pmap.c revision 123710
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 123710 2003-12-22 01:01:32Z 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 vm_page_lock_queues(); 1068 vm_page_flag_clear(m, PG_BUSY); 1069 vm_page_unlock_queues(); 1070 ptdpg[i++] = m; 1071 } 1072 } 1073 1074 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1075 1076 for (i = 0; i < NPGPTD; i++) { 1077 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1078 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1079 } 1080 1081 mtx_lock_spin(&allpmaps_lock); 1082 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1083 mtx_unlock_spin(&allpmaps_lock); 1084 /* Wire in kernel global address entries. */ 1085 /* XXX copies current process, does not fill in MPPTDI */ 1086 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1087#ifdef SMP 1088 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1089#endif 1090 1091 /* install self-referential address mapping entry(s) */ 1092 for (i = 0; i < NPGPTD; i++) { 1093 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1094 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1095#ifdef PAE 1096 pmap->pm_pdpt[i] = pa | PG_V; 1097#endif 1098 } 1099 1100 pmap->pm_active = 0; 1101 TAILQ_INIT(&pmap->pm_pvlist); 1102 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1103} 1104 1105/* 1106 * Wire in kernel global address entries. To avoid a race condition 1107 * between pmap initialization and pmap_growkernel, this procedure 1108 * should be called after the vmspace is attached to the process 1109 * but before this pmap is activated. 1110 */ 1111void 1112pmap_pinit2(pmap) 1113 struct pmap *pmap; 1114{ 1115 /* XXX: Remove this stub when no longer called */ 1116} 1117 1118/* 1119 * this routine is called if the page table page is not 1120 * mapped correctly. 1121 */ 1122static vm_page_t 1123_pmap_allocpte(pmap, ptepindex) 1124 pmap_t pmap; 1125 unsigned ptepindex; 1126{ 1127 vm_paddr_t ptepa; 1128 vm_page_t m; 1129 1130 /* 1131 * Allocate a page table page. 1132 */ 1133 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1134 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1135 VM_WAIT; 1136 /* 1137 * Indicate the need to retry. While waiting, the page table 1138 * page may have been allocated. 1139 */ 1140 return (NULL); 1141 } 1142 if ((m->flags & PG_ZERO) == 0) 1143 pmap_zero_page(m); 1144 1145 KASSERT(m->queue == PQ_NONE, 1146 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1147 1148 /* 1149 * Increment the hold count for the page table page 1150 * (denoting a new mapping.) 1151 */ 1152 m->hold_count++; 1153 1154 /* 1155 * Map the pagetable page into the process address space, if 1156 * it isn't already there. 1157 */ 1158 1159 pmap->pm_stats.resident_count++; 1160 1161 ptepa = VM_PAGE_TO_PHYS(m); 1162 pmap->pm_pdir[ptepindex] = 1163 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1164 1165 vm_page_lock_queues(); 1166 vm_page_flag_clear(m, PG_ZERO); 1167 vm_page_wakeup(m); 1168 vm_page_unlock_queues(); 1169 1170 return m; 1171} 1172 1173static vm_page_t 1174pmap_allocpte(pmap_t pmap, vm_offset_t va) 1175{ 1176 unsigned ptepindex; 1177 pd_entry_t ptepa; 1178 vm_page_t m; 1179 1180 /* 1181 * Calculate pagetable page index 1182 */ 1183 ptepindex = va >> PDRSHIFT; 1184retry: 1185 /* 1186 * Get the page directory entry 1187 */ 1188 ptepa = pmap->pm_pdir[ptepindex]; 1189 1190 /* 1191 * This supports switching from a 4MB page to a 1192 * normal 4K page. 1193 */ 1194 if (ptepa & PG_PS) { 1195 pmap->pm_pdir[ptepindex] = 0; 1196 ptepa = 0; 1197 pmap_invalidate_all(kernel_pmap); 1198 } 1199 1200 /* 1201 * If the page table page is mapped, we just increment the 1202 * hold count, and activate it. 1203 */ 1204 if (ptepa) { 1205 m = PHYS_TO_VM_PAGE(ptepa); 1206 m->hold_count++; 1207 } else { 1208 /* 1209 * Here if the pte page isn't mapped, or if it has 1210 * been deallocated. 1211 */ 1212 m = _pmap_allocpte(pmap, ptepindex); 1213 if (m == NULL) 1214 goto retry; 1215 } 1216 return (m); 1217} 1218 1219 1220/*************************************************** 1221* Pmap allocation/deallocation routines. 1222 ***************************************************/ 1223 1224#ifdef SMP 1225/* 1226 * Deal with a SMP shootdown of other users of the pmap that we are 1227 * trying to dispose of. This can be a bit hairy. 1228 */ 1229static u_int *lazymask; 1230static u_int lazyptd; 1231static volatile u_int lazywait; 1232 1233void pmap_lazyfix_action(void); 1234 1235void 1236pmap_lazyfix_action(void) 1237{ 1238 u_int mymask = PCPU_GET(cpumask); 1239 1240 if (rcr3() == lazyptd) 1241 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1242 atomic_clear_int(lazymask, mymask); 1243 atomic_store_rel_int(&lazywait, 1); 1244} 1245 1246static void 1247pmap_lazyfix_self(u_int mymask) 1248{ 1249 1250 if (rcr3() == lazyptd) 1251 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1252 atomic_clear_int(lazymask, mymask); 1253} 1254 1255 1256static void 1257pmap_lazyfix(pmap_t pmap) 1258{ 1259 u_int mymask = PCPU_GET(cpumask); 1260 u_int mask; 1261 register u_int spins; 1262 1263 while ((mask = pmap->pm_active) != 0) { 1264 spins = 50000000; 1265 mask = mask & -mask; /* Find least significant set bit */ 1266 mtx_lock_spin(&lazypmap_lock); 1267#ifdef PAE 1268 lazyptd = vtophys(pmap->pm_pdpt); 1269#else 1270 lazyptd = vtophys(pmap->pm_pdir); 1271#endif 1272 if (mask == mymask) { 1273 lazymask = &pmap->pm_active; 1274 pmap_lazyfix_self(mymask); 1275 } else { 1276 atomic_store_rel_int((u_int *)&lazymask, 1277 (u_int)&pmap->pm_active); 1278 atomic_store_rel_int(&lazywait, 0); 1279 ipi_selected(mask, IPI_LAZYPMAP); 1280 while (lazywait == 0) { 1281 ia32_pause(); 1282 if (--spins == 0) 1283 break; 1284 } 1285 } 1286 mtx_unlock_spin(&lazypmap_lock); 1287 if (spins == 0) 1288 printf("pmap_lazyfix: spun for 50000000\n"); 1289 } 1290} 1291 1292#else /* SMP */ 1293 1294/* 1295 * Cleaning up on uniprocessor is easy. For various reasons, we're 1296 * unlikely to have to even execute this code, including the fact 1297 * that the cleanup is deferred until the parent does a wait(2), which 1298 * means that another userland process has run. 1299 */ 1300static void 1301pmap_lazyfix(pmap_t pmap) 1302{ 1303 u_int cr3; 1304 1305 cr3 = vtophys(pmap->pm_pdir); 1306 if (cr3 == rcr3()) { 1307 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1308 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1309 } 1310} 1311#endif /* SMP */ 1312 1313/* 1314 * Release any resources held by the given physical map. 1315 * Called when a pmap initialized by pmap_pinit is being released. 1316 * Should only be called if the map contains no valid mappings. 1317 */ 1318void 1319pmap_release(pmap_t pmap) 1320{ 1321 vm_page_t m, ptdpg[NPGPTD]; 1322 int i; 1323 1324 KASSERT(pmap->pm_stats.resident_count == 0, 1325 ("pmap_release: pmap resident count %ld != 0", 1326 pmap->pm_stats.resident_count)); 1327 1328 pmap_lazyfix(pmap); 1329 mtx_lock_spin(&allpmaps_lock); 1330 LIST_REMOVE(pmap, pm_list); 1331 mtx_unlock_spin(&allpmaps_lock); 1332 1333 for (i = 0; i < NPGPTD; i++) 1334 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1335 1336 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1337 sizeof(*pmap->pm_pdir)); 1338#ifdef SMP 1339 pmap->pm_pdir[MPPTDI] = 0; 1340#endif 1341 1342 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1343 1344 vm_page_lock_queues(); 1345 for (i = 0; i < NPGPTD; i++) { 1346 m = ptdpg[i]; 1347#ifdef PAE 1348 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1349 ("pmap_release: got wrong ptd page")); 1350#endif 1351 m->wire_count--; 1352 atomic_subtract_int(&cnt.v_wire_count, 1); 1353 vm_page_busy(m); 1354 vm_page_free_zero(m); 1355 } 1356 vm_page_unlock_queues(); 1357} 1358 1359static int 1360kvm_size(SYSCTL_HANDLER_ARGS) 1361{ 1362 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1363 1364 return sysctl_handle_long(oidp, &ksize, 0, req); 1365} 1366SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1367 0, 0, kvm_size, "IU", "Size of KVM"); 1368 1369static int 1370kvm_free(SYSCTL_HANDLER_ARGS) 1371{ 1372 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1373 1374 return sysctl_handle_long(oidp, &kfree, 0, req); 1375} 1376SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1377 0, 0, kvm_free, "IU", "Amount of KVM free"); 1378 1379/* 1380 * grow the number of kernel page table entries, if needed 1381 */ 1382void 1383pmap_growkernel(vm_offset_t addr) 1384{ 1385 struct pmap *pmap; 1386 int s; 1387 vm_paddr_t ptppaddr; 1388 vm_page_t nkpg; 1389 pd_entry_t newpdir; 1390 pt_entry_t *pde; 1391 1392 s = splhigh(); 1393 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1394 if (kernel_vm_end == 0) { 1395 kernel_vm_end = KERNBASE; 1396 nkpt = 0; 1397 while (pdir_pde(PTD, kernel_vm_end)) { 1398 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1399 nkpt++; 1400 } 1401 } 1402 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1403 while (kernel_vm_end < addr) { 1404 if (pdir_pde(PTD, kernel_vm_end)) { 1405 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1406 continue; 1407 } 1408 1409 /* 1410 * This index is bogus, but out of the way 1411 */ 1412 nkpg = vm_page_alloc(NULL, nkpt, 1413 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1414 if (!nkpg) 1415 panic("pmap_growkernel: no memory to grow kernel"); 1416 1417 nkpt++; 1418 1419 pmap_zero_page(nkpg); 1420 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1421 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1422 pdir_pde(PTD, kernel_vm_end) = newpdir; 1423 1424 mtx_lock_spin(&allpmaps_lock); 1425 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1426 pde = pmap_pde(pmap, kernel_vm_end); 1427 pde_store(pde, newpdir); 1428 } 1429 mtx_unlock_spin(&allpmaps_lock); 1430 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1431 } 1432 splx(s); 1433} 1434 1435 1436/*************************************************** 1437 * page management routines. 1438 ***************************************************/ 1439 1440/* 1441 * free the pv_entry back to the free list 1442 */ 1443static PMAP_INLINE void 1444free_pv_entry(pv_entry_t pv) 1445{ 1446 pv_entry_count--; 1447 uma_zfree(pvzone, pv); 1448} 1449 1450/* 1451 * get a new pv_entry, allocating a block from the system 1452 * when needed. 1453 * the memory allocation is performed bypassing the malloc code 1454 * because of the possibility of allocations at interrupt time. 1455 */ 1456static pv_entry_t 1457get_pv_entry(void) 1458{ 1459 pv_entry_count++; 1460 if (pv_entry_high_water && 1461 (pv_entry_count > pv_entry_high_water) && 1462 (pmap_pagedaemon_waken == 0)) { 1463 pmap_pagedaemon_waken = 1; 1464 wakeup (&vm_pages_needed); 1465 } 1466 return uma_zalloc(pvzone, M_NOWAIT); 1467} 1468 1469/* 1470 * If it is the first entry on the list, it is actually 1471 * in the header and we must copy the following entry up 1472 * to the header. Otherwise we must search the list for 1473 * the entry. In either case we free the now unused entry. 1474 */ 1475 1476static int 1477pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1478{ 1479 pv_entry_t pv; 1480 int rtval; 1481 int s; 1482 1483 s = splvm(); 1484 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1485 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1486 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1487 if (pmap == pv->pv_pmap && va == pv->pv_va) 1488 break; 1489 } 1490 } else { 1491 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1492 if (va == pv->pv_va) 1493 break; 1494 } 1495 } 1496 1497 rtval = 0; 1498 if (pv) { 1499 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1500 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1501 m->md.pv_list_count--; 1502 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1503 vm_page_flag_clear(m, PG_WRITEABLE); 1504 1505 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1506 free_pv_entry(pv); 1507 } 1508 1509 splx(s); 1510 return rtval; 1511} 1512 1513/* 1514 * Create a pv entry for page at pa for 1515 * (pmap, va). 1516 */ 1517static void 1518pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1519{ 1520 1521 int s; 1522 pv_entry_t pv; 1523 1524 s = splvm(); 1525 pv = get_pv_entry(); 1526 pv->pv_va = va; 1527 pv->pv_pmap = pmap; 1528 pv->pv_ptem = mpte; 1529 1530 vm_page_lock_queues(); 1531 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1532 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1533 m->md.pv_list_count++; 1534 1535 vm_page_unlock_queues(); 1536 splx(s); 1537} 1538 1539/* 1540 * pmap_remove_pte: do the things to unmap a page in a process 1541 */ 1542static int 1543pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1544{ 1545 pt_entry_t oldpte; 1546 vm_page_t m, mpte; 1547 1548 oldpte = pte_load_clear(ptq); 1549 if (oldpte & PG_W) 1550 pmap->pm_stats.wired_count -= 1; 1551 /* 1552 * Machines that don't support invlpg, also don't support 1553 * PG_G. 1554 */ 1555 if (oldpte & PG_G) 1556 pmap_invalidate_page(kernel_pmap, va); 1557 pmap->pm_stats.resident_count -= 1; 1558 if (oldpte & PG_MANAGED) { 1559 m = PHYS_TO_VM_PAGE(oldpte); 1560 if (oldpte & PG_M) { 1561#if defined(PMAP_DIAGNOSTIC) 1562 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1563 printf( 1564 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1565 va, oldpte); 1566 } 1567#endif 1568 if (pmap_track_modified(va)) 1569 vm_page_dirty(m); 1570 } 1571 if (oldpte & PG_A) 1572 vm_page_flag_set(m, PG_REFERENCED); 1573 return pmap_remove_entry(pmap, m, va); 1574 } else { 1575 mpte = PHYS_TO_VM_PAGE(*pmap_pde(pmap, va)); 1576 return pmap_unuse_pt(pmap, va, mpte); 1577 } 1578} 1579 1580/* 1581 * Remove a single page from a process address space 1582 */ 1583static void 1584pmap_remove_page(pmap_t pmap, vm_offset_t va) 1585{ 1586 pt_entry_t *pte; 1587 1588 if ((pte = pmap_pte(pmap, va)) == NULL || *pte == 0) 1589 return; 1590 pmap_remove_pte(pmap, pte, va); 1591 pmap_invalidate_page(pmap, va); 1592} 1593 1594/* 1595 * Remove the given range of addresses from the specified map. 1596 * 1597 * It is assumed that the start and end are properly 1598 * rounded to the page size. 1599 */ 1600void 1601pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1602{ 1603 vm_offset_t pdnxt; 1604 pd_entry_t ptpaddr; 1605 pt_entry_t *pte; 1606 int anyvalid; 1607 1608 if (pmap == NULL) 1609 return; 1610 1611 if (pmap->pm_stats.resident_count == 0) 1612 return; 1613 1614 /* 1615 * special handling of removing one page. a very 1616 * common operation and easy to short circuit some 1617 * code. 1618 */ 1619 if ((sva + PAGE_SIZE == eva) && 1620 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1621 pmap_remove_page(pmap, sva); 1622 return; 1623 } 1624 1625 anyvalid = 0; 1626 1627 for (; sva < eva; sva = pdnxt) { 1628 unsigned pdirindex; 1629 1630 /* 1631 * Calculate index for next page table. 1632 */ 1633 pdnxt = (sva + NBPDR) & ~PDRMASK; 1634 if (pmap->pm_stats.resident_count == 0) 1635 break; 1636 1637 pdirindex = sva >> PDRSHIFT; 1638 ptpaddr = pmap->pm_pdir[pdirindex]; 1639 1640 /* 1641 * Weed out invalid mappings. Note: we assume that the page 1642 * directory table is always allocated, and in kernel virtual. 1643 */ 1644 if (ptpaddr == 0) 1645 continue; 1646 1647 /* 1648 * Check for large page. 1649 */ 1650 if ((ptpaddr & PG_PS) != 0) { 1651 pmap->pm_pdir[pdirindex] = 0; 1652 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1653 anyvalid = 1; 1654 continue; 1655 } 1656 1657 /* 1658 * Limit our scan to either the end of the va represented 1659 * by the current page table page, or to the end of the 1660 * range being removed. 1661 */ 1662 if (pdnxt > eva) 1663 pdnxt = eva; 1664 1665 for (; sva != pdnxt; sva += PAGE_SIZE) { 1666 if ((pte = pmap_pte(pmap, sva)) == NULL || 1667 *pte == 0) 1668 continue; 1669 anyvalid = 1; 1670 if (pmap_remove_pte(pmap, pte, sva)) 1671 break; 1672 } 1673 } 1674 1675 if (anyvalid) 1676 pmap_invalidate_all(pmap); 1677} 1678 1679/* 1680 * Routine: pmap_remove_all 1681 * Function: 1682 * Removes this physical page from 1683 * all physical maps in which it resides. 1684 * Reflects back modify bits to the pager. 1685 * 1686 * Notes: 1687 * Original versions of this routine were very 1688 * inefficient because they iteratively called 1689 * pmap_remove (slow...) 1690 */ 1691 1692void 1693pmap_remove_all(vm_page_t m) 1694{ 1695 register pv_entry_t pv; 1696 pt_entry_t *pte, tpte; 1697 int s; 1698 1699#if defined(PMAP_DIAGNOSTIC) 1700 /* 1701 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1702 */ 1703 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1704 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1705 VM_PAGE_TO_PHYS(m)); 1706 } 1707#endif 1708 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1709 s = splvm(); 1710 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1711 pv->pv_pmap->pm_stats.resident_count--; 1712 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1713 tpte = pte_load_clear(pte); 1714 if (tpte & PG_W) 1715 pv->pv_pmap->pm_stats.wired_count--; 1716 if (tpte & PG_A) 1717 vm_page_flag_set(m, PG_REFERENCED); 1718 1719 /* 1720 * Update the vm_page_t clean and reference bits. 1721 */ 1722 if (tpte & PG_M) { 1723#if defined(PMAP_DIAGNOSTIC) 1724 if (pmap_nw_modified((pt_entry_t) tpte)) { 1725 printf( 1726 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1727 pv->pv_va, tpte); 1728 } 1729#endif 1730 if (pmap_track_modified(pv->pv_va)) 1731 vm_page_dirty(m); 1732 } 1733 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1734 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1735 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1736 m->md.pv_list_count--; 1737 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 1738 free_pv_entry(pv); 1739 } 1740 vm_page_flag_clear(m, PG_WRITEABLE); 1741 splx(s); 1742} 1743 1744/* 1745 * Set the physical protection on the 1746 * specified range of this map as requested. 1747 */ 1748void 1749pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1750{ 1751 vm_offset_t pdnxt; 1752 pd_entry_t ptpaddr; 1753 int anychanged; 1754 1755 if (pmap == NULL) 1756 return; 1757 1758 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1759 pmap_remove(pmap, sva, eva); 1760 return; 1761 } 1762 1763 if (prot & VM_PROT_WRITE) 1764 return; 1765 1766 anychanged = 0; 1767 1768 for (; sva < eva; sva = pdnxt) { 1769 unsigned pdirindex; 1770 1771 pdnxt = (sva + NBPDR) & ~PDRMASK; 1772 1773 pdirindex = sva >> PDRSHIFT; 1774 ptpaddr = pmap->pm_pdir[pdirindex]; 1775 1776 /* 1777 * Weed out invalid mappings. Note: we assume that the page 1778 * directory table is always allocated, and in kernel virtual. 1779 */ 1780 if (ptpaddr == 0) 1781 continue; 1782 1783 /* 1784 * Check for large page. 1785 */ 1786 if ((ptpaddr & PG_PS) != 0) { 1787 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1788 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1789 anychanged = 1; 1790 continue; 1791 } 1792 1793 if (pdnxt > eva) 1794 pdnxt = eva; 1795 1796 for (; sva != pdnxt; sva += PAGE_SIZE) { 1797 pt_entry_t pbits; 1798 pt_entry_t *pte; 1799 vm_page_t m; 1800 1801 if ((pte = pmap_pte(pmap, sva)) == NULL) 1802 continue; 1803 pbits = *pte; 1804 if (pbits & PG_MANAGED) { 1805 m = NULL; 1806 if (pbits & PG_A) { 1807 m = PHYS_TO_VM_PAGE(pbits); 1808 vm_page_flag_set(m, PG_REFERENCED); 1809 pbits &= ~PG_A; 1810 } 1811 if ((pbits & PG_M) != 0 && 1812 pmap_track_modified(sva)) { 1813 if (m == NULL) 1814 m = PHYS_TO_VM_PAGE(pbits); 1815 vm_page_dirty(m); 1816 pbits &= ~PG_M; 1817 } 1818 } 1819 1820 pbits &= ~PG_RW; 1821 1822 if (pbits != *pte) { 1823 pte_store(pte, pbits); 1824 anychanged = 1; 1825 } 1826 } 1827 } 1828 if (anychanged) 1829 pmap_invalidate_all(pmap); 1830} 1831 1832/* 1833 * Insert the given physical page (p) at 1834 * the specified virtual address (v) in the 1835 * target physical map with the protection requested. 1836 * 1837 * If specified, the page will be wired down, meaning 1838 * that the related pte can not be reclaimed. 1839 * 1840 * NB: This is the only routine which MAY NOT lazy-evaluate 1841 * or lose information. That is, this routine must actually 1842 * insert this page into the given map NOW. 1843 */ 1844void 1845pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1846 boolean_t wired) 1847{ 1848 vm_paddr_t pa; 1849 register pt_entry_t *pte; 1850 vm_paddr_t opa; 1851 pt_entry_t origpte, newpte; 1852 vm_page_t mpte; 1853 1854 if (pmap == NULL) 1855 return; 1856 1857 va &= PG_FRAME; 1858#ifdef PMAP_DIAGNOSTIC 1859 if (va > VM_MAX_KERNEL_ADDRESS) 1860 panic("pmap_enter: toobig"); 1861 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1862 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1863#endif 1864 1865 mpte = NULL; 1866 /* 1867 * In the case that a page table page is not 1868 * resident, we are creating it here. 1869 */ 1870 if (va < VM_MAXUSER_ADDRESS) { 1871 mpte = pmap_allocpte(pmap, va); 1872 } 1873#if 0 && defined(PMAP_DIAGNOSTIC) 1874 else { 1875 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1876 origpte = *pdeaddr; 1877 if ((origpte & PG_V) == 0) { 1878 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1879 pmap->pm_pdir[PTDPTDI], origpte, va); 1880 } 1881 } 1882#endif 1883 1884 pte = pmap_pte(pmap, va); 1885 1886 /* 1887 * Page Directory table entry not valid, we need a new PT page 1888 */ 1889 if (pte == NULL) { 1890 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1891 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1892 } 1893 1894 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 1895 origpte = *pte; 1896 opa = origpte & PG_FRAME; 1897 1898 if (origpte & PG_PS) { 1899 /* 1900 * Yes, I know this will truncate upper address bits for PAE, 1901 * but I'm actually more interested in the lower bits 1902 */ 1903 printf("pmap_enter: va %p, pte %p, origpte %p\n", 1904 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 1905 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1906 } 1907 1908 /* 1909 * Mapping has not changed, must be protection or wiring change. 1910 */ 1911 if (origpte && (opa == pa)) { 1912 /* 1913 * Wiring change, just update stats. We don't worry about 1914 * wiring PT pages as they remain resident as long as there 1915 * are valid mappings in them. Hence, if a user page is wired, 1916 * the PT page will be also. 1917 */ 1918 if (wired && ((origpte & PG_W) == 0)) 1919 pmap->pm_stats.wired_count++; 1920 else if (!wired && (origpte & PG_W)) 1921 pmap->pm_stats.wired_count--; 1922 1923#if defined(PMAP_DIAGNOSTIC) 1924 if (pmap_nw_modified((pt_entry_t) origpte)) { 1925 printf( 1926 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 1927 va, origpte); 1928 } 1929#endif 1930 1931 /* 1932 * Remove extra pte reference 1933 */ 1934 if (mpte) 1935 mpte->hold_count--; 1936 1937 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) { 1938 if ((origpte & PG_RW) == 0) { 1939 pte_store(pte, origpte | PG_RW); 1940 pmap_invalidate_page(pmap, va); 1941 } 1942 return; 1943 } 1944 1945 /* 1946 * We might be turning off write access to the page, 1947 * so we go ahead and sense modify status. 1948 */ 1949 if (origpte & PG_MANAGED) { 1950 if ((origpte & PG_M) && pmap_track_modified(va)) { 1951 vm_page_t om; 1952 om = PHYS_TO_VM_PAGE(opa); 1953 vm_page_dirty(om); 1954 } 1955 pa |= PG_MANAGED; 1956 } 1957 goto validate; 1958 } 1959 /* 1960 * Mapping has changed, invalidate old range and fall through to 1961 * handle validating new mapping. 1962 */ 1963 if (opa) { 1964 int err; 1965 vm_page_lock_queues(); 1966 err = pmap_remove_pte(pmap, pte, va); 1967 vm_page_unlock_queues(); 1968 if (err) 1969 panic("pmap_enter: pte vanished, va: 0x%x", va); 1970 } 1971 1972 /* 1973 * Enter on the PV list if part of our managed memory. Note that we 1974 * raise IPL while manipulating pv_table since pmap_enter can be 1975 * called at interrupt time. 1976 */ 1977 if (pmap_initialized && 1978 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 1979 pmap_insert_entry(pmap, va, mpte, m); 1980 pa |= PG_MANAGED; 1981 } 1982 1983 /* 1984 * Increment counters 1985 */ 1986 pmap->pm_stats.resident_count++; 1987 if (wired) 1988 pmap->pm_stats.wired_count++; 1989 1990validate: 1991 /* 1992 * Now validate mapping with desired protection/wiring. 1993 */ 1994 newpte = (pt_entry_t)(pa | PG_V); 1995 if ((prot & VM_PROT_WRITE) != 0) 1996 newpte |= PG_RW; 1997 if (wired) 1998 newpte |= PG_W; 1999 if (va < VM_MAXUSER_ADDRESS) 2000 newpte |= PG_U; 2001 if (pmap == kernel_pmap) 2002 newpte |= pgeflag; 2003 2004 /* 2005 * if the mapping or permission bits are different, we need 2006 * to update the pte. 2007 */ 2008 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2009 pte_store(pte, newpte | PG_A); 2010 /*if (origpte)*/ { 2011 pmap_invalidate_page(pmap, va); 2012 } 2013 } 2014} 2015 2016/* 2017 * this code makes some *MAJOR* assumptions: 2018 * 1. Current pmap & pmap exists. 2019 * 2. Not wired. 2020 * 3. Read access. 2021 * 4. No page table pages. 2022 * 5. Tlbflush is deferred to calling procedure. 2023 * 6. Page IS managed. 2024 * but is *MUCH* faster than pmap_enter... 2025 */ 2026 2027vm_page_t 2028pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2029{ 2030 pt_entry_t *pte; 2031 vm_paddr_t pa; 2032 2033 /* 2034 * In the case that a page table page is not 2035 * resident, we are creating it here. 2036 */ 2037 if (va < VM_MAXUSER_ADDRESS) { 2038 unsigned ptepindex; 2039 pd_entry_t ptepa; 2040 2041 /* 2042 * Calculate pagetable page index 2043 */ 2044 ptepindex = va >> PDRSHIFT; 2045 if (mpte && (mpte->pindex == ptepindex)) { 2046 mpte->hold_count++; 2047 } else { 2048retry: 2049 /* 2050 * Get the page directory entry 2051 */ 2052 ptepa = pmap->pm_pdir[ptepindex]; 2053 2054 /* 2055 * If the page table page is mapped, we just increment 2056 * the hold count, and activate it. 2057 */ 2058 if (ptepa) { 2059 if (ptepa & PG_PS) 2060 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2061 mpte = PHYS_TO_VM_PAGE(ptepa); 2062 mpte->hold_count++; 2063 } else { 2064 mpte = _pmap_allocpte(pmap, ptepindex); 2065 if (mpte == NULL) 2066 goto retry; 2067 } 2068 } 2069 } else { 2070 mpte = NULL; 2071 } 2072 2073 /* 2074 * This call to vtopte makes the assumption that we are 2075 * entering the page into the current pmap. In order to support 2076 * quick entry into any pmap, one would likely use pmap_pte_quick. 2077 * But that isn't as quick as vtopte. 2078 */ 2079 pte = vtopte(va); 2080 if (*pte) { 2081 if (mpte != NULL) { 2082 vm_page_lock_queues(); 2083 pmap_unwire_pte_hold(pmap, mpte); 2084 vm_page_unlock_queues(); 2085 } 2086 return 0; 2087 } 2088 2089 /* 2090 * Enter on the PV list if part of our managed memory. Note that we 2091 * raise IPL while manipulating pv_table since pmap_enter can be 2092 * called at interrupt time. 2093 */ 2094 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2095 pmap_insert_entry(pmap, va, mpte, m); 2096 2097 /* 2098 * Increment counters 2099 */ 2100 pmap->pm_stats.resident_count++; 2101 2102 pa = VM_PAGE_TO_PHYS(m); 2103 2104 /* 2105 * Now validate mapping with RO protection 2106 */ 2107 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2108 pte_store(pte, pa | PG_V | PG_U); 2109 else 2110 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2111 2112 return mpte; 2113} 2114 2115/* 2116 * Make a temporary mapping for a physical address. This is only intended 2117 * to be used for panic dumps. 2118 */ 2119void * 2120pmap_kenter_temporary(vm_offset_t pa, int i) 2121{ 2122 vm_offset_t va; 2123 2124 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2125 pmap_kenter(va, pa); 2126#ifndef I386_CPU 2127 invlpg(va); 2128#else 2129 invltlb(); 2130#endif 2131 return ((void *)crashdumpmap); 2132} 2133 2134/* 2135 * This code maps large physical mmap regions into the 2136 * processor address space. Note that some shortcuts 2137 * are taken, but the code works. 2138 */ 2139void 2140pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2141 vm_object_t object, vm_pindex_t pindex, 2142 vm_size_t size) 2143{ 2144 vm_page_t p; 2145 2146 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2147 KASSERT(object->type == OBJT_DEVICE, 2148 ("pmap_object_init_pt: non-device object")); 2149 if (pseflag && 2150 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2151 int i; 2152 vm_page_t m[1]; 2153 unsigned int ptepindex; 2154 int npdes; 2155 pd_entry_t ptepa; 2156 2157 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2158 return; 2159retry: 2160 p = vm_page_lookup(object, pindex); 2161 if (p != NULL) { 2162 vm_page_lock_queues(); 2163 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2164 goto retry; 2165 } else { 2166 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2167 if (p == NULL) 2168 return; 2169 m[0] = p; 2170 2171 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2172 vm_page_lock_queues(); 2173 vm_page_free(p); 2174 vm_page_unlock_queues(); 2175 return; 2176 } 2177 2178 p = vm_page_lookup(object, pindex); 2179 vm_page_lock_queues(); 2180 vm_page_wakeup(p); 2181 } 2182 vm_page_unlock_queues(); 2183 2184 ptepa = VM_PAGE_TO_PHYS(p); 2185 if (ptepa & (NBPDR - 1)) 2186 return; 2187 2188 p->valid = VM_PAGE_BITS_ALL; 2189 2190 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2191 npdes = size >> PDRSHIFT; 2192 for(i = 0; i < npdes; i++) { 2193 pde_store(&pmap->pm_pdir[ptepindex], 2194 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2195 ptepa += NBPDR; 2196 ptepindex += 1; 2197 } 2198 pmap_invalidate_all(pmap); 2199 } 2200} 2201 2202/* 2203 * Routine: pmap_change_wiring 2204 * Function: Change the wiring attribute for a map/virtual-address 2205 * pair. 2206 * In/out conditions: 2207 * The mapping must already exist in the pmap. 2208 */ 2209void 2210pmap_change_wiring(pmap, va, wired) 2211 register pmap_t pmap; 2212 vm_offset_t va; 2213 boolean_t wired; 2214{ 2215 register pt_entry_t *pte; 2216 2217 if (pmap == NULL) 2218 return; 2219 2220 pte = pmap_pte(pmap, va); 2221 2222 if (wired && !pmap_pte_w(pte)) 2223 pmap->pm_stats.wired_count++; 2224 else if (!wired && pmap_pte_w(pte)) 2225 pmap->pm_stats.wired_count--; 2226 2227 /* 2228 * Wiring is not a hardware characteristic so there is no need to 2229 * invalidate TLB. 2230 */ 2231 pmap_pte_set_w(pte, wired); 2232} 2233 2234 2235 2236/* 2237 * Copy the range specified by src_addr/len 2238 * from the source map to the range dst_addr/len 2239 * in the destination map. 2240 * 2241 * This routine is only advisory and need not do anything. 2242 */ 2243 2244void 2245pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2246 vm_offset_t src_addr) 2247{ 2248 vm_offset_t addr; 2249 vm_offset_t end_addr = src_addr + len; 2250 vm_offset_t pdnxt; 2251 vm_page_t m; 2252 2253 if (dst_addr != src_addr) 2254 return; 2255 2256 if (!pmap_is_current(src_pmap)) 2257 return; 2258 2259 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2260 pt_entry_t *src_pte, *dst_pte; 2261 vm_page_t dstmpte, srcmpte; 2262 pd_entry_t srcptepaddr; 2263 unsigned ptepindex; 2264 2265 if (addr >= UPT_MIN_ADDRESS) 2266 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2267 2268 /* 2269 * Don't let optional prefaulting of pages make us go 2270 * way below the low water mark of free pages or way 2271 * above high water mark of used pv entries. 2272 */ 2273 if (cnt.v_free_count < cnt.v_free_reserved || 2274 pv_entry_count > pv_entry_high_water) 2275 break; 2276 2277 pdnxt = (addr + NBPDR) & ~PDRMASK; 2278 ptepindex = addr >> PDRSHIFT; 2279 2280 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2281 if (srcptepaddr == 0) 2282 continue; 2283 2284 if (srcptepaddr & PG_PS) { 2285 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2286 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2287 dst_pmap->pm_stats.resident_count += 2288 NBPDR / PAGE_SIZE; 2289 } 2290 continue; 2291 } 2292 2293 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2294 if (srcmpte->hold_count == 0 || (srcmpte->flags & PG_BUSY)) 2295 continue; 2296 2297 if (pdnxt > end_addr) 2298 pdnxt = end_addr; 2299 2300 src_pte = vtopte(addr); 2301 while (addr < pdnxt) { 2302 pt_entry_t ptetemp; 2303 ptetemp = *src_pte; 2304 /* 2305 * we only virtual copy managed pages 2306 */ 2307 if ((ptetemp & PG_MANAGED) != 0) { 2308 /* 2309 * We have to check after allocpte for the 2310 * pte still being around... allocpte can 2311 * block. 2312 */ 2313 dstmpte = pmap_allocpte(dst_pmap, addr); 2314 dst_pte = pmap_pte(dst_pmap, addr); 2315 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2316 /* 2317 * Clear the modified and 2318 * accessed (referenced) bits 2319 * during the copy. 2320 */ 2321 m = PHYS_TO_VM_PAGE(ptetemp); 2322 *dst_pte = ptetemp & ~(PG_M | PG_A); 2323 dst_pmap->pm_stats.resident_count++; 2324 pmap_insert_entry(dst_pmap, addr, 2325 dstmpte, m); 2326 } else { 2327 vm_page_lock_queues(); 2328 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2329 vm_page_unlock_queues(); 2330 } 2331 if (dstmpte->hold_count >= srcmpte->hold_count) 2332 break; 2333 } 2334 addr += PAGE_SIZE; 2335 src_pte++; 2336 } 2337 } 2338} 2339 2340#ifdef SMP 2341 2342/* 2343 * pmap_zpi_switchout*() 2344 * 2345 * These functions allow us to avoid doing IPIs alltogether in certain 2346 * temporary page-mapping situations (page zeroing). Instead to deal 2347 * with being preempted and moved onto a different cpu we invalidate 2348 * the page when the scheduler switches us in. This does not occur 2349 * very often so we remain relatively optimal with very little effort. 2350 */ 2351static void 2352pmap_zpi_switchout12(void) 2353{ 2354 invlpg((u_int)CADDR1); 2355 invlpg((u_int)CADDR2); 2356} 2357 2358static void 2359pmap_zpi_switchout2(void) 2360{ 2361 invlpg((u_int)CADDR2); 2362} 2363 2364static void 2365pmap_zpi_switchout3(void) 2366{ 2367 invlpg((u_int)CADDR3); 2368} 2369 2370#endif 2371 2372static __inline void 2373pagezero(void *page) 2374{ 2375#if defined(I686_CPU) 2376 if (cpu_class == CPUCLASS_686) { 2377#if defined(CPU_ENABLE_SSE) 2378 if (cpu_feature & CPUID_SSE2) 2379 sse2_pagezero(page); 2380 else 2381#endif 2382 i686_pagezero(page); 2383 } else 2384#endif 2385 bzero(page, PAGE_SIZE); 2386} 2387 2388static __inline void 2389invlcaddr(void *caddr) 2390{ 2391#ifdef I386_CPU 2392 invltlb(); 2393#else 2394 invlpg((u_int)caddr); 2395#endif 2396} 2397 2398/* 2399 * pmap_zero_page zeros the specified hardware page by mapping 2400 * the page into KVM and using bzero to clear its contents. 2401 */ 2402void 2403pmap_zero_page(vm_page_t m) 2404{ 2405 2406 mtx_lock(&CMAPCADDR12_lock); 2407 if (*CMAP2) 2408 panic("pmap_zero_page: CMAP2 busy"); 2409#ifdef SMP 2410 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2; 2411#endif 2412 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2413#ifdef SMP 2414 invlpg((u_int)CADDR2); 2415#endif 2416 pagezero(CADDR2); 2417 *CMAP2 = 0; 2418 invlcaddr(CADDR2); 2419#ifdef SMP 2420 curthread->td_pcb->pcb_switchout = NULL; 2421#endif 2422 mtx_unlock(&CMAPCADDR12_lock); 2423} 2424 2425/* 2426 * pmap_zero_page_area zeros the specified hardware page by mapping 2427 * the page into KVM and using bzero to clear its contents. 2428 * 2429 * off and size may not cover an area beyond a single hardware page. 2430 */ 2431void 2432pmap_zero_page_area(vm_page_t m, int off, int size) 2433{ 2434 2435 mtx_lock(&CMAPCADDR12_lock); 2436 if (*CMAP2) 2437 panic("pmap_zero_page: CMAP2 busy"); 2438#ifdef SMP 2439 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2; 2440#endif 2441 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2442#ifdef SMP 2443 invlpg((u_int)CADDR2); 2444#endif 2445 if (off == 0 && size == PAGE_SIZE) 2446 pagezero(CADDR2); 2447 else 2448 bzero((char *)CADDR2 + off, size); 2449 *CMAP2 = 0; 2450 invlcaddr(CADDR2); 2451#ifdef SMP 2452 curthread->td_pcb->pcb_switchout = NULL; 2453#endif 2454 mtx_unlock(&CMAPCADDR12_lock); 2455} 2456 2457/* 2458 * pmap_zero_page_idle zeros the specified hardware page by mapping 2459 * the page into KVM and using bzero to clear its contents. This 2460 * is intended to be called from the vm_pagezero process only and 2461 * outside of Giant. 2462 */ 2463void 2464pmap_zero_page_idle(vm_page_t m) 2465{ 2466 2467 if (*CMAP3) 2468 panic("pmap_zero_page: CMAP3 busy"); 2469#ifdef SMP 2470 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout3; 2471#endif 2472 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2473#ifdef SMP 2474 invlpg((u_int)CADDR3); 2475#endif 2476 pagezero(CADDR3); 2477 *CMAP3 = 0; 2478 invlcaddr(CADDR3); 2479#ifdef SMP 2480 curthread->td_pcb->pcb_switchout = NULL; 2481#endif 2482} 2483 2484/* 2485 * pmap_copy_page copies the specified (machine independent) 2486 * page by mapping the page into virtual memory and using 2487 * bcopy to copy the page, one machine dependent page at a 2488 * time. 2489 */ 2490void 2491pmap_copy_page(vm_page_t src, vm_page_t dst) 2492{ 2493 2494 mtx_lock(&CMAPCADDR12_lock); 2495 if (*CMAP1) 2496 panic("pmap_copy_page: CMAP1 busy"); 2497 if (*CMAP2) 2498 panic("pmap_copy_page: CMAP2 busy"); 2499#ifdef SMP 2500 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout12; 2501#endif 2502 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2503 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2504#ifdef SMP 2505 invlpg((u_int)CADDR1); 2506 invlpg((u_int)CADDR2); 2507#endif 2508 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2509 *CMAP1 = 0; 2510 *CMAP2 = 0; 2511#ifdef I386_CPU 2512 invltlb(); 2513#else 2514 invlpg((u_int)CADDR1); 2515 invlpg((u_int)CADDR2); 2516#endif 2517#ifdef SMP 2518 curthread->td_pcb->pcb_switchout = NULL; 2519#endif 2520 mtx_unlock(&CMAPCADDR12_lock); 2521} 2522 2523/* 2524 * Returns true if the pmap's pv is one of the first 2525 * 16 pvs linked to from this page. This count may 2526 * be changed upwards or downwards in the future; it 2527 * is only necessary that true be returned for a small 2528 * subset of pmaps for proper page aging. 2529 */ 2530boolean_t 2531pmap_page_exists_quick(pmap, m) 2532 pmap_t pmap; 2533 vm_page_t m; 2534{ 2535 pv_entry_t pv; 2536 int loops = 0; 2537 int s; 2538 2539 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2540 return FALSE; 2541 2542 s = splvm(); 2543 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2544 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2545 if (pv->pv_pmap == pmap) { 2546 splx(s); 2547 return TRUE; 2548 } 2549 loops++; 2550 if (loops >= 16) 2551 break; 2552 } 2553 splx(s); 2554 return (FALSE); 2555} 2556 2557#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2558/* 2559 * Remove all pages from specified address space 2560 * this aids process exit speeds. Also, this code 2561 * is special cased for current process only, but 2562 * can have the more generic (and slightly slower) 2563 * mode enabled. This is much faster than pmap_remove 2564 * in the case of running down an entire address space. 2565 */ 2566void 2567pmap_remove_pages(pmap, sva, eva) 2568 pmap_t pmap; 2569 vm_offset_t sva, eva; 2570{ 2571 pt_entry_t *pte, tpte; 2572 vm_page_t m; 2573 pv_entry_t pv, npv; 2574 int s; 2575 2576#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2577 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2578 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2579 return; 2580 } 2581#endif 2582 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2583 s = splvm(); 2584 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2585 2586 if (pv->pv_va >= eva || pv->pv_va < sva) { 2587 npv = TAILQ_NEXT(pv, pv_plist); 2588 continue; 2589 } 2590 2591#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2592 pte = vtopte(pv->pv_va); 2593#else 2594 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2595#endif 2596 tpte = *pte; 2597 2598 if (tpte == 0) { 2599 printf("TPTE at %p IS ZERO @ VA %08x\n", 2600 pte, pv->pv_va); 2601 panic("bad pte"); 2602 } 2603 2604/* 2605 * We cannot remove wired pages from a process' mapping at this time 2606 */ 2607 if (tpte & PG_W) { 2608 npv = TAILQ_NEXT(pv, pv_plist); 2609 continue; 2610 } 2611 2612 m = PHYS_TO_VM_PAGE(tpte); 2613 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2614 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2615 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2616 2617 KASSERT(m < &vm_page_array[vm_page_array_size], 2618 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2619 2620 pv->pv_pmap->pm_stats.resident_count--; 2621 2622 pte_clear(pte); 2623 2624 /* 2625 * Update the vm_page_t clean and reference bits. 2626 */ 2627 if (tpte & PG_M) { 2628 vm_page_dirty(m); 2629 } 2630 2631 npv = TAILQ_NEXT(pv, pv_plist); 2632 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2633 2634 m->md.pv_list_count--; 2635 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2636 if (TAILQ_FIRST(&m->md.pv_list) == NULL) { 2637 vm_page_flag_clear(m, PG_WRITEABLE); 2638 } 2639 2640 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2641 free_pv_entry(pv); 2642 } 2643 splx(s); 2644 pmap_invalidate_all(pmap); 2645} 2646 2647/* 2648 * pmap_is_modified: 2649 * 2650 * Return whether or not the specified physical page was modified 2651 * in any physical maps. 2652 */ 2653boolean_t 2654pmap_is_modified(vm_page_t m) 2655{ 2656 pv_entry_t pv; 2657 pt_entry_t *pte; 2658 int s; 2659 2660 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2661 return FALSE; 2662 2663 s = splvm(); 2664 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2665 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2666 /* 2667 * if the bit being tested is the modified bit, then 2668 * mark clean_map and ptes as never 2669 * modified. 2670 */ 2671 if (!pmap_track_modified(pv->pv_va)) 2672 continue; 2673#if defined(PMAP_DIAGNOSTIC) 2674 if (!pv->pv_pmap) { 2675 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 2676 continue; 2677 } 2678#endif 2679 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2680 if (*pte & PG_M) { 2681 splx(s); 2682 return TRUE; 2683 } 2684 } 2685 splx(s); 2686 return (FALSE); 2687} 2688 2689/* 2690 * pmap_is_prefaultable: 2691 * 2692 * Return whether or not the specified virtual address is elgible 2693 * for prefault. 2694 */ 2695boolean_t 2696pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2697{ 2698 pt_entry_t *pte; 2699 2700 if ((*pmap_pde(pmap, addr)) == 0) 2701 return (FALSE); 2702 pte = vtopte(addr); 2703 if (*pte) 2704 return (FALSE); 2705 return (TRUE); 2706} 2707 2708/* 2709 * Clear the given bit in each of the given page's ptes. 2710 */ 2711static __inline void 2712pmap_clear_ptes(vm_page_t m, int bit) 2713{ 2714 register pv_entry_t pv; 2715 pt_entry_t pbits, *pte; 2716 int s; 2717 2718 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) || 2719 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2720 return; 2721 2722 s = splvm(); 2723 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2724 /* 2725 * Loop over all current mappings setting/clearing as appropos If 2726 * setting RO do we need to clear the VAC? 2727 */ 2728 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2729 /* 2730 * don't write protect pager mappings 2731 */ 2732 if (bit == PG_RW) { 2733 if (!pmap_track_modified(pv->pv_va)) 2734 continue; 2735 } 2736 2737#if defined(PMAP_DIAGNOSTIC) 2738 if (!pv->pv_pmap) { 2739 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 2740 continue; 2741 } 2742#endif 2743 2744 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2745 pbits = *pte; 2746 if (pbits & bit) { 2747 if (bit == PG_RW) { 2748 if (pbits & PG_M) { 2749 vm_page_dirty(m); 2750 } 2751 pte_store(pte, pbits & ~(PG_M|PG_RW)); 2752 } else { 2753 pte_store(pte, pbits & ~bit); 2754 } 2755 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2756 } 2757 } 2758 if (bit == PG_RW) 2759 vm_page_flag_clear(m, PG_WRITEABLE); 2760 splx(s); 2761} 2762 2763/* 2764 * pmap_page_protect: 2765 * 2766 * Lower the permission for all mappings to a given page. 2767 */ 2768void 2769pmap_page_protect(vm_page_t m, vm_prot_t prot) 2770{ 2771 if ((prot & VM_PROT_WRITE) == 0) { 2772 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2773 pmap_clear_ptes(m, PG_RW); 2774 } else { 2775 pmap_remove_all(m); 2776 } 2777 } 2778} 2779 2780/* 2781 * pmap_ts_referenced: 2782 * 2783 * Return a count of reference bits for a page, clearing those bits. 2784 * It is not necessary for every reference bit to be cleared, but it 2785 * is necessary that 0 only be returned when there are truly no 2786 * reference bits set. 2787 * 2788 * XXX: The exact number of bits to check and clear is a matter that 2789 * should be tested and standardized at some point in the future for 2790 * optimal aging of shared pages. 2791 */ 2792int 2793pmap_ts_referenced(vm_page_t m) 2794{ 2795 register pv_entry_t pv, pvf, pvn; 2796 pt_entry_t *pte; 2797 pt_entry_t v; 2798 int s; 2799 int rtval = 0; 2800 2801 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2802 return (rtval); 2803 2804 s = splvm(); 2805 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2806 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2807 2808 pvf = pv; 2809 2810 do { 2811 pvn = TAILQ_NEXT(pv, pv_list); 2812 2813 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2814 2815 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2816 2817 if (!pmap_track_modified(pv->pv_va)) 2818 continue; 2819 2820 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2821 2822 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2823 pte_store(pte, v & ~PG_A); 2824 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2825 2826 rtval++; 2827 if (rtval > 4) { 2828 break; 2829 } 2830 } 2831 } while ((pv = pvn) != NULL && pv != pvf); 2832 } 2833 splx(s); 2834 2835 return (rtval); 2836} 2837 2838/* 2839 * Clear the modify bits on the specified physical page. 2840 */ 2841void 2842pmap_clear_modify(vm_page_t m) 2843{ 2844 pmap_clear_ptes(m, PG_M); 2845} 2846 2847/* 2848 * pmap_clear_reference: 2849 * 2850 * Clear the reference bit on the specified physical page. 2851 */ 2852void 2853pmap_clear_reference(vm_page_t m) 2854{ 2855 pmap_clear_ptes(m, PG_A); 2856} 2857 2858/* 2859 * Miscellaneous support routines follow 2860 */ 2861 2862/* 2863 * Map a set of physical memory pages into the kernel virtual 2864 * address space. Return a pointer to where it is mapped. This 2865 * routine is intended to be used for mapping device memory, 2866 * NOT real memory. 2867 */ 2868void * 2869pmap_mapdev(pa, size) 2870 vm_paddr_t pa; 2871 vm_size_t size; 2872{ 2873 vm_offset_t va, tmpva, offset; 2874 2875 offset = pa & PAGE_MASK; 2876 size = roundup(offset + size, PAGE_SIZE); 2877 pa = pa & PG_FRAME; 2878 2879 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2880 va = KERNBASE + pa; 2881 else 2882 va = kmem_alloc_nofault(kernel_map, size); 2883 if (!va) 2884 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2885 2886 for (tmpva = va; size > 0; ) { 2887 pmap_kenter(tmpva, pa); 2888 size -= PAGE_SIZE; 2889 tmpva += PAGE_SIZE; 2890 pa += PAGE_SIZE; 2891 } 2892 pmap_invalidate_range(kernel_pmap, va, tmpva); 2893 return ((void *)(va + offset)); 2894} 2895 2896void 2897pmap_unmapdev(va, size) 2898 vm_offset_t va; 2899 vm_size_t size; 2900{ 2901 vm_offset_t base, offset, tmpva; 2902 2903 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2904 return; 2905 base = va & PG_FRAME; 2906 offset = va & PAGE_MASK; 2907 size = roundup(offset + size, PAGE_SIZE); 2908 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2909 pmap_kremove(tmpva); 2910 pmap_invalidate_range(kernel_pmap, va, tmpva); 2911 kmem_free(kernel_map, base, size); 2912} 2913 2914/* 2915 * perform the pmap work for mincore 2916 */ 2917int 2918pmap_mincore(pmap, addr) 2919 pmap_t pmap; 2920 vm_offset_t addr; 2921{ 2922 pt_entry_t *ptep, pte; 2923 vm_page_t m; 2924 int val = 0; 2925 2926 ptep = pmap_pte(pmap, addr); 2927 if (ptep == 0) { 2928 return 0; 2929 } 2930 2931 if ((pte = *ptep) != 0) { 2932 vm_paddr_t pa; 2933 2934 val = MINCORE_INCORE; 2935 if ((pte & PG_MANAGED) == 0) 2936 return val; 2937 2938 pa = pte & PG_FRAME; 2939 2940 m = PHYS_TO_VM_PAGE(pa); 2941 2942 /* 2943 * Modified by us 2944 */ 2945 if (pte & PG_M) 2946 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2947 else { 2948 /* 2949 * Modified by someone else 2950 */ 2951 vm_page_lock_queues(); 2952 if (m->dirty || pmap_is_modified(m)) 2953 val |= MINCORE_MODIFIED_OTHER; 2954 vm_page_unlock_queues(); 2955 } 2956 /* 2957 * Referenced by us 2958 */ 2959 if (pte & PG_A) 2960 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2961 else { 2962 /* 2963 * Referenced by someone else 2964 */ 2965 vm_page_lock_queues(); 2966 if ((m->flags & PG_REFERENCED) || 2967 pmap_ts_referenced(m)) { 2968 val |= MINCORE_REFERENCED_OTHER; 2969 vm_page_flag_set(m, PG_REFERENCED); 2970 } 2971 vm_page_unlock_queues(); 2972 } 2973 } 2974 return val; 2975} 2976 2977void 2978pmap_activate(struct thread *td) 2979{ 2980 struct proc *p = td->td_proc; 2981 pmap_t pmap, oldpmap; 2982 u_int32_t cr3; 2983 2984 critical_enter(); 2985 pmap = vmspace_pmap(td->td_proc->p_vmspace); 2986 oldpmap = PCPU_GET(curpmap); 2987#if defined(SMP) 2988 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 2989 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 2990#else 2991 oldpmap->pm_active &= ~1; 2992 pmap->pm_active |= 1; 2993#endif 2994#ifdef PAE 2995 cr3 = vtophys(pmap->pm_pdpt); 2996#else 2997 cr3 = vtophys(pmap->pm_pdir); 2998#endif 2999 /* XXXKSE this is wrong. 3000 * pmap_activate is for the current thread on the current cpu 3001 */ 3002 if (p->p_flag & P_SA) { 3003 /* Make sure all other cr3 entries are updated. */ 3004 /* what if they are running? XXXKSE (maybe abort them) */ 3005 FOREACH_THREAD_IN_PROC(p, td) { 3006 td->td_pcb->pcb_cr3 = cr3; 3007 } 3008 } else { 3009 td->td_pcb->pcb_cr3 = cr3; 3010 } 3011 load_cr3(cr3); 3012 PCPU_SET(curpmap, pmap); 3013 critical_exit(); 3014} 3015 3016vm_offset_t 3017pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3018{ 3019 3020 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3021 return addr; 3022 } 3023 3024 addr = (addr + PDRMASK) & ~PDRMASK; 3025 return addr; 3026} 3027 3028 3029#if defined(PMAP_DEBUG) 3030pmap_pid_dump(int pid) 3031{ 3032 pmap_t pmap; 3033 struct proc *p; 3034 int npte = 0; 3035 int index; 3036 3037 sx_slock(&allproc_lock); 3038 LIST_FOREACH(p, &allproc, p_list) { 3039 if (p->p_pid != pid) 3040 continue; 3041 3042 if (p->p_vmspace) { 3043 int i,j; 3044 index = 0; 3045 pmap = vmspace_pmap(p->p_vmspace); 3046 for (i = 0; i < NPDEPTD; i++) { 3047 pd_entry_t *pde; 3048 pt_entry_t *pte; 3049 vm_offset_t base = i << PDRSHIFT; 3050 3051 pde = &pmap->pm_pdir[i]; 3052 if (pde && pmap_pde_v(pde)) { 3053 for (j = 0; j < NPTEPG; j++) { 3054 vm_offset_t va = base + (j << PAGE_SHIFT); 3055 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3056 if (index) { 3057 index = 0; 3058 printf("\n"); 3059 } 3060 sx_sunlock(&allproc_lock); 3061 return npte; 3062 } 3063 pte = pmap_pte(pmap, va); 3064 if (pte && pmap_pte_v(pte)) { 3065 pt_entry_t pa; 3066 vm_page_t m; 3067 pa = *pte; 3068 m = PHYS_TO_VM_PAGE(pa); 3069 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3070 va, pa, m->hold_count, m->wire_count, m->flags); 3071 npte++; 3072 index++; 3073 if (index >= 2) { 3074 index = 0; 3075 printf("\n"); 3076 } else { 3077 printf(" "); 3078 } 3079 } 3080 } 3081 } 3082 } 3083 } 3084 } 3085 sx_sunlock(&allproc_lock); 3086 return npte; 3087} 3088#endif 3089 3090#if defined(DEBUG) 3091 3092static void pads(pmap_t pm); 3093void pmap_pvdump(vm_offset_t pa); 3094 3095/* print address space of pmap*/ 3096static void 3097pads(pm) 3098 pmap_t pm; 3099{ 3100 int i, j; 3101 vm_paddr_t va; 3102 pt_entry_t *ptep; 3103 3104 if (pm == kernel_pmap) 3105 return; 3106 for (i = 0; i < NPDEPTD; i++) 3107 if (pm->pm_pdir[i]) 3108 for (j = 0; j < NPTEPG; j++) { 3109 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3110 if (pm == kernel_pmap && va < KERNBASE) 3111 continue; 3112 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3113 continue; 3114 ptep = pmap_pte(pm, va); 3115 if (pmap_pte_v(ptep)) 3116 printf("%x:%x ", va, *ptep); 3117 }; 3118 3119} 3120 3121void 3122pmap_pvdump(pa) 3123 vm_paddr_t pa; 3124{ 3125 pv_entry_t pv; 3126 vm_page_t m; 3127 3128 printf("pa %x", pa); 3129 m = PHYS_TO_VM_PAGE(pa); 3130 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3131 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3132 pads(pv->pv_pmap); 3133 } 3134 printf(" "); 3135} 3136#endif 3137