pmap.c revision 123925
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 123925 2003-12-28 05:10:21Z 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_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 175struct pmap kernel_pmap_store; 176LIST_HEAD(pmaplist, pmap); 177static struct pmaplist allpmaps; 178static struct mtx allpmaps_lock; 179#ifdef SMP 180static struct mtx lazypmap_lock; 181#endif 182 183vm_paddr_t avail_start; /* PA of first available physical page */ 184vm_paddr_t avail_end; /* PA of last available physical page */ 185vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 186vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 187static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 188int pgeflag = 0; /* PG_G or-in */ 189int pseflag = 0; /* PG_PS or-in */ 190 191static int nkpt; 192vm_offset_t kernel_vm_end; 193extern u_int32_t KERNend; 194 195#ifdef PAE 196static uma_zone_t pdptzone; 197#endif 198 199/* 200 * Data for the pv entry allocation mechanism 201 */ 202static uma_zone_t pvzone; 203static struct vm_object pvzone_obj; 204static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 205int pmap_pagedaemon_waken; 206 207/* 208 * All those kernel PT submaps that BSD is so fond of 209 */ 210pt_entry_t *CMAP1 = 0; 211static pt_entry_t *CMAP2, *CMAP3, *ptmmap; 212caddr_t CADDR1 = 0, ptvmmap = 0; 213static caddr_t CADDR2, CADDR3; 214static struct mtx CMAPCADDR12_lock; 215static pt_entry_t *msgbufmap; 216struct msgbuf *msgbufp = 0; 217 218/* 219 * Crashdump maps. 220 */ 221static pt_entry_t *pt_crashdumpmap; 222static caddr_t crashdumpmap; 223 224#ifdef SMP 225extern pt_entry_t *SMPpt; 226#endif 227static pt_entry_t *PMAP1 = 0, *PMAP2; 228static pt_entry_t *PADDR1 = 0, *PADDR2; 229 230static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 231static pv_entry_t get_pv_entry(void); 232static void pmap_clear_ptes(vm_page_t m, int bit) 233 __always_inline; 234 235static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 236static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 237static int pmap_remove_entry(struct pmap *pmap, vm_page_t m, 238 vm_offset_t va); 239static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, 240 vm_page_t mpte, vm_page_t m); 241 242static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va); 243 244static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex); 245static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 246static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t); 247static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 248#ifdef PAE 249static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 250#endif 251 252CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 253CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 254 255/* 256 * Move the kernel virtual free pointer to the next 257 * 4MB. This is used to help improve performance 258 * by using a large (4MB) page for much of the kernel 259 * (.text, .data, .bss) 260 */ 261static vm_offset_t 262pmap_kmem_choose(vm_offset_t addr) 263{ 264 vm_offset_t newaddr = addr; 265 266#ifndef DISABLE_PSE 267 if (cpu_feature & CPUID_PSE) 268 newaddr = (addr + PDRMASK) & ~PDRMASK; 269#endif 270 return newaddr; 271} 272 273/* 274 * Bootstrap the system enough to run with virtual memory. 275 * 276 * On the i386 this is called after mapping has already been enabled 277 * and just syncs the pmap module with what has already been done. 278 * [We can't call it easily with mapping off since the kernel is not 279 * mapped with PA == VA, hence we would have to relocate every address 280 * from the linked base (virtual) address "KERNBASE" to the actual 281 * (physical) address starting relative to 0] 282 */ 283void 284pmap_bootstrap(firstaddr, loadaddr) 285 vm_paddr_t firstaddr; 286 vm_paddr_t loadaddr; 287{ 288 vm_offset_t va; 289 pt_entry_t *pte; 290 int i; 291 292 avail_start = firstaddr; 293 294 /* 295 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 296 * large. It should instead be correctly calculated in locore.s and 297 * not based on 'first' (which is a physical address, not a virtual 298 * address, for the start of unused physical memory). The kernel 299 * page tables are NOT double mapped and thus should not be included 300 * in this calculation. 301 */ 302 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 303 virtual_avail = pmap_kmem_choose(virtual_avail); 304 305 virtual_end = VM_MAX_KERNEL_ADDRESS; 306 307 /* 308 * Initialize the kernel pmap (which is statically allocated). 309 */ 310 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 311#ifdef PAE 312 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 313#endif 314 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 315 TAILQ_INIT(&kernel_pmap->pm_pvlist); 316 LIST_INIT(&allpmaps); 317#ifdef SMP 318 mtx_init(&lazypmap_lock, "lazypmap", NULL, MTX_SPIN); 319#endif 320 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 321 mtx_lock_spin(&allpmaps_lock); 322 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 323 mtx_unlock_spin(&allpmaps_lock); 324 nkpt = NKPT; 325 326 /* 327 * Reserve some special page table entries/VA space for temporary 328 * mapping of pages. 329 */ 330#define SYSMAP(c, p, v, n) \ 331 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 332 333 va = virtual_avail; 334 pte = vtopte(va); 335 336 /* 337 * CMAP1/CMAP2 are used for zeroing and copying pages. 338 * CMAP3 is used for the idle process page zeroing. 339 */ 340 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 341 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 342 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 343 *CMAP3 = 0; 344 345 mtx_init(&CMAPCADDR12_lock, "CMAPCADDR12", NULL, MTX_DEF); 346 347 /* 348 * Crashdump maps. 349 */ 350 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 351 352 /* 353 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 354 * XXX ptmmap is not used. 355 */ 356 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 357 358 /* 359 * msgbufp is used to map the system message buffer. 360 * XXX msgbufmap is not used. 361 */ 362 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 363 atop(round_page(MSGBUF_SIZE))) 364 365 /* 366 * ptemap is used for pmap_pte_quick 367 */ 368 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 369 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1); 370 371 virtual_avail = va; 372 373 *CMAP1 = *CMAP2 = 0; 374 for (i = 0; i < NKPT; i++) 375 PTD[i] = 0; 376 377 /* Turn on PG_G on kernel page(s) */ 378 pmap_set_pg(); 379} 380 381/* 382 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 383 */ 384void 385pmap_set_pg(void) 386{ 387 pd_entry_t pdir; 388 pt_entry_t *pte; 389 vm_offset_t va, endva; 390 int i; 391 392 if (pgeflag == 0) 393 return; 394 395 i = KERNLOAD/NBPDR; 396 endva = KERNBASE + KERNend; 397 398 if (pseflag) { 399 va = KERNBASE + KERNLOAD; 400 while (va < endva) { 401 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 402 pdir |= pgeflag; 403 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 404 invltlb(); /* Play it safe, invltlb() every time */ 405 i++; 406 va += NBPDR; 407 } 408 } else { 409 va = (vm_offset_t)btext; 410 while (va < endva) { 411 pte = vtopte(va); 412 if (*pte) 413 *pte |= pgeflag; 414 invltlb(); /* Play it safe, invltlb() every time */ 415 va += PAGE_SIZE; 416 } 417 } 418} 419 420#ifdef PAE 421static void * 422pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 423{ 424 *flags = UMA_SLAB_PRIV; 425 return (contigmalloc(PAGE_SIZE, NULL, 0, 0x0ULL, 0xffffffffULL, 1, 0)); 426} 427#endif 428 429/* 430 * Initialize the pmap module. 431 * Called by vm_init, to initialize any structures that the pmap 432 * system needs to map virtual memory. 433 * pmap_init has been enhanced to support in a fairly consistant 434 * way, discontiguous physical memory. 435 */ 436void 437pmap_init(phys_start, phys_end) 438 vm_paddr_t phys_start, phys_end; 439{ 440 int i; 441 442 /* 443 * Allocate memory for random pmap data structures. Includes the 444 * pv_head_table. 445 */ 446 447 for(i = 0; i < vm_page_array_size; i++) { 448 vm_page_t m; 449 450 m = &vm_page_array[i]; 451 TAILQ_INIT(&m->md.pv_list); 452 m->md.pv_list_count = 0; 453 } 454 455 /* 456 * init the pv free list 457 */ 458 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL, 459 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 460 uma_prealloc(pvzone, MINPV); 461 462#ifdef PAE 463 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 464 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 465 UMA_ZONE_VM | UMA_ZONE_NOFREE); 466 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 467#endif 468 469 /* 470 * Now it is safe to enable pv_table recording. 471 */ 472 pmap_initialized = TRUE; 473} 474 475/* 476 * Initialize the address space (zone) for the pv_entries. Set a 477 * high water mark so that the system can recover from excessive 478 * numbers of pv entries. 479 */ 480void 481pmap_init2() 482{ 483 int shpgperproc = PMAP_SHPGPERPROC; 484 485 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 486 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 487 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 488 pv_entry_high_water = 9 * (pv_entry_max / 10); 489 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 490} 491 492 493/*************************************************** 494 * Low level helper routines..... 495 ***************************************************/ 496 497#if defined(PMAP_DIAGNOSTIC) 498 499/* 500 * This code checks for non-writeable/modified pages. 501 * This should be an invalid condition. 502 */ 503static int 504pmap_nw_modified(pt_entry_t ptea) 505{ 506 int pte; 507 508 pte = (int) ptea; 509 510 if ((pte & (PG_M|PG_RW)) == PG_M) 511 return 1; 512 else 513 return 0; 514} 515#endif 516 517 518/* 519 * this routine defines the region(s) of memory that should 520 * not be tested for the modified bit. 521 */ 522static PMAP_INLINE int 523pmap_track_modified(vm_offset_t va) 524{ 525 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 526 return 1; 527 else 528 return 0; 529} 530 531#ifdef I386_CPU 532/* 533 * i386 only has "invalidate everything" and no SMP to worry about. 534 */ 535PMAP_INLINE void 536pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 537{ 538 539 if (pmap == kernel_pmap || pmap->pm_active) 540 invltlb(); 541} 542 543PMAP_INLINE void 544pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 545{ 546 547 if (pmap == kernel_pmap || pmap->pm_active) 548 invltlb(); 549} 550 551PMAP_INLINE void 552pmap_invalidate_all(pmap_t pmap) 553{ 554 555 if (pmap == kernel_pmap || pmap->pm_active) 556 invltlb(); 557} 558#else /* !I386_CPU */ 559#ifdef SMP 560/* 561 * For SMP, these functions have to use the IPI mechanism for coherence. 562 */ 563void 564pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 565{ 566 u_int cpumask; 567 u_int other_cpus; 568 569 if (smp_started) { 570 if (!(read_eflags() & PSL_I)) 571 panic("%s: interrupts disabled", __func__); 572 mtx_lock_spin(&smp_tlb_mtx); 573 } else 574 critical_enter(); 575 /* 576 * We need to disable interrupt preemption but MUST NOT have 577 * interrupts disabled here. 578 * XXX we may need to hold schedlock to get a coherent pm_active 579 * XXX critical sections disable interrupts again 580 */ 581 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 582 invlpg(va); 583 smp_invlpg(va); 584 } else { 585 cpumask = PCPU_GET(cpumask); 586 other_cpus = PCPU_GET(other_cpus); 587 if (pmap->pm_active & cpumask) 588 invlpg(va); 589 if (pmap->pm_active & other_cpus) 590 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 591 } 592 if (smp_started) 593 mtx_unlock_spin(&smp_tlb_mtx); 594 else 595 critical_exit(); 596} 597 598void 599pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 600{ 601 u_int cpumask; 602 u_int other_cpus; 603 vm_offset_t addr; 604 605 if (smp_started) { 606 if (!(read_eflags() & PSL_I)) 607 panic("%s: interrupts disabled", __func__); 608 mtx_lock_spin(&smp_tlb_mtx); 609 } else 610 critical_enter(); 611 /* 612 * We need to disable interrupt preemption but MUST NOT have 613 * interrupts disabled here. 614 * XXX we may need to hold schedlock to get a coherent pm_active 615 * XXX critical sections disable interrupts again 616 */ 617 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 618 for (addr = sva; addr < eva; addr += PAGE_SIZE) 619 invlpg(addr); 620 smp_invlpg_range(sva, eva); 621 } else { 622 cpumask = PCPU_GET(cpumask); 623 other_cpus = PCPU_GET(other_cpus); 624 if (pmap->pm_active & cpumask) 625 for (addr = sva; addr < eva; addr += PAGE_SIZE) 626 invlpg(addr); 627 if (pmap->pm_active & other_cpus) 628 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 629 sva, eva); 630 } 631 if (smp_started) 632 mtx_unlock_spin(&smp_tlb_mtx); 633 else 634 critical_exit(); 635} 636 637void 638pmap_invalidate_all(pmap_t pmap) 639{ 640 u_int cpumask; 641 u_int other_cpus; 642 643 if (smp_started) { 644 if (!(read_eflags() & PSL_I)) 645 panic("%s: interrupts disabled", __func__); 646 mtx_lock_spin(&smp_tlb_mtx); 647 } else 648 critical_enter(); 649 /* 650 * We need to disable interrupt preemption but MUST NOT have 651 * interrupts disabled here. 652 * XXX we may need to hold schedlock to get a coherent pm_active 653 * XXX critical sections disable interrupts again 654 */ 655 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 656 invltlb(); 657 smp_invltlb(); 658 } else { 659 cpumask = PCPU_GET(cpumask); 660 other_cpus = PCPU_GET(other_cpus); 661 if (pmap->pm_active & cpumask) 662 invltlb(); 663 if (pmap->pm_active & other_cpus) 664 smp_masked_invltlb(pmap->pm_active & other_cpus); 665 } 666 if (smp_started) 667 mtx_unlock_spin(&smp_tlb_mtx); 668 else 669 critical_exit(); 670} 671#else /* !SMP */ 672/* 673 * Normal, non-SMP, 486+ invalidation functions. 674 * We inline these within pmap.c for speed. 675 */ 676PMAP_INLINE void 677pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 678{ 679 680 if (pmap == kernel_pmap || pmap->pm_active) 681 invlpg(va); 682} 683 684PMAP_INLINE void 685pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 686{ 687 vm_offset_t addr; 688 689 if (pmap == kernel_pmap || pmap->pm_active) 690 for (addr = sva; addr < eva; addr += PAGE_SIZE) 691 invlpg(addr); 692} 693 694PMAP_INLINE void 695pmap_invalidate_all(pmap_t pmap) 696{ 697 698 if (pmap == kernel_pmap || pmap->pm_active) 699 invltlb(); 700} 701#endif /* !SMP */ 702#endif /* !I386_CPU */ 703 704/* 705 * Are we current address space or kernel? N.B. We return FALSE when 706 * a pmap's page table is in use because a kernel thread is borrowing 707 * it. The borrowed page table can change spontaneously, making any 708 * dependence on its continued use subject to a race condition. 709 */ 710static __inline int 711pmap_is_current(pmap_t pmap) 712{ 713 714 return (pmap == kernel_pmap || 715 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 716 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 717} 718 719/* 720 * If the given pmap is not the current pmap, Giant must be held. 721 */ 722pt_entry_t * 723pmap_pte(pmap_t pmap, vm_offset_t va) 724{ 725 pd_entry_t newpf; 726 pd_entry_t *pde; 727 728 pde = pmap_pde(pmap, va); 729 if (*pde & PG_PS) 730 return (pde); 731 if (*pde != 0) { 732 /* are we current address space or kernel? */ 733 if (pmap_is_current(pmap)) 734 return (vtopte(va)); 735 GIANT_REQUIRED; 736 newpf = *pde & PG_FRAME; 737 if ((*PMAP2 & PG_FRAME) != newpf) { 738 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; 739 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 740 } 741 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); 742 } 743 return (0); 744} 745 746/* 747 * Super fast pmap_pte routine best used when scanning 748 * the pv lists. This eliminates many coarse-grained 749 * invltlb calls. Note that many of the pv list 750 * scans are across different pmaps. It is very wasteful 751 * to do an entire invltlb for checking a single mapping. 752 * 753 * If the given pmap is not the current pmap, vm_page_queue_mtx 754 * must be held. 755 */ 756static pt_entry_t * 757pmap_pte_quick(pmap_t pmap, vm_offset_t va) 758{ 759 pd_entry_t newpf; 760 pd_entry_t *pde; 761 762 pde = pmap_pde(pmap, va); 763 if (*pde & PG_PS) 764 return (pde); 765 if (*pde != 0) { 766 /* are we current address space or kernel? */ 767 if (pmap_is_current(pmap)) 768 return (vtopte(va)); 769 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 770 newpf = *pde & PG_FRAME; 771 if ((*PMAP1 & PG_FRAME) != newpf) { 772 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; 773 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR1); 774 } 775 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); 776 } 777 return (0); 778} 779 780/* 781 * Routine: pmap_extract 782 * Function: 783 * Extract the physical page address associated 784 * with the given map/virtual_address pair. 785 */ 786vm_paddr_t 787pmap_extract(pmap, va) 788 register pmap_t pmap; 789 vm_offset_t va; 790{ 791 vm_paddr_t rtval; 792 pt_entry_t *pte; 793 pd_entry_t pde; 794 795 if (pmap == 0) 796 return 0; 797 pde = pmap->pm_pdir[va >> PDRSHIFT]; 798 if (pde != 0) { 799 if ((pde & PG_PS) != 0) { 800 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 801 return rtval; 802 } 803 pte = pmap_pte(pmap, va); 804 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK)); 805 return rtval; 806 } 807 return 0; 808 809} 810 811/* 812 * Routine: pmap_extract_and_hold 813 * Function: 814 * Atomically extract and hold the physical page 815 * with the given pmap and virtual address pair 816 * if that mapping permits the given protection. 817 */ 818vm_page_t 819pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 820{ 821 vm_paddr_t pa; 822 vm_page_t m; 823 824 m = NULL; 825 mtx_lock(&Giant); 826 if ((pa = pmap_extract(pmap, va)) != 0) { 827 m = PHYS_TO_VM_PAGE(pa); 828 vm_page_lock_queues(); 829 vm_page_hold(m); 830 vm_page_unlock_queues(); 831 } 832 mtx_unlock(&Giant); 833 return (m); 834} 835 836/*************************************************** 837 * Low level mapping routines..... 838 ***************************************************/ 839 840/* 841 * Add a wired page to the kva. 842 * Note: not SMP coherent. 843 */ 844PMAP_INLINE void 845pmap_kenter(vm_offset_t va, vm_paddr_t pa) 846{ 847 pt_entry_t *pte; 848 849 pte = vtopte(va); 850 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 851} 852 853/* 854 * Remove a page from the kernel pagetables. 855 * Note: not SMP coherent. 856 */ 857PMAP_INLINE void 858pmap_kremove(vm_offset_t va) 859{ 860 pt_entry_t *pte; 861 862 pte = vtopte(va); 863 pte_clear(pte); 864} 865 866/* 867 * Used to map a range of physical addresses into kernel 868 * virtual address space. 869 * 870 * The value passed in '*virt' is a suggested virtual address for 871 * the mapping. Architectures which can support a direct-mapped 872 * physical to virtual region can return the appropriate address 873 * within that region, leaving '*virt' unchanged. Other 874 * architectures should map the pages starting at '*virt' and 875 * update '*virt' with the first usable address after the mapped 876 * region. 877 */ 878vm_offset_t 879pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 880{ 881 vm_offset_t va, sva; 882 883 va = sva = *virt; 884 while (start < end) { 885 pmap_kenter(va, start); 886 va += PAGE_SIZE; 887 start += PAGE_SIZE; 888 } 889 pmap_invalidate_range(kernel_pmap, sva, va); 890 *virt = va; 891 return (sva); 892} 893 894 895/* 896 * Add a list of wired pages to the kva 897 * this routine is only used for temporary 898 * kernel mappings that do not need to have 899 * page modification or references recorded. 900 * Note that old mappings are simply written 901 * over. The page *must* be wired. 902 * Note: SMP coherent. Uses a ranged shootdown IPI. 903 */ 904void 905pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 906{ 907 vm_offset_t va; 908 909 va = sva; 910 while (count-- > 0) { 911 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 912 va += PAGE_SIZE; 913 m++; 914 } 915 pmap_invalidate_range(kernel_pmap, sva, va); 916} 917 918/* 919 * This routine tears out page mappings from the 920 * kernel -- it is meant only for temporary mappings. 921 * Note: SMP coherent. Uses a ranged shootdown IPI. 922 */ 923void 924pmap_qremove(vm_offset_t sva, int count) 925{ 926 vm_offset_t va; 927 928 va = sva; 929 while (count-- > 0) { 930 pmap_kremove(va); 931 va += PAGE_SIZE; 932 } 933 pmap_invalidate_range(kernel_pmap, sva, va); 934} 935 936/*************************************************** 937 * Page table page management routines..... 938 ***************************************************/ 939 940/* 941 * This routine unholds page table pages, and if the hold count 942 * drops to zero, then it decrements the wire count. 943 */ 944static int 945_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 946{ 947 948 while (vm_page_sleep_if_busy(m, FALSE, "pmuwpt")) 949 vm_page_lock_queues(); 950 951 if (m->hold_count == 0) { 952 vm_offset_t pteva; 953 /* 954 * unmap the page table page 955 */ 956 pmap->pm_pdir[m->pindex] = 0; 957 --pmap->pm_stats.resident_count; 958 /* 959 * We never unwire a kernel page table page, making a 960 * check for the kernel_pmap unnecessary. 961 */ 962 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)) { 963 /* 964 * Do an invltlb to make the invalidated mapping 965 * take effect immediately. 966 */ 967 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 968 pmap_invalidate_page(pmap, pteva); 969 } 970 971 /* 972 * If the page is finally unwired, simply free it. 973 */ 974 --m->wire_count; 975 if (m->wire_count == 0) { 976 vm_page_busy(m); 977 vm_page_free_zero(m); 978 atomic_subtract_int(&cnt.v_wire_count, 1); 979 } 980 return 1; 981 } 982 return 0; 983} 984 985static PMAP_INLINE int 986pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 987{ 988 vm_page_unhold(m); 989 if (m->hold_count == 0) 990 return _pmap_unwire_pte_hold(pmap, m); 991 else 992 return 0; 993} 994 995/* 996 * After removing a page table entry, this routine is used to 997 * conditionally free the page, and manage the hold/wire counts. 998 */ 999static int 1000pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 1001{ 1002 1003 if (va >= VM_MAXUSER_ADDRESS) 1004 return 0; 1005 1006 return pmap_unwire_pte_hold(pmap, mpte); 1007} 1008 1009void 1010pmap_pinit0(pmap) 1011 struct pmap *pmap; 1012{ 1013 1014 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1015#ifdef PAE 1016 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1017#endif 1018 pmap->pm_active = 0; 1019 PCPU_SET(curpmap, pmap); 1020 TAILQ_INIT(&pmap->pm_pvlist); 1021 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1022 mtx_lock_spin(&allpmaps_lock); 1023 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1024 mtx_unlock_spin(&allpmaps_lock); 1025} 1026 1027/* 1028 * Initialize a preallocated and zeroed pmap structure, 1029 * such as one in a vmspace structure. 1030 */ 1031void 1032pmap_pinit(pmap) 1033 register struct pmap *pmap; 1034{ 1035 vm_page_t m, ptdpg[NPGPTD]; 1036 vm_paddr_t pa; 1037 static int color; 1038 int i; 1039 1040 /* 1041 * No need to allocate page table space yet but we do need a valid 1042 * page directory table. 1043 */ 1044 if (pmap->pm_pdir == NULL) { 1045 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1046 NBPTD); 1047#ifdef PAE 1048 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1049 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1050 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1051 ("pmap_pinit: pdpt misaligned")); 1052 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1053 ("pmap_pinit: pdpt above 4g")); 1054#endif 1055 } 1056 1057 /* 1058 * allocate the page directory page(s) 1059 */ 1060 for (i = 0; i < NPGPTD;) { 1061 m = vm_page_alloc(NULL, color++, 1062 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1063 VM_ALLOC_ZERO); 1064 if (m == NULL) 1065 VM_WAIT; 1066 else { 1067 ptdpg[i++] = m; 1068 } 1069 } 1070 1071 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1072 1073 for (i = 0; i < NPGPTD; i++) { 1074 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1075 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1076 } 1077 1078 mtx_lock_spin(&allpmaps_lock); 1079 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1080 mtx_unlock_spin(&allpmaps_lock); 1081 /* Wire in kernel global address entries. */ 1082 /* XXX copies current process, does not fill in MPPTDI */ 1083 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1084#ifdef SMP 1085 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1086#endif 1087 1088 /* install self-referential address mapping entry(s) */ 1089 for (i = 0; i < NPGPTD; i++) { 1090 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1091 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1092#ifdef PAE 1093 pmap->pm_pdpt[i] = pa | PG_V; 1094#endif 1095 } 1096 1097 pmap->pm_active = 0; 1098 TAILQ_INIT(&pmap->pm_pvlist); 1099 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1100} 1101 1102/* 1103 * Wire in kernel global address entries. To avoid a race condition 1104 * between pmap initialization and pmap_growkernel, this procedure 1105 * should be called after the vmspace is attached to the process 1106 * but before this pmap is activated. 1107 */ 1108void 1109pmap_pinit2(pmap) 1110 struct pmap *pmap; 1111{ 1112 /* XXX: Remove this stub when no longer called */ 1113} 1114 1115/* 1116 * this routine is called if the page table page is not 1117 * mapped correctly. 1118 */ 1119static vm_page_t 1120_pmap_allocpte(pmap, ptepindex) 1121 pmap_t pmap; 1122 unsigned ptepindex; 1123{ 1124 vm_paddr_t ptepa; 1125 vm_page_t m; 1126 1127 /* 1128 * Allocate a page table page. 1129 */ 1130 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1131 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1132 VM_WAIT; 1133 /* 1134 * Indicate the need to retry. While waiting, the page table 1135 * page may have been allocated. 1136 */ 1137 return (NULL); 1138 } 1139 if ((m->flags & PG_ZERO) == 0) 1140 pmap_zero_page(m); 1141 1142 KASSERT(m->queue == PQ_NONE, 1143 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1144 1145 /* 1146 * Increment the hold count for the page table page 1147 * (denoting a new mapping.) 1148 */ 1149 m->hold_count++; 1150 1151 /* 1152 * Map the pagetable page into the process address space, if 1153 * it isn't already there. 1154 */ 1155 1156 pmap->pm_stats.resident_count++; 1157 1158 ptepa = VM_PAGE_TO_PHYS(m); 1159 pmap->pm_pdir[ptepindex] = 1160 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1161 1162 vm_page_lock_queues(); 1163 vm_page_flag_clear(m, PG_ZERO); 1164 vm_page_wakeup(m); 1165 vm_page_unlock_queues(); 1166 1167 return m; 1168} 1169 1170static vm_page_t 1171pmap_allocpte(pmap_t pmap, vm_offset_t va) 1172{ 1173 unsigned ptepindex; 1174 pd_entry_t ptepa; 1175 vm_page_t m; 1176 1177 /* 1178 * Calculate pagetable page index 1179 */ 1180 ptepindex = va >> PDRSHIFT; 1181retry: 1182 /* 1183 * Get the page directory entry 1184 */ 1185 ptepa = pmap->pm_pdir[ptepindex]; 1186 1187 /* 1188 * This supports switching from a 4MB page to a 1189 * normal 4K page. 1190 */ 1191 if (ptepa & PG_PS) { 1192 pmap->pm_pdir[ptepindex] = 0; 1193 ptepa = 0; 1194 pmap_invalidate_all(kernel_pmap); 1195 } 1196 1197 /* 1198 * If the page table page is mapped, we just increment the 1199 * hold count, and activate it. 1200 */ 1201 if (ptepa) { 1202 m = PHYS_TO_VM_PAGE(ptepa); 1203 m->hold_count++; 1204 } else { 1205 /* 1206 * Here if the pte page isn't mapped, or if it has 1207 * been deallocated. 1208 */ 1209 m = _pmap_allocpte(pmap, ptepindex); 1210 if (m == NULL) 1211 goto retry; 1212 } 1213 return (m); 1214} 1215 1216 1217/*************************************************** 1218* Pmap allocation/deallocation routines. 1219 ***************************************************/ 1220 1221#ifdef SMP 1222/* 1223 * Deal with a SMP shootdown of other users of the pmap that we are 1224 * trying to dispose of. This can be a bit hairy. 1225 */ 1226static u_int *lazymask; 1227static u_int lazyptd; 1228static volatile u_int lazywait; 1229 1230void pmap_lazyfix_action(void); 1231 1232void 1233pmap_lazyfix_action(void) 1234{ 1235 u_int mymask = PCPU_GET(cpumask); 1236 1237 if (rcr3() == lazyptd) 1238 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1239 atomic_clear_int(lazymask, mymask); 1240 atomic_store_rel_int(&lazywait, 1); 1241} 1242 1243static void 1244pmap_lazyfix_self(u_int mymask) 1245{ 1246 1247 if (rcr3() == lazyptd) 1248 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1249 atomic_clear_int(lazymask, mymask); 1250} 1251 1252 1253static void 1254pmap_lazyfix(pmap_t pmap) 1255{ 1256 u_int mymask = PCPU_GET(cpumask); 1257 u_int mask; 1258 register u_int spins; 1259 1260 while ((mask = pmap->pm_active) != 0) { 1261 spins = 50000000; 1262 mask = mask & -mask; /* Find least significant set bit */ 1263 mtx_lock_spin(&lazypmap_lock); 1264#ifdef PAE 1265 lazyptd = vtophys(pmap->pm_pdpt); 1266#else 1267 lazyptd = vtophys(pmap->pm_pdir); 1268#endif 1269 if (mask == mymask) { 1270 lazymask = &pmap->pm_active; 1271 pmap_lazyfix_self(mymask); 1272 } else { 1273 atomic_store_rel_int((u_int *)&lazymask, 1274 (u_int)&pmap->pm_active); 1275 atomic_store_rel_int(&lazywait, 0); 1276 ipi_selected(mask, IPI_LAZYPMAP); 1277 while (lazywait == 0) { 1278 ia32_pause(); 1279 if (--spins == 0) 1280 break; 1281 } 1282 } 1283 mtx_unlock_spin(&lazypmap_lock); 1284 if (spins == 0) 1285 printf("pmap_lazyfix: spun for 50000000\n"); 1286 } 1287} 1288 1289#else /* SMP */ 1290 1291/* 1292 * Cleaning up on uniprocessor is easy. For various reasons, we're 1293 * unlikely to have to even execute this code, including the fact 1294 * that the cleanup is deferred until the parent does a wait(2), which 1295 * means that another userland process has run. 1296 */ 1297static void 1298pmap_lazyfix(pmap_t pmap) 1299{ 1300 u_int cr3; 1301 1302 cr3 = vtophys(pmap->pm_pdir); 1303 if (cr3 == rcr3()) { 1304 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1305 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1306 } 1307} 1308#endif /* SMP */ 1309 1310/* 1311 * Release any resources held by the given physical map. 1312 * Called when a pmap initialized by pmap_pinit is being released. 1313 * Should only be called if the map contains no valid mappings. 1314 */ 1315void 1316pmap_release(pmap_t pmap) 1317{ 1318 vm_page_t m, ptdpg[NPGPTD]; 1319 int i; 1320 1321 KASSERT(pmap->pm_stats.resident_count == 0, 1322 ("pmap_release: pmap resident count %ld != 0", 1323 pmap->pm_stats.resident_count)); 1324 1325 pmap_lazyfix(pmap); 1326 mtx_lock_spin(&allpmaps_lock); 1327 LIST_REMOVE(pmap, pm_list); 1328 mtx_unlock_spin(&allpmaps_lock); 1329 1330 for (i = 0; i < NPGPTD; i++) 1331 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1332 1333 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1334 sizeof(*pmap->pm_pdir)); 1335#ifdef SMP 1336 pmap->pm_pdir[MPPTDI] = 0; 1337#endif 1338 1339 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1340 1341 vm_page_lock_queues(); 1342 for (i = 0; i < NPGPTD; i++) { 1343 m = ptdpg[i]; 1344#ifdef PAE 1345 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1346 ("pmap_release: got wrong ptd page")); 1347#endif 1348 m->wire_count--; 1349 atomic_subtract_int(&cnt.v_wire_count, 1); 1350 vm_page_free_zero(m); 1351 } 1352 vm_page_unlock_queues(); 1353} 1354 1355static int 1356kvm_size(SYSCTL_HANDLER_ARGS) 1357{ 1358 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1359 1360 return sysctl_handle_long(oidp, &ksize, 0, req); 1361} 1362SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1363 0, 0, kvm_size, "IU", "Size of KVM"); 1364 1365static int 1366kvm_free(SYSCTL_HANDLER_ARGS) 1367{ 1368 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1369 1370 return sysctl_handle_long(oidp, &kfree, 0, req); 1371} 1372SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1373 0, 0, kvm_free, "IU", "Amount of KVM free"); 1374 1375/* 1376 * grow the number of kernel page table entries, if needed 1377 */ 1378void 1379pmap_growkernel(vm_offset_t addr) 1380{ 1381 struct pmap *pmap; 1382 int s; 1383 vm_paddr_t ptppaddr; 1384 vm_page_t nkpg; 1385 pd_entry_t newpdir; 1386 pt_entry_t *pde; 1387 1388 s = splhigh(); 1389 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1390 if (kernel_vm_end == 0) { 1391 kernel_vm_end = KERNBASE; 1392 nkpt = 0; 1393 while (pdir_pde(PTD, kernel_vm_end)) { 1394 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1395 nkpt++; 1396 } 1397 } 1398 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1399 while (kernel_vm_end < addr) { 1400 if (pdir_pde(PTD, kernel_vm_end)) { 1401 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1402 continue; 1403 } 1404 1405 /* 1406 * This index is bogus, but out of the way 1407 */ 1408 nkpg = vm_page_alloc(NULL, nkpt, 1409 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1410 if (!nkpg) 1411 panic("pmap_growkernel: no memory to grow kernel"); 1412 1413 nkpt++; 1414 1415 pmap_zero_page(nkpg); 1416 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1417 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1418 pdir_pde(PTD, kernel_vm_end) = newpdir; 1419 1420 mtx_lock_spin(&allpmaps_lock); 1421 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1422 pde = pmap_pde(pmap, kernel_vm_end); 1423 pde_store(pde, newpdir); 1424 } 1425 mtx_unlock_spin(&allpmaps_lock); 1426 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1427 } 1428 splx(s); 1429} 1430 1431 1432/*************************************************** 1433 * page management routines. 1434 ***************************************************/ 1435 1436/* 1437 * free the pv_entry back to the free list 1438 */ 1439static PMAP_INLINE void 1440free_pv_entry(pv_entry_t pv) 1441{ 1442 pv_entry_count--; 1443 uma_zfree(pvzone, pv); 1444} 1445 1446/* 1447 * get a new pv_entry, allocating a block from the system 1448 * when needed. 1449 * the memory allocation is performed bypassing the malloc code 1450 * because of the possibility of allocations at interrupt time. 1451 */ 1452static pv_entry_t 1453get_pv_entry(void) 1454{ 1455 pv_entry_count++; 1456 if (pv_entry_high_water && 1457 (pv_entry_count > pv_entry_high_water) && 1458 (pmap_pagedaemon_waken == 0)) { 1459 pmap_pagedaemon_waken = 1; 1460 wakeup (&vm_pages_needed); 1461 } 1462 return uma_zalloc(pvzone, M_NOWAIT); 1463} 1464 1465/* 1466 * If it is the first entry on the list, it is actually 1467 * in the header and we must copy the following entry up 1468 * to the header. Otherwise we must search the list for 1469 * the entry. In either case we free the now unused entry. 1470 */ 1471 1472static int 1473pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1474{ 1475 pv_entry_t pv; 1476 int rtval; 1477 int s; 1478 1479 s = splvm(); 1480 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1481 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1482 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1483 if (pmap == pv->pv_pmap && va == pv->pv_va) 1484 break; 1485 } 1486 } else { 1487 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1488 if (va == pv->pv_va) 1489 break; 1490 } 1491 } 1492 1493 rtval = 0; 1494 if (pv) { 1495 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1496 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1497 m->md.pv_list_count--; 1498 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1499 vm_page_flag_clear(m, PG_WRITEABLE); 1500 1501 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1502 free_pv_entry(pv); 1503 } 1504 1505 splx(s); 1506 return rtval; 1507} 1508 1509/* 1510 * Create a pv entry for page at pa for 1511 * (pmap, va). 1512 */ 1513static void 1514pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1515{ 1516 1517 int s; 1518 pv_entry_t pv; 1519 1520 s = splvm(); 1521 pv = get_pv_entry(); 1522 pv->pv_va = va; 1523 pv->pv_pmap = pmap; 1524 pv->pv_ptem = mpte; 1525 1526 vm_page_lock_queues(); 1527 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1528 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1529 m->md.pv_list_count++; 1530 1531 vm_page_unlock_queues(); 1532 splx(s); 1533} 1534 1535/* 1536 * pmap_remove_pte: do the things to unmap a page in a process 1537 */ 1538static int 1539pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1540{ 1541 pt_entry_t oldpte; 1542 vm_page_t m, mpte; 1543 1544 oldpte = pte_load_clear(ptq); 1545 if (oldpte & PG_W) 1546 pmap->pm_stats.wired_count -= 1; 1547 /* 1548 * Machines that don't support invlpg, also don't support 1549 * PG_G. 1550 */ 1551 if (oldpte & PG_G) 1552 pmap_invalidate_page(kernel_pmap, va); 1553 pmap->pm_stats.resident_count -= 1; 1554 if (oldpte & PG_MANAGED) { 1555 m = PHYS_TO_VM_PAGE(oldpte); 1556 if (oldpte & PG_M) { 1557#if defined(PMAP_DIAGNOSTIC) 1558 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1559 printf( 1560 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1561 va, oldpte); 1562 } 1563#endif 1564 if (pmap_track_modified(va)) 1565 vm_page_dirty(m); 1566 } 1567 if (oldpte & PG_A) 1568 vm_page_flag_set(m, PG_REFERENCED); 1569 return pmap_remove_entry(pmap, m, va); 1570 } else { 1571 mpte = PHYS_TO_VM_PAGE(*pmap_pde(pmap, va)); 1572 return pmap_unuse_pt(pmap, va, mpte); 1573 } 1574} 1575 1576/* 1577 * Remove a single page from a process address space 1578 */ 1579static void 1580pmap_remove_page(pmap_t pmap, vm_offset_t va) 1581{ 1582 pt_entry_t *pte; 1583 1584 if ((pte = pmap_pte(pmap, va)) == NULL || *pte == 0) 1585 return; 1586 pmap_remove_pte(pmap, pte, va); 1587 pmap_invalidate_page(pmap, va); 1588} 1589 1590/* 1591 * Remove the given range of addresses from the specified map. 1592 * 1593 * It is assumed that the start and end are properly 1594 * rounded to the page size. 1595 */ 1596void 1597pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1598{ 1599 vm_offset_t pdnxt; 1600 pd_entry_t ptpaddr; 1601 pt_entry_t *pte; 1602 int anyvalid; 1603 1604 if (pmap == NULL) 1605 return; 1606 1607 if (pmap->pm_stats.resident_count == 0) 1608 return; 1609 1610 /* 1611 * special handling of removing one page. a very 1612 * common operation and easy to short circuit some 1613 * code. 1614 */ 1615 if ((sva + PAGE_SIZE == eva) && 1616 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1617 pmap_remove_page(pmap, sva); 1618 return; 1619 } 1620 1621 anyvalid = 0; 1622 1623 for (; sva < eva; sva = pdnxt) { 1624 unsigned pdirindex; 1625 1626 /* 1627 * Calculate index for next page table. 1628 */ 1629 pdnxt = (sva + NBPDR) & ~PDRMASK; 1630 if (pmap->pm_stats.resident_count == 0) 1631 break; 1632 1633 pdirindex = sva >> PDRSHIFT; 1634 ptpaddr = pmap->pm_pdir[pdirindex]; 1635 1636 /* 1637 * Weed out invalid mappings. Note: we assume that the page 1638 * directory table is always allocated, and in kernel virtual. 1639 */ 1640 if (ptpaddr == 0) 1641 continue; 1642 1643 /* 1644 * Check for large page. 1645 */ 1646 if ((ptpaddr & PG_PS) != 0) { 1647 pmap->pm_pdir[pdirindex] = 0; 1648 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1649 anyvalid = 1; 1650 continue; 1651 } 1652 1653 /* 1654 * Limit our scan to either the end of the va represented 1655 * by the current page table page, or to the end of the 1656 * range being removed. 1657 */ 1658 if (pdnxt > eva) 1659 pdnxt = eva; 1660 1661 for (; sva != pdnxt; sva += PAGE_SIZE) { 1662 if ((pte = pmap_pte(pmap, sva)) == NULL || 1663 *pte == 0) 1664 continue; 1665 anyvalid = 1; 1666 if (pmap_remove_pte(pmap, pte, sva)) 1667 break; 1668 } 1669 } 1670 1671 if (anyvalid) 1672 pmap_invalidate_all(pmap); 1673} 1674 1675/* 1676 * Routine: pmap_remove_all 1677 * Function: 1678 * Removes this physical page from 1679 * all physical maps in which it resides. 1680 * Reflects back modify bits to the pager. 1681 * 1682 * Notes: 1683 * Original versions of this routine were very 1684 * inefficient because they iteratively called 1685 * pmap_remove (slow...) 1686 */ 1687 1688void 1689pmap_remove_all(vm_page_t m) 1690{ 1691 register pv_entry_t pv; 1692 pt_entry_t *pte, tpte; 1693 int s; 1694 1695#if defined(PMAP_DIAGNOSTIC) 1696 /* 1697 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1698 */ 1699 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1700 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1701 VM_PAGE_TO_PHYS(m)); 1702 } 1703#endif 1704 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1705 s = splvm(); 1706 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1707 pv->pv_pmap->pm_stats.resident_count--; 1708 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1709 tpte = pte_load_clear(pte); 1710 if (tpte & PG_W) 1711 pv->pv_pmap->pm_stats.wired_count--; 1712 if (tpte & PG_A) 1713 vm_page_flag_set(m, PG_REFERENCED); 1714 1715 /* 1716 * Update the vm_page_t clean and reference bits. 1717 */ 1718 if (tpte & PG_M) { 1719#if defined(PMAP_DIAGNOSTIC) 1720 if (pmap_nw_modified((pt_entry_t) tpte)) { 1721 printf( 1722 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1723 pv->pv_va, tpte); 1724 } 1725#endif 1726 if (pmap_track_modified(pv->pv_va)) 1727 vm_page_dirty(m); 1728 } 1729 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1730 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1731 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1732 m->md.pv_list_count--; 1733 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 1734 free_pv_entry(pv); 1735 } 1736 vm_page_flag_clear(m, PG_WRITEABLE); 1737 splx(s); 1738} 1739 1740/* 1741 * Set the physical protection on the 1742 * specified range of this map as requested. 1743 */ 1744void 1745pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1746{ 1747 vm_offset_t pdnxt; 1748 pd_entry_t ptpaddr; 1749 int anychanged; 1750 1751 if (pmap == NULL) 1752 return; 1753 1754 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1755 pmap_remove(pmap, sva, eva); 1756 return; 1757 } 1758 1759 if (prot & VM_PROT_WRITE) 1760 return; 1761 1762 anychanged = 0; 1763 1764 for (; sva < eva; sva = pdnxt) { 1765 unsigned pdirindex; 1766 1767 pdnxt = (sva + NBPDR) & ~PDRMASK; 1768 1769 pdirindex = sva >> PDRSHIFT; 1770 ptpaddr = pmap->pm_pdir[pdirindex]; 1771 1772 /* 1773 * Weed out invalid mappings. Note: we assume that the page 1774 * directory table is always allocated, and in kernel virtual. 1775 */ 1776 if (ptpaddr == 0) 1777 continue; 1778 1779 /* 1780 * Check for large page. 1781 */ 1782 if ((ptpaddr & PG_PS) != 0) { 1783 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1784 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1785 anychanged = 1; 1786 continue; 1787 } 1788 1789 if (pdnxt > eva) 1790 pdnxt = eva; 1791 1792 for (; sva != pdnxt; sva += PAGE_SIZE) { 1793 pt_entry_t pbits; 1794 pt_entry_t *pte; 1795 vm_page_t m; 1796 1797 if ((pte = pmap_pte(pmap, sva)) == NULL) 1798 continue; 1799 pbits = *pte; 1800 if (pbits & PG_MANAGED) { 1801 m = NULL; 1802 if (pbits & PG_A) { 1803 m = PHYS_TO_VM_PAGE(pbits); 1804 vm_page_flag_set(m, PG_REFERENCED); 1805 pbits &= ~PG_A; 1806 } 1807 if ((pbits & PG_M) != 0 && 1808 pmap_track_modified(sva)) { 1809 if (m == NULL) 1810 m = PHYS_TO_VM_PAGE(pbits); 1811 vm_page_dirty(m); 1812 pbits &= ~PG_M; 1813 } 1814 } 1815 1816 pbits &= ~PG_RW; 1817 1818 if (pbits != *pte) { 1819 pte_store(pte, pbits); 1820 anychanged = 1; 1821 } 1822 } 1823 } 1824 if (anychanged) 1825 pmap_invalidate_all(pmap); 1826} 1827 1828/* 1829 * Insert the given physical page (p) at 1830 * the specified virtual address (v) in the 1831 * target physical map with the protection requested. 1832 * 1833 * If specified, the page will be wired down, meaning 1834 * that the related pte can not be reclaimed. 1835 * 1836 * NB: This is the only routine which MAY NOT lazy-evaluate 1837 * or lose information. That is, this routine must actually 1838 * insert this page into the given map NOW. 1839 */ 1840void 1841pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1842 boolean_t wired) 1843{ 1844 vm_paddr_t pa; 1845 register pt_entry_t *pte; 1846 vm_paddr_t opa; 1847 pt_entry_t origpte, newpte; 1848 vm_page_t mpte; 1849 1850 if (pmap == NULL) 1851 return; 1852 1853 va &= PG_FRAME; 1854#ifdef PMAP_DIAGNOSTIC 1855 if (va > VM_MAX_KERNEL_ADDRESS) 1856 panic("pmap_enter: toobig"); 1857 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1858 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1859#endif 1860 1861 mpte = NULL; 1862 /* 1863 * In the case that a page table page is not 1864 * resident, we are creating it here. 1865 */ 1866 if (va < VM_MAXUSER_ADDRESS) { 1867 mpte = pmap_allocpte(pmap, va); 1868 } 1869#if 0 && defined(PMAP_DIAGNOSTIC) 1870 else { 1871 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1872 origpte = *pdeaddr; 1873 if ((origpte & PG_V) == 0) { 1874 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1875 pmap->pm_pdir[PTDPTDI], origpte, va); 1876 } 1877 } 1878#endif 1879 1880 pte = pmap_pte(pmap, va); 1881 1882 /* 1883 * Page Directory table entry not valid, we need a new PT page 1884 */ 1885 if (pte == NULL) { 1886 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1887 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1888 } 1889 1890 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 1891 origpte = *pte; 1892 opa = origpte & PG_FRAME; 1893 1894 if (origpte & PG_PS) { 1895 /* 1896 * Yes, I know this will truncate upper address bits for PAE, 1897 * but I'm actually more interested in the lower bits 1898 */ 1899 printf("pmap_enter: va %p, pte %p, origpte %p\n", 1900 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 1901 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1902 } 1903 1904 /* 1905 * Mapping has not changed, must be protection or wiring change. 1906 */ 1907 if (origpte && (opa == pa)) { 1908 /* 1909 * Wiring change, just update stats. We don't worry about 1910 * wiring PT pages as they remain resident as long as there 1911 * are valid mappings in them. Hence, if a user page is wired, 1912 * the PT page will be also. 1913 */ 1914 if (wired && ((origpte & PG_W) == 0)) 1915 pmap->pm_stats.wired_count++; 1916 else if (!wired && (origpte & PG_W)) 1917 pmap->pm_stats.wired_count--; 1918 1919#if defined(PMAP_DIAGNOSTIC) 1920 if (pmap_nw_modified((pt_entry_t) origpte)) { 1921 printf( 1922 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 1923 va, origpte); 1924 } 1925#endif 1926 1927 /* 1928 * Remove extra pte reference 1929 */ 1930 if (mpte) 1931 mpte->hold_count--; 1932 1933 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) { 1934 if ((origpte & PG_RW) == 0) { 1935 pte_store(pte, origpte | PG_RW); 1936 pmap_invalidate_page(pmap, va); 1937 } 1938 return; 1939 } 1940 1941 /* 1942 * We might be turning off write access to the page, 1943 * so we go ahead and sense modify status. 1944 */ 1945 if (origpte & PG_MANAGED) { 1946 if ((origpte & PG_M) && pmap_track_modified(va)) { 1947 vm_page_t om; 1948 om = PHYS_TO_VM_PAGE(opa); 1949 vm_page_dirty(om); 1950 } 1951 pa |= PG_MANAGED; 1952 } 1953 goto validate; 1954 } 1955 /* 1956 * Mapping has changed, invalidate old range and fall through to 1957 * handle validating new mapping. 1958 */ 1959 if (opa) { 1960 int err; 1961 vm_page_lock_queues(); 1962 err = pmap_remove_pte(pmap, pte, va); 1963 vm_page_unlock_queues(); 1964 if (err) 1965 panic("pmap_enter: pte vanished, va: 0x%x", va); 1966 } 1967 1968 /* 1969 * Enter on the PV list if part of our managed memory. Note that we 1970 * raise IPL while manipulating pv_table since pmap_enter can be 1971 * called at interrupt time. 1972 */ 1973 if (pmap_initialized && 1974 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 1975 pmap_insert_entry(pmap, va, mpte, m); 1976 pa |= PG_MANAGED; 1977 } 1978 1979 /* 1980 * Increment counters 1981 */ 1982 pmap->pm_stats.resident_count++; 1983 if (wired) 1984 pmap->pm_stats.wired_count++; 1985 1986validate: 1987 /* 1988 * Now validate mapping with desired protection/wiring. 1989 */ 1990 newpte = (pt_entry_t)(pa | PG_V); 1991 if ((prot & VM_PROT_WRITE) != 0) 1992 newpte |= PG_RW; 1993 if (wired) 1994 newpte |= PG_W; 1995 if (va < VM_MAXUSER_ADDRESS) 1996 newpte |= PG_U; 1997 if (pmap == kernel_pmap) 1998 newpte |= pgeflag; 1999 2000 /* 2001 * if the mapping or permission bits are different, we need 2002 * to update the pte. 2003 */ 2004 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2005 pte_store(pte, newpte | PG_A); 2006 /*if (origpte)*/ { 2007 pmap_invalidate_page(pmap, va); 2008 } 2009 } 2010} 2011 2012/* 2013 * this code makes some *MAJOR* assumptions: 2014 * 1. Current pmap & pmap exists. 2015 * 2. Not wired. 2016 * 3. Read access. 2017 * 4. No page table pages. 2018 * 5. Tlbflush is deferred to calling procedure. 2019 * 6. Page IS managed. 2020 * but is *MUCH* faster than pmap_enter... 2021 */ 2022 2023vm_page_t 2024pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2025{ 2026 pt_entry_t *pte; 2027 vm_paddr_t pa; 2028 2029 /* 2030 * In the case that a page table page is not 2031 * resident, we are creating it here. 2032 */ 2033 if (va < VM_MAXUSER_ADDRESS) { 2034 unsigned ptepindex; 2035 pd_entry_t ptepa; 2036 2037 /* 2038 * Calculate pagetable page index 2039 */ 2040 ptepindex = va >> PDRSHIFT; 2041 if (mpte && (mpte->pindex == ptepindex)) { 2042 mpte->hold_count++; 2043 } else { 2044retry: 2045 /* 2046 * Get the page directory entry 2047 */ 2048 ptepa = pmap->pm_pdir[ptepindex]; 2049 2050 /* 2051 * If the page table page is mapped, we just increment 2052 * the hold count, and activate it. 2053 */ 2054 if (ptepa) { 2055 if (ptepa & PG_PS) 2056 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2057 mpte = PHYS_TO_VM_PAGE(ptepa); 2058 mpte->hold_count++; 2059 } else { 2060 mpte = _pmap_allocpte(pmap, ptepindex); 2061 if (mpte == NULL) 2062 goto retry; 2063 } 2064 } 2065 } else { 2066 mpte = NULL; 2067 } 2068 2069 /* 2070 * This call to vtopte makes the assumption that we are 2071 * entering the page into the current pmap. In order to support 2072 * quick entry into any pmap, one would likely use pmap_pte_quick. 2073 * But that isn't as quick as vtopte. 2074 */ 2075 pte = vtopte(va); 2076 if (*pte) { 2077 if (mpte != NULL) { 2078 vm_page_lock_queues(); 2079 pmap_unwire_pte_hold(pmap, mpte); 2080 vm_page_unlock_queues(); 2081 } 2082 return 0; 2083 } 2084 2085 /* 2086 * Enter on the PV list if part of our managed memory. Note that we 2087 * raise IPL while manipulating pv_table since pmap_enter can be 2088 * called at interrupt time. 2089 */ 2090 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2091 pmap_insert_entry(pmap, va, mpte, m); 2092 2093 /* 2094 * Increment counters 2095 */ 2096 pmap->pm_stats.resident_count++; 2097 2098 pa = VM_PAGE_TO_PHYS(m); 2099 2100 /* 2101 * Now validate mapping with RO protection 2102 */ 2103 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2104 pte_store(pte, pa | PG_V | PG_U); 2105 else 2106 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2107 2108 return mpte; 2109} 2110 2111/* 2112 * Make a temporary mapping for a physical address. This is only intended 2113 * to be used for panic dumps. 2114 */ 2115void * 2116pmap_kenter_temporary(vm_offset_t pa, int i) 2117{ 2118 vm_offset_t va; 2119 2120 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2121 pmap_kenter(va, pa); 2122#ifndef I386_CPU 2123 invlpg(va); 2124#else 2125 invltlb(); 2126#endif 2127 return ((void *)crashdumpmap); 2128} 2129 2130/* 2131 * This code maps large physical mmap regions into the 2132 * processor address space. Note that some shortcuts 2133 * are taken, but the code works. 2134 */ 2135void 2136pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2137 vm_object_t object, vm_pindex_t pindex, 2138 vm_size_t size) 2139{ 2140 vm_page_t p; 2141 2142 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2143 KASSERT(object->type == OBJT_DEVICE, 2144 ("pmap_object_init_pt: non-device object")); 2145 if (pseflag && 2146 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2147 int i; 2148 vm_page_t m[1]; 2149 unsigned int ptepindex; 2150 int npdes; 2151 pd_entry_t ptepa; 2152 2153 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2154 return; 2155retry: 2156 p = vm_page_lookup(object, pindex); 2157 if (p != NULL) { 2158 vm_page_lock_queues(); 2159 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2160 goto retry; 2161 } else { 2162 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2163 if (p == NULL) 2164 return; 2165 m[0] = p; 2166 2167 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2168 vm_page_lock_queues(); 2169 vm_page_free(p); 2170 vm_page_unlock_queues(); 2171 return; 2172 } 2173 2174 p = vm_page_lookup(object, pindex); 2175 vm_page_lock_queues(); 2176 vm_page_wakeup(p); 2177 } 2178 vm_page_unlock_queues(); 2179 2180 ptepa = VM_PAGE_TO_PHYS(p); 2181 if (ptepa & (NBPDR - 1)) 2182 return; 2183 2184 p->valid = VM_PAGE_BITS_ALL; 2185 2186 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2187 npdes = size >> PDRSHIFT; 2188 for(i = 0; i < npdes; i++) { 2189 pde_store(&pmap->pm_pdir[ptepindex], 2190 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2191 ptepa += NBPDR; 2192 ptepindex += 1; 2193 } 2194 pmap_invalidate_all(pmap); 2195 } 2196} 2197 2198/* 2199 * Routine: pmap_change_wiring 2200 * Function: Change the wiring attribute for a map/virtual-address 2201 * pair. 2202 * In/out conditions: 2203 * The mapping must already exist in the pmap. 2204 */ 2205void 2206pmap_change_wiring(pmap, va, wired) 2207 register pmap_t pmap; 2208 vm_offset_t va; 2209 boolean_t wired; 2210{ 2211 register pt_entry_t *pte; 2212 2213 if (pmap == NULL) 2214 return; 2215 2216 pte = pmap_pte(pmap, va); 2217 2218 if (wired && !pmap_pte_w(pte)) 2219 pmap->pm_stats.wired_count++; 2220 else if (!wired && pmap_pte_w(pte)) 2221 pmap->pm_stats.wired_count--; 2222 2223 /* 2224 * Wiring is not a hardware characteristic so there is no need to 2225 * invalidate TLB. 2226 */ 2227 pmap_pte_set_w(pte, wired); 2228} 2229 2230 2231 2232/* 2233 * Copy the range specified by src_addr/len 2234 * from the source map to the range dst_addr/len 2235 * in the destination map. 2236 * 2237 * This routine is only advisory and need not do anything. 2238 */ 2239 2240void 2241pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2242 vm_offset_t src_addr) 2243{ 2244 vm_offset_t addr; 2245 vm_offset_t end_addr = src_addr + len; 2246 vm_offset_t pdnxt; 2247 vm_page_t m; 2248 2249 if (dst_addr != src_addr) 2250 return; 2251 2252 if (!pmap_is_current(src_pmap)) 2253 return; 2254 2255 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2256 pt_entry_t *src_pte, *dst_pte; 2257 vm_page_t dstmpte, srcmpte; 2258 pd_entry_t srcptepaddr; 2259 unsigned ptepindex; 2260 2261 if (addr >= UPT_MIN_ADDRESS) 2262 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2263 2264 /* 2265 * Don't let optional prefaulting of pages make us go 2266 * way below the low water mark of free pages or way 2267 * above high water mark of used pv entries. 2268 */ 2269 if (cnt.v_free_count < cnt.v_free_reserved || 2270 pv_entry_count > pv_entry_high_water) 2271 break; 2272 2273 pdnxt = (addr + NBPDR) & ~PDRMASK; 2274 ptepindex = addr >> PDRSHIFT; 2275 2276 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2277 if (srcptepaddr == 0) 2278 continue; 2279 2280 if (srcptepaddr & PG_PS) { 2281 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2282 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2283 dst_pmap->pm_stats.resident_count += 2284 NBPDR / PAGE_SIZE; 2285 } 2286 continue; 2287 } 2288 2289 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2290 if (srcmpte->hold_count == 0 || (srcmpte->flags & PG_BUSY)) 2291 continue; 2292 2293 if (pdnxt > end_addr) 2294 pdnxt = end_addr; 2295 2296 src_pte = vtopte(addr); 2297 while (addr < pdnxt) { 2298 pt_entry_t ptetemp; 2299 ptetemp = *src_pte; 2300 /* 2301 * we only virtual copy managed pages 2302 */ 2303 if ((ptetemp & PG_MANAGED) != 0) { 2304 /* 2305 * We have to check after allocpte for the 2306 * pte still being around... allocpte can 2307 * block. 2308 */ 2309 dstmpte = pmap_allocpte(dst_pmap, addr); 2310 dst_pte = pmap_pte(dst_pmap, addr); 2311 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2312 /* 2313 * Clear the modified and 2314 * accessed (referenced) bits 2315 * during the copy. 2316 */ 2317 m = PHYS_TO_VM_PAGE(ptetemp); 2318 *dst_pte = ptetemp & ~(PG_M | PG_A); 2319 dst_pmap->pm_stats.resident_count++; 2320 pmap_insert_entry(dst_pmap, addr, 2321 dstmpte, m); 2322 } else { 2323 vm_page_lock_queues(); 2324 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2325 vm_page_unlock_queues(); 2326 } 2327 if (dstmpte->hold_count >= srcmpte->hold_count) 2328 break; 2329 } 2330 addr += PAGE_SIZE; 2331 src_pte++; 2332 } 2333 } 2334} 2335 2336#ifdef SMP 2337 2338/* 2339 * pmap_zpi_switchout*() 2340 * 2341 * These functions allow us to avoid doing IPIs alltogether in certain 2342 * temporary page-mapping situations (page zeroing). Instead to deal 2343 * with being preempted and moved onto a different cpu we invalidate 2344 * the page when the scheduler switches us in. This does not occur 2345 * very often so we remain relatively optimal with very little effort. 2346 */ 2347static void 2348pmap_zpi_switchout12(void) 2349{ 2350 invlpg((u_int)CADDR1); 2351 invlpg((u_int)CADDR2); 2352} 2353 2354static void 2355pmap_zpi_switchout2(void) 2356{ 2357 invlpg((u_int)CADDR2); 2358} 2359 2360static void 2361pmap_zpi_switchout3(void) 2362{ 2363 invlpg((u_int)CADDR3); 2364} 2365 2366#endif 2367 2368static __inline void 2369pagezero(void *page) 2370{ 2371#if defined(I686_CPU) 2372 if (cpu_class == CPUCLASS_686) { 2373#if defined(CPU_ENABLE_SSE) 2374 if (cpu_feature & CPUID_SSE2) 2375 sse2_pagezero(page); 2376 else 2377#endif 2378 i686_pagezero(page); 2379 } else 2380#endif 2381 bzero(page, PAGE_SIZE); 2382} 2383 2384static __inline void 2385invlcaddr(void *caddr) 2386{ 2387#ifdef I386_CPU 2388 invltlb(); 2389#else 2390 invlpg((u_int)caddr); 2391#endif 2392} 2393 2394/* 2395 * pmap_zero_page zeros the specified hardware page by mapping 2396 * the page into KVM and using bzero to clear its contents. 2397 */ 2398void 2399pmap_zero_page(vm_page_t m) 2400{ 2401 2402 mtx_lock(&CMAPCADDR12_lock); 2403 if (*CMAP2) 2404 panic("pmap_zero_page: CMAP2 busy"); 2405#ifdef SMP 2406 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2; 2407#endif 2408 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2409#ifdef SMP 2410 invlpg((u_int)CADDR2); 2411#endif 2412 pagezero(CADDR2); 2413 *CMAP2 = 0; 2414 invlcaddr(CADDR2); 2415#ifdef SMP 2416 curthread->td_pcb->pcb_switchout = NULL; 2417#endif 2418 mtx_unlock(&CMAPCADDR12_lock); 2419} 2420 2421/* 2422 * pmap_zero_page_area zeros the specified hardware page by mapping 2423 * the page into KVM and using bzero to clear its contents. 2424 * 2425 * off and size may not cover an area beyond a single hardware page. 2426 */ 2427void 2428pmap_zero_page_area(vm_page_t m, int off, int size) 2429{ 2430 2431 mtx_lock(&CMAPCADDR12_lock); 2432 if (*CMAP2) 2433 panic("pmap_zero_page: CMAP2 busy"); 2434#ifdef SMP 2435 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2; 2436#endif 2437 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2438#ifdef SMP 2439 invlpg((u_int)CADDR2); 2440#endif 2441 if (off == 0 && size == PAGE_SIZE) 2442 pagezero(CADDR2); 2443 else 2444 bzero((char *)CADDR2 + off, size); 2445 *CMAP2 = 0; 2446 invlcaddr(CADDR2); 2447#ifdef SMP 2448 curthread->td_pcb->pcb_switchout = NULL; 2449#endif 2450 mtx_unlock(&CMAPCADDR12_lock); 2451} 2452 2453/* 2454 * pmap_zero_page_idle zeros the specified hardware page by mapping 2455 * the page into KVM and using bzero to clear its contents. This 2456 * is intended to be called from the vm_pagezero process only and 2457 * outside of Giant. 2458 */ 2459void 2460pmap_zero_page_idle(vm_page_t m) 2461{ 2462 2463 if (*CMAP3) 2464 panic("pmap_zero_page: CMAP3 busy"); 2465#ifdef SMP 2466 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout3; 2467#endif 2468 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2469#ifdef SMP 2470 invlpg((u_int)CADDR3); 2471#endif 2472 pagezero(CADDR3); 2473 *CMAP3 = 0; 2474 invlcaddr(CADDR3); 2475#ifdef SMP 2476 curthread->td_pcb->pcb_switchout = NULL; 2477#endif 2478} 2479 2480/* 2481 * pmap_copy_page copies the specified (machine independent) 2482 * page by mapping the page into virtual memory and using 2483 * bcopy to copy the page, one machine dependent page at a 2484 * time. 2485 */ 2486void 2487pmap_copy_page(vm_page_t src, vm_page_t dst) 2488{ 2489 2490 mtx_lock(&CMAPCADDR12_lock); 2491 if (*CMAP1) 2492 panic("pmap_copy_page: CMAP1 busy"); 2493 if (*CMAP2) 2494 panic("pmap_copy_page: CMAP2 busy"); 2495#ifdef SMP 2496 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout12; 2497#endif 2498 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2499 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2500#ifdef SMP 2501 invlpg((u_int)CADDR1); 2502 invlpg((u_int)CADDR2); 2503#endif 2504 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2505 *CMAP1 = 0; 2506 *CMAP2 = 0; 2507#ifdef I386_CPU 2508 invltlb(); 2509#else 2510 invlpg((u_int)CADDR1); 2511 invlpg((u_int)CADDR2); 2512#endif 2513#ifdef SMP 2514 curthread->td_pcb->pcb_switchout = NULL; 2515#endif 2516 mtx_unlock(&CMAPCADDR12_lock); 2517} 2518 2519/* 2520 * Returns true if the pmap's pv is one of the first 2521 * 16 pvs linked to from this page. This count may 2522 * be changed upwards or downwards in the future; it 2523 * is only necessary that true be returned for a small 2524 * subset of pmaps for proper page aging. 2525 */ 2526boolean_t 2527pmap_page_exists_quick(pmap, m) 2528 pmap_t pmap; 2529 vm_page_t m; 2530{ 2531 pv_entry_t pv; 2532 int loops = 0; 2533 int s; 2534 2535 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2536 return FALSE; 2537 2538 s = splvm(); 2539 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2540 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2541 if (pv->pv_pmap == pmap) { 2542 splx(s); 2543 return TRUE; 2544 } 2545 loops++; 2546 if (loops >= 16) 2547 break; 2548 } 2549 splx(s); 2550 return (FALSE); 2551} 2552 2553#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2554/* 2555 * Remove all pages from specified address space 2556 * this aids process exit speeds. Also, this code 2557 * is special cased for current process only, but 2558 * can have the more generic (and slightly slower) 2559 * mode enabled. This is much faster than pmap_remove 2560 * in the case of running down an entire address space. 2561 */ 2562void 2563pmap_remove_pages(pmap, sva, eva) 2564 pmap_t pmap; 2565 vm_offset_t sva, eva; 2566{ 2567 pt_entry_t *pte, tpte; 2568 vm_page_t m; 2569 pv_entry_t pv, npv; 2570 int s; 2571 2572#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2573 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2574 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2575 return; 2576 } 2577#endif 2578 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2579 s = splvm(); 2580 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2581 2582 if (pv->pv_va >= eva || pv->pv_va < sva) { 2583 npv = TAILQ_NEXT(pv, pv_plist); 2584 continue; 2585 } 2586 2587#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2588 pte = vtopte(pv->pv_va); 2589#else 2590 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2591#endif 2592 tpte = *pte; 2593 2594 if (tpte == 0) { 2595 printf("TPTE at %p IS ZERO @ VA %08x\n", 2596 pte, pv->pv_va); 2597 panic("bad pte"); 2598 } 2599 2600/* 2601 * We cannot remove wired pages from a process' mapping at this time 2602 */ 2603 if (tpte & PG_W) { 2604 npv = TAILQ_NEXT(pv, pv_plist); 2605 continue; 2606 } 2607 2608 m = PHYS_TO_VM_PAGE(tpte); 2609 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2610 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2611 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2612 2613 KASSERT(m < &vm_page_array[vm_page_array_size], 2614 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2615 2616 pv->pv_pmap->pm_stats.resident_count--; 2617 2618 pte_clear(pte); 2619 2620 /* 2621 * Update the vm_page_t clean and reference bits. 2622 */ 2623 if (tpte & PG_M) { 2624 vm_page_dirty(m); 2625 } 2626 2627 npv = TAILQ_NEXT(pv, pv_plist); 2628 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2629 2630 m->md.pv_list_count--; 2631 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2632 if (TAILQ_FIRST(&m->md.pv_list) == NULL) { 2633 vm_page_flag_clear(m, PG_WRITEABLE); 2634 } 2635 2636 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2637 free_pv_entry(pv); 2638 } 2639 splx(s); 2640 pmap_invalidate_all(pmap); 2641} 2642 2643/* 2644 * pmap_is_modified: 2645 * 2646 * Return whether or not the specified physical page was modified 2647 * in any physical maps. 2648 */ 2649boolean_t 2650pmap_is_modified(vm_page_t m) 2651{ 2652 pv_entry_t pv; 2653 pt_entry_t *pte; 2654 int s; 2655 2656 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2657 return FALSE; 2658 2659 s = splvm(); 2660 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2661 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2662 /* 2663 * if the bit being tested is the modified bit, then 2664 * mark clean_map and ptes as never 2665 * modified. 2666 */ 2667 if (!pmap_track_modified(pv->pv_va)) 2668 continue; 2669#if defined(PMAP_DIAGNOSTIC) 2670 if (!pv->pv_pmap) { 2671 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 2672 continue; 2673 } 2674#endif 2675 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2676 if (*pte & PG_M) { 2677 splx(s); 2678 return TRUE; 2679 } 2680 } 2681 splx(s); 2682 return (FALSE); 2683} 2684 2685/* 2686 * pmap_is_prefaultable: 2687 * 2688 * Return whether or not the specified virtual address is elgible 2689 * for prefault. 2690 */ 2691boolean_t 2692pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2693{ 2694 pt_entry_t *pte; 2695 2696 if ((*pmap_pde(pmap, addr)) == 0) 2697 return (FALSE); 2698 pte = vtopte(addr); 2699 if (*pte) 2700 return (FALSE); 2701 return (TRUE); 2702} 2703 2704/* 2705 * Clear the given bit in each of the given page's ptes. 2706 */ 2707static __inline void 2708pmap_clear_ptes(vm_page_t m, int bit) 2709{ 2710 register pv_entry_t pv; 2711 pt_entry_t pbits, *pte; 2712 int s; 2713 2714 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) || 2715 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2716 return; 2717 2718 s = splvm(); 2719 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2720 /* 2721 * Loop over all current mappings setting/clearing as appropos If 2722 * setting RO do we need to clear the VAC? 2723 */ 2724 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2725 /* 2726 * don't write protect pager mappings 2727 */ 2728 if (bit == PG_RW) { 2729 if (!pmap_track_modified(pv->pv_va)) 2730 continue; 2731 } 2732 2733#if defined(PMAP_DIAGNOSTIC) 2734 if (!pv->pv_pmap) { 2735 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 2736 continue; 2737 } 2738#endif 2739 2740 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2741 pbits = *pte; 2742 if (pbits & bit) { 2743 if (bit == PG_RW) { 2744 if (pbits & PG_M) { 2745 vm_page_dirty(m); 2746 } 2747 pte_store(pte, pbits & ~(PG_M|PG_RW)); 2748 } else { 2749 pte_store(pte, pbits & ~bit); 2750 } 2751 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2752 } 2753 } 2754 if (bit == PG_RW) 2755 vm_page_flag_clear(m, PG_WRITEABLE); 2756 splx(s); 2757} 2758 2759/* 2760 * pmap_page_protect: 2761 * 2762 * Lower the permission for all mappings to a given page. 2763 */ 2764void 2765pmap_page_protect(vm_page_t m, vm_prot_t prot) 2766{ 2767 if ((prot & VM_PROT_WRITE) == 0) { 2768 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2769 pmap_clear_ptes(m, PG_RW); 2770 } else { 2771 pmap_remove_all(m); 2772 } 2773 } 2774} 2775 2776/* 2777 * pmap_ts_referenced: 2778 * 2779 * Return a count of reference bits for a page, clearing those bits. 2780 * It is not necessary for every reference bit to be cleared, but it 2781 * is necessary that 0 only be returned when there are truly no 2782 * reference bits set. 2783 * 2784 * XXX: The exact number of bits to check and clear is a matter that 2785 * should be tested and standardized at some point in the future for 2786 * optimal aging of shared pages. 2787 */ 2788int 2789pmap_ts_referenced(vm_page_t m) 2790{ 2791 register pv_entry_t pv, pvf, pvn; 2792 pt_entry_t *pte; 2793 pt_entry_t v; 2794 int s; 2795 int rtval = 0; 2796 2797 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2798 return (rtval); 2799 2800 s = splvm(); 2801 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2802 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2803 2804 pvf = pv; 2805 2806 do { 2807 pvn = TAILQ_NEXT(pv, pv_list); 2808 2809 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2810 2811 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2812 2813 if (!pmap_track_modified(pv->pv_va)) 2814 continue; 2815 2816 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2817 2818 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2819 pte_store(pte, v & ~PG_A); 2820 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2821 2822 rtval++; 2823 if (rtval > 4) { 2824 break; 2825 } 2826 } 2827 } while ((pv = pvn) != NULL && pv != pvf); 2828 } 2829 splx(s); 2830 2831 return (rtval); 2832} 2833 2834/* 2835 * Clear the modify bits on the specified physical page. 2836 */ 2837void 2838pmap_clear_modify(vm_page_t m) 2839{ 2840 pmap_clear_ptes(m, PG_M); 2841} 2842 2843/* 2844 * pmap_clear_reference: 2845 * 2846 * Clear the reference bit on the specified physical page. 2847 */ 2848void 2849pmap_clear_reference(vm_page_t m) 2850{ 2851 pmap_clear_ptes(m, PG_A); 2852} 2853 2854/* 2855 * Miscellaneous support routines follow 2856 */ 2857 2858/* 2859 * Map a set of physical memory pages into the kernel virtual 2860 * address space. Return a pointer to where it is mapped. This 2861 * routine is intended to be used for mapping device memory, 2862 * NOT real memory. 2863 */ 2864void * 2865pmap_mapdev(pa, size) 2866 vm_paddr_t pa; 2867 vm_size_t size; 2868{ 2869 vm_offset_t va, tmpva, offset; 2870 2871 offset = pa & PAGE_MASK; 2872 size = roundup(offset + size, PAGE_SIZE); 2873 pa = pa & PG_FRAME; 2874 2875 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2876 va = KERNBASE + pa; 2877 else 2878 va = kmem_alloc_nofault(kernel_map, size); 2879 if (!va) 2880 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2881 2882 for (tmpva = va; size > 0; ) { 2883 pmap_kenter(tmpva, pa); 2884 size -= PAGE_SIZE; 2885 tmpva += PAGE_SIZE; 2886 pa += PAGE_SIZE; 2887 } 2888 pmap_invalidate_range(kernel_pmap, va, tmpva); 2889 return ((void *)(va + offset)); 2890} 2891 2892void 2893pmap_unmapdev(va, size) 2894 vm_offset_t va; 2895 vm_size_t size; 2896{ 2897 vm_offset_t base, offset, tmpva; 2898 2899 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2900 return; 2901 base = va & PG_FRAME; 2902 offset = va & PAGE_MASK; 2903 size = roundup(offset + size, PAGE_SIZE); 2904 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2905 pmap_kremove(tmpva); 2906 pmap_invalidate_range(kernel_pmap, va, tmpva); 2907 kmem_free(kernel_map, base, size); 2908} 2909 2910/* 2911 * perform the pmap work for mincore 2912 */ 2913int 2914pmap_mincore(pmap, addr) 2915 pmap_t pmap; 2916 vm_offset_t addr; 2917{ 2918 pt_entry_t *ptep, pte; 2919 vm_page_t m; 2920 int val = 0; 2921 2922 ptep = pmap_pte(pmap, addr); 2923 if (ptep == 0) { 2924 return 0; 2925 } 2926 2927 if ((pte = *ptep) != 0) { 2928 vm_paddr_t pa; 2929 2930 val = MINCORE_INCORE; 2931 if ((pte & PG_MANAGED) == 0) 2932 return val; 2933 2934 pa = pte & PG_FRAME; 2935 2936 m = PHYS_TO_VM_PAGE(pa); 2937 2938 /* 2939 * Modified by us 2940 */ 2941 if (pte & PG_M) 2942 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2943 else { 2944 /* 2945 * Modified by someone else 2946 */ 2947 vm_page_lock_queues(); 2948 if (m->dirty || pmap_is_modified(m)) 2949 val |= MINCORE_MODIFIED_OTHER; 2950 vm_page_unlock_queues(); 2951 } 2952 /* 2953 * Referenced by us 2954 */ 2955 if (pte & PG_A) 2956 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2957 else { 2958 /* 2959 * Referenced by someone else 2960 */ 2961 vm_page_lock_queues(); 2962 if ((m->flags & PG_REFERENCED) || 2963 pmap_ts_referenced(m)) { 2964 val |= MINCORE_REFERENCED_OTHER; 2965 vm_page_flag_set(m, PG_REFERENCED); 2966 } 2967 vm_page_unlock_queues(); 2968 } 2969 } 2970 return val; 2971} 2972 2973void 2974pmap_activate(struct thread *td) 2975{ 2976 struct proc *p = td->td_proc; 2977 pmap_t pmap, oldpmap; 2978 u_int32_t cr3; 2979 2980 critical_enter(); 2981 pmap = vmspace_pmap(td->td_proc->p_vmspace); 2982 oldpmap = PCPU_GET(curpmap); 2983#if defined(SMP) 2984 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 2985 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 2986#else 2987 oldpmap->pm_active &= ~1; 2988 pmap->pm_active |= 1; 2989#endif 2990#ifdef PAE 2991 cr3 = vtophys(pmap->pm_pdpt); 2992#else 2993 cr3 = vtophys(pmap->pm_pdir); 2994#endif 2995 /* XXXKSE this is wrong. 2996 * pmap_activate is for the current thread on the current cpu 2997 */ 2998 if (p->p_flag & P_SA) { 2999 /* Make sure all other cr3 entries are updated. */ 3000 /* what if they are running? XXXKSE (maybe abort them) */ 3001 FOREACH_THREAD_IN_PROC(p, td) { 3002 td->td_pcb->pcb_cr3 = cr3; 3003 } 3004 } else { 3005 td->td_pcb->pcb_cr3 = cr3; 3006 } 3007 load_cr3(cr3); 3008 PCPU_SET(curpmap, pmap); 3009 critical_exit(); 3010} 3011 3012vm_offset_t 3013pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3014{ 3015 3016 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3017 return addr; 3018 } 3019 3020 addr = (addr + PDRMASK) & ~PDRMASK; 3021 return addr; 3022} 3023 3024 3025#if defined(PMAP_DEBUG) 3026pmap_pid_dump(int pid) 3027{ 3028 pmap_t pmap; 3029 struct proc *p; 3030 int npte = 0; 3031 int index; 3032 3033 sx_slock(&allproc_lock); 3034 LIST_FOREACH(p, &allproc, p_list) { 3035 if (p->p_pid != pid) 3036 continue; 3037 3038 if (p->p_vmspace) { 3039 int i,j; 3040 index = 0; 3041 pmap = vmspace_pmap(p->p_vmspace); 3042 for (i = 0; i < NPDEPTD; i++) { 3043 pd_entry_t *pde; 3044 pt_entry_t *pte; 3045 vm_offset_t base = i << PDRSHIFT; 3046 3047 pde = &pmap->pm_pdir[i]; 3048 if (pde && pmap_pde_v(pde)) { 3049 for (j = 0; j < NPTEPG; j++) { 3050 vm_offset_t va = base + (j << PAGE_SHIFT); 3051 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3052 if (index) { 3053 index = 0; 3054 printf("\n"); 3055 } 3056 sx_sunlock(&allproc_lock); 3057 return npte; 3058 } 3059 pte = pmap_pte(pmap, va); 3060 if (pte && pmap_pte_v(pte)) { 3061 pt_entry_t pa; 3062 vm_page_t m; 3063 pa = *pte; 3064 m = PHYS_TO_VM_PAGE(pa); 3065 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3066 va, pa, m->hold_count, m->wire_count, m->flags); 3067 npte++; 3068 index++; 3069 if (index >= 2) { 3070 index = 0; 3071 printf("\n"); 3072 } else { 3073 printf(" "); 3074 } 3075 } 3076 } 3077 } 3078 } 3079 } 3080 } 3081 sx_sunlock(&allproc_lock); 3082 return npte; 3083} 3084#endif 3085 3086#if defined(DEBUG) 3087 3088static void pads(pmap_t pm); 3089void pmap_pvdump(vm_offset_t pa); 3090 3091/* print address space of pmap*/ 3092static void 3093pads(pm) 3094 pmap_t pm; 3095{ 3096 int i, j; 3097 vm_paddr_t va; 3098 pt_entry_t *ptep; 3099 3100 if (pm == kernel_pmap) 3101 return; 3102 for (i = 0; i < NPDEPTD; i++) 3103 if (pm->pm_pdir[i]) 3104 for (j = 0; j < NPTEPG; j++) { 3105 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3106 if (pm == kernel_pmap && va < KERNBASE) 3107 continue; 3108 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3109 continue; 3110 ptep = pmap_pte(pm, va); 3111 if (pmap_pte_v(ptep)) 3112 printf("%x:%x ", va, *ptep); 3113 }; 3114 3115} 3116 3117void 3118pmap_pvdump(pa) 3119 vm_paddr_t pa; 3120{ 3121 pv_entry_t pv; 3122 vm_page_t m; 3123 3124 printf("pa %x", pa); 3125 m = PHYS_TO_VM_PAGE(pa); 3126 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3127 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3128 pads(pv->pv_pmap); 3129 } 3130 printf(" "); 3131} 3132#endif 3133