pmap.c revision 88742
1/* 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1994 John S. Dyson 5 * All rights reserved. 6 * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * the Systems Programming Group of the University of Utah Computer 11 * Science Department and William Jolitz of UUNET Technologies Inc. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by the University of 24 * California, Berkeley and its contributors. 25 * 4. Neither the name of the University nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 39 * SUCH DAMAGE. 40 * 41 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 42 * $FreeBSD: head/sys/amd64/amd64/pmap.c 88742 2001-12-31 20:02:46Z dillon $ 43 */ 44 45/* 46 * Manages physical address maps. 47 * 48 * In addition to hardware address maps, this 49 * module is called upon to provide software-use-only 50 * maps which may or may not be stored in the same 51 * form as hardware maps. These pseudo-maps are 52 * used to store intermediate results from copy 53 * operations to and from address spaces. 54 * 55 * Since the information managed by this module is 56 * also stored by the logical address mapping module, 57 * this module may throw away valid virtual-to-physical 58 * mappings at almost any time. However, invalidations 59 * of virtual-to-physical mappings must be done as 60 * requested. 61 * 62 * In order to cope with hardware architectures which 63 * make virtual-to-physical map invalidates expensive, 64 * this module may delay invalidate or reduced protection 65 * operations until such time as they are actually 66 * necessary. This module is given full information as 67 * to which processors are currently using which maps, 68 * and to when physical maps must be made correct. 69 */ 70 71#include "opt_disable_pse.h" 72#include "opt_pmap.h" 73#include "opt_msgbuf.h" 74#include "opt_kstack_pages.h" 75 76#include <sys/param.h> 77#include <sys/systm.h> 78#include <sys/kernel.h> 79#include <sys/lock.h> 80#include <sys/mman.h> 81#include <sys/msgbuf.h> 82#include <sys/mutex.h> 83#include <sys/proc.h> 84#include <sys/sx.h> 85#include <sys/user.h> 86#include <sys/vmmeter.h> 87#include <sys/sysctl.h> 88 89#include <vm/vm.h> 90#include <vm/vm_param.h> 91#include <vm/vm_kern.h> 92#include <vm/vm_page.h> 93#include <vm/vm_map.h> 94#include <vm/vm_object.h> 95#include <vm/vm_extern.h> 96#include <vm/vm_pageout.h> 97#include <vm/vm_pager.h> 98#include <vm/vm_zone.h> 99 100#include <machine/cputypes.h> 101#include <machine/md_var.h> 102#include <machine/specialreg.h> 103#if defined(SMP) || defined(APIC_IO) 104#include <machine/smp.h> 105#include <machine/apic.h> 106#include <machine/segments.h> 107#include <machine/tss.h> 108#endif /* SMP || APIC_IO */ 109 110#define PMAP_KEEP_PDIRS 111#ifndef PMAP_SHPGPERPROC 112#define PMAP_SHPGPERPROC 200 113#endif 114 115#if defined(DIAGNOSTIC) 116#define PMAP_DIAGNOSTIC 117#endif 118 119#define MINPV 2048 120 121#if !defined(PMAP_DIAGNOSTIC) 122#define PMAP_INLINE __inline 123#else 124#define PMAP_INLINE 125#endif 126 127/* 128 * Get PDEs and PTEs for user/kernel address space 129 */ 130#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 131#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 132 133#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 134#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 135#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 136#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 137#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 138 139#define pmap_pte_set_w(pte, v) ((v)?(*(int *)pte |= PG_W):(*(int *)pte &= ~PG_W)) 140#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 141 142/* 143 * Given a map and a machine independent protection code, 144 * convert to a vax protection code. 145 */ 146#define pte_prot(m, p) (protection_codes[p]) 147static int protection_codes[8]; 148 149static struct pmap kernel_pmap_store; 150pmap_t kernel_pmap; 151LIST_HEAD(pmaplist, pmap); 152struct pmaplist allpmaps; 153 154vm_offset_t avail_start; /* PA of first available physical page */ 155vm_offset_t avail_end; /* PA of last available physical page */ 156vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 157vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 158static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 159static int pgeflag; /* PG_G or-in */ 160static int pseflag; /* PG_PS or-in */ 161 162static vm_object_t kptobj; 163 164static int nkpt; 165vm_offset_t kernel_vm_end; 166 167/* 168 * Data for the pv entry allocation mechanism 169 */ 170static vm_zone_t pvzone; 171static struct vm_zone pvzone_store; 172static struct vm_object pvzone_obj; 173static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 174static int pmap_pagedaemon_waken = 0; 175static struct pv_entry *pvinit; 176 177/* 178 * All those kernel PT submaps that BSD is so fond of 179 */ 180pt_entry_t *CMAP1 = 0; 181static pt_entry_t *CMAP2, *ptmmap; 182caddr_t CADDR1 = 0, ptvmmap = 0; 183static caddr_t CADDR2; 184static pt_entry_t *msgbufmap; 185struct msgbuf *msgbufp = 0; 186 187/* 188 * Crashdump maps. 189 */ 190static pt_entry_t *pt_crashdumpmap; 191static caddr_t crashdumpmap; 192 193#ifdef SMP 194extern pt_entry_t *SMPpt; 195#endif 196static pt_entry_t *PMAP1 = 0; 197static pt_entry_t *PADDR1 = 0; 198 199static PMAP_INLINE void free_pv_entry __P((pv_entry_t pv)); 200static unsigned * get_ptbase __P((pmap_t pmap)); 201static pv_entry_t get_pv_entry __P((void)); 202static void i386_protection_init __P((void)); 203static __inline void pmap_changebit __P((vm_page_t m, int bit, boolean_t setem)); 204 205static void pmap_remove_all __P((vm_page_t m)); 206static vm_page_t pmap_enter_quick __P((pmap_t pmap, vm_offset_t va, 207 vm_page_t m, vm_page_t mpte)); 208static int pmap_remove_pte __P((pmap_t pmap, unsigned *ptq, vm_offset_t sva)); 209static void pmap_remove_page __P((struct pmap *pmap, vm_offset_t va)); 210static int pmap_remove_entry __P((struct pmap *pmap, vm_page_t m, 211 vm_offset_t va)); 212static boolean_t pmap_testbit __P((vm_page_t m, int bit)); 213static void pmap_insert_entry __P((pmap_t pmap, vm_offset_t va, 214 vm_page_t mpte, vm_page_t m)); 215 216static vm_page_t pmap_allocpte __P((pmap_t pmap, vm_offset_t va)); 217 218static int pmap_release_free_page __P((pmap_t pmap, vm_page_t p)); 219static vm_page_t _pmap_allocpte __P((pmap_t pmap, unsigned ptepindex)); 220static unsigned * pmap_pte_quick __P((pmap_t pmap, vm_offset_t va)); 221static vm_page_t pmap_page_lookup __P((vm_object_t object, vm_pindex_t pindex)); 222static int pmap_unuse_pt __P((pmap_t, vm_offset_t, vm_page_t)); 223static vm_offset_t pmap_kmem_choose(vm_offset_t addr); 224 225static unsigned pdir4mb; 226 227/* 228 * Routine: pmap_pte 229 * Function: 230 * Extract the page table entry associated 231 * with the given map/virtual_address pair. 232 */ 233 234PMAP_INLINE unsigned * 235pmap_pte(pmap, va) 236 register pmap_t pmap; 237 vm_offset_t va; 238{ 239 pd_entry_t *pdeaddr; 240 241 if (pmap) { 242 pdeaddr = pmap_pde(pmap, va); 243 if (*pdeaddr & PG_PS) 244 return pdeaddr; 245 if (*pdeaddr) { 246 return get_ptbase(pmap) + i386_btop(va); 247 } 248 } 249 return (0); 250} 251 252/* 253 * Move the kernel virtual free pointer to the next 254 * 4MB. This is used to help improve performance 255 * by using a large (4MB) page for much of the kernel 256 * (.text, .data, .bss) 257 */ 258static vm_offset_t 259pmap_kmem_choose(vm_offset_t addr) 260{ 261 vm_offset_t newaddr = addr; 262#ifndef DISABLE_PSE 263 if (cpu_feature & CPUID_PSE) { 264 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); 265 } 266#endif 267 return newaddr; 268} 269 270/* 271 * Bootstrap the system enough to run with virtual memory. 272 * 273 * On the i386 this is called after mapping has already been enabled 274 * and just syncs the pmap module with what has already been done. 275 * [We can't call it easily with mapping off since the kernel is not 276 * mapped with PA == VA, hence we would have to relocate every address 277 * from the linked base (virtual) address "KERNBASE" to the actual 278 * (physical) address starting relative to 0] 279 */ 280void 281pmap_bootstrap(firstaddr, loadaddr) 282 vm_offset_t firstaddr; 283 vm_offset_t loadaddr; 284{ 285 vm_offset_t va; 286 pt_entry_t *pte; 287 int i; 288 289 avail_start = firstaddr; 290 291 /* 292 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 293 * large. It should instead be correctly calculated in locore.s and 294 * not based on 'first' (which is a physical address, not a virtual 295 * address, for the start of unused physical memory). The kernel 296 * page tables are NOT double mapped and thus should not be included 297 * in this calculation. 298 */ 299 virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 300 virtual_avail = pmap_kmem_choose(virtual_avail); 301 302 virtual_end = VM_MAX_KERNEL_ADDRESS; 303 304 /* 305 * Initialize protection array. 306 */ 307 i386_protection_init(); 308 309 /* 310 * The kernel's pmap is statically allocated so we don't have to use 311 * pmap_create, which is unlikely to work correctly at this part of 312 * the boot sequence (XXX and which no longer exists). 313 */ 314 kernel_pmap = &kernel_pmap_store; 315 316 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 317 kernel_pmap->pm_count = 1; 318 kernel_pmap->pm_active = -1; /* don't allow deactivation */ 319 TAILQ_INIT(&kernel_pmap->pm_pvlist); 320 LIST_INIT(&allpmaps); 321 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 322 nkpt = NKPT; 323 324 /* 325 * Reserve some special page table entries/VA space for temporary 326 * mapping of pages. 327 */ 328#define SYSMAP(c, p, v, n) \ 329 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 330 331 va = virtual_avail; 332 pte = (pt_entry_t *) pmap_pte(kernel_pmap, va); 333 334 /* 335 * CMAP1/CMAP2 are used for zeroing and copying pages. 336 */ 337 SYSMAP(caddr_t, CMAP1, CADDR1, 1) 338 SYSMAP(caddr_t, CMAP2, CADDR2, 1) 339 340 /* 341 * Crashdump maps. 342 */ 343 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 344 345 /* 346 * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 347 * XXX ptmmap is not used. 348 */ 349 SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 350 351 /* 352 * msgbufp is used to map the system message buffer. 353 * XXX msgbufmap is not used. 354 */ 355 SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 356 atop(round_page(MSGBUF_SIZE))) 357 358 /* 359 * ptemap is used for pmap_pte_quick 360 */ 361 SYSMAP(unsigned *, PMAP1, PADDR1, 1); 362 363 virtual_avail = va; 364 365 *CMAP1 = *CMAP2 = 0; 366 for (i = 0; i < NKPT; i++) 367 PTD[i] = 0; 368 369 pgeflag = 0; 370#if !defined(SMP) /* XXX - see also mp_machdep.c */ 371 if (cpu_feature & CPUID_PGE) { 372 pgeflag = PG_G; 373 } 374#endif 375 376/* 377 * Initialize the 4MB page size flag 378 */ 379 pseflag = 0; 380/* 381 * The 4MB page version of the initial 382 * kernel page mapping. 383 */ 384 pdir4mb = 0; 385 386#if !defined(DISABLE_PSE) 387 if (cpu_feature & CPUID_PSE) { 388 unsigned ptditmp; 389 /* 390 * Note that we have enabled PSE mode 391 */ 392 pseflag = PG_PS; 393 ptditmp = *(PTmap + i386_btop(KERNBASE)); 394 ptditmp &= ~(NBPDR - 1); 395 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag; 396 pdir4mb = ptditmp; 397 398#if !defined(SMP) 399 /* 400 * Enable the PSE mode. 401 */ 402 load_cr4(rcr4() | CR4_PSE); 403 404 /* 405 * We can do the mapping here for the single processor 406 * case. We simply ignore the old page table page from 407 * now on. 408 */ 409 /* 410 * For SMP, we still need 4K pages to bootstrap APs, 411 * PSE will be enabled as soon as all APs are up. 412 */ 413 PTD[KPTDI] = (pd_entry_t) ptditmp; 414 kernel_pmap->pm_pdir[KPTDI] = (pd_entry_t) ptditmp; 415 invltlb(); 416#endif 417 } 418#endif 419 420#ifdef SMP 421 if (cpu_apic_address == 0) 422 panic("pmap_bootstrap: no local apic! (non-SMP hardware?)"); 423 424 /* local apic is mapped on last page */ 425 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag | 426 (cpu_apic_address & PG_FRAME)); 427#endif 428 429 invltlb(); 430} 431 432#ifdef SMP 433/* 434 * Set 4mb pdir for mp startup 435 */ 436void 437pmap_set_opt(void) 438{ 439 if (pseflag && (cpu_feature & CPUID_PSE)) { 440 load_cr4(rcr4() | CR4_PSE); 441 if (pdir4mb && PCPU_GET(cpuid) == 0) { /* only on BSP */ 442 kernel_pmap->pm_pdir[KPTDI] = 443 PTD[KPTDI] = (pd_entry_t)pdir4mb; 444 cpu_invltlb(); 445 } 446 } 447} 448#endif 449 450/* 451 * Initialize the pmap module. 452 * Called by vm_init, to initialize any structures that the pmap 453 * system needs to map virtual memory. 454 * pmap_init has been enhanced to support in a fairly consistant 455 * way, discontiguous physical memory. 456 */ 457void 458pmap_init(phys_start, phys_end) 459 vm_offset_t phys_start, phys_end; 460{ 461 int i; 462 int initial_pvs; 463 464 /* 465 * object for kernel page table pages 466 */ 467 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE); 468 469 /* 470 * Allocate memory for random pmap data structures. Includes the 471 * pv_head_table. 472 */ 473 474 for(i = 0; i < vm_page_array_size; i++) { 475 vm_page_t m; 476 477 m = &vm_page_array[i]; 478 TAILQ_INIT(&m->md.pv_list); 479 m->md.pv_list_count = 0; 480 } 481 482 /* 483 * init the pv free list 484 */ 485 initial_pvs = vm_page_array_size; 486 if (initial_pvs < MINPV) 487 initial_pvs = MINPV; 488 pvzone = &pvzone_store; 489 pvinit = (struct pv_entry *) kmem_alloc(kernel_map, 490 initial_pvs * sizeof (struct pv_entry)); 491 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit, 492 vm_page_array_size); 493 494 /* 495 * Now it is safe to enable pv_table recording. 496 */ 497 pmap_initialized = TRUE; 498} 499 500/* 501 * Initialize the address space (zone) for the pv_entries. Set a 502 * high water mark so that the system can recover from excessive 503 * numbers of pv entries. 504 */ 505void 506pmap_init2() 507{ 508 int shpgperproc = PMAP_SHPGPERPROC; 509 510 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 511 pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 512 pv_entry_high_water = 9 * (pv_entry_max / 10); 513 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1); 514} 515 516 517/*************************************************** 518 * Low level helper routines..... 519 ***************************************************/ 520 521#if defined(PMAP_DIAGNOSTIC) 522 523/* 524 * This code checks for non-writeable/modified pages. 525 * This should be an invalid condition. 526 */ 527static int 528pmap_nw_modified(pt_entry_t ptea) 529{ 530 int pte; 531 532 pte = (int) ptea; 533 534 if ((pte & (PG_M|PG_RW)) == PG_M) 535 return 1; 536 else 537 return 0; 538} 539#endif 540 541 542/* 543 * this routine defines the region(s) of memory that should 544 * not be tested for the modified bit. 545 */ 546static PMAP_INLINE int 547pmap_track_modified(vm_offset_t va) 548{ 549 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 550 return 1; 551 else 552 return 0; 553} 554 555static PMAP_INLINE void 556invltlb_1pg(vm_offset_t va) 557{ 558#ifdef I386_CPU 559 invltlb(); 560#else 561 invlpg(va); 562#endif 563} 564 565static __inline void 566pmap_invalidate_page(pmap_t pmap, vm_offset_t va) 567{ 568#if defined(SMP) 569 if (pmap->pm_active & (1 << PCPU_GET(cpuid))) 570 cpu_invlpg((void *)va); 571 if (pmap->pm_active & PCPU_GET(other_cpus)) 572 smp_invltlb(); 573#else 574 if (pmap->pm_active) 575 invltlb_1pg(va); 576#endif 577} 578 579static __inline void 580pmap_invalidate_all(pmap_t pmap) 581{ 582#if defined(SMP) 583 if (pmap->pm_active & (1 << PCPU_GET(cpuid))) 584 cpu_invltlb(); 585 if (pmap->pm_active & PCPU_GET(other_cpus)) 586 smp_invltlb(); 587#else 588 if (pmap->pm_active) 589 invltlb(); 590#endif 591} 592 593/* 594 * Return an address which is the base of the Virtual mapping of 595 * all the PTEs for the given pmap. Note this doesn't say that 596 * all the PTEs will be present or that the pages there are valid. 597 * The PTEs are made available by the recursive mapping trick. 598 * It will map in the alternate PTE space if needed. 599 */ 600static pt_entry_t * 601get_ptbase(pmap) 602 pmap_t pmap; 603{ 604 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME; 605 606 /* are we current address space or kernel? */ 607 if (pmap == kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) { 608 return PTmap; 609 } 610 /* otherwise, we are alternate address space */ 611 if (frame != (((unsigned) APTDpde) & PG_FRAME)) { 612 APTDpde = (pd_entry_t) (frame | PG_RW | PG_V); 613#if defined(SMP) 614 /* The page directory is not shared between CPUs */ 615 cpu_invltlb(); 616#else 617 invltlb(); 618#endif 619 } 620 return APTmap; 621} 622 623/* 624 * Super fast pmap_pte routine best used when scanning 625 * the pv lists. This eliminates many coarse-grained 626 * invltlb calls. Note that many of the pv list 627 * scans are across different pmaps. It is very wasteful 628 * to do an entire invltlb for checking a single mapping. 629 */ 630 631static pt_entry_t * 632pmap_pte_quick(pmap, va) 633 register pmap_t pmap; 634 vm_offset_t va; 635{ 636 pd_entry_t pde, newpf; 637 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) { 638 pd_entry_t frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME; 639 unsigned index = i386_btop(va); 640 /* are we current address space or kernel? */ 641 if ((pmap == kernel_pmap) || 642 (frame == (((unsigned) PTDpde) & PG_FRAME))) { 643 return PTmap + index; 644 } 645 newpf = pde & PG_FRAME; 646 if (((*PMAP1) & PG_FRAME) != newpf) { 647 *PMAP1 = newpf | PG_RW | PG_V; 648 invltlb_1pg((vm_offset_t) PADDR1); 649 } 650 return PADDR1 + (index & (NPTEPG - 1)); 651 } 652 return (0); 653} 654 655/* 656 * Routine: pmap_extract 657 * Function: 658 * Extract the physical page address associated 659 * with the given map/virtual_address pair. 660 */ 661vm_offset_t 662pmap_extract(pmap, va) 663 register pmap_t pmap; 664 vm_offset_t va; 665{ 666 vm_offset_t rtval; /* XXX FIXME */ 667 vm_offset_t pdirindex; 668 pdirindex = va >> PDRSHIFT; 669 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) { 670 pt_entry_t *pte; 671 if ((rtval & PG_PS) != 0) { 672 rtval &= ~(NBPDR - 1); 673 rtval |= va & (NBPDR - 1); 674 return rtval; 675 } 676 pte = get_ptbase(pmap) + i386_btop(va); 677 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK)); 678 return rtval; 679 } 680 return 0; 681 682} 683 684/*************************************************** 685 * Low level mapping routines..... 686 ***************************************************/ 687 688/* 689 * add a wired page to the kva 690 * note that in order for the mapping to take effect -- you 691 * should do a invltlb after doing the pmap_kenter... 692 */ 693PMAP_INLINE void 694pmap_kenter(vm_offset_t va, vm_offset_t pa) 695{ 696 pt_entry_t *pte; 697 pt_entry_t npte, opte; 698 699 npte = pa | PG_RW | PG_V | pgeflag; 700 pte = vtopte(va); 701 opte = *pte; 702 *pte = npte; 703#ifdef SMP 704 invlpg(va); 705#else 706 invltlb_1pg(va); 707#endif 708} 709 710/* 711 * remove a page from the kernel pagetables 712 */ 713PMAP_INLINE void 714pmap_kremove(vm_offset_t va) 715{ 716 register pt_entry_t *pte; 717 718 pte = vtopte(va); 719 *pte = 0; 720#ifdef SMP 721 invlpg(va); 722#else 723 invltlb_1pg(va); 724#endif 725} 726 727/* 728 * Used to map a range of physical addresses into kernel 729 * virtual address space. 730 * 731 * The value passed in '*virt' is a suggested virtual address for 732 * the mapping. Architectures which can support a direct-mapped 733 * physical to virtual region can return the appropriate address 734 * within that region, leaving '*virt' unchanged. Other 735 * architectures should map the pages starting at '*virt' and 736 * update '*virt' with the first usable address after the mapped 737 * region. 738 */ 739vm_offset_t 740pmap_map(vm_offset_t *virt, vm_offset_t start, vm_offset_t end, int prot) 741{ 742 vm_offset_t sva = *virt; 743 vm_offset_t va = sva; 744 while (start < end) { 745 pmap_kenter(va, start); 746 va += PAGE_SIZE; 747 start += PAGE_SIZE; 748 } 749 *virt = va; 750 return (sva); 751} 752 753 754/* 755 * Add a list of wired pages to the kva 756 * this routine is only used for temporary 757 * kernel mappings that do not need to have 758 * page modification or references recorded. 759 * Note that old mappings are simply written 760 * over. The page *must* be wired. 761 */ 762void 763pmap_qenter(vm_offset_t va, vm_page_t *m, int count) 764{ 765 vm_offset_t end_va; 766 767 end_va = va + count * PAGE_SIZE; 768 769 while (va < end_va) { 770 pt_entry_t *pte; 771 772 pte = vtopte(va); 773 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag; 774#ifdef SMP 775 cpu_invlpg((void *)va); 776#else 777 invltlb_1pg(va); 778#endif 779 va += PAGE_SIZE; 780 m++; 781 } 782#ifdef SMP 783 smp_invltlb(); 784#endif 785} 786 787/* 788 * this routine jerks page mappings from the 789 * kernel -- it is meant only for temporary mappings. 790 */ 791void 792pmap_qremove(vm_offset_t va, int count) 793{ 794 vm_offset_t end_va; 795 796 end_va = va + count*PAGE_SIZE; 797 798 while (va < end_va) { 799 pt_entry_t *pte; 800 801 pte = vtopte(va); 802 *pte = 0; 803#ifdef SMP 804 cpu_invlpg((void *)va); 805#else 806 invltlb_1pg(va); 807#endif 808 va += PAGE_SIZE; 809 } 810#ifdef SMP 811 smp_invltlb(); 812#endif 813} 814 815static vm_page_t 816pmap_page_lookup(vm_object_t object, vm_pindex_t pindex) 817{ 818 vm_page_t m; 819retry: 820 m = vm_page_lookup(object, pindex); 821 if (m && vm_page_sleep_busy(m, FALSE, "pplookp")) 822 goto retry; 823 return m; 824} 825 826/* 827 * Create the Uarea stack for a new process. 828 * This routine directly affects the fork perf for a process. 829 */ 830void 831pmap_new_proc(struct proc *p) 832{ 833#ifdef I386_CPU 834 int updateneeded = 0; 835#endif 836 int i; 837 vm_object_t upobj; 838 vm_offset_t up; 839 vm_page_t m; 840 pt_entry_t *ptek, oldpte; 841 842 /* 843 * allocate object for the upages 844 */ 845 upobj = p->p_upages_obj; 846 if (upobj == NULL) { 847 upobj = vm_object_allocate(OBJT_DEFAULT, UAREA_PAGES); 848 p->p_upages_obj = upobj; 849 } 850 851 /* get a kernel virtual address for the U area for this thread */ 852 up = (vm_offset_t)p->p_uarea; 853 if (up == 0) { 854 up = kmem_alloc_nofault(kernel_map, UAREA_PAGES * PAGE_SIZE); 855 if (up == 0) 856 panic("pmap_new_proc: upage allocation failed"); 857 p->p_uarea = (struct user *)up; 858 } 859 860 ptek = vtopte(up); 861 862 for (i = 0; i < UAREA_PAGES; i++) { 863 /* 864 * Get a kernel stack page 865 */ 866 m = vm_page_grab(upobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 867 868 /* 869 * Wire the page 870 */ 871 m->wire_count++; 872 cnt.v_wire_count++; 873 874 oldpte = *(ptek + i); 875 /* 876 * Enter the page into the kernel address space. 877 */ 878 *(ptek + i) = VM_PAGE_TO_PHYS(m) | PG_RW | PG_V | pgeflag; 879 if (oldpte) { 880#ifdef I386_CPU 881 updateneeded = 1; 882#else 883 invlpg(up + i * PAGE_SIZE); 884#endif 885 } 886 887 vm_page_wakeup(m); 888 vm_page_flag_clear(m, PG_ZERO); 889 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE); 890 m->valid = VM_PAGE_BITS_ALL; 891 } 892#ifdef I386_CPU 893 if (updateneeded) 894 invltlb(); 895#endif 896} 897 898/* 899 * Dispose the U-Area for a process that has exited. 900 * This routine directly impacts the exit perf of a process. 901 */ 902void 903pmap_dispose_proc(p) 904 struct proc *p; 905{ 906 int i; 907 vm_object_t upobj; 908 vm_offset_t up; 909 vm_page_t m; 910 pt_entry_t *ptek, oldpte; 911 912 upobj = p->p_upages_obj; 913 up = (vm_offset_t)p->p_uarea; 914 ptek = vtopte(up); 915 for (i = 0; i < UAREA_PAGES; i++) { 916 m = vm_page_lookup(upobj, i); 917 if (m == NULL) 918 panic("pmap_dispose_proc: upage already missing?"); 919 vm_page_busy(m); 920 oldpte = *(ptek + i); 921 *(ptek + i) = 0; 922#ifndef I386_CPU 923 invlpg(up + i * PAGE_SIZE); 924#endif 925 vm_page_unwire(m, 0); 926 vm_page_free(m); 927 } 928#ifdef I386_CPU 929 invltlb(); 930#endif 931} 932 933/* 934 * Allow the U_AREA for a process to be prejudicially paged out. 935 */ 936void 937pmap_swapout_proc(p) 938 struct proc *p; 939{ 940 int i; 941 vm_object_t upobj; 942 vm_offset_t up; 943 vm_page_t m; 944 945 upobj = p->p_upages_obj; 946 up = (vm_offset_t)p->p_uarea; 947 for (i = 0; i < UAREA_PAGES; i++) { 948 m = vm_page_lookup(upobj, i); 949 if (m == NULL) 950 panic("pmap_swapout_proc: upage already missing?"); 951 vm_page_dirty(m); 952 vm_page_unwire(m, 0); 953 pmap_kremove(up + i * PAGE_SIZE); 954 } 955} 956 957/* 958 * Bring the U-Area for a specified process back in. 959 */ 960void 961pmap_swapin_proc(p) 962 struct proc *p; 963{ 964 int i, rv; 965 vm_object_t upobj; 966 vm_offset_t up; 967 vm_page_t m; 968 969 upobj = p->p_upages_obj; 970 up = (vm_offset_t)p->p_uarea; 971 for (i = 0; i < UAREA_PAGES; i++) { 972 m = vm_page_grab(upobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 973 pmap_kenter(up + i * PAGE_SIZE, VM_PAGE_TO_PHYS(m)); 974 if (m->valid != VM_PAGE_BITS_ALL) { 975 rv = vm_pager_get_pages(upobj, &m, 1, 0); 976 if (rv != VM_PAGER_OK) 977 panic("pmap_swapin_proc: cannot get upage for proc: %d\n", p->p_pid); 978 m = vm_page_lookup(upobj, i); 979 m->valid = VM_PAGE_BITS_ALL; 980 } 981 vm_page_wire(m); 982 vm_page_wakeup(m); 983 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE); 984 } 985} 986 987/* 988 * Create the kernel stack (including pcb for i386) for a new thread. 989 * This routine directly affects the fork perf for a process and 990 * create performance for a thread. 991 */ 992void 993pmap_new_thread(struct thread *td) 994{ 995#ifdef I386_CPU 996 int updateneeded = 0; 997#endif 998 int i; 999 vm_object_t ksobj; 1000 vm_page_t m; 1001 vm_offset_t ks; 1002 pt_entry_t *ptek, oldpte; 1003 1004 /* 1005 * allocate object for the kstack 1006 */ 1007 ksobj = td->td_kstack_obj; 1008 if (ksobj == NULL) { 1009 ksobj = vm_object_allocate(OBJT_DEFAULT, KSTACK_PAGES); 1010 td->td_kstack_obj = ksobj; 1011 } 1012 1013#ifdef KSTACK_GUARD 1014 /* get a kernel virtual address for the kstack for this thread */ 1015 ks = td->td_kstack; 1016 if (ks == 0) { 1017 ks = kmem_alloc_nofault(kernel_map, 1018 (KSTACK_PAGES + 1) * PAGE_SIZE); 1019 if (ks == 0) 1020 panic("pmap_new_thread: kstack allocation failed"); 1021 ks += PAGE_SIZE; 1022 td->td_kstack = ks; 1023 } 1024 1025 ptek = vtopte(ks - PAGE_SIZE); 1026 oldpte = *ptek; 1027 *ptek = 0; 1028 if (oldpte) { 1029#ifdef I386_CPU 1030 updateneeded = 1; 1031#else 1032 invlpg(ks - PAGE_SIZE); 1033#endif 1034 } 1035 ptek++; 1036#else 1037 /* get a kernel virtual address for the kstack for this thread */ 1038 ks = td->td_kstack; 1039 if (ks == 0) { 1040 ks = kmem_alloc_nofault(kernel_map, KSTACK_PAGES * PAGE_SIZE); 1041 if (ks == 0) 1042 panic("pmap_new_thread: kstack allocation failed"); 1043 td->td_kstack = ks; 1044 } 1045 ptek = vtopte(ks); 1046#endif 1047 for (i = 0; i < KSTACK_PAGES; i++) { 1048 /* 1049 * Get a kernel stack page 1050 */ 1051 m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1052 1053 /* 1054 * Wire the page 1055 */ 1056 m->wire_count++; 1057 cnt.v_wire_count++; 1058 1059 oldpte = *(ptek + i); 1060 /* 1061 * Enter the page into the kernel address space. 1062 */ 1063 *(ptek + i) = VM_PAGE_TO_PHYS(m) | PG_RW | PG_V | pgeflag; 1064 if (oldpte) { 1065#ifdef I386_CPU 1066 updateneeded = 1; 1067#else 1068 invlpg(ks + i * PAGE_SIZE); 1069#endif 1070 } 1071 1072 vm_page_wakeup(m); 1073 vm_page_flag_clear(m, PG_ZERO); 1074 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE); 1075 m->valid = VM_PAGE_BITS_ALL; 1076 } 1077#ifdef I386_CPU 1078 if (updateneeded) 1079 invltlb(); 1080#endif 1081} 1082 1083/* 1084 * Dispose the kernel stack for a thread that has exited. 1085 * This routine directly impacts the exit perf of a process and thread. 1086 */ 1087void 1088pmap_dispose_thread(td) 1089 struct thread *td; 1090{ 1091 int i; 1092 vm_object_t ksobj; 1093 vm_offset_t ks; 1094 vm_page_t m; 1095 pt_entry_t *ptek, oldpte; 1096 1097 ksobj = td->td_kstack_obj; 1098 ks = td->td_kstack; 1099 ptek = vtopte(ks); 1100 for (i = 0; i < KSTACK_PAGES; i++) { 1101 m = vm_page_lookup(ksobj, i); 1102 if (m == NULL) 1103 panic("pmap_dispose_thread: kstack already missing?"); 1104 vm_page_busy(m); 1105 oldpte = *(ptek + i); 1106 *(ptek + i) = 0; 1107#ifndef I386_CPU 1108 invlpg(ks + i * PAGE_SIZE); 1109#endif 1110 vm_page_unwire(m, 0); 1111 vm_page_free(m); 1112 } 1113#ifdef I386_CPU 1114 invltlb(); 1115#endif 1116} 1117 1118/* 1119 * Allow the Kernel stack for a thread to be prejudicially paged out. 1120 */ 1121void 1122pmap_swapout_thread(td) 1123 struct thread *td; 1124{ 1125 int i; 1126 vm_object_t ksobj; 1127 vm_offset_t ks; 1128 vm_page_t m; 1129 1130 ksobj = td->td_kstack_obj; 1131 ks = td->td_kstack; 1132 for (i = 0; i < KSTACK_PAGES; i++) { 1133 m = vm_page_lookup(ksobj, i); 1134 if (m == NULL) 1135 panic("pmap_swapout_thread: kstack already missing?"); 1136 vm_page_dirty(m); 1137 vm_page_unwire(m, 0); 1138 pmap_kremove(ks + i * PAGE_SIZE); 1139 } 1140} 1141 1142/* 1143 * Bring the kernel stack for a specified thread back in. 1144 */ 1145void 1146pmap_swapin_thread(td) 1147 struct thread *td; 1148{ 1149 int i, rv; 1150 vm_object_t ksobj; 1151 vm_offset_t ks; 1152 vm_page_t m; 1153 1154 ksobj = td->td_kstack_obj; 1155 ks = td->td_kstack; 1156 for (i = 0; i < KSTACK_PAGES; i++) { 1157 m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1158 pmap_kenter(ks + i * PAGE_SIZE, VM_PAGE_TO_PHYS(m)); 1159 if (m->valid != VM_PAGE_BITS_ALL) { 1160 rv = vm_pager_get_pages(ksobj, &m, 1, 0); 1161 if (rv != VM_PAGER_OK) 1162 panic("pmap_swapin_thread: cannot get kstack for proc: %d\n", td->td_proc->p_pid); 1163 m = vm_page_lookup(ksobj, i); 1164 m->valid = VM_PAGE_BITS_ALL; 1165 } 1166 vm_page_wire(m); 1167 vm_page_wakeup(m); 1168 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE); 1169 } 1170} 1171 1172/*************************************************** 1173 * Page table page management routines..... 1174 ***************************************************/ 1175 1176/* 1177 * This routine unholds page table pages, and if the hold count 1178 * drops to zero, then it decrements the wire count. 1179 */ 1180static int 1181_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1182{ 1183 1184 while (vm_page_sleep_busy(m, FALSE, "pmuwpt")) 1185 ; 1186 1187 if (m->hold_count == 0) { 1188 vm_offset_t pteva; 1189 /* 1190 * unmap the page table page 1191 */ 1192 pmap->pm_pdir[m->pindex] = 0; 1193 --pmap->pm_stats.resident_count; 1194 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == 1195 (PTDpde & PG_FRAME)) { 1196 /* 1197 * Do a invltlb to make the invalidated mapping 1198 * take effect immediately. 1199 */ 1200 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 1201 pmap_invalidate_page(pmap, pteva); 1202 } 1203 1204 if (pmap->pm_ptphint == m) 1205 pmap->pm_ptphint = NULL; 1206 1207 /* 1208 * If the page is finally unwired, simply free it. 1209 */ 1210 --m->wire_count; 1211 if (m->wire_count == 0) { 1212 1213 vm_page_flash(m); 1214 vm_page_busy(m); 1215 vm_page_free_zero(m); 1216 --cnt.v_wire_count; 1217 } 1218 return 1; 1219 } 1220 return 0; 1221} 1222 1223static PMAP_INLINE int 1224pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 1225{ 1226 vm_page_unhold(m); 1227 if (m->hold_count == 0) 1228 return _pmap_unwire_pte_hold(pmap, m); 1229 else 1230 return 0; 1231} 1232 1233/* 1234 * After removing a page table entry, this routine is used to 1235 * conditionally free the page, and manage the hold/wire counts. 1236 */ 1237static int 1238pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 1239{ 1240 unsigned ptepindex; 1241 if (va >= VM_MAXUSER_ADDRESS) 1242 return 0; 1243 1244 if (mpte == NULL) { 1245 ptepindex = (va >> PDRSHIFT); 1246 if (pmap->pm_ptphint && 1247 (pmap->pm_ptphint->pindex == ptepindex)) { 1248 mpte = pmap->pm_ptphint; 1249 } else { 1250 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 1251 pmap->pm_ptphint = mpte; 1252 } 1253 } 1254 1255 return pmap_unwire_pte_hold(pmap, mpte); 1256} 1257 1258void 1259pmap_pinit0(pmap) 1260 struct pmap *pmap; 1261{ 1262 pmap->pm_pdir = 1263 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE); 1264 pmap_kenter((vm_offset_t) pmap->pm_pdir, (vm_offset_t) IdlePTD); 1265 pmap->pm_count = 1; 1266 pmap->pm_ptphint = NULL; 1267 pmap->pm_active = 0; 1268 TAILQ_INIT(&pmap->pm_pvlist); 1269 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1270 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1271} 1272 1273/* 1274 * Initialize a preallocated and zeroed pmap structure, 1275 * such as one in a vmspace structure. 1276 */ 1277void 1278pmap_pinit(pmap) 1279 register struct pmap *pmap; 1280{ 1281 vm_page_t ptdpg; 1282 1283 /* 1284 * No need to allocate page table space yet but we do need a valid 1285 * page directory table. 1286 */ 1287 if (pmap->pm_pdir == NULL) 1288 pmap->pm_pdir = 1289 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE); 1290 1291 /* 1292 * allocate object for the ptes 1293 */ 1294 if (pmap->pm_pteobj == NULL) 1295 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1); 1296 1297 /* 1298 * allocate the page directory page 1299 */ 1300 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI, 1301 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1302 1303 ptdpg->wire_count = 1; 1304 ++cnt.v_wire_count; 1305 1306 1307 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY); /* not usually mapped*/ 1308 ptdpg->valid = VM_PAGE_BITS_ALL; 1309 1310 pmap_kenter((vm_offset_t) pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg)); 1311 if ((ptdpg->flags & PG_ZERO) == 0) 1312 bzero(pmap->pm_pdir, PAGE_SIZE); 1313 1314 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1315 /* Wire in kernel global address entries. */ 1316 /* XXX copies current process, does not fill in MPPTDI */ 1317 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE); 1318#ifdef SMP 1319 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1320#endif 1321 1322 /* install self-referential address mapping entry */ 1323 pmap->pm_pdir[PTDPTDI] = 1324 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M; 1325 1326 pmap->pm_count = 1; 1327 pmap->pm_active = 0; 1328 pmap->pm_ptphint = NULL; 1329 TAILQ_INIT(&pmap->pm_pvlist); 1330 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1331} 1332 1333/* 1334 * Wire in kernel global address entries. To avoid a race condition 1335 * between pmap initialization and pmap_growkernel, this procedure 1336 * should be called after the vmspace is attached to the process 1337 * but before this pmap is activated. 1338 */ 1339void 1340pmap_pinit2(pmap) 1341 struct pmap *pmap; 1342{ 1343 /* XXX: Remove this stub when no longer called */ 1344} 1345 1346static int 1347pmap_release_free_page(pmap_t pmap, vm_page_t p) 1348{ 1349 pd_entry_t *pde = pmap->pm_pdir; 1350 /* 1351 * This code optimizes the case of freeing non-busy 1352 * page-table pages. Those pages are zero now, and 1353 * might as well be placed directly into the zero queue. 1354 */ 1355 if (vm_page_sleep_busy(p, FALSE, "pmaprl")) 1356 return 0; 1357 1358 vm_page_busy(p); 1359 1360 /* 1361 * Remove the page table page from the processes address space. 1362 */ 1363 pde[p->pindex] = 0; 1364 pmap->pm_stats.resident_count--; 1365 1366 if (p->hold_count) { 1367 panic("pmap_release: freeing held page table page"); 1368 } 1369 /* 1370 * Page directory pages need to have the kernel 1371 * stuff cleared, so they can go into the zero queue also. 1372 */ 1373 if (p->pindex == PTDPTDI) { 1374 bzero(pde + KPTDI, nkpt * PTESIZE); 1375#ifdef SMP 1376 pde[MPPTDI] = 0; 1377#endif 1378 pde[APTDPTDI] = 0; 1379 pmap_kremove((vm_offset_t) pmap->pm_pdir); 1380 } 1381 1382 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex)) 1383 pmap->pm_ptphint = NULL; 1384 1385 p->wire_count--; 1386 cnt.v_wire_count--; 1387 vm_page_free_zero(p); 1388 return 1; 1389} 1390 1391/* 1392 * this routine is called if the page table page is not 1393 * mapped correctly. 1394 */ 1395static vm_page_t 1396_pmap_allocpte(pmap, ptepindex) 1397 pmap_t pmap; 1398 unsigned ptepindex; 1399{ 1400 vm_offset_t pteva, ptepa; /* XXXPA */ 1401 vm_page_t m; 1402 1403 /* 1404 * Find or fabricate a new pagetable page 1405 */ 1406 m = vm_page_grab(pmap->pm_pteobj, ptepindex, 1407 VM_ALLOC_ZERO | VM_ALLOC_RETRY); 1408 1409 KASSERT(m->queue == PQ_NONE, 1410 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1411 1412 if (m->wire_count == 0) 1413 cnt.v_wire_count++; 1414 m->wire_count++; 1415 1416 /* 1417 * Increment the hold count for the page table page 1418 * (denoting a new mapping.) 1419 */ 1420 m->hold_count++; 1421 1422 /* 1423 * Map the pagetable page into the process address space, if 1424 * it isn't already there. 1425 */ 1426 1427 pmap->pm_stats.resident_count++; 1428 1429 ptepa = VM_PAGE_TO_PHYS(m); 1430 pmap->pm_pdir[ptepindex] = 1431 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1432 1433 /* 1434 * Set the page table hint 1435 */ 1436 pmap->pm_ptphint = m; 1437 1438 /* 1439 * Try to use the new mapping, but if we cannot, then 1440 * do it with the routine that maps the page explicitly. 1441 */ 1442 if ((m->flags & PG_ZERO) == 0) { 1443 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == 1444 (PTDpde & PG_FRAME)) { 1445 pteva = VM_MAXUSER_ADDRESS + i386_ptob(ptepindex); 1446 bzero((caddr_t) pteva, PAGE_SIZE); 1447 } else { 1448 pmap_zero_page(ptepa); 1449 } 1450 } 1451 1452 m->valid = VM_PAGE_BITS_ALL; 1453 vm_page_flag_clear(m, PG_ZERO); 1454 vm_page_flag_set(m, PG_MAPPED); 1455 vm_page_wakeup(m); 1456 1457 return m; 1458} 1459 1460static vm_page_t 1461pmap_allocpte(pmap_t pmap, vm_offset_t va) 1462{ 1463 unsigned ptepindex; 1464 pd_entry_t ptepa; 1465 vm_page_t m; 1466 1467 /* 1468 * Calculate pagetable page index 1469 */ 1470 ptepindex = va >> PDRSHIFT; 1471 1472 /* 1473 * Get the page directory entry 1474 */ 1475 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex]; 1476 1477 /* 1478 * This supports switching from a 4MB page to a 1479 * normal 4K page. 1480 */ 1481 if (ptepa & PG_PS) { 1482 pmap->pm_pdir[ptepindex] = 0; 1483 ptepa = 0; 1484 invltlb(); 1485 } 1486 1487 /* 1488 * If the page table page is mapped, we just increment the 1489 * hold count, and activate it. 1490 */ 1491 if (ptepa) { 1492 /* 1493 * In order to get the page table page, try the 1494 * hint first. 1495 */ 1496 if (pmap->pm_ptphint && 1497 (pmap->pm_ptphint->pindex == ptepindex)) { 1498 m = pmap->pm_ptphint; 1499 } else { 1500 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 1501 pmap->pm_ptphint = m; 1502 } 1503 m->hold_count++; 1504 return m; 1505 } 1506 /* 1507 * Here if the pte page isn't mapped, or if it has been deallocated. 1508 */ 1509 return _pmap_allocpte(pmap, ptepindex); 1510} 1511 1512 1513/*************************************************** 1514* Pmap allocation/deallocation routines. 1515 ***************************************************/ 1516 1517/* 1518 * Release any resources held by the given physical map. 1519 * Called when a pmap initialized by pmap_pinit is being released. 1520 * Should only be called if the map contains no valid mappings. 1521 */ 1522void 1523pmap_release(pmap_t pmap) 1524{ 1525 vm_page_t p,n,ptdpg; 1526 vm_object_t object = pmap->pm_pteobj; 1527 int curgeneration; 1528 1529#if defined(DIAGNOSTIC) 1530 if (object->ref_count != 1) 1531 panic("pmap_release: pteobj reference count != 1"); 1532#endif 1533 1534 ptdpg = NULL; 1535 LIST_REMOVE(pmap, pm_list); 1536retry: 1537 curgeneration = object->generation; 1538 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = n) { 1539 n = TAILQ_NEXT(p, listq); 1540 if (p->pindex == PTDPTDI) { 1541 ptdpg = p; 1542 continue; 1543 } 1544 while (1) { 1545 if (!pmap_release_free_page(pmap, p) && 1546 (object->generation != curgeneration)) 1547 goto retry; 1548 } 1549 } 1550 1551 if (ptdpg && !pmap_release_free_page(pmap, ptdpg)) 1552 goto retry; 1553} 1554 1555static int 1556kvm_size(SYSCTL_HANDLER_ARGS) 1557{ 1558 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1559 1560 return sysctl_handle_long(oidp, &ksize, 0, req); 1561} 1562SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1563 0, 0, kvm_size, "IU", "Size of KVM"); 1564 1565static int 1566kvm_free(SYSCTL_HANDLER_ARGS) 1567{ 1568 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1569 1570 return sysctl_handle_long(oidp, &kfree, 0, req); 1571} 1572SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1573 0, 0, kvm_free, "IU", "Amount of KVM free"); 1574 1575/* 1576 * grow the number of kernel page table entries, if needed 1577 */ 1578void 1579pmap_growkernel(vm_offset_t addr) 1580{ 1581 struct pmap *pmap; 1582 int s; 1583 vm_offset_t ptppaddr; 1584 vm_page_t nkpg; 1585 pd_entry_t newpdir; 1586 1587 s = splhigh(); 1588 if (kernel_vm_end == 0) { 1589 kernel_vm_end = KERNBASE; 1590 nkpt = 0; 1591 while (pdir_pde(PTD, kernel_vm_end)) { 1592 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1593 nkpt++; 1594 } 1595 } 1596 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1597 while (kernel_vm_end < addr) { 1598 if (pdir_pde(PTD, kernel_vm_end)) { 1599 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1600 continue; 1601 } 1602 1603 /* 1604 * This index is bogus, but out of the way 1605 */ 1606 nkpg = vm_page_alloc(kptobj, nkpt, VM_ALLOC_SYSTEM); 1607 if (!nkpg) 1608 panic("pmap_growkernel: no memory to grow kernel"); 1609 1610 nkpt++; 1611 1612 vm_page_wire(nkpg); 1613 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1614 pmap_zero_page(ptppaddr); 1615 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1616 pdir_pde(PTD, kernel_vm_end) = newpdir; 1617 1618 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1619 *pmap_pde(pmap, kernel_vm_end) = newpdir; 1620 } 1621 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1622 } 1623 splx(s); 1624} 1625 1626/* 1627 * Retire the given physical map from service. 1628 * Should only be called if the map contains 1629 * no valid mappings. 1630 */ 1631void 1632pmap_destroy(pmap_t pmap) 1633{ 1634 int count; 1635 1636 if (pmap == NULL) 1637 return; 1638 1639 count = --pmap->pm_count; 1640 if (count == 0) { 1641 pmap_release(pmap); 1642 panic("destroying a pmap is not yet implemented"); 1643 } 1644} 1645 1646/* 1647 * Add a reference to the specified pmap. 1648 */ 1649void 1650pmap_reference(pmap_t pmap) 1651{ 1652 if (pmap != NULL) { 1653 pmap->pm_count++; 1654 } 1655} 1656 1657/*************************************************** 1658* page management routines. 1659 ***************************************************/ 1660 1661/* 1662 * free the pv_entry back to the free list 1663 */ 1664static PMAP_INLINE void 1665free_pv_entry(pv_entry_t pv) 1666{ 1667 pv_entry_count--; 1668 zfree(pvzone, pv); 1669} 1670 1671/* 1672 * get a new pv_entry, allocating a block from the system 1673 * when needed. 1674 * the memory allocation is performed bypassing the malloc code 1675 * because of the possibility of allocations at interrupt time. 1676 */ 1677static pv_entry_t 1678get_pv_entry(void) 1679{ 1680 pv_entry_count++; 1681 if (pv_entry_high_water && 1682 (pv_entry_count > pv_entry_high_water) && 1683 (pmap_pagedaemon_waken == 0)) { 1684 pmap_pagedaemon_waken = 1; 1685 wakeup (&vm_pages_needed); 1686 } 1687 return zalloc(pvzone); 1688} 1689 1690/* 1691 * This routine is very drastic, but can save the system 1692 * in a pinch. 1693 */ 1694void 1695pmap_collect() 1696{ 1697 int i; 1698 vm_page_t m; 1699 static int warningdone = 0; 1700 1701 if (pmap_pagedaemon_waken == 0) 1702 return; 1703 1704 if (warningdone < 5) { 1705 printf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n"); 1706 warningdone++; 1707 } 1708 1709 for(i = 0; i < vm_page_array_size; i++) { 1710 m = &vm_page_array[i]; 1711 if (m->wire_count || m->hold_count || m->busy || 1712 (m->flags & (PG_BUSY | PG_UNMANAGED))) 1713 continue; 1714 pmap_remove_all(m); 1715 } 1716 pmap_pagedaemon_waken = 0; 1717} 1718 1719 1720/* 1721 * If it is the first entry on the list, it is actually 1722 * in the header and we must copy the following entry up 1723 * to the header. Otherwise we must search the list for 1724 * the entry. In either case we free the now unused entry. 1725 */ 1726 1727static int 1728pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1729{ 1730 pv_entry_t pv; 1731 int rtval; 1732 int s; 1733 1734 s = splvm(); 1735 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1736 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1737 if (pmap == pv->pv_pmap && va == pv->pv_va) 1738 break; 1739 } 1740 } else { 1741 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1742 if (va == pv->pv_va) 1743 break; 1744 } 1745 } 1746 1747 rtval = 0; 1748 if (pv) { 1749 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1750 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1751 m->md.pv_list_count--; 1752 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1753 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 1754 1755 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1756 free_pv_entry(pv); 1757 } 1758 1759 splx(s); 1760 return rtval; 1761} 1762 1763/* 1764 * Create a pv entry for page at pa for 1765 * (pmap, va). 1766 */ 1767static void 1768pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1769{ 1770 1771 int s; 1772 pv_entry_t pv; 1773 1774 s = splvm(); 1775 pv = get_pv_entry(); 1776 pv->pv_va = va; 1777 pv->pv_pmap = pmap; 1778 pv->pv_ptem = mpte; 1779 1780 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1781 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1782 m->md.pv_list_count++; 1783 1784 splx(s); 1785} 1786 1787/* 1788 * pmap_remove_pte: do the things to unmap a page in a process 1789 */ 1790static int 1791pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1792{ 1793 pt_entry_t oldpte; 1794 vm_page_t m; 1795 1796 oldpte = atomic_readandclear_int(ptq); 1797 if (oldpte & PG_W) 1798 pmap->pm_stats.wired_count -= 1; 1799 /* 1800 * Machines that don't support invlpg, also don't support 1801 * PG_G. 1802 */ 1803 if (oldpte & PG_G) 1804 invlpg(va); 1805 pmap->pm_stats.resident_count -= 1; 1806 if (oldpte & PG_MANAGED) { 1807 m = PHYS_TO_VM_PAGE(oldpte); 1808 if (oldpte & PG_M) { 1809#if defined(PMAP_DIAGNOSTIC) 1810 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1811 printf( 1812 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1813 va, oldpte); 1814 } 1815#endif 1816 if (pmap_track_modified(va)) 1817 vm_page_dirty(m); 1818 } 1819 if (oldpte & PG_A) 1820 vm_page_flag_set(m, PG_REFERENCED); 1821 return pmap_remove_entry(pmap, m, va); 1822 } else { 1823 return pmap_unuse_pt(pmap, va, NULL); 1824 } 1825 1826 return 0; 1827} 1828 1829/* 1830 * Remove a single page from a process address space 1831 */ 1832static void 1833pmap_remove_page(pmap_t pmap, vm_offset_t va) 1834{ 1835 register pt_entry_t *ptq; 1836 1837 /* 1838 * if there is no pte for this address, just skip it!!! 1839 */ 1840 if (*pmap_pde(pmap, va) == 0) { 1841 return; 1842 } 1843 1844 /* 1845 * get a local va for mappings for this pmap. 1846 */ 1847 ptq = get_ptbase(pmap) + i386_btop(va); 1848 if (*ptq) { 1849 (void) pmap_remove_pte(pmap, ptq, va); 1850 pmap_invalidate_page(pmap, va); 1851 } 1852 return; 1853} 1854 1855/* 1856 * Remove the given range of addresses from the specified map. 1857 * 1858 * It is assumed that the start and end are properly 1859 * rounded to the page size. 1860 */ 1861void 1862pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1863{ 1864 register pt_entry_t *ptbase; 1865 vm_offset_t pdnxt; 1866 pd_entry_t ptpaddr; 1867 vm_offset_t sindex, eindex; 1868 int anyvalid; 1869 1870 if (pmap == NULL) 1871 return; 1872 1873 if (pmap->pm_stats.resident_count == 0) 1874 return; 1875 1876 /* 1877 * special handling of removing one page. a very 1878 * common operation and easy to short circuit some 1879 * code. 1880 */ 1881 if ((sva + PAGE_SIZE == eva) && 1882 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1883 pmap_remove_page(pmap, sva); 1884 return; 1885 } 1886 1887 anyvalid = 0; 1888 1889 /* 1890 * Get a local virtual address for the mappings that are being 1891 * worked with. 1892 */ 1893 ptbase = get_ptbase(pmap); 1894 1895 sindex = i386_btop(sva); 1896 eindex = i386_btop(eva); 1897 1898 for (; sindex < eindex; sindex = pdnxt) { 1899 unsigned pdirindex; 1900 1901 /* 1902 * Calculate index for next page table. 1903 */ 1904 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1)); 1905 if (pmap->pm_stats.resident_count == 0) 1906 break; 1907 1908 pdirindex = sindex / NPDEPG; 1909 ptpaddr = pmap->pm_pdir[pdirindex]; 1910 if ((ptpaddr & PG_PS) != 0) { 1911 pmap->pm_pdir[pdirindex] = 0; 1912 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1913 anyvalid++; 1914 continue; 1915 } 1916 1917 /* 1918 * Weed out invalid mappings. Note: we assume that the page 1919 * directory table is always allocated, and in kernel virtual. 1920 */ 1921 if (ptpaddr == 0) 1922 continue; 1923 1924 /* 1925 * Limit our scan to either the end of the va represented 1926 * by the current page table page, or to the end of the 1927 * range being removed. 1928 */ 1929 if (pdnxt > eindex) { 1930 pdnxt = eindex; 1931 } 1932 1933 for (; sindex != pdnxt; sindex++) { 1934 vm_offset_t va; 1935 if (ptbase[sindex] == 0) { 1936 continue; 1937 } 1938 va = i386_ptob(sindex); 1939 1940 anyvalid++; 1941 if (pmap_remove_pte(pmap, 1942 ptbase + sindex, va)) 1943 break; 1944 } 1945 } 1946 1947 if (anyvalid) 1948 pmap_invalidate_all(pmap); 1949} 1950 1951/* 1952 * Routine: pmap_remove_all 1953 * Function: 1954 * Removes this physical page from 1955 * all physical maps in which it resides. 1956 * Reflects back modify bits to the pager. 1957 * 1958 * Notes: 1959 * Original versions of this routine were very 1960 * inefficient because they iteratively called 1961 * pmap_remove (slow...) 1962 */ 1963 1964static void 1965pmap_remove_all(vm_page_t m) 1966{ 1967 register pv_entry_t pv; 1968 pt_entry_t *pte, tpte; 1969 int s; 1970 1971#if defined(PMAP_DIAGNOSTIC) 1972 /* 1973 * XXX this makes pmap_page_protect(NONE) illegal for non-managed 1974 * pages! 1975 */ 1976 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1977 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%x", VM_PAGE_TO_PHYS(m)); 1978 } 1979#endif 1980 1981 s = splvm(); 1982 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1983 pv->pv_pmap->pm_stats.resident_count--; 1984 1985 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1986 1987 tpte = atomic_readandclear_int(pte); 1988 if (tpte & PG_W) 1989 pv->pv_pmap->pm_stats.wired_count--; 1990 1991 if (tpte & PG_A) 1992 vm_page_flag_set(m, PG_REFERENCED); 1993 1994 /* 1995 * Update the vm_page_t clean and reference bits. 1996 */ 1997 if (tpte & PG_M) { 1998#if defined(PMAP_DIAGNOSTIC) 1999 if (pmap_nw_modified((pt_entry_t) tpte)) { 2000 printf( 2001 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 2002 pv->pv_va, tpte); 2003 } 2004#endif 2005 if (pmap_track_modified(pv->pv_va)) 2006 vm_page_dirty(m); 2007 } 2008 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2009 2010 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2011 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2012 m->md.pv_list_count--; 2013 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2014 free_pv_entry(pv); 2015 } 2016 2017 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 2018 2019 splx(s); 2020} 2021 2022/* 2023 * Set the physical protection on the 2024 * specified range of this map as requested. 2025 */ 2026void 2027pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 2028{ 2029 register pt_entry_t *ptbase; 2030 vm_offset_t pdnxt; 2031 pd_entry_t ptpaddr; 2032 vm_pindex_t sindex, eindex; 2033 int anychanged; 2034 2035 if (pmap == NULL) 2036 return; 2037 2038 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 2039 pmap_remove(pmap, sva, eva); 2040 return; 2041 } 2042 2043 if (prot & VM_PROT_WRITE) 2044 return; 2045 2046 anychanged = 0; 2047 2048 ptbase = get_ptbase(pmap); 2049 2050 sindex = i386_btop(sva); 2051 eindex = i386_btop(eva); 2052 2053 for (; sindex < eindex; sindex = pdnxt) { 2054 2055 unsigned pdirindex; 2056 2057 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1)); 2058 2059 pdirindex = sindex / NPDEPG; 2060 ptpaddr = pmap->pm_pdir[pdirindex]; 2061 if ((ptpaddr & PG_PS) != 0) { 2062 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 2063 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 2064 anychanged++; 2065 continue; 2066 } 2067 2068 /* 2069 * Weed out invalid mappings. Note: we assume that the page 2070 * directory table is always allocated, and in kernel virtual. 2071 */ 2072 if (ptpaddr == 0) 2073 continue; 2074 2075 if (pdnxt > eindex) { 2076 pdnxt = eindex; 2077 } 2078 2079 for (; sindex != pdnxt; sindex++) { 2080 2081 pt_entry_t pbits; 2082 vm_page_t m; 2083 2084 pbits = ptbase[sindex]; 2085 2086 if (pbits & PG_MANAGED) { 2087 m = NULL; 2088 if (pbits & PG_A) { 2089 m = PHYS_TO_VM_PAGE(pbits); 2090 vm_page_flag_set(m, PG_REFERENCED); 2091 pbits &= ~PG_A; 2092 } 2093 if (pbits & PG_M) { 2094 if (pmap_track_modified(i386_ptob(sindex))) { 2095 if (m == NULL) 2096 m = PHYS_TO_VM_PAGE(pbits); 2097 vm_page_dirty(m); 2098 pbits &= ~PG_M; 2099 } 2100 } 2101 } 2102 2103 pbits &= ~PG_RW; 2104 2105 if (pbits != ptbase[sindex]) { 2106 ptbase[sindex] = pbits; 2107 anychanged = 1; 2108 } 2109 } 2110 } 2111 if (anychanged) 2112 pmap_invalidate_all(pmap); 2113} 2114 2115/* 2116 * Insert the given physical page (p) at 2117 * the specified virtual address (v) in the 2118 * target physical map with the protection requested. 2119 * 2120 * If specified, the page will be wired down, meaning 2121 * that the related pte can not be reclaimed. 2122 * 2123 * NB: This is the only routine which MAY NOT lazy-evaluate 2124 * or lose information. That is, this routine must actually 2125 * insert this page into the given map NOW. 2126 */ 2127void 2128pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 2129 boolean_t wired) 2130{ 2131 vm_offset_t pa; 2132 register pt_entry_t *pte; 2133 vm_offset_t opa; 2134 pt_entry_t origpte, newpte; 2135 vm_page_t mpte; 2136 2137 if (pmap == NULL) 2138 return; 2139 2140 va &= PG_FRAME; 2141#ifdef PMAP_DIAGNOSTIC 2142 if (va > VM_MAX_KERNEL_ADDRESS) 2143 panic("pmap_enter: toobig"); 2144 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 2145 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 2146#endif 2147 2148 mpte = NULL; 2149 /* 2150 * In the case that a page table page is not 2151 * resident, we are creating it here. 2152 */ 2153 if (va < VM_MAXUSER_ADDRESS) { 2154 mpte = pmap_allocpte(pmap, va); 2155 } 2156#if 0 && defined(PMAP_DIAGNOSTIC) 2157 else { 2158 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 2159 if (((origpte = *pdeaddr) & PG_V) == 0) { 2160 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 2161 pmap->pm_pdir[PTDPTDI], origpte, va); 2162 } 2163 } 2164#endif 2165 2166 pte = pmap_pte(pmap, va); 2167 2168 /* 2169 * Page Directory table entry not valid, we need a new PT page 2170 */ 2171 if (pte == NULL) { 2172 panic("pmap_enter: invalid page directory, pdir=%p, va=0x%x\n", 2173 (void *)pmap->pm_pdir[PTDPTDI], va); 2174 } 2175 2176 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 2177 origpte = *(vm_offset_t *)pte; 2178 opa = origpte & PG_FRAME; 2179 2180 if (origpte & PG_PS) 2181 panic("pmap_enter: attempted pmap_enter on 4MB page"); 2182 2183 /* 2184 * Mapping has not changed, must be protection or wiring change. 2185 */ 2186 if (origpte && (opa == pa)) { 2187 /* 2188 * Wiring change, just update stats. We don't worry about 2189 * wiring PT pages as they remain resident as long as there 2190 * are valid mappings in them. Hence, if a user page is wired, 2191 * the PT page will be also. 2192 */ 2193 if (wired && ((origpte & PG_W) == 0)) 2194 pmap->pm_stats.wired_count++; 2195 else if (!wired && (origpte & PG_W)) 2196 pmap->pm_stats.wired_count--; 2197 2198#if defined(PMAP_DIAGNOSTIC) 2199 if (pmap_nw_modified((pt_entry_t) origpte)) { 2200 printf( 2201 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 2202 va, origpte); 2203 } 2204#endif 2205 2206 /* 2207 * Remove extra pte reference 2208 */ 2209 if (mpte) 2210 mpte->hold_count--; 2211 2212 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) { 2213 if ((origpte & PG_RW) == 0) { 2214 *pte |= PG_RW; 2215#ifdef SMP 2216 cpu_invlpg((void *)va); 2217 if (pmap->pm_active & PCPU_GET(other_cpus)) 2218 smp_invltlb(); 2219#else 2220 invltlb_1pg(va); 2221#endif 2222 } 2223 return; 2224 } 2225 2226 /* 2227 * We might be turning off write access to the page, 2228 * so we go ahead and sense modify status. 2229 */ 2230 if (origpte & PG_MANAGED) { 2231 if ((origpte & PG_M) && pmap_track_modified(va)) { 2232 vm_page_t om; 2233 om = PHYS_TO_VM_PAGE(opa); 2234 vm_page_dirty(om); 2235 } 2236 pa |= PG_MANAGED; 2237 } 2238 goto validate; 2239 } 2240 /* 2241 * Mapping has changed, invalidate old range and fall through to 2242 * handle validating new mapping. 2243 */ 2244 if (opa) { 2245 int err; 2246 err = pmap_remove_pte(pmap, pte, va); 2247 if (err) 2248 panic("pmap_enter: pte vanished, va: 0x%x", va); 2249 } 2250 2251 /* 2252 * Enter on the PV list if part of our managed memory. Note that we 2253 * raise IPL while manipulating pv_table since pmap_enter can be 2254 * called at interrupt time. 2255 */ 2256 if (pmap_initialized && 2257 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 2258 pmap_insert_entry(pmap, va, mpte, m); 2259 pa |= PG_MANAGED; 2260 } 2261 2262 /* 2263 * Increment counters 2264 */ 2265 pmap->pm_stats.resident_count++; 2266 if (wired) 2267 pmap->pm_stats.wired_count++; 2268 2269validate: 2270 /* 2271 * Now validate mapping with desired protection/wiring. 2272 */ 2273 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V); 2274 2275 if (wired) 2276 newpte |= PG_W; 2277 if (va < VM_MAXUSER_ADDRESS) 2278 newpte |= PG_U; 2279 if (pmap == kernel_pmap) 2280 newpte |= pgeflag; 2281 2282 /* 2283 * if the mapping or permission bits are different, we need 2284 * to update the pte. 2285 */ 2286 if ((origpte & ~(PG_M|PG_A)) != newpte) { 2287 *pte = newpte | PG_A; 2288 /*if (origpte)*/ { 2289#ifdef SMP 2290 cpu_invlpg((void *)va); 2291 if (pmap->pm_active & PCPU_GET(other_cpus)) 2292 smp_invltlb(); 2293#else 2294 invltlb_1pg(va); 2295#endif 2296 } 2297 } 2298} 2299 2300/* 2301 * this code makes some *MAJOR* assumptions: 2302 * 1. Current pmap & pmap exists. 2303 * 2. Not wired. 2304 * 3. Read access. 2305 * 4. No page table pages. 2306 * 5. Tlbflush is deferred to calling procedure. 2307 * 6. Page IS managed. 2308 * but is *MUCH* faster than pmap_enter... 2309 */ 2310 2311static vm_page_t 2312pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2313{ 2314 pt_entry_t *pte; 2315 vm_offset_t pa; 2316 2317 /* 2318 * In the case that a page table page is not 2319 * resident, we are creating it here. 2320 */ 2321 if (va < VM_MAXUSER_ADDRESS) { 2322 unsigned ptepindex; 2323 pd_entry_t ptepa; 2324 2325 /* 2326 * Calculate pagetable page index 2327 */ 2328 ptepindex = va >> PDRSHIFT; 2329 if (mpte && (mpte->pindex == ptepindex)) { 2330 mpte->hold_count++; 2331 } else { 2332retry: 2333 /* 2334 * Get the page directory entry 2335 */ 2336 ptepa = pmap->pm_pdir[ptepindex]; 2337 2338 /* 2339 * If the page table page is mapped, we just increment 2340 * the hold count, and activate it. 2341 */ 2342 if (ptepa) { 2343 if (ptepa & PG_PS) 2344 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2345 if (pmap->pm_ptphint && 2346 (pmap->pm_ptphint->pindex == ptepindex)) { 2347 mpte = pmap->pm_ptphint; 2348 } else { 2349 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex); 2350 pmap->pm_ptphint = mpte; 2351 } 2352 if (mpte == NULL) 2353 goto retry; 2354 mpte->hold_count++; 2355 } else { 2356 mpte = _pmap_allocpte(pmap, ptepindex); 2357 } 2358 } 2359 } else { 2360 mpte = NULL; 2361 } 2362 2363 /* 2364 * This call to vtopte makes the assumption that we are 2365 * entering the page into the current pmap. In order to support 2366 * quick entry into any pmap, one would likely use pmap_pte_quick. 2367 * But that isn't as quick as vtopte. 2368 */ 2369 pte = vtopte(va); 2370 if (*pte) { 2371 if (mpte) 2372 pmap_unwire_pte_hold(pmap, mpte); 2373 return 0; 2374 } 2375 2376 /* 2377 * Enter on the PV list if part of our managed memory. Note that we 2378 * raise IPL while manipulating pv_table since pmap_enter can be 2379 * called at interrupt time. 2380 */ 2381 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2382 pmap_insert_entry(pmap, va, mpte, m); 2383 2384 /* 2385 * Increment counters 2386 */ 2387 pmap->pm_stats.resident_count++; 2388 2389 pa = VM_PAGE_TO_PHYS(m); 2390 2391 /* 2392 * Now validate mapping with RO protection 2393 */ 2394 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2395 *pte = pa | PG_V | PG_U; 2396 else 2397 *pte = pa | PG_V | PG_U | PG_MANAGED; 2398 2399 return mpte; 2400} 2401 2402/* 2403 * Make a temporary mapping for a physical address. This is only intended 2404 * to be used for panic dumps. 2405 */ 2406void * 2407pmap_kenter_temporary(vm_offset_t pa, int i) 2408{ 2409 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa); 2410 return ((void *)crashdumpmap); 2411} 2412 2413#define MAX_INIT_PT (96) 2414/* 2415 * pmap_object_init_pt preloads the ptes for a given object 2416 * into the specified pmap. This eliminates the blast of soft 2417 * faults on process startup and immediately after an mmap. 2418 */ 2419void 2420pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2421 vm_object_t object, vm_pindex_t pindex, 2422 vm_size_t size, int limit) 2423{ 2424 vm_offset_t tmpidx; 2425 int psize; 2426 vm_page_t p, mpte; 2427 int objpgs; 2428 2429 if (pmap == NULL || object == NULL) 2430 return; 2431 2432 /* 2433 * This code maps large physical mmap regions into the 2434 * processor address space. Note that some shortcuts 2435 * are taken, but the code works. 2436 */ 2437 if (pseflag && (object->type == OBJT_DEVICE) && 2438 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2439 int i; 2440 vm_page_t m[1]; 2441 unsigned int ptepindex; 2442 int npdes; 2443 pd_entry_t ptepa; 2444 2445 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2446 return; 2447 2448retry: 2449 p = vm_page_lookup(object, pindex); 2450 if (p && vm_page_sleep_busy(p, FALSE, "init4p")) 2451 goto retry; 2452 2453 if (p == NULL) { 2454 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2455 if (p == NULL) 2456 return; 2457 m[0] = p; 2458 2459 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2460 vm_page_free(p); 2461 return; 2462 } 2463 2464 p = vm_page_lookup(object, pindex); 2465 vm_page_wakeup(p); 2466 } 2467 2468 ptepa = VM_PAGE_TO_PHYS(p); 2469 if (ptepa & (NBPDR - 1)) { 2470 return; 2471 } 2472 2473 p->valid = VM_PAGE_BITS_ALL; 2474 2475 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2476 npdes = size >> PDRSHIFT; 2477 for(i = 0; i < npdes; i++) { 2478 pmap->pm_pdir[ptepindex] = 2479 ptepa | PG_U | PG_RW | PG_V | PG_PS; 2480 ptepa += NBPDR; 2481 ptepindex += 1; 2482 } 2483 vm_page_flag_set(p, PG_MAPPED); 2484 invltlb(); 2485 return; 2486 } 2487 2488 psize = i386_btop(size); 2489 2490 if ((object->type != OBJT_VNODE) || 2491 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) && 2492 (object->resident_page_count > MAX_INIT_PT))) { 2493 return; 2494 } 2495 2496 if (psize + pindex > object->size) { 2497 if (object->size < pindex) 2498 return; 2499 psize = object->size - pindex; 2500 } 2501 2502 mpte = NULL; 2503 /* 2504 * if we are processing a major portion of the object, then scan the 2505 * entire thing. 2506 */ 2507 if (psize > (object->resident_page_count >> 2)) { 2508 objpgs = psize; 2509 2510 for (p = TAILQ_FIRST(&object->memq); 2511 ((objpgs > 0) && (p != NULL)); 2512 p = TAILQ_NEXT(p, listq)) { 2513 2514 tmpidx = p->pindex; 2515 if (tmpidx < pindex) { 2516 continue; 2517 } 2518 tmpidx -= pindex; 2519 if (tmpidx >= psize) { 2520 continue; 2521 } 2522 /* 2523 * don't allow an madvise to blow away our really 2524 * free pages allocating pv entries. 2525 */ 2526 if ((limit & MAP_PREFAULT_MADVISE) && 2527 cnt.v_free_count < cnt.v_free_reserved) { 2528 break; 2529 } 2530 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2531 (p->busy == 0) && 2532 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2533 if ((p->queue - p->pc) == PQ_CACHE) 2534 vm_page_deactivate(p); 2535 vm_page_busy(p); 2536 mpte = pmap_enter_quick(pmap, 2537 addr + i386_ptob(tmpidx), p, mpte); 2538 vm_page_flag_set(p, PG_MAPPED); 2539 vm_page_wakeup(p); 2540 } 2541 objpgs -= 1; 2542 } 2543 } else { 2544 /* 2545 * else lookup the pages one-by-one. 2546 */ 2547 for (tmpidx = 0; tmpidx < psize; tmpidx += 1) { 2548 /* 2549 * don't allow an madvise to blow away our really 2550 * free pages allocating pv entries. 2551 */ 2552 if ((limit & MAP_PREFAULT_MADVISE) && 2553 cnt.v_free_count < cnt.v_free_reserved) { 2554 break; 2555 } 2556 p = vm_page_lookup(object, tmpidx + pindex); 2557 if (p && 2558 ((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2559 (p->busy == 0) && 2560 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2561 if ((p->queue - p->pc) == PQ_CACHE) 2562 vm_page_deactivate(p); 2563 vm_page_busy(p); 2564 mpte = pmap_enter_quick(pmap, 2565 addr + i386_ptob(tmpidx), p, mpte); 2566 vm_page_flag_set(p, PG_MAPPED); 2567 vm_page_wakeup(p); 2568 } 2569 } 2570 } 2571 return; 2572} 2573 2574/* 2575 * pmap_prefault provides a quick way of clustering 2576 * pagefaults into a processes address space. It is a "cousin" 2577 * of pmap_object_init_pt, except it runs at page fault time instead 2578 * of mmap time. 2579 */ 2580#define PFBAK 4 2581#define PFFOR 4 2582#define PAGEORDER_SIZE (PFBAK+PFFOR) 2583 2584static int pmap_prefault_pageorder[] = { 2585 -PAGE_SIZE, PAGE_SIZE, 2586 -2 * PAGE_SIZE, 2 * PAGE_SIZE, 2587 -3 * PAGE_SIZE, 3 * PAGE_SIZE 2588 -4 * PAGE_SIZE, 4 * PAGE_SIZE 2589}; 2590 2591void 2592pmap_prefault(pmap, addra, entry) 2593 pmap_t pmap; 2594 vm_offset_t addra; 2595 vm_map_entry_t entry; 2596{ 2597 int i; 2598 vm_offset_t starta; 2599 vm_offset_t addr; 2600 vm_pindex_t pindex; 2601 vm_page_t m, mpte; 2602 vm_object_t object; 2603 2604 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) 2605 return; 2606 2607 object = entry->object.vm_object; 2608 2609 starta = addra - PFBAK * PAGE_SIZE; 2610 if (starta < entry->start) { 2611 starta = entry->start; 2612 } else if (starta > addra) { 2613 starta = 0; 2614 } 2615 2616 mpte = NULL; 2617 for (i = 0; i < PAGEORDER_SIZE; i++) { 2618 vm_object_t lobject; 2619 pt_entry_t *pte; 2620 2621 addr = addra + pmap_prefault_pageorder[i]; 2622 if (addr > addra + (PFFOR * PAGE_SIZE)) 2623 addr = 0; 2624 2625 if (addr < starta || addr >= entry->end) 2626 continue; 2627 2628 if ((*pmap_pde(pmap, addr)) == NULL) 2629 continue; 2630 2631 pte = vtopte(addr); 2632 if (*pte) 2633 continue; 2634 2635 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2636 lobject = object; 2637 for (m = vm_page_lookup(lobject, pindex); 2638 (!m && (lobject->type == OBJT_DEFAULT) && (lobject->backing_object)); 2639 lobject = lobject->backing_object) { 2640 if (lobject->backing_object_offset & PAGE_MASK) 2641 break; 2642 pindex += (lobject->backing_object_offset >> PAGE_SHIFT); 2643 m = vm_page_lookup(lobject->backing_object, pindex); 2644 } 2645 2646 /* 2647 * give-up when a page is not in memory 2648 */ 2649 if (m == NULL) 2650 break; 2651 2652 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2653 (m->busy == 0) && 2654 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2655 2656 if ((m->queue - m->pc) == PQ_CACHE) { 2657 vm_page_deactivate(m); 2658 } 2659 vm_page_busy(m); 2660 mpte = pmap_enter_quick(pmap, addr, m, mpte); 2661 vm_page_flag_set(m, PG_MAPPED); 2662 vm_page_wakeup(m); 2663 } 2664 } 2665} 2666 2667/* 2668 * Routine: pmap_change_wiring 2669 * Function: Change the wiring attribute for a map/virtual-address 2670 * pair. 2671 * In/out conditions: 2672 * The mapping must already exist in the pmap. 2673 */ 2674void 2675pmap_change_wiring(pmap, va, wired) 2676 register pmap_t pmap; 2677 vm_offset_t va; 2678 boolean_t wired; 2679{ 2680 register pt_entry_t *pte; 2681 2682 if (pmap == NULL) 2683 return; 2684 2685 pte = pmap_pte(pmap, va); 2686 2687 if (wired && !pmap_pte_w(pte)) 2688 pmap->pm_stats.wired_count++; 2689 else if (!wired && pmap_pte_w(pte)) 2690 pmap->pm_stats.wired_count--; 2691 2692 /* 2693 * Wiring is not a hardware characteristic so there is no need to 2694 * invalidate TLB. 2695 */ 2696 pmap_pte_set_w(pte, wired); 2697} 2698 2699 2700 2701/* 2702 * Copy the range specified by src_addr/len 2703 * from the source map to the range dst_addr/len 2704 * in the destination map. 2705 * 2706 * This routine is only advisory and need not do anything. 2707 */ 2708 2709void 2710pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2711 vm_offset_t src_addr) 2712{ 2713 vm_offset_t addr; 2714 vm_offset_t end_addr = src_addr + len; 2715 vm_offset_t pdnxt; 2716 pd_entry_t src_frame, dst_frame; 2717 vm_page_t m; 2718 pd_entry_t saved_pde; 2719 2720 if (dst_addr != src_addr) 2721 return; 2722 2723 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME; 2724 if (src_frame != (PTDpde & PG_FRAME)) 2725 return; 2726 2727 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME; 2728 if (dst_frame != (APTDpde & PG_FRAME)) { 2729 APTDpde = dst_frame | PG_RW | PG_V; 2730#if defined(SMP) 2731 /* The page directory is not shared between CPUs */ 2732 cpu_invltlb(); 2733#else 2734 invltlb(); 2735#endif 2736 } 2737 saved_pde = APTDpde & (PG_FRAME | PG_RW | PG_V); 2738 for(addr = src_addr; addr < end_addr; addr = pdnxt) { 2739 pt_entry_t *src_pte, *dst_pte; 2740 vm_page_t dstmpte, srcmpte; 2741 pd_entry_t srcptepaddr; 2742 unsigned ptepindex; 2743 2744 if (addr >= UPT_MIN_ADDRESS) 2745 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2746 2747 /* 2748 * Don't let optional prefaulting of pages make us go 2749 * way below the low water mark of free pages or way 2750 * above high water mark of used pv entries. 2751 */ 2752 if (cnt.v_free_count < cnt.v_free_reserved || 2753 pv_entry_count > pv_entry_high_water) 2754 break; 2755 2756 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1)); 2757 ptepindex = addr >> PDRSHIFT; 2758 2759 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2760 if (srcptepaddr == 0) 2761 continue; 2762 2763 if (srcptepaddr & PG_PS) { 2764 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2765 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2766 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; 2767 } 2768 continue; 2769 } 2770 2771 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex); 2772 if ((srcmpte == NULL) || 2773 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY)) 2774 continue; 2775 2776 if (pdnxt > end_addr) 2777 pdnxt = end_addr; 2778 2779 src_pte = vtopte(addr); 2780 dst_pte = avtopte(addr); 2781 while (addr < pdnxt) { 2782 pt_entry_t ptetemp; 2783 ptetemp = *src_pte; 2784 /* 2785 * we only virtual copy managed pages 2786 */ 2787 if ((ptetemp & PG_MANAGED) != 0) { 2788 /* 2789 * We have to check after allocpte for the 2790 * pte still being around... allocpte can 2791 * block. 2792 */ 2793 dstmpte = pmap_allocpte(dst_pmap, addr); 2794 if ((APTDpde & PG_FRAME) != 2795 (saved_pde & PG_FRAME)) { 2796 APTDpde = saved_pde; 2797printf ("IT HAPPENNED!"); 2798#if defined(SMP) 2799 cpu_invltlb(); 2800#else 2801 invltlb(); 2802#endif 2803 } 2804 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2805 /* 2806 * Clear the modified and 2807 * accessed (referenced) bits 2808 * during the copy. 2809 */ 2810 m = PHYS_TO_VM_PAGE(ptetemp); 2811 *dst_pte = ptetemp & ~(PG_M | PG_A); 2812 dst_pmap->pm_stats.resident_count++; 2813 pmap_insert_entry(dst_pmap, addr, 2814 dstmpte, m); 2815 } else { 2816 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2817 } 2818 if (dstmpte->hold_count >= srcmpte->hold_count) 2819 break; 2820 } 2821 addr += PAGE_SIZE; 2822 src_pte++; 2823 dst_pte++; 2824 } 2825 } 2826} 2827 2828/* 2829 * Routine: pmap_kernel 2830 * Function: 2831 * Returns the physical map handle for the kernel. 2832 */ 2833pmap_t 2834pmap_kernel() 2835{ 2836 return (kernel_pmap); 2837} 2838 2839/* 2840 * pmap_zero_page zeros the specified hardware page by mapping 2841 * the page into KVM and using bzero to clear its contents. 2842 */ 2843void 2844pmap_zero_page(vm_offset_t phys) 2845{ 2846 2847 if (*CMAP2) 2848 panic("pmap_zero_page: CMAP2 busy"); 2849 2850 *CMAP2 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M; 2851 invltlb_1pg((vm_offset_t)CADDR2); 2852 2853#if defined(I686_CPU) 2854 if (cpu_class == CPUCLASS_686) 2855 i686_pagezero(CADDR2); 2856 else 2857#endif 2858 bzero(CADDR2, PAGE_SIZE); 2859 *CMAP2 = 0; 2860} 2861 2862/* 2863 * pmap_zero_page_area zeros the specified hardware page by mapping 2864 * the page into KVM and using bzero to clear its contents. 2865 * 2866 * off and size may not cover an area beyond a single hardware page. 2867 */ 2868void 2869pmap_zero_page_area(vm_offset_t phys, int off, int size) 2870{ 2871 2872 if (*CMAP2) 2873 panic("pmap_zero_page: CMAP2 busy"); 2874 2875 *CMAP2 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M; 2876 invltlb_1pg((vm_offset_t)CADDR2); 2877 2878#if defined(I686_CPU) 2879 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE) 2880 i686_pagezero(CADDR2); 2881 else 2882#endif 2883 bzero((char *)CADDR2 + off, size); 2884 *CMAP2 = 0; 2885} 2886 2887/* 2888 * pmap_copy_page copies the specified (machine independent) 2889 * page by mapping the page into virtual memory and using 2890 * bcopy to copy the page, one machine dependent page at a 2891 * time. 2892 */ 2893void 2894pmap_copy_page(vm_offset_t src, vm_offset_t dst) 2895{ 2896 2897 if (*CMAP1) 2898 panic("pmap_copy_page: CMAP1 busy"); 2899 if (*CMAP2) 2900 panic("pmap_copy_page: CMAP2 busy"); 2901 2902 *CMAP1 = PG_V | (src & PG_FRAME) | PG_A; 2903 *CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M; 2904#ifdef I386_CPU 2905 invltlb(); 2906#else 2907 invlpg((u_int)CADDR1); 2908 invlpg((u_int)CADDR2); 2909#endif 2910 2911 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2912 2913 *CMAP1 = 0; 2914 *CMAP2 = 0; 2915} 2916 2917 2918/* 2919 * Routine: pmap_pageable 2920 * Function: 2921 * Make the specified pages (by pmap, offset) 2922 * pageable (or not) as requested. 2923 * 2924 * A page which is not pageable may not take 2925 * a fault; therefore, its page table entry 2926 * must remain valid for the duration. 2927 * 2928 * This routine is merely advisory; pmap_enter 2929 * will specify that these pages are to be wired 2930 * down (or not) as appropriate. 2931 */ 2932void 2933pmap_pageable(pmap, sva, eva, pageable) 2934 pmap_t pmap; 2935 vm_offset_t sva, eva; 2936 boolean_t pageable; 2937{ 2938} 2939 2940/* 2941 * this routine returns true if a physical page resides 2942 * in the given pmap. 2943 */ 2944boolean_t 2945pmap_page_exists(pmap, m) 2946 pmap_t pmap; 2947 vm_page_t m; 2948{ 2949 register pv_entry_t pv; 2950 int s; 2951 2952 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2953 return FALSE; 2954 2955 s = splvm(); 2956 2957 /* 2958 * Not found, check current mappings returning immediately if found. 2959 */ 2960 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2961 if (pv->pv_pmap == pmap) { 2962 splx(s); 2963 return TRUE; 2964 } 2965 } 2966 splx(s); 2967 return (FALSE); 2968} 2969 2970#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2971/* 2972 * Remove all pages from specified address space 2973 * this aids process exit speeds. Also, this code 2974 * is special cased for current process only, but 2975 * can have the more generic (and slightly slower) 2976 * mode enabled. This is much faster than pmap_remove 2977 * in the case of running down an entire address space. 2978 */ 2979void 2980pmap_remove_pages(pmap, sva, eva) 2981 pmap_t pmap; 2982 vm_offset_t sva, eva; 2983{ 2984 pt_entry_t *pte, tpte; 2985 vm_page_t m; 2986 pv_entry_t pv, npv; 2987 int s; 2988 2989#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2990 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2991 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2992 return; 2993 } 2994#endif 2995 2996 s = splvm(); 2997 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2998 2999 if (pv->pv_va >= eva || pv->pv_va < sva) { 3000 npv = TAILQ_NEXT(pv, pv_plist); 3001 continue; 3002 } 3003 3004#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 3005 pte = vtopte(pv->pv_va); 3006#else 3007 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3008#endif 3009 tpte = *pte; 3010 3011 if (tpte == 0) { 3012 printf("TPTE at %p IS ZERO @ VA %08x\n", 3013 pte, pv->pv_va); 3014 panic("bad pte"); 3015 } 3016 3017/* 3018 * We cannot remove wired pages from a process' mapping at this time 3019 */ 3020 if (tpte & PG_W) { 3021 npv = TAILQ_NEXT(pv, pv_plist); 3022 continue; 3023 } 3024 3025 m = PHYS_TO_VM_PAGE(tpte); 3026 KASSERT(m->phys_addr == (tpte & PG_FRAME), 3027 ("vm_page_t %p phys_addr mismatch %08x %08x", 3028 m, m->phys_addr, tpte)); 3029 3030 KASSERT(m < &vm_page_array[vm_page_array_size], 3031 ("pmap_remove_pages: bad tpte %x", tpte)); 3032 3033 pv->pv_pmap->pm_stats.resident_count--; 3034 3035 *pte = 0; 3036 3037 /* 3038 * Update the vm_page_t clean and reference bits. 3039 */ 3040 if (tpte & PG_M) { 3041 vm_page_dirty(m); 3042 } 3043 3044 npv = TAILQ_NEXT(pv, pv_plist); 3045 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 3046 3047 m->md.pv_list_count--; 3048 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 3049 if (TAILQ_FIRST(&m->md.pv_list) == NULL) { 3050 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); 3051 } 3052 3053 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 3054 free_pv_entry(pv); 3055 } 3056 splx(s); 3057 pmap_invalidate_all(pmap); 3058} 3059 3060/* 3061 * pmap_testbit tests bits in pte's 3062 * note that the testbit/changebit routines are inline, 3063 * and a lot of things compile-time evaluate. 3064 */ 3065static boolean_t 3066pmap_testbit(m, bit) 3067 vm_page_t m; 3068 int bit; 3069{ 3070 pv_entry_t pv; 3071 pt_entry_t *pte; 3072 int s; 3073 3074 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3075 return FALSE; 3076 3077 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 3078 return FALSE; 3079 3080 s = splvm(); 3081 3082 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3083 /* 3084 * if the bit being tested is the modified bit, then 3085 * mark clean_map and ptes as never 3086 * modified. 3087 */ 3088 if (bit & (PG_A|PG_M)) { 3089 if (!pmap_track_modified(pv->pv_va)) 3090 continue; 3091 } 3092 3093#if defined(PMAP_DIAGNOSTIC) 3094 if (!pv->pv_pmap) { 3095 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 3096 continue; 3097 } 3098#endif 3099 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3100 if (*pte & bit) { 3101 splx(s); 3102 return TRUE; 3103 } 3104 } 3105 splx(s); 3106 return (FALSE); 3107} 3108 3109/* 3110 * this routine is used to modify bits in ptes 3111 */ 3112static __inline void 3113pmap_changebit(vm_page_t m, int bit, boolean_t setem) 3114{ 3115 register pv_entry_t pv; 3116 register pt_entry_t *pte; 3117 int s; 3118 3119 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3120 return; 3121 3122 s = splvm(); 3123 3124 /* 3125 * Loop over all current mappings setting/clearing as appropos If 3126 * setting RO do we need to clear the VAC? 3127 */ 3128 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3129 /* 3130 * don't write protect pager mappings 3131 */ 3132 if (!setem && (bit == PG_RW)) { 3133 if (!pmap_track_modified(pv->pv_va)) 3134 continue; 3135 } 3136 3137#if defined(PMAP_DIAGNOSTIC) 3138 if (!pv->pv_pmap) { 3139 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 3140 continue; 3141 } 3142#endif 3143 3144 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3145 3146 if (setem) { 3147 *pte |= bit; 3148 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3149 } else { 3150 pt_entry_t pbits = *pte; 3151 if (pbits & bit) { 3152 if (bit == PG_RW) { 3153 if (pbits & PG_M) { 3154 vm_page_dirty(m); 3155 } 3156 *pte = pbits & ~(PG_M|PG_RW); 3157 } else { 3158 *pte = pbits & ~bit; 3159 } 3160 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3161 } 3162 } 3163 } 3164 splx(s); 3165} 3166 3167/* 3168 * pmap_page_protect: 3169 * 3170 * Lower the permission for all mappings to a given page. 3171 */ 3172void 3173pmap_page_protect(vm_page_t m, vm_prot_t prot) 3174{ 3175 if ((prot & VM_PROT_WRITE) == 0) { 3176 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 3177 pmap_changebit(m, PG_RW, FALSE); 3178 } else { 3179 pmap_remove_all(m); 3180 } 3181 } 3182} 3183 3184vm_offset_t 3185pmap_phys_address(ppn) 3186 int ppn; 3187{ 3188 return (i386_ptob(ppn)); 3189} 3190 3191/* 3192 * pmap_ts_referenced: 3193 * 3194 * Return the count of reference bits for a page, clearing all of them. 3195 */ 3196int 3197pmap_ts_referenced(vm_page_t m) 3198{ 3199 register pv_entry_t pv, pvf, pvn; 3200 pt_entry_t *pte; 3201 int s; 3202 int rtval = 0; 3203 3204 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 3205 return (rtval); 3206 3207 s = splvm(); 3208 3209 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 3210 3211 pvf = pv; 3212 3213 do { 3214 pvn = TAILQ_NEXT(pv, pv_list); 3215 3216 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 3217 3218 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 3219 3220 if (!pmap_track_modified(pv->pv_va)) 3221 continue; 3222 3223 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 3224 3225 if (pte && (*pte & PG_A)) { 3226 *pte &= ~PG_A; 3227 3228 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 3229 3230 rtval++; 3231 if (rtval > 4) { 3232 break; 3233 } 3234 } 3235 } while ((pv = pvn) != NULL && pv != pvf); 3236 } 3237 splx(s); 3238 3239 return (rtval); 3240} 3241 3242/* 3243 * pmap_is_modified: 3244 * 3245 * Return whether or not the specified physical page was modified 3246 * in any physical maps. 3247 */ 3248boolean_t 3249pmap_is_modified(vm_page_t m) 3250{ 3251 return pmap_testbit(m, PG_M); 3252} 3253 3254/* 3255 * Clear the modify bits on the specified physical page. 3256 */ 3257void 3258pmap_clear_modify(vm_page_t m) 3259{ 3260 pmap_changebit(m, PG_M, FALSE); 3261} 3262 3263/* 3264 * pmap_clear_reference: 3265 * 3266 * Clear the reference bit on the specified physical page. 3267 */ 3268void 3269pmap_clear_reference(vm_page_t m) 3270{ 3271 pmap_changebit(m, PG_A, FALSE); 3272} 3273 3274/* 3275 * Miscellaneous support routines follow 3276 */ 3277 3278static void 3279i386_protection_init() 3280{ 3281 register int *kp, prot; 3282 3283 kp = protection_codes; 3284 for (prot = 0; prot < 8; prot++) { 3285 switch (prot) { 3286 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE: 3287 /* 3288 * Read access is also 0. There isn't any execute bit, 3289 * so just make it readable. 3290 */ 3291 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE: 3292 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE: 3293 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE: 3294 *kp++ = 0; 3295 break; 3296 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE: 3297 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE: 3298 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE: 3299 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE: 3300 *kp++ = PG_RW; 3301 break; 3302 } 3303 } 3304} 3305 3306/* 3307 * Map a set of physical memory pages into the kernel virtual 3308 * address space. Return a pointer to where it is mapped. This 3309 * routine is intended to be used for mapping device memory, 3310 * NOT real memory. 3311 */ 3312void * 3313pmap_mapdev(pa, size) 3314 vm_offset_t pa; 3315 vm_size_t size; 3316{ 3317 vm_offset_t va, tmpva, offset; 3318 pt_entry_t *pte; 3319 3320 offset = pa & PAGE_MASK; 3321 size = roundup(offset + size, PAGE_SIZE); 3322 3323 GIANT_REQUIRED; 3324 3325 va = kmem_alloc_pageable(kernel_map, size); 3326 if (!va) 3327 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 3328 3329 pa = pa & PG_FRAME; 3330 for (tmpva = va; size > 0;) { 3331 pte = vtopte(tmpva); 3332 *pte = pa | PG_RW | PG_V | pgeflag; 3333 size -= PAGE_SIZE; 3334 tmpva += PAGE_SIZE; 3335 pa += PAGE_SIZE; 3336 } 3337 invltlb(); 3338 3339 return ((void *)(va + offset)); 3340} 3341 3342void 3343pmap_unmapdev(va, size) 3344 vm_offset_t va; 3345 vm_size_t size; 3346{ 3347 vm_offset_t base, offset; 3348 3349 base = va & PG_FRAME; 3350 offset = va & PAGE_MASK; 3351 size = roundup(offset + size, PAGE_SIZE); 3352 kmem_free(kernel_map, base, size); 3353} 3354 3355/* 3356 * perform the pmap work for mincore 3357 */ 3358int 3359pmap_mincore(pmap, addr) 3360 pmap_t pmap; 3361 vm_offset_t addr; 3362{ 3363 pt_entry_t *ptep, pte; 3364 vm_page_t m; 3365 int val = 0; 3366 3367 ptep = pmap_pte(pmap, addr); 3368 if (ptep == 0) { 3369 return 0; 3370 } 3371 3372 if ((pte = *ptep) != 0) { 3373 vm_offset_t pa; 3374 3375 val = MINCORE_INCORE; 3376 if ((pte & PG_MANAGED) == 0) 3377 return val; 3378 3379 pa = pte & PG_FRAME; 3380 3381 m = PHYS_TO_VM_PAGE(pa); 3382 3383 /* 3384 * Modified by us 3385 */ 3386 if (pte & PG_M) 3387 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 3388 /* 3389 * Modified by someone 3390 */ 3391 else if (m->dirty || pmap_is_modified(m)) 3392 val |= MINCORE_MODIFIED_OTHER; 3393 /* 3394 * Referenced by us 3395 */ 3396 if (pte & PG_A) 3397 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 3398 3399 /* 3400 * Referenced by someone 3401 */ 3402 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) { 3403 val |= MINCORE_REFERENCED_OTHER; 3404 vm_page_flag_set(m, PG_REFERENCED); 3405 } 3406 } 3407 return val; 3408} 3409 3410void 3411pmap_activate(struct thread *td) 3412{ 3413 struct proc *p = td->td_proc; 3414 pmap_t pmap; 3415 u_int32_t cr3; 3416 3417 pmap = vmspace_pmap(td->td_proc->p_vmspace); 3418#if defined(SMP) 3419 pmap->pm_active |= 1 << PCPU_GET(cpuid); 3420#else 3421 pmap->pm_active |= 1; 3422#endif 3423#if defined(SWTCH_OPTIM_STATS) 3424 tlb_flush_count++; 3425#endif 3426 cr3 = vtophys(pmap->pm_pdir); 3427 /* XXXKSE this is wrong. 3428 * pmap_activate is for the current thread on the current cpu 3429 */ 3430 if (p->p_flag & P_KSES) { 3431 /* Make sure all other cr3 entries are updated. */ 3432 /* what if they are running? XXXKSE (maybe abort them) */ 3433 FOREACH_THREAD_IN_PROC(p, td) { 3434 td->td_pcb->pcb_cr3 = cr3; 3435 } 3436 } else { 3437 td->td_pcb->pcb_cr3 = cr3; 3438 } 3439 load_cr3(cr3); 3440} 3441 3442vm_offset_t 3443pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3444{ 3445 3446 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3447 return addr; 3448 } 3449 3450 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); 3451 return addr; 3452} 3453 3454 3455#if defined(PMAP_DEBUG) 3456pmap_pid_dump(int pid) 3457{ 3458 pmap_t pmap; 3459 struct proc *p; 3460 int npte = 0; 3461 int index; 3462 3463 sx_slock(&allproc_lock); 3464 LIST_FOREACH(p, &allproc, p_list) { 3465 if (p->p_pid != pid) 3466 continue; 3467 3468 if (p->p_vmspace) { 3469 int i,j; 3470 index = 0; 3471 pmap = vmspace_pmap(p->p_vmspace); 3472 for (i = 0; i < NPDEPG; i++) { 3473 pd_entry_t *pde; 3474 pt_entry_t *pte; 3475 vm_offset_t base = i << PDRSHIFT; 3476 3477 pde = &pmap->pm_pdir[i]; 3478 if (pde && pmap_pde_v(pde)) { 3479 for (j = 0; j < NPTEPG; j++) { 3480 vm_offset_t va = base + (j << PAGE_SHIFT); 3481 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3482 if (index) { 3483 index = 0; 3484 printf("\n"); 3485 } 3486 sx_sunlock(&allproc_lock); 3487 return npte; 3488 } 3489 pte = pmap_pte_quick(pmap, va); 3490 if (pte && pmap_pte_v(pte)) { 3491 pt_entry_t pa; 3492 vm_page_t m; 3493 pa = *pte; 3494 m = PHYS_TO_VM_PAGE(pa); 3495 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3496 va, pa, m->hold_count, m->wire_count, m->flags); 3497 npte++; 3498 index++; 3499 if (index >= 2) { 3500 index = 0; 3501 printf("\n"); 3502 } else { 3503 printf(" "); 3504 } 3505 } 3506 } 3507 } 3508 } 3509 } 3510 } 3511 sx_sunlock(&allproc_lock); 3512 return npte; 3513} 3514#endif 3515 3516#if defined(DEBUG) 3517 3518static void pads __P((pmap_t pm)); 3519void pmap_pvdump __P((vm_offset_t pa)); 3520 3521/* print address space of pmap*/ 3522static void 3523pads(pm) 3524 pmap_t pm; 3525{ 3526 int i, j; 3527 vm_offset_t va; 3528 pt_entry_t *ptep; 3529 3530 if (pm == kernel_pmap) 3531 return; 3532 for (i = 0; i < NPDEPG; i++) 3533 if (pm->pm_pdir[i]) 3534 for (j = 0; j < NPTEPG; j++) { 3535 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3536 if (pm == kernel_pmap && va < KERNBASE) 3537 continue; 3538 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3539 continue; 3540 ptep = pmap_pte_quick(pm, va); 3541 if (pmap_pte_v(ptep)) 3542 printf("%x:%x ", va, *ptep); 3543 }; 3544 3545} 3546 3547void 3548pmap_pvdump(pa) 3549 vm_offset_t pa; 3550{ 3551 pv_entry_t pv; 3552 vm_page_t m; 3553 3554 printf("pa %x", pa); 3555 m = PHYS_TO_VM_PAGE(pa); 3556 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3557 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3558 pads(pv->pv_pmap); 3559 } 3560 printf(" "); 3561} 3562#endif 3563