pmap.c revision 121996
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 121996 2003-11-03 22:32:04Z jhb $"); 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_pmap.h" 105#include "opt_msgbuf.h" 106#include "opt_kstack_pages.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/mman.h> 113#include <sys/msgbuf.h> 114#include <sys/mutex.h> 115#include <sys/proc.h> 116#include <sys/sx.h> 117#include <sys/user.h> 118#include <sys/vmmeter.h> 119#include <sys/sysctl.h> 120#ifdef SMP 121#include <sys/smp.h> 122#endif 123 124#include <vm/vm.h> 125#include <vm/vm_param.h> 126#include <vm/vm_kern.h> 127#include <vm/vm_page.h> 128#include <vm/vm_map.h> 129#include <vm/vm_object.h> 130#include <vm/vm_extern.h> 131#include <vm/vm_pageout.h> 132#include <vm/vm_pager.h> 133#include <vm/uma.h> 134 135#include <machine/cpu.h> 136#include <machine/cputypes.h> 137#include <machine/md_var.h> 138#include <machine/specialreg.h> 139#ifdef SMP 140#include <machine/smp.h> 141#endif 142 143#define PMAP_KEEP_PDIRS 144#ifndef PMAP_SHPGPERPROC 145#define PMAP_SHPGPERPROC 200 146#endif 147 148#if defined(DIAGNOSTIC) 149#define PMAP_DIAGNOSTIC 150#endif 151 152#define MINPV 2048 153 154#if !defined(PMAP_DIAGNOSTIC) 155#define PMAP_INLINE __inline 156#else 157#define PMAP_INLINE 158#endif 159 160/* 161 * Get PDEs and PTEs for user/kernel address space 162 */ 163#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 164#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 165 166#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 167#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 168#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 169#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 170#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 171 172#define pmap_pte_set_w(pte, v) ((v)?(*(int *)pte |= PG_W):(*(int *)pte &= ~PG_W)) 173#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 174 175/* 176 * Given a map and a machine independent protection code, 177 * convert to a vax protection code. 178 */ 179#define pte_prot(m, p) (protection_codes[p]) 180static int protection_codes[8]; 181 182struct pmap kernel_pmap_store; 183LIST_HEAD(pmaplist, pmap); 184static struct pmaplist allpmaps; 185static struct mtx allpmaps_lock; 186#ifdef SMP 187static struct mtx lazypmap_lock; 188#endif 189 190vm_paddr_t avail_start; /* PA of first available physical page */ 191vm_paddr_t avail_end; /* PA of last available physical page */ 192vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 193vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 194static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 195int pgeflag = 0; /* PG_G or-in */ 196int pseflag = 0; /* PG_PS or-in */ 197 198static int nkpt; 199vm_offset_t kernel_vm_end; 200extern u_int32_t KERNend; 201 202#ifdef PAE 203static uma_zone_t pdptzone; 204#endif 205 206/* 207 * Data for the pv entry allocation mechanism 208 */ 209static uma_zone_t pvzone; 210static struct vm_object pvzone_obj; 211static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 212int pmap_pagedaemon_waken; 213 214/* 215 * All those kernel PT submaps that BSD is so fond of 216 */ 217pt_entry_t *CMAP1 = 0; 218static pt_entry_t *CMAP2, *CMAP3, *ptmmap; 219caddr_t CADDR1 = 0, ptvmmap = 0; 220static caddr_t CADDR2, CADDR3; 221static struct mtx CMAPCADDR12_lock; 222static pt_entry_t *msgbufmap; 223struct msgbuf *msgbufp = 0; 224 225/* 226 * Crashdump maps. 227 */ 228static pt_entry_t *pt_crashdumpmap; 229static caddr_t crashdumpmap; 230 231#ifdef SMP 232extern pt_entry_t *SMPpt; 233#endif 234static pt_entry_t *PMAP1 = 0; 235static pt_entry_t *PADDR1 = 0; 236 237static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 238static pv_entry_t get_pv_entry(void); 239static void i386_protection_init(void); 240static void pmap_clear_ptes(vm_page_t m, int bit) 241 __always_inline; 242 243static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 244static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 245static int pmap_remove_entry(struct pmap *pmap, vm_page_t m, 246 vm_offset_t va); 247static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, 248 vm_page_t mpte, vm_page_t m); 249 250static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va); 251 252static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex); 253static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t); 254static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 255static void *pmap_pv_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 256#ifdef PAE 257static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 258#endif 259 260CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 261CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 262 263/* 264 * Move the kernel virtual free pointer to the next 265 * 4MB. This is used to help improve performance 266 * by using a large (4MB) page for much of the kernel 267 * (.text, .data, .bss) 268 */ 269static vm_offset_t 270pmap_kmem_choose(vm_offset_t addr) 271{ 272 vm_offset_t newaddr = addr; 273 274#ifndef DISABLE_PSE 275 if (cpu_feature & CPUID_PSE) 276 newaddr = (addr + PDRMASK) & ~PDRMASK; 277#endif 278 return newaddr; 279} 280 281/* 282 * Bootstrap the system enough to run with virtual memory. 283 * 284 * On the i386 this is called after mapping has already been enabled 285 * and just syncs the pmap module with what has already been done. 286 * [We can't call it easily with mapping off since the kernel is not 287 * mapped with PA == VA, hence we would have to relocate every address 288 * from the linked base (virtual) address "KERNBASE" to the actual 289 * (physical) address starting relative to 0] 290 */ 291void 292pmap_bootstrap(firstaddr, loadaddr) 293 vm_paddr_t firstaddr; 294 vm_paddr_t loadaddr; 295{ 296 vm_offset_t va; 297 pt_entry_t *pte; 298 int i; 299 300 avail_start = firstaddr; 301 302 /* 303 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 304 * large. It should instead be correctly calculated in locore.s and 305 * not based on 'first' (which is a physical address, not a virtual 306 * address, for the start of unused physical memory). The kernel 307 * page tables are NOT double mapped and thus should not be included 308 * in this calculation. 309 */ 310 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 311 virtual_avail = pmap_kmem_choose(virtual_avail); 312 313 virtual_end = VM_MAX_KERNEL_ADDRESS; 314 315 /* 316 * Initialize protection array. 317 */ 318 i386_protection_init(); 319 320 /* 321 * Initialize the kernel pmap (which is statically allocated). 322 */ 323 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 324#ifdef PAE 325 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 326#endif 327 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 328 TAILQ_INIT(&kernel_pmap->pm_pvlist); 329 LIST_INIT(&allpmaps); 330#ifdef SMP 331 mtx_init(&lazypmap_lock, "lazypmap", NULL, MTX_SPIN); 332#endif 333 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 334 mtx_lock_spin(&allpmaps_lock); 335 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 336 mtx_unlock_spin(&allpmaps_lock); 337 nkpt = NKPT; 338 339 /* 340 * Reserve some special page table entries/VA space for temporary 341 * mapping of pages. 342 */ 343#define SYSMAP(c, p, v, n) \ 344 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 345 346 va = virtual_avail; 347 pte = vtopte(va); 348 349 /* 350 * CMAP1/CMAP2 are used for zeroing and copying pages. 351 * CMAP3 is used for the idle process page zeroing. 352 */ 353 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 354 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 355 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 356 *CMAP3 = 0; 357 358 mtx_init(&CMAPCADDR12_lock, "CMAPCADDR12", NULL, MTX_DEF); 359 360 /* 361 * Crashdump maps. 362 */ 363 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 364 365 /* 366 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 367 * XXX ptmmap is not used. 368 */ 369 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 370 371 /* 372 * msgbufp is used to map the system message buffer. 373 * XXX msgbufmap is not used. 374 */ 375 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 376 atop(round_page(MSGBUF_SIZE))) 377 378 /* 379 * ptemap is used for pmap_pte_quick 380 */ 381 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 382 383 virtual_avail = va; 384 385 *CMAP1 = *CMAP2 = 0; 386 for (i = 0; i < NKPT; i++) 387 PTD[i] = 0; 388 389 /* Turn on PG_G on kernel page(s) */ 390 pmap_set_pg(); 391} 392 393/* 394 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 395 */ 396void 397pmap_set_pg(void) 398{ 399 pd_entry_t pdir; 400 pt_entry_t *pte; 401 vm_offset_t va, endva; 402 int i; 403 404 if (pgeflag == 0) 405 return; 406 407 i = KERNLOAD/NBPDR; 408 endva = KERNBASE + KERNend; 409 410 if (pseflag) { 411 va = KERNBASE + KERNLOAD; 412 while (va < endva) { 413 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 414 pdir |= pgeflag; 415 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 416 invltlb(); /* Play it safe, invltlb() every time */ 417 i++; 418 va += NBPDR; 419 } 420 } else { 421 va = (vm_offset_t)btext; 422 while (va < endva) { 423 pte = vtopte(va); 424 if (*pte) 425 *pte |= pgeflag; 426 invltlb(); /* Play it safe, invltlb() every time */ 427 va += PAGE_SIZE; 428 } 429 } 430} 431 432static void * 433pmap_pv_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 434{ 435 *flags = UMA_SLAB_PRIV; 436 return (void *)kmem_alloc(kernel_map, bytes); 437} 438 439#ifdef PAE 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, NULL, 0, 0x0ULL, 0xffffffffULL, 1, 0)); 445} 446#endif 447 448/* 449 * Initialize the pmap module. 450 * Called by vm_init, to initialize any structures that the pmap 451 * system needs to map virtual memory. 452 * pmap_init has been enhanced to support in a fairly consistant 453 * way, discontiguous physical memory. 454 */ 455void 456pmap_init(phys_start, phys_end) 457 vm_paddr_t phys_start, phys_end; 458{ 459 int i; 460 int initial_pvs; 461 462 /* 463 * Allocate memory for random pmap data structures. Includes the 464 * pv_head_table. 465 */ 466 467 for(i = 0; i < vm_page_array_size; i++) { 468 vm_page_t m; 469 470 m = &vm_page_array[i]; 471 TAILQ_INIT(&m->md.pv_list); 472 m->md.pv_list_count = 0; 473 } 474 475 /* 476 * init the pv free list 477 */ 478 initial_pvs = vm_page_array_size; 479 if (initial_pvs < MINPV) 480 initial_pvs = MINPV; 481 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL, 482 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 483 uma_zone_set_allocf(pvzone, pmap_pv_allocf); 484 uma_prealloc(pvzone, initial_pvs); 485 486#ifdef PAE 487 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 488 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 489 UMA_ZONE_VM | UMA_ZONE_NOFREE); 490 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 491#endif 492 493 /* 494 * Now it is safe to enable pv_table recording. 495 */ 496 pmap_initialized = TRUE; 497} 498 499/* 500 * Initialize the address space (zone) for the pv_entries. Set a 501 * high water mark so that the system can recover from excessive 502 * numbers of pv entries. 503 */ 504void 505pmap_init2() 506{ 507 int shpgperproc = PMAP_SHPGPERPROC; 508 509 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 510 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 511 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 512 pv_entry_high_water = 9 * (pv_entry_max / 10); 513 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 514} 515 516 517/*************************************************** 518 * Low level helper routines..... 519 ***************************************************/ 520 521#if defined(PMAP_DIAGNOSTIC) 522 523/* 524 * This code checks for non-writeable/modified pages. 525 * This should be an invalid condition. 526 */ 527static int 528pmap_nw_modified(pt_entry_t ptea) 529{ 530 int pte; 531 532 pte = (int) ptea; 533 534 if ((pte & (PG_M|PG_RW)) == PG_M) 535 return 1; 536 else 537 return 0; 538} 539#endif 540 541 542/* 543 * this routine defines the region(s) of memory that should 544 * not be tested for the modified bit. 545 */ 546static PMAP_INLINE int 547pmap_track_modified(vm_offset_t va) 548{ 549 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 550 return 1; 551 else 552 return 0; 553} 554 555#ifdef I386_CPU 556/* 557 * i386 only has "invalidate everything" and no SMP to worry about. 558 */ 559PMAP_INLINE void 560pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 561{ 562 563 if (pmap == kernel_pmap || pmap->pm_active) 564 invltlb(); 565} 566 567PMAP_INLINE void 568pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 569{ 570 571 if (pmap == kernel_pmap || pmap->pm_active) 572 invltlb(); 573} 574 575PMAP_INLINE void 576pmap_invalidate_all(pmap_t pmap) 577{ 578 579 if (pmap == kernel_pmap || pmap->pm_active) 580 invltlb(); 581} 582#else /* !I386_CPU */ 583#ifdef SMP 584/* 585 * For SMP, these functions have to use the IPI mechanism for coherence. 586 */ 587void 588pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 589{ 590 u_int cpumask; 591 u_int other_cpus; 592 593 if (smp_started) { 594 if (!(read_eflags() & PSL_I)) 595 panic("%s: interrupts disabled", __func__); 596 mtx_lock_spin(&smp_tlb_mtx); 597 } else 598 critical_enter(); 599 /* 600 * We need to disable interrupt preemption but MUST NOT have 601 * interrupts disabled here. 602 * XXX we may need to hold schedlock to get a coherent pm_active 603 * XXX critical sections disable interrupts again 604 */ 605 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 606 invlpg(va); 607 smp_invlpg(va); 608 } else { 609 cpumask = PCPU_GET(cpumask); 610 other_cpus = PCPU_GET(other_cpus); 611 if (pmap->pm_active & cpumask) 612 invlpg(va); 613 if (pmap->pm_active & other_cpus) 614 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 615 } 616 if (smp_started) 617 mtx_unlock_spin(&smp_tlb_mtx); 618 else 619 critical_exit(); 620} 621 622void 623pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 624{ 625 u_int cpumask; 626 u_int other_cpus; 627 vm_offset_t addr; 628 629 if (smp_started) { 630 if (!(read_eflags() & PSL_I)) 631 panic("%s: interrupts disabled", __func__); 632 mtx_lock_spin(&smp_tlb_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 for (addr = sva; addr < eva; addr += PAGE_SIZE) 643 invlpg(addr); 644 smp_invlpg_range(sva, eva); 645 } else { 646 cpumask = PCPU_GET(cpumask); 647 other_cpus = PCPU_GET(other_cpus); 648 if (pmap->pm_active & cpumask) 649 for (addr = sva; addr < eva; addr += PAGE_SIZE) 650 invlpg(addr); 651 if (pmap->pm_active & other_cpus) 652 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 653 sva, eva); 654 } 655 if (smp_started) 656 mtx_unlock_spin(&smp_tlb_mtx); 657 else 658 critical_exit(); 659} 660 661void 662pmap_invalidate_all(pmap_t pmap) 663{ 664 u_int cpumask; 665 u_int other_cpus; 666 667 if (smp_started) { 668 if (!(read_eflags() & PSL_I)) 669 panic("%s: interrupts disabled", __func__); 670 mtx_lock_spin(&smp_tlb_mtx); 671 } else 672 critical_enter(); 673 /* 674 * We need to disable interrupt preemption but MUST NOT have 675 * interrupts disabled here. 676 * XXX we may need to hold schedlock to get a coherent pm_active 677 * XXX critical sections disable interrupts again 678 */ 679 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 680 invltlb(); 681 smp_invltlb(); 682 } else { 683 cpumask = PCPU_GET(cpumask); 684 other_cpus = PCPU_GET(other_cpus); 685 if (pmap->pm_active & cpumask) 686 invltlb(); 687 if (pmap->pm_active & other_cpus) 688 smp_masked_invltlb(pmap->pm_active & other_cpus); 689 } 690 if (smp_started) 691 mtx_unlock_spin(&smp_tlb_mtx); 692 else 693 critical_exit(); 694} 695#else /* !SMP */ 696/* 697 * Normal, non-SMP, 486+ invalidation functions. 698 * We inline these within pmap.c for speed. 699 */ 700PMAP_INLINE void 701pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 702{ 703 704 if (pmap == kernel_pmap || pmap->pm_active) 705 invlpg(va); 706} 707 708PMAP_INLINE void 709pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 710{ 711 vm_offset_t addr; 712 713 if (pmap == kernel_pmap || pmap->pm_active) 714 for (addr = sva; addr < eva; addr += PAGE_SIZE) 715 invlpg(addr); 716} 717 718PMAP_INLINE void 719pmap_invalidate_all(pmap_t pmap) 720{ 721 722 if (pmap == kernel_pmap || pmap->pm_active) 723 invltlb(); 724} 725#endif /* !SMP */ 726#endif /* !I386_CPU */ 727 728/* 729 * Are we current address space or kernel? N.B. We return FALSE when 730 * a pmap's page table is in use because a kernel thread is borrowing 731 * it. The borrowed page table can change spontaneously, making any 732 * dependence on its continued use subject to a race condition. 733 */ 734static int 735pmap_is_current(pmap_t pmap) 736{ 737 return (pmap == kernel_pmap || 738 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 739 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 740} 741 742/* 743 * Super fast pmap_pte routine best used when scanning 744 * the pv lists. This eliminates many coarse-grained 745 * invltlb calls. Note that many of the pv list 746 * scans are across different pmaps. It is very wasteful 747 * to do an entire invltlb for checking a single mapping. 748 */ 749pt_entry_t * 750pmap_pte_quick(pmap, va) 751 register pmap_t pmap; 752 vm_offset_t va; 753{ 754 pd_entry_t newpf; 755 pd_entry_t *pde; 756 757 pde = pmap_pde(pmap, va); 758 if (*pde & PG_PS) 759 return (pde); 760 if (*pde != 0) { 761 /* are we current address space or kernel? */ 762 if (pmap_is_current(pmap)) 763 return vtopte(va); 764 newpf = *pde & PG_FRAME; 765 if (((*PMAP1) & PG_FRAME) != newpf) { 766 *PMAP1 = newpf | PG_RW | PG_V; 767 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR1); 768 } 769 return PADDR1 + (i386_btop(va) & (NPTEPG - 1)); 770 } 771 return (0); 772} 773 774/* 775 * Routine: pmap_extract 776 * Function: 777 * Extract the physical page address associated 778 * with the given map/virtual_address pair. 779 */ 780vm_paddr_t 781pmap_extract(pmap, va) 782 register pmap_t pmap; 783 vm_offset_t va; 784{ 785 vm_paddr_t rtval; 786 pt_entry_t *pte; 787 pd_entry_t pde; 788 789 if (pmap == 0) 790 return 0; 791 pde = pmap->pm_pdir[va >> PDRSHIFT]; 792 if (pde != 0) { 793 if ((pde & PG_PS) != 0) { 794 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 795 return rtval; 796 } 797 pte = pmap_pte_quick(pmap, va); 798 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK)); 799 return rtval; 800 } 801 return 0; 802 803} 804 805/* 806 * Routine: pmap_extract_and_hold 807 * Function: 808 * Atomically extract and hold the physical page 809 * with the given pmap and virtual address pair 810 * if that mapping permits the given protection. 811 */ 812vm_page_t 813pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 814{ 815 vm_paddr_t pa; 816 vm_page_t m; 817 818 m = NULL; 819 mtx_lock(&Giant); 820 if ((pa = pmap_extract(pmap, va)) != 0) { 821 m = PHYS_TO_VM_PAGE(pa); 822 vm_page_lock_queues(); 823 vm_page_hold(m); 824 vm_page_unlock_queues(); 825 } 826 mtx_unlock(&Giant); 827 return (m); 828} 829 830/*************************************************** 831 * Low level mapping routines..... 832 ***************************************************/ 833 834/* 835 * Add a wired page to the kva. 836 * Note: not SMP coherent. 837 */ 838PMAP_INLINE void 839pmap_kenter(vm_offset_t va, vm_paddr_t pa) 840{ 841 pt_entry_t *pte; 842 843 pte = vtopte(va); 844 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 845} 846 847/* 848 * Remove a page from the kernel pagetables. 849 * Note: not SMP coherent. 850 */ 851PMAP_INLINE void 852pmap_kremove(vm_offset_t va) 853{ 854 pt_entry_t *pte; 855 856 pte = vtopte(va); 857 pte_clear(pte); 858} 859 860/* 861 * Used to map a range of physical addresses into kernel 862 * virtual address space. 863 * 864 * The value passed in '*virt' is a suggested virtual address for 865 * the mapping. Architectures which can support a direct-mapped 866 * physical to virtual region can return the appropriate address 867 * within that region, leaving '*virt' unchanged. Other 868 * architectures should map the pages starting at '*virt' and 869 * update '*virt' with the first usable address after the mapped 870 * region. 871 */ 872vm_offset_t 873pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 874{ 875 vm_offset_t va, sva; 876 877 va = sva = *virt; 878 while (start < end) { 879 pmap_kenter(va, start); 880 va += PAGE_SIZE; 881 start += PAGE_SIZE; 882 } 883 pmap_invalidate_range(kernel_pmap, sva, va); 884 *virt = va; 885 return (sva); 886} 887 888 889/* 890 * Add a list of wired pages to the kva 891 * this routine is only used for temporary 892 * kernel mappings that do not need to have 893 * page modification or references recorded. 894 * Note that old mappings are simply written 895 * over. The page *must* be wired. 896 * Note: SMP coherent. Uses a ranged shootdown IPI. 897 */ 898void 899pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 900{ 901 vm_offset_t va; 902 903 va = sva; 904 while (count-- > 0) { 905 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 906 va += PAGE_SIZE; 907 m++; 908 } 909 pmap_invalidate_range(kernel_pmap, sva, va); 910} 911 912/* 913 * This routine tears out page mappings from the 914 * kernel -- it is meant only for temporary mappings. 915 * Note: SMP coherent. Uses a ranged shootdown IPI. 916 */ 917void 918pmap_qremove(vm_offset_t sva, int count) 919{ 920 vm_offset_t va; 921 922 va = sva; 923 while (count-- > 0) { 924 pmap_kremove(va); 925 va += PAGE_SIZE; 926 } 927 pmap_invalidate_range(kernel_pmap, sva, va); 928} 929 930/*************************************************** 931 * Page table page management routines..... 932 ***************************************************/ 933 934/* 935 * This routine unholds page table pages, and if the hold count 936 * drops to zero, then it decrements the wire count. 937 */ 938static int 939_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 940{ 941 942 while (vm_page_sleep_if_busy(m, FALSE, "pmuwpt")) 943 vm_page_lock_queues(); 944 945 if (m->hold_count == 0) { 946 vm_offset_t pteva; 947 /* 948 * unmap the page table page 949 */ 950 pmap->pm_pdir[m->pindex] = 0; 951 --pmap->pm_stats.resident_count; 952 /* 953 * We never unwire a kernel page table page, making a 954 * check for the kernel_pmap unnecessary. 955 */ 956 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)) { 957 /* 958 * Do an invltlb to make the invalidated mapping 959 * take effect immediately. 960 */ 961 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 962 pmap_invalidate_page(pmap, pteva); 963 } 964 965 /* 966 * If the page is finally unwired, simply free it. 967 */ 968 --m->wire_count; 969 if (m->wire_count == 0) { 970 vm_page_busy(m); 971 vm_page_free_zero(m); 972 atomic_subtract_int(&cnt.v_wire_count, 1); 973 } 974 return 1; 975 } 976 return 0; 977} 978 979static PMAP_INLINE int 980pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 981{ 982 vm_page_unhold(m); 983 if (m->hold_count == 0) 984 return _pmap_unwire_pte_hold(pmap, m); 985 else 986 return 0; 987} 988 989/* 990 * After removing a page table entry, this routine is used to 991 * conditionally free the page, and manage the hold/wire counts. 992 */ 993static int 994pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 995{ 996 997 if (va >= VM_MAXUSER_ADDRESS) 998 return 0; 999 1000 return pmap_unwire_pte_hold(pmap, mpte); 1001} 1002 1003void 1004pmap_pinit0(pmap) 1005 struct pmap *pmap; 1006{ 1007 1008 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1009#ifdef PAE 1010 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1011#endif 1012 pmap->pm_active = 0; 1013 PCPU_SET(curpmap, pmap); 1014 TAILQ_INIT(&pmap->pm_pvlist); 1015 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1016 mtx_lock_spin(&allpmaps_lock); 1017 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1018 mtx_unlock_spin(&allpmaps_lock); 1019} 1020 1021/* 1022 * Initialize a preallocated and zeroed pmap structure, 1023 * such as one in a vmspace structure. 1024 */ 1025void 1026pmap_pinit(pmap) 1027 register struct pmap *pmap; 1028{ 1029 vm_page_t m, ptdpg[NPGPTD]; 1030 vm_paddr_t pa; 1031 static int color; 1032 int i; 1033 1034 /* 1035 * No need to allocate page table space yet but we do need a valid 1036 * page directory table. 1037 */ 1038 if (pmap->pm_pdir == NULL) { 1039 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1040 NBPTD); 1041#ifdef PAE 1042 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1043 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1044 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1045 ("pmap_pinit: pdpt misaligned")); 1046 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1047 ("pmap_pinit: pdpt above 4g")); 1048#endif 1049 } 1050 1051 /* 1052 * allocate the page directory page(s) 1053 */ 1054 for (i = 0; i < NPGPTD;) { 1055 m = vm_page_alloc(NULL, color++, 1056 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1057 VM_ALLOC_ZERO); 1058 if (m == NULL) 1059 VM_WAIT; 1060 else { 1061 vm_page_lock_queues(); 1062 vm_page_flag_clear(m, PG_BUSY); 1063 vm_page_unlock_queues(); 1064 ptdpg[i++] = m; 1065 } 1066 } 1067 1068 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1069 1070 for (i = 0; i < NPGPTD; i++) { 1071 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1072 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1073 } 1074 1075 mtx_lock_spin(&allpmaps_lock); 1076 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1077 mtx_unlock_spin(&allpmaps_lock); 1078 /* Wire in kernel global address entries. */ 1079 /* XXX copies current process, does not fill in MPPTDI */ 1080 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1081#ifdef SMP 1082 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1083#endif 1084 1085 /* install self-referential address mapping entry(s) */ 1086 for (i = 0; i < NPGPTD; i++) { 1087 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1088 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1089#ifdef PAE 1090 pmap->pm_pdpt[i] = pa | PG_V; 1091#endif 1092 } 1093 1094 pmap->pm_active = 0; 1095 TAILQ_INIT(&pmap->pm_pvlist); 1096 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1097} 1098 1099/* 1100 * Wire in kernel global address entries. To avoid a race condition 1101 * between pmap initialization and pmap_growkernel, this procedure 1102 * should be called after the vmspace is attached to the process 1103 * but before this pmap is activated. 1104 */ 1105void 1106pmap_pinit2(pmap) 1107 struct pmap *pmap; 1108{ 1109 /* XXX: Remove this stub when no longer called */ 1110} 1111 1112/* 1113 * this routine is called if the page table page is not 1114 * mapped correctly. 1115 */ 1116static vm_page_t 1117_pmap_allocpte(pmap, ptepindex) 1118 pmap_t pmap; 1119 unsigned ptepindex; 1120{ 1121 vm_paddr_t ptepa; 1122 vm_page_t m; 1123 1124 /* 1125 * Allocate a page table page. 1126 */ 1127 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1128 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1129 VM_WAIT; 1130 /* 1131 * Indicate the need to retry. While waiting, the page table 1132 * page may have been allocated. 1133 */ 1134 return (NULL); 1135 } 1136 if ((m->flags & PG_ZERO) == 0) 1137 pmap_zero_page(m); 1138 1139 KASSERT(m->queue == PQ_NONE, 1140 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1141 1142 /* 1143 * Increment the hold count for the page table page 1144 * (denoting a new mapping.) 1145 */ 1146 m->hold_count++; 1147 1148 /* 1149 * Map the pagetable page into the process address space, if 1150 * it isn't already there. 1151 */ 1152 1153 pmap->pm_stats.resident_count++; 1154 1155 ptepa = VM_PAGE_TO_PHYS(m); 1156 pmap->pm_pdir[ptepindex] = 1157 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1158 1159 vm_page_lock_queues(); 1160 vm_page_flag_clear(m, PG_ZERO); 1161 vm_page_wakeup(m); 1162 vm_page_unlock_queues(); 1163 1164 return m; 1165} 1166 1167static vm_page_t 1168pmap_allocpte(pmap_t pmap, vm_offset_t va) 1169{ 1170 unsigned ptepindex; 1171 pd_entry_t ptepa; 1172 vm_page_t m; 1173 1174 /* 1175 * Calculate pagetable page index 1176 */ 1177 ptepindex = va >> PDRSHIFT; 1178retry: 1179 /* 1180 * Get the page directory entry 1181 */ 1182 ptepa = pmap->pm_pdir[ptepindex]; 1183 1184 /* 1185 * This supports switching from a 4MB page to a 1186 * normal 4K page. 1187 */ 1188 if (ptepa & PG_PS) { 1189 pmap->pm_pdir[ptepindex] = 0; 1190 ptepa = 0; 1191 pmap_invalidate_all(kernel_pmap); 1192 } 1193 1194 /* 1195 * If the page table page is mapped, we just increment the 1196 * hold count, and activate it. 1197 */ 1198 if (ptepa) { 1199 m = PHYS_TO_VM_PAGE(ptepa); 1200 m->hold_count++; 1201 } else { 1202 /* 1203 * Here if the pte page isn't mapped, or if it has 1204 * been deallocated. 1205 */ 1206 m = _pmap_allocpte(pmap, ptepindex); 1207 if (m == NULL) 1208 goto retry; 1209 } 1210 return (m); 1211} 1212 1213 1214/*************************************************** 1215* Pmap allocation/deallocation routines. 1216 ***************************************************/ 1217 1218#ifdef SMP 1219/* 1220 * Deal with a SMP shootdown of other users of the pmap that we are 1221 * trying to dispose of. This can be a bit hairy. 1222 */ 1223static u_int *lazymask; 1224static u_int lazyptd; 1225static volatile u_int lazywait; 1226 1227void pmap_lazyfix_action(void); 1228 1229void 1230pmap_lazyfix_action(void) 1231{ 1232 u_int mymask = PCPU_GET(cpumask); 1233 1234 if (rcr3() == lazyptd) 1235 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1236 atomic_clear_int(lazymask, mymask); 1237 atomic_store_rel_int(&lazywait, 1); 1238} 1239 1240static void 1241pmap_lazyfix_self(u_int mymask) 1242{ 1243 1244 if (rcr3() == lazyptd) 1245 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1246 atomic_clear_int(lazymask, mymask); 1247} 1248 1249 1250static void 1251pmap_lazyfix(pmap_t pmap) 1252{ 1253 u_int mymask = PCPU_GET(cpumask); 1254 u_int mask; 1255 register u_int spins; 1256 1257 while ((mask = pmap->pm_active) != 0) { 1258 spins = 50000000; 1259 mask = mask & -mask; /* Find least significant set bit */ 1260 mtx_lock_spin(&lazypmap_lock); 1261#ifdef PAE 1262 lazyptd = vtophys(pmap->pm_pdpt); 1263#else 1264 lazyptd = vtophys(pmap->pm_pdir); 1265#endif 1266 if (mask == mymask) { 1267 lazymask = &pmap->pm_active; 1268 pmap_lazyfix_self(mymask); 1269 } else { 1270 atomic_store_rel_int((u_int *)&lazymask, 1271 (u_int)&pmap->pm_active); 1272 atomic_store_rel_int(&lazywait, 0); 1273 ipi_selected(mask, IPI_LAZYPMAP); 1274 while (lazywait == 0) { 1275 ia32_pause(); 1276 if (--spins == 0) 1277 break; 1278 } 1279 } 1280 mtx_unlock_spin(&lazypmap_lock); 1281 if (spins == 0) 1282 printf("pmap_lazyfix: spun for 50000000\n"); 1283 } 1284} 1285 1286#else /* SMP */ 1287 1288/* 1289 * Cleaning up on uniprocessor is easy. For various reasons, we're 1290 * unlikely to have to even execute this code, including the fact 1291 * that the cleanup is deferred until the parent does a wait(2), which 1292 * means that another userland process has run. 1293 */ 1294static void 1295pmap_lazyfix(pmap_t pmap) 1296{ 1297 u_int cr3; 1298 1299 cr3 = vtophys(pmap->pm_pdir); 1300 if (cr3 == rcr3()) { 1301 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1302 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1303 } 1304} 1305#endif /* SMP */ 1306 1307/* 1308 * Release any resources held by the given physical map. 1309 * Called when a pmap initialized by pmap_pinit is being released. 1310 * Should only be called if the map contains no valid mappings. 1311 */ 1312void 1313pmap_release(pmap_t pmap) 1314{ 1315 vm_page_t m, ptdpg[NPGPTD]; 1316 int i; 1317 1318 KASSERT(pmap->pm_stats.resident_count == 0, 1319 ("pmap_release: pmap resident count %ld != 0", 1320 pmap->pm_stats.resident_count)); 1321 1322 pmap_lazyfix(pmap); 1323 mtx_lock_spin(&allpmaps_lock); 1324 LIST_REMOVE(pmap, pm_list); 1325 mtx_unlock_spin(&allpmaps_lock); 1326 1327 for (i = 0; i < NPGPTD; i++) 1328 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1329 1330 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1331 sizeof(*pmap->pm_pdir)); 1332#ifdef SMP 1333 pmap->pm_pdir[MPPTDI] = 0; 1334#endif 1335 1336 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1337 1338 vm_page_lock_queues(); 1339 for (i = 0; i < NPGPTD; i++) { 1340 m = ptdpg[i]; 1341#ifdef PAE 1342 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1343 ("pmap_release: got wrong ptd page")); 1344#endif 1345 m->wire_count--; 1346 atomic_subtract_int(&cnt.v_wire_count, 1); 1347 vm_page_busy(m); 1348 vm_page_free_zero(m); 1349 } 1350 vm_page_unlock_queues(); 1351} 1352 1353static int 1354kvm_size(SYSCTL_HANDLER_ARGS) 1355{ 1356 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1357 1358 return sysctl_handle_long(oidp, &ksize, 0, req); 1359} 1360SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1361 0, 0, kvm_size, "IU", "Size of KVM"); 1362 1363static int 1364kvm_free(SYSCTL_HANDLER_ARGS) 1365{ 1366 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1367 1368 return sysctl_handle_long(oidp, &kfree, 0, req); 1369} 1370SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1371 0, 0, kvm_free, "IU", "Amount of KVM free"); 1372 1373/* 1374 * grow the number of kernel page table entries, if needed 1375 */ 1376void 1377pmap_growkernel(vm_offset_t addr) 1378{ 1379 struct pmap *pmap; 1380 int s; 1381 vm_paddr_t ptppaddr; 1382 vm_page_t nkpg; 1383 pd_entry_t newpdir; 1384 pt_entry_t *pde; 1385 1386 s = splhigh(); 1387 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1388 if (kernel_vm_end == 0) { 1389 kernel_vm_end = KERNBASE; 1390 nkpt = 0; 1391 while (pdir_pde(PTD, kernel_vm_end)) { 1392 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1393 nkpt++; 1394 } 1395 } 1396 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1397 while (kernel_vm_end < addr) { 1398 if (pdir_pde(PTD, kernel_vm_end)) { 1399 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1400 continue; 1401 } 1402 1403 /* 1404 * This index is bogus, but out of the way 1405 */ 1406 nkpg = vm_page_alloc(NULL, nkpt, 1407 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1408 if (!nkpg) 1409 panic("pmap_growkernel: no memory to grow kernel"); 1410 1411 nkpt++; 1412 1413 pmap_zero_page(nkpg); 1414 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1415 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1416 pdir_pde(PTD, kernel_vm_end) = newpdir; 1417 1418 mtx_lock_spin(&allpmaps_lock); 1419 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1420 pde = pmap_pde(pmap, kernel_vm_end); 1421 pde_store(pde, newpdir); 1422 } 1423 mtx_unlock_spin(&allpmaps_lock); 1424 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1425 } 1426 splx(s); 1427} 1428 1429 1430/*************************************************** 1431 * page management routines. 1432 ***************************************************/ 1433 1434/* 1435 * free the pv_entry back to the free list 1436 */ 1437static PMAP_INLINE void 1438free_pv_entry(pv_entry_t pv) 1439{ 1440 pv_entry_count--; 1441 uma_zfree(pvzone, pv); 1442} 1443 1444/* 1445 * get a new pv_entry, allocating a block from the system 1446 * when needed. 1447 * the memory allocation is performed bypassing the malloc code 1448 * because of the possibility of allocations at interrupt time. 1449 */ 1450static pv_entry_t 1451get_pv_entry(void) 1452{ 1453 pv_entry_count++; 1454 if (pv_entry_high_water && 1455 (pv_entry_count > pv_entry_high_water) && 1456 (pmap_pagedaemon_waken == 0)) { 1457 pmap_pagedaemon_waken = 1; 1458 wakeup (&vm_pages_needed); 1459 } 1460 return uma_zalloc(pvzone, M_NOWAIT); 1461} 1462 1463/* 1464 * If it is the first entry on the list, it is actually 1465 * in the header and we must copy the following entry up 1466 * to the header. Otherwise we must search the list for 1467 * the entry. In either case we free the now unused entry. 1468 */ 1469 1470static int 1471pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1472{ 1473 pv_entry_t pv; 1474 int rtval; 1475 int s; 1476 1477 s = splvm(); 1478 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1479 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1480 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1481 if (pmap == pv->pv_pmap && va == pv->pv_va) 1482 break; 1483 } 1484 } else { 1485 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1486 if (va == pv->pv_va) 1487 break; 1488 } 1489 } 1490 1491 rtval = 0; 1492 if (pv) { 1493 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1494 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1495 m->md.pv_list_count--; 1496 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1497 vm_page_flag_clear(m, PG_WRITEABLE); 1498 1499 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1500 free_pv_entry(pv); 1501 } 1502 1503 splx(s); 1504 return rtval; 1505} 1506 1507/* 1508 * Create a pv entry for page at pa for 1509 * (pmap, va). 1510 */ 1511static void 1512pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1513{ 1514 1515 int s; 1516 pv_entry_t pv; 1517 1518 s = splvm(); 1519 pv = get_pv_entry(); 1520 pv->pv_va = va; 1521 pv->pv_pmap = pmap; 1522 pv->pv_ptem = mpte; 1523 1524 vm_page_lock_queues(); 1525 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1526 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1527 m->md.pv_list_count++; 1528 1529 vm_page_unlock_queues(); 1530 splx(s); 1531} 1532 1533/* 1534 * pmap_remove_pte: do the things to unmap a page in a process 1535 */ 1536static int 1537pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1538{ 1539 pt_entry_t oldpte; 1540 vm_page_t m, mpte; 1541 1542 oldpte = pte_load_clear(ptq); 1543 if (oldpte & PG_W) 1544 pmap->pm_stats.wired_count -= 1; 1545 /* 1546 * Machines that don't support invlpg, also don't support 1547 * PG_G. 1548 */ 1549 if (oldpte & PG_G) 1550 pmap_invalidate_page(kernel_pmap, va); 1551 pmap->pm_stats.resident_count -= 1; 1552 if (oldpte & PG_MANAGED) { 1553 m = PHYS_TO_VM_PAGE(oldpte); 1554 if (oldpte & PG_M) { 1555#if defined(PMAP_DIAGNOSTIC) 1556 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1557 printf( 1558 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1559 va, oldpte); 1560 } 1561#endif 1562 if (pmap_track_modified(va)) 1563 vm_page_dirty(m); 1564 } 1565 if (oldpte & PG_A) 1566 vm_page_flag_set(m, PG_REFERENCED); 1567 return pmap_remove_entry(pmap, m, va); 1568 } else { 1569 mpte = PHYS_TO_VM_PAGE(*pmap_pde(pmap, va)); 1570 return pmap_unuse_pt(pmap, va, mpte); 1571 } 1572} 1573 1574/* 1575 * Remove a single page from a process address space 1576 */ 1577static void 1578pmap_remove_page(pmap_t pmap, vm_offset_t va) 1579{ 1580 pt_entry_t *pte; 1581 1582 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) 1583 return; 1584 pmap_remove_pte(pmap, pte, va); 1585 pmap_invalidate_page(pmap, va); 1586} 1587 1588/* 1589 * Remove the given range of addresses from the specified map. 1590 * 1591 * It is assumed that the start and end are properly 1592 * rounded to the page size. 1593 */ 1594void 1595pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1596{ 1597 vm_offset_t pdnxt; 1598 pd_entry_t ptpaddr; 1599 pt_entry_t *pte; 1600 int anyvalid; 1601 1602 if (pmap == NULL) 1603 return; 1604 1605 if (pmap->pm_stats.resident_count == 0) 1606 return; 1607 1608 /* 1609 * special handling of removing one page. a very 1610 * common operation and easy to short circuit some 1611 * code. 1612 */ 1613 if ((sva + PAGE_SIZE == eva) && 1614 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1615 pmap_remove_page(pmap, sva); 1616 return; 1617 } 1618 1619 anyvalid = 0; 1620 1621 for (; sva < eva; sva = pdnxt) { 1622 unsigned pdirindex; 1623 1624 /* 1625 * Calculate index for next page table. 1626 */ 1627 pdnxt = (sva + NBPDR) & ~PDRMASK; 1628 if (pmap->pm_stats.resident_count == 0) 1629 break; 1630 1631 pdirindex = sva >> PDRSHIFT; 1632 ptpaddr = pmap->pm_pdir[pdirindex]; 1633 1634 /* 1635 * Weed out invalid mappings. Note: we assume that the page 1636 * directory table is always allocated, and in kernel virtual. 1637 */ 1638 if (ptpaddr == 0) 1639 continue; 1640 1641 /* 1642 * Check for large page. 1643 */ 1644 if ((ptpaddr & PG_PS) != 0) { 1645 pmap->pm_pdir[pdirindex] = 0; 1646 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1647 anyvalid = 1; 1648 continue; 1649 } 1650 1651 /* 1652 * Limit our scan to either the end of the va represented 1653 * by the current page table page, or to the end of the 1654 * range being removed. 1655 */ 1656 if (pdnxt > eva) 1657 pdnxt = eva; 1658 1659 for (; sva != pdnxt; sva += PAGE_SIZE) { 1660 if ((pte = pmap_pte_quick(pmap, sva)) == NULL || 1661 *pte == 0) 1662 continue; 1663 anyvalid = 1; 1664 if (pmap_remove_pte(pmap, pte, sva)) 1665 break; 1666 } 1667 } 1668 1669 if (anyvalid) 1670 pmap_invalidate_all(pmap); 1671} 1672 1673/* 1674 * Routine: pmap_remove_all 1675 * Function: 1676 * Removes this physical page from 1677 * all physical maps in which it resides. 1678 * Reflects back modify bits to the pager. 1679 * 1680 * Notes: 1681 * Original versions of this routine were very 1682 * inefficient because they iteratively called 1683 * pmap_remove (slow...) 1684 */ 1685 1686void 1687pmap_remove_all(vm_page_t m) 1688{ 1689 register pv_entry_t pv; 1690 pt_entry_t *pte, tpte; 1691 int s; 1692 1693#if defined(PMAP_DIAGNOSTIC) 1694 /* 1695 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1696 */ 1697 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1698 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1699 VM_PAGE_TO_PHYS(m)); 1700 } 1701#endif 1702 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1703 s = splvm(); 1704 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1705 pv->pv_pmap->pm_stats.resident_count--; 1706 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1707 tpte = pte_load_clear(pte); 1708 if (tpte & PG_W) 1709 pv->pv_pmap->pm_stats.wired_count--; 1710 if (tpte & PG_A) 1711 vm_page_flag_set(m, PG_REFERENCED); 1712 1713 /* 1714 * Update the vm_page_t clean and reference bits. 1715 */ 1716 if (tpte & PG_M) { 1717#if defined(PMAP_DIAGNOSTIC) 1718 if (pmap_nw_modified((pt_entry_t) tpte)) { 1719 printf( 1720 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1721 pv->pv_va, tpte); 1722 } 1723#endif 1724 if (pmap_track_modified(pv->pv_va)) 1725 vm_page_dirty(m); 1726 } 1727 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1728 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1729 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1730 m->md.pv_list_count--; 1731 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 1732 free_pv_entry(pv); 1733 } 1734 vm_page_flag_clear(m, PG_WRITEABLE); 1735 splx(s); 1736} 1737 1738/* 1739 * Set the physical protection on the 1740 * specified range of this map as requested. 1741 */ 1742void 1743pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1744{ 1745 vm_offset_t pdnxt; 1746 pd_entry_t ptpaddr; 1747 int anychanged; 1748 1749 if (pmap == NULL) 1750 return; 1751 1752 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1753 pmap_remove(pmap, sva, eva); 1754 return; 1755 } 1756 1757 if (prot & VM_PROT_WRITE) 1758 return; 1759 1760 anychanged = 0; 1761 1762 for (; sva < eva; sva = pdnxt) { 1763 unsigned 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 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1783 anychanged = 1; 1784 continue; 1785 } 1786 1787 if (pdnxt > eva) 1788 pdnxt = eva; 1789 1790 for (; sva != pdnxt; sva += PAGE_SIZE) { 1791 pt_entry_t pbits; 1792 pt_entry_t *pte; 1793 vm_page_t m; 1794 1795 if ((pte = pmap_pte_quick(pmap, sva)) == NULL) 1796 continue; 1797 pbits = *pte; 1798 if (pbits & PG_MANAGED) { 1799 m = NULL; 1800 if (pbits & PG_A) { 1801 m = PHYS_TO_VM_PAGE(pbits); 1802 vm_page_flag_set(m, PG_REFERENCED); 1803 pbits &= ~PG_A; 1804 } 1805 if ((pbits & PG_M) != 0 && 1806 pmap_track_modified(sva)) { 1807 if (m == NULL) 1808 m = PHYS_TO_VM_PAGE(pbits); 1809 vm_page_dirty(m); 1810 pbits &= ~PG_M; 1811 } 1812 } 1813 1814 pbits &= ~PG_RW; 1815 1816 if (pbits != *pte) { 1817 pte_store(pte, pbits); 1818 anychanged = 1; 1819 } 1820 } 1821 } 1822 if (anychanged) 1823 pmap_invalidate_all(pmap); 1824} 1825 1826/* 1827 * Insert the given physical page (p) at 1828 * the specified virtual address (v) in the 1829 * target physical map with the protection requested. 1830 * 1831 * If specified, the page will be wired down, meaning 1832 * that the related pte can not be reclaimed. 1833 * 1834 * NB: This is the only routine which MAY NOT lazy-evaluate 1835 * or lose information. That is, this routine must actually 1836 * insert this page into the given map NOW. 1837 */ 1838void 1839pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1840 boolean_t wired) 1841{ 1842 vm_paddr_t pa; 1843 register pt_entry_t *pte; 1844 vm_paddr_t opa; 1845 pt_entry_t origpte, newpte; 1846 vm_page_t mpte; 1847 1848 if (pmap == NULL) 1849 return; 1850 1851 va &= PG_FRAME; 1852#ifdef PMAP_DIAGNOSTIC 1853 if (va > VM_MAX_KERNEL_ADDRESS) 1854 panic("pmap_enter: toobig"); 1855 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1856 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1857#endif 1858 1859 mpte = NULL; 1860 /* 1861 * In the case that a page table page is not 1862 * resident, we are creating it here. 1863 */ 1864 if (va < VM_MAXUSER_ADDRESS) { 1865 mpte = pmap_allocpte(pmap, va); 1866 } 1867#if 0 && defined(PMAP_DIAGNOSTIC) 1868 else { 1869 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1870 origpte = *pdeaddr; 1871 if ((origpte & PG_V) == 0) { 1872 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1873 pmap->pm_pdir[PTDPTDI], origpte, va); 1874 } 1875 } 1876#endif 1877 1878 pte = pmap_pte_quick(pmap, va); 1879 1880 /* 1881 * Page Directory table entry not valid, we need a new PT page 1882 */ 1883 if (pte == NULL) { 1884 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1885 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1886 } 1887 1888 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 1889 origpte = *pte; 1890 opa = origpte & PG_FRAME; 1891 1892 if (origpte & PG_PS) { 1893 /* 1894 * Yes, I know this will truncate upper address bits for PAE, 1895 * but I'm actually more interested in the lower bits 1896 */ 1897 printf("pmap_enter: va %p, pte %p, origpte %p\n", 1898 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 1899 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1900 } 1901 1902 /* 1903 * Mapping has not changed, must be protection or wiring change. 1904 */ 1905 if (origpte && (opa == pa)) { 1906 /* 1907 * Wiring change, just update stats. We don't worry about 1908 * wiring PT pages as they remain resident as long as there 1909 * are valid mappings in them. Hence, if a user page is wired, 1910 * the PT page will be also. 1911 */ 1912 if (wired && ((origpte & PG_W) == 0)) 1913 pmap->pm_stats.wired_count++; 1914 else if (!wired && (origpte & PG_W)) 1915 pmap->pm_stats.wired_count--; 1916 1917#if defined(PMAP_DIAGNOSTIC) 1918 if (pmap_nw_modified((pt_entry_t) origpte)) { 1919 printf( 1920 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 1921 va, origpte); 1922 } 1923#endif 1924 1925 /* 1926 * Remove extra pte reference 1927 */ 1928 if (mpte) 1929 mpte->hold_count--; 1930 1931 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) { 1932 if ((origpte & PG_RW) == 0) { 1933 pte_store(pte, origpte | PG_RW); 1934 pmap_invalidate_page(pmap, va); 1935 } 1936 return; 1937 } 1938 1939 /* 1940 * We might be turning off write access to the page, 1941 * so we go ahead and sense modify status. 1942 */ 1943 if (origpte & PG_MANAGED) { 1944 if ((origpte & PG_M) && pmap_track_modified(va)) { 1945 vm_page_t om; 1946 om = PHYS_TO_VM_PAGE(opa); 1947 vm_page_dirty(om); 1948 } 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 int err; 1959 vm_page_lock_queues(); 1960 err = pmap_remove_pte(pmap, pte, va); 1961 vm_page_unlock_queues(); 1962 if (err) 1963 panic("pmap_enter: pte vanished, va: 0x%x", va); 1964 } 1965 1966 /* 1967 * Enter on the PV list if part of our managed memory. Note that we 1968 * raise IPL while manipulating pv_table since pmap_enter can be 1969 * called at interrupt time. 1970 */ 1971 if (pmap_initialized && 1972 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 1973 pmap_insert_entry(pmap, va, mpte, m); 1974 pa |= PG_MANAGED; 1975 } 1976 1977 /* 1978 * Increment counters 1979 */ 1980 pmap->pm_stats.resident_count++; 1981 if (wired) 1982 pmap->pm_stats.wired_count++; 1983 1984validate: 1985 /* 1986 * Now validate mapping with desired protection/wiring. 1987 */ 1988 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V); 1989 1990 if (wired) 1991 newpte |= PG_W; 1992 if (va < VM_MAXUSER_ADDRESS) 1993 newpte |= PG_U; 1994 if (pmap == kernel_pmap) 1995 newpte |= pgeflag; 1996 1997 /* 1998 * if the mapping or permission bits are different, we need 1999 * to update the pte. 2000 */ 2001 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2002 pte_store(pte, newpte | PG_A); 2003 /*if (origpte)*/ { 2004 pmap_invalidate_page(pmap, va); 2005 } 2006 } 2007} 2008 2009/* 2010 * this code makes some *MAJOR* assumptions: 2011 * 1. Current pmap & pmap exists. 2012 * 2. Not wired. 2013 * 3. Read access. 2014 * 4. No page table pages. 2015 * 5. Tlbflush is deferred to calling procedure. 2016 * 6. Page IS managed. 2017 * but is *MUCH* faster than pmap_enter... 2018 */ 2019 2020vm_page_t 2021pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2022{ 2023 pt_entry_t *pte; 2024 vm_paddr_t pa; 2025 2026 /* 2027 * In the case that a page table page is not 2028 * resident, we are creating it here. 2029 */ 2030 if (va < VM_MAXUSER_ADDRESS) { 2031 unsigned ptepindex; 2032 pd_entry_t ptepa; 2033 2034 /* 2035 * Calculate pagetable page index 2036 */ 2037 ptepindex = va >> PDRSHIFT; 2038 if (mpte && (mpte->pindex == ptepindex)) { 2039 mpte->hold_count++; 2040 } else { 2041retry: 2042 /* 2043 * Get the page directory entry 2044 */ 2045 ptepa = pmap->pm_pdir[ptepindex]; 2046 2047 /* 2048 * If the page table page is mapped, we just increment 2049 * the hold count, and activate it. 2050 */ 2051 if (ptepa) { 2052 if (ptepa & PG_PS) 2053 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2054 mpte = PHYS_TO_VM_PAGE(ptepa); 2055 mpte->hold_count++; 2056 } else { 2057 mpte = _pmap_allocpte(pmap, ptepindex); 2058 if (mpte == NULL) 2059 goto retry; 2060 } 2061 } 2062 } else { 2063 mpte = NULL; 2064 } 2065 2066 /* 2067 * This call to vtopte makes the assumption that we are 2068 * entering the page into the current pmap. In order to support 2069 * quick entry into any pmap, one would likely use pmap_pte_quick. 2070 * But that isn't as quick as vtopte. 2071 */ 2072 pte = vtopte(va); 2073 if (*pte) { 2074 if (mpte != NULL) { 2075 vm_page_lock_queues(); 2076 pmap_unwire_pte_hold(pmap, mpte); 2077 vm_page_unlock_queues(); 2078 } 2079 return 0; 2080 } 2081 2082 /* 2083 * Enter on the PV list if part of our managed memory. Note that we 2084 * raise IPL while manipulating pv_table since pmap_enter can be 2085 * called at interrupt time. 2086 */ 2087 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2088 pmap_insert_entry(pmap, va, mpte, m); 2089 2090 /* 2091 * Increment counters 2092 */ 2093 pmap->pm_stats.resident_count++; 2094 2095 pa = VM_PAGE_TO_PHYS(m); 2096 2097 /* 2098 * Now validate mapping with RO protection 2099 */ 2100 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2101 pte_store(pte, pa | PG_V | PG_U); 2102 else 2103 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2104 2105 return mpte; 2106} 2107 2108/* 2109 * Make a temporary mapping for a physical address. This is only intended 2110 * to be used for panic dumps. 2111 */ 2112void * 2113pmap_kenter_temporary(vm_offset_t pa, int i) 2114{ 2115 vm_offset_t va; 2116 2117 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2118 pmap_kenter(va, pa); 2119#ifndef I386_CPU 2120 invlpg(va); 2121#else 2122 invltlb(); 2123#endif 2124 return ((void *)crashdumpmap); 2125} 2126 2127/* 2128 * This code maps large physical mmap regions into the 2129 * processor address space. Note that some shortcuts 2130 * are taken, but the code works. 2131 */ 2132void 2133pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2134 vm_object_t object, vm_pindex_t pindex, 2135 vm_size_t size) 2136{ 2137 vm_page_t p; 2138 2139 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2140 KASSERT(object->type == OBJT_DEVICE, 2141 ("pmap_object_init_pt: non-device object")); 2142 if (pseflag && 2143 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2144 int i; 2145 vm_page_t m[1]; 2146 unsigned int ptepindex; 2147 int npdes; 2148 pd_entry_t ptepa; 2149 2150 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2151 return; 2152retry: 2153 p = vm_page_lookup(object, pindex); 2154 if (p != NULL) { 2155 vm_page_lock_queues(); 2156 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2157 goto retry; 2158 } else { 2159 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2160 if (p == NULL) 2161 return; 2162 m[0] = p; 2163 2164 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2165 vm_page_lock_queues(); 2166 vm_page_free(p); 2167 vm_page_unlock_queues(); 2168 return; 2169 } 2170 2171 p = vm_page_lookup(object, pindex); 2172 vm_page_lock_queues(); 2173 vm_page_wakeup(p); 2174 } 2175 vm_page_unlock_queues(); 2176 2177 ptepa = VM_PAGE_TO_PHYS(p); 2178 if (ptepa & (NBPDR - 1)) 2179 return; 2180 2181 p->valid = VM_PAGE_BITS_ALL; 2182 2183 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2184 npdes = size >> PDRSHIFT; 2185 for(i = 0; i < npdes; i++) { 2186 pde_store(&pmap->pm_pdir[ptepindex], 2187 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2188 ptepa += NBPDR; 2189 ptepindex += 1; 2190 } 2191 pmap_invalidate_all(pmap); 2192 } 2193} 2194 2195/* 2196 * Routine: pmap_change_wiring 2197 * Function: Change the wiring attribute for a map/virtual-address 2198 * pair. 2199 * In/out conditions: 2200 * The mapping must already exist in the pmap. 2201 */ 2202void 2203pmap_change_wiring(pmap, va, wired) 2204 register pmap_t pmap; 2205 vm_offset_t va; 2206 boolean_t wired; 2207{ 2208 register pt_entry_t *pte; 2209 2210 if (pmap == NULL) 2211 return; 2212 2213 pte = pmap_pte_quick(pmap, va); 2214 2215 if (wired && !pmap_pte_w(pte)) 2216 pmap->pm_stats.wired_count++; 2217 else if (!wired && pmap_pte_w(pte)) 2218 pmap->pm_stats.wired_count--; 2219 2220 /* 2221 * Wiring is not a hardware characteristic so there is no need to 2222 * invalidate TLB. 2223 */ 2224 pmap_pte_set_w(pte, wired); 2225} 2226 2227 2228 2229/* 2230 * Copy the range specified by src_addr/len 2231 * from the source map to the range dst_addr/len 2232 * in the destination map. 2233 * 2234 * This routine is only advisory and need not do anything. 2235 */ 2236 2237void 2238pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2239 vm_offset_t src_addr) 2240{ 2241 vm_offset_t addr; 2242 vm_offset_t end_addr = src_addr + len; 2243 vm_offset_t pdnxt; 2244 vm_page_t m; 2245 2246 if (dst_addr != src_addr) 2247 return; 2248 2249 if (!pmap_is_current(src_pmap)) 2250 return; 2251 2252 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2253 pt_entry_t *src_pte, *dst_pte; 2254 vm_page_t dstmpte, srcmpte; 2255 pd_entry_t srcptepaddr; 2256 unsigned ptepindex; 2257 2258 if (addr >= UPT_MIN_ADDRESS) 2259 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2260 2261 /* 2262 * Don't let optional prefaulting of pages make us go 2263 * way below the low water mark of free pages or way 2264 * above high water mark of used pv entries. 2265 */ 2266 if (cnt.v_free_count < cnt.v_free_reserved || 2267 pv_entry_count > pv_entry_high_water) 2268 break; 2269 2270 pdnxt = (addr + NBPDR) & ~PDRMASK; 2271 ptepindex = addr >> PDRSHIFT; 2272 2273 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2274 if (srcptepaddr == 0) 2275 continue; 2276 2277 if (srcptepaddr & PG_PS) { 2278 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2279 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2280 dst_pmap->pm_stats.resident_count += 2281 NBPDR / PAGE_SIZE; 2282 } 2283 continue; 2284 } 2285 2286 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2287 if (srcmpte->hold_count == 0 || (srcmpte->flags & PG_BUSY)) 2288 continue; 2289 2290 if (pdnxt > end_addr) 2291 pdnxt = end_addr; 2292 2293 src_pte = vtopte(addr); 2294 while (addr < pdnxt) { 2295 pt_entry_t ptetemp; 2296 ptetemp = *src_pte; 2297 /* 2298 * we only virtual copy managed pages 2299 */ 2300 if ((ptetemp & PG_MANAGED) != 0) { 2301 /* 2302 * We have to check after allocpte for the 2303 * pte still being around... allocpte can 2304 * block. 2305 */ 2306 dstmpte = pmap_allocpte(dst_pmap, addr); 2307 dst_pte = pmap_pte_quick(dst_pmap, addr); 2308 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2309 /* 2310 * Clear the modified and 2311 * accessed (referenced) bits 2312 * during the copy. 2313 */ 2314 m = PHYS_TO_VM_PAGE(ptetemp); 2315 *dst_pte = ptetemp & ~(PG_M | PG_A); 2316 dst_pmap->pm_stats.resident_count++; 2317 pmap_insert_entry(dst_pmap, addr, 2318 dstmpte, m); 2319 } else { 2320 vm_page_lock_queues(); 2321 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2322 vm_page_unlock_queues(); 2323 } 2324 if (dstmpte->hold_count >= srcmpte->hold_count) 2325 break; 2326 } 2327 addr += PAGE_SIZE; 2328 src_pte++; 2329 } 2330 } 2331} 2332 2333#ifdef SMP 2334 2335/* 2336 * pmap_zpi_switchout*() 2337 * 2338 * These functions allow us to avoid doing IPIs alltogether in certain 2339 * temporary page-mapping situations (page zeroing). Instead to deal 2340 * with being preempted and moved onto a different cpu we invalidate 2341 * the page when the scheduler switches us in. This does not occur 2342 * very often so we remain relatively optimal with very little effort. 2343 */ 2344static void 2345pmap_zpi_switchout12(void) 2346{ 2347 invlpg((u_int)CADDR1); 2348 invlpg((u_int)CADDR2); 2349} 2350 2351static void 2352pmap_zpi_switchout2(void) 2353{ 2354 invlpg((u_int)CADDR2); 2355} 2356 2357static void 2358pmap_zpi_switchout3(void) 2359{ 2360 invlpg((u_int)CADDR3); 2361} 2362 2363#endif 2364 2365static __inline void 2366pagezero(void *page) 2367{ 2368#if defined(I686_CPU) 2369 if (cpu_class == CPUCLASS_686) { 2370#if defined(CPU_ENABLE_SSE) 2371 if (cpu_feature & CPUID_SSE2) 2372 sse2_pagezero(page); 2373 else 2374#endif 2375 i686_pagezero(page); 2376 } else 2377#endif 2378 bzero(page, PAGE_SIZE); 2379} 2380 2381static __inline void 2382invlcaddr(void *caddr) 2383{ 2384#ifdef I386_CPU 2385 invltlb(); 2386#else 2387 invlpg((u_int)caddr); 2388#endif 2389} 2390 2391/* 2392 * pmap_zero_page zeros the specified hardware page by mapping 2393 * the page into KVM and using bzero to clear its contents. 2394 */ 2395void 2396pmap_zero_page(vm_page_t m) 2397{ 2398 2399 mtx_lock(&CMAPCADDR12_lock); 2400 if (*CMAP2) 2401 panic("pmap_zero_page: CMAP2 busy"); 2402#ifdef SMP 2403 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2; 2404#endif 2405 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2406#ifdef SMP 2407 invlpg((u_int)CADDR2); 2408#endif 2409 pagezero(CADDR2); 2410 *CMAP2 = 0; 2411 invlcaddr(CADDR2); 2412#ifdef SMP 2413 curthread->td_pcb->pcb_switchout = NULL; 2414#endif 2415 mtx_unlock(&CMAPCADDR12_lock); 2416} 2417 2418/* 2419 * pmap_zero_page_area zeros the specified hardware page by mapping 2420 * the page into KVM and using bzero to clear its contents. 2421 * 2422 * off and size may not cover an area beyond a single hardware page. 2423 */ 2424void 2425pmap_zero_page_area(vm_page_t m, int off, int size) 2426{ 2427 2428 mtx_lock(&CMAPCADDR12_lock); 2429 if (*CMAP2) 2430 panic("pmap_zero_page: CMAP2 busy"); 2431#ifdef SMP 2432 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2; 2433#endif 2434 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2435#ifdef SMP 2436 invlpg((u_int)CADDR2); 2437#endif 2438 if (off == 0 && size == PAGE_SIZE) 2439 pagezero(CADDR2); 2440 else 2441 bzero((char *)CADDR2 + off, size); 2442 *CMAP2 = 0; 2443 invlcaddr(CADDR2); 2444#ifdef SMP 2445 curthread->td_pcb->pcb_switchout = NULL; 2446#endif 2447 mtx_unlock(&CMAPCADDR12_lock); 2448} 2449 2450/* 2451 * pmap_zero_page_idle zeros the specified hardware page by mapping 2452 * the page into KVM and using bzero to clear its contents. This 2453 * is intended to be called from the vm_pagezero process only and 2454 * outside of Giant. 2455 */ 2456void 2457pmap_zero_page_idle(vm_page_t m) 2458{ 2459 2460 if (*CMAP3) 2461 panic("pmap_zero_page: CMAP3 busy"); 2462#ifdef SMP 2463 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout3; 2464#endif 2465 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2466#ifdef SMP 2467 invlpg((u_int)CADDR3); 2468#endif 2469 pagezero(CADDR3); 2470 *CMAP3 = 0; 2471 invlcaddr(CADDR3); 2472#ifdef SMP 2473 curthread->td_pcb->pcb_switchout = NULL; 2474#endif 2475} 2476 2477/* 2478 * pmap_copy_page copies the specified (machine independent) 2479 * page by mapping the page into virtual memory and using 2480 * bcopy to copy the page, one machine dependent page at a 2481 * time. 2482 */ 2483void 2484pmap_copy_page(vm_page_t src, vm_page_t dst) 2485{ 2486 2487 mtx_lock(&CMAPCADDR12_lock); 2488 if (*CMAP1) 2489 panic("pmap_copy_page: CMAP1 busy"); 2490 if (*CMAP2) 2491 panic("pmap_copy_page: CMAP2 busy"); 2492#ifdef SMP 2493 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout12; 2494#endif 2495 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2496 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2497#ifdef SMP 2498 invlpg((u_int)CADDR1); 2499 invlpg((u_int)CADDR2); 2500#endif 2501 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2502 *CMAP1 = 0; 2503 *CMAP2 = 0; 2504#ifdef I386_CPU 2505 invltlb(); 2506#else 2507 invlpg((u_int)CADDR1); 2508 invlpg((u_int)CADDR2); 2509#endif 2510#ifdef SMP 2511 curthread->td_pcb->pcb_switchout = NULL; 2512#endif 2513 mtx_unlock(&CMAPCADDR12_lock); 2514} 2515 2516/* 2517 * Returns true if the pmap's pv is one of the first 2518 * 16 pvs linked to from this page. This count may 2519 * be changed upwards or downwards in the future; it 2520 * is only necessary that true be returned for a small 2521 * subset of pmaps for proper page aging. 2522 */ 2523boolean_t 2524pmap_page_exists_quick(pmap, m) 2525 pmap_t pmap; 2526 vm_page_t m; 2527{ 2528 pv_entry_t pv; 2529 int loops = 0; 2530 int s; 2531 2532 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2533 return FALSE; 2534 2535 s = splvm(); 2536 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2537 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2538 if (pv->pv_pmap == pmap) { 2539 splx(s); 2540 return TRUE; 2541 } 2542 loops++; 2543 if (loops >= 16) 2544 break; 2545 } 2546 splx(s); 2547 return (FALSE); 2548} 2549 2550#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2551/* 2552 * Remove all pages from specified address space 2553 * this aids process exit speeds. Also, this code 2554 * is special cased for current process only, but 2555 * can have the more generic (and slightly slower) 2556 * mode enabled. This is much faster than pmap_remove 2557 * in the case of running down an entire address space. 2558 */ 2559void 2560pmap_remove_pages(pmap, sva, eva) 2561 pmap_t pmap; 2562 vm_offset_t sva, eva; 2563{ 2564 pt_entry_t *pte, tpte; 2565 vm_page_t m; 2566 pv_entry_t pv, npv; 2567 int s; 2568 2569#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2570 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2571 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2572 return; 2573 } 2574#endif 2575 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2576 s = splvm(); 2577 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2578 2579 if (pv->pv_va >= eva || pv->pv_va < sva) { 2580 npv = TAILQ_NEXT(pv, pv_plist); 2581 continue; 2582 } 2583 2584#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2585 pte = vtopte(pv->pv_va); 2586#else 2587 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2588#endif 2589 tpte = *pte; 2590 2591 if (tpte == 0) { 2592 printf("TPTE at %p IS ZERO @ VA %08x\n", 2593 pte, pv->pv_va); 2594 panic("bad pte"); 2595 } 2596 2597/* 2598 * We cannot remove wired pages from a process' mapping at this time 2599 */ 2600 if (tpte & PG_W) { 2601 npv = TAILQ_NEXT(pv, pv_plist); 2602 continue; 2603 } 2604 2605 m = PHYS_TO_VM_PAGE(tpte); 2606 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2607 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2608 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2609 2610 KASSERT(m < &vm_page_array[vm_page_array_size], 2611 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2612 2613 pv->pv_pmap->pm_stats.resident_count--; 2614 2615 pte_clear(pte); 2616 2617 /* 2618 * Update the vm_page_t clean and reference bits. 2619 */ 2620 if (tpte & PG_M) { 2621 vm_page_dirty(m); 2622 } 2623 2624 npv = TAILQ_NEXT(pv, pv_plist); 2625 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2626 2627 m->md.pv_list_count--; 2628 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2629 if (TAILQ_FIRST(&m->md.pv_list) == NULL) { 2630 vm_page_flag_clear(m, PG_WRITEABLE); 2631 } 2632 2633 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2634 free_pv_entry(pv); 2635 } 2636 splx(s); 2637 pmap_invalidate_all(pmap); 2638} 2639 2640/* 2641 * pmap_is_modified: 2642 * 2643 * Return whether or not the specified physical page was modified 2644 * in any physical maps. 2645 */ 2646boolean_t 2647pmap_is_modified(vm_page_t m) 2648{ 2649 pv_entry_t pv; 2650 pt_entry_t *pte; 2651 int s; 2652 2653 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2654 return FALSE; 2655 2656 s = splvm(); 2657 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2658 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2659 /* 2660 * if the bit being tested is the modified bit, then 2661 * mark clean_map and ptes as never 2662 * modified. 2663 */ 2664 if (!pmap_track_modified(pv->pv_va)) 2665 continue; 2666#if defined(PMAP_DIAGNOSTIC) 2667 if (!pv->pv_pmap) { 2668 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 2669 continue; 2670 } 2671#endif 2672 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2673 if (*pte & PG_M) { 2674 splx(s); 2675 return TRUE; 2676 } 2677 } 2678 splx(s); 2679 return (FALSE); 2680} 2681 2682/* 2683 * pmap_is_prefaultable: 2684 * 2685 * Return whether or not the specified virtual address is elgible 2686 * for prefault. 2687 */ 2688boolean_t 2689pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2690{ 2691 pt_entry_t *pte; 2692 2693 if ((*pmap_pde(pmap, addr)) == 0) 2694 return (FALSE); 2695 pte = vtopte(addr); 2696 if (*pte) 2697 return (FALSE); 2698 return (TRUE); 2699} 2700 2701/* 2702 * Clear the given bit in each of the given page's ptes. 2703 */ 2704static __inline void 2705pmap_clear_ptes(vm_page_t m, int bit) 2706{ 2707 register pv_entry_t pv; 2708 pt_entry_t pbits, *pte; 2709 int s; 2710 2711 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) || 2712 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2713 return; 2714 2715 s = splvm(); 2716 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2717 /* 2718 * Loop over all current mappings setting/clearing as appropos If 2719 * setting RO do we need to clear the VAC? 2720 */ 2721 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2722 /* 2723 * don't write protect pager mappings 2724 */ 2725 if (bit == PG_RW) { 2726 if (!pmap_track_modified(pv->pv_va)) 2727 continue; 2728 } 2729 2730#if defined(PMAP_DIAGNOSTIC) 2731 if (!pv->pv_pmap) { 2732 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 2733 continue; 2734 } 2735#endif 2736 2737 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2738 pbits = *pte; 2739 if (pbits & bit) { 2740 if (bit == PG_RW) { 2741 if (pbits & PG_M) { 2742 vm_page_dirty(m); 2743 } 2744 pte_store(pte, pbits & ~(PG_M|PG_RW)); 2745 } else { 2746 pte_store(pte, pbits & ~bit); 2747 } 2748 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2749 } 2750 } 2751 if (bit == PG_RW) 2752 vm_page_flag_clear(m, PG_WRITEABLE); 2753 splx(s); 2754} 2755 2756/* 2757 * pmap_page_protect: 2758 * 2759 * Lower the permission for all mappings to a given page. 2760 */ 2761void 2762pmap_page_protect(vm_page_t m, vm_prot_t prot) 2763{ 2764 if ((prot & VM_PROT_WRITE) == 0) { 2765 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2766 pmap_clear_ptes(m, PG_RW); 2767 } else { 2768 pmap_remove_all(m); 2769 } 2770 } 2771} 2772 2773/* 2774 * pmap_ts_referenced: 2775 * 2776 * Return a count of reference bits for a page, clearing those bits. 2777 * It is not necessary for every reference bit to be cleared, but it 2778 * is necessary that 0 only be returned when there are truly no 2779 * reference bits set. 2780 * 2781 * XXX: The exact number of bits to check and clear is a matter that 2782 * should be tested and standardized at some point in the future for 2783 * optimal aging of shared pages. 2784 */ 2785int 2786pmap_ts_referenced(vm_page_t m) 2787{ 2788 register pv_entry_t pv, pvf, pvn; 2789 pt_entry_t *pte; 2790 pt_entry_t v; 2791 int s; 2792 int rtval = 0; 2793 2794 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2795 return (rtval); 2796 2797 s = splvm(); 2798 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2799 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2800 2801 pvf = pv; 2802 2803 do { 2804 pvn = TAILQ_NEXT(pv, pv_list); 2805 2806 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2807 2808 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2809 2810 if (!pmap_track_modified(pv->pv_va)) 2811 continue; 2812 2813 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2814 2815 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2816 pte_store(pte, v & ~PG_A); 2817 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2818 2819 rtval++; 2820 if (rtval > 4) { 2821 break; 2822 } 2823 } 2824 } while ((pv = pvn) != NULL && pv != pvf); 2825 } 2826 splx(s); 2827 2828 return (rtval); 2829} 2830 2831/* 2832 * Clear the modify bits on the specified physical page. 2833 */ 2834void 2835pmap_clear_modify(vm_page_t m) 2836{ 2837 pmap_clear_ptes(m, PG_M); 2838} 2839 2840/* 2841 * pmap_clear_reference: 2842 * 2843 * Clear the reference bit on the specified physical page. 2844 */ 2845void 2846pmap_clear_reference(vm_page_t m) 2847{ 2848 pmap_clear_ptes(m, PG_A); 2849} 2850 2851/* 2852 * Miscellaneous support routines follow 2853 */ 2854 2855static void 2856i386_protection_init() 2857{ 2858 register int *kp, prot; 2859 2860 kp = protection_codes; 2861 for (prot = 0; prot < 8; prot++) { 2862 switch (prot) { 2863 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE: 2864 /* 2865 * Read access is also 0. There isn't any execute bit, 2866 * so just make it readable. 2867 */ 2868 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE: 2869 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE: 2870 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE: 2871 *kp++ = 0; 2872 break; 2873 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE: 2874 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE: 2875 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE: 2876 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE: 2877 *kp++ = PG_RW; 2878 break; 2879 } 2880 } 2881} 2882 2883/* 2884 * Map a set of physical memory pages into the kernel virtual 2885 * address space. Return a pointer to where it is mapped. This 2886 * routine is intended to be used for mapping device memory, 2887 * NOT real memory. 2888 */ 2889void * 2890pmap_mapdev(pa, size) 2891 vm_paddr_t pa; 2892 vm_size_t size; 2893{ 2894 vm_offset_t va, tmpva, offset; 2895 2896 offset = pa & PAGE_MASK; 2897 size = roundup(offset + size, PAGE_SIZE); 2898 pa = pa & PG_FRAME; 2899 2900 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2901 va = KERNBASE + pa; 2902 else 2903 va = kmem_alloc_nofault(kernel_map, size); 2904 if (!va) 2905 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2906 2907 for (tmpva = va; size > 0; ) { 2908 pmap_kenter(tmpva, pa); 2909 size -= PAGE_SIZE; 2910 tmpva += PAGE_SIZE; 2911 pa += PAGE_SIZE; 2912 } 2913 pmap_invalidate_range(kernel_pmap, va, tmpva); 2914 return ((void *)(va + offset)); 2915} 2916 2917void 2918pmap_unmapdev(va, size) 2919 vm_offset_t va; 2920 vm_size_t size; 2921{ 2922 vm_offset_t base, offset, tmpva; 2923 2924 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2925 return; 2926 base = va & PG_FRAME; 2927 offset = va & PAGE_MASK; 2928 size = roundup(offset + size, PAGE_SIZE); 2929 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2930 pmap_kremove(tmpva); 2931 pmap_invalidate_range(kernel_pmap, va, tmpva); 2932 kmem_free(kernel_map, base, size); 2933} 2934 2935/* 2936 * perform the pmap work for mincore 2937 */ 2938int 2939pmap_mincore(pmap, addr) 2940 pmap_t pmap; 2941 vm_offset_t addr; 2942{ 2943 pt_entry_t *ptep, pte; 2944 vm_page_t m; 2945 int val = 0; 2946 2947 ptep = pmap_pte_quick(pmap, addr); 2948 if (ptep == 0) { 2949 return 0; 2950 } 2951 2952 if ((pte = *ptep) != 0) { 2953 vm_paddr_t pa; 2954 2955 val = MINCORE_INCORE; 2956 if ((pte & PG_MANAGED) == 0) 2957 return val; 2958 2959 pa = pte & PG_FRAME; 2960 2961 m = PHYS_TO_VM_PAGE(pa); 2962 2963 /* 2964 * Modified by us 2965 */ 2966 if (pte & PG_M) 2967 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2968 else { 2969 /* 2970 * Modified by someone else 2971 */ 2972 vm_page_lock_queues(); 2973 if (m->dirty || pmap_is_modified(m)) 2974 val |= MINCORE_MODIFIED_OTHER; 2975 vm_page_unlock_queues(); 2976 } 2977 /* 2978 * Referenced by us 2979 */ 2980 if (pte & PG_A) 2981 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2982 else { 2983 /* 2984 * Referenced by someone else 2985 */ 2986 vm_page_lock_queues(); 2987 if ((m->flags & PG_REFERENCED) || 2988 pmap_ts_referenced(m)) { 2989 val |= MINCORE_REFERENCED_OTHER; 2990 vm_page_flag_set(m, PG_REFERENCED); 2991 } 2992 vm_page_unlock_queues(); 2993 } 2994 } 2995 return val; 2996} 2997 2998void 2999pmap_activate(struct thread *td) 3000{ 3001 struct proc *p = td->td_proc; 3002 pmap_t pmap, oldpmap; 3003 u_int32_t cr3; 3004 3005 critical_enter(); 3006 pmap = vmspace_pmap(td->td_proc->p_vmspace); 3007 oldpmap = PCPU_GET(curpmap); 3008#if defined(SMP) 3009 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 3010 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 3011#else 3012 oldpmap->pm_active &= ~1; 3013 pmap->pm_active |= 1; 3014#endif 3015#ifdef PAE 3016 cr3 = vtophys(pmap->pm_pdpt); 3017#else 3018 cr3 = vtophys(pmap->pm_pdir); 3019#endif 3020 /* XXXKSE this is wrong. 3021 * pmap_activate is for the current thread on the current cpu 3022 */ 3023 if (p->p_flag & P_SA) { 3024 /* Make sure all other cr3 entries are updated. */ 3025 /* what if they are running? XXXKSE (maybe abort them) */ 3026 FOREACH_THREAD_IN_PROC(p, td) { 3027 td->td_pcb->pcb_cr3 = cr3; 3028 } 3029 } else { 3030 td->td_pcb->pcb_cr3 = cr3; 3031 } 3032 load_cr3(cr3); 3033 PCPU_SET(curpmap, pmap); 3034 critical_exit(); 3035} 3036 3037vm_offset_t 3038pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3039{ 3040 3041 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3042 return addr; 3043 } 3044 3045 addr = (addr + PDRMASK) & ~PDRMASK; 3046 return addr; 3047} 3048 3049 3050#if defined(PMAP_DEBUG) 3051pmap_pid_dump(int pid) 3052{ 3053 pmap_t pmap; 3054 struct proc *p; 3055 int npte = 0; 3056 int index; 3057 3058 sx_slock(&allproc_lock); 3059 LIST_FOREACH(p, &allproc, p_list) { 3060 if (p->p_pid != pid) 3061 continue; 3062 3063 if (p->p_vmspace) { 3064 int i,j; 3065 index = 0; 3066 pmap = vmspace_pmap(p->p_vmspace); 3067 for (i = 0; i < NPDEPTD; i++) { 3068 pd_entry_t *pde; 3069 pt_entry_t *pte; 3070 vm_offset_t base = i << PDRSHIFT; 3071 3072 pde = &pmap->pm_pdir[i]; 3073 if (pde && pmap_pde_v(pde)) { 3074 for (j = 0; j < NPTEPG; j++) { 3075 vm_offset_t va = base + (j << PAGE_SHIFT); 3076 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3077 if (index) { 3078 index = 0; 3079 printf("\n"); 3080 } 3081 sx_sunlock(&allproc_lock); 3082 return npte; 3083 } 3084 pte = pmap_pte_quick(pmap, va); 3085 if (pte && pmap_pte_v(pte)) { 3086 pt_entry_t pa; 3087 vm_page_t m; 3088 pa = *pte; 3089 m = PHYS_TO_VM_PAGE(pa); 3090 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3091 va, pa, m->hold_count, m->wire_count, m->flags); 3092 npte++; 3093 index++; 3094 if (index >= 2) { 3095 index = 0; 3096 printf("\n"); 3097 } else { 3098 printf(" "); 3099 } 3100 } 3101 } 3102 } 3103 } 3104 } 3105 } 3106 sx_sunlock(&allproc_lock); 3107 return npte; 3108} 3109#endif 3110 3111#if defined(DEBUG) 3112 3113static void pads(pmap_t pm); 3114void pmap_pvdump(vm_offset_t pa); 3115 3116/* print address space of pmap*/ 3117static void 3118pads(pm) 3119 pmap_t pm; 3120{ 3121 int i, j; 3122 vm_paddr_t va; 3123 pt_entry_t *ptep; 3124 3125 if (pm == kernel_pmap) 3126 return; 3127 for (i = 0; i < NPDEPTD; i++) 3128 if (pm->pm_pdir[i]) 3129 for (j = 0; j < NPTEPG; j++) { 3130 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3131 if (pm == kernel_pmap && va < KERNBASE) 3132 continue; 3133 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3134 continue; 3135 ptep = pmap_pte_quick(pm, va); 3136 if (pmap_pte_v(ptep)) 3137 printf("%x:%x ", va, *ptep); 3138 }; 3139 3140} 3141 3142void 3143pmap_pvdump(pa) 3144 vm_paddr_t pa; 3145{ 3146 pv_entry_t pv; 3147 vm_page_t m; 3148 3149 printf("pa %x", pa); 3150 m = PHYS_TO_VM_PAGE(pa); 3151 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3152 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3153 pads(pv->pv_pmap); 3154 } 3155 printf(" "); 3156} 3157#endif 3158