pmap.c revision 130560
1/*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1994 John S. Dyson 5 * All rights reserved. 6 * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * the Systems Programming Group of the University of Utah Computer 11 * Science Department and William Jolitz of UUNET Technologies Inc. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by the University of 24 * California, Berkeley and its contributors. 25 * 4. Neither the name of the University nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 39 * SUCH DAMAGE. 40 * 41 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 42 */ 43/*- 44 * Copyright (c) 2003 Networks Associates Technology, Inc. 45 * All rights reserved. 46 * 47 * This software was developed for the FreeBSD Project by Jake Burkholder, 48 * Safeport Network Services, and Network Associates Laboratories, the 49 * Security Research Division of Network Associates, Inc. under 50 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA 51 * CHATS research program. 52 * 53 * Redistribution and use in source and binary forms, with or without 54 * modification, are permitted provided that the following conditions 55 * are met: 56 * 1. Redistributions of source code must retain the above copyright 57 * notice, this list of conditions and the following disclaimer. 58 * 2. Redistributions in binary form must reproduce the above copyright 59 * notice, this list of conditions and the following disclaimer in the 60 * documentation and/or other materials provided with the distribution. 61 * 62 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 63 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 64 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 65 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 66 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 67 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 68 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 69 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 70 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 71 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 72 * SUCH DAMAGE. 73 */ 74 75#include <sys/cdefs.h> 76__FBSDID("$FreeBSD: head/sys/i386/i386/pmap.c 130560 2004-06-16 05:42:44Z alc $"); 77 78/* 79 * Manages physical address maps. 80 * 81 * In addition to hardware address maps, this 82 * module is called upon to provide software-use-only 83 * maps which may or may not be stored in the same 84 * form as hardware maps. These pseudo-maps are 85 * used to store intermediate results from copy 86 * operations to and from address spaces. 87 * 88 * Since the information managed by this module is 89 * also stored by the logical address mapping module, 90 * this module may throw away valid virtual-to-physical 91 * mappings at almost any time. However, invalidations 92 * of virtual-to-physical mappings must be done as 93 * requested. 94 * 95 * In order to cope with hardware architectures which 96 * make virtual-to-physical map invalidates expensive, 97 * this module may delay invalidate or reduced protection 98 * operations until such time as they are actually 99 * necessary. This module is given full information as 100 * to which processors are currently using which maps, 101 * and to when physical maps must be made correct. 102 */ 103 104#include "opt_cpu.h" 105#include "opt_pmap.h" 106#include "opt_msgbuf.h" 107#include "opt_kstack_pages.h" 108 109#include <sys/param.h> 110#include <sys/systm.h> 111#include <sys/kernel.h> 112#include <sys/lock.h> 113#include <sys/mman.h> 114#include <sys/msgbuf.h> 115#include <sys/mutex.h> 116#include <sys/proc.h> 117#include <sys/sx.h> 118#include <sys/user.h> 119#include <sys/vmmeter.h> 120#include <sys/sched.h> 121#include <sys/sysctl.h> 122#ifdef SMP 123#include <sys/smp.h> 124#endif 125 126#include <vm/vm.h> 127#include <vm/vm_param.h> 128#include <vm/vm_kern.h> 129#include <vm/vm_page.h> 130#include <vm/vm_map.h> 131#include <vm/vm_object.h> 132#include <vm/vm_extern.h> 133#include <vm/vm_pageout.h> 134#include <vm/vm_pager.h> 135#include <vm/uma.h> 136 137#include <machine/cpu.h> 138#include <machine/cputypes.h> 139#include <machine/md_var.h> 140#include <machine/specialreg.h> 141#ifdef SMP 142#include <machine/smp.h> 143#endif 144 145#if !defined(CPU_ENABLE_SSE) && defined(I686_CPU) 146#define CPU_ENABLE_SSE 147#endif 148#if defined(CPU_DISABLE_SSE) 149#undef CPU_ENABLE_SSE 150#endif 151 152#ifndef PMAP_SHPGPERPROC 153#define PMAP_SHPGPERPROC 200 154#endif 155 156#if defined(DIAGNOSTIC) 157#define PMAP_DIAGNOSTIC 158#endif 159 160#define MINPV 2048 161 162#if !defined(PMAP_DIAGNOSTIC) 163#define PMAP_INLINE __inline 164#else 165#define PMAP_INLINE 166#endif 167 168/* 169 * Get PDEs and PTEs for user/kernel address space 170 */ 171#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 172#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 173 174#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 175#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 176#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 177#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 178#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 179 180#define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \ 181 atomic_clear_int((u_int *)(pte), PG_W)) 182#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 183 184struct pmap kernel_pmap_store; 185LIST_HEAD(pmaplist, pmap); 186static struct pmaplist allpmaps; 187static struct mtx allpmaps_lock; 188#ifdef SMP 189static struct mtx lazypmap_lock; 190#endif 191 192vm_paddr_t avail_end; /* PA of last available physical page */ 193vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 194vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 195static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 196int pgeflag = 0; /* PG_G or-in */ 197int pseflag = 0; /* PG_PS or-in */ 198 199static int nkpt; 200vm_offset_t kernel_vm_end; 201extern u_int32_t KERNend; 202 203#ifdef PAE 204static uma_zone_t pdptzone; 205#endif 206 207/* 208 * Data for the pv entry allocation mechanism 209 */ 210static uma_zone_t pvzone; 211static struct vm_object pvzone_obj; 212static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 213int pmap_pagedaemon_waken; 214 215/* 216 * All those kernel PT submaps that BSD is so fond of 217 */ 218pt_entry_t *CMAP1 = 0; 219static pt_entry_t *CMAP2, *CMAP3, *ptmmap; 220caddr_t CADDR1 = 0, ptvmmap = 0; 221static caddr_t CADDR2, CADDR3; 222static struct mtx CMAPCADDR12_lock; 223static pt_entry_t *msgbufmap; 224struct msgbuf *msgbufp = 0; 225 226/* 227 * Crashdump maps. 228 */ 229static pt_entry_t *pt_crashdumpmap; 230static caddr_t crashdumpmap; 231 232#ifdef SMP 233extern pt_entry_t *SMPpt; 234#endif 235static pt_entry_t *PMAP1 = 0, *PMAP2; 236static pt_entry_t *PADDR1 = 0, *PADDR2; 237#ifdef SMP 238static int PMAP1cpu; 239static int PMAP1changedcpu; 240SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD, 241 &PMAP1changedcpu, 0, 242 "Number of times pmap_pte_quick changed CPU with same PMAP1"); 243#endif 244static int PMAP1changed; 245SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD, 246 &PMAP1changed, 0, 247 "Number of times pmap_pte_quick changed PMAP1"); 248static int PMAP1unchanged; 249SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD, 250 &PMAP1unchanged, 0, 251 "Number of times pmap_pte_quick didn't change PMAP1"); 252 253static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 254static pv_entry_t get_pv_entry(void); 255static void pmap_clear_ptes(vm_page_t m, int bit); 256 257static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 258static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 259static int pmap_remove_entry(struct pmap *pmap, vm_page_t m, 260 vm_offset_t va); 261static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, 262 vm_page_t mpte, vm_page_t m); 263 264static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va); 265 266static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex); 267static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 268static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t); 269static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 270#ifdef PAE 271static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 272#endif 273 274CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 275CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 276 277/* 278 * Move the kernel virtual free pointer to the next 279 * 4MB. This is used to help improve performance 280 * by using a large (4MB) page for much of the kernel 281 * (.text, .data, .bss) 282 */ 283static vm_offset_t 284pmap_kmem_choose(vm_offset_t addr) 285{ 286 vm_offset_t newaddr = addr; 287 288#ifndef DISABLE_PSE 289 if (cpu_feature & CPUID_PSE) 290 newaddr = (addr + PDRMASK) & ~PDRMASK; 291#endif 292 return newaddr; 293} 294 295/* 296 * Bootstrap the system enough to run with virtual memory. 297 * 298 * On the i386 this is called after mapping has already been enabled 299 * and just syncs the pmap module with what has already been done. 300 * [We can't call it easily with mapping off since the kernel is not 301 * mapped with PA == VA, hence we would have to relocate every address 302 * from the linked base (virtual) address "KERNBASE" to the actual 303 * (physical) address starting relative to 0] 304 */ 305void 306pmap_bootstrap(firstaddr, loadaddr) 307 vm_paddr_t firstaddr; 308 vm_paddr_t loadaddr; 309{ 310 vm_offset_t va; 311 pt_entry_t *pte; 312 int i; 313 314 /* 315 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 316 * large. It should instead be correctly calculated in locore.s and 317 * not based on 'first' (which is a physical address, not a virtual 318 * address, for the start of unused physical memory). The kernel 319 * page tables are NOT double mapped and thus should not be included 320 * in this calculation. 321 */ 322 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 323 virtual_avail = pmap_kmem_choose(virtual_avail); 324 325 virtual_end = VM_MAX_KERNEL_ADDRESS; 326 327 /* 328 * Initialize the kernel pmap (which is statically allocated). 329 */ 330 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 331#ifdef PAE 332 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 333#endif 334 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 335 TAILQ_INIT(&kernel_pmap->pm_pvlist); 336 LIST_INIT(&allpmaps); 337#ifdef SMP 338 mtx_init(&lazypmap_lock, "lazypmap", NULL, MTX_SPIN); 339#endif 340 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 341 mtx_lock_spin(&allpmaps_lock); 342 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 343 mtx_unlock_spin(&allpmaps_lock); 344 nkpt = NKPT; 345 346 /* 347 * Reserve some special page table entries/VA space for temporary 348 * mapping of pages. 349 */ 350#define SYSMAP(c, p, v, n) \ 351 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 352 353 va = virtual_avail; 354 pte = vtopte(va); 355 356 /* 357 * CMAP1/CMAP2 are used for zeroing and copying pages. 358 * CMAP3 is used for the idle process page zeroing. 359 */ 360 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 361 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 362 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 363 *CMAP3 = 0; 364 365 mtx_init(&CMAPCADDR12_lock, "CMAPCADDR12", NULL, MTX_DEF); 366 367 /* 368 * Crashdump maps. 369 */ 370 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 371 372 /* 373 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 374 * XXX ptmmap is not used. 375 */ 376 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 377 378 /* 379 * msgbufp is used to map the system message buffer. 380 * XXX msgbufmap is not used. 381 */ 382 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 383 atop(round_page(MSGBUF_SIZE))) 384 385 /* 386 * ptemap is used for pmap_pte_quick 387 */ 388 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 389 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1); 390 391 virtual_avail = va; 392 393 *CMAP1 = *CMAP2 = 0; 394 for (i = 0; i < NKPT; i++) 395 PTD[i] = 0; 396 397 /* Turn on PG_G on kernel page(s) */ 398 pmap_set_pg(); 399} 400 401/* 402 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 403 */ 404void 405pmap_set_pg(void) 406{ 407 pd_entry_t pdir; 408 pt_entry_t *pte; 409 vm_offset_t va, endva; 410 int i; 411 412 if (pgeflag == 0) 413 return; 414 415 i = KERNLOAD/NBPDR; 416 endva = KERNBASE + KERNend; 417 418 if (pseflag) { 419 va = KERNBASE + KERNLOAD; 420 while (va < endva) { 421 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 422 pdir |= pgeflag; 423 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 424 invltlb(); /* Play it safe, invltlb() every time */ 425 i++; 426 va += NBPDR; 427 } 428 } else { 429 va = (vm_offset_t)btext; 430 while (va < endva) { 431 pte = vtopte(va); 432 if (*pte) 433 *pte |= pgeflag; 434 invltlb(); /* Play it safe, invltlb() every time */ 435 va += PAGE_SIZE; 436 } 437 } 438} 439 440#ifdef PAE 441static void * 442pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 443{ 444 *flags = UMA_SLAB_PRIV; 445 return (contigmalloc(PAGE_SIZE, NULL, 0, 0x0ULL, 0xffffffffULL, 1, 0)); 446} 447#endif 448 449/* 450 * Initialize the pmap module. 451 * Called by vm_init, to initialize any structures that the pmap 452 * system needs to map virtual memory. 453 * pmap_init has been enhanced to support in a fairly consistant 454 * way, discontiguous physical memory. 455 */ 456void 457pmap_init(void) 458{ 459 int i; 460 461 /* 462 * Allocate memory for random pmap data structures. Includes the 463 * pv_head_table. 464 */ 465 466 for(i = 0; i < vm_page_array_size; i++) { 467 vm_page_t m; 468 469 m = &vm_page_array[i]; 470 TAILQ_INIT(&m->md.pv_list); 471 m->md.pv_list_count = 0; 472 } 473 474 /* 475 * init the pv free list 476 */ 477 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL, 478 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 479 uma_prealloc(pvzone, MINPV); 480 481#ifdef PAE 482 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 483 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 484 UMA_ZONE_VM | UMA_ZONE_NOFREE); 485 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 486#endif 487 488 /* 489 * Now it is safe to enable pv_table recording. 490 */ 491 pmap_initialized = TRUE; 492} 493 494/* 495 * Initialize the address space (zone) for the pv_entries. Set a 496 * high water mark so that the system can recover from excessive 497 * numbers of pv entries. 498 */ 499void 500pmap_init2() 501{ 502 int shpgperproc = PMAP_SHPGPERPROC; 503 504 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 505 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 506 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 507 pv_entry_high_water = 9 * (pv_entry_max / 10); 508 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 509} 510 511 512/*************************************************** 513 * Low level helper routines..... 514 ***************************************************/ 515 516#if defined(PMAP_DIAGNOSTIC) 517 518/* 519 * This code checks for non-writeable/modified pages. 520 * This should be an invalid condition. 521 */ 522static int 523pmap_nw_modified(pt_entry_t ptea) 524{ 525 int pte; 526 527 pte = (int) ptea; 528 529 if ((pte & (PG_M|PG_RW)) == PG_M) 530 return 1; 531 else 532 return 0; 533} 534#endif 535 536 537/* 538 * this routine defines the region(s) of memory that should 539 * not be tested for the modified bit. 540 */ 541static PMAP_INLINE int 542pmap_track_modified(vm_offset_t va) 543{ 544 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 545 return 1; 546 else 547 return 0; 548} 549 550#ifdef I386_CPU 551/* 552 * i386 only has "invalidate everything" and no SMP to worry about. 553 */ 554PMAP_INLINE void 555pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 556{ 557 558 if (pmap == kernel_pmap || pmap->pm_active) 559 invltlb(); 560} 561 562PMAP_INLINE void 563pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 564{ 565 566 if (pmap == kernel_pmap || pmap->pm_active) 567 invltlb(); 568} 569 570PMAP_INLINE void 571pmap_invalidate_all(pmap_t pmap) 572{ 573 574 if (pmap == kernel_pmap || pmap->pm_active) 575 invltlb(); 576} 577#else /* !I386_CPU */ 578#ifdef SMP 579/* 580 * For SMP, these functions have to use the IPI mechanism for coherence. 581 */ 582void 583pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 584{ 585 u_int cpumask; 586 u_int other_cpus; 587 588 if (smp_started) { 589 if (!(read_eflags() & PSL_I)) 590 panic("%s: interrupts disabled", __func__); 591 mtx_lock_spin(&smp_tlb_mtx); 592 } else 593 critical_enter(); 594 /* 595 * We need to disable interrupt preemption but MUST NOT have 596 * interrupts disabled here. 597 * XXX we may need to hold schedlock to get a coherent pm_active 598 * XXX critical sections disable interrupts again 599 */ 600 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 601 invlpg(va); 602 smp_invlpg(va); 603 } else { 604 cpumask = PCPU_GET(cpumask); 605 other_cpus = PCPU_GET(other_cpus); 606 if (pmap->pm_active & cpumask) 607 invlpg(va); 608 if (pmap->pm_active & other_cpus) 609 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 610 } 611 if (smp_started) 612 mtx_unlock_spin(&smp_tlb_mtx); 613 else 614 critical_exit(); 615} 616 617void 618pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 619{ 620 u_int cpumask; 621 u_int other_cpus; 622 vm_offset_t addr; 623 624 if (smp_started) { 625 if (!(read_eflags() & PSL_I)) 626 panic("%s: interrupts disabled", __func__); 627 mtx_lock_spin(&smp_tlb_mtx); 628 } else 629 critical_enter(); 630 /* 631 * We need to disable interrupt preemption but MUST NOT have 632 * interrupts disabled here. 633 * XXX we may need to hold schedlock to get a coherent pm_active 634 * XXX critical sections disable interrupts again 635 */ 636 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 637 for (addr = sva; addr < eva; addr += PAGE_SIZE) 638 invlpg(addr); 639 smp_invlpg_range(sva, eva); 640 } else { 641 cpumask = PCPU_GET(cpumask); 642 other_cpus = PCPU_GET(other_cpus); 643 if (pmap->pm_active & cpumask) 644 for (addr = sva; addr < eva; addr += PAGE_SIZE) 645 invlpg(addr); 646 if (pmap->pm_active & other_cpus) 647 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 648 sva, eva); 649 } 650 if (smp_started) 651 mtx_unlock_spin(&smp_tlb_mtx); 652 else 653 critical_exit(); 654} 655 656void 657pmap_invalidate_all(pmap_t pmap) 658{ 659 u_int cpumask; 660 u_int other_cpus; 661 662 if (smp_started) { 663 if (!(read_eflags() & PSL_I)) 664 panic("%s: interrupts disabled", __func__); 665 mtx_lock_spin(&smp_tlb_mtx); 666 } else 667 critical_enter(); 668 /* 669 * We need to disable interrupt preemption but MUST NOT have 670 * interrupts disabled here. 671 * XXX we may need to hold schedlock to get a coherent pm_active 672 * XXX critical sections disable interrupts again 673 */ 674 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 675 invltlb(); 676 smp_invltlb(); 677 } else { 678 cpumask = PCPU_GET(cpumask); 679 other_cpus = PCPU_GET(other_cpus); 680 if (pmap->pm_active & cpumask) 681 invltlb(); 682 if (pmap->pm_active & other_cpus) 683 smp_masked_invltlb(pmap->pm_active & other_cpus); 684 } 685 if (smp_started) 686 mtx_unlock_spin(&smp_tlb_mtx); 687 else 688 critical_exit(); 689} 690#else /* !SMP */ 691/* 692 * Normal, non-SMP, 486+ invalidation functions. 693 * We inline these within pmap.c for speed. 694 */ 695PMAP_INLINE void 696pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 697{ 698 699 if (pmap == kernel_pmap || pmap->pm_active) 700 invlpg(va); 701} 702 703PMAP_INLINE void 704pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 705{ 706 vm_offset_t addr; 707 708 if (pmap == kernel_pmap || pmap->pm_active) 709 for (addr = sva; addr < eva; addr += PAGE_SIZE) 710 invlpg(addr); 711} 712 713PMAP_INLINE void 714pmap_invalidate_all(pmap_t pmap) 715{ 716 717 if (pmap == kernel_pmap || pmap->pm_active) 718 invltlb(); 719} 720#endif /* !SMP */ 721#endif /* !I386_CPU */ 722 723/* 724 * Are we current address space or kernel? N.B. We return FALSE when 725 * a pmap's page table is in use because a kernel thread is borrowing 726 * it. The borrowed page table can change spontaneously, making any 727 * dependence on its continued use subject to a race condition. 728 */ 729static __inline int 730pmap_is_current(pmap_t pmap) 731{ 732 733 return (pmap == kernel_pmap || 734 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 735 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 736} 737 738/* 739 * If the given pmap is not the current pmap, Giant must be held. 740 */ 741pt_entry_t * 742pmap_pte(pmap_t pmap, vm_offset_t va) 743{ 744 pd_entry_t newpf; 745 pd_entry_t *pde; 746 747 pde = pmap_pde(pmap, va); 748 if (*pde & PG_PS) 749 return (pde); 750 if (*pde != 0) { 751 /* are we current address space or kernel? */ 752 if (pmap_is_current(pmap)) 753 return (vtopte(va)); 754 GIANT_REQUIRED; 755 newpf = *pde & PG_FRAME; 756 if ((*PMAP2 & PG_FRAME) != newpf) { 757 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; 758 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 759 } 760 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); 761 } 762 return (0); 763} 764 765static __inline void 766invlcaddr(void *caddr) 767{ 768#ifdef I386_CPU 769 invltlb(); 770#else 771 invlpg((u_int)caddr); 772#endif 773} 774 775/* 776 * Super fast pmap_pte routine best used when scanning 777 * the pv lists. This eliminates many coarse-grained 778 * invltlb calls. Note that many of the pv list 779 * scans are across different pmaps. It is very wasteful 780 * to do an entire invltlb for checking a single mapping. 781 * 782 * If the given pmap is not the current pmap, vm_page_queue_mtx 783 * must be held and curthread pinned to a CPU. 784 */ 785static pt_entry_t * 786pmap_pte_quick(pmap_t pmap, vm_offset_t va) 787{ 788 pd_entry_t newpf; 789 pd_entry_t *pde; 790 791 pde = pmap_pde(pmap, va); 792 if (*pde & PG_PS) 793 return (pde); 794 if (*pde != 0) { 795 /* are we current address space or kernel? */ 796 if (pmap_is_current(pmap)) 797 return (vtopte(va)); 798 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 799 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 800 newpf = *pde & PG_FRAME; 801 if ((*PMAP1 & PG_FRAME) != newpf) { 802 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; 803#ifdef SMP 804 PMAP1cpu = PCPU_GET(cpuid); 805#endif 806 invlcaddr(PADDR1); 807 PMAP1changed++; 808 } else 809#ifdef SMP 810 if (PMAP1cpu != PCPU_GET(cpuid)) { 811 PMAP1cpu = PCPU_GET(cpuid); 812 invlcaddr(PADDR1); 813 PMAP1changedcpu++; 814 } else 815#endif 816 PMAP1unchanged++; 817 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); 818 } 819 return (0); 820} 821 822/* 823 * Routine: pmap_extract 824 * Function: 825 * Extract the physical page address associated 826 * with the given map/virtual_address pair. 827 */ 828vm_paddr_t 829pmap_extract(pmap, va) 830 register pmap_t pmap; 831 vm_offset_t va; 832{ 833 vm_paddr_t rtval; 834 pt_entry_t *pte; 835 pd_entry_t pde; 836 837 if (pmap == 0) 838 return 0; 839 pde = pmap->pm_pdir[va >> PDRSHIFT]; 840 if (pde != 0) { 841 if ((pde & PG_PS) != 0) { 842 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 843 return rtval; 844 } 845 pte = pmap_pte(pmap, va); 846 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK)); 847 return rtval; 848 } 849 return 0; 850 851} 852 853/* 854 * Routine: pmap_extract_and_hold 855 * Function: 856 * Atomically extract and hold the physical page 857 * with the given pmap and virtual address pair 858 * if that mapping permits the given protection. 859 */ 860vm_page_t 861pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 862{ 863 vm_paddr_t pa; 864 vm_page_t m; 865 866 m = NULL; 867 mtx_lock(&Giant); 868 if ((pa = pmap_extract(pmap, va)) != 0) { 869 m = PHYS_TO_VM_PAGE(pa); 870 vm_page_lock_queues(); 871 vm_page_hold(m); 872 vm_page_unlock_queues(); 873 } 874 mtx_unlock(&Giant); 875 return (m); 876} 877 878/*************************************************** 879 * Low level mapping routines..... 880 ***************************************************/ 881 882/* 883 * Add a wired page to the kva. 884 * Note: not SMP coherent. 885 */ 886PMAP_INLINE void 887pmap_kenter(vm_offset_t va, vm_paddr_t pa) 888{ 889 pt_entry_t *pte; 890 891 pte = vtopte(va); 892 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 893} 894 895/* 896 * Remove a page from the kernel pagetables. 897 * Note: not SMP coherent. 898 */ 899PMAP_INLINE void 900pmap_kremove(vm_offset_t va) 901{ 902 pt_entry_t *pte; 903 904 pte = vtopte(va); 905 pte_clear(pte); 906} 907 908/* 909 * Used to map a range of physical addresses into kernel 910 * virtual address space. 911 * 912 * The value passed in '*virt' is a suggested virtual address for 913 * the mapping. Architectures which can support a direct-mapped 914 * physical to virtual region can return the appropriate address 915 * within that region, leaving '*virt' unchanged. Other 916 * architectures should map the pages starting at '*virt' and 917 * update '*virt' with the first usable address after the mapped 918 * region. 919 */ 920vm_offset_t 921pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 922{ 923 vm_offset_t va, sva; 924 925 va = sva = *virt; 926 while (start < end) { 927 pmap_kenter(va, start); 928 va += PAGE_SIZE; 929 start += PAGE_SIZE; 930 } 931 pmap_invalidate_range(kernel_pmap, sva, va); 932 *virt = va; 933 return (sva); 934} 935 936 937/* 938 * Add a list of wired pages to the kva 939 * this routine is only used for temporary 940 * kernel mappings that do not need to have 941 * page modification or references recorded. 942 * Note that old mappings are simply written 943 * over. The page *must* be wired. 944 * Note: SMP coherent. Uses a ranged shootdown IPI. 945 */ 946void 947pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 948{ 949 vm_offset_t va; 950 951 va = sva; 952 while (count-- > 0) { 953 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 954 va += PAGE_SIZE; 955 m++; 956 } 957 pmap_invalidate_range(kernel_pmap, sva, va); 958} 959 960/* 961 * This routine tears out page mappings from the 962 * kernel -- it is meant only for temporary mappings. 963 * Note: SMP coherent. Uses a ranged shootdown IPI. 964 */ 965void 966pmap_qremove(vm_offset_t sva, int count) 967{ 968 vm_offset_t va; 969 970 va = sva; 971 while (count-- > 0) { 972 pmap_kremove(va); 973 va += PAGE_SIZE; 974 } 975 pmap_invalidate_range(kernel_pmap, sva, va); 976} 977 978/*************************************************** 979 * Page table page management routines..... 980 ***************************************************/ 981 982/* 983 * This routine unholds page table pages, and if the hold count 984 * drops to zero, then it decrements the wire count. 985 */ 986static int 987_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 988{ 989 990 while (vm_page_sleep_if_busy(m, FALSE, "pmuwpt")) 991 vm_page_lock_queues(); 992 993 if (m->hold_count == 0) { 994 vm_offset_t pteva; 995 /* 996 * unmap the page table page 997 */ 998 pmap->pm_pdir[m->pindex] = 0; 999 --pmap->pm_stats.resident_count; 1000 /* 1001 * We never unwire a kernel page table page, making a 1002 * check for the kernel_pmap unnecessary. 1003 */ 1004 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)) { 1005 /* 1006 * Do an invltlb to make the invalidated mapping 1007 * take effect immediately. 1008 */ 1009 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1010 pmap_invalidate_page(pmap, pteva); 1011 } 1012 1013 /* 1014 * If the page is finally unwired, simply free it. 1015 */ 1016 --m->wire_count; 1017 if (m->wire_count == 0) { 1018 vm_page_busy(m); 1019 vm_page_free_zero(m); 1020 atomic_subtract_int(&cnt.v_wire_count, 1); 1021 } 1022 return 1; 1023 } 1024 return 0; 1025} 1026 1027static PMAP_INLINE int 1028pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1029{ 1030 vm_page_unhold(m); 1031 if (m->hold_count == 0) 1032 return _pmap_unwire_pte_hold(pmap, m); 1033 else 1034 return 0; 1035} 1036 1037/* 1038 * After removing a page table entry, this routine is used to 1039 * conditionally free the page, and manage the hold/wire counts. 1040 */ 1041static int 1042pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 1043{ 1044 1045 if (va >= VM_MAXUSER_ADDRESS) 1046 return 0; 1047 1048 return pmap_unwire_pte_hold(pmap, mpte); 1049} 1050 1051void 1052pmap_pinit0(pmap) 1053 struct pmap *pmap; 1054{ 1055 1056 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1057#ifdef PAE 1058 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1059#endif 1060 pmap->pm_active = 0; 1061 PCPU_SET(curpmap, pmap); 1062 TAILQ_INIT(&pmap->pm_pvlist); 1063 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1064 mtx_lock_spin(&allpmaps_lock); 1065 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1066 mtx_unlock_spin(&allpmaps_lock); 1067} 1068 1069/* 1070 * Initialize a preallocated and zeroed pmap structure, 1071 * such as one in a vmspace structure. 1072 */ 1073void 1074pmap_pinit(pmap) 1075 register struct pmap *pmap; 1076{ 1077 vm_page_t m, ptdpg[NPGPTD]; 1078 vm_paddr_t pa; 1079 static int color; 1080 int i; 1081 1082 /* 1083 * No need to allocate page table space yet but we do need a valid 1084 * page directory table. 1085 */ 1086 if (pmap->pm_pdir == NULL) { 1087 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1088 NBPTD); 1089#ifdef PAE 1090 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1091 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1092 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1093 ("pmap_pinit: pdpt misaligned")); 1094 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1095 ("pmap_pinit: pdpt above 4g")); 1096#endif 1097 } 1098 1099 /* 1100 * allocate the page directory page(s) 1101 */ 1102 for (i = 0; i < NPGPTD;) { 1103 m = vm_page_alloc(NULL, color++, 1104 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1105 VM_ALLOC_ZERO); 1106 if (m == NULL) 1107 VM_WAIT; 1108 else { 1109 ptdpg[i++] = m; 1110 } 1111 } 1112 1113 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1114 1115 for (i = 0; i < NPGPTD; i++) { 1116 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1117 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1118 } 1119 1120 mtx_lock_spin(&allpmaps_lock); 1121 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1122 mtx_unlock_spin(&allpmaps_lock); 1123 /* Wire in kernel global address entries. */ 1124 /* XXX copies current process, does not fill in MPPTDI */ 1125 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1126#ifdef SMP 1127 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1128#endif 1129 1130 /* install self-referential address mapping entry(s) */ 1131 for (i = 0; i < NPGPTD; i++) { 1132 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1133 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1134#ifdef PAE 1135 pmap->pm_pdpt[i] = pa | PG_V; 1136#endif 1137 } 1138 1139 pmap->pm_active = 0; 1140 TAILQ_INIT(&pmap->pm_pvlist); 1141 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1142} 1143 1144/* 1145 * this routine is called if the page table page is not 1146 * mapped correctly. 1147 */ 1148static vm_page_t 1149_pmap_allocpte(pmap, ptepindex) 1150 pmap_t pmap; 1151 unsigned ptepindex; 1152{ 1153 vm_paddr_t ptepa; 1154 vm_page_t m; 1155 1156 /* 1157 * Allocate a page table page. 1158 */ 1159 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1160 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1161 VM_WAIT; 1162 /* 1163 * Indicate the need to retry. While waiting, the page table 1164 * page may have been allocated. 1165 */ 1166 return (NULL); 1167 } 1168 if ((m->flags & PG_ZERO) == 0) 1169 pmap_zero_page(m); 1170 1171 KASSERT(m->queue == PQ_NONE, 1172 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1173 1174 /* 1175 * Increment the hold count for the page table page 1176 * (denoting a new mapping.) 1177 */ 1178 m->hold_count++; 1179 1180 /* 1181 * Map the pagetable page into the process address space, if 1182 * it isn't already there. 1183 */ 1184 1185 pmap->pm_stats.resident_count++; 1186 1187 ptepa = VM_PAGE_TO_PHYS(m); 1188 pmap->pm_pdir[ptepindex] = 1189 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1190 1191 vm_page_lock_queues(); 1192 vm_page_wakeup(m); 1193 vm_page_unlock_queues(); 1194 1195 return m; 1196} 1197 1198static vm_page_t 1199pmap_allocpte(pmap_t pmap, vm_offset_t va) 1200{ 1201 unsigned ptepindex; 1202 pd_entry_t ptepa; 1203 vm_page_t m; 1204 1205 /* 1206 * Calculate pagetable page index 1207 */ 1208 ptepindex = va >> PDRSHIFT; 1209retry: 1210 /* 1211 * Get the page directory entry 1212 */ 1213 ptepa = pmap->pm_pdir[ptepindex]; 1214 1215 /* 1216 * This supports switching from a 4MB page to a 1217 * normal 4K page. 1218 */ 1219 if (ptepa & PG_PS) { 1220 pmap->pm_pdir[ptepindex] = 0; 1221 ptepa = 0; 1222 pmap_invalidate_all(kernel_pmap); 1223 } 1224 1225 /* 1226 * If the page table page is mapped, we just increment the 1227 * hold count, and activate it. 1228 */ 1229 if (ptepa) { 1230 m = PHYS_TO_VM_PAGE(ptepa); 1231 m->hold_count++; 1232 } else { 1233 /* 1234 * Here if the pte page isn't mapped, or if it has 1235 * been deallocated. 1236 */ 1237 m = _pmap_allocpte(pmap, ptepindex); 1238 if (m == NULL) 1239 goto retry; 1240 } 1241 return (m); 1242} 1243 1244 1245/*************************************************** 1246* Pmap allocation/deallocation routines. 1247 ***************************************************/ 1248 1249#ifdef SMP 1250/* 1251 * Deal with a SMP shootdown of other users of the pmap that we are 1252 * trying to dispose of. This can be a bit hairy. 1253 */ 1254static u_int *lazymask; 1255static u_int lazyptd; 1256static volatile u_int lazywait; 1257 1258void pmap_lazyfix_action(void); 1259 1260void 1261pmap_lazyfix_action(void) 1262{ 1263 u_int mymask = PCPU_GET(cpumask); 1264 1265 if (rcr3() == lazyptd) 1266 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1267 atomic_clear_int(lazymask, mymask); 1268 atomic_store_rel_int(&lazywait, 1); 1269} 1270 1271static void 1272pmap_lazyfix_self(u_int mymask) 1273{ 1274 1275 if (rcr3() == lazyptd) 1276 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1277 atomic_clear_int(lazymask, mymask); 1278} 1279 1280 1281static void 1282pmap_lazyfix(pmap_t pmap) 1283{ 1284 u_int mymask = PCPU_GET(cpumask); 1285 u_int mask; 1286 register u_int spins; 1287 1288 while ((mask = pmap->pm_active) != 0) { 1289 spins = 50000000; 1290 mask = mask & -mask; /* Find least significant set bit */ 1291 mtx_lock_spin(&lazypmap_lock); 1292#ifdef PAE 1293 lazyptd = vtophys(pmap->pm_pdpt); 1294#else 1295 lazyptd = vtophys(pmap->pm_pdir); 1296#endif 1297 if (mask == mymask) { 1298 lazymask = &pmap->pm_active; 1299 pmap_lazyfix_self(mymask); 1300 } else { 1301 atomic_store_rel_int((u_int *)&lazymask, 1302 (u_int)&pmap->pm_active); 1303 atomic_store_rel_int(&lazywait, 0); 1304 ipi_selected(mask, IPI_LAZYPMAP); 1305 while (lazywait == 0) { 1306 ia32_pause(); 1307 if (--spins == 0) 1308 break; 1309 } 1310 } 1311 mtx_unlock_spin(&lazypmap_lock); 1312 if (spins == 0) 1313 printf("pmap_lazyfix: spun for 50000000\n"); 1314 } 1315} 1316 1317#else /* SMP */ 1318 1319/* 1320 * Cleaning up on uniprocessor is easy. For various reasons, we're 1321 * unlikely to have to even execute this code, including the fact 1322 * that the cleanup is deferred until the parent does a wait(2), which 1323 * means that another userland process has run. 1324 */ 1325static void 1326pmap_lazyfix(pmap_t pmap) 1327{ 1328 u_int cr3; 1329 1330 cr3 = vtophys(pmap->pm_pdir); 1331 if (cr3 == rcr3()) { 1332 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1333 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1334 } 1335} 1336#endif /* SMP */ 1337 1338/* 1339 * Release any resources held by the given physical map. 1340 * Called when a pmap initialized by pmap_pinit is being released. 1341 * Should only be called if the map contains no valid mappings. 1342 */ 1343void 1344pmap_release(pmap_t pmap) 1345{ 1346 vm_page_t m, ptdpg[NPGPTD]; 1347 int i; 1348 1349 KASSERT(pmap->pm_stats.resident_count == 0, 1350 ("pmap_release: pmap resident count %ld != 0", 1351 pmap->pm_stats.resident_count)); 1352 1353 pmap_lazyfix(pmap); 1354 mtx_lock_spin(&allpmaps_lock); 1355 LIST_REMOVE(pmap, pm_list); 1356 mtx_unlock_spin(&allpmaps_lock); 1357 1358 for (i = 0; i < NPGPTD; i++) 1359 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1360 1361 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1362 sizeof(*pmap->pm_pdir)); 1363#ifdef SMP 1364 pmap->pm_pdir[MPPTDI] = 0; 1365#endif 1366 1367 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1368 1369 vm_page_lock_queues(); 1370 for (i = 0; i < NPGPTD; i++) { 1371 m = ptdpg[i]; 1372#ifdef PAE 1373 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1374 ("pmap_release: got wrong ptd page")); 1375#endif 1376 m->wire_count--; 1377 atomic_subtract_int(&cnt.v_wire_count, 1); 1378 vm_page_free_zero(m); 1379 } 1380 vm_page_unlock_queues(); 1381} 1382 1383static int 1384kvm_size(SYSCTL_HANDLER_ARGS) 1385{ 1386 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1387 1388 return sysctl_handle_long(oidp, &ksize, 0, req); 1389} 1390SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1391 0, 0, kvm_size, "IU", "Size of KVM"); 1392 1393static int 1394kvm_free(SYSCTL_HANDLER_ARGS) 1395{ 1396 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1397 1398 return sysctl_handle_long(oidp, &kfree, 0, req); 1399} 1400SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1401 0, 0, kvm_free, "IU", "Amount of KVM free"); 1402 1403/* 1404 * grow the number of kernel page table entries, if needed 1405 */ 1406void 1407pmap_growkernel(vm_offset_t addr) 1408{ 1409 struct pmap *pmap; 1410 vm_paddr_t ptppaddr; 1411 vm_page_t nkpg; 1412 pd_entry_t newpdir; 1413 pt_entry_t *pde; 1414 1415 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1416 if (kernel_vm_end == 0) { 1417 kernel_vm_end = KERNBASE; 1418 nkpt = 0; 1419 while (pdir_pde(PTD, kernel_vm_end)) { 1420 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1421 nkpt++; 1422 } 1423 } 1424 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1425 while (kernel_vm_end < addr) { 1426 if (pdir_pde(PTD, kernel_vm_end)) { 1427 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1428 continue; 1429 } 1430 1431 /* 1432 * This index is bogus, but out of the way 1433 */ 1434 nkpg = vm_page_alloc(NULL, nkpt, 1435 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1436 if (!nkpg) 1437 panic("pmap_growkernel: no memory to grow kernel"); 1438 1439 nkpt++; 1440 1441 pmap_zero_page(nkpg); 1442 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1443 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1444 pdir_pde(PTD, kernel_vm_end) = newpdir; 1445 1446 mtx_lock_spin(&allpmaps_lock); 1447 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1448 pde = pmap_pde(pmap, kernel_vm_end); 1449 pde_store(pde, newpdir); 1450 } 1451 mtx_unlock_spin(&allpmaps_lock); 1452 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1453 } 1454} 1455 1456 1457/*************************************************** 1458 * page management routines. 1459 ***************************************************/ 1460 1461/* 1462 * free the pv_entry back to the free list 1463 */ 1464static PMAP_INLINE void 1465free_pv_entry(pv_entry_t pv) 1466{ 1467 pv_entry_count--; 1468 uma_zfree(pvzone, pv); 1469} 1470 1471/* 1472 * get a new pv_entry, allocating a block from the system 1473 * when needed. 1474 * the memory allocation is performed bypassing the malloc code 1475 * because of the possibility of allocations at interrupt time. 1476 */ 1477static pv_entry_t 1478get_pv_entry(void) 1479{ 1480 pv_entry_count++; 1481 if (pv_entry_high_water && 1482 (pv_entry_count > pv_entry_high_water) && 1483 (pmap_pagedaemon_waken == 0)) { 1484 pmap_pagedaemon_waken = 1; 1485 wakeup (&vm_pages_needed); 1486 } 1487 return uma_zalloc(pvzone, M_NOWAIT); 1488} 1489 1490 1491static int 1492pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1493{ 1494 pv_entry_t pv; 1495 int rtval; 1496 1497 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1498 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1499 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1500 if (pmap == pv->pv_pmap && va == pv->pv_va) 1501 break; 1502 } 1503 } else { 1504 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1505 if (va == pv->pv_va) 1506 break; 1507 } 1508 } 1509 1510 rtval = 0; 1511 if (pv) { 1512 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1513 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1514 m->md.pv_list_count--; 1515 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1516 vm_page_flag_clear(m, PG_WRITEABLE); 1517 1518 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1519 free_pv_entry(pv); 1520 } 1521 1522 return rtval; 1523} 1524 1525/* 1526 * Create a pv entry for page at pa for 1527 * (pmap, va). 1528 */ 1529static void 1530pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1531{ 1532 1533 pv_entry_t pv; 1534 1535 pv = get_pv_entry(); 1536 pv->pv_va = va; 1537 pv->pv_pmap = pmap; 1538 pv->pv_ptem = mpte; 1539 1540 vm_page_lock_queues(); 1541 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1542 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1543 m->md.pv_list_count++; 1544 1545 vm_page_unlock_queues(); 1546} 1547 1548/* 1549 * pmap_remove_pte: do the things to unmap a page in a process 1550 */ 1551static int 1552pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1553{ 1554 pt_entry_t oldpte; 1555 vm_page_t m, mpte; 1556 1557 oldpte = pte_load_clear(ptq); 1558 if (oldpte & PG_W) 1559 pmap->pm_stats.wired_count -= 1; 1560 /* 1561 * Machines that don't support invlpg, also don't support 1562 * PG_G. 1563 */ 1564 if (oldpte & PG_G) 1565 pmap_invalidate_page(kernel_pmap, va); 1566 pmap->pm_stats.resident_count -= 1; 1567 if (oldpte & PG_MANAGED) { 1568 m = PHYS_TO_VM_PAGE(oldpte); 1569 if (oldpte & PG_M) { 1570#if defined(PMAP_DIAGNOSTIC) 1571 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1572 printf( 1573 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1574 va, oldpte); 1575 } 1576#endif 1577 if (pmap_track_modified(va)) 1578 vm_page_dirty(m); 1579 } 1580 if (oldpte & PG_A) 1581 vm_page_flag_set(m, PG_REFERENCED); 1582 return pmap_remove_entry(pmap, m, va); 1583 } else { 1584 mpte = PHYS_TO_VM_PAGE(*pmap_pde(pmap, va)); 1585 return pmap_unuse_pt(pmap, va, mpte); 1586 } 1587} 1588 1589/* 1590 * Remove a single page from a process address space 1591 */ 1592static void 1593pmap_remove_page(pmap_t pmap, vm_offset_t va) 1594{ 1595 pt_entry_t *pte; 1596 1597 if ((pte = pmap_pte(pmap, va)) == NULL || *pte == 0) 1598 return; 1599 pmap_remove_pte(pmap, pte, va); 1600 pmap_invalidate_page(pmap, va); 1601} 1602 1603/* 1604 * Remove the given range of addresses from the specified map. 1605 * 1606 * It is assumed that the start and end are properly 1607 * rounded to the page size. 1608 */ 1609void 1610pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1611{ 1612 vm_offset_t pdnxt; 1613 pd_entry_t ptpaddr; 1614 pt_entry_t *pte; 1615 int anyvalid; 1616 1617 if (pmap == NULL) 1618 return; 1619 1620 if (pmap->pm_stats.resident_count == 0) 1621 return; 1622 1623 /* 1624 * special handling of removing one page. a very 1625 * common operation and easy to short circuit some 1626 * code. 1627 */ 1628 if ((sva + PAGE_SIZE == eva) && 1629 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1630 pmap_remove_page(pmap, sva); 1631 return; 1632 } 1633 1634 anyvalid = 0; 1635 1636 for (; sva < eva; sva = pdnxt) { 1637 unsigned pdirindex; 1638 1639 /* 1640 * Calculate index for next page table. 1641 */ 1642 pdnxt = (sva + NBPDR) & ~PDRMASK; 1643 if (pmap->pm_stats.resident_count == 0) 1644 break; 1645 1646 pdirindex = sva >> PDRSHIFT; 1647 ptpaddr = pmap->pm_pdir[pdirindex]; 1648 1649 /* 1650 * Weed out invalid mappings. Note: we assume that the page 1651 * directory table is always allocated, and in kernel virtual. 1652 */ 1653 if (ptpaddr == 0) 1654 continue; 1655 1656 /* 1657 * Check for large page. 1658 */ 1659 if ((ptpaddr & PG_PS) != 0) { 1660 pmap->pm_pdir[pdirindex] = 0; 1661 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1662 anyvalid = 1; 1663 continue; 1664 } 1665 1666 /* 1667 * Limit our scan to either the end of the va represented 1668 * by the current page table page, or to the end of the 1669 * range being removed. 1670 */ 1671 if (pdnxt > eva) 1672 pdnxt = eva; 1673 1674 for (; sva != pdnxt; sva += PAGE_SIZE) { 1675 if ((pte = pmap_pte(pmap, sva)) == NULL || 1676 *pte == 0) 1677 continue; 1678 anyvalid = 1; 1679 if (pmap_remove_pte(pmap, pte, sva)) 1680 break; 1681 } 1682 } 1683 1684 if (anyvalid) 1685 pmap_invalidate_all(pmap); 1686} 1687 1688/* 1689 * Routine: pmap_remove_all 1690 * Function: 1691 * Removes this physical page from 1692 * all physical maps in which it resides. 1693 * Reflects back modify bits to the pager. 1694 * 1695 * Notes: 1696 * Original versions of this routine were very 1697 * inefficient because they iteratively called 1698 * pmap_remove (slow...) 1699 */ 1700 1701void 1702pmap_remove_all(vm_page_t m) 1703{ 1704 register pv_entry_t pv; 1705 pt_entry_t *pte, tpte; 1706 1707#if defined(PMAP_DIAGNOSTIC) 1708 /* 1709 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1710 */ 1711 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1712 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1713 VM_PAGE_TO_PHYS(m)); 1714 } 1715#endif 1716 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1717 sched_pin(); 1718 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1719 pv->pv_pmap->pm_stats.resident_count--; 1720 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1721 tpte = pte_load_clear(pte); 1722 if (tpte & PG_W) 1723 pv->pv_pmap->pm_stats.wired_count--; 1724 if (tpte & PG_A) 1725 vm_page_flag_set(m, PG_REFERENCED); 1726 1727 /* 1728 * Update the vm_page_t clean and reference bits. 1729 */ 1730 if (tpte & PG_M) { 1731#if defined(PMAP_DIAGNOSTIC) 1732 if (pmap_nw_modified((pt_entry_t) tpte)) { 1733 printf( 1734 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1735 pv->pv_va, tpte); 1736 } 1737#endif 1738 if (pmap_track_modified(pv->pv_va)) 1739 vm_page_dirty(m); 1740 } 1741 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1742 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1743 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1744 m->md.pv_list_count--; 1745 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 1746 free_pv_entry(pv); 1747 } 1748 vm_page_flag_clear(m, PG_WRITEABLE); 1749 sched_unpin(); 1750} 1751 1752/* 1753 * Set the physical protection on the 1754 * specified range of this map as requested. 1755 */ 1756void 1757pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1758{ 1759 vm_offset_t pdnxt; 1760 pd_entry_t ptpaddr; 1761 int anychanged; 1762 1763 if (pmap == NULL) 1764 return; 1765 1766 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1767 pmap_remove(pmap, sva, eva); 1768 return; 1769 } 1770 1771 if (prot & VM_PROT_WRITE) 1772 return; 1773 1774 anychanged = 0; 1775 1776 for (; sva < eva; sva = pdnxt) { 1777 unsigned pdirindex; 1778 1779 pdnxt = (sva + NBPDR) & ~PDRMASK; 1780 1781 pdirindex = sva >> PDRSHIFT; 1782 ptpaddr = pmap->pm_pdir[pdirindex]; 1783 1784 /* 1785 * Weed out invalid mappings. Note: we assume that the page 1786 * directory table is always allocated, and in kernel virtual. 1787 */ 1788 if (ptpaddr == 0) 1789 continue; 1790 1791 /* 1792 * Check for large page. 1793 */ 1794 if ((ptpaddr & PG_PS) != 0) { 1795 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1796 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1797 anychanged = 1; 1798 continue; 1799 } 1800 1801 if (pdnxt > eva) 1802 pdnxt = eva; 1803 1804 for (; sva != pdnxt; sva += PAGE_SIZE) { 1805 pt_entry_t pbits; 1806 pt_entry_t *pte; 1807 vm_page_t m; 1808 1809 if ((pte = pmap_pte(pmap, sva)) == NULL) 1810 continue; 1811 pbits = *pte; 1812 if (pbits & PG_MANAGED) { 1813 m = NULL; 1814 if (pbits & PG_A) { 1815 m = PHYS_TO_VM_PAGE(pbits); 1816 vm_page_flag_set(m, PG_REFERENCED); 1817 pbits &= ~PG_A; 1818 } 1819 if ((pbits & PG_M) != 0 && 1820 pmap_track_modified(sva)) { 1821 if (m == NULL) 1822 m = PHYS_TO_VM_PAGE(pbits); 1823 vm_page_dirty(m); 1824 pbits &= ~PG_M; 1825 } 1826 } 1827 1828 pbits &= ~PG_RW; 1829 1830 if (pbits != *pte) { 1831 pte_store(pte, pbits); 1832 anychanged = 1; 1833 } 1834 } 1835 } 1836 if (anychanged) 1837 pmap_invalidate_all(pmap); 1838} 1839 1840/* 1841 * Insert the given physical page (p) at 1842 * the specified virtual address (v) in the 1843 * target physical map with the protection requested. 1844 * 1845 * If specified, the page will be wired down, meaning 1846 * that the related pte can not be reclaimed. 1847 * 1848 * NB: This is the only routine which MAY NOT lazy-evaluate 1849 * or lose information. That is, this routine must actually 1850 * insert this page into the given map NOW. 1851 */ 1852void 1853pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1854 boolean_t wired) 1855{ 1856 vm_paddr_t pa; 1857 register pt_entry_t *pte; 1858 vm_paddr_t opa; 1859 pt_entry_t origpte, newpte; 1860 vm_page_t mpte; 1861 1862 if (pmap == NULL) 1863 return; 1864 1865 va &= PG_FRAME; 1866#ifdef PMAP_DIAGNOSTIC 1867 if (va > VM_MAX_KERNEL_ADDRESS) 1868 panic("pmap_enter: toobig"); 1869 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1870 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1871#endif 1872 1873 mpte = NULL; 1874 /* 1875 * In the case that a page table page is not 1876 * resident, we are creating it here. 1877 */ 1878 if (va < VM_MAXUSER_ADDRESS) { 1879 mpte = pmap_allocpte(pmap, va); 1880 } 1881#if 0 && defined(PMAP_DIAGNOSTIC) 1882 else { 1883 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1884 origpte = *pdeaddr; 1885 if ((origpte & PG_V) == 0) { 1886 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1887 pmap->pm_pdir[PTDPTDI], origpte, va); 1888 } 1889 } 1890#endif 1891 1892 pte = pmap_pte(pmap, va); 1893 1894 /* 1895 * Page Directory table entry not valid, we need a new PT page 1896 */ 1897 if (pte == NULL) { 1898 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1899 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1900 } 1901 1902 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 1903 origpte = *pte; 1904 opa = origpte & PG_FRAME; 1905 1906 if (origpte & PG_PS) { 1907 /* 1908 * Yes, I know this will truncate upper address bits for PAE, 1909 * but I'm actually more interested in the lower bits 1910 */ 1911 printf("pmap_enter: va %p, pte %p, origpte %p\n", 1912 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 1913 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1914 } 1915 1916 /* 1917 * Mapping has not changed, must be protection or wiring change. 1918 */ 1919 if (origpte && (opa == pa)) { 1920 /* 1921 * Wiring change, just update stats. We don't worry about 1922 * wiring PT pages as they remain resident as long as there 1923 * are valid mappings in them. Hence, if a user page is wired, 1924 * the PT page will be also. 1925 */ 1926 if (wired && ((origpte & PG_W) == 0)) 1927 pmap->pm_stats.wired_count++; 1928 else if (!wired && (origpte & PG_W)) 1929 pmap->pm_stats.wired_count--; 1930 1931#if defined(PMAP_DIAGNOSTIC) 1932 if (pmap_nw_modified((pt_entry_t) origpte)) { 1933 printf( 1934 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 1935 va, origpte); 1936 } 1937#endif 1938 1939 /* 1940 * Remove extra pte reference 1941 */ 1942 if (mpte) 1943 mpte->hold_count--; 1944 1945 /* 1946 * We might be turning off write access to the page, 1947 * so we go ahead and sense modify status. 1948 */ 1949 if (origpte & PG_MANAGED) { 1950 if ((origpte & PG_M) && pmap_track_modified(va)) { 1951 vm_page_t om; 1952 om = PHYS_TO_VM_PAGE(opa); 1953 vm_page_dirty(om); 1954 } 1955 pa |= PG_MANAGED; 1956 } 1957 goto validate; 1958 } 1959 /* 1960 * Mapping has changed, invalidate old range and fall through to 1961 * handle validating new mapping. 1962 */ 1963 if (opa) { 1964 int err; 1965 vm_page_lock_queues(); 1966 err = pmap_remove_pte(pmap, pte, va); 1967 vm_page_unlock_queues(); 1968 if (err) 1969 panic("pmap_enter: pte vanished, va: 0x%x", va); 1970 } 1971 1972 /* 1973 * Enter on the PV list if part of our managed memory. Note that we 1974 * raise IPL while manipulating pv_table since pmap_enter can be 1975 * called at interrupt time. 1976 */ 1977 if (pmap_initialized && 1978 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 1979 pmap_insert_entry(pmap, va, mpte, m); 1980 pa |= PG_MANAGED; 1981 } 1982 1983 /* 1984 * Increment counters 1985 */ 1986 pmap->pm_stats.resident_count++; 1987 if (wired) 1988 pmap->pm_stats.wired_count++; 1989 1990validate: 1991 /* 1992 * Now validate mapping with desired protection/wiring. 1993 */ 1994 newpte = (pt_entry_t)(pa | PG_V); 1995 if ((prot & VM_PROT_WRITE) != 0) 1996 newpte |= PG_RW; 1997 if (wired) 1998 newpte |= PG_W; 1999 if (va < VM_MAXUSER_ADDRESS) 2000 newpte |= PG_U; 2001 if (pmap == kernel_pmap) 2002 newpte |= pgeflag; 2003 2004 /* 2005 * if the mapping or permission bits are different, we need 2006 * to update the pte. 2007 */ 2008 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2009 pte_store(pte, newpte | PG_A); 2010 /*if (origpte)*/ { 2011 pmap_invalidate_page(pmap, va); 2012 } 2013 } 2014} 2015 2016/* 2017 * this code makes some *MAJOR* assumptions: 2018 * 1. Current pmap & pmap exists. 2019 * 2. Not wired. 2020 * 3. Read access. 2021 * 4. No page table pages. 2022 * 5. Tlbflush is deferred to calling procedure. 2023 * 6. Page IS managed. 2024 * but is *MUCH* faster than pmap_enter... 2025 */ 2026 2027vm_page_t 2028pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2029{ 2030 pt_entry_t *pte; 2031 vm_paddr_t pa; 2032 2033 /* 2034 * In the case that a page table page is not 2035 * resident, we are creating it here. 2036 */ 2037 if (va < VM_MAXUSER_ADDRESS) { 2038 unsigned ptepindex; 2039 pd_entry_t ptepa; 2040 2041 /* 2042 * Calculate pagetable page index 2043 */ 2044 ptepindex = va >> PDRSHIFT; 2045 if (mpte && (mpte->pindex == ptepindex)) { 2046 mpte->hold_count++; 2047 } else { 2048retry: 2049 /* 2050 * Get the page directory entry 2051 */ 2052 ptepa = pmap->pm_pdir[ptepindex]; 2053 2054 /* 2055 * If the page table page is mapped, we just increment 2056 * the hold count, and activate it. 2057 */ 2058 if (ptepa) { 2059 if (ptepa & PG_PS) 2060 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2061 mpte = PHYS_TO_VM_PAGE(ptepa); 2062 mpte->hold_count++; 2063 } else { 2064 mpte = _pmap_allocpte(pmap, ptepindex); 2065 if (mpte == NULL) 2066 goto retry; 2067 } 2068 } 2069 } else { 2070 mpte = NULL; 2071 } 2072 2073 /* 2074 * This call to vtopte makes the assumption that we are 2075 * entering the page into the current pmap. In order to support 2076 * quick entry into any pmap, one would likely use pmap_pte_quick. 2077 * But that isn't as quick as vtopte. 2078 */ 2079 pte = vtopte(va); 2080 if (*pte) { 2081 if (mpte != NULL) { 2082 vm_page_lock_queues(); 2083 pmap_unwire_pte_hold(pmap, mpte); 2084 vm_page_unlock_queues(); 2085 } 2086 return 0; 2087 } 2088 2089 /* 2090 * Enter on the PV list if part of our managed memory. Note that we 2091 * raise IPL while manipulating pv_table since pmap_enter can be 2092 * called at interrupt time. 2093 */ 2094 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2095 pmap_insert_entry(pmap, va, mpte, m); 2096 2097 /* 2098 * Increment counters 2099 */ 2100 pmap->pm_stats.resident_count++; 2101 2102 pa = VM_PAGE_TO_PHYS(m); 2103 2104 /* 2105 * Now validate mapping with RO protection 2106 */ 2107 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2108 pte_store(pte, pa | PG_V | PG_U); 2109 else 2110 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2111 2112 return mpte; 2113} 2114 2115/* 2116 * Make a temporary mapping for a physical address. This is only intended 2117 * to be used for panic dumps. 2118 */ 2119void * 2120pmap_kenter_temporary(vm_paddr_t pa, int i) 2121{ 2122 vm_offset_t va; 2123 2124 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2125 pmap_kenter(va, pa); 2126#ifndef I386_CPU 2127 invlpg(va); 2128#else 2129 invltlb(); 2130#endif 2131 return ((void *)crashdumpmap); 2132} 2133 2134/* 2135 * This code maps large physical mmap regions into the 2136 * processor address space. Note that some shortcuts 2137 * are taken, but the code works. 2138 */ 2139void 2140pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2141 vm_object_t object, vm_pindex_t pindex, 2142 vm_size_t size) 2143{ 2144 vm_page_t p; 2145 2146 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2147 KASSERT(object->type == OBJT_DEVICE, 2148 ("pmap_object_init_pt: non-device object")); 2149 if (pseflag && 2150 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2151 int i; 2152 vm_page_t m[1]; 2153 unsigned int ptepindex; 2154 int npdes; 2155 pd_entry_t ptepa; 2156 2157 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2158 return; 2159retry: 2160 p = vm_page_lookup(object, pindex); 2161 if (p != NULL) { 2162 vm_page_lock_queues(); 2163 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2164 goto retry; 2165 } else { 2166 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2167 if (p == NULL) 2168 return; 2169 m[0] = p; 2170 2171 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2172 vm_page_lock_queues(); 2173 vm_page_free(p); 2174 vm_page_unlock_queues(); 2175 return; 2176 } 2177 2178 p = vm_page_lookup(object, pindex); 2179 vm_page_lock_queues(); 2180 vm_page_wakeup(p); 2181 } 2182 vm_page_unlock_queues(); 2183 2184 ptepa = VM_PAGE_TO_PHYS(p); 2185 if (ptepa & (NBPDR - 1)) 2186 return; 2187 2188 p->valid = VM_PAGE_BITS_ALL; 2189 2190 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2191 npdes = size >> PDRSHIFT; 2192 for(i = 0; i < npdes; i++) { 2193 pde_store(&pmap->pm_pdir[ptepindex], 2194 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2195 ptepa += NBPDR; 2196 ptepindex += 1; 2197 } 2198 pmap_invalidate_all(pmap); 2199 } 2200} 2201 2202/* 2203 * Routine: pmap_change_wiring 2204 * Function: Change the wiring attribute for a map/virtual-address 2205 * pair. 2206 * In/out conditions: 2207 * The mapping must already exist in the pmap. 2208 */ 2209void 2210pmap_change_wiring(pmap, va, wired) 2211 register pmap_t pmap; 2212 vm_offset_t va; 2213 boolean_t wired; 2214{ 2215 register pt_entry_t *pte; 2216 2217 if (pmap == NULL) 2218 return; 2219 2220 pte = pmap_pte(pmap, va); 2221 2222 if (wired && !pmap_pte_w(pte)) 2223 pmap->pm_stats.wired_count++; 2224 else if (!wired && pmap_pte_w(pte)) 2225 pmap->pm_stats.wired_count--; 2226 2227 /* 2228 * Wiring is not a hardware characteristic so there is no need to 2229 * invalidate TLB. 2230 */ 2231 pmap_pte_set_w(pte, wired); 2232} 2233 2234 2235 2236/* 2237 * Copy the range specified by src_addr/len 2238 * from the source map to the range dst_addr/len 2239 * in the destination map. 2240 * 2241 * This routine is only advisory and need not do anything. 2242 */ 2243 2244void 2245pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2246 vm_offset_t src_addr) 2247{ 2248 vm_offset_t addr; 2249 vm_offset_t end_addr = src_addr + len; 2250 vm_offset_t pdnxt; 2251 vm_page_t m; 2252 2253 if (dst_addr != src_addr) 2254 return; 2255 2256 if (!pmap_is_current(src_pmap)) 2257 return; 2258 2259 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2260 pt_entry_t *src_pte, *dst_pte; 2261 vm_page_t dstmpte, srcmpte; 2262 pd_entry_t srcptepaddr; 2263 unsigned ptepindex; 2264 2265 if (addr >= UPT_MIN_ADDRESS) 2266 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2267 2268 /* 2269 * Don't let optional prefaulting of pages make us go 2270 * way below the low water mark of free pages or way 2271 * above high water mark of used pv entries. 2272 */ 2273 if (cnt.v_free_count < cnt.v_free_reserved || 2274 pv_entry_count > pv_entry_high_water) 2275 break; 2276 2277 pdnxt = (addr + NBPDR) & ~PDRMASK; 2278 ptepindex = addr >> PDRSHIFT; 2279 2280 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2281 if (srcptepaddr == 0) 2282 continue; 2283 2284 if (srcptepaddr & PG_PS) { 2285 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2286 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2287 dst_pmap->pm_stats.resident_count += 2288 NBPDR / PAGE_SIZE; 2289 } 2290 continue; 2291 } 2292 2293 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2294 if (srcmpte->hold_count == 0 || (srcmpte->flags & PG_BUSY)) 2295 continue; 2296 2297 if (pdnxt > end_addr) 2298 pdnxt = end_addr; 2299 2300 src_pte = vtopte(addr); 2301 while (addr < pdnxt) { 2302 pt_entry_t ptetemp; 2303 ptetemp = *src_pte; 2304 /* 2305 * we only virtual copy managed pages 2306 */ 2307 if ((ptetemp & PG_MANAGED) != 0) { 2308 /* 2309 * We have to check after allocpte for the 2310 * pte still being around... allocpte can 2311 * block. 2312 */ 2313 dstmpte = pmap_allocpte(dst_pmap, addr); 2314 dst_pte = pmap_pte(dst_pmap, addr); 2315 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2316 /* 2317 * Clear the modified and 2318 * accessed (referenced) bits 2319 * during the copy. 2320 */ 2321 m = PHYS_TO_VM_PAGE(ptetemp); 2322 *dst_pte = ptetemp & ~(PG_M | PG_A); 2323 dst_pmap->pm_stats.resident_count++; 2324 pmap_insert_entry(dst_pmap, addr, 2325 dstmpte, m); 2326 } else { 2327 vm_page_lock_queues(); 2328 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2329 vm_page_unlock_queues(); 2330 } 2331 if (dstmpte->hold_count >= srcmpte->hold_count) 2332 break; 2333 } 2334 addr += PAGE_SIZE; 2335 src_pte++; 2336 } 2337 } 2338} 2339 2340static __inline void 2341pagezero(void *page) 2342{ 2343#if defined(I686_CPU) 2344 if (cpu_class == CPUCLASS_686) { 2345#if defined(CPU_ENABLE_SSE) 2346 if (cpu_feature & CPUID_SSE2) 2347 sse2_pagezero(page); 2348 else 2349#endif 2350 i686_pagezero(page); 2351 } else 2352#endif 2353 bzero(page, PAGE_SIZE); 2354} 2355 2356/* 2357 * pmap_zero_page zeros the specified hardware page by mapping 2358 * the page into KVM and using bzero to clear its contents. 2359 */ 2360void 2361pmap_zero_page(vm_page_t m) 2362{ 2363 2364 mtx_lock(&CMAPCADDR12_lock); 2365 if (*CMAP2) 2366 panic("pmap_zero_page: CMAP2 busy"); 2367 sched_pin(); 2368 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2369 invlcaddr(CADDR2); 2370 pagezero(CADDR2); 2371 *CMAP2 = 0; 2372 sched_unpin(); 2373 mtx_unlock(&CMAPCADDR12_lock); 2374} 2375 2376/* 2377 * pmap_zero_page_area zeros the specified hardware page by mapping 2378 * the page into KVM and using bzero to clear its contents. 2379 * 2380 * off and size may not cover an area beyond a single hardware page. 2381 */ 2382void 2383pmap_zero_page_area(vm_page_t m, int off, int size) 2384{ 2385 2386 mtx_lock(&CMAPCADDR12_lock); 2387 if (*CMAP2) 2388 panic("pmap_zero_page: CMAP2 busy"); 2389 sched_pin(); 2390 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2391 invlcaddr(CADDR2); 2392 if (off == 0 && size == PAGE_SIZE) 2393 pagezero(CADDR2); 2394 else 2395 bzero((char *)CADDR2 + off, size); 2396 *CMAP2 = 0; 2397 sched_unpin(); 2398 mtx_unlock(&CMAPCADDR12_lock); 2399} 2400 2401/* 2402 * pmap_zero_page_idle zeros the specified hardware page by mapping 2403 * the page into KVM and using bzero to clear its contents. This 2404 * is intended to be called from the vm_pagezero process only and 2405 * outside of Giant. 2406 */ 2407void 2408pmap_zero_page_idle(vm_page_t m) 2409{ 2410 2411 if (*CMAP3) 2412 panic("pmap_zero_page: CMAP3 busy"); 2413 sched_pin(); 2414 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2415 invlcaddr(CADDR3); 2416 pagezero(CADDR3); 2417 *CMAP3 = 0; 2418 sched_unpin(); 2419} 2420 2421/* 2422 * pmap_copy_page copies the specified (machine independent) 2423 * page by mapping the page into virtual memory and using 2424 * bcopy to copy the page, one machine dependent page at a 2425 * time. 2426 */ 2427void 2428pmap_copy_page(vm_page_t src, vm_page_t dst) 2429{ 2430 2431 mtx_lock(&CMAPCADDR12_lock); 2432 if (*CMAP1) 2433 panic("pmap_copy_page: CMAP1 busy"); 2434 if (*CMAP2) 2435 panic("pmap_copy_page: CMAP2 busy"); 2436 sched_pin(); 2437#ifdef I386_CPU 2438 invltlb(); 2439#else 2440 invlpg((u_int)CADDR1); 2441 invlpg((u_int)CADDR2); 2442#endif 2443 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2444 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2445 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2446 *CMAP1 = 0; 2447 *CMAP2 = 0; 2448 sched_unpin(); 2449 mtx_unlock(&CMAPCADDR12_lock); 2450} 2451 2452/* 2453 * Returns true if the pmap's pv is one of the first 2454 * 16 pvs linked to from this page. This count may 2455 * be changed upwards or downwards in the future; it 2456 * is only necessary that true be returned for a small 2457 * subset of pmaps for proper page aging. 2458 */ 2459boolean_t 2460pmap_page_exists_quick(pmap, m) 2461 pmap_t pmap; 2462 vm_page_t m; 2463{ 2464 pv_entry_t pv; 2465 int loops = 0; 2466 2467 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2468 return FALSE; 2469 2470 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2471 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2472 if (pv->pv_pmap == pmap) { 2473 return TRUE; 2474 } 2475 loops++; 2476 if (loops >= 16) 2477 break; 2478 } 2479 return (FALSE); 2480} 2481 2482#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2483/* 2484 * Remove all pages from specified address space 2485 * this aids process exit speeds. Also, this code 2486 * is special cased for current process only, but 2487 * can have the more generic (and slightly slower) 2488 * mode enabled. This is much faster than pmap_remove 2489 * in the case of running down an entire address space. 2490 */ 2491void 2492pmap_remove_pages(pmap, sva, eva) 2493 pmap_t pmap; 2494 vm_offset_t sva, eva; 2495{ 2496 pt_entry_t *pte, tpte; 2497 vm_page_t m; 2498 pv_entry_t pv, npv; 2499 2500#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2501 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2502 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2503 return; 2504 } 2505#endif 2506 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2507 sched_pin(); 2508 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2509 2510 if (pv->pv_va >= eva || pv->pv_va < sva) { 2511 npv = TAILQ_NEXT(pv, pv_plist); 2512 continue; 2513 } 2514 2515#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2516 pte = vtopte(pv->pv_va); 2517#else 2518 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2519#endif 2520 tpte = *pte; 2521 2522 if (tpte == 0) { 2523 printf("TPTE at %p IS ZERO @ VA %08x\n", 2524 pte, pv->pv_va); 2525 panic("bad pte"); 2526 } 2527 2528/* 2529 * We cannot remove wired pages from a process' mapping at this time 2530 */ 2531 if (tpte & PG_W) { 2532 npv = TAILQ_NEXT(pv, pv_plist); 2533 continue; 2534 } 2535 2536 m = PHYS_TO_VM_PAGE(tpte); 2537 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2538 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2539 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2540 2541 KASSERT(m < &vm_page_array[vm_page_array_size], 2542 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2543 2544 pv->pv_pmap->pm_stats.resident_count--; 2545 2546 pte_clear(pte); 2547 2548 /* 2549 * Update the vm_page_t clean and reference bits. 2550 */ 2551 if (tpte & PG_M) { 2552 vm_page_dirty(m); 2553 } 2554 2555 npv = TAILQ_NEXT(pv, pv_plist); 2556 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2557 2558 m->md.pv_list_count--; 2559 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2560 if (TAILQ_FIRST(&m->md.pv_list) == NULL) { 2561 vm_page_flag_clear(m, PG_WRITEABLE); 2562 } 2563 2564 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2565 free_pv_entry(pv); 2566 } 2567 sched_unpin(); 2568 pmap_invalidate_all(pmap); 2569} 2570 2571/* 2572 * pmap_is_modified: 2573 * 2574 * Return whether or not the specified physical page was modified 2575 * in any physical maps. 2576 */ 2577boolean_t 2578pmap_is_modified(vm_page_t m) 2579{ 2580 pv_entry_t pv; 2581 pt_entry_t *pte; 2582 2583 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2584 return FALSE; 2585 2586 sched_pin(); 2587 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2588 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2589 /* 2590 * if the bit being tested is the modified bit, then 2591 * mark clean_map and ptes as never 2592 * modified. 2593 */ 2594 if (!pmap_track_modified(pv->pv_va)) 2595 continue; 2596#if defined(PMAP_DIAGNOSTIC) 2597 if (!pv->pv_pmap) { 2598 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 2599 continue; 2600 } 2601#endif 2602 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2603 if (*pte & PG_M) { 2604 sched_unpin(); 2605 return TRUE; 2606 } 2607 } 2608 sched_unpin(); 2609 return (FALSE); 2610} 2611 2612/* 2613 * pmap_is_prefaultable: 2614 * 2615 * Return whether or not the specified virtual address is elgible 2616 * for prefault. 2617 */ 2618boolean_t 2619pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2620{ 2621 pt_entry_t *pte; 2622 2623 if ((*pmap_pde(pmap, addr)) == 0) 2624 return (FALSE); 2625 pte = vtopte(addr); 2626 if (*pte) 2627 return (FALSE); 2628 return (TRUE); 2629} 2630 2631/* 2632 * Clear the given bit in each of the given page's ptes. 2633 */ 2634static __inline void 2635pmap_clear_ptes(vm_page_t m, int bit) 2636{ 2637 register pv_entry_t pv; 2638 pt_entry_t pbits, *pte; 2639 2640 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) || 2641 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2642 return; 2643 2644 sched_pin(); 2645 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2646 /* 2647 * Loop over all current mappings setting/clearing as appropos If 2648 * setting RO do we need to clear the VAC? 2649 */ 2650 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2651 /* 2652 * don't write protect pager mappings 2653 */ 2654 if (bit == PG_RW) { 2655 if (!pmap_track_modified(pv->pv_va)) 2656 continue; 2657 } 2658 2659#if defined(PMAP_DIAGNOSTIC) 2660 if (!pv->pv_pmap) { 2661 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 2662 continue; 2663 } 2664#endif 2665 2666 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2667 pbits = *pte; 2668 if (pbits & bit) { 2669 if (bit == PG_RW) { 2670 if (pbits & PG_M) { 2671 vm_page_dirty(m); 2672 } 2673 pte_store(pte, pbits & ~(PG_M|PG_RW)); 2674 } else { 2675 pte_store(pte, pbits & ~bit); 2676 } 2677 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2678 } 2679 } 2680 if (bit == PG_RW) 2681 vm_page_flag_clear(m, PG_WRITEABLE); 2682 sched_unpin(); 2683} 2684 2685/* 2686 * pmap_page_protect: 2687 * 2688 * Lower the permission for all mappings to a given page. 2689 */ 2690void 2691pmap_page_protect(vm_page_t m, vm_prot_t prot) 2692{ 2693 if ((prot & VM_PROT_WRITE) == 0) { 2694 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2695 pmap_clear_ptes(m, PG_RW); 2696 } else { 2697 pmap_remove_all(m); 2698 } 2699 } 2700} 2701 2702/* 2703 * pmap_ts_referenced: 2704 * 2705 * Return a count of reference bits for a page, clearing those bits. 2706 * It is not necessary for every reference bit to be cleared, but it 2707 * is necessary that 0 only be returned when there are truly no 2708 * reference bits set. 2709 * 2710 * XXX: The exact number of bits to check and clear is a matter that 2711 * should be tested and standardized at some point in the future for 2712 * optimal aging of shared pages. 2713 */ 2714int 2715pmap_ts_referenced(vm_page_t m) 2716{ 2717 register pv_entry_t pv, pvf, pvn; 2718 pt_entry_t *pte; 2719 pt_entry_t v; 2720 int rtval = 0; 2721 2722 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2723 return (rtval); 2724 2725 sched_pin(); 2726 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2727 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2728 2729 pvf = pv; 2730 2731 do { 2732 pvn = TAILQ_NEXT(pv, pv_list); 2733 2734 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2735 2736 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2737 2738 if (!pmap_track_modified(pv->pv_va)) 2739 continue; 2740 2741 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2742 2743 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2744 atomic_clear_int((u_int *)pte, PG_A); 2745 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2746 2747 rtval++; 2748 if (rtval > 4) { 2749 break; 2750 } 2751 } 2752 } while ((pv = pvn) != NULL && pv != pvf); 2753 } 2754 sched_unpin(); 2755 2756 return (rtval); 2757} 2758 2759/* 2760 * Clear the modify bits on the specified physical page. 2761 */ 2762void 2763pmap_clear_modify(vm_page_t m) 2764{ 2765 pmap_clear_ptes(m, PG_M); 2766} 2767 2768/* 2769 * pmap_clear_reference: 2770 * 2771 * Clear the reference bit on the specified physical page. 2772 */ 2773void 2774pmap_clear_reference(vm_page_t m) 2775{ 2776 pmap_clear_ptes(m, PG_A); 2777} 2778 2779/* 2780 * Miscellaneous support routines follow 2781 */ 2782 2783/* 2784 * Map a set of physical memory pages into the kernel virtual 2785 * address space. Return a pointer to where it is mapped. This 2786 * routine is intended to be used for mapping device memory, 2787 * NOT real memory. 2788 */ 2789void * 2790pmap_mapdev(pa, size) 2791 vm_paddr_t pa; 2792 vm_size_t size; 2793{ 2794 vm_offset_t va, tmpva, offset; 2795 2796 offset = pa & PAGE_MASK; 2797 size = roundup(offset + size, PAGE_SIZE); 2798 pa = pa & PG_FRAME; 2799 2800 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2801 va = KERNBASE + pa; 2802 else 2803 va = kmem_alloc_nofault(kernel_map, size); 2804 if (!va) 2805 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2806 2807 for (tmpva = va; size > 0; ) { 2808 pmap_kenter(tmpva, pa); 2809 size -= PAGE_SIZE; 2810 tmpva += PAGE_SIZE; 2811 pa += PAGE_SIZE; 2812 } 2813 pmap_invalidate_range(kernel_pmap, va, tmpva); 2814 return ((void *)(va + offset)); 2815} 2816 2817void 2818pmap_unmapdev(va, size) 2819 vm_offset_t va; 2820 vm_size_t size; 2821{ 2822 vm_offset_t base, offset, tmpva; 2823 2824 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2825 return; 2826 base = va & PG_FRAME; 2827 offset = va & PAGE_MASK; 2828 size = roundup(offset + size, PAGE_SIZE); 2829 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2830 pmap_kremove(tmpva); 2831 pmap_invalidate_range(kernel_pmap, va, tmpva); 2832 kmem_free(kernel_map, base, size); 2833} 2834 2835/* 2836 * perform the pmap work for mincore 2837 */ 2838int 2839pmap_mincore(pmap, addr) 2840 pmap_t pmap; 2841 vm_offset_t addr; 2842{ 2843 pt_entry_t *ptep, pte; 2844 vm_page_t m; 2845 int val = 0; 2846 2847 ptep = pmap_pte(pmap, addr); 2848 if (ptep == 0) { 2849 return 0; 2850 } 2851 2852 if ((pte = *ptep) != 0) { 2853 vm_paddr_t pa; 2854 2855 val = MINCORE_INCORE; 2856 if ((pte & PG_MANAGED) == 0) 2857 return val; 2858 2859 pa = pte & PG_FRAME; 2860 2861 m = PHYS_TO_VM_PAGE(pa); 2862 2863 /* 2864 * Modified by us 2865 */ 2866 if (pte & PG_M) 2867 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2868 else { 2869 /* 2870 * Modified by someone else 2871 */ 2872 vm_page_lock_queues(); 2873 if (m->dirty || pmap_is_modified(m)) 2874 val |= MINCORE_MODIFIED_OTHER; 2875 vm_page_unlock_queues(); 2876 } 2877 /* 2878 * Referenced by us 2879 */ 2880 if (pte & PG_A) 2881 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2882 else { 2883 /* 2884 * Referenced by someone else 2885 */ 2886 vm_page_lock_queues(); 2887 if ((m->flags & PG_REFERENCED) || 2888 pmap_ts_referenced(m)) { 2889 val |= MINCORE_REFERENCED_OTHER; 2890 vm_page_flag_set(m, PG_REFERENCED); 2891 } 2892 vm_page_unlock_queues(); 2893 } 2894 } 2895 return val; 2896} 2897 2898void 2899pmap_activate(struct thread *td) 2900{ 2901 struct proc *p = td->td_proc; 2902 pmap_t pmap, oldpmap; 2903 u_int32_t cr3; 2904 2905 critical_enter(); 2906 pmap = vmspace_pmap(td->td_proc->p_vmspace); 2907 oldpmap = PCPU_GET(curpmap); 2908#if defined(SMP) 2909 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 2910 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 2911#else 2912 oldpmap->pm_active &= ~1; 2913 pmap->pm_active |= 1; 2914#endif 2915#ifdef PAE 2916 cr3 = vtophys(pmap->pm_pdpt); 2917#else 2918 cr3 = vtophys(pmap->pm_pdir); 2919#endif 2920 /* XXXKSE this is wrong. 2921 * pmap_activate is for the current thread on the current cpu 2922 */ 2923 if (p->p_flag & P_SA) { 2924 /* Make sure all other cr3 entries are updated. */ 2925 /* what if they are running? XXXKSE (maybe abort them) */ 2926 FOREACH_THREAD_IN_PROC(p, td) { 2927 td->td_pcb->pcb_cr3 = cr3; 2928 } 2929 } else { 2930 td->td_pcb->pcb_cr3 = cr3; 2931 } 2932 load_cr3(cr3); 2933 PCPU_SET(curpmap, pmap); 2934 critical_exit(); 2935} 2936 2937vm_offset_t 2938pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 2939{ 2940 2941 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 2942 return addr; 2943 } 2944 2945 addr = (addr + PDRMASK) & ~PDRMASK; 2946 return addr; 2947} 2948 2949 2950#if defined(PMAP_DEBUG) 2951pmap_pid_dump(int pid) 2952{ 2953 pmap_t pmap; 2954 struct proc *p; 2955 int npte = 0; 2956 int index; 2957 2958 sx_slock(&allproc_lock); 2959 LIST_FOREACH(p, &allproc, p_list) { 2960 if (p->p_pid != pid) 2961 continue; 2962 2963 if (p->p_vmspace) { 2964 int i,j; 2965 index = 0; 2966 pmap = vmspace_pmap(p->p_vmspace); 2967 for (i = 0; i < NPDEPTD; i++) { 2968 pd_entry_t *pde; 2969 pt_entry_t *pte; 2970 vm_offset_t base = i << PDRSHIFT; 2971 2972 pde = &pmap->pm_pdir[i]; 2973 if (pde && pmap_pde_v(pde)) { 2974 for (j = 0; j < NPTEPG; j++) { 2975 vm_offset_t va = base + (j << PAGE_SHIFT); 2976 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 2977 if (index) { 2978 index = 0; 2979 printf("\n"); 2980 } 2981 sx_sunlock(&allproc_lock); 2982 return npte; 2983 } 2984 pte = pmap_pte(pmap, va); 2985 if (pte && pmap_pte_v(pte)) { 2986 pt_entry_t pa; 2987 vm_page_t m; 2988 pa = *pte; 2989 m = PHYS_TO_VM_PAGE(pa); 2990 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 2991 va, pa, m->hold_count, m->wire_count, m->flags); 2992 npte++; 2993 index++; 2994 if (index >= 2) { 2995 index = 0; 2996 printf("\n"); 2997 } else { 2998 printf(" "); 2999 } 3000 } 3001 } 3002 } 3003 } 3004 } 3005 } 3006 sx_sunlock(&allproc_lock); 3007 return npte; 3008} 3009#endif 3010 3011#if defined(DEBUG) 3012 3013static void pads(pmap_t pm); 3014void pmap_pvdump(vm_offset_t pa); 3015 3016/* print address space of pmap*/ 3017static void 3018pads(pm) 3019 pmap_t pm; 3020{ 3021 int i, j; 3022 vm_paddr_t va; 3023 pt_entry_t *ptep; 3024 3025 if (pm == kernel_pmap) 3026 return; 3027 for (i = 0; i < NPDEPTD; i++) 3028 if (pm->pm_pdir[i]) 3029 for (j = 0; j < NPTEPG; j++) { 3030 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3031 if (pm == kernel_pmap && va < KERNBASE) 3032 continue; 3033 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3034 continue; 3035 ptep = pmap_pte(pm, va); 3036 if (pmap_pte_v(ptep)) 3037 printf("%x:%x ", va, *ptep); 3038 }; 3039 3040} 3041 3042void 3043pmap_pvdump(pa) 3044 vm_paddr_t pa; 3045{ 3046 pv_entry_t pv; 3047 vm_page_t m; 3048 3049 printf("pa %x", pa); 3050 m = PHYS_TO_VM_PAGE(pa); 3051 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3052 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3053 pads(pv->pv_pmap); 3054 } 3055 printf(" "); 3056} 3057#endif 3058