pmap.c revision 134393
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 134393 2004-08-27 19:06:17Z alc $"); 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"); 247 248static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 249static pv_entry_t get_pv_entry(void); 250static void pmap_clear_ptes(vm_page_t m, int bit); 251 252static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 253static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 254static int pmap_remove_entry(struct pmap *pmap, vm_page_t m, 255 vm_offset_t va); 256static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); 257 258static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va); 259 260static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex); 261static int _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m); 262static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 263static int pmap_unuse_pt(pmap_t, vm_offset_t); 264static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 265#ifdef PAE 266static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 267#endif 268 269CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 270CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 271 272/* 273 * Move the kernel virtual free pointer to the next 274 * 4MB. This is used to help improve performance 275 * by using a large (4MB) page for much of the kernel 276 * (.text, .data, .bss) 277 */ 278static vm_offset_t 279pmap_kmem_choose(vm_offset_t addr) 280{ 281 vm_offset_t newaddr = addr; 282 283#ifndef DISABLE_PSE 284 if (cpu_feature & CPUID_PSE) 285 newaddr = (addr + PDRMASK) & ~PDRMASK; 286#endif 287 return newaddr; 288} 289 290/* 291 * Bootstrap the system enough to run with virtual memory. 292 * 293 * On the i386 this is called after mapping has already been enabled 294 * and just syncs the pmap module with what has already been done. 295 * [We can't call it easily with mapping off since the kernel is not 296 * mapped with PA == VA, hence we would have to relocate every address 297 * from the linked base (virtual) address "KERNBASE" to the actual 298 * (physical) address starting relative to 0] 299 */ 300void 301pmap_bootstrap(firstaddr, loadaddr) 302 vm_paddr_t firstaddr; 303 vm_paddr_t loadaddr; 304{ 305 vm_offset_t va; 306 pt_entry_t *pte, *unused; 307 int i; 308 309 /* 310 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 311 * large. It should instead be correctly calculated in locore.s and 312 * not based on 'first' (which is a physical address, not a virtual 313 * address, for the start of unused physical memory). The kernel 314 * page tables are NOT double mapped and thus should not be included 315 * in this calculation. 316 */ 317 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 318 virtual_avail = pmap_kmem_choose(virtual_avail); 319 320 virtual_end = VM_MAX_KERNEL_ADDRESS; 321 322 /* 323 * Initialize the kernel pmap (which is statically allocated). 324 */ 325 PMAP_LOCK_INIT(kernel_pmap); 326 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 327#ifdef PAE 328 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 329#endif 330 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 331 TAILQ_INIT(&kernel_pmap->pm_pvlist); 332 LIST_INIT(&allpmaps); 333 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 334 mtx_lock_spin(&allpmaps_lock); 335 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 336 mtx_unlock_spin(&allpmaps_lock); 337 nkpt = NKPT; 338 339 /* 340 * Reserve some special page table entries/VA space for temporary 341 * mapping of pages. 342 */ 343#define SYSMAP(c, p, v, n) \ 344 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 345 346 va = virtual_avail; 347 pte = vtopte(va); 348 349 /* 350 * CMAP1/CMAP2 are used for zeroing and copying pages. 351 * CMAP3 is used for the idle process page zeroing. 352 */ 353 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 354 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 355 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 356 *CMAP3 = 0; 357 358 mtx_init(&CMAPCADDR12_lock, "CMAPCADDR12", NULL, MTX_DEF); 359 360 /* 361 * Crashdump maps. 362 */ 363 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS) 364 365 /* 366 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 367 */ 368 SYSMAP(caddr_t, unused, ptvmmap, 1) 369 370 /* 371 * msgbufp is used to map the system message buffer. 372 */ 373 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE))) 374 375 /* 376 * ptemap is used for pmap_pte_quick 377 */ 378 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 379 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1); 380 381 virtual_avail = va; 382 383 *CMAP1 = *CMAP2 = 0; 384 for (i = 0; i < NKPT; i++) 385 PTD[i] = 0; 386 387 /* Turn on PG_G on kernel page(s) */ 388 pmap_set_pg(); 389} 390 391/* 392 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 393 */ 394void 395pmap_set_pg(void) 396{ 397 pd_entry_t pdir; 398 pt_entry_t *pte; 399 vm_offset_t va, endva; 400 int i; 401 402 if (pgeflag == 0) 403 return; 404 405 i = KERNLOAD/NBPDR; 406 endva = KERNBASE + KERNend; 407 408 if (pseflag) { 409 va = KERNBASE + KERNLOAD; 410 while (va < endva) { 411 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 412 pdir |= pgeflag; 413 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 414 invltlb(); /* Play it safe, invltlb() every time */ 415 i++; 416 va += NBPDR; 417 } 418 } else { 419 va = (vm_offset_t)btext; 420 while (va < endva) { 421 pte = vtopte(va); 422 if (*pte) 423 *pte |= pgeflag; 424 invltlb(); /* Play it safe, invltlb() every time */ 425 va += PAGE_SIZE; 426 } 427 } 428} 429 430#ifdef PAE 431 432static MALLOC_DEFINE(M_PMAPPDPT, "pmap", "pmap pdpt"); 433 434static void * 435pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 436{ 437 *flags = UMA_SLAB_PRIV; 438 return (contigmalloc(PAGE_SIZE, M_PMAPPDPT, 0, 0x0ULL, 0xffffffffULL, 439 1, 0)); 440} 441#endif 442 443/* 444 * Initialize the pmap module. 445 * Called by vm_init, to initialize any structures that the pmap 446 * system needs to map virtual memory. 447 * pmap_init has been enhanced to support in a fairly consistant 448 * way, discontiguous physical memory. 449 */ 450void 451pmap_init(void) 452{ 453 int i; 454 455 /* 456 * Allocate memory for random pmap data structures. Includes the 457 * pv_head_table. 458 */ 459 460 for(i = 0; i < vm_page_array_size; i++) { 461 vm_page_t m; 462 463 m = &vm_page_array[i]; 464 TAILQ_INIT(&m->md.pv_list); 465 m->md.pv_list_count = 0; 466 } 467 468 /* 469 * init the pv free list 470 */ 471 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL, 472 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 473 uma_prealloc(pvzone, MINPV); 474 475#ifdef PAE 476 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 477 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 478 UMA_ZONE_VM | UMA_ZONE_NOFREE); 479 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 480#endif 481 482 /* 483 * Now it is safe to enable pv_table recording. 484 */ 485 pmap_initialized = TRUE; 486} 487 488/* 489 * Initialize the address space (zone) for the pv_entries. Set a 490 * high water mark so that the system can recover from excessive 491 * numbers of pv entries. 492 */ 493void 494pmap_init2() 495{ 496 int shpgperproc = PMAP_SHPGPERPROC; 497 498 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 499 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 500 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 501 pv_entry_high_water = 9 * (pv_entry_max / 10); 502 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 503} 504 505 506/*************************************************** 507 * Low level helper routines..... 508 ***************************************************/ 509 510#if defined(PMAP_DIAGNOSTIC) 511 512/* 513 * This code checks for non-writeable/modified pages. 514 * This should be an invalid condition. 515 */ 516static int 517pmap_nw_modified(pt_entry_t ptea) 518{ 519 int pte; 520 521 pte = (int) ptea; 522 523 if ((pte & (PG_M|PG_RW)) == PG_M) 524 return 1; 525 else 526 return 0; 527} 528#endif 529 530 531/* 532 * this routine defines the region(s) of memory that should 533 * not be tested for the modified bit. 534 */ 535static PMAP_INLINE int 536pmap_track_modified(vm_offset_t va) 537{ 538 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 539 return 1; 540 else 541 return 0; 542} 543 544#ifdef I386_CPU 545/* 546 * i386 only has "invalidate everything" and no SMP to worry about. 547 */ 548PMAP_INLINE void 549pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 550{ 551 552 if (pmap == kernel_pmap || pmap->pm_active) 553 invltlb(); 554} 555 556PMAP_INLINE void 557pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 558{ 559 560 if (pmap == kernel_pmap || pmap->pm_active) 561 invltlb(); 562} 563 564PMAP_INLINE void 565pmap_invalidate_all(pmap_t pmap) 566{ 567 568 if (pmap == kernel_pmap || pmap->pm_active) 569 invltlb(); 570} 571#else /* !I386_CPU */ 572#ifdef SMP 573/* 574 * For SMP, these functions have to use the IPI mechanism for coherence. 575 */ 576void 577pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 578{ 579 u_int cpumask; 580 u_int other_cpus; 581 582 if (smp_started) { 583 if (!(read_eflags() & PSL_I)) 584 panic("%s: interrupts disabled", __func__); 585 mtx_lock_spin(&smp_rv_mtx); 586 } else 587 critical_enter(); 588 /* 589 * We need to disable interrupt preemption but MUST NOT have 590 * interrupts disabled here. 591 * XXX we may need to hold schedlock to get a coherent pm_active 592 * XXX critical sections disable interrupts again 593 */ 594 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 595 invlpg(va); 596 smp_invlpg(va); 597 } else { 598 cpumask = PCPU_GET(cpumask); 599 other_cpus = PCPU_GET(other_cpus); 600 if (pmap->pm_active & cpumask) 601 invlpg(va); 602 if (pmap->pm_active & other_cpus) 603 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 604 } 605 if (smp_started) 606 mtx_unlock_spin(&smp_rv_mtx); 607 else 608 critical_exit(); 609} 610 611void 612pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 613{ 614 u_int cpumask; 615 u_int other_cpus; 616 vm_offset_t addr; 617 618 if (smp_started) { 619 if (!(read_eflags() & PSL_I)) 620 panic("%s: interrupts disabled", __func__); 621 mtx_lock_spin(&smp_rv_mtx); 622 } else 623 critical_enter(); 624 /* 625 * We need to disable interrupt preemption but MUST NOT have 626 * interrupts disabled here. 627 * XXX we may need to hold schedlock to get a coherent pm_active 628 * XXX critical sections disable interrupts again 629 */ 630 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 631 for (addr = sva; addr < eva; addr += PAGE_SIZE) 632 invlpg(addr); 633 smp_invlpg_range(sva, eva); 634 } else { 635 cpumask = PCPU_GET(cpumask); 636 other_cpus = PCPU_GET(other_cpus); 637 if (pmap->pm_active & cpumask) 638 for (addr = sva; addr < eva; addr += PAGE_SIZE) 639 invlpg(addr); 640 if (pmap->pm_active & other_cpus) 641 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 642 sva, eva); 643 } 644 if (smp_started) 645 mtx_unlock_spin(&smp_rv_mtx); 646 else 647 critical_exit(); 648} 649 650void 651pmap_invalidate_all(pmap_t pmap) 652{ 653 u_int cpumask; 654 u_int other_cpus; 655 656 if (smp_started) { 657 if (!(read_eflags() & PSL_I)) 658 panic("%s: interrupts disabled", __func__); 659 mtx_lock_spin(&smp_rv_mtx); 660 } else 661 critical_enter(); 662 /* 663 * We need to disable interrupt preemption but MUST NOT have 664 * interrupts disabled here. 665 * XXX we may need to hold schedlock to get a coherent pm_active 666 * XXX critical sections disable interrupts again 667 */ 668 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 669 invltlb(); 670 smp_invltlb(); 671 } else { 672 cpumask = PCPU_GET(cpumask); 673 other_cpus = PCPU_GET(other_cpus); 674 if (pmap->pm_active & cpumask) 675 invltlb(); 676 if (pmap->pm_active & other_cpus) 677 smp_masked_invltlb(pmap->pm_active & other_cpus); 678 } 679 if (smp_started) 680 mtx_unlock_spin(&smp_rv_mtx); 681 else 682 critical_exit(); 683} 684#else /* !SMP */ 685/* 686 * Normal, non-SMP, 486+ invalidation functions. 687 * We inline these within pmap.c for speed. 688 */ 689PMAP_INLINE void 690pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 691{ 692 693 if (pmap == kernel_pmap || pmap->pm_active) 694 invlpg(va); 695} 696 697PMAP_INLINE void 698pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 699{ 700 vm_offset_t addr; 701 702 if (pmap == kernel_pmap || pmap->pm_active) 703 for (addr = sva; addr < eva; addr += PAGE_SIZE) 704 invlpg(addr); 705} 706 707PMAP_INLINE void 708pmap_invalidate_all(pmap_t pmap) 709{ 710 711 if (pmap == kernel_pmap || pmap->pm_active) 712 invltlb(); 713} 714#endif /* !SMP */ 715#endif /* !I386_CPU */ 716 717/* 718 * Are we current address space or kernel? N.B. We return FALSE when 719 * a pmap's page table is in use because a kernel thread is borrowing 720 * it. The borrowed page table can change spontaneously, making any 721 * dependence on its continued use subject to a race condition. 722 */ 723static __inline int 724pmap_is_current(pmap_t pmap) 725{ 726 727 return (pmap == kernel_pmap || 728 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 729 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 730} 731 732/* 733 * If the given pmap is not the current pmap, Giant must be held. 734 */ 735pt_entry_t * 736pmap_pte(pmap_t pmap, vm_offset_t va) 737{ 738 pd_entry_t newpf; 739 pd_entry_t *pde; 740 741 pde = pmap_pde(pmap, va); 742 if (*pde & PG_PS) 743 return (pde); 744 if (*pde != 0) { 745 /* are we current address space or kernel? */ 746 if (pmap_is_current(pmap)) 747 return (vtopte(va)); 748 GIANT_REQUIRED; 749 newpf = *pde & PG_FRAME; 750 if ((*PMAP2 & PG_FRAME) != newpf) { 751 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; 752 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 753 } 754 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); 755 } 756 return (0); 757} 758 759static __inline void 760invlcaddr(void *caddr) 761{ 762#ifdef I386_CPU 763 invltlb(); 764#else 765 invlpg((u_int)caddr); 766#endif 767} 768 769/* 770 * Super fast pmap_pte routine best used when scanning 771 * the pv lists. This eliminates many coarse-grained 772 * invltlb calls. Note that many of the pv list 773 * scans are across different pmaps. It is very wasteful 774 * to do an entire invltlb for checking a single mapping. 775 * 776 * If the given pmap is not the current pmap, vm_page_queue_mtx 777 * must be held and curthread pinned to a CPU. 778 */ 779static pt_entry_t * 780pmap_pte_quick(pmap_t pmap, vm_offset_t va) 781{ 782 pd_entry_t newpf; 783 pd_entry_t *pde; 784 785 pde = pmap_pde(pmap, va); 786 if (*pde & PG_PS) 787 return (pde); 788 if (*pde != 0) { 789 /* are we current address space or kernel? */ 790 if (pmap_is_current(pmap)) 791 return (vtopte(va)); 792 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 793 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 794 newpf = *pde & PG_FRAME; 795 if ((*PMAP1 & PG_FRAME) != newpf) { 796 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; 797#ifdef SMP 798 PMAP1cpu = PCPU_GET(cpuid); 799#endif 800 invlcaddr(PADDR1); 801 PMAP1changed++; 802 } else 803#ifdef SMP 804 if (PMAP1cpu != PCPU_GET(cpuid)) { 805 PMAP1cpu = PCPU_GET(cpuid); 806 invlcaddr(PADDR1); 807 PMAP1changedcpu++; 808 } else 809#endif 810 PMAP1unchanged++; 811 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); 812 } 813 return (0); 814} 815 816/* 817 * Routine: pmap_extract 818 * Function: 819 * Extract the physical page address associated 820 * with the given map/virtual_address pair. 821 */ 822vm_paddr_t 823pmap_extract(pmap_t pmap, vm_offset_t va) 824{ 825 vm_paddr_t rtval; 826 pt_entry_t *pte; 827 pd_entry_t pde; 828 829 rtval = 0; 830 PMAP_LOCK(pmap); 831 pde = pmap->pm_pdir[va >> PDRSHIFT]; 832 if (pde != 0) { 833 if ((pde & PG_PS) != 0) { 834 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 835 PMAP_UNLOCK(pmap); 836 return rtval; 837 } 838 pte = pmap_pte(pmap, va); 839 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK); 840 } 841 PMAP_UNLOCK(pmap); 842 return (rtval); 843} 844 845/* 846 * Routine: pmap_extract_and_hold 847 * Function: 848 * Atomically extract and hold the physical page 849 * with the given pmap and virtual address pair 850 * if that mapping permits the given protection. 851 */ 852vm_page_t 853pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 854{ 855 pd_entry_t pde; 856 pt_entry_t pte; 857 vm_page_t m; 858 859 m = NULL; 860 vm_page_lock_queues(); 861 PMAP_LOCK(pmap); 862 pde = *pmap_pde(pmap, va); 863 if (pde != 0) { 864 if (pde & PG_PS) { 865 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) { 866 m = PHYS_TO_VM_PAGE((pde & ~PDRMASK) | 867 (va & PDRMASK)); 868 vm_page_hold(m); 869 } 870 } else { 871 sched_pin(); 872 pte = *pmap_pte_quick(pmap, va); 873 if (pte != 0 && 874 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) { 875 m = PHYS_TO_VM_PAGE(pte & PG_FRAME); 876 vm_page_hold(m); 877 } 878 sched_unpin(); 879 } 880 } 881 vm_page_unlock_queues(); 882 PMAP_UNLOCK(pmap); 883 return (m); 884} 885 886/*************************************************** 887 * Low level mapping routines..... 888 ***************************************************/ 889 890/* 891 * Add a wired page to the kva. 892 * Note: not SMP coherent. 893 */ 894PMAP_INLINE void 895pmap_kenter(vm_offset_t va, vm_paddr_t pa) 896{ 897 pt_entry_t *pte; 898 899 pte = vtopte(va); 900 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 901} 902 903/* 904 * Remove a page from the kernel pagetables. 905 * Note: not SMP coherent. 906 */ 907PMAP_INLINE void 908pmap_kremove(vm_offset_t va) 909{ 910 pt_entry_t *pte; 911 912 pte = vtopte(va); 913 pte_clear(pte); 914} 915 916/* 917 * Used to map a range of physical addresses into kernel 918 * virtual address space. 919 * 920 * The value passed in '*virt' is a suggested virtual address for 921 * the mapping. Architectures which can support a direct-mapped 922 * physical to virtual region can return the appropriate address 923 * within that region, leaving '*virt' unchanged. Other 924 * architectures should map the pages starting at '*virt' and 925 * update '*virt' with the first usable address after the mapped 926 * region. 927 */ 928vm_offset_t 929pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 930{ 931 vm_offset_t va, sva; 932 933 va = sva = *virt; 934 while (start < end) { 935 pmap_kenter(va, start); 936 va += PAGE_SIZE; 937 start += PAGE_SIZE; 938 } 939 pmap_invalidate_range(kernel_pmap, sva, va); 940 *virt = va; 941 return (sva); 942} 943 944 945/* 946 * Add a list of wired pages to the kva 947 * this routine is only used for temporary 948 * kernel mappings that do not need to have 949 * page modification or references recorded. 950 * Note that old mappings are simply written 951 * over. The page *must* be wired. 952 * Note: SMP coherent. Uses a ranged shootdown IPI. 953 */ 954void 955pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 956{ 957 vm_offset_t va; 958 959 va = sva; 960 while (count-- > 0) { 961 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 962 va += PAGE_SIZE; 963 m++; 964 } 965 pmap_invalidate_range(kernel_pmap, sva, va); 966} 967 968/* 969 * This routine tears out page mappings from the 970 * kernel -- it is meant only for temporary mappings. 971 * Note: SMP coherent. Uses a ranged shootdown IPI. 972 */ 973void 974pmap_qremove(vm_offset_t sva, int count) 975{ 976 vm_offset_t va; 977 978 va = sva; 979 while (count-- > 0) { 980 pmap_kremove(va); 981 va += PAGE_SIZE; 982 } 983 pmap_invalidate_range(kernel_pmap, sva, va); 984} 985 986/*************************************************** 987 * Page table page management routines..... 988 ***************************************************/ 989 990/* 991 * This routine unholds page table pages, and if the hold count 992 * drops to zero, then it decrements the wire count. 993 */ 994static PMAP_INLINE int 995pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 996{ 997 998 vm_page_unhold(m); 999 if (m->hold_count == 0) 1000 return _pmap_unwire_pte_hold(pmap, m); 1001 else 1002 return 0; 1003} 1004 1005static int 1006_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1007{ 1008 vm_offset_t pteva; 1009 1010 /* 1011 * unmap the page table page 1012 */ 1013 pmap->pm_pdir[m->pindex] = 0; 1014 --pmap->pm_stats.resident_count; 1015 /* 1016 * We never unwire a kernel page table page, making a 1017 * check for the kernel_pmap unnecessary. 1018 */ 1019 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)) { 1020 /* 1021 * Do an invltlb to make the invalidated mapping 1022 * take effect immediately. 1023 */ 1024 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1025 pmap_invalidate_page(pmap, pteva); 1026 } 1027 1028 /* 1029 * If the page is finally unwired, simply free it. 1030 */ 1031 --m->wire_count; 1032 if (m->wire_count == 0) { 1033 vm_page_free_zero(m); 1034 atomic_subtract_int(&cnt.v_wire_count, 1); 1035 } 1036 return 1; 1037} 1038 1039/* 1040 * After removing a page table entry, this routine is used to 1041 * conditionally free the page, and manage the hold/wire counts. 1042 */ 1043static int 1044pmap_unuse_pt(pmap_t pmap, vm_offset_t va) 1045{ 1046 pd_entry_t ptepde; 1047 vm_page_t mpte; 1048 1049 if (va >= VM_MAXUSER_ADDRESS) 1050 return 0; 1051 ptepde = *pmap_pde(pmap, va); 1052 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); 1053 return pmap_unwire_pte_hold(pmap, mpte); 1054} 1055 1056void 1057pmap_pinit0(pmap) 1058 struct pmap *pmap; 1059{ 1060 1061 PMAP_LOCK_INIT(pmap); 1062 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1063#ifdef PAE 1064 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1065#endif 1066 pmap->pm_active = 0; 1067 PCPU_SET(curpmap, pmap); 1068 TAILQ_INIT(&pmap->pm_pvlist); 1069 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1070 mtx_lock_spin(&allpmaps_lock); 1071 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1072 mtx_unlock_spin(&allpmaps_lock); 1073} 1074 1075/* 1076 * Initialize a preallocated and zeroed pmap structure, 1077 * such as one in a vmspace structure. 1078 */ 1079void 1080pmap_pinit(pmap) 1081 register struct pmap *pmap; 1082{ 1083 vm_page_t m, ptdpg[NPGPTD]; 1084 vm_paddr_t pa; 1085 static int color; 1086 int i; 1087 1088 PMAP_LOCK_INIT(pmap); 1089 1090 /* 1091 * No need to allocate page table space yet but we do need a valid 1092 * page directory table. 1093 */ 1094 if (pmap->pm_pdir == NULL) { 1095 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1096 NBPTD); 1097#ifdef PAE 1098 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1099 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1100 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1101 ("pmap_pinit: pdpt misaligned")); 1102 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1103 ("pmap_pinit: pdpt above 4g")); 1104#endif 1105 } 1106 1107 /* 1108 * allocate the page directory page(s) 1109 */ 1110 for (i = 0; i < NPGPTD;) { 1111 m = vm_page_alloc(NULL, color++, 1112 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1113 VM_ALLOC_ZERO); 1114 if (m == NULL) 1115 VM_WAIT; 1116 else { 1117 ptdpg[i++] = m; 1118 } 1119 } 1120 1121 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1122 1123 for (i = 0; i < NPGPTD; i++) { 1124 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1125 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1126 } 1127 1128 mtx_lock_spin(&allpmaps_lock); 1129 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1130 mtx_unlock_spin(&allpmaps_lock); 1131 /* Wire in kernel global address entries. */ 1132 /* XXX copies current process, does not fill in MPPTDI */ 1133 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1134#ifdef SMP 1135 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1136#endif 1137 1138 /* install self-referential address mapping entry(s) */ 1139 for (i = 0; i < NPGPTD; i++) { 1140 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1141 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1142#ifdef PAE 1143 pmap->pm_pdpt[i] = pa | PG_V; 1144#endif 1145 } 1146 1147 pmap->pm_active = 0; 1148 TAILQ_INIT(&pmap->pm_pvlist); 1149 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1150} 1151 1152/* 1153 * this routine is called if the page table page is not 1154 * mapped correctly. 1155 */ 1156static vm_page_t 1157_pmap_allocpte(pmap, ptepindex) 1158 pmap_t pmap; 1159 unsigned ptepindex; 1160{ 1161 vm_paddr_t ptepa; 1162 vm_page_t m; 1163 1164 /* 1165 * Allocate a page table page. 1166 */ 1167 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1168 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1169 PMAP_UNLOCK(pmap); 1170 vm_page_unlock_queues(); 1171 VM_WAIT; 1172 vm_page_lock_queues(); 1173 PMAP_LOCK(pmap); 1174 1175 /* 1176 * Indicate the need to retry. While waiting, the page table 1177 * page may have been allocated. 1178 */ 1179 return (NULL); 1180 } 1181 if ((m->flags & PG_ZERO) == 0) 1182 pmap_zero_page(m); 1183 1184 KASSERT(m->queue == PQ_NONE, 1185 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1186 1187 /* 1188 * Increment the hold count for the page table page 1189 * (denoting a new mapping.) 1190 */ 1191 m->hold_count++; 1192 1193 /* 1194 * Map the pagetable page into the process address space, if 1195 * it isn't already there. 1196 */ 1197 1198 pmap->pm_stats.resident_count++; 1199 1200 ptepa = VM_PAGE_TO_PHYS(m); 1201 pmap->pm_pdir[ptepindex] = 1202 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1203 1204 return m; 1205} 1206 1207static vm_page_t 1208pmap_allocpte(pmap_t pmap, vm_offset_t va) 1209{ 1210 unsigned ptepindex; 1211 pd_entry_t ptepa; 1212 vm_page_t m; 1213 1214 /* 1215 * Calculate pagetable page index 1216 */ 1217 ptepindex = va >> PDRSHIFT; 1218retry: 1219 /* 1220 * Get the page directory entry 1221 */ 1222 ptepa = pmap->pm_pdir[ptepindex]; 1223 1224 /* 1225 * This supports switching from a 4MB page to a 1226 * normal 4K page. 1227 */ 1228 if (ptepa & PG_PS) { 1229 pmap->pm_pdir[ptepindex] = 0; 1230 ptepa = 0; 1231 pmap_invalidate_all(kernel_pmap); 1232 } 1233 1234 /* 1235 * If the page table page is mapped, we just increment the 1236 * hold count, and activate it. 1237 */ 1238 if (ptepa) { 1239 m = PHYS_TO_VM_PAGE(ptepa); 1240 m->hold_count++; 1241 } else { 1242 /* 1243 * Here if the pte page isn't mapped, or if it has 1244 * been deallocated. 1245 */ 1246 m = _pmap_allocpte(pmap, ptepindex); 1247 if (m == NULL) 1248 goto retry; 1249 } 1250 return (m); 1251} 1252 1253 1254/*************************************************** 1255* Pmap allocation/deallocation routines. 1256 ***************************************************/ 1257 1258#ifdef SMP 1259/* 1260 * Deal with a SMP shootdown of other users of the pmap that we are 1261 * trying to dispose of. This can be a bit hairy. 1262 */ 1263static u_int *lazymask; 1264static u_int lazyptd; 1265static volatile u_int lazywait; 1266 1267void pmap_lazyfix_action(void); 1268 1269void 1270pmap_lazyfix_action(void) 1271{ 1272 u_int mymask = PCPU_GET(cpumask); 1273 1274 if (rcr3() == lazyptd) 1275 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1276 atomic_clear_int(lazymask, mymask); 1277 atomic_store_rel_int(&lazywait, 1); 1278} 1279 1280static void 1281pmap_lazyfix_self(u_int mymask) 1282{ 1283 1284 if (rcr3() == lazyptd) 1285 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1286 atomic_clear_int(lazymask, mymask); 1287} 1288 1289 1290static void 1291pmap_lazyfix(pmap_t pmap) 1292{ 1293 u_int mymask = PCPU_GET(cpumask); 1294 u_int mask; 1295 register u_int spins; 1296 1297 while ((mask = pmap->pm_active) != 0) { 1298 spins = 50000000; 1299 mask = mask & -mask; /* Find least significant set bit */ 1300 mtx_lock_spin(&smp_rv_mtx); 1301#ifdef PAE 1302 lazyptd = vtophys(pmap->pm_pdpt); 1303#else 1304 lazyptd = vtophys(pmap->pm_pdir); 1305#endif 1306 if (mask == mymask) { 1307 lazymask = &pmap->pm_active; 1308 pmap_lazyfix_self(mymask); 1309 } else { 1310 atomic_store_rel_int((u_int *)&lazymask, 1311 (u_int)&pmap->pm_active); 1312 atomic_store_rel_int(&lazywait, 0); 1313 ipi_selected(mask, IPI_LAZYPMAP); 1314 while (lazywait == 0) { 1315 ia32_pause(); 1316 if (--spins == 0) 1317 break; 1318 } 1319 } 1320 mtx_unlock_spin(&smp_rv_mtx); 1321 if (spins == 0) 1322 printf("pmap_lazyfix: spun for 50000000\n"); 1323 } 1324} 1325 1326#else /* SMP */ 1327 1328/* 1329 * Cleaning up on uniprocessor is easy. For various reasons, we're 1330 * unlikely to have to even execute this code, including the fact 1331 * that the cleanup is deferred until the parent does a wait(2), which 1332 * means that another userland process has run. 1333 */ 1334static void 1335pmap_lazyfix(pmap_t pmap) 1336{ 1337 u_int cr3; 1338 1339 cr3 = vtophys(pmap->pm_pdir); 1340 if (cr3 == rcr3()) { 1341 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1342 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1343 } 1344} 1345#endif /* SMP */ 1346 1347/* 1348 * Release any resources held by the given physical map. 1349 * Called when a pmap initialized by pmap_pinit is being released. 1350 * Should only be called if the map contains no valid mappings. 1351 */ 1352void 1353pmap_release(pmap_t pmap) 1354{ 1355 vm_page_t m, ptdpg[NPGPTD]; 1356 int i; 1357 1358 KASSERT(pmap->pm_stats.resident_count == 0, 1359 ("pmap_release: pmap resident count %ld != 0", 1360 pmap->pm_stats.resident_count)); 1361 1362 pmap_lazyfix(pmap); 1363 mtx_lock_spin(&allpmaps_lock); 1364 LIST_REMOVE(pmap, pm_list); 1365 mtx_unlock_spin(&allpmaps_lock); 1366 1367 for (i = 0; i < NPGPTD; i++) 1368 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1369 1370 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1371 sizeof(*pmap->pm_pdir)); 1372#ifdef SMP 1373 pmap->pm_pdir[MPPTDI] = 0; 1374#endif 1375 1376 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1377 1378 vm_page_lock_queues(); 1379 for (i = 0; i < NPGPTD; i++) { 1380 m = ptdpg[i]; 1381#ifdef PAE 1382 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1383 ("pmap_release: got wrong ptd page")); 1384#endif 1385 m->wire_count--; 1386 atomic_subtract_int(&cnt.v_wire_count, 1); 1387 vm_page_free_zero(m); 1388 } 1389 vm_page_unlock_queues(); 1390 PMAP_LOCK_DESTROY(pmap); 1391} 1392 1393static int 1394kvm_size(SYSCTL_HANDLER_ARGS) 1395{ 1396 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1397 1398 return sysctl_handle_long(oidp, &ksize, 0, req); 1399} 1400SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1401 0, 0, kvm_size, "IU", "Size of KVM"); 1402 1403static int 1404kvm_free(SYSCTL_HANDLER_ARGS) 1405{ 1406 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1407 1408 return sysctl_handle_long(oidp, &kfree, 0, req); 1409} 1410SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1411 0, 0, kvm_free, "IU", "Amount of KVM free"); 1412 1413/* 1414 * grow the number of kernel page table entries, if needed 1415 */ 1416void 1417pmap_growkernel(vm_offset_t addr) 1418{ 1419 struct pmap *pmap; 1420 vm_paddr_t ptppaddr; 1421 vm_page_t nkpg; 1422 pd_entry_t newpdir; 1423 pt_entry_t *pde; 1424 1425 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1426 if (kernel_vm_end == 0) { 1427 kernel_vm_end = KERNBASE; 1428 nkpt = 0; 1429 while (pdir_pde(PTD, kernel_vm_end)) { 1430 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1431 nkpt++; 1432 } 1433 } 1434 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1435 while (kernel_vm_end < addr) { 1436 if (pdir_pde(PTD, kernel_vm_end)) { 1437 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1438 continue; 1439 } 1440 1441 /* 1442 * This index is bogus, but out of the way 1443 */ 1444 nkpg = vm_page_alloc(NULL, nkpt, 1445 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1446 if (!nkpg) 1447 panic("pmap_growkernel: no memory to grow kernel"); 1448 1449 nkpt++; 1450 1451 pmap_zero_page(nkpg); 1452 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1453 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1454 pdir_pde(PTD, kernel_vm_end) = newpdir; 1455 1456 mtx_lock_spin(&allpmaps_lock); 1457 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1458 pde = pmap_pde(pmap, kernel_vm_end); 1459 pde_store(pde, newpdir); 1460 } 1461 mtx_unlock_spin(&allpmaps_lock); 1462 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1463 } 1464} 1465 1466 1467/*************************************************** 1468 * page management routines. 1469 ***************************************************/ 1470 1471/* 1472 * free the pv_entry back to the free list 1473 */ 1474static PMAP_INLINE void 1475free_pv_entry(pv_entry_t pv) 1476{ 1477 pv_entry_count--; 1478 uma_zfree(pvzone, pv); 1479} 1480 1481/* 1482 * get a new pv_entry, allocating a block from the system 1483 * when needed. 1484 * the memory allocation is performed bypassing the malloc code 1485 * because of the possibility of allocations at interrupt time. 1486 */ 1487static pv_entry_t 1488get_pv_entry(void) 1489{ 1490 pv_entry_count++; 1491 if (pv_entry_high_water && 1492 (pv_entry_count > pv_entry_high_water) && 1493 (pmap_pagedaemon_waken == 0)) { 1494 pmap_pagedaemon_waken = 1; 1495 wakeup (&vm_pages_needed); 1496 } 1497 return uma_zalloc(pvzone, M_NOWAIT); 1498} 1499 1500 1501static int 1502pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1503{ 1504 pv_entry_t pv; 1505 int rtval; 1506 1507 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1508 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1509 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1510 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1511 if (pmap == pv->pv_pmap && va == pv->pv_va) 1512 break; 1513 } 1514 } else { 1515 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1516 if (va == pv->pv_va) 1517 break; 1518 } 1519 } 1520 1521 rtval = 0; 1522 if (pv) { 1523 rtval = pmap_unuse_pt(pmap, va); 1524 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1525 m->md.pv_list_count--; 1526 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1527 vm_page_flag_clear(m, PG_WRITEABLE); 1528 1529 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1530 free_pv_entry(pv); 1531 } 1532 1533 return rtval; 1534} 1535 1536/* 1537 * Create a pv entry for page at pa for 1538 * (pmap, va). 1539 */ 1540static void 1541pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) 1542{ 1543 pv_entry_t pv; 1544 1545 pv = get_pv_entry(); 1546 pv->pv_va = va; 1547 pv->pv_pmap = pmap; 1548 1549 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1550 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1551 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1552 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1553 m->md.pv_list_count++; 1554} 1555 1556/* 1557 * pmap_remove_pte: do the things to unmap a page in a process 1558 */ 1559static int 1560pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1561{ 1562 pt_entry_t oldpte; 1563 vm_page_t m; 1564 1565 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1566 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1567 oldpte = pte_load_clear(ptq); 1568 if (oldpte & PG_W) 1569 pmap->pm_stats.wired_count -= 1; 1570 /* 1571 * Machines that don't support invlpg, also don't support 1572 * PG_G. 1573 */ 1574 if (oldpte & PG_G) 1575 pmap_invalidate_page(kernel_pmap, va); 1576 pmap->pm_stats.resident_count -= 1; 1577 if (oldpte & PG_MANAGED) { 1578 m = PHYS_TO_VM_PAGE(oldpte); 1579 if (oldpte & PG_M) { 1580#if defined(PMAP_DIAGNOSTIC) 1581 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1582 printf( 1583 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1584 va, oldpte); 1585 } 1586#endif 1587 if (pmap_track_modified(va)) 1588 vm_page_dirty(m); 1589 } 1590 if (oldpte & PG_A) 1591 vm_page_flag_set(m, PG_REFERENCED); 1592 return pmap_remove_entry(pmap, m, va); 1593 } else { 1594 return pmap_unuse_pt(pmap, va); 1595 } 1596} 1597 1598/* 1599 * Remove a single page from a process address space 1600 */ 1601static void 1602pmap_remove_page(pmap_t pmap, vm_offset_t va) 1603{ 1604 pt_entry_t *pte; 1605 1606 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1607 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 1608 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1609 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) 1610 return; 1611 pmap_remove_pte(pmap, pte, va); 1612 pmap_invalidate_page(pmap, va); 1613} 1614 1615/* 1616 * Remove the given range of addresses from the specified map. 1617 * 1618 * It is assumed that the start and end are properly 1619 * rounded to the page size. 1620 */ 1621void 1622pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1623{ 1624 vm_offset_t pdnxt; 1625 pd_entry_t ptpaddr; 1626 pt_entry_t *pte; 1627 int anyvalid; 1628 1629 /* 1630 * Perform an unsynchronized read. This is, however, safe. 1631 */ 1632 if (pmap->pm_stats.resident_count == 0) 1633 return; 1634 1635 vm_page_lock_queues(); 1636 sched_pin(); 1637 PMAP_LOCK(pmap); 1638 1639 /* 1640 * special handling of removing one page. a very 1641 * common operation and easy to short circuit some 1642 * code. 1643 */ 1644 if ((sva + PAGE_SIZE == eva) && 1645 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1646 pmap_remove_page(pmap, sva); 1647 goto out; 1648 } 1649 1650 anyvalid = 0; 1651 1652 for (; sva < eva; sva = pdnxt) { 1653 unsigned pdirindex; 1654 1655 /* 1656 * Calculate index for next page table. 1657 */ 1658 pdnxt = (sva + NBPDR) & ~PDRMASK; 1659 if (pmap->pm_stats.resident_count == 0) 1660 break; 1661 1662 pdirindex = sva >> PDRSHIFT; 1663 ptpaddr = pmap->pm_pdir[pdirindex]; 1664 1665 /* 1666 * Weed out invalid mappings. Note: we assume that the page 1667 * directory table is always allocated, and in kernel virtual. 1668 */ 1669 if (ptpaddr == 0) 1670 continue; 1671 1672 /* 1673 * Check for large page. 1674 */ 1675 if ((ptpaddr & PG_PS) != 0) { 1676 pmap->pm_pdir[pdirindex] = 0; 1677 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1678 anyvalid = 1; 1679 continue; 1680 } 1681 1682 /* 1683 * Limit our scan to either the end of the va represented 1684 * by the current page table page, or to the end of the 1685 * range being removed. 1686 */ 1687 if (pdnxt > eva) 1688 pdnxt = eva; 1689 1690 for (; sva != pdnxt; sva += PAGE_SIZE) { 1691 if ((pte = pmap_pte_quick(pmap, sva)) == NULL || 1692 *pte == 0) 1693 continue; 1694 anyvalid = 1; 1695 if (pmap_remove_pte(pmap, pte, sva)) 1696 break; 1697 } 1698 } 1699 1700 if (anyvalid) 1701 pmap_invalidate_all(pmap); 1702out: 1703 sched_unpin(); 1704 vm_page_unlock_queues(); 1705 PMAP_UNLOCK(pmap); 1706} 1707 1708/* 1709 * Routine: pmap_remove_all 1710 * Function: 1711 * Removes this physical page from 1712 * all physical maps in which it resides. 1713 * Reflects back modify bits to the pager. 1714 * 1715 * Notes: 1716 * Original versions of this routine were very 1717 * inefficient because they iteratively called 1718 * pmap_remove (slow...) 1719 */ 1720 1721void 1722pmap_remove_all(vm_page_t m) 1723{ 1724 register pv_entry_t pv; 1725 pt_entry_t *pte, tpte; 1726 1727#if defined(PMAP_DIAGNOSTIC) 1728 /* 1729 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1730 */ 1731 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1732 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1733 VM_PAGE_TO_PHYS(m)); 1734 } 1735#endif 1736 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1737 sched_pin(); 1738 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1739 PMAP_LOCK(pv->pv_pmap); 1740 pv->pv_pmap->pm_stats.resident_count--; 1741 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1742 tpte = pte_load_clear(pte); 1743 if (tpte & PG_W) 1744 pv->pv_pmap->pm_stats.wired_count--; 1745 if (tpte & PG_A) 1746 vm_page_flag_set(m, PG_REFERENCED); 1747 1748 /* 1749 * Update the vm_page_t clean and reference bits. 1750 */ 1751 if (tpte & PG_M) { 1752#if defined(PMAP_DIAGNOSTIC) 1753 if (pmap_nw_modified((pt_entry_t) tpte)) { 1754 printf( 1755 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1756 pv->pv_va, tpte); 1757 } 1758#endif 1759 if (pmap_track_modified(pv->pv_va)) 1760 vm_page_dirty(m); 1761 } 1762 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1763 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1764 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1765 m->md.pv_list_count--; 1766 pmap_unuse_pt(pv->pv_pmap, pv->pv_va); 1767 PMAP_UNLOCK(pv->pv_pmap); 1768 free_pv_entry(pv); 1769 } 1770 vm_page_flag_clear(m, PG_WRITEABLE); 1771 sched_unpin(); 1772} 1773 1774/* 1775 * Set the physical protection on the 1776 * specified range of this map as requested. 1777 */ 1778void 1779pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1780{ 1781 vm_offset_t pdnxt; 1782 pd_entry_t ptpaddr; 1783 int anychanged; 1784 1785 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1786 pmap_remove(pmap, sva, eva); 1787 return; 1788 } 1789 1790 if (prot & VM_PROT_WRITE) 1791 return; 1792 1793 anychanged = 0; 1794 1795 vm_page_lock_queues(); 1796 sched_pin(); 1797 PMAP_LOCK(pmap); 1798 for (; sva < eva; sva = pdnxt) { 1799 unsigned pdirindex; 1800 1801 pdnxt = (sva + NBPDR) & ~PDRMASK; 1802 1803 pdirindex = sva >> PDRSHIFT; 1804 ptpaddr = pmap->pm_pdir[pdirindex]; 1805 1806 /* 1807 * Weed out invalid mappings. Note: we assume that the page 1808 * directory table is always allocated, and in kernel virtual. 1809 */ 1810 if (ptpaddr == 0) 1811 continue; 1812 1813 /* 1814 * Check for large page. 1815 */ 1816 if ((ptpaddr & PG_PS) != 0) { 1817 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1818 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1819 anychanged = 1; 1820 continue; 1821 } 1822 1823 if (pdnxt > eva) 1824 pdnxt = eva; 1825 1826 for (; sva != pdnxt; sva += PAGE_SIZE) { 1827 pt_entry_t pbits; 1828 pt_entry_t *pte; 1829 vm_page_t m; 1830 1831 if ((pte = pmap_pte_quick(pmap, sva)) == NULL) 1832 continue; 1833 pbits = *pte; 1834 if (pbits & PG_MANAGED) { 1835 m = NULL; 1836 if (pbits & PG_A) { 1837 m = PHYS_TO_VM_PAGE(pbits); 1838 vm_page_flag_set(m, PG_REFERENCED); 1839 pbits &= ~PG_A; 1840 } 1841 if ((pbits & PG_M) != 0 && 1842 pmap_track_modified(sva)) { 1843 if (m == NULL) 1844 m = PHYS_TO_VM_PAGE(pbits); 1845 vm_page_dirty(m); 1846 pbits &= ~PG_M; 1847 } 1848 } 1849 1850 pbits &= ~PG_RW; 1851 1852 if (pbits != *pte) { 1853 pte_store(pte, pbits); 1854 anychanged = 1; 1855 } 1856 } 1857 } 1858 if (anychanged) 1859 pmap_invalidate_all(pmap); 1860 sched_unpin(); 1861 vm_page_unlock_queues(); 1862 PMAP_UNLOCK(pmap); 1863} 1864 1865/* 1866 * Insert the given physical page (p) at 1867 * the specified virtual address (v) in the 1868 * target physical map with the protection requested. 1869 * 1870 * If specified, the page will be wired down, meaning 1871 * that the related pte can not be reclaimed. 1872 * 1873 * NB: This is the only routine which MAY NOT lazy-evaluate 1874 * or lose information. That is, this routine must actually 1875 * insert this page into the given map NOW. 1876 */ 1877void 1878pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1879 boolean_t wired) 1880{ 1881 vm_paddr_t pa; 1882 register pt_entry_t *pte; 1883 vm_paddr_t opa; 1884 pt_entry_t origpte, newpte; 1885 vm_page_t mpte; 1886 1887 va &= PG_FRAME; 1888#ifdef PMAP_DIAGNOSTIC 1889 if (va > VM_MAX_KERNEL_ADDRESS) 1890 panic("pmap_enter: toobig"); 1891 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1892 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1893#endif 1894 1895 mpte = NULL; 1896 1897 vm_page_lock_queues(); 1898 PMAP_LOCK(pmap); 1899 sched_pin(); 1900 1901 /* 1902 * In the case that a page table page is not 1903 * resident, we are creating it here. 1904 */ 1905 if (va < VM_MAXUSER_ADDRESS) { 1906 mpte = pmap_allocpte(pmap, va); 1907 } 1908#if 0 && defined(PMAP_DIAGNOSTIC) 1909 else { 1910 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1911 origpte = *pdeaddr; 1912 if ((origpte & PG_V) == 0) { 1913 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1914 pmap->pm_pdir[PTDPTDI], origpte, va); 1915 } 1916 } 1917#endif 1918 1919 pte = pmap_pte_quick(pmap, va); 1920 1921 /* 1922 * Page Directory table entry not valid, we need a new PT page 1923 */ 1924 if (pte == NULL) { 1925 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1926 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1927 } 1928 1929 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 1930 origpte = *pte; 1931 opa = origpte & PG_FRAME; 1932 1933 if (origpte & PG_PS) { 1934 /* 1935 * Yes, I know this will truncate upper address bits for PAE, 1936 * but I'm actually more interested in the lower bits 1937 */ 1938 printf("pmap_enter: va %p, pte %p, origpte %p\n", 1939 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 1940 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1941 } 1942 1943 /* 1944 * Mapping has not changed, must be protection or wiring change. 1945 */ 1946 if (origpte && (opa == pa)) { 1947 /* 1948 * Wiring change, just update stats. We don't worry about 1949 * wiring PT pages as they remain resident as long as there 1950 * are valid mappings in them. Hence, if a user page is wired, 1951 * the PT page will be also. 1952 */ 1953 if (wired && ((origpte & PG_W) == 0)) 1954 pmap->pm_stats.wired_count++; 1955 else if (!wired && (origpte & PG_W)) 1956 pmap->pm_stats.wired_count--; 1957 1958#if defined(PMAP_DIAGNOSTIC) 1959 if (pmap_nw_modified((pt_entry_t) origpte)) { 1960 printf( 1961 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 1962 va, origpte); 1963 } 1964#endif 1965 1966 /* 1967 * Remove extra pte reference 1968 */ 1969 if (mpte) 1970 mpte->hold_count--; 1971 1972 /* 1973 * We might be turning off write access to the page, 1974 * so we go ahead and sense modify status. 1975 */ 1976 if (origpte & PG_MANAGED) { 1977 if ((origpte & PG_M) && pmap_track_modified(va)) { 1978 vm_page_t om; 1979 om = PHYS_TO_VM_PAGE(opa); 1980 vm_page_dirty(om); 1981 } 1982 pa |= PG_MANAGED; 1983 } 1984 goto validate; 1985 } 1986 /* 1987 * Mapping has changed, invalidate old range and fall through to 1988 * handle validating new mapping. 1989 */ 1990 if (opa) { 1991 int err; 1992 err = pmap_remove_pte(pmap, pte, va); 1993 if (err) 1994 panic("pmap_enter: pte vanished, va: 0x%x", va); 1995 } 1996 1997 /* 1998 * Enter on the PV list if part of our managed memory. Note that we 1999 * raise IPL while manipulating pv_table since pmap_enter can be 2000 * called at interrupt time. 2001 */ 2002 if (pmap_initialized && 2003 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2004 pmap_insert_entry(pmap, va, m); 2005 pa |= PG_MANAGED; 2006 } 2007 2008 /* 2009 * Increment counters 2010 */ 2011 pmap->pm_stats.resident_count++; 2012 if (wired) 2013 pmap->pm_stats.wired_count++; 2014 2015validate: 2016 /* 2017 * Now validate mapping with desired protection/wiring. 2018 */ 2019 newpte = (pt_entry_t)(pa | PG_V); 2020 if ((prot & VM_PROT_WRITE) != 0) 2021 newpte |= PG_RW; 2022 if (wired) 2023 newpte |= PG_W; 2024 if (va < VM_MAXUSER_ADDRESS) 2025 newpte |= PG_U; 2026 if (pmap == kernel_pmap) 2027 newpte |= pgeflag; 2028 2029 /* 2030 * if the mapping or permission bits are different, we need 2031 * to update the pte. 2032 */ 2033 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2034 pte_store(pte, newpte | PG_A); 2035 /*if (origpte)*/ { 2036 pmap_invalidate_page(pmap, va); 2037 } 2038 } 2039 sched_unpin(); 2040 vm_page_unlock_queues(); 2041 PMAP_UNLOCK(pmap); 2042} 2043 2044/* 2045 * this code makes some *MAJOR* assumptions: 2046 * 1. Current pmap & pmap exists. 2047 * 2. Not wired. 2048 * 3. Read access. 2049 * 4. No page table pages. 2050 * 5. Tlbflush is deferred to calling procedure. 2051 * 6. Page IS managed. 2052 * but is *MUCH* faster than pmap_enter... 2053 */ 2054 2055vm_page_t 2056pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2057{ 2058 pt_entry_t *pte; 2059 vm_paddr_t pa; 2060 2061 vm_page_lock_queues(); 2062 PMAP_LOCK(pmap); 2063 2064 /* 2065 * In the case that a page table page is not 2066 * resident, we are creating it here. 2067 */ 2068 if (va < VM_MAXUSER_ADDRESS) { 2069 unsigned ptepindex; 2070 pd_entry_t ptepa; 2071 2072 /* 2073 * Calculate pagetable page index 2074 */ 2075 ptepindex = va >> PDRSHIFT; 2076 if (mpte && (mpte->pindex == ptepindex)) { 2077 mpte->hold_count++; 2078 } else { 2079retry: 2080 /* 2081 * Get the page directory entry 2082 */ 2083 ptepa = pmap->pm_pdir[ptepindex]; 2084 2085 /* 2086 * If the page table page is mapped, we just increment 2087 * the hold count, and activate it. 2088 */ 2089 if (ptepa) { 2090 if (ptepa & PG_PS) 2091 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2092 mpte = PHYS_TO_VM_PAGE(ptepa); 2093 mpte->hold_count++; 2094 } else { 2095 mpte = _pmap_allocpte(pmap, ptepindex); 2096 if (mpte == NULL) 2097 goto retry; 2098 } 2099 } 2100 } else { 2101 mpte = NULL; 2102 } 2103 2104 /* 2105 * This call to vtopte makes the assumption that we are 2106 * entering the page into the current pmap. In order to support 2107 * quick entry into any pmap, one would likely use pmap_pte_quick. 2108 * But that isn't as quick as vtopte. 2109 */ 2110 pte = vtopte(va); 2111 if (*pte) { 2112 if (mpte != NULL) { 2113 pmap_unwire_pte_hold(pmap, mpte); 2114 mpte = NULL; 2115 } 2116 goto out; 2117 } 2118 2119 /* 2120 * Enter on the PV list if part of our managed memory. Note that we 2121 * raise IPL while manipulating pv_table since pmap_enter can be 2122 * called at interrupt time. 2123 */ 2124 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2125 pmap_insert_entry(pmap, va, m); 2126 2127 /* 2128 * Increment counters 2129 */ 2130 pmap->pm_stats.resident_count++; 2131 2132 pa = VM_PAGE_TO_PHYS(m); 2133 2134 /* 2135 * Now validate mapping with RO protection 2136 */ 2137 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2138 pte_store(pte, pa | PG_V | PG_U); 2139 else 2140 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2141out: 2142 vm_page_unlock_queues(); 2143 PMAP_UNLOCK(pmap); 2144 return mpte; 2145} 2146 2147/* 2148 * Make a temporary mapping for a physical address. This is only intended 2149 * to be used for panic dumps. 2150 */ 2151void * 2152pmap_kenter_temporary(vm_paddr_t pa, int i) 2153{ 2154 vm_offset_t va; 2155 2156 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2157 pmap_kenter(va, pa); 2158#ifndef I386_CPU 2159 invlpg(va); 2160#else 2161 invltlb(); 2162#endif 2163 return ((void *)crashdumpmap); 2164} 2165 2166/* 2167 * This code maps large physical mmap regions into the 2168 * processor address space. Note that some shortcuts 2169 * are taken, but the code works. 2170 */ 2171void 2172pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2173 vm_object_t object, vm_pindex_t pindex, 2174 vm_size_t size) 2175{ 2176 vm_page_t p; 2177 2178 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2179 KASSERT(object->type == OBJT_DEVICE, 2180 ("pmap_object_init_pt: non-device object")); 2181 if (pseflag && 2182 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2183 int i; 2184 vm_page_t m[1]; 2185 unsigned int ptepindex; 2186 int npdes; 2187 pd_entry_t ptepa; 2188 2189 PMAP_LOCK(pmap); 2190 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2191 goto out; 2192 PMAP_UNLOCK(pmap); 2193retry: 2194 p = vm_page_lookup(object, pindex); 2195 if (p != NULL) { 2196 vm_page_lock_queues(); 2197 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2198 goto retry; 2199 } else { 2200 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2201 if (p == NULL) 2202 return; 2203 m[0] = p; 2204 2205 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2206 vm_page_lock_queues(); 2207 vm_page_free(p); 2208 vm_page_unlock_queues(); 2209 return; 2210 } 2211 2212 p = vm_page_lookup(object, pindex); 2213 vm_page_lock_queues(); 2214 vm_page_wakeup(p); 2215 } 2216 vm_page_unlock_queues(); 2217 2218 ptepa = VM_PAGE_TO_PHYS(p); 2219 if (ptepa & (NBPDR - 1)) 2220 return; 2221 2222 p->valid = VM_PAGE_BITS_ALL; 2223 2224 PMAP_LOCK(pmap); 2225 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2226 npdes = size >> PDRSHIFT; 2227 for(i = 0; i < npdes; i++) { 2228 pde_store(&pmap->pm_pdir[ptepindex], 2229 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2230 ptepa += NBPDR; 2231 ptepindex += 1; 2232 } 2233 pmap_invalidate_all(pmap); 2234out: 2235 PMAP_UNLOCK(pmap); 2236 } 2237} 2238 2239/* 2240 * Routine: pmap_change_wiring 2241 * Function: Change the wiring attribute for a map/virtual-address 2242 * pair. 2243 * In/out conditions: 2244 * The mapping must already exist in the pmap. 2245 */ 2246void 2247pmap_change_wiring(pmap, va, wired) 2248 register pmap_t pmap; 2249 vm_offset_t va; 2250 boolean_t wired; 2251{ 2252 register pt_entry_t *pte; 2253 2254 PMAP_LOCK(pmap); 2255 pte = pmap_pte(pmap, va); 2256 2257 if (wired && !pmap_pte_w(pte)) 2258 pmap->pm_stats.wired_count++; 2259 else if (!wired && pmap_pte_w(pte)) 2260 pmap->pm_stats.wired_count--; 2261 2262 /* 2263 * Wiring is not a hardware characteristic so there is no need to 2264 * invalidate TLB. 2265 */ 2266 pmap_pte_set_w(pte, wired); 2267 PMAP_UNLOCK(pmap); 2268} 2269 2270 2271 2272/* 2273 * Copy the range specified by src_addr/len 2274 * from the source map to the range dst_addr/len 2275 * in the destination map. 2276 * 2277 * This routine is only advisory and need not do anything. 2278 */ 2279 2280void 2281pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2282 vm_offset_t src_addr) 2283{ 2284 vm_offset_t addr; 2285 vm_offset_t end_addr = src_addr + len; 2286 vm_offset_t pdnxt; 2287 vm_page_t m; 2288 2289 if (dst_addr != src_addr) 2290 return; 2291 2292 if (!pmap_is_current(src_pmap)) 2293 return; 2294 2295 vm_page_lock_queues(); 2296 PMAP_LOCK(dst_pmap); 2297 sched_pin(); 2298 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2299 pt_entry_t *src_pte, *dst_pte; 2300 vm_page_t dstmpte, srcmpte; 2301 pd_entry_t srcptepaddr; 2302 unsigned ptepindex; 2303 2304 if (addr >= UPT_MIN_ADDRESS) 2305 panic("pmap_copy: invalid to pmap_copy page tables"); 2306 2307 /* 2308 * Don't let optional prefaulting of pages make us go 2309 * way below the low water mark of free pages or way 2310 * above high water mark of used pv entries. 2311 */ 2312 if (cnt.v_free_count < cnt.v_free_reserved || 2313 pv_entry_count > pv_entry_high_water) 2314 break; 2315 2316 pdnxt = (addr + NBPDR) & ~PDRMASK; 2317 ptepindex = addr >> PDRSHIFT; 2318 2319 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2320 if (srcptepaddr == 0) 2321 continue; 2322 2323 if (srcptepaddr & PG_PS) { 2324 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2325 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2326 dst_pmap->pm_stats.resident_count += 2327 NBPDR / PAGE_SIZE; 2328 } 2329 continue; 2330 } 2331 2332 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2333 if (srcmpte->hold_count == 0) 2334 panic("pmap_copy: source page table page is unused"); 2335 2336 if (pdnxt > end_addr) 2337 pdnxt = end_addr; 2338 2339 src_pte = vtopte(addr); 2340 while (addr < pdnxt) { 2341 pt_entry_t ptetemp; 2342 ptetemp = *src_pte; 2343 /* 2344 * we only virtual copy managed pages 2345 */ 2346 if ((ptetemp & PG_MANAGED) != 0) { 2347 /* 2348 * We have to check after allocpte for the 2349 * pte still being around... allocpte can 2350 * block. 2351 */ 2352 dstmpte = pmap_allocpte(dst_pmap, addr); 2353 dst_pte = pmap_pte_quick(dst_pmap, addr); 2354 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2355 /* 2356 * Clear the modified and 2357 * accessed (referenced) bits 2358 * during the copy. 2359 */ 2360 m = PHYS_TO_VM_PAGE(ptetemp); 2361 *dst_pte = ptetemp & ~(PG_M | PG_A); 2362 dst_pmap->pm_stats.resident_count++; 2363 pmap_insert_entry(dst_pmap, addr, m); 2364 } else 2365 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2366 if (dstmpte->hold_count >= srcmpte->hold_count) 2367 break; 2368 } 2369 addr += PAGE_SIZE; 2370 src_pte++; 2371 } 2372 } 2373 sched_unpin(); 2374 vm_page_unlock_queues(); 2375 PMAP_UNLOCK(dst_pmap); 2376} 2377 2378static __inline void 2379pagezero(void *page) 2380{ 2381#if defined(I686_CPU) 2382 if (cpu_class == CPUCLASS_686) { 2383#if defined(CPU_ENABLE_SSE) 2384 if (cpu_feature & CPUID_SSE2) 2385 sse2_pagezero(page); 2386 else 2387#endif 2388 i686_pagezero(page); 2389 } else 2390#endif 2391 bzero(page, PAGE_SIZE); 2392} 2393 2394/* 2395 * pmap_zero_page zeros the specified hardware page by mapping 2396 * the page into KVM and using bzero to clear its contents. 2397 */ 2398void 2399pmap_zero_page(vm_page_t m) 2400{ 2401 2402 mtx_lock(&CMAPCADDR12_lock); 2403 if (*CMAP2) 2404 panic("pmap_zero_page: CMAP2 busy"); 2405 sched_pin(); 2406 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2407 invlcaddr(CADDR2); 2408 pagezero(CADDR2); 2409 *CMAP2 = 0; 2410 sched_unpin(); 2411 mtx_unlock(&CMAPCADDR12_lock); 2412} 2413 2414/* 2415 * pmap_zero_page_area zeros the specified hardware page by mapping 2416 * the page into KVM and using bzero to clear its contents. 2417 * 2418 * off and size may not cover an area beyond a single hardware page. 2419 */ 2420void 2421pmap_zero_page_area(vm_page_t m, int off, int size) 2422{ 2423 2424 mtx_lock(&CMAPCADDR12_lock); 2425 if (*CMAP2) 2426 panic("pmap_zero_page: CMAP2 busy"); 2427 sched_pin(); 2428 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2429 invlcaddr(CADDR2); 2430 if (off == 0 && size == PAGE_SIZE) 2431 pagezero(CADDR2); 2432 else 2433 bzero((char *)CADDR2 + off, size); 2434 *CMAP2 = 0; 2435 sched_unpin(); 2436 mtx_unlock(&CMAPCADDR12_lock); 2437} 2438 2439/* 2440 * pmap_zero_page_idle zeros the specified hardware page by mapping 2441 * the page into KVM and using bzero to clear its contents. This 2442 * is intended to be called from the vm_pagezero process only and 2443 * outside of Giant. 2444 */ 2445void 2446pmap_zero_page_idle(vm_page_t m) 2447{ 2448 2449 if (*CMAP3) 2450 panic("pmap_zero_page: CMAP3 busy"); 2451 sched_pin(); 2452 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2453 invlcaddr(CADDR3); 2454 pagezero(CADDR3); 2455 *CMAP3 = 0; 2456 sched_unpin(); 2457} 2458 2459/* 2460 * pmap_copy_page copies the specified (machine independent) 2461 * page by mapping the page into virtual memory and using 2462 * bcopy to copy the page, one machine dependent page at a 2463 * time. 2464 */ 2465void 2466pmap_copy_page(vm_page_t src, vm_page_t dst) 2467{ 2468 2469 mtx_lock(&CMAPCADDR12_lock); 2470 if (*CMAP1) 2471 panic("pmap_copy_page: CMAP1 busy"); 2472 if (*CMAP2) 2473 panic("pmap_copy_page: CMAP2 busy"); 2474 sched_pin(); 2475#ifdef I386_CPU 2476 invltlb(); 2477#else 2478 invlpg((u_int)CADDR1); 2479 invlpg((u_int)CADDR2); 2480#endif 2481 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2482 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2483 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2484 *CMAP1 = 0; 2485 *CMAP2 = 0; 2486 sched_unpin(); 2487 mtx_unlock(&CMAPCADDR12_lock); 2488} 2489 2490/* 2491 * Returns true if the pmap's pv is one of the first 2492 * 16 pvs linked to from this page. This count may 2493 * be changed upwards or downwards in the future; it 2494 * is only necessary that true be returned for a small 2495 * subset of pmaps for proper page aging. 2496 */ 2497boolean_t 2498pmap_page_exists_quick(pmap, m) 2499 pmap_t pmap; 2500 vm_page_t m; 2501{ 2502 pv_entry_t pv; 2503 int loops = 0; 2504 2505 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2506 return FALSE; 2507 2508 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2509 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2510 if (pv->pv_pmap == pmap) { 2511 return TRUE; 2512 } 2513 loops++; 2514 if (loops >= 16) 2515 break; 2516 } 2517 return (FALSE); 2518} 2519 2520#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2521/* 2522 * Remove all pages from specified address space 2523 * this aids process exit speeds. Also, this code 2524 * is special cased for current process only, but 2525 * can have the more generic (and slightly slower) 2526 * mode enabled. This is much faster than pmap_remove 2527 * in the case of running down an entire address space. 2528 */ 2529void 2530pmap_remove_pages(pmap, sva, eva) 2531 pmap_t pmap; 2532 vm_offset_t sva, eva; 2533{ 2534 pt_entry_t *pte, tpte; 2535 vm_page_t m; 2536 pv_entry_t pv, npv; 2537 2538#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2539 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2540 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2541 return; 2542 } 2543#endif 2544 vm_page_lock_queues(); 2545 PMAP_LOCK(pmap); 2546 sched_pin(); 2547 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2548 2549 if (pv->pv_va >= eva || pv->pv_va < sva) { 2550 npv = TAILQ_NEXT(pv, pv_plist); 2551 continue; 2552 } 2553 2554#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2555 pte = vtopte(pv->pv_va); 2556#else 2557 pte = pmap_pte_quick(pmap, pv->pv_va); 2558#endif 2559 tpte = *pte; 2560 2561 if (tpte == 0) { 2562 printf("TPTE at %p IS ZERO @ VA %08x\n", 2563 pte, pv->pv_va); 2564 panic("bad pte"); 2565 } 2566 2567/* 2568 * We cannot remove wired pages from a process' mapping at this time 2569 */ 2570 if (tpte & PG_W) { 2571 npv = TAILQ_NEXT(pv, pv_plist); 2572 continue; 2573 } 2574 2575 m = PHYS_TO_VM_PAGE(tpte); 2576 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2577 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2578 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2579 2580 KASSERT(m < &vm_page_array[vm_page_array_size], 2581 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2582 2583 pmap->pm_stats.resident_count--; 2584 2585 pte_clear(pte); 2586 2587 /* 2588 * Update the vm_page_t clean and reference bits. 2589 */ 2590 if (tpte & PG_M) { 2591 vm_page_dirty(m); 2592 } 2593 2594 npv = TAILQ_NEXT(pv, pv_plist); 2595 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 2596 2597 m->md.pv_list_count--; 2598 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2599 if (TAILQ_EMPTY(&m->md.pv_list)) 2600 vm_page_flag_clear(m, PG_WRITEABLE); 2601 2602 pmap_unuse_pt(pmap, pv->pv_va); 2603 free_pv_entry(pv); 2604 } 2605 sched_unpin(); 2606 pmap_invalidate_all(pmap); 2607 PMAP_UNLOCK(pmap); 2608 vm_page_unlock_queues(); 2609} 2610 2611/* 2612 * pmap_is_modified: 2613 * 2614 * Return whether or not the specified physical page was modified 2615 * in any physical maps. 2616 */ 2617boolean_t 2618pmap_is_modified(vm_page_t m) 2619{ 2620 pv_entry_t pv; 2621 pt_entry_t *pte; 2622 boolean_t rv; 2623 2624 rv = FALSE; 2625 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2626 return (rv); 2627 2628 sched_pin(); 2629 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2630 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2631 /* 2632 * if the bit being tested is the modified bit, then 2633 * mark clean_map and ptes as never 2634 * modified. 2635 */ 2636 if (!pmap_track_modified(pv->pv_va)) 2637 continue; 2638#if defined(PMAP_DIAGNOSTIC) 2639 if (!pv->pv_pmap) { 2640 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 2641 continue; 2642 } 2643#endif 2644 PMAP_LOCK(pv->pv_pmap); 2645 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2646 rv = (*pte & PG_M) != 0; 2647 PMAP_UNLOCK(pv->pv_pmap); 2648 if (rv) 2649 break; 2650 } 2651 sched_unpin(); 2652 return (rv); 2653} 2654 2655/* 2656 * pmap_is_prefaultable: 2657 * 2658 * Return whether or not the specified virtual address is elgible 2659 * for prefault. 2660 */ 2661boolean_t 2662pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2663{ 2664 pt_entry_t *pte; 2665 boolean_t rv; 2666 2667 rv = FALSE; 2668 PMAP_LOCK(pmap); 2669 if (*pmap_pde(pmap, addr)) { 2670 pte = vtopte(addr); 2671 rv = *pte == 0; 2672 } 2673 PMAP_UNLOCK(pmap); 2674 return (rv); 2675} 2676 2677/* 2678 * Clear the given bit in each of the given page's ptes. 2679 */ 2680static __inline void 2681pmap_clear_ptes(vm_page_t m, int bit) 2682{ 2683 register pv_entry_t pv; 2684 pt_entry_t pbits, *pte; 2685 2686 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) || 2687 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2688 return; 2689 2690 sched_pin(); 2691 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2692 /* 2693 * Loop over all current mappings setting/clearing as appropos If 2694 * setting RO do we need to clear the VAC? 2695 */ 2696 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2697 /* 2698 * don't write protect pager mappings 2699 */ 2700 if (bit == PG_RW) { 2701 if (!pmap_track_modified(pv->pv_va)) 2702 continue; 2703 } 2704 2705#if defined(PMAP_DIAGNOSTIC) 2706 if (!pv->pv_pmap) { 2707 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 2708 continue; 2709 } 2710#endif 2711 2712 PMAP_LOCK(pv->pv_pmap); 2713 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2714 pbits = *pte; 2715 if (pbits & bit) { 2716 if (bit == PG_RW) { 2717 if (pbits & PG_M) { 2718 vm_page_dirty(m); 2719 } 2720 pte_store(pte, pbits & ~(PG_M|PG_RW)); 2721 } else { 2722 pte_store(pte, pbits & ~bit); 2723 } 2724 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2725 } 2726 PMAP_UNLOCK(pv->pv_pmap); 2727 } 2728 if (bit == PG_RW) 2729 vm_page_flag_clear(m, PG_WRITEABLE); 2730 sched_unpin(); 2731} 2732 2733/* 2734 * pmap_page_protect: 2735 * 2736 * Lower the permission for all mappings to a given page. 2737 */ 2738void 2739pmap_page_protect(vm_page_t m, vm_prot_t prot) 2740{ 2741 if ((prot & VM_PROT_WRITE) == 0) { 2742 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2743 pmap_clear_ptes(m, PG_RW); 2744 } else { 2745 pmap_remove_all(m); 2746 } 2747 } 2748} 2749 2750/* 2751 * pmap_ts_referenced: 2752 * 2753 * Return a count of reference bits for a page, clearing those bits. 2754 * It is not necessary for every reference bit to be cleared, but it 2755 * is necessary that 0 only be returned when there are truly no 2756 * reference bits set. 2757 * 2758 * XXX: The exact number of bits to check and clear is a matter that 2759 * should be tested and standardized at some point in the future for 2760 * optimal aging of shared pages. 2761 */ 2762int 2763pmap_ts_referenced(vm_page_t m) 2764{ 2765 register pv_entry_t pv, pvf, pvn; 2766 pt_entry_t *pte; 2767 pt_entry_t v; 2768 int rtval = 0; 2769 2770 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2771 return (rtval); 2772 2773 sched_pin(); 2774 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2775 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2776 2777 pvf = pv; 2778 2779 do { 2780 pvn = TAILQ_NEXT(pv, pv_list); 2781 2782 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2783 2784 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2785 2786 if (!pmap_track_modified(pv->pv_va)) 2787 continue; 2788 2789 PMAP_LOCK(pv->pv_pmap); 2790 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2791 2792 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2793 atomic_clear_int((u_int *)pte, PG_A); 2794 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2795 2796 rtval++; 2797 if (rtval > 4) { 2798 PMAP_UNLOCK(pv->pv_pmap); 2799 break; 2800 } 2801 } 2802 PMAP_UNLOCK(pv->pv_pmap); 2803 } while ((pv = pvn) != NULL && pv != pvf); 2804 } 2805 sched_unpin(); 2806 2807 return (rtval); 2808} 2809 2810/* 2811 * Clear the modify bits on the specified physical page. 2812 */ 2813void 2814pmap_clear_modify(vm_page_t m) 2815{ 2816 pmap_clear_ptes(m, PG_M); 2817} 2818 2819/* 2820 * pmap_clear_reference: 2821 * 2822 * Clear the reference bit on the specified physical page. 2823 */ 2824void 2825pmap_clear_reference(vm_page_t m) 2826{ 2827 pmap_clear_ptes(m, PG_A); 2828} 2829 2830/* 2831 * Miscellaneous support routines follow 2832 */ 2833 2834/* 2835 * Map a set of physical memory pages into the kernel virtual 2836 * address space. Return a pointer to where it is mapped. This 2837 * routine is intended to be used for mapping device memory, 2838 * NOT real memory. 2839 */ 2840void * 2841pmap_mapdev(pa, size) 2842 vm_paddr_t pa; 2843 vm_size_t size; 2844{ 2845 vm_offset_t va, tmpva, offset; 2846 2847 offset = pa & PAGE_MASK; 2848 size = roundup(offset + size, PAGE_SIZE); 2849 pa = pa & PG_FRAME; 2850 2851 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2852 va = KERNBASE + pa; 2853 else 2854 va = kmem_alloc_nofault(kernel_map, size); 2855 if (!va) 2856 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2857 2858 for (tmpva = va; size > 0; ) { 2859 pmap_kenter(tmpva, pa); 2860 size -= PAGE_SIZE; 2861 tmpva += PAGE_SIZE; 2862 pa += PAGE_SIZE; 2863 } 2864 pmap_invalidate_range(kernel_pmap, va, tmpva); 2865 return ((void *)(va + offset)); 2866} 2867 2868void 2869pmap_unmapdev(va, size) 2870 vm_offset_t va; 2871 vm_size_t size; 2872{ 2873 vm_offset_t base, offset, tmpva; 2874 2875 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2876 return; 2877 base = va & PG_FRAME; 2878 offset = va & PAGE_MASK; 2879 size = roundup(offset + size, PAGE_SIZE); 2880 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2881 pmap_kremove(tmpva); 2882 pmap_invalidate_range(kernel_pmap, va, tmpva); 2883 kmem_free(kernel_map, base, size); 2884} 2885 2886/* 2887 * perform the pmap work for mincore 2888 */ 2889int 2890pmap_mincore(pmap, addr) 2891 pmap_t pmap; 2892 vm_offset_t addr; 2893{ 2894 pt_entry_t *ptep, pte; 2895 vm_page_t m; 2896 int val = 0; 2897 2898 PMAP_LOCK(pmap); 2899 ptep = pmap_pte(pmap, addr); 2900 pte = (ptep != NULL) ? *ptep : 0; 2901 PMAP_UNLOCK(pmap); 2902 2903 if (pte != 0) { 2904 vm_paddr_t pa; 2905 2906 val = MINCORE_INCORE; 2907 if ((pte & PG_MANAGED) == 0) 2908 return val; 2909 2910 pa = pte & PG_FRAME; 2911 2912 m = PHYS_TO_VM_PAGE(pa); 2913 2914 /* 2915 * Modified by us 2916 */ 2917 if (pte & PG_M) 2918 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2919 else { 2920 /* 2921 * Modified by someone else 2922 */ 2923 vm_page_lock_queues(); 2924 if (m->dirty || pmap_is_modified(m)) 2925 val |= MINCORE_MODIFIED_OTHER; 2926 vm_page_unlock_queues(); 2927 } 2928 /* 2929 * Referenced by us 2930 */ 2931 if (pte & PG_A) 2932 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2933 else { 2934 /* 2935 * Referenced by someone else 2936 */ 2937 vm_page_lock_queues(); 2938 if ((m->flags & PG_REFERENCED) || 2939 pmap_ts_referenced(m)) { 2940 val |= MINCORE_REFERENCED_OTHER; 2941 vm_page_flag_set(m, PG_REFERENCED); 2942 } 2943 vm_page_unlock_queues(); 2944 } 2945 } 2946 return val; 2947} 2948 2949void 2950pmap_activate(struct thread *td) 2951{ 2952 struct proc *p = td->td_proc; 2953 pmap_t pmap, oldpmap; 2954 u_int32_t cr3; 2955 2956 critical_enter(); 2957 pmap = vmspace_pmap(td->td_proc->p_vmspace); 2958 oldpmap = PCPU_GET(curpmap); 2959#if defined(SMP) 2960 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 2961 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 2962#else 2963 oldpmap->pm_active &= ~1; 2964 pmap->pm_active |= 1; 2965#endif 2966#ifdef PAE 2967 cr3 = vtophys(pmap->pm_pdpt); 2968#else 2969 cr3 = vtophys(pmap->pm_pdir); 2970#endif 2971 /* XXXKSE this is wrong. 2972 * pmap_activate is for the current thread on the current cpu 2973 */ 2974 if (p->p_flag & P_SA) { 2975 /* Make sure all other cr3 entries are updated. */ 2976 /* what if they are running? XXXKSE (maybe abort them) */ 2977 FOREACH_THREAD_IN_PROC(p, td) { 2978 td->td_pcb->pcb_cr3 = cr3; 2979 } 2980 } else { 2981 td->td_pcb->pcb_cr3 = cr3; 2982 } 2983 load_cr3(cr3); 2984 PCPU_SET(curpmap, pmap); 2985 critical_exit(); 2986} 2987 2988vm_offset_t 2989pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 2990{ 2991 2992 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 2993 return addr; 2994 } 2995 2996 addr = (addr + PDRMASK) & ~PDRMASK; 2997 return addr; 2998} 2999 3000 3001#if defined(PMAP_DEBUG) 3002pmap_pid_dump(int pid) 3003{ 3004 pmap_t pmap; 3005 struct proc *p; 3006 int npte = 0; 3007 int index; 3008 3009 sx_slock(&allproc_lock); 3010 LIST_FOREACH(p, &allproc, p_list) { 3011 if (p->p_pid != pid) 3012 continue; 3013 3014 if (p->p_vmspace) { 3015 int i,j; 3016 index = 0; 3017 pmap = vmspace_pmap(p->p_vmspace); 3018 for (i = 0; i < NPDEPTD; i++) { 3019 pd_entry_t *pde; 3020 pt_entry_t *pte; 3021 vm_offset_t base = i << PDRSHIFT; 3022 3023 pde = &pmap->pm_pdir[i]; 3024 if (pde && pmap_pde_v(pde)) { 3025 for (j = 0; j < NPTEPG; j++) { 3026 vm_offset_t va = base + (j << PAGE_SHIFT); 3027 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3028 if (index) { 3029 index = 0; 3030 printf("\n"); 3031 } 3032 sx_sunlock(&allproc_lock); 3033 return npte; 3034 } 3035 pte = pmap_pte(pmap, va); 3036 if (pte && pmap_pte_v(pte)) { 3037 pt_entry_t pa; 3038 vm_page_t m; 3039 pa = *pte; 3040 m = PHYS_TO_VM_PAGE(pa); 3041 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3042 va, pa, m->hold_count, m->wire_count, m->flags); 3043 npte++; 3044 index++; 3045 if (index >= 2) { 3046 index = 0; 3047 printf("\n"); 3048 } else { 3049 printf(" "); 3050 } 3051 } 3052 } 3053 } 3054 } 3055 } 3056 } 3057 sx_sunlock(&allproc_lock); 3058 return npte; 3059} 3060#endif 3061 3062#if defined(DEBUG) 3063 3064static void pads(pmap_t pm); 3065void pmap_pvdump(vm_offset_t pa); 3066 3067/* print address space of pmap*/ 3068static void 3069pads(pm) 3070 pmap_t pm; 3071{ 3072 int i, j; 3073 vm_paddr_t va; 3074 pt_entry_t *ptep; 3075 3076 if (pm == kernel_pmap) 3077 return; 3078 for (i = 0; i < NPDEPTD; i++) 3079 if (pm->pm_pdir[i]) 3080 for (j = 0; j < NPTEPG; j++) { 3081 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3082 if (pm == kernel_pmap && va < KERNBASE) 3083 continue; 3084 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3085 continue; 3086 ptep = pmap_pte(pm, va); 3087 if (pmap_pte_v(ptep)) 3088 printf("%x:%x ", va, *ptep); 3089 }; 3090 3091} 3092 3093void 3094pmap_pvdump(pa) 3095 vm_paddr_t pa; 3096{ 3097 pv_entry_t pv; 3098 vm_page_t m; 3099 3100 printf("pa %x", pa); 3101 m = PHYS_TO_VM_PAGE(pa); 3102 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3103 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3104 pads(pv->pv_pmap); 3105 } 3106 printf(" "); 3107} 3108#endif 3109