pmap.c revision 157394
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 * Copyright (c) 2005 Alan L. Cox <alc@cs.rice.edu> 9 * All rights reserved. 10 * 11 * This code is derived from software contributed to Berkeley by 12 * the Systems Programming Group of the University of Utah Computer 13 * Science Department and William Jolitz of UUNET Technologies Inc. 14 * 15 * Redistribution and use in source and binary forms, with or without 16 * modification, are permitted provided that the following conditions 17 * are met: 18 * 1. Redistributions of source code must retain the above copyright 19 * notice, this list of conditions and the following disclaimer. 20 * 2. Redistributions in binary form must reproduce the above copyright 21 * notice, this list of conditions and the following disclaimer in the 22 * documentation and/or other materials provided with the distribution. 23 * 3. All advertising materials mentioning features or use of this software 24 * must display the following acknowledgement: 25 * This product includes software developed by the University of 26 * California, Berkeley and its contributors. 27 * 4. Neither the name of the University nor the names of its contributors 28 * may be used to endorse or promote products derived from this software 29 * without specific prior written permission. 30 * 31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 41 * SUCH DAMAGE. 42 * 43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 44 */ 45/*- 46 * Copyright (c) 2003 Networks Associates Technology, Inc. 47 * All rights reserved. 48 * 49 * This software was developed for the FreeBSD Project by Jake Burkholder, 50 * Safeport Network Services, and Network Associates Laboratories, the 51 * Security Research Division of Network Associates, Inc. under 52 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA 53 * CHATS research program. 54 * 55 * Redistribution and use in source and binary forms, with or without 56 * modification, are permitted provided that the following conditions 57 * are met: 58 * 1. Redistributions of source code must retain the above copyright 59 * notice, this list of conditions and the following disclaimer. 60 * 2. Redistributions in binary form must reproduce the above copyright 61 * notice, this list of conditions and the following disclaimer in the 62 * documentation and/or other materials provided with the distribution. 63 * 64 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 65 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 66 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 67 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 68 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 69 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 70 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 71 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 72 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 73 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 74 * SUCH DAMAGE. 75 */ 76 77#include <sys/cdefs.h> 78__FBSDID("$FreeBSD: head/sys/i386/i386/pmap.c 157394 2006-04-02 05:45:05Z alc $"); 79 80/* 81 * Manages physical address maps. 82 * 83 * In addition to hardware address maps, this 84 * module is called upon to provide software-use-only 85 * maps which may or may not be stored in the same 86 * form as hardware maps. These pseudo-maps are 87 * used to store intermediate results from copy 88 * operations to and from address spaces. 89 * 90 * Since the information managed by this module is 91 * also stored by the logical address mapping module, 92 * this module may throw away valid virtual-to-physical 93 * mappings at almost any time. However, invalidations 94 * of virtual-to-physical mappings must be done as 95 * requested. 96 * 97 * In order to cope with hardware architectures which 98 * make virtual-to-physical map invalidates expensive, 99 * this module may delay invalidate or reduced protection 100 * operations until such time as they are actually 101 * necessary. This module is given full information as 102 * to which processors are currently using which maps, 103 * and to when physical maps must be made correct. 104 */ 105 106#include "opt_cpu.h" 107#include "opt_pmap.h" 108#include "opt_msgbuf.h" 109#include "opt_smp.h" 110#include "opt_xbox.h" 111 112#include <sys/param.h> 113#include <sys/systm.h> 114#include <sys/kernel.h> 115#include <sys/lock.h> 116#include <sys/malloc.h> 117#include <sys/mman.h> 118#include <sys/msgbuf.h> 119#include <sys/mutex.h> 120#include <sys/proc.h> 121#include <sys/sx.h> 122#include <sys/vmmeter.h> 123#include <sys/sched.h> 124#include <sys/sysctl.h> 125#ifdef SMP 126#include <sys/smp.h> 127#endif 128 129#include <vm/vm.h> 130#include <vm/vm_param.h> 131#include <vm/vm_kern.h> 132#include <vm/vm_page.h> 133#include <vm/vm_map.h> 134#include <vm/vm_object.h> 135#include <vm/vm_extern.h> 136#include <vm/vm_pageout.h> 137#include <vm/vm_pager.h> 138#include <vm/uma.h> 139 140#include <machine/cpu.h> 141#include <machine/cputypes.h> 142#include <machine/md_var.h> 143#include <machine/pcb.h> 144#include <machine/specialreg.h> 145#ifdef SMP 146#include <machine/smp.h> 147#endif 148 149#ifdef XBOX 150#include <machine/xbox.h> 151#endif 152 153#if !defined(CPU_DISABLE_SSE) && defined(I686_CPU) 154#define CPU_ENABLE_SSE 155#endif 156 157#ifndef PMAP_SHPGPERPROC 158#define PMAP_SHPGPERPROC 200 159#endif 160 161#if defined(DIAGNOSTIC) 162#define PMAP_DIAGNOSTIC 163#endif 164 165#if !defined(PMAP_DIAGNOSTIC) 166#define PMAP_INLINE __inline 167#else 168#define PMAP_INLINE 169#endif 170 171/* 172 * Get PDEs and PTEs for user/kernel address space 173 */ 174#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 175#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 176 177#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 178#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 179#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 180#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 181#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 182 183#define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \ 184 atomic_clear_int((u_int *)(pte), PG_W)) 185#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 186 187struct pmap kernel_pmap_store; 188LIST_HEAD(pmaplist, pmap); 189static struct pmaplist allpmaps; 190static struct mtx allpmaps_lock; 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) */ 195int pgeflag = 0; /* PG_G or-in */ 196int pseflag = 0; /* PG_PS or-in */ 197 198static int nkpt; 199vm_offset_t kernel_vm_end; 200extern u_int32_t KERNend; 201 202#ifdef PAE 203static uma_zone_t pdptzone; 204#endif 205 206/* 207 * Data for the pv entry allocation mechanism 208 */ 209static uma_zone_t pvzone; 210static struct vm_object pvzone_obj; 211static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 212 213/* 214 * All those kernel PT submaps that BSD is so fond of 215 */ 216struct sysmaps { 217 struct mtx lock; 218 pt_entry_t *CMAP1; 219 pt_entry_t *CMAP2; 220 caddr_t CADDR1; 221 caddr_t CADDR2; 222}; 223static struct sysmaps sysmaps_pcpu[MAXCPU]; 224pt_entry_t *CMAP1 = 0; 225static pt_entry_t *CMAP3; 226caddr_t CADDR1 = 0, ptvmmap = 0; 227static caddr_t CADDR3; 228struct msgbuf *msgbufp = 0; 229 230/* 231 * Crashdump maps. 232 */ 233static caddr_t crashdumpmap; 234 235#ifdef SMP 236extern pt_entry_t *SMPpt; 237#endif 238static pt_entry_t *PMAP1 = 0, *PMAP2; 239static pt_entry_t *PADDR1 = 0, *PADDR2; 240#ifdef SMP 241static int PMAP1cpu; 242static int PMAP1changedcpu; 243SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD, 244 &PMAP1changedcpu, 0, 245 "Number of times pmap_pte_quick changed CPU with same PMAP1"); 246#endif 247static int PMAP1changed; 248SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD, 249 &PMAP1changed, 0, 250 "Number of times pmap_pte_quick changed PMAP1"); 251static int PMAP1unchanged; 252SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD, 253 &PMAP1unchanged, 0, 254 "Number of times pmap_pte_quick didn't change PMAP1"); 255static struct mtx PMAP2mutex; 256 257static PMAP_INLINE void free_pv_entry(pv_entry_t pv); 258static pv_entry_t get_pv_entry(pmap_t locked_pmap); 259static void pmap_clear_ptes(vm_page_t m, int bit); 260 261static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 262static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 263static void pmap_remove_entry(struct pmap *pmap, vm_page_t m, 264 vm_offset_t va); 265static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); 266static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, 267 vm_page_t m); 268 269static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags); 270 271static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags); 272static int _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m); 273static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 274static void pmap_pte_release(pt_entry_t *pte); 275static int pmap_unuse_pt(pmap_t, vm_offset_t); 276static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 277#ifdef PAE 278static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 279#endif 280 281CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 282CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 283 284/* 285 * Move the kernel virtual free pointer to the next 286 * 4MB. This is used to help improve performance 287 * by using a large (4MB) page for much of the kernel 288 * (.text, .data, .bss) 289 */ 290static vm_offset_t 291pmap_kmem_choose(vm_offset_t addr) 292{ 293 vm_offset_t newaddr = addr; 294 295#ifndef DISABLE_PSE 296 if (cpu_feature & CPUID_PSE) 297 newaddr = (addr + PDRMASK) & ~PDRMASK; 298#endif 299 return newaddr; 300} 301 302/* 303 * Bootstrap the system enough to run with virtual memory. 304 * 305 * On the i386 this is called after mapping has already been enabled 306 * and just syncs the pmap module with what has already been done. 307 * [We can't call it easily with mapping off since the kernel is not 308 * mapped with PA == VA, hence we would have to relocate every address 309 * from the linked base (virtual) address "KERNBASE" to the actual 310 * (physical) address starting relative to 0] 311 */ 312void 313pmap_bootstrap(firstaddr, loadaddr) 314 vm_paddr_t firstaddr; 315 vm_paddr_t loadaddr; 316{ 317 vm_offset_t va; 318 pt_entry_t *pte, *unused; 319 struct sysmaps *sysmaps; 320 int i; 321 322 /* 323 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 324 * large. It should instead be correctly calculated in locore.s and 325 * not based on 'first' (which is a physical address, not a virtual 326 * address, for the start of unused physical memory). The kernel 327 * page tables are NOT double mapped and thus should not be included 328 * in this calculation. 329 */ 330 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 331 virtual_avail = pmap_kmem_choose(virtual_avail); 332 333 virtual_end = VM_MAX_KERNEL_ADDRESS; 334 335 /* 336 * Initialize the kernel pmap (which is statically allocated). 337 */ 338 PMAP_LOCK_INIT(kernel_pmap); 339 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 340#ifdef PAE 341 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 342#endif 343 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 344 TAILQ_INIT(&kernel_pmap->pm_pvlist); 345 LIST_INIT(&allpmaps); 346 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 347 mtx_lock_spin(&allpmaps_lock); 348 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 349 mtx_unlock_spin(&allpmaps_lock); 350 nkpt = NKPT; 351 352 /* 353 * Reserve some special page table entries/VA space for temporary 354 * mapping of pages. 355 */ 356#define SYSMAP(c, p, v, n) \ 357 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 358 359 va = virtual_avail; 360 pte = vtopte(va); 361 362 /* 363 * CMAP1/CMAP2 are used for zeroing and copying pages. 364 * CMAP3 is used for the idle process page zeroing. 365 */ 366 for (i = 0; i < MAXCPU; i++) { 367 sysmaps = &sysmaps_pcpu[i]; 368 mtx_init(&sysmaps->lock, "SYSMAPS", NULL, MTX_DEF); 369 SYSMAP(caddr_t, sysmaps->CMAP1, sysmaps->CADDR1, 1) 370 SYSMAP(caddr_t, sysmaps->CMAP2, sysmaps->CADDR2, 1) 371 } 372 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 373 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 374 *CMAP3 = 0; 375 376 /* 377 * Crashdump maps. 378 */ 379 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS) 380 381 /* 382 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 383 */ 384 SYSMAP(caddr_t, unused, ptvmmap, 1) 385 386 /* 387 * msgbufp is used to map the system message buffer. 388 */ 389 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE))) 390 391 /* 392 * ptemap is used for pmap_pte_quick 393 */ 394 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 395 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1); 396 397 mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF); 398 399 virtual_avail = va; 400 401 *CMAP1 = 0; 402 403#ifdef XBOX 404 /* FIXME: This is gross, but needed for the XBOX. Since we are in such 405 * an early stadium, we cannot yet neatly map video memory ... :-( 406 * Better fixes are very welcome! */ 407 if (!arch_i386_is_xbox) 408#endif 409 for (i = 0; i < NKPT; i++) 410 PTD[i] = 0; 411 412 /* Turn on PG_G on kernel page(s) */ 413 pmap_set_pg(); 414} 415 416/* 417 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 418 */ 419void 420pmap_set_pg(void) 421{ 422 pd_entry_t pdir; 423 pt_entry_t *pte; 424 vm_offset_t va, endva; 425 int i; 426 427 if (pgeflag == 0) 428 return; 429 430 i = KERNLOAD/NBPDR; 431 endva = KERNBASE + KERNend; 432 433 if (pseflag) { 434 va = KERNBASE + KERNLOAD; 435 while (va < endva) { 436 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 437 pdir |= pgeflag; 438 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 439 invltlb(); /* Play it safe, invltlb() every time */ 440 i++; 441 va += NBPDR; 442 } 443 } else { 444 va = (vm_offset_t)btext; 445 while (va < endva) { 446 pte = vtopte(va); 447 if (*pte) 448 *pte |= pgeflag; 449 invltlb(); /* Play it safe, invltlb() every time */ 450 va += PAGE_SIZE; 451 } 452 } 453} 454 455/* 456 * Initialize a vm_page's machine-dependent fields. 457 */ 458void 459pmap_page_init(vm_page_t m) 460{ 461 462 TAILQ_INIT(&m->md.pv_list); 463 m->md.pv_list_count = 0; 464} 465 466#ifdef PAE 467 468static MALLOC_DEFINE(M_PMAPPDPT, "pmap", "pmap pdpt"); 469 470static void * 471pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 472{ 473 *flags = UMA_SLAB_PRIV; 474 return (contigmalloc(PAGE_SIZE, M_PMAPPDPT, 0, 0x0ULL, 0xffffffffULL, 475 1, 0)); 476} 477#endif 478 479/* 480 * Initialize the pmap module. 481 * Called by vm_init, to initialize any structures that the pmap 482 * system needs to map virtual memory. 483 */ 484void 485pmap_init(void) 486{ 487 int shpgperproc = PMAP_SHPGPERPROC; 488 489 /* 490 * Initialize the address space (zone) for the pv entries. Set a 491 * high water mark so that the system can recover from excessive 492 * numbers of pv entries. 493 */ 494 pvzone = uma_zcreate("PV ENTRY", sizeof(struct pv_entry), NULL, NULL, 495 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 496 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 497 pv_entry_max = shpgperproc * maxproc + cnt.v_page_count; 498 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 499 pv_entry_high_water = 9 * (pv_entry_max / 10); 500 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 501 502#ifdef PAE 503 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 504 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 505 UMA_ZONE_VM | UMA_ZONE_NOFREE); 506 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 507#endif 508} 509 510 511/*************************************************** 512 * Low level helper routines..... 513 ***************************************************/ 514 515 516/* 517 * this routine defines the region(s) of memory that should 518 * not be tested for the modified bit. 519 */ 520static PMAP_INLINE int 521pmap_track_modified(vm_offset_t va) 522{ 523 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 524 return 1; 525 else 526 return 0; 527} 528 529#ifdef SMP 530/* 531 * For SMP, these functions have to use the IPI mechanism for coherence. 532 */ 533void 534pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 535{ 536 u_int cpumask; 537 u_int other_cpus; 538 539 if (smp_started) { 540 if (!(read_eflags() & PSL_I)) 541 panic("%s: interrupts disabled", __func__); 542 mtx_lock_spin(&smp_ipi_mtx); 543 } else 544 critical_enter(); 545 /* 546 * We need to disable interrupt preemption but MUST NOT have 547 * interrupts disabled here. 548 * XXX we may need to hold schedlock to get a coherent pm_active 549 * XXX critical sections disable interrupts again 550 */ 551 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 552 invlpg(va); 553 smp_invlpg(va); 554 } else { 555 cpumask = PCPU_GET(cpumask); 556 other_cpus = PCPU_GET(other_cpus); 557 if (pmap->pm_active & cpumask) 558 invlpg(va); 559 if (pmap->pm_active & other_cpus) 560 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 561 } 562 if (smp_started) 563 mtx_unlock_spin(&smp_ipi_mtx); 564 else 565 critical_exit(); 566} 567 568void 569pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 570{ 571 u_int cpumask; 572 u_int other_cpus; 573 vm_offset_t addr; 574 575 if (smp_started) { 576 if (!(read_eflags() & PSL_I)) 577 panic("%s: interrupts disabled", __func__); 578 mtx_lock_spin(&smp_ipi_mtx); 579 } else 580 critical_enter(); 581 /* 582 * We need to disable interrupt preemption but MUST NOT have 583 * interrupts disabled here. 584 * XXX we may need to hold schedlock to get a coherent pm_active 585 * XXX critical sections disable interrupts again 586 */ 587 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 588 for (addr = sva; addr < eva; addr += PAGE_SIZE) 589 invlpg(addr); 590 smp_invlpg_range(sva, eva); 591 } else { 592 cpumask = PCPU_GET(cpumask); 593 other_cpus = PCPU_GET(other_cpus); 594 if (pmap->pm_active & cpumask) 595 for (addr = sva; addr < eva; addr += PAGE_SIZE) 596 invlpg(addr); 597 if (pmap->pm_active & other_cpus) 598 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 599 sva, eva); 600 } 601 if (smp_started) 602 mtx_unlock_spin(&smp_ipi_mtx); 603 else 604 critical_exit(); 605} 606 607void 608pmap_invalidate_all(pmap_t pmap) 609{ 610 u_int cpumask; 611 u_int other_cpus; 612 613 if (smp_started) { 614 if (!(read_eflags() & PSL_I)) 615 panic("%s: interrupts disabled", __func__); 616 mtx_lock_spin(&smp_ipi_mtx); 617 } else 618 critical_enter(); 619 /* 620 * We need to disable interrupt preemption but MUST NOT have 621 * interrupts disabled here. 622 * XXX we may need to hold schedlock to get a coherent pm_active 623 * XXX critical sections disable interrupts again 624 */ 625 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 626 invltlb(); 627 smp_invltlb(); 628 } else { 629 cpumask = PCPU_GET(cpumask); 630 other_cpus = PCPU_GET(other_cpus); 631 if (pmap->pm_active & cpumask) 632 invltlb(); 633 if (pmap->pm_active & other_cpus) 634 smp_masked_invltlb(pmap->pm_active & other_cpus); 635 } 636 if (smp_started) 637 mtx_unlock_spin(&smp_ipi_mtx); 638 else 639 critical_exit(); 640} 641#else /* !SMP */ 642/* 643 * Normal, non-SMP, 486+ invalidation functions. 644 * We inline these within pmap.c for speed. 645 */ 646PMAP_INLINE void 647pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 648{ 649 650 if (pmap == kernel_pmap || pmap->pm_active) 651 invlpg(va); 652} 653 654PMAP_INLINE void 655pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 656{ 657 vm_offset_t addr; 658 659 if (pmap == kernel_pmap || pmap->pm_active) 660 for (addr = sva; addr < eva; addr += PAGE_SIZE) 661 invlpg(addr); 662} 663 664PMAP_INLINE void 665pmap_invalidate_all(pmap_t pmap) 666{ 667 668 if (pmap == kernel_pmap || pmap->pm_active) 669 invltlb(); 670} 671#endif /* !SMP */ 672 673/* 674 * Are we current address space or kernel? N.B. We return FALSE when 675 * a pmap's page table is in use because a kernel thread is borrowing 676 * it. The borrowed page table can change spontaneously, making any 677 * dependence on its continued use subject to a race condition. 678 */ 679static __inline int 680pmap_is_current(pmap_t pmap) 681{ 682 683 return (pmap == kernel_pmap || 684 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 685 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 686} 687 688/* 689 * If the given pmap is not the current or kernel pmap, the returned pte must 690 * be released by passing it to pmap_pte_release(). 691 */ 692pt_entry_t * 693pmap_pte(pmap_t pmap, vm_offset_t va) 694{ 695 pd_entry_t newpf; 696 pd_entry_t *pde; 697 698 pde = pmap_pde(pmap, va); 699 if (*pde & PG_PS) 700 return (pde); 701 if (*pde != 0) { 702 /* are we current address space or kernel? */ 703 if (pmap_is_current(pmap)) 704 return (vtopte(va)); 705 mtx_lock(&PMAP2mutex); 706 newpf = *pde & PG_FRAME; 707 if ((*PMAP2 & PG_FRAME) != newpf) { 708 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; 709 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 710 } 711 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); 712 } 713 return (0); 714} 715 716/* 717 * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte 718 * being NULL. 719 */ 720static __inline void 721pmap_pte_release(pt_entry_t *pte) 722{ 723 724 if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2) 725 mtx_unlock(&PMAP2mutex); 726} 727 728static __inline void 729invlcaddr(void *caddr) 730{ 731 732 invlpg((u_int)caddr); 733} 734 735/* 736 * Super fast pmap_pte routine best used when scanning 737 * the pv lists. This eliminates many coarse-grained 738 * invltlb calls. Note that many of the pv list 739 * scans are across different pmaps. It is very wasteful 740 * to do an entire invltlb for checking a single mapping. 741 * 742 * If the given pmap is not the current pmap, vm_page_queue_mtx 743 * must be held and curthread pinned to a CPU. 744 */ 745static pt_entry_t * 746pmap_pte_quick(pmap_t pmap, vm_offset_t va) 747{ 748 pd_entry_t newpf; 749 pd_entry_t *pde; 750 751 pde = pmap_pde(pmap, va); 752 if (*pde & PG_PS) 753 return (pde); 754 if (*pde != 0) { 755 /* are we current address space or kernel? */ 756 if (pmap_is_current(pmap)) 757 return (vtopte(va)); 758 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 759 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 760 newpf = *pde & PG_FRAME; 761 if ((*PMAP1 & PG_FRAME) != newpf) { 762 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; 763#ifdef SMP 764 PMAP1cpu = PCPU_GET(cpuid); 765#endif 766 invlcaddr(PADDR1); 767 PMAP1changed++; 768 } else 769#ifdef SMP 770 if (PMAP1cpu != PCPU_GET(cpuid)) { 771 PMAP1cpu = PCPU_GET(cpuid); 772 invlcaddr(PADDR1); 773 PMAP1changedcpu++; 774 } else 775#endif 776 PMAP1unchanged++; 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_t pmap, vm_offset_t va) 790{ 791 vm_paddr_t rtval; 792 pt_entry_t *pte; 793 pd_entry_t pde; 794 795 rtval = 0; 796 PMAP_LOCK(pmap); 797 pde = pmap->pm_pdir[va >> PDRSHIFT]; 798 if (pde != 0) { 799 if ((pde & PG_PS) != 0) { 800 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 801 PMAP_UNLOCK(pmap); 802 return rtval; 803 } 804 pte = pmap_pte(pmap, va); 805 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK); 806 pmap_pte_release(pte); 807 } 808 PMAP_UNLOCK(pmap); 809 return (rtval); 810} 811 812/* 813 * Routine: pmap_extract_and_hold 814 * Function: 815 * Atomically extract and hold the physical page 816 * with the given pmap and virtual address pair 817 * if that mapping permits the given protection. 818 */ 819vm_page_t 820pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 821{ 822 pd_entry_t pde; 823 pt_entry_t pte; 824 vm_page_t m; 825 826 m = NULL; 827 vm_page_lock_queues(); 828 PMAP_LOCK(pmap); 829 pde = *pmap_pde(pmap, va); 830 if (pde != 0) { 831 if (pde & PG_PS) { 832 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) { 833 m = PHYS_TO_VM_PAGE((pde & ~PDRMASK) | 834 (va & PDRMASK)); 835 vm_page_hold(m); 836 } 837 } else { 838 sched_pin(); 839 pte = *pmap_pte_quick(pmap, va); 840 if (pte != 0 && 841 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) { 842 m = PHYS_TO_VM_PAGE(pte & PG_FRAME); 843 vm_page_hold(m); 844 } 845 sched_unpin(); 846 } 847 } 848 vm_page_unlock_queues(); 849 PMAP_UNLOCK(pmap); 850 return (m); 851} 852 853/*************************************************** 854 * Low level mapping routines..... 855 ***************************************************/ 856 857/* 858 * Add a wired page to the kva. 859 * Note: not SMP coherent. 860 */ 861PMAP_INLINE void 862pmap_kenter(vm_offset_t va, vm_paddr_t pa) 863{ 864 pt_entry_t *pte; 865 866 pte = vtopte(va); 867 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 868} 869 870/* 871 * Remove a page from the kernel pagetables. 872 * Note: not SMP coherent. 873 */ 874PMAP_INLINE void 875pmap_kremove(vm_offset_t va) 876{ 877 pt_entry_t *pte; 878 879 pte = vtopte(va); 880 pte_clear(pte); 881} 882 883/* 884 * Used to map a range of physical addresses into kernel 885 * virtual address space. 886 * 887 * The value passed in '*virt' is a suggested virtual address for 888 * the mapping. Architectures which can support a direct-mapped 889 * physical to virtual region can return the appropriate address 890 * within that region, leaving '*virt' unchanged. Other 891 * architectures should map the pages starting at '*virt' and 892 * update '*virt' with the first usable address after the mapped 893 * region. 894 */ 895vm_offset_t 896pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 897{ 898 vm_offset_t va, sva; 899 900 va = sva = *virt; 901 while (start < end) { 902 pmap_kenter(va, start); 903 va += PAGE_SIZE; 904 start += PAGE_SIZE; 905 } 906 pmap_invalidate_range(kernel_pmap, sva, va); 907 *virt = va; 908 return (sva); 909} 910 911 912/* 913 * Add a list of wired pages to the kva 914 * this routine is only used for temporary 915 * kernel mappings that do not need to have 916 * page modification or references recorded. 917 * Note that old mappings are simply written 918 * over. The page *must* be wired. 919 * Note: SMP coherent. Uses a ranged shootdown IPI. 920 */ 921void 922pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 923{ 924 vm_offset_t va; 925 926 va = sva; 927 while (count-- > 0) { 928 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 929 va += PAGE_SIZE; 930 m++; 931 } 932 pmap_invalidate_range(kernel_pmap, sva, va); 933} 934 935/* 936 * This routine tears out page mappings from the 937 * kernel -- it is meant only for temporary mappings. 938 * Note: SMP coherent. Uses a ranged shootdown IPI. 939 */ 940void 941pmap_qremove(vm_offset_t sva, int count) 942{ 943 vm_offset_t va; 944 945 va = sva; 946 while (count-- > 0) { 947 pmap_kremove(va); 948 va += PAGE_SIZE; 949 } 950 pmap_invalidate_range(kernel_pmap, sva, va); 951} 952 953/*************************************************** 954 * Page table page management routines..... 955 ***************************************************/ 956 957/* 958 * This routine unholds page table pages, and if the hold count 959 * drops to zero, then it decrements the wire count. 960 */ 961static PMAP_INLINE int 962pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 963{ 964 965 --m->wire_count; 966 if (m->wire_count == 0) 967 return _pmap_unwire_pte_hold(pmap, m); 968 else 969 return 0; 970} 971 972static int 973_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 974{ 975 vm_offset_t pteva; 976 977 /* 978 * unmap the page table page 979 */ 980 pmap->pm_pdir[m->pindex] = 0; 981 --pmap->pm_stats.resident_count; 982 983 /* 984 * Do an invltlb to make the invalidated mapping 985 * take effect immediately. 986 */ 987 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 988 pmap_invalidate_page(pmap, pteva); 989 990 vm_page_free_zero(m); 991 atomic_subtract_int(&cnt.v_wire_count, 1); 992 return 1; 993} 994 995/* 996 * After removing a page table entry, this routine is used to 997 * conditionally free the page, and manage the hold/wire counts. 998 */ 999static int 1000pmap_unuse_pt(pmap_t pmap, vm_offset_t va) 1001{ 1002 pd_entry_t ptepde; 1003 vm_page_t mpte; 1004 1005 if (va >= VM_MAXUSER_ADDRESS) 1006 return 0; 1007 ptepde = *pmap_pde(pmap, va); 1008 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); 1009 return pmap_unwire_pte_hold(pmap, mpte); 1010} 1011 1012void 1013pmap_pinit0(pmap) 1014 struct pmap *pmap; 1015{ 1016 1017 PMAP_LOCK_INIT(pmap); 1018 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1019#ifdef PAE 1020 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1021#endif 1022 pmap->pm_active = 0; 1023 PCPU_SET(curpmap, pmap); 1024 TAILQ_INIT(&pmap->pm_pvlist); 1025 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1026 mtx_lock_spin(&allpmaps_lock); 1027 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1028 mtx_unlock_spin(&allpmaps_lock); 1029} 1030 1031/* 1032 * Initialize a preallocated and zeroed pmap structure, 1033 * such as one in a vmspace structure. 1034 */ 1035void 1036pmap_pinit(pmap) 1037 register struct pmap *pmap; 1038{ 1039 vm_page_t m, ptdpg[NPGPTD]; 1040 vm_paddr_t pa; 1041 static int color; 1042 int i; 1043 1044 PMAP_LOCK_INIT(pmap); 1045 1046 /* 1047 * No need to allocate page table space yet but we do need a valid 1048 * page directory table. 1049 */ 1050 if (pmap->pm_pdir == NULL) { 1051 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1052 NBPTD); 1053#ifdef PAE 1054 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1055 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1056 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1057 ("pmap_pinit: pdpt misaligned")); 1058 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1059 ("pmap_pinit: pdpt above 4g")); 1060#endif 1061 } 1062 1063 /* 1064 * allocate the page directory page(s) 1065 */ 1066 for (i = 0; i < NPGPTD;) { 1067 m = vm_page_alloc(NULL, color++, 1068 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1069 VM_ALLOC_ZERO); 1070 if (m == NULL) 1071 VM_WAIT; 1072 else { 1073 ptdpg[i++] = m; 1074 } 1075 } 1076 1077 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1078 1079 for (i = 0; i < NPGPTD; i++) { 1080 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1081 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1082 } 1083 1084 mtx_lock_spin(&allpmaps_lock); 1085 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1086 mtx_unlock_spin(&allpmaps_lock); 1087 /* Wire in kernel global address entries. */ 1088 /* XXX copies current process, does not fill in MPPTDI */ 1089 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1090#ifdef SMP 1091 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1092#endif 1093 1094 /* install self-referential address mapping entry(s) */ 1095 for (i = 0; i < NPGPTD; i++) { 1096 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1097 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1098#ifdef PAE 1099 pmap->pm_pdpt[i] = pa | PG_V; 1100#endif 1101 } 1102 1103 pmap->pm_active = 0; 1104 TAILQ_INIT(&pmap->pm_pvlist); 1105 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1106} 1107 1108/* 1109 * this routine is called if the page table page is not 1110 * mapped correctly. 1111 */ 1112static vm_page_t 1113_pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags) 1114{ 1115 vm_paddr_t ptepa; 1116 vm_page_t m; 1117 1118 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT || 1119 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK, 1120 ("_pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK")); 1121 1122 /* 1123 * Allocate a page table page. 1124 */ 1125 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1126 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1127 if (flags & M_WAITOK) { 1128 PMAP_UNLOCK(pmap); 1129 vm_page_unlock_queues(); 1130 VM_WAIT; 1131 vm_page_lock_queues(); 1132 PMAP_LOCK(pmap); 1133 } 1134 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 /* 1145 * Map the pagetable page into the process address space, if 1146 * it isn't already there. 1147 */ 1148 1149 pmap->pm_stats.resident_count++; 1150 1151 ptepa = VM_PAGE_TO_PHYS(m); 1152 pmap->pm_pdir[ptepindex] = 1153 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1154 1155 return m; 1156} 1157 1158static vm_page_t 1159pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags) 1160{ 1161 unsigned ptepindex; 1162 pd_entry_t ptepa; 1163 vm_page_t m; 1164 1165 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT || 1166 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK, 1167 ("pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK")); 1168 1169 /* 1170 * Calculate pagetable page index 1171 */ 1172 ptepindex = va >> PDRSHIFT; 1173retry: 1174 /* 1175 * Get the page directory entry 1176 */ 1177 ptepa = pmap->pm_pdir[ptepindex]; 1178 1179 /* 1180 * This supports switching from a 4MB page to a 1181 * normal 4K page. 1182 */ 1183 if (ptepa & PG_PS) { 1184 pmap->pm_pdir[ptepindex] = 0; 1185 ptepa = 0; 1186 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1187 pmap_invalidate_all(kernel_pmap); 1188 } 1189 1190 /* 1191 * If the page table page is mapped, we just increment the 1192 * hold count, and activate it. 1193 */ 1194 if (ptepa) { 1195 m = PHYS_TO_VM_PAGE(ptepa); 1196 m->wire_count++; 1197 } else { 1198 /* 1199 * Here if the pte page isn't mapped, or if it has 1200 * been deallocated. 1201 */ 1202 m = _pmap_allocpte(pmap, ptepindex, flags); 1203 if (m == NULL && (flags & M_WAITOK)) 1204 goto retry; 1205 } 1206 return (m); 1207} 1208 1209 1210/*************************************************** 1211* Pmap allocation/deallocation routines. 1212 ***************************************************/ 1213 1214#ifdef SMP 1215/* 1216 * Deal with a SMP shootdown of other users of the pmap that we are 1217 * trying to dispose of. This can be a bit hairy. 1218 */ 1219static u_int *lazymask; 1220static u_int lazyptd; 1221static volatile u_int lazywait; 1222 1223void pmap_lazyfix_action(void); 1224 1225void 1226pmap_lazyfix_action(void) 1227{ 1228 u_int mymask = PCPU_GET(cpumask); 1229 1230#ifdef COUNT_IPIS 1231 *ipi_lazypmap_counts[PCPU_GET(cpuid)]++; 1232#endif 1233 if (rcr3() == lazyptd) 1234 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1235 atomic_clear_int(lazymask, mymask); 1236 atomic_store_rel_int(&lazywait, 1); 1237} 1238 1239static void 1240pmap_lazyfix_self(u_int mymask) 1241{ 1242 1243 if (rcr3() == lazyptd) 1244 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1245 atomic_clear_int(lazymask, mymask); 1246} 1247 1248 1249static void 1250pmap_lazyfix(pmap_t pmap) 1251{ 1252 u_int mymask; 1253 u_int mask; 1254 register u_int spins; 1255 1256 while ((mask = pmap->pm_active) != 0) { 1257 spins = 50000000; 1258 mask = mask & -mask; /* Find least significant set bit */ 1259 mtx_lock_spin(&smp_ipi_mtx); 1260#ifdef PAE 1261 lazyptd = vtophys(pmap->pm_pdpt); 1262#else 1263 lazyptd = vtophys(pmap->pm_pdir); 1264#endif 1265 mymask = PCPU_GET(cpumask); 1266 if (mask == mymask) { 1267 lazymask = &pmap->pm_active; 1268 pmap_lazyfix_self(mymask); 1269 } else { 1270 atomic_store_rel_int((u_int *)&lazymask, 1271 (u_int)&pmap->pm_active); 1272 atomic_store_rel_int(&lazywait, 0); 1273 ipi_selected(mask, IPI_LAZYPMAP); 1274 while (lazywait == 0) { 1275 ia32_pause(); 1276 if (--spins == 0) 1277 break; 1278 } 1279 } 1280 mtx_unlock_spin(&smp_ipi_mtx); 1281 if (spins == 0) 1282 printf("pmap_lazyfix: spun for 50000000\n"); 1283 } 1284} 1285 1286#else /* SMP */ 1287 1288/* 1289 * Cleaning up on uniprocessor is easy. For various reasons, we're 1290 * unlikely to have to even execute this code, including the fact 1291 * that the cleanup is deferred until the parent does a wait(2), which 1292 * means that another userland process has run. 1293 */ 1294static void 1295pmap_lazyfix(pmap_t pmap) 1296{ 1297 u_int cr3; 1298 1299 cr3 = vtophys(pmap->pm_pdir); 1300 if (cr3 == rcr3()) { 1301 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1302 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1303 } 1304} 1305#endif /* SMP */ 1306 1307/* 1308 * Release any resources held by the given physical map. 1309 * Called when a pmap initialized by pmap_pinit is being released. 1310 * Should only be called if the map contains no valid mappings. 1311 */ 1312void 1313pmap_release(pmap_t pmap) 1314{ 1315 vm_page_t m, ptdpg[NPGPTD]; 1316 int i; 1317 1318 KASSERT(pmap->pm_stats.resident_count == 0, 1319 ("pmap_release: pmap resident count %ld != 0", 1320 pmap->pm_stats.resident_count)); 1321 1322 pmap_lazyfix(pmap); 1323 mtx_lock_spin(&allpmaps_lock); 1324 LIST_REMOVE(pmap, pm_list); 1325 mtx_unlock_spin(&allpmaps_lock); 1326 1327 for (i = 0; i < NPGPTD; i++) 1328 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1329 1330 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1331 sizeof(*pmap->pm_pdir)); 1332#ifdef SMP 1333 pmap->pm_pdir[MPPTDI] = 0; 1334#endif 1335 1336 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1337 1338 vm_page_lock_queues(); 1339 for (i = 0; i < NPGPTD; i++) { 1340 m = ptdpg[i]; 1341#ifdef PAE 1342 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1343 ("pmap_release: got wrong ptd page")); 1344#endif 1345 m->wire_count--; 1346 atomic_subtract_int(&cnt.v_wire_count, 1); 1347 vm_page_free_zero(m); 1348 } 1349 vm_page_unlock_queues(); 1350 PMAP_LOCK_DESTROY(pmap); 1351} 1352 1353static int 1354kvm_size(SYSCTL_HANDLER_ARGS) 1355{ 1356 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1357 1358 return sysctl_handle_long(oidp, &ksize, 0, req); 1359} 1360SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1361 0, 0, kvm_size, "IU", "Size of KVM"); 1362 1363static int 1364kvm_free(SYSCTL_HANDLER_ARGS) 1365{ 1366 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1367 1368 return sysctl_handle_long(oidp, &kfree, 0, req); 1369} 1370SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1371 0, 0, kvm_free, "IU", "Amount of KVM free"); 1372 1373/* 1374 * grow the number of kernel page table entries, if needed 1375 */ 1376void 1377pmap_growkernel(vm_offset_t addr) 1378{ 1379 struct pmap *pmap; 1380 vm_paddr_t ptppaddr; 1381 vm_page_t nkpg; 1382 pd_entry_t newpdir; 1383 pt_entry_t *pde; 1384 1385 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1386 if (kernel_vm_end == 0) { 1387 kernel_vm_end = KERNBASE; 1388 nkpt = 0; 1389 while (pdir_pde(PTD, kernel_vm_end)) { 1390 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1391 nkpt++; 1392 if (kernel_vm_end - 1 >= kernel_map->max_offset) { 1393 kernel_vm_end = kernel_map->max_offset; 1394 break; 1395 } 1396 } 1397 } 1398 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1399 if (addr - 1 >= kernel_map->max_offset) 1400 addr = kernel_map->max_offset; 1401 while (kernel_vm_end < addr) { 1402 if (pdir_pde(PTD, kernel_vm_end)) { 1403 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1404 if (kernel_vm_end - 1 >= kernel_map->max_offset) { 1405 kernel_vm_end = kernel_map->max_offset; 1406 break; 1407 } 1408 continue; 1409 } 1410 1411 /* 1412 * This index is bogus, but out of the way 1413 */ 1414 nkpg = vm_page_alloc(NULL, nkpt, 1415 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1416 if (!nkpg) 1417 panic("pmap_growkernel: no memory to grow kernel"); 1418 1419 nkpt++; 1420 1421 pmap_zero_page(nkpg); 1422 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1423 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1424 pdir_pde(PTD, kernel_vm_end) = newpdir; 1425 1426 mtx_lock_spin(&allpmaps_lock); 1427 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1428 pde = pmap_pde(pmap, kernel_vm_end); 1429 pde_store(pde, newpdir); 1430 } 1431 mtx_unlock_spin(&allpmaps_lock); 1432 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1433 if (kernel_vm_end - 1 >= kernel_map->max_offset) { 1434 kernel_vm_end = kernel_map->max_offset; 1435 break; 1436 } 1437 } 1438} 1439 1440 1441/*************************************************** 1442 * page management routines. 1443 ***************************************************/ 1444 1445/* 1446 * free the pv_entry back to the free list 1447 */ 1448static PMAP_INLINE void 1449free_pv_entry(pv_entry_t pv) 1450{ 1451 pv_entry_count--; 1452 uma_zfree(pvzone, pv); 1453} 1454 1455/* 1456 * get a new pv_entry, allocating a block from the system 1457 * when needed. 1458 */ 1459static pv_entry_t 1460get_pv_entry(pmap_t locked_pmap) 1461{ 1462 static const struct timeval printinterval = { 60, 0 }; 1463 static struct timeval lastprint; 1464 struct vpgqueues *vpq; 1465 pmap_t pmap; 1466 pt_entry_t *pte, tpte; 1467 pv_entry_t allocated_pv, next_pv, pv; 1468 vm_offset_t va; 1469 vm_page_t m; 1470 1471 PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED); 1472 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1473 allocated_pv = uma_zalloc(pvzone, M_NOWAIT); 1474 if (allocated_pv != NULL) { 1475 pv_entry_count++; 1476 if (pv_entry_count > pv_entry_high_water) 1477 pagedaemon_wakeup(); 1478 else 1479 return (allocated_pv); 1480 } 1481 1482 /* 1483 * Reclaim pv entries: At first, destroy mappings to inactive 1484 * pages. After that, if a pv entry is still needed, destroy 1485 * mappings to active pages. 1486 */ 1487 if (ratecheck(&lastprint, &printinterval)) 1488 printf("Approaching the limit on PV entries, " 1489 "increase the vm.pmap.shpgperproc tunable.\n"); 1490 vpq = &vm_page_queues[PQ_INACTIVE]; 1491retry: 1492 sched_pin(); 1493 TAILQ_FOREACH(m, &vpq->pl, pageq) { 1494 if (m->hold_count || m->busy || (m->flags & PG_BUSY)) 1495 continue; 1496 TAILQ_FOREACH_SAFE(pv, &m->md.pv_list, pv_list, next_pv) { 1497 va = pv->pv_va; 1498 pmap = pv->pv_pmap; 1499 /* Avoid deadlock and lock recursion. */ 1500 if (pmap > locked_pmap) 1501 PMAP_LOCK(pmap); 1502 else if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap)) 1503 continue; 1504 pmap->pm_stats.resident_count--; 1505 pte = pmap_pte_quick(pmap, va); 1506 tpte = pte_load_clear(pte); 1507 KASSERT((tpte & PG_W) == 0, 1508 ("get_pv_entry: wired pte %#jx", (uintmax_t)tpte)); 1509 if (tpte & PG_A) 1510 vm_page_flag_set(m, PG_REFERENCED); 1511 if (tpte & PG_M) { 1512 KASSERT((tpte & PG_RW), 1513 ("get_pv_entry: modified page not writable: va: %#x, pte: %#jx", 1514 va, (uintmax_t)tpte)); 1515 if (pmap_track_modified(va)) 1516 vm_page_dirty(m); 1517 } 1518 pmap_invalidate_page(pmap, va); 1519 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1520 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1521 if (TAILQ_EMPTY(&m->md.pv_list)) 1522 vm_page_flag_clear(m, PG_WRITEABLE); 1523 m->md.pv_list_count--; 1524 pmap_unuse_pt(pmap, va); 1525 if (pmap != locked_pmap) 1526 PMAP_UNLOCK(pmap); 1527 if (allocated_pv == NULL) 1528 allocated_pv = pv; 1529 else 1530 free_pv_entry(pv); 1531 } 1532 } 1533 sched_unpin(); 1534 if (allocated_pv == NULL) { 1535 if (vpq == &vm_page_queues[PQ_INACTIVE]) { 1536 vpq = &vm_page_queues[PQ_ACTIVE]; 1537 goto retry; 1538 } 1539 panic("get_pv_entry: increase the vm.pmap.shpgperproc tunable"); 1540 } 1541 return (allocated_pv); 1542} 1543 1544static void 1545pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1546{ 1547 pv_entry_t pv; 1548 1549 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1550 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1551 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1552 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1553 if (pmap == pv->pv_pmap && va == pv->pv_va) 1554 break; 1555 } 1556 } else { 1557 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1558 if (va == pv->pv_va) 1559 break; 1560 } 1561 } 1562 KASSERT(pv != NULL, ("pmap_remove_entry: pv not found")); 1563 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1564 m->md.pv_list_count--; 1565 if (TAILQ_EMPTY(&m->md.pv_list)) 1566 vm_page_flag_clear(m, PG_WRITEABLE); 1567 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1568 free_pv_entry(pv); 1569} 1570 1571/* 1572 * Create a pv entry for page at pa for 1573 * (pmap, va). 1574 */ 1575static void 1576pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) 1577{ 1578 pv_entry_t pv; 1579 1580 pv = get_pv_entry(pmap); 1581 pv->pv_va = va; 1582 pv->pv_pmap = pmap; 1583 1584 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1585 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1586 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1587 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1588 m->md.pv_list_count++; 1589} 1590 1591/* 1592 * Conditionally create a pv entry. 1593 */ 1594static boolean_t 1595pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) 1596{ 1597 pv_entry_t pv; 1598 1599 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1600 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1601 if (pv_entry_count < pv_entry_high_water && 1602 (pv = uma_zalloc(pvzone, M_NOWAIT)) != NULL) { 1603 pv_entry_count++; 1604 pv->pv_va = va; 1605 pv->pv_pmap = pmap; 1606 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1607 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1608 m->md.pv_list_count++; 1609 return (TRUE); 1610 } else 1611 return (FALSE); 1612} 1613 1614/* 1615 * pmap_remove_pte: do the things to unmap a page in a process 1616 */ 1617static int 1618pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1619{ 1620 pt_entry_t oldpte; 1621 vm_page_t m; 1622 1623 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1624 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1625 oldpte = pte_load_clear(ptq); 1626 if (oldpte & PG_W) 1627 pmap->pm_stats.wired_count -= 1; 1628 /* 1629 * Machines that don't support invlpg, also don't support 1630 * PG_G. 1631 */ 1632 if (oldpte & PG_G) 1633 pmap_invalidate_page(kernel_pmap, va); 1634 pmap->pm_stats.resident_count -= 1; 1635 if (oldpte & PG_MANAGED) { 1636 m = PHYS_TO_VM_PAGE(oldpte); 1637 if (oldpte & PG_M) { 1638 KASSERT((oldpte & PG_RW), 1639 ("pmap_remove_pte: modified page not writable: va: %#x, pte: %#jx", 1640 va, (uintmax_t)oldpte)); 1641 if (pmap_track_modified(va)) 1642 vm_page_dirty(m); 1643 } 1644 if (oldpte & PG_A) 1645 vm_page_flag_set(m, PG_REFERENCED); 1646 pmap_remove_entry(pmap, m, va); 1647 } 1648 return (pmap_unuse_pt(pmap, va)); 1649} 1650 1651/* 1652 * Remove a single page from a process address space 1653 */ 1654static void 1655pmap_remove_page(pmap_t pmap, vm_offset_t va) 1656{ 1657 pt_entry_t *pte; 1658 1659 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1660 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 1661 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1662 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) 1663 return; 1664 pmap_remove_pte(pmap, pte, va); 1665 pmap_invalidate_page(pmap, va); 1666} 1667 1668/* 1669 * Remove the given range of addresses from the specified map. 1670 * 1671 * It is assumed that the start and end are properly 1672 * rounded to the page size. 1673 */ 1674void 1675pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1676{ 1677 vm_offset_t pdnxt; 1678 pd_entry_t ptpaddr; 1679 pt_entry_t *pte; 1680 int anyvalid; 1681 1682 /* 1683 * Perform an unsynchronized read. This is, however, safe. 1684 */ 1685 if (pmap->pm_stats.resident_count == 0) 1686 return; 1687 1688 anyvalid = 0; 1689 1690 vm_page_lock_queues(); 1691 sched_pin(); 1692 PMAP_LOCK(pmap); 1693 1694 /* 1695 * special handling of removing one page. a very 1696 * common operation and easy to short circuit some 1697 * code. 1698 */ 1699 if ((sva + PAGE_SIZE == eva) && 1700 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1701 pmap_remove_page(pmap, sva); 1702 goto out; 1703 } 1704 1705 for (; sva < eva; sva = pdnxt) { 1706 unsigned pdirindex; 1707 1708 /* 1709 * Calculate index for next page table. 1710 */ 1711 pdnxt = (sva + NBPDR) & ~PDRMASK; 1712 if (pmap->pm_stats.resident_count == 0) 1713 break; 1714 1715 pdirindex = sva >> PDRSHIFT; 1716 ptpaddr = pmap->pm_pdir[pdirindex]; 1717 1718 /* 1719 * Weed out invalid mappings. Note: we assume that the page 1720 * directory table is always allocated, and in kernel virtual. 1721 */ 1722 if (ptpaddr == 0) 1723 continue; 1724 1725 /* 1726 * Check for large page. 1727 */ 1728 if ((ptpaddr & PG_PS) != 0) { 1729 pmap->pm_pdir[pdirindex] = 0; 1730 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1731 anyvalid = 1; 1732 continue; 1733 } 1734 1735 /* 1736 * Limit our scan to either the end of the va represented 1737 * by the current page table page, or to the end of the 1738 * range being removed. 1739 */ 1740 if (pdnxt > eva) 1741 pdnxt = eva; 1742 1743 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, 1744 sva += PAGE_SIZE) { 1745 if (*pte == 0) 1746 continue; 1747 1748 /* 1749 * The TLB entry for a PG_G mapping is invalidated 1750 * by pmap_remove_pte(). 1751 */ 1752 if ((*pte & PG_G) == 0) 1753 anyvalid = 1; 1754 if (pmap_remove_pte(pmap, pte, sva)) 1755 break; 1756 } 1757 } 1758out: 1759 sched_unpin(); 1760 vm_page_unlock_queues(); 1761 if (anyvalid) 1762 pmap_invalidate_all(pmap); 1763 PMAP_UNLOCK(pmap); 1764} 1765 1766/* 1767 * Routine: pmap_remove_all 1768 * Function: 1769 * Removes this physical page from 1770 * all physical maps in which it resides. 1771 * Reflects back modify bits to the pager. 1772 * 1773 * Notes: 1774 * Original versions of this routine were very 1775 * inefficient because they iteratively called 1776 * pmap_remove (slow...) 1777 */ 1778 1779void 1780pmap_remove_all(vm_page_t m) 1781{ 1782 register pv_entry_t pv; 1783 pt_entry_t *pte, tpte; 1784 1785#if defined(PMAP_DIAGNOSTIC) 1786 /* 1787 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1788 */ 1789 if (m->flags & PG_FICTITIOUS) { 1790 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1791 VM_PAGE_TO_PHYS(m)); 1792 } 1793#endif 1794 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1795 sched_pin(); 1796 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1797 PMAP_LOCK(pv->pv_pmap); 1798 pv->pv_pmap->pm_stats.resident_count--; 1799 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1800 tpte = pte_load_clear(pte); 1801 if (tpte & PG_W) 1802 pv->pv_pmap->pm_stats.wired_count--; 1803 if (tpte & PG_A) 1804 vm_page_flag_set(m, PG_REFERENCED); 1805 1806 /* 1807 * Update the vm_page_t clean and reference bits. 1808 */ 1809 if (tpte & PG_M) { 1810 KASSERT((tpte & PG_RW), 1811 ("pmap_remove_all: modified page not writable: va: %#x, pte: %#jx", 1812 pv->pv_va, (uintmax_t)tpte)); 1813 if (pmap_track_modified(pv->pv_va)) 1814 vm_page_dirty(m); 1815 } 1816 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1817 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1818 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1819 m->md.pv_list_count--; 1820 pmap_unuse_pt(pv->pv_pmap, pv->pv_va); 1821 PMAP_UNLOCK(pv->pv_pmap); 1822 free_pv_entry(pv); 1823 } 1824 vm_page_flag_clear(m, PG_WRITEABLE); 1825 sched_unpin(); 1826} 1827 1828/* 1829 * Set the physical protection on the 1830 * specified range of this map as requested. 1831 */ 1832void 1833pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1834{ 1835 vm_offset_t pdnxt; 1836 pd_entry_t ptpaddr; 1837 pt_entry_t *pte; 1838 int anychanged; 1839 1840 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1841 pmap_remove(pmap, sva, eva); 1842 return; 1843 } 1844 1845 if (prot & VM_PROT_WRITE) 1846 return; 1847 1848 anychanged = 0; 1849 1850 vm_page_lock_queues(); 1851 sched_pin(); 1852 PMAP_LOCK(pmap); 1853 for (; sva < eva; sva = pdnxt) { 1854 unsigned obits, pbits, pdirindex; 1855 1856 pdnxt = (sva + NBPDR) & ~PDRMASK; 1857 1858 pdirindex = sva >> PDRSHIFT; 1859 ptpaddr = pmap->pm_pdir[pdirindex]; 1860 1861 /* 1862 * Weed out invalid mappings. Note: we assume that the page 1863 * directory table is always allocated, and in kernel virtual. 1864 */ 1865 if (ptpaddr == 0) 1866 continue; 1867 1868 /* 1869 * Check for large page. 1870 */ 1871 if ((ptpaddr & PG_PS) != 0) { 1872 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1873 anychanged = 1; 1874 continue; 1875 } 1876 1877 if (pdnxt > eva) 1878 pdnxt = eva; 1879 1880 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, 1881 sva += PAGE_SIZE) { 1882 vm_page_t m; 1883 1884retry: 1885 /* 1886 * Regardless of whether a pte is 32 or 64 bits in 1887 * size, PG_RW, PG_A, and PG_M are among the least 1888 * significant 32 bits. 1889 */ 1890 obits = pbits = *(u_int *)pte; 1891 if (pbits & PG_MANAGED) { 1892 m = NULL; 1893 if (pbits & PG_A) { 1894 m = PHYS_TO_VM_PAGE(*pte); 1895 vm_page_flag_set(m, PG_REFERENCED); 1896 pbits &= ~PG_A; 1897 } 1898 if ((pbits & PG_M) != 0 && 1899 pmap_track_modified(sva)) { 1900 if (m == NULL) 1901 m = PHYS_TO_VM_PAGE(*pte); 1902 vm_page_dirty(m); 1903 } 1904 } 1905 1906 pbits &= ~(PG_RW | PG_M); 1907 1908 if (pbits != obits) { 1909 if (!atomic_cmpset_int((u_int *)pte, obits, 1910 pbits)) 1911 goto retry; 1912 if (obits & PG_G) 1913 pmap_invalidate_page(pmap, sva); 1914 else 1915 anychanged = 1; 1916 } 1917 } 1918 } 1919 sched_unpin(); 1920 vm_page_unlock_queues(); 1921 if (anychanged) 1922 pmap_invalidate_all(pmap); 1923 PMAP_UNLOCK(pmap); 1924} 1925 1926/* 1927 * Insert the given physical page (p) at 1928 * the specified virtual address (v) in the 1929 * target physical map with the protection requested. 1930 * 1931 * If specified, the page will be wired down, meaning 1932 * that the related pte can not be reclaimed. 1933 * 1934 * NB: This is the only routine which MAY NOT lazy-evaluate 1935 * or lose information. That is, this routine must actually 1936 * insert this page into the given map NOW. 1937 */ 1938void 1939pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1940 boolean_t wired) 1941{ 1942 vm_paddr_t pa; 1943 register pt_entry_t *pte; 1944 vm_paddr_t opa; 1945 pt_entry_t origpte, newpte; 1946 vm_page_t mpte, om; 1947 boolean_t invlva; 1948 1949 va &= PG_FRAME; 1950#ifdef PMAP_DIAGNOSTIC 1951 if (va > VM_MAX_KERNEL_ADDRESS) 1952 panic("pmap_enter: toobig"); 1953 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1954 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1955#endif 1956 1957 mpte = NULL; 1958 1959 vm_page_lock_queues(); 1960 PMAP_LOCK(pmap); 1961 sched_pin(); 1962 1963 /* 1964 * In the case that a page table page is not 1965 * resident, we are creating it here. 1966 */ 1967 if (va < VM_MAXUSER_ADDRESS) { 1968 mpte = pmap_allocpte(pmap, va, M_WAITOK); 1969 } 1970#if 0 && defined(PMAP_DIAGNOSTIC) 1971 else { 1972 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1973 origpte = *pdeaddr; 1974 if ((origpte & PG_V) == 0) { 1975 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1976 pmap->pm_pdir[PTDPTDI], origpte, va); 1977 } 1978 } 1979#endif 1980 1981 pte = pmap_pte_quick(pmap, va); 1982 1983 /* 1984 * Page Directory table entry not valid, we need a new PT page 1985 */ 1986 if (pte == NULL) { 1987 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1988 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1989 } 1990 1991 pa = VM_PAGE_TO_PHYS(m); 1992 om = NULL; 1993 origpte = *pte; 1994 opa = origpte & PG_FRAME; 1995 1996 if (origpte & PG_PS) { 1997 /* 1998 * Yes, I know this will truncate upper address bits for PAE, 1999 * but I'm actually more interested in the lower bits 2000 */ 2001 printf("pmap_enter: va %p, pte %p, origpte %p\n", 2002 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 2003 panic("pmap_enter: attempted pmap_enter on 4MB page"); 2004 } 2005 2006 /* 2007 * Mapping has not changed, must be protection or wiring change. 2008 */ 2009 if (origpte && (opa == pa)) { 2010 /* 2011 * Wiring change, just update stats. We don't worry about 2012 * wiring PT pages as they remain resident as long as there 2013 * are valid mappings in them. Hence, if a user page is wired, 2014 * the PT page will be also. 2015 */ 2016 if (wired && ((origpte & PG_W) == 0)) 2017 pmap->pm_stats.wired_count++; 2018 else if (!wired && (origpte & PG_W)) 2019 pmap->pm_stats.wired_count--; 2020 2021 /* 2022 * Remove extra pte reference 2023 */ 2024 if (mpte) 2025 mpte->wire_count--; 2026 2027 /* 2028 * We might be turning off write access to the page, 2029 * so we go ahead and sense modify status. 2030 */ 2031 if (origpte & PG_MANAGED) { 2032 om = m; 2033 pa |= PG_MANAGED; 2034 } 2035 goto validate; 2036 } 2037 /* 2038 * Mapping has changed, invalidate old range and fall through to 2039 * handle validating new mapping. 2040 */ 2041 if (opa) { 2042 if (origpte & PG_W) 2043 pmap->pm_stats.wired_count--; 2044 if (origpte & PG_MANAGED) { 2045 om = PHYS_TO_VM_PAGE(opa); 2046 pmap_remove_entry(pmap, om, va); 2047 } 2048 if (mpte != NULL) { 2049 mpte->wire_count--; 2050 KASSERT(mpte->wire_count > 0, 2051 ("pmap_enter: missing reference to page table page," 2052 " va: 0x%x", va)); 2053 } 2054 } else 2055 pmap->pm_stats.resident_count++; 2056 2057 /* 2058 * Enter on the PV list if part of our managed memory. 2059 */ 2060 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0) { 2061 pmap_insert_entry(pmap, va, m); 2062 pa |= PG_MANAGED; 2063 } 2064 2065 /* 2066 * Increment counters 2067 */ 2068 if (wired) 2069 pmap->pm_stats.wired_count++; 2070 2071validate: 2072 /* 2073 * Now validate mapping with desired protection/wiring. 2074 */ 2075 newpte = (pt_entry_t)(pa | PG_V); 2076 if ((prot & VM_PROT_WRITE) != 0) 2077 newpte |= PG_RW; 2078 if (wired) 2079 newpte |= PG_W; 2080 if (va < VM_MAXUSER_ADDRESS) 2081 newpte |= PG_U; 2082 if (pmap == kernel_pmap) 2083 newpte |= pgeflag; 2084 2085 /* 2086 * if the mapping or permission bits are different, we need 2087 * to update the pte. 2088 */ 2089 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2090 if (origpte & PG_V) { 2091 invlva = FALSE; 2092 origpte = pte_load_store(pte, newpte | PG_A); 2093 if (origpte & PG_A) { 2094 if (origpte & PG_MANAGED) 2095 vm_page_flag_set(om, PG_REFERENCED); 2096 if (opa != VM_PAGE_TO_PHYS(m)) 2097 invlva = TRUE; 2098 } 2099 if (origpte & PG_M) { 2100 KASSERT((origpte & PG_RW), 2101 ("pmap_enter: modified page not writable: va: %#x, pte: %#jx", 2102 va, (uintmax_t)origpte)); 2103 if ((origpte & PG_MANAGED) && 2104 pmap_track_modified(va)) 2105 vm_page_dirty(om); 2106 if ((prot & VM_PROT_WRITE) == 0) 2107 invlva = TRUE; 2108 } 2109 if (invlva) 2110 pmap_invalidate_page(pmap, va); 2111 } else 2112 pte_store(pte, newpte | PG_A); 2113 } 2114 sched_unpin(); 2115 vm_page_unlock_queues(); 2116 PMAP_UNLOCK(pmap); 2117} 2118 2119/* 2120 * this code makes some *MAJOR* assumptions: 2121 * 1. Current pmap & pmap exists. 2122 * 2. Not wired. 2123 * 3. Read access. 2124 * 4. No page table pages. 2125 * but is *MUCH* faster than pmap_enter... 2126 */ 2127 2128vm_page_t 2129pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 2130 vm_page_t mpte) 2131{ 2132 pt_entry_t *pte; 2133 vm_paddr_t pa; 2134 2135 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2136 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2137 PMAP_LOCK(pmap); 2138 2139 /* 2140 * In the case that a page table page is not 2141 * resident, we are creating it here. 2142 */ 2143 if (va < VM_MAXUSER_ADDRESS) { 2144 unsigned ptepindex; 2145 pd_entry_t ptepa; 2146 2147 /* 2148 * Calculate pagetable page index 2149 */ 2150 ptepindex = va >> PDRSHIFT; 2151 if (mpte && (mpte->pindex == ptepindex)) { 2152 mpte->wire_count++; 2153 } else { 2154retry: 2155 /* 2156 * Get the page directory entry 2157 */ 2158 ptepa = pmap->pm_pdir[ptepindex]; 2159 2160 /* 2161 * If the page table page is mapped, we just increment 2162 * the hold count, and activate it. 2163 */ 2164 if (ptepa) { 2165 if (ptepa & PG_PS) 2166 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2167 mpte = PHYS_TO_VM_PAGE(ptepa); 2168 mpte->wire_count++; 2169 } else { 2170 mpte = _pmap_allocpte(pmap, ptepindex, 2171 M_NOWAIT); 2172 if (mpte == NULL) { 2173 PMAP_UNLOCK(pmap); 2174 vm_page_busy(m); 2175 vm_page_unlock_queues(); 2176 VM_OBJECT_UNLOCK(m->object); 2177 VM_WAIT; 2178 VM_OBJECT_LOCK(m->object); 2179 vm_page_lock_queues(); 2180 vm_page_wakeup(m); 2181 PMAP_LOCK(pmap); 2182 goto retry; 2183 } 2184 } 2185 } 2186 } else { 2187 mpte = NULL; 2188 } 2189 2190 /* 2191 * This call to vtopte makes the assumption that we are 2192 * entering the page into the current pmap. In order to support 2193 * quick entry into any pmap, one would likely use pmap_pte_quick. 2194 * But that isn't as quick as vtopte. 2195 */ 2196 pte = vtopte(va); 2197 if (*pte) { 2198 if (mpte != NULL) { 2199 pmap_unwire_pte_hold(pmap, mpte); 2200 mpte = NULL; 2201 } 2202 goto out; 2203 } 2204 2205 /* 2206 * Enter on the PV list if part of our managed memory. Note that we 2207 * raise IPL while manipulating pv_table since pmap_enter can be 2208 * called at interrupt time. 2209 */ 2210 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2211 pmap_insert_entry(pmap, va, m); 2212 2213 /* 2214 * Increment counters 2215 */ 2216 pmap->pm_stats.resident_count++; 2217 2218 pa = VM_PAGE_TO_PHYS(m); 2219 2220 /* 2221 * Now validate mapping with RO protection 2222 */ 2223 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2224 pte_store(pte, pa | PG_V | PG_U); 2225 else 2226 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2227out: 2228 PMAP_UNLOCK(pmap); 2229 return mpte; 2230} 2231 2232/* 2233 * Make a temporary mapping for a physical address. This is only intended 2234 * to be used for panic dumps. 2235 */ 2236void * 2237pmap_kenter_temporary(vm_paddr_t pa, int i) 2238{ 2239 vm_offset_t va; 2240 2241 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2242 pmap_kenter(va, pa); 2243 invlpg(va); 2244 return ((void *)crashdumpmap); 2245} 2246 2247/* 2248 * This code maps large physical mmap regions into the 2249 * processor address space. Note that some shortcuts 2250 * are taken, but the code works. 2251 */ 2252void 2253pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2254 vm_object_t object, vm_pindex_t pindex, 2255 vm_size_t size) 2256{ 2257 vm_page_t p; 2258 2259 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2260 KASSERT(object->type == OBJT_DEVICE, 2261 ("pmap_object_init_pt: non-device object")); 2262 if (pseflag && 2263 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2264 int i; 2265 vm_page_t m[1]; 2266 unsigned int ptepindex; 2267 int npdes; 2268 pd_entry_t ptepa; 2269 2270 PMAP_LOCK(pmap); 2271 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2272 goto out; 2273 PMAP_UNLOCK(pmap); 2274retry: 2275 p = vm_page_lookup(object, pindex); 2276 if (p != NULL) { 2277 vm_page_lock_queues(); 2278 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2279 goto retry; 2280 } else { 2281 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2282 if (p == NULL) 2283 return; 2284 m[0] = p; 2285 2286 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2287 vm_page_lock_queues(); 2288 vm_page_free(p); 2289 vm_page_unlock_queues(); 2290 return; 2291 } 2292 2293 p = vm_page_lookup(object, pindex); 2294 vm_page_lock_queues(); 2295 vm_page_wakeup(p); 2296 } 2297 vm_page_unlock_queues(); 2298 2299 ptepa = VM_PAGE_TO_PHYS(p); 2300 if (ptepa & (NBPDR - 1)) 2301 return; 2302 2303 p->valid = VM_PAGE_BITS_ALL; 2304 2305 PMAP_LOCK(pmap); 2306 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2307 npdes = size >> PDRSHIFT; 2308 for(i = 0; i < npdes; i++) { 2309 pde_store(&pmap->pm_pdir[ptepindex], 2310 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2311 ptepa += NBPDR; 2312 ptepindex += 1; 2313 } 2314 pmap_invalidate_all(pmap); 2315out: 2316 PMAP_UNLOCK(pmap); 2317 } 2318} 2319 2320/* 2321 * Routine: pmap_change_wiring 2322 * Function: Change the wiring attribute for a map/virtual-address 2323 * pair. 2324 * In/out conditions: 2325 * The mapping must already exist in the pmap. 2326 */ 2327void 2328pmap_change_wiring(pmap, va, wired) 2329 register pmap_t pmap; 2330 vm_offset_t va; 2331 boolean_t wired; 2332{ 2333 register pt_entry_t *pte; 2334 2335 PMAP_LOCK(pmap); 2336 pte = pmap_pte(pmap, va); 2337 2338 if (wired && !pmap_pte_w(pte)) 2339 pmap->pm_stats.wired_count++; 2340 else if (!wired && pmap_pte_w(pte)) 2341 pmap->pm_stats.wired_count--; 2342 2343 /* 2344 * Wiring is not a hardware characteristic so there is no need to 2345 * invalidate TLB. 2346 */ 2347 pmap_pte_set_w(pte, wired); 2348 pmap_pte_release(pte); 2349 PMAP_UNLOCK(pmap); 2350} 2351 2352 2353 2354/* 2355 * Copy the range specified by src_addr/len 2356 * from the source map to the range dst_addr/len 2357 * in the destination map. 2358 * 2359 * This routine is only advisory and need not do anything. 2360 */ 2361 2362void 2363pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2364 vm_offset_t src_addr) 2365{ 2366 vm_offset_t addr; 2367 vm_offset_t end_addr = src_addr + len; 2368 vm_offset_t pdnxt; 2369 2370 if (dst_addr != src_addr) 2371 return; 2372 2373 if (!pmap_is_current(src_pmap)) 2374 return; 2375 2376 vm_page_lock_queues(); 2377 if (dst_pmap < src_pmap) { 2378 PMAP_LOCK(dst_pmap); 2379 PMAP_LOCK(src_pmap); 2380 } else { 2381 PMAP_LOCK(src_pmap); 2382 PMAP_LOCK(dst_pmap); 2383 } 2384 sched_pin(); 2385 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2386 pt_entry_t *src_pte, *dst_pte; 2387 vm_page_t dstmpte, srcmpte; 2388 pd_entry_t srcptepaddr; 2389 unsigned ptepindex; 2390 2391 if (addr >= UPT_MIN_ADDRESS) 2392 panic("pmap_copy: invalid to pmap_copy page tables"); 2393 2394 pdnxt = (addr + NBPDR) & ~PDRMASK; 2395 ptepindex = addr >> PDRSHIFT; 2396 2397 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2398 if (srcptepaddr == 0) 2399 continue; 2400 2401 if (srcptepaddr & PG_PS) { 2402 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2403 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2404 dst_pmap->pm_stats.resident_count += 2405 NBPDR / PAGE_SIZE; 2406 } 2407 continue; 2408 } 2409 2410 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2411 if (srcmpte->wire_count == 0) 2412 panic("pmap_copy: source page table page is unused"); 2413 2414 if (pdnxt > end_addr) 2415 pdnxt = end_addr; 2416 2417 src_pte = vtopte(addr); 2418 while (addr < pdnxt) { 2419 pt_entry_t ptetemp; 2420 ptetemp = *src_pte; 2421 /* 2422 * we only virtual copy managed pages 2423 */ 2424 if ((ptetemp & PG_MANAGED) != 0) { 2425 /* 2426 * We have to check after allocpte for the 2427 * pte still being around... allocpte can 2428 * block. 2429 */ 2430 dstmpte = pmap_allocpte(dst_pmap, addr, 2431 M_NOWAIT); 2432 if (dstmpte == NULL) 2433 break; 2434 dst_pte = pmap_pte_quick(dst_pmap, addr); 2435 if (*dst_pte == 0 && 2436 pmap_try_insert_pv_entry(dst_pmap, addr, 2437 PHYS_TO_VM_PAGE(ptetemp & PG_FRAME))) { 2438 /* 2439 * Clear the modified and 2440 * accessed (referenced) bits 2441 * during the copy. 2442 */ 2443 *dst_pte = ptetemp & ~(PG_M | PG_A); 2444 dst_pmap->pm_stats.resident_count++; 2445 } else 2446 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2447 if (dstmpte->wire_count >= srcmpte->wire_count) 2448 break; 2449 } 2450 addr += PAGE_SIZE; 2451 src_pte++; 2452 } 2453 } 2454 sched_unpin(); 2455 vm_page_unlock_queues(); 2456 PMAP_UNLOCK(src_pmap); 2457 PMAP_UNLOCK(dst_pmap); 2458} 2459 2460static __inline void 2461pagezero(void *page) 2462{ 2463#if defined(I686_CPU) 2464 if (cpu_class == CPUCLASS_686) { 2465#if defined(CPU_ENABLE_SSE) 2466 if (cpu_feature & CPUID_SSE2) 2467 sse2_pagezero(page); 2468 else 2469#endif 2470 i686_pagezero(page); 2471 } else 2472#endif 2473 bzero(page, PAGE_SIZE); 2474} 2475 2476/* 2477 * pmap_zero_page zeros the specified hardware page by mapping 2478 * the page into KVM and using bzero to clear its contents. 2479 */ 2480void 2481pmap_zero_page(vm_page_t m) 2482{ 2483 struct sysmaps *sysmaps; 2484 2485 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2486 mtx_lock(&sysmaps->lock); 2487 if (*sysmaps->CMAP2) 2488 panic("pmap_zero_page: CMAP2 busy"); 2489 sched_pin(); 2490 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2491 invlcaddr(sysmaps->CADDR2); 2492 pagezero(sysmaps->CADDR2); 2493 *sysmaps->CMAP2 = 0; 2494 sched_unpin(); 2495 mtx_unlock(&sysmaps->lock); 2496} 2497 2498/* 2499 * pmap_zero_page_area zeros the specified hardware page by mapping 2500 * the page into KVM and using bzero to clear its contents. 2501 * 2502 * off and size may not cover an area beyond a single hardware page. 2503 */ 2504void 2505pmap_zero_page_area(vm_page_t m, int off, int size) 2506{ 2507 struct sysmaps *sysmaps; 2508 2509 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2510 mtx_lock(&sysmaps->lock); 2511 if (*sysmaps->CMAP2) 2512 panic("pmap_zero_page: CMAP2 busy"); 2513 sched_pin(); 2514 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2515 invlcaddr(sysmaps->CADDR2); 2516 if (off == 0 && size == PAGE_SIZE) 2517 pagezero(sysmaps->CADDR2); 2518 else 2519 bzero((char *)sysmaps->CADDR2 + off, size); 2520 *sysmaps->CMAP2 = 0; 2521 sched_unpin(); 2522 mtx_unlock(&sysmaps->lock); 2523} 2524 2525/* 2526 * pmap_zero_page_idle zeros the specified hardware page by mapping 2527 * the page into KVM and using bzero to clear its contents. This 2528 * is intended to be called from the vm_pagezero process only and 2529 * outside of Giant. 2530 */ 2531void 2532pmap_zero_page_idle(vm_page_t m) 2533{ 2534 2535 if (*CMAP3) 2536 panic("pmap_zero_page: CMAP3 busy"); 2537 sched_pin(); 2538 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2539 invlcaddr(CADDR3); 2540 pagezero(CADDR3); 2541 *CMAP3 = 0; 2542 sched_unpin(); 2543} 2544 2545/* 2546 * pmap_copy_page copies the specified (machine independent) 2547 * page by mapping the page into virtual memory and using 2548 * bcopy to copy the page, one machine dependent page at a 2549 * time. 2550 */ 2551void 2552pmap_copy_page(vm_page_t src, vm_page_t dst) 2553{ 2554 struct sysmaps *sysmaps; 2555 2556 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2557 mtx_lock(&sysmaps->lock); 2558 if (*sysmaps->CMAP1) 2559 panic("pmap_copy_page: CMAP1 busy"); 2560 if (*sysmaps->CMAP2) 2561 panic("pmap_copy_page: CMAP2 busy"); 2562 sched_pin(); 2563 invlpg((u_int)sysmaps->CADDR1); 2564 invlpg((u_int)sysmaps->CADDR2); 2565 *sysmaps->CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2566 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2567 bcopy(sysmaps->CADDR1, sysmaps->CADDR2, PAGE_SIZE); 2568 *sysmaps->CMAP1 = 0; 2569 *sysmaps->CMAP2 = 0; 2570 sched_unpin(); 2571 mtx_unlock(&sysmaps->lock); 2572} 2573 2574/* 2575 * Returns true if the pmap's pv is one of the first 2576 * 16 pvs linked to from this page. This count may 2577 * be changed upwards or downwards in the future; it 2578 * is only necessary that true be returned for a small 2579 * subset of pmaps for proper page aging. 2580 */ 2581boolean_t 2582pmap_page_exists_quick(pmap, m) 2583 pmap_t pmap; 2584 vm_page_t m; 2585{ 2586 pv_entry_t pv; 2587 int loops = 0; 2588 2589 if (m->flags & PG_FICTITIOUS) 2590 return FALSE; 2591 2592 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2593 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2594 if (pv->pv_pmap == pmap) { 2595 return TRUE; 2596 } 2597 loops++; 2598 if (loops >= 16) 2599 break; 2600 } 2601 return (FALSE); 2602} 2603 2604#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2605/* 2606 * Remove all pages from specified address space 2607 * this aids process exit speeds. Also, this code 2608 * is special cased for current process only, but 2609 * can have the more generic (and slightly slower) 2610 * mode enabled. This is much faster than pmap_remove 2611 * in the case of running down an entire address space. 2612 */ 2613void 2614pmap_remove_pages(pmap, sva, eva) 2615 pmap_t pmap; 2616 vm_offset_t sva, eva; 2617{ 2618 pt_entry_t *pte, tpte; 2619 vm_page_t m; 2620 pv_entry_t pv, npv; 2621 2622#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2623 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) { 2624 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2625 return; 2626 } 2627#endif 2628 vm_page_lock_queues(); 2629 PMAP_LOCK(pmap); 2630 sched_pin(); 2631 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2632 2633 if (pv->pv_va >= eva || pv->pv_va < sva) { 2634 npv = TAILQ_NEXT(pv, pv_plist); 2635 continue; 2636 } 2637 2638#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2639 pte = vtopte(pv->pv_va); 2640#else 2641 pte = pmap_pte_quick(pmap, pv->pv_va); 2642#endif 2643 tpte = *pte; 2644 2645 if (tpte == 0) { 2646 printf("TPTE at %p IS ZERO @ VA %08x\n", 2647 pte, pv->pv_va); 2648 panic("bad pte"); 2649 } 2650 2651/* 2652 * We cannot remove wired pages from a process' mapping at this time 2653 */ 2654 if (tpte & PG_W) { 2655 npv = TAILQ_NEXT(pv, pv_plist); 2656 continue; 2657 } 2658 2659 m = PHYS_TO_VM_PAGE(tpte); 2660 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2661 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2662 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2663 2664 KASSERT(m < &vm_page_array[vm_page_array_size], 2665 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2666 2667 pmap->pm_stats.resident_count--; 2668 2669 pte_clear(pte); 2670 2671 /* 2672 * Update the vm_page_t clean and reference bits. 2673 */ 2674 if (tpte & PG_M) { 2675 vm_page_dirty(m); 2676 } 2677 2678 npv = TAILQ_NEXT(pv, pv_plist); 2679 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 2680 2681 m->md.pv_list_count--; 2682 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2683 if (TAILQ_EMPTY(&m->md.pv_list)) 2684 vm_page_flag_clear(m, PG_WRITEABLE); 2685 2686 pmap_unuse_pt(pmap, pv->pv_va); 2687 free_pv_entry(pv); 2688 } 2689 sched_unpin(); 2690 pmap_invalidate_all(pmap); 2691 PMAP_UNLOCK(pmap); 2692 vm_page_unlock_queues(); 2693} 2694 2695/* 2696 * pmap_is_modified: 2697 * 2698 * Return whether or not the specified physical page was modified 2699 * in any physical maps. 2700 */ 2701boolean_t 2702pmap_is_modified(vm_page_t m) 2703{ 2704 pv_entry_t pv; 2705 pt_entry_t *pte; 2706 boolean_t rv; 2707 2708 rv = FALSE; 2709 if (m->flags & PG_FICTITIOUS) 2710 return (rv); 2711 2712 sched_pin(); 2713 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2714 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2715 /* 2716 * if the bit being tested is the modified bit, then 2717 * mark clean_map and ptes as never 2718 * modified. 2719 */ 2720 if (!pmap_track_modified(pv->pv_va)) 2721 continue; 2722 PMAP_LOCK(pv->pv_pmap); 2723 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2724 rv = (*pte & PG_M) != 0; 2725 PMAP_UNLOCK(pv->pv_pmap); 2726 if (rv) 2727 break; 2728 } 2729 sched_unpin(); 2730 return (rv); 2731} 2732 2733/* 2734 * pmap_is_prefaultable: 2735 * 2736 * Return whether or not the specified virtual address is elgible 2737 * for prefault. 2738 */ 2739boolean_t 2740pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2741{ 2742 pt_entry_t *pte; 2743 boolean_t rv; 2744 2745 rv = FALSE; 2746 PMAP_LOCK(pmap); 2747 if (*pmap_pde(pmap, addr)) { 2748 pte = vtopte(addr); 2749 rv = *pte == 0; 2750 } 2751 PMAP_UNLOCK(pmap); 2752 return (rv); 2753} 2754 2755/* 2756 * Clear the given bit in each of the given page's ptes. The bit is 2757 * expressed as a 32-bit mask. Consequently, if the pte is 64 bits in 2758 * size, only a bit within the least significant 32 can be cleared. 2759 */ 2760static __inline void 2761pmap_clear_ptes(vm_page_t m, int bit) 2762{ 2763 register pv_entry_t pv; 2764 pt_entry_t pbits, *pte; 2765 2766 if ((m->flags & PG_FICTITIOUS) || 2767 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2768 return; 2769 2770 sched_pin(); 2771 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2772 /* 2773 * Loop over all current mappings setting/clearing as appropos If 2774 * setting RO do we need to clear the VAC? 2775 */ 2776 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2777 /* 2778 * don't write protect pager mappings 2779 */ 2780 if (bit == PG_RW) { 2781 if (!pmap_track_modified(pv->pv_va)) 2782 continue; 2783 } 2784 2785 PMAP_LOCK(pv->pv_pmap); 2786 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2787retry: 2788 pbits = *pte; 2789 if (pbits & bit) { 2790 if (bit == PG_RW) { 2791 /* 2792 * Regardless of whether a pte is 32 or 64 bits 2793 * in size, PG_RW and PG_M are among the least 2794 * significant 32 bits. 2795 */ 2796 if (!atomic_cmpset_int((u_int *)pte, pbits, 2797 pbits & ~(PG_RW | PG_M))) 2798 goto retry; 2799 if (pbits & PG_M) { 2800 vm_page_dirty(m); 2801 } 2802 } else { 2803 atomic_clear_int((u_int *)pte, bit); 2804 } 2805 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2806 } 2807 PMAP_UNLOCK(pv->pv_pmap); 2808 } 2809 if (bit == PG_RW) 2810 vm_page_flag_clear(m, PG_WRITEABLE); 2811 sched_unpin(); 2812} 2813 2814/* 2815 * pmap_page_protect: 2816 * 2817 * Lower the permission for all mappings to a given page. 2818 */ 2819void 2820pmap_page_protect(vm_page_t m, vm_prot_t prot) 2821{ 2822 if ((prot & VM_PROT_WRITE) == 0) { 2823 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2824 pmap_clear_ptes(m, PG_RW); 2825 } else { 2826 pmap_remove_all(m); 2827 } 2828 } 2829} 2830 2831/* 2832 * pmap_ts_referenced: 2833 * 2834 * Return a count of reference bits for a page, clearing those bits. 2835 * It is not necessary for every reference bit to be cleared, but it 2836 * is necessary that 0 only be returned when there are truly no 2837 * reference bits set. 2838 * 2839 * XXX: The exact number of bits to check and clear is a matter that 2840 * should be tested and standardized at some point in the future for 2841 * optimal aging of shared pages. 2842 */ 2843int 2844pmap_ts_referenced(vm_page_t m) 2845{ 2846 register pv_entry_t pv, pvf, pvn; 2847 pt_entry_t *pte; 2848 pt_entry_t v; 2849 int rtval = 0; 2850 2851 if (m->flags & PG_FICTITIOUS) 2852 return (rtval); 2853 2854 sched_pin(); 2855 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2856 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2857 2858 pvf = pv; 2859 2860 do { 2861 pvn = TAILQ_NEXT(pv, pv_list); 2862 2863 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2864 2865 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2866 2867 if (!pmap_track_modified(pv->pv_va)) 2868 continue; 2869 2870 PMAP_LOCK(pv->pv_pmap); 2871 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2872 2873 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2874 atomic_clear_int((u_int *)pte, PG_A); 2875 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2876 2877 rtval++; 2878 if (rtval > 4) { 2879 PMAP_UNLOCK(pv->pv_pmap); 2880 break; 2881 } 2882 } 2883 PMAP_UNLOCK(pv->pv_pmap); 2884 } while ((pv = pvn) != NULL && pv != pvf); 2885 } 2886 sched_unpin(); 2887 2888 return (rtval); 2889} 2890 2891/* 2892 * Clear the modify bits on the specified physical page. 2893 */ 2894void 2895pmap_clear_modify(vm_page_t m) 2896{ 2897 pmap_clear_ptes(m, PG_M); 2898} 2899 2900/* 2901 * pmap_clear_reference: 2902 * 2903 * Clear the reference bit on the specified physical page. 2904 */ 2905void 2906pmap_clear_reference(vm_page_t m) 2907{ 2908 pmap_clear_ptes(m, PG_A); 2909} 2910 2911/* 2912 * Miscellaneous support routines follow 2913 */ 2914 2915/* 2916 * Map a set of physical memory pages into the kernel virtual 2917 * address space. Return a pointer to where it is mapped. This 2918 * routine is intended to be used for mapping device memory, 2919 * NOT real memory. 2920 */ 2921void * 2922pmap_mapdev(pa, size) 2923 vm_paddr_t pa; 2924 vm_size_t size; 2925{ 2926 vm_offset_t va, tmpva, offset; 2927 2928 offset = pa & PAGE_MASK; 2929 size = roundup(offset + size, PAGE_SIZE); 2930 pa = pa & PG_FRAME; 2931 2932 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2933 va = KERNBASE + pa; 2934 else 2935 va = kmem_alloc_nofault(kernel_map, size); 2936 if (!va) 2937 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2938 2939 for (tmpva = va; size > 0; ) { 2940 pmap_kenter(tmpva, pa); 2941 size -= PAGE_SIZE; 2942 tmpva += PAGE_SIZE; 2943 pa += PAGE_SIZE; 2944 } 2945 pmap_invalidate_range(kernel_pmap, va, tmpva); 2946 return ((void *)(va + offset)); 2947} 2948 2949void 2950pmap_unmapdev(va, size) 2951 vm_offset_t va; 2952 vm_size_t size; 2953{ 2954 vm_offset_t base, offset, tmpva; 2955 2956 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2957 return; 2958 base = va & PG_FRAME; 2959 offset = va & PAGE_MASK; 2960 size = roundup(offset + size, PAGE_SIZE); 2961 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2962 pmap_kremove(tmpva); 2963 pmap_invalidate_range(kernel_pmap, va, tmpva); 2964 kmem_free(kernel_map, base, size); 2965} 2966 2967/* 2968 * perform the pmap work for mincore 2969 */ 2970int 2971pmap_mincore(pmap, addr) 2972 pmap_t pmap; 2973 vm_offset_t addr; 2974{ 2975 pt_entry_t *ptep, pte; 2976 vm_page_t m; 2977 int val = 0; 2978 2979 PMAP_LOCK(pmap); 2980 ptep = pmap_pte(pmap, addr); 2981 pte = (ptep != NULL) ? *ptep : 0; 2982 pmap_pte_release(ptep); 2983 PMAP_UNLOCK(pmap); 2984 2985 if (pte != 0) { 2986 vm_paddr_t pa; 2987 2988 val = MINCORE_INCORE; 2989 if ((pte & PG_MANAGED) == 0) 2990 return val; 2991 2992 pa = pte & PG_FRAME; 2993 2994 m = PHYS_TO_VM_PAGE(pa); 2995 2996 /* 2997 * Modified by us 2998 */ 2999 if (pte & PG_M) 3000 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 3001 else { 3002 /* 3003 * Modified by someone else 3004 */ 3005 vm_page_lock_queues(); 3006 if (m->dirty || pmap_is_modified(m)) 3007 val |= MINCORE_MODIFIED_OTHER; 3008 vm_page_unlock_queues(); 3009 } 3010 /* 3011 * Referenced by us 3012 */ 3013 if (pte & PG_A) 3014 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 3015 else { 3016 /* 3017 * Referenced by someone else 3018 */ 3019 vm_page_lock_queues(); 3020 if ((m->flags & PG_REFERENCED) || 3021 pmap_ts_referenced(m)) { 3022 val |= MINCORE_REFERENCED_OTHER; 3023 vm_page_flag_set(m, PG_REFERENCED); 3024 } 3025 vm_page_unlock_queues(); 3026 } 3027 } 3028 return val; 3029} 3030 3031void 3032pmap_activate(struct thread *td) 3033{ 3034 pmap_t pmap, oldpmap; 3035 u_int32_t cr3; 3036 3037 critical_enter(); 3038 pmap = vmspace_pmap(td->td_proc->p_vmspace); 3039 oldpmap = PCPU_GET(curpmap); 3040#if defined(SMP) 3041 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 3042 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 3043#else 3044 oldpmap->pm_active &= ~1; 3045 pmap->pm_active |= 1; 3046#endif 3047#ifdef PAE 3048 cr3 = vtophys(pmap->pm_pdpt); 3049#else 3050 cr3 = vtophys(pmap->pm_pdir); 3051#endif 3052 /* 3053 * pmap_activate is for the current thread on the current cpu 3054 */ 3055 td->td_pcb->pcb_cr3 = cr3; 3056 load_cr3(cr3); 3057 PCPU_SET(curpmap, pmap); 3058 critical_exit(); 3059} 3060 3061vm_offset_t 3062pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3063{ 3064 3065 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3066 return addr; 3067 } 3068 3069 addr = (addr + PDRMASK) & ~PDRMASK; 3070 return addr; 3071} 3072 3073 3074#if defined(PMAP_DEBUG) 3075pmap_pid_dump(int pid) 3076{ 3077 pmap_t pmap; 3078 struct proc *p; 3079 int npte = 0; 3080 int index; 3081 3082 sx_slock(&allproc_lock); 3083 LIST_FOREACH(p, &allproc, p_list) { 3084 if (p->p_pid != pid) 3085 continue; 3086 3087 if (p->p_vmspace) { 3088 int i,j; 3089 index = 0; 3090 pmap = vmspace_pmap(p->p_vmspace); 3091 for (i = 0; i < NPDEPTD; i++) { 3092 pd_entry_t *pde; 3093 pt_entry_t *pte; 3094 vm_offset_t base = i << PDRSHIFT; 3095 3096 pde = &pmap->pm_pdir[i]; 3097 if (pde && pmap_pde_v(pde)) { 3098 for (j = 0; j < NPTEPG; j++) { 3099 vm_offset_t va = base + (j << PAGE_SHIFT); 3100 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3101 if (index) { 3102 index = 0; 3103 printf("\n"); 3104 } 3105 sx_sunlock(&allproc_lock); 3106 return npte; 3107 } 3108 pte = pmap_pte(pmap, va); 3109 if (pte && pmap_pte_v(pte)) { 3110 pt_entry_t pa; 3111 vm_page_t m; 3112 pa = *pte; 3113 m = PHYS_TO_VM_PAGE(pa); 3114 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3115 va, pa, m->hold_count, m->wire_count, m->flags); 3116 npte++; 3117 index++; 3118 if (index >= 2) { 3119 index = 0; 3120 printf("\n"); 3121 } else { 3122 printf(" "); 3123 } 3124 } 3125 } 3126 } 3127 } 3128 } 3129 } 3130 sx_sunlock(&allproc_lock); 3131 return npte; 3132} 3133#endif 3134 3135#if defined(DEBUG) 3136 3137static void pads(pmap_t pm); 3138void pmap_pvdump(vm_offset_t pa); 3139 3140/* print address space of pmap*/ 3141static void 3142pads(pm) 3143 pmap_t pm; 3144{ 3145 int i, j; 3146 vm_paddr_t va; 3147 pt_entry_t *ptep; 3148 3149 if (pm == kernel_pmap) 3150 return; 3151 for (i = 0; i < NPDEPTD; i++) 3152 if (pm->pm_pdir[i]) 3153 for (j = 0; j < NPTEPG; j++) { 3154 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3155 if (pm == kernel_pmap && va < KERNBASE) 3156 continue; 3157 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3158 continue; 3159 ptep = pmap_pte(pm, va); 3160 if (pmap_pte_v(ptep)) 3161 printf("%x:%x ", va, *ptep); 3162 }; 3163 3164} 3165 3166void 3167pmap_pvdump(pa) 3168 vm_paddr_t pa; 3169{ 3170 pv_entry_t pv; 3171 vm_page_t m; 3172 3173 printf("pa %x", pa); 3174 m = PHYS_TO_VM_PAGE(pa); 3175 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3176 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3177 pads(pv->pv_pmap); 3178 } 3179 printf(" "); 3180} 3181#endif 3182