pmap.c revision 110781
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 110781 2003-02-13 02:42:06Z 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); 154static struct pmaplist allpmaps; 155static struct mtx allpmaps_lock; 156 157vm_offset_t avail_start; /* PA of first available physical page */ 158vm_offset_t avail_end; /* PA of last available physical page */ 159vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 160vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 161static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 162static int pgeflag; /* PG_G or-in */ 163static int pseflag; /* PG_PS or-in */ 164 165static int nkpt; 166vm_offset_t kernel_vm_end; 167extern u_int32_t KERNend; 168 169/* 170 * Data for the pv entry allocation mechanism 171 */ 172static uma_zone_t pvzone; 173static struct vm_object pvzone_obj; 174static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 175int pmap_pagedaemon_waken; 176 177/* 178 * All those kernel PT submaps that BSD is so fond of 179 */ 180pt_entry_t *CMAP1 = 0; 181static pt_entry_t *CMAP2, *CMAP3, *ptmmap; 182caddr_t CADDR1 = 0, ptvmmap = 0; 183static caddr_t CADDR2, CADDR3; 184static pt_entry_t *msgbufmap; 185struct msgbuf *msgbufp = 0; 186 187/* 188 * Crashdump maps. 189 */ 190static pt_entry_t *pt_crashdumpmap; 191static caddr_t crashdumpmap; 192 193#ifdef SMP 194extern pt_entry_t *SMPpt; 195#endif 196static pt_entry_t *PMAP1 = 0; 197static pt_entry_t *PADDR1 = 0; 198 199static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 200static pt_entry_t *get_ptbase(pmap_t pmap); 201static pv_entry_t get_pv_entry(void); 202static void i386_protection_init(void); 203static __inline void pmap_changebit(vm_page_t m, int bit, boolean_t setem); 204 205static vm_page_t pmap_enter_quick(pmap_t pmap, vm_offset_t va, 206 vm_page_t m, vm_page_t mpte); 207static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 208static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 209static int pmap_remove_entry(struct pmap *pmap, vm_page_t m, 210 vm_offset_t va); 211static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, 212 vm_page_t mpte, vm_page_t m); 213 214static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va); 215 216static int pmap_release_free_page(pmap_t pmap, vm_page_t p); 217static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex); 218static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 219static vm_page_t pmap_page_lookup(vm_object_t object, vm_pindex_t pindex); 220static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t); 221static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 222static void *pmap_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 223 224static pd_entry_t pdir4mb; 225 226/* 227 * Routine: pmap_pte 228 * Function: 229 * Extract the page table entry associated 230 * with the given map/virtual_address pair. 231 */ 232 233PMAP_INLINE pt_entry_t * 234pmap_pte(pmap, va) 235 register pmap_t pmap; 236 vm_offset_t va; 237{ 238 pd_entry_t *pdeaddr; 239 240 if (pmap) { 241 pdeaddr = pmap_pde(pmap, va); 242 if (*pdeaddr & PG_PS) 243 return pdeaddr; 244 if (*pdeaddr) { 245 return get_ptbase(pmap) + i386_btop(va); 246 } 247 } 248 return (0); 249} 250 251/* 252 * Move the kernel virtual free pointer to the next 253 * 4MB. This is used to help improve performance 254 * by using a large (4MB) page for much of the kernel 255 * (.text, .data, .bss) 256 */ 257static vm_offset_t 258pmap_kmem_choose(vm_offset_t addr) 259{ 260 vm_offset_t newaddr = addr; 261 262#ifdef I686_CPU_not /* Problem seems to have gone away */ 263 /* Deal with un-resolved Pentium4 issues */ 264 if (cpu_class == CPUCLASS_686 && 265 strcmp(cpu_vendor, "GenuineIntel") == 0 && 266 (cpu_id & 0xf00) == 0xf00) 267 return newaddr; 268#endif 269#ifndef DISABLE_PSE 270 if (cpu_feature & CPUID_PSE) 271 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); 272#endif 273 return newaddr; 274} 275 276/* 277 * Bootstrap the system enough to run with virtual memory. 278 * 279 * On the i386 this is called after mapping has already been enabled 280 * and just syncs the pmap module with what has already been done. 281 * [We can't call it easily with mapping off since the kernel is not 282 * mapped with PA == VA, hence we would have to relocate every address 283 * from the linked base (virtual) address "KERNBASE" to the actual 284 * (physical) address starting relative to 0] 285 */ 286void 287pmap_bootstrap(firstaddr, loadaddr) 288 vm_offset_t firstaddr; 289 vm_offset_t loadaddr; 290{ 291 vm_offset_t va; 292 pt_entry_t *pte; 293 int i; 294 295 avail_start = firstaddr; 296 297 /* 298 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 299 * large. It should instead be correctly calculated in locore.s and 300 * not based on 'first' (which is a physical address, not a virtual 301 * address, for the start of unused physical memory). The kernel 302 * page tables are NOT double mapped and thus should not be included 303 * in this calculation. 304 */ 305 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 306 virtual_avail = pmap_kmem_choose(virtual_avail); 307 308 virtual_end = VM_MAX_KERNEL_ADDRESS; 309 310 /* 311 * Initialize protection array. 312 */ 313 i386_protection_init(); 314 315 /* 316 * Initialize the kernel pmap (which is statically allocated). 317 */ 318 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 319 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 320 TAILQ_INIT(&kernel_pmap->pm_pvlist); 321 LIST_INIT(&allpmaps); 322 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 323 mtx_lock_spin(&allpmaps_lock); 324 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 325 mtx_unlock_spin(&allpmaps_lock); 326 nkpt = NKPT; 327 328 /* 329 * Reserve some special page table entries/VA space for temporary 330 * mapping of pages. 331 */ 332#define SYSMAP(c, p, v, n) \ 333 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 334 335 va = virtual_avail; 336 pte = (pt_entry_t *) pmap_pte(kernel_pmap, va); 337 338 /* 339 * CMAP1/CMAP2 are used for zeroing and copying pages. 340 * CMAP3 is used for the idle process page zeroing. 341 */ 342 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 343 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 344 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 345 346 /* 347 * Crashdump maps. 348 */ 349 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 350 351 /* 352 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 353 * XXX ptmmap is not used. 354 */ 355 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 356 357 /* 358 * msgbufp is used to map the system message buffer. 359 * XXX msgbufmap is not used. 360 */ 361 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 362 atop(round_page(MSGBUF_SIZE))) 363 364 /* 365 * ptemap is used for pmap_pte_quick 366 */ 367 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 368 369 virtual_avail = va; 370 371 *CMAP1 = *CMAP2 = 0; 372 for (i = 0; i < NKPT; i++) 373 PTD[i] = 0; 374 375 pgeflag = 0; 376#ifndef DISABLE_PG_G 377 if (cpu_feature & CPUID_PGE) 378 pgeflag = PG_G; 379#endif 380#ifdef I686_CPU_not /* Problem seems to have gone away */ 381 /* Deal with un-resolved Pentium4 issues */ 382 if (cpu_class == CPUCLASS_686 && 383 strcmp(cpu_vendor, "GenuineIntel") == 0 && 384 (cpu_id & 0xf00) == 0xf00) { 385 printf("Warning: Pentium 4 cpu: PG_G disabled (global flag)\n"); 386 pgeflag = 0; 387 } 388#endif 389 390/* 391 * Initialize the 4MB page size flag 392 */ 393 pseflag = 0; 394/* 395 * The 4MB page version of the initial 396 * kernel page mapping. 397 */ 398 pdir4mb = 0; 399 400#ifndef DISABLE_PSE 401 if (cpu_feature & CPUID_PSE) 402 pseflag = PG_PS; 403#endif 404#ifdef I686_CPU_not /* Problem seems to have gone away */ 405 /* Deal with un-resolved Pentium4 issues */ 406 if (cpu_class == CPUCLASS_686 && 407 strcmp(cpu_vendor, "GenuineIntel") == 0 && 408 (cpu_id & 0xf00) == 0xf00) { 409 printf("Warning: Pentium 4 cpu: PG_PS disabled (4MB pages)\n"); 410 pseflag = 0; 411 } 412#endif 413#ifndef DISABLE_PSE 414 if (pseflag) { 415 pd_entry_t ptditmp; 416 /* 417 * Note that we have enabled PSE mode 418 */ 419 ptditmp = *(PTmap + i386_btop(KERNBASE)); 420 ptditmp &= ~(NBPDR - 1); 421 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag; 422 pdir4mb = ptditmp; 423 } 424#endif 425#ifndef SMP 426 /* 427 * Turn on PGE/PSE. SMP does this later on since the 428 * 4K page tables are required for AP boot (for now). 429 * XXX fixme. 430 */ 431 pmap_set_opt(); 432#endif 433#ifdef SMP 434 if (cpu_apic_address == 0) 435 panic("pmap_bootstrap: no local apic! (non-SMP hardware?)"); 436 437 /* local apic is mapped on last page */ 438 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag | 439 (cpu_apic_address & PG_FRAME)); 440#endif 441 invltlb(); 442} 443 444/* 445 * Enable 4MB page mode for MP startup. Turn on PG_G support. 446 * BSP will run this after all the AP's have started up. 447 */ 448void 449pmap_set_opt(void) 450{ 451 pt_entry_t *pte; 452 vm_offset_t va, endva; 453 454 if (pgeflag && (cpu_feature & CPUID_PGE)) { 455 load_cr4(rcr4() | CR4_PGE); 456 invltlb(); /* Insurance */ 457 } 458#ifndef DISABLE_PSE 459 if (pseflag && (cpu_feature & CPUID_PSE)) { 460 load_cr4(rcr4() | CR4_PSE); 461 invltlb(); /* Insurance */ 462 } 463#endif 464 if (PCPU_GET(cpuid) == 0) { 465#ifndef DISABLE_PSE 466 if (pdir4mb) { 467 kernel_pmap->pm_pdir[KPTDI] = PTD[KPTDI] = pdir4mb; 468 invltlb(); /* Insurance */ 469 } 470#endif 471 if (pgeflag) { 472 /* Turn on PG_G for text, data, bss pages. */ 473 va = (vm_offset_t)btext; 474#ifndef DISABLE_PSE 475 if (pseflag && (cpu_feature & CPUID_PSE)) { 476 if (va < KERNBASE + (1 << PDRSHIFT)) 477 va = KERNBASE + (1 << PDRSHIFT); 478 } 479#endif 480 endva = KERNBASE + KERNend; 481 while (va < endva) { 482 pte = vtopte(va); 483 if (*pte) 484 *pte |= pgeflag; 485 va += PAGE_SIZE; 486 } 487 invltlb(); /* Insurance */ 488 } 489 /* 490 * We do not need to broadcast the invltlb here, because 491 * each AP does it the moment it is released from the boot 492 * lock. See ap_init(). 493 */ 494 } 495} 496 497static void * 498pmap_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 499{ 500 *flags = UMA_SLAB_PRIV; 501 return (void *)kmem_alloc(kernel_map, bytes); 502} 503 504/* 505 * Initialize the pmap module. 506 * Called by vm_init, to initialize any structures that the pmap 507 * system needs to map virtual memory. 508 * pmap_init has been enhanced to support in a fairly consistant 509 * way, discontiguous physical memory. 510 */ 511void 512pmap_init(phys_start, phys_end) 513 vm_offset_t phys_start, phys_end; 514{ 515 int i; 516 int initial_pvs; 517 518 /* 519 * Allocate memory for random pmap data structures. Includes the 520 * pv_head_table. 521 */ 522 523 for(i = 0; i < vm_page_array_size; i++) { 524 vm_page_t m; 525 526 m = &vm_page_array[i]; 527 TAILQ_INIT(&m->md.pv_list); 528 m->md.pv_list_count = 0; 529 } 530 531 /* 532 * init the pv free list 533 */ 534 initial_pvs = vm_page_array_size; 535 if (initial_pvs < MINPV) 536 initial_pvs = MINPV; 537 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL, 538 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM); 539 uma_zone_set_allocf(pvzone, pmap_allocf); 540 uma_prealloc(pvzone, initial_pvs); 541 542 /* 543 * Now it is safe to enable pv_table recording. 544 */ 545 pmap_initialized = TRUE; 546} 547 548/* 549 * Initialize the address space (zone) for the pv_entries. Set a 550 * high water mark so that the system can recover from excessive 551 * numbers of pv entries. 552 */ 553void 554pmap_init2() 555{ 556 int shpgperproc = PMAP_SHPGPERPROC; 557 558 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 559 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 560 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 561 pv_entry_high_water = 9 * (pv_entry_max / 10); 562 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 563} 564 565 566/*************************************************** 567 * Low level helper routines..... 568 ***************************************************/ 569 570#if defined(PMAP_DIAGNOSTIC) 571 572/* 573 * This code checks for non-writeable/modified pages. 574 * This should be an invalid condition. 575 */ 576static int 577pmap_nw_modified(pt_entry_t ptea) 578{ 579 int pte; 580 581 pte = (int) ptea; 582 583 if ((pte & (PG_M|PG_RW)) == PG_M) 584 return 1; 585 else 586 return 0; 587} 588#endif 589 590 591/* 592 * this routine defines the region(s) of memory that should 593 * not be tested for the modified bit. 594 */ 595static PMAP_INLINE int 596pmap_track_modified(vm_offset_t va) 597{ 598 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 599 return 1; 600 else 601 return 0; 602} 603 604#ifdef I386_CPU 605/* 606 * i386 only has "invalidate everything" and no SMP to worry about. 607 */ 608PMAP_INLINE void 609pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 610{ 611 612 if (pmap == kernel_pmap || pmap->pm_active) 613 invltlb(); 614} 615 616PMAP_INLINE void 617pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 618{ 619 620 if (pmap == kernel_pmap || pmap->pm_active) 621 invltlb(); 622} 623 624PMAP_INLINE void 625pmap_invalidate_all(pmap_t pmap) 626{ 627 628 if (pmap == kernel_pmap || pmap->pm_active) 629 invltlb(); 630} 631#else /* !I386_CPU */ 632#ifdef SMP 633/* 634 * For SMP, these functions have to use the IPI mechanism for coherence. 635 */ 636void 637pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 638{ 639 u_int cpumask; 640 u_int other_cpus; 641 642 critical_enter(); 643 /* 644 * We need to disable interrupt preemption but MUST NOT have 645 * interrupts disabled here. 646 * XXX we may need to hold schedlock to get a coherent pm_active 647 */ 648 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) { 649 invlpg(va); 650 smp_invlpg(va); 651 } else { 652 cpumask = PCPU_GET(cpumask); 653 other_cpus = PCPU_GET(other_cpus); 654 if (pmap->pm_active & cpumask) 655 invlpg(va); 656 if (pmap->pm_active & other_cpus) 657 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 658 } 659 critical_exit(); 660} 661 662void 663pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 664{ 665 u_int cpumask; 666 u_int other_cpus; 667 vm_offset_t addr; 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 for (addr = sva; addr < eva; addr += PAGE_SIZE) 677 invlpg(addr); 678 smp_invlpg_range(sva, eva); 679 } else { 680 cpumask = PCPU_GET(cpumask); 681 other_cpus = PCPU_GET(other_cpus); 682 if (pmap->pm_active & cpumask) 683 for (addr = sva; addr < eva; addr += PAGE_SIZE) 684 invlpg(addr); 685 if (pmap->pm_active & other_cpus) 686 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 687 sva, eva); 688 } 689 critical_exit(); 690} 691 692void 693pmap_invalidate_all(pmap_t pmap) 694{ 695 u_int cpumask; 696 u_int other_cpus; 697 698#ifdef SWTCH_OPTIM_STATS 699 tlb_flush_count++; 700#endif 701 critical_enter(); 702 /* 703 * We need to disable interrupt preemption but MUST NOT have 704 * interrupts disabled here. 705 * XXX we may need to hold schedlock to get a coherent pm_active 706 */ 707 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) { 708 invltlb(); 709 smp_invltlb(); 710 } else { 711 cpumask = PCPU_GET(cpumask); 712 other_cpus = PCPU_GET(other_cpus); 713 if (pmap->pm_active & cpumask) 714 invltlb(); 715 if (pmap->pm_active & other_cpus) 716 smp_masked_invltlb(pmap->pm_active & other_cpus); 717 } 718 critical_exit(); 719} 720#else /* !SMP */ 721/* 722 * Normal, non-SMP, 486+ invalidation functions. 723 * We inline these within pmap.c for speed. 724 */ 725PMAP_INLINE void 726pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 727{ 728 729 if (pmap == kernel_pmap || pmap->pm_active) 730 invlpg(va); 731} 732 733PMAP_INLINE void 734pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 735{ 736 vm_offset_t addr; 737 738 if (pmap == kernel_pmap || pmap->pm_active) 739 for (addr = sva; addr < eva; addr += PAGE_SIZE) 740 invlpg(addr); 741} 742 743PMAP_INLINE void 744pmap_invalidate_all(pmap_t pmap) 745{ 746 747 if (pmap == kernel_pmap || pmap->pm_active) 748 invltlb(); 749} 750#endif /* !SMP */ 751#endif /* !I386_CPU */ 752 753/* 754 * Return an address which is the base of the Virtual mapping of 755 * all the PTEs for the given pmap. Note this doesn't say that 756 * all the PTEs will be present or that the pages there are valid. 757 * The PTEs are made available by the recursive mapping trick. 758 * It will map in the alternate PTE space if needed. 759 */ 760static pt_entry_t * 761get_ptbase(pmap) 762 pmap_t pmap; 763{ 764 pd_entry_t frame; 765 766 /* are we current address space or kernel? */ 767 if (pmap == kernel_pmap) 768 return PTmap; 769 frame = pmap->pm_pdir[PTDPTDI] & PG_FRAME; 770 if (frame == (PTDpde & PG_FRAME)) 771 return PTmap; 772 /* otherwise, we are alternate address space */ 773 if (frame != (APTDpde & PG_FRAME)) { 774 APTDpde = (pd_entry_t) (frame | PG_RW | PG_V); 775 pmap_invalidate_all(kernel_pmap); /* XXX Bandaid */ 776 } 777 return APTmap; 778} 779 780/* 781 * Super fast pmap_pte routine best used when scanning 782 * the pv lists. This eliminates many coarse-grained 783 * invltlb calls. Note that many of the pv list 784 * scans are across different pmaps. It is very wasteful 785 * to do an entire invltlb for checking a single mapping. 786 */ 787 788static pt_entry_t * 789pmap_pte_quick(pmap, va) 790 register pmap_t pmap; 791 vm_offset_t va; 792{ 793 pd_entry_t pde, newpf; 794 pde = pmap->pm_pdir[va >> PDRSHIFT]; 795 if (pde != 0) { 796 pd_entry_t frame = pmap->pm_pdir[PTDPTDI] & PG_FRAME; 797 unsigned index = i386_btop(va); 798 /* are we current address space or kernel? */ 799 if (pmap == kernel_pmap || frame == (PTDpde & PG_FRAME)) 800 return PTmap + index; 801 newpf = pde & PG_FRAME; 802 if (((*PMAP1) & PG_FRAME) != newpf) { 803 *PMAP1 = newpf | PG_RW | PG_V; 804 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR1); 805 } 806 return PADDR1 + (index & (NPTEPG - 1)); 807 } 808 return (0); 809} 810 811/* 812 * Routine: pmap_extract 813 * Function: 814 * Extract the physical page address associated 815 * with the given map/virtual_address pair. 816 */ 817vm_offset_t 818pmap_extract(pmap, va) 819 register pmap_t pmap; 820 vm_offset_t va; 821{ 822 vm_offset_t rtval; /* XXX FIXME */ 823 vm_offset_t pdirindex; 824 825 if (pmap == 0) 826 return 0; 827 pdirindex = va >> PDRSHIFT; 828 rtval = pmap->pm_pdir[pdirindex]; 829 if (rtval != 0) { 830 pt_entry_t *pte; 831 if ((rtval & PG_PS) != 0) { 832 rtval &= ~(NBPDR - 1); 833 rtval |= va & (NBPDR - 1); 834 return rtval; 835 } 836 pte = get_ptbase(pmap) + i386_btop(va); 837 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK)); 838 return rtval; 839 } 840 return 0; 841 842} 843 844/*************************************************** 845 * Low level mapping routines..... 846 ***************************************************/ 847 848/* 849 * Add a wired page to the kva. 850 * Note: not SMP coherent. 851 */ 852PMAP_INLINE void 853pmap_kenter(vm_offset_t va, vm_offset_t pa) 854{ 855 pt_entry_t *pte; 856 857 pte = vtopte(va); 858 *pte = pa | PG_RW | PG_V | pgeflag; 859} 860 861/* 862 * Remove a page from the kernel pagetables. 863 * Note: not SMP coherent. 864 */ 865PMAP_INLINE void 866pmap_kremove(vm_offset_t va) 867{ 868 pt_entry_t *pte; 869 870 pte = vtopte(va); 871 *pte = 0; 872} 873 874/* 875 * Used to map a range of physical addresses into kernel 876 * virtual address space. 877 * 878 * The value passed in '*virt' is a suggested virtual address for 879 * the mapping. Architectures which can support a direct-mapped 880 * physical to virtual region can return the appropriate address 881 * within that region, leaving '*virt' unchanged. Other 882 * architectures should map the pages starting at '*virt' and 883 * update '*virt' with the first usable address after the mapped 884 * region. 885 */ 886vm_offset_t 887pmap_map(vm_offset_t *virt, vm_offset_t start, vm_offset_t end, int prot) 888{ 889 vm_offset_t va, sva; 890 891 va = sva = *virt; 892 while (start < end) { 893 pmap_kenter(va, start); 894 va += PAGE_SIZE; 895 start += PAGE_SIZE; 896 } 897 pmap_invalidate_range(kernel_pmap, sva, va); 898 *virt = va; 899 return (sva); 900} 901 902 903/* 904 * Add a list of wired pages to the kva 905 * this routine is only used for temporary 906 * kernel mappings that do not need to have 907 * page modification or references recorded. 908 * Note that old mappings are simply written 909 * over. The page *must* be wired. 910 * Note: SMP coherent. Uses a ranged shootdown IPI. 911 */ 912void 913pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 914{ 915 vm_offset_t va; 916 917 va = sva; 918 while (count-- > 0) { 919 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 920 va += PAGE_SIZE; 921 m++; 922 } 923 pmap_invalidate_range(kernel_pmap, sva, va); 924} 925 926/* 927 * This routine tears out page mappings from the 928 * kernel -- it is meant only for temporary mappings. 929 * Note: SMP coherent. Uses a ranged shootdown IPI. 930 */ 931void 932pmap_qremove(vm_offset_t sva, int count) 933{ 934 vm_offset_t va; 935 936 va = sva; 937 while (count-- > 0) { 938 pmap_kremove(va); 939 va += PAGE_SIZE; 940 } 941 pmap_invalidate_range(kernel_pmap, sva, va); 942} 943 944static vm_page_t 945pmap_page_lookup(vm_object_t object, vm_pindex_t pindex) 946{ 947 vm_page_t m; 948 949retry: 950 m = vm_page_lookup(object, pindex); 951 if (m != NULL) { 952 vm_page_lock_queues(); 953 if (vm_page_sleep_if_busy(m, FALSE, "pplookp")) 954 goto retry; 955 vm_page_unlock_queues(); 956 } 957 return m; 958} 959 960#ifndef KSTACK_MAX_PAGES 961#define KSTACK_MAX_PAGES 32 962#endif 963 964/* 965 * Create the kernel stack (including pcb for i386) for a new thread. 966 * This routine directly affects the fork perf for a process and 967 * create performance for a thread. 968 */ 969void 970pmap_new_thread(struct thread *td, int pages) 971{ 972 int i; 973 vm_page_t ma[KSTACK_MAX_PAGES]; 974 vm_object_t ksobj; 975 vm_page_t m; 976 vm_offset_t ks; 977 978 /* Bounds check */ 979 if (pages <= 1) 980 pages = KSTACK_PAGES; 981 else if (pages > KSTACK_MAX_PAGES) 982 pages = KSTACK_MAX_PAGES; 983 984 /* 985 * allocate object for the kstack 986 */ 987 ksobj = vm_object_allocate(OBJT_DEFAULT, pages); 988 td->td_kstack_obj = ksobj; 989 990 /* get a kernel virtual address for the kstack for this thread */ 991#ifdef KSTACK_GUARD 992 ks = kmem_alloc_nofault(kernel_map, (pages + 1) * PAGE_SIZE); 993 if (ks == 0) 994 panic("pmap_new_thread: kstack allocation failed"); 995 if (*vtopte(ks) != 0) 996 pmap_qremove(ks, 1); 997 ks += PAGE_SIZE; 998 td->td_kstack = ks; 999#else 1000 /* get a kernel virtual address for the kstack for this thread */ 1001 ks = kmem_alloc_nofault(kernel_map, pages * PAGE_SIZE); 1002 if (ks == 0) 1003 panic("pmap_new_thread: kstack allocation failed"); 1004 td->td_kstack = ks; 1005#endif 1006 /* 1007 * Knowing the number of pages allocated is useful when you 1008 * want to deallocate them. 1009 */ 1010 td->td_kstack_pages = pages; 1011 1012 /* 1013 * For the length of the stack, link in a real page of ram for each 1014 * page of stack. 1015 */ 1016 for (i = 0; i < pages; i++) { 1017 /* 1018 * Get a kernel stack page 1019 */ 1020 m = vm_page_grab(ksobj, i, 1021 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED); 1022 ma[i] = m; 1023 1024 vm_page_lock_queues(); 1025 vm_page_wakeup(m); 1026 vm_page_flag_clear(m, PG_ZERO); 1027 m->valid = VM_PAGE_BITS_ALL; 1028 vm_page_unlock_queues(); 1029 } 1030 pmap_qenter(ks, ma, pages); 1031} 1032 1033/* 1034 * Dispose the kernel stack for a thread that has exited. 1035 * This routine directly impacts the exit perf of a process and thread. 1036 */ 1037void 1038pmap_dispose_thread(td) 1039 struct thread *td; 1040{ 1041 int i; 1042 int pages; 1043 vm_object_t ksobj; 1044 vm_offset_t ks; 1045 vm_page_t m; 1046 1047 pages = td->td_kstack_pages; 1048 ksobj = td->td_kstack_obj; 1049 ks = td->td_kstack; 1050 pmap_qremove(ks, pages); 1051 for (i = 0; i < pages; i++) { 1052 m = vm_page_lookup(ksobj, i); 1053 if (m == NULL) 1054 panic("pmap_dispose_thread: kstack already missing?"); 1055 vm_page_lock_queues(); 1056 vm_page_busy(m); 1057 vm_page_unwire(m, 0); 1058 vm_page_free(m); 1059 vm_page_unlock_queues(); 1060 } 1061 /* 1062 * Free the space that this stack was mapped to in the kernel 1063 * address map. 1064 */ 1065#ifdef KSTACK_GUARD 1066 kmem_free(kernel_map, ks - PAGE_SIZE, (pages + 1) * PAGE_SIZE); 1067#else 1068 kmem_free(kernel_map, ks, pages * PAGE_SIZE); 1069#endif 1070 vm_object_deallocate(ksobj); 1071} 1072 1073/* 1074 * Set up a variable sized alternate kstack. Though it may look MI, it may 1075 * need to be different on certain arches like ia64. 1076 */ 1077void 1078pmap_new_altkstack(struct thread *td, int pages) 1079{ 1080 /* shuffle the original stack */ 1081 td->td_altkstack_obj = td->td_kstack_obj; 1082 td->td_altkstack = td->td_kstack; 1083 td->td_altkstack_pages = td->td_kstack_pages; 1084 1085 pmap_new_thread(td, pages); 1086} 1087 1088void 1089pmap_dispose_altkstack(td) 1090 struct thread *td; 1091{ 1092 pmap_dispose_thread(td); 1093 1094 /* restore the original kstack */ 1095 td->td_kstack = td->td_altkstack; 1096 td->td_kstack_obj = td->td_altkstack_obj; 1097 td->td_kstack_pages = td->td_altkstack_pages; 1098 td->td_altkstack = 0; 1099 td->td_altkstack_obj = NULL; 1100 td->td_altkstack_pages = 0; 1101} 1102 1103/* 1104 * Allow the Kernel stack for a thread to be prejudicially paged out. 1105 */ 1106void 1107pmap_swapout_thread(td) 1108 struct thread *td; 1109{ 1110 int i; 1111 int pages; 1112 vm_object_t ksobj; 1113 vm_offset_t ks; 1114 vm_page_t m; 1115 1116 pages = td->td_kstack_pages; 1117 ksobj = td->td_kstack_obj; 1118 ks = td->td_kstack; 1119 pmap_qremove(ks, pages); 1120 for (i = 0; i < pages; i++) { 1121 m = vm_page_lookup(ksobj, i); 1122 if (m == NULL) 1123 panic("pmap_swapout_thread: kstack already missing?"); 1124 vm_page_lock_queues(); 1125 vm_page_dirty(m); 1126 vm_page_unwire(m, 0); 1127 vm_page_unlock_queues(); 1128 } 1129} 1130 1131/* 1132 * Bring the kernel stack for a specified thread back in. 1133 */ 1134void 1135pmap_swapin_thread(td) 1136 struct thread *td; 1137{ 1138 int i, rv; 1139 int pages; 1140 vm_page_t ma[KSTACK_MAX_PAGES]; 1141 vm_object_t ksobj; 1142 vm_offset_t ks; 1143 vm_page_t m; 1144 1145 pages = td->td_kstack_pages; 1146 ksobj = td->td_kstack_obj; 1147 ks = td->td_kstack; 1148 for (i = 0; i < pages; i++) { 1149 m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1150 if (m->valid != VM_PAGE_BITS_ALL) { 1151 rv = vm_pager_get_pages(ksobj, &m, 1, 0); 1152 if (rv != VM_PAGER_OK) 1153 panic("pmap_swapin_thread: cannot get kstack for proc: %d\n", td->td_proc->p_pid); 1154 m = vm_page_lookup(ksobj, i); 1155 m->valid = VM_PAGE_BITS_ALL; 1156 } 1157 ma[i] = m; 1158 vm_page_lock_queues(); 1159 vm_page_wire(m); 1160 vm_page_wakeup(m); 1161 vm_page_unlock_queues(); 1162 } 1163 pmap_qenter(ks, ma, pages); 1164} 1165 1166/*************************************************** 1167 * Page table page management routines..... 1168 ***************************************************/ 1169 1170/* 1171 * This routine unholds page table pages, and if the hold count 1172 * drops to zero, then it decrements the wire count. 1173 */ 1174static int 1175_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1176{ 1177 1178 while (vm_page_sleep_if_busy(m, FALSE, "pmuwpt")) 1179 vm_page_lock_queues(); 1180 1181 if (m->hold_count == 0) { 1182 vm_offset_t pteva; 1183 /* 1184 * unmap the page table page 1185 */ 1186 pmap->pm_pdir[m->pindex] = 0; 1187 --pmap->pm_stats.resident_count; 1188 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == 1189 (PTDpde & PG_FRAME)) { 1190 /* 1191 * Do an invltlb to make the invalidated mapping 1192 * take effect immediately. 1193 */ 1194 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1195 pmap_invalidate_page(pmap, pteva); 1196 } 1197 1198 if (pmap->pm_ptphint == m) 1199 pmap->pm_ptphint = NULL; 1200 1201 /* 1202 * If the page is finally unwired, simply free it. 1203 */ 1204 --m->wire_count; 1205 if (m->wire_count == 0) { 1206 vm_page_busy(m); 1207 vm_page_free_zero(m); 1208 atomic_subtract_int(&cnt.v_wire_count, 1); 1209 } 1210 return 1; 1211 } 1212 return 0; 1213} 1214 1215static PMAP_INLINE int 1216pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1217{ 1218 vm_page_unhold(m); 1219 if (m->hold_count == 0) 1220 return _pmap_unwire_pte_hold(pmap, m); 1221 else 1222 return 0; 1223} 1224 1225/* 1226 * After removing a page table entry, this routine is used to 1227 * conditionally free the page, and manage the hold/wire counts. 1228 */ 1229static int 1230pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 1231{ 1232 unsigned ptepindex; 1233 if (va >= VM_MAXUSER_ADDRESS) 1234 return 0; 1235 1236 if (mpte == NULL) { 1237 ptepindex = (va >> PDRSHIFT); 1238 if (pmap->pm_ptphint && 1239 (pmap->pm_ptphint->pindex == ptepindex)) { 1240 mpte = pmap->pm_ptphint; 1241 } else { 1242 while ((mpte = vm_page_lookup(pmap->pm_pteobj, ptepindex)) != NULL && 1243 vm_page_sleep_if_busy(mpte, FALSE, "pulook")) 1244 vm_page_lock_queues(); 1245 pmap->pm_ptphint = mpte; 1246 } 1247 } 1248 1249 return pmap_unwire_pte_hold(pmap, mpte); 1250} 1251 1252void 1253pmap_pinit0(pmap) 1254 struct pmap *pmap; 1255{ 1256 pmap->pm_pdir = 1257 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE); 1258 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t)IdlePTD); 1259#ifndef I386_CPU 1260 invlpg((vm_offset_t)pmap->pm_pdir); 1261#else 1262 invltlb(); 1263#endif 1264 pmap->pm_ptphint = NULL; 1265 pmap->pm_active = 0; 1266 TAILQ_INIT(&pmap->pm_pvlist); 1267 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1268 mtx_lock_spin(&allpmaps_lock); 1269 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1270 mtx_unlock_spin(&allpmaps_lock); 1271} 1272 1273/* 1274 * Initialize a preallocated and zeroed pmap structure, 1275 * such as one in a vmspace structure. 1276 */ 1277void 1278pmap_pinit(pmap) 1279 register struct pmap *pmap; 1280{ 1281 vm_page_t ptdpg; 1282 1283 /* 1284 * No need to allocate page table space yet but we do need a valid 1285 * page directory table. 1286 */ 1287 if (pmap->pm_pdir == NULL) 1288 pmap->pm_pdir = 1289 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE); 1290 1291 /* 1292 * allocate object for the ptes 1293 */ 1294 if (pmap->pm_pteobj == NULL) 1295 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1); 1296 1297 /* 1298 * allocate the page directory page 1299 */ 1300 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI, 1301 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED | VM_ALLOC_ZERO); 1302 vm_page_lock_queues(); 1303 vm_page_flag_clear(ptdpg, PG_BUSY); 1304 ptdpg->valid = VM_PAGE_BITS_ALL; 1305 vm_page_unlock_queues(); 1306 1307 pmap_qenter((vm_offset_t) pmap->pm_pdir, &ptdpg, 1); 1308 if ((ptdpg->flags & PG_ZERO) == 0) 1309 bzero(pmap->pm_pdir, PAGE_SIZE); 1310 1311 mtx_lock_spin(&allpmaps_lock); 1312 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1313 mtx_unlock_spin(&allpmaps_lock); 1314 /* Wire in kernel global address entries. */ 1315 /* XXX copies current process, does not fill in MPPTDI */ 1316 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE); 1317#ifdef SMP 1318 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1319#endif 1320 1321 /* install self-referential address mapping entry */ 1322 pmap->pm_pdir[PTDPTDI] = 1323 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M; 1324 1325 pmap->pm_active = 0; 1326 pmap->pm_ptphint = NULL; 1327 TAILQ_INIT(&pmap->pm_pvlist); 1328 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1329} 1330 1331/* 1332 * Wire in kernel global address entries. To avoid a race condition 1333 * between pmap initialization and pmap_growkernel, this procedure 1334 * should be called after the vmspace is attached to the process 1335 * but before this pmap is activated. 1336 */ 1337void 1338pmap_pinit2(pmap) 1339 struct pmap *pmap; 1340{ 1341 /* XXX: Remove this stub when no longer called */ 1342} 1343 1344static int 1345pmap_release_free_page(pmap_t pmap, vm_page_t p) 1346{ 1347 pd_entry_t *pde = pmap->pm_pdir; 1348 1349 /* 1350 * This code optimizes the case of freeing non-busy 1351 * page-table pages. Those pages are zero now, and 1352 * might as well be placed directly into the zero queue. 1353 */ 1354 vm_page_lock_queues(); 1355 if (vm_page_sleep_if_busy(p, FALSE, "pmaprl")) 1356 return (0); 1357 vm_page_busy(p); 1358 1359 /* 1360 * Remove the page table page from the processes address space. 1361 */ 1362 pde[p->pindex] = 0; 1363 pmap->pm_stats.resident_count--; 1364 1365 if (p->hold_count) { 1366 panic("pmap_release: freeing held page table page"); 1367 } 1368 /* 1369 * Page directory pages need to have the kernel 1370 * stuff cleared, so they can go into the zero queue also. 1371 */ 1372 if (p->pindex == PTDPTDI) { 1373 bzero(pde + KPTDI, nkpt * PTESIZE); 1374#ifdef SMP 1375 pde[MPPTDI] = 0; 1376#endif 1377 pde[APTDPTDI] = 0; 1378 pmap_kremove((vm_offset_t) pmap->pm_pdir); 1379 } 1380 1381 if (pmap->pm_ptphint == p) 1382 pmap->pm_ptphint = NULL; 1383 1384 p->wire_count--; 1385 atomic_subtract_int(&cnt.v_wire_count, 1); 1386 vm_page_free_zero(p); 1387 vm_page_unlock_queues(); 1388 return 1; 1389} 1390 1391/* 1392 * this routine is called if the page table page is not 1393 * mapped correctly. 1394 */ 1395static vm_page_t 1396_pmap_allocpte(pmap, ptepindex) 1397 pmap_t pmap; 1398 unsigned ptepindex; 1399{ 1400 vm_offset_t pteva, ptepa; /* XXXPA */ 1401 vm_page_t m; 1402 1403 /* 1404 * Find or fabricate a new pagetable page 1405 */ 1406 m = vm_page_grab(pmap->pm_pteobj, ptepindex, 1407 VM_ALLOC_WIRED | VM_ALLOC_ZERO | VM_ALLOC_RETRY); 1408 1409 KASSERT(m->queue == PQ_NONE, 1410 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1411 1412 /* 1413 * Increment the hold count for the page table page 1414 * (denoting a new mapping.) 1415 */ 1416 m->hold_count++; 1417 1418 /* 1419 * Map the pagetable page into the process address space, if 1420 * it isn't already there. 1421 */ 1422 1423 pmap->pm_stats.resident_count++; 1424 1425 ptepa = VM_PAGE_TO_PHYS(m); 1426 pmap->pm_pdir[ptepindex] = 1427 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1428 1429 /* 1430 * Set the page table hint 1431 */ 1432 pmap->pm_ptphint = m; 1433 1434 /* 1435 * Try to use the new mapping, but if we cannot, then 1436 * do it with the routine that maps the page explicitly. 1437 */ 1438 if ((m->flags & PG_ZERO) == 0) { 1439 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == 1440 (PTDpde & PG_FRAME)) { 1441 pteva = VM_MAXUSER_ADDRESS + i386_ptob(ptepindex); 1442 bzero((caddr_t) pteva, PAGE_SIZE); 1443 } else { 1444 pmap_zero_page(m); 1445 } 1446 } 1447 vm_page_lock_queues(); 1448 m->valid = VM_PAGE_BITS_ALL; 1449 vm_page_flag_clear(m, PG_ZERO); 1450 vm_page_wakeup(m); 1451 vm_page_unlock_queues(); 1452 1453 return m; 1454} 1455 1456static vm_page_t 1457pmap_allocpte(pmap_t pmap, vm_offset_t va) 1458{ 1459 unsigned ptepindex; 1460 pd_entry_t ptepa; 1461 vm_page_t m; 1462 1463 /* 1464 * Calculate pagetable page index 1465 */ 1466 ptepindex = va >> PDRSHIFT; 1467 1468 /* 1469 * Get the page directory entry 1470 */ 1471 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex]; 1472 1473 /* 1474 * This supports switching from a 4MB page to a 1475 * normal 4K page. 1476 */ 1477 if (ptepa & PG_PS) { 1478 pmap->pm_pdir[ptepindex] = 0; 1479 ptepa = 0; 1480 pmap_invalidate_all(kernel_pmap); 1481 } 1482 1483 /* 1484 * If the page table page is mapped, we just increment the 1485 * hold count, and activate it. 1486 */ 1487 if (ptepa) { 1488 /* 1489 * In order to get the page table page, try the 1490 * hint first. 1491 */ 1492 if (pmap->pm_ptphint && 1493 (pmap->pm_ptphint->pindex == ptepindex)) { 1494 m = pmap->pm_ptphint; 1495 } else { 1496 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 1497 pmap->pm_ptphint = m; 1498 } 1499 m->hold_count++; 1500 return m; 1501 } 1502 /* 1503 * Here if the pte page isn't mapped, or if it has been deallocated. 1504 */ 1505 return _pmap_allocpte(pmap, ptepindex); 1506} 1507 1508 1509/*************************************************** 1510* Pmap allocation/deallocation routines. 1511 ***************************************************/ 1512 1513/* 1514 * Release any resources held by the given physical map. 1515 * Called when a pmap initialized by pmap_pinit is being released. 1516 * Should only be called if the map contains no valid mappings. 1517 */ 1518void 1519pmap_release(pmap_t pmap) 1520{ 1521 vm_page_t p,n,ptdpg; 1522 vm_object_t object = pmap->pm_pteobj; 1523 int curgeneration; 1524 1525#if defined(DIAGNOSTIC) 1526 if (object->ref_count != 1) 1527 panic("pmap_release: pteobj reference count != 1"); 1528#endif 1529 1530 ptdpg = NULL; 1531 mtx_lock_spin(&allpmaps_lock); 1532 LIST_REMOVE(pmap, pm_list); 1533 mtx_unlock_spin(&allpmaps_lock); 1534retry: 1535 curgeneration = object->generation; 1536 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = n) { 1537 n = TAILQ_NEXT(p, listq); 1538 if (p->pindex == PTDPTDI) { 1539 ptdpg = p; 1540 continue; 1541 } 1542 while (1) { 1543 if (!pmap_release_free_page(pmap, p) && 1544 (object->generation != curgeneration)) 1545 goto retry; 1546 } 1547 } 1548 1549 if (ptdpg && !pmap_release_free_page(pmap, ptdpg)) 1550 goto retry; 1551} 1552 1553static int 1554kvm_size(SYSCTL_HANDLER_ARGS) 1555{ 1556 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1557 1558 return sysctl_handle_long(oidp, &ksize, 0, req); 1559} 1560SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1561 0, 0, kvm_size, "IU", "Size of KVM"); 1562 1563static int 1564kvm_free(SYSCTL_HANDLER_ARGS) 1565{ 1566 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1567 1568 return sysctl_handle_long(oidp, &kfree, 0, req); 1569} 1570SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1571 0, 0, kvm_free, "IU", "Amount of KVM free"); 1572 1573/* 1574 * grow the number of kernel page table entries, if needed 1575 */ 1576void 1577pmap_growkernel(vm_offset_t addr) 1578{ 1579 struct pmap *pmap; 1580 int s; 1581 vm_offset_t ptppaddr; 1582 vm_page_t nkpg; 1583 pd_entry_t newpdir; 1584 1585 s = splhigh(); 1586 if (kernel_vm_end == 0) { 1587 kernel_vm_end = KERNBASE; 1588 nkpt = 0; 1589 while (pdir_pde(PTD, kernel_vm_end)) { 1590 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1591 nkpt++; 1592 } 1593 } 1594 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1595 while (kernel_vm_end < addr) { 1596 if (pdir_pde(PTD, kernel_vm_end)) { 1597 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1598 continue; 1599 } 1600 1601 /* 1602 * This index is bogus, but out of the way 1603 */ 1604 nkpg = vm_page_alloc(NULL, nkpt, 1605 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1606 if (!nkpg) 1607 panic("pmap_growkernel: no memory to grow kernel"); 1608 1609 nkpt++; 1610 1611 pmap_zero_page(nkpg); 1612 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1613 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1614 pdir_pde(PTD, kernel_vm_end) = newpdir; 1615 1616 mtx_lock_spin(&allpmaps_lock); 1617 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1618 *pmap_pde(pmap, kernel_vm_end) = newpdir; 1619 } 1620 mtx_unlock_spin(&allpmaps_lock); 1621 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1622 } 1623 splx(s); 1624} 1625 1626 1627/*************************************************** 1628 * page management routines. 1629 ***************************************************/ 1630 1631/* 1632 * free the pv_entry back to the free list 1633 */ 1634static PMAP_INLINE void 1635free_pv_entry(pv_entry_t pv) 1636{ 1637 pv_entry_count--; 1638 uma_zfree(pvzone, pv); 1639} 1640 1641/* 1642 * get a new pv_entry, allocating a block from the system 1643 * when needed. 1644 * the memory allocation is performed bypassing the malloc code 1645 * because of the possibility of allocations at interrupt time. 1646 */ 1647static pv_entry_t 1648get_pv_entry(void) 1649{ 1650 pv_entry_count++; 1651 if (pv_entry_high_water && 1652 (pv_entry_count > pv_entry_high_water) && 1653 (pmap_pagedaemon_waken == 0)) { 1654 pmap_pagedaemon_waken = 1; 1655 wakeup (&vm_pages_needed); 1656 } 1657 return uma_zalloc(pvzone, M_NOWAIT); 1658} 1659 1660/* 1661 * If it is the first entry on the list, it is actually 1662 * in the header and we must copy the following entry up 1663 * to the header. Otherwise we must search the list for 1664 * the entry. In either case we free the now unused entry. 1665 */ 1666 1667static int 1668pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1669{ 1670 pv_entry_t pv; 1671 int rtval; 1672 int s; 1673 1674 s = splvm(); 1675 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1676 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1677 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1678 if (pmap == pv->pv_pmap && va == pv->pv_va) 1679 break; 1680 } 1681 } else { 1682 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1683 if (va == pv->pv_va) 1684 break; 1685 } 1686 } 1687 1688 rtval = 0; 1689 if (pv) { 1690 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1691 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1692 m->md.pv_list_count--; 1693 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1694 vm_page_flag_clear(m, PG_WRITEABLE); 1695 1696 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1697 free_pv_entry(pv); 1698 } 1699 1700 splx(s); 1701 return rtval; 1702} 1703 1704/* 1705 * Create a pv entry for page at pa for 1706 * (pmap, va). 1707 */ 1708static void 1709pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1710{ 1711 1712 int s; 1713 pv_entry_t pv; 1714 1715 s = splvm(); 1716 pv = get_pv_entry(); 1717 pv->pv_va = va; 1718 pv->pv_pmap = pmap; 1719 pv->pv_ptem = mpte; 1720 1721 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1722 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1723 m->md.pv_list_count++; 1724 1725 splx(s); 1726} 1727 1728/* 1729 * pmap_remove_pte: do the things to unmap a page in a process 1730 */ 1731static int 1732pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1733{ 1734 pt_entry_t oldpte; 1735 vm_page_t m; 1736 1737 oldpte = atomic_readandclear_int(ptq); 1738 if (oldpte & PG_W) 1739 pmap->pm_stats.wired_count -= 1; 1740 /* 1741 * Machines that don't support invlpg, also don't support 1742 * PG_G. 1743 */ 1744 if (oldpte & PG_G) 1745 pmap_invalidate_page(kernel_pmap, va); 1746 pmap->pm_stats.resident_count -= 1; 1747 if (oldpte & PG_MANAGED) { 1748 m = PHYS_TO_VM_PAGE(oldpte); 1749 if (oldpte & PG_M) { 1750#if defined(PMAP_DIAGNOSTIC) 1751 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1752 printf( 1753 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1754 va, oldpte); 1755 } 1756#endif 1757 if (pmap_track_modified(va)) 1758 vm_page_dirty(m); 1759 } 1760 if (oldpte & PG_A) 1761 vm_page_flag_set(m, PG_REFERENCED); 1762 return pmap_remove_entry(pmap, m, va); 1763 } else { 1764 return pmap_unuse_pt(pmap, va, NULL); 1765 } 1766 1767 return 0; 1768} 1769 1770/* 1771 * Remove a single page from a process address space 1772 */ 1773static void 1774pmap_remove_page(pmap_t pmap, vm_offset_t va) 1775{ 1776 register pt_entry_t *ptq; 1777 1778 /* 1779 * if there is no pte for this address, just skip it!!! 1780 */ 1781 if (*pmap_pde(pmap, va) == 0) { 1782 return; 1783 } 1784 1785 /* 1786 * get a local va for mappings for this pmap. 1787 */ 1788 ptq = get_ptbase(pmap) + i386_btop(va); 1789 if (*ptq) { 1790 (void) pmap_remove_pte(pmap, ptq, va); 1791 pmap_invalidate_page(pmap, va); 1792 } 1793 return; 1794} 1795 1796/* 1797 * Remove the given range of addresses from the specified map. 1798 * 1799 * It is assumed that the start and end are properly 1800 * rounded to the page size. 1801 */ 1802void 1803pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1804{ 1805 register pt_entry_t *ptbase; 1806 vm_offset_t pdnxt; 1807 pd_entry_t ptpaddr; 1808 vm_offset_t sindex, eindex; 1809 int anyvalid; 1810 1811 if (pmap == NULL) 1812 return; 1813 1814 if (pmap->pm_stats.resident_count == 0) 1815 return; 1816 1817 /* 1818 * special handling of removing one page. a very 1819 * common operation and easy to short circuit some 1820 * code. 1821 */ 1822 if ((sva + PAGE_SIZE == eva) && 1823 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1824 pmap_remove_page(pmap, sva); 1825 return; 1826 } 1827 1828 anyvalid = 0; 1829 1830 /* 1831 * Get a local virtual address for the mappings that are being 1832 * worked with. 1833 */ 1834 ptbase = get_ptbase(pmap); 1835 1836 sindex = i386_btop(sva); 1837 eindex = i386_btop(eva); 1838 1839 for (; sindex < eindex; sindex = pdnxt) { 1840 unsigned pdirindex; 1841 1842 /* 1843 * Calculate index for next page table. 1844 */ 1845 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1)); 1846 if (pmap->pm_stats.resident_count == 0) 1847 break; 1848 1849 pdirindex = sindex / NPDEPG; 1850 ptpaddr = pmap->pm_pdir[pdirindex]; 1851 if ((ptpaddr & PG_PS) != 0) { 1852 pmap->pm_pdir[pdirindex] = 0; 1853 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1854 anyvalid++; 1855 continue; 1856 } 1857 1858 /* 1859 * Weed out invalid mappings. Note: we assume that the page 1860 * directory table is always allocated, and in kernel virtual. 1861 */ 1862 if (ptpaddr == 0) 1863 continue; 1864 1865 /* 1866 * Limit our scan to either the end of the va represented 1867 * by the current page table page, or to the end of the 1868 * range being removed. 1869 */ 1870 if (pdnxt > eindex) { 1871 pdnxt = eindex; 1872 } 1873 1874 for (; sindex != pdnxt; sindex++) { 1875 vm_offset_t va; 1876 if (ptbase[sindex] == 0) { 1877 continue; 1878 } 1879 va = i386_ptob(sindex); 1880 1881 anyvalid++; 1882 if (pmap_remove_pte(pmap, ptbase + sindex, va)) 1883 break; 1884 } 1885 } 1886 1887 if (anyvalid) 1888 pmap_invalidate_all(pmap); 1889} 1890 1891/* 1892 * Routine: pmap_remove_all 1893 * Function: 1894 * Removes this physical page from 1895 * all physical maps in which it resides. 1896 * Reflects back modify bits to the pager. 1897 * 1898 * Notes: 1899 * Original versions of this routine were very 1900 * inefficient because they iteratively called 1901 * pmap_remove (slow...) 1902 */ 1903 1904void 1905pmap_remove_all(vm_page_t m) 1906{ 1907 register pv_entry_t pv; 1908 pt_entry_t *pte, tpte; 1909 int s; 1910 1911#if defined(PMAP_DIAGNOSTIC) 1912 /* 1913 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1914 */ 1915 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1916 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1917 VM_PAGE_TO_PHYS(m)); 1918 } 1919#endif 1920 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1921 s = splvm(); 1922 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1923 pv->pv_pmap->pm_stats.resident_count--; 1924 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1925 tpte = atomic_readandclear_int(pte); 1926 if (tpte & PG_W) 1927 pv->pv_pmap->pm_stats.wired_count--; 1928 if (tpte & PG_A) 1929 vm_page_flag_set(m, PG_REFERENCED); 1930 1931 /* 1932 * Update the vm_page_t clean and reference bits. 1933 */ 1934 if (tpte & PG_M) { 1935#if defined(PMAP_DIAGNOSTIC) 1936 if (pmap_nw_modified((pt_entry_t) tpte)) { 1937 printf( 1938 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1939 pv->pv_va, tpte); 1940 } 1941#endif 1942 if (pmap_track_modified(pv->pv_va)) 1943 vm_page_dirty(m); 1944 } 1945 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1946 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1947 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1948 m->md.pv_list_count--; 1949 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 1950 free_pv_entry(pv); 1951 } 1952 vm_page_flag_clear(m, PG_WRITEABLE); 1953 splx(s); 1954} 1955 1956/* 1957 * Set the physical protection on the 1958 * specified range of this map as requested. 1959 */ 1960void 1961pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1962{ 1963 register pt_entry_t *ptbase; 1964 vm_offset_t pdnxt; 1965 pd_entry_t ptpaddr; 1966 vm_offset_t sindex, eindex; 1967 int anychanged; 1968 1969 if (pmap == NULL) 1970 return; 1971 1972 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1973 pmap_remove(pmap, sva, eva); 1974 return; 1975 } 1976 1977 if (prot & VM_PROT_WRITE) 1978 return; 1979 1980 anychanged = 0; 1981 1982 ptbase = get_ptbase(pmap); 1983 1984 sindex = i386_btop(sva); 1985 eindex = i386_btop(eva); 1986 1987 for (; sindex < eindex; sindex = pdnxt) { 1988 1989 unsigned pdirindex; 1990 1991 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1)); 1992 1993 pdirindex = sindex / NPDEPG; 1994 ptpaddr = pmap->pm_pdir[pdirindex]; 1995 if ((ptpaddr & PG_PS) != 0) { 1996 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1997 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1998 anychanged++; 1999 continue; 2000 } 2001 2002 /* 2003 * Weed out invalid mappings. Note: we assume that the page 2004 * directory table is always allocated, and in kernel virtual. 2005 */ 2006 if (ptpaddr == 0) 2007 continue; 2008 2009 if (pdnxt > eindex) { 2010 pdnxt = eindex; 2011 } 2012 2013 for (; sindex != pdnxt; sindex++) { 2014 2015 pt_entry_t pbits; 2016 vm_page_t m; 2017 2018 pbits = ptbase[sindex]; 2019 2020 if (pbits & PG_MANAGED) { 2021 m = NULL; 2022 if (pbits & PG_A) { 2023 m = PHYS_TO_VM_PAGE(pbits); 2024 vm_page_flag_set(m, PG_REFERENCED); 2025 pbits &= ~PG_A; 2026 } 2027 if (pbits & PG_M) { 2028 if (pmap_track_modified(i386_ptob(sindex))) { 2029 if (m == NULL) 2030 m = PHYS_TO_VM_PAGE(pbits); 2031 vm_page_dirty(m); 2032 pbits &= ~PG_M; 2033 } 2034 } 2035 } 2036 2037 pbits &= ~PG_RW; 2038 2039 if (pbits != ptbase[sindex]) { 2040 ptbase[sindex] = pbits; 2041 anychanged = 1; 2042 } 2043 } 2044 } 2045 if (anychanged) 2046 pmap_invalidate_all(pmap); 2047} 2048 2049/* 2050 * Insert the given physical page (p) at 2051 * the specified virtual address (v) in the 2052 * target physical map with the protection requested. 2053 * 2054 * If specified, the page will be wired down, meaning 2055 * that the related pte can not be reclaimed. 2056 * 2057 * NB: This is the only routine which MAY NOT lazy-evaluate 2058 * or lose information. That is, this routine must actually 2059 * insert this page into the given map NOW. 2060 */ 2061void 2062pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 2063 boolean_t wired) 2064{ 2065 vm_offset_t pa; 2066 register pt_entry_t *pte; 2067 vm_offset_t opa; 2068 pt_entry_t origpte, newpte; 2069 vm_page_t mpte; 2070 2071 if (pmap == NULL) 2072 return; 2073 2074 va &= PG_FRAME; 2075#ifdef PMAP_DIAGNOSTIC 2076 if (va > VM_MAX_KERNEL_ADDRESS) 2077 panic("pmap_enter: toobig"); 2078 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 2079 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 2080#endif 2081 2082 mpte = NULL; 2083 /* 2084 * In the case that a page table page is not 2085 * resident, we are creating it here. 2086 */ 2087 if (va < VM_MAXUSER_ADDRESS) { 2088 mpte = pmap_allocpte(pmap, va); 2089 } 2090#if 0 && defined(PMAP_DIAGNOSTIC) 2091 else { 2092 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 2093 origpte = *pdeaddr; 2094 if ((origpte & PG_V) == 0) { 2095 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 2096 pmap->pm_pdir[PTDPTDI], origpte, va); 2097 } 2098 } 2099#endif 2100 2101 pte = pmap_pte(pmap, va); 2102 2103 /* 2104 * Page Directory table entry not valid, we need a new PT page 2105 */ 2106 if (pte == NULL) { 2107 panic("pmap_enter: invalid page directory, pdir=%p, va=0x%x\n", 2108 (void *)pmap->pm_pdir[PTDPTDI], va); 2109 } 2110 2111 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 2112 origpte = *(vm_offset_t *)pte; 2113 opa = origpte & PG_FRAME; 2114 2115 if (origpte & PG_PS) 2116 panic("pmap_enter: attempted pmap_enter on 4MB page"); 2117 2118 /* 2119 * Mapping has not changed, must be protection or wiring change. 2120 */ 2121 if (origpte && (opa == pa)) { 2122 /* 2123 * Wiring change, just update stats. We don't worry about 2124 * wiring PT pages as they remain resident as long as there 2125 * are valid mappings in them. Hence, if a user page is wired, 2126 * the PT page will be also. 2127 */ 2128 if (wired && ((origpte & PG_W) == 0)) 2129 pmap->pm_stats.wired_count++; 2130 else if (!wired && (origpte & PG_W)) 2131 pmap->pm_stats.wired_count--; 2132 2133#if defined(PMAP_DIAGNOSTIC) 2134 if (pmap_nw_modified((pt_entry_t) origpte)) { 2135 printf( 2136 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 2137 va, origpte); 2138 } 2139#endif 2140 2141 /* 2142 * Remove extra pte reference 2143 */ 2144 if (mpte) 2145 mpte->hold_count--; 2146 2147 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) { 2148 if ((origpte & PG_RW) == 0) { 2149 *pte |= PG_RW; 2150 pmap_invalidate_page(pmap, va); 2151 } 2152 return; 2153 } 2154 2155 /* 2156 * We might be turning off write access to the page, 2157 * so we go ahead and sense modify status. 2158 */ 2159 if (origpte & PG_MANAGED) { 2160 if ((origpte & PG_M) && pmap_track_modified(va)) { 2161 vm_page_t om; 2162 om = PHYS_TO_VM_PAGE(opa); 2163 vm_page_dirty(om); 2164 } 2165 pa |= PG_MANAGED; 2166 } 2167 goto validate; 2168 } 2169 /* 2170 * Mapping has changed, invalidate old range and fall through to 2171 * handle validating new mapping. 2172 */ 2173 if (opa) { 2174 int err; 2175 vm_page_lock_queues(); 2176 err = pmap_remove_pte(pmap, pte, va); 2177 vm_page_unlock_queues(); 2178 if (err) 2179 panic("pmap_enter: pte vanished, va: 0x%x", va); 2180 } 2181 2182 /* 2183 * Enter on the PV list if part of our managed memory. Note that we 2184 * raise IPL while manipulating pv_table since pmap_enter can be 2185 * called at interrupt time. 2186 */ 2187 if (pmap_initialized && 2188 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2189 pmap_insert_entry(pmap, va, mpte, m); 2190 pa |= PG_MANAGED; 2191 } 2192 2193 /* 2194 * Increment counters 2195 */ 2196 pmap->pm_stats.resident_count++; 2197 if (wired) 2198 pmap->pm_stats.wired_count++; 2199 2200validate: 2201 /* 2202 * Now validate mapping with desired protection/wiring. 2203 */ 2204 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V); 2205 2206 if (wired) 2207 newpte |= PG_W; 2208 if (va < VM_MAXUSER_ADDRESS) 2209 newpte |= PG_U; 2210 if (pmap == kernel_pmap) 2211 newpte |= pgeflag; 2212 2213 /* 2214 * if the mapping or permission bits are different, we need 2215 * to update the pte. 2216 */ 2217 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2218 *pte = newpte | PG_A; 2219 /*if (origpte)*/ { 2220 pmap_invalidate_page(pmap, va); 2221 } 2222 } 2223} 2224 2225/* 2226 * this code makes some *MAJOR* assumptions: 2227 * 1. Current pmap & pmap exists. 2228 * 2. Not wired. 2229 * 3. Read access. 2230 * 4. No page table pages. 2231 * 5. Tlbflush is deferred to calling procedure. 2232 * 6. Page IS managed. 2233 * but is *MUCH* faster than pmap_enter... 2234 */ 2235 2236static vm_page_t 2237pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2238{ 2239 pt_entry_t *pte; 2240 vm_offset_t pa; 2241 2242 /* 2243 * In the case that a page table page is not 2244 * resident, we are creating it here. 2245 */ 2246 if (va < VM_MAXUSER_ADDRESS) { 2247 unsigned ptepindex; 2248 pd_entry_t ptepa; 2249 2250 /* 2251 * Calculate pagetable page index 2252 */ 2253 ptepindex = va >> PDRSHIFT; 2254 if (mpte && (mpte->pindex == ptepindex)) { 2255 mpte->hold_count++; 2256 } else { 2257retry: 2258 /* 2259 * Get the page directory entry 2260 */ 2261 ptepa = pmap->pm_pdir[ptepindex]; 2262 2263 /* 2264 * If the page table page is mapped, we just increment 2265 * the hold count, and activate it. 2266 */ 2267 if (ptepa) { 2268 if (ptepa & PG_PS) 2269 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2270 if (pmap->pm_ptphint && 2271 (pmap->pm_ptphint->pindex == ptepindex)) { 2272 mpte = pmap->pm_ptphint; 2273 } else { 2274 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 2275 pmap->pm_ptphint = mpte; 2276 } 2277 if (mpte == NULL) 2278 goto retry; 2279 mpte->hold_count++; 2280 } else { 2281 mpte = _pmap_allocpte(pmap, ptepindex); 2282 } 2283 } 2284 } else { 2285 mpte = NULL; 2286 } 2287 2288 /* 2289 * This call to vtopte makes the assumption that we are 2290 * entering the page into the current pmap. In order to support 2291 * quick entry into any pmap, one would likely use pmap_pte_quick. 2292 * But that isn't as quick as vtopte. 2293 */ 2294 pte = vtopte(va); 2295 if (*pte) { 2296 if (mpte != NULL) { 2297 vm_page_lock_queues(); 2298 pmap_unwire_pte_hold(pmap, mpte); 2299 vm_page_unlock_queues(); 2300 } 2301 return 0; 2302 } 2303 2304 /* 2305 * Enter on the PV list if part of our managed memory. Note that we 2306 * raise IPL while manipulating pv_table since pmap_enter can be 2307 * called at interrupt time. 2308 */ 2309 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2310 pmap_insert_entry(pmap, va, mpte, m); 2311 2312 /* 2313 * Increment counters 2314 */ 2315 pmap->pm_stats.resident_count++; 2316 2317 pa = VM_PAGE_TO_PHYS(m); 2318 2319 /* 2320 * Now validate mapping with RO protection 2321 */ 2322 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2323 *pte = pa | PG_V | PG_U; 2324 else 2325 *pte = pa | PG_V | PG_U | PG_MANAGED; 2326 2327 return mpte; 2328} 2329 2330/* 2331 * Make a temporary mapping for a physical address. This is only intended 2332 * to be used for panic dumps. 2333 */ 2334void * 2335pmap_kenter_temporary(vm_offset_t pa, int i) 2336{ 2337 vm_offset_t va; 2338 2339 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2340 pmap_kenter(va, pa); 2341#ifndef I386_CPU 2342 invlpg(va); 2343#else 2344 invltlb(); 2345#endif 2346 return ((void *)crashdumpmap); 2347} 2348 2349#define MAX_INIT_PT (96) 2350/* 2351 * pmap_object_init_pt preloads the ptes for a given object 2352 * into the specified pmap. This eliminates the blast of soft 2353 * faults on process startup and immediately after an mmap. 2354 */ 2355void 2356pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2357 vm_object_t object, vm_pindex_t pindex, 2358 vm_size_t size, int limit) 2359{ 2360 vm_offset_t tmpidx; 2361 int psize; 2362 vm_page_t p, mpte; 2363 2364 if (pmap == NULL || object == NULL) 2365 return; 2366 2367 /* 2368 * This code maps large physical mmap regions into the 2369 * processor address space. Note that some shortcuts 2370 * are taken, but the code works. 2371 */ 2372 if (pseflag && (object->type == OBJT_DEVICE) && 2373 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2374 int i; 2375 vm_page_t m[1]; 2376 unsigned int ptepindex; 2377 int npdes; 2378 pd_entry_t ptepa; 2379 2380 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2381 return; 2382 2383retry: 2384 p = vm_page_lookup(object, pindex); 2385 if (p != NULL) { 2386 vm_page_lock_queues(); 2387 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2388 goto retry; 2389 } else { 2390 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2391 if (p == NULL) 2392 return; 2393 m[0] = p; 2394 2395 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2396 vm_page_lock_queues(); 2397 vm_page_free(p); 2398 vm_page_unlock_queues(); 2399 return; 2400 } 2401 2402 p = vm_page_lookup(object, pindex); 2403 vm_page_lock_queues(); 2404 vm_page_wakeup(p); 2405 } 2406 vm_page_unlock_queues(); 2407 2408 ptepa = VM_PAGE_TO_PHYS(p); 2409 if (ptepa & (NBPDR - 1)) { 2410 return; 2411 } 2412 2413 p->valid = VM_PAGE_BITS_ALL; 2414 2415 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2416 npdes = size >> PDRSHIFT; 2417 for(i = 0; i < npdes; i++) { 2418 pmap->pm_pdir[ptepindex] = 2419 ptepa | PG_U | PG_RW | PG_V | PG_PS; 2420 ptepa += NBPDR; 2421 ptepindex += 1; 2422 } 2423 pmap_invalidate_all(kernel_pmap); 2424 return; 2425 } 2426 2427 psize = i386_btop(size); 2428 2429 if ((object->type != OBJT_VNODE) || 2430 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) && 2431 (object->resident_page_count > MAX_INIT_PT))) { 2432 return; 2433 } 2434 2435 if (psize + pindex > object->size) { 2436 if (object->size < pindex) 2437 return; 2438 psize = object->size - pindex; 2439 } 2440 2441 mpte = NULL; 2442 2443 if ((p = TAILQ_FIRST(&object->memq)) != NULL) { 2444 if (p->pindex < pindex) { 2445 p = vm_page_splay(pindex, object->root); 2446 if ((object->root = p)->pindex < pindex) 2447 p = TAILQ_NEXT(p, listq); 2448 } 2449 } 2450 /* 2451 * Assert: the variable p is either (1) the page with the 2452 * least pindex greater than or equal to the parameter pindex 2453 * or (2) NULL. 2454 */ 2455 for (; 2456 p != NULL && (tmpidx = p->pindex - pindex) < psize; 2457 p = TAILQ_NEXT(p, listq)) { 2458 /* 2459 * don't allow an madvise to blow away our really 2460 * free pages allocating pv entries. 2461 */ 2462 if ((limit & MAP_PREFAULT_MADVISE) && 2463 cnt.v_free_count < cnt.v_free_reserved) { 2464 break; 2465 } 2466 vm_page_lock_queues(); 2467 if ((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL && 2468 (p->busy == 0) && 2469 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2470 if ((p->queue - p->pc) == PQ_CACHE) 2471 vm_page_deactivate(p); 2472 vm_page_busy(p); 2473 vm_page_unlock_queues(); 2474 mpte = pmap_enter_quick(pmap, 2475 addr + i386_ptob(tmpidx), p, mpte); 2476 vm_page_lock_queues(); 2477 vm_page_wakeup(p); 2478 } 2479 vm_page_unlock_queues(); 2480 } 2481 return; 2482} 2483 2484/* 2485 * pmap_prefault provides a quick way of clustering 2486 * pagefaults into a processes address space. It is a "cousin" 2487 * of pmap_object_init_pt, except it runs at page fault time instead 2488 * of mmap time. 2489 */ 2490#define PFBAK 4 2491#define PFFOR 4 2492#define PAGEORDER_SIZE (PFBAK+PFFOR) 2493 2494static int pmap_prefault_pageorder[] = { 2495 -1 * PAGE_SIZE, 1 * PAGE_SIZE, 2496 -2 * PAGE_SIZE, 2 * PAGE_SIZE, 2497 -3 * PAGE_SIZE, 3 * PAGE_SIZE, 2498 -4 * PAGE_SIZE, 4 * PAGE_SIZE 2499}; 2500 2501void 2502pmap_prefault(pmap, addra, entry) 2503 pmap_t pmap; 2504 vm_offset_t addra; 2505 vm_map_entry_t entry; 2506{ 2507 int i; 2508 vm_offset_t starta; 2509 vm_offset_t addr; 2510 vm_pindex_t pindex; 2511 vm_page_t m, mpte; 2512 vm_object_t object; 2513 2514 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) 2515 return; 2516 2517 object = entry->object.vm_object; 2518 2519 starta = addra - PFBAK * PAGE_SIZE; 2520 if (starta < entry->start) { 2521 starta = entry->start; 2522 } else if (starta > addra) { 2523 starta = 0; 2524 } 2525 2526 mpte = NULL; 2527 for (i = 0; i < PAGEORDER_SIZE; i++) { 2528 vm_object_t lobject; 2529 pt_entry_t *pte; 2530 2531 addr = addra + pmap_prefault_pageorder[i]; 2532 if (addr > addra + (PFFOR * PAGE_SIZE)) 2533 addr = 0; 2534 2535 if (addr < starta || addr >= entry->end) 2536 continue; 2537 2538 if ((*pmap_pde(pmap, addr)) == 0) 2539 continue; 2540 2541 pte = vtopte(addr); 2542 if (*pte) 2543 continue; 2544 2545 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2546 lobject = object; 2547 for (m = vm_page_lookup(lobject, pindex); 2548 (!m && (lobject->type == OBJT_DEFAULT) && (lobject->backing_object)); 2549 lobject = lobject->backing_object) { 2550 if (lobject->backing_object_offset & PAGE_MASK) 2551 break; 2552 pindex += (lobject->backing_object_offset >> PAGE_SHIFT); 2553 m = vm_page_lookup(lobject->backing_object, pindex); 2554 } 2555 2556 /* 2557 * give-up when a page is not in memory 2558 */ 2559 if (m == NULL) 2560 break; 2561 vm_page_lock_queues(); 2562 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2563 (m->busy == 0) && 2564 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2565 2566 if ((m->queue - m->pc) == PQ_CACHE) { 2567 vm_page_deactivate(m); 2568 } 2569 vm_page_busy(m); 2570 vm_page_unlock_queues(); 2571 mpte = pmap_enter_quick(pmap, addr, m, mpte); 2572 vm_page_lock_queues(); 2573 vm_page_wakeup(m); 2574 } 2575 vm_page_unlock_queues(); 2576 } 2577} 2578 2579/* 2580 * Routine: pmap_change_wiring 2581 * Function: Change the wiring attribute for a map/virtual-address 2582 * pair. 2583 * In/out conditions: 2584 * The mapping must already exist in the pmap. 2585 */ 2586void 2587pmap_change_wiring(pmap, va, wired) 2588 register pmap_t pmap; 2589 vm_offset_t va; 2590 boolean_t wired; 2591{ 2592 register pt_entry_t *pte; 2593 2594 if (pmap == NULL) 2595 return; 2596 2597 pte = pmap_pte(pmap, va); 2598 2599 if (wired && !pmap_pte_w(pte)) 2600 pmap->pm_stats.wired_count++; 2601 else if (!wired && pmap_pte_w(pte)) 2602 pmap->pm_stats.wired_count--; 2603 2604 /* 2605 * Wiring is not a hardware characteristic so there is no need to 2606 * invalidate TLB. 2607 */ 2608 pmap_pte_set_w(pte, wired); 2609} 2610 2611 2612 2613/* 2614 * Copy the range specified by src_addr/len 2615 * from the source map to the range dst_addr/len 2616 * in the destination map. 2617 * 2618 * This routine is only advisory and need not do anything. 2619 */ 2620 2621void 2622pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2623 vm_offset_t src_addr) 2624{ 2625 vm_offset_t addr; 2626 vm_offset_t end_addr = src_addr + len; 2627 vm_offset_t pdnxt; 2628 pd_entry_t src_frame, dst_frame; 2629 vm_page_t m; 2630 2631 if (dst_addr != src_addr) 2632 return; 2633 2634 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME; 2635 if (src_frame != (PTDpde & PG_FRAME)) 2636 return; 2637 2638 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME; 2639 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2640 pt_entry_t *src_pte, *dst_pte; 2641 vm_page_t dstmpte, srcmpte; 2642 pd_entry_t srcptepaddr; 2643 unsigned ptepindex; 2644 2645 if (addr >= UPT_MIN_ADDRESS) 2646 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2647 2648 /* 2649 * Don't let optional prefaulting of pages make us go 2650 * way below the low water mark of free pages or way 2651 * above high water mark of used pv entries. 2652 */ 2653 if (cnt.v_free_count < cnt.v_free_reserved || 2654 pv_entry_count > pv_entry_high_water) 2655 break; 2656 2657 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1)); 2658 ptepindex = addr >> PDRSHIFT; 2659 2660 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2661 if (srcptepaddr == 0) 2662 continue; 2663 2664 if (srcptepaddr & PG_PS) { 2665 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2666 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2667 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; 2668 } 2669 continue; 2670 } 2671 2672 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex); 2673 if ((srcmpte == NULL) || 2674 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY)) 2675 continue; 2676 2677 if (pdnxt > end_addr) 2678 pdnxt = end_addr; 2679 2680 /* 2681 * Have to recheck this before every avtopte() call below 2682 * in case we have blocked and something else used APTDpde. 2683 */ 2684 if (dst_frame != (APTDpde & PG_FRAME)) { 2685 APTDpde = dst_frame | PG_RW | PG_V; 2686 pmap_invalidate_all(kernel_pmap); /* XXX Bandaid */ 2687 } 2688 src_pte = vtopte(addr); 2689 dst_pte = avtopte(addr); 2690 while (addr < pdnxt) { 2691 pt_entry_t ptetemp; 2692 ptetemp = *src_pte; 2693 /* 2694 * we only virtual copy managed pages 2695 */ 2696 if ((ptetemp & PG_MANAGED) != 0) { 2697 /* 2698 * We have to check after allocpte for the 2699 * pte still being around... allocpte can 2700 * block. 2701 */ 2702 dstmpte = pmap_allocpte(dst_pmap, addr); 2703 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2704 /* 2705 * Clear the modified and 2706 * accessed (referenced) bits 2707 * during the copy. 2708 */ 2709 m = PHYS_TO_VM_PAGE(ptetemp); 2710 *dst_pte = ptetemp & ~(PG_M | PG_A); 2711 dst_pmap->pm_stats.resident_count++; 2712 pmap_insert_entry(dst_pmap, addr, 2713 dstmpte, m); 2714 } else { 2715 vm_page_lock_queues(); 2716 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2717 vm_page_unlock_queues(); 2718 } 2719 if (dstmpte->hold_count >= srcmpte->hold_count) 2720 break; 2721 } 2722 addr += PAGE_SIZE; 2723 src_pte++; 2724 dst_pte++; 2725 } 2726 } 2727} 2728 2729#ifdef SMP 2730 2731/* 2732 * pmap_zpi_switchin*() 2733 * 2734 * These functions allow us to avoid doing IPIs alltogether in certain 2735 * temporary page-mapping situations (page zeroing). Instead to deal 2736 * with being preempted and moved onto a different cpu we invalidate 2737 * the page when the scheduler switches us in. This does not occur 2738 * very often so we remain relatively optimal with very little effort. 2739 */ 2740static void 2741pmap_zpi_switchin12(void) 2742{ 2743 invlpg((u_int)CADDR1); 2744 invlpg((u_int)CADDR2); 2745} 2746 2747static void 2748pmap_zpi_switchin2(void) 2749{ 2750 invlpg((u_int)CADDR2); 2751} 2752 2753static void 2754pmap_zpi_switchin3(void) 2755{ 2756 invlpg((u_int)CADDR3); 2757} 2758 2759#endif 2760 2761/* 2762 * pmap_zero_page zeros the specified hardware page by mapping 2763 * the page into KVM and using bzero to clear its contents. 2764 */ 2765void 2766pmap_zero_page(vm_page_t m) 2767{ 2768 vm_offset_t phys; 2769 2770 phys = VM_PAGE_TO_PHYS(m); 2771 if (*CMAP2) 2772 panic("pmap_zero_page: CMAP2 busy"); 2773 *CMAP2 = PG_V | PG_RW | phys | PG_A | PG_M; 2774#ifdef I386_CPU 2775 invltlb(); 2776#else 2777#ifdef SMP 2778 curthread->td_switchin = pmap_zpi_switchin2; 2779#endif 2780 invlpg((u_int)CADDR2); 2781#endif 2782#if defined(I686_CPU) 2783 if (cpu_class == CPUCLASS_686) 2784 i686_pagezero(CADDR2); 2785 else 2786#endif 2787 bzero(CADDR2, PAGE_SIZE); 2788#ifdef SMP 2789 curthread->td_switchin = NULL; 2790#endif 2791 *CMAP2 = 0; 2792} 2793 2794/* 2795 * pmap_zero_page_area zeros the specified hardware page by mapping 2796 * the page into KVM and using bzero to clear its contents. 2797 * 2798 * off and size may not cover an area beyond a single hardware page. 2799 */ 2800void 2801pmap_zero_page_area(vm_page_t m, int off, int size) 2802{ 2803 vm_offset_t phys; 2804 2805 phys = VM_PAGE_TO_PHYS(m); 2806 if (*CMAP2) 2807 panic("pmap_zero_page: CMAP2 busy"); 2808 *CMAP2 = PG_V | PG_RW | phys | PG_A | PG_M; 2809#ifdef I386_CPU 2810 invltlb(); 2811#else 2812#ifdef SMP 2813 curthread->td_switchin = pmap_zpi_switchin2; 2814#endif 2815 invlpg((u_int)CADDR2); 2816#endif 2817#if defined(I686_CPU) 2818 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE) 2819 i686_pagezero(CADDR2); 2820 else 2821#endif 2822 bzero((char *)CADDR2 + off, size); 2823#ifdef SMP 2824 curthread->td_switchin = NULL; 2825#endif 2826 *CMAP2 = 0; 2827} 2828 2829/* 2830 * pmap_zero_page_idle zeros the specified hardware page by mapping 2831 * the page into KVM and using bzero to clear its contents. This 2832 * is intended to be called from the vm_pagezero process only and 2833 * outside of Giant. 2834 */ 2835void 2836pmap_zero_page_idle(vm_page_t m) 2837{ 2838 vm_offset_t phys; 2839 2840 phys = VM_PAGE_TO_PHYS(m); 2841 if (*CMAP3) 2842 panic("pmap_zero_page: CMAP3 busy"); 2843 *CMAP3 = PG_V | PG_RW | phys | PG_A | PG_M; 2844#ifdef I386_CPU 2845 invltlb(); 2846#else 2847#ifdef SMP 2848 curthread->td_switchin = pmap_zpi_switchin3; 2849#endif 2850 invlpg((u_int)CADDR3); 2851#endif 2852#if defined(I686_CPU) 2853 if (cpu_class == CPUCLASS_686) 2854 i686_pagezero(CADDR3); 2855 else 2856#endif 2857 bzero(CADDR3, PAGE_SIZE); 2858#ifdef SMP 2859 curthread->td_switchin = NULL; 2860#endif 2861 *CMAP3 = 0; 2862} 2863 2864/* 2865 * pmap_copy_page copies the specified (machine independent) 2866 * page by mapping the page into virtual memory and using 2867 * bcopy to copy the page, one machine dependent page at a 2868 * time. 2869 */ 2870void 2871pmap_copy_page(vm_page_t src, vm_page_t dst) 2872{ 2873 2874 if (*CMAP1) 2875 panic("pmap_copy_page: CMAP1 busy"); 2876 if (*CMAP2) 2877 panic("pmap_copy_page: CMAP2 busy"); 2878 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2879 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2880#ifdef I386_CPU 2881 invltlb(); 2882#else 2883#ifdef SMP 2884 curthread->td_switchin = pmap_zpi_switchin12; 2885#endif 2886 invlpg((u_int)CADDR1); 2887 invlpg((u_int)CADDR2); 2888#endif 2889 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2890#ifdef SMP 2891 curthread->td_switchin = NULL; 2892#endif 2893 *CMAP1 = 0; 2894 *CMAP2 = 0; 2895} 2896 2897/* 2898 * Returns true if the pmap's pv is one of the first 2899 * 16 pvs linked to from this page. This count may 2900 * be changed upwards or downwards in the future; it 2901 * is only necessary that true be returned for a small 2902 * subset of pmaps for proper page aging. 2903 */ 2904boolean_t 2905pmap_page_exists_quick(pmap, m) 2906 pmap_t pmap; 2907 vm_page_t m; 2908{ 2909 pv_entry_t pv; 2910 int loops = 0; 2911 int s; 2912 2913 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2914 return FALSE; 2915 2916 s = splvm(); 2917 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2918 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2919 if (pv->pv_pmap == pmap) { 2920 splx(s); 2921 return TRUE; 2922 } 2923 loops++; 2924 if (loops >= 16) 2925 break; 2926 } 2927 splx(s); 2928 return (FALSE); 2929} 2930 2931#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2932/* 2933 * Remove all pages from specified address space 2934 * this aids process exit speeds. Also, this code 2935 * is special cased for current process only, but 2936 * can have the more generic (and slightly slower) 2937 * mode enabled. This is much faster than pmap_remove 2938 * in the case of running down an entire address space. 2939 */ 2940void 2941pmap_remove_pages(pmap, sva, eva) 2942 pmap_t pmap; 2943 vm_offset_t sva, eva; 2944{ 2945 pt_entry_t *pte, tpte; 2946 vm_page_t m; 2947 pv_entry_t pv, npv; 2948 int s; 2949 2950#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2951 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2952 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2953 return; 2954 } 2955#endif 2956 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2957 s = splvm(); 2958 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2959 2960 if (pv->pv_va >= eva || pv->pv_va < sva) { 2961 npv = TAILQ_NEXT(pv, pv_plist); 2962 continue; 2963 } 2964 2965#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2966 pte = vtopte(pv->pv_va); 2967#else 2968 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2969#endif 2970 tpte = *pte; 2971 2972 if (tpte == 0) { 2973 printf("TPTE at %p IS ZERO @ VA %08x\n", 2974 pte, pv->pv_va); 2975 panic("bad pte"); 2976 } 2977 2978/* 2979 * We cannot remove wired pages from a process' mapping at this time 2980 */ 2981 if (tpte & PG_W) { 2982 npv = TAILQ_NEXT(pv, pv_plist); 2983 continue; 2984 } 2985 2986 m = PHYS_TO_VM_PAGE(tpte); 2987 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2988 ("vm_page_t %p phys_addr mismatch %08x %08x", 2989 m, m->phys_addr, tpte)); 2990 2991 KASSERT(m < &vm_page_array[vm_page_array_size], 2992 ("pmap_remove_pages: bad tpte %x", tpte)); 2993 2994 pv->pv_pmap->pm_stats.resident_count--; 2995 2996 *pte = 0; 2997 2998 /* 2999 * Update the vm_page_t clean and reference bits. 3000 */ 3001 if (tpte & PG_M) { 3002 vm_page_dirty(m); 3003 } 3004 3005 npv = TAILQ_NEXT(pv, pv_plist); 3006 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 3007 3008 m->md.pv_list_count--; 3009 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 3010 if (TAILQ_FIRST(&m->md.pv_list) == NULL) { 3011 vm_page_flag_clear(m, PG_WRITEABLE); 3012 } 3013 3014 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 3015 free_pv_entry(pv); 3016 } 3017 splx(s); 3018 pmap_invalidate_all(pmap); 3019} 3020 3021/* 3022 * pmap_is_modified: 3023 * 3024 * Return whether or not the specified physical page was modified 3025 * in any physical maps. 3026 */ 3027boolean_t 3028pmap_is_modified(vm_page_t m) 3029{ 3030 pv_entry_t pv; 3031 pt_entry_t *pte; 3032 int s; 3033 3034 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3035 return FALSE; 3036 3037 s = splvm(); 3038 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3039 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3040 /* 3041 * if the bit being tested is the modified bit, then 3042 * mark clean_map and ptes as never 3043 * modified. 3044 */ 3045 if (!pmap_track_modified(pv->pv_va)) 3046 continue; 3047#if defined(PMAP_DIAGNOSTIC) 3048 if (!pv->pv_pmap) { 3049 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 3050 continue; 3051 } 3052#endif 3053 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3054 if (*pte & PG_M) { 3055 splx(s); 3056 return TRUE; 3057 } 3058 } 3059 splx(s); 3060 return (FALSE); 3061} 3062 3063/* 3064 * this routine is used to modify bits in ptes 3065 */ 3066static __inline void 3067pmap_changebit(vm_page_t m, int bit, boolean_t setem) 3068{ 3069 register pv_entry_t pv; 3070 register pt_entry_t *pte; 3071 int s; 3072 3073 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) || 3074 (!setem && bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 3075 return; 3076 3077 s = splvm(); 3078 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3079 /* 3080 * Loop over all current mappings setting/clearing as appropos If 3081 * setting RO do we need to clear the VAC? 3082 */ 3083 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3084 /* 3085 * don't write protect pager mappings 3086 */ 3087 if (!setem && (bit == PG_RW)) { 3088 if (!pmap_track_modified(pv->pv_va)) 3089 continue; 3090 } 3091 3092#if defined(PMAP_DIAGNOSTIC) 3093 if (!pv->pv_pmap) { 3094 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 3095 continue; 3096 } 3097#endif 3098 3099 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3100 3101 if (setem) { 3102 *pte |= bit; 3103 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3104 } else { 3105 pt_entry_t pbits = *pte; 3106 if (pbits & bit) { 3107 if (bit == PG_RW) { 3108 if (pbits & PG_M) { 3109 vm_page_dirty(m); 3110 } 3111 *pte = pbits & ~(PG_M|PG_RW); 3112 } else { 3113 *pte = pbits & ~bit; 3114 } 3115 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3116 } 3117 } 3118 } 3119 if (!setem && bit == PG_RW) 3120 vm_page_flag_clear(m, PG_WRITEABLE); 3121 splx(s); 3122} 3123 3124/* 3125 * pmap_page_protect: 3126 * 3127 * Lower the permission for all mappings to a given page. 3128 */ 3129void 3130pmap_page_protect(vm_page_t m, vm_prot_t prot) 3131{ 3132 if ((prot & VM_PROT_WRITE) == 0) { 3133 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 3134 pmap_changebit(m, PG_RW, FALSE); 3135 } else { 3136 pmap_remove_all(m); 3137 } 3138 } 3139} 3140 3141vm_offset_t 3142pmap_phys_address(ppn) 3143 int ppn; 3144{ 3145 return (i386_ptob(ppn)); 3146} 3147 3148/* 3149 * pmap_ts_referenced: 3150 * 3151 * Return a count of reference bits for a page, clearing those bits. 3152 * It is not necessary for every reference bit to be cleared, but it 3153 * is necessary that 0 only be returned when there are truly no 3154 * reference bits set. 3155 * 3156 * XXX: The exact number of bits to check and clear is a matter that 3157 * should be tested and standardized at some point in the future for 3158 * optimal aging of shared pages. 3159 */ 3160int 3161pmap_ts_referenced(vm_page_t m) 3162{ 3163 register pv_entry_t pv, pvf, pvn; 3164 pt_entry_t *pte; 3165 int s; 3166 int rtval = 0; 3167 3168 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3169 return (rtval); 3170 3171 s = splvm(); 3172 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3173 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 3174 3175 pvf = pv; 3176 3177 do { 3178 pvn = TAILQ_NEXT(pv, pv_list); 3179 3180 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 3181 3182 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 3183 3184 if (!pmap_track_modified(pv->pv_va)) 3185 continue; 3186 3187 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3188 3189 if (pte && (*pte & PG_A)) { 3190 *pte &= ~PG_A; 3191 3192 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3193 3194 rtval++; 3195 if (rtval > 4) { 3196 break; 3197 } 3198 } 3199 } while ((pv = pvn) != NULL && pv != pvf); 3200 } 3201 splx(s); 3202 3203 return (rtval); 3204} 3205 3206/* 3207 * Clear the modify bits on the specified physical page. 3208 */ 3209void 3210pmap_clear_modify(vm_page_t m) 3211{ 3212 pmap_changebit(m, PG_M, FALSE); 3213} 3214 3215/* 3216 * pmap_clear_reference: 3217 * 3218 * Clear the reference bit on the specified physical page. 3219 */ 3220void 3221pmap_clear_reference(vm_page_t m) 3222{ 3223 pmap_changebit(m, PG_A, FALSE); 3224} 3225 3226/* 3227 * Miscellaneous support routines follow 3228 */ 3229 3230static void 3231i386_protection_init() 3232{ 3233 register int *kp, prot; 3234 3235 kp = protection_codes; 3236 for (prot = 0; prot < 8; prot++) { 3237 switch (prot) { 3238 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE: 3239 /* 3240 * Read access is also 0. There isn't any execute bit, 3241 * so just make it readable. 3242 */ 3243 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE: 3244 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE: 3245 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE: 3246 *kp++ = 0; 3247 break; 3248 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE: 3249 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE: 3250 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE: 3251 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE: 3252 *kp++ = PG_RW; 3253 break; 3254 } 3255 } 3256} 3257 3258/* 3259 * Map a set of physical memory pages into the kernel virtual 3260 * address space. Return a pointer to where it is mapped. This 3261 * routine is intended to be used for mapping device memory, 3262 * NOT real memory. 3263 */ 3264void * 3265pmap_mapdev(pa, size) 3266 vm_offset_t pa; 3267 vm_size_t size; 3268{ 3269 vm_offset_t va, tmpva, offset; 3270 3271 offset = pa & PAGE_MASK; 3272 size = roundup(offset + size, PAGE_SIZE); 3273 3274 GIANT_REQUIRED; 3275 3276 va = kmem_alloc_pageable(kernel_map, size); 3277 if (!va) 3278 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 3279 3280 pa = pa & PG_FRAME; 3281 for (tmpva = va; size > 0; ) { 3282 pmap_kenter(tmpva, pa); 3283 size -= PAGE_SIZE; 3284 tmpva += PAGE_SIZE; 3285 pa += PAGE_SIZE; 3286 } 3287 pmap_invalidate_range(kernel_pmap, va, tmpva); 3288 return ((void *)(va + offset)); 3289} 3290 3291void 3292pmap_unmapdev(va, size) 3293 vm_offset_t va; 3294 vm_size_t size; 3295{ 3296 vm_offset_t base, offset, tmpva; 3297 pt_entry_t *pte; 3298 3299 base = va & PG_FRAME; 3300 offset = va & PAGE_MASK; 3301 size = roundup(offset + size, PAGE_SIZE); 3302 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) { 3303 pte = vtopte(tmpva); 3304 *pte = 0; 3305 } 3306 pmap_invalidate_range(kernel_pmap, va, tmpva); 3307 kmem_free(kernel_map, base, size); 3308} 3309 3310/* 3311 * perform the pmap work for mincore 3312 */ 3313int 3314pmap_mincore(pmap, addr) 3315 pmap_t pmap; 3316 vm_offset_t addr; 3317{ 3318 pt_entry_t *ptep, pte; 3319 vm_page_t m; 3320 int val = 0; 3321 3322 ptep = pmap_pte(pmap, addr); 3323 if (ptep == 0) { 3324 return 0; 3325 } 3326 3327 if ((pte = *ptep) != 0) { 3328 vm_offset_t pa; 3329 3330 val = MINCORE_INCORE; 3331 if ((pte & PG_MANAGED) == 0) 3332 return val; 3333 3334 pa = pte & PG_FRAME; 3335 3336 m = PHYS_TO_VM_PAGE(pa); 3337 3338 /* 3339 * Modified by us 3340 */ 3341 if (pte & PG_M) 3342 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 3343 else { 3344 /* 3345 * Modified by someone else 3346 */ 3347 vm_page_lock_queues(); 3348 if (m->dirty || pmap_is_modified(m)) 3349 val |= MINCORE_MODIFIED_OTHER; 3350 vm_page_unlock_queues(); 3351 } 3352 /* 3353 * Referenced by us 3354 */ 3355 if (pte & PG_A) 3356 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 3357 else { 3358 /* 3359 * Referenced by someone else 3360 */ 3361 vm_page_lock_queues(); 3362 if ((m->flags & PG_REFERENCED) || 3363 pmap_ts_referenced(m)) { 3364 val |= MINCORE_REFERENCED_OTHER; 3365 vm_page_flag_set(m, PG_REFERENCED); 3366 } 3367 vm_page_unlock_queues(); 3368 } 3369 } 3370 return val; 3371} 3372 3373void 3374pmap_activate(struct thread *td) 3375{ 3376 struct proc *p = td->td_proc; 3377 pmap_t pmap; 3378 u_int32_t cr3; 3379 3380 pmap = vmspace_pmap(td->td_proc->p_vmspace); 3381#if defined(SMP) 3382 pmap->pm_active |= PCPU_GET(cpumask); 3383#else 3384 pmap->pm_active |= 1; 3385#endif 3386 cr3 = vtophys(pmap->pm_pdir); 3387 /* XXXKSE this is wrong. 3388 * pmap_activate is for the current thread on the current cpu 3389 */ 3390 if (p->p_flag & P_KSES) { 3391 /* Make sure all other cr3 entries are updated. */ 3392 /* what if they are running? XXXKSE (maybe abort them) */ 3393 FOREACH_THREAD_IN_PROC(p, td) { 3394 td->td_pcb->pcb_cr3 = cr3; 3395 } 3396 } else { 3397 td->td_pcb->pcb_cr3 = cr3; 3398 } 3399 load_cr3(cr3); 3400#ifdef SWTCH_OPTIM_STATS 3401 tlb_flush_count++; 3402#endif 3403} 3404 3405vm_offset_t 3406pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3407{ 3408 3409 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3410 return addr; 3411 } 3412 3413 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); 3414 return addr; 3415} 3416 3417 3418#if defined(PMAP_DEBUG) 3419pmap_pid_dump(int pid) 3420{ 3421 pmap_t pmap; 3422 struct proc *p; 3423 int npte = 0; 3424 int index; 3425 3426 sx_slock(&allproc_lock); 3427 LIST_FOREACH(p, &allproc, p_list) { 3428 if (p->p_pid != pid) 3429 continue; 3430 3431 if (p->p_vmspace) { 3432 int i,j; 3433 index = 0; 3434 pmap = vmspace_pmap(p->p_vmspace); 3435 for (i = 0; i < NPDEPG; i++) { 3436 pd_entry_t *pde; 3437 pt_entry_t *pte; 3438 vm_offset_t base = i << PDRSHIFT; 3439 3440 pde = &pmap->pm_pdir[i]; 3441 if (pde && pmap_pde_v(pde)) { 3442 for (j = 0; j < NPTEPG; j++) { 3443 vm_offset_t va = base + (j << PAGE_SHIFT); 3444 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3445 if (index) { 3446 index = 0; 3447 printf("\n"); 3448 } 3449 sx_sunlock(&allproc_lock); 3450 return npte; 3451 } 3452 pte = pmap_pte_quick(pmap, va); 3453 if (pte && pmap_pte_v(pte)) { 3454 pt_entry_t pa; 3455 vm_page_t m; 3456 pa = *pte; 3457 m = PHYS_TO_VM_PAGE(pa); 3458 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3459 va, pa, m->hold_count, m->wire_count, m->flags); 3460 npte++; 3461 index++; 3462 if (index >= 2) { 3463 index = 0; 3464 printf("\n"); 3465 } else { 3466 printf(" "); 3467 } 3468 } 3469 } 3470 } 3471 } 3472 } 3473 } 3474 sx_sunlock(&allproc_lock); 3475 return npte; 3476} 3477#endif 3478 3479#if defined(DEBUG) 3480 3481static void pads(pmap_t pm); 3482void pmap_pvdump(vm_offset_t pa); 3483 3484/* print address space of pmap*/ 3485static void 3486pads(pm) 3487 pmap_t pm; 3488{ 3489 int i, j; 3490 vm_offset_t va; 3491 pt_entry_t *ptep; 3492 3493 if (pm == kernel_pmap) 3494 return; 3495 for (i = 0; i < NPDEPG; i++) 3496 if (pm->pm_pdir[i]) 3497 for (j = 0; j < NPTEPG; j++) { 3498 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3499 if (pm == kernel_pmap && va < KERNBASE) 3500 continue; 3501 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3502 continue; 3503 ptep = pmap_pte_quick(pm, va); 3504 if (pmap_pte_v(ptep)) 3505 printf("%x:%x ", va, *ptep); 3506 }; 3507 3508} 3509 3510void 3511pmap_pvdump(pa) 3512 vm_offset_t pa; 3513{ 3514 pv_entry_t pv; 3515 vm_page_t m; 3516 3517 printf("pa %x", pa); 3518 m = PHYS_TO_VM_PAGE(pa); 3519 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3520 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3521 pads(pv->pv_pmap); 3522 } 3523 printf(" "); 3524} 3525#endif 3526