pmap.c revision 152219
12116Sjkh/*- 22116Sjkh * Copyright (c) 1991 Regents of the University of California. 32116Sjkh * All rights reserved. 42116Sjkh * Copyright (c) 1994 John S. Dyson 52116Sjkh * All rights reserved. 62116Sjkh * Copyright (c) 1994 David Greenman 72116Sjkh * All rights reserved. 82116Sjkh * Copyright (c) 2005 Alan L. Cox <alc@cs.rice.edu> 92116Sjkh * All rights reserved. 102116Sjkh * 118870Srgrimes * This code is derived from software contributed to Berkeley by 122116Sjkh * the Systems Programming Group of the University of Utah Computer 132116Sjkh * Science Department and William Jolitz of UUNET Technologies Inc. 142116Sjkh * 152116Sjkh * Redistribution and use in source and binary forms, with or without 162116Sjkh * modification, are permitted provided that the following conditions 1750476Speter * are met: 182116Sjkh * 1. Redistributions of source code must retain the above copyright 192116Sjkh * notice, this list of conditions and the following disclaimer. 202116Sjkh * 2. Redistributions in binary form must reproduce the above copyright 212116Sjkh * notice, this list of conditions and the following disclaimer in the 222116Sjkh * documentation and/or other materials provided with the distribution. 232116Sjkh * 3. All advertising materials mentioning features or use of this software 242116Sjkh * must display the following acknowledgement: 252116Sjkh * This product includes software developed by the University of 262116Sjkh * California, Berkeley and its contributors. 272116Sjkh * 4. Neither the name of the University nor the names of its contributors 282116Sjkh * may be used to endorse or promote products derived from this software 29152947Sbde * without specific prior written permission. 30152947Sbde * 31152947Sbde * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 32152947Sbde * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 332116Sjkh * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 34176032Sbde * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 35176032Sbde * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 36176032Sbde * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 37176032Sbde * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 38176032Sbde * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 39176032Sbde * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 402116Sjkh * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 41175468Sdas * SUCH DAMAGE. 42175468Sdas * 4397407Salfred * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 44142369Sdas */ 452116Sjkh/*- 46142369Sdas * Copyright (c) 2003 Networks Associates Technology, Inc. 4717141Sjkh * All rights reserved. 482116Sjkh * 492116Sjkh * This software was developed for the FreeBSD Project by Jake Burkholder, 502116Sjkh * Safeport Network Services, and Network Associates Laboratories, the 512116Sjkh * Security Research Division of Network Associates, Inc. under 52142369Sdas * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA 532116Sjkh * CHATS research program. 542116Sjkh * 552116Sjkh * Redistribution and use in source and binary forms, with or without 562116Sjkh * modification, are permitted provided that the following conditions 572116Sjkh * are met: 582116Sjkh * 1. Redistributions of source code must retain the above copyright 592116Sjkh * notice, this list of conditions and the following disclaimer. 602116Sjkh * 2. Redistributions in binary form must reproduce the above copyright 612116Sjkh * notice, this list of conditions and the following disclaimer in the 622116Sjkh * documentation and/or other materials provided with the distribution. 632116Sjkh * 642116Sjkh * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 658870Srgrimes * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 662116Sjkh * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 67142369Sdas * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 682116Sjkh * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 692116Sjkh * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 702116Sjkh * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 71142369Sdas * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 72142369Sdas * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 732116Sjkh * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 74142369Sdas * SUCH DAMAGE. 758870Srgrimes */ 762116Sjkh 772116Sjkh#include <sys/cdefs.h> 782116Sjkh__FBSDID("$FreeBSD: head/sys/i386/i386/pmap.c 152219 2005-11-09 03:55:40Z imp $"); 792116Sjkh 802116Sjkh/* 812116Sjkh * Manages physical address maps. 822116Sjkh * 83176032Sbde * In addition to hardware address maps, this 84142369Sdas * module is called upon to provide software-use-only 85142369Sdas * maps which may or may not be stored in the same 862116Sjkh * form as hardware maps. These pseudo-maps are 872116Sjkh * used to store intermediate results from copy 882116Sjkh * operations to and from address spaces. 892116Sjkh * 902116Sjkh * Since the information managed by this module is 912116Sjkh * also stored by the logical address mapping module, 922116Sjkh * this module may throw away valid virtual-to-physical 932116Sjkh * mappings at almost any time. However, invalidations 942116Sjkh * of virtual-to-physical mappings must be done as 952116Sjkh * requested. 962116Sjkh * 972116Sjkh * 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_xbox.h" 110 111#include <sys/param.h> 112#include <sys/systm.h> 113#include <sys/kernel.h> 114#include <sys/lock.h> 115#include <sys/malloc.h> 116#include <sys/mman.h> 117#include <sys/msgbuf.h> 118#include <sys/mutex.h> 119#include <sys/proc.h> 120#include <sys/sx.h> 121#include <sys/vmmeter.h> 122#include <sys/sched.h> 123#include <sys/sysctl.h> 124#ifdef SMP 125#include <sys/smp.h> 126#endif 127 128#include <vm/vm.h> 129#include <vm/vm_param.h> 130#include <vm/vm_kern.h> 131#include <vm/vm_page.h> 132#include <vm/vm_map.h> 133#include <vm/vm_object.h> 134#include <vm/vm_extern.h> 135#include <vm/vm_pageout.h> 136#include <vm/vm_pager.h> 137#include <vm/uma.h> 138 139#include <machine/cpu.h> 140#include <machine/cputypes.h> 141#include <machine/md_var.h> 142#include <machine/pcb.h> 143#include <machine/specialreg.h> 144#ifdef SMP 145#include <machine/smp.h> 146#endif 147 148#ifdef XBOX 149#include <machine/xbox.h> 150#endif 151 152#if !defined(CPU_DISABLE_SSE) && defined(I686_CPU) 153#define CPU_ENABLE_SSE 154#endif 155 156#ifndef PMAP_SHPGPERPROC 157#define PMAP_SHPGPERPROC 200 158#endif 159 160#if defined(DIAGNOSTIC) 161#define PMAP_DIAGNOSTIC 162#endif 163 164#if !defined(PMAP_DIAGNOSTIC) 165#define PMAP_INLINE __inline 166#else 167#define PMAP_INLINE 168#endif 169 170/* 171 * Get PDEs and PTEs for user/kernel address space 172 */ 173#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 174#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 175 176#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 177#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 178#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 179#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 180#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 181 182#define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \ 183 atomic_clear_int((u_int *)(pte), PG_W)) 184#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 185 186struct pmap kernel_pmap_store; 187LIST_HEAD(pmaplist, pmap); 188static struct pmaplist allpmaps; 189static struct mtx allpmaps_lock; 190 191vm_paddr_t avail_end; /* PA of last available physical page */ 192vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 193vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 194int pgeflag = 0; /* PG_G or-in */ 195int pseflag = 0; /* PG_PS or-in */ 196 197static int nkpt; 198vm_offset_t kernel_vm_end; 199extern u_int32_t KERNend; 200 201#ifdef PAE 202static uma_zone_t pdptzone; 203#endif 204 205/* 206 * Data for the pv entry allocation mechanism 207 */ 208static uma_zone_t pvzone; 209static struct vm_object pvzone_obj; 210static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 211int pmap_pagedaemon_waken; 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(void); 259static pv_entry_t pv_entry_reclaim(pmap_t locked_pmap); 260static void pmap_clear_ptes(vm_page_t m, int bit); 261 262static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 263static void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 264static void pmap_remove_entry(struct pmap *pmap, vm_page_t m, 265 vm_offset_t va); 266static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); 267 268static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags); 269 270static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags); 271static int _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m); 272static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 273static void pmap_pte_release(pt_entry_t *pte); 274static int pmap_unuse_pt(pmap_t, vm_offset_t); 275static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 276#ifdef PAE 277static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 278#endif 279 280CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 281CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 282 283/* 284 * Move the kernel virtual free pointer to the next 285 * 4MB. This is used to help improve performance 286 * by using a large (4MB) page for much of the kernel 287 * (.text, .data, .bss) 288 */ 289static vm_offset_t 290pmap_kmem_choose(vm_offset_t addr) 291{ 292 vm_offset_t newaddr = addr; 293 294#ifndef DISABLE_PSE 295 if (cpu_feature & CPUID_PSE) 296 newaddr = (addr + PDRMASK) & ~PDRMASK; 297#endif 298 return newaddr; 299} 300 301/* 302 * Bootstrap the system enough to run with virtual memory. 303 * 304 * On the i386 this is called after mapping has already been enabled 305 * and just syncs the pmap module with what has already been done. 306 * [We can't call it easily with mapping off since the kernel is not 307 * mapped with PA == VA, hence we would have to relocate every address 308 * from the linked base (virtual) address "KERNBASE" to the actual 309 * (physical) address starting relative to 0] 310 */ 311void 312pmap_bootstrap(firstaddr, loadaddr) 313 vm_paddr_t firstaddr; 314 vm_paddr_t loadaddr; 315{ 316 vm_offset_t va; 317 pt_entry_t *pte, *unused; 318 struct sysmaps *sysmaps; 319 int i; 320 321 /* 322 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 323 * large. It should instead be correctly calculated in locore.s and 324 * not based on 'first' (which is a physical address, not a virtual 325 * address, for the start of unused physical memory). The kernel 326 * page tables are NOT double mapped and thus should not be included 327 * in this calculation. 328 */ 329 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 330 virtual_avail = pmap_kmem_choose(virtual_avail); 331 332 virtual_end = VM_MAX_KERNEL_ADDRESS; 333 334 /* 335 * Initialize the kernel pmap (which is statically allocated). 336 */ 337 PMAP_LOCK_INIT(kernel_pmap); 338 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 339#ifdef PAE 340 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 341#endif 342 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 343 TAILQ_INIT(&kernel_pmap->pm_pvlist); 344 LIST_INIT(&allpmaps); 345 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 346 mtx_lock_spin(&allpmaps_lock); 347 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 348 mtx_unlock_spin(&allpmaps_lock); 349 nkpt = NKPT; 350 351 /* 352 * Reserve some special page table entries/VA space for temporary 353 * mapping of pages. 354 */ 355#define SYSMAP(c, p, v, n) \ 356 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 357 358 va = virtual_avail; 359 pte = vtopte(va); 360 361 /* 362 * CMAP1/CMAP2 are used for zeroing and copying pages. 363 * CMAP3 is used for the idle process page zeroing. 364 */ 365 for (i = 0; i < MAXCPU; i++) { 366 sysmaps = &sysmaps_pcpu[i]; 367 mtx_init(&sysmaps->lock, "SYSMAPS", NULL, MTX_DEF); 368 SYSMAP(caddr_t, sysmaps->CMAP1, sysmaps->CADDR1, 1) 369 SYSMAP(caddr_t, sysmaps->CMAP2, sysmaps->CADDR2, 1) 370 } 371 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 372 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 373 *CMAP3 = 0; 374 375 /* 376 * Crashdump maps. 377 */ 378 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS) 379 380 /* 381 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 382 */ 383 SYSMAP(caddr_t, unused, ptvmmap, 1) 384 385 /* 386 * msgbufp is used to map the system message buffer. 387 */ 388 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE))) 389 390 /* 391 * ptemap is used for pmap_pte_quick 392 */ 393 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 394 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1); 395 396 mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF); 397 398 virtual_avail = va; 399 400 *CMAP1 = 0; 401 402#ifdef XBOX 403 /* FIXME: This is gross, but needed for the XBOX. Since we are in such 404 * an early stadium, we cannot yet neatly map video memory ... :-( 405 * Better fixes are very welcome! */ 406 if (!arch_i386_is_xbox) 407#endif 408 for (i = 0; i < NKPT; i++) 409 PTD[i] = 0; 410 411 /* Turn on PG_G on kernel page(s) */ 412 pmap_set_pg(); 413} 414 415/* 416 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 417 */ 418void 419pmap_set_pg(void) 420{ 421 pd_entry_t pdir; 422 pt_entry_t *pte; 423 vm_offset_t va, endva; 424 int i; 425 426 if (pgeflag == 0) 427 return; 428 429 i = KERNLOAD/NBPDR; 430 endva = KERNBASE + KERNend; 431 432 if (pseflag) { 433 va = KERNBASE + KERNLOAD; 434 while (va < endva) { 435 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 436 pdir |= pgeflag; 437 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 438 invltlb(); /* Play it safe, invltlb() every time */ 439 i++; 440 va += NBPDR; 441 } 442 } else { 443 va = (vm_offset_t)btext; 444 while (va < endva) { 445 pte = vtopte(va); 446 if (*pte) 447 *pte |= pgeflag; 448 invltlb(); /* Play it safe, invltlb() every time */ 449 va += PAGE_SIZE; 450 } 451 } 452} 453 454/* 455 * Initialize a vm_page's machine-dependent fields. 456 */ 457void 458pmap_page_init(vm_page_t m) 459{ 460 461 TAILQ_INIT(&m->md.pv_list); 462 m->md.pv_list_count = 0; 463} 464 465#ifdef PAE 466 467static MALLOC_DEFINE(M_PMAPPDPT, "pmap", "pmap pdpt"); 468 469static void * 470pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 471{ 472 *flags = UMA_SLAB_PRIV; 473 return (contigmalloc(PAGE_SIZE, M_PMAPPDPT, 0, 0x0ULL, 0xffffffffULL, 474 1, 0)); 475} 476#endif 477 478/* 479 * Initialize the pmap module. 480 * Called by vm_init, to initialize any structures that the pmap 481 * system needs to map virtual memory. 482 */ 483void 484pmap_init(void) 485{ 486 int shpgperproc = PMAP_SHPGPERPROC; 487 488 /* 489 * Initialize the address space (zone) for the pv entries. Set a 490 * high water mark so that the system can recover from excessive 491 * numbers of pv entries. 492 */ 493 pvzone = uma_zcreate("PV ENTRY", sizeof(struct pv_entry), NULL, NULL, 494 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 495 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 496 pv_entry_max = shpgperproc * maxproc + cnt.v_page_count; 497 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 498 pv_entry_high_water = 9 * (pv_entry_max / 10); 499 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 500 501#ifdef PAE 502 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 503 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 504 UMA_ZONE_VM | UMA_ZONE_NOFREE); 505 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 506#endif 507} 508 509void 510pmap_init2() 511{ 512} 513 514 515/*************************************************** 516 * Low level helper routines..... 517 ***************************************************/ 518 519 520/* 521 * this routine defines the region(s) of memory that should 522 * not be tested for the modified bit. 523 */ 524static PMAP_INLINE int 525pmap_track_modified(vm_offset_t va) 526{ 527 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 528 return 1; 529 else 530 return 0; 531} 532 533#ifdef SMP 534/* 535 * For SMP, these functions have to use the IPI mechanism for coherence. 536 */ 537void 538pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 539{ 540 u_int cpumask; 541 u_int other_cpus; 542 543 if (smp_started) { 544 if (!(read_eflags() & PSL_I)) 545 panic("%s: interrupts disabled", __func__); 546 mtx_lock_spin(&smp_ipi_mtx); 547 } else 548 critical_enter(); 549 /* 550 * We need to disable interrupt preemption but MUST NOT have 551 * interrupts disabled here. 552 * XXX we may need to hold schedlock to get a coherent pm_active 553 * XXX critical sections disable interrupts again 554 */ 555 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 556 invlpg(va); 557 smp_invlpg(va); 558 } else { 559 cpumask = PCPU_GET(cpumask); 560 other_cpus = PCPU_GET(other_cpus); 561 if (pmap->pm_active & cpumask) 562 invlpg(va); 563 if (pmap->pm_active & other_cpus) 564 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 565 } 566 if (smp_started) 567 mtx_unlock_spin(&smp_ipi_mtx); 568 else 569 critical_exit(); 570} 571 572void 573pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 574{ 575 u_int cpumask; 576 u_int other_cpus; 577 vm_offset_t addr; 578 579 if (smp_started) { 580 if (!(read_eflags() & PSL_I)) 581 panic("%s: interrupts disabled", __func__); 582 mtx_lock_spin(&smp_ipi_mtx); 583 } else 584 critical_enter(); 585 /* 586 * We need to disable interrupt preemption but MUST NOT have 587 * interrupts disabled here. 588 * XXX we may need to hold schedlock to get a coherent pm_active 589 * XXX critical sections disable interrupts again 590 */ 591 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 592 for (addr = sva; addr < eva; addr += PAGE_SIZE) 593 invlpg(addr); 594 smp_invlpg_range(sva, eva); 595 } else { 596 cpumask = PCPU_GET(cpumask); 597 other_cpus = PCPU_GET(other_cpus); 598 if (pmap->pm_active & cpumask) 599 for (addr = sva; addr < eva; addr += PAGE_SIZE) 600 invlpg(addr); 601 if (pmap->pm_active & other_cpus) 602 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 603 sva, eva); 604 } 605 if (smp_started) 606 mtx_unlock_spin(&smp_ipi_mtx); 607 else 608 critical_exit(); 609} 610 611void 612pmap_invalidate_all(pmap_t pmap) 613{ 614 u_int cpumask; 615 u_int other_cpus; 616 617 if (smp_started) { 618 if (!(read_eflags() & PSL_I)) 619 panic("%s: interrupts disabled", __func__); 620 mtx_lock_spin(&smp_ipi_mtx); 621 } else 622 critical_enter(); 623 /* 624 * We need to disable interrupt preemption but MUST NOT have 625 * interrupts disabled here. 626 * XXX we may need to hold schedlock to get a coherent pm_active 627 * XXX critical sections disable interrupts again 628 */ 629 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 630 invltlb(); 631 smp_invltlb(); 632 } else { 633 cpumask = PCPU_GET(cpumask); 634 other_cpus = PCPU_GET(other_cpus); 635 if (pmap->pm_active & cpumask) 636 invltlb(); 637 if (pmap->pm_active & other_cpus) 638 smp_masked_invltlb(pmap->pm_active & other_cpus); 639 } 640 if (smp_started) 641 mtx_unlock_spin(&smp_ipi_mtx); 642 else 643 critical_exit(); 644} 645#else /* !SMP */ 646/* 647 * Normal, non-SMP, 486+ invalidation functions. 648 * We inline these within pmap.c for speed. 649 */ 650PMAP_INLINE void 651pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 652{ 653 654 if (pmap == kernel_pmap || pmap->pm_active) 655 invlpg(va); 656} 657 658PMAP_INLINE void 659pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 660{ 661 vm_offset_t addr; 662 663 if (pmap == kernel_pmap || pmap->pm_active) 664 for (addr = sva; addr < eva; addr += PAGE_SIZE) 665 invlpg(addr); 666} 667 668PMAP_INLINE void 669pmap_invalidate_all(pmap_t pmap) 670{ 671 672 if (pmap == kernel_pmap || pmap->pm_active) 673 invltlb(); 674} 675#endif /* !SMP */ 676 677/* 678 * Are we current address space or kernel? N.B. We return FALSE when 679 * a pmap's page table is in use because a kernel thread is borrowing 680 * it. The borrowed page table can change spontaneously, making any 681 * dependence on its continued use subject to a race condition. 682 */ 683static __inline int 684pmap_is_current(pmap_t pmap) 685{ 686 687 return (pmap == kernel_pmap || 688 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 689 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 690} 691 692/* 693 * If the given pmap is not the current or kernel pmap, the returned pte must 694 * be released by passing it to pmap_pte_release(). 695 */ 696pt_entry_t * 697pmap_pte(pmap_t pmap, vm_offset_t va) 698{ 699 pd_entry_t newpf; 700 pd_entry_t *pde; 701 702 pde = pmap_pde(pmap, va); 703 if (*pde & PG_PS) 704 return (pde); 705 if (*pde != 0) { 706 /* are we current address space or kernel? */ 707 if (pmap_is_current(pmap)) 708 return (vtopte(va)); 709 mtx_lock(&PMAP2mutex); 710 newpf = *pde & PG_FRAME; 711 if ((*PMAP2 & PG_FRAME) != newpf) { 712 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; 713 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 714 } 715 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); 716 } 717 return (0); 718} 719 720/* 721 * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte 722 * being NULL. 723 */ 724static __inline void 725pmap_pte_release(pt_entry_t *pte) 726{ 727 728 if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2) 729 mtx_unlock(&PMAP2mutex); 730} 731 732static __inline void 733invlcaddr(void *caddr) 734{ 735 736 invlpg((u_int)caddr); 737} 738 739/* 740 * Super fast pmap_pte routine best used when scanning 741 * the pv lists. This eliminates many coarse-grained 742 * invltlb calls. Note that many of the pv list 743 * scans are across different pmaps. It is very wasteful 744 * to do an entire invltlb for checking a single mapping. 745 * 746 * If the given pmap is not the current pmap, vm_page_queue_mtx 747 * must be held and curthread pinned to a CPU. 748 */ 749static pt_entry_t * 750pmap_pte_quick(pmap_t pmap, vm_offset_t va) 751{ 752 pd_entry_t newpf; 753 pd_entry_t *pde; 754 755 pde = pmap_pde(pmap, va); 756 if (*pde & PG_PS) 757 return (pde); 758 if (*pde != 0) { 759 /* are we current address space or kernel? */ 760 if (pmap_is_current(pmap)) 761 return (vtopte(va)); 762 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 763 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 764 newpf = *pde & PG_FRAME; 765 if ((*PMAP1 & PG_FRAME) != newpf) { 766 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; 767#ifdef SMP 768 PMAP1cpu = PCPU_GET(cpuid); 769#endif 770 invlcaddr(PADDR1); 771 PMAP1changed++; 772 } else 773#ifdef SMP 774 if (PMAP1cpu != PCPU_GET(cpuid)) { 775 PMAP1cpu = PCPU_GET(cpuid); 776 invlcaddr(PADDR1); 777 PMAP1changedcpu++; 778 } else 779#endif 780 PMAP1unchanged++; 781 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); 782 } 783 return (0); 784} 785 786/* 787 * Routine: pmap_extract 788 * Function: 789 * Extract the physical page address associated 790 * with the given map/virtual_address pair. 791 */ 792vm_paddr_t 793pmap_extract(pmap_t pmap, vm_offset_t va) 794{ 795 vm_paddr_t rtval; 796 pt_entry_t *pte; 797 pd_entry_t pde; 798 799 rtval = 0; 800 PMAP_LOCK(pmap); 801 pde = pmap->pm_pdir[va >> PDRSHIFT]; 802 if (pde != 0) { 803 if ((pde & PG_PS) != 0) { 804 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 805 PMAP_UNLOCK(pmap); 806 return rtval; 807 } 808 pte = pmap_pte(pmap, va); 809 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK); 810 pmap_pte_release(pte); 811 } 812 PMAP_UNLOCK(pmap); 813 return (rtval); 814} 815 816/* 817 * Routine: pmap_extract_and_hold 818 * Function: 819 * Atomically extract and hold the physical page 820 * with the given pmap and virtual address pair 821 * if that mapping permits the given protection. 822 */ 823vm_page_t 824pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 825{ 826 pd_entry_t pde; 827 pt_entry_t pte; 828 vm_page_t m; 829 830 m = NULL; 831 vm_page_lock_queues(); 832 PMAP_LOCK(pmap); 833 pde = *pmap_pde(pmap, va); 834 if (pde != 0) { 835 if (pde & PG_PS) { 836 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) { 837 m = PHYS_TO_VM_PAGE((pde & ~PDRMASK) | 838 (va & PDRMASK)); 839 vm_page_hold(m); 840 } 841 } else { 842 sched_pin(); 843 pte = *pmap_pte_quick(pmap, va); 844 if (pte != 0 && 845 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) { 846 m = PHYS_TO_VM_PAGE(pte & PG_FRAME); 847 vm_page_hold(m); 848 } 849 sched_unpin(); 850 } 851 } 852 vm_page_unlock_queues(); 853 PMAP_UNLOCK(pmap); 854 return (m); 855} 856 857/*************************************************** 858 * Low level mapping routines..... 859 ***************************************************/ 860 861/* 862 * Add a wired page to the kva. 863 * Note: not SMP coherent. 864 */ 865PMAP_INLINE void 866pmap_kenter(vm_offset_t va, vm_paddr_t pa) 867{ 868 pt_entry_t *pte; 869 870 pte = vtopte(va); 871 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 872} 873 874/* 875 * Remove a page from the kernel pagetables. 876 * Note: not SMP coherent. 877 */ 878PMAP_INLINE void 879pmap_kremove(vm_offset_t va) 880{ 881 pt_entry_t *pte; 882 883 pte = vtopte(va); 884 pte_clear(pte); 885} 886 887/* 888 * Used to map a range of physical addresses into kernel 889 * virtual address space. 890 * 891 * The value passed in '*virt' is a suggested virtual address for 892 * the mapping. Architectures which can support a direct-mapped 893 * physical to virtual region can return the appropriate address 894 * within that region, leaving '*virt' unchanged. Other 895 * architectures should map the pages starting at '*virt' and 896 * update '*virt' with the first usable address after the mapped 897 * region. 898 */ 899vm_offset_t 900pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 901{ 902 vm_offset_t va, sva; 903 904 va = sva = *virt; 905 while (start < end) { 906 pmap_kenter(va, start); 907 va += PAGE_SIZE; 908 start += PAGE_SIZE; 909 } 910 pmap_invalidate_range(kernel_pmap, sva, va); 911 *virt = va; 912 return (sva); 913} 914 915 916/* 917 * Add a list of wired pages to the kva 918 * this routine is only used for temporary 919 * kernel mappings that do not need to have 920 * page modification or references recorded. 921 * Note that old mappings are simply written 922 * over. The page *must* be wired. 923 * Note: SMP coherent. Uses a ranged shootdown IPI. 924 */ 925void 926pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 927{ 928 vm_offset_t va; 929 930 va = sva; 931 while (count-- > 0) { 932 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 933 va += PAGE_SIZE; 934 m++; 935 } 936 pmap_invalidate_range(kernel_pmap, sva, va); 937} 938 939/* 940 * This routine tears out page mappings from the 941 * kernel -- it is meant only for temporary mappings. 942 * Note: SMP coherent. Uses a ranged shootdown IPI. 943 */ 944void 945pmap_qremove(vm_offset_t sva, int count) 946{ 947 vm_offset_t va; 948 949 va = sva; 950 while (count-- > 0) { 951 pmap_kremove(va); 952 va += PAGE_SIZE; 953 } 954 pmap_invalidate_range(kernel_pmap, sva, va); 955} 956 957/*************************************************** 958 * Page table page management routines..... 959 ***************************************************/ 960 961/* 962 * This routine unholds page table pages, and if the hold count 963 * drops to zero, then it decrements the wire count. 964 */ 965static PMAP_INLINE int 966pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 967{ 968 969 --m->wire_count; 970 if (m->wire_count == 0) 971 return _pmap_unwire_pte_hold(pmap, m); 972 else 973 return 0; 974} 975 976static int 977_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 978{ 979 vm_offset_t pteva; 980 981 /* 982 * unmap the page table page 983 */ 984 pmap->pm_pdir[m->pindex] = 0; 985 --pmap->pm_stats.resident_count; 986 987 /* 988 * Do an invltlb to make the invalidated mapping 989 * take effect immediately. 990 */ 991 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 992 pmap_invalidate_page(pmap, pteva); 993 994 vm_page_free_zero(m); 995 atomic_subtract_int(&cnt.v_wire_count, 1); 996 return 1; 997} 998 999/* 1000 * After removing a page table entry, this routine is used to 1001 * conditionally free the page, and manage the hold/wire counts. 1002 */ 1003static int 1004pmap_unuse_pt(pmap_t pmap, vm_offset_t va) 1005{ 1006 pd_entry_t ptepde; 1007 vm_page_t mpte; 1008 1009 if (va >= VM_MAXUSER_ADDRESS) 1010 return 0; 1011 ptepde = *pmap_pde(pmap, va); 1012 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); 1013 return pmap_unwire_pte_hold(pmap, mpte); 1014} 1015 1016void 1017pmap_pinit0(pmap) 1018 struct pmap *pmap; 1019{ 1020 1021 PMAP_LOCK_INIT(pmap); 1022 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1023#ifdef PAE 1024 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1025#endif 1026 pmap->pm_active = 0; 1027 PCPU_SET(curpmap, pmap); 1028 TAILQ_INIT(&pmap->pm_pvlist); 1029 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1030 mtx_lock_spin(&allpmaps_lock); 1031 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1032 mtx_unlock_spin(&allpmaps_lock); 1033} 1034 1035/* 1036 * Initialize a preallocated and zeroed pmap structure, 1037 * such as one in a vmspace structure. 1038 */ 1039void 1040pmap_pinit(pmap) 1041 register struct pmap *pmap; 1042{ 1043 vm_page_t m, ptdpg[NPGPTD]; 1044 vm_paddr_t pa; 1045 static int color; 1046 int i; 1047 1048 PMAP_LOCK_INIT(pmap); 1049 1050 /* 1051 * No need to allocate page table space yet but we do need a valid 1052 * page directory table. 1053 */ 1054 if (pmap->pm_pdir == NULL) { 1055 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1056 NBPTD); 1057#ifdef PAE 1058 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1059 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1060 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1061 ("pmap_pinit: pdpt misaligned")); 1062 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1063 ("pmap_pinit: pdpt above 4g")); 1064#endif 1065 } 1066 1067 /* 1068 * allocate the page directory page(s) 1069 */ 1070 for (i = 0; i < NPGPTD;) { 1071 m = vm_page_alloc(NULL, color++, 1072 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1073 VM_ALLOC_ZERO); 1074 if (m == NULL) 1075 VM_WAIT; 1076 else { 1077 ptdpg[i++] = m; 1078 } 1079 } 1080 1081 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1082 1083 for (i = 0; i < NPGPTD; i++) { 1084 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1085 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1086 } 1087 1088 mtx_lock_spin(&allpmaps_lock); 1089 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1090 mtx_unlock_spin(&allpmaps_lock); 1091 /* Wire in kernel global address entries. */ 1092 /* XXX copies current process, does not fill in MPPTDI */ 1093 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1094#ifdef SMP 1095 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1096#endif 1097 1098 /* install self-referential address mapping entry(s) */ 1099 for (i = 0; i < NPGPTD; i++) { 1100 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1101 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1102#ifdef PAE 1103 pmap->pm_pdpt[i] = pa | PG_V; 1104#endif 1105 } 1106 1107 pmap->pm_active = 0; 1108 TAILQ_INIT(&pmap->pm_pvlist); 1109 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1110} 1111 1112/* 1113 * this routine is called if the page table page is not 1114 * mapped correctly. 1115 */ 1116static vm_page_t 1117_pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags) 1118{ 1119 vm_paddr_t ptepa; 1120 vm_page_t m; 1121 1122 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT || 1123 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK, 1124 ("_pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK")); 1125 1126 /* 1127 * Allocate a page table page. 1128 */ 1129 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1130 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1131 if (flags & M_WAITOK) { 1132 PMAP_UNLOCK(pmap); 1133 vm_page_unlock_queues(); 1134 VM_WAIT; 1135 vm_page_lock_queues(); 1136 PMAP_LOCK(pmap); 1137 } 1138 1139 /* 1140 * Indicate the need to retry. While waiting, the page table 1141 * page may have been allocated. 1142 */ 1143 return (NULL); 1144 } 1145 if ((m->flags & PG_ZERO) == 0) 1146 pmap_zero_page(m); 1147 1148 /* 1149 * Map the pagetable page into the process address space, if 1150 * it isn't already there. 1151 */ 1152 1153 pmap->pm_stats.resident_count++; 1154 1155 ptepa = VM_PAGE_TO_PHYS(m); 1156 pmap->pm_pdir[ptepindex] = 1157 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1158 1159 return m; 1160} 1161 1162static vm_page_t 1163pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags) 1164{ 1165 unsigned ptepindex; 1166 pd_entry_t ptepa; 1167 vm_page_t m; 1168 1169 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT || 1170 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK, 1171 ("pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK")); 1172 1173 /* 1174 * Calculate pagetable page index 1175 */ 1176 ptepindex = va >> PDRSHIFT; 1177retry: 1178 /* 1179 * Get the page directory entry 1180 */ 1181 ptepa = pmap->pm_pdir[ptepindex]; 1182 1183 /* 1184 * This supports switching from a 4MB page to a 1185 * normal 4K page. 1186 */ 1187 if (ptepa & PG_PS) { 1188 pmap->pm_pdir[ptepindex] = 0; 1189 ptepa = 0; 1190 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1191 pmap_invalidate_all(kernel_pmap); 1192 } 1193 1194 /* 1195 * If the page table page is mapped, we just increment the 1196 * hold count, and activate it. 1197 */ 1198 if (ptepa) { 1199 m = PHYS_TO_VM_PAGE(ptepa); 1200 m->wire_count++; 1201 } else { 1202 /* 1203 * Here if the pte page isn't mapped, or if it has 1204 * been deallocated. 1205 */ 1206 m = _pmap_allocpte(pmap, ptepindex, flags); 1207 if (m == NULL && (flags & M_WAITOK)) 1208 goto retry; 1209 } 1210 return (m); 1211} 1212 1213 1214/*************************************************** 1215* Pmap allocation/deallocation routines. 1216 ***************************************************/ 1217 1218#ifdef SMP 1219/* 1220 * Deal with a SMP shootdown of other users of the pmap that we are 1221 * trying to dispose of. This can be a bit hairy. 1222 */ 1223static u_int *lazymask; 1224static u_int lazyptd; 1225static volatile u_int lazywait; 1226 1227void pmap_lazyfix_action(void); 1228 1229void 1230pmap_lazyfix_action(void) 1231{ 1232 u_int mymask = PCPU_GET(cpumask); 1233 1234 if (rcr3() == lazyptd) 1235 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1236 atomic_clear_int(lazymask, mymask); 1237 atomic_store_rel_int(&lazywait, 1); 1238} 1239 1240static void 1241pmap_lazyfix_self(u_int mymask) 1242{ 1243 1244 if (rcr3() == lazyptd) 1245 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1246 atomic_clear_int(lazymask, mymask); 1247} 1248 1249 1250static void 1251pmap_lazyfix(pmap_t pmap) 1252{ 1253 u_int mymask; 1254 u_int mask; 1255 register u_int spins; 1256 1257 while ((mask = pmap->pm_active) != 0) { 1258 spins = 50000000; 1259 mask = mask & -mask; /* Find least significant set bit */ 1260 mtx_lock_spin(&smp_ipi_mtx); 1261#ifdef PAE 1262 lazyptd = vtophys(pmap->pm_pdpt); 1263#else 1264 lazyptd = vtophys(pmap->pm_pdir); 1265#endif 1266 mymask = PCPU_GET(cpumask); 1267 if (mask == mymask) { 1268 lazymask = &pmap->pm_active; 1269 pmap_lazyfix_self(mymask); 1270 } else { 1271 atomic_store_rel_int((u_int *)&lazymask, 1272 (u_int)&pmap->pm_active); 1273 atomic_store_rel_int(&lazywait, 0); 1274 ipi_selected(mask, IPI_LAZYPMAP); 1275 while (lazywait == 0) { 1276 ia32_pause(); 1277 if (--spins == 0) 1278 break; 1279 } 1280 } 1281 mtx_unlock_spin(&smp_ipi_mtx); 1282 if (spins == 0) 1283 printf("pmap_lazyfix: spun for 50000000\n"); 1284 } 1285} 1286 1287#else /* SMP */ 1288 1289/* 1290 * Cleaning up on uniprocessor is easy. For various reasons, we're 1291 * unlikely to have to even execute this code, including the fact 1292 * that the cleanup is deferred until the parent does a wait(2), which 1293 * means that another userland process has run. 1294 */ 1295static void 1296pmap_lazyfix(pmap_t pmap) 1297{ 1298 u_int cr3; 1299 1300 cr3 = vtophys(pmap->pm_pdir); 1301 if (cr3 == rcr3()) { 1302 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1303 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1304 } 1305} 1306#endif /* SMP */ 1307 1308/* 1309 * Release any resources held by the given physical map. 1310 * Called when a pmap initialized by pmap_pinit is being released. 1311 * Should only be called if the map contains no valid mappings. 1312 */ 1313void 1314pmap_release(pmap_t pmap) 1315{ 1316 vm_page_t m, ptdpg[NPGPTD]; 1317 int i; 1318 1319 KASSERT(pmap->pm_stats.resident_count == 0, 1320 ("pmap_release: pmap resident count %ld != 0", 1321 pmap->pm_stats.resident_count)); 1322 1323 pmap_lazyfix(pmap); 1324 mtx_lock_spin(&allpmaps_lock); 1325 LIST_REMOVE(pmap, pm_list); 1326 mtx_unlock_spin(&allpmaps_lock); 1327 1328 for (i = 0; i < NPGPTD; i++) 1329 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1330 1331 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1332 sizeof(*pmap->pm_pdir)); 1333#ifdef SMP 1334 pmap->pm_pdir[MPPTDI] = 0; 1335#endif 1336 1337 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1338 1339 vm_page_lock_queues(); 1340 for (i = 0; i < NPGPTD; i++) { 1341 m = ptdpg[i]; 1342#ifdef PAE 1343 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1344 ("pmap_release: got wrong ptd page")); 1345#endif 1346 m->wire_count--; 1347 atomic_subtract_int(&cnt.v_wire_count, 1); 1348 vm_page_free_zero(m); 1349 } 1350 vm_page_unlock_queues(); 1351 PMAP_LOCK_DESTROY(pmap); 1352} 1353 1354static int 1355kvm_size(SYSCTL_HANDLER_ARGS) 1356{ 1357 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1358 1359 return sysctl_handle_long(oidp, &ksize, 0, req); 1360} 1361SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1362 0, 0, kvm_size, "IU", "Size of KVM"); 1363 1364static int 1365kvm_free(SYSCTL_HANDLER_ARGS) 1366{ 1367 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1368 1369 return sysctl_handle_long(oidp, &kfree, 0, req); 1370} 1371SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1372 0, 0, kvm_free, "IU", "Amount of KVM free"); 1373 1374/* 1375 * grow the number of kernel page table entries, if needed 1376 */ 1377void 1378pmap_growkernel(vm_offset_t addr) 1379{ 1380 struct pmap *pmap; 1381 vm_paddr_t ptppaddr; 1382 vm_page_t nkpg; 1383 pd_entry_t newpdir; 1384 pt_entry_t *pde; 1385 1386 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1387 if (kernel_vm_end == 0) { 1388 kernel_vm_end = KERNBASE; 1389 nkpt = 0; 1390 while (pdir_pde(PTD, kernel_vm_end)) { 1391 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1392 nkpt++; 1393 } 1394 } 1395 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1396 while (kernel_vm_end < addr) { 1397 if (pdir_pde(PTD, kernel_vm_end)) { 1398 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1399 continue; 1400 } 1401 1402 /* 1403 * This index is bogus, but out of the way 1404 */ 1405 nkpg = vm_page_alloc(NULL, nkpt, 1406 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1407 if (!nkpg) 1408 panic("pmap_growkernel: no memory to grow kernel"); 1409 1410 nkpt++; 1411 1412 pmap_zero_page(nkpg); 1413 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1414 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1415 pdir_pde(PTD, kernel_vm_end) = newpdir; 1416 1417 mtx_lock_spin(&allpmaps_lock); 1418 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1419 pde = pmap_pde(pmap, kernel_vm_end); 1420 pde_store(pde, newpdir); 1421 } 1422 mtx_unlock_spin(&allpmaps_lock); 1423 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1424 } 1425} 1426 1427 1428/*************************************************** 1429 * page management routines. 1430 ***************************************************/ 1431 1432/* 1433 * free the pv_entry back to the free list 1434 */ 1435static PMAP_INLINE void 1436free_pv_entry(pv_entry_t pv) 1437{ 1438 pv_entry_count--; 1439 uma_zfree(pvzone, pv); 1440} 1441 1442/* 1443 * get a new pv_entry, allocating a block from the system 1444 * when needed. 1445 * the memory allocation is performed bypassing the malloc code 1446 * because of the possibility of allocations at interrupt time. 1447 */ 1448static pv_entry_t 1449get_pv_entry(void) 1450{ 1451 pv_entry_count++; 1452 if ((pv_entry_count > pv_entry_high_water) && 1453 (pmap_pagedaemon_waken == 0)) { 1454 pmap_pagedaemon_waken = 1; 1455 wakeup (&vm_pages_needed); 1456 } 1457 return uma_zalloc(pvzone, M_NOWAIT); 1458} 1459 1460/* 1461 * Reclaim a pv entry by removing a mapping to an inactive page. 1462 */ 1463static pv_entry_t 1464pv_entry_reclaim(pmap_t locked_pmap) 1465{ 1466 pmap_t pmap; 1467 pt_entry_t *pte, tpte; 1468 pv_entry_t pv; 1469 vm_offset_t va; 1470 vm_page_t m; 1471 1472 PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED); 1473 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1474 sched_pin(); 1475 TAILQ_FOREACH(m, &vm_page_queues[PQ_INACTIVE].pl, pageq) { 1476 if (m->hold_count || m->busy || (m->flags & PG_BUSY)) 1477 continue; 1478 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1479 va = pv->pv_va; 1480 pmap = pv->pv_pmap; 1481 if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap)) 1482 continue; 1483 pmap->pm_stats.resident_count--; 1484 pte = pmap_pte_quick(pmap, va); 1485 tpte = pte_load_clear(pte); 1486 KASSERT((tpte & PG_W) == 0, 1487 ("pv_entry_reclaim: wired pte %#jx", 1488 (uintmax_t)tpte)); 1489 if (tpte & PG_A) 1490 vm_page_flag_set(m, PG_REFERENCED); 1491 if (tpte & PG_M) { 1492 KASSERT((tpte & PG_RW), 1493 ("pv_entry_reclaim: modified page not writable: va: %#x, pte: %#jx", 1494 va, (uintmax_t)tpte)); 1495 if (pmap_track_modified(va)) 1496 vm_page_dirty(m); 1497 } 1498 pmap_invalidate_page(pmap, va); 1499 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1500 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1501 if (TAILQ_EMPTY(&m->md.pv_list)) 1502 vm_page_flag_clear(m, PG_WRITEABLE); 1503 m->md.pv_list_count--; 1504 pmap_unuse_pt(pmap, va); 1505 if (pmap != locked_pmap) 1506 PMAP_UNLOCK(pmap); 1507 sched_unpin(); 1508 return (pv); 1509 } 1510 } 1511 sched_unpin(); 1512 panic("pv_entry_reclaim: increase vm.pmap.shpgperproc"); 1513} 1514 1515static void 1516pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1517{ 1518 pv_entry_t pv; 1519 1520 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1521 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1522 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1523 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1524 if (pmap == pv->pv_pmap && va == pv->pv_va) 1525 break; 1526 } 1527 } else { 1528 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1529 if (va == pv->pv_va) 1530 break; 1531 } 1532 } 1533 KASSERT(pv != NULL, ("pmap_remove_entry: pv not found")); 1534 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1535 m->md.pv_list_count--; 1536 if (TAILQ_EMPTY(&m->md.pv_list)) 1537 vm_page_flag_clear(m, PG_WRITEABLE); 1538 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1539 free_pv_entry(pv); 1540} 1541 1542/* 1543 * Create a pv entry for page at pa for 1544 * (pmap, va). 1545 */ 1546static void 1547pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) 1548{ 1549 pv_entry_t pv; 1550 1551 pv = get_pv_entry(); 1552 if (pv == NULL) { 1553 pv_entry_count--; 1554 pv = pv_entry_reclaim(pmap); 1555 } 1556 pv->pv_va = va; 1557 pv->pv_pmap = pmap; 1558 1559 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1560 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1561 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1562 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1563 m->md.pv_list_count++; 1564} 1565 1566/* 1567 * pmap_remove_pte: do the things to unmap a page in a process 1568 */ 1569static int 1570pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1571{ 1572 pt_entry_t oldpte; 1573 vm_page_t m; 1574 1575 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1576 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1577 oldpte = pte_load_clear(ptq); 1578 if (oldpte & PG_W) 1579 pmap->pm_stats.wired_count -= 1; 1580 /* 1581 * Machines that don't support invlpg, also don't support 1582 * PG_G. 1583 */ 1584 if (oldpte & PG_G) 1585 pmap_invalidate_page(kernel_pmap, va); 1586 pmap->pm_stats.resident_count -= 1; 1587 if (oldpte & PG_MANAGED) { 1588 m = PHYS_TO_VM_PAGE(oldpte); 1589 if (oldpte & PG_M) { 1590 KASSERT((oldpte & PG_RW), 1591 ("pmap_remove_pte: modified page not writable: va: %#x, pte: %#jx", 1592 va, (uintmax_t)oldpte)); 1593 if (pmap_track_modified(va)) 1594 vm_page_dirty(m); 1595 } 1596 if (oldpte & PG_A) 1597 vm_page_flag_set(m, PG_REFERENCED); 1598 pmap_remove_entry(pmap, m, va); 1599 } 1600 return (pmap_unuse_pt(pmap, va)); 1601} 1602 1603/* 1604 * Remove a single page from a process address space 1605 */ 1606static void 1607pmap_remove_page(pmap_t pmap, vm_offset_t va) 1608{ 1609 pt_entry_t *pte; 1610 1611 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1612 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 1613 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1614 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) 1615 return; 1616 pmap_remove_pte(pmap, pte, va); 1617 pmap_invalidate_page(pmap, va); 1618} 1619 1620/* 1621 * Remove the given range of addresses from the specified map. 1622 * 1623 * It is assumed that the start and end are properly 1624 * rounded to the page size. 1625 */ 1626void 1627pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1628{ 1629 vm_offset_t pdnxt; 1630 pd_entry_t ptpaddr; 1631 pt_entry_t *pte; 1632 int anyvalid; 1633 1634 /* 1635 * Perform an unsynchronized read. This is, however, safe. 1636 */ 1637 if (pmap->pm_stats.resident_count == 0) 1638 return; 1639 1640 anyvalid = 0; 1641 1642 vm_page_lock_queues(); 1643 sched_pin(); 1644 PMAP_LOCK(pmap); 1645 1646 /* 1647 * special handling of removing one page. a very 1648 * common operation and easy to short circuit some 1649 * code. 1650 */ 1651 if ((sva + PAGE_SIZE == eva) && 1652 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1653 pmap_remove_page(pmap, sva); 1654 goto out; 1655 } 1656 1657 for (; sva < eva; sva = pdnxt) { 1658 unsigned pdirindex; 1659 1660 /* 1661 * Calculate index for next page table. 1662 */ 1663 pdnxt = (sva + NBPDR) & ~PDRMASK; 1664 if (pmap->pm_stats.resident_count == 0) 1665 break; 1666 1667 pdirindex = sva >> PDRSHIFT; 1668 ptpaddr = pmap->pm_pdir[pdirindex]; 1669 1670 /* 1671 * Weed out invalid mappings. Note: we assume that the page 1672 * directory table is always allocated, and in kernel virtual. 1673 */ 1674 if (ptpaddr == 0) 1675 continue; 1676 1677 /* 1678 * Check for large page. 1679 */ 1680 if ((ptpaddr & PG_PS) != 0) { 1681 pmap->pm_pdir[pdirindex] = 0; 1682 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1683 anyvalid = 1; 1684 continue; 1685 } 1686 1687 /* 1688 * Limit our scan to either the end of the va represented 1689 * by the current page table page, or to the end of the 1690 * range being removed. 1691 */ 1692 if (pdnxt > eva) 1693 pdnxt = eva; 1694 1695 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, 1696 sva += PAGE_SIZE) { 1697 if (*pte == 0) 1698 continue; 1699 anyvalid = 1; 1700 if (pmap_remove_pte(pmap, pte, sva)) 1701 break; 1702 } 1703 } 1704out: 1705 sched_unpin(); 1706 vm_page_unlock_queues(); 1707 if (anyvalid) 1708 pmap_invalidate_all(pmap); 1709 PMAP_UNLOCK(pmap); 1710} 1711 1712/* 1713 * Routine: pmap_remove_all 1714 * Function: 1715 * Removes this physical page from 1716 * all physical maps in which it resides. 1717 * Reflects back modify bits to the pager. 1718 * 1719 * Notes: 1720 * Original versions of this routine were very 1721 * inefficient because they iteratively called 1722 * pmap_remove (slow...) 1723 */ 1724 1725void 1726pmap_remove_all(vm_page_t m) 1727{ 1728 register pv_entry_t pv; 1729 pt_entry_t *pte, tpte; 1730 1731#if defined(PMAP_DIAGNOSTIC) 1732 /* 1733 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1734 */ 1735 if (m->flags & PG_FICTITIOUS) { 1736 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1737 VM_PAGE_TO_PHYS(m)); 1738 } 1739#endif 1740 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1741 sched_pin(); 1742 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1743 PMAP_LOCK(pv->pv_pmap); 1744 pv->pv_pmap->pm_stats.resident_count--; 1745 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1746 tpte = pte_load_clear(pte); 1747 if (tpte & PG_W) 1748 pv->pv_pmap->pm_stats.wired_count--; 1749 if (tpte & PG_A) 1750 vm_page_flag_set(m, PG_REFERENCED); 1751 1752 /* 1753 * Update the vm_page_t clean and reference bits. 1754 */ 1755 if (tpte & PG_M) { 1756 KASSERT((tpte & PG_RW), 1757 ("pmap_remove_all: modified page not writable: va: %#x, pte: %#jx", 1758 pv->pv_va, (uintmax_t)tpte)); 1759 if (pmap_track_modified(pv->pv_va)) 1760 vm_page_dirty(m); 1761 } 1762 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1763 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1764 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1765 m->md.pv_list_count--; 1766 pmap_unuse_pt(pv->pv_pmap, pv->pv_va); 1767 PMAP_UNLOCK(pv->pv_pmap); 1768 free_pv_entry(pv); 1769 } 1770 vm_page_flag_clear(m, PG_WRITEABLE); 1771 sched_unpin(); 1772} 1773 1774/* 1775 * Set the physical protection on the 1776 * specified range of this map as requested. 1777 */ 1778void 1779pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1780{ 1781 vm_offset_t pdnxt; 1782 pd_entry_t ptpaddr; 1783 pt_entry_t *pte; 1784 int anychanged; 1785 1786 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1787 pmap_remove(pmap, sva, eva); 1788 return; 1789 } 1790 1791 if (prot & VM_PROT_WRITE) 1792 return; 1793 1794 anychanged = 0; 1795 1796 vm_page_lock_queues(); 1797 sched_pin(); 1798 PMAP_LOCK(pmap); 1799 for (; sva < eva; sva = pdnxt) { 1800 unsigned obits, pbits, pdirindex; 1801 1802 pdnxt = (sva + NBPDR) & ~PDRMASK; 1803 1804 pdirindex = sva >> PDRSHIFT; 1805 ptpaddr = pmap->pm_pdir[pdirindex]; 1806 1807 /* 1808 * Weed out invalid mappings. Note: we assume that the page 1809 * directory table is always allocated, and in kernel virtual. 1810 */ 1811 if (ptpaddr == 0) 1812 continue; 1813 1814 /* 1815 * Check for large page. 1816 */ 1817 if ((ptpaddr & PG_PS) != 0) { 1818 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1819 anychanged = 1; 1820 continue; 1821 } 1822 1823 if (pdnxt > eva) 1824 pdnxt = eva; 1825 1826 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, 1827 sva += PAGE_SIZE) { 1828 vm_page_t m; 1829 1830retry: 1831 /* 1832 * Regardless of whether a pte is 32 or 64 bits in 1833 * size, PG_RW, PG_A, and PG_M are among the least 1834 * significant 32 bits. 1835 */ 1836 obits = pbits = *(u_int *)pte; 1837 if (pbits & PG_MANAGED) { 1838 m = NULL; 1839 if (pbits & PG_A) { 1840 m = PHYS_TO_VM_PAGE(*pte); 1841 vm_page_flag_set(m, PG_REFERENCED); 1842 pbits &= ~PG_A; 1843 } 1844 if ((pbits & PG_M) != 0 && 1845 pmap_track_modified(sva)) { 1846 if (m == NULL) 1847 m = PHYS_TO_VM_PAGE(*pte); 1848 vm_page_dirty(m); 1849 } 1850 } 1851 1852 pbits &= ~(PG_RW | PG_M); 1853 1854 if (pbits != obits) { 1855 if (!atomic_cmpset_int((u_int *)pte, obits, 1856 pbits)) 1857 goto retry; 1858 if (obits & PG_G) 1859 pmap_invalidate_page(pmap, sva); 1860 else 1861 anychanged = 1; 1862 } 1863 } 1864 } 1865 sched_unpin(); 1866 vm_page_unlock_queues(); 1867 if (anychanged) 1868 pmap_invalidate_all(pmap); 1869 PMAP_UNLOCK(pmap); 1870} 1871 1872/* 1873 * Insert the given physical page (p) at 1874 * the specified virtual address (v) in the 1875 * target physical map with the protection requested. 1876 * 1877 * If specified, the page will be wired down, meaning 1878 * that the related pte can not be reclaimed. 1879 * 1880 * NB: This is the only routine which MAY NOT lazy-evaluate 1881 * or lose information. That is, this routine must actually 1882 * insert this page into the given map NOW. 1883 */ 1884void 1885pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1886 boolean_t wired) 1887{ 1888 vm_paddr_t pa; 1889 register pt_entry_t *pte; 1890 vm_paddr_t opa; 1891 pt_entry_t origpte, newpte; 1892 vm_page_t mpte, om; 1893 boolean_t invlva; 1894 1895 va &= PG_FRAME; 1896#ifdef PMAP_DIAGNOSTIC 1897 if (va > VM_MAX_KERNEL_ADDRESS) 1898 panic("pmap_enter: toobig"); 1899 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1900 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1901#endif 1902 1903 mpte = NULL; 1904 1905 vm_page_lock_queues(); 1906 PMAP_LOCK(pmap); 1907 sched_pin(); 1908 1909 /* 1910 * In the case that a page table page is not 1911 * resident, we are creating it here. 1912 */ 1913 if (va < VM_MAXUSER_ADDRESS) { 1914 mpte = pmap_allocpte(pmap, va, M_WAITOK); 1915 } 1916#if 0 && defined(PMAP_DIAGNOSTIC) 1917 else { 1918 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1919 origpte = *pdeaddr; 1920 if ((origpte & PG_V) == 0) { 1921 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1922 pmap->pm_pdir[PTDPTDI], origpte, va); 1923 } 1924 } 1925#endif 1926 1927 pte = pmap_pte_quick(pmap, va); 1928 1929 /* 1930 * Page Directory table entry not valid, we need a new PT page 1931 */ 1932 if (pte == NULL) { 1933 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1934 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1935 } 1936 1937 pa = VM_PAGE_TO_PHYS(m); 1938 om = NULL; 1939 origpte = *pte; 1940 opa = origpte & PG_FRAME; 1941 1942 if (origpte & PG_PS) { 1943 /* 1944 * Yes, I know this will truncate upper address bits for PAE, 1945 * but I'm actually more interested in the lower bits 1946 */ 1947 printf("pmap_enter: va %p, pte %p, origpte %p\n", 1948 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 1949 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1950 } 1951 1952 /* 1953 * Mapping has not changed, must be protection or wiring change. 1954 */ 1955 if (origpte && (opa == pa)) { 1956 /* 1957 * Wiring change, just update stats. We don't worry about 1958 * wiring PT pages as they remain resident as long as there 1959 * are valid mappings in them. Hence, if a user page is wired, 1960 * the PT page will be also. 1961 */ 1962 if (wired && ((origpte & PG_W) == 0)) 1963 pmap->pm_stats.wired_count++; 1964 else if (!wired && (origpte & PG_W)) 1965 pmap->pm_stats.wired_count--; 1966 1967 /* 1968 * Remove extra pte reference 1969 */ 1970 if (mpte) 1971 mpte->wire_count--; 1972 1973 /* 1974 * We might be turning off write access to the page, 1975 * so we go ahead and sense modify status. 1976 */ 1977 if (origpte & PG_MANAGED) { 1978 om = m; 1979 pa |= PG_MANAGED; 1980 } 1981 goto validate; 1982 } 1983 /* 1984 * Mapping has changed, invalidate old range and fall through to 1985 * handle validating new mapping. 1986 */ 1987 if (opa) { 1988 if (origpte & PG_W) 1989 pmap->pm_stats.wired_count--; 1990 if (origpte & PG_MANAGED) { 1991 om = PHYS_TO_VM_PAGE(opa); 1992 pmap_remove_entry(pmap, om, va); 1993 } 1994 if (mpte != NULL) { 1995 mpte->wire_count--; 1996 KASSERT(mpte->wire_count > 0, 1997 ("pmap_enter: missing reference to page table page," 1998 " va: 0x%x", va)); 1999 } 2000 } else 2001 pmap->pm_stats.resident_count++; 2002 2003 /* 2004 * Enter on the PV list if part of our managed memory. 2005 */ 2006 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0) { 2007 pmap_insert_entry(pmap, va, m); 2008 pa |= PG_MANAGED; 2009 } 2010 2011 /* 2012 * Increment counters 2013 */ 2014 if (wired) 2015 pmap->pm_stats.wired_count++; 2016 2017validate: 2018 /* 2019 * Now validate mapping with desired protection/wiring. 2020 */ 2021 newpte = (pt_entry_t)(pa | PG_V); 2022 if ((prot & VM_PROT_WRITE) != 0) 2023 newpte |= PG_RW; 2024 if (wired) 2025 newpte |= PG_W; 2026 if (va < VM_MAXUSER_ADDRESS) 2027 newpte |= PG_U; 2028 if (pmap == kernel_pmap) 2029 newpte |= pgeflag; 2030 2031 /* 2032 * if the mapping or permission bits are different, we need 2033 * to update the pte. 2034 */ 2035 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2036 if (origpte & PG_V) { 2037 invlva = FALSE; 2038 origpte = pte_load_store(pte, newpte | PG_A); 2039 if (origpte & PG_A) { 2040 if (origpte & PG_MANAGED) 2041 vm_page_flag_set(om, PG_REFERENCED); 2042 if (opa != VM_PAGE_TO_PHYS(m)) 2043 invlva = TRUE; 2044 } 2045 if (origpte & PG_M) { 2046 KASSERT((origpte & PG_RW), 2047 ("pmap_enter: modified page not writable: va: %#x, pte: %#jx", 2048 va, (uintmax_t)origpte)); 2049 if ((origpte & PG_MANAGED) && 2050 pmap_track_modified(va)) 2051 vm_page_dirty(om); 2052 if ((prot & VM_PROT_WRITE) == 0) 2053 invlva = TRUE; 2054 } 2055 if (invlva) 2056 pmap_invalidate_page(pmap, va); 2057 } else 2058 pte_store(pte, newpte | PG_A); 2059 } 2060 sched_unpin(); 2061 vm_page_unlock_queues(); 2062 PMAP_UNLOCK(pmap); 2063} 2064 2065/* 2066 * this code makes some *MAJOR* assumptions: 2067 * 1. Current pmap & pmap exists. 2068 * 2. Not wired. 2069 * 3. Read access. 2070 * 4. No page table pages. 2071 * but is *MUCH* faster than pmap_enter... 2072 */ 2073 2074vm_page_t 2075pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 2076 vm_page_t mpte) 2077{ 2078 pt_entry_t *pte; 2079 vm_paddr_t pa; 2080 2081 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2082 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2083 PMAP_LOCK(pmap); 2084 2085 /* 2086 * In the case that a page table page is not 2087 * resident, we are creating it here. 2088 */ 2089 if (va < VM_MAXUSER_ADDRESS) { 2090 unsigned ptepindex; 2091 pd_entry_t ptepa; 2092 2093 /* 2094 * Calculate pagetable page index 2095 */ 2096 ptepindex = va >> PDRSHIFT; 2097 if (mpte && (mpte->pindex == ptepindex)) { 2098 mpte->wire_count++; 2099 } else { 2100retry: 2101 /* 2102 * Get the page directory entry 2103 */ 2104 ptepa = pmap->pm_pdir[ptepindex]; 2105 2106 /* 2107 * If the page table page is mapped, we just increment 2108 * the hold count, and activate it. 2109 */ 2110 if (ptepa) { 2111 if (ptepa & PG_PS) 2112 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2113 mpte = PHYS_TO_VM_PAGE(ptepa); 2114 mpte->wire_count++; 2115 } else { 2116 mpte = _pmap_allocpte(pmap, ptepindex, 2117 M_NOWAIT); 2118 if (mpte == NULL) { 2119 PMAP_UNLOCK(pmap); 2120 vm_page_busy(m); 2121 vm_page_unlock_queues(); 2122 VM_OBJECT_UNLOCK(m->object); 2123 VM_WAIT; 2124 VM_OBJECT_LOCK(m->object); 2125 vm_page_lock_queues(); 2126 vm_page_wakeup(m); 2127 PMAP_LOCK(pmap); 2128 goto retry; 2129 } 2130 } 2131 } 2132 } else { 2133 mpte = NULL; 2134 } 2135 2136 /* 2137 * This call to vtopte makes the assumption that we are 2138 * entering the page into the current pmap. In order to support 2139 * quick entry into any pmap, one would likely use pmap_pte_quick. 2140 * But that isn't as quick as vtopte. 2141 */ 2142 pte = vtopte(va); 2143 if (*pte) { 2144 if (mpte != NULL) { 2145 pmap_unwire_pte_hold(pmap, mpte); 2146 mpte = NULL; 2147 } 2148 goto out; 2149 } 2150 2151 /* 2152 * Enter on the PV list if part of our managed memory. Note that we 2153 * raise IPL while manipulating pv_table since pmap_enter can be 2154 * called at interrupt time. 2155 */ 2156 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2157 pmap_insert_entry(pmap, va, m); 2158 2159 /* 2160 * Increment counters 2161 */ 2162 pmap->pm_stats.resident_count++; 2163 2164 pa = VM_PAGE_TO_PHYS(m); 2165 2166 /* 2167 * Now validate mapping with RO protection 2168 */ 2169 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2170 pte_store(pte, pa | PG_V | PG_U); 2171 else 2172 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2173out: 2174 PMAP_UNLOCK(pmap); 2175 return mpte; 2176} 2177 2178/* 2179 * Make a temporary mapping for a physical address. This is only intended 2180 * to be used for panic dumps. 2181 */ 2182void * 2183pmap_kenter_temporary(vm_paddr_t pa, int i) 2184{ 2185 vm_offset_t va; 2186 2187 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2188 pmap_kenter(va, pa); 2189 invlpg(va); 2190 return ((void *)crashdumpmap); 2191} 2192 2193/* 2194 * This code maps large physical mmap regions into the 2195 * processor address space. Note that some shortcuts 2196 * are taken, but the code works. 2197 */ 2198void 2199pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2200 vm_object_t object, vm_pindex_t pindex, 2201 vm_size_t size) 2202{ 2203 vm_page_t p; 2204 2205 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2206 KASSERT(object->type == OBJT_DEVICE, 2207 ("pmap_object_init_pt: non-device object")); 2208 if (pseflag && 2209 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2210 int i; 2211 vm_page_t m[1]; 2212 unsigned int ptepindex; 2213 int npdes; 2214 pd_entry_t ptepa; 2215 2216 PMAP_LOCK(pmap); 2217 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2218 goto out; 2219 PMAP_UNLOCK(pmap); 2220retry: 2221 p = vm_page_lookup(object, pindex); 2222 if (p != NULL) { 2223 vm_page_lock_queues(); 2224 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2225 goto retry; 2226 } else { 2227 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2228 if (p == NULL) 2229 return; 2230 m[0] = p; 2231 2232 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2233 vm_page_lock_queues(); 2234 vm_page_free(p); 2235 vm_page_unlock_queues(); 2236 return; 2237 } 2238 2239 p = vm_page_lookup(object, pindex); 2240 vm_page_lock_queues(); 2241 vm_page_wakeup(p); 2242 } 2243 vm_page_unlock_queues(); 2244 2245 ptepa = VM_PAGE_TO_PHYS(p); 2246 if (ptepa & (NBPDR - 1)) 2247 return; 2248 2249 p->valid = VM_PAGE_BITS_ALL; 2250 2251 PMAP_LOCK(pmap); 2252 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2253 npdes = size >> PDRSHIFT; 2254 for(i = 0; i < npdes; i++) { 2255 pde_store(&pmap->pm_pdir[ptepindex], 2256 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2257 ptepa += NBPDR; 2258 ptepindex += 1; 2259 } 2260 pmap_invalidate_all(pmap); 2261out: 2262 PMAP_UNLOCK(pmap); 2263 } 2264} 2265 2266/* 2267 * Routine: pmap_change_wiring 2268 * Function: Change the wiring attribute for a map/virtual-address 2269 * pair. 2270 * In/out conditions: 2271 * The mapping must already exist in the pmap. 2272 */ 2273void 2274pmap_change_wiring(pmap, va, wired) 2275 register pmap_t pmap; 2276 vm_offset_t va; 2277 boolean_t wired; 2278{ 2279 register pt_entry_t *pte; 2280 2281 PMAP_LOCK(pmap); 2282 pte = pmap_pte(pmap, va); 2283 2284 if (wired && !pmap_pte_w(pte)) 2285 pmap->pm_stats.wired_count++; 2286 else if (!wired && pmap_pte_w(pte)) 2287 pmap->pm_stats.wired_count--; 2288 2289 /* 2290 * Wiring is not a hardware characteristic so there is no need to 2291 * invalidate TLB. 2292 */ 2293 pmap_pte_set_w(pte, wired); 2294 pmap_pte_release(pte); 2295 PMAP_UNLOCK(pmap); 2296} 2297 2298 2299 2300/* 2301 * Copy the range specified by src_addr/len 2302 * from the source map to the range dst_addr/len 2303 * in the destination map. 2304 * 2305 * This routine is only advisory and need not do anything. 2306 */ 2307 2308void 2309pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2310 vm_offset_t src_addr) 2311{ 2312 vm_offset_t addr; 2313 vm_offset_t end_addr = src_addr + len; 2314 vm_offset_t pdnxt; 2315 vm_page_t m; 2316 2317 if (dst_addr != src_addr) 2318 return; 2319 2320 if (!pmap_is_current(src_pmap)) 2321 return; 2322 2323 vm_page_lock_queues(); 2324 if (dst_pmap < src_pmap) { 2325 PMAP_LOCK(dst_pmap); 2326 PMAP_LOCK(src_pmap); 2327 } else { 2328 PMAP_LOCK(src_pmap); 2329 PMAP_LOCK(dst_pmap); 2330 } 2331 sched_pin(); 2332 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2333 pt_entry_t *src_pte, *dst_pte; 2334 vm_page_t dstmpte, srcmpte; 2335 pd_entry_t srcptepaddr; 2336 unsigned ptepindex; 2337 2338 if (addr >= UPT_MIN_ADDRESS) 2339 panic("pmap_copy: invalid to pmap_copy page tables"); 2340 2341 /* 2342 * Don't let optional prefaulting of pages make us go 2343 * way below the low water mark of free pages or way 2344 * above high water mark of used pv entries. 2345 */ 2346 if (cnt.v_free_count < cnt.v_free_reserved || 2347 pv_entry_count > pv_entry_high_water) 2348 break; 2349 2350 pdnxt = (addr + NBPDR) & ~PDRMASK; 2351 ptepindex = addr >> PDRSHIFT; 2352 2353 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2354 if (srcptepaddr == 0) 2355 continue; 2356 2357 if (srcptepaddr & PG_PS) { 2358 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2359 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2360 dst_pmap->pm_stats.resident_count += 2361 NBPDR / PAGE_SIZE; 2362 } 2363 continue; 2364 } 2365 2366 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2367 if (srcmpte->wire_count == 0) 2368 panic("pmap_copy: source page table page is unused"); 2369 2370 if (pdnxt > end_addr) 2371 pdnxt = end_addr; 2372 2373 src_pte = vtopte(addr); 2374 while (addr < pdnxt) { 2375 pt_entry_t ptetemp; 2376 ptetemp = *src_pte; 2377 /* 2378 * we only virtual copy managed pages 2379 */ 2380 if ((ptetemp & PG_MANAGED) != 0) { 2381 /* 2382 * We have to check after allocpte for the 2383 * pte still being around... allocpte can 2384 * block. 2385 */ 2386 dstmpte = pmap_allocpte(dst_pmap, addr, 2387 M_NOWAIT); 2388 if (dstmpte == NULL) 2389 break; 2390 dst_pte = pmap_pte_quick(dst_pmap, addr); 2391 if (*dst_pte == 0) { 2392 /* 2393 * Clear the modified and 2394 * accessed (referenced) bits 2395 * during the copy. 2396 */ 2397 m = PHYS_TO_VM_PAGE(ptetemp); 2398 *dst_pte = ptetemp & ~(PG_M | PG_A); 2399 dst_pmap->pm_stats.resident_count++; 2400 pmap_insert_entry(dst_pmap, addr, m); 2401 } else 2402 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2403 if (dstmpte->wire_count >= srcmpte->wire_count) 2404 break; 2405 } 2406 addr += PAGE_SIZE; 2407 src_pte++; 2408 } 2409 } 2410 sched_unpin(); 2411 vm_page_unlock_queues(); 2412 PMAP_UNLOCK(src_pmap); 2413 PMAP_UNLOCK(dst_pmap); 2414} 2415 2416static __inline void 2417pagezero(void *page) 2418{ 2419#if defined(I686_CPU) 2420 if (cpu_class == CPUCLASS_686) { 2421#if defined(CPU_ENABLE_SSE) 2422 if (cpu_feature & CPUID_SSE2) 2423 sse2_pagezero(page); 2424 else 2425#endif 2426 i686_pagezero(page); 2427 } else 2428#endif 2429 bzero(page, PAGE_SIZE); 2430} 2431 2432/* 2433 * pmap_zero_page zeros the specified hardware page by mapping 2434 * the page into KVM and using bzero to clear its contents. 2435 */ 2436void 2437pmap_zero_page(vm_page_t m) 2438{ 2439 struct sysmaps *sysmaps; 2440 2441 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2442 mtx_lock(&sysmaps->lock); 2443 if (*sysmaps->CMAP2) 2444 panic("pmap_zero_page: CMAP2 busy"); 2445 sched_pin(); 2446 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2447 invlcaddr(sysmaps->CADDR2); 2448 pagezero(sysmaps->CADDR2); 2449 *sysmaps->CMAP2 = 0; 2450 sched_unpin(); 2451 mtx_unlock(&sysmaps->lock); 2452} 2453 2454/* 2455 * pmap_zero_page_area zeros the specified hardware page by mapping 2456 * the page into KVM and using bzero to clear its contents. 2457 * 2458 * off and size may not cover an area beyond a single hardware page. 2459 */ 2460void 2461pmap_zero_page_area(vm_page_t m, int off, int size) 2462{ 2463 struct sysmaps *sysmaps; 2464 2465 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2466 mtx_lock(&sysmaps->lock); 2467 if (*sysmaps->CMAP2) 2468 panic("pmap_zero_page: CMAP2 busy"); 2469 sched_pin(); 2470 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2471 invlcaddr(sysmaps->CADDR2); 2472 if (off == 0 && size == PAGE_SIZE) 2473 pagezero(sysmaps->CADDR2); 2474 else 2475 bzero((char *)sysmaps->CADDR2 + off, size); 2476 *sysmaps->CMAP2 = 0; 2477 sched_unpin(); 2478 mtx_unlock(&sysmaps->lock); 2479} 2480 2481/* 2482 * pmap_zero_page_idle zeros the specified hardware page by mapping 2483 * the page into KVM and using bzero to clear its contents. This 2484 * is intended to be called from the vm_pagezero process only and 2485 * outside of Giant. 2486 */ 2487void 2488pmap_zero_page_idle(vm_page_t m) 2489{ 2490 2491 if (*CMAP3) 2492 panic("pmap_zero_page: CMAP3 busy"); 2493 sched_pin(); 2494 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2495 invlcaddr(CADDR3); 2496 pagezero(CADDR3); 2497 *CMAP3 = 0; 2498 sched_unpin(); 2499} 2500 2501/* 2502 * pmap_copy_page copies the specified (machine independent) 2503 * page by mapping the page into virtual memory and using 2504 * bcopy to copy the page, one machine dependent page at a 2505 * time. 2506 */ 2507void 2508pmap_copy_page(vm_page_t src, vm_page_t dst) 2509{ 2510 struct sysmaps *sysmaps; 2511 2512 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 2513 mtx_lock(&sysmaps->lock); 2514 if (*sysmaps->CMAP1) 2515 panic("pmap_copy_page: CMAP1 busy"); 2516 if (*sysmaps->CMAP2) 2517 panic("pmap_copy_page: CMAP2 busy"); 2518 sched_pin(); 2519 invlpg((u_int)sysmaps->CADDR1); 2520 invlpg((u_int)sysmaps->CADDR2); 2521 *sysmaps->CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2522 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2523 bcopy(sysmaps->CADDR1, sysmaps->CADDR2, PAGE_SIZE); 2524 *sysmaps->CMAP1 = 0; 2525 *sysmaps->CMAP2 = 0; 2526 sched_unpin(); 2527 mtx_unlock(&sysmaps->lock); 2528} 2529 2530/* 2531 * Returns true if the pmap's pv is one of the first 2532 * 16 pvs linked to from this page. This count may 2533 * be changed upwards or downwards in the future; it 2534 * is only necessary that true be returned for a small 2535 * subset of pmaps for proper page aging. 2536 */ 2537boolean_t 2538pmap_page_exists_quick(pmap, m) 2539 pmap_t pmap; 2540 vm_page_t m; 2541{ 2542 pv_entry_t pv; 2543 int loops = 0; 2544 2545 if (m->flags & PG_FICTITIOUS) 2546 return FALSE; 2547 2548 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2549 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2550 if (pv->pv_pmap == pmap) { 2551 return TRUE; 2552 } 2553 loops++; 2554 if (loops >= 16) 2555 break; 2556 } 2557 return (FALSE); 2558} 2559 2560#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2561/* 2562 * Remove all pages from specified address space 2563 * this aids process exit speeds. Also, this code 2564 * is special cased for current process only, but 2565 * can have the more generic (and slightly slower) 2566 * mode enabled. This is much faster than pmap_remove 2567 * in the case of running down an entire address space. 2568 */ 2569void 2570pmap_remove_pages(pmap, sva, eva) 2571 pmap_t pmap; 2572 vm_offset_t sva, eva; 2573{ 2574 pt_entry_t *pte, tpte; 2575 vm_page_t m; 2576 pv_entry_t pv, npv; 2577 2578#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2579 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) { 2580 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2581 return; 2582 } 2583#endif 2584 vm_page_lock_queues(); 2585 PMAP_LOCK(pmap); 2586 sched_pin(); 2587 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2588 2589 if (pv->pv_va >= eva || pv->pv_va < sva) { 2590 npv = TAILQ_NEXT(pv, pv_plist); 2591 continue; 2592 } 2593 2594#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2595 pte = vtopte(pv->pv_va); 2596#else 2597 pte = pmap_pte_quick(pmap, pv->pv_va); 2598#endif 2599 tpte = *pte; 2600 2601 if (tpte == 0) { 2602 printf("TPTE at %p IS ZERO @ VA %08x\n", 2603 pte, pv->pv_va); 2604 panic("bad pte"); 2605 } 2606 2607/* 2608 * We cannot remove wired pages from a process' mapping at this time 2609 */ 2610 if (tpte & PG_W) { 2611 npv = TAILQ_NEXT(pv, pv_plist); 2612 continue; 2613 } 2614 2615 m = PHYS_TO_VM_PAGE(tpte); 2616 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2617 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2618 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2619 2620 KASSERT(m < &vm_page_array[vm_page_array_size], 2621 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2622 2623 pmap->pm_stats.resident_count--; 2624 2625 pte_clear(pte); 2626 2627 /* 2628 * Update the vm_page_t clean and reference bits. 2629 */ 2630 if (tpte & PG_M) { 2631 vm_page_dirty(m); 2632 } 2633 2634 npv = TAILQ_NEXT(pv, pv_plist); 2635 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 2636 2637 m->md.pv_list_count--; 2638 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2639 if (TAILQ_EMPTY(&m->md.pv_list)) 2640 vm_page_flag_clear(m, PG_WRITEABLE); 2641 2642 pmap_unuse_pt(pmap, pv->pv_va); 2643 free_pv_entry(pv); 2644 } 2645 sched_unpin(); 2646 pmap_invalidate_all(pmap); 2647 PMAP_UNLOCK(pmap); 2648 vm_page_unlock_queues(); 2649} 2650 2651/* 2652 * pmap_is_modified: 2653 * 2654 * Return whether or not the specified physical page was modified 2655 * in any physical maps. 2656 */ 2657boolean_t 2658pmap_is_modified(vm_page_t m) 2659{ 2660 pv_entry_t pv; 2661 pt_entry_t *pte; 2662 boolean_t rv; 2663 2664 rv = FALSE; 2665 if (m->flags & PG_FICTITIOUS) 2666 return (rv); 2667 2668 sched_pin(); 2669 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2670 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2671 /* 2672 * if the bit being tested is the modified bit, then 2673 * mark clean_map and ptes as never 2674 * modified. 2675 */ 2676 if (!pmap_track_modified(pv->pv_va)) 2677 continue; 2678 PMAP_LOCK(pv->pv_pmap); 2679 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2680 rv = (*pte & PG_M) != 0; 2681 PMAP_UNLOCK(pv->pv_pmap); 2682 if (rv) 2683 break; 2684 } 2685 sched_unpin(); 2686 return (rv); 2687} 2688 2689/* 2690 * pmap_is_prefaultable: 2691 * 2692 * Return whether or not the specified virtual address is elgible 2693 * for prefault. 2694 */ 2695boolean_t 2696pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2697{ 2698 pt_entry_t *pte; 2699 boolean_t rv; 2700 2701 rv = FALSE; 2702 PMAP_LOCK(pmap); 2703 if (*pmap_pde(pmap, addr)) { 2704 pte = vtopte(addr); 2705 rv = *pte == 0; 2706 } 2707 PMAP_UNLOCK(pmap); 2708 return (rv); 2709} 2710 2711/* 2712 * Clear the given bit in each of the given page's ptes. The bit is 2713 * expressed as a 32-bit mask. Consequently, if the pte is 64 bits in 2714 * size, only a bit within the least significant 32 can be cleared. 2715 */ 2716static __inline void 2717pmap_clear_ptes(vm_page_t m, int bit) 2718{ 2719 register pv_entry_t pv; 2720 pt_entry_t pbits, *pte; 2721 2722 if ((m->flags & PG_FICTITIOUS) || 2723 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2724 return; 2725 2726 sched_pin(); 2727 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2728 /* 2729 * Loop over all current mappings setting/clearing as appropos If 2730 * setting RO do we need to clear the VAC? 2731 */ 2732 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2733 /* 2734 * don't write protect pager mappings 2735 */ 2736 if (bit == PG_RW) { 2737 if (!pmap_track_modified(pv->pv_va)) 2738 continue; 2739 } 2740 2741 PMAP_LOCK(pv->pv_pmap); 2742 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2743retry: 2744 pbits = *pte; 2745 if (pbits & bit) { 2746 if (bit == PG_RW) { 2747 /* 2748 * Regardless of whether a pte is 32 or 64 bits 2749 * in size, PG_RW and PG_M are among the least 2750 * significant 32 bits. 2751 */ 2752 if (!atomic_cmpset_int((u_int *)pte, pbits, 2753 pbits & ~(PG_RW | PG_M))) 2754 goto retry; 2755 if (pbits & PG_M) { 2756 vm_page_dirty(m); 2757 } 2758 } else { 2759 atomic_clear_int((u_int *)pte, bit); 2760 } 2761 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2762 } 2763 PMAP_UNLOCK(pv->pv_pmap); 2764 } 2765 if (bit == PG_RW) 2766 vm_page_flag_clear(m, PG_WRITEABLE); 2767 sched_unpin(); 2768} 2769 2770/* 2771 * pmap_page_protect: 2772 * 2773 * Lower the permission for all mappings to a given page. 2774 */ 2775void 2776pmap_page_protect(vm_page_t m, vm_prot_t prot) 2777{ 2778 if ((prot & VM_PROT_WRITE) == 0) { 2779 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2780 pmap_clear_ptes(m, PG_RW); 2781 } else { 2782 pmap_remove_all(m); 2783 } 2784 } 2785} 2786 2787/* 2788 * pmap_ts_referenced: 2789 * 2790 * Return a count of reference bits for a page, clearing those bits. 2791 * It is not necessary for every reference bit to be cleared, but it 2792 * is necessary that 0 only be returned when there are truly no 2793 * reference bits set. 2794 * 2795 * XXX: The exact number of bits to check and clear is a matter that 2796 * should be tested and standardized at some point in the future for 2797 * optimal aging of shared pages. 2798 */ 2799int 2800pmap_ts_referenced(vm_page_t m) 2801{ 2802 register pv_entry_t pv, pvf, pvn; 2803 pt_entry_t *pte; 2804 pt_entry_t v; 2805 int rtval = 0; 2806 2807 if (m->flags & PG_FICTITIOUS) 2808 return (rtval); 2809 2810 sched_pin(); 2811 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2812 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2813 2814 pvf = pv; 2815 2816 do { 2817 pvn = TAILQ_NEXT(pv, pv_list); 2818 2819 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2820 2821 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2822 2823 if (!pmap_track_modified(pv->pv_va)) 2824 continue; 2825 2826 PMAP_LOCK(pv->pv_pmap); 2827 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2828 2829 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2830 atomic_clear_int((u_int *)pte, PG_A); 2831 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2832 2833 rtval++; 2834 if (rtval > 4) { 2835 PMAP_UNLOCK(pv->pv_pmap); 2836 break; 2837 } 2838 } 2839 PMAP_UNLOCK(pv->pv_pmap); 2840 } while ((pv = pvn) != NULL && pv != pvf); 2841 } 2842 sched_unpin(); 2843 2844 return (rtval); 2845} 2846 2847/* 2848 * Clear the modify bits on the specified physical page. 2849 */ 2850void 2851pmap_clear_modify(vm_page_t m) 2852{ 2853 pmap_clear_ptes(m, PG_M); 2854} 2855 2856/* 2857 * pmap_clear_reference: 2858 * 2859 * Clear the reference bit on the specified physical page. 2860 */ 2861void 2862pmap_clear_reference(vm_page_t m) 2863{ 2864 pmap_clear_ptes(m, PG_A); 2865} 2866 2867/* 2868 * Miscellaneous support routines follow 2869 */ 2870 2871/* 2872 * Map a set of physical memory pages into the kernel virtual 2873 * address space. Return a pointer to where it is mapped. This 2874 * routine is intended to be used for mapping device memory, 2875 * NOT real memory. 2876 */ 2877void * 2878pmap_mapdev(pa, size) 2879 vm_paddr_t pa; 2880 vm_size_t size; 2881{ 2882 vm_offset_t va, tmpva, offset; 2883 2884 offset = pa & PAGE_MASK; 2885 size = roundup(offset + size, PAGE_SIZE); 2886 pa = pa & PG_FRAME; 2887 2888 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2889 va = KERNBASE + pa; 2890 else 2891 va = kmem_alloc_nofault(kernel_map, size); 2892 if (!va) 2893 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2894 2895 for (tmpva = va; size > 0; ) { 2896 pmap_kenter(tmpva, pa); 2897 size -= PAGE_SIZE; 2898 tmpva += PAGE_SIZE; 2899 pa += PAGE_SIZE; 2900 } 2901 pmap_invalidate_range(kernel_pmap, va, tmpva); 2902 return ((void *)(va + offset)); 2903} 2904 2905void 2906pmap_unmapdev(va, size) 2907 vm_offset_t va; 2908 vm_size_t size; 2909{ 2910 vm_offset_t base, offset, tmpva; 2911 2912 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2913 return; 2914 base = va & PG_FRAME; 2915 offset = va & PAGE_MASK; 2916 size = roundup(offset + size, PAGE_SIZE); 2917 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2918 pmap_kremove(tmpva); 2919 pmap_invalidate_range(kernel_pmap, va, tmpva); 2920 kmem_free(kernel_map, base, size); 2921} 2922 2923/* 2924 * perform the pmap work for mincore 2925 */ 2926int 2927pmap_mincore(pmap, addr) 2928 pmap_t pmap; 2929 vm_offset_t addr; 2930{ 2931 pt_entry_t *ptep, pte; 2932 vm_page_t m; 2933 int val = 0; 2934 2935 PMAP_LOCK(pmap); 2936 ptep = pmap_pte(pmap, addr); 2937 pte = (ptep != NULL) ? *ptep : 0; 2938 pmap_pte_release(ptep); 2939 PMAP_UNLOCK(pmap); 2940 2941 if (pte != 0) { 2942 vm_paddr_t pa; 2943 2944 val = MINCORE_INCORE; 2945 if ((pte & PG_MANAGED) == 0) 2946 return val; 2947 2948 pa = pte & PG_FRAME; 2949 2950 m = PHYS_TO_VM_PAGE(pa); 2951 2952 /* 2953 * Modified by us 2954 */ 2955 if (pte & PG_M) 2956 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2957 else { 2958 /* 2959 * Modified by someone else 2960 */ 2961 vm_page_lock_queues(); 2962 if (m->dirty || pmap_is_modified(m)) 2963 val |= MINCORE_MODIFIED_OTHER; 2964 vm_page_unlock_queues(); 2965 } 2966 /* 2967 * Referenced by us 2968 */ 2969 if (pte & PG_A) 2970 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2971 else { 2972 /* 2973 * Referenced by someone else 2974 */ 2975 vm_page_lock_queues(); 2976 if ((m->flags & PG_REFERENCED) || 2977 pmap_ts_referenced(m)) { 2978 val |= MINCORE_REFERENCED_OTHER; 2979 vm_page_flag_set(m, PG_REFERENCED); 2980 } 2981 vm_page_unlock_queues(); 2982 } 2983 } 2984 return val; 2985} 2986 2987void 2988pmap_activate(struct thread *td) 2989{ 2990 struct proc *p = td->td_proc; 2991 pmap_t pmap, oldpmap; 2992 u_int32_t cr3; 2993 2994 critical_enter(); 2995 pmap = vmspace_pmap(td->td_proc->p_vmspace); 2996 oldpmap = PCPU_GET(curpmap); 2997#if defined(SMP) 2998 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 2999 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 3000#else 3001 oldpmap->pm_active &= ~1; 3002 pmap->pm_active |= 1; 3003#endif 3004#ifdef PAE 3005 cr3 = vtophys(pmap->pm_pdpt); 3006#else 3007 cr3 = vtophys(pmap->pm_pdir); 3008#endif 3009 /* XXXKSE this is wrong. 3010 * pmap_activate is for the current thread on the current cpu 3011 */ 3012 if (p->p_flag & P_SA) { 3013 /* Make sure all other cr3 entries are updated. */ 3014 /* what if they are running? XXXKSE (maybe abort them) */ 3015 FOREACH_THREAD_IN_PROC(p, td) { 3016 td->td_pcb->pcb_cr3 = cr3; 3017 } 3018 } else { 3019 td->td_pcb->pcb_cr3 = cr3; 3020 } 3021 load_cr3(cr3); 3022 PCPU_SET(curpmap, pmap); 3023 critical_exit(); 3024} 3025 3026vm_offset_t 3027pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3028{ 3029 3030 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3031 return addr; 3032 } 3033 3034 addr = (addr + PDRMASK) & ~PDRMASK; 3035 return addr; 3036} 3037 3038 3039#if defined(PMAP_DEBUG) 3040pmap_pid_dump(int pid) 3041{ 3042 pmap_t pmap; 3043 struct proc *p; 3044 int npte = 0; 3045 int index; 3046 3047 sx_slock(&allproc_lock); 3048 LIST_FOREACH(p, &allproc, p_list) { 3049 if (p->p_pid != pid) 3050 continue; 3051 3052 if (p->p_vmspace) { 3053 int i,j; 3054 index = 0; 3055 pmap = vmspace_pmap(p->p_vmspace); 3056 for (i = 0; i < NPDEPTD; i++) { 3057 pd_entry_t *pde; 3058 pt_entry_t *pte; 3059 vm_offset_t base = i << PDRSHIFT; 3060 3061 pde = &pmap->pm_pdir[i]; 3062 if (pde && pmap_pde_v(pde)) { 3063 for (j = 0; j < NPTEPG; j++) { 3064 vm_offset_t va = base + (j << PAGE_SHIFT); 3065 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3066 if (index) { 3067 index = 0; 3068 printf("\n"); 3069 } 3070 sx_sunlock(&allproc_lock); 3071 return npte; 3072 } 3073 pte = pmap_pte(pmap, va); 3074 if (pte && pmap_pte_v(pte)) { 3075 pt_entry_t pa; 3076 vm_page_t m; 3077 pa = *pte; 3078 m = PHYS_TO_VM_PAGE(pa); 3079 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3080 va, pa, m->hold_count, m->wire_count, m->flags); 3081 npte++; 3082 index++; 3083 if (index >= 2) { 3084 index = 0; 3085 printf("\n"); 3086 } else { 3087 printf(" "); 3088 } 3089 } 3090 } 3091 } 3092 } 3093 } 3094 } 3095 sx_sunlock(&allproc_lock); 3096 return npte; 3097} 3098#endif 3099 3100#if defined(DEBUG) 3101 3102static void pads(pmap_t pm); 3103void pmap_pvdump(vm_offset_t pa); 3104 3105/* print address space of pmap*/ 3106static void 3107pads(pm) 3108 pmap_t pm; 3109{ 3110 int i, j; 3111 vm_paddr_t va; 3112 pt_entry_t *ptep; 3113 3114 if (pm == kernel_pmap) 3115 return; 3116 for (i = 0; i < NPDEPTD; i++) 3117 if (pm->pm_pdir[i]) 3118 for (j = 0; j < NPTEPG; j++) { 3119 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3120 if (pm == kernel_pmap && va < KERNBASE) 3121 continue; 3122 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3123 continue; 3124 ptep = pmap_pte(pm, va); 3125 if (pmap_pte_v(ptep)) 3126 printf("%x:%x ", va, *ptep); 3127 }; 3128 3129} 3130 3131void 3132pmap_pvdump(pa) 3133 vm_paddr_t pa; 3134{ 3135 pv_entry_t pv; 3136 vm_page_t m; 3137 3138 printf("pa %x", pa); 3139 m = PHYS_TO_VM_PAGE(pa); 3140 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3141 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3142 pads(pv->pv_pmap); 3143 } 3144 printf(" "); 3145} 3146#endif 3147