pmap.c revision 149532
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 149532 2005-08-27 19:51:14Z 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" 107 108#include <sys/param.h> 109#include <sys/systm.h> 110#include <sys/kernel.h> 111#include <sys/lock.h> 112#include <sys/malloc.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/vmmeter.h> 119#include <sys/sched.h> 120#include <sys/sysctl.h> 121#ifdef SMP 122#include <sys/smp.h> 123#endif 124 125#include <vm/vm.h> 126#include <vm/vm_param.h> 127#include <vm/vm_kern.h> 128#include <vm/vm_page.h> 129#include <vm/vm_map.h> 130#include <vm/vm_object.h> 131#include <vm/vm_extern.h> 132#include <vm/vm_pageout.h> 133#include <vm/vm_pager.h> 134#include <vm/uma.h> 135 136#include <machine/cpu.h> 137#include <machine/cputypes.h> 138#include <machine/md_var.h> 139#include <machine/pcb.h> 140#include <machine/specialreg.h> 141#ifdef SMP 142#include <machine/smp.h> 143#endif 144 145#if !defined(CPU_DISABLE_SSE) && defined(I686_CPU) 146#define CPU_ENABLE_SSE 147#endif 148 149#ifndef PMAP_SHPGPERPROC 150#define PMAP_SHPGPERPROC 200 151#endif 152 153#if defined(DIAGNOSTIC) 154#define PMAP_DIAGNOSTIC 155#endif 156 157#define MINPV 2048 158 159#if !defined(PMAP_DIAGNOSTIC) 160#define PMAP_INLINE __inline 161#else 162#define PMAP_INLINE 163#endif 164 165/* 166 * Get PDEs and PTEs for user/kernel address space 167 */ 168#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 169#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 170 171#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 172#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 173#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 174#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 175#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 176 177#define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \ 178 atomic_clear_int((u_int *)(pte), PG_W)) 179#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 180 181struct pmap kernel_pmap_store; 182LIST_HEAD(pmaplist, pmap); 183static struct pmaplist allpmaps; 184static struct mtx allpmaps_lock; 185 186vm_paddr_t avail_end; /* PA of last available physical page */ 187vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 188vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 189int pgeflag = 0; /* PG_G or-in */ 190int pseflag = 0; /* PG_PS or-in */ 191 192static int nkpt; 193vm_offset_t kernel_vm_end; 194extern u_int32_t KERNend; 195 196#ifdef PAE 197static uma_zone_t pdptzone; 198#endif 199 200/* 201 * Data for the pv entry allocation mechanism 202 */ 203static uma_zone_t pvzone; 204static struct vm_object pvzone_obj; 205static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 206int pmap_pagedaemon_waken; 207 208/* 209 * All those kernel PT submaps that BSD is so fond of 210 */ 211struct sysmaps { 212 struct mtx lock; 213 pt_entry_t *CMAP1; 214 pt_entry_t *CMAP2; 215 caddr_t CADDR1; 216 caddr_t CADDR2; 217}; 218static struct sysmaps sysmaps_pcpu[MAXCPU]; 219pt_entry_t *CMAP1 = 0; 220static pt_entry_t *CMAP3; 221caddr_t CADDR1 = 0, ptvmmap = 0; 222static caddr_t CADDR3; 223struct msgbuf *msgbufp = 0; 224 225/* 226 * Crashdump maps. 227 */ 228static caddr_t crashdumpmap; 229 230#ifdef SMP 231extern pt_entry_t *SMPpt; 232#endif 233static pt_entry_t *PMAP1 = 0, *PMAP2; 234static pt_entry_t *PADDR1 = 0, *PADDR2; 235#ifdef SMP 236static int PMAP1cpu; 237static int PMAP1changedcpu; 238SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD, 239 &PMAP1changedcpu, 0, 240 "Number of times pmap_pte_quick changed CPU with same PMAP1"); 241#endif 242static int PMAP1changed; 243SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD, 244 &PMAP1changed, 0, 245 "Number of times pmap_pte_quick changed PMAP1"); 246static int PMAP1unchanged; 247SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD, 248 &PMAP1unchanged, 0, 249 "Number of times pmap_pte_quick didn't change PMAP1"); 250static struct mtx PMAP2mutex; 251 252static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 253static pv_entry_t get_pv_entry(void); 254static void pmap_clear_ptes(vm_page_t m, int bit); 255 256static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 257static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 258static void pmap_remove_entry(struct pmap *pmap, vm_page_t m, 259 vm_offset_t va); 260static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); 261 262static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags); 263 264static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags); 265static int _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m); 266static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 267static void pmap_pte_release(pt_entry_t *pte); 268static int pmap_unuse_pt(pmap_t, vm_offset_t); 269static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 270#ifdef PAE 271static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 272#endif 273 274CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 275CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 276 277/* 278 * Move the kernel virtual free pointer to the next 279 * 4MB. This is used to help improve performance 280 * by using a large (4MB) page for much of the kernel 281 * (.text, .data, .bss) 282 */ 283static vm_offset_t 284pmap_kmem_choose(vm_offset_t addr) 285{ 286 vm_offset_t newaddr = addr; 287 288#ifndef DISABLE_PSE 289 if (cpu_feature & CPUID_PSE) 290 newaddr = (addr + PDRMASK) & ~PDRMASK; 291#endif 292 return newaddr; 293} 294 295/* 296 * Bootstrap the system enough to run with virtual memory. 297 * 298 * On the i386 this is called after mapping has already been enabled 299 * and just syncs the pmap module with what has already been done. 300 * [We can't call it easily with mapping off since the kernel is not 301 * mapped with PA == VA, hence we would have to relocate every address 302 * from the linked base (virtual) address "KERNBASE" to the actual 303 * (physical) address starting relative to 0] 304 */ 305void 306pmap_bootstrap(firstaddr, loadaddr) 307 vm_paddr_t firstaddr; 308 vm_paddr_t loadaddr; 309{ 310 vm_offset_t va; 311 pt_entry_t *pte, *unused; 312 struct sysmaps *sysmaps; 313 int i; 314 315 /* 316 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 317 * large. It should instead be correctly calculated in locore.s and 318 * not based on 'first' (which is a physical address, not a virtual 319 * address, for the start of unused physical memory). The kernel 320 * page tables are NOT double mapped and thus should not be included 321 * in this calculation. 322 */ 323 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 324 virtual_avail = pmap_kmem_choose(virtual_avail); 325 326 virtual_end = VM_MAX_KERNEL_ADDRESS; 327 328 /* 329 * Initialize the kernel pmap (which is statically allocated). 330 */ 331 PMAP_LOCK_INIT(kernel_pmap); 332 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 333#ifdef PAE 334 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 335#endif 336 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 337 TAILQ_INIT(&kernel_pmap->pm_pvlist); 338 LIST_INIT(&allpmaps); 339 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 340 mtx_lock_spin(&allpmaps_lock); 341 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 342 mtx_unlock_spin(&allpmaps_lock); 343 nkpt = NKPT; 344 345 /* 346 * Reserve some special page table entries/VA space for temporary 347 * mapping of pages. 348 */ 349#define SYSMAP(c, p, v, n) \ 350 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 351 352 va = virtual_avail; 353 pte = vtopte(va); 354 355 /* 356 * CMAP1/CMAP2 are used for zeroing and copying pages. 357 * CMAP3 is used for the idle process page zeroing. 358 */ 359 for (i = 0; i < MAXCPU; i++) { 360 sysmaps = &sysmaps_pcpu[i]; 361 mtx_init(&sysmaps->lock, "SYSMAPS", NULL, MTX_DEF); 362 SYSMAP(caddr_t, sysmaps->CMAP1, sysmaps->CADDR1, 1) 363 SYSMAP(caddr_t, sysmaps->CMAP2, sysmaps->CADDR2, 1) 364 } 365 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 366 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 367 *CMAP3 = 0; 368 369 /* 370 * Crashdump maps. 371 */ 372 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS) 373 374 /* 375 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 376 */ 377 SYSMAP(caddr_t, unused, ptvmmap, 1) 378 379 /* 380 * msgbufp is used to map the system message buffer. 381 */ 382 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE))) 383 384 /* 385 * ptemap is used for pmap_pte_quick 386 */ 387 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 388 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1); 389 390 mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF); 391 392 virtual_avail = va; 393 394 *CMAP1 = 0; 395 for (i = 0; i < NKPT; i++) 396 PTD[i] = 0; 397 398 /* Turn on PG_G on kernel page(s) */ 399 pmap_set_pg(); 400} 401 402/* 403 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 404 */ 405void 406pmap_set_pg(void) 407{ 408 pd_entry_t pdir; 409 pt_entry_t *pte; 410 vm_offset_t va, endva; 411 int i; 412 413 if (pgeflag == 0) 414 return; 415 416 i = KERNLOAD/NBPDR; 417 endva = KERNBASE + KERNend; 418 419 if (pseflag) { 420 va = KERNBASE + KERNLOAD; 421 while (va < endva) { 422 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 423 pdir |= pgeflag; 424 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 425 invltlb(); /* Play it safe, invltlb() every time */ 426 i++; 427 va += NBPDR; 428 } 429 } else { 430 va = (vm_offset_t)btext; 431 while (va < endva) { 432 pte = vtopte(va); 433 if (*pte) 434 *pte |= pgeflag; 435 invltlb(); /* Play it safe, invltlb() every time */ 436 va += PAGE_SIZE; 437 } 438 } 439} 440 441/* 442 * Initialize a vm_page's machine-dependent fields. 443 */ 444void 445pmap_page_init(vm_page_t m) 446{ 447 448 TAILQ_INIT(&m->md.pv_list); 449 m->md.pv_list_count = 0; 450} 451 452#ifdef PAE 453 454static MALLOC_DEFINE(M_PMAPPDPT, "pmap", "pmap pdpt"); 455 456static void * 457pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 458{ 459 *flags = UMA_SLAB_PRIV; 460 return (contigmalloc(PAGE_SIZE, M_PMAPPDPT, 0, 0x0ULL, 0xffffffffULL, 461 1, 0)); 462} 463#endif 464 465/* 466 * Initialize the pmap module. 467 * Called by vm_init, to initialize any structures that the pmap 468 * system needs to map virtual memory. 469 */ 470void 471pmap_init(void) 472{ 473 474 /* 475 * init the pv free list 476 */ 477 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL, 478 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 479 uma_prealloc(pvzone, MINPV); 480 481#ifdef PAE 482 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 483 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 484 UMA_ZONE_VM | UMA_ZONE_NOFREE); 485 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 486#endif 487} 488 489/* 490 * Initialize the address space (zone) for the pv_entries. Set a 491 * high water mark so that the system can recover from excessive 492 * numbers of pv entries. 493 */ 494void 495pmap_init2() 496{ 497 int shpgperproc = PMAP_SHPGPERPROC; 498 499 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 500 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 501 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 502 pv_entry_high_water = 9 * (pv_entry_max / 10); 503 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 504} 505 506 507/*************************************************** 508 * Low level helper routines..... 509 ***************************************************/ 510 511#if defined(PMAP_DIAGNOSTIC) 512 513/* 514 * This code checks for non-writeable/modified pages. 515 * This should be an invalid condition. 516 */ 517static int 518pmap_nw_modified(pt_entry_t ptea) 519{ 520 int pte; 521 522 pte = (int) ptea; 523 524 if ((pte & (PG_M|PG_RW)) == PG_M) 525 return 1; 526 else 527 return 0; 528} 529#endif 530 531 532/* 533 * this routine defines the region(s) of memory that should 534 * not be tested for the modified bit. 535 */ 536static PMAP_INLINE int 537pmap_track_modified(vm_offset_t va) 538{ 539 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 540 return 1; 541 else 542 return 0; 543} 544 545#ifdef SMP 546/* 547 * For SMP, these functions have to use the IPI mechanism for coherence. 548 */ 549void 550pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 551{ 552 u_int cpumask; 553 u_int other_cpus; 554 555 if (smp_started) { 556 if (!(read_eflags() & PSL_I)) 557 panic("%s: interrupts disabled", __func__); 558 mtx_lock_spin(&smp_ipi_mtx); 559 } else 560 critical_enter(); 561 /* 562 * We need to disable interrupt preemption but MUST NOT have 563 * interrupts disabled here. 564 * XXX we may need to hold schedlock to get a coherent pm_active 565 * XXX critical sections disable interrupts again 566 */ 567 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 568 invlpg(va); 569 smp_invlpg(va); 570 } else { 571 cpumask = PCPU_GET(cpumask); 572 other_cpus = PCPU_GET(other_cpus); 573 if (pmap->pm_active & cpumask) 574 invlpg(va); 575 if (pmap->pm_active & other_cpus) 576 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 577 } 578 if (smp_started) 579 mtx_unlock_spin(&smp_ipi_mtx); 580 else 581 critical_exit(); 582} 583 584void 585pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 586{ 587 u_int cpumask; 588 u_int other_cpus; 589 vm_offset_t addr; 590 591 if (smp_started) { 592 if (!(read_eflags() & PSL_I)) 593 panic("%s: interrupts disabled", __func__); 594 mtx_lock_spin(&smp_ipi_mtx); 595 } else 596 critical_enter(); 597 /* 598 * We need to disable interrupt preemption but MUST NOT have 599 * interrupts disabled here. 600 * XXX we may need to hold schedlock to get a coherent pm_active 601 * XXX critical sections disable interrupts again 602 */ 603 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 604 for (addr = sva; addr < eva; addr += PAGE_SIZE) 605 invlpg(addr); 606 smp_invlpg_range(sva, eva); 607 } else { 608 cpumask = PCPU_GET(cpumask); 609 other_cpus = PCPU_GET(other_cpus); 610 if (pmap->pm_active & cpumask) 611 for (addr = sva; addr < eva; addr += PAGE_SIZE) 612 invlpg(addr); 613 if (pmap->pm_active & other_cpus) 614 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 615 sva, eva); 616 } 617 if (smp_started) 618 mtx_unlock_spin(&smp_ipi_mtx); 619 else 620 critical_exit(); 621} 622 623void 624pmap_invalidate_all(pmap_t pmap) 625{ 626 u_int cpumask; 627 u_int other_cpus; 628 629 if (smp_started) { 630 if (!(read_eflags() & PSL_I)) 631 panic("%s: interrupts disabled", __func__); 632 mtx_lock_spin(&smp_ipi_mtx); 633 } else 634 critical_enter(); 635 /* 636 * We need to disable interrupt preemption but MUST NOT have 637 * interrupts disabled here. 638 * XXX we may need to hold schedlock to get a coherent pm_active 639 * XXX critical sections disable interrupts again 640 */ 641 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 642 invltlb(); 643 smp_invltlb(); 644 } else { 645 cpumask = PCPU_GET(cpumask); 646 other_cpus = PCPU_GET(other_cpus); 647 if (pmap->pm_active & cpumask) 648 invltlb(); 649 if (pmap->pm_active & other_cpus) 650 smp_masked_invltlb(pmap->pm_active & other_cpus); 651 } 652 if (smp_started) 653 mtx_unlock_spin(&smp_ipi_mtx); 654 else 655 critical_exit(); 656} 657#else /* !SMP */ 658/* 659 * Normal, non-SMP, 486+ invalidation functions. 660 * We inline these within pmap.c for speed. 661 */ 662PMAP_INLINE void 663pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 664{ 665 666 if (pmap == kernel_pmap || pmap->pm_active) 667 invlpg(va); 668} 669 670PMAP_INLINE void 671pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 672{ 673 vm_offset_t addr; 674 675 if (pmap == kernel_pmap || pmap->pm_active) 676 for (addr = sva; addr < eva; addr += PAGE_SIZE) 677 invlpg(addr); 678} 679 680PMAP_INLINE void 681pmap_invalidate_all(pmap_t pmap) 682{ 683 684 if (pmap == kernel_pmap || pmap->pm_active) 685 invltlb(); 686} 687#endif /* !SMP */ 688 689/* 690 * Are we current address space or kernel? N.B. We return FALSE when 691 * a pmap's page table is in use because a kernel thread is borrowing 692 * it. The borrowed page table can change spontaneously, making any 693 * dependence on its continued use subject to a race condition. 694 */ 695static __inline int 696pmap_is_current(pmap_t pmap) 697{ 698 699 return (pmap == kernel_pmap || 700 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 701 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 702} 703 704/* 705 * If the given pmap is not the current or kernel pmap, the returned pte must 706 * be released by passing it to pmap_pte_release(). 707 */ 708pt_entry_t * 709pmap_pte(pmap_t pmap, vm_offset_t va) 710{ 711 pd_entry_t newpf; 712 pd_entry_t *pde; 713 714 pde = pmap_pde(pmap, va); 715 if (*pde & PG_PS) 716 return (pde); 717 if (*pde != 0) { 718 /* are we current address space or kernel? */ 719 if (pmap_is_current(pmap)) 720 return (vtopte(va)); 721 mtx_lock(&PMAP2mutex); 722 newpf = *pde & PG_FRAME; 723 if ((*PMAP2 & PG_FRAME) != newpf) { 724 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; 725 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 726 } 727 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); 728 } 729 return (0); 730} 731 732/* 733 * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte 734 * being NULL. 735 */ 736static __inline void 737pmap_pte_release(pt_entry_t *pte) 738{ 739 740 if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2) 741 mtx_unlock(&PMAP2mutex); 742} 743 744static __inline void 745invlcaddr(void *caddr) 746{ 747 748 invlpg((u_int)caddr); 749} 750 751/* 752 * Super fast pmap_pte routine best used when scanning 753 * the pv lists. This eliminates many coarse-grained 754 * invltlb calls. Note that many of the pv list 755 * scans are across different pmaps. It is very wasteful 756 * to do an entire invltlb for checking a single mapping. 757 * 758 * If the given pmap is not the current pmap, vm_page_queue_mtx 759 * must be held and curthread pinned to a CPU. 760 */ 761static pt_entry_t * 762pmap_pte_quick(pmap_t pmap, vm_offset_t va) 763{ 764 pd_entry_t newpf; 765 pd_entry_t *pde; 766 767 pde = pmap_pde(pmap, va); 768 if (*pde & PG_PS) 769 return (pde); 770 if (*pde != 0) { 771 /* are we current address space or kernel? */ 772 if (pmap_is_current(pmap)) 773 return (vtopte(va)); 774 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 775 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 776 newpf = *pde & PG_FRAME; 777 if ((*PMAP1 & PG_FRAME) != newpf) { 778 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; 779#ifdef SMP 780 PMAP1cpu = PCPU_GET(cpuid); 781#endif 782 invlcaddr(PADDR1); 783 PMAP1changed++; 784 } else 785#ifdef SMP 786 if (PMAP1cpu != PCPU_GET(cpuid)) { 787 PMAP1cpu = PCPU_GET(cpuid); 788 invlcaddr(PADDR1); 789 PMAP1changedcpu++; 790 } else 791#endif 792 PMAP1unchanged++; 793 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); 794 } 795 return (0); 796} 797 798/* 799 * Routine: pmap_extract 800 * Function: 801 * Extract the physical page address associated 802 * with the given map/virtual_address pair. 803 */ 804vm_paddr_t 805pmap_extract(pmap_t pmap, vm_offset_t va) 806{ 807 vm_paddr_t rtval; 808 pt_entry_t *pte; 809 pd_entry_t pde; 810 811 rtval = 0; 812 PMAP_LOCK(pmap); 813 pde = pmap->pm_pdir[va >> PDRSHIFT]; 814 if (pde != 0) { 815 if ((pde & PG_PS) != 0) { 816 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 817 PMAP_UNLOCK(pmap); 818 return rtval; 819 } 820 pte = pmap_pte(pmap, va); 821 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK); 822 pmap_pte_release(pte); 823 } 824 PMAP_UNLOCK(pmap); 825 return (rtval); 826} 827 828/* 829 * Routine: pmap_extract_and_hold 830 * Function: 831 * Atomically extract and hold the physical page 832 * with the given pmap and virtual address pair 833 * if that mapping permits the given protection. 834 */ 835vm_page_t 836pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 837{ 838 pd_entry_t pde; 839 pt_entry_t pte; 840 vm_page_t m; 841 842 m = NULL; 843 vm_page_lock_queues(); 844 PMAP_LOCK(pmap); 845 pde = *pmap_pde(pmap, va); 846 if (pde != 0) { 847 if (pde & PG_PS) { 848 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) { 849 m = PHYS_TO_VM_PAGE((pde & ~PDRMASK) | 850 (va & PDRMASK)); 851 vm_page_hold(m); 852 } 853 } else { 854 sched_pin(); 855 pte = *pmap_pte_quick(pmap, va); 856 if (pte != 0 && 857 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) { 858 m = PHYS_TO_VM_PAGE(pte & PG_FRAME); 859 vm_page_hold(m); 860 } 861 sched_unpin(); 862 } 863 } 864 vm_page_unlock_queues(); 865 PMAP_UNLOCK(pmap); 866 return (m); 867} 868 869/*************************************************** 870 * Low level mapping routines..... 871 ***************************************************/ 872 873/* 874 * Add a wired page to the kva. 875 * Note: not SMP coherent. 876 */ 877PMAP_INLINE void 878pmap_kenter(vm_offset_t va, vm_paddr_t pa) 879{ 880 pt_entry_t *pte; 881 882 pte = vtopte(va); 883 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 884} 885 886/* 887 * Remove a page from the kernel pagetables. 888 * Note: not SMP coherent. 889 */ 890PMAP_INLINE void 891pmap_kremove(vm_offset_t va) 892{ 893 pt_entry_t *pte; 894 895 pte = vtopte(va); 896 pte_clear(pte); 897} 898 899/* 900 * Used to map a range of physical addresses into kernel 901 * virtual address space. 902 * 903 * The value passed in '*virt' is a suggested virtual address for 904 * the mapping. Architectures which can support a direct-mapped 905 * physical to virtual region can return the appropriate address 906 * within that region, leaving '*virt' unchanged. Other 907 * architectures should map the pages starting at '*virt' and 908 * update '*virt' with the first usable address after the mapped 909 * region. 910 */ 911vm_offset_t 912pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 913{ 914 vm_offset_t va, sva; 915 916 va = sva = *virt; 917 while (start < end) { 918 pmap_kenter(va, start); 919 va += PAGE_SIZE; 920 start += PAGE_SIZE; 921 } 922 pmap_invalidate_range(kernel_pmap, sva, va); 923 *virt = va; 924 return (sva); 925} 926 927 928/* 929 * Add a list of wired pages to the kva 930 * this routine is only used for temporary 931 * kernel mappings that do not need to have 932 * page modification or references recorded. 933 * Note that old mappings are simply written 934 * over. The page *must* be wired. 935 * Note: SMP coherent. Uses a ranged shootdown IPI. 936 */ 937void 938pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 939{ 940 vm_offset_t va; 941 942 va = sva; 943 while (count-- > 0) { 944 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 945 va += PAGE_SIZE; 946 m++; 947 } 948 pmap_invalidate_range(kernel_pmap, sva, va); 949} 950 951/* 952 * This routine tears out page mappings from the 953 * kernel -- it is meant only for temporary mappings. 954 * Note: SMP coherent. Uses a ranged shootdown IPI. 955 */ 956void 957pmap_qremove(vm_offset_t sva, int count) 958{ 959 vm_offset_t va; 960 961 va = sva; 962 while (count-- > 0) { 963 pmap_kremove(va); 964 va += PAGE_SIZE; 965 } 966 pmap_invalidate_range(kernel_pmap, sva, va); 967} 968 969/*************************************************** 970 * Page table page management routines..... 971 ***************************************************/ 972 973/* 974 * This routine unholds page table pages, and if the hold count 975 * drops to zero, then it decrements the wire count. 976 */ 977static PMAP_INLINE int 978pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 979{ 980 981 --m->wire_count; 982 if (m->wire_count == 0) 983 return _pmap_unwire_pte_hold(pmap, m); 984 else 985 return 0; 986} 987 988static int 989_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 990{ 991 vm_offset_t pteva; 992 993 /* 994 * unmap the page table page 995 */ 996 pmap->pm_pdir[m->pindex] = 0; 997 --pmap->pm_stats.resident_count; 998 999 /* 1000 * Do an invltlb to make the invalidated mapping 1001 * take effect immediately. 1002 */ 1003 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1004 pmap_invalidate_page(pmap, pteva); 1005 1006 vm_page_free_zero(m); 1007 atomic_subtract_int(&cnt.v_wire_count, 1); 1008 return 1; 1009} 1010 1011/* 1012 * After removing a page table entry, this routine is used to 1013 * conditionally free the page, and manage the hold/wire counts. 1014 */ 1015static int 1016pmap_unuse_pt(pmap_t pmap, vm_offset_t va) 1017{ 1018 pd_entry_t ptepde; 1019 vm_page_t mpte; 1020 1021 if (va >= VM_MAXUSER_ADDRESS) 1022 return 0; 1023 ptepde = *pmap_pde(pmap, va); 1024 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); 1025 return pmap_unwire_pte_hold(pmap, mpte); 1026} 1027 1028void 1029pmap_pinit0(pmap) 1030 struct pmap *pmap; 1031{ 1032 1033 PMAP_LOCK_INIT(pmap); 1034 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1035#ifdef PAE 1036 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1037#endif 1038 pmap->pm_active = 0; 1039 PCPU_SET(curpmap, pmap); 1040 TAILQ_INIT(&pmap->pm_pvlist); 1041 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1042 mtx_lock_spin(&allpmaps_lock); 1043 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1044 mtx_unlock_spin(&allpmaps_lock); 1045} 1046 1047/* 1048 * Initialize a preallocated and zeroed pmap structure, 1049 * such as one in a vmspace structure. 1050 */ 1051void 1052pmap_pinit(pmap) 1053 register struct pmap *pmap; 1054{ 1055 vm_page_t m, ptdpg[NPGPTD]; 1056 vm_paddr_t pa; 1057 static int color; 1058 int i; 1059 1060 PMAP_LOCK_INIT(pmap); 1061 1062 /* 1063 * No need to allocate page table space yet but we do need a valid 1064 * page directory table. 1065 */ 1066 if (pmap->pm_pdir == NULL) { 1067 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1068 NBPTD); 1069#ifdef PAE 1070 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1071 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1072 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1073 ("pmap_pinit: pdpt misaligned")); 1074 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1075 ("pmap_pinit: pdpt above 4g")); 1076#endif 1077 } 1078 1079 /* 1080 * allocate the page directory page(s) 1081 */ 1082 for (i = 0; i < NPGPTD;) { 1083 m = vm_page_alloc(NULL, color++, 1084 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1085 VM_ALLOC_ZERO); 1086 if (m == NULL) 1087 VM_WAIT; 1088 else { 1089 ptdpg[i++] = m; 1090 } 1091 } 1092 1093 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1094 1095 for (i = 0; i < NPGPTD; i++) { 1096 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1097 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1098 } 1099 1100 mtx_lock_spin(&allpmaps_lock); 1101 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1102 mtx_unlock_spin(&allpmaps_lock); 1103 /* Wire in kernel global address entries. */ 1104 /* XXX copies current process, does not fill in MPPTDI */ 1105 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1106#ifdef SMP 1107 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1108#endif 1109 1110 /* install self-referential address mapping entry(s) */ 1111 for (i = 0; i < NPGPTD; i++) { 1112 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1113 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1114#ifdef PAE 1115 pmap->pm_pdpt[i] = pa | PG_V; 1116#endif 1117 } 1118 1119 pmap->pm_active = 0; 1120 TAILQ_INIT(&pmap->pm_pvlist); 1121 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1122} 1123 1124/* 1125 * this routine is called if the page table page is not 1126 * mapped correctly. 1127 */ 1128static vm_page_t 1129_pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags) 1130{ 1131 vm_paddr_t ptepa; 1132 vm_page_t m; 1133 1134 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT || 1135 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK, 1136 ("_pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK")); 1137 1138 /* 1139 * Allocate a page table page. 1140 */ 1141 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1142 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1143 if (flags & M_WAITOK) { 1144 PMAP_UNLOCK(pmap); 1145 vm_page_unlock_queues(); 1146 VM_WAIT; 1147 vm_page_lock_queues(); 1148 PMAP_LOCK(pmap); 1149 } 1150 1151 /* 1152 * Indicate the need to retry. While waiting, the page table 1153 * page may have been allocated. 1154 */ 1155 return (NULL); 1156 } 1157 if ((m->flags & PG_ZERO) == 0) 1158 pmap_zero_page(m); 1159 1160 /* 1161 * Map the pagetable page into the process address space, if 1162 * it isn't already there. 1163 */ 1164 1165 pmap->pm_stats.resident_count++; 1166 1167 ptepa = VM_PAGE_TO_PHYS(m); 1168 pmap->pm_pdir[ptepindex] = 1169 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1170 1171 return m; 1172} 1173 1174static vm_page_t 1175pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags) 1176{ 1177 unsigned ptepindex; 1178 pd_entry_t ptepa; 1179 vm_page_t m; 1180 1181 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT || 1182 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK, 1183 ("pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK")); 1184 1185 /* 1186 * Calculate pagetable page index 1187 */ 1188 ptepindex = va >> PDRSHIFT; 1189retry: 1190 /* 1191 * Get the page directory entry 1192 */ 1193 ptepa = pmap->pm_pdir[ptepindex]; 1194 1195 /* 1196 * This supports switching from a 4MB page to a 1197 * normal 4K page. 1198 */ 1199 if (ptepa & PG_PS) { 1200 pmap->pm_pdir[ptepindex] = 0; 1201 ptepa = 0; 1202 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1203 pmap_invalidate_all(kernel_pmap); 1204 } 1205 1206 /* 1207 * If the page table page is mapped, we just increment the 1208 * hold count, and activate it. 1209 */ 1210 if (ptepa) { 1211 m = PHYS_TO_VM_PAGE(ptepa); 1212 m->wire_count++; 1213 } else { 1214 /* 1215 * Here if the pte page isn't mapped, or if it has 1216 * been deallocated. 1217 */ 1218 m = _pmap_allocpte(pmap, ptepindex, flags); 1219 if (m == NULL && (flags & M_WAITOK)) 1220 goto retry; 1221 } 1222 return (m); 1223} 1224 1225 1226/*************************************************** 1227* Pmap allocation/deallocation routines. 1228 ***************************************************/ 1229 1230#ifdef SMP 1231/* 1232 * Deal with a SMP shootdown of other users of the pmap that we are 1233 * trying to dispose of. This can be a bit hairy. 1234 */ 1235static u_int *lazymask; 1236static u_int lazyptd; 1237static volatile u_int lazywait; 1238 1239void pmap_lazyfix_action(void); 1240 1241void 1242pmap_lazyfix_action(void) 1243{ 1244 u_int mymask = PCPU_GET(cpumask); 1245 1246 if (rcr3() == lazyptd) 1247 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1248 atomic_clear_int(lazymask, mymask); 1249 atomic_store_rel_int(&lazywait, 1); 1250} 1251 1252static void 1253pmap_lazyfix_self(u_int mymask) 1254{ 1255 1256 if (rcr3() == lazyptd) 1257 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1258 atomic_clear_int(lazymask, mymask); 1259} 1260 1261 1262static void 1263pmap_lazyfix(pmap_t pmap) 1264{ 1265 u_int mymask; 1266 u_int mask; 1267 register u_int spins; 1268 1269 while ((mask = pmap->pm_active) != 0) { 1270 spins = 50000000; 1271 mask = mask & -mask; /* Find least significant set bit */ 1272 mtx_lock_spin(&smp_ipi_mtx); 1273#ifdef PAE 1274 lazyptd = vtophys(pmap->pm_pdpt); 1275#else 1276 lazyptd = vtophys(pmap->pm_pdir); 1277#endif 1278 mymask = PCPU_GET(cpumask); 1279 if (mask == mymask) { 1280 lazymask = &pmap->pm_active; 1281 pmap_lazyfix_self(mymask); 1282 } else { 1283 atomic_store_rel_int((u_int *)&lazymask, 1284 (u_int)&pmap->pm_active); 1285 atomic_store_rel_int(&lazywait, 0); 1286 ipi_selected(mask, IPI_LAZYPMAP); 1287 while (lazywait == 0) { 1288 ia32_pause(); 1289 if (--spins == 0) 1290 break; 1291 } 1292 } 1293 mtx_unlock_spin(&smp_ipi_mtx); 1294 if (spins == 0) 1295 printf("pmap_lazyfix: spun for 50000000\n"); 1296 } 1297} 1298 1299#else /* SMP */ 1300 1301/* 1302 * Cleaning up on uniprocessor is easy. For various reasons, we're 1303 * unlikely to have to even execute this code, including the fact 1304 * that the cleanup is deferred until the parent does a wait(2), which 1305 * means that another userland process has run. 1306 */ 1307static void 1308pmap_lazyfix(pmap_t pmap) 1309{ 1310 u_int cr3; 1311 1312 cr3 = vtophys(pmap->pm_pdir); 1313 if (cr3 == rcr3()) { 1314 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1315 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1316 } 1317} 1318#endif /* SMP */ 1319 1320/* 1321 * Release any resources held by the given physical map. 1322 * Called when a pmap initialized by pmap_pinit is being released. 1323 * Should only be called if the map contains no valid mappings. 1324 */ 1325void 1326pmap_release(pmap_t pmap) 1327{ 1328 vm_page_t m, ptdpg[NPGPTD]; 1329 int i; 1330 1331 KASSERT(pmap->pm_stats.resident_count == 0, 1332 ("pmap_release: pmap resident count %ld != 0", 1333 pmap->pm_stats.resident_count)); 1334 1335 pmap_lazyfix(pmap); 1336 mtx_lock_spin(&allpmaps_lock); 1337 LIST_REMOVE(pmap, pm_list); 1338 mtx_unlock_spin(&allpmaps_lock); 1339 1340 for (i = 0; i < NPGPTD; i++) 1341 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1342 1343 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1344 sizeof(*pmap->pm_pdir)); 1345#ifdef SMP 1346 pmap->pm_pdir[MPPTDI] = 0; 1347#endif 1348 1349 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1350 1351 vm_page_lock_queues(); 1352 for (i = 0; i < NPGPTD; i++) { 1353 m = ptdpg[i]; 1354#ifdef PAE 1355 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1356 ("pmap_release: got wrong ptd page")); 1357#endif 1358 m->wire_count--; 1359 atomic_subtract_int(&cnt.v_wire_count, 1); 1360 vm_page_free_zero(m); 1361 } 1362 vm_page_unlock_queues(); 1363 PMAP_LOCK_DESTROY(pmap); 1364} 1365 1366static int 1367kvm_size(SYSCTL_HANDLER_ARGS) 1368{ 1369 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1370 1371 return sysctl_handle_long(oidp, &ksize, 0, req); 1372} 1373SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1374 0, 0, kvm_size, "IU", "Size of KVM"); 1375 1376static int 1377kvm_free(SYSCTL_HANDLER_ARGS) 1378{ 1379 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1380 1381 return sysctl_handle_long(oidp, &kfree, 0, req); 1382} 1383SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1384 0, 0, kvm_free, "IU", "Amount of KVM free"); 1385 1386/* 1387 * grow the number of kernel page table entries, if needed 1388 */ 1389void 1390pmap_growkernel(vm_offset_t addr) 1391{ 1392 struct pmap *pmap; 1393 vm_paddr_t ptppaddr; 1394 vm_page_t nkpg; 1395 pd_entry_t newpdir; 1396 pt_entry_t *pde; 1397 1398 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1399 if (kernel_vm_end == 0) { 1400 kernel_vm_end = KERNBASE; 1401 nkpt = 0; 1402 while (pdir_pde(PTD, kernel_vm_end)) { 1403 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1404 nkpt++; 1405 } 1406 } 1407 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1408 while (kernel_vm_end < addr) { 1409 if (pdir_pde(PTD, kernel_vm_end)) { 1410 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1411 continue; 1412 } 1413 1414 /* 1415 * This index is bogus, but out of the way 1416 */ 1417 nkpg = vm_page_alloc(NULL, nkpt, 1418 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1419 if (!nkpg) 1420 panic("pmap_growkernel: no memory to grow kernel"); 1421 1422 nkpt++; 1423 1424 pmap_zero_page(nkpg); 1425 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1426 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1427 pdir_pde(PTD, kernel_vm_end) = newpdir; 1428 1429 mtx_lock_spin(&allpmaps_lock); 1430 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1431 pde = pmap_pde(pmap, kernel_vm_end); 1432 pde_store(pde, newpdir); 1433 } 1434 mtx_unlock_spin(&allpmaps_lock); 1435 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1436 } 1437} 1438 1439 1440/*************************************************** 1441 * page management routines. 1442 ***************************************************/ 1443 1444/* 1445 * free the pv_entry back to the free list 1446 */ 1447static PMAP_INLINE void 1448free_pv_entry(pv_entry_t pv) 1449{ 1450 pv_entry_count--; 1451 uma_zfree(pvzone, pv); 1452} 1453 1454/* 1455 * get a new pv_entry, allocating a block from the system 1456 * when needed. 1457 * the memory allocation is performed bypassing the malloc code 1458 * because of the possibility of allocations at interrupt time. 1459 */ 1460static pv_entry_t 1461get_pv_entry(void) 1462{ 1463 pv_entry_count++; 1464 if (pv_entry_high_water && 1465 (pv_entry_count > pv_entry_high_water) && 1466 (pmap_pagedaemon_waken == 0)) { 1467 pmap_pagedaemon_waken = 1; 1468 wakeup (&vm_pages_needed); 1469 } 1470 return uma_zalloc(pvzone, M_NOWAIT); 1471} 1472 1473 1474static void 1475pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1476{ 1477 pv_entry_t pv; 1478 1479 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1480 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1481 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1482 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1483 if (pmap == pv->pv_pmap && va == pv->pv_va) 1484 break; 1485 } 1486 } else { 1487 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1488 if (va == pv->pv_va) 1489 break; 1490 } 1491 } 1492 KASSERT(pv != NULL, ("pmap_remove_entry: pv not found")); 1493 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1494 m->md.pv_list_count--; 1495 if (TAILQ_EMPTY(&m->md.pv_list)) 1496 vm_page_flag_clear(m, PG_WRITEABLE); 1497 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1498 free_pv_entry(pv); 1499} 1500 1501/* 1502 * Create a pv entry for page at pa for 1503 * (pmap, va). 1504 */ 1505static void 1506pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) 1507{ 1508 pv_entry_t pv; 1509 1510 pv = get_pv_entry(); 1511 pv->pv_va = va; 1512 pv->pv_pmap = pmap; 1513 1514 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1515 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1516 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1517 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1518 m->md.pv_list_count++; 1519} 1520 1521/* 1522 * pmap_remove_pte: do the things to unmap a page in a process 1523 */ 1524static int 1525pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1526{ 1527 pt_entry_t oldpte; 1528 vm_page_t m; 1529 1530 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1531 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1532 oldpte = pte_load_clear(ptq); 1533 if (oldpte & PG_W) 1534 pmap->pm_stats.wired_count -= 1; 1535 /* 1536 * Machines that don't support invlpg, also don't support 1537 * PG_G. 1538 */ 1539 if (oldpte & PG_G) 1540 pmap_invalidate_page(kernel_pmap, va); 1541 pmap->pm_stats.resident_count -= 1; 1542 if (oldpte & PG_MANAGED) { 1543 m = PHYS_TO_VM_PAGE(oldpte); 1544 if (oldpte & PG_M) { 1545#if defined(PMAP_DIAGNOSTIC) 1546 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1547 printf( 1548 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1549 va, oldpte); 1550 } 1551#endif 1552 if (pmap_track_modified(va)) 1553 vm_page_dirty(m); 1554 } 1555 if (oldpte & PG_A) 1556 vm_page_flag_set(m, PG_REFERENCED); 1557 pmap_remove_entry(pmap, m, va); 1558 } 1559 return (pmap_unuse_pt(pmap, va)); 1560} 1561 1562/* 1563 * Remove a single page from a process address space 1564 */ 1565static void 1566pmap_remove_page(pmap_t pmap, vm_offset_t va) 1567{ 1568 pt_entry_t *pte; 1569 1570 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1571 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 1572 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1573 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) 1574 return; 1575 pmap_remove_pte(pmap, pte, va); 1576 pmap_invalidate_page(pmap, va); 1577} 1578 1579/* 1580 * Remove the given range of addresses from the specified map. 1581 * 1582 * It is assumed that the start and end are properly 1583 * rounded to the page size. 1584 */ 1585void 1586pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1587{ 1588 vm_offset_t pdnxt; 1589 pd_entry_t ptpaddr; 1590 pt_entry_t *pte; 1591 int anyvalid; 1592 1593 /* 1594 * Perform an unsynchronized read. This is, however, safe. 1595 */ 1596 if (pmap->pm_stats.resident_count == 0) 1597 return; 1598 1599 anyvalid = 0; 1600 1601 vm_page_lock_queues(); 1602 sched_pin(); 1603 PMAP_LOCK(pmap); 1604 1605 /* 1606 * special handling of removing one page. a very 1607 * common operation and easy to short circuit some 1608 * code. 1609 */ 1610 if ((sva + PAGE_SIZE == eva) && 1611 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1612 pmap_remove_page(pmap, sva); 1613 goto out; 1614 } 1615 1616 for (; sva < eva; sva = pdnxt) { 1617 unsigned pdirindex; 1618 1619 /* 1620 * Calculate index for next page table. 1621 */ 1622 pdnxt = (sva + NBPDR) & ~PDRMASK; 1623 if (pmap->pm_stats.resident_count == 0) 1624 break; 1625 1626 pdirindex = sva >> PDRSHIFT; 1627 ptpaddr = pmap->pm_pdir[pdirindex]; 1628 1629 /* 1630 * Weed out invalid mappings. Note: we assume that the page 1631 * directory table is always allocated, and in kernel virtual. 1632 */ 1633 if (ptpaddr == 0) 1634 continue; 1635 1636 /* 1637 * Check for large page. 1638 */ 1639 if ((ptpaddr & PG_PS) != 0) { 1640 pmap->pm_pdir[pdirindex] = 0; 1641 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1642 anyvalid = 1; 1643 continue; 1644 } 1645 1646 /* 1647 * Limit our scan to either the end of the va represented 1648 * by the current page table page, or to the end of the 1649 * range being removed. 1650 */ 1651 if (pdnxt > eva) 1652 pdnxt = eva; 1653 1654 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, 1655 sva += PAGE_SIZE) { 1656 if (*pte == 0) 1657 continue; 1658 anyvalid = 1; 1659 if (pmap_remove_pte(pmap, pte, sva)) 1660 break; 1661 } 1662 } 1663out: 1664 sched_unpin(); 1665 vm_page_unlock_queues(); 1666 if (anyvalid) 1667 pmap_invalidate_all(pmap); 1668 PMAP_UNLOCK(pmap); 1669} 1670 1671/* 1672 * Routine: pmap_remove_all 1673 * Function: 1674 * Removes this physical page from 1675 * all physical maps in which it resides. 1676 * Reflects back modify bits to the pager. 1677 * 1678 * Notes: 1679 * Original versions of this routine were very 1680 * inefficient because they iteratively called 1681 * pmap_remove (slow...) 1682 */ 1683 1684void 1685pmap_remove_all(vm_page_t m) 1686{ 1687 register pv_entry_t pv; 1688 pt_entry_t *pte, tpte; 1689 1690#if defined(PMAP_DIAGNOSTIC) 1691 /* 1692 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1693 */ 1694 if (m->flags & PG_FICTITIOUS) { 1695 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1696 VM_PAGE_TO_PHYS(m)); 1697 } 1698#endif 1699 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1700 sched_pin(); 1701 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1702 PMAP_LOCK(pv->pv_pmap); 1703 pv->pv_pmap->pm_stats.resident_count--; 1704 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1705 tpte = pte_load_clear(pte); 1706 if (tpte & PG_W) 1707 pv->pv_pmap->pm_stats.wired_count--; 1708 if (tpte & PG_A) 1709 vm_page_flag_set(m, PG_REFERENCED); 1710 1711 /* 1712 * Update the vm_page_t clean and reference bits. 1713 */ 1714 if (tpte & PG_M) { 1715#if defined(PMAP_DIAGNOSTIC) 1716 if (pmap_nw_modified((pt_entry_t) tpte)) { 1717 printf( 1718 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1719 pv->pv_va, tpte); 1720 } 1721#endif 1722 if (pmap_track_modified(pv->pv_va)) 1723 vm_page_dirty(m); 1724 } 1725 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1726 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1727 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1728 m->md.pv_list_count--; 1729 pmap_unuse_pt(pv->pv_pmap, pv->pv_va); 1730 PMAP_UNLOCK(pv->pv_pmap); 1731 free_pv_entry(pv); 1732 } 1733 vm_page_flag_clear(m, PG_WRITEABLE); 1734 sched_unpin(); 1735} 1736 1737/* 1738 * Set the physical protection on the 1739 * specified range of this map as requested. 1740 */ 1741void 1742pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1743{ 1744 vm_offset_t pdnxt; 1745 pd_entry_t ptpaddr; 1746 pt_entry_t *pte; 1747 int anychanged; 1748 1749 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1750 pmap_remove(pmap, sva, eva); 1751 return; 1752 } 1753 1754 if (prot & VM_PROT_WRITE) 1755 return; 1756 1757 anychanged = 0; 1758 1759 vm_page_lock_queues(); 1760 sched_pin(); 1761 PMAP_LOCK(pmap); 1762 for (; sva < eva; sva = pdnxt) { 1763 unsigned obits, pbits, pdirindex; 1764 1765 pdnxt = (sva + NBPDR) & ~PDRMASK; 1766 1767 pdirindex = sva >> PDRSHIFT; 1768 ptpaddr = pmap->pm_pdir[pdirindex]; 1769 1770 /* 1771 * Weed out invalid mappings. Note: we assume that the page 1772 * directory table is always allocated, and in kernel virtual. 1773 */ 1774 if (ptpaddr == 0) 1775 continue; 1776 1777 /* 1778 * Check for large page. 1779 */ 1780 if ((ptpaddr & PG_PS) != 0) { 1781 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1782 anychanged = 1; 1783 continue; 1784 } 1785 1786 if (pdnxt > eva) 1787 pdnxt = eva; 1788 1789 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, 1790 sva += PAGE_SIZE) { 1791 vm_page_t m; 1792 1793retry: 1794 /* 1795 * Regardless of whether a pte is 32 or 64 bits in 1796 * size, PG_RW, PG_A, and PG_M are among the least 1797 * significant 32 bits. 1798 */ 1799 obits = pbits = *(u_int *)pte; 1800 if (pbits & PG_MANAGED) { 1801 m = NULL; 1802 if (pbits & PG_A) { 1803 m = PHYS_TO_VM_PAGE(*pte); 1804 vm_page_flag_set(m, PG_REFERENCED); 1805 pbits &= ~PG_A; 1806 } 1807 if ((pbits & PG_M) != 0 && 1808 pmap_track_modified(sva)) { 1809 if (m == NULL) 1810 m = PHYS_TO_VM_PAGE(*pte); 1811 vm_page_dirty(m); 1812 } 1813 } 1814 1815 pbits &= ~(PG_RW | PG_M); 1816 1817 if (pbits != obits) { 1818 if (!atomic_cmpset_int((u_int *)pte, obits, 1819 pbits)) 1820 goto retry; 1821 if (obits & PG_G) 1822 pmap_invalidate_page(pmap, sva); 1823 else 1824 anychanged = 1; 1825 } 1826 } 1827 } 1828 sched_unpin(); 1829 vm_page_unlock_queues(); 1830 if (anychanged) 1831 pmap_invalidate_all(pmap); 1832 PMAP_UNLOCK(pmap); 1833} 1834 1835/* 1836 * Insert the given physical page (p) at 1837 * the specified virtual address (v) in the 1838 * target physical map with the protection requested. 1839 * 1840 * If specified, the page will be wired down, meaning 1841 * that the related pte can not be reclaimed. 1842 * 1843 * NB: This is the only routine which MAY NOT lazy-evaluate 1844 * or lose information. That is, this routine must actually 1845 * insert this page into the given map NOW. 1846 */ 1847void 1848pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1849 boolean_t wired) 1850{ 1851 vm_paddr_t pa; 1852 register pt_entry_t *pte; 1853 vm_paddr_t opa; 1854 pt_entry_t origpte, newpte; 1855 vm_page_t mpte, om; 1856 1857 va &= PG_FRAME; 1858#ifdef PMAP_DIAGNOSTIC 1859 if (va > VM_MAX_KERNEL_ADDRESS) 1860 panic("pmap_enter: toobig"); 1861 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1862 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1863#endif 1864 1865 mpte = NULL; 1866 1867 vm_page_lock_queues(); 1868 PMAP_LOCK(pmap); 1869 sched_pin(); 1870 1871 /* 1872 * In the case that a page table page is not 1873 * resident, we are creating it here. 1874 */ 1875 if (va < VM_MAXUSER_ADDRESS) { 1876 mpte = pmap_allocpte(pmap, va, M_WAITOK); 1877 } 1878#if 0 && defined(PMAP_DIAGNOSTIC) 1879 else { 1880 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1881 origpte = *pdeaddr; 1882 if ((origpte & PG_V) == 0) { 1883 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1884 pmap->pm_pdir[PTDPTDI], origpte, va); 1885 } 1886 } 1887#endif 1888 1889 pte = pmap_pte_quick(pmap, va); 1890 1891 /* 1892 * Page Directory table entry not valid, we need a new PT page 1893 */ 1894 if (pte == NULL) { 1895 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1896 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1897 } 1898 1899 pa = VM_PAGE_TO_PHYS(m); 1900 om = NULL; 1901 origpte = *pte; 1902 opa = origpte & PG_FRAME; 1903 1904 if (origpte & PG_PS) { 1905 /* 1906 * Yes, I know this will truncate upper address bits for PAE, 1907 * but I'm actually more interested in the lower bits 1908 */ 1909 printf("pmap_enter: va %p, pte %p, origpte %p\n", 1910 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 1911 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1912 } 1913 1914 /* 1915 * Mapping has not changed, must be protection or wiring change. 1916 */ 1917 if (origpte && (opa == pa)) { 1918 /* 1919 * Wiring change, just update stats. We don't worry about 1920 * wiring PT pages as they remain resident as long as there 1921 * are valid mappings in them. Hence, if a user page is wired, 1922 * the PT page will be also. 1923 */ 1924 if (wired && ((origpte & PG_W) == 0)) 1925 pmap->pm_stats.wired_count++; 1926 else if (!wired && (origpte & PG_W)) 1927 pmap->pm_stats.wired_count--; 1928 1929#if defined(PMAP_DIAGNOSTIC) 1930 if (pmap_nw_modified((pt_entry_t) origpte)) { 1931 printf( 1932 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 1933 va, origpte); 1934 } 1935#endif 1936 1937 /* 1938 * Remove extra pte reference 1939 */ 1940 if (mpte) 1941 mpte->wire_count--; 1942 1943 /* 1944 * We might be turning off write access to the page, 1945 * so we go ahead and sense modify status. 1946 */ 1947 if (origpte & PG_MANAGED) { 1948 om = m; 1949 pa |= PG_MANAGED; 1950 } 1951 goto validate; 1952 } 1953 /* 1954 * Mapping has changed, invalidate old range and fall through to 1955 * handle validating new mapping. 1956 */ 1957 if (opa) { 1958 if (origpte & PG_W) 1959 pmap->pm_stats.wired_count--; 1960 if (origpte & PG_MANAGED) { 1961 om = PHYS_TO_VM_PAGE(opa); 1962 pmap_remove_entry(pmap, om, va); 1963 } 1964 if (mpte != NULL) { 1965 mpte->wire_count--; 1966 KASSERT(mpte->wire_count > 0, 1967 ("pmap_enter: missing reference to page table page," 1968 " va: 0x%x", va)); 1969 } 1970 } else 1971 pmap->pm_stats.resident_count++; 1972 1973 /* 1974 * Enter on the PV list if part of our managed memory. 1975 */ 1976 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0) { 1977 pmap_insert_entry(pmap, va, m); 1978 pa |= PG_MANAGED; 1979 } 1980 1981 /* 1982 * Increment counters 1983 */ 1984 if (wired) 1985 pmap->pm_stats.wired_count++; 1986 1987validate: 1988 /* 1989 * Now validate mapping with desired protection/wiring. 1990 */ 1991 newpte = (pt_entry_t)(pa | PG_V); 1992 if ((prot & VM_PROT_WRITE) != 0) 1993 newpte |= PG_RW; 1994 if (wired) 1995 newpte |= PG_W; 1996 if (va < VM_MAXUSER_ADDRESS) 1997 newpte |= PG_U; 1998 if (pmap == kernel_pmap) 1999 newpte |= pgeflag; 2000 2001 /* 2002 * if the mapping or permission bits are different, we need 2003 * to update the pte. 2004 */ 2005 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2006 if (origpte & PG_MANAGED) { 2007 origpte = pte_load_store(pte, newpte | PG_A); 2008 if ((origpte & PG_M) && pmap_track_modified(va)) 2009 vm_page_dirty(om); 2010 if (origpte & PG_A) 2011 vm_page_flag_set(om, PG_REFERENCED); 2012 } else 2013 pte_store(pte, newpte | PG_A); 2014 if (origpte) { 2015 pmap_invalidate_page(pmap, va); 2016 } 2017 } 2018 sched_unpin(); 2019 vm_page_unlock_queues(); 2020 PMAP_UNLOCK(pmap); 2021} 2022 2023/* 2024 * this code makes some *MAJOR* assumptions: 2025 * 1. Current pmap & pmap exists. 2026 * 2. Not wired. 2027 * 3. Read access. 2028 * 4. No page table pages. 2029 * 6. Page IS managed. 2030 * but is *MUCH* faster than pmap_enter... 2031 */ 2032 2033vm_page_t 2034pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2035{ 2036 pt_entry_t *pte; 2037 vm_paddr_t pa; 2038 2039 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2040 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2041 PMAP_LOCK(pmap); 2042 2043 /* 2044 * In the case that a page table page is not 2045 * resident, we are creating it here. 2046 */ 2047 if (va < VM_MAXUSER_ADDRESS) { 2048 unsigned ptepindex; 2049 pd_entry_t ptepa; 2050 2051 /* 2052 * Calculate pagetable page index 2053 */ 2054 ptepindex = va >> PDRSHIFT; 2055 if (mpte && (mpte->pindex == ptepindex)) { 2056 mpte->wire_count++; 2057 } else { 2058retry: 2059 /* 2060 * Get the page directory entry 2061 */ 2062 ptepa = pmap->pm_pdir[ptepindex]; 2063 2064 /* 2065 * If the page table page is mapped, we just increment 2066 * the hold count, and activate it. 2067 */ 2068 if (ptepa) { 2069 if (ptepa & PG_PS) 2070 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2071 mpte = PHYS_TO_VM_PAGE(ptepa); 2072 mpte->wire_count++; 2073 } else { 2074 mpte = _pmap_allocpte(pmap, ptepindex, 2075 M_NOWAIT); 2076 if (mpte == NULL) { 2077 PMAP_UNLOCK(pmap); 2078 vm_page_busy(m); 2079 vm_page_unlock_queues(); 2080 VM_OBJECT_UNLOCK(m->object); 2081 VM_WAIT; 2082 VM_OBJECT_LOCK(m->object); 2083 vm_page_lock_queues(); 2084 vm_page_wakeup(m); 2085 PMAP_LOCK(pmap); 2086 goto retry; 2087 } 2088 } 2089 } 2090 } else { 2091 mpte = NULL; 2092 } 2093 2094 /* 2095 * This call to vtopte makes the assumption that we are 2096 * entering the page into the current pmap. In order to support 2097 * quick entry into any pmap, one would likely use pmap_pte_quick. 2098 * But that isn't as quick as vtopte. 2099 */ 2100 pte = vtopte(va); 2101 if (*pte) { 2102 if (mpte != NULL) { 2103 pmap_unwire_pte_hold(pmap, mpte); 2104 mpte = NULL; 2105 } 2106 goto out; 2107 } 2108 2109 /* 2110 * Enter on the PV list if part of our managed memory. Note that we 2111 * raise IPL while manipulating pv_table since pmap_enter can be 2112 * called at interrupt time. 2113 */ 2114 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2115 pmap_insert_entry(pmap, va, m); 2116 2117 /* 2118 * Increment counters 2119 */ 2120 pmap->pm_stats.resident_count++; 2121 2122 pa = VM_PAGE_TO_PHYS(m); 2123 2124 /* 2125 * Now validate mapping with RO protection 2126 */ 2127 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2128 pte_store(pte, pa | PG_V | PG_U); 2129 else 2130 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2131out: 2132 PMAP_UNLOCK(pmap); 2133 return mpte; 2134} 2135 2136/* 2137 * Make a temporary mapping for a physical address. This is only intended 2138 * to be used for panic dumps. 2139 */ 2140void * 2141pmap_kenter_temporary(vm_paddr_t pa, int i) 2142{ 2143 vm_offset_t va; 2144 2145 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2146 pmap_kenter(va, pa); 2147 invlpg(va); 2148 return ((void *)crashdumpmap); 2149} 2150 2151/* 2152 * This code maps large physical mmap regions into the 2153 * processor address space. Note that some shortcuts 2154 * are taken, but the code works. 2155 */ 2156void 2157pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2158 vm_object_t object, vm_pindex_t pindex, 2159 vm_size_t size) 2160{ 2161 vm_page_t p; 2162 2163 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2164 KASSERT(object->type == OBJT_DEVICE, 2165 ("pmap_object_init_pt: non-device object")); 2166 if (pseflag && 2167 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2168 int i; 2169 vm_page_t m[1]; 2170 unsigned int ptepindex; 2171 int npdes; 2172 pd_entry_t ptepa; 2173 2174 PMAP_LOCK(pmap); 2175 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2176 goto out; 2177 PMAP_UNLOCK(pmap); 2178retry: 2179 p = vm_page_lookup(object, pindex); 2180 if (p != NULL) { 2181 vm_page_lock_queues(); 2182 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2183 goto retry; 2184 } else { 2185 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2186 if (p == NULL) 2187 return; 2188 m[0] = p; 2189 2190 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2191 vm_page_lock_queues(); 2192 vm_page_free(p); 2193 vm_page_unlock_queues(); 2194 return; 2195 } 2196 2197 p = vm_page_lookup(object, pindex); 2198 vm_page_lock_queues(); 2199 vm_page_wakeup(p); 2200 } 2201 vm_page_unlock_queues(); 2202 2203 ptepa = VM_PAGE_TO_PHYS(p); 2204 if (ptepa & (NBPDR - 1)) 2205 return; 2206 2207 p->valid = VM_PAGE_BITS_ALL; 2208 2209 PMAP_LOCK(pmap); 2210 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2211 npdes = size >> PDRSHIFT; 2212 for(i = 0; i < npdes; i++) { 2213 pde_store(&pmap->pm_pdir[ptepindex], 2214 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2215 ptepa += NBPDR; 2216 ptepindex += 1; 2217 } 2218 pmap_invalidate_all(pmap); 2219out: 2220 PMAP_UNLOCK(pmap); 2221 } 2222} 2223 2224/* 2225 * Routine: pmap_change_wiring 2226 * Function: Change the wiring attribute for a map/virtual-address 2227 * pair. 2228 * In/out conditions: 2229 * The mapping must already exist in the pmap. 2230 */ 2231void 2232pmap_change_wiring(pmap, va, wired) 2233 register pmap_t pmap; 2234 vm_offset_t va; 2235 boolean_t wired; 2236{ 2237 register pt_entry_t *pte; 2238 2239 PMAP_LOCK(pmap); 2240 pte = pmap_pte(pmap, va); 2241 2242 if (wired && !pmap_pte_w(pte)) 2243 pmap->pm_stats.wired_count++; 2244 else if (!wired && pmap_pte_w(pte)) 2245 pmap->pm_stats.wired_count--; 2246 2247 /* 2248 * Wiring is not a hardware characteristic so there is no need to 2249 * invalidate TLB. 2250 */ 2251 pmap_pte_set_w(pte, wired); 2252 pmap_pte_release(pte); 2253 PMAP_UNLOCK(pmap); 2254} 2255 2256 2257 2258/* 2259 * Copy the range specified by src_addr/len 2260 * from the source map to the range dst_addr/len 2261 * in the destination map. 2262 * 2263 * This routine is only advisory and need not do anything. 2264 */ 2265 2266void 2267pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2268 vm_offset_t src_addr) 2269{ 2270 vm_offset_t addr; 2271 vm_offset_t end_addr = src_addr + len; 2272 vm_offset_t pdnxt; 2273 vm_page_t m; 2274 2275 if (dst_addr != src_addr) 2276 return; 2277 2278 if (!pmap_is_current(src_pmap)) 2279 return; 2280 2281 vm_page_lock_queues(); 2282 if (dst_pmap < src_pmap) { 2283 PMAP_LOCK(dst_pmap); 2284 PMAP_LOCK(src_pmap); 2285 } else { 2286 PMAP_LOCK(src_pmap); 2287 PMAP_LOCK(dst_pmap); 2288 } 2289 sched_pin(); 2290 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2291 pt_entry_t *src_pte, *dst_pte; 2292 vm_page_t dstmpte, srcmpte; 2293 pd_entry_t srcptepaddr; 2294 unsigned ptepindex; 2295 2296 if (addr >= UPT_MIN_ADDRESS) 2297 panic("pmap_copy: invalid to pmap_copy page tables"); 2298 2299 /* 2300 * Don't let optional prefaulting of pages make us go 2301 * way below the low water mark of free pages or way 2302 * above high water mark of used pv entries. 2303 */ 2304 if (cnt.v_free_count < cnt.v_free_reserved || 2305 pv_entry_count > pv_entry_high_water) 2306 break; 2307 2308 pdnxt = (addr + NBPDR) & ~PDRMASK; 2309 ptepindex = addr >> PDRSHIFT; 2310 2311 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2312 if (srcptepaddr == 0) 2313 continue; 2314 2315 if (srcptepaddr & PG_PS) { 2316 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2317 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2318 dst_pmap->pm_stats.resident_count += 2319 NBPDR / PAGE_SIZE; 2320 } 2321 continue; 2322 } 2323 2324 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2325 if (srcmpte->wire_count == 0) 2326 panic("pmap_copy: source page table page is unused"); 2327 2328 if (pdnxt > end_addr) 2329 pdnxt = end_addr; 2330 2331 src_pte = vtopte(addr); 2332 while (addr < pdnxt) { 2333 pt_entry_t ptetemp; 2334 ptetemp = *src_pte; 2335 /* 2336 * we only virtual copy managed pages 2337 */ 2338 if ((ptetemp & PG_MANAGED) != 0) { 2339 /* 2340 * We have to check after allocpte for the 2341 * pte still being around... allocpte can 2342 * block. 2343 */ 2344 dstmpte = pmap_allocpte(dst_pmap, addr, 2345 M_NOWAIT); 2346 if (dstmpte == NULL) 2347 break; 2348 dst_pte = pmap_pte_quick(dst_pmap, addr); 2349 if (*dst_pte == 0) { 2350 /* 2351 * Clear the modified and 2352 * accessed (referenced) bits 2353 * during the copy. 2354 */ 2355 m = PHYS_TO_VM_PAGE(ptetemp); 2356 *dst_pte = ptetemp & ~(PG_M | PG_A); 2357 dst_pmap->pm_stats.resident_count++; 2358 pmap_insert_entry(dst_pmap, addr, m); 2359 } else 2360 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2361 if (dstmpte->wire_count >= srcmpte->wire_count) 2362 break; 2363 } 2364 addr += PAGE_SIZE; 2365 src_pte++; 2366 } 2367 } 2368 sched_unpin(); 2369 vm_page_unlock_queues(); 2370 PMAP_UNLOCK(src_pmap); 2371 PMAP_UNLOCK(dst_pmap); 2372} 2373 2374static __inline void 2375pagezero(void *page) 2376{ 2377#if defined(I686_CPU) 2378 if (cpu_class == CPUCLASS_686) { 2379#if defined(CPU_ENABLE_SSE) 2380 if (cpu_feature & CPUID_SSE2) 2381 sse2_pagezero(page); 2382 else 2383#endif 2384 i686_pagezero(page); 2385 } else 2386#endif 2387 bzero(page, PAGE_SIZE); 2388} 2389 2390/* 2391 * pmap_zero_page zeros the specified hardware page by mapping 2392 * the page into KVM and using bzero to clear its contents. 2393 */ 2394void 2395pmap_zero_page(vm_page_t m) 2396{ 2397 struct sysmaps *sysmaps; 2398 2399 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2400 mtx_lock(&sysmaps->lock); 2401 if (*sysmaps->CMAP2) 2402 panic("pmap_zero_page: CMAP2 busy"); 2403 sched_pin(); 2404 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2405 invlcaddr(sysmaps->CADDR2); 2406 pagezero(sysmaps->CADDR2); 2407 *sysmaps->CMAP2 = 0; 2408 sched_unpin(); 2409 mtx_unlock(&sysmaps->lock); 2410} 2411 2412/* 2413 * pmap_zero_page_area zeros the specified hardware page by mapping 2414 * the page into KVM and using bzero to clear its contents. 2415 * 2416 * off and size may not cover an area beyond a single hardware page. 2417 */ 2418void 2419pmap_zero_page_area(vm_page_t m, int off, int size) 2420{ 2421 struct sysmaps *sysmaps; 2422 2423 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2424 mtx_lock(&sysmaps->lock); 2425 if (*sysmaps->CMAP2) 2426 panic("pmap_zero_page: CMAP2 busy"); 2427 sched_pin(); 2428 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2429 invlcaddr(sysmaps->CADDR2); 2430 if (off == 0 && size == PAGE_SIZE) 2431 pagezero(sysmaps->CADDR2); 2432 else 2433 bzero((char *)sysmaps->CADDR2 + off, size); 2434 *sysmaps->CMAP2 = 0; 2435 sched_unpin(); 2436 mtx_unlock(&sysmaps->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 struct sysmaps *sysmaps; 2469 2470 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2471 mtx_lock(&sysmaps->lock); 2472 if (*sysmaps->CMAP1) 2473 panic("pmap_copy_page: CMAP1 busy"); 2474 if (*sysmaps->CMAP2) 2475 panic("pmap_copy_page: CMAP2 busy"); 2476 sched_pin(); 2477 invlpg((u_int)sysmaps->CADDR1); 2478 invlpg((u_int)sysmaps->CADDR2); 2479 *sysmaps->CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2480 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2481 bcopy(sysmaps->CADDR1, sysmaps->CADDR2, PAGE_SIZE); 2482 *sysmaps->CMAP1 = 0; 2483 *sysmaps->CMAP2 = 0; 2484 sched_unpin(); 2485 mtx_unlock(&sysmaps->lock); 2486} 2487 2488/* 2489 * Returns true if the pmap's pv is one of the first 2490 * 16 pvs linked to from this page. This count may 2491 * be changed upwards or downwards in the future; it 2492 * is only necessary that true be returned for a small 2493 * subset of pmaps for proper page aging. 2494 */ 2495boolean_t 2496pmap_page_exists_quick(pmap, m) 2497 pmap_t pmap; 2498 vm_page_t m; 2499{ 2500 pv_entry_t pv; 2501 int loops = 0; 2502 2503 if (m->flags & PG_FICTITIOUS) 2504 return FALSE; 2505 2506 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2507 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2508 if (pv->pv_pmap == pmap) { 2509 return TRUE; 2510 } 2511 loops++; 2512 if (loops >= 16) 2513 break; 2514 } 2515 return (FALSE); 2516} 2517 2518#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2519/* 2520 * Remove all pages from specified address space 2521 * this aids process exit speeds. Also, this code 2522 * is special cased for current process only, but 2523 * can have the more generic (and slightly slower) 2524 * mode enabled. This is much faster than pmap_remove 2525 * in the case of running down an entire address space. 2526 */ 2527void 2528pmap_remove_pages(pmap, sva, eva) 2529 pmap_t pmap; 2530 vm_offset_t sva, eva; 2531{ 2532 pt_entry_t *pte, tpte; 2533 vm_page_t m; 2534 pv_entry_t pv, npv; 2535 2536#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2537 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) { 2538 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2539 return; 2540 } 2541#endif 2542 vm_page_lock_queues(); 2543 PMAP_LOCK(pmap); 2544 sched_pin(); 2545 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2546 2547 if (pv->pv_va >= eva || pv->pv_va < sva) { 2548 npv = TAILQ_NEXT(pv, pv_plist); 2549 continue; 2550 } 2551 2552#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2553 pte = vtopte(pv->pv_va); 2554#else 2555 pte = pmap_pte_quick(pmap, pv->pv_va); 2556#endif 2557 tpte = *pte; 2558 2559 if (tpte == 0) { 2560 printf("TPTE at %p IS ZERO @ VA %08x\n", 2561 pte, pv->pv_va); 2562 panic("bad pte"); 2563 } 2564 2565/* 2566 * We cannot remove wired pages from a process' mapping at this time 2567 */ 2568 if (tpte & PG_W) { 2569 npv = TAILQ_NEXT(pv, pv_plist); 2570 continue; 2571 } 2572 2573 m = PHYS_TO_VM_PAGE(tpte); 2574 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2575 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2576 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2577 2578 KASSERT(m < &vm_page_array[vm_page_array_size], 2579 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2580 2581 pmap->pm_stats.resident_count--; 2582 2583 pte_clear(pte); 2584 2585 /* 2586 * Update the vm_page_t clean and reference bits. 2587 */ 2588 if (tpte & PG_M) { 2589 vm_page_dirty(m); 2590 } 2591 2592 npv = TAILQ_NEXT(pv, pv_plist); 2593 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 2594 2595 m->md.pv_list_count--; 2596 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2597 if (TAILQ_EMPTY(&m->md.pv_list)) 2598 vm_page_flag_clear(m, PG_WRITEABLE); 2599 2600 pmap_unuse_pt(pmap, pv->pv_va); 2601 free_pv_entry(pv); 2602 } 2603 sched_unpin(); 2604 pmap_invalidate_all(pmap); 2605 PMAP_UNLOCK(pmap); 2606 vm_page_unlock_queues(); 2607} 2608 2609/* 2610 * pmap_is_modified: 2611 * 2612 * Return whether or not the specified physical page was modified 2613 * in any physical maps. 2614 */ 2615boolean_t 2616pmap_is_modified(vm_page_t m) 2617{ 2618 pv_entry_t pv; 2619 pt_entry_t *pte; 2620 boolean_t rv; 2621 2622 rv = FALSE; 2623 if (m->flags & PG_FICTITIOUS) 2624 return (rv); 2625 2626 sched_pin(); 2627 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2628 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2629 /* 2630 * if the bit being tested is the modified bit, then 2631 * mark clean_map and ptes as never 2632 * modified. 2633 */ 2634 if (!pmap_track_modified(pv->pv_va)) 2635 continue; 2636 PMAP_LOCK(pv->pv_pmap); 2637 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2638 rv = (*pte & PG_M) != 0; 2639 PMAP_UNLOCK(pv->pv_pmap); 2640 if (rv) 2641 break; 2642 } 2643 sched_unpin(); 2644 return (rv); 2645} 2646 2647/* 2648 * pmap_is_prefaultable: 2649 * 2650 * Return whether or not the specified virtual address is elgible 2651 * for prefault. 2652 */ 2653boolean_t 2654pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2655{ 2656 pt_entry_t *pte; 2657 boolean_t rv; 2658 2659 rv = FALSE; 2660 PMAP_LOCK(pmap); 2661 if (*pmap_pde(pmap, addr)) { 2662 pte = vtopte(addr); 2663 rv = *pte == 0; 2664 } 2665 PMAP_UNLOCK(pmap); 2666 return (rv); 2667} 2668 2669/* 2670 * Clear the given bit in each of the given page's ptes. The bit is 2671 * expressed as a 32-bit mask. Consequently, if the pte is 64 bits in 2672 * size, only a bit within the least significant 32 can be cleared. 2673 */ 2674static __inline void 2675pmap_clear_ptes(vm_page_t m, int bit) 2676{ 2677 register pv_entry_t pv; 2678 pt_entry_t pbits, *pte; 2679 2680 if ((m->flags & PG_FICTITIOUS) || 2681 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2682 return; 2683 2684 sched_pin(); 2685 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2686 /* 2687 * Loop over all current mappings setting/clearing as appropos If 2688 * setting RO do we need to clear the VAC? 2689 */ 2690 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2691 /* 2692 * don't write protect pager mappings 2693 */ 2694 if (bit == PG_RW) { 2695 if (!pmap_track_modified(pv->pv_va)) 2696 continue; 2697 } 2698 2699 PMAP_LOCK(pv->pv_pmap); 2700 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2701retry: 2702 pbits = *pte; 2703 if (pbits & bit) { 2704 if (bit == PG_RW) { 2705 /* 2706 * Regardless of whether a pte is 32 or 64 bits 2707 * in size, PG_RW and PG_M are among the least 2708 * significant 32 bits. 2709 */ 2710 if (!atomic_cmpset_int((u_int *)pte, pbits, 2711 pbits & ~(PG_RW | PG_M))) 2712 goto retry; 2713 if (pbits & PG_M) { 2714 vm_page_dirty(m); 2715 } 2716 } else { 2717 atomic_clear_int((u_int *)pte, bit); 2718 } 2719 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2720 } 2721 PMAP_UNLOCK(pv->pv_pmap); 2722 } 2723 if (bit == PG_RW) 2724 vm_page_flag_clear(m, PG_WRITEABLE); 2725 sched_unpin(); 2726} 2727 2728/* 2729 * pmap_page_protect: 2730 * 2731 * Lower the permission for all mappings to a given page. 2732 */ 2733void 2734pmap_page_protect(vm_page_t m, vm_prot_t prot) 2735{ 2736 if ((prot & VM_PROT_WRITE) == 0) { 2737 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2738 pmap_clear_ptes(m, PG_RW); 2739 } else { 2740 pmap_remove_all(m); 2741 } 2742 } 2743} 2744 2745/* 2746 * pmap_ts_referenced: 2747 * 2748 * Return a count of reference bits for a page, clearing those bits. 2749 * It is not necessary for every reference bit to be cleared, but it 2750 * is necessary that 0 only be returned when there are truly no 2751 * reference bits set. 2752 * 2753 * XXX: The exact number of bits to check and clear is a matter that 2754 * should be tested and standardized at some point in the future for 2755 * optimal aging of shared pages. 2756 */ 2757int 2758pmap_ts_referenced(vm_page_t m) 2759{ 2760 register pv_entry_t pv, pvf, pvn; 2761 pt_entry_t *pte; 2762 pt_entry_t v; 2763 int rtval = 0; 2764 2765 if (m->flags & PG_FICTITIOUS) 2766 return (rtval); 2767 2768 sched_pin(); 2769 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2770 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2771 2772 pvf = pv; 2773 2774 do { 2775 pvn = TAILQ_NEXT(pv, pv_list); 2776 2777 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2778 2779 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2780 2781 if (!pmap_track_modified(pv->pv_va)) 2782 continue; 2783 2784 PMAP_LOCK(pv->pv_pmap); 2785 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2786 2787 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2788 atomic_clear_int((u_int *)pte, PG_A); 2789 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2790 2791 rtval++; 2792 if (rtval > 4) { 2793 PMAP_UNLOCK(pv->pv_pmap); 2794 break; 2795 } 2796 } 2797 PMAP_UNLOCK(pv->pv_pmap); 2798 } while ((pv = pvn) != NULL && pv != pvf); 2799 } 2800 sched_unpin(); 2801 2802 return (rtval); 2803} 2804 2805/* 2806 * Clear the modify bits on the specified physical page. 2807 */ 2808void 2809pmap_clear_modify(vm_page_t m) 2810{ 2811 pmap_clear_ptes(m, PG_M); 2812} 2813 2814/* 2815 * pmap_clear_reference: 2816 * 2817 * Clear the reference bit on the specified physical page. 2818 */ 2819void 2820pmap_clear_reference(vm_page_t m) 2821{ 2822 pmap_clear_ptes(m, PG_A); 2823} 2824 2825/* 2826 * Miscellaneous support routines follow 2827 */ 2828 2829/* 2830 * Map a set of physical memory pages into the kernel virtual 2831 * address space. Return a pointer to where it is mapped. This 2832 * routine is intended to be used for mapping device memory, 2833 * NOT real memory. 2834 */ 2835void * 2836pmap_mapdev(pa, size) 2837 vm_paddr_t pa; 2838 vm_size_t size; 2839{ 2840 vm_offset_t va, tmpva, offset; 2841 2842 offset = pa & PAGE_MASK; 2843 size = roundup(offset + size, PAGE_SIZE); 2844 pa = pa & PG_FRAME; 2845 2846 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2847 va = KERNBASE + pa; 2848 else 2849 va = kmem_alloc_nofault(kernel_map, size); 2850 if (!va) 2851 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2852 2853 for (tmpva = va; size > 0; ) { 2854 pmap_kenter(tmpva, pa); 2855 size -= PAGE_SIZE; 2856 tmpva += PAGE_SIZE; 2857 pa += PAGE_SIZE; 2858 } 2859 pmap_invalidate_range(kernel_pmap, va, tmpva); 2860 return ((void *)(va + offset)); 2861} 2862 2863void 2864pmap_unmapdev(va, size) 2865 vm_offset_t va; 2866 vm_size_t size; 2867{ 2868 vm_offset_t base, offset, tmpva; 2869 2870 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2871 return; 2872 base = va & PG_FRAME; 2873 offset = va & PAGE_MASK; 2874 size = roundup(offset + size, PAGE_SIZE); 2875 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2876 pmap_kremove(tmpva); 2877 pmap_invalidate_range(kernel_pmap, va, tmpva); 2878 kmem_free(kernel_map, base, size); 2879} 2880 2881/* 2882 * perform the pmap work for mincore 2883 */ 2884int 2885pmap_mincore(pmap, addr) 2886 pmap_t pmap; 2887 vm_offset_t addr; 2888{ 2889 pt_entry_t *ptep, pte; 2890 vm_page_t m; 2891 int val = 0; 2892 2893 PMAP_LOCK(pmap); 2894 ptep = pmap_pte(pmap, addr); 2895 pte = (ptep != NULL) ? *ptep : 0; 2896 pmap_pte_release(ptep); 2897 PMAP_UNLOCK(pmap); 2898 2899 if (pte != 0) { 2900 vm_paddr_t pa; 2901 2902 val = MINCORE_INCORE; 2903 if ((pte & PG_MANAGED) == 0) 2904 return val; 2905 2906 pa = pte & PG_FRAME; 2907 2908 m = PHYS_TO_VM_PAGE(pa); 2909 2910 /* 2911 * Modified by us 2912 */ 2913 if (pte & PG_M) 2914 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2915 else { 2916 /* 2917 * Modified by someone else 2918 */ 2919 vm_page_lock_queues(); 2920 if (m->dirty || pmap_is_modified(m)) 2921 val |= MINCORE_MODIFIED_OTHER; 2922 vm_page_unlock_queues(); 2923 } 2924 /* 2925 * Referenced by us 2926 */ 2927 if (pte & PG_A) 2928 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2929 else { 2930 /* 2931 * Referenced by someone else 2932 */ 2933 vm_page_lock_queues(); 2934 if ((m->flags & PG_REFERENCED) || 2935 pmap_ts_referenced(m)) { 2936 val |= MINCORE_REFERENCED_OTHER; 2937 vm_page_flag_set(m, PG_REFERENCED); 2938 } 2939 vm_page_unlock_queues(); 2940 } 2941 } 2942 return val; 2943} 2944 2945void 2946pmap_activate(struct thread *td) 2947{ 2948 struct proc *p = td->td_proc; 2949 pmap_t pmap, oldpmap; 2950 u_int32_t cr3; 2951 2952 critical_enter(); 2953 pmap = vmspace_pmap(td->td_proc->p_vmspace); 2954 oldpmap = PCPU_GET(curpmap); 2955#if defined(SMP) 2956 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 2957 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 2958#else 2959 oldpmap->pm_active &= ~1; 2960 pmap->pm_active |= 1; 2961#endif 2962#ifdef PAE 2963 cr3 = vtophys(pmap->pm_pdpt); 2964#else 2965 cr3 = vtophys(pmap->pm_pdir); 2966#endif 2967 /* XXXKSE this is wrong. 2968 * pmap_activate is for the current thread on the current cpu 2969 */ 2970 if (p->p_flag & P_SA) { 2971 /* Make sure all other cr3 entries are updated. */ 2972 /* what if they are running? XXXKSE (maybe abort them) */ 2973 FOREACH_THREAD_IN_PROC(p, td) { 2974 td->td_pcb->pcb_cr3 = cr3; 2975 } 2976 } else { 2977 td->td_pcb->pcb_cr3 = cr3; 2978 } 2979 load_cr3(cr3); 2980 PCPU_SET(curpmap, pmap); 2981 critical_exit(); 2982} 2983 2984vm_offset_t 2985pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 2986{ 2987 2988 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 2989 return addr; 2990 } 2991 2992 addr = (addr + PDRMASK) & ~PDRMASK; 2993 return addr; 2994} 2995 2996 2997#if defined(PMAP_DEBUG) 2998pmap_pid_dump(int pid) 2999{ 3000 pmap_t pmap; 3001 struct proc *p; 3002 int npte = 0; 3003 int index; 3004 3005 sx_slock(&allproc_lock); 3006 LIST_FOREACH(p, &allproc, p_list) { 3007 if (p->p_pid != pid) 3008 continue; 3009 3010 if (p->p_vmspace) { 3011 int i,j; 3012 index = 0; 3013 pmap = vmspace_pmap(p->p_vmspace); 3014 for (i = 0; i < NPDEPTD; i++) { 3015 pd_entry_t *pde; 3016 pt_entry_t *pte; 3017 vm_offset_t base = i << PDRSHIFT; 3018 3019 pde = &pmap->pm_pdir[i]; 3020 if (pde && pmap_pde_v(pde)) { 3021 for (j = 0; j < NPTEPG; j++) { 3022 vm_offset_t va = base + (j << PAGE_SHIFT); 3023 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3024 if (index) { 3025 index = 0; 3026 printf("\n"); 3027 } 3028 sx_sunlock(&allproc_lock); 3029 return npte; 3030 } 3031 pte = pmap_pte(pmap, va); 3032 if (pte && pmap_pte_v(pte)) { 3033 pt_entry_t pa; 3034 vm_page_t m; 3035 pa = *pte; 3036 m = PHYS_TO_VM_PAGE(pa); 3037 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3038 va, pa, m->hold_count, m->wire_count, m->flags); 3039 npte++; 3040 index++; 3041 if (index >= 2) { 3042 index = 0; 3043 printf("\n"); 3044 } else { 3045 printf(" "); 3046 } 3047 } 3048 } 3049 } 3050 } 3051 } 3052 } 3053 sx_sunlock(&allproc_lock); 3054 return npte; 3055} 3056#endif 3057 3058#if defined(DEBUG) 3059 3060static void pads(pmap_t pm); 3061void pmap_pvdump(vm_offset_t pa); 3062 3063/* print address space of pmap*/ 3064static void 3065pads(pm) 3066 pmap_t pm; 3067{ 3068 int i, j; 3069 vm_paddr_t va; 3070 pt_entry_t *ptep; 3071 3072 if (pm == kernel_pmap) 3073 return; 3074 for (i = 0; i < NPDEPTD; i++) 3075 if (pm->pm_pdir[i]) 3076 for (j = 0; j < NPTEPG; j++) { 3077 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3078 if (pm == kernel_pmap && va < KERNBASE) 3079 continue; 3080 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3081 continue; 3082 ptep = pmap_pte(pm, va); 3083 if (pmap_pte_v(ptep)) 3084 printf("%x:%x ", va, *ptep); 3085 }; 3086 3087} 3088 3089void 3090pmap_pvdump(pa) 3091 vm_paddr_t pa; 3092{ 3093 pv_entry_t pv; 3094 vm_page_t m; 3095 3096 printf("pa %x", pa); 3097 m = PHYS_TO_VM_PAGE(pa); 3098 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3099 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3100 pads(pv->pv_pmap); 3101 } 3102 printf(" "); 3103} 3104#endif 3105