printf(" "); 3137} 3138#endif 3139