pmap.c revision 125308
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 125308 2004-02-01 20:14:00Z 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#define PMAP_KEEP_PDIRS 153#ifndef PMAP_SHPGPERPROC 154#define PMAP_SHPGPERPROC 200 155#endif 156 157#if defined(DIAGNOSTIC) 158#define PMAP_DIAGNOSTIC 159#endif 160 161#define MINPV 2048 162 163#if !defined(PMAP_DIAGNOSTIC) 164#define PMAP_INLINE __inline 165#else 166#define PMAP_INLINE 167#endif 168 169/* 170 * Get PDEs and PTEs for user/kernel address space 171 */ 172#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 173#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 174 175#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 176#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 177#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 178#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 179#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 180 181#define pmap_pte_set_w(pte, v) ((v)?(*(int *)pte |= PG_W):(*(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_start; /* PA of first available physical page */ 193vm_paddr_t avail_end; /* PA of last available physical page */ 194vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 195vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 196static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 197int pgeflag = 0; /* PG_G or-in */ 198int pseflag = 0; /* PG_PS or-in */ 199 200static int nkpt; 201vm_offset_t kernel_vm_end; 202extern u_int32_t KERNend; 203 204#ifdef PAE 205static uma_zone_t pdptzone; 206#endif 207 208/* 209 * Data for the pv entry allocation mechanism 210 */ 211static uma_zone_t pvzone; 212static struct vm_object pvzone_obj; 213static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 214int pmap_pagedaemon_waken; 215 216/* 217 * All those kernel PT submaps that BSD is so fond of 218 */ 219pt_entry_t *CMAP1 = 0; 220static pt_entry_t *CMAP2, *CMAP3, *ptmmap; 221caddr_t CADDR1 = 0, ptvmmap = 0; 222static caddr_t CADDR2, CADDR3; 223static struct mtx CMAPCADDR12_lock; 224static pt_entry_t *msgbufmap; 225struct msgbuf *msgbufp = 0; 226 227/* 228 * Crashdump maps. 229 */ 230static pt_entry_t *pt_crashdumpmap; 231static caddr_t crashdumpmap; 232 233#ifdef SMP 234extern pt_entry_t *SMPpt; 235#endif 236static pt_entry_t *PMAP1 = 0, *PMAP2; 237static pt_entry_t *PADDR1 = 0, *PADDR2; 238#ifdef SMP 239static int PMAP1cpu; 240static int PMAP1changedcpu; 241SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD, 242 &PMAP1changedcpu, 0, 243 "Number of times pmap_pte_quick changed CPU with same PMAP1"); 244#endif 245static int PMAP1changed; 246SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD, 247 &PMAP1changed, 0, 248 "Number of times pmap_pte_quick changed PMAP1"); 249static int PMAP1unchanged; 250SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD, 251 &PMAP1unchanged, 0, 252 "Number of times pmap_pte_quick didn't change PMAP1"); 253 254static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 255static pv_entry_t get_pv_entry(void); 256static void pmap_clear_ptes(vm_page_t m, int bit) 257 __always_inline; 258 259static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 260static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 261static int pmap_remove_entry(struct pmap *pmap, vm_page_t m, 262 vm_offset_t va); 263static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, 264 vm_page_t mpte, vm_page_t m); 265 266static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va); 267 268static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex); 269static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 270static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t); 271static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 272#ifdef PAE 273static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 274#endif 275 276CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 277CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 278 279/* 280 * Move the kernel virtual free pointer to the next 281 * 4MB. This is used to help improve performance 282 * by using a large (4MB) page for much of the kernel 283 * (.text, .data, .bss) 284 */ 285static vm_offset_t 286pmap_kmem_choose(vm_offset_t addr) 287{ 288 vm_offset_t newaddr = addr; 289 290#ifndef DISABLE_PSE 291 if (cpu_feature & CPUID_PSE) 292 newaddr = (addr + PDRMASK) & ~PDRMASK; 293#endif 294 return newaddr; 295} 296 297/* 298 * Bootstrap the system enough to run with virtual memory. 299 * 300 * On the i386 this is called after mapping has already been enabled 301 * and just syncs the pmap module with what has already been done. 302 * [We can't call it easily with mapping off since the kernel is not 303 * mapped with PA == VA, hence we would have to relocate every address 304 * from the linked base (virtual) address "KERNBASE" to the actual 305 * (physical) address starting relative to 0] 306 */ 307void 308pmap_bootstrap(firstaddr, loadaddr) 309 vm_paddr_t firstaddr; 310 vm_paddr_t loadaddr; 311{ 312 vm_offset_t va; 313 pt_entry_t *pte; 314 int i; 315 316 avail_start = firstaddr; 317 318 /* 319 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 320 * large. It should instead be correctly calculated in locore.s and 321 * not based on 'first' (which is a physical address, not a virtual 322 * address, for the start of unused physical memory). The kernel 323 * page tables are NOT double mapped and thus should not be included 324 * in this calculation. 325 */ 326 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 327 virtual_avail = pmap_kmem_choose(virtual_avail); 328 329 virtual_end = VM_MAX_KERNEL_ADDRESS; 330 331 /* 332 * Initialize the kernel pmap (which is statically allocated). 333 */ 334 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 335#ifdef PAE 336 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 337#endif 338 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 339 TAILQ_INIT(&kernel_pmap->pm_pvlist); 340 LIST_INIT(&allpmaps); 341#ifdef SMP 342 mtx_init(&lazypmap_lock, "lazypmap", NULL, MTX_SPIN); 343#endif 344 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 345 mtx_lock_spin(&allpmaps_lock); 346 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 347 mtx_unlock_spin(&allpmaps_lock); 348 nkpt = NKPT; 349 350 /* 351 * Reserve some special page table entries/VA space for temporary 352 * mapping of pages. 353 */ 354#define SYSMAP(c, p, v, n) \ 355 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 356 357 va = virtual_avail; 358 pte = vtopte(va); 359 360 /* 361 * CMAP1/CMAP2 are used for zeroing and copying pages. 362 * CMAP3 is used for the idle process page zeroing. 363 */ 364 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 365 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 366 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 367 *CMAP3 = 0; 368 369 mtx_init(&CMAPCADDR12_lock, "CMAPCADDR12", NULL, MTX_DEF); 370 371 /* 372 * Crashdump maps. 373 */ 374 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 375 376 /* 377 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 378 * XXX ptmmap is not used. 379 */ 380 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 381 382 /* 383 * msgbufp is used to map the system message buffer. 384 * XXX msgbufmap is not used. 385 */ 386 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 387 atop(round_page(MSGBUF_SIZE))) 388 389 /* 390 * ptemap is used for pmap_pte_quick 391 */ 392 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 393 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1); 394 395 virtual_avail = va; 396 397 *CMAP1 = *CMAP2 = 0; 398 for (i = 0; i < NKPT; i++) 399 PTD[i] = 0; 400 401 /* Turn on PG_G on kernel page(s) */ 402 pmap_set_pg(); 403} 404 405/* 406 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 407 */ 408void 409pmap_set_pg(void) 410{ 411 pd_entry_t pdir; 412 pt_entry_t *pte; 413 vm_offset_t va, endva; 414 int i; 415 416 if (pgeflag == 0) 417 return; 418 419 i = KERNLOAD/NBPDR; 420 endva = KERNBASE + KERNend; 421 422 if (pseflag) { 423 va = KERNBASE + KERNLOAD; 424 while (va < endva) { 425 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 426 pdir |= pgeflag; 427 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 428 invltlb(); /* Play it safe, invltlb() every time */ 429 i++; 430 va += NBPDR; 431 } 432 } else { 433 va = (vm_offset_t)btext; 434 while (va < endva) { 435 pte = vtopte(va); 436 if (*pte) 437 *pte |= pgeflag; 438 invltlb(); /* Play it safe, invltlb() every time */ 439 va += PAGE_SIZE; 440 } 441 } 442} 443 444#ifdef PAE 445static void * 446pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 447{ 448 *flags = UMA_SLAB_PRIV; 449 return (contigmalloc(PAGE_SIZE, NULL, 0, 0x0ULL, 0xffffffffULL, 1, 0)); 450} 451#endif 452 453/* 454 * Initialize the pmap module. 455 * Called by vm_init, to initialize any structures that the pmap 456 * system needs to map virtual memory. 457 * pmap_init has been enhanced to support in a fairly consistant 458 * way, discontiguous physical memory. 459 */ 460void 461pmap_init(phys_start, phys_end) 462 vm_paddr_t phys_start, phys_end; 463{ 464 int i; 465 466 /* 467 * Allocate memory for random pmap data structures. Includes the 468 * pv_head_table. 469 */ 470 471 for(i = 0; i < vm_page_array_size; i++) { 472 vm_page_t m; 473 474 m = &vm_page_array[i]; 475 TAILQ_INIT(&m->md.pv_list); 476 m->md.pv_list_count = 0; 477 } 478 479 /* 480 * init the pv free list 481 */ 482 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL, 483 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 484 uma_prealloc(pvzone, MINPV); 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 __inline int 735pmap_is_current(pmap_t pmap) 736{ 737 738 return (pmap == kernel_pmap || 739 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 740 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 741} 742 743/* 744 * If the given pmap is not the current pmap, Giant must be held. 745 */ 746pt_entry_t * 747pmap_pte(pmap_t pmap, vm_offset_t va) 748{ 749 pd_entry_t newpf; 750 pd_entry_t *pde; 751 752 pde = pmap_pde(pmap, va); 753 if (*pde & PG_PS) 754 return (pde); 755 if (*pde != 0) { 756 /* are we current address space or kernel? */ 757 if (pmap_is_current(pmap)) 758 return (vtopte(va)); 759 GIANT_REQUIRED; 760 newpf = *pde & PG_FRAME; 761 if ((*PMAP2 & PG_FRAME) != newpf) { 762 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; 763 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 764 } 765 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); 766 } 767 return (0); 768} 769 770static __inline void 771invlcaddr(void *caddr) 772{ 773#ifdef I386_CPU 774 invltlb(); 775#else 776 invlpg((u_int)caddr); 777#endif 778} 779 780/* 781 * Super fast pmap_pte routine best used when scanning 782 * the pv lists. This eliminates many coarse-grained 783 * invltlb calls. Note that many of the pv list 784 * scans are across different pmaps. It is very wasteful 785 * to do an entire invltlb for checking a single mapping. 786 * 787 * If the given pmap is not the current pmap, vm_page_queue_mtx 788 * must be held and curthread pinned to a CPU. 789 */ 790static pt_entry_t * 791pmap_pte_quick(pmap_t pmap, vm_offset_t va) 792{ 793 pd_entry_t newpf; 794 pd_entry_t *pde; 795 796 pde = pmap_pde(pmap, va); 797 if (*pde & PG_PS) 798 return (pde); 799 if (*pde != 0) { 800 /* are we current address space or kernel? */ 801 if (pmap_is_current(pmap)) 802 return (vtopte(va)); 803 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 804 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 805 newpf = *pde & PG_FRAME; 806 if ((*PMAP1 & PG_FRAME) != newpf) { 807 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; 808#ifdef SMP 809 PMAP1cpu = PCPU_GET(cpuid); 810#endif 811 invlcaddr(PADDR1); 812 PMAP1changed++; 813 } else 814#ifdef SMP 815 if (PMAP1cpu != PCPU_GET(cpuid)) { 816 PMAP1cpu = PCPU_GET(cpuid); 817 invlcaddr(PADDR1); 818 PMAP1changedcpu++; 819 } else 820#endif 821 PMAP1unchanged++; 822 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); 823 } 824 return (0); 825} 826 827/* 828 * Routine: pmap_extract 829 * Function: 830 * Extract the physical page address associated 831 * with the given map/virtual_address pair. 832 */ 833vm_paddr_t 834pmap_extract(pmap, va) 835 register pmap_t pmap; 836 vm_offset_t va; 837{ 838 vm_paddr_t rtval; 839 pt_entry_t *pte; 840 pd_entry_t pde; 841 842 if (pmap == 0) 843 return 0; 844 pde = pmap->pm_pdir[va >> PDRSHIFT]; 845 if (pde != 0) { 846 if ((pde & PG_PS) != 0) { 847 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 848 return rtval; 849 } 850 pte = pmap_pte(pmap, va); 851 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK)); 852 return rtval; 853 } 854 return 0; 855 856} 857 858/* 859 * Routine: pmap_extract_and_hold 860 * Function: 861 * Atomically extract and hold the physical page 862 * with the given pmap and virtual address pair 863 * if that mapping permits the given protection. 864 */ 865vm_page_t 866pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 867{ 868 vm_paddr_t pa; 869 vm_page_t m; 870 871 m = NULL; 872 mtx_lock(&Giant); 873 if ((pa = pmap_extract(pmap, va)) != 0) { 874 m = PHYS_TO_VM_PAGE(pa); 875 vm_page_lock_queues(); 876 vm_page_hold(m); 877 vm_page_unlock_queues(); 878 } 879 mtx_unlock(&Giant); 880 return (m); 881} 882 883/*************************************************** 884 * Low level mapping routines..... 885 ***************************************************/ 886 887/* 888 * Add a wired page to the kva. 889 * Note: not SMP coherent. 890 */ 891PMAP_INLINE void 892pmap_kenter(vm_offset_t va, vm_paddr_t pa) 893{ 894 pt_entry_t *pte; 895 896 pte = vtopte(va); 897 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 898} 899 900/* 901 * Remove a page from the kernel pagetables. 902 * Note: not SMP coherent. 903 */ 904PMAP_INLINE void 905pmap_kremove(vm_offset_t va) 906{ 907 pt_entry_t *pte; 908 909 pte = vtopte(va); 910 pte_clear(pte); 911} 912 913/* 914 * Used to map a range of physical addresses into kernel 915 * virtual address space. 916 * 917 * The value passed in '*virt' is a suggested virtual address for 918 * the mapping. Architectures which can support a direct-mapped 919 * physical to virtual region can return the appropriate address 920 * within that region, leaving '*virt' unchanged. Other 921 * architectures should map the pages starting at '*virt' and 922 * update '*virt' with the first usable address after the mapped 923 * region. 924 */ 925vm_offset_t 926pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 927{ 928 vm_offset_t va, sva; 929 930 va = sva = *virt; 931 while (start < end) { 932 pmap_kenter(va, start); 933 va += PAGE_SIZE; 934 start += PAGE_SIZE; 935 } 936 pmap_invalidate_range(kernel_pmap, sva, va); 937 *virt = va; 938 return (sva); 939} 940 941 942/* 943 * Add a list of wired pages to the kva 944 * this routine is only used for temporary 945 * kernel mappings that do not need to have 946 * page modification or references recorded. 947 * Note that old mappings are simply written 948 * over. The page *must* be wired. 949 * Note: SMP coherent. Uses a ranged shootdown IPI. 950 */ 951void 952pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 953{ 954 vm_offset_t va; 955 956 va = sva; 957 while (count-- > 0) { 958 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 959 va += PAGE_SIZE; 960 m++; 961 } 962 pmap_invalidate_range(kernel_pmap, sva, va); 963} 964 965/* 966 * This routine tears out page mappings from the 967 * kernel -- it is meant only for temporary mappings. 968 * Note: SMP coherent. Uses a ranged shootdown IPI. 969 */ 970void 971pmap_qremove(vm_offset_t sva, int count) 972{ 973 vm_offset_t va; 974 975 va = sva; 976 while (count-- > 0) { 977 pmap_kremove(va); 978 va += PAGE_SIZE; 979 } 980 pmap_invalidate_range(kernel_pmap, sva, va); 981} 982 983/*************************************************** 984 * Page table page management routines..... 985 ***************************************************/ 986 987/* 988 * This routine unholds page table pages, and if the hold count 989 * drops to zero, then it decrements the wire count. 990 */ 991static int 992_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 993{ 994 995 while (vm_page_sleep_if_busy(m, FALSE, "pmuwpt")) 996 vm_page_lock_queues(); 997 998 if (m->hold_count == 0) { 999 vm_offset_t pteva; 1000 /* 1001 * unmap the page table page 1002 */ 1003 pmap->pm_pdir[m->pindex] = 0; 1004 --pmap->pm_stats.resident_count; 1005 /* 1006 * We never unwire a kernel page table page, making a 1007 * check for the kernel_pmap unnecessary. 1008 */ 1009 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)) { 1010 /* 1011 * Do an invltlb to make the invalidated mapping 1012 * take effect immediately. 1013 */ 1014 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1015 pmap_invalidate_page(pmap, pteva); 1016 } 1017 1018 /* 1019 * If the page is finally unwired, simply free it. 1020 */ 1021 --m->wire_count; 1022 if (m->wire_count == 0) { 1023 vm_page_busy(m); 1024 vm_page_free_zero(m); 1025 atomic_subtract_int(&cnt.v_wire_count, 1); 1026 } 1027 return 1; 1028 } 1029 return 0; 1030} 1031 1032static PMAP_INLINE int 1033pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1034{ 1035 vm_page_unhold(m); 1036 if (m->hold_count == 0) 1037 return _pmap_unwire_pte_hold(pmap, m); 1038 else 1039 return 0; 1040} 1041 1042/* 1043 * After removing a page table entry, this routine is used to 1044 * conditionally free the page, and manage the hold/wire counts. 1045 */ 1046static int 1047pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 1048{ 1049 1050 if (va >= VM_MAXUSER_ADDRESS) 1051 return 0; 1052 1053 return pmap_unwire_pte_hold(pmap, mpte); 1054} 1055 1056void 1057pmap_pinit0(pmap) 1058 struct pmap *pmap; 1059{ 1060 1061 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1062#ifdef PAE 1063 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1064#endif 1065 pmap->pm_active = 0; 1066 PCPU_SET(curpmap, pmap); 1067 TAILQ_INIT(&pmap->pm_pvlist); 1068 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1069 mtx_lock_spin(&allpmaps_lock); 1070 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1071 mtx_unlock_spin(&allpmaps_lock); 1072} 1073 1074/* 1075 * Initialize a preallocated and zeroed pmap structure, 1076 * such as one in a vmspace structure. 1077 */ 1078void 1079pmap_pinit(pmap) 1080 register struct pmap *pmap; 1081{ 1082 vm_page_t m, ptdpg[NPGPTD]; 1083 vm_paddr_t pa; 1084 static int color; 1085 int i; 1086 1087 /* 1088 * No need to allocate page table space yet but we do need a valid 1089 * page directory table. 1090 */ 1091 if (pmap->pm_pdir == NULL) { 1092 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1093 NBPTD); 1094#ifdef PAE 1095 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1096 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1097 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1098 ("pmap_pinit: pdpt misaligned")); 1099 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1100 ("pmap_pinit: pdpt above 4g")); 1101#endif 1102 } 1103 1104 /* 1105 * allocate the page directory page(s) 1106 */ 1107 for (i = 0; i < NPGPTD;) { 1108 m = vm_page_alloc(NULL, color++, 1109 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1110 VM_ALLOC_ZERO); 1111 if (m == NULL) 1112 VM_WAIT; 1113 else { 1114 ptdpg[i++] = m; 1115 } 1116 } 1117 1118 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1119 1120 for (i = 0; i < NPGPTD; i++) { 1121 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1122 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1123 } 1124 1125 mtx_lock_spin(&allpmaps_lock); 1126 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1127 mtx_unlock_spin(&allpmaps_lock); 1128 /* Wire in kernel global address entries. */ 1129 /* XXX copies current process, does not fill in MPPTDI */ 1130 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1131#ifdef SMP 1132 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1133#endif 1134 1135 /* install self-referential address mapping entry(s) */ 1136 for (i = 0; i < NPGPTD; i++) { 1137 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1138 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1139#ifdef PAE 1140 pmap->pm_pdpt[i] = pa | PG_V; 1141#endif 1142 } 1143 1144 pmap->pm_active = 0; 1145 TAILQ_INIT(&pmap->pm_pvlist); 1146 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1147} 1148 1149/* 1150 * Wire in kernel global address entries. To avoid a race condition 1151 * between pmap initialization and pmap_growkernel, this procedure 1152 * should be called after the vmspace is attached to the process 1153 * but before this pmap is activated. 1154 */ 1155void 1156pmap_pinit2(pmap) 1157 struct pmap *pmap; 1158{ 1159 /* XXX: Remove this stub when no longer called */ 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 VM_WAIT; 1180 /* 1181 * Indicate the need to retry. While waiting, the page table 1182 * page may have been allocated. 1183 */ 1184 return (NULL); 1185 } 1186 if ((m->flags & PG_ZERO) == 0) 1187 pmap_zero_page(m); 1188 1189 KASSERT(m->queue == PQ_NONE, 1190 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1191 1192 /* 1193 * Increment the hold count for the page table page 1194 * (denoting a new mapping.) 1195 */ 1196 m->hold_count++; 1197 1198 /* 1199 * Map the pagetable page into the process address space, if 1200 * it isn't already there. 1201 */ 1202 1203 pmap->pm_stats.resident_count++; 1204 1205 ptepa = VM_PAGE_TO_PHYS(m); 1206 pmap->pm_pdir[ptepindex] = 1207 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1208 1209 vm_page_lock_queues(); 1210 vm_page_wakeup(m); 1211 vm_page_unlock_queues(); 1212 1213 return m; 1214} 1215 1216static vm_page_t 1217pmap_allocpte(pmap_t pmap, vm_offset_t va) 1218{ 1219 unsigned ptepindex; 1220 pd_entry_t ptepa; 1221 vm_page_t m; 1222 1223 /* 1224 * Calculate pagetable page index 1225 */ 1226 ptepindex = va >> PDRSHIFT; 1227retry: 1228 /* 1229 * Get the page directory entry 1230 */ 1231 ptepa = pmap->pm_pdir[ptepindex]; 1232 1233 /* 1234 * This supports switching from a 4MB page to a 1235 * normal 4K page. 1236 */ 1237 if (ptepa & PG_PS) { 1238 pmap->pm_pdir[ptepindex] = 0; 1239 ptepa = 0; 1240 pmap_invalidate_all(kernel_pmap); 1241 } 1242 1243 /* 1244 * If the page table page is mapped, we just increment the 1245 * hold count, and activate it. 1246 */ 1247 if (ptepa) { 1248 m = PHYS_TO_VM_PAGE(ptepa); 1249 m->hold_count++; 1250 } else { 1251 /* 1252 * Here if the pte page isn't mapped, or if it has 1253 * been deallocated. 1254 */ 1255 m = _pmap_allocpte(pmap, ptepindex); 1256 if (m == NULL) 1257 goto retry; 1258 } 1259 return (m); 1260} 1261 1262 1263/*************************************************** 1264* Pmap allocation/deallocation routines. 1265 ***************************************************/ 1266 1267#ifdef SMP 1268/* 1269 * Deal with a SMP shootdown of other users of the pmap that we are 1270 * trying to dispose of. This can be a bit hairy. 1271 */ 1272static u_int *lazymask; 1273static u_int lazyptd; 1274static volatile u_int lazywait; 1275 1276void pmap_lazyfix_action(void); 1277 1278void 1279pmap_lazyfix_action(void) 1280{ 1281 u_int mymask = PCPU_GET(cpumask); 1282 1283 if (rcr3() == lazyptd) 1284 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1285 atomic_clear_int(lazymask, mymask); 1286 atomic_store_rel_int(&lazywait, 1); 1287} 1288 1289static void 1290pmap_lazyfix_self(u_int mymask) 1291{ 1292 1293 if (rcr3() == lazyptd) 1294 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1295 atomic_clear_int(lazymask, mymask); 1296} 1297 1298 1299static void 1300pmap_lazyfix(pmap_t pmap) 1301{ 1302 u_int mymask = PCPU_GET(cpumask); 1303 u_int mask; 1304 register u_int spins; 1305 1306 while ((mask = pmap->pm_active) != 0) { 1307 spins = 50000000; 1308 mask = mask & -mask; /* Find least significant set bit */ 1309 mtx_lock_spin(&lazypmap_lock); 1310#ifdef PAE 1311 lazyptd = vtophys(pmap->pm_pdpt); 1312#else 1313 lazyptd = vtophys(pmap->pm_pdir); 1314#endif 1315 if (mask == mymask) { 1316 lazymask = &pmap->pm_active; 1317 pmap_lazyfix_self(mymask); 1318 } else { 1319 atomic_store_rel_int((u_int *)&lazymask, 1320 (u_int)&pmap->pm_active); 1321 atomic_store_rel_int(&lazywait, 0); 1322 ipi_selected(mask, IPI_LAZYPMAP); 1323 while (lazywait == 0) { 1324 ia32_pause(); 1325 if (--spins == 0) 1326 break; 1327 } 1328 } 1329 mtx_unlock_spin(&lazypmap_lock); 1330 if (spins == 0) 1331 printf("pmap_lazyfix: spun for 50000000\n"); 1332 } 1333} 1334 1335#else /* SMP */ 1336 1337/* 1338 * Cleaning up on uniprocessor is easy. For various reasons, we're 1339 * unlikely to have to even execute this code, including the fact 1340 * that the cleanup is deferred until the parent does a wait(2), which 1341 * means that another userland process has run. 1342 */ 1343static void 1344pmap_lazyfix(pmap_t pmap) 1345{ 1346 u_int cr3; 1347 1348 cr3 = vtophys(pmap->pm_pdir); 1349 if (cr3 == rcr3()) { 1350 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1351 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1352 } 1353} 1354#endif /* SMP */ 1355 1356/* 1357 * Release any resources held by the given physical map. 1358 * Called when a pmap initialized by pmap_pinit is being released. 1359 * Should only be called if the map contains no valid mappings. 1360 */ 1361void 1362pmap_release(pmap_t pmap) 1363{ 1364 vm_page_t m, ptdpg[NPGPTD]; 1365 int i; 1366 1367 KASSERT(pmap->pm_stats.resident_count == 0, 1368 ("pmap_release: pmap resident count %ld != 0", 1369 pmap->pm_stats.resident_count)); 1370 1371 pmap_lazyfix(pmap); 1372 mtx_lock_spin(&allpmaps_lock); 1373 LIST_REMOVE(pmap, pm_list); 1374 mtx_unlock_spin(&allpmaps_lock); 1375 1376 for (i = 0; i < NPGPTD; i++) 1377 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1378 1379 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1380 sizeof(*pmap->pm_pdir)); 1381#ifdef SMP 1382 pmap->pm_pdir[MPPTDI] = 0; 1383#endif 1384 1385 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1386 1387 vm_page_lock_queues(); 1388 for (i = 0; i < NPGPTD; i++) { 1389 m = ptdpg[i]; 1390#ifdef PAE 1391 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1392 ("pmap_release: got wrong ptd page")); 1393#endif 1394 m->wire_count--; 1395 atomic_subtract_int(&cnt.v_wire_count, 1); 1396 vm_page_free_zero(m); 1397 } 1398 vm_page_unlock_queues(); 1399} 1400 1401static int 1402kvm_size(SYSCTL_HANDLER_ARGS) 1403{ 1404 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1405 1406 return sysctl_handle_long(oidp, &ksize, 0, req); 1407} 1408SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1409 0, 0, kvm_size, "IU", "Size of KVM"); 1410 1411static int 1412kvm_free(SYSCTL_HANDLER_ARGS) 1413{ 1414 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1415 1416 return sysctl_handle_long(oidp, &kfree, 0, req); 1417} 1418SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1419 0, 0, kvm_free, "IU", "Amount of KVM free"); 1420 1421/* 1422 * grow the number of kernel page table entries, if needed 1423 */ 1424void 1425pmap_growkernel(vm_offset_t addr) 1426{ 1427 struct pmap *pmap; 1428 int s; 1429 vm_paddr_t ptppaddr; 1430 vm_page_t nkpg; 1431 pd_entry_t newpdir; 1432 pt_entry_t *pde; 1433 1434 s = splhigh(); 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 splx(s); 1475} 1476 1477 1478/*************************************************** 1479 * page management routines. 1480 ***************************************************/ 1481 1482/* 1483 * free the pv_entry back to the free list 1484 */ 1485static PMAP_INLINE void 1486free_pv_entry(pv_entry_t pv) 1487{ 1488 pv_entry_count--; 1489 uma_zfree(pvzone, pv); 1490} 1491 1492/* 1493 * get a new pv_entry, allocating a block from the system 1494 * when needed. 1495 * the memory allocation is performed bypassing the malloc code 1496 * because of the possibility of allocations at interrupt time. 1497 */ 1498static pv_entry_t 1499get_pv_entry(void) 1500{ 1501 pv_entry_count++; 1502 if (pv_entry_high_water && 1503 (pv_entry_count > pv_entry_high_water) && 1504 (pmap_pagedaemon_waken == 0)) { 1505 pmap_pagedaemon_waken = 1; 1506 wakeup (&vm_pages_needed); 1507 } 1508 return uma_zalloc(pvzone, M_NOWAIT); 1509} 1510 1511/* 1512 * If it is the first entry on the list, it is actually 1513 * in the header and we must copy the following entry up 1514 * to the header. Otherwise we must search the list for 1515 * the entry. In either case we free the now unused entry. 1516 */ 1517 1518static int 1519pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1520{ 1521 pv_entry_t pv; 1522 int rtval; 1523 int s; 1524 1525 s = splvm(); 1526 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1527 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1528 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1529 if (pmap == pv->pv_pmap && va == pv->pv_va) 1530 break; 1531 } 1532 } else { 1533 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1534 if (va == pv->pv_va) 1535 break; 1536 } 1537 } 1538 1539 rtval = 0; 1540 if (pv) { 1541 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1542 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1543 m->md.pv_list_count--; 1544 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1545 vm_page_flag_clear(m, PG_WRITEABLE); 1546 1547 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1548 free_pv_entry(pv); 1549 } 1550 1551 splx(s); 1552 return rtval; 1553} 1554 1555/* 1556 * Create a pv entry for page at pa for 1557 * (pmap, va). 1558 */ 1559static void 1560pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1561{ 1562 1563 int s; 1564 pv_entry_t pv; 1565 1566 s = splvm(); 1567 pv = get_pv_entry(); 1568 pv->pv_va = va; 1569 pv->pv_pmap = pmap; 1570 pv->pv_ptem = mpte; 1571 1572 vm_page_lock_queues(); 1573 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1574 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1575 m->md.pv_list_count++; 1576 1577 vm_page_unlock_queues(); 1578 splx(s); 1579} 1580 1581/* 1582 * pmap_remove_pte: do the things to unmap a page in a process 1583 */ 1584static int 1585pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1586{ 1587 pt_entry_t oldpte; 1588 vm_page_t m, mpte; 1589 1590 oldpte = pte_load_clear(ptq); 1591 if (oldpte & PG_W) 1592 pmap->pm_stats.wired_count -= 1; 1593 /* 1594 * Machines that don't support invlpg, also don't support 1595 * PG_G. 1596 */ 1597 if (oldpte & PG_G) 1598 pmap_invalidate_page(kernel_pmap, va); 1599 pmap->pm_stats.resident_count -= 1; 1600 if (oldpte & PG_MANAGED) { 1601 m = PHYS_TO_VM_PAGE(oldpte); 1602 if (oldpte & PG_M) { 1603#if defined(PMAP_DIAGNOSTIC) 1604 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1605 printf( 1606 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1607 va, oldpte); 1608 } 1609#endif 1610 if (pmap_track_modified(va)) 1611 vm_page_dirty(m); 1612 } 1613 if (oldpte & PG_A) 1614 vm_page_flag_set(m, PG_REFERENCED); 1615 return pmap_remove_entry(pmap, m, va); 1616 } else { 1617 mpte = PHYS_TO_VM_PAGE(*pmap_pde(pmap, va)); 1618 return pmap_unuse_pt(pmap, va, mpte); 1619 } 1620} 1621 1622/* 1623 * Remove a single page from a process address space 1624 */ 1625static void 1626pmap_remove_page(pmap_t pmap, vm_offset_t va) 1627{ 1628 pt_entry_t *pte; 1629 1630 if ((pte = pmap_pte(pmap, va)) == NULL || *pte == 0) 1631 return; 1632 pmap_remove_pte(pmap, pte, va); 1633 pmap_invalidate_page(pmap, va); 1634} 1635 1636/* 1637 * Remove the given range of addresses from the specified map. 1638 * 1639 * It is assumed that the start and end are properly 1640 * rounded to the page size. 1641 */ 1642void 1643pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1644{ 1645 vm_offset_t pdnxt; 1646 pd_entry_t ptpaddr; 1647 pt_entry_t *pte; 1648 int anyvalid; 1649 1650 if (pmap == NULL) 1651 return; 1652 1653 if (pmap->pm_stats.resident_count == 0) 1654 return; 1655 1656 /* 1657 * special handling of removing one page. a very 1658 * common operation and easy to short circuit some 1659 * code. 1660 */ 1661 if ((sva + PAGE_SIZE == eva) && 1662 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1663 pmap_remove_page(pmap, sva); 1664 return; 1665 } 1666 1667 anyvalid = 0; 1668 1669 for (; sva < eva; sva = pdnxt) { 1670 unsigned pdirindex; 1671 1672 /* 1673 * Calculate index for next page table. 1674 */ 1675 pdnxt = (sva + NBPDR) & ~PDRMASK; 1676 if (pmap->pm_stats.resident_count == 0) 1677 break; 1678 1679 pdirindex = sva >> PDRSHIFT; 1680 ptpaddr = pmap->pm_pdir[pdirindex]; 1681 1682 /* 1683 * Weed out invalid mappings. Note: we assume that the page 1684 * directory table is always allocated, and in kernel virtual. 1685 */ 1686 if (ptpaddr == 0) 1687 continue; 1688 1689 /* 1690 * Check for large page. 1691 */ 1692 if ((ptpaddr & PG_PS) != 0) { 1693 pmap->pm_pdir[pdirindex] = 0; 1694 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1695 anyvalid = 1; 1696 continue; 1697 } 1698 1699 /* 1700 * Limit our scan to either the end of the va represented 1701 * by the current page table page, or to the end of the 1702 * range being removed. 1703 */ 1704 if (pdnxt > eva) 1705 pdnxt = eva; 1706 1707 for (; sva != pdnxt; sva += PAGE_SIZE) { 1708 if ((pte = pmap_pte(pmap, sva)) == NULL || 1709 *pte == 0) 1710 continue; 1711 anyvalid = 1; 1712 if (pmap_remove_pte(pmap, pte, sva)) 1713 break; 1714 } 1715 } 1716 1717 if (anyvalid) 1718 pmap_invalidate_all(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 int s; 1740 1741#if defined(PMAP_DIAGNOSTIC) 1742 /* 1743 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1744 */ 1745 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1746 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1747 VM_PAGE_TO_PHYS(m)); 1748 } 1749#endif 1750 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1751 s = splvm(); 1752 sched_pin(); 1753 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1754 pv->pv_pmap->pm_stats.resident_count--; 1755 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1756 tpte = pte_load_clear(pte); 1757 if (tpte & PG_W) 1758 pv->pv_pmap->pm_stats.wired_count--; 1759 if (tpte & PG_A) 1760 vm_page_flag_set(m, PG_REFERENCED); 1761 1762 /* 1763 * Update the vm_page_t clean and reference bits. 1764 */ 1765 if (tpte & PG_M) { 1766#if defined(PMAP_DIAGNOSTIC) 1767 if (pmap_nw_modified((pt_entry_t) tpte)) { 1768 printf( 1769 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1770 pv->pv_va, tpte); 1771 } 1772#endif 1773 if (pmap_track_modified(pv->pv_va)) 1774 vm_page_dirty(m); 1775 } 1776 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1777 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1778 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1779 m->md.pv_list_count--; 1780 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 1781 free_pv_entry(pv); 1782 } 1783 vm_page_flag_clear(m, PG_WRITEABLE); 1784 sched_unpin(); 1785 splx(s); 1786} 1787 1788/* 1789 * Set the physical protection on the 1790 * specified range of this map as requested. 1791 */ 1792void 1793pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1794{ 1795 vm_offset_t pdnxt; 1796 pd_entry_t ptpaddr; 1797 int anychanged; 1798 1799 if (pmap == NULL) 1800 return; 1801 1802 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1803 pmap_remove(pmap, sva, eva); 1804 return; 1805 } 1806 1807 if (prot & VM_PROT_WRITE) 1808 return; 1809 1810 anychanged = 0; 1811 1812 for (; sva < eva; sva = pdnxt) { 1813 unsigned pdirindex; 1814 1815 pdnxt = (sva + NBPDR) & ~PDRMASK; 1816 1817 pdirindex = sva >> PDRSHIFT; 1818 ptpaddr = pmap->pm_pdir[pdirindex]; 1819 1820 /* 1821 * Weed out invalid mappings. Note: we assume that the page 1822 * directory table is always allocated, and in kernel virtual. 1823 */ 1824 if (ptpaddr == 0) 1825 continue; 1826 1827 /* 1828 * Check for large page. 1829 */ 1830 if ((ptpaddr & PG_PS) != 0) { 1831 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1832 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1833 anychanged = 1; 1834 continue; 1835 } 1836 1837 if (pdnxt > eva) 1838 pdnxt = eva; 1839 1840 for (; sva != pdnxt; sva += PAGE_SIZE) { 1841 pt_entry_t pbits; 1842 pt_entry_t *pte; 1843 vm_page_t m; 1844 1845 if ((pte = pmap_pte(pmap, sva)) == NULL) 1846 continue; 1847 pbits = *pte; 1848 if (pbits & PG_MANAGED) { 1849 m = NULL; 1850 if (pbits & PG_A) { 1851 m = PHYS_TO_VM_PAGE(pbits); 1852 vm_page_flag_set(m, PG_REFERENCED); 1853 pbits &= ~PG_A; 1854 } 1855 if ((pbits & PG_M) != 0 && 1856 pmap_track_modified(sva)) { 1857 if (m == NULL) 1858 m = PHYS_TO_VM_PAGE(pbits); 1859 vm_page_dirty(m); 1860 pbits &= ~PG_M; 1861 } 1862 } 1863 1864 pbits &= ~PG_RW; 1865 1866 if (pbits != *pte) { 1867 pte_store(pte, pbits); 1868 anychanged = 1; 1869 } 1870 } 1871 } 1872 if (anychanged) 1873 pmap_invalidate_all(pmap); 1874} 1875 1876/* 1877 * Insert the given physical page (p) at 1878 * the specified virtual address (v) in the 1879 * target physical map with the protection requested. 1880 * 1881 * If specified, the page will be wired down, meaning 1882 * that the related pte can not be reclaimed. 1883 * 1884 * NB: This is the only routine which MAY NOT lazy-evaluate 1885 * or lose information. That is, this routine must actually 1886 * insert this page into the given map NOW. 1887 */ 1888void 1889pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1890 boolean_t wired) 1891{ 1892 vm_paddr_t pa; 1893 register pt_entry_t *pte; 1894 vm_paddr_t opa; 1895 pt_entry_t origpte, newpte; 1896 vm_page_t mpte; 1897 1898 if (pmap == NULL) 1899 return; 1900 1901 va &= PG_FRAME; 1902#ifdef PMAP_DIAGNOSTIC 1903 if (va > VM_MAX_KERNEL_ADDRESS) 1904 panic("pmap_enter: toobig"); 1905 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1906 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1907#endif 1908 1909 mpte = NULL; 1910 /* 1911 * In the case that a page table page is not 1912 * resident, we are creating it here. 1913 */ 1914 if (va < VM_MAXUSER_ADDRESS) { 1915 mpte = pmap_allocpte(pmap, va); 1916 } 1917#if 0 && defined(PMAP_DIAGNOSTIC) 1918 else { 1919 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1920 origpte = *pdeaddr; 1921 if ((origpte & PG_V) == 0) { 1922 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1923 pmap->pm_pdir[PTDPTDI], origpte, va); 1924 } 1925 } 1926#endif 1927 1928 pte = pmap_pte(pmap, va); 1929 1930 /* 1931 * Page Directory table entry not valid, we need a new PT page 1932 */ 1933 if (pte == NULL) { 1934 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1935 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1936 } 1937 1938 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 1939 origpte = *pte; 1940 opa = origpte & PG_FRAME; 1941 1942 if (origpte & PG_PS) { 1943 /* 1944 * Yes, I know this will truncate upper address bits for PAE, 1945 * but I'm actually more interested in the lower bits 1946 */ 1947 printf("pmap_enter: va %p, pte %p, origpte %p\n", 1948 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 1949 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1950 } 1951 1952 /* 1953 * Mapping has not changed, must be protection or wiring change. 1954 */ 1955 if (origpte && (opa == pa)) { 1956 /* 1957 * Wiring change, just update stats. We don't worry about 1958 * wiring PT pages as they remain resident as long as there 1959 * are valid mappings in them. Hence, if a user page is wired, 1960 * the PT page will be also. 1961 */ 1962 if (wired && ((origpte & PG_W) == 0)) 1963 pmap->pm_stats.wired_count++; 1964 else if (!wired && (origpte & PG_W)) 1965 pmap->pm_stats.wired_count--; 1966 1967#if defined(PMAP_DIAGNOSTIC) 1968 if (pmap_nw_modified((pt_entry_t) origpte)) { 1969 printf( 1970 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 1971 va, origpte); 1972 } 1973#endif 1974 1975 /* 1976 * Remove extra pte reference 1977 */ 1978 if (mpte) 1979 mpte->hold_count--; 1980 1981 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) { 1982 if ((origpte & PG_RW) == 0) { 1983 pte_store(pte, origpte | PG_RW); 1984 pmap_invalidate_page(pmap, va); 1985 } 1986 return; 1987 } 1988 1989 /* 1990 * We might be turning off write access to the page, 1991 * so we go ahead and sense modify status. 1992 */ 1993 if (origpte & PG_MANAGED) { 1994 if ((origpte & PG_M) && pmap_track_modified(va)) { 1995 vm_page_t om; 1996 om = PHYS_TO_VM_PAGE(opa); 1997 vm_page_dirty(om); 1998 } 1999 pa |= PG_MANAGED; 2000 } 2001 goto validate; 2002 } 2003 /* 2004 * Mapping has changed, invalidate old range and fall through to 2005 * handle validating new mapping. 2006 */ 2007 if (opa) { 2008 int err; 2009 vm_page_lock_queues(); 2010 err = pmap_remove_pte(pmap, pte, va); 2011 vm_page_unlock_queues(); 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, mpte, 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} 2059 2060/* 2061 * this code makes some *MAJOR* assumptions: 2062 * 1. Current pmap & pmap exists. 2063 * 2. Not wired. 2064 * 3. Read access. 2065 * 4. No page table pages. 2066 * 5. Tlbflush is deferred to calling procedure. 2067 * 6. Page IS managed. 2068 * but is *MUCH* faster than pmap_enter... 2069 */ 2070 2071vm_page_t 2072pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2073{ 2074 pt_entry_t *pte; 2075 vm_paddr_t pa; 2076 2077 /* 2078 * In the case that a page table page is not 2079 * resident, we are creating it here. 2080 */ 2081 if (va < VM_MAXUSER_ADDRESS) { 2082 unsigned ptepindex; 2083 pd_entry_t ptepa; 2084 2085 /* 2086 * Calculate pagetable page index 2087 */ 2088 ptepindex = va >> PDRSHIFT; 2089 if (mpte && (mpte->pindex == ptepindex)) { 2090 mpte->hold_count++; 2091 } else { 2092retry: 2093 /* 2094 * Get the page directory entry 2095 */ 2096 ptepa = pmap->pm_pdir[ptepindex]; 2097 2098 /* 2099 * If the page table page is mapped, we just increment 2100 * the hold count, and activate it. 2101 */ 2102 if (ptepa) { 2103 if (ptepa & PG_PS) 2104 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2105 mpte = PHYS_TO_VM_PAGE(ptepa); 2106 mpte->hold_count++; 2107 } else { 2108 mpte = _pmap_allocpte(pmap, ptepindex); 2109 if (mpte == NULL) 2110 goto retry; 2111 } 2112 } 2113 } else { 2114 mpte = NULL; 2115 } 2116 2117 /* 2118 * This call to vtopte makes the assumption that we are 2119 * entering the page into the current pmap. In order to support 2120 * quick entry into any pmap, one would likely use pmap_pte_quick. 2121 * But that isn't as quick as vtopte. 2122 */ 2123 pte = vtopte(va); 2124 if (*pte) { 2125 if (mpte != NULL) { 2126 vm_page_lock_queues(); 2127 pmap_unwire_pte_hold(pmap, mpte); 2128 vm_page_unlock_queues(); 2129 } 2130 return 0; 2131 } 2132 2133 /* 2134 * Enter on the PV list if part of our managed memory. Note that we 2135 * raise IPL while manipulating pv_table since pmap_enter can be 2136 * called at interrupt time. 2137 */ 2138 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2139 pmap_insert_entry(pmap, va, mpte, m); 2140 2141 /* 2142 * Increment counters 2143 */ 2144 pmap->pm_stats.resident_count++; 2145 2146 pa = VM_PAGE_TO_PHYS(m); 2147 2148 /* 2149 * Now validate mapping with RO protection 2150 */ 2151 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2152 pte_store(pte, pa | PG_V | PG_U); 2153 else 2154 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2155 2156 return mpte; 2157} 2158 2159/* 2160 * Make a temporary mapping for a physical address. This is only intended 2161 * to be used for panic dumps. 2162 */ 2163void * 2164pmap_kenter_temporary(vm_offset_t pa, int i) 2165{ 2166 vm_offset_t va; 2167 2168 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2169 pmap_kenter(va, pa); 2170#ifndef I386_CPU 2171 invlpg(va); 2172#else 2173 invltlb(); 2174#endif 2175 return ((void *)crashdumpmap); 2176} 2177 2178/* 2179 * This code maps large physical mmap regions into the 2180 * processor address space. Note that some shortcuts 2181 * are taken, but the code works. 2182 */ 2183void 2184pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2185 vm_object_t object, vm_pindex_t pindex, 2186 vm_size_t size) 2187{ 2188 vm_page_t p; 2189 2190 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2191 KASSERT(object->type == OBJT_DEVICE, 2192 ("pmap_object_init_pt: non-device object")); 2193 if (pseflag && 2194 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2195 int i; 2196 vm_page_t m[1]; 2197 unsigned int ptepindex; 2198 int npdes; 2199 pd_entry_t ptepa; 2200 2201 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2202 return; 2203retry: 2204 p = vm_page_lookup(object, pindex); 2205 if (p != NULL) { 2206 vm_page_lock_queues(); 2207 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2208 goto retry; 2209 } else { 2210 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2211 if (p == NULL) 2212 return; 2213 m[0] = p; 2214 2215 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2216 vm_page_lock_queues(); 2217 vm_page_free(p); 2218 vm_page_unlock_queues(); 2219 return; 2220 } 2221 2222 p = vm_page_lookup(object, pindex); 2223 vm_page_lock_queues(); 2224 vm_page_wakeup(p); 2225 } 2226 vm_page_unlock_queues(); 2227 2228 ptepa = VM_PAGE_TO_PHYS(p); 2229 if (ptepa & (NBPDR - 1)) 2230 return; 2231 2232 p->valid = VM_PAGE_BITS_ALL; 2233 2234 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2235 npdes = size >> PDRSHIFT; 2236 for(i = 0; i < npdes; i++) { 2237 pde_store(&pmap->pm_pdir[ptepindex], 2238 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2239 ptepa += NBPDR; 2240 ptepindex += 1; 2241 } 2242 pmap_invalidate_all(pmap); 2243 } 2244} 2245 2246/* 2247 * Routine: pmap_change_wiring 2248 * Function: Change the wiring attribute for a map/virtual-address 2249 * pair. 2250 * In/out conditions: 2251 * The mapping must already exist in the pmap. 2252 */ 2253void 2254pmap_change_wiring(pmap, va, wired) 2255 register pmap_t pmap; 2256 vm_offset_t va; 2257 boolean_t wired; 2258{ 2259 register pt_entry_t *pte; 2260 2261 if (pmap == NULL) 2262 return; 2263 2264 pte = pmap_pte(pmap, va); 2265 2266 if (wired && !pmap_pte_w(pte)) 2267 pmap->pm_stats.wired_count++; 2268 else if (!wired && pmap_pte_w(pte)) 2269 pmap->pm_stats.wired_count--; 2270 2271 /* 2272 * Wiring is not a hardware characteristic so there is no need to 2273 * invalidate TLB. 2274 */ 2275 pmap_pte_set_w(pte, wired); 2276} 2277 2278 2279 2280/* 2281 * Copy the range specified by src_addr/len 2282 * from the source map to the range dst_addr/len 2283 * in the destination map. 2284 * 2285 * This routine is only advisory and need not do anything. 2286 */ 2287 2288void 2289pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2290 vm_offset_t src_addr) 2291{ 2292 vm_offset_t addr; 2293 vm_offset_t end_addr = src_addr + len; 2294 vm_offset_t pdnxt; 2295 vm_page_t m; 2296 2297 if (dst_addr != src_addr) 2298 return; 2299 2300 if (!pmap_is_current(src_pmap)) 2301 return; 2302 2303 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2304 pt_entry_t *src_pte, *dst_pte; 2305 vm_page_t dstmpte, srcmpte; 2306 pd_entry_t srcptepaddr; 2307 unsigned ptepindex; 2308 2309 if (addr >= UPT_MIN_ADDRESS) 2310 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2311 2312 /* 2313 * Don't let optional prefaulting of pages make us go 2314 * way below the low water mark of free pages or way 2315 * above high water mark of used pv entries. 2316 */ 2317 if (cnt.v_free_count < cnt.v_free_reserved || 2318 pv_entry_count > pv_entry_high_water) 2319 break; 2320 2321 pdnxt = (addr + NBPDR) & ~PDRMASK; 2322 ptepindex = addr >> PDRSHIFT; 2323 2324 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2325 if (srcptepaddr == 0) 2326 continue; 2327 2328 if (srcptepaddr & PG_PS) { 2329 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2330 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2331 dst_pmap->pm_stats.resident_count += 2332 NBPDR / PAGE_SIZE; 2333 } 2334 continue; 2335 } 2336 2337 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2338 if (srcmpte->hold_count == 0 || (srcmpte->flags & PG_BUSY)) 2339 continue; 2340 2341 if (pdnxt > end_addr) 2342 pdnxt = end_addr; 2343 2344 src_pte = vtopte(addr); 2345 while (addr < pdnxt) { 2346 pt_entry_t ptetemp; 2347 ptetemp = *src_pte; 2348 /* 2349 * we only virtual copy managed pages 2350 */ 2351 if ((ptetemp & PG_MANAGED) != 0) { 2352 /* 2353 * We have to check after allocpte for the 2354 * pte still being around... allocpte can 2355 * block. 2356 */ 2357 dstmpte = pmap_allocpte(dst_pmap, addr); 2358 dst_pte = pmap_pte(dst_pmap, addr); 2359 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2360 /* 2361 * Clear the modified and 2362 * accessed (referenced) bits 2363 * during the copy. 2364 */ 2365 m = PHYS_TO_VM_PAGE(ptetemp); 2366 *dst_pte = ptetemp & ~(PG_M | PG_A); 2367 dst_pmap->pm_stats.resident_count++; 2368 pmap_insert_entry(dst_pmap, addr, 2369 dstmpte, m); 2370 } else { 2371 vm_page_lock_queues(); 2372 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2373 vm_page_unlock_queues(); 2374 } 2375 if (dstmpte->hold_count >= srcmpte->hold_count) 2376 break; 2377 } 2378 addr += PAGE_SIZE; 2379 src_pte++; 2380 } 2381 } 2382} 2383 2384static __inline void 2385pagezero(void *page) 2386{ 2387#if defined(I686_CPU) 2388 if (cpu_class == CPUCLASS_686) { 2389#if defined(CPU_ENABLE_SSE) 2390 if (cpu_feature & CPUID_SSE2) 2391 sse2_pagezero(page); 2392 else 2393#endif 2394 i686_pagezero(page); 2395 } else 2396#endif 2397 bzero(page, PAGE_SIZE); 2398} 2399 2400/* 2401 * pmap_zero_page zeros the specified hardware page by mapping 2402 * the page into KVM and using bzero to clear its contents. 2403 */ 2404void 2405pmap_zero_page(vm_page_t m) 2406{ 2407 2408 mtx_lock(&CMAPCADDR12_lock); 2409 if (*CMAP2) 2410 panic("pmap_zero_page: CMAP2 busy"); 2411 sched_pin(); 2412 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2413 invlcaddr(CADDR2); 2414 pagezero(CADDR2); 2415 *CMAP2 = 0; 2416 sched_unpin(); 2417 mtx_unlock(&CMAPCADDR12_lock); 2418} 2419 2420/* 2421 * pmap_zero_page_area zeros the specified hardware page by mapping 2422 * the page into KVM and using bzero to clear its contents. 2423 * 2424 * off and size may not cover an area beyond a single hardware page. 2425 */ 2426void 2427pmap_zero_page_area(vm_page_t m, int off, int size) 2428{ 2429 2430 mtx_lock(&CMAPCADDR12_lock); 2431 if (*CMAP2) 2432 panic("pmap_zero_page: CMAP2 busy"); 2433 sched_pin(); 2434 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2435 invlcaddr(CADDR2); 2436 if (off == 0 && size == PAGE_SIZE) 2437 pagezero(CADDR2); 2438 else 2439 bzero((char *)CADDR2 + off, size); 2440 *CMAP2 = 0; 2441 sched_unpin(); 2442 mtx_unlock(&CMAPCADDR12_lock); 2443} 2444 2445/* 2446 * pmap_zero_page_idle zeros the specified hardware page by mapping 2447 * the page into KVM and using bzero to clear its contents. This 2448 * is intended to be called from the vm_pagezero process only and 2449 * outside of Giant. 2450 */ 2451void 2452pmap_zero_page_idle(vm_page_t m) 2453{ 2454 2455 if (*CMAP3) 2456 panic("pmap_zero_page: CMAP3 busy"); 2457 sched_pin(); 2458 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2459 invlcaddr(CADDR3); 2460 pagezero(CADDR3); 2461 *CMAP3 = 0; 2462 sched_unpin(); 2463} 2464 2465/* 2466 * pmap_copy_page copies the specified (machine independent) 2467 * page by mapping the page into virtual memory and using 2468 * bcopy to copy the page, one machine dependent page at a 2469 * time. 2470 */ 2471void 2472pmap_copy_page(vm_page_t src, vm_page_t dst) 2473{ 2474 2475 mtx_lock(&CMAPCADDR12_lock); 2476 if (*CMAP1) 2477 panic("pmap_copy_page: CMAP1 busy"); 2478 if (*CMAP2) 2479 panic("pmap_copy_page: CMAP2 busy"); 2480 sched_pin(); 2481#ifdef I386_CPU 2482 invltlb(); 2483#else 2484 invlpg((u_int)CADDR1); 2485 invlpg((u_int)CADDR2); 2486#endif 2487 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2488 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2489 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2490 *CMAP1 = 0; 2491 *CMAP2 = 0; 2492 sched_unpin(); 2493 mtx_unlock(&CMAPCADDR12_lock); 2494} 2495 2496/* 2497 * Returns true if the pmap's pv is one of the first 2498 * 16 pvs linked to from this page. This count may 2499 * be changed upwards or downwards in the future; it 2500 * is only necessary that true be returned for a small 2501 * subset of pmaps for proper page aging. 2502 */ 2503boolean_t 2504pmap_page_exists_quick(pmap, m) 2505 pmap_t pmap; 2506 vm_page_t m; 2507{ 2508 pv_entry_t pv; 2509 int loops = 0; 2510 int s; 2511 2512 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2513 return FALSE; 2514 2515 s = splvm(); 2516 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2517 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2518 if (pv->pv_pmap == pmap) { 2519 splx(s); 2520 return TRUE; 2521 } 2522 loops++; 2523 if (loops >= 16) 2524 break; 2525 } 2526 splx(s); 2527 return (FALSE); 2528} 2529 2530#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2531/* 2532 * Remove all pages from specified address space 2533 * this aids process exit speeds. Also, this code 2534 * is special cased for current process only, but 2535 * can have the more generic (and slightly slower) 2536 * mode enabled. This is much faster than pmap_remove 2537 * in the case of running down an entire address space. 2538 */ 2539void 2540pmap_remove_pages(pmap, sva, eva) 2541 pmap_t pmap; 2542 vm_offset_t sva, eva; 2543{ 2544 pt_entry_t *pte, tpte; 2545 vm_page_t m; 2546 pv_entry_t pv, npv; 2547 int s; 2548 2549#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2550 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2551 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2552 return; 2553 } 2554#endif 2555 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2556 s = splvm(); 2557 sched_pin(); 2558 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2559 2560 if (pv->pv_va >= eva || pv->pv_va < sva) { 2561 npv = TAILQ_NEXT(pv, pv_plist); 2562 continue; 2563 } 2564 2565#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2566 pte = vtopte(pv->pv_va); 2567#else 2568 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2569#endif 2570 tpte = *pte; 2571 2572 if (tpte == 0) { 2573 printf("TPTE at %p IS ZERO @ VA %08x\n", 2574 pte, pv->pv_va); 2575 panic("bad pte"); 2576 } 2577 2578/* 2579 * We cannot remove wired pages from a process' mapping at this time 2580 */ 2581 if (tpte & PG_W) { 2582 npv = TAILQ_NEXT(pv, pv_plist); 2583 continue; 2584 } 2585 2586 m = PHYS_TO_VM_PAGE(tpte); 2587 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2588 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2589 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2590 2591 KASSERT(m < &vm_page_array[vm_page_array_size], 2592 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2593 2594 pv->pv_pmap->pm_stats.resident_count--; 2595 2596 pte_clear(pte); 2597 2598 /* 2599 * Update the vm_page_t clean and reference bits. 2600 */ 2601 if (tpte & PG_M) { 2602 vm_page_dirty(m); 2603 } 2604 2605 npv = TAILQ_NEXT(pv, pv_plist); 2606 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2607 2608 m->md.pv_list_count--; 2609 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2610 if (TAILQ_FIRST(&m->md.pv_list) == NULL) { 2611 vm_page_flag_clear(m, PG_WRITEABLE); 2612 } 2613 2614 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2615 free_pv_entry(pv); 2616 } 2617 sched_unpin(); 2618 splx(s); 2619 pmap_invalidate_all(pmap); 2620} 2621 2622/* 2623 * pmap_is_modified: 2624 * 2625 * Return whether or not the specified physical page was modified 2626 * in any physical maps. 2627 */ 2628boolean_t 2629pmap_is_modified(vm_page_t m) 2630{ 2631 pv_entry_t pv; 2632 pt_entry_t *pte; 2633 int s; 2634 2635 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2636 return FALSE; 2637 2638 s = splvm(); 2639 sched_pin(); 2640 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2641 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2642 /* 2643 * if the bit being tested is the modified bit, then 2644 * mark clean_map and ptes as never 2645 * modified. 2646 */ 2647 if (!pmap_track_modified(pv->pv_va)) 2648 continue; 2649#if defined(PMAP_DIAGNOSTIC) 2650 if (!pv->pv_pmap) { 2651 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 2652 continue; 2653 } 2654#endif 2655 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2656 if (*pte & PG_M) { 2657 sched_unpin(); 2658 splx(s); 2659 return TRUE; 2660 } 2661 } 2662 sched_unpin(); 2663 splx(s); 2664 return (FALSE); 2665} 2666 2667/* 2668 * pmap_is_prefaultable: 2669 * 2670 * Return whether or not the specified virtual address is elgible 2671 * for prefault. 2672 */ 2673boolean_t 2674pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2675{ 2676 pt_entry_t *pte; 2677 2678 if ((*pmap_pde(pmap, addr)) == 0) 2679 return (FALSE); 2680 pte = vtopte(addr); 2681 if (*pte) 2682 return (FALSE); 2683 return (TRUE); 2684} 2685 2686/* 2687 * Clear the given bit in each of the given page's ptes. 2688 */ 2689static __inline void 2690pmap_clear_ptes(vm_page_t m, int bit) 2691{ 2692 register pv_entry_t pv; 2693 pt_entry_t pbits, *pte; 2694 int s; 2695 2696 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) || 2697 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2698 return; 2699 2700 s = splvm(); 2701 sched_pin(); 2702 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2703 /* 2704 * Loop over all current mappings setting/clearing as appropos If 2705 * setting RO do we need to clear the VAC? 2706 */ 2707 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2708 /* 2709 * don't write protect pager mappings 2710 */ 2711 if (bit == PG_RW) { 2712 if (!pmap_track_modified(pv->pv_va)) 2713 continue; 2714 } 2715 2716#if defined(PMAP_DIAGNOSTIC) 2717 if (!pv->pv_pmap) { 2718 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 2719 continue; 2720 } 2721#endif 2722 2723 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2724 pbits = *pte; 2725 if (pbits & bit) { 2726 if (bit == PG_RW) { 2727 if (pbits & PG_M) { 2728 vm_page_dirty(m); 2729 } 2730 pte_store(pte, pbits & ~(PG_M|PG_RW)); 2731 } else { 2732 pte_store(pte, pbits & ~bit); 2733 } 2734 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2735 } 2736 } 2737 if (bit == PG_RW) 2738 vm_page_flag_clear(m, PG_WRITEABLE); 2739 sched_unpin(); 2740 splx(s); 2741} 2742 2743/* 2744 * pmap_page_protect: 2745 * 2746 * Lower the permission for all mappings to a given page. 2747 */ 2748void 2749pmap_page_protect(vm_page_t m, vm_prot_t prot) 2750{ 2751 if ((prot & VM_PROT_WRITE) == 0) { 2752 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2753 pmap_clear_ptes(m, PG_RW); 2754 } else { 2755 pmap_remove_all(m); 2756 } 2757 } 2758} 2759 2760/* 2761 * pmap_ts_referenced: 2762 * 2763 * Return a count of reference bits for a page, clearing those bits. 2764 * It is not necessary for every reference bit to be cleared, but it 2765 * is necessary that 0 only be returned when there are truly no 2766 * reference bits set. 2767 * 2768 * XXX: The exact number of bits to check and clear is a matter that 2769 * should be tested and standardized at some point in the future for 2770 * optimal aging of shared pages. 2771 */ 2772int 2773pmap_ts_referenced(vm_page_t m) 2774{ 2775 register pv_entry_t pv, pvf, pvn; 2776 pt_entry_t *pte; 2777 pt_entry_t v; 2778 int s; 2779 int rtval = 0; 2780 2781 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2782 return (rtval); 2783 2784 s = splvm(); 2785 sched_pin(); 2786 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2787 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2788 2789 pvf = pv; 2790 2791 do { 2792 pvn = TAILQ_NEXT(pv, pv_list); 2793 2794 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2795 2796 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2797 2798 if (!pmap_track_modified(pv->pv_va)) 2799 continue; 2800 2801 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2802 2803 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2804 pte_store(pte, v & ~PG_A); 2805 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2806 2807 rtval++; 2808 if (rtval > 4) { 2809 break; 2810 } 2811 } 2812 } while ((pv = pvn) != NULL && pv != pvf); 2813 } 2814 sched_unpin(); 2815 splx(s); 2816 2817 return (rtval); 2818} 2819 2820/* 2821 * Clear the modify bits on the specified physical page. 2822 */ 2823void 2824pmap_clear_modify(vm_page_t m) 2825{ 2826 pmap_clear_ptes(m, PG_M); 2827} 2828 2829/* 2830 * pmap_clear_reference: 2831 * 2832 * Clear the reference bit on the specified physical page. 2833 */ 2834void 2835pmap_clear_reference(vm_page_t m) 2836{ 2837 pmap_clear_ptes(m, PG_A); 2838} 2839 2840/* 2841 * Miscellaneous support routines follow 2842 */ 2843 2844/* 2845 * Map a set of physical memory pages into the kernel virtual 2846 * address space. Return a pointer to where it is mapped. This 2847 * routine is intended to be used for mapping device memory, 2848 * NOT real memory. 2849 */ 2850void * 2851pmap_mapdev(pa, size) 2852 vm_paddr_t pa; 2853 vm_size_t size; 2854{ 2855 vm_offset_t va, tmpva, offset; 2856 2857 offset = pa & PAGE_MASK; 2858 size = roundup(offset + size, PAGE_SIZE); 2859 pa = pa & PG_FRAME; 2860 2861 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2862 va = KERNBASE + pa; 2863 else 2864 va = kmem_alloc_nofault(kernel_map, size); 2865 if (!va) 2866 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2867 2868 for (tmpva = va; size > 0; ) { 2869 pmap_kenter(tmpva, pa); 2870 size -= PAGE_SIZE; 2871 tmpva += PAGE_SIZE; 2872 pa += PAGE_SIZE; 2873 } 2874 pmap_invalidate_range(kernel_pmap, va, tmpva); 2875 return ((void *)(va + offset)); 2876} 2877 2878void 2879pmap_unmapdev(va, size) 2880 vm_offset_t va; 2881 vm_size_t size; 2882{ 2883 vm_offset_t base, offset, tmpva; 2884 2885 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2886 return; 2887 base = va & PG_FRAME; 2888 offset = va & PAGE_MASK; 2889 size = roundup(offset + size, PAGE_SIZE); 2890 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2891 pmap_kremove(tmpva); 2892 pmap_invalidate_range(kernel_pmap, va, tmpva); 2893 kmem_free(kernel_map, base, size); 2894} 2895 2896/* 2897 * perform the pmap work for mincore 2898 */ 2899int 2900pmap_mincore(pmap, addr) 2901 pmap_t pmap; 2902 vm_offset_t addr; 2903{ 2904 pt_entry_t *ptep, pte; 2905 vm_page_t m; 2906 int val = 0; 2907 2908 ptep = pmap_pte(pmap, addr); 2909 if (ptep == 0) { 2910 return 0; 2911 } 2912 2913 if ((pte = *ptep) != 0) { 2914 vm_paddr_t pa; 2915 2916 val = MINCORE_INCORE; 2917 if ((pte & PG_MANAGED) == 0) 2918 return val; 2919 2920 pa = pte & PG_FRAME; 2921 2922 m = PHYS_TO_VM_PAGE(pa); 2923 2924 /* 2925 * Modified by us 2926 */ 2927 if (pte & PG_M) 2928 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2929 else { 2930 /* 2931 * Modified by someone else 2932 */ 2933 vm_page_lock_queues(); 2934 if (m->dirty || pmap_is_modified(m)) 2935 val |= MINCORE_MODIFIED_OTHER; 2936 vm_page_unlock_queues(); 2937 } 2938 /* 2939 * Referenced by us 2940 */ 2941 if (pte & PG_A) 2942 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2943 else { 2944 /* 2945 * Referenced by someone else 2946 */ 2947 vm_page_lock_queues(); 2948 if ((m->flags & PG_REFERENCED) || 2949 pmap_ts_referenced(m)) { 2950 val |= MINCORE_REFERENCED_OTHER; 2951 vm_page_flag_set(m, PG_REFERENCED); 2952 } 2953 vm_page_unlock_queues(); 2954 } 2955 } 2956 return val; 2957} 2958 2959void 2960pmap_activate(struct thread *td) 2961{ 2962 struct proc *p = td->td_proc; 2963 pmap_t pmap, oldpmap; 2964 u_int32_t cr3; 2965 2966 critical_enter(); 2967 pmap = vmspace_pmap(td->td_proc->p_vmspace); 2968 oldpmap = PCPU_GET(curpmap); 2969#if defined(SMP) 2970 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 2971 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 2972#else 2973 oldpmap->pm_active &= ~1; 2974 pmap->pm_active |= 1; 2975#endif 2976#ifdef PAE 2977 cr3 = vtophys(pmap->pm_pdpt); 2978#else 2979 cr3 = vtophys(pmap->pm_pdir); 2980#endif 2981 /* XXXKSE this is wrong. 2982 * pmap_activate is for the current thread on the current cpu 2983 */ 2984 if (p->p_flag & P_SA) { 2985 /* Make sure all other cr3 entries are updated. */ 2986 /* what if they are running? XXXKSE (maybe abort them) */ 2987 FOREACH_THREAD_IN_PROC(p, td) { 2988 td->td_pcb->pcb_cr3 = cr3; 2989 } 2990 } else { 2991 td->td_pcb->pcb_cr3 = cr3; 2992 } 2993 load_cr3(cr3); 2994 PCPU_SET(curpmap, pmap); 2995 critical_exit(); 2996} 2997 2998vm_offset_t 2999pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3000{ 3001 3002 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3003 return addr; 3004 } 3005 3006 addr = (addr + PDRMASK) & ~PDRMASK; 3007 return addr; 3008} 3009 3010 3011#if defined(PMAP_DEBUG) 3012pmap_pid_dump(int pid) 3013{ 3014 pmap_t pmap; 3015 struct proc *p; 3016 int npte = 0; 3017 int index; 3018 3019 sx_slock(&allproc_lock); 3020 LIST_FOREACH(p, &allproc, p_list) { 3021 if (p->p_pid != pid) 3022 continue; 3023 3024 if (p->p_vmspace) { 3025 int i,j; 3026 index = 0; 3027 pmap = vmspace_pmap(p->p_vmspace); 3028 for (i = 0; i < NPDEPTD; i++) { 3029 pd_entry_t *pde; 3030 pt_entry_t *pte; 3031 vm_offset_t base = i << PDRSHIFT; 3032 3033 pde = &pmap->pm_pdir[i]; 3034 if (pde && pmap_pde_v(pde)) { 3035 for (j = 0; j < NPTEPG; j++) { 3036 vm_offset_t va = base + (j << PAGE_SHIFT); 3037 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3038 if (index) { 3039 index = 0; 3040 printf("\n"); 3041 } 3042 sx_sunlock(&allproc_lock); 3043 return npte; 3044 } 3045 pte = pmap_pte(pmap, va); 3046 if (pte && pmap_pte_v(pte)) { 3047 pt_entry_t pa; 3048 vm_page_t m; 3049 pa = *pte; 3050 m = PHYS_TO_VM_PAGE(pa); 3051 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3052 va, pa, m->hold_count, m->wire_count, m->flags); 3053 npte++; 3054 index++; 3055 if (index >= 2) { 3056 index = 0; 3057 printf("\n"); 3058 } else { 3059 printf(" "); 3060 } 3061 } 3062 } 3063 } 3064 } 3065 } 3066 } 3067 sx_sunlock(&allproc_lock); 3068 return npte; 3069} 3070#endif 3071 3072#if defined(DEBUG) 3073 3074static void pads(pmap_t pm); 3075void pmap_pvdump(vm_offset_t pa); 3076 3077/* print address space of pmap*/ 3078static void 3079pads(pm) 3080 pmap_t pm; 3081{ 3082 int i, j; 3083 vm_paddr_t va; 3084 pt_entry_t *ptep; 3085 3086 if (pm == kernel_pmap) 3087 return; 3088 for (i = 0; i < NPDEPTD; i++) 3089 if (pm->pm_pdir[i]) 3090 for (j = 0; j < NPTEPG; j++) { 3091 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3092 if (pm == kernel_pmap && va < KERNBASE) 3093 continue; 3094 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3095 continue; 3096 ptep = pmap_pte(pm, va); 3097 if (pmap_pte_v(ptep)) 3098 printf("%x:%x ", va, *ptep); 3099 }; 3100 3101} 3102 3103void 3104pmap_pvdump(pa) 3105 vm_paddr_t pa; 3106{ 3107 pv_entry_t pv; 3108 vm_page_t m; 3109 3110 printf("pa %x", pa); 3111 m = PHYS_TO_VM_PAGE(pa); 3112 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3113 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3114 pads(pv->pv_pmap); 3115 } 3116 printf(" "); 3117} 3118#endif 3119