pmap.c revision 91358
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 * $FreeBSD: head/sys/i386/i386/pmap.c 91358 2002-02-27 06:05:24Z peter $ 43 */ 44 45/* 46 * Manages physical address maps. 47 * 48 * In addition to hardware address maps, this 49 * module is called upon to provide software-use-only 50 * maps which may or may not be stored in the same 51 * form as hardware maps. These pseudo-maps are 52 * used to store intermediate results from copy 53 * operations to and from address spaces. 54 * 55 * Since the information managed by this module is 56 * also stored by the logical address mapping module, 57 * this module may throw away valid virtual-to-physical 58 * mappings at almost any time. However, invalidations 59 * of virtual-to-physical mappings must be done as 60 * requested. 61 * 62 * In order to cope with hardware architectures which 63 * make virtual-to-physical map invalidates expensive, 64 * this module may delay invalidate or reduced protection 65 * operations until such time as they are actually 66 * necessary. This module is given full information as 67 * to which processors are currently using which maps, 68 * and to when physical maps must be made correct. 69 */ 70 71#include "opt_disable_pse.h" 72#include "opt_pmap.h" 73#include "opt_msgbuf.h" 74#include "opt_kstack_pages.h" 75 76#include <sys/param.h> 77#include <sys/systm.h> 78#include <sys/kernel.h> 79#include <sys/lock.h> 80#include <sys/mman.h> 81#include <sys/msgbuf.h> 82#include <sys/mutex.h> 83#include <sys/proc.h> 84#include <sys/sx.h> 85#include <sys/user.h> 86#include <sys/vmmeter.h> 87#include <sys/sysctl.h> 88#if defined(SMP) 89#include <sys/smp.h> 90#endif 91 92#include <vm/vm.h> 93#include <vm/vm_param.h> 94#include <vm/vm_kern.h> 95#include <vm/vm_page.h> 96#include <vm/vm_map.h> 97#include <vm/vm_object.h> 98#include <vm/vm_extern.h> 99#include <vm/vm_pageout.h> 100#include <vm/vm_pager.h> 101#include <vm/vm_zone.h> 102 103#include <machine/cputypes.h> 104#include <machine/md_var.h> 105#include <machine/specialreg.h> 106#if defined(SMP) || defined(APIC_IO) 107#include <machine/apic.h> 108#include <machine/segments.h> 109#include <machine/tss.h> 110#endif /* SMP || APIC_IO */ 111 112#define PMAP_KEEP_PDIRS 113#ifndef PMAP_SHPGPERPROC 114#define PMAP_SHPGPERPROC 200 115#endif 116 117#if defined(DIAGNOSTIC) 118#define PMAP_DIAGNOSTIC 119#endif 120 121#define MINPV 2048 122 123#if !defined(PMAP_DIAGNOSTIC) 124#define PMAP_INLINE __inline 125#else 126#define PMAP_INLINE 127#endif 128 129/* 130 * Get PDEs and PTEs for user/kernel address space 131 */ 132#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 133#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 134 135#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 136#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 137#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 138#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 139#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 140 141#define pmap_pte_set_w(pte, v) ((v)?(*(int *)pte |= PG_W):(*(int *)pte &= ~PG_W)) 142#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 143 144/* 145 * Given a map and a machine independent protection code, 146 * convert to a vax protection code. 147 */ 148#define pte_prot(m, p) (protection_codes[p]) 149static int protection_codes[8]; 150 151static struct pmap kernel_pmap_store; 152pmap_t kernel_pmap; 153LIST_HEAD(pmaplist, pmap); 154struct pmaplist allpmaps; 155 156vm_offset_t avail_start; /* PA of first available physical page */ 157vm_offset_t avail_end; /* PA of last available physical page */ 158vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 159vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 160static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 161static int pgeflag; /* PG_G or-in */ 162static int pseflag; /* PG_PS or-in */ 163 164static vm_object_t kptobj; 165 166static int nkpt; 167vm_offset_t kernel_vm_end; 168 169/* 170 * Data for the pv entry allocation mechanism 171 */ 172static vm_zone_t pvzone; 173static struct vm_zone pvzone_store; 174static struct vm_object pvzone_obj; 175static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 176static int pmap_pagedaemon_waken = 0; 177static struct pv_entry *pvinit; 178 179/* 180 * All those kernel PT submaps that BSD is so fond of 181 */ 182pt_entry_t *CMAP1 = 0; 183static pt_entry_t *CMAP2, *ptmmap; 184caddr_t CADDR1 = 0, ptvmmap = 0; 185static caddr_t CADDR2; 186static pt_entry_t *msgbufmap; 187struct msgbuf *msgbufp = 0; 188 189/* 190 * Crashdump maps. 191 */ 192static pt_entry_t *pt_crashdumpmap; 193static caddr_t crashdumpmap; 194 195#ifdef SMP 196extern pt_entry_t *SMPpt; 197#endif 198static pt_entry_t *PMAP1 = 0; 199static pt_entry_t *PADDR1 = 0; 200 201static PMAP_INLINE void free_pv_entry __P((pv_entry_t pv)); 202static pt_entry_t *get_ptbase __P((pmap_t pmap)); 203static pv_entry_t get_pv_entry __P((void)); 204static void i386_protection_init __P((void)); 205static __inline void pmap_changebit __P((vm_page_t m, int bit, boolean_t setem)); 206 207static void pmap_remove_all __P((vm_page_t m)); 208static vm_page_t pmap_enter_quick __P((pmap_t pmap, vm_offset_t va, 209 vm_page_t m, vm_page_t mpte)); 210static int pmap_remove_pte __P((pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva)); 211static void pmap_remove_page __P((struct pmap *pmap, vm_offset_t va)); 212static int pmap_remove_entry __P((struct pmap *pmap, vm_page_t m, 213 vm_offset_t va)); 214static boolean_t pmap_testbit __P((vm_page_t m, int bit)); 215static void pmap_insert_entry __P((pmap_t pmap, vm_offset_t va, 216 vm_page_t mpte, vm_page_t m)); 217 218static vm_page_t pmap_allocpte __P((pmap_t pmap, vm_offset_t va)); 219 220static int pmap_release_free_page __P((pmap_t pmap, vm_page_t p)); 221static vm_page_t _pmap_allocpte __P((pmap_t pmap, unsigned ptepindex)); 222static pt_entry_t *pmap_pte_quick __P((pmap_t pmap, vm_offset_t va)); 223static vm_page_t pmap_page_lookup __P((vm_object_t object, vm_pindex_t pindex)); 224static int pmap_unuse_pt __P((pmap_t, vm_offset_t, vm_page_t)); 225static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 226 227static pd_entry_t pdir4mb; 228 229/* 230 * Routine: pmap_pte 231 * Function: 232 * Extract the page table entry associated 233 * with the given map/virtual_address pair. 234 */ 235 236PMAP_INLINE pt_entry_t * 237pmap_pte(pmap, va) 238 register pmap_t pmap; 239 vm_offset_t va; 240{ 241 pd_entry_t *pdeaddr; 242 243 if (pmap) { 244 pdeaddr = pmap_pde(pmap, va); 245 if (*pdeaddr & PG_PS) 246 return pdeaddr; 247 if (*pdeaddr) { 248 return get_ptbase(pmap) + i386_btop(va); 249 } 250 } 251 return (0); 252} 253 254/* 255 * Move the kernel virtual free pointer to the next 256 * 4MB. This is used to help improve performance 257 * by using a large (4MB) page for much of the kernel 258 * (.text, .data, .bss) 259 */ 260static vm_offset_t 261pmap_kmem_choose(vm_offset_t addr) 262{ 263 vm_offset_t newaddr = addr; 264 265#ifndef DISABLE_PSE 266 if (cpu_feature & CPUID_PSE) 267 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); 268#endif 269 return newaddr; 270} 271 272/* 273 * Bootstrap the system enough to run with virtual memory. 274 * 275 * On the i386 this is called after mapping has already been enabled 276 * and just syncs the pmap module with what has already been done. 277 * [We can't call it easily with mapping off since the kernel is not 278 * mapped with PA == VA, hence we would have to relocate every address 279 * from the linked base (virtual) address "KERNBASE" to the actual 280 * (physical) address starting relative to 0] 281 */ 282void 283pmap_bootstrap(firstaddr, loadaddr) 284 vm_offset_t firstaddr; 285 vm_offset_t loadaddr; 286{ 287 vm_offset_t va; 288 pt_entry_t *pte; 289 int i; 290 291 avail_start = firstaddr; 292 293 /* 294 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 295 * large. It should instead be correctly calculated in locore.s and 296 * not based on 'first' (which is a physical address, not a virtual 297 * address, for the start of unused physical memory). The kernel 298 * page tables are NOT double mapped and thus should not be included 299 * in this calculation. 300 */ 301 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 302 virtual_avail = pmap_kmem_choose(virtual_avail); 303 304 virtual_end = VM_MAX_KERNEL_ADDRESS; 305 306 /* 307 * Initialize protection array. 308 */ 309 i386_protection_init(); 310 311 /* 312 * The kernel's pmap is statically allocated so we don't have to use 313 * pmap_create, which is unlikely to work correctly at this part of 314 * the boot sequence (XXX and which no longer exists). 315 */ 316 kernel_pmap = &kernel_pmap_store; 317 318 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 319 kernel_pmap->pm_count = 1; 320 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 321 TAILQ_INIT(&kernel_pmap->pm_pvlist); 322 LIST_INIT(&allpmaps); 323 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 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 = (pt_entry_t *) pmap_pte(kernel_pmap, va); 335 336 /* 337 * CMAP1/CMAP2 are used for zeroing and copying pages. 338 */ 339 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 340 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 341 342 /* 343 * Crashdump maps. 344 */ 345 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 346 347 /* 348 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 349 * XXX ptmmap is not used. 350 */ 351 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 352 353 /* 354 * msgbufp is used to map the system message buffer. 355 * XXX msgbufmap is not used. 356 */ 357 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 358 atop(round_page(MSGBUF_SIZE))) 359 360 /* 361 * ptemap is used for pmap_pte_quick 362 */ 363 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 364 365 virtual_avail = va; 366 367 *CMAP1 = *CMAP2 = 0; 368 for (i = 0; i < NKPT; i++) 369 PTD[i] = 0; 370 371 pgeflag = 0; 372#if /* !defined(SMP) || */ defined(ENABLE_PG_G) 373 if (cpu_feature & CPUID_PGE) 374 pgeflag = PG_G; 375#endif 376 377/* 378 * Initialize the 4MB page size flag 379 */ 380 pseflag = 0; 381/* 382 * The 4MB page version of the initial 383 * kernel page mapping. 384 */ 385 pdir4mb = 0; 386 387#ifndef DISABLE_PSE 388 if (cpu_feature & CPUID_PSE) { 389 pd_entry_t ptditmp; 390 /* 391 * Note that we have enabled PSE mode 392 */ 393 pseflag = PG_PS; 394 ptditmp = *(PTmap + i386_btop(KERNBASE)); 395 ptditmp &= ~(NBPDR - 1); 396 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag; 397 pdir4mb = ptditmp; 398 } 399#endif 400#ifndef SMP 401 /* 402 * Turn on PGE/PSE. SMP does this later on since the 403 * 4K page tables are required for AP boot (for now). 404 * XXX fixme. 405 */ 406 pmap_set_opt(); 407#endif 408#ifdef SMP 409 if (cpu_apic_address == 0) 410 panic("pmap_bootstrap: no local apic! (non-SMP hardware?)"); 411 /* local apic is mapped on last page */ 412 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag | 413 (cpu_apic_address & PG_FRAME)); 414#endif 415 cpu_invltlb(); 416} 417 418/* 419 * Enable 4MB page mode for MP startup. Turn on PG_G support. 420 * BSP will run this after all the AP's have started up. 421 */ 422void 423pmap_set_opt(void) 424{ 425 pt_entry_t *pte; 426 vm_offset_t va; 427 428 if (pgeflag && (cpu_feature & CPUID_PGE)) 429 load_cr4(rcr4() | CR4_PGE); 430#ifndef DISABLE_PSE 431 if (pseflag && (cpu_feature & CPUID_PSE)) 432 load_cr4(rcr4() | CR4_PSE); 433#endif 434 if (PCPU_GET(cpuid) == 0) { 435#ifndef DISABLE_PSE 436 if (pdir4mb) 437 kernel_pmap->pm_pdir[KPTDI] = PTD[KPTDI] = pdir4mb; 438#endif 439 if (pgeflag) { 440 /* XXX see earlier comments about virtual_avail */ 441 for (va = KERNBASE; va < virtual_avail; va += PAGE_SIZE) 442 { 443 pte = vtopte(va); 444 if (*pte) 445 *pte |= pgeflag; 446 } 447 } 448 /* 449 * for SMP, this will cause all cpus to reload again, which 450 * is actually what we want since they now have CR4_PGE on. 451 */ 452 invltlb(); 453 } else 454 cpu_invltlb(); 455} 456 457/* 458 * Initialize the pmap module. 459 * Called by vm_init, to initialize any structures that the pmap 460 * system needs to map virtual memory. 461 * pmap_init has been enhanced to support in a fairly consistant 462 * way, discontiguous physical memory. 463 */ 464void 465pmap_init(phys_start, phys_end) 466 vm_offset_t phys_start, phys_end; 467{ 468 int i; 469 int initial_pvs; 470 471 /* 472 * object for kernel page table pages 473 */ 474 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE); 475 476 /* 477 * Allocate memory for random pmap data structures. Includes the 478 * pv_head_table. 479 */ 480 481 for(i = 0; i < vm_page_array_size; i++) { 482 vm_page_t m; 483 484 m = &vm_page_array[i]; 485 TAILQ_INIT(&m->md.pv_list); 486 m->md.pv_list_count = 0; 487 } 488 489 /* 490 * init the pv free list 491 */ 492 initial_pvs = vm_page_array_size; 493 if (initial_pvs < MINPV) 494 initial_pvs = MINPV; 495 pvzone = &pvzone_store; 496 pvinit = (struct pv_entry *) kmem_alloc(kernel_map, 497 initial_pvs * sizeof (struct pv_entry)); 498 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit, 499 vm_page_array_size); 500 501 /* 502 * Now it is safe to enable pv_table recording. 503 */ 504 pmap_initialized = TRUE; 505} 506 507/* 508 * Initialize the address space (zone) for the pv_entries. Set a 509 * high water mark so that the system can recover from excessive 510 * numbers of pv entries. 511 */ 512void 513pmap_init2() 514{ 515 int shpgperproc = PMAP_SHPGPERPROC; 516 517 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 518 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 519 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 520 pv_entry_high_water = 9 * (pv_entry_max / 10); 521 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1); 522} 523 524 525/*************************************************** 526 * Low level helper routines..... 527 ***************************************************/ 528 529#if defined(PMAP_DIAGNOSTIC) 530 531/* 532 * This code checks for non-writeable/modified pages. 533 * This should be an invalid condition. 534 */ 535static int 536pmap_nw_modified(pt_entry_t ptea) 537{ 538 int pte; 539 540 pte = (int) ptea; 541 542 if ((pte & (PG_M|PG_RW)) == PG_M) 543 return 1; 544 else 545 return 0; 546} 547#endif 548 549 550/* 551 * this routine defines the region(s) of memory that should 552 * not be tested for the modified bit. 553 */ 554static PMAP_INLINE int 555pmap_track_modified(vm_offset_t va) 556{ 557 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 558 return 1; 559 else 560 return 0; 561} 562 563static __inline void 564pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 565{ 566#if defined(SMP) 567 u_int cpumask; 568 u_int other_cpus; 569 struct thread *td; 570 571 td = curthread; 572 critical_enter(); 573 /* 574 * We need to disable interrupt preemption but MUST NOT have 575 * interrupts disabled here. 576 * XXX we may need to hold schedlock to get a coherent pm_active 577 */ 578 if (td->td_critnest == 1) 579 cpu_critical_exit(td->td_savecrit); 580 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) { 581 invlpg(va); /* global */ 582 } else { 583 cpumask = PCPU_GET(cpumask); 584 other_cpus = PCPU_GET(other_cpus); 585 if (pmap->pm_active & cpumask) 586 cpu_invlpg(va); 587 if (pmap->pm_active & other_cpus) 588 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 589 } 590 critical_exit(); 591#else 592 if (pmap->pm_active) 593 cpu_invlpg(va); 594#endif 595} 596 597static __inline void 598pmap_invalidate_all(pmap_t pmap) 599{ 600#if defined(SMP) 601 u_int cpumask; 602 u_int other_cpus; 603 struct thread *td; 604 605 td = curthread; 606 critical_enter(); 607 /* 608 * We need to disable interrupt preemption but MUST NOT have 609 * interrupts disabled here. 610 * XXX we may need to hold schedlock to get a coherent pm_active 611 */ 612 if (td->td_critnest == 1) 613 cpu_critical_exit(td->td_savecrit); 614 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) { 615 invltlb(); /* global */ 616 } else { 617 cpumask = PCPU_GET(cpumask); 618 other_cpus = PCPU_GET(other_cpus); 619 if (pmap->pm_active & cpumask) 620 cpu_invltlb(); 621 if (pmap->pm_active & other_cpus) 622 smp_masked_invltlb(pmap->pm_active & other_cpus); 623 } 624 critical_exit(); 625#else 626 if (pmap->pm_active) 627 invltlb(); 628#endif 629} 630 631/* 632 * Return an address which is the base of the Virtual mapping of 633 * all the PTEs for the given pmap. Note this doesn't say that 634 * all the PTEs will be present or that the pages there are valid. 635 * The PTEs are made available by the recursive mapping trick. 636 * It will map in the alternate PTE space if needed. 637 */ 638static pt_entry_t * 639get_ptbase(pmap) 640 pmap_t pmap; 641{ 642 pd_entry_t frame = pmap->pm_pdir[PTDPTDI] & PG_FRAME; 643 644 /* are we current address space or kernel? */ 645 if (pmap == kernel_pmap || frame == (PTDpde & PG_FRAME)) 646 return PTmap; 647 /* otherwise, we are alternate address space */ 648 if (frame != (APTDpde & PG_FRAME)) { 649 APTDpde = (pd_entry_t) (frame | PG_RW | PG_V); 650 invltlb(); 651 } 652 return APTmap; 653} 654 655/* 656 * Super fast pmap_pte routine best used when scanning 657 * the pv lists. This eliminates many coarse-grained 658 * invltlb calls. Note that many of the pv list 659 * scans are across different pmaps. It is very wasteful 660 * to do an entire invltlb for checking a single mapping. 661 */ 662 663static pt_entry_t * 664pmap_pte_quick(pmap, va) 665 register pmap_t pmap; 666 vm_offset_t va; 667{ 668 pd_entry_t pde, newpf; 669 pde = pmap->pm_pdir[va >> PDRSHIFT]; 670 if (pde != 0) { 671 pd_entry_t frame = pmap->pm_pdir[PTDPTDI] & PG_FRAME; 672 unsigned index = i386_btop(va); 673 /* are we current address space or kernel? */ 674 if (pmap == kernel_pmap || frame == (PTDpde & PG_FRAME)) 675 return PTmap + index; 676 newpf = pde & PG_FRAME; 677 if (((*PMAP1) & PG_FRAME) != newpf) { 678 *PMAP1 = newpf | PG_RW | PG_V; 679 pmap_invalidate_page(pmap, (vm_offset_t) PADDR1); 680 } 681 return PADDR1 + (index & (NPTEPG - 1)); 682 } 683 return (0); 684} 685 686/* 687 * Routine: pmap_extract 688 * Function: 689 * Extract the physical page address associated 690 * with the given map/virtual_address pair. 691 */ 692vm_offset_t 693pmap_extract(pmap, va) 694 register pmap_t pmap; 695 vm_offset_t va; 696{ 697 vm_offset_t rtval; /* XXX FIXME */ 698 vm_offset_t pdirindex; 699 700 if (pmap == 0) 701 return 0; 702 pdirindex = va >> PDRSHIFT; 703 rtval = pmap->pm_pdir[pdirindex]; 704 if (rtval != 0) { 705 pt_entry_t *pte; 706 if ((rtval & PG_PS) != 0) { 707 rtval &= ~(NBPDR - 1); 708 rtval |= va & (NBPDR - 1); 709 return rtval; 710 } 711 pte = get_ptbase(pmap) + i386_btop(va); 712 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK)); 713 return rtval; 714 } 715 return 0; 716 717} 718 719/*************************************************** 720 * Low level mapping routines..... 721 ***************************************************/ 722 723/* 724 * add a wired page to the kva 725 */ 726PMAP_INLINE void 727pmap_kenter(vm_offset_t va, vm_offset_t pa) 728{ 729 pt_entry_t *pte; 730 731 pte = vtopte(va); 732 *pte = pa | PG_RW | PG_V | pgeflag; 733 invlpg(va); 734} 735 736/* 737 * remove a page from the kernel pagetables 738 */ 739PMAP_INLINE void 740pmap_kremove(vm_offset_t va) 741{ 742 pt_entry_t *pte; 743 744 pte = vtopte(va); 745 *pte = 0; 746 invlpg(va); 747} 748 749/* 750 * Used to map a range of physical addresses into kernel 751 * virtual address space. 752 * 753 * The value passed in '*virt' is a suggested virtual address for 754 * the mapping. Architectures which can support a direct-mapped 755 * physical to virtual region can return the appropriate address 756 * within that region, leaving '*virt' unchanged. Other 757 * architectures should map the pages starting at '*virt' and 758 * update '*virt' with the first usable address after the mapped 759 * region. 760 */ 761vm_offset_t 762pmap_map(vm_offset_t *virt, vm_offset_t start, vm_offset_t end, int prot) 763{ 764 vm_offset_t va, sva; 765 766 va = sva = *virt; 767 while (start < end) { 768 pmap_kenter(va, start); 769 va += PAGE_SIZE; 770 start += PAGE_SIZE; 771 } 772 invlpg_range(sva, end); 773 *virt = va; 774 return (sva); 775} 776 777 778/* 779 * Add a list of wired pages to the kva 780 * this routine is only used for temporary 781 * kernel mappings that do not need to have 782 * page modification or references recorded. 783 * Note that old mappings are simply written 784 * over. The page *must* be wired. 785 */ 786void 787pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 788{ 789 vm_offset_t va, end_va; 790 791 va = sva; 792 end_va = va + count * PAGE_SIZE; 793 794 while (va < end_va) { 795 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 796 va += PAGE_SIZE; 797 m++; 798 } 799 invlpg_range(sva, end_va); 800} 801 802/* 803 * this routine jerks page mappings from the 804 * kernel -- it is meant only for temporary mappings. 805 */ 806void 807pmap_qremove(vm_offset_t sva, int count) 808{ 809 vm_offset_t va, end_va; 810 811 va = sva; 812 end_va = va + count * PAGE_SIZE; 813 814 while (va < end_va) { 815 pmap_kremove(va); 816 va += PAGE_SIZE; 817 } 818 invlpg_range(sva, end_va); 819} 820 821static vm_page_t 822pmap_page_lookup(vm_object_t object, vm_pindex_t pindex) 823{ 824 vm_page_t m; 825retry: 826 m = vm_page_lookup(object, pindex); 827 if (m && vm_page_sleep_busy(m, FALSE, "pplookp")) 828 goto retry; 829 return m; 830} 831 832/* 833 * Create the Uarea stack for a new process. 834 * This routine directly affects the fork perf for a process. 835 */ 836void 837pmap_new_proc(struct proc *p) 838{ 839 int i; 840 vm_page_t ma[UAREA_PAGES]; 841 vm_object_t upobj; 842 vm_offset_t up; 843 vm_page_t m; 844 845 /* 846 * allocate object for the upages 847 */ 848 upobj = p->p_upages_obj; 849 if (upobj == NULL) { 850 upobj = vm_object_allocate(OBJT_DEFAULT, UAREA_PAGES); 851 p->p_upages_obj = upobj; 852 } 853 854 /* get a kernel virtual address for the U area for this thread */ 855 up = (vm_offset_t)p->p_uarea; 856 if (up == 0) { 857 up = kmem_alloc_nofault(kernel_map, UAREA_PAGES * PAGE_SIZE); 858 if (up == 0) 859 panic("pmap_new_proc: upage allocation failed"); 860 p->p_uarea = (struct user *)up; 861 } 862 863 for (i = 0; i < UAREA_PAGES; i++) { 864 /* 865 * Get a kernel stack page 866 */ 867 m = vm_page_grab(upobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 868 ma[i] = m; 869 870 /* 871 * Wire the page 872 */ 873 m->wire_count++; 874 cnt.v_wire_count++; 875 876 vm_page_wakeup(m); 877 vm_page_flag_clear(m, PG_ZERO); 878 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE); 879 m->valid = VM_PAGE_BITS_ALL; 880 } 881 pmap_qenter(up, ma, UAREA_PAGES); 882} 883 884/* 885 * Dispose the U-Area for a process that has exited. 886 * This routine directly impacts the exit perf of a process. 887 */ 888void 889pmap_dispose_proc(p) 890 struct proc *p; 891{ 892 int i; 893 vm_object_t upobj; 894 vm_offset_t up; 895 vm_page_t m; 896 897 upobj = p->p_upages_obj; 898 up = (vm_offset_t)p->p_uarea; 899 pmap_qremove(up, UAREA_PAGES); 900 for (i = 0; i < UAREA_PAGES; i++) { 901 m = vm_page_lookup(upobj, i); 902 if (m == NULL) 903 panic("pmap_dispose_proc: upage already missing?"); 904 vm_page_busy(m); 905 vm_page_unwire(m, 0); 906 vm_page_free(m); 907 } 908} 909 910/* 911 * Allow the U_AREA for a process to be prejudicially paged out. 912 */ 913void 914pmap_swapout_proc(p) 915 struct proc *p; 916{ 917 int i; 918 vm_object_t upobj; 919 vm_offset_t up; 920 vm_page_t m; 921 922 upobj = p->p_upages_obj; 923 up = (vm_offset_t)p->p_uarea; 924 pmap_qremove(up, UAREA_PAGES); 925 for (i = 0; i < UAREA_PAGES; i++) { 926 m = vm_page_lookup(upobj, i); 927 if (m == NULL) 928 panic("pmap_swapout_proc: upage already missing?"); 929 vm_page_dirty(m); 930 vm_page_unwire(m, 0); 931 } 932} 933 934/* 935 * Bring the U-Area for a specified process back in. 936 */ 937void 938pmap_swapin_proc(p) 939 struct proc *p; 940{ 941 int i, rv; 942 vm_page_t ma[UAREA_PAGES]; 943 vm_object_t upobj; 944 vm_offset_t up; 945 vm_page_t m; 946 947 upobj = p->p_upages_obj; 948 up = (vm_offset_t)p->p_uarea; 949 for (i = 0; i < UAREA_PAGES; i++) { 950 m = vm_page_grab(upobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 951 if (m->valid != VM_PAGE_BITS_ALL) { 952 rv = vm_pager_get_pages(upobj, &m, 1, 0); 953 if (rv != VM_PAGER_OK) 954 panic("pmap_swapin_proc: cannot get upage for proc: %d\n", p->p_pid); 955 m = vm_page_lookup(upobj, i); 956 m->valid = VM_PAGE_BITS_ALL; 957 } 958 ma[i] = m; 959 vm_page_wire(m); 960 vm_page_wakeup(m); 961 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE); 962 } 963 pmap_qenter(up, ma, UAREA_PAGES); 964} 965 966/* 967 * Create the kernel stack (including pcb for i386) for a new thread. 968 * This routine directly affects the fork perf for a process and 969 * create performance for a thread. 970 */ 971void 972pmap_new_thread(struct thread *td) 973{ 974 int i; 975 vm_page_t ma[KSTACK_PAGES]; 976 vm_object_t ksobj; 977 vm_page_t m; 978 vm_offset_t ks; 979 980 /* 981 * allocate object for the kstack 982 */ 983 ksobj = td->td_kstack_obj; 984 if (ksobj == NULL) { 985 ksobj = vm_object_allocate(OBJT_DEFAULT, KSTACK_PAGES); 986 td->td_kstack_obj = ksobj; 987 } 988 989 /* get a kernel virtual address for the kstack for this thread */ 990 ks = td->td_kstack; 991#ifdef KSTACK_GUARD 992 if (ks == 0) { 993 ks = kmem_alloc_nofault(kernel_map, 994 (KSTACK_PAGES + 1) * PAGE_SIZE); 995 if (ks == 0) 996 panic("pmap_new_thread: kstack allocation failed"); 997 if (*vtopte(ks) != 0) 998 pmap_qremove(ks, 1); 999 ks += PAGE_SIZE; 1000 td->td_kstack = ks; 1001 } 1002#else 1003 if (ks == 0) { 1004 ks = kmem_alloc_nofault(kernel_map, KSTACK_PAGES * PAGE_SIZE); 1005 if (ks == 0) 1006 panic("pmap_new_thread: kstack allocation failed"); 1007 td->td_kstack = ks; 1008 } 1009#endif 1010 for (i = 0; i < KSTACK_PAGES; i++) { 1011 /* 1012 * Get a kernel stack page 1013 */ 1014 m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1015 ma[i] = m; 1016 1017 /* 1018 * Wire the page 1019 */ 1020 m->wire_count++; 1021 cnt.v_wire_count++; 1022 1023 vm_page_wakeup(m); 1024 vm_page_flag_clear(m, PG_ZERO); 1025 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE); 1026 m->valid = VM_PAGE_BITS_ALL; 1027 } 1028 pmap_qenter(ks, ma, KSTACK_PAGES); 1029} 1030 1031/* 1032 * Dispose the kernel stack for a thread that has exited. 1033 * This routine directly impacts the exit perf of a process and thread. 1034 */ 1035void 1036pmap_dispose_thread(td) 1037 struct thread *td; 1038{ 1039 int i; 1040 vm_object_t ksobj; 1041 vm_offset_t ks; 1042 vm_page_t m; 1043 1044 ksobj = td->td_kstack_obj; 1045 ks = td->td_kstack; 1046 pmap_qremove(ks, KSTACK_PAGES); 1047 for (i = 0; i < KSTACK_PAGES; i++) { 1048 m = vm_page_lookup(ksobj, i); 1049 if (m == NULL) 1050 panic("pmap_dispose_thread: kstack already missing?"); 1051 vm_page_busy(m); 1052 vm_page_unwire(m, 0); 1053 vm_page_free(m); 1054 } 1055} 1056 1057/* 1058 * Allow the Kernel stack for a thread to be prejudicially paged out. 1059 */ 1060void 1061pmap_swapout_thread(td) 1062 struct thread *td; 1063{ 1064 int i; 1065 vm_object_t ksobj; 1066 vm_offset_t ks; 1067 vm_page_t m; 1068 1069 ksobj = td->td_kstack_obj; 1070 ks = td->td_kstack; 1071 pmap_qremove(ks, KSTACK_PAGES); 1072 for (i = 0; i < KSTACK_PAGES; i++) { 1073 m = vm_page_lookup(ksobj, i); 1074 if (m == NULL) 1075 panic("pmap_swapout_thread: kstack already missing?"); 1076 vm_page_dirty(m); 1077 vm_page_unwire(m, 0); 1078 } 1079} 1080 1081/* 1082 * Bring the kernel stack for a specified thread back in. 1083 */ 1084void 1085pmap_swapin_thread(td) 1086 struct thread *td; 1087{ 1088 int i, rv; 1089 vm_page_t ma[KSTACK_PAGES]; 1090 vm_object_t ksobj; 1091 vm_offset_t ks; 1092 vm_page_t m; 1093 1094 ksobj = td->td_kstack_obj; 1095 ks = td->td_kstack; 1096 for (i = 0; i < KSTACK_PAGES; i++) { 1097 m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1098 if (m->valid != VM_PAGE_BITS_ALL) { 1099 rv = vm_pager_get_pages(ksobj, &m, 1, 0); 1100 if (rv != VM_PAGER_OK) 1101 panic("pmap_swapin_thread: cannot get kstack for proc: %d\n", td->td_proc->p_pid); 1102 m = vm_page_lookup(ksobj, i); 1103 m->valid = VM_PAGE_BITS_ALL; 1104 } 1105 ma[i] = m; 1106 vm_page_wire(m); 1107 vm_page_wakeup(m); 1108 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE); 1109 } 1110 pmap_qenter(ks, ma, KSTACK_PAGES); 1111} 1112 1113/*************************************************** 1114 * Page table page management routines..... 1115 ***************************************************/ 1116 1117/* 1118 * This routine unholds page table pages, and if the hold count 1119 * drops to zero, then it decrements the wire count. 1120 */ 1121static int 1122_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1123{ 1124 1125 while (vm_page_sleep_busy(m, FALSE, "pmuwpt")) 1126 ; 1127 1128 if (m->hold_count == 0) { 1129 vm_offset_t pteva; 1130 /* 1131 * unmap the page table page 1132 */ 1133 pmap->pm_pdir[m->pindex] = 0; 1134 --pmap->pm_stats.resident_count; 1135 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == 1136 (PTDpde & PG_FRAME)) { 1137 /* 1138 * Do a invltlb to make the invalidated mapping 1139 * take effect immediately. 1140 */ 1141 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1142 pmap_invalidate_page(pmap, pteva); 1143 } 1144 1145 if (pmap->pm_ptphint == m) 1146 pmap->pm_ptphint = NULL; 1147 1148 /* 1149 * If the page is finally unwired, simply free it. 1150 */ 1151 --m->wire_count; 1152 if (m->wire_count == 0) { 1153 1154 vm_page_flash(m); 1155 vm_page_busy(m); 1156 vm_page_free_zero(m); 1157 --cnt.v_wire_count; 1158 } 1159 return 1; 1160 } 1161 return 0; 1162} 1163 1164static PMAP_INLINE int 1165pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1166{ 1167 vm_page_unhold(m); 1168 if (m->hold_count == 0) 1169 return _pmap_unwire_pte_hold(pmap, m); 1170 else 1171 return 0; 1172} 1173 1174/* 1175 * After removing a page table entry, this routine is used to 1176 * conditionally free the page, and manage the hold/wire counts. 1177 */ 1178static int 1179pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 1180{ 1181 unsigned ptepindex; 1182 if (va >= VM_MAXUSER_ADDRESS) 1183 return 0; 1184 1185 if (mpte == NULL) { 1186 ptepindex = (va >> PDRSHIFT); 1187 if (pmap->pm_ptphint && 1188 (pmap->pm_ptphint->pindex == ptepindex)) { 1189 mpte = pmap->pm_ptphint; 1190 } else { 1191 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 1192 pmap->pm_ptphint = mpte; 1193 } 1194 } 1195 1196 return pmap_unwire_pte_hold(pmap, mpte); 1197} 1198 1199void 1200pmap_pinit0(pmap) 1201 struct pmap *pmap; 1202{ 1203 pmap->pm_pdir = 1204 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE); 1205 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t)IdlePTD); 1206 invlpg((vm_offset_t)pmap->pm_pdir); 1207 pmap->pm_count = 1; 1208 pmap->pm_ptphint = NULL; 1209 pmap->pm_active = 0; 1210 TAILQ_INIT(&pmap->pm_pvlist); 1211 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1212 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1213} 1214 1215/* 1216 * Initialize a preallocated and zeroed pmap structure, 1217 * such as one in a vmspace structure. 1218 */ 1219void 1220pmap_pinit(pmap) 1221 register struct pmap *pmap; 1222{ 1223 vm_page_t ptdpg; 1224 1225 /* 1226 * No need to allocate page table space yet but we do need a valid 1227 * page directory table. 1228 */ 1229 if (pmap->pm_pdir == NULL) 1230 pmap->pm_pdir = 1231 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE); 1232 1233 /* 1234 * allocate object for the ptes 1235 */ 1236 if (pmap->pm_pteobj == NULL) 1237 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1); 1238 1239 /* 1240 * allocate the page directory page 1241 */ 1242 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI, 1243 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1244 1245 ptdpg->wire_count = 1; 1246 ++cnt.v_wire_count; 1247 1248 1249 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY); /* not usually mapped*/ 1250 ptdpg->valid = VM_PAGE_BITS_ALL; 1251 1252 pmap_qenter((vm_offset_t) pmap->pm_pdir, &ptdpg, 1); 1253 if ((ptdpg->flags & PG_ZERO) == 0) 1254 bzero(pmap->pm_pdir, PAGE_SIZE); 1255 1256 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1257 /* Wire in kernel global address entries. */ 1258 /* XXX copies current process, does not fill in MPPTDI */ 1259 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE); 1260#ifdef SMP 1261 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1262#endif 1263 1264 /* install self-referential address mapping entry */ 1265 pmap->pm_pdir[PTDPTDI] = 1266 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M; 1267 1268 pmap->pm_count = 1; 1269 pmap->pm_active = 0; 1270 pmap->pm_ptphint = NULL; 1271 TAILQ_INIT(&pmap->pm_pvlist); 1272 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1273} 1274 1275/* 1276 * Wire in kernel global address entries. To avoid a race condition 1277 * between pmap initialization and pmap_growkernel, this procedure 1278 * should be called after the vmspace is attached to the process 1279 * but before this pmap is activated. 1280 */ 1281void 1282pmap_pinit2(pmap) 1283 struct pmap *pmap; 1284{ 1285 /* XXX: Remove this stub when no longer called */ 1286} 1287 1288static int 1289pmap_release_free_page(pmap_t pmap, vm_page_t p) 1290{ 1291 pd_entry_t *pde = pmap->pm_pdir; 1292 /* 1293 * This code optimizes the case of freeing non-busy 1294 * page-table pages. Those pages are zero now, and 1295 * might as well be placed directly into the zero queue. 1296 */ 1297 if (vm_page_sleep_busy(p, FALSE, "pmaprl")) 1298 return 0; 1299 1300 vm_page_busy(p); 1301 1302 /* 1303 * Remove the page table page from the processes address space. 1304 */ 1305 pde[p->pindex] = 0; 1306 pmap->pm_stats.resident_count--; 1307 1308 if (p->hold_count) { 1309 panic("pmap_release: freeing held page table page"); 1310 } 1311 /* 1312 * Page directory pages need to have the kernel 1313 * stuff cleared, so they can go into the zero queue also. 1314 */ 1315 if (p->pindex == PTDPTDI) { 1316 bzero(pde + KPTDI, nkpt * PTESIZE); 1317#ifdef SMP 1318 pde[MPPTDI] = 0; 1319#endif 1320 pde[APTDPTDI] = 0; 1321 pmap_kremove((vm_offset_t) pmap->pm_pdir); 1322 } 1323 1324 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex)) 1325 pmap->pm_ptphint = NULL; 1326 1327 p->wire_count--; 1328 cnt.v_wire_count--; 1329 vm_page_free_zero(p); 1330 return 1; 1331} 1332 1333/* 1334 * this routine is called if the page table page is not 1335 * mapped correctly. 1336 */ 1337static vm_page_t 1338_pmap_allocpte(pmap, ptepindex) 1339 pmap_t pmap; 1340 unsigned ptepindex; 1341{ 1342 vm_offset_t pteva, ptepa; /* XXXPA */ 1343 vm_page_t m; 1344 1345 /* 1346 * Find or fabricate a new pagetable page 1347 */ 1348 m = vm_page_grab(pmap->pm_pteobj, ptepindex, 1349 VM_ALLOC_ZERO | VM_ALLOC_RETRY); 1350 1351 KASSERT(m->queue == PQ_NONE, 1352 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1353 1354 if (m->wire_count == 0) 1355 cnt.v_wire_count++; 1356 m->wire_count++; 1357 1358 /* 1359 * Increment the hold count for the page table page 1360 * (denoting a new mapping.) 1361 */ 1362 m->hold_count++; 1363 1364 /* 1365 * Map the pagetable page into the process address space, if 1366 * it isn't already there. 1367 */ 1368 1369 pmap->pm_stats.resident_count++; 1370 1371 ptepa = VM_PAGE_TO_PHYS(m); 1372 pmap->pm_pdir[ptepindex] = 1373 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1374 1375 /* 1376 * Set the page table hint 1377 */ 1378 pmap->pm_ptphint = m; 1379 1380 /* 1381 * Try to use the new mapping, but if we cannot, then 1382 * do it with the routine that maps the page explicitly. 1383 */ 1384 if ((m->flags & PG_ZERO) == 0) { 1385 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == 1386 (PTDpde & PG_FRAME)) { 1387 pteva = VM_MAXUSER_ADDRESS + i386_ptob(ptepindex); 1388 bzero((caddr_t) pteva, PAGE_SIZE); 1389 } else { 1390 pmap_zero_page(ptepa); 1391 } 1392 } 1393 1394 m->valid = VM_PAGE_BITS_ALL; 1395 vm_page_flag_clear(m, PG_ZERO); 1396 vm_page_flag_set(m, PG_MAPPED); 1397 vm_page_wakeup(m); 1398 1399 return m; 1400} 1401 1402static vm_page_t 1403pmap_allocpte(pmap_t pmap, vm_offset_t va) 1404{ 1405 unsigned ptepindex; 1406 pd_entry_t ptepa; 1407 vm_page_t m; 1408 1409 /* 1410 * Calculate pagetable page index 1411 */ 1412 ptepindex = va >> PDRSHIFT; 1413 1414 /* 1415 * Get the page directory entry 1416 */ 1417 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex]; 1418 1419 /* 1420 * This supports switching from a 4MB page to a 1421 * normal 4K page. 1422 */ 1423 if (ptepa & PG_PS) { 1424 pmap->pm_pdir[ptepindex] = 0; 1425 ptepa = 0; 1426 invltlb(); 1427 } 1428 1429 /* 1430 * If the page table page is mapped, we just increment the 1431 * hold count, and activate it. 1432 */ 1433 if (ptepa) { 1434 /* 1435 * In order to get the page table page, try the 1436 * hint first. 1437 */ 1438 if (pmap->pm_ptphint && 1439 (pmap->pm_ptphint->pindex == ptepindex)) { 1440 m = pmap->pm_ptphint; 1441 } else { 1442 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 1443 pmap->pm_ptphint = m; 1444 } 1445 m->hold_count++; 1446 return m; 1447 } 1448 /* 1449 * Here if the pte page isn't mapped, or if it has been deallocated. 1450 */ 1451 return _pmap_allocpte(pmap, ptepindex); 1452} 1453 1454 1455/*************************************************** 1456* Pmap allocation/deallocation routines. 1457 ***************************************************/ 1458 1459/* 1460 * Release any resources held by the given physical map. 1461 * Called when a pmap initialized by pmap_pinit is being released. 1462 * Should only be called if the map contains no valid mappings. 1463 */ 1464void 1465pmap_release(pmap_t pmap) 1466{ 1467 vm_page_t p,n,ptdpg; 1468 vm_object_t object = pmap->pm_pteobj; 1469 int curgeneration; 1470 1471#if defined(DIAGNOSTIC) 1472 if (object->ref_count != 1) 1473 panic("pmap_release: pteobj reference count != 1"); 1474#endif 1475 1476 ptdpg = NULL; 1477 LIST_REMOVE(pmap, pm_list); 1478retry: 1479 curgeneration = object->generation; 1480 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = n) { 1481 n = TAILQ_NEXT(p, listq); 1482 if (p->pindex == PTDPTDI) { 1483 ptdpg = p; 1484 continue; 1485 } 1486 while (1) { 1487 if (!pmap_release_free_page(pmap, p) && 1488 (object->generation != curgeneration)) 1489 goto retry; 1490 } 1491 } 1492 1493 if (ptdpg && !pmap_release_free_page(pmap, ptdpg)) 1494 goto retry; 1495} 1496 1497static int 1498kvm_size(SYSCTL_HANDLER_ARGS) 1499{ 1500 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1501 1502 return sysctl_handle_long(oidp, &ksize, 0, req); 1503} 1504SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1505 0, 0, kvm_size, "IU", "Size of KVM"); 1506 1507static int 1508kvm_free(SYSCTL_HANDLER_ARGS) 1509{ 1510 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1511 1512 return sysctl_handle_long(oidp, &kfree, 0, req); 1513} 1514SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1515 0, 0, kvm_free, "IU", "Amount of KVM free"); 1516 1517/* 1518 * grow the number of kernel page table entries, if needed 1519 */ 1520void 1521pmap_growkernel(vm_offset_t addr) 1522{ 1523 struct pmap *pmap; 1524 int s; 1525 vm_offset_t ptppaddr; 1526 vm_page_t nkpg; 1527 pd_entry_t newpdir; 1528 1529 s = splhigh(); 1530 if (kernel_vm_end == 0) { 1531 kernel_vm_end = KERNBASE; 1532 nkpt = 0; 1533 while (pdir_pde(PTD, kernel_vm_end)) { 1534 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1535 nkpt++; 1536 } 1537 } 1538 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1539 while (kernel_vm_end < addr) { 1540 if (pdir_pde(PTD, kernel_vm_end)) { 1541 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1542 continue; 1543 } 1544 1545 /* 1546 * This index is bogus, but out of the way 1547 */ 1548 nkpg = vm_page_alloc(kptobj, nkpt, VM_ALLOC_SYSTEM); 1549 if (!nkpg) 1550 panic("pmap_growkernel: no memory to grow kernel"); 1551 1552 nkpt++; 1553 1554 vm_page_wire(nkpg); 1555 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1556 pmap_zero_page(ptppaddr); 1557 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1558 pdir_pde(PTD, kernel_vm_end) = newpdir; 1559 1560 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1561 *pmap_pde(pmap, kernel_vm_end) = newpdir; 1562 } 1563 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1564 } 1565 splx(s); 1566} 1567 1568/* 1569 * Retire the given physical map from service. 1570 * Should only be called if the map contains 1571 * no valid mappings. 1572 */ 1573void 1574pmap_destroy(pmap_t pmap) 1575{ 1576 int count; 1577 1578 if (pmap == NULL) 1579 return; 1580 1581 count = --pmap->pm_count; 1582 if (count == 0) { 1583 pmap_release(pmap); 1584 panic("destroying a pmap is not yet implemented"); 1585 } 1586} 1587 1588/* 1589 * Add a reference to the specified pmap. 1590 */ 1591void 1592pmap_reference(pmap_t pmap) 1593{ 1594 if (pmap != NULL) { 1595 pmap->pm_count++; 1596 } 1597} 1598 1599/*************************************************** 1600* page management routines. 1601 ***************************************************/ 1602 1603/* 1604 * free the pv_entry back to the free list 1605 */ 1606static PMAP_INLINE void 1607free_pv_entry(pv_entry_t pv) 1608{ 1609 pv_entry_count--; 1610 zfree(pvzone, pv); 1611} 1612 1613/* 1614 * get a new pv_entry, allocating a block from the system 1615 * when needed. 1616 * the memory allocation is performed bypassing the malloc code 1617 * because of the possibility of allocations at interrupt time. 1618 */ 1619static pv_entry_t 1620get_pv_entry(void) 1621{ 1622 pv_entry_count++; 1623 if (pv_entry_high_water && 1624 (pv_entry_count > pv_entry_high_water) && 1625 (pmap_pagedaemon_waken == 0)) { 1626 pmap_pagedaemon_waken = 1; 1627 wakeup (&vm_pages_needed); 1628 } 1629 return zalloc(pvzone); 1630} 1631 1632/* 1633 * This routine is very drastic, but can save the system 1634 * in a pinch. 1635 */ 1636void 1637pmap_collect() 1638{ 1639 int i; 1640 vm_page_t m; 1641 static int warningdone = 0; 1642 1643 if (pmap_pagedaemon_waken == 0) 1644 return; 1645 1646 if (warningdone < 5) { 1647 printf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n"); 1648 warningdone++; 1649 } 1650 1651 for(i = 0; i < vm_page_array_size; i++) { 1652 m = &vm_page_array[i]; 1653 if (m->wire_count || m->hold_count || m->busy || 1654 (m->flags & (PG_BUSY | PG_UNMANAGED))) 1655 continue; 1656 pmap_remove_all(m); 1657 } 1658 pmap_pagedaemon_waken = 0; 1659} 1660 1661 1662/* 1663 * If it is the first entry on the list, it is actually 1664 * in the header and we must copy the following entry up 1665 * to the header. Otherwise we must search the list for 1666 * the entry. In either case we free the now unused entry. 1667 */ 1668 1669static int 1670pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1671{ 1672 pv_entry_t pv; 1673 int rtval; 1674 int s; 1675 1676 s = splvm(); 1677 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1678 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1679 if (pmap == pv->pv_pmap && va == pv->pv_va) 1680 break; 1681 } 1682 } else { 1683 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1684 if (va == pv->pv_va) 1685 break; 1686 } 1687 } 1688 1689 rtval = 0; 1690 if (pv) { 1691 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1692 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1693 m->md.pv_list_count--; 1694 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1695 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 1696 1697 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1698 free_pv_entry(pv); 1699 } 1700 1701 splx(s); 1702 return rtval; 1703} 1704 1705/* 1706 * Create a pv entry for page at pa for 1707 * (pmap, va). 1708 */ 1709static void 1710pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1711{ 1712 1713 int s; 1714 pv_entry_t pv; 1715 1716 s = splvm(); 1717 pv = get_pv_entry(); 1718 pv->pv_va = va; 1719 pv->pv_pmap = pmap; 1720 pv->pv_ptem = mpte; 1721 1722 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1723 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1724 m->md.pv_list_count++; 1725 1726 splx(s); 1727} 1728 1729/* 1730 * pmap_remove_pte: do the things to unmap a page in a process 1731 */ 1732static int 1733pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1734{ 1735 pt_entry_t oldpte; 1736 vm_page_t m; 1737 1738 oldpte = atomic_readandclear_int(ptq); 1739 if (oldpte & PG_W) 1740 pmap->pm_stats.wired_count -= 1; 1741 /* 1742 * Machines that don't support invlpg, also don't support 1743 * PG_G. 1744 */ 1745 if (oldpte & PG_G) 1746 invlpg(va); 1747 pmap->pm_stats.resident_count -= 1; 1748 if (oldpte & PG_MANAGED) { 1749 m = PHYS_TO_VM_PAGE(oldpte); 1750 if (oldpte & PG_M) { 1751#if defined(PMAP_DIAGNOSTIC) 1752 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1753 printf( 1754 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1755 va, oldpte); 1756 } 1757#endif 1758 if (pmap_track_modified(va)) 1759 vm_page_dirty(m); 1760 } 1761 if (oldpte & PG_A) 1762 vm_page_flag_set(m, PG_REFERENCED); 1763 return pmap_remove_entry(pmap, m, va); 1764 } else { 1765 return pmap_unuse_pt(pmap, va, NULL); 1766 } 1767 1768 return 0; 1769} 1770 1771/* 1772 * Remove a single page from a process address space 1773 */ 1774static void 1775pmap_remove_page(pmap_t pmap, vm_offset_t va) 1776{ 1777 register pt_entry_t *ptq; 1778 1779 /* 1780 * if there is no pte for this address, just skip it!!! 1781 */ 1782 if (*pmap_pde(pmap, va) == 0) { 1783 return; 1784 } 1785 1786 /* 1787 * get a local va for mappings for this pmap. 1788 */ 1789 ptq = get_ptbase(pmap) + i386_btop(va); 1790 if (*ptq) { 1791 (void) pmap_remove_pte(pmap, ptq, va); 1792 pmap_invalidate_page(pmap, va); 1793 } 1794 return; 1795} 1796 1797/* 1798 * Remove the given range of addresses from the specified map. 1799 * 1800 * It is assumed that the start and end are properly 1801 * rounded to the page size. 1802 */ 1803void 1804pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1805{ 1806 register pt_entry_t *ptbase; 1807 vm_offset_t pdnxt; 1808 pd_entry_t ptpaddr; 1809 vm_offset_t sindex, eindex; 1810 int anyvalid; 1811 1812 if (pmap == NULL) 1813 return; 1814 1815 if (pmap->pm_stats.resident_count == 0) 1816 return; 1817 1818 /* 1819 * special handling of removing one page. a very 1820 * common operation and easy to short circuit some 1821 * code. 1822 */ 1823 if ((sva + PAGE_SIZE == eva) && 1824 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1825 pmap_remove_page(pmap, sva); 1826 return; 1827 } 1828 1829 anyvalid = 0; 1830 1831 /* 1832 * Get a local virtual address for the mappings that are being 1833 * worked with. 1834 */ 1835 ptbase = get_ptbase(pmap); 1836 1837 sindex = i386_btop(sva); 1838 eindex = i386_btop(eva); 1839 1840 for (; sindex < eindex; sindex = pdnxt) { 1841 unsigned pdirindex; 1842 1843 /* 1844 * Calculate index for next page table. 1845 */ 1846 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1)); 1847 if (pmap->pm_stats.resident_count == 0) 1848 break; 1849 1850 pdirindex = sindex / NPDEPG; 1851 ptpaddr = pmap->pm_pdir[pdirindex]; 1852 if ((ptpaddr & PG_PS) != 0) { 1853 pmap->pm_pdir[pdirindex] = 0; 1854 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1855 anyvalid++; 1856 continue; 1857 } 1858 1859 /* 1860 * Weed out invalid mappings. Note: we assume that the page 1861 * directory table is always allocated, and in kernel virtual. 1862 */ 1863 if (ptpaddr == 0) 1864 continue; 1865 1866 /* 1867 * Limit our scan to either the end of the va represented 1868 * by the current page table page, or to the end of the 1869 * range being removed. 1870 */ 1871 if (pdnxt > eindex) { 1872 pdnxt = eindex; 1873 } 1874 1875 for (; sindex != pdnxt; sindex++) { 1876 vm_offset_t va; 1877 if (ptbase[sindex] == 0) { 1878 continue; 1879 } 1880 va = i386_ptob(sindex); 1881 1882 anyvalid++; 1883 if (pmap_remove_pte(pmap, 1884 ptbase + sindex, va)) 1885 break; 1886 } 1887 } 1888 1889 if (anyvalid) 1890 pmap_invalidate_all(pmap); 1891} 1892 1893/* 1894 * Routine: pmap_remove_all 1895 * Function: 1896 * Removes this physical page from 1897 * all physical maps in which it resides. 1898 * Reflects back modify bits to the pager. 1899 * 1900 * Notes: 1901 * Original versions of this routine were very 1902 * inefficient because they iteratively called 1903 * pmap_remove (slow...) 1904 */ 1905 1906static void 1907pmap_remove_all(vm_page_t m) 1908{ 1909 register pv_entry_t pv; 1910 pt_entry_t *pte, tpte; 1911 int s; 1912 1913#if defined(PMAP_DIAGNOSTIC) 1914 /* 1915 * XXX this makes pmap_page_protect(NONE) illegal for non-managed 1916 * pages! 1917 */ 1918 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1919 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%x", VM_PAGE_TO_PHYS(m)); 1920 } 1921#endif 1922 1923 s = splvm(); 1924 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1925 pv->pv_pmap->pm_stats.resident_count--; 1926 1927 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1928 1929 tpte = atomic_readandclear_int(pte); 1930 if (tpte & PG_W) 1931 pv->pv_pmap->pm_stats.wired_count--; 1932 1933 if (tpte & PG_A) 1934 vm_page_flag_set(m, PG_REFERENCED); 1935 1936 /* 1937 * Update the vm_page_t clean and reference bits. 1938 */ 1939 if (tpte & PG_M) { 1940#if defined(PMAP_DIAGNOSTIC) 1941 if (pmap_nw_modified((pt_entry_t) tpte)) { 1942 printf( 1943 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1944 pv->pv_va, tpte); 1945 } 1946#endif 1947 if (pmap_track_modified(pv->pv_va)) 1948 vm_page_dirty(m); 1949 } 1950 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1951 1952 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1953 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1954 m->md.pv_list_count--; 1955 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 1956 free_pv_entry(pv); 1957 } 1958 1959 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 1960 1961 splx(s); 1962} 1963 1964/* 1965 * Set the physical protection on the 1966 * specified range of this map as requested. 1967 */ 1968void 1969pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1970{ 1971 register pt_entry_t *ptbase; 1972 vm_offset_t pdnxt; 1973 pd_entry_t ptpaddr; 1974 vm_pindex_t sindex, eindex; 1975 int anychanged; 1976 1977 if (pmap == NULL) 1978 return; 1979 1980 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1981 pmap_remove(pmap, sva, eva); 1982 return; 1983 } 1984 1985 if (prot & VM_PROT_WRITE) 1986 return; 1987 1988 anychanged = 0; 1989 1990 ptbase = get_ptbase(pmap); 1991 1992 sindex = i386_btop(sva); 1993 eindex = i386_btop(eva); 1994 1995 for (; sindex < eindex; sindex = pdnxt) { 1996 1997 unsigned pdirindex; 1998 1999 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1)); 2000 2001 pdirindex = sindex / NPDEPG; 2002 ptpaddr = pmap->pm_pdir[pdirindex]; 2003 if ((ptpaddr & PG_PS) != 0) { 2004 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 2005 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 2006 anychanged++; 2007 continue; 2008 } 2009 2010 /* 2011 * Weed out invalid mappings. Note: we assume that the page 2012 * directory table is always allocated, and in kernel virtual. 2013 */ 2014 if (ptpaddr == 0) 2015 continue; 2016 2017 if (pdnxt > eindex) { 2018 pdnxt = eindex; 2019 } 2020 2021 for (; sindex != pdnxt; sindex++) { 2022 2023 pt_entry_t pbits; 2024 vm_page_t m; 2025 2026 pbits = ptbase[sindex]; 2027 2028 if (pbits & PG_MANAGED) { 2029 m = NULL; 2030 if (pbits & PG_A) { 2031 m = PHYS_TO_VM_PAGE(pbits); 2032 vm_page_flag_set(m, PG_REFERENCED); 2033 pbits &= ~PG_A; 2034 } 2035 if (pbits & PG_M) { 2036 if (pmap_track_modified(i386_ptob(sindex))) { 2037 if (m == NULL) 2038 m = PHYS_TO_VM_PAGE(pbits); 2039 vm_page_dirty(m); 2040 pbits &= ~PG_M; 2041 } 2042 } 2043 } 2044 2045 pbits &= ~PG_RW; 2046 2047 if (pbits != ptbase[sindex]) { 2048 ptbase[sindex] = pbits; 2049 anychanged = 1; 2050 } 2051 } 2052 } 2053 if (anychanged) 2054 pmap_invalidate_all(pmap); 2055} 2056 2057/* 2058 * Insert the given physical page (p) at 2059 * the specified virtual address (v) in the 2060 * target physical map with the protection requested. 2061 * 2062 * If specified, the page will be wired down, meaning 2063 * that the related pte can not be reclaimed. 2064 * 2065 * NB: This is the only routine which MAY NOT lazy-evaluate 2066 * or lose information. That is, this routine must actually 2067 * insert this page into the given map NOW. 2068 */ 2069void 2070pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 2071 boolean_t wired) 2072{ 2073 vm_offset_t pa; 2074 register pt_entry_t *pte; 2075 vm_offset_t opa; 2076 pt_entry_t origpte, newpte; 2077 vm_page_t mpte; 2078 2079 if (pmap == NULL) 2080 return; 2081 2082 va &= PG_FRAME; 2083#ifdef PMAP_DIAGNOSTIC 2084 if (va > VM_MAX_KERNEL_ADDRESS) 2085 panic("pmap_enter: toobig"); 2086 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 2087 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 2088#endif 2089 2090 mpte = NULL; 2091 /* 2092 * In the case that a page table page is not 2093 * resident, we are creating it here. 2094 */ 2095 if (va < VM_MAXUSER_ADDRESS) { 2096 mpte = pmap_allocpte(pmap, va); 2097 } 2098#if 0 && defined(PMAP_DIAGNOSTIC) 2099 else { 2100 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 2101 origpte = *pdeaddr; 2102 if ((origpte & PG_V) == 0) { 2103 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 2104 pmap->pm_pdir[PTDPTDI], origpte, va); 2105 } 2106 } 2107#endif 2108 2109 pte = pmap_pte(pmap, va); 2110 2111 /* 2112 * Page Directory table entry not valid, we need a new PT page 2113 */ 2114 if (pte == NULL) { 2115 panic("pmap_enter: invalid page directory, pdir=%p, va=0x%x\n", 2116 (void *)pmap->pm_pdir[PTDPTDI], va); 2117 } 2118 2119 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 2120 origpte = *(vm_offset_t *)pte; 2121 opa = origpte & PG_FRAME; 2122 2123 if (origpte & PG_PS) 2124 panic("pmap_enter: attempted pmap_enter on 4MB page"); 2125 2126 /* 2127 * Mapping has not changed, must be protection or wiring change. 2128 */ 2129 if (origpte && (opa == pa)) { 2130 /* 2131 * Wiring change, just update stats. We don't worry about 2132 * wiring PT pages as they remain resident as long as there 2133 * are valid mappings in them. Hence, if a user page is wired, 2134 * the PT page will be also. 2135 */ 2136 if (wired && ((origpte & PG_W) == 0)) 2137 pmap->pm_stats.wired_count++; 2138 else if (!wired && (origpte & PG_W)) 2139 pmap->pm_stats.wired_count--; 2140 2141#if defined(PMAP_DIAGNOSTIC) 2142 if (pmap_nw_modified((pt_entry_t) origpte)) { 2143 printf( 2144 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 2145 va, origpte); 2146 } 2147#endif 2148 2149 /* 2150 * Remove extra pte reference 2151 */ 2152 if (mpte) 2153 mpte->hold_count--; 2154 2155 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) { 2156 if ((origpte & PG_RW) == 0) { 2157 *pte |= PG_RW; 2158 pmap_invalidate_page(pmap, va); 2159 } 2160 return; 2161 } 2162 2163 /* 2164 * We might be turning off write access to the page, 2165 * so we go ahead and sense modify status. 2166 */ 2167 if (origpte & PG_MANAGED) { 2168 if ((origpte & PG_M) && pmap_track_modified(va)) { 2169 vm_page_t om; 2170 om = PHYS_TO_VM_PAGE(opa); 2171 vm_page_dirty(om); 2172 } 2173 pa |= PG_MANAGED; 2174 } 2175 goto validate; 2176 } 2177 /* 2178 * Mapping has changed, invalidate old range and fall through to 2179 * handle validating new mapping. 2180 */ 2181 if (opa) { 2182 int err; 2183 err = pmap_remove_pte(pmap, pte, va); 2184 if (err) 2185 panic("pmap_enter: pte vanished, va: 0x%x", va); 2186 } 2187 2188 /* 2189 * Enter on the PV list if part of our managed memory. Note that we 2190 * raise IPL while manipulating pv_table since pmap_enter can be 2191 * called at interrupt time. 2192 */ 2193 if (pmap_initialized && 2194 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2195 pmap_insert_entry(pmap, va, mpte, m); 2196 pa |= PG_MANAGED; 2197 } 2198 2199 /* 2200 * Increment counters 2201 */ 2202 pmap->pm_stats.resident_count++; 2203 if (wired) 2204 pmap->pm_stats.wired_count++; 2205 2206validate: 2207 /* 2208 * Now validate mapping with desired protection/wiring. 2209 */ 2210 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V); 2211 2212 if (wired) 2213 newpte |= PG_W; 2214 if (va < VM_MAXUSER_ADDRESS) 2215 newpte |= PG_U; 2216 if (pmap == kernel_pmap) 2217 newpte |= pgeflag; 2218 2219 /* 2220 * if the mapping or permission bits are different, we need 2221 * to update the pte. 2222 */ 2223 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2224 *pte = newpte | PG_A; 2225 /*if (origpte)*/ { 2226 pmap_invalidate_page(pmap, va); 2227 } 2228 } 2229} 2230 2231/* 2232 * this code makes some *MAJOR* assumptions: 2233 * 1. Current pmap & pmap exists. 2234 * 2. Not wired. 2235 * 3. Read access. 2236 * 4. No page table pages. 2237 * 5. Tlbflush is deferred to calling procedure. 2238 * 6. Page IS managed. 2239 * but is *MUCH* faster than pmap_enter... 2240 */ 2241 2242static vm_page_t 2243pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2244{ 2245 pt_entry_t *pte; 2246 vm_offset_t pa; 2247 2248 /* 2249 * In the case that a page table page is not 2250 * resident, we are creating it here. 2251 */ 2252 if (va < VM_MAXUSER_ADDRESS) { 2253 unsigned ptepindex; 2254 pd_entry_t ptepa; 2255 2256 /* 2257 * Calculate pagetable page index 2258 */ 2259 ptepindex = va >> PDRSHIFT; 2260 if (mpte && (mpte->pindex == ptepindex)) { 2261 mpte->hold_count++; 2262 } else { 2263retry: 2264 /* 2265 * Get the page directory entry 2266 */ 2267 ptepa = pmap->pm_pdir[ptepindex]; 2268 2269 /* 2270 * If the page table page is mapped, we just increment 2271 * the hold count, and activate it. 2272 */ 2273 if (ptepa) { 2274 if (ptepa & PG_PS) 2275 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2276 if (pmap->pm_ptphint && 2277 (pmap->pm_ptphint->pindex == ptepindex)) { 2278 mpte = pmap->pm_ptphint; 2279 } else { 2280 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 2281 pmap->pm_ptphint = mpte; 2282 } 2283 if (mpte == NULL) 2284 goto retry; 2285 mpte->hold_count++; 2286 } else { 2287 mpte = _pmap_allocpte(pmap, ptepindex); 2288 } 2289 } 2290 } else { 2291 mpte = NULL; 2292 } 2293 2294 /* 2295 * This call to vtopte makes the assumption that we are 2296 * entering the page into the current pmap. In order to support 2297 * quick entry into any pmap, one would likely use pmap_pte_quick. 2298 * But that isn't as quick as vtopte. 2299 */ 2300 pte = vtopte(va); 2301 if (*pte) { 2302 if (mpte) 2303 pmap_unwire_pte_hold(pmap, mpte); 2304 return 0; 2305 } 2306 2307 /* 2308 * Enter on the PV list if part of our managed memory. Note that we 2309 * raise IPL while manipulating pv_table since pmap_enter can be 2310 * called at interrupt time. 2311 */ 2312 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2313 pmap_insert_entry(pmap, va, mpte, m); 2314 2315 /* 2316 * Increment counters 2317 */ 2318 pmap->pm_stats.resident_count++; 2319 2320 pa = VM_PAGE_TO_PHYS(m); 2321 2322 /* 2323 * Now validate mapping with RO protection 2324 */ 2325 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2326 *pte = pa | PG_V | PG_U; 2327 else 2328 *pte = pa | PG_V | PG_U | PG_MANAGED; 2329 2330 return mpte; 2331} 2332 2333/* 2334 * Make a temporary mapping for a physical address. This is only intended 2335 * to be used for panic dumps. 2336 */ 2337void * 2338pmap_kenter_temporary(vm_offset_t pa, int i) 2339{ 2340 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa); 2341 invlpg((vm_offset_t)crashdumpmap + (i * PAGE_SIZE)); 2342 return ((void *)crashdumpmap); 2343} 2344 2345#define MAX_INIT_PT (96) 2346/* 2347 * pmap_object_init_pt preloads the ptes for a given object 2348 * into the specified pmap. This eliminates the blast of soft 2349 * faults on process startup and immediately after an mmap. 2350 */ 2351void 2352pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2353 vm_object_t object, vm_pindex_t pindex, 2354 vm_size_t size, int limit) 2355{ 2356 vm_offset_t tmpidx; 2357 int psize; 2358 vm_page_t p, mpte; 2359 int objpgs; 2360 2361 if (pmap == NULL || object == NULL) 2362 return; 2363 2364 /* 2365 * This code maps large physical mmap regions into the 2366 * processor address space. Note that some shortcuts 2367 * are taken, but the code works. 2368 */ 2369 if (pseflag && (object->type == OBJT_DEVICE) && 2370 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2371 int i; 2372 vm_page_t m[1]; 2373 unsigned int ptepindex; 2374 int npdes; 2375 pd_entry_t ptepa; 2376 2377 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2378 return; 2379 2380retry: 2381 p = vm_page_lookup(object, pindex); 2382 if (p && vm_page_sleep_busy(p, FALSE, "init4p")) 2383 goto retry; 2384 2385 if (p == NULL) { 2386 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2387 if (p == NULL) 2388 return; 2389 m[0] = p; 2390 2391 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2392 vm_page_free(p); 2393 return; 2394 } 2395 2396 p = vm_page_lookup(object, pindex); 2397 vm_page_wakeup(p); 2398 } 2399 2400 ptepa = VM_PAGE_TO_PHYS(p); 2401 if (ptepa & (NBPDR - 1)) { 2402 return; 2403 } 2404 2405 p->valid = VM_PAGE_BITS_ALL; 2406 2407 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2408 npdes = size >> PDRSHIFT; 2409 for(i = 0; i < npdes; i++) { 2410 pmap->pm_pdir[ptepindex] = 2411 ptepa | PG_U | PG_RW | PG_V | PG_PS; 2412 ptepa += NBPDR; 2413 ptepindex += 1; 2414 } 2415 vm_page_flag_set(p, PG_MAPPED); 2416 invltlb(); 2417 return; 2418 } 2419 2420 psize = i386_btop(size); 2421 2422 if ((object->type != OBJT_VNODE) || 2423 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) && 2424 (object->resident_page_count > MAX_INIT_PT))) { 2425 return; 2426 } 2427 2428 if (psize + pindex > object->size) { 2429 if (object->size < pindex) 2430 return; 2431 psize = object->size - pindex; 2432 } 2433 2434 mpte = NULL; 2435 /* 2436 * if we are processing a major portion of the object, then scan the 2437 * entire thing. 2438 */ 2439 if (psize > (object->resident_page_count >> 2)) { 2440 objpgs = psize; 2441 2442 for (p = TAILQ_FIRST(&object->memq); 2443 ((objpgs > 0) && (p != NULL)); 2444 p = TAILQ_NEXT(p, listq)) { 2445 2446 tmpidx = p->pindex; 2447 if (tmpidx < pindex) { 2448 continue; 2449 } 2450 tmpidx -= pindex; 2451 if (tmpidx >= psize) { 2452 continue; 2453 } 2454 /* 2455 * don't allow an madvise to blow away our really 2456 * free pages allocating pv entries. 2457 */ 2458 if ((limit & MAP_PREFAULT_MADVISE) && 2459 cnt.v_free_count < cnt.v_free_reserved) { 2460 break; 2461 } 2462 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2463 (p->busy == 0) && 2464 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2465 if ((p->queue - p->pc) == PQ_CACHE) 2466 vm_page_deactivate(p); 2467 vm_page_busy(p); 2468 mpte = pmap_enter_quick(pmap, 2469 addr + i386_ptob(tmpidx), p, mpte); 2470 vm_page_flag_set(p, PG_MAPPED); 2471 vm_page_wakeup(p); 2472 } 2473 objpgs -= 1; 2474 } 2475 } else { 2476 /* 2477 * else lookup the pages one-by-one. 2478 */ 2479 for (tmpidx = 0; tmpidx < psize; tmpidx += 1) { 2480 /* 2481 * don't allow an madvise to blow away our really 2482 * free pages allocating pv entries. 2483 */ 2484 if ((limit & MAP_PREFAULT_MADVISE) && 2485 cnt.v_free_count < cnt.v_free_reserved) { 2486 break; 2487 } 2488 p = vm_page_lookup(object, tmpidx + pindex); 2489 if (p && 2490 ((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2491 (p->busy == 0) && 2492 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2493 if ((p->queue - p->pc) == PQ_CACHE) 2494 vm_page_deactivate(p); 2495 vm_page_busy(p); 2496 mpte = pmap_enter_quick(pmap, 2497 addr + i386_ptob(tmpidx), p, mpte); 2498 vm_page_flag_set(p, PG_MAPPED); 2499 vm_page_wakeup(p); 2500 } 2501 } 2502 } 2503 return; 2504} 2505 2506/* 2507 * pmap_prefault provides a quick way of clustering 2508 * pagefaults into a processes address space. It is a "cousin" 2509 * of pmap_object_init_pt, except it runs at page fault time instead 2510 * of mmap time. 2511 */ 2512#define PFBAK 4 2513#define PFFOR 4 2514#define PAGEORDER_SIZE (PFBAK+PFFOR) 2515 2516static int pmap_prefault_pageorder[] = { 2517 -PAGE_SIZE, PAGE_SIZE, 2518 -2 * PAGE_SIZE, 2 * PAGE_SIZE, 2519 -3 * PAGE_SIZE, 3 * PAGE_SIZE 2520 -4 * PAGE_SIZE, 4 * PAGE_SIZE 2521}; 2522 2523void 2524pmap_prefault(pmap, addra, entry) 2525 pmap_t pmap; 2526 vm_offset_t addra; 2527 vm_map_entry_t entry; 2528{ 2529 int i; 2530 vm_offset_t starta; 2531 vm_offset_t addr; 2532 vm_pindex_t pindex; 2533 vm_page_t m, mpte; 2534 vm_object_t object; 2535 2536 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) 2537 return; 2538 2539 object = entry->object.vm_object; 2540 2541 starta = addra - PFBAK * PAGE_SIZE; 2542 if (starta < entry->start) { 2543 starta = entry->start; 2544 } else if (starta > addra) { 2545 starta = 0; 2546 } 2547 2548 mpte = NULL; 2549 for (i = 0; i < PAGEORDER_SIZE; i++) { 2550 vm_object_t lobject; 2551 pt_entry_t *pte; 2552 2553 addr = addra + pmap_prefault_pageorder[i]; 2554 if (addr > addra + (PFFOR * PAGE_SIZE)) 2555 addr = 0; 2556 2557 if (addr < starta || addr >= entry->end) 2558 continue; 2559 2560 if ((*pmap_pde(pmap, addr)) == NULL) 2561 continue; 2562 2563 pte = vtopte(addr); 2564 if (*pte) 2565 continue; 2566 2567 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2568 lobject = object; 2569 for (m = vm_page_lookup(lobject, pindex); 2570 (!m && (lobject->type == OBJT_DEFAULT) && (lobject->backing_object)); 2571 lobject = lobject->backing_object) { 2572 if (lobject->backing_object_offset & PAGE_MASK) 2573 break; 2574 pindex += (lobject->backing_object_offset >> PAGE_SHIFT); 2575 m = vm_page_lookup(lobject->backing_object, pindex); 2576 } 2577 2578 /* 2579 * give-up when a page is not in memory 2580 */ 2581 if (m == NULL) 2582 break; 2583 2584 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2585 (m->busy == 0) && 2586 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2587 2588 if ((m->queue - m->pc) == PQ_CACHE) { 2589 vm_page_deactivate(m); 2590 } 2591 vm_page_busy(m); 2592 mpte = pmap_enter_quick(pmap, addr, m, mpte); 2593 vm_page_flag_set(m, PG_MAPPED); 2594 vm_page_wakeup(m); 2595 } 2596 } 2597} 2598 2599/* 2600 * Routine: pmap_change_wiring 2601 * Function: Change the wiring attribute for a map/virtual-address 2602 * pair. 2603 * In/out conditions: 2604 * The mapping must already exist in the pmap. 2605 */ 2606void 2607pmap_change_wiring(pmap, va, wired) 2608 register pmap_t pmap; 2609 vm_offset_t va; 2610 boolean_t wired; 2611{ 2612 register pt_entry_t *pte; 2613 2614 if (pmap == NULL) 2615 return; 2616 2617 pte = pmap_pte(pmap, va); 2618 2619 if (wired && !pmap_pte_w(pte)) 2620 pmap->pm_stats.wired_count++; 2621 else if (!wired && pmap_pte_w(pte)) 2622 pmap->pm_stats.wired_count--; 2623 2624 /* 2625 * Wiring is not a hardware characteristic so there is no need to 2626 * invalidate TLB. 2627 */ 2628 pmap_pte_set_w(pte, wired); 2629} 2630 2631 2632 2633/* 2634 * Copy the range specified by src_addr/len 2635 * from the source map to the range dst_addr/len 2636 * in the destination map. 2637 * 2638 * This routine is only advisory and need not do anything. 2639 */ 2640 2641void 2642pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2643 vm_offset_t src_addr) 2644{ 2645 vm_offset_t addr; 2646 vm_offset_t end_addr = src_addr + len; 2647 vm_offset_t pdnxt; 2648 pd_entry_t src_frame, dst_frame; 2649 vm_page_t m; 2650 2651 if (dst_addr != src_addr) 2652 return; 2653 2654 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME; 2655 if (src_frame != (PTDpde & PG_FRAME)) 2656 return; 2657 2658 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME; 2659 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2660 pt_entry_t *src_pte, *dst_pte; 2661 vm_page_t dstmpte, srcmpte; 2662 pd_entry_t srcptepaddr; 2663 unsigned ptepindex; 2664 2665 if (addr >= UPT_MIN_ADDRESS) 2666 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2667 2668 /* 2669 * Don't let optional prefaulting of pages make us go 2670 * way below the low water mark of free pages or way 2671 * above high water mark of used pv entries. 2672 */ 2673 if (cnt.v_free_count < cnt.v_free_reserved || 2674 pv_entry_count > pv_entry_high_water) 2675 break; 2676 2677 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1)); 2678 ptepindex = addr >> PDRSHIFT; 2679 2680 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2681 if (srcptepaddr == 0) 2682 continue; 2683 2684 if (srcptepaddr & PG_PS) { 2685 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2686 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2687 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; 2688 } 2689 continue; 2690 } 2691 2692 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex); 2693 if ((srcmpte == NULL) || 2694 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY)) 2695 continue; 2696 2697 if (pdnxt > end_addr) 2698 pdnxt = end_addr; 2699 2700 /* 2701 * Have to recheck this before every avtopte() call below 2702 * in case we have blocked and something else used APTDpde. 2703 */ 2704 if (dst_frame != (APTDpde & PG_FRAME)) { 2705 APTDpde = dst_frame | PG_RW | PG_V; 2706 invltlb(); 2707 } 2708 src_pte = vtopte(addr); 2709 dst_pte = avtopte(addr); 2710 while (addr < pdnxt) { 2711 pt_entry_t ptetemp; 2712 ptetemp = *src_pte; 2713 /* 2714 * we only virtual copy managed pages 2715 */ 2716 if ((ptetemp & PG_MANAGED) != 0) { 2717 /* 2718 * We have to check after allocpte for the 2719 * pte still being around... allocpte can 2720 * block. 2721 */ 2722 dstmpte = pmap_allocpte(dst_pmap, addr); 2723 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2724 /* 2725 * Clear the modified and 2726 * accessed (referenced) bits 2727 * during the copy. 2728 */ 2729 m = PHYS_TO_VM_PAGE(ptetemp); 2730 *dst_pte = ptetemp & ~(PG_M | PG_A); 2731 dst_pmap->pm_stats.resident_count++; 2732 pmap_insert_entry(dst_pmap, addr, 2733 dstmpte, m); 2734 } else { 2735 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2736 } 2737 if (dstmpte->hold_count >= srcmpte->hold_count) 2738 break; 2739 } 2740 addr += PAGE_SIZE; 2741 src_pte++; 2742 dst_pte++; 2743 } 2744 } 2745} 2746 2747/* 2748 * Routine: pmap_kernel 2749 * Function: 2750 * Returns the physical map handle for the kernel. 2751 */ 2752pmap_t 2753pmap_kernel() 2754{ 2755 return (kernel_pmap); 2756} 2757 2758/* 2759 * pmap_zero_page zeros the specified hardware page by mapping 2760 * the page into KVM and using bzero to clear its contents. 2761 */ 2762void 2763pmap_zero_page(vm_offset_t phys) 2764{ 2765 2766#ifdef SMP 2767 /* XXX overkill, we only want to disable migration here */ 2768 /* XXX or maybe not. down the track we have reentrancy issues */ 2769 critical_enter(); 2770#endif 2771 if (*CMAP2) 2772 panic("pmap_zero_page: CMAP2 busy"); 2773 *CMAP2 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M; 2774 cpu_invlpg((vm_offset_t)CADDR2); /* SMP: local cpu only */ 2775#if defined(I686_CPU) 2776 if (cpu_class == CPUCLASS_686) 2777 i686_pagezero(CADDR2); 2778 else 2779#endif 2780 bzero(CADDR2, PAGE_SIZE); 2781 *CMAP2 = 0; 2782#ifdef SMP 2783 critical_exit(); 2784#endif 2785} 2786 2787/* 2788 * pmap_zero_page_area zeros the specified hardware page by mapping 2789 * the page into KVM and using bzero to clear its contents. 2790 * 2791 * off and size may not cover an area beyond a single hardware page. 2792 */ 2793void 2794pmap_zero_page_area(vm_offset_t phys, int off, int size) 2795{ 2796 2797#ifdef SMP 2798 /* XXX overkill, we only want to disable migration here */ 2799 /* XXX or maybe not. down the track we have reentrancy issues */ 2800 critical_enter(); 2801#endif 2802 if (*CMAP2) 2803 panic("pmap_zero_page: CMAP2 busy"); 2804 *CMAP2 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M; 2805 cpu_invlpg((vm_offset_t)CADDR2); /* SMP: local cpu only */ 2806#if defined(I686_CPU) 2807 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE) 2808 i686_pagezero(CADDR2); 2809 else 2810#endif 2811 bzero((char *)CADDR2 + off, size); 2812 *CMAP2 = 0; 2813#ifdef SMP 2814 critical_exit(); 2815#endif 2816} 2817 2818/* 2819 * pmap_copy_page copies the specified (machine independent) 2820 * page by mapping the page into virtual memory and using 2821 * bcopy to copy the page, one machine dependent page at a 2822 * time. 2823 */ 2824void 2825pmap_copy_page(vm_offset_t src, vm_offset_t dst) 2826{ 2827 2828#ifdef SMP 2829 /* XXX overkill, we only want to disable migration here */ 2830 /* XXX or maybe not. down the track we have reentrancy issues */ 2831 critical_enter(); 2832#endif 2833 if (*CMAP1) 2834 panic("pmap_copy_page: CMAP1 busy"); 2835 if (*CMAP2) 2836 panic("pmap_copy_page: CMAP2 busy"); 2837 2838 *CMAP1 = PG_V | (src & PG_FRAME) | PG_A; 2839 *CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M; 2840 cpu_invlpg((u_int)CADDR1); /* SMP: local only */ 2841 cpu_invlpg((u_int)CADDR2); /* SMP: local only */ 2842 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2843 *CMAP1 = 0; 2844 *CMAP2 = 0; 2845#ifdef SMP 2846 critical_exit(); 2847#endif 2848} 2849 2850 2851/* 2852 * Routine: pmap_pageable 2853 * Function: 2854 * Make the specified pages (by pmap, offset) 2855 * pageable (or not) as requested. 2856 * 2857 * A page which is not pageable may not take 2858 * a fault; therefore, its page table entry 2859 * must remain valid for the duration. 2860 * 2861 * This routine is merely advisory; pmap_enter 2862 * will specify that these pages are to be wired 2863 * down (or not) as appropriate. 2864 */ 2865void 2866pmap_pageable(pmap, sva, eva, pageable) 2867 pmap_t pmap; 2868 vm_offset_t sva, eva; 2869 boolean_t pageable; 2870{ 2871} 2872 2873/* 2874 * this routine returns true if a physical page resides 2875 * in the given pmap. 2876 */ 2877boolean_t 2878pmap_page_exists(pmap, m) 2879 pmap_t pmap; 2880 vm_page_t m; 2881{ 2882 register pv_entry_t pv; 2883 int s; 2884 2885 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2886 return FALSE; 2887 2888 s = splvm(); 2889 2890 /* 2891 * Not found, check current mappings returning immediately if found. 2892 */ 2893 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2894 if (pv->pv_pmap == pmap) { 2895 splx(s); 2896 return TRUE; 2897 } 2898 } 2899 splx(s); 2900 return (FALSE); 2901} 2902 2903#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2904/* 2905 * Remove all pages from specified address space 2906 * this aids process exit speeds. Also, this code 2907 * is special cased for current process only, but 2908 * can have the more generic (and slightly slower) 2909 * mode enabled. This is much faster than pmap_remove 2910 * in the case of running down an entire address space. 2911 */ 2912void 2913pmap_remove_pages(pmap, sva, eva) 2914 pmap_t pmap; 2915 vm_offset_t sva, eva; 2916{ 2917 pt_entry_t *pte, tpte; 2918 vm_page_t m; 2919 pv_entry_t pv, npv; 2920 int s; 2921 2922#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2923 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2924 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2925 return; 2926 } 2927#endif 2928 2929 s = splvm(); 2930 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2931 2932 if (pv->pv_va >= eva || pv->pv_va < sva) { 2933 npv = TAILQ_NEXT(pv, pv_plist); 2934 continue; 2935 } 2936 2937#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2938 pte = vtopte(pv->pv_va); 2939#else 2940 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2941#endif 2942 tpte = *pte; 2943 2944 if (tpte == 0) { 2945 printf("TPTE at %p IS ZERO @ VA %08x\n", 2946 pte, pv->pv_va); 2947 panic("bad peter"); 2948 } 2949 2950/* 2951 * We cannot remove wired pages from a process' mapping at this time 2952 */ 2953 if (tpte & PG_W) { 2954 npv = TAILQ_NEXT(pv, pv_plist); 2955 continue; 2956 } 2957 2958 m = PHYS_TO_VM_PAGE(tpte); 2959 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2960 ("vm_page_t %p phys_addr mismatch %08x %08x", 2961 m, m->phys_addr, tpte)); 2962 2963 KASSERT(m < &vm_page_array[vm_page_array_size], 2964 ("pmap_remove_pages: bad tpte %x", tpte)); 2965 2966 pv->pv_pmap->pm_stats.resident_count--; 2967 2968 *pte = 0; 2969 2970 /* 2971 * Update the vm_page_t clean and reference bits. 2972 */ 2973 if (tpte & PG_M) { 2974 vm_page_dirty(m); 2975 } 2976 2977 npv = TAILQ_NEXT(pv, pv_plist); 2978 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2979 2980 m->md.pv_list_count--; 2981 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2982 if (TAILQ_FIRST(&m->md.pv_list) == NULL) { 2983 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 2984 } 2985 2986 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2987 free_pv_entry(pv); 2988 } 2989 splx(s); 2990 pmap_invalidate_all(pmap); 2991} 2992 2993/* 2994 * pmap_testbit tests bits in pte's 2995 * note that the testbit/changebit routines are inline, 2996 * and a lot of things compile-time evaluate. 2997 */ 2998static boolean_t 2999pmap_testbit(m, bit) 3000 vm_page_t m; 3001 int bit; 3002{ 3003 pv_entry_t pv; 3004 pt_entry_t *pte; 3005 int s; 3006 3007 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3008 return FALSE; 3009 3010 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 3011 return FALSE; 3012 3013 s = splvm(); 3014 3015 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3016 /* 3017 * if the bit being tested is the modified bit, then 3018 * mark clean_map and ptes as never 3019 * modified. 3020 */ 3021 if (bit & (PG_A|PG_M)) { 3022 if (!pmap_track_modified(pv->pv_va)) 3023 continue; 3024 } 3025 3026#if defined(PMAP_DIAGNOSTIC) 3027 if (!pv->pv_pmap) { 3028 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 3029 continue; 3030 } 3031#endif 3032 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3033 if (*pte & bit) { 3034 splx(s); 3035 return TRUE; 3036 } 3037 } 3038 splx(s); 3039 return (FALSE); 3040} 3041 3042/* 3043 * this routine is used to modify bits in ptes 3044 */ 3045static __inline void 3046pmap_changebit(vm_page_t m, int bit, boolean_t setem) 3047{ 3048 register pv_entry_t pv; 3049 register pt_entry_t *pte; 3050 int s; 3051 3052 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3053 return; 3054 3055 s = splvm(); 3056 3057 /* 3058 * Loop over all current mappings setting/clearing as appropos If 3059 * setting RO do we need to clear the VAC? 3060 */ 3061 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3062 /* 3063 * don't write protect pager mappings 3064 */ 3065 if (!setem && (bit == PG_RW)) { 3066 if (!pmap_track_modified(pv->pv_va)) 3067 continue; 3068 } 3069 3070#if defined(PMAP_DIAGNOSTIC) 3071 if (!pv->pv_pmap) { 3072 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 3073 continue; 3074 } 3075#endif 3076 3077 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3078 3079 if (setem) { 3080 *pte |= bit; 3081 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3082 } else { 3083 pt_entry_t pbits = *pte; 3084 if (pbits & bit) { 3085 if (bit == PG_RW) { 3086 if (pbits & PG_M) { 3087 vm_page_dirty(m); 3088 } 3089 *pte = pbits & ~(PG_M|PG_RW); 3090 } else { 3091 *pte = pbits & ~bit; 3092 } 3093 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3094 } 3095 } 3096 } 3097 splx(s); 3098} 3099 3100/* 3101 * pmap_page_protect: 3102 * 3103 * Lower the permission for all mappings to a given page. 3104 */ 3105void 3106pmap_page_protect(vm_page_t m, vm_prot_t prot) 3107{ 3108 if ((prot & VM_PROT_WRITE) == 0) { 3109 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 3110 pmap_changebit(m, PG_RW, FALSE); 3111 } else { 3112 pmap_remove_all(m); 3113 } 3114 } 3115} 3116 3117vm_offset_t 3118pmap_phys_address(ppn) 3119 int ppn; 3120{ 3121 return (i386_ptob(ppn)); 3122} 3123 3124/* 3125 * pmap_ts_referenced: 3126 * 3127 * Return the count of reference bits for a page, clearing all of them. 3128 */ 3129int 3130pmap_ts_referenced(vm_page_t m) 3131{ 3132 register pv_entry_t pv, pvf, pvn; 3133 pt_entry_t *pte; 3134 int s; 3135 int rtval = 0; 3136 3137 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3138 return (rtval); 3139 3140 s = splvm(); 3141 3142 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 3143 3144 pvf = pv; 3145 3146 do { 3147 pvn = TAILQ_NEXT(pv, pv_list); 3148 3149 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 3150 3151 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 3152 3153 if (!pmap_track_modified(pv->pv_va)) 3154 continue; 3155 3156 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3157 3158 if (pte && (*pte & PG_A)) { 3159 *pte &= ~PG_A; 3160 3161 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3162 3163 rtval++; 3164 if (rtval > 4) { 3165 break; 3166 } 3167 } 3168 } while ((pv = pvn) != NULL && pv != pvf); 3169 } 3170 splx(s); 3171 3172 return (rtval); 3173} 3174 3175/* 3176 * pmap_is_modified: 3177 * 3178 * Return whether or not the specified physical page was modified 3179 * in any physical maps. 3180 */ 3181boolean_t 3182pmap_is_modified(vm_page_t m) 3183{ 3184 return pmap_testbit(m, PG_M); 3185} 3186 3187/* 3188 * Clear the modify bits on the specified physical page. 3189 */ 3190void 3191pmap_clear_modify(vm_page_t m) 3192{ 3193 pmap_changebit(m, PG_M, FALSE); 3194} 3195 3196/* 3197 * pmap_clear_reference: 3198 * 3199 * Clear the reference bit on the specified physical page. 3200 */ 3201void 3202pmap_clear_reference(vm_page_t m) 3203{ 3204 pmap_changebit(m, PG_A, FALSE); 3205} 3206 3207/* 3208 * Miscellaneous support routines follow 3209 */ 3210 3211static void 3212i386_protection_init() 3213{ 3214 register int *kp, prot; 3215 3216 kp = protection_codes; 3217 for (prot = 0; prot < 8; prot++) { 3218 switch (prot) { 3219 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE: 3220 /* 3221 * Read access is also 0. There isn't any execute bit, 3222 * so just make it readable. 3223 */ 3224 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE: 3225 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE: 3226 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE: 3227 *kp++ = 0; 3228 break; 3229 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE: 3230 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE: 3231 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE: 3232 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE: 3233 *kp++ = PG_RW; 3234 break; 3235 } 3236 } 3237} 3238 3239/* 3240 * Map a set of physical memory pages into the kernel virtual 3241 * address space. Return a pointer to where it is mapped. This 3242 * routine is intended to be used for mapping device memory, 3243 * NOT real memory. 3244 */ 3245void * 3246pmap_mapdev(pa, size) 3247 vm_offset_t pa; 3248 vm_size_t size; 3249{ 3250 vm_offset_t va, tmpva, offset; 3251 pt_entry_t *pte; 3252 3253 offset = pa & PAGE_MASK; 3254 size = roundup(offset + size, PAGE_SIZE); 3255 3256 GIANT_REQUIRED; 3257 3258 va = kmem_alloc_pageable(kernel_map, size); 3259 if (!va) 3260 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 3261 3262 pa = pa & PG_FRAME; 3263 for (tmpva = va; size > 0; ) { 3264 pte = vtopte(tmpva); 3265 *pte = pa | PG_RW | PG_V | pgeflag; 3266 size -= PAGE_SIZE; 3267 tmpva += PAGE_SIZE; 3268 } 3269 invlpg_range(va, tmpva); 3270 3271 return ((void *)(va + offset)); 3272} 3273 3274void 3275pmap_unmapdev(va, size) 3276 vm_offset_t va; 3277 vm_size_t size; 3278{ 3279 vm_offset_t base, offset, tmpva; 3280 pt_entry_t *pte; 3281 3282 base = va & PG_FRAME; 3283 offset = va & PAGE_MASK; 3284 size = roundup(offset + size, PAGE_SIZE); 3285 3286 for (tmpva = base; size > 0; ) { 3287 pte = vtopte(tmpva); 3288 *pte = 0; 3289 size -= PAGE_SIZE; 3290 tmpva += PAGE_SIZE; 3291 } 3292 invlpg_range(va, tmpva); 3293 kmem_free(kernel_map, base, size); 3294} 3295 3296/* 3297 * perform the pmap work for mincore 3298 */ 3299int 3300pmap_mincore(pmap, addr) 3301 pmap_t pmap; 3302 vm_offset_t addr; 3303{ 3304 pt_entry_t *ptep, pte; 3305 vm_page_t m; 3306 int val = 0; 3307 3308 ptep = pmap_pte(pmap, addr); 3309 if (ptep == 0) { 3310 return 0; 3311 } 3312 3313 if ((pte = *ptep) != 0) { 3314 vm_offset_t pa; 3315 3316 val = MINCORE_INCORE; 3317 if ((pte & PG_MANAGED) == 0) 3318 return val; 3319 3320 pa = pte & PG_FRAME; 3321 3322 m = PHYS_TO_VM_PAGE(pa); 3323 3324 /* 3325 * Modified by us 3326 */ 3327 if (pte & PG_M) 3328 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 3329 /* 3330 * Modified by someone 3331 */ 3332 else if (m->dirty || pmap_is_modified(m)) 3333 val |= MINCORE_MODIFIED_OTHER; 3334 /* 3335 * Referenced by us 3336 */ 3337 if (pte & PG_A) 3338 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 3339 3340 /* 3341 * Referenced by someone 3342 */ 3343 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) { 3344 val |= MINCORE_REFERENCED_OTHER; 3345 vm_page_flag_set(m, PG_REFERENCED); 3346 } 3347 } 3348 return val; 3349} 3350 3351void 3352pmap_activate(struct thread *td) 3353{ 3354 struct proc *p = td->td_proc; 3355 pmap_t pmap; 3356 u_int32_t cr3; 3357 3358 pmap = vmspace_pmap(td->td_proc->p_vmspace); 3359#if defined(SMP) 3360 pmap->pm_active |= PCPU_GET(cpumask); 3361#else 3362 pmap->pm_active |= 1; 3363#endif 3364#if defined(SWTCH_OPTIM_STATS) 3365 tlb_flush_count++; 3366#endif 3367 cr3 = vtophys(pmap->pm_pdir); 3368 /* XXXKSE this is wrong. 3369 * pmap_activate is for the current thread on the current cpu 3370 */ 3371 if (p->p_flag & P_KSES) { 3372 /* Make sure all other cr3 entries are updated. */ 3373 /* what if they are running? XXXKSE (maybe abort them) */ 3374 FOREACH_THREAD_IN_PROC(p, td) { 3375 td->td_pcb->pcb_cr3 = cr3; 3376 } 3377 } else { 3378 td->td_pcb->pcb_cr3 = cr3; 3379 } 3380 load_cr3(cr3); 3381} 3382 3383vm_offset_t 3384pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3385{ 3386 3387 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3388 return addr; 3389 } 3390 3391 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); 3392 return addr; 3393} 3394 3395 3396#if defined(PMAP_DEBUG) 3397pmap_pid_dump(int pid) 3398{ 3399 pmap_t pmap; 3400 struct proc *p; 3401 int npte = 0; 3402 int index; 3403 3404 sx_slock(&allproc_lock); 3405 LIST_FOREACH(p, &allproc, p_list) { 3406 if (p->p_pid != pid) 3407 continue; 3408 3409 if (p->p_vmspace) { 3410 int i,j; 3411 index = 0; 3412 pmap = vmspace_pmap(p->p_vmspace); 3413 for (i = 0; i < NPDEPG; i++) { 3414 pd_entry_t *pde; 3415 pt_entry_t *pte; 3416 vm_offset_t base = i << PDRSHIFT; 3417 3418 pde = &pmap->pm_pdir[i]; 3419 if (pde && pmap_pde_v(pde)) { 3420 for (j = 0; j < NPTEPG; j++) { 3421 vm_offset_t va = base + (j << PAGE_SHIFT); 3422 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3423 if (index) { 3424 index = 0; 3425 printf("\n"); 3426 } 3427 sx_sunlock(&allproc_lock); 3428 return npte; 3429 } 3430 pte = pmap_pte_quick(pmap, va); 3431 if (pte && pmap_pte_v(pte)) { 3432 pt_entry_t pa; 3433 vm_page_t m; 3434 pa = *pte; 3435 m = PHYS_TO_VM_PAGE(pa); 3436 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3437 va, pa, m->hold_count, m->wire_count, m->flags); 3438 npte++; 3439 index++; 3440 if (index >= 2) { 3441 index = 0; 3442 printf("\n"); 3443 } else { 3444 printf(" "); 3445 } 3446 } 3447 } 3448 } 3449 } 3450 } 3451 } 3452 sx_sunlock(&allproc_lock); 3453 return npte; 3454} 3455#endif 3456 3457#if defined(DEBUG) 3458 3459static void pads __P((pmap_t pm)); 3460void pmap_pvdump __P((vm_offset_t pa)); 3461 3462/* print address space of pmap*/ 3463static void 3464pads(pm) 3465 pmap_t pm; 3466{ 3467 int i, j; 3468 vm_offset_t va; 3469 pt_entry_t *ptep; 3470 3471 if (pm == kernel_pmap) 3472 return; 3473 for (i = 0; i < NPDEPG; i++) 3474 if (pm->pm_pdir[i]) 3475 for (j = 0; j < NPTEPG; j++) { 3476 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3477 if (pm == kernel_pmap && va < KERNBASE) 3478 continue; 3479 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3480 continue; 3481 ptep = pmap_pte_quick(pm, va); 3482 if (pmap_pte_v(ptep)) 3483 printf("%x:%x ", va, *ptep); 3484 }; 3485 3486} 3487 3488void 3489pmap_pvdump(pa) 3490 vm_offset_t pa; 3491{ 3492 pv_entry_t pv; 3493 vm_page_t m; 3494 3495 printf("pa %x", pa); 3496 m = PHYS_TO_VM_PAGE(pa); 3497 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3498 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3499 pads(pv->pv_pmap); 3500 } 3501 printf(" "); 3502} 3503#endif 3504