pmap.c revision 120734
14Srgrimes/*- 2233207Stijl * Copyright (c) 1991 Regents of the University of California. 34Srgrimes * All rights reserved. 4233207Stijl * Copyright (c) 1994 John S. Dyson 54Srgrimes * All rights reserved. 6233207Stijl * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * the Systems Programming Group of the University of Utah Computer 11 * Science Department and William Jolitz of UUNET Technologies Inc. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by the University of 24 * California, Berkeley and its contributors. 25 * 4. Neither the name of the University nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 39 * SUCH DAMAGE. 40 * 41 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 42 */ 43/*- 44 * Copyright (c) 2003 Networks Associates Technology, Inc. 45 * All rights reserved. 46 * 47 * This software was developed for the FreeBSD Project by Jake Burkholder, 48 * Safeport Network Services, and Network Associates Laboratories, the 49 * Security Research Division of Network Associates, Inc. under 50 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA 51 * CHATS research program. 52 * 53 * Redistribution and use in source and binary forms, with or without 54 * modification, are permitted provided that the following conditions 55 * are met: 56 * 1. Redistributions of source code must retain the above copyright 57 * notice, this list of conditions and the following disclaimer. 58 * 2. Redistributions in binary form must reproduce the above copyright 59 * notice, this list of conditions and the following disclaimer in the 60 * documentation and/or other materials provided with the distribution. 61 * 62 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 63 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 64 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 65 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 66 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 67 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 68 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 69 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 70 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 71 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 72 * SUCH DAMAGE. 73 */ 74 75#include <sys/cdefs.h> 76__FBSDID("$FreeBSD: head/sys/i386/i386/pmap.c 120734 2003-10-04 13:07:53Z jeff $"); 77 78/* 79 * Manages physical address maps. 80 * 81 * In addition to hardware address maps, this 82 * module is called upon to provide software-use-only 83 * maps which may or may not be stored in the same 84 * form as hardware maps. These pseudo-maps are 85 * used to store intermediate results from copy 86 * operations to and from address spaces. 87 * 88 * Since the information managed by this module is 89 * also stored by the logical address mapping module, 90 * this module may throw away valid virtual-to-physical 91 * mappings at almost any time. However, invalidations 92 * of virtual-to-physical mappings must be done as 93 * requested. 94 * 95 * In order to cope with hardware architectures which 96 * make virtual-to-physical map invalidates expensive, 97 * this module may delay invalidate or reduced protection 98 * operations until such time as they are actually 99 * necessary. This module is given full information as 100 * to which processors are currently using which maps, 101 * and to when physical maps must be made correct. 102 */ 103 104#include "opt_pmap.h" 105#include "opt_msgbuf.h" 106#include "opt_kstack_pages.h" 107 108#include <sys/param.h> 109#include <sys/systm.h> 110#include <sys/kernel.h> 111#include <sys/lock.h> 112#include <sys/mman.h> 113#include <sys/msgbuf.h> 114#include <sys/mutex.h> 115#include <sys/proc.h> 116#include <sys/sx.h> 117#include <sys/user.h> 118#include <sys/vmmeter.h> 119#include <sys/sysctl.h> 120#ifdef SMP 121#include <sys/smp.h> 122#endif 123 124#include <vm/vm.h> 125#include <vm/vm_param.h> 126#include <vm/vm_kern.h> 127#include <vm/vm_page.h> 128#include <vm/vm_map.h> 129#include <vm/vm_object.h> 130#include <vm/vm_extern.h> 131#include <vm/vm_pageout.h> 132#include <vm/vm_pager.h> 133#include <vm/uma.h> 134 135#include <machine/cpu.h> 136#include <machine/cputypes.h> 137#include <machine/md_var.h> 138#include <machine/specialreg.h> 139#if defined(SMP) || defined(APIC_IO) 140#include <machine/smp.h> 141#include <machine/apic.h> 142#include <machine/segments.h> 143#include <machine/tss.h> 144#endif /* SMP || APIC_IO */ 145 146#define PMAP_KEEP_PDIRS 147#ifndef PMAP_SHPGPERPROC 148#define PMAP_SHPGPERPROC 200 149#endif 150 151#if defined(DIAGNOSTIC) 152#define PMAP_DIAGNOSTIC 153#endif 154 155#define MINPV 2048 156 157#if !defined(PMAP_DIAGNOSTIC) 158#define PMAP_INLINE __inline 159#else 160#define PMAP_INLINE 161#endif 162 163/* 164 * Get PDEs and PTEs for user/kernel address space 165 */ 166#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 167#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 168 169#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 170#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 171#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 172#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 173#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 174 175#define pmap_pte_set_w(pte, v) ((v)?(*(int *)pte |= PG_W):(*(int *)pte &= ~PG_W)) 176#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 177 178/* 179 * Given a map and a machine independent protection code, 180 * convert to a vax protection code. 181 */ 182#define pte_prot(m, p) (protection_codes[p]) 183static int protection_codes[8]; 184 185struct pmap kernel_pmap_store; 186LIST_HEAD(pmaplist, pmap); 187static struct pmaplist allpmaps; 188static struct mtx allpmaps_lock; 189#ifdef SMP 190static struct mtx lazypmap_lock; 191#endif 192 193vm_paddr_t avail_start; /* PA of first available physical page */ 194vm_paddr_t avail_end; /* PA of last available physical page */ 195vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 196vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 197static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 198int pgeflag = 0; /* PG_G or-in */ 199int pseflag = 0; /* PG_PS or-in */ 200 201static int nkpt; 202vm_offset_t kernel_vm_end; 203extern u_int32_t KERNend; 204 205#ifdef PAE 206static uma_zone_t pdptzone; 207#endif 208 209/* 210 * Data for the pv entry allocation mechanism 211 */ 212static uma_zone_t pvzone; 213static struct vm_object pvzone_obj; 214static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 215int pmap_pagedaemon_waken; 216 217/* 218 * All those kernel PT submaps that BSD is so fond of 219 */ 220pt_entry_t *CMAP1 = 0; 221static pt_entry_t *CMAP2, *CMAP3, *ptmmap; 222caddr_t CADDR1 = 0, ptvmmap = 0; 223static caddr_t CADDR2, CADDR3; 224static struct mtx CMAPCADDR12_lock; 225static pt_entry_t *msgbufmap; 226struct msgbuf *msgbufp = 0; 227 228/* 229 * Crashdump maps. 230 */ 231static pt_entry_t *pt_crashdumpmap; 232static caddr_t crashdumpmap; 233 234#ifdef SMP 235extern pt_entry_t *SMPpt; 236#endif 237static pt_entry_t *PMAP1 = 0; 238static pt_entry_t *PADDR1 = 0; 239 240static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 241static pv_entry_t get_pv_entry(void); 242static void i386_protection_init(void); 243static void pmap_clear_ptes(vm_page_t m, int bit) 244 __always_inline; 245 246static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 247static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 248static int pmap_remove_entry(struct pmap *pmap, vm_page_t m, 249 vm_offset_t va); 250static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, 251 vm_page_t mpte, vm_page_t m); 252 253static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va); 254 255static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex); 256static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t); 257static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 258static void *pmap_pv_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 259#ifdef PAE 260static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 261#endif 262 263CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 264CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 265 266/* 267 * Move the kernel virtual free pointer to the next 268 * 4MB. This is used to help improve performance 269 * by using a large (4MB) page for much of the kernel 270 * (.text, .data, .bss) 271 */ 272static vm_offset_t 273pmap_kmem_choose(vm_offset_t addr) 274{ 275 vm_offset_t newaddr = addr; 276 277#ifdef I686_CPU_not /* Problem seems to have gone away */ 278 /* Deal with un-resolved Pentium4 issues */ 279 if (cpu_class == CPUCLASS_686 && 280 strcmp(cpu_vendor, "GenuineIntel") == 0 && 281 (cpu_id & 0xf00) == 0xf00) 282 return newaddr; 283#endif 284#ifndef DISABLE_PSE 285 if (cpu_feature & CPUID_PSE) 286 newaddr = (addr + PDRMASK) & ~PDRMASK; 287#endif 288 return newaddr; 289} 290 291/* 292 * Bootstrap the system enough to run with virtual memory. 293 * 294 * On the i386 this is called after mapping has already been enabled 295 * and just syncs the pmap module with what has already been done. 296 * [We can't call it easily with mapping off since the kernel is not 297 * mapped with PA == VA, hence we would have to relocate every address 298 * from the linked base (virtual) address "KERNBASE" to the actual 299 * (physical) address starting relative to 0] 300 */ 301void 302pmap_bootstrap(firstaddr, loadaddr) 303 vm_paddr_t firstaddr; 304 vm_paddr_t loadaddr; 305{ 306 vm_offset_t va; 307 pt_entry_t *pte; 308 int i; 309 310 avail_start = firstaddr; 311 312 /* 313 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 314 * large. It should instead be correctly calculated in locore.s and 315 * not based on 'first' (which is a physical address, not a virtual 316 * address, for the start of unused physical memory). The kernel 317 * page tables are NOT double mapped and thus should not be included 318 * in this calculation. 319 */ 320 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 321 virtual_avail = pmap_kmem_choose(virtual_avail); 322 323 virtual_end = VM_MAX_KERNEL_ADDRESS; 324 325 /* 326 * Initialize protection array. 327 */ 328 i386_protection_init(); 329 330 /* 331 * Initialize the kernel pmap (which is statically allocated). 332 */ 333 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 334#ifdef PAE 335 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 336#endif 337 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 338 TAILQ_INIT(&kernel_pmap->pm_pvlist); 339 LIST_INIT(&allpmaps); 340#ifdef SMP 341 mtx_init(&lazypmap_lock, "lazypmap", NULL, MTX_SPIN); 342#endif 343 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 344 mtx_lock_spin(&allpmaps_lock); 345 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 346 mtx_unlock_spin(&allpmaps_lock); 347 nkpt = NKPT; 348 349 /* 350 * Reserve some special page table entries/VA space for temporary 351 * mapping of pages. 352 */ 353#define SYSMAP(c, p, v, n) \ 354 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 355 356 va = virtual_avail; 357 pte = vtopte(va); 358 359 /* 360 * CMAP1/CMAP2 are used for zeroing and copying pages. 361 * CMAP3 is used for the idle process page zeroing. 362 */ 363 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 364 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 365 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 366 367 mtx_init(&CMAPCADDR12_lock, "CMAPCADDR12", NULL, MTX_DEF); 368 369 /* 370 * Crashdump maps. 371 */ 372 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 373 374 /* 375 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 376 * XXX ptmmap is not used. 377 */ 378 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 379 380 /* 381 * msgbufp is used to map the system message buffer. 382 * XXX msgbufmap is not used. 383 */ 384 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 385 atop(round_page(MSGBUF_SIZE))) 386 387 /* 388 * ptemap is used for pmap_pte_quick 389 */ 390 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 391 392 virtual_avail = va; 393 394 *CMAP1 = *CMAP2 = 0; 395 for (i = 0; i < NKPT; i++) 396 PTD[i] = 0; 397 398#ifdef I686_CPU_not /* Problem seems to have gone away */ 399 /* Deal with un-resolved Pentium4 issues */ 400 if (cpu_class == CPUCLASS_686 && 401 strcmp(cpu_vendor, "GenuineIntel") == 0 && 402 (cpu_id & 0xf00) == 0xf00) { 403 printf("Warning: Pentium 4 cpu: PG_G disabled (global flag)\n"); 404 pgeflag = 0; 405 } 406#endif 407 408#ifdef I686_CPU_not /* Problem seems to have gone away */ 409 /* Deal with un-resolved Pentium4 issues */ 410 if (cpu_class == CPUCLASS_686 && 411 strcmp(cpu_vendor, "GenuineIntel") == 0 && 412 (cpu_id & 0xf00) == 0xf00) { 413 printf("Warning: Pentium 4 cpu: PG_PS disabled (4MB pages)\n"); 414 pseflag = 0; 415 } 416#endif 417 418 /* Turn on PG_G on kernel page(s) */ 419 pmap_set_pg(); 420 421#ifdef SMP 422 if (cpu_apic_address == 0) 423 panic("pmap_bootstrap: no local apic! (non-SMP hardware?)"); 424 425 /* local apic is mapped on last page */ 426 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag | 427 (cpu_apic_address & PG_FRAME)); 428#endif 429 invltlb(); 430} 431 432/* 433 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 434 */ 435void 436pmap_set_pg(void) 437{ 438 pd_entry_t pdir; 439 pt_entry_t *pte; 440 vm_offset_t va, endva; 441 int i; 442 443 if (pgeflag == 0) 444 return; 445 446 i = KERNLOAD/NBPDR; 447 endva = KERNBASE + KERNend; 448 449 if (pseflag) { 450 va = KERNBASE + KERNLOAD; 451 while (va < endva) { 452 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 453 pdir |= pgeflag; 454 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 455 invltlb(); /* Play it safe, invltlb() every time */ 456 i++; 457 va += NBPDR; 458 } 459 } else { 460 va = (vm_offset_t)btext; 461 while (va < endva) { 462 pte = vtopte(va); 463 if (*pte) 464 *pte |= pgeflag; 465 invltlb(); /* Play it safe, invltlb() every time */ 466 va += PAGE_SIZE; 467 } 468 } 469} 470 471static void * 472pmap_pv_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 473{ 474 *flags = UMA_SLAB_PRIV; 475 return (void *)kmem_alloc(kernel_map, bytes); 476} 477 478#ifdef PAE 479static void * 480pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 481{ 482 *flags = UMA_SLAB_PRIV; 483 return (contigmalloc(PAGE_SIZE, NULL, 0, 0x0ULL, 0xffffffffULL, 1, 0)); 484} 485#endif 486 487/* 488 * Initialize the pmap module. 489 * Called by vm_init, to initialize any structures that the pmap 490 * system needs to map virtual memory. 491 * pmap_init has been enhanced to support in a fairly consistant 492 * way, discontiguous physical memory. 493 */ 494void 495pmap_init(phys_start, phys_end) 496 vm_paddr_t phys_start, phys_end; 497{ 498 int i; 499 int initial_pvs; 500 501 /* 502 * Allocate memory for random pmap data structures. Includes the 503 * pv_head_table. 504 */ 505 506 for(i = 0; i < vm_page_array_size; i++) { 507 vm_page_t m; 508 509 m = &vm_page_array[i]; 510 TAILQ_INIT(&m->md.pv_list); 511 m->md.pv_list_count = 0; 512 } 513 514 /* 515 * init the pv free list 516 */ 517 initial_pvs = vm_page_array_size; 518 if (initial_pvs < MINPV) 519 initial_pvs = MINPV; 520 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL, 521 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 522 uma_zone_set_allocf(pvzone, pmap_pv_allocf); 523 uma_prealloc(pvzone, initial_pvs); 524 525#ifdef PAE 526 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 527 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 528 UMA_ZONE_VM | UMA_ZONE_NOFREE); 529 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 530#endif 531 532 /* 533 * Now it is safe to enable pv_table recording. 534 */ 535 pmap_initialized = TRUE; 536} 537 538/* 539 * Initialize the address space (zone) for the pv_entries. Set a 540 * high water mark so that the system can recover from excessive 541 * numbers of pv entries. 542 */ 543void 544pmap_init2() 545{ 546 int shpgperproc = PMAP_SHPGPERPROC; 547 548 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 549 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 550 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 551 pv_entry_high_water = 9 * (pv_entry_max / 10); 552 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 553} 554 555 556/*************************************************** 557 * Low level helper routines..... 558 ***************************************************/ 559 560#if defined(PMAP_DIAGNOSTIC) 561 562/* 563 * This code checks for non-writeable/modified pages. 564 * This should be an invalid condition. 565 */ 566static int 567pmap_nw_modified(pt_entry_t ptea) 568{ 569 int pte; 570 571 pte = (int) ptea; 572 573 if ((pte & (PG_M|PG_RW)) == PG_M) 574 return 1; 575 else 576 return 0; 577} 578#endif 579 580 581/* 582 * this routine defines the region(s) of memory that should 583 * not be tested for the modified bit. 584 */ 585static PMAP_INLINE int 586pmap_track_modified(vm_offset_t va) 587{ 588 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 589 return 1; 590 else 591 return 0; 592} 593 594#ifdef I386_CPU 595/* 596 * i386 only has "invalidate everything" and no SMP to worry about. 597 */ 598PMAP_INLINE void 599pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 600{ 601 602 if (pmap == kernel_pmap || pmap->pm_active) 603 invltlb(); 604} 605 606PMAP_INLINE void 607pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 608{ 609 610 if (pmap == kernel_pmap || pmap->pm_active) 611 invltlb(); 612} 613 614PMAP_INLINE void 615pmap_invalidate_all(pmap_t pmap) 616{ 617 618 if (pmap == kernel_pmap || pmap->pm_active) 619 invltlb(); 620} 621#else /* !I386_CPU */ 622#ifdef SMP 623/* 624 * For SMP, these functions have to use the IPI mechanism for coherence. 625 */ 626void 627pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 628{ 629 u_int cpumask; 630 u_int other_cpus; 631 632 critical_enter(); 633 /* 634 * We need to disable interrupt preemption but MUST NOT have 635 * interrupts disabled here. 636 * XXX we may need to hold schedlock to get a coherent pm_active 637 */ 638 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) { 639 invlpg(va); 640 smp_invlpg(va); 641 } else { 642 cpumask = PCPU_GET(cpumask); 643 other_cpus = PCPU_GET(other_cpus); 644 if (pmap->pm_active & cpumask) 645 invlpg(va); 646 if (pmap->pm_active & other_cpus) 647 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 648 } 649 critical_exit(); 650} 651 652void 653pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 654{ 655 u_int cpumask; 656 u_int other_cpus; 657 vm_offset_t addr; 658 659 critical_enter(); 660 /* 661 * We need to disable interrupt preemption but MUST NOT have 662 * interrupts disabled here. 663 * XXX we may need to hold schedlock to get a coherent pm_active 664 */ 665 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) { 666 for (addr = sva; addr < eva; addr += PAGE_SIZE) 667 invlpg(addr); 668 smp_invlpg_range(sva, eva); 669 } else { 670 cpumask = PCPU_GET(cpumask); 671 other_cpus = PCPU_GET(other_cpus); 672 if (pmap->pm_active & cpumask) 673 for (addr = sva; addr < eva; addr += PAGE_SIZE) 674 invlpg(addr); 675 if (pmap->pm_active & other_cpus) 676 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 677 sva, eva); 678 } 679 critical_exit(); 680} 681 682void 683pmap_invalidate_all(pmap_t pmap) 684{ 685 u_int cpumask; 686 u_int other_cpus; 687 688 critical_enter(); 689 /* 690 * We need to disable interrupt preemption but MUST NOT have 691 * interrupts disabled here. 692 * XXX we may need to hold schedlock to get a coherent pm_active 693 */ 694 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) { 695 invltlb(); 696 smp_invltlb(); 697 } else { 698 cpumask = PCPU_GET(cpumask); 699 other_cpus = PCPU_GET(other_cpus); 700 if (pmap->pm_active & cpumask) 701 invltlb(); 702 if (pmap->pm_active & other_cpus) 703 smp_masked_invltlb(pmap->pm_active & other_cpus); 704 } 705 critical_exit(); 706} 707#else /* !SMP */ 708/* 709 * Normal, non-SMP, 486+ invalidation functions. 710 * We inline these within pmap.c for speed. 711 */ 712PMAP_INLINE void 713pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 714{ 715 716 if (pmap == kernel_pmap || pmap->pm_active) 717 invlpg(va); 718} 719 720PMAP_INLINE void 721pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 722{ 723 vm_offset_t addr; 724 725 if (pmap == kernel_pmap || pmap->pm_active) 726 for (addr = sva; addr < eva; addr += PAGE_SIZE) 727 invlpg(addr); 728} 729 730PMAP_INLINE void 731pmap_invalidate_all(pmap_t pmap) 732{ 733 734 if (pmap == kernel_pmap || pmap->pm_active) 735 invltlb(); 736} 737#endif /* !SMP */ 738#endif /* !I386_CPU */ 739 740/* 741 * Are we current address space or kernel? 742 */ 743static __inline int 744pmap_is_current(pmap_t pmap) 745{ 746 return (pmap == kernel_pmap || 747 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)); 748} 749 750/* 751 * Super fast pmap_pte routine best used when scanning 752 * the pv lists. This eliminates many coarse-grained 753 * invltlb calls. Note that many of the pv list 754 * scans are across different pmaps. It is very wasteful 755 * to do an entire invltlb for checking a single mapping. 756 */ 757pt_entry_t * 758pmap_pte_quick(pmap, va) 759 register pmap_t pmap; 760 vm_offset_t va; 761{ 762 pd_entry_t newpf; 763 pd_entry_t *pde; 764 765 pde = pmap_pde(pmap, va); 766 if (*pde & PG_PS) 767 return (pde); 768 if (*pde != 0) { 769 /* are we current address space or kernel? */ 770 if (pmap_is_current(pmap)) 771 return vtopte(va); 772 newpf = *pde & PG_FRAME; 773 if (((*PMAP1) & PG_FRAME) != newpf) { 774 *PMAP1 = newpf | PG_RW | PG_V; 775 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR1); 776 } 777 return PADDR1 + (i386_btop(va) & (NPTEPG - 1)); 778 } 779 return (0); 780} 781 782/* 783 * Routine: pmap_extract 784 * Function: 785 * Extract the physical page address associated 786 * with the given map/virtual_address pair. 787 */ 788vm_paddr_t 789pmap_extract(pmap, va) 790 register pmap_t pmap; 791 vm_offset_t va; 792{ 793 vm_paddr_t rtval; 794 pt_entry_t *pte; 795 pd_entry_t pde; 796 797 if (pmap == 0) 798 return 0; 799 pde = pmap->pm_pdir[va >> PDRSHIFT]; 800 if (pde != 0) { 801 if ((pde & PG_PS) != 0) { 802 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 803 return rtval; 804 } 805 pte = pmap_pte_quick(pmap, va); 806 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK)); 807 return rtval; 808 } 809 return 0; 810 811} 812 813/* 814 * Routine: pmap_extract_and_hold 815 * Function: 816 * Atomically extract and hold the physical page 817 * with the given pmap and virtual address pair 818 * if that mapping permits the given protection. 819 */ 820vm_page_t 821pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 822{ 823 vm_paddr_t pa; 824 vm_page_t m; 825 826 m = NULL; 827 mtx_lock(&Giant); 828 if ((pa = pmap_extract(pmap, va)) != 0) { 829 m = PHYS_TO_VM_PAGE(pa); 830 vm_page_lock_queues(); 831 vm_page_hold(m); 832 vm_page_unlock_queues(); 833 } 834 mtx_unlock(&Giant); 835 return (m); 836} 837 838/*************************************************** 839 * Low level mapping routines..... 840 ***************************************************/ 841 842/* 843 * Add a wired page to the kva. 844 * Note: not SMP coherent. 845 */ 846PMAP_INLINE void 847pmap_kenter(vm_offset_t va, vm_paddr_t pa) 848{ 849 pt_entry_t *pte; 850 851 pte = vtopte(va); 852 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 853} 854 855/* 856 * Remove a page from the kernel pagetables. 857 * Note: not SMP coherent. 858 */ 859PMAP_INLINE void 860pmap_kremove(vm_offset_t va) 861{ 862 pt_entry_t *pte; 863 864 pte = vtopte(va); 865 pte_clear(pte); 866} 867 868/* 869 * Used to map a range of physical addresses into kernel 870 * virtual address space. 871 * 872 * The value passed in '*virt' is a suggested virtual address for 873 * the mapping. Architectures which can support a direct-mapped 874 * physical to virtual region can return the appropriate address 875 * within that region, leaving '*virt' unchanged. Other 876 * architectures should map the pages starting at '*virt' and 877 * update '*virt' with the first usable address after the mapped 878 * region. 879 */ 880vm_offset_t 881pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 882{ 883 vm_offset_t va, sva; 884 885 va = sva = *virt; 886 while (start < end) { 887 pmap_kenter(va, start); 888 va += PAGE_SIZE; 889 start += PAGE_SIZE; 890 } 891 pmap_invalidate_range(kernel_pmap, sva, va); 892 *virt = va; 893 return (sva); 894} 895 896 897/* 898 * Add a list of wired pages to the kva 899 * this routine is only used for temporary 900 * kernel mappings that do not need to have 901 * page modification or references recorded. 902 * Note that old mappings are simply written 903 * over. The page *must* be wired. 904 * Note: SMP coherent. Uses a ranged shootdown IPI. 905 */ 906void 907pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 908{ 909 vm_offset_t va; 910 911 va = sva; 912 while (count-- > 0) { 913 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 914 va += PAGE_SIZE; 915 m++; 916 } 917 pmap_invalidate_range(kernel_pmap, sva, va); 918} 919 920/* 921 * This routine tears out page mappings from the 922 * kernel -- it is meant only for temporary mappings. 923 * Note: SMP coherent. Uses a ranged shootdown IPI. 924 */ 925void 926pmap_qremove(vm_offset_t sva, int count) 927{ 928 vm_offset_t va; 929 930 va = sva; 931 while (count-- > 0) { 932 pmap_kremove(va); 933 va += PAGE_SIZE; 934 } 935 pmap_invalidate_range(kernel_pmap, sva, va); 936} 937 938/*************************************************** 939 * Page table page management routines..... 940 ***************************************************/ 941 942/* 943 * This routine unholds page table pages, and if the hold count 944 * drops to zero, then it decrements the wire count. 945 */ 946static int 947_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 948{ 949 950 while (vm_page_sleep_if_busy(m, FALSE, "pmuwpt")) 951 vm_page_lock_queues(); 952 953 if (m->hold_count == 0) { 954 vm_offset_t pteva; 955 /* 956 * unmap the page table page 957 */ 958 pmap->pm_pdir[m->pindex] = 0; 959 --pmap->pm_stats.resident_count; 960 if (pmap_is_current(pmap)) { 961 /* 962 * Do an invltlb to make the invalidated mapping 963 * take effect immediately. 964 */ 965 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 966 pmap_invalidate_page(pmap, pteva); 967 } 968 969 /* 970 * If the page is finally unwired, simply free it. 971 */ 972 --m->wire_count; 973 if (m->wire_count == 0) { 974 vm_page_busy(m); 975 vm_page_free_zero(m); 976 atomic_subtract_int(&cnt.v_wire_count, 1); 977 } 978 return 1; 979 } 980 return 0; 981} 982 983static PMAP_INLINE int 984pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 985{ 986 vm_page_unhold(m); 987 if (m->hold_count == 0) 988 return _pmap_unwire_pte_hold(pmap, m); 989 else 990 return 0; 991} 992 993/* 994 * After removing a page table entry, this routine is used to 995 * conditionally free the page, and manage the hold/wire counts. 996 */ 997static int 998pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 999{ 1000 1001 if (va >= VM_MAXUSER_ADDRESS) 1002 return 0; 1003 1004 return pmap_unwire_pte_hold(pmap, mpte); 1005} 1006 1007void 1008pmap_pinit0(pmap) 1009 struct pmap *pmap; 1010{ 1011 1012 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1013#ifdef PAE 1014 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1015#endif 1016 pmap->pm_active = 0; 1017 PCPU_SET(curpmap, pmap); 1018 TAILQ_INIT(&pmap->pm_pvlist); 1019 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1020 mtx_lock_spin(&allpmaps_lock); 1021 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1022 mtx_unlock_spin(&allpmaps_lock); 1023} 1024 1025/* 1026 * Initialize a preallocated and zeroed pmap structure, 1027 * such as one in a vmspace structure. 1028 */ 1029void 1030pmap_pinit(pmap) 1031 register struct pmap *pmap; 1032{ 1033 vm_page_t m, ptdpg[NPGPTD]; 1034 vm_paddr_t pa; 1035 static int color; 1036 int i; 1037 1038 /* 1039 * No need to allocate page table space yet but we do need a valid 1040 * page directory table. 1041 */ 1042 if (pmap->pm_pdir == NULL) { 1043 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1044 NBPTD); 1045#ifdef PAE 1046 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1047 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1048 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1049 ("pmap_pinit: pdpt misaligned")); 1050 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1051 ("pmap_pinit: pdpt above 4g")); 1052#endif 1053 } 1054 1055 /* 1056 * allocate the page directory page(s) 1057 */ 1058 for (i = 0; i < NPGPTD;) { 1059 m = vm_page_alloc(NULL, color++, 1060 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1061 VM_ALLOC_ZERO); 1062 if (m == NULL) 1063 VM_WAIT; 1064 else { 1065 vm_page_lock_queues(); 1066 vm_page_flag_clear(m, PG_BUSY); 1067 m->valid = VM_PAGE_BITS_ALL; 1068 vm_page_unlock_queues(); 1069 ptdpg[i++] = m; 1070 } 1071 } 1072 1073 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1074 1075 for (i = 0; i < NPGPTD; i++) { 1076 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1077 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1078 } 1079 1080 mtx_lock_spin(&allpmaps_lock); 1081 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1082 mtx_unlock_spin(&allpmaps_lock); 1083 /* Wire in kernel global address entries. */ 1084 /* XXX copies current process, does not fill in MPPTDI */ 1085 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1086#ifdef SMP 1087 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1088#endif 1089 1090 /* install self-referential address mapping entry(s) */ 1091 for (i = 0; i < NPGPTD; i++) { 1092 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1093 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1094#ifdef PAE 1095 pmap->pm_pdpt[i] = pa | PG_V; 1096#endif 1097 } 1098 1099 pmap->pm_active = 0; 1100 TAILQ_INIT(&pmap->pm_pvlist); 1101 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1102} 1103 1104/* 1105 * Wire in kernel global address entries. To avoid a race condition 1106 * between pmap initialization and pmap_growkernel, this procedure 1107 * should be called after the vmspace is attached to the process 1108 * but before this pmap is activated. 1109 */ 1110void 1111pmap_pinit2(pmap) 1112 struct pmap *pmap; 1113{ 1114 /* XXX: Remove this stub when no longer called */ 1115} 1116 1117/* 1118 * this routine is called if the page table page is not 1119 * mapped correctly. 1120 */ 1121static vm_page_t 1122_pmap_allocpte(pmap, ptepindex) 1123 pmap_t pmap; 1124 unsigned ptepindex; 1125{ 1126 vm_paddr_t ptepa; 1127 vm_page_t m; 1128 1129 /* 1130 * Allocate a page table page. 1131 */ 1132 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1133 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1134 VM_WAIT; 1135 /* 1136 * Indicate the need to retry. While waiting, the page table 1137 * page may have been allocated. 1138 */ 1139 return (NULL); 1140 } 1141 if ((m->flags & PG_ZERO) == 0) 1142 pmap_zero_page(m); 1143 1144 KASSERT(m->queue == PQ_NONE, 1145 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1146 1147 /* 1148 * Increment the hold count for the page table page 1149 * (denoting a new mapping.) 1150 */ 1151 m->hold_count++; 1152 1153 /* 1154 * Map the pagetable page into the process address space, if 1155 * it isn't already there. 1156 */ 1157 1158 pmap->pm_stats.resident_count++; 1159 1160 ptepa = VM_PAGE_TO_PHYS(m); 1161 pmap->pm_pdir[ptepindex] = 1162 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1163 1164 vm_page_lock_queues(); 1165 m->valid = VM_PAGE_BITS_ALL; 1166 vm_page_flag_clear(m, PG_ZERO); 1167 vm_page_wakeup(m); 1168 vm_page_unlock_queues(); 1169 1170 return m; 1171} 1172 1173static vm_page_t 1174pmap_allocpte(pmap_t pmap, vm_offset_t va) 1175{ 1176 unsigned ptepindex; 1177 pd_entry_t ptepa; 1178 vm_page_t m; 1179 1180 /* 1181 * Calculate pagetable page index 1182 */ 1183 ptepindex = va >> PDRSHIFT; 1184retry: 1185 /* 1186 * Get the page directory entry 1187 */ 1188 ptepa = pmap->pm_pdir[ptepindex]; 1189 1190 /* 1191 * This supports switching from a 4MB page to a 1192 * normal 4K page. 1193 */ 1194 if (ptepa & PG_PS) { 1195 pmap->pm_pdir[ptepindex] = 0; 1196 ptepa = 0; 1197 pmap_invalidate_all(kernel_pmap); 1198 } 1199 1200 /* 1201 * If the page table page is mapped, we just increment the 1202 * hold count, and activate it. 1203 */ 1204 if (ptepa) { 1205 m = PHYS_TO_VM_PAGE(ptepa); 1206 m->hold_count++; 1207 } else { 1208 /* 1209 * Here if the pte page isn't mapped, or if it has 1210 * been deallocated. 1211 */ 1212 m = _pmap_allocpte(pmap, ptepindex); 1213 if (m == NULL) 1214 goto retry; 1215 } 1216 return (m); 1217} 1218 1219 1220/*************************************************** 1221* Pmap allocation/deallocation routines. 1222 ***************************************************/ 1223 1224#ifdef SMP 1225/* 1226 * Deal with a SMP shootdown of other users of the pmap that we are 1227 * trying to dispose of. This can be a bit hairy. 1228 */ 1229static u_int *lazymask; 1230static u_int lazyptd; 1231static volatile u_int lazywait; 1232 1233void pmap_lazyfix_action(void); 1234 1235void 1236pmap_lazyfix_action(void) 1237{ 1238 u_int mymask = PCPU_GET(cpumask); 1239 1240 if (rcr3() == lazyptd) 1241 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1242 atomic_clear_int(lazymask, mymask); 1243 atomic_store_rel_int(&lazywait, 1); 1244} 1245 1246static void 1247pmap_lazyfix_self(u_int mymask) 1248{ 1249 1250 if (rcr3() == lazyptd) 1251 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1252 atomic_clear_int(lazymask, mymask); 1253} 1254 1255 1256static void 1257pmap_lazyfix(pmap_t pmap) 1258{ 1259 u_int mymask = PCPU_GET(cpumask); 1260 u_int mask; 1261 register u_int spins; 1262 1263 while ((mask = pmap->pm_active) != 0) { 1264 spins = 50000000; 1265 mask = mask & -mask; /* Find least significant set bit */ 1266 mtx_lock_spin(&lazypmap_lock); 1267#ifdef PAE 1268 lazyptd = vtophys(pmap->pm_pdpt); 1269#else 1270 lazyptd = vtophys(pmap->pm_pdir); 1271#endif 1272 if (mask == mymask) { 1273 lazymask = &pmap->pm_active; 1274 pmap_lazyfix_self(mymask); 1275 } else { 1276 atomic_store_rel_int((u_int *)&lazymask, 1277 (u_int)&pmap->pm_active); 1278 atomic_store_rel_int(&lazywait, 0); 1279 ipi_selected(mask, IPI_LAZYPMAP); 1280 while (lazywait == 0) { 1281 ia32_pause(); 1282 if (--spins == 0) 1283 break; 1284 } 1285 } 1286 mtx_unlock_spin(&lazypmap_lock); 1287 if (spins == 0) 1288 printf("pmap_lazyfix: spun for 50000000\n"); 1289 } 1290} 1291 1292#else /* SMP */ 1293 1294/* 1295 * Cleaning up on uniprocessor is easy. For various reasons, we're 1296 * unlikely to have to even execute this code, including the fact 1297 * that the cleanup is deferred until the parent does a wait(2), which 1298 * means that another userland process has run. 1299 */ 1300static void 1301pmap_lazyfix(pmap_t pmap) 1302{ 1303 u_int cr3; 1304 1305 cr3 = vtophys(pmap->pm_pdir); 1306 if (cr3 == rcr3()) { 1307 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1308 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1309 } 1310} 1311#endif /* SMP */ 1312 1313/* 1314 * Release any resources held by the given physical map. 1315 * Called when a pmap initialized by pmap_pinit is being released. 1316 * Should only be called if the map contains no valid mappings. 1317 */ 1318void 1319pmap_release(pmap_t pmap) 1320{ 1321 vm_page_t m, ptdpg[NPGPTD]; 1322 int i; 1323 1324 KASSERT(pmap->pm_stats.resident_count == 0, 1325 ("pmap_release: pmap resident count %ld != 0", 1326 pmap->pm_stats.resident_count)); 1327 1328 pmap_lazyfix(pmap); 1329 mtx_lock_spin(&allpmaps_lock); 1330 LIST_REMOVE(pmap, pm_list); 1331 mtx_unlock_spin(&allpmaps_lock); 1332 1333 for (i = 0; i < NPGPTD; i++) 1334 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1335 1336 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1337 sizeof(*pmap->pm_pdir)); 1338#ifdef SMP 1339 pmap->pm_pdir[MPPTDI] = 0; 1340#endif 1341 1342 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1343 1344 vm_page_lock_queues(); 1345 for (i = 0; i < NPGPTD; i++) { 1346 m = ptdpg[i]; 1347#ifdef PAE 1348 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1349 ("pmap_release: got wrong ptd page")); 1350#endif 1351 m->wire_count--; 1352 atomic_subtract_int(&cnt.v_wire_count, 1); 1353 vm_page_busy(m); 1354 vm_page_free_zero(m); 1355 } 1356 vm_page_unlock_queues(); 1357} 1358 1359static int 1360kvm_size(SYSCTL_HANDLER_ARGS) 1361{ 1362 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1363 1364 return sysctl_handle_long(oidp, &ksize, 0, req); 1365} 1366SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1367 0, 0, kvm_size, "IU", "Size of KVM"); 1368 1369static int 1370kvm_free(SYSCTL_HANDLER_ARGS) 1371{ 1372 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1373 1374 return sysctl_handle_long(oidp, &kfree, 0, req); 1375} 1376SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1377 0, 0, kvm_free, "IU", "Amount of KVM free"); 1378 1379/* 1380 * grow the number of kernel page table entries, if needed 1381 */ 1382void 1383pmap_growkernel(vm_offset_t addr) 1384{ 1385 struct pmap *pmap; 1386 int s; 1387 vm_paddr_t ptppaddr; 1388 vm_page_t nkpg; 1389 pd_entry_t newpdir; 1390 pt_entry_t *pde; 1391 1392 s = splhigh(); 1393 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1394 if (kernel_vm_end == 0) { 1395 kernel_vm_end = KERNBASE; 1396 nkpt = 0; 1397 while (pdir_pde(PTD, kernel_vm_end)) { 1398 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1399 nkpt++; 1400 } 1401 } 1402 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1403 while (kernel_vm_end < addr) { 1404 if (pdir_pde(PTD, kernel_vm_end)) { 1405 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1406 continue; 1407 } 1408 1409 /* 1410 * This index is bogus, but out of the way 1411 */ 1412 nkpg = vm_page_alloc(NULL, nkpt, 1413 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1414 if (!nkpg) 1415 panic("pmap_growkernel: no memory to grow kernel"); 1416 1417 nkpt++; 1418 1419 pmap_zero_page(nkpg); 1420 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1421 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1422 pdir_pde(PTD, kernel_vm_end) = newpdir; 1423 1424 mtx_lock_spin(&allpmaps_lock); 1425 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1426 pde = pmap_pde(pmap, kernel_vm_end); 1427 pde_store(pde, newpdir); 1428 } 1429 mtx_unlock_spin(&allpmaps_lock); 1430 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1431 } 1432 splx(s); 1433} 1434 1435 1436/*************************************************** 1437 * page management routines. 1438 ***************************************************/ 1439 1440/* 1441 * free the pv_entry back to the free list 1442 */ 1443static PMAP_INLINE void 1444free_pv_entry(pv_entry_t pv) 1445{ 1446 pv_entry_count--; 1447 uma_zfree(pvzone, pv); 1448} 1449 1450/* 1451 * get a new pv_entry, allocating a block from the system 1452 * when needed. 1453 * the memory allocation is performed bypassing the malloc code 1454 * because of the possibility of allocations at interrupt time. 1455 */ 1456static pv_entry_t 1457get_pv_entry(void) 1458{ 1459 pv_entry_count++; 1460 if (pv_entry_high_water && 1461 (pv_entry_count > pv_entry_high_water) && 1462 (pmap_pagedaemon_waken == 0)) { 1463 pmap_pagedaemon_waken = 1; 1464 wakeup (&vm_pages_needed); 1465 } 1466 return uma_zalloc(pvzone, M_NOWAIT); 1467} 1468 1469/* 1470 * If it is the first entry on the list, it is actually 1471 * in the header and we must copy the following entry up 1472 * to the header. Otherwise we must search the list for 1473 * the entry. In either case we free the now unused entry. 1474 */ 1475 1476static int 1477pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1478{ 1479 pv_entry_t pv; 1480 int rtval; 1481 int s; 1482 1483 s = splvm(); 1484 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1485 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1486 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1487 if (pmap == pv->pv_pmap && va == pv->pv_va) 1488 break; 1489 } 1490 } else { 1491 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1492 if (va == pv->pv_va) 1493 break; 1494 } 1495 } 1496 1497 rtval = 0; 1498 if (pv) { 1499 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1500 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1501 m->md.pv_list_count--; 1502 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1503 vm_page_flag_clear(m, PG_WRITEABLE); 1504 1505 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1506 free_pv_entry(pv); 1507 } 1508 1509 splx(s); 1510 return rtval; 1511} 1512 1513/* 1514 * Create a pv entry for page at pa for 1515 * (pmap, va). 1516 */ 1517static void 1518pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1519{ 1520 1521 int s; 1522 pv_entry_t pv; 1523 1524 s = splvm(); 1525 pv = get_pv_entry(); 1526 pv->pv_va = va; 1527 pv->pv_pmap = pmap; 1528 pv->pv_ptem = mpte; 1529 1530 vm_page_lock_queues(); 1531 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1532 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1533 m->md.pv_list_count++; 1534 1535 vm_page_unlock_queues(); 1536 splx(s); 1537} 1538 1539/* 1540 * pmap_remove_pte: do the things to unmap a page in a process 1541 */ 1542static int 1543pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1544{ 1545 pt_entry_t oldpte; 1546 vm_page_t m, mpte; 1547 1548 oldpte = pte_load_clear(ptq); 1549 if (oldpte & PG_W) 1550 pmap->pm_stats.wired_count -= 1; 1551 /* 1552 * Machines that don't support invlpg, also don't support 1553 * PG_G. 1554 */ 1555 if (oldpte & PG_G) 1556 pmap_invalidate_page(kernel_pmap, va); 1557 pmap->pm_stats.resident_count -= 1; 1558 if (oldpte & PG_MANAGED) { 1559 m = PHYS_TO_VM_PAGE(oldpte); 1560 if (oldpte & PG_M) { 1561#if defined(PMAP_DIAGNOSTIC) 1562 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1563 printf( 1564 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1565 va, oldpte); 1566 } 1567#endif 1568 if (pmap_track_modified(va)) 1569 vm_page_dirty(m); 1570 } 1571 if (oldpte & PG_A) 1572 vm_page_flag_set(m, PG_REFERENCED); 1573 return pmap_remove_entry(pmap, m, va); 1574 } else { 1575 mpte = PHYS_TO_VM_PAGE(*pmap_pde(pmap, va)); 1576 return pmap_unuse_pt(pmap, va, mpte); 1577 } 1578} 1579 1580/* 1581 * Remove a single page from a process address space 1582 */ 1583static void 1584pmap_remove_page(pmap_t pmap, vm_offset_t va) 1585{ 1586 pt_entry_t *pte; 1587 1588 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) 1589 return; 1590 pmap_remove_pte(pmap, pte, va); 1591 pmap_invalidate_page(pmap, va); 1592} 1593 1594/* 1595 * Remove the given range of addresses from the specified map. 1596 * 1597 * It is assumed that the start and end are properly 1598 * rounded to the page size. 1599 */ 1600void 1601pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1602{ 1603 vm_offset_t pdnxt; 1604 pd_entry_t ptpaddr; 1605 pt_entry_t *pte; 1606 int anyvalid; 1607 1608 if (pmap == NULL) 1609 return; 1610 1611 if (pmap->pm_stats.resident_count == 0) 1612 return; 1613 1614 /* 1615 * special handling of removing one page. a very 1616 * common operation and easy to short circuit some 1617 * code. 1618 */ 1619 if ((sva + PAGE_SIZE == eva) && 1620 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1621 pmap_remove_page(pmap, sva); 1622 return; 1623 } 1624 1625 anyvalid = 0; 1626 1627 for (; sva < eva; sva = pdnxt) { 1628 unsigned pdirindex; 1629 1630 /* 1631 * Calculate index for next page table. 1632 */ 1633 pdnxt = (sva + NBPDR) & ~PDRMASK; 1634 if (pmap->pm_stats.resident_count == 0) 1635 break; 1636 1637 pdirindex = sva >> PDRSHIFT; 1638 ptpaddr = pmap->pm_pdir[pdirindex]; 1639 1640 /* 1641 * Weed out invalid mappings. Note: we assume that the page 1642 * directory table is always allocated, and in kernel virtual. 1643 */ 1644 if (ptpaddr == 0) 1645 continue; 1646 1647 /* 1648 * Check for large page. 1649 */ 1650 if ((ptpaddr & PG_PS) != 0) { 1651 pmap->pm_pdir[pdirindex] = 0; 1652 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1653 anyvalid = 1; 1654 continue; 1655 } 1656 1657 /* 1658 * Limit our scan to either the end of the va represented 1659 * by the current page table page, or to the end of the 1660 * range being removed. 1661 */ 1662 if (pdnxt > eva) 1663 pdnxt = eva; 1664 1665 for (; sva != pdnxt; sva += PAGE_SIZE) { 1666 if ((pte = pmap_pte_quick(pmap, sva)) == NULL || 1667 *pte == 0) 1668 continue; 1669 anyvalid = 1; 1670 if (pmap_remove_pte(pmap, pte, sva)) 1671 break; 1672 } 1673 } 1674 1675 if (anyvalid) 1676 pmap_invalidate_all(pmap); 1677} 1678 1679/* 1680 * Routine: pmap_remove_all 1681 * Function: 1682 * Removes this physical page from 1683 * all physical maps in which it resides. 1684 * Reflects back modify bits to the pager. 1685 * 1686 * Notes: 1687 * Original versions of this routine were very 1688 * inefficient because they iteratively called 1689 * pmap_remove (slow...) 1690 */ 1691 1692void 1693pmap_remove_all(vm_page_t m) 1694{ 1695 register pv_entry_t pv; 1696 pt_entry_t *pte, tpte; 1697 int s; 1698 1699#if defined(PMAP_DIAGNOSTIC) 1700 /* 1701 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1702 */ 1703 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1704 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1705 VM_PAGE_TO_PHYS(m)); 1706 } 1707#endif 1708 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1709 s = splvm(); 1710 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1711 pv->pv_pmap->pm_stats.resident_count--; 1712 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1713 tpte = pte_load_clear(pte); 1714 if (tpte & PG_W) 1715 pv->pv_pmap->pm_stats.wired_count--; 1716 if (tpte & PG_A) 1717 vm_page_flag_set(m, PG_REFERENCED); 1718 1719 /* 1720 * Update the vm_page_t clean and reference bits. 1721 */ 1722 if (tpte & PG_M) { 1723#if defined(PMAP_DIAGNOSTIC) 1724 if (pmap_nw_modified((pt_entry_t) tpte)) { 1725 printf( 1726 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1727 pv->pv_va, tpte); 1728 } 1729#endif 1730 if (pmap_track_modified(pv->pv_va)) 1731 vm_page_dirty(m); 1732 } 1733 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1734 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1735 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1736 m->md.pv_list_count--; 1737 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 1738 free_pv_entry(pv); 1739 } 1740 vm_page_flag_clear(m, PG_WRITEABLE); 1741 splx(s); 1742} 1743 1744/* 1745 * Set the physical protection on the 1746 * specified range of this map as requested. 1747 */ 1748void 1749pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1750{ 1751 vm_offset_t pdnxt; 1752 pd_entry_t ptpaddr; 1753 int anychanged; 1754 1755 if (pmap == NULL) 1756 return; 1757 1758 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1759 pmap_remove(pmap, sva, eva); 1760 return; 1761 } 1762 1763 if (prot & VM_PROT_WRITE) 1764 return; 1765 1766 anychanged = 0; 1767 1768 for (; sva < eva; sva = pdnxt) { 1769 unsigned pdirindex; 1770 1771 pdnxt = (sva + NBPDR) & ~PDRMASK; 1772 1773 pdirindex = sva >> PDRSHIFT; 1774 ptpaddr = pmap->pm_pdir[pdirindex]; 1775 1776 /* 1777 * Weed out invalid mappings. Note: we assume that the page 1778 * directory table is always allocated, and in kernel virtual. 1779 */ 1780 if (ptpaddr == 0) 1781 continue; 1782 1783 /* 1784 * Check for large page. 1785 */ 1786 if ((ptpaddr & PG_PS) != 0) { 1787 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1788 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1789 anychanged = 1; 1790 continue; 1791 } 1792 1793 if (pdnxt > eva) 1794 pdnxt = eva; 1795 1796 for (; sva != pdnxt; sva += PAGE_SIZE) { 1797 pt_entry_t pbits; 1798 pt_entry_t *pte; 1799 vm_page_t m; 1800 1801 if ((pte = pmap_pte_quick(pmap, sva)) == NULL) 1802 continue; 1803 pbits = *pte; 1804 if (pbits & PG_MANAGED) { 1805 m = NULL; 1806 if (pbits & PG_A) { 1807 m = PHYS_TO_VM_PAGE(pbits); 1808 vm_page_flag_set(m, PG_REFERENCED); 1809 pbits &= ~PG_A; 1810 } 1811 if ((pbits & PG_M) != 0 && 1812 pmap_track_modified(sva)) { 1813 if (m == NULL) 1814 m = PHYS_TO_VM_PAGE(pbits); 1815 vm_page_dirty(m); 1816 pbits &= ~PG_M; 1817 } 1818 } 1819 1820 pbits &= ~PG_RW; 1821 1822 if (pbits != *pte) { 1823 pte_store(pte, pbits); 1824 anychanged = 1; 1825 } 1826 } 1827 } 1828 if (anychanged) 1829 pmap_invalidate_all(pmap); 1830} 1831 1832/* 1833 * Insert the given physical page (p) at 1834 * the specified virtual address (v) in the 1835 * target physical map with the protection requested. 1836 * 1837 * If specified, the page will be wired down, meaning 1838 * that the related pte can not be reclaimed. 1839 * 1840 * NB: This is the only routine which MAY NOT lazy-evaluate 1841 * or lose information. That is, this routine must actually 1842 * insert this page into the given map NOW. 1843 */ 1844void 1845pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1846 boolean_t wired) 1847{ 1848 vm_paddr_t pa; 1849 register pt_entry_t *pte; 1850 vm_paddr_t opa; 1851 pt_entry_t origpte, newpte; 1852 vm_page_t mpte; 1853 1854 if (pmap == NULL) 1855 return; 1856 1857 va &= PG_FRAME; 1858#ifdef PMAP_DIAGNOSTIC 1859 if (va > VM_MAX_KERNEL_ADDRESS) 1860 panic("pmap_enter: toobig"); 1861 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1862 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1863#endif 1864 1865 mpte = NULL; 1866 /* 1867 * In the case that a page table page is not 1868 * resident, we are creating it here. 1869 */ 1870 if (va < VM_MAXUSER_ADDRESS) { 1871 mpte = pmap_allocpte(pmap, va); 1872 } 1873#if 0 && defined(PMAP_DIAGNOSTIC) 1874 else { 1875 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1876 origpte = *pdeaddr; 1877 if ((origpte & PG_V) == 0) { 1878 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1879 pmap->pm_pdir[PTDPTDI], origpte, va); 1880 } 1881 } 1882#endif 1883 1884 pte = pmap_pte_quick(pmap, va); 1885 1886 /* 1887 * Page Directory table entry not valid, we need a new PT page 1888 */ 1889 if (pte == NULL) { 1890 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1891 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1892 } 1893 1894 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 1895 origpte = *pte; 1896 opa = origpte & PG_FRAME; 1897 1898 if (origpte & PG_PS) 1899 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1900 1901 /* 1902 * Mapping has not changed, must be protection or wiring change. 1903 */ 1904 if (origpte && (opa == pa)) { 1905 /* 1906 * Wiring change, just update stats. We don't worry about 1907 * wiring PT pages as they remain resident as long as there 1908 * are valid mappings in them. Hence, if a user page is wired, 1909 * the PT page will be also. 1910 */ 1911 if (wired && ((origpte & PG_W) == 0)) 1912 pmap->pm_stats.wired_count++; 1913 else if (!wired && (origpte & PG_W)) 1914 pmap->pm_stats.wired_count--; 1915 1916#if defined(PMAP_DIAGNOSTIC) 1917 if (pmap_nw_modified((pt_entry_t) origpte)) { 1918 printf( 1919 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 1920 va, origpte); 1921 } 1922#endif 1923 1924 /* 1925 * Remove extra pte reference 1926 */ 1927 if (mpte) 1928 mpte->hold_count--; 1929 1930 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) { 1931 if ((origpte & PG_RW) == 0) { 1932 pte_store(pte, origpte | PG_RW); 1933 pmap_invalidate_page(pmap, va); 1934 } 1935 return; 1936 } 1937 1938 /* 1939 * We might be turning off write access to the page, 1940 * so we go ahead and sense modify status. 1941 */ 1942 if (origpte & PG_MANAGED) { 1943 if ((origpte & PG_M) && pmap_track_modified(va)) { 1944 vm_page_t om; 1945 om = PHYS_TO_VM_PAGE(opa); 1946 vm_page_dirty(om); 1947 } 1948 pa |= PG_MANAGED; 1949 } 1950 goto validate; 1951 } 1952 /* 1953 * Mapping has changed, invalidate old range and fall through to 1954 * handle validating new mapping. 1955 */ 1956 if (opa) { 1957 int err; 1958 vm_page_lock_queues(); 1959 err = pmap_remove_pte(pmap, pte, va); 1960 vm_page_unlock_queues(); 1961 if (err) 1962 panic("pmap_enter: pte vanished, va: 0x%x", va); 1963 } 1964 1965 /* 1966 * Enter on the PV list if part of our managed memory. Note that we 1967 * raise IPL while manipulating pv_table since pmap_enter can be 1968 * called at interrupt time. 1969 */ 1970 if (pmap_initialized && 1971 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 1972 pmap_insert_entry(pmap, va, mpte, m); 1973 pa |= PG_MANAGED; 1974 } 1975 1976 /* 1977 * Increment counters 1978 */ 1979 pmap->pm_stats.resident_count++; 1980 if (wired) 1981 pmap->pm_stats.wired_count++; 1982 1983validate: 1984 /* 1985 * Now validate mapping with desired protection/wiring. 1986 */ 1987 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V); 1988 1989 if (wired) 1990 newpte |= PG_W; 1991 if (va < VM_MAXUSER_ADDRESS) 1992 newpte |= PG_U; 1993 if (pmap == kernel_pmap) 1994 newpte |= pgeflag; 1995 1996 /* 1997 * if the mapping or permission bits are different, we need 1998 * to update the pte. 1999 */ 2000 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2001 pte_store(pte, newpte | PG_A); 2002 /*if (origpte)*/ { 2003 pmap_invalidate_page(pmap, va); 2004 } 2005 } 2006} 2007 2008/* 2009 * this code makes some *MAJOR* assumptions: 2010 * 1. Current pmap & pmap exists. 2011 * 2. Not wired. 2012 * 3. Read access. 2013 * 4. No page table pages. 2014 * 5. Tlbflush is deferred to calling procedure. 2015 * 6. Page IS managed. 2016 * but is *MUCH* faster than pmap_enter... 2017 */ 2018 2019vm_page_t 2020pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2021{ 2022 pt_entry_t *pte; 2023 vm_paddr_t pa; 2024 2025 /* 2026 * In the case that a page table page is not 2027 * resident, we are creating it here. 2028 */ 2029 if (va < VM_MAXUSER_ADDRESS) { 2030 unsigned ptepindex; 2031 pd_entry_t ptepa; 2032 2033 /* 2034 * Calculate pagetable page index 2035 */ 2036 ptepindex = va >> PDRSHIFT; 2037 if (mpte && (mpte->pindex == ptepindex)) { 2038 mpte->hold_count++; 2039 } else { 2040retry: 2041 /* 2042 * Get the page directory entry 2043 */ 2044 ptepa = pmap->pm_pdir[ptepindex]; 2045 2046 /* 2047 * If the page table page is mapped, we just increment 2048 * the hold count, and activate it. 2049 */ 2050 if (ptepa) { 2051 if (ptepa & PG_PS) 2052 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2053 mpte = PHYS_TO_VM_PAGE(ptepa); 2054 mpte->hold_count++; 2055 } else { 2056 mpte = _pmap_allocpte(pmap, ptepindex); 2057 if (mpte == NULL) 2058 goto retry; 2059 } 2060 } 2061 } else { 2062 mpte = NULL; 2063 } 2064 2065 /* 2066 * This call to vtopte makes the assumption that we are 2067 * entering the page into the current pmap. In order to support 2068 * quick entry into any pmap, one would likely use pmap_pte_quick. 2069 * But that isn't as quick as vtopte. 2070 */ 2071 pte = vtopte(va); 2072 if (*pte) { 2073 if (mpte != NULL) { 2074 vm_page_lock_queues(); 2075 pmap_unwire_pte_hold(pmap, mpte); 2076 vm_page_unlock_queues(); 2077 } 2078 return 0; 2079 } 2080 2081 /* 2082 * Enter on the PV list if part of our managed memory. Note that we 2083 * raise IPL while manipulating pv_table since pmap_enter can be 2084 * called at interrupt time. 2085 */ 2086 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2087 pmap_insert_entry(pmap, va, mpte, m); 2088 2089 /* 2090 * Increment counters 2091 */ 2092 pmap->pm_stats.resident_count++; 2093 2094 pa = VM_PAGE_TO_PHYS(m); 2095 2096 /* 2097 * Now validate mapping with RO protection 2098 */ 2099 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2100 pte_store(pte, pa | PG_V | PG_U); 2101 else 2102 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2103 2104 return mpte; 2105} 2106 2107/* 2108 * Make a temporary mapping for a physical address. This is only intended 2109 * to be used for panic dumps. 2110 */ 2111void * 2112pmap_kenter_temporary(vm_offset_t pa, int i) 2113{ 2114 vm_offset_t va; 2115 2116 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2117 pmap_kenter(va, pa); 2118#ifndef I386_CPU 2119 invlpg(va); 2120#else 2121 invltlb(); 2122#endif 2123 return ((void *)crashdumpmap); 2124} 2125 2126/* 2127 * This code maps large physical mmap regions into the 2128 * processor address space. Note that some shortcuts 2129 * are taken, but the code works. 2130 */ 2131void 2132pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2133 vm_object_t object, vm_pindex_t pindex, 2134 vm_size_t size) 2135{ 2136 vm_page_t p; 2137 2138 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2139 KASSERT(object->type == OBJT_DEVICE, 2140 ("pmap_object_init_pt: non-device object")); 2141 if (pseflag && 2142 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2143 int i; 2144 vm_page_t m[1]; 2145 unsigned int ptepindex; 2146 int npdes; 2147 pd_entry_t ptepa; 2148 2149 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2150 return; 2151retry: 2152 p = vm_page_lookup(object, pindex); 2153 if (p != NULL) { 2154 vm_page_lock_queues(); 2155 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2156 goto retry; 2157 } else { 2158 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2159 if (p == NULL) 2160 return; 2161 m[0] = p; 2162 2163 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2164 vm_page_lock_queues(); 2165 vm_page_free(p); 2166 vm_page_unlock_queues(); 2167 return; 2168 } 2169 2170 p = vm_page_lookup(object, pindex); 2171 vm_page_lock_queues(); 2172 vm_page_wakeup(p); 2173 } 2174 vm_page_unlock_queues(); 2175 2176 ptepa = VM_PAGE_TO_PHYS(p); 2177 if (ptepa & (NBPDR - 1)) 2178 return; 2179 2180 p->valid = VM_PAGE_BITS_ALL; 2181 2182 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2183 npdes = size >> PDRSHIFT; 2184 for(i = 0; i < npdes; i++) { 2185 pde_store(&pmap->pm_pdir[ptepindex], 2186 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2187 ptepa += NBPDR; 2188 ptepindex += 1; 2189 } 2190 pmap_invalidate_all(pmap); 2191 } 2192} 2193 2194/* 2195 * Routine: pmap_change_wiring 2196 * Function: Change the wiring attribute for a map/virtual-address 2197 * pair. 2198 * In/out conditions: 2199 * The mapping must already exist in the pmap. 2200 */ 2201void 2202pmap_change_wiring(pmap, va, wired) 2203 register pmap_t pmap; 2204 vm_offset_t va; 2205 boolean_t wired; 2206{ 2207 register pt_entry_t *pte; 2208 2209 if (pmap == NULL) 2210 return; 2211 2212 pte = pmap_pte_quick(pmap, va); 2213 2214 if (wired && !pmap_pte_w(pte)) 2215 pmap->pm_stats.wired_count++; 2216 else if (!wired && pmap_pte_w(pte)) 2217 pmap->pm_stats.wired_count--; 2218 2219 /* 2220 * Wiring is not a hardware characteristic so there is no need to 2221 * invalidate TLB. 2222 */ 2223 pmap_pte_set_w(pte, wired); 2224} 2225 2226 2227 2228/* 2229 * Copy the range specified by src_addr/len 2230 * from the source map to the range dst_addr/len 2231 * in the destination map. 2232 * 2233 * This routine is only advisory and need not do anything. 2234 */ 2235 2236void 2237pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2238 vm_offset_t src_addr) 2239{ 2240 vm_offset_t addr; 2241 vm_offset_t end_addr = src_addr + len; 2242 vm_offset_t pdnxt; 2243 vm_page_t m; 2244 2245 if (dst_addr != src_addr) 2246 return; 2247 2248 if (!pmap_is_current(src_pmap)) 2249 return; 2250 2251 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2252 pt_entry_t *src_pte, *dst_pte; 2253 vm_page_t dstmpte, srcmpte; 2254 pd_entry_t srcptepaddr; 2255 unsigned ptepindex; 2256 2257 if (addr >= UPT_MIN_ADDRESS) 2258 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2259 2260 /* 2261 * Don't let optional prefaulting of pages make us go 2262 * way below the low water mark of free pages or way 2263 * above high water mark of used pv entries. 2264 */ 2265 if (cnt.v_free_count < cnt.v_free_reserved || 2266 pv_entry_count > pv_entry_high_water) 2267 break; 2268 2269 pdnxt = (addr + NBPDR) & ~PDRMASK; 2270 ptepindex = addr >> PDRSHIFT; 2271 2272 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2273 if (srcptepaddr == 0) 2274 continue; 2275 2276 if (srcptepaddr & PG_PS) { 2277 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2278 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2279 dst_pmap->pm_stats.resident_count += 2280 NBPDR / PAGE_SIZE; 2281 } 2282 continue; 2283 } 2284 2285 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2286 if (srcmpte->hold_count == 0 || (srcmpte->flags & PG_BUSY)) 2287 continue; 2288 2289 if (pdnxt > end_addr) 2290 pdnxt = end_addr; 2291 2292 src_pte = vtopte(addr); 2293 while (addr < pdnxt) { 2294 pt_entry_t ptetemp; 2295 ptetemp = *src_pte; 2296 /* 2297 * we only virtual copy managed pages 2298 */ 2299 if ((ptetemp & PG_MANAGED) != 0) { 2300 /* 2301 * We have to check after allocpte for the 2302 * pte still being around... allocpte can 2303 * block. 2304 */ 2305 dstmpte = pmap_allocpte(dst_pmap, addr); 2306 dst_pte = pmap_pte_quick(dst_pmap, addr); 2307 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2308 /* 2309 * Clear the modified and 2310 * accessed (referenced) bits 2311 * during the copy. 2312 */ 2313 m = PHYS_TO_VM_PAGE(ptetemp); 2314 *dst_pte = ptetemp & ~(PG_M | PG_A); 2315 dst_pmap->pm_stats.resident_count++; 2316 pmap_insert_entry(dst_pmap, addr, 2317 dstmpte, m); 2318 } else { 2319 vm_page_lock_queues(); 2320 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2321 vm_page_unlock_queues(); 2322 } 2323 if (dstmpte->hold_count >= srcmpte->hold_count) 2324 break; 2325 } 2326 addr += PAGE_SIZE; 2327 src_pte++; 2328 } 2329 } 2330} 2331 2332#ifdef SMP 2333 2334/* 2335 * pmap_zpi_switchout*() 2336 * 2337 * These functions allow us to avoid doing IPIs alltogether in certain 2338 * temporary page-mapping situations (page zeroing). Instead to deal 2339 * with being preempted and moved onto a different cpu we invalidate 2340 * the page when the scheduler switches us in. This does not occur 2341 * very often so we remain relatively optimal with very little effort. 2342 */ 2343static void 2344pmap_zpi_switchout12(void) 2345{ 2346 invlpg((u_int)CADDR1); 2347 invlpg((u_int)CADDR2); 2348} 2349 2350static void 2351pmap_zpi_switchout2(void) 2352{ 2353 invlpg((u_int)CADDR2); 2354} 2355 2356static void 2357pmap_zpi_switchout3(void) 2358{ 2359 invlpg((u_int)CADDR3); 2360} 2361 2362#endif 2363 2364static __inline void 2365pagezero(void *page) 2366{ 2367#if defined(I686_CPU) 2368 if (cpu_class == CPUCLASS_686) { 2369#if defined(CPU_ENABLE_SSE) 2370 if (cpu_feature & CPUID_SSE2) 2371 sse2_pagezero(page); 2372 else 2373#endif 2374 i686_pagezero(page); 2375 } else 2376#endif 2377 bzero(page, PAGE_SIZE); 2378} 2379 2380static __inline void 2381invlcaddr(void *caddr) 2382{ 2383#ifdef I386_CPU 2384 invltlb(); 2385#else 2386 invlpg((u_int)caddr); 2387#endif 2388} 2389 2390/* 2391 * pmap_zero_page zeros the specified hardware page by mapping 2392 * the page into KVM and using bzero to clear its contents. 2393 */ 2394void 2395pmap_zero_page(vm_page_t m) 2396{ 2397 2398 mtx_lock(&CMAPCADDR12_lock); 2399 if (*CMAP2) 2400 panic("pmap_zero_page: CMAP2 busy"); 2401#ifdef SMP 2402 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2; 2403#endif 2404 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2405 pagezero(CADDR2); 2406 *CMAP2 = 0; 2407 invlcaddr(CADDR2); 2408#ifdef SMP 2409 curthread->td_pcb->pcb_switchout = NULL; 2410#endif 2411 mtx_unlock(&CMAPCADDR12_lock); 2412} 2413 2414/* 2415 * pmap_zero_page_area zeros the specified hardware page by mapping 2416 * the page into KVM and using bzero to clear its contents. 2417 * 2418 * off and size may not cover an area beyond a single hardware page. 2419 */ 2420void 2421pmap_zero_page_area(vm_page_t m, int off, int size) 2422{ 2423 2424 mtx_lock(&CMAPCADDR12_lock); 2425 if (*CMAP2) 2426 panic("pmap_zero_page: CMAP2 busy"); 2427#ifdef SMP 2428 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2; 2429#endif 2430 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2431 if (off == 0 && size == PAGE_SIZE) 2432 pagezero(CADDR2); 2433 else 2434 bzero((char *)CADDR2 + off, size); 2435 *CMAP2 = 0; 2436 invlcaddr(CADDR2); 2437#ifdef SMP 2438 curthread->td_pcb->pcb_switchout = NULL; 2439#endif 2440 mtx_unlock(&CMAPCADDR12_lock); 2441} 2442 2443/* 2444 * pmap_zero_page_idle zeros the specified hardware page by mapping 2445 * the page into KVM and using bzero to clear its contents. This 2446 * is intended to be called from the vm_pagezero process only and 2447 * outside of Giant. 2448 */ 2449void 2450pmap_zero_page_idle(vm_page_t m) 2451{ 2452 2453 if (*CMAP3) 2454 panic("pmap_zero_page: CMAP3 busy"); 2455#ifdef SMP 2456 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout3; 2457#endif 2458 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2459 pagezero(CADDR3); 2460 *CMAP3 = 0; 2461 invlcaddr(CADDR3); 2462#ifdef SMP 2463 curthread->td_pcb->pcb_switchout = NULL; 2464#endif 2465} 2466 2467/* 2468 * pmap_copy_page copies the specified (machine independent) 2469 * page by mapping the page into virtual memory and using 2470 * bcopy to copy the page, one machine dependent page at a 2471 * time. 2472 */ 2473void 2474pmap_copy_page(vm_page_t src, vm_page_t dst) 2475{ 2476 2477 mtx_lock(&CMAPCADDR12_lock); 2478 if (*CMAP1) 2479 panic("pmap_copy_page: CMAP1 busy"); 2480 if (*CMAP2) 2481 panic("pmap_copy_page: CMAP2 busy"); 2482#ifdef SMP 2483 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout12; 2484#endif 2485 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2486 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2487 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2488 *CMAP1 = 0; 2489 *CMAP2 = 0; 2490#ifdef I386_CPU 2491 invltlb(); 2492#else 2493 invlpg((u_int)CADDR1); 2494 invlpg((u_int)CADDR2); 2495#endif 2496#ifdef SMP 2497 curthread->td_pcb->pcb_switchout = NULL; 2498#endif 2499 mtx_unlock(&CMAPCADDR12_lock); 2500} 2501 2502/* 2503 * Returns true if the pmap's pv is one of the first 2504 * 16 pvs linked to from this page. This count may 2505 * be changed upwards or downwards in the future; it 2506 * is only necessary that true be returned for a small 2507 * subset of pmaps for proper page aging. 2508 */ 2509boolean_t 2510pmap_page_exists_quick(pmap, m) 2511 pmap_t pmap; 2512 vm_page_t m; 2513{ 2514 pv_entry_t pv; 2515 int loops = 0; 2516 int s; 2517 2518 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2519 return FALSE; 2520 2521 s = splvm(); 2522 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2523 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2524 if (pv->pv_pmap == pmap) { 2525 splx(s); 2526 return TRUE; 2527 } 2528 loops++; 2529 if (loops >= 16) 2530 break; 2531 } 2532 splx(s); 2533 return (FALSE); 2534} 2535 2536#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2537/* 2538 * Remove all pages from specified address space 2539 * this aids process exit speeds. Also, this code 2540 * is special cased for current process only, but 2541 * can have the more generic (and slightly slower) 2542 * mode enabled. This is much faster than pmap_remove 2543 * in the case of running down an entire address space. 2544 */ 2545void 2546pmap_remove_pages(pmap, sva, eva) 2547 pmap_t pmap; 2548 vm_offset_t sva, eva; 2549{ 2550 pt_entry_t *pte, tpte; 2551 vm_page_t m; 2552 pv_entry_t pv, npv; 2553 int s; 2554 2555#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2556 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2557 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2558 return; 2559 } 2560#endif 2561 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2562 s = splvm(); 2563 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2564 2565 if (pv->pv_va >= eva || pv->pv_va < sva) { 2566 npv = TAILQ_NEXT(pv, pv_plist); 2567 continue; 2568 } 2569 2570#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2571 pte = vtopte(pv->pv_va); 2572#else 2573 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2574#endif 2575 tpte = *pte; 2576 2577 if (tpte == 0) { 2578 printf("TPTE at %p IS ZERO @ VA %08x\n", 2579 pte, pv->pv_va); 2580 panic("bad pte"); 2581 } 2582 2583/* 2584 * We cannot remove wired pages from a process' mapping at this time 2585 */ 2586 if (tpte & PG_W) { 2587 npv = TAILQ_NEXT(pv, pv_plist); 2588 continue; 2589 } 2590 2591 m = PHYS_TO_VM_PAGE(tpte); 2592 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2593 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2594 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2595 2596 KASSERT(m < &vm_page_array[vm_page_array_size], 2597 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2598 2599 pv->pv_pmap->pm_stats.resident_count--; 2600 2601 pte_clear(pte); 2602 2603 /* 2604 * Update the vm_page_t clean and reference bits. 2605 */ 2606 if (tpte & PG_M) { 2607 vm_page_dirty(m); 2608 } 2609 2610 npv = TAILQ_NEXT(pv, pv_plist); 2611 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2612 2613 m->md.pv_list_count--; 2614 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2615 if (TAILQ_FIRST(&m->md.pv_list) == NULL) { 2616 vm_page_flag_clear(m, PG_WRITEABLE); 2617 } 2618 2619 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2620 free_pv_entry(pv); 2621 } 2622 splx(s); 2623 pmap_invalidate_all(pmap); 2624} 2625 2626/* 2627 * pmap_is_modified: 2628 * 2629 * Return whether or not the specified physical page was modified 2630 * in any physical maps. 2631 */ 2632boolean_t 2633pmap_is_modified(vm_page_t m) 2634{ 2635 pv_entry_t pv; 2636 pt_entry_t *pte; 2637 int s; 2638 2639 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2640 return FALSE; 2641 2642 s = splvm(); 2643 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2644 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2645 /* 2646 * if the bit being tested is the modified bit, then 2647 * mark clean_map and ptes as never 2648 * modified. 2649 */ 2650 if (!pmap_track_modified(pv->pv_va)) 2651 continue; 2652#if defined(PMAP_DIAGNOSTIC) 2653 if (!pv->pv_pmap) { 2654 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 2655 continue; 2656 } 2657#endif 2658 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2659 if (*pte & PG_M) { 2660 splx(s); 2661 return TRUE; 2662 } 2663 } 2664 splx(s); 2665 return (FALSE); 2666} 2667 2668/* 2669 * pmap_is_prefaultable: 2670 * 2671 * Return whether or not the specified virtual address is elgible 2672 * for prefault. 2673 */ 2674boolean_t 2675pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2676{ 2677 pt_entry_t *pte; 2678 2679 if ((*pmap_pde(pmap, addr)) == 0) 2680 return (FALSE); 2681 pte = vtopte(addr); 2682 if (*pte) 2683 return (FALSE); 2684 return (TRUE); 2685} 2686 2687/* 2688 * Clear the given bit in each of the given page's ptes. 2689 */ 2690static __inline void 2691pmap_clear_ptes(vm_page_t m, int bit) 2692{ 2693 register pv_entry_t pv; 2694 pt_entry_t pbits, *pte; 2695 int s; 2696 2697 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) || 2698 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2699 return; 2700 2701 s = splvm(); 2702 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2703 /* 2704 * Loop over all current mappings setting/clearing as appropos If 2705 * setting RO do we need to clear the VAC? 2706 */ 2707 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2708 /* 2709 * don't write protect pager mappings 2710 */ 2711 if (bit == PG_RW) { 2712 if (!pmap_track_modified(pv->pv_va)) 2713 continue; 2714 } 2715 2716#if defined(PMAP_DIAGNOSTIC) 2717 if (!pv->pv_pmap) { 2718 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 2719 continue; 2720 } 2721#endif 2722 2723 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2724 pbits = *pte; 2725 if (pbits & bit) { 2726 if (bit == PG_RW) { 2727 if (pbits & PG_M) { 2728 vm_page_dirty(m); 2729 } 2730 pte_store(pte, pbits & ~(PG_M|PG_RW)); 2731 } else { 2732 pte_store(pte, pbits & ~bit); 2733 } 2734 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2735 } 2736 } 2737 if (bit == PG_RW) 2738 vm_page_flag_clear(m, PG_WRITEABLE); 2739 splx(s); 2740} 2741 2742/* 2743 * pmap_page_protect: 2744 * 2745 * Lower the permission for all mappings to a given page. 2746 */ 2747void 2748pmap_page_protect(vm_page_t m, vm_prot_t prot) 2749{ 2750 if ((prot & VM_PROT_WRITE) == 0) { 2751 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2752 pmap_clear_ptes(m, PG_RW); 2753 } else { 2754 pmap_remove_all(m); 2755 } 2756 } 2757} 2758 2759/* 2760 * pmap_ts_referenced: 2761 * 2762 * Return a count of reference bits for a page, clearing those bits. 2763 * It is not necessary for every reference bit to be cleared, but it 2764 * is necessary that 0 only be returned when there are truly no 2765 * reference bits set. 2766 * 2767 * XXX: The exact number of bits to check and clear is a matter that 2768 * should be tested and standardized at some point in the future for 2769 * optimal aging of shared pages. 2770 */ 2771int 2772pmap_ts_referenced(vm_page_t m) 2773{ 2774 register pv_entry_t pv, pvf, pvn; 2775 pt_entry_t *pte; 2776 pt_entry_t v; 2777 int s; 2778 int rtval = 0; 2779 2780 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2781 return (rtval); 2782 2783 s = splvm(); 2784 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2785 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2786 2787 pvf = pv; 2788 2789 do { 2790 pvn = TAILQ_NEXT(pv, pv_list); 2791 2792 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2793 2794 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2795 2796 if (!pmap_track_modified(pv->pv_va)) 2797 continue; 2798 2799 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2800 2801 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2802 pte_store(pte, v & ~PG_A); 2803 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2804 2805 rtval++; 2806 if (rtval > 4) { 2807 break; 2808 } 2809 } 2810 } while ((pv = pvn) != NULL && pv != pvf); 2811 } 2812 splx(s); 2813 2814 return (rtval); 2815} 2816 2817/* 2818 * Clear the modify bits on the specified physical page. 2819 */ 2820void 2821pmap_clear_modify(vm_page_t m) 2822{ 2823 pmap_clear_ptes(m, PG_M); 2824} 2825 2826/* 2827 * pmap_clear_reference: 2828 * 2829 * Clear the reference bit on the specified physical page. 2830 */ 2831void 2832pmap_clear_reference(vm_page_t m) 2833{ 2834 pmap_clear_ptes(m, PG_A); 2835} 2836 2837/* 2838 * Miscellaneous support routines follow 2839 */ 2840 2841static void 2842i386_protection_init() 2843{ 2844 register int *kp, prot; 2845 2846 kp = protection_codes; 2847 for (prot = 0; prot < 8; prot++) { 2848 switch (prot) { 2849 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE: 2850 /* 2851 * Read access is also 0. There isn't any execute bit, 2852 * so just make it readable. 2853 */ 2854 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE: 2855 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE: 2856 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE: 2857 *kp++ = 0; 2858 break; 2859 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE: 2860 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE: 2861 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE: 2862 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE: 2863 *kp++ = PG_RW; 2864 break; 2865 } 2866 } 2867} 2868 2869/* 2870 * Map a set of physical memory pages into the kernel virtual 2871 * address space. Return a pointer to where it is mapped. This 2872 * routine is intended to be used for mapping device memory, 2873 * NOT real memory. 2874 */ 2875void * 2876pmap_mapdev(pa, size) 2877 vm_paddr_t pa; 2878 vm_size_t size; 2879{ 2880 vm_offset_t va, tmpva, offset; 2881 2882 offset = pa & PAGE_MASK; 2883 size = roundup(offset + size, PAGE_SIZE); 2884 2885 va = kmem_alloc_nofault(kernel_map, size); 2886 if (!va) 2887 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2888 2889 pa = pa & PG_FRAME; 2890 for (tmpva = va; size > 0; ) { 2891 pmap_kenter(tmpva, pa); 2892 size -= PAGE_SIZE; 2893 tmpva += PAGE_SIZE; 2894 pa += PAGE_SIZE; 2895 } 2896 pmap_invalidate_range(kernel_pmap, va, tmpva); 2897 return ((void *)(va + offset)); 2898} 2899 2900void 2901pmap_unmapdev(va, size) 2902 vm_offset_t va; 2903 vm_size_t size; 2904{ 2905 vm_offset_t base, offset, tmpva; 2906 pt_entry_t *pte; 2907 2908 base = va & PG_FRAME; 2909 offset = va & PAGE_MASK; 2910 size = roundup(offset + size, PAGE_SIZE); 2911 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) { 2912 pte = vtopte(tmpva); 2913 pte_clear(pte); 2914 } 2915 pmap_invalidate_range(kernel_pmap, va, tmpva); 2916 kmem_free(kernel_map, base, size); 2917} 2918 2919/* 2920 * perform the pmap work for mincore 2921 */ 2922int 2923pmap_mincore(pmap, addr) 2924 pmap_t pmap; 2925 vm_offset_t addr; 2926{ 2927 pt_entry_t *ptep, pte; 2928 vm_page_t m; 2929 int val = 0; 2930 2931 ptep = pmap_pte_quick(pmap, addr); 2932 if (ptep == 0) { 2933 return 0; 2934 } 2935 2936 if ((pte = *ptep) != 0) { 2937 vm_paddr_t pa; 2938 2939 val = MINCORE_INCORE; 2940 if ((pte & PG_MANAGED) == 0) 2941 return val; 2942 2943 pa = pte & PG_FRAME; 2944 2945 m = PHYS_TO_VM_PAGE(pa); 2946 2947 /* 2948 * Modified by us 2949 */ 2950 if (pte & PG_M) 2951 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2952 else { 2953 /* 2954 * Modified by someone else 2955 */ 2956 vm_page_lock_queues(); 2957 if (m->dirty || pmap_is_modified(m)) 2958 val |= MINCORE_MODIFIED_OTHER; 2959 vm_page_unlock_queues(); 2960 } 2961 /* 2962 * Referenced by us 2963 */ 2964 if (pte & PG_A) 2965 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2966 else { 2967 /* 2968 * Referenced by someone else 2969 */ 2970 vm_page_lock_queues(); 2971 if ((m->flags & PG_REFERENCED) || 2972 pmap_ts_referenced(m)) { 2973 val |= MINCORE_REFERENCED_OTHER; 2974 vm_page_flag_set(m, PG_REFERENCED); 2975 } 2976 vm_page_unlock_queues(); 2977 } 2978 } 2979 return val; 2980} 2981 2982void 2983pmap_activate(struct thread *td) 2984{ 2985 struct proc *p = td->td_proc; 2986 pmap_t pmap, oldpmap; 2987 u_int32_t cr3; 2988 2989 critical_enter(); 2990 pmap = vmspace_pmap(td->td_proc->p_vmspace); 2991 oldpmap = PCPU_GET(curpmap); 2992#if defined(SMP) 2993 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 2994 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 2995#else 2996 oldpmap->pm_active &= ~1; 2997 pmap->pm_active |= 1; 2998#endif 2999#ifdef PAE 3000 cr3 = vtophys(pmap->pm_pdpt); 3001#else 3002 cr3 = vtophys(pmap->pm_pdir); 3003#endif 3004 /* XXXKSE this is wrong. 3005 * pmap_activate is for the current thread on the current cpu 3006 */ 3007 if (p->p_flag & P_SA) { 3008 /* Make sure all other cr3 entries are updated. */ 3009 /* what if they are running? XXXKSE (maybe abort them) */ 3010 FOREACH_THREAD_IN_PROC(p, td) { 3011 td->td_pcb->pcb_cr3 = cr3; 3012 } 3013 } else { 3014 td->td_pcb->pcb_cr3 = cr3; 3015 } 3016 load_cr3(cr3); 3017 PCPU_SET(curpmap, pmap); 3018 critical_exit(); 3019} 3020 3021vm_offset_t 3022pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3023{ 3024 3025 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3026 return addr; 3027 } 3028 3029 addr = (addr + PDRMASK) & ~PDRMASK; 3030 return addr; 3031} 3032 3033 3034#if defined(PMAP_DEBUG) 3035pmap_pid_dump(int pid) 3036{ 3037 pmap_t pmap; 3038 struct proc *p; 3039 int npte = 0; 3040 int index; 3041 3042 sx_slock(&allproc_lock); 3043 LIST_FOREACH(p, &allproc, p_list) { 3044 if (p->p_pid != pid) 3045 continue; 3046 3047 if (p->p_vmspace) { 3048 int i,j; 3049 index = 0; 3050 pmap = vmspace_pmap(p->p_vmspace); 3051 for (i = 0; i < NPDEPTD; i++) { 3052 pd_entry_t *pde; 3053 pt_entry_t *pte; 3054 vm_offset_t base = i << PDRSHIFT; 3055 3056 pde = &pmap->pm_pdir[i]; 3057 if (pde && pmap_pde_v(pde)) { 3058 for (j = 0; j < NPTEPG; j++) { 3059 vm_offset_t va = base + (j << PAGE_SHIFT); 3060 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3061 if (index) { 3062 index = 0; 3063 printf("\n"); 3064 } 3065 sx_sunlock(&allproc_lock); 3066 return npte; 3067 } 3068 pte = pmap_pte_quick(pmap, va); 3069 if (pte && pmap_pte_v(pte)) { 3070 pt_entry_t pa; 3071 vm_page_t m; 3072 pa = *pte; 3073 m = PHYS_TO_VM_PAGE(pa); 3074 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3075 va, pa, m->hold_count, m->wire_count, m->flags); 3076 npte++; 3077 index++; 3078 if (index >= 2) { 3079 index = 0; 3080 printf("\n"); 3081 } else { 3082 printf(" "); 3083 } 3084 } 3085 } 3086 } 3087 } 3088 } 3089 } 3090 sx_sunlock(&allproc_lock); 3091 return npte; 3092} 3093#endif 3094 3095#if defined(DEBUG) 3096 3097static void pads(pmap_t pm); 3098void pmap_pvdump(vm_offset_t pa); 3099 3100/* print address space of pmap*/ 3101static void 3102pads(pm) 3103 pmap_t pm; 3104{ 3105 int i, j; 3106 vm_paddr_t va; 3107 pt_entry_t *ptep; 3108 3109 if (pm == kernel_pmap) 3110 return; 3111 for (i = 0; i < NPDEPTD; i++) 3112 if (pm->pm_pdir[i]) 3113 for (j = 0; j < NPTEPG; j++) { 3114 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3115 if (pm == kernel_pmap && va < KERNBASE) 3116 continue; 3117 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3118 continue; 3119 ptep = pmap_pte_quick(pm, va); 3120 if (pmap_pte_v(ptep)) 3121 printf("%x:%x ", va, *ptep); 3122 }; 3123 3124} 3125 3126void 3127pmap_pvdump(pa) 3128 vm_paddr_t pa; 3129{ 3130 pv_entry_t pv; 3131 vm_page_t m; 3132 3133 printf("pa %x", pa); 3134 m = PHYS_TO_VM_PAGE(pa); 3135 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3136 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3137 pads(pv->pv_pmap); 3138 } 3139 printf(" "); 3140} 3141#endif 3142