pmap.c revision 99930
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 99930 2002-07-13 21:03:08Z 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_pmap.h" 72#include "opt_msgbuf.h" 73#include "opt_kstack_pages.h" 74 75#include <sys/param.h> 76#include <sys/systm.h> 77#include <sys/kernel.h> 78#include <sys/lock.h> 79#include <sys/mman.h> 80#include <sys/msgbuf.h> 81#include <sys/mutex.h> 82#include <sys/proc.h> 83#include <sys/sx.h> 84#include <sys/user.h> 85#include <sys/vmmeter.h> 86#include <sys/sysctl.h> 87#ifdef SMP 88#include <sys/smp.h> 89#endif 90 91#include <vm/vm.h> 92#include <vm/vm_param.h> 93#include <vm/vm_kern.h> 94#include <vm/vm_page.h> 95#include <vm/vm_map.h> 96#include <vm/vm_object.h> 97#include <vm/vm_extern.h> 98#include <vm/vm_pageout.h> 99#include <vm/vm_pager.h> 100#include <vm/uma.h> 101 102#include <machine/cpu.h> 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/smp.h> 108#include <machine/apic.h> 109#include <machine/segments.h> 110#include <machine/tss.h> 111#endif /* SMP || APIC_IO */ 112 113#define PMAP_KEEP_PDIRS 114#ifndef PMAP_SHPGPERPROC 115#define PMAP_SHPGPERPROC 200 116#endif 117 118#if defined(DIAGNOSTIC) 119#define PMAP_DIAGNOSTIC 120#endif 121 122#define MINPV 2048 123 124#if !defined(PMAP_DIAGNOSTIC) 125#define PMAP_INLINE __inline 126#else 127#define PMAP_INLINE 128#endif 129 130/* 131 * Get PDEs and PTEs for user/kernel address space 132 */ 133#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 134#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 135 136#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 137#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 138#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 139#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 140#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 141 142#define pmap_pte_set_w(pte, v) ((v)?(*(int *)pte |= PG_W):(*(int *)pte &= ~PG_W)) 143#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 144 145/* 146 * Given a map and a machine independent protection code, 147 * convert to a vax protection code. 148 */ 149#define pte_prot(m, p) (protection_codes[p]) 150static int protection_codes[8]; 151 152struct pmap kernel_pmap_store; 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; 168extern u_int32_t KERNend; 169 170/* 171 * Data for the pv entry allocation mechanism 172 */ 173static uma_zone_t pvzone; 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; 177 178/* 179 * All those kernel PT submaps that BSD is so fond of 180 */ 181pt_entry_t *CMAP1 = 0; 182static pt_entry_t *CMAP2, *CMAP3, *ptmmap; 183caddr_t CADDR1 = 0, ptvmmap = 0; 184static caddr_t CADDR2, CADDR3; 185static pt_entry_t *msgbufmap; 186struct msgbuf *msgbufp = 0; 187 188/* 189 * Crashdump maps. 190 */ 191static pt_entry_t *pt_crashdumpmap; 192static caddr_t crashdumpmap; 193 194#ifdef SMP 195extern pt_entry_t *SMPpt; 196#endif 197static pt_entry_t *PMAP1 = 0; 198static pt_entry_t *PADDR1 = 0; 199 200static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 201static pt_entry_t *get_ptbase(pmap_t pmap); 202static pv_entry_t get_pv_entry(void); 203static void i386_protection_init(void); 204static __inline void pmap_changebit(vm_page_t m, int bit, boolean_t setem); 205 206static void pmap_remove_all(vm_page_t m); 207static vm_page_t pmap_enter_quick(pmap_t pmap, vm_offset_t va, 208 vm_page_t m, vm_page_t mpte); 209static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 210static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 211static int pmap_remove_entry(struct pmap *pmap, vm_page_t m, 212 vm_offset_t va); 213static boolean_t pmap_testbit(vm_page_t m, int bit); 214static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, 215 vm_page_t mpte, vm_page_t m); 216 217static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va); 218 219static int pmap_release_free_page(pmap_t pmap, vm_page_t p); 220static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex); 221static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 222static vm_page_t pmap_page_lookup(vm_object_t object, vm_pindex_t pindex); 223static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t); 224static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 225static void *pmap_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 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 * Initialize the kernel pmap (which is statically allocated). 313 */ 314 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 315 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 316 TAILQ_INIT(&kernel_pmap->pm_pvlist); 317 LIST_INIT(&allpmaps); 318 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 319 nkpt = NKPT; 320 321 /* 322 * Reserve some special page table entries/VA space for temporary 323 * mapping of pages. 324 */ 325#define SYSMAP(c, p, v, n) \ 326 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 327 328 va = virtual_avail; 329 pte = (pt_entry_t *) pmap_pte(kernel_pmap, va); 330 331 /* 332 * CMAP1/CMAP2 are used for zeroing and copying pages. 333 * CMAP3 is used for the idle process page zeroing. 334 */ 335 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 336 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 337 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 338 339 /* 340 * Crashdump maps. 341 */ 342 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 343 344 /* 345 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 346 * XXX ptmmap is not used. 347 */ 348 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 349 350 /* 351 * msgbufp is used to map the system message buffer. 352 * XXX msgbufmap is not used. 353 */ 354 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 355 atop(round_page(MSGBUF_SIZE))) 356 357 /* 358 * ptemap is used for pmap_pte_quick 359 */ 360 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 361 362 virtual_avail = va; 363 364 *CMAP1 = *CMAP2 = 0; 365 for (i = 0; i < NKPT; i++) 366 PTD[i] = 0; 367 368 pgeflag = 0; 369#ifndef DISABLE_PG_G 370 if (cpu_feature & CPUID_PGE) 371 pgeflag = PG_G; 372#endif 373 374/* 375 * Initialize the 4MB page size flag 376 */ 377 pseflag = 0; 378/* 379 * The 4MB page version of the initial 380 * kernel page mapping. 381 */ 382 pdir4mb = 0; 383 384#ifndef DISABLE_PSE 385 if (cpu_feature & CPUID_PSE) { 386 pd_entry_t ptditmp; 387 /* 388 * Note that we have enabled PSE mode 389 */ 390 pseflag = PG_PS; 391 ptditmp = *(PTmap + i386_btop(KERNBASE)); 392 ptditmp &= ~(NBPDR - 1); 393 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag; 394 pdir4mb = ptditmp; 395 } 396#endif 397#ifndef SMP 398 /* 399 * Turn on PGE/PSE. SMP does this later on since the 400 * 4K page tables are required for AP boot (for now). 401 * XXX fixme. 402 */ 403 pmap_set_opt(); 404#endif 405#ifdef SMP 406 if (cpu_apic_address == 0) 407 panic("pmap_bootstrap: no local apic! (non-SMP hardware?)"); 408 409 /* local apic is mapped on last page */ 410 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag | 411 (cpu_apic_address & PG_FRAME)); 412#endif 413 invltlb(); 414} 415 416/* 417 * Enable 4MB page mode for MP startup. Turn on PG_G support. 418 * BSP will run this after all the AP's have started up. 419 */ 420void 421pmap_set_opt(void) 422{ 423 pt_entry_t *pte; 424 vm_offset_t va, endva; 425 426 if (pgeflag && (cpu_feature & CPUID_PGE)) { 427 load_cr4(rcr4() | CR4_PGE); 428 invltlb(); /* Insurance */ 429 } 430#ifndef DISABLE_PSE 431 if (pseflag && (cpu_feature & CPUID_PSE)) { 432 load_cr4(rcr4() | CR4_PSE); 433 invltlb(); /* Insurance */ 434 } 435#endif 436 if (PCPU_GET(cpuid) == 0) { 437#ifndef DISABLE_PSE 438 if (pdir4mb) { 439 kernel_pmap->pm_pdir[KPTDI] = PTD[KPTDI] = pdir4mb; 440 invltlb(); /* Insurance */ 441 } 442#endif 443 if (pgeflag) { 444 /* Turn on PG_G for text, data, bss pages. */ 445 va = (vm_offset_t)btext; 446 endva = KERNBASE + KERNend; 447 while (va < endva) { 448 pte = vtopte(va); 449 if (*pte) 450 *pte |= pgeflag; 451 va += PAGE_SIZE; 452 } 453 invltlb(); /* Insurance */ 454 } 455 /* 456 * We do not need to broadcast the invltlb here, because 457 * each AP does it the moment it is released from the boot 458 * lock. See ap_init(). 459 */ 460 } 461} 462 463void * 464pmap_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 465{ 466 *flags = UMA_SLAB_PRIV; 467 return (void *)kmem_alloc(kernel_map, bytes); 468} 469 470/* 471 * Initialize the pmap module. 472 * Called by vm_init, to initialize any structures that the pmap 473 * system needs to map virtual memory. 474 * pmap_init has been enhanced to support in a fairly consistant 475 * way, discontiguous physical memory. 476 */ 477void 478pmap_init(phys_start, phys_end) 479 vm_offset_t phys_start, phys_end; 480{ 481 int i; 482 int initial_pvs; 483 484 /* 485 * object for kernel page table pages 486 */ 487 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE); 488 489 /* 490 * Allocate memory for random pmap data structures. Includes the 491 * pv_head_table. 492 */ 493 494 for(i = 0; i < vm_page_array_size; i++) { 495 vm_page_t m; 496 497 m = &vm_page_array[i]; 498 TAILQ_INIT(&m->md.pv_list); 499 m->md.pv_list_count = 0; 500 } 501 502 /* 503 * init the pv free list 504 */ 505 initial_pvs = vm_page_array_size; 506 if (initial_pvs < MINPV) 507 initial_pvs = MINPV; 508 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL, 509 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM); 510 uma_zone_set_allocf(pvzone, pmap_allocf); 511 uma_prealloc(pvzone, initial_pvs); 512 513 /* 514 * Now it is safe to enable pv_table recording. 515 */ 516 pmap_initialized = TRUE; 517} 518 519/* 520 * Initialize the address space (zone) for the pv_entries. Set a 521 * high water mark so that the system can recover from excessive 522 * numbers of pv entries. 523 */ 524void 525pmap_init2() 526{ 527 int shpgperproc = PMAP_SHPGPERPROC; 528 529 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 530 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 531 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 532 pv_entry_high_water = 9 * (pv_entry_max / 10); 533 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 534} 535 536 537/*************************************************** 538 * Low level helper routines..... 539 ***************************************************/ 540 541#if defined(PMAP_DIAGNOSTIC) 542 543/* 544 * This code checks for non-writeable/modified pages. 545 * This should be an invalid condition. 546 */ 547static int 548pmap_nw_modified(pt_entry_t ptea) 549{ 550 int pte; 551 552 pte = (int) ptea; 553 554 if ((pte & (PG_M|PG_RW)) == PG_M) 555 return 1; 556 else 557 return 0; 558} 559#endif 560 561 562/* 563 * this routine defines the region(s) of memory that should 564 * not be tested for the modified bit. 565 */ 566static PMAP_INLINE int 567pmap_track_modified(vm_offset_t va) 568{ 569 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 570 return 1; 571 else 572 return 0; 573} 574 575#ifdef I386_CPU 576/* 577 * i386 only has "invalidate everything" and no SMP to worry about. 578 */ 579PMAP_INLINE void 580pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 581{ 582 583 if (pmap == kernel_pmap || pmap->pm_active) 584 invltlb(); 585} 586 587PMAP_INLINE void 588pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 589{ 590 591 if (pmap == kernel_pmap || pmap->pm_active) 592 invltlb(); 593} 594 595PMAP_INLINE void 596pmap_invalidate_all(pmap_t pmap) 597{ 598 599 if (pmap == kernel_pmap || pmap->pm_active) 600 invltlb(); 601} 602#else /* !I386_CPU */ 603#ifdef SMP 604/* 605 * For SMP, these functions have to use the IPI mechanism for coherence. 606 */ 607void 608pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 609{ 610 u_int cpumask; 611 u_int other_cpus; 612 613 critical_enter(); 614 /* 615 * We need to disable interrupt preemption but MUST NOT have 616 * interrupts disabled here. 617 * XXX we may need to hold schedlock to get a coherent pm_active 618 */ 619 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) { 620 invlpg(va); 621 smp_invlpg(va); 622 } else { 623 cpumask = PCPU_GET(cpumask); 624 other_cpus = PCPU_GET(other_cpus); 625 if (pmap->pm_active & cpumask) 626 invlpg(va); 627 if (pmap->pm_active & other_cpus) 628 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 629 } 630 critical_exit(); 631} 632 633void 634pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 635{ 636 u_int cpumask; 637 u_int other_cpus; 638 vm_offset_t addr; 639 640 critical_enter(); 641 /* 642 * We need to disable interrupt preemption but MUST NOT have 643 * interrupts disabled here. 644 * XXX we may need to hold schedlock to get a coherent pm_active 645 */ 646 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) { 647 for (addr = sva; addr < eva; addr += PAGE_SIZE) 648 invlpg(addr); 649 smp_invlpg_range(sva, eva); 650 } else { 651 cpumask = PCPU_GET(cpumask); 652 other_cpus = PCPU_GET(other_cpus); 653 if (pmap->pm_active & cpumask) 654 for (addr = sva; addr < eva; addr += PAGE_SIZE) 655 invlpg(addr); 656 if (pmap->pm_active & other_cpus) 657 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 658 sva, eva); 659 } 660 critical_exit(); 661} 662 663void 664pmap_invalidate_all(pmap_t pmap) 665{ 666 u_int cpumask; 667 u_int other_cpus; 668 669 critical_enter(); 670 /* 671 * We need to disable interrupt preemption but MUST NOT have 672 * interrupts disabled here. 673 * XXX we may need to hold schedlock to get a coherent pm_active 674 */ 675 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) { 676 invltlb(); 677 smp_invltlb(); 678 } else { 679 cpumask = PCPU_GET(cpumask); 680 other_cpus = PCPU_GET(other_cpus); 681 if (pmap->pm_active & cpumask) 682 invltlb(); 683 if (pmap->pm_active & other_cpus) 684 smp_masked_invltlb(pmap->pm_active & other_cpus); 685 } 686 critical_exit(); 687} 688#else /* !SMP */ 689/* 690 * Normal, non-SMP, 486+ invalidation functions. 691 * We inline these within pmap.c for speed. 692 */ 693PMAP_INLINE void 694pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 695{ 696 697 if (pmap == kernel_pmap || pmap->pm_active) 698 invlpg(va); 699} 700 701PMAP_INLINE void 702pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 703{ 704 vm_offset_t addr; 705 706 if (pmap == kernel_pmap || pmap->pm_active) 707 for (addr = sva; addr < eva; addr += PAGE_SIZE) 708 invlpg(addr); 709} 710 711PMAP_INLINE void 712pmap_invalidate_all(pmap_t pmap) 713{ 714 715 if (pmap == kernel_pmap || pmap->pm_active) 716 invltlb(); 717} 718#endif /* !SMP */ 719#endif /* !I386_CPU */ 720 721/* 722 * Return an address which is the base of the Virtual mapping of 723 * all the PTEs for the given pmap. Note this doesn't say that 724 * all the PTEs will be present or that the pages there are valid. 725 * The PTEs are made available by the recursive mapping trick. 726 * It will map in the alternate PTE space if needed. 727 */ 728static pt_entry_t * 729get_ptbase(pmap) 730 pmap_t pmap; 731{ 732 pd_entry_t frame; 733 734 /* are we current address space or kernel? */ 735 if (pmap == kernel_pmap) 736 return PTmap; 737 frame = pmap->pm_pdir[PTDPTDI] & PG_FRAME; 738 if (frame == (PTDpde & PG_FRAME)) 739 return PTmap; 740 /* otherwise, we are alternate address space */ 741 if (frame != (APTDpde & PG_FRAME)) { 742 APTDpde = (pd_entry_t) (frame | PG_RW | PG_V); 743 pmap_invalidate_all(kernel_pmap); /* XXX Bandaid */ 744 } 745 return APTmap; 746} 747 748/* 749 * Super fast pmap_pte routine best used when scanning 750 * the pv lists. This eliminates many coarse-grained 751 * invltlb calls. Note that many of the pv list 752 * scans are across different pmaps. It is very wasteful 753 * to do an entire invltlb for checking a single mapping. 754 */ 755 756static pt_entry_t * 757pmap_pte_quick(pmap, va) 758 register pmap_t pmap; 759 vm_offset_t va; 760{ 761 pd_entry_t pde, newpf; 762 pde = pmap->pm_pdir[va >> PDRSHIFT]; 763 if (pde != 0) { 764 pd_entry_t frame = pmap->pm_pdir[PTDPTDI] & PG_FRAME; 765 unsigned index = i386_btop(va); 766 /* are we current address space or kernel? */ 767 if (pmap == kernel_pmap || frame == (PTDpde & PG_FRAME)) 768 return PTmap + index; 769 newpf = pde & PG_FRAME; 770 if (((*PMAP1) & PG_FRAME) != newpf) { 771 *PMAP1 = newpf | PG_RW | PG_V; 772 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR1); 773 } 774 return PADDR1 + (index & (NPTEPG - 1)); 775 } 776 return (0); 777} 778 779/* 780 * Routine: pmap_extract 781 * Function: 782 * Extract the physical page address associated 783 * with the given map/virtual_address pair. 784 */ 785vm_offset_t 786pmap_extract(pmap, va) 787 register pmap_t pmap; 788 vm_offset_t va; 789{ 790 vm_offset_t rtval; /* XXX FIXME */ 791 vm_offset_t pdirindex; 792 793 if (pmap == 0) 794 return 0; 795 pdirindex = va >> PDRSHIFT; 796 rtval = pmap->pm_pdir[pdirindex]; 797 if (rtval != 0) { 798 pt_entry_t *pte; 799 if ((rtval & PG_PS) != 0) { 800 rtval &= ~(NBPDR - 1); 801 rtval |= va & (NBPDR - 1); 802 return rtval; 803 } 804 pte = get_ptbase(pmap) + i386_btop(va); 805 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK)); 806 return rtval; 807 } 808 return 0; 809 810} 811 812/*************************************************** 813 * Low level mapping routines..... 814 ***************************************************/ 815 816/* 817 * Add a wired page to the kva. 818 * Note: not SMP coherent. 819 */ 820PMAP_INLINE void 821pmap_kenter(vm_offset_t va, vm_offset_t pa) 822{ 823 pt_entry_t *pte; 824 825 pte = vtopte(va); 826 *pte = pa | PG_RW | PG_V | pgeflag; 827} 828 829/* 830 * Remove a page from the kernel pagetables. 831 * Note: not SMP coherent. 832 */ 833PMAP_INLINE void 834pmap_kremove(vm_offset_t va) 835{ 836 pt_entry_t *pte; 837 838 pte = vtopte(va); 839 *pte = 0; 840} 841 842/* 843 * Used to map a range of physical addresses into kernel 844 * virtual address space. 845 * 846 * The value passed in '*virt' is a suggested virtual address for 847 * the mapping. Architectures which can support a direct-mapped 848 * physical to virtual region can return the appropriate address 849 * within that region, leaving '*virt' unchanged. Other 850 * architectures should map the pages starting at '*virt' and 851 * update '*virt' with the first usable address after the mapped 852 * region. 853 */ 854vm_offset_t 855pmap_map(vm_offset_t *virt, vm_offset_t start, vm_offset_t end, int prot) 856{ 857 vm_offset_t va, sva; 858 859 va = sva = *virt; 860 while (start < end) { 861 pmap_kenter(va, start); 862 va += PAGE_SIZE; 863 start += PAGE_SIZE; 864 } 865 pmap_invalidate_range(kernel_pmap, sva, va); 866 *virt = va; 867 return (sva); 868} 869 870 871/* 872 * Add a list of wired pages to the kva 873 * this routine is only used for temporary 874 * kernel mappings that do not need to have 875 * page modification or references recorded. 876 * Note that old mappings are simply written 877 * over. The page *must* be wired. 878 * Note: SMP coherent. Uses a ranged shootdown IPI. 879 */ 880void 881pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 882{ 883 vm_offset_t va; 884 885 va = sva; 886 while (count-- > 0) { 887 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 888 va += PAGE_SIZE; 889 m++; 890 } 891 pmap_invalidate_range(kernel_pmap, sva, va); 892} 893 894/* 895 * This routine tears out page mappings from the 896 * kernel -- it is meant only for temporary mappings. 897 * Note: SMP coherent. Uses a ranged shootdown IPI. 898 */ 899void 900pmap_qremove(vm_offset_t sva, int count) 901{ 902 vm_offset_t va; 903 904 va = sva; 905 while (count-- > 0) { 906 pmap_kremove(va); 907 va += PAGE_SIZE; 908 } 909 pmap_invalidate_range(kernel_pmap, sva, va); 910} 911 912static vm_page_t 913pmap_page_lookup(vm_object_t object, vm_pindex_t pindex) 914{ 915 vm_page_t m; 916 917retry: 918 m = vm_page_lookup(object, pindex); 919 if (m && vm_page_sleep_busy(m, FALSE, "pplookp")) 920 goto retry; 921 return m; 922} 923 924/* 925 * Create the kernel stack (including pcb for i386) for a new thread. 926 * This routine directly affects the fork perf for a process and 927 * create performance for a thread. 928 */ 929void 930pmap_new_thread(struct thread *td) 931{ 932 int i; 933 vm_page_t ma[KSTACK_PAGES]; 934 vm_object_t ksobj; 935 vm_page_t m; 936 vm_offset_t ks; 937 938 /* 939 * allocate object for the kstack 940 */ 941 ksobj = vm_object_allocate(OBJT_DEFAULT, KSTACK_PAGES); 942 td->td_kstack_obj = ksobj; 943 944 /* get a kernel virtual address for the kstack for this thread */ 945#ifdef KSTACK_GUARD 946 ks = kmem_alloc_nofault(kernel_map, (KSTACK_PAGES + 1) * PAGE_SIZE); 947 if (ks == 0) 948 panic("pmap_new_thread: kstack allocation failed"); 949 if (*vtopte(ks) != 0) 950 pmap_qremove(ks, 1); 951 ks += PAGE_SIZE; 952 td->td_kstack = ks; 953#else 954 /* get a kernel virtual address for the kstack for this thread */ 955 ks = kmem_alloc_nofault(kernel_map, KSTACK_PAGES * PAGE_SIZE); 956 if (ks == 0) 957 panic("pmap_new_thread: kstack allocation failed"); 958 td->td_kstack = ks; 959#endif 960 /* 961 * For the length of the stack, link in a real page of ram for each 962 * page of stack. 963 */ 964 for (i = 0; i < KSTACK_PAGES; i++) { 965 /* 966 * Get a kernel stack page 967 */ 968 m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 969 ma[i] = m; 970 971 /* 972 * Wire the page 973 */ 974 m->wire_count++; 975 cnt.v_wire_count++; 976 977 vm_page_wakeup(m); 978 vm_page_flag_clear(m, PG_ZERO); 979 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE); 980 m->valid = VM_PAGE_BITS_ALL; 981 } 982 pmap_qenter(ks, ma, KSTACK_PAGES); 983} 984 985/* 986 * Dispose the kernel stack for a thread that has exited. 987 * This routine directly impacts the exit perf of a process and thread. 988 */ 989void 990pmap_dispose_thread(td) 991 struct thread *td; 992{ 993 int i; 994 vm_object_t ksobj; 995 vm_offset_t ks; 996 vm_page_t m; 997 998 ksobj = td->td_kstack_obj; 999 ks = td->td_kstack; 1000 pmap_qremove(ks, KSTACK_PAGES); 1001 for (i = 0; i < KSTACK_PAGES; i++) { 1002 m = vm_page_lookup(ksobj, i); 1003 if (m == NULL) 1004 panic("pmap_dispose_thread: kstack already missing?"); 1005 vm_page_lock_queues(); 1006 vm_page_busy(m); 1007 vm_page_unwire(m, 0); 1008 vm_page_free(m); 1009 vm_page_unlock_queues(); 1010 } 1011 /* 1012 * Free the space that this stack was mapped to in the kernel 1013 * address map. 1014 */ 1015#ifdef KSTACK_GUARD 1016 kmem_free(kernel_map, ks - PAGE_SIZE, (KSTACK_PAGES + 1) * PAGE_SIZE); 1017#else 1018 kmem_free(kernel_map, ks, KSTACK_PAGES * PAGE_SIZE); 1019#endif 1020 vm_object_deallocate(ksobj); 1021} 1022 1023/* 1024 * Allow the Kernel stack for a thread to be prejudicially paged out. 1025 */ 1026void 1027pmap_swapout_thread(td) 1028 struct thread *td; 1029{ 1030 int i; 1031 vm_object_t ksobj; 1032 vm_offset_t ks; 1033 vm_page_t m; 1034 1035 ksobj = td->td_kstack_obj; 1036 ks = td->td_kstack; 1037 pmap_qremove(ks, KSTACK_PAGES); 1038 for (i = 0; i < KSTACK_PAGES; i++) { 1039 m = vm_page_lookup(ksobj, i); 1040 if (m == NULL) 1041 panic("pmap_swapout_thread: kstack already missing?"); 1042 vm_page_lock_queues(); 1043 vm_page_dirty(m); 1044 vm_page_unwire(m, 0); 1045 vm_page_unlock_queues(); 1046 } 1047} 1048 1049/* 1050 * Bring the kernel stack for a specified thread back in. 1051 */ 1052void 1053pmap_swapin_thread(td) 1054 struct thread *td; 1055{ 1056 int i, rv; 1057 vm_page_t ma[KSTACK_PAGES]; 1058 vm_object_t ksobj; 1059 vm_offset_t ks; 1060 vm_page_t m; 1061 1062 ksobj = td->td_kstack_obj; 1063 ks = td->td_kstack; 1064 for (i = 0; i < KSTACK_PAGES; i++) { 1065 m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1066 if (m->valid != VM_PAGE_BITS_ALL) { 1067 rv = vm_pager_get_pages(ksobj, &m, 1, 0); 1068 if (rv != VM_PAGER_OK) 1069 panic("pmap_swapin_thread: cannot get kstack for proc: %d\n", td->td_proc->p_pid); 1070 m = vm_page_lookup(ksobj, i); 1071 m->valid = VM_PAGE_BITS_ALL; 1072 } 1073 ma[i] = m; 1074 vm_page_wire(m); 1075 vm_page_wakeup(m); 1076 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE); 1077 } 1078 pmap_qenter(ks, ma, KSTACK_PAGES); 1079} 1080 1081/*************************************************** 1082 * Page table page management routines..... 1083 ***************************************************/ 1084 1085/* 1086 * This routine unholds page table pages, and if the hold count 1087 * drops to zero, then it decrements the wire count. 1088 */ 1089static int 1090_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1091{ 1092 1093 while (vm_page_sleep_busy(m, FALSE, "pmuwpt")) 1094 ; 1095 1096 if (m->hold_count == 0) { 1097 vm_offset_t pteva; 1098 /* 1099 * unmap the page table page 1100 */ 1101 pmap->pm_pdir[m->pindex] = 0; 1102 --pmap->pm_stats.resident_count; 1103 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == 1104 (PTDpde & PG_FRAME)) { 1105 /* 1106 * Do a invltlb to make the invalidated mapping 1107 * take effect immediately. 1108 */ 1109 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1110 pmap_invalidate_page(pmap, pteva); 1111 } 1112 1113 if (pmap->pm_ptphint == m) 1114 pmap->pm_ptphint = NULL; 1115 1116 /* 1117 * If the page is finally unwired, simply free it. 1118 */ 1119 --m->wire_count; 1120 if (m->wire_count == 0) { 1121 1122 vm_page_flash(m); 1123 vm_page_busy(m); 1124 vm_page_free_zero(m); 1125 --cnt.v_wire_count; 1126 } 1127 return 1; 1128 } 1129 return 0; 1130} 1131 1132static PMAP_INLINE int 1133pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1134{ 1135 vm_page_unhold(m); 1136 if (m->hold_count == 0) 1137 return _pmap_unwire_pte_hold(pmap, m); 1138 else 1139 return 0; 1140} 1141 1142/* 1143 * After removing a page table entry, this routine is used to 1144 * conditionally free the page, and manage the hold/wire counts. 1145 */ 1146static int 1147pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 1148{ 1149 unsigned ptepindex; 1150 if (va >= VM_MAXUSER_ADDRESS) 1151 return 0; 1152 1153 if (mpte == NULL) { 1154 ptepindex = (va >> PDRSHIFT); 1155 if (pmap->pm_ptphint && 1156 (pmap->pm_ptphint->pindex == ptepindex)) { 1157 mpte = pmap->pm_ptphint; 1158 } else { 1159 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 1160 pmap->pm_ptphint = mpte; 1161 } 1162 } 1163 1164 return pmap_unwire_pte_hold(pmap, mpte); 1165} 1166 1167void 1168pmap_pinit0(pmap) 1169 struct pmap *pmap; 1170{ 1171 pmap->pm_pdir = 1172 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE); 1173 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t)IdlePTD); 1174 invlpg((vm_offset_t)pmap->pm_pdir); 1175 pmap->pm_ptphint = NULL; 1176 pmap->pm_active = 0; 1177 TAILQ_INIT(&pmap->pm_pvlist); 1178 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1179 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1180} 1181 1182/* 1183 * Initialize a preallocated and zeroed pmap structure, 1184 * such as one in a vmspace structure. 1185 */ 1186void 1187pmap_pinit(pmap) 1188 register struct pmap *pmap; 1189{ 1190 vm_page_t ptdpg; 1191 1192 /* 1193 * No need to allocate page table space yet but we do need a valid 1194 * page directory table. 1195 */ 1196 if (pmap->pm_pdir == NULL) 1197 pmap->pm_pdir = 1198 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE); 1199 1200 /* 1201 * allocate object for the ptes 1202 */ 1203 if (pmap->pm_pteobj == NULL) 1204 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1); 1205 1206 /* 1207 * allocate the page directory page 1208 */ 1209 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI, 1210 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1211 1212 ptdpg->wire_count = 1; 1213 ++cnt.v_wire_count; 1214 1215 1216 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY); /* not usually mapped*/ 1217 ptdpg->valid = VM_PAGE_BITS_ALL; 1218 1219 pmap_qenter((vm_offset_t) pmap->pm_pdir, &ptdpg, 1); 1220 if ((ptdpg->flags & PG_ZERO) == 0) 1221 bzero(pmap->pm_pdir, PAGE_SIZE); 1222 1223 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1224 /* Wire in kernel global address entries. */ 1225 /* XXX copies current process, does not fill in MPPTDI */ 1226 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE); 1227#ifdef SMP 1228 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1229#endif 1230 1231 /* install self-referential address mapping entry */ 1232 pmap->pm_pdir[PTDPTDI] = 1233 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M; 1234 1235 pmap->pm_active = 0; 1236 pmap->pm_ptphint = NULL; 1237 TAILQ_INIT(&pmap->pm_pvlist); 1238 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1239} 1240 1241/* 1242 * Wire in kernel global address entries. To avoid a race condition 1243 * between pmap initialization and pmap_growkernel, this procedure 1244 * should be called after the vmspace is attached to the process 1245 * but before this pmap is activated. 1246 */ 1247void 1248pmap_pinit2(pmap) 1249 struct pmap *pmap; 1250{ 1251 /* XXX: Remove this stub when no longer called */ 1252} 1253 1254static int 1255pmap_release_free_page(pmap_t pmap, vm_page_t p) 1256{ 1257 pd_entry_t *pde = pmap->pm_pdir; 1258 /* 1259 * This code optimizes the case of freeing non-busy 1260 * page-table pages. Those pages are zero now, and 1261 * might as well be placed directly into the zero queue. 1262 */ 1263 if (vm_page_sleep_busy(p, FALSE, "pmaprl")) 1264 return 0; 1265 1266 vm_page_busy(p); 1267 1268 /* 1269 * Remove the page table page from the processes address space. 1270 */ 1271 pde[p->pindex] = 0; 1272 pmap->pm_stats.resident_count--; 1273 1274 if (p->hold_count) { 1275 panic("pmap_release: freeing held page table page"); 1276 } 1277 /* 1278 * Page directory pages need to have the kernel 1279 * stuff cleared, so they can go into the zero queue also. 1280 */ 1281 if (p->pindex == PTDPTDI) { 1282 bzero(pde + KPTDI, nkpt * PTESIZE); 1283#ifdef SMP 1284 pde[MPPTDI] = 0; 1285#endif 1286 pde[APTDPTDI] = 0; 1287 pmap_kremove((vm_offset_t) pmap->pm_pdir); 1288 } 1289 1290 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex)) 1291 pmap->pm_ptphint = NULL; 1292 1293 p->wire_count--; 1294 cnt.v_wire_count--; 1295 vm_page_free_zero(p); 1296 return 1; 1297} 1298 1299/* 1300 * this routine is called if the page table page is not 1301 * mapped correctly. 1302 */ 1303static vm_page_t 1304_pmap_allocpte(pmap, ptepindex) 1305 pmap_t pmap; 1306 unsigned ptepindex; 1307{ 1308 vm_offset_t pteva, ptepa; /* XXXPA */ 1309 vm_page_t m; 1310 1311 /* 1312 * Find or fabricate a new pagetable page 1313 */ 1314 m = vm_page_grab(pmap->pm_pteobj, ptepindex, 1315 VM_ALLOC_ZERO | VM_ALLOC_RETRY); 1316 1317 KASSERT(m->queue == PQ_NONE, 1318 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1319 1320 if (m->wire_count == 0) 1321 cnt.v_wire_count++; 1322 m->wire_count++; 1323 1324 /* 1325 * Increment the hold count for the page table page 1326 * (denoting a new mapping.) 1327 */ 1328 m->hold_count++; 1329 1330 /* 1331 * Map the pagetable page into the process address space, if 1332 * it isn't already there. 1333 */ 1334 1335 pmap->pm_stats.resident_count++; 1336 1337 ptepa = VM_PAGE_TO_PHYS(m); 1338 pmap->pm_pdir[ptepindex] = 1339 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1340 1341 /* 1342 * Set the page table hint 1343 */ 1344 pmap->pm_ptphint = m; 1345 1346 /* 1347 * Try to use the new mapping, but if we cannot, then 1348 * do it with the routine that maps the page explicitly. 1349 */ 1350 if ((m->flags & PG_ZERO) == 0) { 1351 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == 1352 (PTDpde & PG_FRAME)) { 1353 pteva = VM_MAXUSER_ADDRESS + i386_ptob(ptepindex); 1354 bzero((caddr_t) pteva, PAGE_SIZE); 1355 } else { 1356 pmap_zero_page(m); 1357 } 1358 } 1359 1360 m->valid = VM_PAGE_BITS_ALL; 1361 vm_page_flag_clear(m, PG_ZERO); 1362 vm_page_flag_set(m, PG_MAPPED); 1363 vm_page_wakeup(m); 1364 1365 return m; 1366} 1367 1368static vm_page_t 1369pmap_allocpte(pmap_t pmap, vm_offset_t va) 1370{ 1371 unsigned ptepindex; 1372 pd_entry_t ptepa; 1373 vm_page_t m; 1374 1375 /* 1376 * Calculate pagetable page index 1377 */ 1378 ptepindex = va >> PDRSHIFT; 1379 1380 /* 1381 * Get the page directory entry 1382 */ 1383 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex]; 1384 1385 /* 1386 * This supports switching from a 4MB page to a 1387 * normal 4K page. 1388 */ 1389 if (ptepa & PG_PS) { 1390 pmap->pm_pdir[ptepindex] = 0; 1391 ptepa = 0; 1392 pmap_invalidate_all(kernel_pmap); 1393 } 1394 1395 /* 1396 * If the page table page is mapped, we just increment the 1397 * hold count, and activate it. 1398 */ 1399 if (ptepa) { 1400 /* 1401 * In order to get the page table page, try the 1402 * hint first. 1403 */ 1404 if (pmap->pm_ptphint && 1405 (pmap->pm_ptphint->pindex == ptepindex)) { 1406 m = pmap->pm_ptphint; 1407 } else { 1408 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 1409 pmap->pm_ptphint = m; 1410 } 1411 m->hold_count++; 1412 return m; 1413 } 1414 /* 1415 * Here if the pte page isn't mapped, or if it has been deallocated. 1416 */ 1417 return _pmap_allocpte(pmap, ptepindex); 1418} 1419 1420 1421/*************************************************** 1422* Pmap allocation/deallocation routines. 1423 ***************************************************/ 1424 1425/* 1426 * Release any resources held by the given physical map. 1427 * Called when a pmap initialized by pmap_pinit is being released. 1428 * Should only be called if the map contains no valid mappings. 1429 */ 1430void 1431pmap_release(pmap_t pmap) 1432{ 1433 vm_page_t p,n,ptdpg; 1434 vm_object_t object = pmap->pm_pteobj; 1435 int curgeneration; 1436 1437#if defined(DIAGNOSTIC) 1438 if (object->ref_count != 1) 1439 panic("pmap_release: pteobj reference count != 1"); 1440#endif 1441 1442 ptdpg = NULL; 1443 LIST_REMOVE(pmap, pm_list); 1444retry: 1445 curgeneration = object->generation; 1446 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = n) { 1447 n = TAILQ_NEXT(p, listq); 1448 if (p->pindex == PTDPTDI) { 1449 ptdpg = p; 1450 continue; 1451 } 1452 while (1) { 1453 if (!pmap_release_free_page(pmap, p) && 1454 (object->generation != curgeneration)) 1455 goto retry; 1456 } 1457 } 1458 1459 if (ptdpg && !pmap_release_free_page(pmap, ptdpg)) 1460 goto retry; 1461} 1462 1463static int 1464kvm_size(SYSCTL_HANDLER_ARGS) 1465{ 1466 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1467 1468 return sysctl_handle_long(oidp, &ksize, 0, req); 1469} 1470SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1471 0, 0, kvm_size, "IU", "Size of KVM"); 1472 1473static int 1474kvm_free(SYSCTL_HANDLER_ARGS) 1475{ 1476 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1477 1478 return sysctl_handle_long(oidp, &kfree, 0, req); 1479} 1480SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1481 0, 0, kvm_free, "IU", "Amount of KVM free"); 1482 1483/* 1484 * grow the number of kernel page table entries, if needed 1485 */ 1486void 1487pmap_growkernel(vm_offset_t addr) 1488{ 1489 struct pmap *pmap; 1490 int s; 1491 vm_offset_t ptppaddr; 1492 vm_page_t nkpg; 1493 pd_entry_t newpdir; 1494 1495 s = splhigh(); 1496 if (kernel_vm_end == 0) { 1497 kernel_vm_end = KERNBASE; 1498 nkpt = 0; 1499 while (pdir_pde(PTD, kernel_vm_end)) { 1500 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1501 nkpt++; 1502 } 1503 } 1504 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1505 while (kernel_vm_end < addr) { 1506 if (pdir_pde(PTD, kernel_vm_end)) { 1507 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1508 continue; 1509 } 1510 1511 /* 1512 * This index is bogus, but out of the way 1513 */ 1514 nkpg = vm_page_alloc(kptobj, nkpt, VM_ALLOC_SYSTEM); 1515 if (!nkpg) 1516 panic("pmap_growkernel: no memory to grow kernel"); 1517 1518 nkpt++; 1519 1520 vm_page_wire(nkpg); 1521 pmap_zero_page(nkpg); 1522 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1523 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1524 pdir_pde(PTD, kernel_vm_end) = newpdir; 1525 1526 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1527 *pmap_pde(pmap, kernel_vm_end) = newpdir; 1528 } 1529 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1530 } 1531 splx(s); 1532} 1533 1534 1535/*************************************************** 1536 * page management routines. 1537 ***************************************************/ 1538 1539/* 1540 * free the pv_entry back to the free list 1541 */ 1542static PMAP_INLINE void 1543free_pv_entry(pv_entry_t pv) 1544{ 1545 pv_entry_count--; 1546 uma_zfree(pvzone, pv); 1547} 1548 1549/* 1550 * get a new pv_entry, allocating a block from the system 1551 * when needed. 1552 * the memory allocation is performed bypassing the malloc code 1553 * because of the possibility of allocations at interrupt time. 1554 */ 1555static pv_entry_t 1556get_pv_entry(void) 1557{ 1558 pv_entry_count++; 1559 if (pv_entry_high_water && 1560 (pv_entry_count > pv_entry_high_water) && 1561 (pmap_pagedaemon_waken == 0)) { 1562 pmap_pagedaemon_waken = 1; 1563 wakeup (&vm_pages_needed); 1564 } 1565 return uma_zalloc(pvzone, M_NOWAIT); 1566} 1567 1568/* 1569 * This routine is very drastic, but can save the system 1570 * in a pinch. 1571 */ 1572void 1573pmap_collect() 1574{ 1575 int i; 1576 vm_page_t m; 1577 static int warningdone = 0; 1578 1579 if (pmap_pagedaemon_waken == 0) 1580 return; 1581 1582 if (warningdone < 5) { 1583 printf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n"); 1584 warningdone++; 1585 } 1586 1587 for(i = 0; i < vm_page_array_size; i++) { 1588 m = &vm_page_array[i]; 1589 if (m->wire_count || m->hold_count || m->busy || 1590 (m->flags & (PG_BUSY | PG_UNMANAGED))) 1591 continue; 1592 pmap_remove_all(m); 1593 } 1594 pmap_pagedaemon_waken = 0; 1595} 1596 1597 1598/* 1599 * If it is the first entry on the list, it is actually 1600 * in the header and we must copy the following entry up 1601 * to the header. Otherwise we must search the list for 1602 * the entry. In either case we free the now unused entry. 1603 */ 1604 1605static int 1606pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1607{ 1608 pv_entry_t pv; 1609 int rtval; 1610 int s; 1611 1612 s = splvm(); 1613 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1614 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1615 if (pmap == pv->pv_pmap && va == pv->pv_va) 1616 break; 1617 } 1618 } else { 1619 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1620 if (va == pv->pv_va) 1621 break; 1622 } 1623 } 1624 1625 rtval = 0; 1626 if (pv) { 1627 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1628 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1629 m->md.pv_list_count--; 1630 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1631 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 1632 1633 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1634 free_pv_entry(pv); 1635 } 1636 1637 splx(s); 1638 return rtval; 1639} 1640 1641/* 1642 * Create a pv entry for page at pa for 1643 * (pmap, va). 1644 */ 1645static void 1646pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1647{ 1648 1649 int s; 1650 pv_entry_t pv; 1651 1652 s = splvm(); 1653 pv = get_pv_entry(); 1654 pv->pv_va = va; 1655 pv->pv_pmap = pmap; 1656 pv->pv_ptem = mpte; 1657 1658 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1659 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1660 m->md.pv_list_count++; 1661 1662 splx(s); 1663} 1664 1665/* 1666 * pmap_remove_pte: do the things to unmap a page in a process 1667 */ 1668static int 1669pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1670{ 1671 pt_entry_t oldpte; 1672 vm_page_t m; 1673 1674 oldpte = atomic_readandclear_int(ptq); 1675 if (oldpte & PG_W) 1676 pmap->pm_stats.wired_count -= 1; 1677 /* 1678 * Machines that don't support invlpg, also don't support 1679 * PG_G. 1680 */ 1681 if (oldpte & PG_G) 1682 pmap_invalidate_page(kernel_pmap, va); 1683 pmap->pm_stats.resident_count -= 1; 1684 if (oldpte & PG_MANAGED) { 1685 m = PHYS_TO_VM_PAGE(oldpte); 1686 if (oldpte & PG_M) { 1687#if defined(PMAP_DIAGNOSTIC) 1688 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1689 printf( 1690 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1691 va, oldpte); 1692 } 1693#endif 1694 if (pmap_track_modified(va)) 1695 vm_page_dirty(m); 1696 } 1697 if (oldpte & PG_A) 1698 vm_page_flag_set(m, PG_REFERENCED); 1699 return pmap_remove_entry(pmap, m, va); 1700 } else { 1701 return pmap_unuse_pt(pmap, va, NULL); 1702 } 1703 1704 return 0; 1705} 1706 1707/* 1708 * Remove a single page from a process address space 1709 */ 1710static void 1711pmap_remove_page(pmap_t pmap, vm_offset_t va) 1712{ 1713 register pt_entry_t *ptq; 1714 1715 /* 1716 * if there is no pte for this address, just skip it!!! 1717 */ 1718 if (*pmap_pde(pmap, va) == 0) { 1719 return; 1720 } 1721 1722 /* 1723 * get a local va for mappings for this pmap. 1724 */ 1725 ptq = get_ptbase(pmap) + i386_btop(va); 1726 if (*ptq) { 1727 (void) pmap_remove_pte(pmap, ptq, va); 1728 pmap_invalidate_page(pmap, va); 1729 } 1730 return; 1731} 1732 1733/* 1734 * Remove the given range of addresses from the specified map. 1735 * 1736 * It is assumed that the start and end are properly 1737 * rounded to the page size. 1738 */ 1739void 1740pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1741{ 1742 register pt_entry_t *ptbase; 1743 vm_offset_t pdnxt; 1744 pd_entry_t ptpaddr; 1745 vm_offset_t sindex, eindex; 1746 int anyvalid; 1747 1748 if (pmap == NULL) 1749 return; 1750 1751 if (pmap->pm_stats.resident_count == 0) 1752 return; 1753 1754 /* 1755 * special handling of removing one page. a very 1756 * common operation and easy to short circuit some 1757 * code. 1758 */ 1759 if ((sva + PAGE_SIZE == eva) && 1760 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1761 pmap_remove_page(pmap, sva); 1762 return; 1763 } 1764 1765 anyvalid = 0; 1766 1767 /* 1768 * Get a local virtual address for the mappings that are being 1769 * worked with. 1770 */ 1771 ptbase = get_ptbase(pmap); 1772 1773 sindex = i386_btop(sva); 1774 eindex = i386_btop(eva); 1775 1776 for (; sindex < eindex; sindex = pdnxt) { 1777 unsigned pdirindex; 1778 1779 /* 1780 * Calculate index for next page table. 1781 */ 1782 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1)); 1783 if (pmap->pm_stats.resident_count == 0) 1784 break; 1785 1786 pdirindex = sindex / NPDEPG; 1787 ptpaddr = pmap->pm_pdir[pdirindex]; 1788 if ((ptpaddr & PG_PS) != 0) { 1789 pmap->pm_pdir[pdirindex] = 0; 1790 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1791 anyvalid++; 1792 continue; 1793 } 1794 1795 /* 1796 * Weed out invalid mappings. Note: we assume that the page 1797 * directory table is always allocated, and in kernel virtual. 1798 */ 1799 if (ptpaddr == 0) 1800 continue; 1801 1802 /* 1803 * Limit our scan to either the end of the va represented 1804 * by the current page table page, or to the end of the 1805 * range being removed. 1806 */ 1807 if (pdnxt > eindex) { 1808 pdnxt = eindex; 1809 } 1810 1811 for (; sindex != pdnxt; sindex++) { 1812 vm_offset_t va; 1813 if (ptbase[sindex] == 0) { 1814 continue; 1815 } 1816 va = i386_ptob(sindex); 1817 1818 anyvalid++; 1819 if (pmap_remove_pte(pmap, 1820 ptbase + sindex, va)) 1821 break; 1822 } 1823 } 1824 1825 if (anyvalid) 1826 pmap_invalidate_all(pmap); 1827} 1828 1829/* 1830 * Routine: pmap_remove_all 1831 * Function: 1832 * Removes this physical page from 1833 * all physical maps in which it resides. 1834 * Reflects back modify bits to the pager. 1835 * 1836 * Notes: 1837 * Original versions of this routine were very 1838 * inefficient because they iteratively called 1839 * pmap_remove (slow...) 1840 */ 1841 1842static void 1843pmap_remove_all(vm_page_t m) 1844{ 1845 register pv_entry_t pv; 1846 pt_entry_t *pte, tpte; 1847 int s; 1848 1849#if defined(PMAP_DIAGNOSTIC) 1850 /* 1851 * XXX this makes pmap_page_protect(NONE) illegal for non-managed 1852 * pages! 1853 */ 1854 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1855 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%x", VM_PAGE_TO_PHYS(m)); 1856 } 1857#endif 1858 1859 s = splvm(); 1860 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1861 pv->pv_pmap->pm_stats.resident_count--; 1862 1863 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1864 1865 tpte = atomic_readandclear_int(pte); 1866 if (tpte & PG_W) 1867 pv->pv_pmap->pm_stats.wired_count--; 1868 1869 if (tpte & PG_A) 1870 vm_page_flag_set(m, PG_REFERENCED); 1871 1872 /* 1873 * Update the vm_page_t clean and reference bits. 1874 */ 1875 if (tpte & PG_M) { 1876#if defined(PMAP_DIAGNOSTIC) 1877 if (pmap_nw_modified((pt_entry_t) tpte)) { 1878 printf( 1879 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1880 pv->pv_va, tpte); 1881 } 1882#endif 1883 if (pmap_track_modified(pv->pv_va)) 1884 vm_page_dirty(m); 1885 } 1886 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1887 1888 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1889 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1890 m->md.pv_list_count--; 1891 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 1892 free_pv_entry(pv); 1893 } 1894 1895 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 1896 1897 splx(s); 1898} 1899 1900/* 1901 * Set the physical protection on the 1902 * specified range of this map as requested. 1903 */ 1904void 1905pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1906{ 1907 register pt_entry_t *ptbase; 1908 vm_offset_t pdnxt; 1909 pd_entry_t ptpaddr; 1910 vm_offset_t sindex, eindex; 1911 int anychanged; 1912 1913 if (pmap == NULL) 1914 return; 1915 1916 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1917 pmap_remove(pmap, sva, eva); 1918 return; 1919 } 1920 1921 if (prot & VM_PROT_WRITE) 1922 return; 1923 1924 anychanged = 0; 1925 1926 ptbase = get_ptbase(pmap); 1927 1928 sindex = i386_btop(sva); 1929 eindex = i386_btop(eva); 1930 1931 for (; sindex < eindex; sindex = pdnxt) { 1932 1933 unsigned pdirindex; 1934 1935 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1)); 1936 1937 pdirindex = sindex / NPDEPG; 1938 ptpaddr = pmap->pm_pdir[pdirindex]; 1939 if ((ptpaddr & PG_PS) != 0) { 1940 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1941 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1942 anychanged++; 1943 continue; 1944 } 1945 1946 /* 1947 * Weed out invalid mappings. Note: we assume that the page 1948 * directory table is always allocated, and in kernel virtual. 1949 */ 1950 if (ptpaddr == 0) 1951 continue; 1952 1953 if (pdnxt > eindex) { 1954 pdnxt = eindex; 1955 } 1956 1957 for (; sindex != pdnxt; sindex++) { 1958 1959 pt_entry_t pbits; 1960 vm_page_t m; 1961 1962 pbits = ptbase[sindex]; 1963 1964 if (pbits & PG_MANAGED) { 1965 m = NULL; 1966 if (pbits & PG_A) { 1967 m = PHYS_TO_VM_PAGE(pbits); 1968 vm_page_flag_set(m, PG_REFERENCED); 1969 pbits &= ~PG_A; 1970 } 1971 if (pbits & PG_M) { 1972 if (pmap_track_modified(i386_ptob(sindex))) { 1973 if (m == NULL) 1974 m = PHYS_TO_VM_PAGE(pbits); 1975 vm_page_dirty(m); 1976 pbits &= ~PG_M; 1977 } 1978 } 1979 } 1980 1981 pbits &= ~PG_RW; 1982 1983 if (pbits != ptbase[sindex]) { 1984 ptbase[sindex] = pbits; 1985 anychanged = 1; 1986 } 1987 } 1988 } 1989 if (anychanged) 1990 pmap_invalidate_all(pmap); 1991} 1992 1993/* 1994 * Insert the given physical page (p) at 1995 * the specified virtual address (v) in the 1996 * target physical map with the protection requested. 1997 * 1998 * If specified, the page will be wired down, meaning 1999 * that the related pte can not be reclaimed. 2000 * 2001 * NB: This is the only routine which MAY NOT lazy-evaluate 2002 * or lose information. That is, this routine must actually 2003 * insert this page into the given map NOW. 2004 */ 2005void 2006pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 2007 boolean_t wired) 2008{ 2009 vm_offset_t pa; 2010 register pt_entry_t *pte; 2011 vm_offset_t opa; 2012 pt_entry_t origpte, newpte; 2013 vm_page_t mpte; 2014 2015 if (pmap == NULL) 2016 return; 2017 2018 va &= PG_FRAME; 2019#ifdef PMAP_DIAGNOSTIC 2020 if (va > VM_MAX_KERNEL_ADDRESS) 2021 panic("pmap_enter: toobig"); 2022 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 2023 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 2024#endif 2025 2026 mpte = NULL; 2027 /* 2028 * In the case that a page table page is not 2029 * resident, we are creating it here. 2030 */ 2031 if (va < VM_MAXUSER_ADDRESS) { 2032 mpte = pmap_allocpte(pmap, va); 2033 } 2034#if 0 && defined(PMAP_DIAGNOSTIC) 2035 else { 2036 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 2037 origpte = *pdeaddr; 2038 if ((origpte & PG_V) == 0) { 2039 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 2040 pmap->pm_pdir[PTDPTDI], origpte, va); 2041 } 2042 } 2043#endif 2044 2045 pte = pmap_pte(pmap, va); 2046 2047 /* 2048 * Page Directory table entry not valid, we need a new PT page 2049 */ 2050 if (pte == NULL) { 2051 panic("pmap_enter: invalid page directory, pdir=%p, va=0x%x\n", 2052 (void *)pmap->pm_pdir[PTDPTDI], va); 2053 } 2054 2055 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 2056 origpte = *(vm_offset_t *)pte; 2057 opa = origpte & PG_FRAME; 2058 2059 if (origpte & PG_PS) 2060 panic("pmap_enter: attempted pmap_enter on 4MB page"); 2061 2062 /* 2063 * Mapping has not changed, must be protection or wiring change. 2064 */ 2065 if (origpte && (opa == pa)) { 2066 /* 2067 * Wiring change, just update stats. We don't worry about 2068 * wiring PT pages as they remain resident as long as there 2069 * are valid mappings in them. Hence, if a user page is wired, 2070 * the PT page will be also. 2071 */ 2072 if (wired && ((origpte & PG_W) == 0)) 2073 pmap->pm_stats.wired_count++; 2074 else if (!wired && (origpte & PG_W)) 2075 pmap->pm_stats.wired_count--; 2076 2077#if defined(PMAP_DIAGNOSTIC) 2078 if (pmap_nw_modified((pt_entry_t) origpte)) { 2079 printf( 2080 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 2081 va, origpte); 2082 } 2083#endif 2084 2085 /* 2086 * Remove extra pte reference 2087 */ 2088 if (mpte) 2089 mpte->hold_count--; 2090 2091 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) { 2092 if ((origpte & PG_RW) == 0) { 2093 *pte |= PG_RW; 2094 pmap_invalidate_page(pmap, va); 2095 } 2096 return; 2097 } 2098 2099 /* 2100 * We might be turning off write access to the page, 2101 * so we go ahead and sense modify status. 2102 */ 2103 if (origpte & PG_MANAGED) { 2104 if ((origpte & PG_M) && pmap_track_modified(va)) { 2105 vm_page_t om; 2106 om = PHYS_TO_VM_PAGE(opa); 2107 vm_page_dirty(om); 2108 } 2109 pa |= PG_MANAGED; 2110 } 2111 goto validate; 2112 } 2113 /* 2114 * Mapping has changed, invalidate old range and fall through to 2115 * handle validating new mapping. 2116 */ 2117 if (opa) { 2118 int err; 2119 err = pmap_remove_pte(pmap, pte, va); 2120 if (err) 2121 panic("pmap_enter: pte vanished, va: 0x%x", va); 2122 } 2123 2124 /* 2125 * Enter on the PV list if part of our managed memory. Note that we 2126 * raise IPL while manipulating pv_table since pmap_enter can be 2127 * called at interrupt time. 2128 */ 2129 if (pmap_initialized && 2130 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2131 pmap_insert_entry(pmap, va, mpte, m); 2132 pa |= PG_MANAGED; 2133 } 2134 2135 /* 2136 * Increment counters 2137 */ 2138 pmap->pm_stats.resident_count++; 2139 if (wired) 2140 pmap->pm_stats.wired_count++; 2141 2142validate: 2143 /* 2144 * Now validate mapping with desired protection/wiring. 2145 */ 2146 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V); 2147 2148 if (wired) 2149 newpte |= PG_W; 2150 if (va < VM_MAXUSER_ADDRESS) 2151 newpte |= PG_U; 2152 if (pmap == kernel_pmap) 2153 newpte |= pgeflag; 2154 2155 /* 2156 * if the mapping or permission bits are different, we need 2157 * to update the pte. 2158 */ 2159 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2160 *pte = newpte | PG_A; 2161 /*if (origpte)*/ { 2162 pmap_invalidate_page(pmap, va); 2163 } 2164 } 2165} 2166 2167/* 2168 * this code makes some *MAJOR* assumptions: 2169 * 1. Current pmap & pmap exists. 2170 * 2. Not wired. 2171 * 3. Read access. 2172 * 4. No page table pages. 2173 * 5. Tlbflush is deferred to calling procedure. 2174 * 6. Page IS managed. 2175 * but is *MUCH* faster than pmap_enter... 2176 */ 2177 2178static vm_page_t 2179pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2180{ 2181 pt_entry_t *pte; 2182 vm_offset_t pa; 2183 2184 /* 2185 * In the case that a page table page is not 2186 * resident, we are creating it here. 2187 */ 2188 if (va < VM_MAXUSER_ADDRESS) { 2189 unsigned ptepindex; 2190 pd_entry_t ptepa; 2191 2192 /* 2193 * Calculate pagetable page index 2194 */ 2195 ptepindex = va >> PDRSHIFT; 2196 if (mpte && (mpte->pindex == ptepindex)) { 2197 mpte->hold_count++; 2198 } else { 2199retry: 2200 /* 2201 * Get the page directory entry 2202 */ 2203 ptepa = pmap->pm_pdir[ptepindex]; 2204 2205 /* 2206 * If the page table page is mapped, we just increment 2207 * the hold count, and activate it. 2208 */ 2209 if (ptepa) { 2210 if (ptepa & PG_PS) 2211 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2212 if (pmap->pm_ptphint && 2213 (pmap->pm_ptphint->pindex == ptepindex)) { 2214 mpte = pmap->pm_ptphint; 2215 } else { 2216 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 2217 pmap->pm_ptphint = mpte; 2218 } 2219 if (mpte == NULL) 2220 goto retry; 2221 mpte->hold_count++; 2222 } else { 2223 mpte = _pmap_allocpte(pmap, ptepindex); 2224 } 2225 } 2226 } else { 2227 mpte = NULL; 2228 } 2229 2230 /* 2231 * This call to vtopte makes the assumption that we are 2232 * entering the page into the current pmap. In order to support 2233 * quick entry into any pmap, one would likely use pmap_pte_quick. 2234 * But that isn't as quick as vtopte. 2235 */ 2236 pte = vtopte(va); 2237 if (*pte) { 2238 if (mpte) 2239 pmap_unwire_pte_hold(pmap, mpte); 2240 return 0; 2241 } 2242 2243 /* 2244 * Enter on the PV list if part of our managed memory. Note that we 2245 * raise IPL while manipulating pv_table since pmap_enter can be 2246 * called at interrupt time. 2247 */ 2248 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2249 pmap_insert_entry(pmap, va, mpte, m); 2250 2251 /* 2252 * Increment counters 2253 */ 2254 pmap->pm_stats.resident_count++; 2255 2256 pa = VM_PAGE_TO_PHYS(m); 2257 2258 /* 2259 * Now validate mapping with RO protection 2260 */ 2261 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2262 *pte = pa | PG_V | PG_U; 2263 else 2264 *pte = pa | PG_V | PG_U | PG_MANAGED; 2265 2266 return mpte; 2267} 2268 2269/* 2270 * Make a temporary mapping for a physical address. This is only intended 2271 * to be used for panic dumps. 2272 */ 2273void * 2274pmap_kenter_temporary(vm_offset_t pa, int i) 2275{ 2276 vm_offset_t va; 2277 2278 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2279 pmap_kenter(va, pa); 2280 invlpg(va); 2281 return ((void *)crashdumpmap); 2282} 2283 2284#define MAX_INIT_PT (96) 2285/* 2286 * pmap_object_init_pt preloads the ptes for a given object 2287 * into the specified pmap. This eliminates the blast of soft 2288 * faults on process startup and immediately after an mmap. 2289 */ 2290void 2291pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2292 vm_object_t object, vm_pindex_t pindex, 2293 vm_size_t size, int limit) 2294{ 2295 vm_offset_t tmpidx; 2296 int psize; 2297 vm_page_t p, mpte; 2298 int objpgs; 2299 2300 if (pmap == NULL || object == NULL) 2301 return; 2302 2303 /* 2304 * This code maps large physical mmap regions into the 2305 * processor address space. Note that some shortcuts 2306 * are taken, but the code works. 2307 */ 2308 if (pseflag && (object->type == OBJT_DEVICE) && 2309 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2310 int i; 2311 vm_page_t m[1]; 2312 unsigned int ptepindex; 2313 int npdes; 2314 pd_entry_t ptepa; 2315 2316 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2317 return; 2318 2319retry: 2320 p = vm_page_lookup(object, pindex); 2321 if (p && vm_page_sleep_busy(p, FALSE, "init4p")) 2322 goto retry; 2323 2324 if (p == NULL) { 2325 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2326 if (p == NULL) 2327 return; 2328 m[0] = p; 2329 2330 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2331 vm_page_free(p); 2332 return; 2333 } 2334 2335 p = vm_page_lookup(object, pindex); 2336 vm_page_wakeup(p); 2337 } 2338 2339 ptepa = VM_PAGE_TO_PHYS(p); 2340 if (ptepa & (NBPDR - 1)) { 2341 return; 2342 } 2343 2344 p->valid = VM_PAGE_BITS_ALL; 2345 2346 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2347 npdes = size >> PDRSHIFT; 2348 for(i = 0; i < npdes; i++) { 2349 pmap->pm_pdir[ptepindex] = 2350 ptepa | PG_U | PG_RW | PG_V | PG_PS; 2351 ptepa += NBPDR; 2352 ptepindex += 1; 2353 } 2354 vm_page_flag_set(p, PG_MAPPED); 2355 pmap_invalidate_all(kernel_pmap); 2356 return; 2357 } 2358 2359 psize = i386_btop(size); 2360 2361 if ((object->type != OBJT_VNODE) || 2362 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) && 2363 (object->resident_page_count > MAX_INIT_PT))) { 2364 return; 2365 } 2366 2367 if (psize + pindex > object->size) { 2368 if (object->size < pindex) 2369 return; 2370 psize = object->size - pindex; 2371 } 2372 2373 mpte = NULL; 2374 /* 2375 * if we are processing a major portion of the object, then scan the 2376 * entire thing. 2377 */ 2378 if (psize > (object->resident_page_count >> 2)) { 2379 objpgs = psize; 2380 2381 for (p = TAILQ_FIRST(&object->memq); 2382 ((objpgs > 0) && (p != NULL)); 2383 p = TAILQ_NEXT(p, listq)) { 2384 2385 if (p->pindex < pindex || p->pindex - pindex >= psize) { 2386 continue; 2387 } 2388 tmpidx = p->pindex - pindex; 2389 /* 2390 * don't allow an madvise to blow away our really 2391 * free pages allocating pv entries. 2392 */ 2393 if ((limit & MAP_PREFAULT_MADVISE) && 2394 cnt.v_free_count < cnt.v_free_reserved) { 2395 break; 2396 } 2397 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2398 (p->busy == 0) && 2399 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2400 if ((p->queue - p->pc) == PQ_CACHE) 2401 vm_page_deactivate(p); 2402 vm_page_busy(p); 2403 mpte = pmap_enter_quick(pmap, 2404 addr + i386_ptob(tmpidx), p, mpte); 2405 vm_page_flag_set(p, PG_MAPPED); 2406 vm_page_wakeup(p); 2407 } 2408 objpgs -= 1; 2409 } 2410 } else { 2411 /* 2412 * else lookup the pages one-by-one. 2413 */ 2414 for (tmpidx = 0; tmpidx < psize; tmpidx += 1) { 2415 /* 2416 * don't allow an madvise to blow away our really 2417 * free pages allocating pv entries. 2418 */ 2419 if ((limit & MAP_PREFAULT_MADVISE) && 2420 cnt.v_free_count < cnt.v_free_reserved) { 2421 break; 2422 } 2423 p = vm_page_lookup(object, tmpidx + pindex); 2424 if (p && 2425 ((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2426 (p->busy == 0) && 2427 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2428 if ((p->queue - p->pc) == PQ_CACHE) 2429 vm_page_deactivate(p); 2430 vm_page_busy(p); 2431 mpte = pmap_enter_quick(pmap, 2432 addr + i386_ptob(tmpidx), p, mpte); 2433 vm_page_flag_set(p, PG_MAPPED); 2434 vm_page_wakeup(p); 2435 } 2436 } 2437 } 2438 return; 2439} 2440 2441/* 2442 * pmap_prefault provides a quick way of clustering 2443 * pagefaults into a processes address space. It is a "cousin" 2444 * of pmap_object_init_pt, except it runs at page fault time instead 2445 * of mmap time. 2446 */ 2447#define PFBAK 4 2448#define PFFOR 4 2449#define PAGEORDER_SIZE (PFBAK+PFFOR) 2450 2451static int pmap_prefault_pageorder[] = { 2452 -PAGE_SIZE, PAGE_SIZE, 2453 -2 * PAGE_SIZE, 2 * PAGE_SIZE, 2454 -3 * PAGE_SIZE, 3 * PAGE_SIZE 2455 -4 * PAGE_SIZE, 4 * PAGE_SIZE 2456}; 2457 2458void 2459pmap_prefault(pmap, addra, entry) 2460 pmap_t pmap; 2461 vm_offset_t addra; 2462 vm_map_entry_t entry; 2463{ 2464 int i; 2465 vm_offset_t starta; 2466 vm_offset_t addr; 2467 vm_pindex_t pindex; 2468 vm_page_t m, mpte; 2469 vm_object_t object; 2470 2471 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) 2472 return; 2473 2474 object = entry->object.vm_object; 2475 2476 starta = addra - PFBAK * PAGE_SIZE; 2477 if (starta < entry->start) { 2478 starta = entry->start; 2479 } else if (starta > addra) { 2480 starta = 0; 2481 } 2482 2483 mpte = NULL; 2484 for (i = 0; i < PAGEORDER_SIZE; i++) { 2485 vm_object_t lobject; 2486 pt_entry_t *pte; 2487 2488 addr = addra + pmap_prefault_pageorder[i]; 2489 if (addr > addra + (PFFOR * PAGE_SIZE)) 2490 addr = 0; 2491 2492 if (addr < starta || addr >= entry->end) 2493 continue; 2494 2495 if ((*pmap_pde(pmap, addr)) == NULL) 2496 continue; 2497 2498 pte = vtopte(addr); 2499 if (*pte) 2500 continue; 2501 2502 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2503 lobject = object; 2504 for (m = vm_page_lookup(lobject, pindex); 2505 (!m && (lobject->type == OBJT_DEFAULT) && (lobject->backing_object)); 2506 lobject = lobject->backing_object) { 2507 if (lobject->backing_object_offset & PAGE_MASK) 2508 break; 2509 pindex += (lobject->backing_object_offset >> PAGE_SHIFT); 2510 m = vm_page_lookup(lobject->backing_object, pindex); 2511 } 2512 2513 /* 2514 * give-up when a page is not in memory 2515 */ 2516 if (m == NULL) 2517 break; 2518 2519 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2520 (m->busy == 0) && 2521 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2522 2523 if ((m->queue - m->pc) == PQ_CACHE) { 2524 vm_page_deactivate(m); 2525 } 2526 vm_page_busy(m); 2527 mpte = pmap_enter_quick(pmap, addr, m, mpte); 2528 vm_page_flag_set(m, PG_MAPPED); 2529 vm_page_wakeup(m); 2530 } 2531 } 2532} 2533 2534/* 2535 * Routine: pmap_change_wiring 2536 * Function: Change the wiring attribute for a map/virtual-address 2537 * pair. 2538 * In/out conditions: 2539 * The mapping must already exist in the pmap. 2540 */ 2541void 2542pmap_change_wiring(pmap, va, wired) 2543 register pmap_t pmap; 2544 vm_offset_t va; 2545 boolean_t wired; 2546{ 2547 register pt_entry_t *pte; 2548 2549 if (pmap == NULL) 2550 return; 2551 2552 pte = pmap_pte(pmap, va); 2553 2554 if (wired && !pmap_pte_w(pte)) 2555 pmap->pm_stats.wired_count++; 2556 else if (!wired && pmap_pte_w(pte)) 2557 pmap->pm_stats.wired_count--; 2558 2559 /* 2560 * Wiring is not a hardware characteristic so there is no need to 2561 * invalidate TLB. 2562 */ 2563 pmap_pte_set_w(pte, wired); 2564} 2565 2566 2567 2568/* 2569 * Copy the range specified by src_addr/len 2570 * from the source map to the range dst_addr/len 2571 * in the destination map. 2572 * 2573 * This routine is only advisory and need not do anything. 2574 */ 2575 2576void 2577pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2578 vm_offset_t src_addr) 2579{ 2580 vm_offset_t addr; 2581 vm_offset_t end_addr = src_addr + len; 2582 vm_offset_t pdnxt; 2583 pd_entry_t src_frame, dst_frame; 2584 vm_page_t m; 2585 2586 if (dst_addr != src_addr) 2587 return; 2588 2589 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME; 2590 if (src_frame != (PTDpde & PG_FRAME)) 2591 return; 2592 2593 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME; 2594 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2595 pt_entry_t *src_pte, *dst_pte; 2596 vm_page_t dstmpte, srcmpte; 2597 pd_entry_t srcptepaddr; 2598 unsigned ptepindex; 2599 2600 if (addr >= UPT_MIN_ADDRESS) 2601 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2602 2603 /* 2604 * Don't let optional prefaulting of pages make us go 2605 * way below the low water mark of free pages or way 2606 * above high water mark of used pv entries. 2607 */ 2608 if (cnt.v_free_count < cnt.v_free_reserved || 2609 pv_entry_count > pv_entry_high_water) 2610 break; 2611 2612 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1)); 2613 ptepindex = addr >> PDRSHIFT; 2614 2615 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2616 if (srcptepaddr == 0) 2617 continue; 2618 2619 if (srcptepaddr & PG_PS) { 2620 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2621 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2622 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; 2623 } 2624 continue; 2625 } 2626 2627 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex); 2628 if ((srcmpte == NULL) || 2629 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY)) 2630 continue; 2631 2632 if (pdnxt > end_addr) 2633 pdnxt = end_addr; 2634 2635 /* 2636 * Have to recheck this before every avtopte() call below 2637 * in case we have blocked and something else used APTDpde. 2638 */ 2639 if (dst_frame != (APTDpde & PG_FRAME)) { 2640 APTDpde = dst_frame | PG_RW | PG_V; 2641 pmap_invalidate_all(kernel_pmap); /* XXX Bandaid */ 2642 } 2643 src_pte = vtopte(addr); 2644 dst_pte = avtopte(addr); 2645 while (addr < pdnxt) { 2646 pt_entry_t ptetemp; 2647 ptetemp = *src_pte; 2648 /* 2649 * we only virtual copy managed pages 2650 */ 2651 if ((ptetemp & PG_MANAGED) != 0) { 2652 /* 2653 * We have to check after allocpte for the 2654 * pte still being around... allocpte can 2655 * block. 2656 */ 2657 dstmpte = pmap_allocpte(dst_pmap, addr); 2658 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2659 /* 2660 * Clear the modified and 2661 * accessed (referenced) bits 2662 * during the copy. 2663 */ 2664 m = PHYS_TO_VM_PAGE(ptetemp); 2665 *dst_pte = ptetemp & ~(PG_M | PG_A); 2666 dst_pmap->pm_stats.resident_count++; 2667 pmap_insert_entry(dst_pmap, addr, 2668 dstmpte, m); 2669 } else { 2670 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2671 } 2672 if (dstmpte->hold_count >= srcmpte->hold_count) 2673 break; 2674 } 2675 addr += PAGE_SIZE; 2676 src_pte++; 2677 dst_pte++; 2678 } 2679 } 2680} 2681 2682#ifdef SMP 2683 2684/* 2685 * pmap_zpi_switchin*() 2686 * 2687 * These functions allow us to avoid doing IPIs alltogether in certain 2688 * temporary page-mapping situations (page zeroing). Instead to deal 2689 * with being preempted and moved onto a different cpu we invalidate 2690 * the page when the scheduler switches us in. This does not occur 2691 * very often so we remain relatively optimal with very little effort. 2692 */ 2693static void 2694pmap_zpi_switchin12(void) 2695{ 2696 invlpg((u_int)CADDR1); 2697 invlpg((u_int)CADDR2); 2698} 2699 2700static void 2701pmap_zpi_switchin2(void) 2702{ 2703 invlpg((u_int)CADDR2); 2704} 2705 2706static void 2707pmap_zpi_switchin3(void) 2708{ 2709 invlpg((u_int)CADDR3); 2710} 2711 2712#endif 2713 2714/* 2715 * pmap_zero_page zeros the specified hardware page by mapping 2716 * the page into KVM and using bzero to clear its contents. 2717 */ 2718void 2719pmap_zero_page(vm_page_t m) 2720{ 2721 vm_offset_t phys; 2722 2723 phys = VM_PAGE_TO_PHYS(m); 2724 if (*CMAP2) 2725 panic("pmap_zero_page: CMAP2 busy"); 2726 *CMAP2 = PG_V | PG_RW | phys | PG_A | PG_M; 2727#ifdef I386_CPU 2728 invltlb(); 2729#else 2730#ifdef SMP 2731 curthread->td_switchin = pmap_zpi_switchin2; 2732#endif 2733 invlpg((u_int)CADDR2); 2734#endif 2735#if defined(I686_CPU) 2736 if (cpu_class == CPUCLASS_686) 2737 i686_pagezero(CADDR2); 2738 else 2739#endif 2740 bzero(CADDR2, PAGE_SIZE); 2741#ifdef SMP 2742 curthread->td_switchin = NULL; 2743#endif 2744 *CMAP2 = 0; 2745} 2746 2747/* 2748 * pmap_zero_page_area zeros the specified hardware page by mapping 2749 * the page into KVM and using bzero to clear its contents. 2750 * 2751 * off and size may not cover an area beyond a single hardware page. 2752 */ 2753void 2754pmap_zero_page_area(vm_page_t m, int off, int size) 2755{ 2756 vm_offset_t phys; 2757 2758 phys = VM_PAGE_TO_PHYS(m); 2759 if (*CMAP2) 2760 panic("pmap_zero_page: CMAP2 busy"); 2761 *CMAP2 = PG_V | PG_RW | phys | PG_A | PG_M; 2762#ifdef I386_CPU 2763 invltlb(); 2764#else 2765#ifdef SMP 2766 curthread->td_switchin = pmap_zpi_switchin2; 2767#endif 2768 invlpg((u_int)CADDR2); 2769#endif 2770#if defined(I686_CPU) 2771 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE) 2772 i686_pagezero(CADDR2); 2773 else 2774#endif 2775 bzero((char *)CADDR2 + off, size); 2776#ifdef SMP 2777 curthread->td_switchin = NULL; 2778#endif 2779 *CMAP2 = 0; 2780} 2781 2782/* 2783 * pmap_zero_page_idle zeros the specified hardware page by mapping 2784 * the page into KVM and using bzero to clear its contents. This 2785 * is intended to be called from the vm_pagezero process only and 2786 * outside of Giant. 2787 */ 2788void 2789pmap_zero_page_idle(vm_page_t m) 2790{ 2791 vm_offset_t phys; 2792 2793 phys = VM_PAGE_TO_PHYS(m); 2794 if (*CMAP3) 2795 panic("pmap_zero_page: CMAP3 busy"); 2796 *CMAP3 = PG_V | PG_RW | phys | PG_A | PG_M; 2797#ifdef I386_CPU 2798 invltlb(); 2799#else 2800#ifdef SMP 2801 curthread->td_switchin = pmap_zpi_switchin3; 2802#endif 2803 invlpg((u_int)CADDR3); 2804#endif 2805#if defined(I686_CPU) 2806 if (cpu_class == CPUCLASS_686) 2807 i686_pagezero(CADDR3); 2808 else 2809#endif 2810 bzero(CADDR3, PAGE_SIZE); 2811#ifdef SMP 2812 curthread->td_switchin = NULL; 2813#endif 2814 *CMAP3 = 0; 2815} 2816 2817/* 2818 * pmap_copy_page copies the specified (machine independent) 2819 * page by mapping the page into virtual memory and using 2820 * bcopy to copy the page, one machine dependent page at a 2821 * time. 2822 */ 2823void 2824pmap_copy_page(vm_page_t src, vm_page_t dst) 2825{ 2826 2827 if (*CMAP1) 2828 panic("pmap_copy_page: CMAP1 busy"); 2829 if (*CMAP2) 2830 panic("pmap_copy_page: CMAP2 busy"); 2831 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2832 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2833#ifdef I386_CPU 2834 invltlb(); 2835#else 2836#ifdef SMP 2837 curthread->td_switchin = pmap_zpi_switchin12; 2838#endif 2839 invlpg((u_int)CADDR1); 2840 invlpg((u_int)CADDR2); 2841#endif 2842 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2843#ifdef SMP 2844 curthread->td_switchin = NULL; 2845#endif 2846 *CMAP1 = 0; 2847 *CMAP2 = 0; 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 * Returns true if the pmap's pv is one of the first 2875 * 16 pvs linked to from this page. This count may 2876 * be changed upwards or downwards in the future; it 2877 * is only necessary that true be returned for a small 2878 * subset of pmaps for proper page aging. 2879 */ 2880boolean_t 2881pmap_page_exists_quick(pmap, m) 2882 pmap_t pmap; 2883 vm_page_t m; 2884{ 2885 pv_entry_t pv; 2886 int loops = 0; 2887 int s; 2888 2889 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2890 return FALSE; 2891 2892 s = splvm(); 2893 2894 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2895 if (pv->pv_pmap == pmap) { 2896 splx(s); 2897 return TRUE; 2898 } 2899 loops++; 2900 if (loops >= 16) 2901 break; 2902 } 2903 splx(s); 2904 return (FALSE); 2905} 2906 2907#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2908/* 2909 * Remove all pages from specified address space 2910 * this aids process exit speeds. Also, this code 2911 * is special cased for current process only, but 2912 * can have the more generic (and slightly slower) 2913 * mode enabled. This is much faster than pmap_remove 2914 * in the case of running down an entire address space. 2915 */ 2916void 2917pmap_remove_pages(pmap, sva, eva) 2918 pmap_t pmap; 2919 vm_offset_t sva, eva; 2920{ 2921 pt_entry_t *pte, tpte; 2922 vm_page_t m; 2923 pv_entry_t pv, npv; 2924 int s; 2925 2926#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2927 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2928 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2929 return; 2930 } 2931#endif 2932 2933 s = splvm(); 2934 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2935 2936 if (pv->pv_va >= eva || pv->pv_va < sva) { 2937 npv = TAILQ_NEXT(pv, pv_plist); 2938 continue; 2939 } 2940 2941#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2942 pte = vtopte(pv->pv_va); 2943#else 2944 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2945#endif 2946 tpte = *pte; 2947 2948 if (tpte == 0) { 2949 printf("TPTE at %p IS ZERO @ VA %08x\n", 2950 pte, pv->pv_va); 2951 panic("bad pte"); 2952 } 2953 2954/* 2955 * We cannot remove wired pages from a process' mapping at this time 2956 */ 2957 if (tpte & PG_W) { 2958 npv = TAILQ_NEXT(pv, pv_plist); 2959 continue; 2960 } 2961 2962 m = PHYS_TO_VM_PAGE(tpte); 2963 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2964 ("vm_page_t %p phys_addr mismatch %08x %08x", 2965 m, m->phys_addr, tpte)); 2966 2967 KASSERT(m < &vm_page_array[vm_page_array_size], 2968 ("pmap_remove_pages: bad tpte %x", tpte)); 2969 2970 pv->pv_pmap->pm_stats.resident_count--; 2971 2972 *pte = 0; 2973 2974 /* 2975 * Update the vm_page_t clean and reference bits. 2976 */ 2977 if (tpte & PG_M) { 2978 vm_page_dirty(m); 2979 } 2980 2981 npv = TAILQ_NEXT(pv, pv_plist); 2982 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2983 2984 m->md.pv_list_count--; 2985 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2986 if (TAILQ_FIRST(&m->md.pv_list) == NULL) { 2987 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 2988 } 2989 2990 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2991 free_pv_entry(pv); 2992 } 2993 splx(s); 2994 pmap_invalidate_all(pmap); 2995} 2996 2997/* 2998 * pmap_testbit tests bits in pte's 2999 * note that the testbit/changebit routines are inline, 3000 * and a lot of things compile-time evaluate. 3001 */ 3002static boolean_t 3003pmap_testbit(m, bit) 3004 vm_page_t m; 3005 int bit; 3006{ 3007 pv_entry_t pv; 3008 pt_entry_t *pte; 3009 int s; 3010 3011 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3012 return FALSE; 3013 3014 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 3015 return FALSE; 3016 3017 s = splvm(); 3018 3019 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3020 /* 3021 * if the bit being tested is the modified bit, then 3022 * mark clean_map and ptes as never 3023 * modified. 3024 */ 3025 if (bit & (PG_A|PG_M)) { 3026 if (!pmap_track_modified(pv->pv_va)) 3027 continue; 3028 } 3029 3030#if defined(PMAP_DIAGNOSTIC) 3031 if (!pv->pv_pmap) { 3032 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 3033 continue; 3034 } 3035#endif 3036 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3037 if (*pte & bit) { 3038 splx(s); 3039 return TRUE; 3040 } 3041 } 3042 splx(s); 3043 return (FALSE); 3044} 3045 3046/* 3047 * this routine is used to modify bits in ptes 3048 */ 3049static __inline void 3050pmap_changebit(vm_page_t m, int bit, boolean_t setem) 3051{ 3052 register pv_entry_t pv; 3053 register pt_entry_t *pte; 3054 int s; 3055 3056 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3057 return; 3058 3059 s = splvm(); 3060 3061 /* 3062 * Loop over all current mappings setting/clearing as appropos If 3063 * setting RO do we need to clear the VAC? 3064 */ 3065 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3066 /* 3067 * don't write protect pager mappings 3068 */ 3069 if (!setem && (bit == PG_RW)) { 3070 if (!pmap_track_modified(pv->pv_va)) 3071 continue; 3072 } 3073 3074#if defined(PMAP_DIAGNOSTIC) 3075 if (!pv->pv_pmap) { 3076 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 3077 continue; 3078 } 3079#endif 3080 3081 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3082 3083 if (setem) { 3084 *pte |= bit; 3085 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3086 } else { 3087 pt_entry_t pbits = *pte; 3088 if (pbits & bit) { 3089 if (bit == PG_RW) { 3090 if (pbits & PG_M) { 3091 vm_page_dirty(m); 3092 } 3093 *pte = pbits & ~(PG_M|PG_RW); 3094 } else { 3095 *pte = pbits & ~bit; 3096 } 3097 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3098 } 3099 } 3100 } 3101 splx(s); 3102} 3103 3104/* 3105 * pmap_page_protect: 3106 * 3107 * Lower the permission for all mappings to a given page. 3108 */ 3109void 3110pmap_page_protect(vm_page_t m, vm_prot_t prot) 3111{ 3112 if ((prot & VM_PROT_WRITE) == 0) { 3113 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 3114 pmap_changebit(m, PG_RW, FALSE); 3115 } else { 3116 pmap_remove_all(m); 3117 } 3118 } 3119} 3120 3121vm_offset_t 3122pmap_phys_address(ppn) 3123 int ppn; 3124{ 3125 return (i386_ptob(ppn)); 3126} 3127 3128/* 3129 * pmap_ts_referenced: 3130 * 3131 * Return a count of reference bits for a page, clearing those bits. 3132 * It is not necessary for every reference bit to be cleared, but it 3133 * is necessary that 0 only be returned when there are truly no 3134 * reference bits set. 3135 * 3136 * XXX: The exact number of bits to check and clear is a matter that 3137 * should be tested and standardized at some point in the future for 3138 * optimal aging of shared pages. 3139 */ 3140int 3141pmap_ts_referenced(vm_page_t m) 3142{ 3143 register pv_entry_t pv, pvf, pvn; 3144 pt_entry_t *pte; 3145 int s; 3146 int rtval = 0; 3147 3148 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3149 return (rtval); 3150 3151 s = splvm(); 3152 3153 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 3154 3155 pvf = pv; 3156 3157 do { 3158 pvn = TAILQ_NEXT(pv, pv_list); 3159 3160 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 3161 3162 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 3163 3164 if (!pmap_track_modified(pv->pv_va)) 3165 continue; 3166 3167 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3168 3169 if (pte && (*pte & PG_A)) { 3170 *pte &= ~PG_A; 3171 3172 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3173 3174 rtval++; 3175 if (rtval > 4) { 3176 break; 3177 } 3178 } 3179 } while ((pv = pvn) != NULL && pv != pvf); 3180 } 3181 splx(s); 3182 3183 return (rtval); 3184} 3185 3186/* 3187 * pmap_is_modified: 3188 * 3189 * Return whether or not the specified physical page was modified 3190 * in any physical maps. 3191 */ 3192boolean_t 3193pmap_is_modified(vm_page_t m) 3194{ 3195 return pmap_testbit(m, PG_M); 3196} 3197 3198/* 3199 * Clear the modify bits on the specified physical page. 3200 */ 3201void 3202pmap_clear_modify(vm_page_t m) 3203{ 3204 pmap_changebit(m, PG_M, FALSE); 3205} 3206 3207/* 3208 * pmap_clear_reference: 3209 * 3210 * Clear the reference bit on the specified physical page. 3211 */ 3212void 3213pmap_clear_reference(vm_page_t m) 3214{ 3215 pmap_changebit(m, PG_A, FALSE); 3216} 3217 3218/* 3219 * Miscellaneous support routines follow 3220 */ 3221 3222static void 3223i386_protection_init() 3224{ 3225 register int *kp, prot; 3226 3227 kp = protection_codes; 3228 for (prot = 0; prot < 8; prot++) { 3229 switch (prot) { 3230 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE: 3231 /* 3232 * Read access is also 0. There isn't any execute bit, 3233 * so just make it readable. 3234 */ 3235 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE: 3236 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE: 3237 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE: 3238 *kp++ = 0; 3239 break; 3240 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE: 3241 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE: 3242 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE: 3243 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE: 3244 *kp++ = PG_RW; 3245 break; 3246 } 3247 } 3248} 3249 3250/* 3251 * Map a set of physical memory pages into the kernel virtual 3252 * address space. Return a pointer to where it is mapped. This 3253 * routine is intended to be used for mapping device memory, 3254 * NOT real memory. 3255 */ 3256void * 3257pmap_mapdev(pa, size) 3258 vm_offset_t pa; 3259 vm_size_t size; 3260{ 3261 vm_offset_t va, tmpva, offset; 3262 pt_entry_t *pte; 3263 3264 offset = pa & PAGE_MASK; 3265 size = roundup(offset + size, PAGE_SIZE); 3266 3267 GIANT_REQUIRED; 3268 3269 va = kmem_alloc_pageable(kernel_map, size); 3270 if (!va) 3271 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 3272 3273 pa = pa & PG_FRAME; 3274 for (tmpva = va; size > 0; ) { 3275 pte = vtopte(tmpva); 3276 *pte = pa | PG_RW | PG_V | pgeflag; 3277 size -= PAGE_SIZE; 3278 tmpva += PAGE_SIZE; 3279 pa += PAGE_SIZE; 3280 } 3281 pmap_invalidate_range(kernel_pmap, va, tmpva); 3282 return ((void *)(va + offset)); 3283} 3284 3285void 3286pmap_unmapdev(va, size) 3287 vm_offset_t va; 3288 vm_size_t size; 3289{ 3290 vm_offset_t base, offset, tmpva; 3291 pt_entry_t *pte; 3292 3293 base = va & PG_FRAME; 3294 offset = va & PAGE_MASK; 3295 size = roundup(offset + size, PAGE_SIZE); 3296 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) { 3297 pte = vtopte(tmpva); 3298 *pte = 0; 3299 } 3300 pmap_invalidate_range(kernel_pmap, va, tmpva); 3301 kmem_free(kernel_map, base, size); 3302} 3303 3304/* 3305 * perform the pmap work for mincore 3306 */ 3307int 3308pmap_mincore(pmap, addr) 3309 pmap_t pmap; 3310 vm_offset_t addr; 3311{ 3312 pt_entry_t *ptep, pte; 3313 vm_page_t m; 3314 int val = 0; 3315 3316 ptep = pmap_pte(pmap, addr); 3317 if (ptep == 0) { 3318 return 0; 3319 } 3320 3321 if ((pte = *ptep) != 0) { 3322 vm_offset_t pa; 3323 3324 val = MINCORE_INCORE; 3325 if ((pte & PG_MANAGED) == 0) 3326 return val; 3327 3328 pa = pte & PG_FRAME; 3329 3330 m = PHYS_TO_VM_PAGE(pa); 3331 3332 /* 3333 * Modified by us 3334 */ 3335 if (pte & PG_M) 3336 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 3337 /* 3338 * Modified by someone 3339 */ 3340 else if (m->dirty || pmap_is_modified(m)) 3341 val |= MINCORE_MODIFIED_OTHER; 3342 /* 3343 * Referenced by us 3344 */ 3345 if (pte & PG_A) 3346 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 3347 3348 /* 3349 * Referenced by someone 3350 */ 3351 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) { 3352 val |= MINCORE_REFERENCED_OTHER; 3353 vm_page_flag_set(m, PG_REFERENCED); 3354 } 3355 } 3356 return val; 3357} 3358 3359void 3360pmap_activate(struct thread *td) 3361{ 3362 struct proc *p = td->td_proc; 3363 pmap_t pmap; 3364 u_int32_t cr3; 3365 3366 pmap = vmspace_pmap(td->td_proc->p_vmspace); 3367#if defined(SMP) 3368 pmap->pm_active |= PCPU_GET(cpumask); 3369#else 3370 pmap->pm_active |= 1; 3371#endif 3372#if defined(SWTCH_OPTIM_STATS) 3373 tlb_flush_count++; 3374#endif 3375 cr3 = vtophys(pmap->pm_pdir); 3376 /* XXXKSE this is wrong. 3377 * pmap_activate is for the current thread on the current cpu 3378 */ 3379 if (p->p_flag & P_KSES) { 3380 /* Make sure all other cr3 entries are updated. */ 3381 /* what if they are running? XXXKSE (maybe abort them) */ 3382 FOREACH_THREAD_IN_PROC(p, td) { 3383 td->td_pcb->pcb_cr3 = cr3; 3384 } 3385 } else { 3386 td->td_pcb->pcb_cr3 = cr3; 3387 } 3388 load_cr3(cr3); 3389} 3390 3391vm_offset_t 3392pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3393{ 3394 3395 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3396 return addr; 3397 } 3398 3399 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); 3400 return addr; 3401} 3402 3403 3404#if defined(PMAP_DEBUG) 3405pmap_pid_dump(int pid) 3406{ 3407 pmap_t pmap; 3408 struct proc *p; 3409 int npte = 0; 3410 int index; 3411 3412 sx_slock(&allproc_lock); 3413 LIST_FOREACH(p, &allproc, p_list) { 3414 if (p->p_pid != pid) 3415 continue; 3416 3417 if (p->p_vmspace) { 3418 int i,j; 3419 index = 0; 3420 pmap = vmspace_pmap(p->p_vmspace); 3421 for (i = 0; i < NPDEPG; i++) { 3422 pd_entry_t *pde; 3423 pt_entry_t *pte; 3424 vm_offset_t base = i << PDRSHIFT; 3425 3426 pde = &pmap->pm_pdir[i]; 3427 if (pde && pmap_pde_v(pde)) { 3428 for (j = 0; j < NPTEPG; j++) { 3429 vm_offset_t va = base + (j << PAGE_SHIFT); 3430 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3431 if (index) { 3432 index = 0; 3433 printf("\n"); 3434 } 3435 sx_sunlock(&allproc_lock); 3436 return npte; 3437 } 3438 pte = pmap_pte_quick(pmap, va); 3439 if (pte && pmap_pte_v(pte)) { 3440 pt_entry_t pa; 3441 vm_page_t m; 3442 pa = *pte; 3443 m = PHYS_TO_VM_PAGE(pa); 3444 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3445 va, pa, m->hold_count, m->wire_count, m->flags); 3446 npte++; 3447 index++; 3448 if (index >= 2) { 3449 index = 0; 3450 printf("\n"); 3451 } else { 3452 printf(" "); 3453 } 3454 } 3455 } 3456 } 3457 } 3458 } 3459 } 3460 sx_sunlock(&allproc_lock); 3461 return npte; 3462} 3463#endif 3464 3465#if defined(DEBUG) 3466 3467static void pads(pmap_t pm); 3468void pmap_pvdump(vm_offset_t pa); 3469 3470/* print address space of pmap*/ 3471static void 3472pads(pm) 3473 pmap_t pm; 3474{ 3475 int i, j; 3476 vm_offset_t va; 3477 pt_entry_t *ptep; 3478 3479 if (pm == kernel_pmap) 3480 return; 3481 for (i = 0; i < NPDEPG; i++) 3482 if (pm->pm_pdir[i]) 3483 for (j = 0; j < NPTEPG; j++) { 3484 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3485 if (pm == kernel_pmap && va < KERNBASE) 3486 continue; 3487 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3488 continue; 3489 ptep = pmap_pte_quick(pm, va); 3490 if (pmap_pte_v(ptep)) 3491 printf("%x:%x ", va, *ptep); 3492 }; 3493 3494} 3495 3496void 3497pmap_pvdump(pa) 3498 vm_offset_t pa; 3499{ 3500 pv_entry_t pv; 3501 vm_page_t m; 3502 3503 printf("pa %x", pa); 3504 m = PHYS_TO_VM_PAGE(pa); 3505 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3506 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3507 pads(pv->pv_pmap); 3508 } 3509 printf(" "); 3510} 3511#endif 3512