pmap.c revision 196707
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) 2005-2008 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 196707 2009-08-31 18:41:13Z jhb $"); 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/ktr.h> 116#include <sys/lock.h> 117#include <sys/malloc.h> 118#include <sys/mman.h> 119#include <sys/msgbuf.h> 120#include <sys/mutex.h> 121#include <sys/proc.h> 122#include <sys/sf_buf.h> 123#include <sys/sx.h> 124#include <sys/vmmeter.h> 125#include <sys/sched.h> 126#include <sys/sysctl.h> 127#ifdef SMP 128#include <sys/smp.h> 129#endif 130 131#include <vm/vm.h> 132#include <vm/vm_param.h> 133#include <vm/vm_kern.h> 134#include <vm/vm_page.h> 135#include <vm/vm_map.h> 136#include <vm/vm_object.h> 137#include <vm/vm_extern.h> 138#include <vm/vm_pageout.h> 139#include <vm/vm_pager.h> 140#include <vm/vm_reserv.h> 141#include <vm/uma.h> 142 143#include <machine/cpu.h> 144#include <machine/cputypes.h> 145#include <machine/md_var.h> 146#include <machine/pcb.h> 147#include <machine/specialreg.h> 148#ifdef SMP 149#include <machine/smp.h> 150#endif 151 152#ifdef XBOX 153#include <machine/xbox.h> 154#endif 155 156#if !defined(CPU_DISABLE_SSE) && defined(I686_CPU) 157#define CPU_ENABLE_SSE 158#endif 159 160#ifndef PMAP_SHPGPERPROC 161#define PMAP_SHPGPERPROC 200 162#endif 163 164#if !defined(DIAGNOSTIC) 165#define PMAP_INLINE __gnu89_inline 166#else 167#define PMAP_INLINE 168#endif 169 170#define PV_STATS 171#ifdef PV_STATS 172#define PV_STAT(x) do { x ; } while (0) 173#else 174#define PV_STAT(x) do { } while (0) 175#endif 176 177#define pa_index(pa) ((pa) >> PDRSHIFT) 178#define pa_to_pvh(pa) (&pv_table[pa_index(pa)]) 179 180/* 181 * Get PDEs and PTEs for user/kernel address space 182 */ 183#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 184#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 185 186#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 187#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 188#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 189#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 190#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 191 192#define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \ 193 atomic_clear_int((u_int *)(pte), PG_W)) 194#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 195 196struct pmap kernel_pmap_store; 197LIST_HEAD(pmaplist, pmap); 198static struct pmaplist allpmaps; 199static struct mtx allpmaps_lock; 200 201vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 202vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 203int pgeflag = 0; /* PG_G or-in */ 204int pseflag = 0; /* PG_PS or-in */ 205 206static int nkpt; 207vm_offset_t kernel_vm_end; 208extern u_int32_t KERNend; 209 210#ifdef PAE 211pt_entry_t pg_nx; 212static uma_zone_t pdptzone; 213#endif 214 215static int pat_works; /* Is page attribute table sane? */ 216 217SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters"); 218 219static int pg_ps_enabled; 220SYSCTL_INT(_vm_pmap, OID_AUTO, pg_ps_enabled, CTLFLAG_RD, &pg_ps_enabled, 0, 221 "Are large page mappings enabled?"); 222 223/* 224 * Data for the pv entry allocation mechanism 225 */ 226static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 227static struct md_page *pv_table; 228static int shpgperproc = PMAP_SHPGPERPROC; 229 230struct pv_chunk *pv_chunkbase; /* KVA block for pv_chunks */ 231int pv_maxchunks; /* How many chunks we have KVA for */ 232vm_offset_t pv_vafree; /* freelist stored in the PTE */ 233 234/* 235 * All those kernel PT submaps that BSD is so fond of 236 */ 237struct sysmaps { 238 struct mtx lock; 239 pt_entry_t *CMAP1; 240 pt_entry_t *CMAP2; 241 caddr_t CADDR1; 242 caddr_t CADDR2; 243}; 244static struct sysmaps sysmaps_pcpu[MAXCPU]; 245pt_entry_t *CMAP1 = 0; 246static pt_entry_t *CMAP3; 247caddr_t CADDR1 = 0, ptvmmap = 0; 248static caddr_t CADDR3; 249struct msgbuf *msgbufp = 0; 250 251/* 252 * Crashdump maps. 253 */ 254static caddr_t crashdumpmap; 255 256static pt_entry_t *PMAP1 = 0, *PMAP2; 257static pt_entry_t *PADDR1 = 0, *PADDR2; 258#ifdef SMP 259static int PMAP1cpu; 260static int PMAP1changedcpu; 261SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD, 262 &PMAP1changedcpu, 0, 263 "Number of times pmap_pte_quick changed CPU with same PMAP1"); 264#endif 265static int PMAP1changed; 266SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD, 267 &PMAP1changed, 0, 268 "Number of times pmap_pte_quick changed PMAP1"); 269static int PMAP1unchanged; 270SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD, 271 &PMAP1unchanged, 0, 272 "Number of times pmap_pte_quick didn't change PMAP1"); 273static struct mtx PMAP2mutex; 274 275static void free_pv_entry(pmap_t pmap, pv_entry_t pv); 276static pv_entry_t get_pv_entry(pmap_t locked_pmap, int try); 277static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa); 278static boolean_t pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa); 279static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa); 280static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va); 281static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, 282 vm_offset_t va); 283static int pmap_pvh_wired_mappings(struct md_page *pvh, int count); 284 285static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va); 286static boolean_t pmap_enter_pde(pmap_t pmap, vm_offset_t va, vm_page_t m, 287 vm_prot_t prot); 288static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, 289 vm_page_t m, vm_prot_t prot, vm_page_t mpte); 290static void pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte); 291static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte); 292static boolean_t pmap_is_modified_pvh(struct md_page *pvh); 293static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode); 294static vm_page_t pmap_lookup_pt_page(pmap_t pmap, vm_offset_t va); 295static void pmap_pde_attr(pd_entry_t *pde, int cache_bits); 296static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va); 297static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, 298 vm_prot_t prot); 299static void pmap_pte_attr(pt_entry_t *pte, int cache_bits); 300static void pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, 301 vm_page_t *free); 302static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva, 303 vm_page_t *free); 304static void pmap_remove_pt_page(pmap_t pmap, vm_page_t mpte); 305static void pmap_remove_page(struct pmap *pmap, vm_offset_t va, 306 vm_page_t *free); 307static void pmap_remove_entry(struct pmap *pmap, vm_page_t m, 308 vm_offset_t va); 309static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); 310static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, 311 vm_page_t m); 312 313static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags); 314 315static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags); 316static int _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, vm_page_t *free); 317static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); 318static void pmap_pte_release(pt_entry_t *pte); 319static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t *); 320static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 321#ifdef PAE 322static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 323#endif 324 325CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 326CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 327 328/* 329 * If you get an error here, then you set KVA_PAGES wrong! See the 330 * description of KVA_PAGES in sys/i386/include/pmap.h. It must be 331 * multiple of 4 for a normal kernel, or a multiple of 8 for a PAE. 332 */ 333CTASSERT(KERNBASE % (1 << 24) == 0); 334 335/* 336 * Move the kernel virtual free pointer to the next 337 * 4MB. This is used to help improve performance 338 * by using a large (4MB) page for much of the kernel 339 * (.text, .data, .bss) 340 */ 341static vm_offset_t 342pmap_kmem_choose(vm_offset_t addr) 343{ 344 vm_offset_t newaddr = addr; 345 346#ifndef DISABLE_PSE 347 if (cpu_feature & CPUID_PSE) 348 newaddr = (addr + PDRMASK) & ~PDRMASK; 349#endif 350 return newaddr; 351} 352 353/* 354 * Bootstrap the system enough to run with virtual memory. 355 * 356 * On the i386 this is called after mapping has already been enabled 357 * and just syncs the pmap module with what has already been done. 358 * [We can't call it easily with mapping off since the kernel is not 359 * mapped with PA == VA, hence we would have to relocate every address 360 * from the linked base (virtual) address "KERNBASE" to the actual 361 * (physical) address starting relative to 0] 362 */ 363void 364pmap_bootstrap(vm_paddr_t firstaddr) 365{ 366 vm_offset_t va; 367 pt_entry_t *pte, *unused; 368 struct sysmaps *sysmaps; 369 int i; 370 371 /* 372 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 373 * large. It should instead be correctly calculated in locore.s and 374 * not based on 'first' (which is a physical address, not a virtual 375 * address, for the start of unused physical memory). The kernel 376 * page tables are NOT double mapped and thus should not be included 377 * in this calculation. 378 */ 379 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 380 virtual_avail = pmap_kmem_choose(virtual_avail); 381 382 virtual_end = VM_MAX_KERNEL_ADDRESS; 383 384 /* 385 * Initialize the kernel pmap (which is statically allocated). 386 */ 387 PMAP_LOCK_INIT(kernel_pmap); 388 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 389#ifdef PAE 390 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 391#endif 392 kernel_pmap->pm_root = NULL; 393 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 394 TAILQ_INIT(&kernel_pmap->pm_pvchunk); 395 LIST_INIT(&allpmaps); 396 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 397 mtx_lock_spin(&allpmaps_lock); 398 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 399 mtx_unlock_spin(&allpmaps_lock); 400 nkpt = NKPT; 401 402 /* 403 * Reserve some special page table entries/VA space for temporary 404 * mapping of pages. 405 */ 406#define SYSMAP(c, p, v, n) \ 407 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 408 409 va = virtual_avail; 410 pte = vtopte(va); 411 412 /* 413 * CMAP1/CMAP2 are used for zeroing and copying pages. 414 * CMAP3 is used for the idle process page zeroing. 415 */ 416 for (i = 0; i < MAXCPU; i++) { 417 sysmaps = &sysmaps_pcpu[i]; 418 mtx_init(&sysmaps->lock, "SYSMAPS", NULL, MTX_DEF); 419 SYSMAP(caddr_t, sysmaps->CMAP1, sysmaps->CADDR1, 1) 420 SYSMAP(caddr_t, sysmaps->CMAP2, sysmaps->CADDR2, 1) 421 } 422 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 423 SYSMAP(caddr_t, CMAP3, CADDR3, 1) 424 *CMAP3 = 0; 425 426 /* 427 * Crashdump maps. 428 */ 429 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS) 430 431 /* 432 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 433 */ 434 SYSMAP(caddr_t, unused, ptvmmap, 1) 435 436 /* 437 * msgbufp is used to map the system message buffer. 438 */ 439 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE))) 440 441 /* 442 * ptemap is used for pmap_pte_quick 443 */ 444 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 445 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1); 446 447 mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF); 448 449 virtual_avail = va; 450 451 *CMAP1 = 0; 452 453 /* 454 * Leave in place an identity mapping (virt == phys) for the low 1 MB 455 * physical memory region that is used by the ACPI wakeup code. This 456 * mapping must not have PG_G set. 457 */ 458#ifdef XBOX 459 /* FIXME: This is gross, but needed for the XBOX. Since we are in such 460 * an early stadium, we cannot yet neatly map video memory ... :-( 461 * Better fixes are very welcome! */ 462 if (!arch_i386_is_xbox) 463#endif 464 for (i = 1; i < NKPT; i++) 465 PTD[i] = 0; 466 467 /* Initialize the PAT MSR if present. */ 468 pmap_init_pat(); 469 470 /* Turn on PG_G on kernel page(s) */ 471 pmap_set_pg(); 472} 473 474/* 475 * Setup the PAT MSR. 476 */ 477void 478pmap_init_pat(void) 479{ 480 uint64_t pat_msr; 481 482 /* Bail if this CPU doesn't implement PAT. */ 483 if (!(cpu_feature & CPUID_PAT)) 484 return; 485 486 if (cpu_vendor_id != CPU_VENDOR_INTEL || 487 (I386_CPU_FAMILY(cpu_id) == 6 && I386_CPU_MODEL(cpu_id) >= 0xe)) { 488 /* 489 * Leave the indices 0-3 at the default of WB, WT, UC, and UC-. 490 * Program 4 and 5 as WP and WC. 491 * Leave 6 and 7 as UC and UC-. 492 */ 493 pat_msr = rdmsr(MSR_PAT); 494 pat_msr &= ~(PAT_MASK(4) | PAT_MASK(5)); 495 pat_msr |= PAT_VALUE(4, PAT_WRITE_PROTECTED) | 496 PAT_VALUE(5, PAT_WRITE_COMBINING); 497 pat_works = 1; 498 } else { 499 /* 500 * Due to some Intel errata, we can only safely use the lower 4 501 * PAT entries. Thus, just replace PAT Index 2 with WC instead 502 * of UC-. 503 * 504 * Intel Pentium III Processor Specification Update 505 * Errata E.27 (Upper Four PAT Entries Not Usable With Mode B 506 * or Mode C Paging) 507 * 508 * Intel Pentium IV Processor Specification Update 509 * Errata N46 (PAT Index MSB May Be Calculated Incorrectly) 510 */ 511 pat_msr = rdmsr(MSR_PAT); 512 pat_msr &= ~PAT_MASK(2); 513 pat_msr |= PAT_VALUE(2, PAT_WRITE_COMBINING); 514 pat_works = 0; 515 } 516 wrmsr(MSR_PAT, pat_msr); 517} 518 519/* 520 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 521 */ 522void 523pmap_set_pg(void) 524{ 525 pd_entry_t pdir; 526 pt_entry_t *pte; 527 vm_offset_t va, endva; 528 int i; 529 530 if (pgeflag == 0) 531 return; 532 533 i = KERNLOAD/NBPDR; 534 endva = KERNBASE + KERNend; 535 536 if (pseflag) { 537 va = KERNBASE + KERNLOAD; 538 while (va < endva) { 539 pdir = kernel_pmap->pm_pdir[KPTDI+i]; 540 pdir |= pgeflag; 541 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 542 invltlb(); /* Play it safe, invltlb() every time */ 543 i++; 544 va += NBPDR; 545 } 546 } else { 547 va = (vm_offset_t)btext; 548 while (va < endva) { 549 pte = vtopte(va); 550 if (*pte) 551 *pte |= pgeflag; 552 invltlb(); /* Play it safe, invltlb() every time */ 553 va += PAGE_SIZE; 554 } 555 } 556} 557 558/* 559 * Initialize a vm_page's machine-dependent fields. 560 */ 561void 562pmap_page_init(vm_page_t m) 563{ 564 565 TAILQ_INIT(&m->md.pv_list); 566 m->md.pat_mode = PAT_WRITE_BACK; 567} 568 569#ifdef PAE 570static void * 571pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 572{ 573 574 /* Inform UMA that this allocator uses kernel_map/object. */ 575 *flags = UMA_SLAB_KERNEL; 576 return ((void *)kmem_alloc_contig(kernel_map, bytes, wait, 0x0ULL, 577 0xffffffffULL, 1, 0, VM_MEMATTR_DEFAULT)); 578} 579#endif 580 581/* 582 * ABuse the pte nodes for unmapped kva to thread a kva freelist through. 583 * Requirements: 584 * - Must deal with pages in order to ensure that none of the PG_* bits 585 * are ever set, PG_V in particular. 586 * - Assumes we can write to ptes without pte_store() atomic ops, even 587 * on PAE systems. This should be ok. 588 * - Assumes nothing will ever test these addresses for 0 to indicate 589 * no mapping instead of correctly checking PG_V. 590 * - Assumes a vm_offset_t will fit in a pte (true for i386). 591 * Because PG_V is never set, there can be no mappings to invalidate. 592 */ 593static vm_offset_t 594pmap_ptelist_alloc(vm_offset_t *head) 595{ 596 pt_entry_t *pte; 597 vm_offset_t va; 598 599 va = *head; 600 if (va == 0) 601 return (va); /* Out of memory */ 602 pte = vtopte(va); 603 *head = *pte; 604 if (*head & PG_V) 605 panic("pmap_ptelist_alloc: va with PG_V set!"); 606 *pte = 0; 607 return (va); 608} 609 610static void 611pmap_ptelist_free(vm_offset_t *head, vm_offset_t va) 612{ 613 pt_entry_t *pte; 614 615 if (va & PG_V) 616 panic("pmap_ptelist_free: freeing va with PG_V set!"); 617 pte = vtopte(va); 618 *pte = *head; /* virtual! PG_V is 0 though */ 619 *head = va; 620} 621 622static void 623pmap_ptelist_init(vm_offset_t *head, void *base, int npages) 624{ 625 int i; 626 vm_offset_t va; 627 628 *head = 0; 629 for (i = npages - 1; i >= 0; i--) { 630 va = (vm_offset_t)base + i * PAGE_SIZE; 631 pmap_ptelist_free(head, va); 632 } 633} 634 635 636/* 637 * Initialize the pmap module. 638 * Called by vm_init, to initialize any structures that the pmap 639 * system needs to map virtual memory. 640 */ 641void 642pmap_init(void) 643{ 644 vm_page_t mpte; 645 vm_size_t s; 646 int i, pv_npg; 647 648 /* 649 * Initialize the vm page array entries for the kernel pmap's 650 * page table pages. 651 */ 652 for (i = 0; i < nkpt; i++) { 653 mpte = PHYS_TO_VM_PAGE(PTD[i + KPTDI] & PG_FRAME); 654 KASSERT(mpte >= vm_page_array && 655 mpte < &vm_page_array[vm_page_array_size], 656 ("pmap_init: page table page is out of range")); 657 mpte->pindex = i + KPTDI; 658 mpte->phys_addr = PTD[i + KPTDI] & PG_FRAME; 659 } 660 661 /* 662 * Initialize the address space (zone) for the pv entries. Set a 663 * high water mark so that the system can recover from excessive 664 * numbers of pv entries. 665 */ 666 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 667 pv_entry_max = shpgperproc * maxproc + cnt.v_page_count; 668 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 669 pv_entry_max = roundup(pv_entry_max, _NPCPV); 670 pv_entry_high_water = 9 * (pv_entry_max / 10); 671 672 /* 673 * Are large page mappings enabled? 674 */ 675 TUNABLE_INT_FETCH("vm.pmap.pg_ps_enabled", &pg_ps_enabled); 676 677 /* 678 * Calculate the size of the pv head table for superpages. 679 */ 680 for (i = 0; phys_avail[i + 1]; i += 2); 681 pv_npg = round_4mpage(phys_avail[(i - 2) + 1]) / NBPDR; 682 683 /* 684 * Allocate memory for the pv head table for superpages. 685 */ 686 s = (vm_size_t)(pv_npg * sizeof(struct md_page)); 687 s = round_page(s); 688 pv_table = (struct md_page *)kmem_alloc(kernel_map, s); 689 for (i = 0; i < pv_npg; i++) 690 TAILQ_INIT(&pv_table[i].pv_list); 691 692 pv_maxchunks = MAX(pv_entry_max / _NPCPV, maxproc); 693 pv_chunkbase = (struct pv_chunk *)kmem_alloc_nofault(kernel_map, 694 PAGE_SIZE * pv_maxchunks); 695 if (pv_chunkbase == NULL) 696 panic("pmap_init: not enough kvm for pv chunks"); 697 pmap_ptelist_init(&pv_vafree, pv_chunkbase, pv_maxchunks); 698#ifdef PAE 699 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 700 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 701 UMA_ZONE_VM | UMA_ZONE_NOFREE); 702 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 703#endif 704} 705 706 707SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_max, CTLFLAG_RD, &pv_entry_max, 0, 708 "Max number of PV entries"); 709SYSCTL_INT(_vm_pmap, OID_AUTO, shpgperproc, CTLFLAG_RD, &shpgperproc, 0, 710 "Page share factor per proc"); 711 712SYSCTL_NODE(_vm_pmap, OID_AUTO, pde, CTLFLAG_RD, 0, 713 "2/4MB page mapping counters"); 714 715static u_long pmap_pde_demotions; 716SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, demotions, CTLFLAG_RD, 717 &pmap_pde_demotions, 0, "2/4MB page demotions"); 718 719static u_long pmap_pde_mappings; 720SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, mappings, CTLFLAG_RD, 721 &pmap_pde_mappings, 0, "2/4MB page mappings"); 722 723static u_long pmap_pde_p_failures; 724SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, p_failures, CTLFLAG_RD, 725 &pmap_pde_p_failures, 0, "2/4MB page promotion failures"); 726 727static u_long pmap_pde_promotions; 728SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, promotions, CTLFLAG_RD, 729 &pmap_pde_promotions, 0, "2/4MB page promotions"); 730 731/*************************************************** 732 * Low level helper routines..... 733 ***************************************************/ 734 735/* 736 * Determine the appropriate bits to set in a PTE or PDE for a specified 737 * caching mode. 738 */ 739int 740pmap_cache_bits(int mode, boolean_t is_pde) 741{ 742 int pat_flag, pat_index, cache_bits; 743 744 /* The PAT bit is different for PTE's and PDE's. */ 745 pat_flag = is_pde ? PG_PDE_PAT : PG_PTE_PAT; 746 747 /* If we don't support PAT, map extended modes to older ones. */ 748 if (!(cpu_feature & CPUID_PAT)) { 749 switch (mode) { 750 case PAT_UNCACHEABLE: 751 case PAT_WRITE_THROUGH: 752 case PAT_WRITE_BACK: 753 break; 754 case PAT_UNCACHED: 755 case PAT_WRITE_COMBINING: 756 case PAT_WRITE_PROTECTED: 757 mode = PAT_UNCACHEABLE; 758 break; 759 } 760 } 761 762 /* Map the caching mode to a PAT index. */ 763 if (pat_works) { 764 switch (mode) { 765 case PAT_UNCACHEABLE: 766 pat_index = 3; 767 break; 768 case PAT_WRITE_THROUGH: 769 pat_index = 1; 770 break; 771 case PAT_WRITE_BACK: 772 pat_index = 0; 773 break; 774 case PAT_UNCACHED: 775 pat_index = 2; 776 break; 777 case PAT_WRITE_COMBINING: 778 pat_index = 5; 779 break; 780 case PAT_WRITE_PROTECTED: 781 pat_index = 4; 782 break; 783 default: 784 panic("Unknown caching mode %d\n", mode); 785 } 786 } else { 787 switch (mode) { 788 case PAT_UNCACHED: 789 case PAT_UNCACHEABLE: 790 case PAT_WRITE_PROTECTED: 791 pat_index = 3; 792 break; 793 case PAT_WRITE_THROUGH: 794 pat_index = 1; 795 break; 796 case PAT_WRITE_BACK: 797 pat_index = 0; 798 break; 799 case PAT_WRITE_COMBINING: 800 pat_index = 2; 801 break; 802 default: 803 panic("Unknown caching mode %d\n", mode); 804 } 805 } 806 807 /* Map the 3-bit index value into the PAT, PCD, and PWT bits. */ 808 cache_bits = 0; 809 if (pat_index & 0x4) 810 cache_bits |= pat_flag; 811 if (pat_index & 0x2) 812 cache_bits |= PG_NC_PCD; 813 if (pat_index & 0x1) 814 cache_bits |= PG_NC_PWT; 815 return (cache_bits); 816} 817#ifdef SMP 818/* 819 * For SMP, these functions have to use the IPI mechanism for coherence. 820 * 821 * N.B.: Before calling any of the following TLB invalidation functions, 822 * the calling processor must ensure that all stores updating a non- 823 * kernel page table are globally performed. Otherwise, another 824 * processor could cache an old, pre-update entry without being 825 * invalidated. This can happen one of two ways: (1) The pmap becomes 826 * active on another processor after its pm_active field is checked by 827 * one of the following functions but before a store updating the page 828 * table is globally performed. (2) The pmap becomes active on another 829 * processor before its pm_active field is checked but due to 830 * speculative loads one of the following functions stills reads the 831 * pmap as inactive on the other processor. 832 * 833 * The kernel page table is exempt because its pm_active field is 834 * immutable. The kernel page table is always active on every 835 * processor. 836 */ 837void 838pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 839{ 840 u_int cpumask; 841 u_int other_cpus; 842 843 sched_pin(); 844 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 845 invlpg(va); 846 smp_invlpg(va); 847 } else { 848 cpumask = PCPU_GET(cpumask); 849 other_cpus = PCPU_GET(other_cpus); 850 if (pmap->pm_active & cpumask) 851 invlpg(va); 852 if (pmap->pm_active & other_cpus) 853 smp_masked_invlpg(pmap->pm_active & other_cpus, va); 854 } 855 sched_unpin(); 856} 857 858void 859pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 860{ 861 u_int cpumask; 862 u_int other_cpus; 863 vm_offset_t addr; 864 865 sched_pin(); 866 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 867 for (addr = sva; addr < eva; addr += PAGE_SIZE) 868 invlpg(addr); 869 smp_invlpg_range(sva, eva); 870 } else { 871 cpumask = PCPU_GET(cpumask); 872 other_cpus = PCPU_GET(other_cpus); 873 if (pmap->pm_active & cpumask) 874 for (addr = sva; addr < eva; addr += PAGE_SIZE) 875 invlpg(addr); 876 if (pmap->pm_active & other_cpus) 877 smp_masked_invlpg_range(pmap->pm_active & other_cpus, 878 sva, eva); 879 } 880 sched_unpin(); 881} 882 883void 884pmap_invalidate_all(pmap_t pmap) 885{ 886 u_int cpumask; 887 u_int other_cpus; 888 889 sched_pin(); 890 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 891 invltlb(); 892 smp_invltlb(); 893 } else { 894 cpumask = PCPU_GET(cpumask); 895 other_cpus = PCPU_GET(other_cpus); 896 if (pmap->pm_active & cpumask) 897 invltlb(); 898 if (pmap->pm_active & other_cpus) 899 smp_masked_invltlb(pmap->pm_active & other_cpus); 900 } 901 sched_unpin(); 902} 903 904void 905pmap_invalidate_cache(void) 906{ 907 908 sched_pin(); 909 wbinvd(); 910 smp_cache_flush(); 911 sched_unpin(); 912} 913#else /* !SMP */ 914/* 915 * Normal, non-SMP, 486+ invalidation functions. 916 * We inline these within pmap.c for speed. 917 */ 918PMAP_INLINE void 919pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 920{ 921 922 if (pmap == kernel_pmap || pmap->pm_active) 923 invlpg(va); 924} 925 926PMAP_INLINE void 927pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 928{ 929 vm_offset_t addr; 930 931 if (pmap == kernel_pmap || pmap->pm_active) 932 for (addr = sva; addr < eva; addr += PAGE_SIZE) 933 invlpg(addr); 934} 935 936PMAP_INLINE void 937pmap_invalidate_all(pmap_t pmap) 938{ 939 940 if (pmap == kernel_pmap || pmap->pm_active) 941 invltlb(); 942} 943 944PMAP_INLINE void 945pmap_invalidate_cache(void) 946{ 947 948 wbinvd(); 949} 950#endif /* !SMP */ 951 952void 953pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva) 954{ 955 956 KASSERT((sva & PAGE_MASK) == 0, 957 ("pmap_invalidate_cache_range: sva not page-aligned")); 958 KASSERT((eva & PAGE_MASK) == 0, 959 ("pmap_invalidate_cache_range: eva not page-aligned")); 960 961 if (cpu_feature & CPUID_SS) 962 ; /* If "Self Snoop" is supported, do nothing. */ 963 else if (cpu_feature & CPUID_CLFSH) { 964 965 /* 966 * Otherwise, do per-cache line flush. Use the mfence 967 * instruction to insure that previous stores are 968 * included in the write-back. The processor 969 * propagates flush to other processors in the cache 970 * coherence domain. 971 */ 972 mfence(); 973 for (; sva < eva; sva += cpu_clflush_line_size) 974 clflush(sva); 975 mfence(); 976 } else { 977 978 /* 979 * No targeted cache flush methods are supported by CPU, 980 * globally invalidate cache as a last resort. 981 */ 982 pmap_invalidate_cache(); 983 } 984} 985 986/* 987 * Are we current address space or kernel? N.B. We return FALSE when 988 * a pmap's page table is in use because a kernel thread is borrowing 989 * it. The borrowed page table can change spontaneously, making any 990 * dependence on its continued use subject to a race condition. 991 */ 992static __inline int 993pmap_is_current(pmap_t pmap) 994{ 995 996 return (pmap == kernel_pmap || 997 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 998 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 999} 1000 1001/* 1002 * If the given pmap is not the current or kernel pmap, the returned pte must 1003 * be released by passing it to pmap_pte_release(). 1004 */ 1005pt_entry_t * 1006pmap_pte(pmap_t pmap, vm_offset_t va) 1007{ 1008 pd_entry_t newpf; 1009 pd_entry_t *pde; 1010 1011 pde = pmap_pde(pmap, va); 1012 if (*pde & PG_PS) 1013 return (pde); 1014 if (*pde != 0) { 1015 /* are we current address space or kernel? */ 1016 if (pmap_is_current(pmap)) 1017 return (vtopte(va)); 1018 mtx_lock(&PMAP2mutex); 1019 newpf = *pde & PG_FRAME; 1020 if ((*PMAP2 & PG_FRAME) != newpf) { 1021 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; 1022 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 1023 } 1024 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); 1025 } 1026 return (0); 1027} 1028 1029/* 1030 * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte 1031 * being NULL. 1032 */ 1033static __inline void 1034pmap_pte_release(pt_entry_t *pte) 1035{ 1036 1037 if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2) 1038 mtx_unlock(&PMAP2mutex); 1039} 1040 1041static __inline void 1042invlcaddr(void *caddr) 1043{ 1044 1045 invlpg((u_int)caddr); 1046} 1047 1048/* 1049 * Super fast pmap_pte routine best used when scanning 1050 * the pv lists. This eliminates many coarse-grained 1051 * invltlb calls. Note that many of the pv list 1052 * scans are across different pmaps. It is very wasteful 1053 * to do an entire invltlb for checking a single mapping. 1054 * 1055 * If the given pmap is not the current pmap, vm_page_queue_mtx 1056 * must be held and curthread pinned to a CPU. 1057 */ 1058static pt_entry_t * 1059pmap_pte_quick(pmap_t pmap, vm_offset_t va) 1060{ 1061 pd_entry_t newpf; 1062 pd_entry_t *pde; 1063 1064 pde = pmap_pde(pmap, va); 1065 if (*pde & PG_PS) 1066 return (pde); 1067 if (*pde != 0) { 1068 /* are we current address space or kernel? */ 1069 if (pmap_is_current(pmap)) 1070 return (vtopte(va)); 1071 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1072 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 1073 newpf = *pde & PG_FRAME; 1074 if ((*PMAP1 & PG_FRAME) != newpf) { 1075 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; 1076#ifdef SMP 1077 PMAP1cpu = PCPU_GET(cpuid); 1078#endif 1079 invlcaddr(PADDR1); 1080 PMAP1changed++; 1081 } else 1082#ifdef SMP 1083 if (PMAP1cpu != PCPU_GET(cpuid)) { 1084 PMAP1cpu = PCPU_GET(cpuid); 1085 invlcaddr(PADDR1); 1086 PMAP1changedcpu++; 1087 } else 1088#endif 1089 PMAP1unchanged++; 1090 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); 1091 } 1092 return (0); 1093} 1094 1095/* 1096 * Routine: pmap_extract 1097 * Function: 1098 * Extract the physical page address associated 1099 * with the given map/virtual_address pair. 1100 */ 1101vm_paddr_t 1102pmap_extract(pmap_t pmap, vm_offset_t va) 1103{ 1104 vm_paddr_t rtval; 1105 pt_entry_t *pte; 1106 pd_entry_t pde; 1107 1108 rtval = 0; 1109 PMAP_LOCK(pmap); 1110 pde = pmap->pm_pdir[va >> PDRSHIFT]; 1111 if (pde != 0) { 1112 if ((pde & PG_PS) != 0) 1113 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK); 1114 else { 1115 pte = pmap_pte(pmap, va); 1116 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK); 1117 pmap_pte_release(pte); 1118 } 1119 } 1120 PMAP_UNLOCK(pmap); 1121 return (rtval); 1122} 1123 1124/* 1125 * Routine: pmap_extract_and_hold 1126 * Function: 1127 * Atomically extract and hold the physical page 1128 * with the given pmap and virtual address pair 1129 * if that mapping permits the given protection. 1130 */ 1131vm_page_t 1132pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 1133{ 1134 pd_entry_t pde; 1135 pt_entry_t pte; 1136 vm_page_t m; 1137 1138 m = NULL; 1139 vm_page_lock_queues(); 1140 PMAP_LOCK(pmap); 1141 pde = *pmap_pde(pmap, va); 1142 if (pde != 0) { 1143 if (pde & PG_PS) { 1144 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) { 1145 m = PHYS_TO_VM_PAGE((pde & PG_PS_FRAME) | 1146 (va & PDRMASK)); 1147 vm_page_hold(m); 1148 } 1149 } else { 1150 sched_pin(); 1151 pte = *pmap_pte_quick(pmap, va); 1152 if (pte != 0 && 1153 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) { 1154 m = PHYS_TO_VM_PAGE(pte & PG_FRAME); 1155 vm_page_hold(m); 1156 } 1157 sched_unpin(); 1158 } 1159 } 1160 vm_page_unlock_queues(); 1161 PMAP_UNLOCK(pmap); 1162 return (m); 1163} 1164 1165/*************************************************** 1166 * Low level mapping routines..... 1167 ***************************************************/ 1168 1169/* 1170 * Add a wired page to the kva. 1171 * Note: not SMP coherent. 1172 */ 1173PMAP_INLINE void 1174pmap_kenter(vm_offset_t va, vm_paddr_t pa) 1175{ 1176 pt_entry_t *pte; 1177 1178 pte = vtopte(va); 1179 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 1180} 1181 1182static __inline void 1183pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode) 1184{ 1185 pt_entry_t *pte; 1186 1187 pte = vtopte(va); 1188 pte_store(pte, pa | PG_RW | PG_V | pgeflag | pmap_cache_bits(mode, 0)); 1189} 1190 1191/* 1192 * Remove a page from the kernel pagetables. 1193 * Note: not SMP coherent. 1194 */ 1195PMAP_INLINE void 1196pmap_kremove(vm_offset_t va) 1197{ 1198 pt_entry_t *pte; 1199 1200 pte = vtopte(va); 1201 pte_clear(pte); 1202} 1203 1204/* 1205 * Used to map a range of physical addresses into kernel 1206 * virtual address space. 1207 * 1208 * The value passed in '*virt' is a suggested virtual address for 1209 * the mapping. Architectures which can support a direct-mapped 1210 * physical to virtual region can return the appropriate address 1211 * within that region, leaving '*virt' unchanged. Other 1212 * architectures should map the pages starting at '*virt' and 1213 * update '*virt' with the first usable address after the mapped 1214 * region. 1215 */ 1216vm_offset_t 1217pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 1218{ 1219 vm_offset_t va, sva; 1220 1221 va = sva = *virt; 1222 while (start < end) { 1223 pmap_kenter(va, start); 1224 va += PAGE_SIZE; 1225 start += PAGE_SIZE; 1226 } 1227 pmap_invalidate_range(kernel_pmap, sva, va); 1228 *virt = va; 1229 return (sva); 1230} 1231 1232 1233/* 1234 * Add a list of wired pages to the kva 1235 * this routine is only used for temporary 1236 * kernel mappings that do not need to have 1237 * page modification or references recorded. 1238 * Note that old mappings are simply written 1239 * over. The page *must* be wired. 1240 * Note: SMP coherent. Uses a ranged shootdown IPI. 1241 */ 1242void 1243pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count) 1244{ 1245 pt_entry_t *endpte, oldpte, *pte; 1246 1247 oldpte = 0; 1248 pte = vtopte(sva); 1249 endpte = pte + count; 1250 while (pte < endpte) { 1251 oldpte |= *pte; 1252 pte_store(pte, VM_PAGE_TO_PHYS(*ma) | pgeflag | 1253 pmap_cache_bits((*ma)->md.pat_mode, 0) | PG_RW | PG_V); 1254 pte++; 1255 ma++; 1256 } 1257 if ((oldpte & PG_V) != 0) 1258 pmap_invalidate_range(kernel_pmap, sva, sva + count * 1259 PAGE_SIZE); 1260} 1261 1262/* 1263 * This routine tears out page mappings from the 1264 * kernel -- it is meant only for temporary mappings. 1265 * Note: SMP coherent. Uses a ranged shootdown IPI. 1266 */ 1267void 1268pmap_qremove(vm_offset_t sva, int count) 1269{ 1270 vm_offset_t va; 1271 1272 va = sva; 1273 while (count-- > 0) { 1274 pmap_kremove(va); 1275 va += PAGE_SIZE; 1276 } 1277 pmap_invalidate_range(kernel_pmap, sva, va); 1278} 1279 1280/*************************************************** 1281 * Page table page management routines..... 1282 ***************************************************/ 1283static __inline void 1284pmap_free_zero_pages(vm_page_t free) 1285{ 1286 vm_page_t m; 1287 1288 while (free != NULL) { 1289 m = free; 1290 free = m->right; 1291 /* Preserve the page's PG_ZERO setting. */ 1292 vm_page_free_toq(m); 1293 } 1294} 1295 1296/* 1297 * Schedule the specified unused page table page to be freed. Specifically, 1298 * add the page to the specified list of pages that will be released to the 1299 * physical memory manager after the TLB has been updated. 1300 */ 1301static __inline void 1302pmap_add_delayed_free_list(vm_page_t m, vm_page_t *free, boolean_t set_PG_ZERO) 1303{ 1304 1305 if (set_PG_ZERO) 1306 m->flags |= PG_ZERO; 1307 else 1308 m->flags &= ~PG_ZERO; 1309 m->right = *free; 1310 *free = m; 1311} 1312 1313/* 1314 * Inserts the specified page table page into the specified pmap's collection 1315 * of idle page table pages. Each of a pmap's page table pages is responsible 1316 * for mapping a distinct range of virtual addresses. The pmap's collection is 1317 * ordered by this virtual address range. 1318 */ 1319static void 1320pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte) 1321{ 1322 vm_page_t root; 1323 1324 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1325 root = pmap->pm_root; 1326 if (root == NULL) { 1327 mpte->left = NULL; 1328 mpte->right = NULL; 1329 } else { 1330 root = vm_page_splay(mpte->pindex, root); 1331 if (mpte->pindex < root->pindex) { 1332 mpte->left = root->left; 1333 mpte->right = root; 1334 root->left = NULL; 1335 } else if (mpte->pindex == root->pindex) 1336 panic("pmap_insert_pt_page: pindex already inserted"); 1337 else { 1338 mpte->right = root->right; 1339 mpte->left = root; 1340 root->right = NULL; 1341 } 1342 } 1343 pmap->pm_root = mpte; 1344} 1345 1346/* 1347 * Looks for a page table page mapping the specified virtual address in the 1348 * specified pmap's collection of idle page table pages. Returns NULL if there 1349 * is no page table page corresponding to the specified virtual address. 1350 */ 1351static vm_page_t 1352pmap_lookup_pt_page(pmap_t pmap, vm_offset_t va) 1353{ 1354 vm_page_t mpte; 1355 vm_pindex_t pindex = va >> PDRSHIFT; 1356 1357 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1358 if ((mpte = pmap->pm_root) != NULL && mpte->pindex != pindex) { 1359 mpte = vm_page_splay(pindex, mpte); 1360 if ((pmap->pm_root = mpte)->pindex != pindex) 1361 mpte = NULL; 1362 } 1363 return (mpte); 1364} 1365 1366/* 1367 * Removes the specified page table page from the specified pmap's collection 1368 * of idle page table pages. The specified page table page must be a member of 1369 * the pmap's collection. 1370 */ 1371static void 1372pmap_remove_pt_page(pmap_t pmap, vm_page_t mpte) 1373{ 1374 vm_page_t root; 1375 1376 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1377 if (mpte != pmap->pm_root) 1378 vm_page_splay(mpte->pindex, pmap->pm_root); 1379 if (mpte->left == NULL) 1380 root = mpte->right; 1381 else { 1382 root = vm_page_splay(mpte->pindex, mpte->left); 1383 root->right = mpte->right; 1384 } 1385 pmap->pm_root = root; 1386} 1387 1388/* 1389 * This routine unholds page table pages, and if the hold count 1390 * drops to zero, then it decrements the wire count. 1391 */ 1392static __inline int 1393pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, vm_page_t *free) 1394{ 1395 1396 --m->wire_count; 1397 if (m->wire_count == 0) 1398 return _pmap_unwire_pte_hold(pmap, m, free); 1399 else 1400 return 0; 1401} 1402 1403static int 1404_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, vm_page_t *free) 1405{ 1406 vm_offset_t pteva; 1407 1408 /* 1409 * unmap the page table page 1410 */ 1411 pmap->pm_pdir[m->pindex] = 0; 1412 --pmap->pm_stats.resident_count; 1413 1414 /* 1415 * This is a release store so that the ordinary store unmapping 1416 * the page table page is globally performed before TLB shoot- 1417 * down is begun. 1418 */ 1419 atomic_subtract_rel_int(&cnt.v_wire_count, 1); 1420 1421 /* 1422 * Do an invltlb to make the invalidated mapping 1423 * take effect immediately. 1424 */ 1425 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1426 pmap_invalidate_page(pmap, pteva); 1427 1428 /* 1429 * Put page on a list so that it is released after 1430 * *ALL* TLB shootdown is done 1431 */ 1432 pmap_add_delayed_free_list(m, free, TRUE); 1433 1434 return 1; 1435} 1436 1437/* 1438 * After removing a page table entry, this routine is used to 1439 * conditionally free the page, and manage the hold/wire counts. 1440 */ 1441static int 1442pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t *free) 1443{ 1444 pd_entry_t ptepde; 1445 vm_page_t mpte; 1446 1447 if (va >= VM_MAXUSER_ADDRESS) 1448 return 0; 1449 ptepde = *pmap_pde(pmap, va); 1450 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); 1451 return pmap_unwire_pte_hold(pmap, mpte, free); 1452} 1453 1454void 1455pmap_pinit0(pmap_t pmap) 1456{ 1457 1458 PMAP_LOCK_INIT(pmap); 1459 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 1460#ifdef PAE 1461 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 1462#endif 1463 pmap->pm_root = NULL; 1464 pmap->pm_active = 0; 1465 PCPU_SET(curpmap, pmap); 1466 TAILQ_INIT(&pmap->pm_pvchunk); 1467 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1468 mtx_lock_spin(&allpmaps_lock); 1469 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1470 mtx_unlock_spin(&allpmaps_lock); 1471} 1472 1473/* 1474 * Initialize a preallocated and zeroed pmap structure, 1475 * such as one in a vmspace structure. 1476 */ 1477int 1478pmap_pinit(pmap_t pmap) 1479{ 1480 vm_page_t m, ptdpg[NPGPTD]; 1481 vm_paddr_t pa; 1482 static int color; 1483 int i; 1484 1485 PMAP_LOCK_INIT(pmap); 1486 1487 /* 1488 * No need to allocate page table space yet but we do need a valid 1489 * page directory table. 1490 */ 1491 if (pmap->pm_pdir == NULL) { 1492 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1493 NBPTD); 1494 1495 if (pmap->pm_pdir == NULL) { 1496 PMAP_LOCK_DESTROY(pmap); 1497 return (0); 1498 } 1499#ifdef PAE 1500 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1501 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1502 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1503 ("pmap_pinit: pdpt misaligned")); 1504 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1505 ("pmap_pinit: pdpt above 4g")); 1506#endif 1507 pmap->pm_root = NULL; 1508 } 1509 KASSERT(pmap->pm_root == NULL, 1510 ("pmap_pinit: pmap has reserved page table page(s)")); 1511 1512 /* 1513 * allocate the page directory page(s) 1514 */ 1515 for (i = 0; i < NPGPTD;) { 1516 m = vm_page_alloc(NULL, color++, 1517 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1518 VM_ALLOC_ZERO); 1519 if (m == NULL) 1520 VM_WAIT; 1521 else { 1522 ptdpg[i++] = m; 1523 } 1524 } 1525 1526 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1527 1528 for (i = 0; i < NPGPTD; i++) { 1529 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1530 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1531 } 1532 1533 mtx_lock_spin(&allpmaps_lock); 1534 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1535 mtx_unlock_spin(&allpmaps_lock); 1536 /* Wire in kernel global address entries. */ 1537 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1538 1539 /* install self-referential address mapping entry(s) */ 1540 for (i = 0; i < NPGPTD; i++) { 1541 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1542 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1543#ifdef PAE 1544 pmap->pm_pdpt[i] = pa | PG_V; 1545#endif 1546 } 1547 1548 pmap->pm_active = 0; 1549 TAILQ_INIT(&pmap->pm_pvchunk); 1550 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1551 1552 return (1); 1553} 1554 1555/* 1556 * this routine is called if the page table page is not 1557 * mapped correctly. 1558 */ 1559static vm_page_t 1560_pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags) 1561{ 1562 vm_paddr_t ptepa; 1563 vm_page_t m; 1564 1565 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT || 1566 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK, 1567 ("_pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK")); 1568 1569 /* 1570 * Allocate a page table page. 1571 */ 1572 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1573 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1574 if (flags & M_WAITOK) { 1575 PMAP_UNLOCK(pmap); 1576 vm_page_unlock_queues(); 1577 VM_WAIT; 1578 vm_page_lock_queues(); 1579 PMAP_LOCK(pmap); 1580 } 1581 1582 /* 1583 * Indicate the need to retry. While waiting, the page table 1584 * page may have been allocated. 1585 */ 1586 return (NULL); 1587 } 1588 if ((m->flags & PG_ZERO) == 0) 1589 pmap_zero_page(m); 1590 1591 /* 1592 * Map the pagetable page into the process address space, if 1593 * it isn't already there. 1594 */ 1595 1596 pmap->pm_stats.resident_count++; 1597 1598 ptepa = VM_PAGE_TO_PHYS(m); 1599 pmap->pm_pdir[ptepindex] = 1600 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1601 1602 return m; 1603} 1604 1605static vm_page_t 1606pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags) 1607{ 1608 unsigned ptepindex; 1609 pd_entry_t ptepa; 1610 vm_page_t m; 1611 1612 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT || 1613 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK, 1614 ("pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK")); 1615 1616 /* 1617 * Calculate pagetable page index 1618 */ 1619 ptepindex = va >> PDRSHIFT; 1620retry: 1621 /* 1622 * Get the page directory entry 1623 */ 1624 ptepa = pmap->pm_pdir[ptepindex]; 1625 1626 /* 1627 * This supports switching from a 4MB page to a 1628 * normal 4K page. 1629 */ 1630 if (ptepa & PG_PS) { 1631 (void)pmap_demote_pde(pmap, &pmap->pm_pdir[ptepindex], va); 1632 ptepa = pmap->pm_pdir[ptepindex]; 1633 } 1634 1635 /* 1636 * If the page table page is mapped, we just increment the 1637 * hold count, and activate it. 1638 */ 1639 if (ptepa) { 1640 m = PHYS_TO_VM_PAGE(ptepa & PG_FRAME); 1641 m->wire_count++; 1642 } else { 1643 /* 1644 * Here if the pte page isn't mapped, or if it has 1645 * been deallocated. 1646 */ 1647 m = _pmap_allocpte(pmap, ptepindex, flags); 1648 if (m == NULL && (flags & M_WAITOK)) 1649 goto retry; 1650 } 1651 return (m); 1652} 1653 1654 1655/*************************************************** 1656* Pmap allocation/deallocation routines. 1657 ***************************************************/ 1658 1659#ifdef SMP 1660/* 1661 * Deal with a SMP shootdown of other users of the pmap that we are 1662 * trying to dispose of. This can be a bit hairy. 1663 */ 1664static cpumask_t *lazymask; 1665static u_int lazyptd; 1666static volatile u_int lazywait; 1667 1668void pmap_lazyfix_action(void); 1669 1670void 1671pmap_lazyfix_action(void) 1672{ 1673 cpumask_t mymask = PCPU_GET(cpumask); 1674 1675#ifdef COUNT_IPIS 1676 (*ipi_lazypmap_counts[PCPU_GET(cpuid)])++; 1677#endif 1678 if (rcr3() == lazyptd) 1679 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1680 atomic_clear_int(lazymask, mymask); 1681 atomic_store_rel_int(&lazywait, 1); 1682} 1683 1684static void 1685pmap_lazyfix_self(cpumask_t mymask) 1686{ 1687 1688 if (rcr3() == lazyptd) 1689 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1690 atomic_clear_int(lazymask, mymask); 1691} 1692 1693 1694static void 1695pmap_lazyfix(pmap_t pmap) 1696{ 1697 cpumask_t mymask, mask; 1698 u_int spins; 1699 1700 while ((mask = pmap->pm_active) != 0) { 1701 spins = 50000000; 1702 mask = mask & -mask; /* Find least significant set bit */ 1703 mtx_lock_spin(&smp_ipi_mtx); 1704#ifdef PAE 1705 lazyptd = vtophys(pmap->pm_pdpt); 1706#else 1707 lazyptd = vtophys(pmap->pm_pdir); 1708#endif 1709 mymask = PCPU_GET(cpumask); 1710 if (mask == mymask) { 1711 lazymask = &pmap->pm_active; 1712 pmap_lazyfix_self(mymask); 1713 } else { 1714 atomic_store_rel_int((u_int *)&lazymask, 1715 (u_int)&pmap->pm_active); 1716 atomic_store_rel_int(&lazywait, 0); 1717 ipi_selected(mask, IPI_LAZYPMAP); 1718 while (lazywait == 0) { 1719 ia32_pause(); 1720 if (--spins == 0) 1721 break; 1722 } 1723 } 1724 mtx_unlock_spin(&smp_ipi_mtx); 1725 if (spins == 0) 1726 printf("pmap_lazyfix: spun for 50000000\n"); 1727 } 1728} 1729 1730#else /* SMP */ 1731 1732/* 1733 * Cleaning up on uniprocessor is easy. For various reasons, we're 1734 * unlikely to have to even execute this code, including the fact 1735 * that the cleanup is deferred until the parent does a wait(2), which 1736 * means that another userland process has run. 1737 */ 1738static void 1739pmap_lazyfix(pmap_t pmap) 1740{ 1741 u_int cr3; 1742 1743 cr3 = vtophys(pmap->pm_pdir); 1744 if (cr3 == rcr3()) { 1745 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1746 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1747 } 1748} 1749#endif /* SMP */ 1750 1751/* 1752 * Release any resources held by the given physical map. 1753 * Called when a pmap initialized by pmap_pinit is being released. 1754 * Should only be called if the map contains no valid mappings. 1755 */ 1756void 1757pmap_release(pmap_t pmap) 1758{ 1759 vm_page_t m, ptdpg[NPGPTD]; 1760 int i; 1761 1762 KASSERT(pmap->pm_stats.resident_count == 0, 1763 ("pmap_release: pmap resident count %ld != 0", 1764 pmap->pm_stats.resident_count)); 1765 KASSERT(pmap->pm_root == NULL, 1766 ("pmap_release: pmap has reserved page table page(s)")); 1767 1768 pmap_lazyfix(pmap); 1769 mtx_lock_spin(&allpmaps_lock); 1770 LIST_REMOVE(pmap, pm_list); 1771 mtx_unlock_spin(&allpmaps_lock); 1772 1773 for (i = 0; i < NPGPTD; i++) 1774 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i] & 1775 PG_FRAME); 1776 1777 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1778 sizeof(*pmap->pm_pdir)); 1779 1780 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1781 1782 for (i = 0; i < NPGPTD; i++) { 1783 m = ptdpg[i]; 1784#ifdef PAE 1785 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1786 ("pmap_release: got wrong ptd page")); 1787#endif 1788 m->wire_count--; 1789 atomic_subtract_int(&cnt.v_wire_count, 1); 1790 vm_page_free_zero(m); 1791 } 1792 PMAP_LOCK_DESTROY(pmap); 1793} 1794 1795static int 1796kvm_size(SYSCTL_HANDLER_ARGS) 1797{ 1798 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1799 1800 return sysctl_handle_long(oidp, &ksize, 0, req); 1801} 1802SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1803 0, 0, kvm_size, "IU", "Size of KVM"); 1804 1805static int 1806kvm_free(SYSCTL_HANDLER_ARGS) 1807{ 1808 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1809 1810 return sysctl_handle_long(oidp, &kfree, 0, req); 1811} 1812SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1813 0, 0, kvm_free, "IU", "Amount of KVM free"); 1814 1815/* 1816 * grow the number of kernel page table entries, if needed 1817 */ 1818void 1819pmap_growkernel(vm_offset_t addr) 1820{ 1821 struct pmap *pmap; 1822 vm_paddr_t ptppaddr; 1823 vm_page_t nkpg; 1824 pd_entry_t newpdir; 1825 pt_entry_t *pde; 1826 1827 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1828 if (kernel_vm_end == 0) { 1829 kernel_vm_end = KERNBASE; 1830 nkpt = 0; 1831 while (pdir_pde(PTD, kernel_vm_end)) { 1832 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1833 nkpt++; 1834 if (kernel_vm_end - 1 >= kernel_map->max_offset) { 1835 kernel_vm_end = kernel_map->max_offset; 1836 break; 1837 } 1838 } 1839 } 1840 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1841 if (addr - 1 >= kernel_map->max_offset) 1842 addr = kernel_map->max_offset; 1843 while (kernel_vm_end < addr) { 1844 if (pdir_pde(PTD, kernel_vm_end)) { 1845 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1846 if (kernel_vm_end - 1 >= kernel_map->max_offset) { 1847 kernel_vm_end = kernel_map->max_offset; 1848 break; 1849 } 1850 continue; 1851 } 1852 1853 nkpg = vm_page_alloc(NULL, kernel_vm_end >> PDRSHIFT, 1854 VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1855 VM_ALLOC_ZERO); 1856 if (nkpg == NULL) 1857 panic("pmap_growkernel: no memory to grow kernel"); 1858 1859 nkpt++; 1860 1861 if ((nkpg->flags & PG_ZERO) == 0) 1862 pmap_zero_page(nkpg); 1863 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1864 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1865 pdir_pde(PTD, kernel_vm_end) = newpdir; 1866 1867 mtx_lock_spin(&allpmaps_lock); 1868 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1869 pde = pmap_pde(pmap, kernel_vm_end); 1870 pde_store(pde, newpdir); 1871 } 1872 mtx_unlock_spin(&allpmaps_lock); 1873 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1874 if (kernel_vm_end - 1 >= kernel_map->max_offset) { 1875 kernel_vm_end = kernel_map->max_offset; 1876 break; 1877 } 1878 } 1879} 1880 1881 1882/*************************************************** 1883 * page management routines. 1884 ***************************************************/ 1885 1886CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE); 1887CTASSERT(_NPCM == 11); 1888 1889static __inline struct pv_chunk * 1890pv_to_chunk(pv_entry_t pv) 1891{ 1892 1893 return (struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK); 1894} 1895 1896#define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap) 1897 1898#define PC_FREE0_9 0xfffffffful /* Free values for index 0 through 9 */ 1899#define PC_FREE10 0x0000fffful /* Free values for index 10 */ 1900 1901static uint32_t pc_freemask[11] = { 1902 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, 1903 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, 1904 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, 1905 PC_FREE0_9, PC_FREE10 1906}; 1907 1908SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0, 1909 "Current number of pv entries"); 1910 1911#ifdef PV_STATS 1912static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail; 1913 1914SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0, 1915 "Current number of pv entry chunks"); 1916SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0, 1917 "Current number of pv entry chunks allocated"); 1918SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0, 1919 "Current number of pv entry chunks frees"); 1920SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0, 1921 "Number of times tried to get a chunk page but failed."); 1922 1923static long pv_entry_frees, pv_entry_allocs; 1924static int pv_entry_spare; 1925 1926SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0, 1927 "Current number of pv entry frees"); 1928SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0, 1929 "Current number of pv entry allocs"); 1930SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0, 1931 "Current number of spare pv entries"); 1932 1933static int pmap_collect_inactive, pmap_collect_active; 1934 1935SYSCTL_INT(_vm_pmap, OID_AUTO, pmap_collect_inactive, CTLFLAG_RD, &pmap_collect_inactive, 0, 1936 "Current number times pmap_collect called on inactive queue"); 1937SYSCTL_INT(_vm_pmap, OID_AUTO, pmap_collect_active, CTLFLAG_RD, &pmap_collect_active, 0, 1938 "Current number times pmap_collect called on active queue"); 1939#endif 1940 1941/* 1942 * We are in a serious low memory condition. Resort to 1943 * drastic measures to free some pages so we can allocate 1944 * another pv entry chunk. This is normally called to 1945 * unmap inactive pages, and if necessary, active pages. 1946 */ 1947static void 1948pmap_collect(pmap_t locked_pmap, struct vpgqueues *vpq) 1949{ 1950 struct md_page *pvh; 1951 pd_entry_t *pde; 1952 pmap_t pmap; 1953 pt_entry_t *pte, tpte; 1954 pv_entry_t next_pv, pv; 1955 vm_offset_t va; 1956 vm_page_t m, free; 1957 1958 sched_pin(); 1959 TAILQ_FOREACH(m, &vpq->pl, pageq) { 1960 if (m->hold_count || m->busy) 1961 continue; 1962 TAILQ_FOREACH_SAFE(pv, &m->md.pv_list, pv_list, next_pv) { 1963 va = pv->pv_va; 1964 pmap = PV_PMAP(pv); 1965 /* Avoid deadlock and lock recursion. */ 1966 if (pmap > locked_pmap) 1967 PMAP_LOCK(pmap); 1968 else if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap)) 1969 continue; 1970 pmap->pm_stats.resident_count--; 1971 pde = pmap_pde(pmap, va); 1972 KASSERT((*pde & PG_PS) == 0, ("pmap_collect: found" 1973 " a 4mpage in page %p's pv list", m)); 1974 pte = pmap_pte_quick(pmap, va); 1975 tpte = pte_load_clear(pte); 1976 KASSERT((tpte & PG_W) == 0, 1977 ("pmap_collect: wired pte %#jx", (uintmax_t)tpte)); 1978 if (tpte & PG_A) 1979 vm_page_flag_set(m, PG_REFERENCED); 1980 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) 1981 vm_page_dirty(m); 1982 free = NULL; 1983 pmap_unuse_pt(pmap, va, &free); 1984 pmap_invalidate_page(pmap, va); 1985 pmap_free_zero_pages(free); 1986 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1987 if (TAILQ_EMPTY(&m->md.pv_list)) { 1988 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); 1989 if (TAILQ_EMPTY(&pvh->pv_list)) 1990 vm_page_flag_clear(m, PG_WRITEABLE); 1991 } 1992 free_pv_entry(pmap, pv); 1993 if (pmap != locked_pmap) 1994 PMAP_UNLOCK(pmap); 1995 } 1996 } 1997 sched_unpin(); 1998} 1999 2000 2001/* 2002 * free the pv_entry back to the free list 2003 */ 2004static void 2005free_pv_entry(pmap_t pmap, pv_entry_t pv) 2006{ 2007 vm_page_t m; 2008 struct pv_chunk *pc; 2009 int idx, field, bit; 2010 2011 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2012 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 2013 PV_STAT(pv_entry_frees++); 2014 PV_STAT(pv_entry_spare++); 2015 pv_entry_count--; 2016 pc = pv_to_chunk(pv); 2017 idx = pv - &pc->pc_pventry[0]; 2018 field = idx / 32; 2019 bit = idx % 32; 2020 pc->pc_map[field] |= 1ul << bit; 2021 /* move to head of list */ 2022 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); 2023 for (idx = 0; idx < _NPCM; idx++) 2024 if (pc->pc_map[idx] != pc_freemask[idx]) { 2025 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); 2026 return; 2027 } 2028 PV_STAT(pv_entry_spare -= _NPCPV); 2029 PV_STAT(pc_chunk_count--); 2030 PV_STAT(pc_chunk_frees++); 2031 /* entire chunk is free, return it */ 2032 m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc)); 2033 pmap_qremove((vm_offset_t)pc, 1); 2034 vm_page_unwire(m, 0); 2035 vm_page_free(m); 2036 pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc); 2037} 2038 2039/* 2040 * get a new pv_entry, allocating a block from the system 2041 * when needed. 2042 */ 2043static pv_entry_t 2044get_pv_entry(pmap_t pmap, int try) 2045{ 2046 static const struct timeval printinterval = { 60, 0 }; 2047 static struct timeval lastprint; 2048 static vm_pindex_t colour; 2049 struct vpgqueues *pq; 2050 int bit, field; 2051 pv_entry_t pv; 2052 struct pv_chunk *pc; 2053 vm_page_t m; 2054 2055 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 2056 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2057 PV_STAT(pv_entry_allocs++); 2058 pv_entry_count++; 2059 if (pv_entry_count > pv_entry_high_water) 2060 if (ratecheck(&lastprint, &printinterval)) 2061 printf("Approaching the limit on PV entries, consider " 2062 "increasing either the vm.pmap.shpgperproc or the " 2063 "vm.pmap.pv_entry_max tunable.\n"); 2064 pq = NULL; 2065retry: 2066 pc = TAILQ_FIRST(&pmap->pm_pvchunk); 2067 if (pc != NULL) { 2068 for (field = 0; field < _NPCM; field++) { 2069 if (pc->pc_map[field]) { 2070 bit = bsfl(pc->pc_map[field]); 2071 break; 2072 } 2073 } 2074 if (field < _NPCM) { 2075 pv = &pc->pc_pventry[field * 32 + bit]; 2076 pc->pc_map[field] &= ~(1ul << bit); 2077 /* If this was the last item, move it to tail */ 2078 for (field = 0; field < _NPCM; field++) 2079 if (pc->pc_map[field] != 0) { 2080 PV_STAT(pv_entry_spare--); 2081 return (pv); /* not full, return */ 2082 } 2083 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); 2084 TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); 2085 PV_STAT(pv_entry_spare--); 2086 return (pv); 2087 } 2088 } 2089 /* 2090 * Access to the ptelist "pv_vafree" is synchronized by the page 2091 * queues lock. If "pv_vafree" is currently non-empty, it will 2092 * remain non-empty until pmap_ptelist_alloc() completes. 2093 */ 2094 if (pv_vafree == 0 || (m = vm_page_alloc(NULL, colour, (pq == 2095 &vm_page_queues[PQ_ACTIVE] ? VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL) | 2096 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) { 2097 if (try) { 2098 pv_entry_count--; 2099 PV_STAT(pc_chunk_tryfail++); 2100 return (NULL); 2101 } 2102 /* 2103 * Reclaim pv entries: At first, destroy mappings to 2104 * inactive pages. After that, if a pv chunk entry 2105 * is still needed, destroy mappings to active pages. 2106 */ 2107 if (pq == NULL) { 2108 PV_STAT(pmap_collect_inactive++); 2109 pq = &vm_page_queues[PQ_INACTIVE]; 2110 } else if (pq == &vm_page_queues[PQ_INACTIVE]) { 2111 PV_STAT(pmap_collect_active++); 2112 pq = &vm_page_queues[PQ_ACTIVE]; 2113 } else 2114 panic("get_pv_entry: increase vm.pmap.shpgperproc"); 2115 pmap_collect(pmap, pq); 2116 goto retry; 2117 } 2118 PV_STAT(pc_chunk_count++); 2119 PV_STAT(pc_chunk_allocs++); 2120 colour++; 2121 pc = (struct pv_chunk *)pmap_ptelist_alloc(&pv_vafree); 2122 pmap_qenter((vm_offset_t)pc, &m, 1); 2123 pc->pc_pmap = pmap; 2124 pc->pc_map[0] = pc_freemask[0] & ~1ul; /* preallocated bit 0 */ 2125 for (field = 1; field < _NPCM; field++) 2126 pc->pc_map[field] = pc_freemask[field]; 2127 pv = &pc->pc_pventry[0]; 2128 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); 2129 PV_STAT(pv_entry_spare += _NPCPV - 1); 2130 return (pv); 2131} 2132 2133static __inline pv_entry_t 2134pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va) 2135{ 2136 pv_entry_t pv; 2137 2138 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2139 TAILQ_FOREACH(pv, &pvh->pv_list, pv_list) { 2140 if (pmap == PV_PMAP(pv) && va == pv->pv_va) { 2141 TAILQ_REMOVE(&pvh->pv_list, pv, pv_list); 2142 break; 2143 } 2144 } 2145 return (pv); 2146} 2147 2148static void 2149pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa) 2150{ 2151 struct md_page *pvh; 2152 pv_entry_t pv; 2153 vm_offset_t va_last; 2154 vm_page_t m; 2155 2156 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2157 KASSERT((pa & PDRMASK) == 0, 2158 ("pmap_pv_demote_pde: pa is not 4mpage aligned")); 2159 2160 /* 2161 * Transfer the 4mpage's pv entry for this mapping to the first 2162 * page's pv list. 2163 */ 2164 pvh = pa_to_pvh(pa); 2165 va = trunc_4mpage(va); 2166 pv = pmap_pvh_remove(pvh, pmap, va); 2167 KASSERT(pv != NULL, ("pmap_pv_demote_pde: pv not found")); 2168 m = PHYS_TO_VM_PAGE(pa); 2169 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2170 /* Instantiate the remaining NPTEPG - 1 pv entries. */ 2171 va_last = va + NBPDR - PAGE_SIZE; 2172 do { 2173 m++; 2174 KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0, 2175 ("pmap_pv_demote_pde: page %p is not managed", m)); 2176 va += PAGE_SIZE; 2177 pmap_insert_entry(pmap, va, m); 2178 } while (va < va_last); 2179} 2180 2181static void 2182pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa) 2183{ 2184 struct md_page *pvh; 2185 pv_entry_t pv; 2186 vm_offset_t va_last; 2187 vm_page_t m; 2188 2189 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2190 KASSERT((pa & PDRMASK) == 0, 2191 ("pmap_pv_promote_pde: pa is not 4mpage aligned")); 2192 2193 /* 2194 * Transfer the first page's pv entry for this mapping to the 2195 * 4mpage's pv list. Aside from avoiding the cost of a call 2196 * to get_pv_entry(), a transfer avoids the possibility that 2197 * get_pv_entry() calls pmap_collect() and that pmap_collect() 2198 * removes one of the mappings that is being promoted. 2199 */ 2200 m = PHYS_TO_VM_PAGE(pa); 2201 va = trunc_4mpage(va); 2202 pv = pmap_pvh_remove(&m->md, pmap, va); 2203 KASSERT(pv != NULL, ("pmap_pv_promote_pde: pv not found")); 2204 pvh = pa_to_pvh(pa); 2205 TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_list); 2206 /* Free the remaining NPTEPG - 1 pv entries. */ 2207 va_last = va + NBPDR - PAGE_SIZE; 2208 do { 2209 m++; 2210 va += PAGE_SIZE; 2211 pmap_pvh_free(&m->md, pmap, va); 2212 } while (va < va_last); 2213} 2214 2215static void 2216pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va) 2217{ 2218 pv_entry_t pv; 2219 2220 pv = pmap_pvh_remove(pvh, pmap, va); 2221 KASSERT(pv != NULL, ("pmap_pvh_free: pv not found")); 2222 free_pv_entry(pmap, pv); 2223} 2224 2225static void 2226pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 2227{ 2228 struct md_page *pvh; 2229 2230 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2231 pmap_pvh_free(&m->md, pmap, va); 2232 if (TAILQ_EMPTY(&m->md.pv_list)) { 2233 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); 2234 if (TAILQ_EMPTY(&pvh->pv_list)) 2235 vm_page_flag_clear(m, PG_WRITEABLE); 2236 } 2237} 2238 2239/* 2240 * Create a pv entry for page at pa for 2241 * (pmap, va). 2242 */ 2243static void 2244pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) 2245{ 2246 pv_entry_t pv; 2247 2248 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 2249 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2250 pv = get_pv_entry(pmap, FALSE); 2251 pv->pv_va = va; 2252 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2253} 2254 2255/* 2256 * Conditionally create a pv entry. 2257 */ 2258static boolean_t 2259pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) 2260{ 2261 pv_entry_t pv; 2262 2263 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 2264 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2265 if (pv_entry_count < pv_entry_high_water && 2266 (pv = get_pv_entry(pmap, TRUE)) != NULL) { 2267 pv->pv_va = va; 2268 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2269 return (TRUE); 2270 } else 2271 return (FALSE); 2272} 2273 2274/* 2275 * Create the pv entries for each of the pages within a superpage. 2276 */ 2277static boolean_t 2278pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa) 2279{ 2280 struct md_page *pvh; 2281 pv_entry_t pv; 2282 2283 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2284 if (pv_entry_count < pv_entry_high_water && 2285 (pv = get_pv_entry(pmap, TRUE)) != NULL) { 2286 pv->pv_va = va; 2287 pvh = pa_to_pvh(pa); 2288 TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_list); 2289 return (TRUE); 2290 } else 2291 return (FALSE); 2292} 2293 2294/* 2295 * Fills a page table page with mappings to consecutive physical pages. 2296 */ 2297static void 2298pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte) 2299{ 2300 pt_entry_t *pte; 2301 2302 for (pte = firstpte; pte < firstpte + NPTEPG; pte++) { 2303 *pte = newpte; 2304 newpte += PAGE_SIZE; 2305 } 2306} 2307 2308/* 2309 * Tries to demote a 2- or 4MB page mapping. If demotion fails, the 2310 * 2- or 4MB page mapping is invalidated. 2311 */ 2312static boolean_t 2313pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) 2314{ 2315 pd_entry_t newpde, oldpde; 2316 pmap_t allpmaps_entry; 2317 pt_entry_t *firstpte, newpte; 2318 vm_paddr_t mptepa; 2319 vm_page_t free, mpte; 2320 2321 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 2322 oldpde = *pde; 2323 KASSERT((oldpde & (PG_PS | PG_V)) == (PG_PS | PG_V), 2324 ("pmap_demote_pde: oldpde is missing PG_PS and/or PG_V")); 2325 mpte = pmap_lookup_pt_page(pmap, va); 2326 if (mpte != NULL) 2327 pmap_remove_pt_page(pmap, mpte); 2328 else { 2329 KASSERT((oldpde & PG_W) == 0, 2330 ("pmap_demote_pde: page table page for a wired mapping" 2331 " is missing")); 2332 2333 /* 2334 * Invalidate the 2- or 4MB page mapping and return 2335 * "failure" if the mapping was never accessed or the 2336 * allocation of the new page table page fails. 2337 */ 2338 if ((oldpde & PG_A) == 0 || (mpte = vm_page_alloc(NULL, 2339 va >> PDRSHIFT, VM_ALLOC_NOOBJ | VM_ALLOC_NORMAL | 2340 VM_ALLOC_WIRED)) == NULL) { 2341 free = NULL; 2342 pmap_remove_pde(pmap, pde, trunc_4mpage(va), &free); 2343 pmap_invalidate_page(pmap, trunc_4mpage(va)); 2344 pmap_free_zero_pages(free); 2345 CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#x" 2346 " in pmap %p", va, pmap); 2347 return (FALSE); 2348 } 2349 if (va < VM_MAXUSER_ADDRESS) 2350 pmap->pm_stats.resident_count++; 2351 } 2352 mptepa = VM_PAGE_TO_PHYS(mpte); 2353 2354 /* 2355 * Temporarily map the page table page (mpte) into the kernel's 2356 * address space at either PADDR1 or PADDR2. 2357 */ 2358 if (curthread->td_pinned > 0 && mtx_owned(&vm_page_queue_mtx)) { 2359 if ((*PMAP1 & PG_FRAME) != mptepa) { 2360 *PMAP1 = mptepa | PG_RW | PG_V | PG_A | PG_M; 2361#ifdef SMP 2362 PMAP1cpu = PCPU_GET(cpuid); 2363#endif 2364 invlcaddr(PADDR1); 2365 PMAP1changed++; 2366 } else 2367#ifdef SMP 2368 if (PMAP1cpu != PCPU_GET(cpuid)) { 2369 PMAP1cpu = PCPU_GET(cpuid); 2370 invlcaddr(PADDR1); 2371 PMAP1changedcpu++; 2372 } else 2373#endif 2374 PMAP1unchanged++; 2375 firstpte = PADDR1; 2376 } else { 2377 mtx_lock(&PMAP2mutex); 2378 if ((*PMAP2 & PG_FRAME) != mptepa) { 2379 *PMAP2 = mptepa | PG_RW | PG_V | PG_A | PG_M; 2380 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); 2381 } 2382 firstpte = PADDR2; 2383 } 2384 newpde = mptepa | PG_M | PG_A | (oldpde & PG_U) | PG_RW | PG_V; 2385 KASSERT((oldpde & PG_A) != 0, 2386 ("pmap_demote_pde: oldpde is missing PG_A")); 2387 KASSERT((oldpde & (PG_M | PG_RW)) != PG_RW, 2388 ("pmap_demote_pde: oldpde is missing PG_M")); 2389 newpte = oldpde & ~PG_PS; 2390 if ((newpte & PG_PDE_PAT) != 0) 2391 newpte ^= PG_PDE_PAT | PG_PTE_PAT; 2392 2393 /* 2394 * If the page table page is new, initialize it. 2395 */ 2396 if (mpte->wire_count == 1) { 2397 mpte->wire_count = NPTEPG; 2398 pmap_fill_ptp(firstpte, newpte); 2399 } 2400 KASSERT((*firstpte & PG_FRAME) == (newpte & PG_FRAME), 2401 ("pmap_demote_pde: firstpte and newpte map different physical" 2402 " addresses")); 2403 2404 /* 2405 * If the mapping has changed attributes, update the page table 2406 * entries. 2407 */ 2408 if ((*firstpte & PG_PTE_PROMOTE) != (newpte & PG_PTE_PROMOTE)) 2409 pmap_fill_ptp(firstpte, newpte); 2410 2411 /* 2412 * Demote the mapping. This pmap is locked. The old PDE has 2413 * PG_A set. If the old PDE has PG_RW set, it also has PG_M 2414 * set. Thus, there is no danger of a race with another 2415 * processor changing the setting of PG_A and/or PG_M between 2416 * the read above and the store below. 2417 */ 2418 if (pmap == kernel_pmap) { 2419 /* 2420 * A harmless race exists between this loop and the bcopy() 2421 * in pmap_pinit() that initializes the kernel segment of 2422 * the new page table. Specifically, that bcopy() may copy 2423 * the new PDE from the PTD, which is first in allpmaps, to 2424 * the new page table before this loop updates that new 2425 * page table. 2426 */ 2427 mtx_lock_spin(&allpmaps_lock); 2428 LIST_FOREACH(allpmaps_entry, &allpmaps, pm_list) { 2429 pde = pmap_pde(allpmaps_entry, va); 2430 KASSERT(*pde == newpde || (*pde & PG_PTE_PROMOTE) == 2431 (oldpde & PG_PTE_PROMOTE), 2432 ("pmap_demote_pde: pde was %#jx, expected %#jx", 2433 (uintmax_t)*pde, (uintmax_t)oldpde)); 2434 pde_store(pde, newpde); 2435 } 2436 mtx_unlock_spin(&allpmaps_lock); 2437 } else 2438 pde_store(pde, newpde); 2439 if (firstpte == PADDR2) 2440 mtx_unlock(&PMAP2mutex); 2441 2442 /* 2443 * Invalidate the recursive mapping of the page table page. 2444 */ 2445 pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va)); 2446 2447 /* 2448 * Demote the pv entry. This depends on the earlier demotion 2449 * of the mapping. Specifically, the (re)creation of a per- 2450 * page pv entry might trigger the execution of pmap_collect(), 2451 * which might reclaim a newly (re)created per-page pv entry 2452 * and destroy the associated mapping. In order to destroy 2453 * the mapping, the PDE must have already changed from mapping 2454 * the 2mpage to referencing the page table page. 2455 */ 2456 if ((oldpde & PG_MANAGED) != 0) 2457 pmap_pv_demote_pde(pmap, va, oldpde & PG_PS_FRAME); 2458 2459 pmap_pde_demotions++; 2460 CTR2(KTR_PMAP, "pmap_demote_pde: success for va %#x" 2461 " in pmap %p", va, pmap); 2462 return (TRUE); 2463} 2464 2465/* 2466 * pmap_remove_pde: do the things to unmap a superpage in a process 2467 */ 2468static void 2469pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, 2470 vm_page_t *free) 2471{ 2472 struct md_page *pvh; 2473 pd_entry_t oldpde; 2474 vm_offset_t eva, va; 2475 vm_page_t m, mpte; 2476 2477 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 2478 KASSERT((sva & PDRMASK) == 0, 2479 ("pmap_remove_pde: sva is not 4mpage aligned")); 2480 oldpde = pte_load_clear(pdq); 2481 if (oldpde & PG_W) 2482 pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; 2483 2484 /* 2485 * Machines that don't support invlpg, also don't support 2486 * PG_G. 2487 */ 2488 if (oldpde & PG_G) 2489 pmap_invalidate_page(kernel_pmap, sva); 2490 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 2491 if (oldpde & PG_MANAGED) { 2492 pvh = pa_to_pvh(oldpde & PG_PS_FRAME); 2493 pmap_pvh_free(pvh, pmap, sva); 2494 eva = sva + NBPDR; 2495 for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); 2496 va < eva; va += PAGE_SIZE, m++) { 2497 if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) 2498 vm_page_dirty(m); 2499 if (oldpde & PG_A) 2500 vm_page_flag_set(m, PG_REFERENCED); 2501 if (TAILQ_EMPTY(&m->md.pv_list) && 2502 TAILQ_EMPTY(&pvh->pv_list)) 2503 vm_page_flag_clear(m, PG_WRITEABLE); 2504 } 2505 } 2506 if (pmap == kernel_pmap) { 2507 if (!pmap_demote_pde(pmap, pdq, sva)) 2508 panic("pmap_remove_pde: failed demotion"); 2509 } else { 2510 mpte = pmap_lookup_pt_page(pmap, sva); 2511 if (mpte != NULL) { 2512 pmap_remove_pt_page(pmap, mpte); 2513 pmap->pm_stats.resident_count--; 2514 KASSERT(mpte->wire_count == NPTEPG, 2515 ("pmap_remove_pde: pte page wire count error")); 2516 mpte->wire_count = 0; 2517 pmap_add_delayed_free_list(mpte, free, FALSE); 2518 atomic_subtract_int(&cnt.v_wire_count, 1); 2519 } 2520 } 2521} 2522 2523/* 2524 * pmap_remove_pte: do the things to unmap a page in a process 2525 */ 2526static int 2527pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va, vm_page_t *free) 2528{ 2529 pt_entry_t oldpte; 2530 vm_page_t m; 2531 2532 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2533 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 2534 oldpte = pte_load_clear(ptq); 2535 if (oldpte & PG_W) 2536 pmap->pm_stats.wired_count -= 1; 2537 /* 2538 * Machines that don't support invlpg, also don't support 2539 * PG_G. 2540 */ 2541 if (oldpte & PG_G) 2542 pmap_invalidate_page(kernel_pmap, va); 2543 pmap->pm_stats.resident_count -= 1; 2544 if (oldpte & PG_MANAGED) { 2545 m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME); 2546 if ((oldpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) 2547 vm_page_dirty(m); 2548 if (oldpte & PG_A) 2549 vm_page_flag_set(m, PG_REFERENCED); 2550 pmap_remove_entry(pmap, m, va); 2551 } 2552 return (pmap_unuse_pt(pmap, va, free)); 2553} 2554 2555/* 2556 * Remove a single page from a process address space 2557 */ 2558static void 2559pmap_remove_page(pmap_t pmap, vm_offset_t va, vm_page_t *free) 2560{ 2561 pt_entry_t *pte; 2562 2563 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2564 KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); 2565 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 2566 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) 2567 return; 2568 pmap_remove_pte(pmap, pte, va, free); 2569 pmap_invalidate_page(pmap, va); 2570} 2571 2572/* 2573 * Remove the given range of addresses from the specified map. 2574 * 2575 * It is assumed that the start and end are properly 2576 * rounded to the page size. 2577 */ 2578void 2579pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 2580{ 2581 vm_offset_t pdnxt; 2582 pd_entry_t ptpaddr; 2583 pt_entry_t *pte; 2584 vm_page_t free = NULL; 2585 int anyvalid; 2586 2587 /* 2588 * Perform an unsynchronized read. This is, however, safe. 2589 */ 2590 if (pmap->pm_stats.resident_count == 0) 2591 return; 2592 2593 anyvalid = 0; 2594 2595 vm_page_lock_queues(); 2596 sched_pin(); 2597 PMAP_LOCK(pmap); 2598 2599 /* 2600 * special handling of removing one page. a very 2601 * common operation and easy to short circuit some 2602 * code. 2603 */ 2604 if ((sva + PAGE_SIZE == eva) && 2605 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 2606 pmap_remove_page(pmap, sva, &free); 2607 goto out; 2608 } 2609 2610 for (; sva < eva; sva = pdnxt) { 2611 unsigned pdirindex; 2612 2613 /* 2614 * Calculate index for next page table. 2615 */ 2616 pdnxt = (sva + NBPDR) & ~PDRMASK; 2617 if (pdnxt < sva) 2618 pdnxt = eva; 2619 if (pmap->pm_stats.resident_count == 0) 2620 break; 2621 2622 pdirindex = sva >> PDRSHIFT; 2623 ptpaddr = pmap->pm_pdir[pdirindex]; 2624 2625 /* 2626 * Weed out invalid mappings. Note: we assume that the page 2627 * directory table is always allocated, and in kernel virtual. 2628 */ 2629 if (ptpaddr == 0) 2630 continue; 2631 2632 /* 2633 * Check for large page. 2634 */ 2635 if ((ptpaddr & PG_PS) != 0) { 2636 /* 2637 * Are we removing the entire large page? If not, 2638 * demote the mapping and fall through. 2639 */ 2640 if (sva + NBPDR == pdnxt && eva >= pdnxt) { 2641 /* 2642 * The TLB entry for a PG_G mapping is 2643 * invalidated by pmap_remove_pde(). 2644 */ 2645 if ((ptpaddr & PG_G) == 0) 2646 anyvalid = 1; 2647 pmap_remove_pde(pmap, 2648 &pmap->pm_pdir[pdirindex], sva, &free); 2649 continue; 2650 } else if (!pmap_demote_pde(pmap, 2651 &pmap->pm_pdir[pdirindex], sva)) { 2652 /* The large page mapping was destroyed. */ 2653 continue; 2654 } 2655 } 2656 2657 /* 2658 * Limit our scan to either the end of the va represented 2659 * by the current page table page, or to the end of the 2660 * range being removed. 2661 */ 2662 if (pdnxt > eva) 2663 pdnxt = eva; 2664 2665 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, 2666 sva += PAGE_SIZE) { 2667 if (*pte == 0) 2668 continue; 2669 2670 /* 2671 * The TLB entry for a PG_G mapping is invalidated 2672 * by pmap_remove_pte(). 2673 */ 2674 if ((*pte & PG_G) == 0) 2675 anyvalid = 1; 2676 if (pmap_remove_pte(pmap, pte, sva, &free)) 2677 break; 2678 } 2679 } 2680out: 2681 sched_unpin(); 2682 if (anyvalid) 2683 pmap_invalidate_all(pmap); 2684 vm_page_unlock_queues(); 2685 PMAP_UNLOCK(pmap); 2686 pmap_free_zero_pages(free); 2687} 2688 2689/* 2690 * Routine: pmap_remove_all 2691 * Function: 2692 * Removes this physical page from 2693 * all physical maps in which it resides. 2694 * Reflects back modify bits to the pager. 2695 * 2696 * Notes: 2697 * Original versions of this routine were very 2698 * inefficient because they iteratively called 2699 * pmap_remove (slow...) 2700 */ 2701 2702void 2703pmap_remove_all(vm_page_t m) 2704{ 2705 struct md_page *pvh; 2706 pv_entry_t pv; 2707 pmap_t pmap; 2708 pt_entry_t *pte, tpte; 2709 pd_entry_t *pde; 2710 vm_offset_t va; 2711 vm_page_t free; 2712 2713 KASSERT((m->flags & PG_FICTITIOUS) == 0, 2714 ("pmap_remove_all: page %p is fictitious", m)); 2715 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2716 sched_pin(); 2717 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); 2718 while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) { 2719 va = pv->pv_va; 2720 pmap = PV_PMAP(pv); 2721 PMAP_LOCK(pmap); 2722 pde = pmap_pde(pmap, va); 2723 (void)pmap_demote_pde(pmap, pde, va); 2724 PMAP_UNLOCK(pmap); 2725 } 2726 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2727 pmap = PV_PMAP(pv); 2728 PMAP_LOCK(pmap); 2729 pmap->pm_stats.resident_count--; 2730 pde = pmap_pde(pmap, pv->pv_va); 2731 KASSERT((*pde & PG_PS) == 0, ("pmap_remove_all: found" 2732 " a 4mpage in page %p's pv list", m)); 2733 pte = pmap_pte_quick(pmap, pv->pv_va); 2734 tpte = pte_load_clear(pte); 2735 if (tpte & PG_W) 2736 pmap->pm_stats.wired_count--; 2737 if (tpte & PG_A) 2738 vm_page_flag_set(m, PG_REFERENCED); 2739 2740 /* 2741 * Update the vm_page_t clean and reference bits. 2742 */ 2743 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) 2744 vm_page_dirty(m); 2745 free = NULL; 2746 pmap_unuse_pt(pmap, pv->pv_va, &free); 2747 pmap_invalidate_page(pmap, pv->pv_va); 2748 pmap_free_zero_pages(free); 2749 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2750 free_pv_entry(pmap, pv); 2751 PMAP_UNLOCK(pmap); 2752 } 2753 vm_page_flag_clear(m, PG_WRITEABLE); 2754 sched_unpin(); 2755} 2756 2757/* 2758 * pmap_protect_pde: do the things to protect a 4mpage in a process 2759 */ 2760static boolean_t 2761pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot) 2762{ 2763 pd_entry_t newpde, oldpde; 2764 vm_offset_t eva, va; 2765 vm_page_t m; 2766 boolean_t anychanged; 2767 2768 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 2769 KASSERT((sva & PDRMASK) == 0, 2770 ("pmap_protect_pde: sva is not 4mpage aligned")); 2771 anychanged = FALSE; 2772retry: 2773 oldpde = newpde = *pde; 2774 if (oldpde & PG_MANAGED) { 2775 eva = sva + NBPDR; 2776 for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); 2777 va < eva; va += PAGE_SIZE, m++) { 2778 /* 2779 * In contrast to the analogous operation on a 4KB page 2780 * mapping, the mapping's PG_A flag is not cleared and 2781 * the page's PG_REFERENCED flag is not set. The 2782 * reason is that pmap_demote_pde() expects that a 2/4MB 2783 * page mapping with a stored page table page has PG_A 2784 * set. 2785 */ 2786 if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) 2787 vm_page_dirty(m); 2788 } 2789 } 2790 if ((prot & VM_PROT_WRITE) == 0) 2791 newpde &= ~(PG_RW | PG_M); 2792#ifdef PAE 2793 if ((prot & VM_PROT_EXECUTE) == 0) 2794 newpde |= pg_nx; 2795#endif 2796 if (newpde != oldpde) { 2797 if (!pde_cmpset(pde, oldpde, newpde)) 2798 goto retry; 2799 if (oldpde & PG_G) 2800 pmap_invalidate_page(pmap, sva); 2801 else 2802 anychanged = TRUE; 2803 } 2804 return (anychanged); 2805} 2806 2807/* 2808 * Set the physical protection on the 2809 * specified range of this map as requested. 2810 */ 2811void 2812pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 2813{ 2814 vm_offset_t pdnxt; 2815 pd_entry_t ptpaddr; 2816 pt_entry_t *pte; 2817 int anychanged; 2818 2819 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 2820 pmap_remove(pmap, sva, eva); 2821 return; 2822 } 2823 2824#ifdef PAE 2825 if ((prot & (VM_PROT_WRITE|VM_PROT_EXECUTE)) == 2826 (VM_PROT_WRITE|VM_PROT_EXECUTE)) 2827 return; 2828#else 2829 if (prot & VM_PROT_WRITE) 2830 return; 2831#endif 2832 2833 anychanged = 0; 2834 2835 vm_page_lock_queues(); 2836 sched_pin(); 2837 PMAP_LOCK(pmap); 2838 for (; sva < eva; sva = pdnxt) { 2839 pt_entry_t obits, pbits; 2840 unsigned pdirindex; 2841 2842 pdnxt = (sva + NBPDR) & ~PDRMASK; 2843 if (pdnxt < sva) 2844 pdnxt = eva; 2845 2846 pdirindex = sva >> PDRSHIFT; 2847 ptpaddr = pmap->pm_pdir[pdirindex]; 2848 2849 /* 2850 * Weed out invalid mappings. Note: we assume that the page 2851 * directory table is always allocated, and in kernel virtual. 2852 */ 2853 if (ptpaddr == 0) 2854 continue; 2855 2856 /* 2857 * Check for large page. 2858 */ 2859 if ((ptpaddr & PG_PS) != 0) { 2860 /* 2861 * Are we protecting the entire large page? If not, 2862 * demote the mapping and fall through. 2863 */ 2864 if (sva + NBPDR == pdnxt && eva >= pdnxt) { 2865 /* 2866 * The TLB entry for a PG_G mapping is 2867 * invalidated by pmap_protect_pde(). 2868 */ 2869 if (pmap_protect_pde(pmap, 2870 &pmap->pm_pdir[pdirindex], sva, prot)) 2871 anychanged = 1; 2872 continue; 2873 } else if (!pmap_demote_pde(pmap, 2874 &pmap->pm_pdir[pdirindex], sva)) { 2875 /* The large page mapping was destroyed. */ 2876 continue; 2877 } 2878 } 2879 2880 if (pdnxt > eva) 2881 pdnxt = eva; 2882 2883 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, 2884 sva += PAGE_SIZE) { 2885 vm_page_t m; 2886 2887retry: 2888 /* 2889 * Regardless of whether a pte is 32 or 64 bits in 2890 * size, PG_RW, PG_A, and PG_M are among the least 2891 * significant 32 bits. 2892 */ 2893 obits = pbits = *pte; 2894 if ((pbits & PG_V) == 0) 2895 continue; 2896 if (pbits & PG_MANAGED) { 2897 m = NULL; 2898 if (pbits & PG_A) { 2899 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME); 2900 vm_page_flag_set(m, PG_REFERENCED); 2901 pbits &= ~PG_A; 2902 } 2903 if ((pbits & (PG_M | PG_RW)) == (PG_M | PG_RW)) { 2904 if (m == NULL) 2905 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME); 2906 vm_page_dirty(m); 2907 } 2908 } 2909 2910 if ((prot & VM_PROT_WRITE) == 0) 2911 pbits &= ~(PG_RW | PG_M); 2912#ifdef PAE 2913 if ((prot & VM_PROT_EXECUTE) == 0) 2914 pbits |= pg_nx; 2915#endif 2916 2917 if (pbits != obits) { 2918#ifdef PAE 2919 if (!atomic_cmpset_64(pte, obits, pbits)) 2920 goto retry; 2921#else 2922 if (!atomic_cmpset_int((u_int *)pte, obits, 2923 pbits)) 2924 goto retry; 2925#endif 2926 if (obits & PG_G) 2927 pmap_invalidate_page(pmap, sva); 2928 else 2929 anychanged = 1; 2930 } 2931 } 2932 } 2933 sched_unpin(); 2934 if (anychanged) 2935 pmap_invalidate_all(pmap); 2936 vm_page_unlock_queues(); 2937 PMAP_UNLOCK(pmap); 2938} 2939 2940/* 2941 * Tries to promote the 512 or 1024, contiguous 4KB page mappings that are 2942 * within a single page table page (PTP) to a single 2- or 4MB page mapping. 2943 * For promotion to occur, two conditions must be met: (1) the 4KB page 2944 * mappings must map aligned, contiguous physical memory and (2) the 4KB page 2945 * mappings must have identical characteristics. 2946 * 2947 * Managed (PG_MANAGED) mappings within the kernel address space are not 2948 * promoted. The reason is that kernel PDEs are replicated in each pmap but 2949 * pmap_clear_ptes() and pmap_ts_referenced() only read the PDE from the kernel 2950 * pmap. 2951 */ 2952static void 2953pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) 2954{ 2955 pd_entry_t newpde; 2956 pmap_t allpmaps_entry; 2957 pt_entry_t *firstpte, oldpte, pa, *pte; 2958 vm_offset_t oldpteva; 2959 vm_page_t mpte; 2960 2961 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 2962 2963 /* 2964 * Examine the first PTE in the specified PTP. Abort if this PTE is 2965 * either invalid, unused, or does not map the first 4KB physical page 2966 * within a 2- or 4MB page. 2967 */ 2968 firstpte = vtopte(trunc_4mpage(va)); 2969setpde: 2970 newpde = *firstpte; 2971 if ((newpde & ((PG_FRAME & PDRMASK) | PG_A | PG_V)) != (PG_A | PG_V)) { 2972 pmap_pde_p_failures++; 2973 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" 2974 " in pmap %p", va, pmap); 2975 return; 2976 } 2977 if ((*firstpte & PG_MANAGED) != 0 && pmap == kernel_pmap) { 2978 pmap_pde_p_failures++; 2979 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" 2980 " in pmap %p", va, pmap); 2981 return; 2982 } 2983 if ((newpde & (PG_M | PG_RW)) == PG_RW) { 2984 /* 2985 * When PG_M is already clear, PG_RW can be cleared without 2986 * a TLB invalidation. 2987 */ 2988 if (!atomic_cmpset_int((u_int *)firstpte, newpde, newpde & 2989 ~PG_RW)) 2990 goto setpde; 2991 newpde &= ~PG_RW; 2992 } 2993 2994 /* 2995 * Examine each of the other PTEs in the specified PTP. Abort if this 2996 * PTE maps an unexpected 4KB physical page or does not have identical 2997 * characteristics to the first PTE. 2998 */ 2999 pa = (newpde & (PG_PS_FRAME | PG_A | PG_V)) + NBPDR - PAGE_SIZE; 3000 for (pte = firstpte + NPTEPG - 1; pte > firstpte; pte--) { 3001setpte: 3002 oldpte = *pte; 3003 if ((oldpte & (PG_FRAME | PG_A | PG_V)) != pa) { 3004 pmap_pde_p_failures++; 3005 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" 3006 " in pmap %p", va, pmap); 3007 return; 3008 } 3009 if ((oldpte & (PG_M | PG_RW)) == PG_RW) { 3010 /* 3011 * When PG_M is already clear, PG_RW can be cleared 3012 * without a TLB invalidation. 3013 */ 3014 if (!atomic_cmpset_int((u_int *)pte, oldpte, 3015 oldpte & ~PG_RW)) 3016 goto setpte; 3017 oldpte &= ~PG_RW; 3018 oldpteva = (oldpte & PG_FRAME & PDRMASK) | 3019 (va & ~PDRMASK); 3020 CTR2(KTR_PMAP, "pmap_promote_pde: protect for va %#x" 3021 " in pmap %p", oldpteva, pmap); 3022 } 3023 if ((oldpte & PG_PTE_PROMOTE) != (newpde & PG_PTE_PROMOTE)) { 3024 pmap_pde_p_failures++; 3025 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" 3026 " in pmap %p", va, pmap); 3027 return; 3028 } 3029 pa -= PAGE_SIZE; 3030 } 3031 3032 /* 3033 * Save the page table page in its current state until the PDE 3034 * mapping the superpage is demoted by pmap_demote_pde() or 3035 * destroyed by pmap_remove_pde(). 3036 */ 3037 mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); 3038 KASSERT(mpte >= vm_page_array && 3039 mpte < &vm_page_array[vm_page_array_size], 3040 ("pmap_promote_pde: page table page is out of range")); 3041 KASSERT(mpte->pindex == va >> PDRSHIFT, 3042 ("pmap_promote_pde: page table page's pindex is wrong")); 3043 pmap_insert_pt_page(pmap, mpte); 3044 3045 /* 3046 * Promote the pv entries. 3047 */ 3048 if ((newpde & PG_MANAGED) != 0) 3049 pmap_pv_promote_pde(pmap, va, newpde & PG_PS_FRAME); 3050 3051 /* 3052 * Propagate the PAT index to its proper position. 3053 */ 3054 if ((newpde & PG_PTE_PAT) != 0) 3055 newpde ^= PG_PDE_PAT | PG_PTE_PAT; 3056 3057 /* 3058 * Map the superpage. 3059 */ 3060 if (pmap == kernel_pmap) { 3061 mtx_lock_spin(&allpmaps_lock); 3062 LIST_FOREACH(allpmaps_entry, &allpmaps, pm_list) { 3063 pde = pmap_pde(allpmaps_entry, va); 3064 pde_store(pde, PG_PS | newpde); 3065 } 3066 mtx_unlock_spin(&allpmaps_lock); 3067 } else 3068 pde_store(pde, PG_PS | newpde); 3069 3070 pmap_pde_promotions++; 3071 CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#x" 3072 " in pmap %p", va, pmap); 3073} 3074 3075/* 3076 * Insert the given physical page (p) at 3077 * the specified virtual address (v) in the 3078 * target physical map with the protection requested. 3079 * 3080 * If specified, the page will be wired down, meaning 3081 * that the related pte can not be reclaimed. 3082 * 3083 * NB: This is the only routine which MAY NOT lazy-evaluate 3084 * or lose information. That is, this routine must actually 3085 * insert this page into the given map NOW. 3086 */ 3087void 3088pmap_enter(pmap_t pmap, vm_offset_t va, vm_prot_t access, vm_page_t m, 3089 vm_prot_t prot, boolean_t wired) 3090{ 3091 vm_paddr_t pa; 3092 pd_entry_t *pde; 3093 pt_entry_t *pte; 3094 vm_paddr_t opa; 3095 pt_entry_t origpte, newpte; 3096 vm_page_t mpte, om; 3097 boolean_t invlva; 3098 3099 va = trunc_page(va); 3100 KASSERT(va <= VM_MAX_KERNEL_ADDRESS, ("pmap_enter: toobig")); 3101 KASSERT(va < UPT_MIN_ADDRESS || va >= UPT_MAX_ADDRESS, 3102 ("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va)); 3103 3104 mpte = NULL; 3105 3106 vm_page_lock_queues(); 3107 PMAP_LOCK(pmap); 3108 sched_pin(); 3109 3110 /* 3111 * In the case that a page table page is not 3112 * resident, we are creating it here. 3113 */ 3114 if (va < VM_MAXUSER_ADDRESS) { 3115 mpte = pmap_allocpte(pmap, va, M_WAITOK); 3116 } 3117 3118 pde = pmap_pde(pmap, va); 3119 if ((*pde & PG_PS) != 0) 3120 panic("pmap_enter: attempted pmap_enter on 4MB page"); 3121 pte = pmap_pte_quick(pmap, va); 3122 3123 /* 3124 * Page Directory table entry not valid, we need a new PT page 3125 */ 3126 if (pte == NULL) { 3127 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x", 3128 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 3129 } 3130 3131 pa = VM_PAGE_TO_PHYS(m); 3132 om = NULL; 3133 origpte = *pte; 3134 opa = origpte & PG_FRAME; 3135 3136 /* 3137 * Mapping has not changed, must be protection or wiring change. 3138 */ 3139 if (origpte && (opa == pa)) { 3140 /* 3141 * Wiring change, just update stats. We don't worry about 3142 * wiring PT pages as they remain resident as long as there 3143 * are valid mappings in them. Hence, if a user page is wired, 3144 * the PT page will be also. 3145 */ 3146 if (wired && ((origpte & PG_W) == 0)) 3147 pmap->pm_stats.wired_count++; 3148 else if (!wired && (origpte & PG_W)) 3149 pmap->pm_stats.wired_count--; 3150 3151 /* 3152 * Remove extra pte reference 3153 */ 3154 if (mpte) 3155 mpte->wire_count--; 3156 3157 /* 3158 * We might be turning off write access to the page, 3159 * so we go ahead and sense modify status. 3160 */ 3161 if (origpte & PG_MANAGED) { 3162 om = m; 3163 pa |= PG_MANAGED; 3164 } 3165 goto validate; 3166 } 3167 /* 3168 * Mapping has changed, invalidate old range and fall through to 3169 * handle validating new mapping. 3170 */ 3171 if (opa) { 3172 if (origpte & PG_W) 3173 pmap->pm_stats.wired_count--; 3174 if (origpte & PG_MANAGED) { 3175 om = PHYS_TO_VM_PAGE(opa); 3176 pmap_remove_entry(pmap, om, va); 3177 } 3178 if (mpte != NULL) { 3179 mpte->wire_count--; 3180 KASSERT(mpte->wire_count > 0, 3181 ("pmap_enter: missing reference to page table page," 3182 " va: 0x%x", va)); 3183 } 3184 } else 3185 pmap->pm_stats.resident_count++; 3186 3187 /* 3188 * Enter on the PV list if part of our managed memory. 3189 */ 3190 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0) { 3191 KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva, 3192 ("pmap_enter: managed mapping within the clean submap")); 3193 pmap_insert_entry(pmap, va, m); 3194 pa |= PG_MANAGED; 3195 } 3196 3197 /* 3198 * Increment counters 3199 */ 3200 if (wired) 3201 pmap->pm_stats.wired_count++; 3202 3203validate: 3204 /* 3205 * Now validate mapping with desired protection/wiring. 3206 */ 3207 newpte = (pt_entry_t)(pa | pmap_cache_bits(m->md.pat_mode, 0) | PG_V); 3208 if ((prot & VM_PROT_WRITE) != 0) { 3209 newpte |= PG_RW; 3210 vm_page_flag_set(m, PG_WRITEABLE); 3211 } 3212#ifdef PAE 3213 if ((prot & VM_PROT_EXECUTE) == 0) 3214 newpte |= pg_nx; 3215#endif 3216 if (wired) 3217 newpte |= PG_W; 3218 if (va < VM_MAXUSER_ADDRESS) 3219 newpte |= PG_U; 3220 if (pmap == kernel_pmap) 3221 newpte |= pgeflag; 3222 3223 /* 3224 * if the mapping or permission bits are different, we need 3225 * to update the pte. 3226 */ 3227 if ((origpte & ~(PG_M|PG_A)) != newpte) { 3228 newpte |= PG_A; 3229 if ((access & VM_PROT_WRITE) != 0) 3230 newpte |= PG_M; 3231 if (origpte & PG_V) { 3232 invlva = FALSE; 3233 origpte = pte_load_store(pte, newpte); 3234 if (origpte & PG_A) { 3235 if (origpte & PG_MANAGED) 3236 vm_page_flag_set(om, PG_REFERENCED); 3237 if (opa != VM_PAGE_TO_PHYS(m)) 3238 invlva = TRUE; 3239#ifdef PAE 3240 if ((origpte & PG_NX) == 0 && 3241 (newpte & PG_NX) != 0) 3242 invlva = TRUE; 3243#endif 3244 } 3245 if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { 3246 if ((origpte & PG_MANAGED) != 0) 3247 vm_page_dirty(om); 3248 if ((prot & VM_PROT_WRITE) == 0) 3249 invlva = TRUE; 3250 } 3251 if (invlva) 3252 pmap_invalidate_page(pmap, va); 3253 } else 3254 pte_store(pte, newpte); 3255 } 3256 3257 /* 3258 * If both the page table page and the reservation are fully 3259 * populated, then attempt promotion. 3260 */ 3261 if ((mpte == NULL || mpte->wire_count == NPTEPG) && 3262 pg_ps_enabled && vm_reserv_level_iffullpop(m) == 0) 3263 pmap_promote_pde(pmap, pde, va); 3264 3265 sched_unpin(); 3266 vm_page_unlock_queues(); 3267 PMAP_UNLOCK(pmap); 3268} 3269 3270/* 3271 * Tries to create a 2- or 4MB page mapping. Returns TRUE if successful and 3272 * FALSE otherwise. Fails if (1) a page table page cannot be allocated without 3273 * blocking, (2) a mapping already exists at the specified virtual address, or 3274 * (3) a pv entry cannot be allocated without reclaiming another pv entry. 3275 */ 3276static boolean_t 3277pmap_enter_pde(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) 3278{ 3279 pd_entry_t *pde, newpde; 3280 3281 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3282 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 3283 pde = pmap_pde(pmap, va); 3284 if (*pde != 0) { 3285 CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" 3286 " in pmap %p", va, pmap); 3287 return (FALSE); 3288 } 3289 newpde = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(m->md.pat_mode, 1) | 3290 PG_PS | PG_V; 3291 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0) { 3292 newpde |= PG_MANAGED; 3293 3294 /* 3295 * Abort this mapping if its PV entry could not be created. 3296 */ 3297 if (!pmap_pv_insert_pde(pmap, va, VM_PAGE_TO_PHYS(m))) { 3298 CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" 3299 " in pmap %p", va, pmap); 3300 return (FALSE); 3301 } 3302 } 3303#ifdef PAE 3304 if ((prot & VM_PROT_EXECUTE) == 0) 3305 newpde |= pg_nx; 3306#endif 3307 if (va < VM_MAXUSER_ADDRESS) 3308 newpde |= PG_U; 3309 3310 /* 3311 * Increment counters. 3312 */ 3313 pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; 3314 3315 /* 3316 * Map the superpage. 3317 */ 3318 pde_store(pde, newpde); 3319 3320 pmap_pde_mappings++; 3321 CTR2(KTR_PMAP, "pmap_enter_pde: success for va %#lx" 3322 " in pmap %p", va, pmap); 3323 return (TRUE); 3324} 3325 3326/* 3327 * Maps a sequence of resident pages belonging to the same object. 3328 * The sequence begins with the given page m_start. This page is 3329 * mapped at the given virtual address start. Each subsequent page is 3330 * mapped at a virtual address that is offset from start by the same 3331 * amount as the page is offset from m_start within the object. The 3332 * last page in the sequence is the page with the largest offset from 3333 * m_start that can be mapped at a virtual address less than the given 3334 * virtual address end. Not every virtual page between start and end 3335 * is mapped; only those for which a resident page exists with the 3336 * corresponding offset from m_start are mapped. 3337 */ 3338void 3339pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, 3340 vm_page_t m_start, vm_prot_t prot) 3341{ 3342 vm_offset_t va; 3343 vm_page_t m, mpte; 3344 vm_pindex_t diff, psize; 3345 3346 VM_OBJECT_LOCK_ASSERT(m_start->object, MA_OWNED); 3347 psize = atop(end - start); 3348 mpte = NULL; 3349 m = m_start; 3350 PMAP_LOCK(pmap); 3351 while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { 3352 va = start + ptoa(diff); 3353 if ((va & PDRMASK) == 0 && va + NBPDR <= end && 3354 (VM_PAGE_TO_PHYS(m) & PDRMASK) == 0 && 3355 pg_ps_enabled && vm_reserv_level_iffullpop(m) == 0 && 3356 pmap_enter_pde(pmap, va, m, prot)) 3357 m = &m[NBPDR / PAGE_SIZE - 1]; 3358 else 3359 mpte = pmap_enter_quick_locked(pmap, va, m, prot, 3360 mpte); 3361 m = TAILQ_NEXT(m, listq); 3362 } 3363 PMAP_UNLOCK(pmap); 3364} 3365 3366/* 3367 * this code makes some *MAJOR* assumptions: 3368 * 1. Current pmap & pmap exists. 3369 * 2. Not wired. 3370 * 3. Read access. 3371 * 4. No page table pages. 3372 * but is *MUCH* faster than pmap_enter... 3373 */ 3374 3375void 3376pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) 3377{ 3378 3379 PMAP_LOCK(pmap); 3380 (void) pmap_enter_quick_locked(pmap, va, m, prot, NULL); 3381 PMAP_UNLOCK(pmap); 3382} 3383 3384static vm_page_t 3385pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, 3386 vm_prot_t prot, vm_page_t mpte) 3387{ 3388 pt_entry_t *pte; 3389 vm_paddr_t pa; 3390 vm_page_t free; 3391 3392 KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva || 3393 (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0, 3394 ("pmap_enter_quick_locked: managed mapping within the clean submap")); 3395 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3396 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 3397 3398 /* 3399 * In the case that a page table page is not 3400 * resident, we are creating it here. 3401 */ 3402 if (va < VM_MAXUSER_ADDRESS) { 3403 unsigned ptepindex; 3404 pd_entry_t ptepa; 3405 3406 /* 3407 * Calculate pagetable page index 3408 */ 3409 ptepindex = va >> PDRSHIFT; 3410 if (mpte && (mpte->pindex == ptepindex)) { 3411 mpte->wire_count++; 3412 } else { 3413 /* 3414 * Get the page directory entry 3415 */ 3416 ptepa = pmap->pm_pdir[ptepindex]; 3417 3418 /* 3419 * If the page table page is mapped, we just increment 3420 * the hold count, and activate it. 3421 */ 3422 if (ptepa) { 3423 if (ptepa & PG_PS) 3424 return (NULL); 3425 mpte = PHYS_TO_VM_PAGE(ptepa & PG_FRAME); 3426 mpte->wire_count++; 3427 } else { 3428 mpte = _pmap_allocpte(pmap, ptepindex, 3429 M_NOWAIT); 3430 if (mpte == NULL) 3431 return (mpte); 3432 } 3433 } 3434 } else { 3435 mpte = NULL; 3436 } 3437 3438 /* 3439 * This call to vtopte makes the assumption that we are 3440 * entering the page into the current pmap. In order to support 3441 * quick entry into any pmap, one would likely use pmap_pte_quick. 3442 * But that isn't as quick as vtopte. 3443 */ 3444 pte = vtopte(va); 3445 if (*pte) { 3446 if (mpte != NULL) { 3447 mpte->wire_count--; 3448 mpte = NULL; 3449 } 3450 return (mpte); 3451 } 3452 3453 /* 3454 * Enter on the PV list if part of our managed memory. 3455 */ 3456 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0 && 3457 !pmap_try_insert_pv_entry(pmap, va, m)) { 3458 if (mpte != NULL) { 3459 free = NULL; 3460 if (pmap_unwire_pte_hold(pmap, mpte, &free)) { 3461 pmap_invalidate_page(pmap, va); 3462 pmap_free_zero_pages(free); 3463 } 3464 3465 mpte = NULL; 3466 } 3467 return (mpte); 3468 } 3469 3470 /* 3471 * Increment counters 3472 */ 3473 pmap->pm_stats.resident_count++; 3474 3475 pa = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(m->md.pat_mode, 0); 3476#ifdef PAE 3477 if ((prot & VM_PROT_EXECUTE) == 0) 3478 pa |= pg_nx; 3479#endif 3480 3481 /* 3482 * Now validate mapping with RO protection 3483 */ 3484 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 3485 pte_store(pte, pa | PG_V | PG_U); 3486 else 3487 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 3488 return mpte; 3489} 3490 3491/* 3492 * Make a temporary mapping for a physical address. This is only intended 3493 * to be used for panic dumps. 3494 */ 3495void * 3496pmap_kenter_temporary(vm_paddr_t pa, int i) 3497{ 3498 vm_offset_t va; 3499 3500 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 3501 pmap_kenter(va, pa); 3502 invlpg(va); 3503 return ((void *)crashdumpmap); 3504} 3505 3506/* 3507 * This code maps large physical mmap regions into the 3508 * processor address space. Note that some shortcuts 3509 * are taken, but the code works. 3510 */ 3511void 3512pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object, 3513 vm_pindex_t pindex, vm_size_t size) 3514{ 3515 pd_entry_t *pde; 3516 vm_paddr_t pa, ptepa; 3517 vm_page_t p; 3518 int pat_mode; 3519 3520 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 3521 KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, 3522 ("pmap_object_init_pt: non-device object")); 3523 if (pseflag && 3524 (addr & (NBPDR - 1)) == 0 && (size & (NBPDR - 1)) == 0) { 3525 if (!vm_object_populate(object, pindex, pindex + atop(size))) 3526 return; 3527 p = vm_page_lookup(object, pindex); 3528 KASSERT(p->valid == VM_PAGE_BITS_ALL, 3529 ("pmap_object_init_pt: invalid page %p", p)); 3530 pat_mode = p->md.pat_mode; 3531 3532 /* 3533 * Abort the mapping if the first page is not physically 3534 * aligned to a 2/4MB page boundary. 3535 */ 3536 ptepa = VM_PAGE_TO_PHYS(p); 3537 if (ptepa & (NBPDR - 1)) 3538 return; 3539 3540 /* 3541 * Skip the first page. Abort the mapping if the rest of 3542 * the pages are not physically contiguous or have differing 3543 * memory attributes. 3544 */ 3545 p = TAILQ_NEXT(p, listq); 3546 for (pa = ptepa + PAGE_SIZE; pa < ptepa + size; 3547 pa += PAGE_SIZE) { 3548 KASSERT(p->valid == VM_PAGE_BITS_ALL, 3549 ("pmap_object_init_pt: invalid page %p", p)); 3550 if (pa != VM_PAGE_TO_PHYS(p) || 3551 pat_mode != p->md.pat_mode) 3552 return; 3553 p = TAILQ_NEXT(p, listq); 3554 } 3555 3556 /* 3557 * Map using 2/4MB pages. Since "ptepa" is 2/4M aligned and 3558 * "size" is a multiple of 2/4M, adding the PAT setting to 3559 * "pa" will not affect the termination of this loop. 3560 */ 3561 PMAP_LOCK(pmap); 3562 for (pa = ptepa | pmap_cache_bits(pat_mode, 1); pa < ptepa + 3563 size; pa += NBPDR) { 3564 pde = pmap_pde(pmap, addr); 3565 if (*pde == 0) { 3566 pde_store(pde, pa | PG_PS | PG_M | PG_A | 3567 PG_U | PG_RW | PG_V); 3568 pmap->pm_stats.resident_count += NBPDR / 3569 PAGE_SIZE; 3570 pmap_pde_mappings++; 3571 } 3572 /* Else continue on if the PDE is already valid. */ 3573 addr += NBPDR; 3574 } 3575 PMAP_UNLOCK(pmap); 3576 } 3577} 3578 3579/* 3580 * Routine: pmap_change_wiring 3581 * Function: Change the wiring attribute for a map/virtual-address 3582 * pair. 3583 * In/out conditions: 3584 * The mapping must already exist in the pmap. 3585 */ 3586void 3587pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired) 3588{ 3589 pd_entry_t *pde; 3590 pt_entry_t *pte; 3591 boolean_t are_queues_locked; 3592 3593 are_queues_locked = FALSE; 3594retry: 3595 PMAP_LOCK(pmap); 3596 pde = pmap_pde(pmap, va); 3597 if ((*pde & PG_PS) != 0) { 3598 if (!wired != ((*pde & PG_W) == 0)) { 3599 if (!are_queues_locked) { 3600 are_queues_locked = TRUE; 3601 if (!mtx_trylock(&vm_page_queue_mtx)) { 3602 PMAP_UNLOCK(pmap); 3603 vm_page_lock_queues(); 3604 goto retry; 3605 } 3606 } 3607 if (!pmap_demote_pde(pmap, pde, va)) 3608 panic("pmap_change_wiring: demotion failed"); 3609 } else 3610 goto out; 3611 } 3612 pte = pmap_pte(pmap, va); 3613 3614 if (wired && !pmap_pte_w(pte)) 3615 pmap->pm_stats.wired_count++; 3616 else if (!wired && pmap_pte_w(pte)) 3617 pmap->pm_stats.wired_count--; 3618 3619 /* 3620 * Wiring is not a hardware characteristic so there is no need to 3621 * invalidate TLB. 3622 */ 3623 pmap_pte_set_w(pte, wired); 3624 pmap_pte_release(pte); 3625out: 3626 if (are_queues_locked) 3627 vm_page_unlock_queues(); 3628 PMAP_UNLOCK(pmap); 3629} 3630 3631 3632 3633/* 3634 * Copy the range specified by src_addr/len 3635 * from the source map to the range dst_addr/len 3636 * in the destination map. 3637 * 3638 * This routine is only advisory and need not do anything. 3639 */ 3640 3641void 3642pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 3643 vm_offset_t src_addr) 3644{ 3645 vm_page_t free; 3646 vm_offset_t addr; 3647 vm_offset_t end_addr = src_addr + len; 3648 vm_offset_t pdnxt; 3649 3650 if (dst_addr != src_addr) 3651 return; 3652 3653 if (!pmap_is_current(src_pmap)) 3654 return; 3655 3656 vm_page_lock_queues(); 3657 if (dst_pmap < src_pmap) { 3658 PMAP_LOCK(dst_pmap); 3659 PMAP_LOCK(src_pmap); 3660 } else { 3661 PMAP_LOCK(src_pmap); 3662 PMAP_LOCK(dst_pmap); 3663 } 3664 sched_pin(); 3665 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 3666 pt_entry_t *src_pte, *dst_pte; 3667 vm_page_t dstmpte, srcmpte; 3668 pd_entry_t srcptepaddr; 3669 unsigned ptepindex; 3670 3671 KASSERT(addr < UPT_MIN_ADDRESS, 3672 ("pmap_copy: invalid to pmap_copy page tables")); 3673 3674 pdnxt = (addr + NBPDR) & ~PDRMASK; 3675 if (pdnxt < addr) 3676 pdnxt = end_addr; 3677 ptepindex = addr >> PDRSHIFT; 3678 3679 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 3680 if (srcptepaddr == 0) 3681 continue; 3682 3683 if (srcptepaddr & PG_PS) { 3684 if (dst_pmap->pm_pdir[ptepindex] == 0 && 3685 ((srcptepaddr & PG_MANAGED) == 0 || 3686 pmap_pv_insert_pde(dst_pmap, addr, srcptepaddr & 3687 PG_PS_FRAME))) { 3688 dst_pmap->pm_pdir[ptepindex] = srcptepaddr & 3689 ~PG_W; 3690 dst_pmap->pm_stats.resident_count += 3691 NBPDR / PAGE_SIZE; 3692 } 3693 continue; 3694 } 3695 3696 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr & PG_FRAME); 3697 KASSERT(srcmpte->wire_count > 0, 3698 ("pmap_copy: source page table page is unused")); 3699 3700 if (pdnxt > end_addr) 3701 pdnxt = end_addr; 3702 3703 src_pte = vtopte(addr); 3704 while (addr < pdnxt) { 3705 pt_entry_t ptetemp; 3706 ptetemp = *src_pte; 3707 /* 3708 * we only virtual copy managed pages 3709 */ 3710 if ((ptetemp & PG_MANAGED) != 0) { 3711 dstmpte = pmap_allocpte(dst_pmap, addr, 3712 M_NOWAIT); 3713 if (dstmpte == NULL) 3714 goto out; 3715 dst_pte = pmap_pte_quick(dst_pmap, addr); 3716 if (*dst_pte == 0 && 3717 pmap_try_insert_pv_entry(dst_pmap, addr, 3718 PHYS_TO_VM_PAGE(ptetemp & PG_FRAME))) { 3719 /* 3720 * Clear the wired, modified, and 3721 * accessed (referenced) bits 3722 * during the copy. 3723 */ 3724 *dst_pte = ptetemp & ~(PG_W | PG_M | 3725 PG_A); 3726 dst_pmap->pm_stats.resident_count++; 3727 } else { 3728 free = NULL; 3729 if (pmap_unwire_pte_hold(dst_pmap, 3730 dstmpte, &free)) { 3731 pmap_invalidate_page(dst_pmap, 3732 addr); 3733 pmap_free_zero_pages(free); 3734 } 3735 goto out; 3736 } 3737 if (dstmpte->wire_count >= srcmpte->wire_count) 3738 break; 3739 } 3740 addr += PAGE_SIZE; 3741 src_pte++; 3742 } 3743 } 3744out: 3745 sched_unpin(); 3746 vm_page_unlock_queues(); 3747 PMAP_UNLOCK(src_pmap); 3748 PMAP_UNLOCK(dst_pmap); 3749} 3750 3751static __inline void 3752pagezero(void *page) 3753{ 3754#if defined(I686_CPU) 3755 if (cpu_class == CPUCLASS_686) { 3756#if defined(CPU_ENABLE_SSE) 3757 if (cpu_feature & CPUID_SSE2) 3758 sse2_pagezero(page); 3759 else 3760#endif 3761 i686_pagezero(page); 3762 } else 3763#endif 3764 bzero(page, PAGE_SIZE); 3765} 3766 3767/* 3768 * pmap_zero_page zeros the specified hardware page by mapping 3769 * the page into KVM and using bzero to clear its contents. 3770 */ 3771void 3772pmap_zero_page(vm_page_t m) 3773{ 3774 struct sysmaps *sysmaps; 3775 3776 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 3777 mtx_lock(&sysmaps->lock); 3778 if (*sysmaps->CMAP2) 3779 panic("pmap_zero_page: CMAP2 busy"); 3780 sched_pin(); 3781 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M | 3782 pmap_cache_bits(m->md.pat_mode, 0); 3783 invlcaddr(sysmaps->CADDR2); 3784 pagezero(sysmaps->CADDR2); 3785 *sysmaps->CMAP2 = 0; 3786 sched_unpin(); 3787 mtx_unlock(&sysmaps->lock); 3788} 3789 3790/* 3791 * pmap_zero_page_area zeros the specified hardware page by mapping 3792 * the page into KVM and using bzero to clear its contents. 3793 * 3794 * off and size may not cover an area beyond a single hardware page. 3795 */ 3796void 3797pmap_zero_page_area(vm_page_t m, int off, int size) 3798{ 3799 struct sysmaps *sysmaps; 3800 3801 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 3802 mtx_lock(&sysmaps->lock); 3803 if (*sysmaps->CMAP2) 3804 panic("pmap_zero_page_area: CMAP2 busy"); 3805 sched_pin(); 3806 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M | 3807 pmap_cache_bits(m->md.pat_mode, 0); 3808 invlcaddr(sysmaps->CADDR2); 3809 if (off == 0 && size == PAGE_SIZE) 3810 pagezero(sysmaps->CADDR2); 3811 else 3812 bzero((char *)sysmaps->CADDR2 + off, size); 3813 *sysmaps->CMAP2 = 0; 3814 sched_unpin(); 3815 mtx_unlock(&sysmaps->lock); 3816} 3817 3818/* 3819 * pmap_zero_page_idle zeros the specified hardware page by mapping 3820 * the page into KVM and using bzero to clear its contents. This 3821 * is intended to be called from the vm_pagezero process only and 3822 * outside of Giant. 3823 */ 3824void 3825pmap_zero_page_idle(vm_page_t m) 3826{ 3827 3828 if (*CMAP3) 3829 panic("pmap_zero_page_idle: CMAP3 busy"); 3830 sched_pin(); 3831 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M | 3832 pmap_cache_bits(m->md.pat_mode, 0); 3833 invlcaddr(CADDR3); 3834 pagezero(CADDR3); 3835 *CMAP3 = 0; 3836 sched_unpin(); 3837} 3838 3839/* 3840 * pmap_copy_page copies the specified (machine independent) 3841 * page by mapping the page into virtual memory and using 3842 * bcopy to copy the page, one machine dependent page at a 3843 * time. 3844 */ 3845void 3846pmap_copy_page(vm_page_t src, vm_page_t dst) 3847{ 3848 struct sysmaps *sysmaps; 3849 3850 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 3851 mtx_lock(&sysmaps->lock); 3852 if (*sysmaps->CMAP1) 3853 panic("pmap_copy_page: CMAP1 busy"); 3854 if (*sysmaps->CMAP2) 3855 panic("pmap_copy_page: CMAP2 busy"); 3856 sched_pin(); 3857 invlpg((u_int)sysmaps->CADDR1); 3858 invlpg((u_int)sysmaps->CADDR2); 3859 *sysmaps->CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A | 3860 pmap_cache_bits(src->md.pat_mode, 0); 3861 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M | 3862 pmap_cache_bits(dst->md.pat_mode, 0); 3863 bcopy(sysmaps->CADDR1, sysmaps->CADDR2, PAGE_SIZE); 3864 *sysmaps->CMAP1 = 0; 3865 *sysmaps->CMAP2 = 0; 3866 sched_unpin(); 3867 mtx_unlock(&sysmaps->lock); 3868} 3869 3870/* 3871 * Returns true if the pmap's pv is one of the first 3872 * 16 pvs linked to from this page. This count may 3873 * be changed upwards or downwards in the future; it 3874 * is only necessary that true be returned for a small 3875 * subset of pmaps for proper page aging. 3876 */ 3877boolean_t 3878pmap_page_exists_quick(pmap_t pmap, vm_page_t m) 3879{ 3880 struct md_page *pvh; 3881 pv_entry_t pv; 3882 int loops = 0; 3883 3884 if (m->flags & PG_FICTITIOUS) 3885 return FALSE; 3886 3887 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3888 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3889 if (PV_PMAP(pv) == pmap) { 3890 return TRUE; 3891 } 3892 loops++; 3893 if (loops >= 16) 3894 break; 3895 } 3896 if (loops < 16) { 3897 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); 3898 TAILQ_FOREACH(pv, &pvh->pv_list, pv_list) { 3899 if (PV_PMAP(pv) == pmap) 3900 return (TRUE); 3901 loops++; 3902 if (loops >= 16) 3903 break; 3904 } 3905 } 3906 return (FALSE); 3907} 3908 3909/* 3910 * pmap_page_wired_mappings: 3911 * 3912 * Return the number of managed mappings to the given physical page 3913 * that are wired. 3914 */ 3915int 3916pmap_page_wired_mappings(vm_page_t m) 3917{ 3918 int count; 3919 3920 count = 0; 3921 if ((m->flags & PG_FICTITIOUS) != 0) 3922 return (count); 3923 count = pmap_pvh_wired_mappings(&m->md, count); 3924 return (pmap_pvh_wired_mappings(pa_to_pvh(VM_PAGE_TO_PHYS(m)), count)); 3925} 3926 3927/* 3928 * pmap_pvh_wired_mappings: 3929 * 3930 * Return the updated number "count" of managed mappings that are wired. 3931 */ 3932static int 3933pmap_pvh_wired_mappings(struct md_page *pvh, int count) 3934{ 3935 pmap_t pmap; 3936 pt_entry_t *pte; 3937 pv_entry_t pv; 3938 3939 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3940 sched_pin(); 3941 TAILQ_FOREACH(pv, &pvh->pv_list, pv_list) { 3942 pmap = PV_PMAP(pv); 3943 PMAP_LOCK(pmap); 3944 pte = pmap_pte_quick(pmap, pv->pv_va); 3945 if ((*pte & PG_W) != 0) 3946 count++; 3947 PMAP_UNLOCK(pmap); 3948 } 3949 sched_unpin(); 3950 return (count); 3951} 3952 3953/* 3954 * Returns TRUE if the given page is mapped individually or as part of 3955 * a 4mpage. Otherwise, returns FALSE. 3956 */ 3957boolean_t 3958pmap_page_is_mapped(vm_page_t m) 3959{ 3960 struct md_page *pvh; 3961 3962 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0) 3963 return (FALSE); 3964 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3965 if (TAILQ_EMPTY(&m->md.pv_list)) { 3966 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); 3967 return (!TAILQ_EMPTY(&pvh->pv_list)); 3968 } else 3969 return (TRUE); 3970} 3971 3972/* 3973 * Remove all pages from specified address space 3974 * this aids process exit speeds. Also, this code 3975 * is special cased for current process only, but 3976 * can have the more generic (and slightly slower) 3977 * mode enabled. This is much faster than pmap_remove 3978 * in the case of running down an entire address space. 3979 */ 3980void 3981pmap_remove_pages(pmap_t pmap) 3982{ 3983 pt_entry_t *pte, tpte; 3984 vm_page_t free = NULL; 3985 vm_page_t m, mpte, mt; 3986 pv_entry_t pv; 3987 struct md_page *pvh; 3988 struct pv_chunk *pc, *npc; 3989 int field, idx; 3990 int32_t bit; 3991 uint32_t inuse, bitmask; 3992 int allfree; 3993 3994 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) { 3995 printf("warning: pmap_remove_pages called with non-current pmap\n"); 3996 return; 3997 } 3998 vm_page_lock_queues(); 3999 PMAP_LOCK(pmap); 4000 sched_pin(); 4001 TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) { 4002 allfree = 1; 4003 for (field = 0; field < _NPCM; field++) { 4004 inuse = (~(pc->pc_map[field])) & pc_freemask[field]; 4005 while (inuse != 0) { 4006 bit = bsfl(inuse); 4007 bitmask = 1UL << bit; 4008 idx = field * 32 + bit; 4009 pv = &pc->pc_pventry[idx]; 4010 inuse &= ~bitmask; 4011 4012 pte = pmap_pde(pmap, pv->pv_va); 4013 tpte = *pte; 4014 if ((tpte & PG_PS) == 0) { 4015 pte = vtopte(pv->pv_va); 4016 tpte = *pte & ~PG_PTE_PAT; 4017 } 4018 4019 if (tpte == 0) { 4020 printf( 4021 "TPTE at %p IS ZERO @ VA %08x\n", 4022 pte, pv->pv_va); 4023 panic("bad pte"); 4024 } 4025 4026/* 4027 * We cannot remove wired pages from a process' mapping at this time 4028 */ 4029 if (tpte & PG_W) { 4030 allfree = 0; 4031 continue; 4032 } 4033 4034 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME); 4035 KASSERT(m->phys_addr == (tpte & PG_FRAME), 4036 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 4037 m, (uintmax_t)m->phys_addr, 4038 (uintmax_t)tpte)); 4039 4040 KASSERT(m < &vm_page_array[vm_page_array_size], 4041 ("pmap_remove_pages: bad tpte %#jx", 4042 (uintmax_t)tpte)); 4043 4044 pte_clear(pte); 4045 4046 /* 4047 * Update the vm_page_t clean/reference bits. 4048 */ 4049 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { 4050 if ((tpte & PG_PS) != 0) { 4051 for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) 4052 vm_page_dirty(mt); 4053 } else 4054 vm_page_dirty(m); 4055 } 4056 4057 /* Mark free */ 4058 PV_STAT(pv_entry_frees++); 4059 PV_STAT(pv_entry_spare++); 4060 pv_entry_count--; 4061 pc->pc_map[field] |= bitmask; 4062 if ((tpte & PG_PS) != 0) { 4063 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 4064 pvh = pa_to_pvh(tpte & PG_PS_FRAME); 4065 TAILQ_REMOVE(&pvh->pv_list, pv, pv_list); 4066 if (TAILQ_EMPTY(&pvh->pv_list)) { 4067 for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) 4068 if (TAILQ_EMPTY(&mt->md.pv_list)) 4069 vm_page_flag_clear(mt, PG_WRITEABLE); 4070 } 4071 mpte = pmap_lookup_pt_page(pmap, pv->pv_va); 4072 if (mpte != NULL) { 4073 pmap_remove_pt_page(pmap, mpte); 4074 pmap->pm_stats.resident_count--; 4075 KASSERT(mpte->wire_count == NPTEPG, 4076 ("pmap_remove_pages: pte page wire count error")); 4077 mpte->wire_count = 0; 4078 pmap_add_delayed_free_list(mpte, &free, FALSE); 4079 atomic_subtract_int(&cnt.v_wire_count, 1); 4080 } 4081 } else { 4082 pmap->pm_stats.resident_count--; 4083 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 4084 if (TAILQ_EMPTY(&m->md.pv_list)) { 4085 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); 4086 if (TAILQ_EMPTY(&pvh->pv_list)) 4087 vm_page_flag_clear(m, PG_WRITEABLE); 4088 } 4089 pmap_unuse_pt(pmap, pv->pv_va, &free); 4090 } 4091 } 4092 } 4093 if (allfree) { 4094 PV_STAT(pv_entry_spare -= _NPCPV); 4095 PV_STAT(pc_chunk_count--); 4096 PV_STAT(pc_chunk_frees++); 4097 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); 4098 m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc)); 4099 pmap_qremove((vm_offset_t)pc, 1); 4100 vm_page_unwire(m, 0); 4101 vm_page_free(m); 4102 pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc); 4103 } 4104 } 4105 sched_unpin(); 4106 pmap_invalidate_all(pmap); 4107 vm_page_unlock_queues(); 4108 PMAP_UNLOCK(pmap); 4109 pmap_free_zero_pages(free); 4110} 4111 4112/* 4113 * pmap_is_modified: 4114 * 4115 * Return whether or not the specified physical page was modified 4116 * in any physical maps. 4117 */ 4118boolean_t 4119pmap_is_modified(vm_page_t m) 4120{ 4121 4122 if (m->flags & PG_FICTITIOUS) 4123 return (FALSE); 4124 if (pmap_is_modified_pvh(&m->md)) 4125 return (TRUE); 4126 return (pmap_is_modified_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m)))); 4127} 4128 4129/* 4130 * Returns TRUE if any of the given mappings were used to modify 4131 * physical memory. Otherwise, returns FALSE. Both page and 2mpage 4132 * mappings are supported. 4133 */ 4134static boolean_t 4135pmap_is_modified_pvh(struct md_page *pvh) 4136{ 4137 pv_entry_t pv; 4138 pt_entry_t *pte; 4139 pmap_t pmap; 4140 boolean_t rv; 4141 4142 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 4143 rv = FALSE; 4144 sched_pin(); 4145 TAILQ_FOREACH(pv, &pvh->pv_list, pv_list) { 4146 pmap = PV_PMAP(pv); 4147 PMAP_LOCK(pmap); 4148 pte = pmap_pte_quick(pmap, pv->pv_va); 4149 rv = (*pte & (PG_M | PG_RW)) == (PG_M | PG_RW); 4150 PMAP_UNLOCK(pmap); 4151 if (rv) 4152 break; 4153 } 4154 sched_unpin(); 4155 return (rv); 4156} 4157 4158/* 4159 * pmap_is_prefaultable: 4160 * 4161 * Return whether or not the specified virtual address is elgible 4162 * for prefault. 4163 */ 4164boolean_t 4165pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 4166{ 4167 pd_entry_t *pde; 4168 pt_entry_t *pte; 4169 boolean_t rv; 4170 4171 rv = FALSE; 4172 PMAP_LOCK(pmap); 4173 pde = pmap_pde(pmap, addr); 4174 if (*pde != 0 && (*pde & PG_PS) == 0) { 4175 pte = vtopte(addr); 4176 rv = *pte == 0; 4177 } 4178 PMAP_UNLOCK(pmap); 4179 return (rv); 4180} 4181 4182/* 4183 * Clear the write and modified bits in each of the given page's mappings. 4184 */ 4185void 4186pmap_remove_write(vm_page_t m) 4187{ 4188 struct md_page *pvh; 4189 pv_entry_t next_pv, pv; 4190 pmap_t pmap; 4191 pd_entry_t *pde; 4192 pt_entry_t oldpte, *pte; 4193 vm_offset_t va; 4194 4195 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 4196 if ((m->flags & PG_FICTITIOUS) != 0 || 4197 (m->flags & PG_WRITEABLE) == 0) 4198 return; 4199 sched_pin(); 4200 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); 4201 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_list, next_pv) { 4202 va = pv->pv_va; 4203 pmap = PV_PMAP(pv); 4204 PMAP_LOCK(pmap); 4205 pde = pmap_pde(pmap, va); 4206 if ((*pde & PG_RW) != 0) 4207 (void)pmap_demote_pde(pmap, pde, va); 4208 PMAP_UNLOCK(pmap); 4209 } 4210 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 4211 pmap = PV_PMAP(pv); 4212 PMAP_LOCK(pmap); 4213 pde = pmap_pde(pmap, pv->pv_va); 4214 KASSERT((*pde & PG_PS) == 0, ("pmap_clear_write: found" 4215 " a 4mpage in page %p's pv list", m)); 4216 pte = pmap_pte_quick(pmap, pv->pv_va); 4217retry: 4218 oldpte = *pte; 4219 if ((oldpte & PG_RW) != 0) { 4220 /* 4221 * Regardless of whether a pte is 32 or 64 bits 4222 * in size, PG_RW and PG_M are among the least 4223 * significant 32 bits. 4224 */ 4225 if (!atomic_cmpset_int((u_int *)pte, oldpte, 4226 oldpte & ~(PG_RW | PG_M))) 4227 goto retry; 4228 if ((oldpte & PG_M) != 0) 4229 vm_page_dirty(m); 4230 pmap_invalidate_page(pmap, pv->pv_va); 4231 } 4232 PMAP_UNLOCK(pmap); 4233 } 4234 vm_page_flag_clear(m, PG_WRITEABLE); 4235 sched_unpin(); 4236} 4237 4238/* 4239 * pmap_ts_referenced: 4240 * 4241 * Return a count of reference bits for a page, clearing those bits. 4242 * It is not necessary for every reference bit to be cleared, but it 4243 * is necessary that 0 only be returned when there are truly no 4244 * reference bits set. 4245 * 4246 * XXX: The exact number of bits to check and clear is a matter that 4247 * should be tested and standardized at some point in the future for 4248 * optimal aging of shared pages. 4249 */ 4250int 4251pmap_ts_referenced(vm_page_t m) 4252{ 4253 struct md_page *pvh; 4254 pv_entry_t pv, pvf, pvn; 4255 pmap_t pmap; 4256 pd_entry_t oldpde, *pde; 4257 pt_entry_t *pte; 4258 vm_offset_t va; 4259 int rtval = 0; 4260 4261 if (m->flags & PG_FICTITIOUS) 4262 return (rtval); 4263 sched_pin(); 4264 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 4265 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); 4266 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_list, pvn) { 4267 va = pv->pv_va; 4268 pmap = PV_PMAP(pv); 4269 PMAP_LOCK(pmap); 4270 pde = pmap_pde(pmap, va); 4271 oldpde = *pde; 4272 if ((oldpde & PG_A) != 0) { 4273 if (pmap_demote_pde(pmap, pde, va)) { 4274 if ((oldpde & PG_W) == 0) { 4275 /* 4276 * Remove the mapping to a single page 4277 * so that a subsequent access may 4278 * repromote. Since the underlying 4279 * page table page is fully populated, 4280 * this removal never frees a page 4281 * table page. 4282 */ 4283 va += VM_PAGE_TO_PHYS(m) - (oldpde & 4284 PG_PS_FRAME); 4285 pmap_remove_page(pmap, va, NULL); 4286 rtval++; 4287 if (rtval > 4) { 4288 PMAP_UNLOCK(pmap); 4289 return (rtval); 4290 } 4291 } 4292 } 4293 } 4294 PMAP_UNLOCK(pmap); 4295 } 4296 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 4297 pvf = pv; 4298 do { 4299 pvn = TAILQ_NEXT(pv, pv_list); 4300 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 4301 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 4302 pmap = PV_PMAP(pv); 4303 PMAP_LOCK(pmap); 4304 pde = pmap_pde(pmap, pv->pv_va); 4305 KASSERT((*pde & PG_PS) == 0, ("pmap_ts_referenced:" 4306 " found a 4mpage in page %p's pv list", m)); 4307 pte = pmap_pte_quick(pmap, pv->pv_va); 4308 if ((*pte & PG_A) != 0) { 4309 atomic_clear_int((u_int *)pte, PG_A); 4310 pmap_invalidate_page(pmap, pv->pv_va); 4311 rtval++; 4312 if (rtval > 4) 4313 pvn = NULL; 4314 } 4315 PMAP_UNLOCK(pmap); 4316 } while ((pv = pvn) != NULL && pv != pvf); 4317 } 4318 sched_unpin(); 4319 return (rtval); 4320} 4321 4322/* 4323 * Clear the modify bits on the specified physical page. 4324 */ 4325void 4326pmap_clear_modify(vm_page_t m) 4327{ 4328 struct md_page *pvh; 4329 pv_entry_t next_pv, pv; 4330 pmap_t pmap; 4331 pd_entry_t oldpde, *pde; 4332 pt_entry_t oldpte, *pte; 4333 vm_offset_t va; 4334 4335 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 4336 if ((m->flags & PG_FICTITIOUS) != 0) 4337 return; 4338 sched_pin(); 4339 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); 4340 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_list, next_pv) { 4341 va = pv->pv_va; 4342 pmap = PV_PMAP(pv); 4343 PMAP_LOCK(pmap); 4344 pde = pmap_pde(pmap, va); 4345 oldpde = *pde; 4346 if ((oldpde & PG_RW) != 0) { 4347 if (pmap_demote_pde(pmap, pde, va)) { 4348 if ((oldpde & PG_W) == 0) { 4349 /* 4350 * Write protect the mapping to a 4351 * single page so that a subsequent 4352 * write access may repromote. 4353 */ 4354 va += VM_PAGE_TO_PHYS(m) - (oldpde & 4355 PG_PS_FRAME); 4356 pte = pmap_pte_quick(pmap, va); 4357 oldpte = *pte; 4358 if ((oldpte & PG_V) != 0) { 4359 /* 4360 * Regardless of whether a pte is 32 or 64 bits 4361 * in size, PG_RW and PG_M are among the least 4362 * significant 32 bits. 4363 */ 4364 while (!atomic_cmpset_int((u_int *)pte, 4365 oldpte, 4366 oldpte & ~(PG_M | PG_RW))) 4367 oldpte = *pte; 4368 vm_page_dirty(m); 4369 pmap_invalidate_page(pmap, va); 4370 } 4371 } 4372 } 4373 } 4374 PMAP_UNLOCK(pmap); 4375 } 4376 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 4377 pmap = PV_PMAP(pv); 4378 PMAP_LOCK(pmap); 4379 pde = pmap_pde(pmap, pv->pv_va); 4380 KASSERT((*pde & PG_PS) == 0, ("pmap_clear_modify: found" 4381 " a 4mpage in page %p's pv list", m)); 4382 pte = pmap_pte_quick(pmap, pv->pv_va); 4383 if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { 4384 /* 4385 * Regardless of whether a pte is 32 or 64 bits 4386 * in size, PG_M is among the least significant 4387 * 32 bits. 4388 */ 4389 atomic_clear_int((u_int *)pte, PG_M); 4390 pmap_invalidate_page(pmap, pv->pv_va); 4391 } 4392 PMAP_UNLOCK(pmap); 4393 } 4394 sched_unpin(); 4395} 4396 4397/* 4398 * pmap_clear_reference: 4399 * 4400 * Clear the reference bit on the specified physical page. 4401 */ 4402void 4403pmap_clear_reference(vm_page_t m) 4404{ 4405 struct md_page *pvh; 4406 pv_entry_t next_pv, pv; 4407 pmap_t pmap; 4408 pd_entry_t oldpde, *pde; 4409 pt_entry_t *pte; 4410 vm_offset_t va; 4411 4412 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 4413 if ((m->flags & PG_FICTITIOUS) != 0) 4414 return; 4415 sched_pin(); 4416 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); 4417 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_list, next_pv) { 4418 va = pv->pv_va; 4419 pmap = PV_PMAP(pv); 4420 PMAP_LOCK(pmap); 4421 pde = pmap_pde(pmap, va); 4422 oldpde = *pde; 4423 if ((oldpde & PG_A) != 0) { 4424 if (pmap_demote_pde(pmap, pde, va)) { 4425 /* 4426 * Remove the mapping to a single page so 4427 * that a subsequent access may repromote. 4428 * Since the underlying page table page is 4429 * fully populated, this removal never frees 4430 * a page table page. 4431 */ 4432 va += VM_PAGE_TO_PHYS(m) - (oldpde & 4433 PG_PS_FRAME); 4434 pmap_remove_page(pmap, va, NULL); 4435 } 4436 } 4437 PMAP_UNLOCK(pmap); 4438 } 4439 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 4440 pmap = PV_PMAP(pv); 4441 PMAP_LOCK(pmap); 4442 pde = pmap_pde(pmap, pv->pv_va); 4443 KASSERT((*pde & PG_PS) == 0, ("pmap_clear_reference: found" 4444 " a 4mpage in page %p's pv list", m)); 4445 pte = pmap_pte_quick(pmap, pv->pv_va); 4446 if ((*pte & PG_A) != 0) { 4447 /* 4448 * Regardless of whether a pte is 32 or 64 bits 4449 * in size, PG_A is among the least significant 4450 * 32 bits. 4451 */ 4452 atomic_clear_int((u_int *)pte, PG_A); 4453 pmap_invalidate_page(pmap, pv->pv_va); 4454 } 4455 PMAP_UNLOCK(pmap); 4456 } 4457 sched_unpin(); 4458} 4459 4460/* 4461 * Miscellaneous support routines follow 4462 */ 4463 4464/* Adjust the cache mode for a 4KB page mapped via a PTE. */ 4465static __inline void 4466pmap_pte_attr(pt_entry_t *pte, int cache_bits) 4467{ 4468 u_int opte, npte; 4469 4470 /* 4471 * The cache mode bits are all in the low 32-bits of the 4472 * PTE, so we can just spin on updating the low 32-bits. 4473 */ 4474 do { 4475 opte = *(u_int *)pte; 4476 npte = opte & ~PG_PTE_CACHE; 4477 npte |= cache_bits; 4478 } while (npte != opte && !atomic_cmpset_int((u_int *)pte, opte, npte)); 4479} 4480 4481/* Adjust the cache mode for a 2/4MB page mapped via a PDE. */ 4482static __inline void 4483pmap_pde_attr(pd_entry_t *pde, int cache_bits) 4484{ 4485 u_int opde, npde; 4486 4487 /* 4488 * The cache mode bits are all in the low 32-bits of the 4489 * PDE, so we can just spin on updating the low 32-bits. 4490 */ 4491 do { 4492 opde = *(u_int *)pde; 4493 npde = opde & ~PG_PDE_CACHE; 4494 npde |= cache_bits; 4495 } while (npde != opde && !atomic_cmpset_int((u_int *)pde, opde, npde)); 4496} 4497 4498/* 4499 * Map a set of physical memory pages into the kernel virtual 4500 * address space. Return a pointer to where it is mapped. This 4501 * routine is intended to be used for mapping device memory, 4502 * NOT real memory. 4503 */ 4504void * 4505pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode) 4506{ 4507 vm_offset_t va, offset; 4508 vm_size_t tmpsize; 4509 4510 offset = pa & PAGE_MASK; 4511 size = roundup(offset + size, PAGE_SIZE); 4512 pa = pa & PG_FRAME; 4513 4514 if (pa < KERNLOAD && pa + size <= KERNLOAD) 4515 va = KERNBASE + pa; 4516 else 4517 va = kmem_alloc_nofault(kernel_map, size); 4518 if (!va) 4519 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 4520 4521 for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE) 4522 pmap_kenter_attr(va + tmpsize, pa + tmpsize, mode); 4523 pmap_invalidate_range(kernel_pmap, va, va + tmpsize); 4524 pmap_invalidate_cache_range(va, va + size); 4525 return ((void *)(va + offset)); 4526} 4527 4528void * 4529pmap_mapdev(vm_paddr_t pa, vm_size_t size) 4530{ 4531 4532 return (pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE)); 4533} 4534 4535void * 4536pmap_mapbios(vm_paddr_t pa, vm_size_t size) 4537{ 4538 4539 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK)); 4540} 4541 4542void 4543pmap_unmapdev(vm_offset_t va, vm_size_t size) 4544{ 4545 vm_offset_t base, offset, tmpva; 4546 4547 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 4548 return; 4549 base = trunc_page(va); 4550 offset = va & PAGE_MASK; 4551 size = roundup(offset + size, PAGE_SIZE); 4552 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 4553 pmap_kremove(tmpva); 4554 pmap_invalidate_range(kernel_pmap, va, tmpva); 4555 kmem_free(kernel_map, base, size); 4556} 4557 4558/* 4559 * Sets the memory attribute for the specified page. 4560 */ 4561void 4562pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma) 4563{ 4564 struct sysmaps *sysmaps; 4565 vm_offset_t sva, eva; 4566 4567 m->md.pat_mode = ma; 4568 if ((m->flags & PG_FICTITIOUS) != 0) 4569 return; 4570 4571 /* 4572 * If "m" is a normal page, flush it from the cache. 4573 * See pmap_invalidate_cache_range(). 4574 * 4575 * First, try to find an existing mapping of the page by sf 4576 * buffer. sf_buf_invalidate_cache() modifies mapping and 4577 * flushes the cache. 4578 */ 4579 if (sf_buf_invalidate_cache(m)) 4580 return; 4581 4582 /* 4583 * If page is not mapped by sf buffer, but CPU does not 4584 * support self snoop, map the page transient and do 4585 * invalidation. In the worst case, whole cache is flushed by 4586 * pmap_invalidate_cache_range(). 4587 */ 4588 if ((cpu_feature & (CPUID_SS|CPUID_CLFSH)) == CPUID_CLFSH) { 4589 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; 4590 mtx_lock(&sysmaps->lock); 4591 if (*sysmaps->CMAP2) 4592 panic("pmap_page_set_memattr: CMAP2 busy"); 4593 sched_pin(); 4594 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | 4595 PG_A | PG_M | pmap_cache_bits(m->md.pat_mode, 0); 4596 invlcaddr(sysmaps->CADDR2); 4597 sva = (vm_offset_t)sysmaps->CADDR2; 4598 eva = sva + PAGE_SIZE; 4599 } else 4600 sva = eva = 0; /* gcc */ 4601 pmap_invalidate_cache_range(sva, eva); 4602 if (sva != 0) { 4603 *sysmaps->CMAP2 = 0; 4604 sched_unpin(); 4605 mtx_unlock(&sysmaps->lock); 4606 } 4607} 4608 4609/* 4610 * Changes the specified virtual address range's memory type to that given by 4611 * the parameter "mode". The specified virtual address range must be 4612 * completely contained within either the kernel map. 4613 * 4614 * Returns zero if the change completed successfully, and either EINVAL or 4615 * ENOMEM if the change failed. Specifically, EINVAL is returned if some part 4616 * of the virtual address range was not mapped, and ENOMEM is returned if 4617 * there was insufficient memory available to complete the change. 4618 */ 4619int 4620pmap_change_attr(vm_offset_t va, vm_size_t size, int mode) 4621{ 4622 vm_offset_t base, offset, tmpva; 4623 pd_entry_t *pde; 4624 pt_entry_t *pte; 4625 int cache_bits_pte, cache_bits_pde; 4626 boolean_t changed; 4627 4628 base = trunc_page(va); 4629 offset = va & PAGE_MASK; 4630 size = roundup(offset + size, PAGE_SIZE); 4631 4632 /* 4633 * Only supported on kernel virtual addresses above the recursive map. 4634 */ 4635 if (base < VM_MIN_KERNEL_ADDRESS) 4636 return (EINVAL); 4637 4638 cache_bits_pde = pmap_cache_bits(mode, 1); 4639 cache_bits_pte = pmap_cache_bits(mode, 0); 4640 changed = FALSE; 4641 4642 /* 4643 * Pages that aren't mapped aren't supported. Also break down 4644 * 2/4MB pages into 4KB pages if required. 4645 */ 4646 PMAP_LOCK(kernel_pmap); 4647 for (tmpva = base; tmpva < base + size; ) { 4648 pde = pmap_pde(kernel_pmap, tmpva); 4649 if (*pde == 0) { 4650 PMAP_UNLOCK(kernel_pmap); 4651 return (EINVAL); 4652 } 4653 if (*pde & PG_PS) { 4654 /* 4655 * If the current 2/4MB page already has 4656 * the required memory type, then we need not 4657 * demote this page. Just increment tmpva to 4658 * the next 2/4MB page frame. 4659 */ 4660 if ((*pde & PG_PDE_CACHE) == cache_bits_pde) { 4661 tmpva = trunc_4mpage(tmpva) + NBPDR; 4662 continue; 4663 } 4664 4665 /* 4666 * If the current offset aligns with a 2/4MB 4667 * page frame and there is at least 2/4MB left 4668 * within the range, then we need not break 4669 * down this page into 4KB pages. 4670 */ 4671 if ((tmpva & PDRMASK) == 0 && 4672 tmpva + PDRMASK < base + size) { 4673 tmpva += NBPDR; 4674 continue; 4675 } 4676 if (!pmap_demote_pde(kernel_pmap, pde, tmpva)) { 4677 PMAP_UNLOCK(kernel_pmap); 4678 return (ENOMEM); 4679 } 4680 } 4681 pte = vtopte(tmpva); 4682 if (*pte == 0) { 4683 PMAP_UNLOCK(kernel_pmap); 4684 return (EINVAL); 4685 } 4686 tmpva += PAGE_SIZE; 4687 } 4688 PMAP_UNLOCK(kernel_pmap); 4689 4690 /* 4691 * Ok, all the pages exist, so run through them updating their 4692 * cache mode if required. 4693 */ 4694 for (tmpva = base; tmpva < base + size; ) { 4695 pde = pmap_pde(kernel_pmap, tmpva); 4696 if (*pde & PG_PS) { 4697 if ((*pde & PG_PDE_CACHE) != cache_bits_pde) { 4698 pmap_pde_attr(pde, cache_bits_pde); 4699 changed = TRUE; 4700 } 4701 tmpva = trunc_4mpage(tmpva) + NBPDR; 4702 } else { 4703 pte = vtopte(tmpva); 4704 if ((*pte & PG_PTE_CACHE) != cache_bits_pte) { 4705 pmap_pte_attr(pte, cache_bits_pte); 4706 changed = TRUE; 4707 } 4708 tmpva += PAGE_SIZE; 4709 } 4710 } 4711 4712 /* 4713 * Flush CPU caches to make sure any data isn't cached that 4714 * shouldn't be, etc. 4715 */ 4716 if (changed) { 4717 pmap_invalidate_range(kernel_pmap, base, tmpva); 4718 pmap_invalidate_cache_range(base, tmpva); 4719 } 4720 return (0); 4721} 4722 4723/* 4724 * perform the pmap work for mincore 4725 */ 4726int 4727pmap_mincore(pmap_t pmap, vm_offset_t addr) 4728{ 4729 pd_entry_t *pdep; 4730 pt_entry_t *ptep, pte; 4731 vm_paddr_t pa; 4732 vm_page_t m; 4733 int val = 0; 4734 4735 PMAP_LOCK(pmap); 4736 pdep = pmap_pde(pmap, addr); 4737 if (*pdep != 0) { 4738 if (*pdep & PG_PS) { 4739 pte = *pdep; 4740 val = MINCORE_SUPER; 4741 /* Compute the physical address of the 4KB page. */ 4742 pa = ((*pdep & PG_PS_FRAME) | (addr & PDRMASK)) & 4743 PG_FRAME; 4744 } else { 4745 ptep = pmap_pte(pmap, addr); 4746 pte = *ptep; 4747 pmap_pte_release(ptep); 4748 pa = pte & PG_FRAME; 4749 } 4750 } else { 4751 pte = 0; 4752 pa = 0; 4753 } 4754 PMAP_UNLOCK(pmap); 4755 4756 if (pte != 0) { 4757 val |= MINCORE_INCORE; 4758 if ((pte & PG_MANAGED) == 0) 4759 return val; 4760 4761 m = PHYS_TO_VM_PAGE(pa); 4762 4763 /* 4764 * Modified by us 4765 */ 4766 if ((pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) 4767 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 4768 else { 4769 /* 4770 * Modified by someone else 4771 */ 4772 vm_page_lock_queues(); 4773 if (m->dirty || pmap_is_modified(m)) 4774 val |= MINCORE_MODIFIED_OTHER; 4775 vm_page_unlock_queues(); 4776 } 4777 /* 4778 * Referenced by us 4779 */ 4780 if (pte & PG_A) 4781 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 4782 else { 4783 /* 4784 * Referenced by someone else 4785 */ 4786 vm_page_lock_queues(); 4787 if ((m->flags & PG_REFERENCED) || 4788 pmap_ts_referenced(m)) { 4789 val |= MINCORE_REFERENCED_OTHER; 4790 vm_page_flag_set(m, PG_REFERENCED); 4791 } 4792 vm_page_unlock_queues(); 4793 } 4794 } 4795 return val; 4796} 4797 4798void 4799pmap_activate(struct thread *td) 4800{ 4801 pmap_t pmap, oldpmap; 4802 u_int32_t cr3; 4803 4804 critical_enter(); 4805 pmap = vmspace_pmap(td->td_proc->p_vmspace); 4806 oldpmap = PCPU_GET(curpmap); 4807#if defined(SMP) 4808 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 4809 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 4810#else 4811 oldpmap->pm_active &= ~1; 4812 pmap->pm_active |= 1; 4813#endif 4814#ifdef PAE 4815 cr3 = vtophys(pmap->pm_pdpt); 4816#else 4817 cr3 = vtophys(pmap->pm_pdir); 4818#endif 4819 /* 4820 * pmap_activate is for the current thread on the current cpu 4821 */ 4822 td->td_pcb->pcb_cr3 = cr3; 4823 load_cr3(cr3); 4824 PCPU_SET(curpmap, pmap); 4825 critical_exit(); 4826} 4827 4828/* 4829 * Increase the starting virtual address of the given mapping if a 4830 * different alignment might result in more superpage mappings. 4831 */ 4832void 4833pmap_align_superpage(vm_object_t object, vm_ooffset_t offset, 4834 vm_offset_t *addr, vm_size_t size) 4835{ 4836 vm_offset_t superpage_offset; 4837 4838 if (size < NBPDR) 4839 return; 4840 if (object != NULL && (object->flags & OBJ_COLORED) != 0) 4841 offset += ptoa(object->pg_color); 4842 superpage_offset = offset & PDRMASK; 4843 if (size - ((NBPDR - superpage_offset) & PDRMASK) < NBPDR || 4844 (*addr & PDRMASK) == superpage_offset) 4845 return; 4846 if ((*addr & PDRMASK) < superpage_offset) 4847 *addr = (*addr & ~PDRMASK) + superpage_offset; 4848 else 4849 *addr = ((*addr + PDRMASK) & ~PDRMASK) + superpage_offset; 4850} 4851 4852 4853#if defined(PMAP_DEBUG) 4854pmap_pid_dump(int pid) 4855{ 4856 pmap_t pmap; 4857 struct proc *p; 4858 int npte = 0; 4859 int index; 4860 4861 sx_slock(&allproc_lock); 4862 FOREACH_PROC_IN_SYSTEM(p) { 4863 if (p->p_pid != pid) 4864 continue; 4865 4866 if (p->p_vmspace) { 4867 int i,j; 4868 index = 0; 4869 pmap = vmspace_pmap(p->p_vmspace); 4870 for (i = 0; i < NPDEPTD; i++) { 4871 pd_entry_t *pde; 4872 pt_entry_t *pte; 4873 vm_offset_t base = i << PDRSHIFT; 4874 4875 pde = &pmap->pm_pdir[i]; 4876 if (pde && pmap_pde_v(pde)) { 4877 for (j = 0; j < NPTEPG; j++) { 4878 vm_offset_t va = base + (j << PAGE_SHIFT); 4879 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 4880 if (index) { 4881 index = 0; 4882 printf("\n"); 4883 } 4884 sx_sunlock(&allproc_lock); 4885 return npte; 4886 } 4887 pte = pmap_pte(pmap, va); 4888 if (pte && pmap_pte_v(pte)) { 4889 pt_entry_t pa; 4890 vm_page_t m; 4891 pa = *pte; 4892 m = PHYS_TO_VM_PAGE(pa & PG_FRAME); 4893 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 4894 va, pa, m->hold_count, m->wire_count, m->flags); 4895 npte++; 4896 index++; 4897 if (index >= 2) { 4898 index = 0; 4899 printf("\n"); 4900 } else { 4901 printf(" "); 4902 } 4903 } 4904 } 4905 } 4906 } 4907 } 4908 } 4909 sx_sunlock(&allproc_lock); 4910 return npte; 4911} 4912#endif 4913 4914#if defined(DEBUG) 4915 4916static void pads(pmap_t pm); 4917void pmap_pvdump(vm_offset_t pa); 4918 4919/* print address space of pmap*/ 4920static void 4921pads(pmap_t pm) 4922{ 4923 int i, j; 4924 vm_paddr_t va; 4925 pt_entry_t *ptep; 4926 4927 if (pm == kernel_pmap) 4928 return; 4929 for (i = 0; i < NPDEPTD; i++) 4930 if (pm->pm_pdir[i]) 4931 for (j = 0; j < NPTEPG; j++) { 4932 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 4933 if (pm == kernel_pmap && va < KERNBASE) 4934 continue; 4935 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 4936 continue; 4937 ptep = pmap_pte(pm, va); 4938 if (pmap_pte_v(ptep)) 4939 printf("%x:%x ", va, *ptep); 4940 }; 4941 4942} 4943 4944void 4945pmap_pvdump(vm_paddr_t pa) 4946{ 4947 pv_entry_t pv; 4948 pmap_t pmap; 4949 vm_page_t m; 4950 4951 printf("pa %x", pa); 4952 m = PHYS_TO_VM_PAGE(pa); 4953 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 4954 pmap = PV_PMAP(pv); 4955 printf(" -> pmap %p, va %x", (void *)pmap, pv->pv_va); 4956 pads(pmap); 4957 } 4958 printf(" "); 4959} 4960#endif 4961