pmap.c revision 133124
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 133124 2004-08-04 18:04:44Z 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#ifdef SMP 189static struct mtx lazypmap_lock; 190#endif 191 192vm_paddr_t avail_end; /* PA of last available physical page */ 193vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 194vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 195static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 196int pgeflag = 0; /* PG_G or-in */ 197int pseflag = 0; /* PG_PS or-in */ 198 199static int nkpt; 200vm_offset_t kernel_vm_end; 201extern u_int32_t KERNend; 202 203#ifdef PAE 204static uma_zone_t pdptzone; 205#endif 206 207/* 208 * Data for the pv entry allocation mechanism 209 */ 210static uma_zone_t pvzone; 211static struct vm_object pvzone_obj; 212static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 213int pmap_pagedaemon_waken; 214 215/* 216 * All those kernel PT submaps that BSD is so fond of 217 */ 218pt_entry_t *CMAP1 = 0; 219static pt_entry_t *CMAP2, *CMAP3; 220caddr_t CADDR1 = 0, ptvmmap = 0; 221static caddr_t CADDR2, CADDR3; 222static struct mtx CMAPCADDR12_lock; 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"); 250 251static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 252static pv_entry_t get_pv_entry(void); 253static void pmap_clear_ptes(vm_page_t m, int bit); 254 255static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 256static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 257static int pmap_remove_entry(struct pmap *pmap, vm_page_t m, 258 vm_offset_t va); 259static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); 260 261static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va); 262 263static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex); 264static int _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m); 265static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 266static int pmap_unuse_pt(pmap_t, vm_offset_t); 267static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 268#ifdef PAE 269static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 270#endif 271 272CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 273CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 274 275/* 276 * Move the kernel virtual free pointer to the next 277 * 4MB. This is used to help improve performance 278 * by using a large (4MB) page for much of the kernel 279 * (.text, .data, .bss) 280 */ 281static vm_offset_t 282pmap_kmem_choose(vm_offset_t addr) 283{ 284 vm_offset_t newaddr = addr; 285 286#ifndef DISABLE_PSE 287 if (cpu_feature & CPUID_PSE) 288 newaddr = (addr + PDRMASK) & ~PDRMASK; 289#endif 290 return newaddr; 291} 292 293/* 294 * Bootstrap the system enough to run with virtual memory. 295 * 296 * On the i386 this is called after mapping has already been enabled 297 * and just syncs the pmap module with what has already been done. 298 * [We can't call it easily with mapping off since the kernel is not 299 * mapped with PA == VA, hence we would have to relocate every address 300 * from the linked base (virtual) address "KERNBASE" to the actual 301 * (physical) address starting relative to 0] 302 */ 303void 304pmap_bootstrap(firstaddr, loadaddr) 305 vm_paddr_t firstaddr; 306 vm_paddr_t loadaddr; 307{ 308 vm_offset_t va; 309 pt_entry_t *pte, *unused; 310 int i; 311 312 /* 313 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 314 * large. It should instead be correctly calculated in locore.s and 315 * not based on 'first' (which is a physical address, not a virtual 316 * address, for the start of unused physical memory). The kernel 317 * page tables are NOT double mapped and thus should not be included 318 * in this calculation. 319 */ 320 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 321 virtual_avail = pmap_kmem_choose(virtual_avail); 322 323 virtual_end = VM_MAX_KERNEL_ADDRESS; 324 325 /* 326 * Initialize the kernel pmap (which is statically allocated). 327 */ 328 PMAP_LOCK_INIT(kernel_pmap); 329 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 330#ifdef PAE 331 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 332#endif 333 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 334 TAILQ_INIT(&kernel_pmap->pm_pvlist); 335 LIST_INIT(&allpmaps); 336#ifdef SMP 337 mtx_init(&lazypmap_lock, "lazypmap", NULL, MTX_SPIN); 338#endif 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 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 360 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 361 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 362 *CMAP3 = 0; 363 364 mtx_init(&CMAPCADDR12_lock, "CMAPCADDR12", NULL, MTX_DEF); 365 366 /* 367 * Crashdump maps. 368 */ 369 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS) 370 371 /* 372 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 373 */ 374 SYSMAP(caddr_t, unused, ptvmmap, 1) 375 376 /* 377 * msgbufp is used to map the system message buffer. 378 */ 379 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE))) 380 381 /* 382 * ptemap is used for pmap_pte_quick 383 */ 384 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 385 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1); 386 387 virtual_avail = va; 388 389 *CMAP1 = *CMAP2 = 0; 390 for (i = 0; i < NKPT; i++) 391 PTD[i] = 0; 392 393 /* Turn on PG_G on kernel page(s) */ 394 pmap_set_pg(); 395} 396 397/* 398 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 399 */ 400void 401pmap_set_pg(void) 402{ 403 pd_entry_t pdir; 404 pt_entry_t *pte; 405 vm_offset_t va, endva; 406 int i; 407 408 if (pgeflag == 0) 409 return; 410 411 i = KERNLOAD/NBPDR; 412 endva = KERNBASE + KERNend; 413 414 if (pseflag) { 415 va = KERNBASE + KERNLOAD; 416 while (va < endva) { 417 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 418 pdir |= pgeflag; 419 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 420 invltlb(); /* Play it safe, invltlb() every time */ 421 i++; 422 va += NBPDR; 423 } 424 } else { 425 va = (vm_offset_t)btext; 426 while (va < endva) { 427 pte = vtopte(va); 428 if (*pte) 429 *pte |= pgeflag; 430 invltlb(); /* Play it safe, invltlb() every time */ 431 va += PAGE_SIZE; 432 } 433 } 434} 435 436#ifdef PAE 437 438static MALLOC_DEFINE(M_PMAPPDPT, "pmap", "pmap pdpt"); 439 440static void * 441pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 442{ 443 *flags = UMA_SLAB_PRIV; 444 return (contigmalloc(PAGE_SIZE, M_PMAPPDPT, 0, 0x0ULL, 0xffffffffULL, 445 1, 0)); 446} 447#endif 448 449/* 450 * Initialize the pmap module. 451 * Called by vm_init, to initialize any structures that the pmap 452 * system needs to map virtual memory. 453 * pmap_init has been enhanced to support in a fairly consistant 454 * way, discontiguous physical memory. 455 */ 456void 457pmap_init(void) 458{ 459 int i; 460 461 /* 462 * Allocate memory for random pmap data structures. Includes the 463 * pv_head_table. 464 */ 465 466 for(i = 0; i < vm_page_array_size; i++) { 467 vm_page_t m; 468 469 m = &vm_page_array[i]; 470 TAILQ_INIT(&m->md.pv_list); 471 m->md.pv_list_count = 0; 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 * Now it is safe to enable pv_table recording. 490 */ 491 pmap_initialized = TRUE; 492} 493 494/* 495 * Initialize the address space (zone) for the pv_entries. Set a 496 * high water mark so that the system can recover from excessive 497 * numbers of pv entries. 498 */ 499void 500pmap_init2() 501{ 502 int shpgperproc = PMAP_SHPGPERPROC; 503 504 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 505 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 506 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 507 pv_entry_high_water = 9 * (pv_entry_max / 10); 508 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 509} 510 511 512/*************************************************** 513 * Low level helper routines..... 514 ***************************************************/ 515 516#if defined(PMAP_DIAGNOSTIC) 517 518/* 519 * This code checks for non-writeable/modified pages. 520 * This should be an invalid condition. 521 */ 522static int 523pmap_nw_modified(pt_entry_t ptea) 524{ 525 int pte; 526 527 pte = (int) ptea; 528 529 if ((pte & (PG_M|PG_RW)) == PG_M) 530 return 1; 531 else 532 return 0; 533} 534#endif 535 536 537/* 538 * this routine defines the region(s) of memory that should 539 * not be tested for the modified bit. 540 */ 541static PMAP_INLINE int 542pmap_track_modified(vm_offset_t va) 543{ 544 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 545 return 1; 546 else 547 return 0; 548} 549 550#ifdef I386_CPU 551/* 552 * i386 only has "invalidate everything" and no SMP to worry about. 553 */ 554PMAP_INLINE void 555pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 556{ 557 558 if (pmap == kernel_pmap || pmap->pm_active) 559 invltlb(); 560} 561 562PMAP_INLINE void 563pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 564{ 565 566 if (pmap == kernel_pmap || pmap->pm_active) 567 invltlb(); 568} 569 570PMAP_INLINE void 571pmap_invalidate_all(pmap_t pmap) 572{ 573 574 if (pmap == kernel_pmap || pmap->pm_active) 575 invltlb(); 576} 577#else /* !I386_CPU */ 578#ifdef SMP 579/* 580 * For SMP, these functions have to use the IPI mechanism for coherence. 581 */ 582void 583pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 584{ 585 u_int cpumask; 586 u_int other_cpus; 587 588 if (smp_started) { 589 if (!(read_eflags() & PSL_I)) 590 panic("%s: interrupts disabled", __func__); 591 mtx_lock_spin(&smp_tlb_mtx); 592 } else 593 critical_enter(); 594 /* 595 * We need to disable interrupt preemption but MUST NOT have 596 * interrupts disabled here. 597 * XXX we may need to hold schedlock to get a coherent pm_active 598 * XXX critical sections disable interrupts again 599 */ 600 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 601 invlpg(va); 602 smp_invlpg(va); 603 } else { 604 cpumask = PCPU_GET(cpumask); 605 other_cpus = PCPU_GET(other_cpus); 606 if (pmap->pm_active & cpumask) 607 invlpg(va); 608 if (pmap->pm_active & other_cpus) 609 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 610 } 611 if (smp_started) 612 mtx_unlock_spin(&smp_tlb_mtx); 613 else 614 critical_exit(); 615} 616 617void 618pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 619{ 620 u_int cpumask; 621 u_int other_cpus; 622 vm_offset_t addr; 623 624 if (smp_started) { 625 if (!(read_eflags() & PSL_I)) 626 panic("%s: interrupts disabled", __func__); 627 mtx_lock_spin(&smp_tlb_mtx); 628 } else 629 critical_enter(); 630 /* 631 * We need to disable interrupt preemption but MUST NOT have 632 * interrupts disabled here. 633 * XXX we may need to hold schedlock to get a coherent pm_active 634 * XXX critical sections disable interrupts again 635 */ 636 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 637 for (addr = sva; addr < eva; addr += PAGE_SIZE) 638 invlpg(addr); 639 smp_invlpg_range(sva, eva); 640 } else { 641 cpumask = PCPU_GET(cpumask); 642 other_cpus = PCPU_GET(other_cpus); 643 if (pmap->pm_active & cpumask) 644 for (addr = sva; addr < eva; addr += PAGE_SIZE) 645 invlpg(addr); 646 if (pmap->pm_active & other_cpus) 647 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 648 sva, eva); 649 } 650 if (smp_started) 651 mtx_unlock_spin(&smp_tlb_mtx); 652 else 653 critical_exit(); 654} 655 656void 657pmap_invalidate_all(pmap_t pmap) 658{ 659 u_int cpumask; 660 u_int other_cpus; 661 662 if (smp_started) { 663 if (!(read_eflags() & PSL_I)) 664 panic("%s: interrupts disabled", __func__); 665 mtx_lock_spin(&smp_tlb_mtx); 666 } else 667 critical_enter(); 668 /* 669 * We need to disable interrupt preemption but MUST NOT have 670 * interrupts disabled here. 671 * XXX we may need to hold schedlock to get a coherent pm_active 672 * XXX critical sections disable interrupts again 673 */ 674 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 675 invltlb(); 676 smp_invltlb(); 677 } else { 678 cpumask = PCPU_GET(cpumask); 679 other_cpus = PCPU_GET(other_cpus); 680 if (pmap->pm_active & cpumask) 681 invltlb(); 682 if (pmap->pm_active & other_cpus) 683 smp_masked_invltlb(pmap->pm_active & other_cpus); 684 } 685 if (smp_started) 686 mtx_unlock_spin(&smp_tlb_mtx); 687 else 688 critical_exit(); 689} 690#else /* !SMP */ 691/* 692 * Normal, non-SMP, 486+ invalidation functions. 693 * We inline these within pmap.c for speed. 694 */ 695PMAP_INLINE void 696pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 697{ 698 699 if (pmap == kernel_pmap || pmap->pm_active) 700 invlpg(va); 701} 702 703PMAP_INLINE void 704pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 705{ 706 vm_offset_t addr; 707 708 if (pmap == kernel_pmap || pmap->pm_active) 709 for (addr = sva; addr < eva; addr += PAGE_SIZE) 710 invlpg(addr); 711} 712 713PMAP_INLINE void 714pmap_invalidate_all(pmap_t pmap) 715{ 716 717 if (pmap == kernel_pmap || pmap->pm_active) 718 invltlb(); 719} 720#endif /* !SMP */ 721#endif /* !I386_CPU */ 722 723/* 724 * Are we current address space or kernel? N.B. We return FALSE when 725 * a pmap's page table is in use because a kernel thread is borrowing 726 * it. The borrowed page table can change spontaneously, making any 727 * dependence on its continued use subject to a race condition. 728 */ 729static __inline int 730pmap_is_current(pmap_t pmap) 731{ 732 733 return (pmap == kernel_pmap || 734 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 735 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 736} 737 738/* 739 * If the given pmap is not the current pmap, Giant must be held. 740 */ 741pt_entry_t * 742pmap_pte(pmap_t pmap, vm_offset_t va) 743{ 744 pd_entry_t newpf; 745 pd_entry_t *pde; 746 747 pde = pmap_pde(pmap, va); 748 if (*pde & PG_PS) 749 return (pde); 750 if (*pde != 0) { 751 /* are we current address space or kernel? */ 752 if (pmap_is_current(pmap)) 753 return (vtopte(va)); 754 GIANT_REQUIRED; 755 newpf = *pde & PG_FRAME; 756 if ((*PMAP2 & PG_FRAME) != newpf) { 757 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; 758 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 759 } 760 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); 761 } 762 return (0); 763} 764 765static __inline void 766invlcaddr(void *caddr) 767{ 768#ifdef I386_CPU 769 invltlb(); 770#else 771 invlpg((u_int)caddr); 772#endif 773} 774 775/* 776 * Super fast pmap_pte routine best used when scanning 777 * the pv lists. This eliminates many coarse-grained 778 * invltlb calls. Note that many of the pv list 779 * scans are across different pmaps. It is very wasteful 780 * to do an entire invltlb for checking a single mapping. 781 * 782 * If the given pmap is not the current pmap, vm_page_queue_mtx 783 * must be held and curthread pinned to a CPU. 784 */ 785static pt_entry_t * 786pmap_pte_quick(pmap_t pmap, vm_offset_t va) 787{ 788 pd_entry_t newpf; 789 pd_entry_t *pde; 790 791 pde = pmap_pde(pmap, va); 792 if (*pde & PG_PS) 793 return (pde); 794 if (*pde != 0) { 795 /* are we current address space or kernel? */ 796 if (pmap_is_current(pmap)) 797 return (vtopte(va)); 798 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 799 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 800 newpf = *pde & PG_FRAME; 801 if ((*PMAP1 & PG_FRAME) != newpf) { 802 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; 803#ifdef SMP 804 PMAP1cpu = PCPU_GET(cpuid); 805#endif 806 invlcaddr(PADDR1); 807 PMAP1changed++; 808 } else 809#ifdef SMP 810 if (PMAP1cpu != PCPU_GET(cpuid)) { 811 PMAP1cpu = PCPU_GET(cpuid); 812 invlcaddr(PADDR1); 813 PMAP1changedcpu++; 814 } else 815#endif 816 PMAP1unchanged++; 817 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); 818 } 819 return (0); 820} 821 822/* 823 * Routine: pmap_extract 824 * Function: 825 * Extract the physical page address associated 826 * with the given map/virtual_address pair. 827 */ 828vm_paddr_t 829pmap_extract(pmap_t pmap, vm_offset_t va) 830{ 831 vm_paddr_t rtval; 832 pt_entry_t *pte; 833 pd_entry_t pde; 834 835 rtval = 0; 836 if (pmap == NULL) 837 return (rtval); 838 PMAP_LOCK(pmap); 839 pde = pmap->pm_pdir[va >> PDRSHIFT]; 840 if (pde != 0) { 841 if ((pde & PG_PS) != 0) { 842 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 843 PMAP_UNLOCK(pmap); 844 return rtval; 845 } 846 pte = pmap_pte(pmap, va); 847 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK); 848 } 849 PMAP_UNLOCK(pmap); 850 return (rtval); 851} 852 853/* 854 * Routine: pmap_extract_and_hold 855 * Function: 856 * Atomically extract and hold the physical page 857 * with the given pmap and virtual address pair 858 * if that mapping permits the given protection. 859 */ 860vm_page_t 861pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 862{ 863 pd_entry_t pde; 864 pt_entry_t pte; 865 vm_page_t m; 866 867 m = NULL; 868 if (pmap == NULL) 869 return (m); 870 vm_page_lock_queues(); 871 PMAP_LOCK(pmap); 872 pde = *pmap_pde(pmap, va); 873 if (pde != 0) { 874 if (pde & PG_PS) { 875 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) { 876 m = PHYS_TO_VM_PAGE((pde & ~PDRMASK) | 877 (va & PDRMASK)); 878 vm_page_hold(m); 879 } 880 } else { 881 sched_pin(); 882 pte = *pmap_pte_quick(pmap, va); 883 if (pte != 0 && 884 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) { 885 m = PHYS_TO_VM_PAGE(pte & PG_FRAME); 886 vm_page_hold(m); 887 } 888 sched_unpin(); 889 } 890 } 891 vm_page_unlock_queues(); 892 PMAP_UNLOCK(pmap); 893 return (m); 894} 895 896/*************************************************** 897 * Low level mapping routines..... 898 ***************************************************/ 899 900/* 901 * Add a wired page to the kva. 902 * Note: not SMP coherent. 903 */ 904PMAP_INLINE void 905pmap_kenter(vm_offset_t va, vm_paddr_t pa) 906{ 907 pt_entry_t *pte; 908 909 pte = vtopte(va); 910 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 911} 912 913/* 914 * Remove a page from the kernel pagetables. 915 * Note: not SMP coherent. 916 */ 917PMAP_INLINE void 918pmap_kremove(vm_offset_t va) 919{ 920 pt_entry_t *pte; 921 922 pte = vtopte(va); 923 pte_clear(pte); 924} 925 926/* 927 * Used to map a range of physical addresses into kernel 928 * virtual address space. 929 * 930 * The value passed in '*virt' is a suggested virtual address for 931 * the mapping. Architectures which can support a direct-mapped 932 * physical to virtual region can return the appropriate address 933 * within that region, leaving '*virt' unchanged. Other 934 * architectures should map the pages starting at '*virt' and 935 * update '*virt' with the first usable address after the mapped 936 * region. 937 */ 938vm_offset_t 939pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 940{ 941 vm_offset_t va, sva; 942 943 va = sva = *virt; 944 while (start < end) { 945 pmap_kenter(va, start); 946 va += PAGE_SIZE; 947 start += PAGE_SIZE; 948 } 949 pmap_invalidate_range(kernel_pmap, sva, va); 950 *virt = va; 951 return (sva); 952} 953 954 955/* 956 * Add a list of wired pages to the kva 957 * this routine is only used for temporary 958 * kernel mappings that do not need to have 959 * page modification or references recorded. 960 * Note that old mappings are simply written 961 * over. The page *must* be wired. 962 * Note: SMP coherent. Uses a ranged shootdown IPI. 963 */ 964void 965pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 966{ 967 vm_offset_t va; 968 969 va = sva; 970 while (count-- > 0) { 971 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 972 va += PAGE_SIZE; 973 m++; 974 } 975 pmap_invalidate_range(kernel_pmap, sva, va); 976} 977 978/* 979 * This routine tears out page mappings from the 980 * kernel -- it is meant only for temporary mappings. 981 * Note: SMP coherent. Uses a ranged shootdown IPI. 982 */ 983void 984pmap_qremove(vm_offset_t sva, int count) 985{ 986 vm_offset_t va; 987 988 va = sva; 989 while (count-- > 0) { 990 pmap_kremove(va); 991 va += PAGE_SIZE; 992 } 993 pmap_invalidate_range(kernel_pmap, sva, va); 994} 995 996/*************************************************** 997 * Page table page management routines..... 998 ***************************************************/ 999 1000/* 1001 * This routine unholds page table pages, and if the hold count 1002 * drops to zero, then it decrements the wire count. 1003 */ 1004static PMAP_INLINE int 1005pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1006{ 1007 1008 vm_page_unhold(m); 1009 if (m->hold_count == 0) 1010 return _pmap_unwire_pte_hold(pmap, m); 1011 else 1012 return 0; 1013} 1014 1015static int 1016_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1017{ 1018 vm_offset_t pteva; 1019 1020 /* 1021 * unmap the page table page 1022 */ 1023 pmap->pm_pdir[m->pindex] = 0; 1024 --pmap->pm_stats.resident_count; 1025 /* 1026 * We never unwire a kernel page table page, making a 1027 * check for the kernel_pmap unnecessary. 1028 */ 1029 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)) { 1030 /* 1031 * Do an invltlb to make the invalidated mapping 1032 * take effect immediately. 1033 */ 1034 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1035 pmap_invalidate_page(pmap, pteva); 1036 } 1037 1038 /* 1039 * If the page is finally unwired, simply free it. 1040 */ 1041 --m->wire_count; 1042 if (m->wire_count == 0) { 1043 vm_page_free_zero(m); 1044 atomic_subtract_int(&cnt.v_wire_count, 1); 1045 } 1046 return 1; 1047} 1048 1049/* 1050 * After removing a page table entry, this routine is used to 1051 * conditionally free the page, and manage the hold/wire counts. 1052 */ 1053static int 1054pmap_unuse_pt(pmap_t pmap, vm_offset_t va) 1055{ 1056 pd_entry_t ptepde; 1057 vm_page_t mpte; 1058 1059 if (va >= VM_MAXUSER_ADDRESS) 1060 return 0; 1061 ptepde = *pmap_pde(pmap, va); 1062 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); 1063 return pmap_unwire_pte_hold(pmap, mpte); 1064} 1065 1066void 1067pmap_pinit0(pmap) 1068 struct pmap *pmap; 1069{ 1070 1071 PMAP_LOCK_INIT(pmap); 1072 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1073#ifdef PAE 1074 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1075#endif 1076 pmap->pm_active = 0; 1077 PCPU_SET(curpmap, pmap); 1078 TAILQ_INIT(&pmap->pm_pvlist); 1079 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1080 mtx_lock_spin(&allpmaps_lock); 1081 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1082 mtx_unlock_spin(&allpmaps_lock); 1083} 1084 1085/* 1086 * Initialize a preallocated and zeroed pmap structure, 1087 * such as one in a vmspace structure. 1088 */ 1089void 1090pmap_pinit(pmap) 1091 register struct pmap *pmap; 1092{ 1093 vm_page_t m, ptdpg[NPGPTD]; 1094 vm_paddr_t pa; 1095 static int color; 1096 int i; 1097 1098 PMAP_LOCK_INIT(pmap); 1099 1100 /* 1101 * No need to allocate page table space yet but we do need a valid 1102 * page directory table. 1103 */ 1104 if (pmap->pm_pdir == NULL) { 1105 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1106 NBPTD); 1107#ifdef PAE 1108 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1109 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1110 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1111 ("pmap_pinit: pdpt misaligned")); 1112 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1113 ("pmap_pinit: pdpt above 4g")); 1114#endif 1115 } 1116 1117 /* 1118 * allocate the page directory page(s) 1119 */ 1120 for (i = 0; i < NPGPTD;) { 1121 m = vm_page_alloc(NULL, color++, 1122 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1123 VM_ALLOC_ZERO); 1124 if (m == NULL) 1125 VM_WAIT; 1126 else { 1127 ptdpg[i++] = m; 1128 } 1129 } 1130 1131 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1132 1133 for (i = 0; i < NPGPTD; i++) { 1134 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1135 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1136 } 1137 1138 mtx_lock_spin(&allpmaps_lock); 1139 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1140 mtx_unlock_spin(&allpmaps_lock); 1141 /* Wire in kernel global address entries. */ 1142 /* XXX copies current process, does not fill in MPPTDI */ 1143 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1144#ifdef SMP 1145 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1146#endif 1147 1148 /* install self-referential address mapping entry(s) */ 1149 for (i = 0; i < NPGPTD; i++) { 1150 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1151 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1152#ifdef PAE 1153 pmap->pm_pdpt[i] = pa | PG_V; 1154#endif 1155 } 1156 1157 pmap->pm_active = 0; 1158 TAILQ_INIT(&pmap->pm_pvlist); 1159 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1160} 1161 1162/* 1163 * this routine is called if the page table page is not 1164 * mapped correctly. 1165 */ 1166static vm_page_t 1167_pmap_allocpte(pmap, ptepindex) 1168 pmap_t pmap; 1169 unsigned ptepindex; 1170{ 1171 vm_paddr_t ptepa; 1172 vm_page_t m; 1173 1174 /* 1175 * Allocate a page table page. 1176 */ 1177 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1178 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1179 PMAP_UNLOCK(pmap); 1180 vm_page_unlock_queues(); 1181 VM_WAIT; 1182 vm_page_lock_queues(); 1183 PMAP_LOCK(pmap); 1184 1185 /* 1186 * Indicate the need to retry. While waiting, the page table 1187 * page may have been allocated. 1188 */ 1189 return (NULL); 1190 } 1191 if ((m->flags & PG_ZERO) == 0) 1192 pmap_zero_page(m); 1193 1194 KASSERT(m->queue == PQ_NONE, 1195 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1196 1197 /* 1198 * Increment the hold count for the page table page 1199 * (denoting a new mapping.) 1200 */ 1201 m->hold_count++; 1202 1203 /* 1204 * Map the pagetable page into the process address space, if 1205 * it isn't already there. 1206 */ 1207 1208 pmap->pm_stats.resident_count++; 1209 1210 ptepa = VM_PAGE_TO_PHYS(m); 1211 pmap->pm_pdir[ptepindex] = 1212 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1213 1214 return m; 1215} 1216 1217static vm_page_t 1218pmap_allocpte(pmap_t pmap, vm_offset_t va) 1219{ 1220 unsigned ptepindex; 1221 pd_entry_t ptepa; 1222 vm_page_t m; 1223 1224 /* 1225 * Calculate pagetable page index 1226 */ 1227 ptepindex = va >> PDRSHIFT; 1228retry: 1229 /* 1230 * Get the page directory entry 1231 */ 1232 ptepa = pmap->pm_pdir[ptepindex]; 1233 1234 /* 1235 * This supports switching from a 4MB page to a 1236 * normal 4K page. 1237 */ 1238 if (ptepa & PG_PS) { 1239 pmap->pm_pdir[ptepindex] = 0; 1240 ptepa = 0; 1241 pmap_invalidate_all(kernel_pmap); 1242 } 1243 1244 /* 1245 * If the page table page is mapped, we just increment the 1246 * hold count, and activate it. 1247 */ 1248 if (ptepa) { 1249 m = PHYS_TO_VM_PAGE(ptepa); 1250 m->hold_count++; 1251 } else { 1252 /* 1253 * Here if the pte page isn't mapped, or if it has 1254 * been deallocated. 1255 */ 1256 m = _pmap_allocpte(pmap, ptepindex); 1257 if (m == NULL) 1258 goto retry; 1259 } 1260 return (m); 1261} 1262 1263 1264/*************************************************** 1265* Pmap allocation/deallocation routines. 1266 ***************************************************/ 1267 1268#ifdef SMP 1269/* 1270 * Deal with a SMP shootdown of other users of the pmap that we are 1271 * trying to dispose of. This can be a bit hairy. 1272 */ 1273static u_int *lazymask; 1274static u_int lazyptd; 1275static volatile u_int lazywait; 1276 1277void pmap_lazyfix_action(void); 1278 1279void 1280pmap_lazyfix_action(void) 1281{ 1282 u_int mymask = PCPU_GET(cpumask); 1283 1284 if (rcr3() == lazyptd) 1285 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1286 atomic_clear_int(lazymask, mymask); 1287 atomic_store_rel_int(&lazywait, 1); 1288} 1289 1290static void 1291pmap_lazyfix_self(u_int mymask) 1292{ 1293 1294 if (rcr3() == lazyptd) 1295 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1296 atomic_clear_int(lazymask, mymask); 1297} 1298 1299 1300static void 1301pmap_lazyfix(pmap_t pmap) 1302{ 1303 u_int mymask = PCPU_GET(cpumask); 1304 u_int mask; 1305 register u_int spins; 1306 1307 while ((mask = pmap->pm_active) != 0) { 1308 spins = 50000000; 1309 mask = mask & -mask; /* Find least significant set bit */ 1310 mtx_lock_spin(&lazypmap_lock); 1311#ifdef PAE 1312 lazyptd = vtophys(pmap->pm_pdpt); 1313#else 1314 lazyptd = vtophys(pmap->pm_pdir); 1315#endif 1316 if (mask == mymask) { 1317 lazymask = &pmap->pm_active; 1318 pmap_lazyfix_self(mymask); 1319 } else { 1320 atomic_store_rel_int((u_int *)&lazymask, 1321 (u_int)&pmap->pm_active); 1322 atomic_store_rel_int(&lazywait, 0); 1323 ipi_selected(mask, IPI_LAZYPMAP); 1324 while (lazywait == 0) { 1325 ia32_pause(); 1326 if (--spins == 0) 1327 break; 1328 } 1329 } 1330 mtx_unlock_spin(&lazypmap_lock); 1331 if (spins == 0) 1332 printf("pmap_lazyfix: spun for 50000000\n"); 1333 } 1334} 1335 1336#else /* SMP */ 1337 1338/* 1339 * Cleaning up on uniprocessor is easy. For various reasons, we're 1340 * unlikely to have to even execute this code, including the fact 1341 * that the cleanup is deferred until the parent does a wait(2), which 1342 * means that another userland process has run. 1343 */ 1344static void 1345pmap_lazyfix(pmap_t pmap) 1346{ 1347 u_int cr3; 1348 1349 cr3 = vtophys(pmap->pm_pdir); 1350 if (cr3 == rcr3()) { 1351 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1352 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1353 } 1354} 1355#endif /* SMP */ 1356 1357/* 1358 * Release any resources held by the given physical map. 1359 * Called when a pmap initialized by pmap_pinit is being released. 1360 * Should only be called if the map contains no valid mappings. 1361 */ 1362void 1363pmap_release(pmap_t pmap) 1364{ 1365 vm_page_t m, ptdpg[NPGPTD]; 1366 int i; 1367 1368 KASSERT(pmap->pm_stats.resident_count == 0, 1369 ("pmap_release: pmap resident count %ld != 0", 1370 pmap->pm_stats.resident_count)); 1371 1372 pmap_lazyfix(pmap); 1373 mtx_lock_spin(&allpmaps_lock); 1374 LIST_REMOVE(pmap, pm_list); 1375 mtx_unlock_spin(&allpmaps_lock); 1376 1377 for (i = 0; i < NPGPTD; i++) 1378 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1379 1380 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1381 sizeof(*pmap->pm_pdir)); 1382#ifdef SMP 1383 pmap->pm_pdir[MPPTDI] = 0; 1384#endif 1385 1386 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1387 1388 vm_page_lock_queues(); 1389 for (i = 0; i < NPGPTD; i++) { 1390 m = ptdpg[i]; 1391#ifdef PAE 1392 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1393 ("pmap_release: got wrong ptd page")); 1394#endif 1395 m->wire_count--; 1396 atomic_subtract_int(&cnt.v_wire_count, 1); 1397 vm_page_free_zero(m); 1398 } 1399 vm_page_unlock_queues(); 1400 PMAP_LOCK_DESTROY(pmap); 1401} 1402 1403static int 1404kvm_size(SYSCTL_HANDLER_ARGS) 1405{ 1406 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1407 1408 return sysctl_handle_long(oidp, &ksize, 0, req); 1409} 1410SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1411 0, 0, kvm_size, "IU", "Size of KVM"); 1412 1413static int 1414kvm_free(SYSCTL_HANDLER_ARGS) 1415{ 1416 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1417 1418 return sysctl_handle_long(oidp, &kfree, 0, req); 1419} 1420SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1421 0, 0, kvm_free, "IU", "Amount of KVM free"); 1422 1423/* 1424 * grow the number of kernel page table entries, if needed 1425 */ 1426void 1427pmap_growkernel(vm_offset_t addr) 1428{ 1429 struct pmap *pmap; 1430 vm_paddr_t ptppaddr; 1431 vm_page_t nkpg; 1432 pd_entry_t newpdir; 1433 pt_entry_t *pde; 1434 1435 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1436 if (kernel_vm_end == 0) { 1437 kernel_vm_end = KERNBASE; 1438 nkpt = 0; 1439 while (pdir_pde(PTD, kernel_vm_end)) { 1440 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1441 nkpt++; 1442 } 1443 } 1444 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1445 while (kernel_vm_end < addr) { 1446 if (pdir_pde(PTD, kernel_vm_end)) { 1447 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1448 continue; 1449 } 1450 1451 /* 1452 * This index is bogus, but out of the way 1453 */ 1454 nkpg = vm_page_alloc(NULL, nkpt, 1455 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1456 if (!nkpg) 1457 panic("pmap_growkernel: no memory to grow kernel"); 1458 1459 nkpt++; 1460 1461 pmap_zero_page(nkpg); 1462 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1463 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1464 pdir_pde(PTD, kernel_vm_end) = newpdir; 1465 1466 mtx_lock_spin(&allpmaps_lock); 1467 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1468 pde = pmap_pde(pmap, kernel_vm_end); 1469 pde_store(pde, newpdir); 1470 } 1471 mtx_unlock_spin(&allpmaps_lock); 1472 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1473 } 1474} 1475 1476 1477/*************************************************** 1478 * page management routines. 1479 ***************************************************/ 1480 1481/* 1482 * free the pv_entry back to the free list 1483 */ 1484static PMAP_INLINE void 1485free_pv_entry(pv_entry_t pv) 1486{ 1487 pv_entry_count--; 1488 uma_zfree(pvzone, pv); 1489} 1490 1491/* 1492 * get a new pv_entry, allocating a block from the system 1493 * when needed. 1494 * the memory allocation is performed bypassing the malloc code 1495 * because of the possibility of allocations at interrupt time. 1496 */ 1497static pv_entry_t 1498get_pv_entry(void) 1499{ 1500 pv_entry_count++; 1501 if (pv_entry_high_water && 1502 (pv_entry_count > pv_entry_high_water) && 1503 (pmap_pagedaemon_waken == 0)) { 1504 pmap_pagedaemon_waken = 1; 1505 wakeup (&vm_pages_needed); 1506 } 1507 return uma_zalloc(pvzone, M_NOWAIT); 1508} 1509 1510 1511static int 1512pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1513{ 1514 pv_entry_t pv; 1515 int rtval; 1516 1517 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1518 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1519 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1520 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1521 if (pmap == pv->pv_pmap && va == pv->pv_va) 1522 break; 1523 } 1524 } else { 1525 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1526 if (va == pv->pv_va) 1527 break; 1528 } 1529 } 1530 1531 rtval = 0; 1532 if (pv) { 1533 rtval = pmap_unuse_pt(pmap, va); 1534 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1535 m->md.pv_list_count--; 1536 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1537 vm_page_flag_clear(m, PG_WRITEABLE); 1538 1539 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1540 free_pv_entry(pv); 1541 } 1542 1543 return rtval; 1544} 1545 1546/* 1547 * Create a pv entry for page at pa for 1548 * (pmap, va). 1549 */ 1550static void 1551pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) 1552{ 1553 pv_entry_t pv; 1554 1555 pv = get_pv_entry(); 1556 pv->pv_va = va; 1557 pv->pv_pmap = pmap; 1558 1559 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1560 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1561 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1562 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1563 m->md.pv_list_count++; 1564} 1565 1566/* 1567 * pmap_remove_pte: do the things to unmap a page in a process 1568 */ 1569static int 1570pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1571{ 1572 pt_entry_t oldpte; 1573 vm_page_t m; 1574 1575 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1576 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1577 oldpte = pte_load_clear(ptq); 1578 if (oldpte & PG_W) 1579 pmap->pm_stats.wired_count -= 1; 1580 /* 1581 * Machines that don't support invlpg, also don't support 1582 * PG_G. 1583 */ 1584 if (oldpte & PG_G) 1585 pmap_invalidate_page(kernel_pmap, va); 1586 pmap->pm_stats.resident_count -= 1; 1587 if (oldpte & PG_MANAGED) { 1588 m = PHYS_TO_VM_PAGE(oldpte); 1589 if (oldpte & PG_M) { 1590#if defined(PMAP_DIAGNOSTIC) 1591 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1592 printf( 1593 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1594 va, oldpte); 1595 } 1596#endif 1597 if (pmap_track_modified(va)) 1598 vm_page_dirty(m); 1599 } 1600 if (oldpte & PG_A) 1601 vm_page_flag_set(m, PG_REFERENCED); 1602 return pmap_remove_entry(pmap, m, va); 1603 } else { 1604 return pmap_unuse_pt(pmap, va); 1605 } 1606} 1607 1608/* 1609 * Remove a single page from a process address space 1610 */ 1611static void 1612pmap_remove_page(pmap_t pmap, vm_offset_t va) 1613{ 1614 pt_entry_t *pte; 1615 1616 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1617 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 1618 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1619 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) 1620 return; 1621 pmap_remove_pte(pmap, pte, va); 1622 pmap_invalidate_page(pmap, va); 1623} 1624 1625/* 1626 * Remove the given range of addresses from the specified map. 1627 * 1628 * It is assumed that the start and end are properly 1629 * rounded to the page size. 1630 */ 1631void 1632pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1633{ 1634 vm_offset_t pdnxt; 1635 pd_entry_t ptpaddr; 1636 pt_entry_t *pte; 1637 int anyvalid; 1638 1639 if (pmap == NULL) 1640 return; 1641 1642 /* 1643 * Perform an unsynchronized read. This is, however, safe. 1644 */ 1645 if (pmap->pm_stats.resident_count == 0) 1646 return; 1647 1648 vm_page_lock_queues(); 1649 sched_pin(); 1650 PMAP_LOCK(pmap); 1651 1652 /* 1653 * special handling of removing one page. a very 1654 * common operation and easy to short circuit some 1655 * code. 1656 */ 1657 if ((sva + PAGE_SIZE == eva) && 1658 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1659 pmap_remove_page(pmap, sva); 1660 goto out; 1661 } 1662 1663 anyvalid = 0; 1664 1665 for (; sva < eva; sva = pdnxt) { 1666 unsigned pdirindex; 1667 1668 /* 1669 * Calculate index for next page table. 1670 */ 1671 pdnxt = (sva + NBPDR) & ~PDRMASK; 1672 if (pmap->pm_stats.resident_count == 0) 1673 break; 1674 1675 pdirindex = sva >> PDRSHIFT; 1676 ptpaddr = pmap->pm_pdir[pdirindex]; 1677 1678 /* 1679 * Weed out invalid mappings. Note: we assume that the page 1680 * directory table is always allocated, and in kernel virtual. 1681 */ 1682 if (ptpaddr == 0) 1683 continue; 1684 1685 /* 1686 * Check for large page. 1687 */ 1688 if ((ptpaddr & PG_PS) != 0) { 1689 pmap->pm_pdir[pdirindex] = 0; 1690 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1691 anyvalid = 1; 1692 continue; 1693 } 1694 1695 /* 1696 * Limit our scan to either the end of the va represented 1697 * by the current page table page, or to the end of the 1698 * range being removed. 1699 */ 1700 if (pdnxt > eva) 1701 pdnxt = eva; 1702 1703 for (; sva != pdnxt; sva += PAGE_SIZE) { 1704 if ((pte = pmap_pte_quick(pmap, sva)) == NULL || 1705 *pte == 0) 1706 continue; 1707 anyvalid = 1; 1708 if (pmap_remove_pte(pmap, pte, sva)) 1709 break; 1710 } 1711 } 1712 1713 if (anyvalid) 1714 pmap_invalidate_all(pmap); 1715out: 1716 sched_unpin(); 1717 vm_page_unlock_queues(); 1718 PMAP_UNLOCK(pmap); 1719} 1720 1721/* 1722 * Routine: pmap_remove_all 1723 * Function: 1724 * Removes this physical page from 1725 * all physical maps in which it resides. 1726 * Reflects back modify bits to the pager. 1727 * 1728 * Notes: 1729 * Original versions of this routine were very 1730 * inefficient because they iteratively called 1731 * pmap_remove (slow...) 1732 */ 1733 1734void 1735pmap_remove_all(vm_page_t m) 1736{ 1737 register pv_entry_t pv; 1738 pt_entry_t *pte, tpte; 1739 1740#if defined(PMAP_DIAGNOSTIC) 1741 /* 1742 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1743 */ 1744 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1745 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1746 VM_PAGE_TO_PHYS(m)); 1747 } 1748#endif 1749 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1750 sched_pin(); 1751 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1752 PMAP_LOCK(pv->pv_pmap); 1753 pv->pv_pmap->pm_stats.resident_count--; 1754 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1755 tpte = pte_load_clear(pte); 1756 if (tpte & PG_W) 1757 pv->pv_pmap->pm_stats.wired_count--; 1758 if (tpte & PG_A) 1759 vm_page_flag_set(m, PG_REFERENCED); 1760 1761 /* 1762 * Update the vm_page_t clean and reference bits. 1763 */ 1764 if (tpte & PG_M) { 1765#if defined(PMAP_DIAGNOSTIC) 1766 if (pmap_nw_modified((pt_entry_t) tpte)) { 1767 printf( 1768 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1769 pv->pv_va, tpte); 1770 } 1771#endif 1772 if (pmap_track_modified(pv->pv_va)) 1773 vm_page_dirty(m); 1774 } 1775 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1776 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1777 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1778 m->md.pv_list_count--; 1779 pmap_unuse_pt(pv->pv_pmap, pv->pv_va); 1780 PMAP_UNLOCK(pv->pv_pmap); 1781 free_pv_entry(pv); 1782 } 1783 vm_page_flag_clear(m, PG_WRITEABLE); 1784 sched_unpin(); 1785} 1786 1787/* 1788 * Set the physical protection on the 1789 * specified range of this map as requested. 1790 */ 1791void 1792pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1793{ 1794 vm_offset_t pdnxt; 1795 pd_entry_t ptpaddr; 1796 int anychanged; 1797 1798 if (pmap == NULL) 1799 return; 1800 1801 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1802 pmap_remove(pmap, sva, eva); 1803 return; 1804 } 1805 1806 if (prot & VM_PROT_WRITE) 1807 return; 1808 1809 anychanged = 0; 1810 1811 vm_page_lock_queues(); 1812 sched_pin(); 1813 PMAP_LOCK(pmap); 1814 for (; sva < eva; sva = pdnxt) { 1815 unsigned pdirindex; 1816 1817 pdnxt = (sva + NBPDR) & ~PDRMASK; 1818 1819 pdirindex = sva >> PDRSHIFT; 1820 ptpaddr = pmap->pm_pdir[pdirindex]; 1821 1822 /* 1823 * Weed out invalid mappings. Note: we assume that the page 1824 * directory table is always allocated, and in kernel virtual. 1825 */ 1826 if (ptpaddr == 0) 1827 continue; 1828 1829 /* 1830 * Check for large page. 1831 */ 1832 if ((ptpaddr & PG_PS) != 0) { 1833 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1834 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1835 anychanged = 1; 1836 continue; 1837 } 1838 1839 if (pdnxt > eva) 1840 pdnxt = eva; 1841 1842 for (; sva != pdnxt; sva += PAGE_SIZE) { 1843 pt_entry_t pbits; 1844 pt_entry_t *pte; 1845 vm_page_t m; 1846 1847 if ((pte = pmap_pte_quick(pmap, sva)) == NULL) 1848 continue; 1849 pbits = *pte; 1850 if (pbits & PG_MANAGED) { 1851 m = NULL; 1852 if (pbits & PG_A) { 1853 m = PHYS_TO_VM_PAGE(pbits); 1854 vm_page_flag_set(m, PG_REFERENCED); 1855 pbits &= ~PG_A; 1856 } 1857 if ((pbits & PG_M) != 0 && 1858 pmap_track_modified(sva)) { 1859 if (m == NULL) 1860 m = PHYS_TO_VM_PAGE(pbits); 1861 vm_page_dirty(m); 1862 pbits &= ~PG_M; 1863 } 1864 } 1865 1866 pbits &= ~PG_RW; 1867 1868 if (pbits != *pte) { 1869 pte_store(pte, pbits); 1870 anychanged = 1; 1871 } 1872 } 1873 } 1874 if (anychanged) 1875 pmap_invalidate_all(pmap); 1876 sched_unpin(); 1877 vm_page_unlock_queues(); 1878 PMAP_UNLOCK(pmap); 1879} 1880 1881/* 1882 * Insert the given physical page (p) at 1883 * the specified virtual address (v) in the 1884 * target physical map with the protection requested. 1885 * 1886 * If specified, the page will be wired down, meaning 1887 * that the related pte can not be reclaimed. 1888 * 1889 * NB: This is the only routine which MAY NOT lazy-evaluate 1890 * or lose information. That is, this routine must actually 1891 * insert this page into the given map NOW. 1892 */ 1893void 1894pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1895 boolean_t wired) 1896{ 1897 vm_paddr_t pa; 1898 register pt_entry_t *pte; 1899 vm_paddr_t opa; 1900 pt_entry_t origpte, newpte; 1901 vm_page_t mpte; 1902 1903 if (pmap == NULL) 1904 return; 1905 1906 va &= PG_FRAME; 1907#ifdef PMAP_DIAGNOSTIC 1908 if (va > VM_MAX_KERNEL_ADDRESS) 1909 panic("pmap_enter: toobig"); 1910 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1911 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1912#endif 1913 1914 mpte = NULL; 1915 1916 vm_page_lock_queues(); 1917 PMAP_LOCK(pmap); 1918 sched_pin(); 1919 1920 /* 1921 * In the case that a page table page is not 1922 * resident, we are creating it here. 1923 */ 1924 if (va < VM_MAXUSER_ADDRESS) { 1925 mpte = pmap_allocpte(pmap, va); 1926 } 1927#if 0 && defined(PMAP_DIAGNOSTIC) 1928 else { 1929 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1930 origpte = *pdeaddr; 1931 if ((origpte & PG_V) == 0) { 1932 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1933 pmap->pm_pdir[PTDPTDI], origpte, va); 1934 } 1935 } 1936#endif 1937 1938 pte = pmap_pte_quick(pmap, va); 1939 1940 /* 1941 * Page Directory table entry not valid, we need a new PT page 1942 */ 1943 if (pte == NULL) { 1944 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1945 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1946 } 1947 1948 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 1949 origpte = *pte; 1950 opa = origpte & PG_FRAME; 1951 1952 if (origpte & PG_PS) { 1953 /* 1954 * Yes, I know this will truncate upper address bits for PAE, 1955 * but I'm actually more interested in the lower bits 1956 */ 1957 printf("pmap_enter: va %p, pte %p, origpte %p\n", 1958 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 1959 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1960 } 1961 1962 /* 1963 * Mapping has not changed, must be protection or wiring change. 1964 */ 1965 if (origpte && (opa == pa)) { 1966 /* 1967 * Wiring change, just update stats. We don't worry about 1968 * wiring PT pages as they remain resident as long as there 1969 * are valid mappings in them. Hence, if a user page is wired, 1970 * the PT page will be also. 1971 */ 1972 if (wired && ((origpte & PG_W) == 0)) 1973 pmap->pm_stats.wired_count++; 1974 else if (!wired && (origpte & PG_W)) 1975 pmap->pm_stats.wired_count--; 1976 1977#if defined(PMAP_DIAGNOSTIC) 1978 if (pmap_nw_modified((pt_entry_t) origpte)) { 1979 printf( 1980 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 1981 va, origpte); 1982 } 1983#endif 1984 1985 /* 1986 * Remove extra pte reference 1987 */ 1988 if (mpte) 1989 mpte->hold_count--; 1990 1991 /* 1992 * We might be turning off write access to the page, 1993 * so we go ahead and sense modify status. 1994 */ 1995 if (origpte & PG_MANAGED) { 1996 if ((origpte & PG_M) && pmap_track_modified(va)) { 1997 vm_page_t om; 1998 om = PHYS_TO_VM_PAGE(opa); 1999 vm_page_dirty(om); 2000 } 2001 pa |= PG_MANAGED; 2002 } 2003 goto validate; 2004 } 2005 /* 2006 * Mapping has changed, invalidate old range and fall through to 2007 * handle validating new mapping. 2008 */ 2009 if (opa) { 2010 int err; 2011 err = pmap_remove_pte(pmap, pte, va); 2012 if (err) 2013 panic("pmap_enter: pte vanished, va: 0x%x", va); 2014 } 2015 2016 /* 2017 * Enter on the PV list if part of our managed memory. Note that we 2018 * raise IPL while manipulating pv_table since pmap_enter can be 2019 * called at interrupt time. 2020 */ 2021 if (pmap_initialized && 2022 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2023 pmap_insert_entry(pmap, va, m); 2024 pa |= PG_MANAGED; 2025 } 2026 2027 /* 2028 * Increment counters 2029 */ 2030 pmap->pm_stats.resident_count++; 2031 if (wired) 2032 pmap->pm_stats.wired_count++; 2033 2034validate: 2035 /* 2036 * Now validate mapping with desired protection/wiring. 2037 */ 2038 newpte = (pt_entry_t)(pa | PG_V); 2039 if ((prot & VM_PROT_WRITE) != 0) 2040 newpte |= PG_RW; 2041 if (wired) 2042 newpte |= PG_W; 2043 if (va < VM_MAXUSER_ADDRESS) 2044 newpte |= PG_U; 2045 if (pmap == kernel_pmap) 2046 newpte |= pgeflag; 2047 2048 /* 2049 * if the mapping or permission bits are different, we need 2050 * to update the pte. 2051 */ 2052 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2053 pte_store(pte, newpte | PG_A); 2054 /*if (origpte)*/ { 2055 pmap_invalidate_page(pmap, va); 2056 } 2057 } 2058 sched_unpin(); 2059 vm_page_unlock_queues(); 2060 PMAP_UNLOCK(pmap); 2061} 2062 2063/* 2064 * this code makes some *MAJOR* assumptions: 2065 * 1. Current pmap & pmap exists. 2066 * 2. Not wired. 2067 * 3. Read access. 2068 * 4. No page table pages. 2069 * 5. Tlbflush is deferred to calling procedure. 2070 * 6. Page IS managed. 2071 * but is *MUCH* faster than pmap_enter... 2072 */ 2073 2074vm_page_t 2075pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2076{ 2077 pt_entry_t *pte; 2078 vm_paddr_t pa; 2079 2080 vm_page_lock_queues(); 2081 PMAP_LOCK(pmap); 2082 2083 /* 2084 * In the case that a page table page is not 2085 * resident, we are creating it here. 2086 */ 2087 if (va < VM_MAXUSER_ADDRESS) { 2088 unsigned ptepindex; 2089 pd_entry_t ptepa; 2090 2091 /* 2092 * Calculate pagetable page index 2093 */ 2094 ptepindex = va >> PDRSHIFT; 2095 if (mpte && (mpte->pindex == ptepindex)) { 2096 mpte->hold_count++; 2097 } else { 2098retry: 2099 /* 2100 * Get the page directory entry 2101 */ 2102 ptepa = pmap->pm_pdir[ptepindex]; 2103 2104 /* 2105 * If the page table page is mapped, we just increment 2106 * the hold count, and activate it. 2107 */ 2108 if (ptepa) { 2109 if (ptepa & PG_PS) 2110 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2111 mpte = PHYS_TO_VM_PAGE(ptepa); 2112 mpte->hold_count++; 2113 } else { 2114 mpte = _pmap_allocpte(pmap, ptepindex); 2115 if (mpte == NULL) 2116 goto retry; 2117 } 2118 } 2119 } else { 2120 mpte = NULL; 2121 } 2122 2123 /* 2124 * This call to vtopte makes the assumption that we are 2125 * entering the page into the current pmap. In order to support 2126 * quick entry into any pmap, one would likely use pmap_pte_quick. 2127 * But that isn't as quick as vtopte. 2128 */ 2129 pte = vtopte(va); 2130 if (*pte) { 2131 if (mpte != NULL) { 2132 pmap_unwire_pte_hold(pmap, mpte); 2133 mpte = NULL; 2134 } 2135 goto out; 2136 } 2137 2138 /* 2139 * Enter on the PV list if part of our managed memory. Note that we 2140 * raise IPL while manipulating pv_table since pmap_enter can be 2141 * called at interrupt time. 2142 */ 2143 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2144 pmap_insert_entry(pmap, va, m); 2145 2146 /* 2147 * Increment counters 2148 */ 2149 pmap->pm_stats.resident_count++; 2150 2151 pa = VM_PAGE_TO_PHYS(m); 2152 2153 /* 2154 * Now validate mapping with RO protection 2155 */ 2156 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2157 pte_store(pte, pa | PG_V | PG_U); 2158 else 2159 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2160out: 2161 vm_page_unlock_queues(); 2162 PMAP_UNLOCK(pmap); 2163 return mpte; 2164} 2165 2166/* 2167 * Make a temporary mapping for a physical address. This is only intended 2168 * to be used for panic dumps. 2169 */ 2170void * 2171pmap_kenter_temporary(vm_paddr_t pa, int i) 2172{ 2173 vm_offset_t va; 2174 2175 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2176 pmap_kenter(va, pa); 2177#ifndef I386_CPU 2178 invlpg(va); 2179#else 2180 invltlb(); 2181#endif 2182 return ((void *)crashdumpmap); 2183} 2184 2185/* 2186 * This code maps large physical mmap regions into the 2187 * processor address space. Note that some shortcuts 2188 * are taken, but the code works. 2189 */ 2190void 2191pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2192 vm_object_t object, vm_pindex_t pindex, 2193 vm_size_t size) 2194{ 2195 vm_page_t p; 2196 2197 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2198 KASSERT(object->type == OBJT_DEVICE, 2199 ("pmap_object_init_pt: non-device object")); 2200 if (pseflag && 2201 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2202 int i; 2203 vm_page_t m[1]; 2204 unsigned int ptepindex; 2205 int npdes; 2206 pd_entry_t ptepa; 2207 2208 PMAP_LOCK(pmap); 2209 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2210 goto out; 2211 PMAP_UNLOCK(pmap); 2212retry: 2213 p = vm_page_lookup(object, pindex); 2214 if (p != NULL) { 2215 vm_page_lock_queues(); 2216 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2217 goto retry; 2218 } else { 2219 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2220 if (p == NULL) 2221 return; 2222 m[0] = p; 2223 2224 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2225 vm_page_lock_queues(); 2226 vm_page_free(p); 2227 vm_page_unlock_queues(); 2228 return; 2229 } 2230 2231 p = vm_page_lookup(object, pindex); 2232 vm_page_lock_queues(); 2233 vm_page_wakeup(p); 2234 } 2235 vm_page_unlock_queues(); 2236 2237 ptepa = VM_PAGE_TO_PHYS(p); 2238 if (ptepa & (NBPDR - 1)) 2239 return; 2240 2241 p->valid = VM_PAGE_BITS_ALL; 2242 2243 PMAP_LOCK(pmap); 2244 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2245 npdes = size >> PDRSHIFT; 2246 for(i = 0; i < npdes; i++) { 2247 pde_store(&pmap->pm_pdir[ptepindex], 2248 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2249 ptepa += NBPDR; 2250 ptepindex += 1; 2251 } 2252 pmap_invalidate_all(pmap); 2253out: 2254 PMAP_UNLOCK(pmap); 2255 } 2256} 2257 2258/* 2259 * Routine: pmap_change_wiring 2260 * Function: Change the wiring attribute for a map/virtual-address 2261 * pair. 2262 * In/out conditions: 2263 * The mapping must already exist in the pmap. 2264 */ 2265void 2266pmap_change_wiring(pmap, va, wired) 2267 register pmap_t pmap; 2268 vm_offset_t va; 2269 boolean_t wired; 2270{ 2271 register pt_entry_t *pte; 2272 2273 if (pmap == NULL) 2274 return; 2275 2276 PMAP_LOCK(pmap); 2277 pte = pmap_pte(pmap, va); 2278 2279 if (wired && !pmap_pte_w(pte)) 2280 pmap->pm_stats.wired_count++; 2281 else if (!wired && pmap_pte_w(pte)) 2282 pmap->pm_stats.wired_count--; 2283 2284 /* 2285 * Wiring is not a hardware characteristic so there is no need to 2286 * invalidate TLB. 2287 */ 2288 pmap_pte_set_w(pte, wired); 2289 PMAP_UNLOCK(pmap); 2290} 2291 2292 2293 2294/* 2295 * Copy the range specified by src_addr/len 2296 * from the source map to the range dst_addr/len 2297 * in the destination map. 2298 * 2299 * This routine is only advisory and need not do anything. 2300 */ 2301 2302void 2303pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2304 vm_offset_t src_addr) 2305{ 2306 vm_offset_t addr; 2307 vm_offset_t end_addr = src_addr + len; 2308 vm_offset_t pdnxt; 2309 vm_page_t m; 2310 2311 if (dst_addr != src_addr) 2312 return; 2313 2314 if (!pmap_is_current(src_pmap)) 2315 return; 2316 2317 vm_page_lock_queues(); 2318 PMAP_LOCK(dst_pmap); 2319 sched_pin(); 2320 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2321 pt_entry_t *src_pte, *dst_pte; 2322 vm_page_t dstmpte, srcmpte; 2323 pd_entry_t srcptepaddr; 2324 unsigned ptepindex; 2325 2326 if (addr >= UPT_MIN_ADDRESS) 2327 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2328 2329 /* 2330 * Don't let optional prefaulting of pages make us go 2331 * way below the low water mark of free pages or way 2332 * above high water mark of used pv entries. 2333 */ 2334 if (cnt.v_free_count < cnt.v_free_reserved || 2335 pv_entry_count > pv_entry_high_water) 2336 break; 2337 2338 pdnxt = (addr + NBPDR) & ~PDRMASK; 2339 ptepindex = addr >> PDRSHIFT; 2340 2341 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2342 if (srcptepaddr == 0) 2343 continue; 2344 2345 if (srcptepaddr & PG_PS) { 2346 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2347 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2348 dst_pmap->pm_stats.resident_count += 2349 NBPDR / PAGE_SIZE; 2350 } 2351 continue; 2352 } 2353 2354 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2355 if (srcmpte->hold_count == 0 || (srcmpte->flags & PG_BUSY)) 2356 continue; 2357 2358 if (pdnxt > end_addr) 2359 pdnxt = end_addr; 2360 2361 src_pte = vtopte(addr); 2362 while (addr < pdnxt) { 2363 pt_entry_t ptetemp; 2364 ptetemp = *src_pte; 2365 /* 2366 * we only virtual copy managed pages 2367 */ 2368 if ((ptetemp & PG_MANAGED) != 0) { 2369 /* 2370 * We have to check after allocpte for the 2371 * pte still being around... allocpte can 2372 * block. 2373 */ 2374 dstmpte = pmap_allocpte(dst_pmap, addr); 2375 dst_pte = pmap_pte_quick(dst_pmap, addr); 2376 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2377 /* 2378 * Clear the modified and 2379 * accessed (referenced) bits 2380 * during the copy. 2381 */ 2382 m = PHYS_TO_VM_PAGE(ptetemp); 2383 *dst_pte = ptetemp & ~(PG_M | PG_A); 2384 dst_pmap->pm_stats.resident_count++; 2385 pmap_insert_entry(dst_pmap, addr, m); 2386 } else 2387 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2388 if (dstmpte->hold_count >= srcmpte->hold_count) 2389 break; 2390 } 2391 addr += PAGE_SIZE; 2392 src_pte++; 2393 } 2394 } 2395 sched_unpin(); 2396 vm_page_unlock_queues(); 2397 PMAP_UNLOCK(dst_pmap); 2398} 2399 2400static __inline void 2401pagezero(void *page) 2402{ 2403#if defined(I686_CPU) 2404 if (cpu_class == CPUCLASS_686) { 2405#if defined(CPU_ENABLE_SSE) 2406 if (cpu_feature & CPUID_SSE2) 2407 sse2_pagezero(page); 2408 else 2409#endif 2410 i686_pagezero(page); 2411 } else 2412#endif 2413 bzero(page, PAGE_SIZE); 2414} 2415 2416/* 2417 * pmap_zero_page zeros the specified hardware page by mapping 2418 * the page into KVM and using bzero to clear its contents. 2419 */ 2420void 2421pmap_zero_page(vm_page_t m) 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 pagezero(CADDR2); 2431 *CMAP2 = 0; 2432 sched_unpin(); 2433 mtx_unlock(&CMAPCADDR12_lock); 2434} 2435 2436/* 2437 * pmap_zero_page_area zeros the specified hardware page by mapping 2438 * the page into KVM and using bzero to clear its contents. 2439 * 2440 * off and size may not cover an area beyond a single hardware page. 2441 */ 2442void 2443pmap_zero_page_area(vm_page_t m, int off, int size) 2444{ 2445 2446 mtx_lock(&CMAPCADDR12_lock); 2447 if (*CMAP2) 2448 panic("pmap_zero_page: CMAP2 busy"); 2449 sched_pin(); 2450 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2451 invlcaddr(CADDR2); 2452 if (off == 0 && size == PAGE_SIZE) 2453 pagezero(CADDR2); 2454 else 2455 bzero((char *)CADDR2 + off, size); 2456 *CMAP2 = 0; 2457 sched_unpin(); 2458 mtx_unlock(&CMAPCADDR12_lock); 2459} 2460 2461/* 2462 * pmap_zero_page_idle zeros the specified hardware page by mapping 2463 * the page into KVM and using bzero to clear its contents. This 2464 * is intended to be called from the vm_pagezero process only and 2465 * outside of Giant. 2466 */ 2467void 2468pmap_zero_page_idle(vm_page_t m) 2469{ 2470 2471 if (*CMAP3) 2472 panic("pmap_zero_page: CMAP3 busy"); 2473 sched_pin(); 2474 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2475 invlcaddr(CADDR3); 2476 pagezero(CADDR3); 2477 *CMAP3 = 0; 2478 sched_unpin(); 2479} 2480 2481/* 2482 * pmap_copy_page copies the specified (machine independent) 2483 * page by mapping the page into virtual memory and using 2484 * bcopy to copy the page, one machine dependent page at a 2485 * time. 2486 */ 2487void 2488pmap_copy_page(vm_page_t src, vm_page_t dst) 2489{ 2490 2491 mtx_lock(&CMAPCADDR12_lock); 2492 if (*CMAP1) 2493 panic("pmap_copy_page: CMAP1 busy"); 2494 if (*CMAP2) 2495 panic("pmap_copy_page: CMAP2 busy"); 2496 sched_pin(); 2497#ifdef I386_CPU 2498 invltlb(); 2499#else 2500 invlpg((u_int)CADDR1); 2501 invlpg((u_int)CADDR2); 2502#endif 2503 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2504 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2505 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2506 *CMAP1 = 0; 2507 *CMAP2 = 0; 2508 sched_unpin(); 2509 mtx_unlock(&CMAPCADDR12_lock); 2510} 2511 2512/* 2513 * Returns true if the pmap's pv is one of the first 2514 * 16 pvs linked to from this page. This count may 2515 * be changed upwards or downwards in the future; it 2516 * is only necessary that true be returned for a small 2517 * subset of pmaps for proper page aging. 2518 */ 2519boolean_t 2520pmap_page_exists_quick(pmap, m) 2521 pmap_t pmap; 2522 vm_page_t m; 2523{ 2524 pv_entry_t pv; 2525 int loops = 0; 2526 2527 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2528 return FALSE; 2529 2530 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2531 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2532 if (pv->pv_pmap == pmap) { 2533 return TRUE; 2534 } 2535 loops++; 2536 if (loops >= 16) 2537 break; 2538 } 2539 return (FALSE); 2540} 2541 2542#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2543/* 2544 * Remove all pages from specified address space 2545 * this aids process exit speeds. Also, this code 2546 * is special cased for current process only, but 2547 * can have the more generic (and slightly slower) 2548 * mode enabled. This is much faster than pmap_remove 2549 * in the case of running down an entire address space. 2550 */ 2551void 2552pmap_remove_pages(pmap, sva, eva) 2553 pmap_t pmap; 2554 vm_offset_t sva, eva; 2555{ 2556 pt_entry_t *pte, tpte; 2557 vm_page_t m; 2558 pv_entry_t pv, npv; 2559 2560#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2561 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2562 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2563 return; 2564 } 2565#endif 2566 vm_page_lock_queues(); 2567 PMAP_LOCK(pmap); 2568 sched_pin(); 2569 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2570 2571 if (pv->pv_va >= eva || pv->pv_va < sva) { 2572 npv = TAILQ_NEXT(pv, pv_plist); 2573 continue; 2574 } 2575 2576#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2577 pte = vtopte(pv->pv_va); 2578#else 2579 pte = pmap_pte_quick(pmap, pv->pv_va); 2580#endif 2581 tpte = *pte; 2582 2583 if (tpte == 0) { 2584 printf("TPTE at %p IS ZERO @ VA %08x\n", 2585 pte, pv->pv_va); 2586 panic("bad pte"); 2587 } 2588 2589/* 2590 * We cannot remove wired pages from a process' mapping at this time 2591 */ 2592 if (tpte & PG_W) { 2593 npv = TAILQ_NEXT(pv, pv_plist); 2594 continue; 2595 } 2596 2597 m = PHYS_TO_VM_PAGE(tpte); 2598 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2599 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2600 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2601 2602 KASSERT(m < &vm_page_array[vm_page_array_size], 2603 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2604 2605 pmap->pm_stats.resident_count--; 2606 2607 pte_clear(pte); 2608 2609 /* 2610 * Update the vm_page_t clean and reference bits. 2611 */ 2612 if (tpte & PG_M) { 2613 vm_page_dirty(m); 2614 } 2615 2616 npv = TAILQ_NEXT(pv, pv_plist); 2617 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 2618 2619 m->md.pv_list_count--; 2620 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2621 if (TAILQ_EMPTY(&m->md.pv_list)) 2622 vm_page_flag_clear(m, PG_WRITEABLE); 2623 2624 pmap_unuse_pt(pmap, pv->pv_va); 2625 free_pv_entry(pv); 2626 } 2627 sched_unpin(); 2628 pmap_invalidate_all(pmap); 2629 PMAP_UNLOCK(pmap); 2630 vm_page_unlock_queues(); 2631} 2632 2633/* 2634 * pmap_is_modified: 2635 * 2636 * Return whether or not the specified physical page was modified 2637 * in any physical maps. 2638 */ 2639boolean_t 2640pmap_is_modified(vm_page_t m) 2641{ 2642 pv_entry_t pv; 2643 pt_entry_t *pte; 2644 boolean_t rv; 2645 2646 rv = FALSE; 2647 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2648 return (rv); 2649 2650 sched_pin(); 2651 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2652 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2653 /* 2654 * if the bit being tested is the modified bit, then 2655 * mark clean_map and ptes as never 2656 * modified. 2657 */ 2658 if (!pmap_track_modified(pv->pv_va)) 2659 continue; 2660#if defined(PMAP_DIAGNOSTIC) 2661 if (!pv->pv_pmap) { 2662 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 2663 continue; 2664 } 2665#endif 2666 PMAP_LOCK(pv->pv_pmap); 2667 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2668 rv = (*pte & PG_M) != 0; 2669 PMAP_UNLOCK(pv->pv_pmap); 2670 if (rv) 2671 break; 2672 } 2673 sched_unpin(); 2674 return (rv); 2675} 2676 2677/* 2678 * pmap_is_prefaultable: 2679 * 2680 * Return whether or not the specified virtual address is elgible 2681 * for prefault. 2682 */ 2683boolean_t 2684pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2685{ 2686 pt_entry_t *pte; 2687 boolean_t rv; 2688 2689 rv = FALSE; 2690 PMAP_LOCK(pmap); 2691 if (*pmap_pde(pmap, addr)) { 2692 pte = vtopte(addr); 2693 rv = *pte == 0; 2694 } 2695 PMAP_UNLOCK(pmap); 2696 return (rv); 2697} 2698 2699/* 2700 * Clear the given bit in each of the given page's ptes. 2701 */ 2702static __inline void 2703pmap_clear_ptes(vm_page_t m, int bit) 2704{ 2705 register pv_entry_t pv; 2706 pt_entry_t pbits, *pte; 2707 2708 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) || 2709 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2710 return; 2711 2712 sched_pin(); 2713 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2714 /* 2715 * Loop over all current mappings setting/clearing as appropos If 2716 * setting RO do we need to clear the VAC? 2717 */ 2718 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2719 /* 2720 * don't write protect pager mappings 2721 */ 2722 if (bit == PG_RW) { 2723 if (!pmap_track_modified(pv->pv_va)) 2724 continue; 2725 } 2726 2727#if defined(PMAP_DIAGNOSTIC) 2728 if (!pv->pv_pmap) { 2729 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 2730 continue; 2731 } 2732#endif 2733 2734 PMAP_LOCK(pv->pv_pmap); 2735 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2736 pbits = *pte; 2737 if (pbits & bit) { 2738 if (bit == PG_RW) { 2739 if (pbits & PG_M) { 2740 vm_page_dirty(m); 2741 } 2742 pte_store(pte, pbits & ~(PG_M|PG_RW)); 2743 } else { 2744 pte_store(pte, pbits & ~bit); 2745 } 2746 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2747 } 2748 PMAP_UNLOCK(pv->pv_pmap); 2749 } 2750 if (bit == PG_RW) 2751 vm_page_flag_clear(m, PG_WRITEABLE); 2752 sched_unpin(); 2753} 2754 2755/* 2756 * pmap_page_protect: 2757 * 2758 * Lower the permission for all mappings to a given page. 2759 */ 2760void 2761pmap_page_protect(vm_page_t m, vm_prot_t prot) 2762{ 2763 if ((prot & VM_PROT_WRITE) == 0) { 2764 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2765 pmap_clear_ptes(m, PG_RW); 2766 } else { 2767 pmap_remove_all(m); 2768 } 2769 } 2770} 2771 2772/* 2773 * pmap_ts_referenced: 2774 * 2775 * Return a count of reference bits for a page, clearing those bits. 2776 * It is not necessary for every reference bit to be cleared, but it 2777 * is necessary that 0 only be returned when there are truly no 2778 * reference bits set. 2779 * 2780 * XXX: The exact number of bits to check and clear is a matter that 2781 * should be tested and standardized at some point in the future for 2782 * optimal aging of shared pages. 2783 */ 2784int 2785pmap_ts_referenced(vm_page_t m) 2786{ 2787 register pv_entry_t pv, pvf, pvn; 2788 pt_entry_t *pte; 2789 pt_entry_t v; 2790 int rtval = 0; 2791 2792 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2793 return (rtval); 2794 2795 sched_pin(); 2796 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2797 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2798 2799 pvf = pv; 2800 2801 do { 2802 pvn = TAILQ_NEXT(pv, pv_list); 2803 2804 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2805 2806 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2807 2808 if (!pmap_track_modified(pv->pv_va)) 2809 continue; 2810 2811 PMAP_LOCK(pv->pv_pmap); 2812 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2813 2814 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2815 atomic_clear_int((u_int *)pte, PG_A); 2816 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2817 2818 rtval++; 2819 if (rtval > 4) { 2820 PMAP_UNLOCK(pv->pv_pmap); 2821 break; 2822 } 2823 } 2824 PMAP_UNLOCK(pv->pv_pmap); 2825 } while ((pv = pvn) != NULL && pv != pvf); 2826 } 2827 sched_unpin(); 2828 2829 return (rtval); 2830} 2831 2832/* 2833 * Clear the modify bits on the specified physical page. 2834 */ 2835void 2836pmap_clear_modify(vm_page_t m) 2837{ 2838 pmap_clear_ptes(m, PG_M); 2839} 2840 2841/* 2842 * pmap_clear_reference: 2843 * 2844 * Clear the reference bit on the specified physical page. 2845 */ 2846void 2847pmap_clear_reference(vm_page_t m) 2848{ 2849 pmap_clear_ptes(m, PG_A); 2850} 2851 2852/* 2853 * Miscellaneous support routines follow 2854 */ 2855 2856/* 2857 * Map a set of physical memory pages into the kernel virtual 2858 * address space. Return a pointer to where it is mapped. This 2859 * routine is intended to be used for mapping device memory, 2860 * NOT real memory. 2861 */ 2862void * 2863pmap_mapdev(pa, size) 2864 vm_paddr_t pa; 2865 vm_size_t size; 2866{ 2867 vm_offset_t va, tmpva, offset; 2868 2869 offset = pa & PAGE_MASK; 2870 size = roundup(offset + size, PAGE_SIZE); 2871 pa = pa & PG_FRAME; 2872 2873 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2874 va = KERNBASE + pa; 2875 else 2876 va = kmem_alloc_nofault(kernel_map, size); 2877 if (!va) 2878 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2879 2880 for (tmpva = va; size > 0; ) { 2881 pmap_kenter(tmpva, pa); 2882 size -= PAGE_SIZE; 2883 tmpva += PAGE_SIZE; 2884 pa += PAGE_SIZE; 2885 } 2886 pmap_invalidate_range(kernel_pmap, va, tmpva); 2887 return ((void *)(va + offset)); 2888} 2889 2890void 2891pmap_unmapdev(va, size) 2892 vm_offset_t va; 2893 vm_size_t size; 2894{ 2895 vm_offset_t base, offset, tmpva; 2896 2897 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2898 return; 2899 base = va & PG_FRAME; 2900 offset = va & PAGE_MASK; 2901 size = roundup(offset + size, PAGE_SIZE); 2902 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2903 pmap_kremove(tmpva); 2904 pmap_invalidate_range(kernel_pmap, va, tmpva); 2905 kmem_free(kernel_map, base, size); 2906} 2907 2908/* 2909 * perform the pmap work for mincore 2910 */ 2911int 2912pmap_mincore(pmap, addr) 2913 pmap_t pmap; 2914 vm_offset_t addr; 2915{ 2916 pt_entry_t *ptep, pte; 2917 vm_page_t m; 2918 int val = 0; 2919 2920 PMAP_LOCK(pmap); 2921 ptep = pmap_pte(pmap, addr); 2922 pte = (ptep != NULL) ? *ptep : 0; 2923 PMAP_UNLOCK(pmap); 2924 2925 if (pte != 0) { 2926 vm_paddr_t pa; 2927 2928 val = MINCORE_INCORE; 2929 if ((pte & PG_MANAGED) == 0) 2930 return val; 2931 2932 pa = pte & PG_FRAME; 2933 2934 m = PHYS_TO_VM_PAGE(pa); 2935 2936 /* 2937 * Modified by us 2938 */ 2939 if (pte & PG_M) 2940 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2941 else { 2942 /* 2943 * Modified by someone else 2944 */ 2945 vm_page_lock_queues(); 2946 if (m->dirty || pmap_is_modified(m)) 2947 val |= MINCORE_MODIFIED_OTHER; 2948 vm_page_unlock_queues(); 2949 } 2950 /* 2951 * Referenced by us 2952 */ 2953 if (pte & PG_A) 2954 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2955 else { 2956 /* 2957 * Referenced by someone else 2958 */ 2959 vm_page_lock_queues(); 2960 if ((m->flags & PG_REFERENCED) || 2961 pmap_ts_referenced(m)) { 2962 val |= MINCORE_REFERENCED_OTHER; 2963 vm_page_flag_set(m, PG_REFERENCED); 2964 } 2965 vm_page_unlock_queues(); 2966 } 2967 } 2968 return val; 2969} 2970 2971void 2972pmap_activate(struct thread *td) 2973{ 2974 struct proc *p = td->td_proc; 2975 pmap_t pmap, oldpmap; 2976 u_int32_t cr3; 2977 2978 critical_enter(); 2979 pmap = vmspace_pmap(td->td_proc->p_vmspace); 2980 oldpmap = PCPU_GET(curpmap); 2981#if defined(SMP) 2982 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 2983 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 2984#else 2985 oldpmap->pm_active &= ~1; 2986 pmap->pm_active |= 1; 2987#endif 2988#ifdef PAE 2989 cr3 = vtophys(pmap->pm_pdpt); 2990#else 2991 cr3 = vtophys(pmap->pm_pdir); 2992#endif 2993 /* XXXKSE this is wrong. 2994 * pmap_activate is for the current thread on the current cpu 2995 */ 2996 if (p->p_flag & P_SA) { 2997 /* Make sure all other cr3 entries are updated. */ 2998 /* what if they are running? XXXKSE (maybe abort them) */ 2999 FOREACH_THREAD_IN_PROC(p, td) { 3000 td->td_pcb->pcb_cr3 = cr3; 3001 } 3002 } else { 3003 td->td_pcb->pcb_cr3 = cr3; 3004 } 3005 load_cr3(cr3); 3006 PCPU_SET(curpmap, pmap); 3007 critical_exit(); 3008} 3009 3010vm_offset_t 3011pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3012{ 3013 3014 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3015 return addr; 3016 } 3017 3018 addr = (addr + PDRMASK) & ~PDRMASK; 3019 return addr; 3020} 3021 3022 3023#if defined(PMAP_DEBUG) 3024pmap_pid_dump(int pid) 3025{ 3026 pmap_t pmap; 3027 struct proc *p; 3028 int npte = 0; 3029 int index; 3030 3031 sx_slock(&allproc_lock); 3032 LIST_FOREACH(p, &allproc, p_list) { 3033 if (p->p_pid != pid) 3034 continue; 3035 3036 if (p->p_vmspace) { 3037 int i,j; 3038 index = 0; 3039 pmap = vmspace_pmap(p->p_vmspace); 3040 for (i = 0; i < NPDEPTD; i++) { 3041 pd_entry_t *pde; 3042 pt_entry_t *pte; 3043 vm_offset_t base = i << PDRSHIFT; 3044 3045 pde = &pmap->pm_pdir[i]; 3046 if (pde && pmap_pde_v(pde)) { 3047 for (j = 0; j < NPTEPG; j++) { 3048 vm_offset_t va = base + (j << PAGE_SHIFT); 3049 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3050 if (index) { 3051 index = 0; 3052 printf("\n"); 3053 } 3054 sx_sunlock(&allproc_lock); 3055 return npte; 3056 } 3057 pte = pmap_pte(pmap, va); 3058 if (pte && pmap_pte_v(pte)) { 3059 pt_entry_t pa; 3060 vm_page_t m; 3061 pa = *pte; 3062 m = PHYS_TO_VM_PAGE(pa); 3063 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3064 va, pa, m->hold_count, m->wire_count, m->flags); 3065 npte++; 3066 index++; 3067 if (index >= 2) { 3068 index = 0; 3069 printf("\n"); 3070 } else { 3071 printf(" "); 3072 } 3073 } 3074 } 3075 } 3076 } 3077 } 3078 } 3079 sx_sunlock(&allproc_lock); 3080 return npte; 3081} 3082#endif 3083 3084#if defined(DEBUG) 3085 3086static void pads(pmap_t pm); 3087void pmap_pvdump(vm_offset_t pa); 3088 3089/* print address space of pmap*/ 3090static void 3091pads(pm) 3092 pmap_t pm; 3093{ 3094 int i, j; 3095 vm_paddr_t va; 3096 pt_entry_t *ptep; 3097 3098 if (pm == kernel_pmap) 3099 return; 3100 for (i = 0; i < NPDEPTD; i++) 3101 if (pm->pm_pdir[i]) 3102 for (j = 0; j < NPTEPG; j++) { 3103 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3104 if (pm == kernel_pmap && va < KERNBASE) 3105 continue; 3106 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3107 continue; 3108 ptep = pmap_pte(pm, va); 3109 if (pmap_pte_v(ptep)) 3110 printf("%x:%x ", va, *ptep); 3111 }; 3112 3113} 3114 3115void 3116pmap_pvdump(pa) 3117 vm_paddr_t pa; 3118{ 3119 pv_entry_t pv; 3120 vm_page_t m; 3121 3122 printf("pa %x", pa); 3123 m = PHYS_TO_VM_PAGE(pa); 3124 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3125 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3126 pads(pv->pv_pmap); 3127 } 3128 printf(" "); 3129} 3130#endif 3131