pmap.c revision 1.36
1 /* $NetBSD: pmap.c,v 1.36 2017/09/07 06:29:47 skrll Exp $ */ 2 3 /*- 4 * Copyright (c) 1998, 2001 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility, 9 * NASA Ames Research Center and by Chris G. Demetriou. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 22 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 23 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 24 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 25 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 26 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 27 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 28 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 29 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 30 * POSSIBILITY OF SUCH DAMAGE. 31 */ 32 33 /* 34 * Copyright (c) 1992, 1993 35 * The Regents of the University of California. All rights reserved. 36 * 37 * This code is derived from software contributed to Berkeley by 38 * the Systems Programming Group of the University of Utah Computer 39 * Science Department and Ralph Campbell. 40 * 41 * Redistribution and use in source and binary forms, with or without 42 * modification, are permitted provided that the following conditions 43 * are met: 44 * 1. Redistributions of source code must retain the above copyright 45 * notice, this list of conditions and the following disclaimer. 46 * 2. Redistributions in binary form must reproduce the above copyright 47 * notice, this list of conditions and the following disclaimer in the 48 * documentation and/or other materials provided with the distribution. 49 * 3. Neither the name of the University nor the names of its contributors 50 * may be used to endorse or promote products derived from this software 51 * without specific prior written permission. 52 * 53 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 54 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 55 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 56 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 57 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 58 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 59 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 60 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 61 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 62 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 63 * SUCH DAMAGE. 64 * 65 * @(#)pmap.c 8.4 (Berkeley) 1/26/94 66 */ 67 68 #include <sys/cdefs.h> 69 70 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.36 2017/09/07 06:29:47 skrll Exp $"); 71 72 /* 73 * Manages physical address maps. 74 * 75 * In addition to hardware address maps, this 76 * module is called upon to provide software-use-only 77 * maps which may or may not be stored in the same 78 * form as hardware maps. These pseudo-maps are 79 * used to store intermediate results from copy 80 * operations to and from address spaces. 81 * 82 * Since the information managed by this module is 83 * also stored by the logical address mapping module, 84 * this module may throw away valid virtual-to-physical 85 * mappings at almost any time. However, invalidations 86 * of virtual-to-physical mappings must be done as 87 * requested. 88 * 89 * In order to cope with hardware architectures which 90 * make virtual-to-physical map invalidates expensive, 91 * this module may delay invalidate or reduced protection 92 * operations until such time as they are actually 93 * necessary. This module is given full information as 94 * to which processors are currently using which maps, 95 * and to when physical maps must be made correct. 96 */ 97 98 #include "opt_modular.h" 99 #include "opt_multiprocessor.h" 100 #include "opt_sysv.h" 101 102 #define __PMAP_PRIVATE 103 104 #include <sys/param.h> 105 #include <sys/atomic.h> 106 #include <sys/buf.h> 107 #include <sys/cpu.h> 108 #include <sys/mutex.h> 109 #include <sys/pool.h> 110 #include <sys/atomic.h> 111 #include <sys/mutex.h> 112 #include <sys/atomic.h> 113 114 #include <uvm/uvm.h> 115 #include <uvm/uvm_physseg.h> 116 117 #if defined(MULTIPROCESSOR) && defined(PMAP_VIRTUAL_CACHE_ALIASES) \ 118 && !defined(PMAP_NO_PV_UNCACHED) 119 #error PMAP_VIRTUAL_CACHE_ALIASES with MULTIPROCESSOR requires \ 120 PMAP_NO_PV_UNCACHED to be defined 121 #endif 122 123 PMAP_COUNTER(remove_kernel_calls, "remove kernel calls"); 124 PMAP_COUNTER(remove_kernel_pages, "kernel pages unmapped"); 125 PMAP_COUNTER(remove_user_calls, "remove user calls"); 126 PMAP_COUNTER(remove_user_pages, "user pages unmapped"); 127 PMAP_COUNTER(remove_flushes, "remove cache flushes"); 128 PMAP_COUNTER(remove_tlb_ops, "remove tlb ops"); 129 PMAP_COUNTER(remove_pvfirst, "remove pv first"); 130 PMAP_COUNTER(remove_pvsearch, "remove pv search"); 131 132 PMAP_COUNTER(prefer_requests, "prefer requests"); 133 PMAP_COUNTER(prefer_adjustments, "prefer adjustments"); 134 135 PMAP_COUNTER(idlezeroed_pages, "pages idle zeroed"); 136 137 PMAP_COUNTER(kenter_pa, "kernel fast mapped pages"); 138 PMAP_COUNTER(kenter_pa_bad, "kernel fast mapped pages (bad color)"); 139 PMAP_COUNTER(kenter_pa_unmanaged, "kernel fast mapped unmanaged pages"); 140 PMAP_COUNTER(kremove_pages, "kernel fast unmapped pages"); 141 142 PMAP_COUNTER(page_cache_evictions, "pages changed to uncacheable"); 143 PMAP_COUNTER(page_cache_restorations, "pages changed to cacheable"); 144 145 PMAP_COUNTER(kernel_mappings_bad, "kernel pages mapped (bad color)"); 146 PMAP_COUNTER(user_mappings_bad, "user pages mapped (bad color)"); 147 PMAP_COUNTER(kernel_mappings, "kernel pages mapped"); 148 PMAP_COUNTER(user_mappings, "user pages mapped"); 149 PMAP_COUNTER(user_mappings_changed, "user mapping changed"); 150 PMAP_COUNTER(kernel_mappings_changed, "kernel mapping changed"); 151 PMAP_COUNTER(uncached_mappings, "uncached pages mapped"); 152 PMAP_COUNTER(unmanaged_mappings, "unmanaged pages mapped"); 153 PMAP_COUNTER(managed_mappings, "managed pages mapped"); 154 PMAP_COUNTER(mappings, "pages mapped"); 155 PMAP_COUNTER(remappings, "pages remapped"); 156 PMAP_COUNTER(unmappings, "pages unmapped"); 157 PMAP_COUNTER(primary_mappings, "page initial mappings"); 158 PMAP_COUNTER(primary_unmappings, "page final unmappings"); 159 PMAP_COUNTER(tlb_hit, "page mapping"); 160 161 PMAP_COUNTER(exec_mappings, "exec pages mapped"); 162 PMAP_COUNTER(exec_synced_mappings, "exec pages synced"); 163 PMAP_COUNTER(exec_synced_remove, "exec pages synced (PR)"); 164 PMAP_COUNTER(exec_synced_clear_modify, "exec pages synced (CM)"); 165 PMAP_COUNTER(exec_synced_page_protect, "exec pages synced (PP)"); 166 PMAP_COUNTER(exec_synced_protect, "exec pages synced (P)"); 167 PMAP_COUNTER(exec_uncached_page_protect, "exec pages uncached (PP)"); 168 PMAP_COUNTER(exec_uncached_clear_modify, "exec pages uncached (CM)"); 169 PMAP_COUNTER(exec_uncached_zero_page, "exec pages uncached (ZP)"); 170 PMAP_COUNTER(exec_uncached_copy_page, "exec pages uncached (CP)"); 171 PMAP_COUNTER(exec_uncached_remove, "exec pages uncached (PR)"); 172 173 PMAP_COUNTER(create, "creates"); 174 PMAP_COUNTER(reference, "references"); 175 PMAP_COUNTER(dereference, "dereferences"); 176 PMAP_COUNTER(destroy, "destroyed"); 177 PMAP_COUNTER(activate, "activations"); 178 PMAP_COUNTER(deactivate, "deactivations"); 179 PMAP_COUNTER(update, "updates"); 180 #ifdef MULTIPROCESSOR 181 PMAP_COUNTER(shootdown_ipis, "shootdown IPIs"); 182 #endif 183 PMAP_COUNTER(unwire, "unwires"); 184 PMAP_COUNTER(copy, "copies"); 185 PMAP_COUNTER(clear_modify, "clear_modifies"); 186 PMAP_COUNTER(protect, "protects"); 187 PMAP_COUNTER(page_protect, "page_protects"); 188 189 #define PMAP_ASID_RESERVED 0 190 CTASSERT(PMAP_ASID_RESERVED == 0); 191 192 #ifndef PMAP_SEGTAB_ALIGN 193 #define PMAP_SEGTAB_ALIGN /* nothing */ 194 #endif 195 #ifdef _LP64 196 pmap_segtab_t pmap_kstart_segtab PMAP_SEGTAB_ALIGN; /* first mid-level segtab for kernel */ 197 #endif 198 pmap_segtab_t pmap_kern_segtab PMAP_SEGTAB_ALIGN = { /* top level segtab for kernel */ 199 #ifdef _LP64 200 .seg_seg[(VM_MIN_KERNEL_ADDRESS & XSEGOFSET) >> SEGSHIFT] = &pmap_kstart_segtab, 201 #endif 202 }; 203 204 struct pmap_kernel kernel_pmap_store = { 205 .kernel_pmap = { 206 .pm_count = 1, 207 .pm_segtab = &pmap_kern_segtab, 208 .pm_minaddr = VM_MIN_KERNEL_ADDRESS, 209 .pm_maxaddr = VM_MAX_KERNEL_ADDRESS, 210 }, 211 }; 212 213 struct pmap * const kernel_pmap_ptr = &kernel_pmap_store.kernel_pmap; 214 215 struct pmap_limits pmap_limits = { /* VA and PA limits */ 216 .virtual_start = VM_MIN_KERNEL_ADDRESS, 217 }; 218 219 #ifdef UVMHIST 220 static struct kern_history_ent pmapexechistbuf[10000]; 221 static struct kern_history_ent pmaphistbuf[10000]; 222 UVMHIST_DEFINE(pmapexechist); 223 UVMHIST_DEFINE(pmaphist); 224 #endif 225 226 /* 227 * The pools from which pmap structures and sub-structures are allocated. 228 */ 229 struct pool pmap_pmap_pool; 230 struct pool pmap_pv_pool; 231 232 #ifndef PMAP_PV_LOWAT 233 #define PMAP_PV_LOWAT 16 234 #endif 235 int pmap_pv_lowat = PMAP_PV_LOWAT; 236 237 bool pmap_initialized = false; 238 #define PMAP_PAGE_COLOROK_P(a, b) \ 239 ((((int)(a) ^ (int)(b)) & pmap_page_colormask) == 0) 240 u_int pmap_page_colormask; 241 242 #define PAGE_IS_MANAGED(pa) (pmap_initialized && uvm_pageismanaged(pa)) 243 244 #define PMAP_IS_ACTIVE(pm) \ 245 ((pm) == pmap_kernel() || \ 246 (pm) == curlwp->l_proc->p_vmspace->vm_map.pmap) 247 248 /* Forward function declarations */ 249 void pmap_page_remove(struct vm_page *); 250 static void pmap_pvlist_check(struct vm_page_md *); 251 void pmap_remove_pv(pmap_t, vaddr_t, struct vm_page *, bool); 252 void pmap_enter_pv(pmap_t, vaddr_t, struct vm_page *, pt_entry_t *, u_int); 253 254 /* 255 * PV table management functions. 256 */ 257 void *pmap_pv_page_alloc(struct pool *, int); 258 void pmap_pv_page_free(struct pool *, void *); 259 260 struct pool_allocator pmap_pv_page_allocator = { 261 pmap_pv_page_alloc, pmap_pv_page_free, 0, 262 }; 263 264 #define pmap_pv_alloc() pool_get(&pmap_pv_pool, PR_NOWAIT) 265 #define pmap_pv_free(pv) pool_put(&pmap_pv_pool, (pv)) 266 267 #if !defined(MULTIPROCESSOR) || !defined(PMAP_MD_NEED_TLB_MISS_LOCK) 268 #define pmap_md_tlb_miss_lock_enter() do { } while(/*CONSTCOND*/0) 269 #define pmap_md_tlb_miss_lock_exit() do { } while(/*CONSTCOND*/0) 270 #endif /* !MULTIPROCESSOR || !PMAP_MD_NEED_TLB_MISS_LOCK */ 271 272 #ifndef MULTIPROCESSOR 273 kmutex_t pmap_pvlist_mutex __cacheline_aligned; 274 #endif 275 276 /* 277 * Debug functions. 278 */ 279 280 #ifdef DEBUG 281 static inline void 282 pmap_asid_check(pmap_t pm, const char *func) 283 { 284 if (!PMAP_IS_ACTIVE(pm)) 285 return; 286 287 struct pmap_asid_info * const pai = PMAP_PAI(pm, cpu_tlb_info(curcpu())); 288 tlb_asid_t asid = tlb_get_asid(); 289 if (asid != pai->pai_asid) 290 panic("%s: inconsistency for active TLB update: %u <-> %u", 291 func, asid, pai->pai_asid); 292 } 293 #endif 294 295 static void 296 pmap_addr_range_check(pmap_t pmap, vaddr_t sva, vaddr_t eva, const char *func) 297 { 298 #ifdef DEBUG 299 if (pmap == pmap_kernel()) { 300 if (sva < VM_MIN_KERNEL_ADDRESS) 301 panic("%s: kva %#"PRIxVADDR" not in range", 302 func, sva); 303 if (eva >= pmap_limits.virtual_end) 304 panic("%s: kva %#"PRIxVADDR" not in range", 305 func, eva); 306 } else { 307 if (eva > VM_MAXUSER_ADDRESS) 308 panic("%s: uva %#"PRIxVADDR" not in range", 309 func, eva); 310 pmap_asid_check(pmap, func); 311 } 312 #endif 313 } 314 315 /* 316 * Misc. functions. 317 */ 318 319 bool 320 pmap_page_clear_attributes(struct vm_page_md *mdpg, u_int clear_attributes) 321 { 322 volatile unsigned long * const attrp = &mdpg->mdpg_attrs; 323 #ifdef MULTIPROCESSOR 324 for (;;) { 325 u_int old_attr = *attrp; 326 if ((old_attr & clear_attributes) == 0) 327 return false; 328 u_int new_attr = old_attr & ~clear_attributes; 329 if (old_attr == atomic_cas_ulong(attrp, old_attr, new_attr)) 330 return true; 331 } 332 #else 333 unsigned long old_attr = *attrp; 334 if ((old_attr & clear_attributes) == 0) 335 return false; 336 *attrp &= ~clear_attributes; 337 return true; 338 #endif 339 } 340 341 void 342 pmap_page_set_attributes(struct vm_page_md *mdpg, u_int set_attributes) 343 { 344 #ifdef MULTIPROCESSOR 345 atomic_or_ulong(&mdpg->mdpg_attrs, set_attributes); 346 #else 347 mdpg->mdpg_attrs |= set_attributes; 348 #endif 349 } 350 351 static void 352 pmap_page_syncicache(struct vm_page *pg) 353 { 354 #ifndef MULTIPROCESSOR 355 struct pmap * const curpmap = curlwp->l_proc->p_vmspace->vm_map.pmap; 356 #endif 357 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 358 pv_entry_t pv = &mdpg->mdpg_first; 359 kcpuset_t *onproc; 360 #ifdef MULTIPROCESSOR 361 kcpuset_create(&onproc, true); 362 KASSERT(onproc != NULL); 363 #else 364 onproc = NULL; 365 #endif 366 VM_PAGEMD_PVLIST_READLOCK(mdpg); 367 pmap_pvlist_check(mdpg); 368 369 if (pv->pv_pmap != NULL) { 370 for (; pv != NULL; pv = pv->pv_next) { 371 #ifdef MULTIPROCESSOR 372 kcpuset_merge(onproc, pv->pv_pmap->pm_onproc); 373 if (kcpuset_match(onproc, kcpuset_running)) { 374 break; 375 } 376 #else 377 if (pv->pv_pmap == curpmap) { 378 onproc = curcpu()->ci_data.cpu_kcpuset; 379 break; 380 } 381 #endif 382 } 383 } 384 pmap_pvlist_check(mdpg); 385 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 386 kpreempt_disable(); 387 pmap_md_page_syncicache(pg, onproc); 388 kpreempt_enable(); 389 #ifdef MULTIPROCESSOR 390 kcpuset_destroy(onproc); 391 #endif 392 } 393 394 /* 395 * Define the initial bounds of the kernel virtual address space. 396 */ 397 void 398 pmap_virtual_space(vaddr_t *vstartp, vaddr_t *vendp) 399 { 400 401 *vstartp = pmap_limits.virtual_start; 402 *vendp = pmap_limits.virtual_end; 403 } 404 405 vaddr_t 406 pmap_growkernel(vaddr_t maxkvaddr) 407 { 408 vaddr_t virtual_end = pmap_limits.virtual_end; 409 maxkvaddr = pmap_round_seg(maxkvaddr) - 1; 410 411 /* 412 * Reserve PTEs for the new KVA space. 413 */ 414 for (; virtual_end < maxkvaddr; virtual_end += NBSEG) { 415 pmap_pte_reserve(pmap_kernel(), virtual_end, 0); 416 } 417 418 /* 419 * Don't exceed VM_MAX_KERNEL_ADDRESS! 420 */ 421 if (virtual_end == 0 || virtual_end > VM_MAX_KERNEL_ADDRESS) 422 virtual_end = VM_MAX_KERNEL_ADDRESS; 423 424 /* 425 * Update new end. 426 */ 427 pmap_limits.virtual_end = virtual_end; 428 return virtual_end; 429 } 430 431 /* 432 * Bootstrap memory allocator (alternative to vm_bootstrap_steal_memory()). 433 * This function allows for early dynamic memory allocation until the virtual 434 * memory system has been bootstrapped. After that point, either kmem_alloc 435 * or malloc should be used. This function works by stealing pages from the 436 * (to be) managed page pool, then implicitly mapping the pages (by using 437 * their k0seg addresses) and zeroing them. 438 * 439 * It may be used once the physical memory segments have been pre-loaded 440 * into the vm_physmem[] array. Early memory allocation MUST use this 441 * interface! This cannot be used after vm_page_startup(), and will 442 * generate a panic if tried. 443 * 444 * Note that this memory will never be freed, and in essence it is wired 445 * down. 446 * 447 * We must adjust *vstartp and/or *vendp iff we use address space 448 * from the kernel virtual address range defined by pmap_virtual_space(). 449 */ 450 vaddr_t 451 pmap_steal_memory(vsize_t size, vaddr_t *vstartp, vaddr_t *vendp) 452 { 453 size_t npgs; 454 paddr_t pa; 455 vaddr_t va; 456 457 uvm_physseg_t maybe_bank = UVM_PHYSSEG_TYPE_INVALID; 458 459 size = round_page(size); 460 npgs = atop(size); 461 462 aprint_debug("%s: need %zu pages\n", __func__, npgs); 463 464 for (uvm_physseg_t bank = uvm_physseg_get_first(); 465 uvm_physseg_valid_p(bank); 466 bank = uvm_physseg_get_next(bank)) { 467 468 if (uvm.page_init_done == true) 469 panic("pmap_steal_memory: called _after_ bootstrap"); 470 471 aprint_debug("%s: seg %"PRIxPHYSSEG": %#"PRIxPADDR" %#"PRIxPADDR" %#"PRIxPADDR" %#"PRIxPADDR"\n", 472 __func__, bank, 473 uvm_physseg_get_avail_start(bank), uvm_physseg_get_start(bank), 474 uvm_physseg_get_avail_end(bank), uvm_physseg_get_end(bank)); 475 476 if (uvm_physseg_get_avail_start(bank) != uvm_physseg_get_start(bank) 477 || uvm_physseg_get_avail_start(bank) >= uvm_physseg_get_avail_end(bank)) { 478 aprint_debug("%s: seg %"PRIxPHYSSEG": bad start\n", __func__, bank); 479 continue; 480 } 481 482 if (uvm_physseg_get_avail_end(bank) - uvm_physseg_get_avail_start(bank) < npgs) { 483 aprint_debug("%s: seg %"PRIxPHYSSEG": too small for %zu pages\n", 484 __func__, bank, npgs); 485 continue; 486 } 487 488 if (!pmap_md_ok_to_steal_p(bank, npgs)) { 489 continue; 490 } 491 492 /* 493 * Always try to allocate from the segment with the least 494 * amount of space left. 495 */ 496 #define VM_PHYSMEM_SPACE(b) ((uvm_physseg_get_avail_end(b)) - (uvm_physseg_get_avail_start(b))) 497 if (uvm_physseg_valid_p(maybe_bank) == false 498 || VM_PHYSMEM_SPACE(bank) < VM_PHYSMEM_SPACE(maybe_bank)) { 499 maybe_bank = bank; 500 } 501 } 502 503 if (uvm_physseg_valid_p(maybe_bank)) { 504 const uvm_physseg_t bank = maybe_bank; 505 506 /* 507 * There are enough pages here; steal them! 508 */ 509 pa = ptoa(uvm_physseg_get_start(bank)); 510 uvm_physseg_unplug(atop(pa), npgs); 511 512 aprint_debug("%s: seg %"PRIxPHYSSEG": %zu pages stolen (%#"PRIxPADDR" left)\n", 513 __func__, bank, npgs, VM_PHYSMEM_SPACE(bank)); 514 515 va = pmap_md_map_poolpage(pa, size); 516 memset((void *)va, 0, size); 517 return va; 518 } 519 520 /* 521 * If we got here, there was no memory left. 522 */ 523 panic("pmap_steal_memory: no memory to steal %zu pages", npgs); 524 } 525 526 /* 527 * Initialize the pmap module. 528 * Called by vm_init, to initialize any structures that the pmap 529 * system needs to map virtual memory. 530 */ 531 void 532 pmap_init(void) 533 { 534 UVMHIST_INIT_STATIC(pmapexechist, pmapexechistbuf); 535 UVMHIST_INIT_STATIC(pmaphist, pmaphistbuf); 536 537 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 538 539 /* 540 * Initialize the segtab lock. 541 */ 542 mutex_init(&pmap_segtab_lock, MUTEX_DEFAULT, IPL_HIGH); 543 544 /* 545 * Set a low water mark on the pv_entry pool, so that we are 546 * more likely to have these around even in extreme memory 547 * starvation. 548 */ 549 pool_setlowat(&pmap_pv_pool, pmap_pv_lowat); 550 551 /* 552 * Set the page colormask but allow pmap_md_init to override it. 553 */ 554 pmap_page_colormask = ptoa(uvmexp.colormask); 555 556 pmap_md_init(); 557 558 /* 559 * Now it is safe to enable pv entry recording. 560 */ 561 pmap_initialized = true; 562 } 563 564 /* 565 * Create and return a physical map. 566 * 567 * If the size specified for the map 568 * is zero, the map is an actual physical 569 * map, and may be referenced by the 570 * hardware. 571 * 572 * If the size specified is non-zero, 573 * the map will be used in software only, and 574 * is bounded by that size. 575 */ 576 pmap_t 577 pmap_create(void) 578 { 579 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 580 PMAP_COUNT(create); 581 582 pmap_t pmap = pool_get(&pmap_pmap_pool, PR_WAITOK); 583 memset(pmap, 0, PMAP_SIZE); 584 585 KASSERT(pmap->pm_pai[0].pai_link.le_prev == NULL); 586 587 pmap->pm_count = 1; 588 pmap->pm_minaddr = VM_MIN_ADDRESS; 589 pmap->pm_maxaddr = VM_MAXUSER_ADDRESS; 590 591 pmap_segtab_init(pmap); 592 593 #ifdef MULTIPROCESSOR 594 kcpuset_create(&pmap->pm_active, true); 595 kcpuset_create(&pmap->pm_onproc, true); 596 KASSERT(pmap->pm_active != NULL); 597 KASSERT(pmap->pm_onproc != NULL); 598 #endif 599 600 UVMHIST_LOG(pmaphist, " <-- done (pmap=%p)", pmap, 0, 0, 0); 601 602 return pmap; 603 } 604 605 /* 606 * Retire the given physical map from service. 607 * Should only be called if the map contains 608 * no valid mappings. 609 */ 610 void 611 pmap_destroy(pmap_t pmap) 612 { 613 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 614 UVMHIST_LOG(pmaphist, "(pmap=%p)", pmap, 0, 0, 0); 615 616 if (atomic_dec_uint_nv(&pmap->pm_count) > 0) { 617 PMAP_COUNT(dereference); 618 UVMHIST_LOG(pmaphist, " <-- done (deref)", 0, 0, 0, 0); 619 return; 620 } 621 622 PMAP_COUNT(destroy); 623 KASSERT(pmap->pm_count == 0); 624 kpreempt_disable(); 625 pmap_md_tlb_miss_lock_enter(); 626 pmap_tlb_asid_release_all(pmap); 627 pmap_segtab_destroy(pmap, NULL, 0); 628 pmap_md_tlb_miss_lock_exit(); 629 630 #ifdef MULTIPROCESSOR 631 kcpuset_destroy(pmap->pm_active); 632 kcpuset_destroy(pmap->pm_onproc); 633 pmap->pm_active = NULL; 634 pmap->pm_onproc = NULL; 635 #endif 636 637 pool_put(&pmap_pmap_pool, pmap); 638 kpreempt_enable(); 639 640 UVMHIST_LOG(pmaphist, " <-- done (freed)", 0, 0, 0, 0); 641 } 642 643 /* 644 * Add a reference to the specified pmap. 645 */ 646 void 647 pmap_reference(pmap_t pmap) 648 { 649 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 650 UVMHIST_LOG(pmaphist, "(pmap=%p)", pmap, 0, 0, 0); 651 PMAP_COUNT(reference); 652 653 if (pmap != NULL) { 654 atomic_inc_uint(&pmap->pm_count); 655 } 656 657 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 658 } 659 660 /* 661 * Make a new pmap (vmspace) active for the given process. 662 */ 663 void 664 pmap_activate(struct lwp *l) 665 { 666 pmap_t pmap = l->l_proc->p_vmspace->vm_map.pmap; 667 668 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 669 UVMHIST_LOG(pmaphist, "(l=%p pmap=%p)", l, pmap, 0, 0); 670 PMAP_COUNT(activate); 671 672 kpreempt_disable(); 673 pmap_md_tlb_miss_lock_enter(); 674 pmap_tlb_asid_acquire(pmap, l); 675 if (l == curlwp) { 676 pmap_segtab_activate(pmap, l); 677 } 678 pmap_md_tlb_miss_lock_exit(); 679 kpreempt_enable(); 680 681 UVMHIST_LOG(pmaphist, " <-- done (%u:%u)", l->l_proc->p_pid, l->l_lid, 682 0, 0); 683 } 684 685 /* 686 * Remove this page from all physical maps in which it resides. 687 * Reflects back modify bits to the pager. 688 */ 689 void 690 pmap_page_remove(struct vm_page *pg) 691 { 692 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 693 694 kpreempt_disable(); 695 VM_PAGEMD_PVLIST_LOCK(mdpg); 696 pmap_pvlist_check(mdpg); 697 698 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 699 700 UVMHIST_LOG(pmapexechist, "pg %p (pa %#"PRIxPADDR") [page removed]: " 701 "execpage cleared", pg, VM_PAGE_TO_PHYS(pg), 702 0, 0); 703 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 704 pmap_page_clear_attributes(mdpg, VM_PAGEMD_EXECPAGE|VM_PAGEMD_UNCACHED); 705 #else 706 pmap_page_clear_attributes(mdpg, VM_PAGEMD_EXECPAGE); 707 #endif 708 PMAP_COUNT(exec_uncached_remove); 709 710 pv_entry_t pv = &mdpg->mdpg_first; 711 if (pv->pv_pmap == NULL) { 712 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 713 kpreempt_enable(); 714 UVMHIST_LOG(pmaphist, " <-- done (empty)", 0, 0, 0, 0); 715 return; 716 } 717 718 pv_entry_t npv; 719 pv_entry_t pvp = NULL; 720 721 for (; pv != NULL; pv = npv) { 722 npv = pv->pv_next; 723 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 724 if (pv->pv_va & PV_KENTER) { 725 UVMHIST_LOG(pmaphist, " pv %p pmap %p va %" 726 PRIxVADDR" skip", pv, pv->pv_pmap, pv->pv_va, 0); 727 728 KASSERT(pv->pv_pmap == pmap_kernel()); 729 730 /* Assume no more - it'll get fixed if there are */ 731 pv->pv_next = NULL; 732 733 /* 734 * pvp is non-null when we already have a PV_KENTER 735 * pv in pvh_first; otherwise we haven't seen a 736 * PV_KENTER pv and we need to copy this one to 737 * pvh_first 738 */ 739 if (pvp) { 740 /* 741 * The previous PV_KENTER pv needs to point to 742 * this PV_KENTER pv 743 */ 744 pvp->pv_next = pv; 745 } else { 746 pv_entry_t fpv = &mdpg->mdpg_first; 747 *fpv = *pv; 748 KASSERT(fpv->pv_pmap == pmap_kernel()); 749 } 750 pvp = pv; 751 continue; 752 } 753 #endif 754 const pmap_t pmap = pv->pv_pmap; 755 vaddr_t va = trunc_page(pv->pv_va); 756 pt_entry_t * const ptep = pmap_pte_lookup(pmap, va); 757 KASSERTMSG(ptep != NULL, "%#"PRIxVADDR " %#"PRIxVADDR, va, 758 pmap_limits.virtual_end); 759 pt_entry_t pte = *ptep; 760 UVMHIST_LOG(pmaphist, " pv %p pmap %p va %"PRIxVADDR 761 " pte %#"PRIxPTE, pv, pmap, va, pte_value(pte)); 762 if (!pte_valid_p(pte)) 763 continue; 764 const bool is_kernel_pmap_p = (pmap == pmap_kernel()); 765 if (is_kernel_pmap_p) { 766 PMAP_COUNT(remove_kernel_pages); 767 } else { 768 PMAP_COUNT(remove_user_pages); 769 } 770 if (pte_wired_p(pte)) 771 pmap->pm_stats.wired_count--; 772 pmap->pm_stats.resident_count--; 773 774 pmap_md_tlb_miss_lock_enter(); 775 const pt_entry_t npte = pte_nv_entry(is_kernel_pmap_p); 776 pte_set(ptep, npte); 777 if (__predict_true(!(pmap->pm_flags & PMAP_DEFERRED_ACTIVATE))) { 778 /* 779 * Flush the TLB for the given address. 780 */ 781 pmap_tlb_invalidate_addr(pmap, va); 782 } 783 pmap_md_tlb_miss_lock_exit(); 784 785 /* 786 * non-null means this is a non-pvh_first pv, so we should 787 * free it. 788 */ 789 if (pvp) { 790 KASSERT(pvp->pv_pmap == pmap_kernel()); 791 KASSERT(pvp->pv_next == NULL); 792 pmap_pv_free(pv); 793 } else { 794 pv->pv_pmap = NULL; 795 pv->pv_next = NULL; 796 } 797 } 798 799 pmap_pvlist_check(mdpg); 800 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 801 kpreempt_enable(); 802 803 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 804 } 805 806 807 /* 808 * Make a previously active pmap (vmspace) inactive. 809 */ 810 void 811 pmap_deactivate(struct lwp *l) 812 { 813 pmap_t pmap = l->l_proc->p_vmspace->vm_map.pmap; 814 815 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 816 UVMHIST_LOG(pmaphist, "(l=%p pmap=%p)", l, pmap, 0, 0); 817 PMAP_COUNT(deactivate); 818 819 kpreempt_disable(); 820 KASSERT(l == curlwp || l->l_cpu == curlwp->l_cpu); 821 pmap_md_tlb_miss_lock_enter(); 822 curcpu()->ci_pmap_user_segtab = PMAP_INVALID_SEGTAB_ADDRESS; 823 #ifdef _LP64 824 curcpu()->ci_pmap_user_seg0tab = NULL; 825 #endif 826 pmap_tlb_asid_deactivate(pmap); 827 pmap_md_tlb_miss_lock_exit(); 828 kpreempt_enable(); 829 830 UVMHIST_LOG(pmaphist, " <-- done (%u:%u)", l->l_proc->p_pid, l->l_lid, 831 0, 0); 832 } 833 834 void 835 pmap_update(struct pmap *pmap) 836 { 837 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 838 UVMHIST_LOG(pmaphist, "(pmap=%p)", pmap, 0, 0, 0); 839 PMAP_COUNT(update); 840 841 kpreempt_disable(); 842 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN) 843 u_int pending = atomic_swap_uint(&pmap->pm_shootdown_pending, 0); 844 if (pending && pmap_tlb_shootdown_bystanders(pmap)) 845 PMAP_COUNT(shootdown_ipis); 846 #endif 847 pmap_md_tlb_miss_lock_enter(); 848 #if defined(DEBUG) && !defined(MULTIPROCESSOR) 849 pmap_tlb_check(pmap, pmap_md_tlb_check_entry); 850 #endif /* DEBUG */ 851 852 /* 853 * If pmap_remove_all was called, we deactivated ourselves and nuked 854 * our ASID. Now we have to reactivate ourselves. 855 */ 856 if (__predict_false(pmap->pm_flags & PMAP_DEFERRED_ACTIVATE)) { 857 pmap->pm_flags ^= PMAP_DEFERRED_ACTIVATE; 858 pmap_tlb_asid_acquire(pmap, curlwp); 859 pmap_segtab_activate(pmap, curlwp); 860 } 861 pmap_md_tlb_miss_lock_exit(); 862 kpreempt_enable(); 863 864 UVMHIST_LOG(pmaphist, " <-- done (%c)", 865 (pmap == pmap_kernel() ? 'k' : 'u'), 0, 0, 0); 866 } 867 868 /* 869 * Remove the given range of addresses from the specified map. 870 * 871 * It is assumed that the start and end are properly 872 * rounded to the page size. 873 */ 874 875 static bool 876 pmap_pte_remove(pmap_t pmap, vaddr_t sva, vaddr_t eva, pt_entry_t *ptep, 877 uintptr_t flags) 878 { 879 const pt_entry_t npte = flags; 880 const bool is_kernel_pmap_p = (pmap == pmap_kernel()); 881 882 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 883 UVMHIST_LOG(pmaphist, "(pmap=%p %cva=%#"PRIxVADDR"..%#"PRIxVADDR, 884 pmap, (is_kernel_pmap_p ? 'k' : 'u'), sva, eva); 885 UVMHIST_LOG(pmaphist, "ptep=%p, flags(npte)=%#"PRIxPTR")", 886 ptep, flags, 0, 0); 887 888 KASSERT(kpreempt_disabled()); 889 890 for (; sva < eva; sva += NBPG, ptep++) { 891 const pt_entry_t pte = *ptep; 892 if (!pte_valid_p(pte)) 893 continue; 894 if (is_kernel_pmap_p) { 895 PMAP_COUNT(remove_kernel_pages); 896 } else { 897 PMAP_COUNT(remove_user_pages); 898 } 899 if (pte_wired_p(pte)) 900 pmap->pm_stats.wired_count--; 901 pmap->pm_stats.resident_count--; 902 struct vm_page * const pg = PHYS_TO_VM_PAGE(pte_to_paddr(pte)); 903 if (__predict_true(pg != NULL)) { 904 pmap_remove_pv(pmap, sva, pg, pte_modified_p(pte)); 905 } 906 pmap_md_tlb_miss_lock_enter(); 907 pte_set(ptep, npte); 908 if (__predict_true(!(pmap->pm_flags & PMAP_DEFERRED_ACTIVATE))) { 909 910 /* 911 * Flush the TLB for the given address. 912 */ 913 pmap_tlb_invalidate_addr(pmap, sva); 914 } 915 pmap_md_tlb_miss_lock_exit(); 916 } 917 918 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 919 920 return false; 921 } 922 923 void 924 pmap_remove(pmap_t pmap, vaddr_t sva, vaddr_t eva) 925 { 926 const bool is_kernel_pmap_p = (pmap == pmap_kernel()); 927 const pt_entry_t npte = pte_nv_entry(is_kernel_pmap_p); 928 929 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 930 UVMHIST_LOG(pmaphist, "(pmap=%p, va=%#"PRIxVADDR"..%#"PRIxVADDR")", 931 pmap, sva, eva, 0); 932 933 if (is_kernel_pmap_p) { 934 PMAP_COUNT(remove_kernel_calls); 935 } else { 936 PMAP_COUNT(remove_user_calls); 937 } 938 #ifdef PMAP_FAULTINFO 939 curpcb->pcb_faultinfo.pfi_faultaddr = 0; 940 curpcb->pcb_faultinfo.pfi_repeats = 0; 941 curpcb->pcb_faultinfo.pfi_faultpte = NULL; 942 #endif 943 kpreempt_disable(); 944 pmap_addr_range_check(pmap, sva, eva, __func__); 945 pmap_pte_process(pmap, sva, eva, pmap_pte_remove, npte); 946 kpreempt_enable(); 947 948 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 949 } 950 951 /* 952 * pmap_page_protect: 953 * 954 * Lower the permission for all mappings to a given page. 955 */ 956 void 957 pmap_page_protect(struct vm_page *pg, vm_prot_t prot) 958 { 959 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 960 pv_entry_t pv; 961 vaddr_t va; 962 963 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 964 UVMHIST_LOG(pmaphist, "(pg=%p (pa %#"PRIxPADDR") prot=%#x)", 965 pg, VM_PAGE_TO_PHYS(pg), prot, 0); 966 PMAP_COUNT(page_protect); 967 968 switch (prot) { 969 case VM_PROT_READ|VM_PROT_WRITE: 970 case VM_PROT_ALL: 971 break; 972 973 /* copy_on_write */ 974 case VM_PROT_READ: 975 case VM_PROT_READ|VM_PROT_EXECUTE: 976 pv = &mdpg->mdpg_first; 977 kpreempt_disable(); 978 VM_PAGEMD_PVLIST_READLOCK(mdpg); 979 pmap_pvlist_check(mdpg); 980 /* 981 * Loop over all current mappings setting/clearing as 982 * appropriate. 983 */ 984 if (pv->pv_pmap != NULL) { 985 while (pv != NULL) { 986 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 987 if (pv->pv_va & PV_KENTER) { 988 pv = pv->pv_next; 989 continue; 990 } 991 #endif 992 const pmap_t pmap = pv->pv_pmap; 993 va = trunc_page(pv->pv_va); 994 const uintptr_t gen = 995 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 996 pmap_protect(pmap, va, va + PAGE_SIZE, prot); 997 KASSERT(pv->pv_pmap == pmap); 998 pmap_update(pmap); 999 if (gen != VM_PAGEMD_PVLIST_READLOCK(mdpg)) { 1000 pv = &mdpg->mdpg_first; 1001 } else { 1002 pv = pv->pv_next; 1003 } 1004 pmap_pvlist_check(mdpg); 1005 } 1006 } 1007 pmap_pvlist_check(mdpg); 1008 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1009 kpreempt_enable(); 1010 break; 1011 1012 /* remove_all */ 1013 default: 1014 pmap_page_remove(pg); 1015 } 1016 1017 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1018 } 1019 1020 static bool 1021 pmap_pte_protect(pmap_t pmap, vaddr_t sva, vaddr_t eva, pt_entry_t *ptep, 1022 uintptr_t flags) 1023 { 1024 const vm_prot_t prot = (flags & VM_PROT_ALL); 1025 1026 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1027 UVMHIST_LOG(pmaphist, "(pmap=%p %cva=%#"PRIxVADDR"..%#"PRIxVADDR")", 1028 pmap, (pmap == pmap_kernel() ? 'k' : 'u'), sva, eva); 1029 UVMHIST_LOG(pmaphist, "ptep=%p, flags(npte)=%#"PRIxPTR")", 1030 ptep, flags, 0, 0); 1031 1032 KASSERT(kpreempt_disabled()); 1033 /* 1034 * Change protection on every valid mapping within this segment. 1035 */ 1036 for (; sva < eva; sva += NBPG, ptep++) { 1037 pt_entry_t pte = *ptep; 1038 if (!pte_valid_p(pte)) 1039 continue; 1040 struct vm_page * const pg = PHYS_TO_VM_PAGE(pte_to_paddr(pte)); 1041 if (pg != NULL && pte_modified_p(pte)) { 1042 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1043 if (VM_PAGEMD_EXECPAGE_P(mdpg)) { 1044 KASSERT(mdpg->mdpg_first.pv_pmap != NULL); 1045 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1046 if (VM_PAGEMD_CACHED_P(mdpg)) { 1047 #endif 1048 UVMHIST_LOG(pmapexechist, 1049 "pg %p (pa %#"PRIxPADDR"): " 1050 "syncicached performed", 1051 pg, VM_PAGE_TO_PHYS(pg), 0, 0); 1052 pmap_page_syncicache(pg); 1053 PMAP_COUNT(exec_synced_protect); 1054 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1055 } 1056 #endif 1057 } 1058 } 1059 pte = pte_prot_downgrade(pte, prot); 1060 if (*ptep != pte) { 1061 pmap_md_tlb_miss_lock_enter(); 1062 pte_set(ptep, pte); 1063 /* 1064 * Update the TLB if needed. 1065 */ 1066 pmap_tlb_update_addr(pmap, sva, pte, PMAP_TLB_NEED_IPI); 1067 pmap_md_tlb_miss_lock_exit(); 1068 } 1069 } 1070 1071 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1072 1073 return false; 1074 } 1075 1076 /* 1077 * Set the physical protection on the 1078 * specified range of this map as requested. 1079 */ 1080 void 1081 pmap_protect(pmap_t pmap, vaddr_t sva, vaddr_t eva, vm_prot_t prot) 1082 { 1083 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1084 UVMHIST_LOG(pmaphist, 1085 "(pmap=%p, va=%#"PRIxVADDR"..%#"PRIxVADDR", prot=%u)", 1086 pmap, sva, eva, prot); 1087 PMAP_COUNT(protect); 1088 1089 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1090 pmap_remove(pmap, sva, eva); 1091 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1092 return; 1093 } 1094 1095 /* 1096 * Change protection on every valid mapping within this segment. 1097 */ 1098 kpreempt_disable(); 1099 pmap_addr_range_check(pmap, sva, eva, __func__); 1100 pmap_pte_process(pmap, sva, eva, pmap_pte_protect, prot); 1101 kpreempt_enable(); 1102 1103 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1104 } 1105 1106 #if defined(PMAP_VIRTUAL_CACHE_ALIASES) && !defined(PMAP_NO_PV_UNCACHED) 1107 /* 1108 * pmap_page_cache: 1109 * 1110 * Change all mappings of a managed page to cached/uncached. 1111 */ 1112 void 1113 pmap_page_cache(struct vm_page *pg, bool cached) 1114 { 1115 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1116 1117 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1118 UVMHIST_LOG(pmaphist, "(pg=%p (pa %#"PRIxPADDR") cached=%d)", 1119 pg, VM_PAGE_TO_PHYS(pg), cached, 0); 1120 1121 KASSERT(kpreempt_disabled()); 1122 KASSERT(VM_PAGEMD_PVLIST_LOCKED_P(mdpg)); 1123 1124 if (cached) { 1125 pmap_page_clear_attributes(mdpg, VM_PAGEMD_UNCACHED); 1126 PMAP_COUNT(page_cache_restorations); 1127 } else { 1128 pmap_page_set_attributes(mdpg, VM_PAGEMD_UNCACHED); 1129 PMAP_COUNT(page_cache_evictions); 1130 } 1131 1132 for (pv_entry_t pv = &mdpg->mdpg_first; pv != NULL; pv = pv->pv_next) { 1133 pmap_t pmap = pv->pv_pmap; 1134 vaddr_t va = trunc_page(pv->pv_va); 1135 1136 KASSERT(pmap != NULL); 1137 KASSERT(pmap != pmap_kernel() || !pmap_md_direct_mapped_vaddr_p(va)); 1138 pt_entry_t * const ptep = pmap_pte_lookup(pmap, va); 1139 if (ptep == NULL) 1140 continue; 1141 pt_entry_t pte = *ptep; 1142 if (pte_valid_p(pte)) { 1143 pte = pte_cached_change(pte, cached); 1144 pmap_md_tlb_miss_lock_enter(); 1145 pte_set(ptep, pte); 1146 pmap_tlb_update_addr(pmap, va, pte, PMAP_TLB_NEED_IPI); 1147 pmap_md_tlb_miss_lock_exit(); 1148 } 1149 } 1150 1151 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1152 } 1153 #endif /* PMAP_VIRTUAL_CACHE_ALIASES && !PMAP_NO_PV_UNCACHED */ 1154 1155 /* 1156 * Insert the given physical page (p) at 1157 * the specified virtual address (v) in the 1158 * target physical map with the protection requested. 1159 * 1160 * If specified, the page will be wired down, meaning 1161 * that the related pte can not be reclaimed. 1162 * 1163 * NB: This is the only routine which MAY NOT lazy-evaluate 1164 * or lose information. That is, this routine must actually 1165 * insert this page into the given map NOW. 1166 */ 1167 int 1168 pmap_enter(pmap_t pmap, vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags) 1169 { 1170 const bool wired = (flags & PMAP_WIRED) != 0; 1171 const bool is_kernel_pmap_p = (pmap == pmap_kernel()); 1172 u_int update_flags = (flags & VM_PROT_ALL) != 0 ? PMAP_TLB_INSERT : 0; 1173 #ifdef UVMHIST 1174 struct kern_history * const histp = 1175 ((prot & VM_PROT_EXECUTE) ? &pmapexechist : &pmaphist); 1176 #endif 1177 1178 UVMHIST_FUNC(__func__); UVMHIST_CALLED(*histp); 1179 UVMHIST_LOG(*histp, "(pmap=%p, va=%#"PRIxVADDR", pa=%#"PRIxPADDR, 1180 pmap, va, pa, 0); 1181 UVMHIST_LOG(*histp, "prot=%#x flags=%#x)", prot, flags, 0, 0); 1182 1183 const bool good_color = PMAP_PAGE_COLOROK_P(pa, va); 1184 if (is_kernel_pmap_p) { 1185 PMAP_COUNT(kernel_mappings); 1186 if (!good_color) 1187 PMAP_COUNT(kernel_mappings_bad); 1188 } else { 1189 PMAP_COUNT(user_mappings); 1190 if (!good_color) 1191 PMAP_COUNT(user_mappings_bad); 1192 } 1193 pmap_addr_range_check(pmap, va, va, __func__); 1194 1195 KASSERTMSG(prot & VM_PROT_READ, "no READ (%#x) in prot %#x", 1196 VM_PROT_READ, prot); 1197 1198 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 1199 struct vm_page_md * const mdpg = (pg ? VM_PAGE_TO_MD(pg) : NULL); 1200 1201 if (pg) { 1202 /* Set page referenced/modified status based on flags */ 1203 if (flags & VM_PROT_WRITE) { 1204 pmap_page_set_attributes(mdpg, VM_PAGEMD_MODIFIED|VM_PAGEMD_REFERENCED); 1205 } else if (flags & VM_PROT_ALL) { 1206 pmap_page_set_attributes(mdpg, VM_PAGEMD_REFERENCED); 1207 } 1208 1209 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1210 if (!VM_PAGEMD_CACHED_P(mdpg)) { 1211 flags |= PMAP_NOCACHE; 1212 PMAP_COUNT(uncached_mappings); 1213 } 1214 #endif 1215 1216 PMAP_COUNT(managed_mappings); 1217 } else { 1218 /* 1219 * Assumption: if it is not part of our managed memory 1220 * then it must be device memory which may be volatile. 1221 */ 1222 if ((flags & PMAP_CACHE_MASK) == 0) 1223 flags |= PMAP_NOCACHE; 1224 PMAP_COUNT(unmanaged_mappings); 1225 } 1226 1227 pt_entry_t npte = pte_make_enter(pa, mdpg, prot, flags, 1228 is_kernel_pmap_p); 1229 1230 kpreempt_disable(); 1231 1232 pt_entry_t * const ptep = pmap_pte_reserve(pmap, va, flags); 1233 if (__predict_false(ptep == NULL)) { 1234 kpreempt_enable(); 1235 UVMHIST_LOG(*histp, " <-- ENOMEM", 0, 0, 0, 0); 1236 return ENOMEM; 1237 } 1238 const pt_entry_t opte = *ptep; 1239 const bool resident = pte_valid_p(opte); 1240 bool remap = false; 1241 if (resident) { 1242 if (pte_to_paddr(opte) != pa) { 1243 KASSERT(!is_kernel_pmap_p); 1244 const pt_entry_t rpte = pte_nv_entry(false); 1245 1246 pmap_addr_range_check(pmap, va, va + NBPG, __func__); 1247 pmap_pte_process(pmap, va, va + NBPG, pmap_pte_remove, 1248 rpte); 1249 PMAP_COUNT(user_mappings_changed); 1250 remap = true; 1251 } 1252 update_flags |= PMAP_TLB_NEED_IPI; 1253 } 1254 1255 if (!resident || remap) { 1256 pmap->pm_stats.resident_count++; 1257 } 1258 1259 /* Done after case that may sleep/return. */ 1260 if (pg) 1261 pmap_enter_pv(pmap, va, pg, &npte, 0); 1262 1263 /* 1264 * Now validate mapping with desired protection/wiring. 1265 * Assume uniform modified and referenced status for all 1266 * MIPS pages in a MACH page. 1267 */ 1268 if (wired) { 1269 pmap->pm_stats.wired_count++; 1270 npte = pte_wire_entry(npte); 1271 } 1272 1273 UVMHIST_LOG(*histp, "new pte %#"PRIxPTE" (pa %#"PRIxPADDR")", 1274 pte_value(npte), pa, 0, 0); 1275 1276 KASSERT(pte_valid_p(npte)); 1277 1278 pmap_md_tlb_miss_lock_enter(); 1279 pte_set(ptep, npte); 1280 pmap_tlb_update_addr(pmap, va, npte, update_flags); 1281 pmap_md_tlb_miss_lock_exit(); 1282 kpreempt_enable(); 1283 1284 if (pg != NULL && (prot == (VM_PROT_READ | VM_PROT_EXECUTE))) { 1285 KASSERT(mdpg != NULL); 1286 PMAP_COUNT(exec_mappings); 1287 if (!VM_PAGEMD_EXECPAGE_P(mdpg) && pte_cached_p(npte)) { 1288 if (!pte_deferred_exec_p(npte)) { 1289 UVMHIST_LOG(*histp, "va=%#"PRIxVADDR 1290 " pg %p: immediate syncicache", 1291 va, pg, 0, 0); 1292 pmap_page_syncicache(pg); 1293 pmap_page_set_attributes(mdpg, 1294 VM_PAGEMD_EXECPAGE); 1295 PMAP_COUNT(exec_synced_mappings); 1296 } else { 1297 UVMHIST_LOG(*histp, "va=%#"PRIxVADDR 1298 " pg %p: defer syncicache: pte %#x", 1299 va, pg, npte, 0); 1300 } 1301 } else { 1302 UVMHIST_LOG(*histp, 1303 "va=%#"PRIxVADDR" pg %p: no syncicache cached %d", 1304 va, pg, pte_cached_p(npte), 0); 1305 } 1306 } else if (pg != NULL && (prot & VM_PROT_EXECUTE)) { 1307 KASSERT(mdpg != NULL); 1308 KASSERT(prot & VM_PROT_WRITE); 1309 PMAP_COUNT(exec_mappings); 1310 pmap_page_syncicache(pg); 1311 pmap_page_clear_attributes(mdpg, VM_PAGEMD_EXECPAGE); 1312 UVMHIST_LOG(*histp, 1313 "va=%#"PRIxVADDR" pg %p: immediate syncicache (writeable)", 1314 va, pg, 0, 0); 1315 } 1316 1317 UVMHIST_LOG(*histp, " <-- 0 (OK)", 0, 0, 0, 0); 1318 return 0; 1319 } 1320 1321 void 1322 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags) 1323 { 1324 pmap_t pmap = pmap_kernel(); 1325 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 1326 struct vm_page_md * const mdpg = (pg ? VM_PAGE_TO_MD(pg) : NULL); 1327 1328 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1329 UVMHIST_LOG(pmaphist, 1330 "(va=%#"PRIxVADDR", pa=%#"PRIxPADDR", prot=%u, flags=%#x)", 1331 va, pa, prot, flags); 1332 PMAP_COUNT(kenter_pa); 1333 1334 if (mdpg == NULL) { 1335 PMAP_COUNT(kenter_pa_unmanaged); 1336 if ((flags & PMAP_CACHE_MASK) == 0) 1337 flags |= PMAP_NOCACHE; 1338 } else { 1339 if ((flags & PMAP_NOCACHE) == 0 && !PMAP_PAGE_COLOROK_P(pa, va)) 1340 PMAP_COUNT(kenter_pa_bad); 1341 } 1342 1343 pt_entry_t npte = pte_make_kenter_pa(pa, mdpg, prot, flags); 1344 kpreempt_disable(); 1345 pt_entry_t * const ptep = pmap_pte_lookup(pmap, va); 1346 KASSERTMSG(ptep != NULL, "%#"PRIxVADDR " %#"PRIxVADDR, va, 1347 pmap_limits.virtual_end); 1348 KASSERT(!pte_valid_p(*ptep)); 1349 1350 /* 1351 * No need to track non-managed pages or PMAP_KMPAGEs pages for aliases 1352 */ 1353 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1354 if (pg != NULL && (flags & PMAP_KMPAGE) == 0 1355 && pmap_md_virtual_cache_aliasing_p()) { 1356 pmap_enter_pv(pmap, va, pg, &npte, PV_KENTER); 1357 } 1358 #endif 1359 1360 /* 1361 * We have the option to force this mapping into the TLB but we 1362 * don't. Instead let the next reference to the page do it. 1363 */ 1364 pmap_md_tlb_miss_lock_enter(); 1365 pte_set(ptep, npte); 1366 pmap_tlb_update_addr(pmap_kernel(), va, npte, 0); 1367 pmap_md_tlb_miss_lock_exit(); 1368 kpreempt_enable(); 1369 #if DEBUG > 1 1370 for (u_int i = 0; i < PAGE_SIZE / sizeof(long); i++) { 1371 if (((long *)va)[i] != ((long *)pa)[i]) 1372 panic("%s: contents (%lx) of va %#"PRIxVADDR 1373 " != contents (%lx) of pa %#"PRIxPADDR, __func__, 1374 ((long *)va)[i], va, ((long *)pa)[i], pa); 1375 } 1376 #endif 1377 1378 UVMHIST_LOG(pmaphist, " <-- done (ptep=%p)", ptep, 0, 0, 0); 1379 } 1380 1381 /* 1382 * Remove the given range of addresses from the kernel map. 1383 * 1384 * It is assumed that the start and end are properly 1385 * rounded to the page size. 1386 */ 1387 1388 static bool 1389 pmap_pte_kremove(pmap_t pmap, vaddr_t sva, vaddr_t eva, pt_entry_t *ptep, 1390 uintptr_t flags) 1391 { 1392 const pt_entry_t new_pte = pte_nv_entry(true); 1393 1394 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1395 UVMHIST_LOG(pmaphist, 1396 "(pmap=%p, sva=%#"PRIxVADDR", eva=%#"PRIxVADDR", ptep=%p)", 1397 pmap, sva, eva, ptep); 1398 1399 KASSERT(kpreempt_disabled()); 1400 1401 for (; sva < eva; sva += NBPG, ptep++) { 1402 pt_entry_t pte = *ptep; 1403 if (!pte_valid_p(pte)) 1404 continue; 1405 1406 PMAP_COUNT(kremove_pages); 1407 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1408 struct vm_page * const pg = PHYS_TO_VM_PAGE(pte_to_paddr(pte)); 1409 if (pg != NULL && pmap_md_virtual_cache_aliasing_p()) { 1410 pmap_remove_pv(pmap, sva, pg, !pte_readonly_p(pte)); 1411 } 1412 #endif 1413 1414 pmap_md_tlb_miss_lock_enter(); 1415 pte_set(ptep, new_pte); 1416 pmap_tlb_invalidate_addr(pmap, sva); 1417 pmap_md_tlb_miss_lock_exit(); 1418 } 1419 1420 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1421 1422 return false; 1423 } 1424 1425 void 1426 pmap_kremove(vaddr_t va, vsize_t len) 1427 { 1428 const vaddr_t sva = trunc_page(va); 1429 const vaddr_t eva = round_page(va + len); 1430 1431 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1432 UVMHIST_LOG(pmaphist, "(va=%#"PRIxVADDR", len=%#"PRIxVSIZE")", 1433 va, len, 0, 0); 1434 1435 kpreempt_disable(); 1436 pmap_pte_process(pmap_kernel(), sva, eva, pmap_pte_kremove, 0); 1437 kpreempt_enable(); 1438 1439 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1440 } 1441 1442 void 1443 pmap_remove_all(struct pmap *pmap) 1444 { 1445 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1446 UVMHIST_LOG(pmaphist, "(pm=%p)", pmap, 0, 0, 0); 1447 1448 KASSERT(pmap != pmap_kernel()); 1449 1450 kpreempt_disable(); 1451 /* 1452 * Free all of our ASIDs which means we can skip doing all the 1453 * tlb_invalidate_addrs(). 1454 */ 1455 pmap_md_tlb_miss_lock_enter(); 1456 #ifdef MULTIPROCESSOR 1457 // This should be the last CPU with this pmap onproc 1458 KASSERT(!kcpuset_isotherset(pmap->pm_onproc, cpu_index(curcpu()))); 1459 if (kcpuset_isset(pmap->pm_onproc, cpu_index(curcpu()))) 1460 #endif 1461 pmap_tlb_asid_deactivate(pmap); 1462 #ifdef MULTIPROCESSOR 1463 KASSERT(kcpuset_iszero(pmap->pm_onproc)); 1464 #endif 1465 pmap_tlb_asid_release_all(pmap); 1466 pmap_md_tlb_miss_lock_exit(); 1467 pmap->pm_flags |= PMAP_DEFERRED_ACTIVATE; 1468 1469 #ifdef PMAP_FAULTINFO 1470 curpcb->pcb_faultinfo.pfi_faultaddr = 0; 1471 curpcb->pcb_faultinfo.pfi_repeats = 0; 1472 curpcb->pcb_faultinfo.pfi_faultpte = NULL; 1473 #endif 1474 kpreempt_enable(); 1475 1476 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1477 } 1478 1479 /* 1480 * Routine: pmap_unwire 1481 * Function: Clear the wired attribute for a map/virtual-address 1482 * pair. 1483 * In/out conditions: 1484 * The mapping must already exist in the pmap. 1485 */ 1486 void 1487 pmap_unwire(pmap_t pmap, vaddr_t va) 1488 { 1489 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1490 UVMHIST_LOG(pmaphist, "(pmap=%p, va=%#"PRIxVADDR")", pmap, va, 0, 0); 1491 PMAP_COUNT(unwire); 1492 1493 /* 1494 * Don't need to flush the TLB since PG_WIRED is only in software. 1495 */ 1496 kpreempt_disable(); 1497 pmap_addr_range_check(pmap, va, va, __func__); 1498 pt_entry_t * const ptep = pmap_pte_lookup(pmap, va); 1499 KASSERTMSG(ptep != NULL, "pmap %p va %#"PRIxVADDR" invalid STE", 1500 pmap, va); 1501 pt_entry_t pte = *ptep; 1502 KASSERTMSG(pte_valid_p(pte), 1503 "pmap %p va %#"PRIxVADDR" invalid PTE %#"PRIxPTE" @ %p", 1504 pmap, va, pte_value(pte), ptep); 1505 1506 if (pte_wired_p(pte)) { 1507 pmap_md_tlb_miss_lock_enter(); 1508 pte_set(ptep, pte_unwire_entry(pte)); 1509 pmap_md_tlb_miss_lock_exit(); 1510 pmap->pm_stats.wired_count--; 1511 } 1512 #ifdef DIAGNOSTIC 1513 else { 1514 printf("%s: wiring for pmap %p va %#"PRIxVADDR" unchanged!\n", 1515 __func__, pmap, va); 1516 } 1517 #endif 1518 kpreempt_enable(); 1519 1520 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1521 } 1522 1523 /* 1524 * Routine: pmap_extract 1525 * Function: 1526 * Extract the physical page address associated 1527 * with the given map/virtual_address pair. 1528 */ 1529 bool 1530 pmap_extract(pmap_t pmap, vaddr_t va, paddr_t *pap) 1531 { 1532 paddr_t pa; 1533 1534 if (pmap == pmap_kernel()) { 1535 if (pmap_md_direct_mapped_vaddr_p(va)) { 1536 pa = pmap_md_direct_mapped_vaddr_to_paddr(va); 1537 goto done; 1538 } 1539 if (pmap_md_io_vaddr_p(va)) 1540 panic("pmap_extract: io address %#"PRIxVADDR"", va); 1541 1542 if (va >= pmap_limits.virtual_end) 1543 panic("%s: illegal kernel mapped address %#"PRIxVADDR, 1544 __func__, va); 1545 } 1546 kpreempt_disable(); 1547 const pt_entry_t * const ptep = pmap_pte_lookup(pmap, va); 1548 if (ptep == NULL || !pte_valid_p(*ptep)) { 1549 kpreempt_enable(); 1550 return false; 1551 } 1552 pa = pte_to_paddr(*ptep) | (va & PGOFSET); 1553 kpreempt_enable(); 1554 done: 1555 if (pap != NULL) { 1556 *pap = pa; 1557 } 1558 return true; 1559 } 1560 1561 /* 1562 * Copy the range specified by src_addr/len 1563 * from the source map to the range dst_addr/len 1564 * in the destination map. 1565 * 1566 * This routine is only advisory and need not do anything. 1567 */ 1568 void 1569 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vaddr_t dst_addr, vsize_t len, 1570 vaddr_t src_addr) 1571 { 1572 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1573 PMAP_COUNT(copy); 1574 } 1575 1576 /* 1577 * pmap_clear_reference: 1578 * 1579 * Clear the reference bit on the specified physical page. 1580 */ 1581 bool 1582 pmap_clear_reference(struct vm_page *pg) 1583 { 1584 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1585 1586 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1587 UVMHIST_LOG(pmaphist, "(pg=%p (pa %#"PRIxPADDR"))", 1588 pg, VM_PAGE_TO_PHYS(pg), 0,0); 1589 1590 bool rv = pmap_page_clear_attributes(mdpg, VM_PAGEMD_REFERENCED); 1591 1592 UVMHIST_LOG(pmaphist, " <-- wasref %u", rv, 0, 0, 0); 1593 1594 return rv; 1595 } 1596 1597 /* 1598 * pmap_is_referenced: 1599 * 1600 * Return whether or not the specified physical page is referenced 1601 * by any physical maps. 1602 */ 1603 bool 1604 pmap_is_referenced(struct vm_page *pg) 1605 { 1606 return VM_PAGEMD_REFERENCED_P(VM_PAGE_TO_MD(pg)); 1607 } 1608 1609 /* 1610 * Clear the modify bits on the specified physical page. 1611 */ 1612 bool 1613 pmap_clear_modify(struct vm_page *pg) 1614 { 1615 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1616 pv_entry_t pv = &mdpg->mdpg_first; 1617 pv_entry_t pv_next; 1618 1619 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1620 UVMHIST_LOG(pmaphist, "(pg=%p (%#"PRIxPADDR"))", 1621 pg, VM_PAGE_TO_PHYS(pg), 0,0); 1622 PMAP_COUNT(clear_modify); 1623 1624 if (VM_PAGEMD_EXECPAGE_P(mdpg)) { 1625 if (pv->pv_pmap == NULL) { 1626 UVMHIST_LOG(pmapexechist, 1627 "pg %p (pa %#"PRIxPADDR"): execpage cleared", 1628 pg, VM_PAGE_TO_PHYS(pg), 0, 0); 1629 pmap_page_clear_attributes(mdpg, VM_PAGEMD_EXECPAGE); 1630 PMAP_COUNT(exec_uncached_clear_modify); 1631 } else { 1632 UVMHIST_LOG(pmapexechist, 1633 "pg %p (pa %#"PRIxPADDR"): syncicache performed", 1634 pg, VM_PAGE_TO_PHYS(pg), 0, 0); 1635 pmap_page_syncicache(pg); 1636 PMAP_COUNT(exec_synced_clear_modify); 1637 } 1638 } 1639 if (!pmap_page_clear_attributes(mdpg, VM_PAGEMD_MODIFIED)) { 1640 UVMHIST_LOG(pmaphist, " <-- false", 0, 0, 0, 0); 1641 return false; 1642 } 1643 if (pv->pv_pmap == NULL) { 1644 UVMHIST_LOG(pmaphist, " <-- true (no mappings)", 0, 0, 0, 0); 1645 return true; 1646 } 1647 1648 /* 1649 * remove write access from any pages that are dirty 1650 * so we can tell if they are written to again later. 1651 * flush the VAC first if there is one. 1652 */ 1653 kpreempt_disable(); 1654 KASSERT(!VM_PAGEMD_PVLIST_LOCKED_P(mdpg)); 1655 VM_PAGEMD_PVLIST_READLOCK(mdpg); 1656 pmap_pvlist_check(mdpg); 1657 for (; pv != NULL; pv = pv_next) { 1658 pmap_t pmap = pv->pv_pmap; 1659 vaddr_t va = trunc_page(pv->pv_va); 1660 1661 pv_next = pv->pv_next; 1662 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1663 if (pv->pv_va & PV_KENTER) 1664 continue; 1665 #endif 1666 pt_entry_t * const ptep = pmap_pte_lookup(pmap, va); 1667 KASSERT(ptep); 1668 pt_entry_t pte = pte_prot_nowrite(*ptep); 1669 if (*ptep == pte) { 1670 continue; 1671 } 1672 KASSERT(pte_valid_p(pte)); 1673 const uintptr_t gen = VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1674 pmap_md_tlb_miss_lock_enter(); 1675 pte_set(ptep, pte); 1676 pmap_tlb_invalidate_addr(pmap, va); 1677 pmap_md_tlb_miss_lock_exit(); 1678 pmap_update(pmap); 1679 if (__predict_false(gen != VM_PAGEMD_PVLIST_READLOCK(mdpg))) { 1680 /* 1681 * The list changed! So restart from the beginning. 1682 */ 1683 pv_next = &mdpg->mdpg_first; 1684 pmap_pvlist_check(mdpg); 1685 } 1686 } 1687 pmap_pvlist_check(mdpg); 1688 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1689 kpreempt_enable(); 1690 1691 UVMHIST_LOG(pmaphist, " <-- true (mappings changed)", 0, 0, 0, 0); 1692 return true; 1693 } 1694 1695 /* 1696 * pmap_is_modified: 1697 * 1698 * Return whether or not the specified physical page is modified 1699 * by any physical maps. 1700 */ 1701 bool 1702 pmap_is_modified(struct vm_page *pg) 1703 { 1704 return VM_PAGEMD_MODIFIED_P(VM_PAGE_TO_MD(pg)); 1705 } 1706 1707 /* 1708 * pmap_set_modified: 1709 * 1710 * Sets the page modified reference bit for the specified page. 1711 */ 1712 void 1713 pmap_set_modified(paddr_t pa) 1714 { 1715 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 1716 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1717 pmap_page_set_attributes(mdpg, VM_PAGEMD_MODIFIED|VM_PAGEMD_REFERENCED); 1718 } 1719 1720 /******************** pv_entry management ********************/ 1721 1722 static void 1723 pmap_pvlist_check(struct vm_page_md *mdpg) 1724 { 1725 #ifdef DEBUG 1726 pv_entry_t pv = &mdpg->mdpg_first; 1727 if (pv->pv_pmap != NULL) { 1728 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1729 const u_int colormask = uvmexp.colormask; 1730 u_int colors = 0; 1731 #endif 1732 for (; pv != NULL; pv = pv->pv_next) { 1733 KASSERT(pv->pv_pmap != pmap_kernel() || !pmap_md_direct_mapped_vaddr_p(pv->pv_va)); 1734 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1735 colors |= __BIT(atop(pv->pv_va) & colormask); 1736 #endif 1737 } 1738 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1739 // Assert that if there is more than 1 color mapped, that the 1740 // page is uncached. 1741 KASSERTMSG(!pmap_md_virtual_cache_aliasing_p() 1742 || colors == 0 || (colors & (colors-1)) == 0 1743 || VM_PAGEMD_UNCACHED_P(mdpg), "colors=%#x uncached=%u", 1744 colors, VM_PAGEMD_UNCACHED_P(mdpg)); 1745 #endif 1746 } else { 1747 KASSERT(pv->pv_next == NULL); 1748 } 1749 #endif /* DEBUG */ 1750 } 1751 1752 /* 1753 * Enter the pmap and virtual address into the 1754 * physical to virtual map table. 1755 */ 1756 void 1757 pmap_enter_pv(pmap_t pmap, vaddr_t va, struct vm_page *pg, pt_entry_t *nptep, 1758 u_int flags) 1759 { 1760 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1761 pv_entry_t pv, npv, apv; 1762 #ifdef UVMHIST 1763 bool first = false; 1764 #endif 1765 1766 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1767 UVMHIST_LOG(pmaphist, 1768 "(pmap=%p va=%#"PRIxVADDR" pg=%p (%#"PRIxPADDR")", 1769 pmap, va, pg, VM_PAGE_TO_PHYS(pg)); 1770 UVMHIST_LOG(pmaphist, "nptep=%p (%#"PRIxPTE"))", 1771 nptep, pte_value(*nptep), 0, 0); 1772 1773 KASSERT(kpreempt_disabled()); 1774 KASSERT(pmap != pmap_kernel() || !pmap_md_direct_mapped_vaddr_p(va)); 1775 KASSERTMSG(pmap != pmap_kernel() || !pmap_md_io_vaddr_p(va), 1776 "va %#"PRIxVADDR, va); 1777 1778 apv = NULL; 1779 VM_PAGEMD_PVLIST_LOCK(mdpg); 1780 again: 1781 pv = &mdpg->mdpg_first; 1782 pmap_pvlist_check(mdpg); 1783 if (pv->pv_pmap == NULL) { 1784 KASSERT(pv->pv_next == NULL); 1785 /* 1786 * No entries yet, use header as the first entry 1787 */ 1788 PMAP_COUNT(primary_mappings); 1789 PMAP_COUNT(mappings); 1790 #ifdef UVMHIST 1791 first = true; 1792 #endif 1793 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1794 KASSERT(VM_PAGEMD_CACHED_P(mdpg)); 1795 // If the new mapping has an incompatible color the last 1796 // mapping of this page, clean the page before using it. 1797 if (!PMAP_PAGE_COLOROK_P(va, pv->pv_va)) { 1798 pmap_md_vca_clean(pg, PMAP_WBINV); 1799 } 1800 #endif 1801 pv->pv_pmap = pmap; 1802 pv->pv_va = va | flags; 1803 } else { 1804 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1805 if (pmap_md_vca_add(pg, va, nptep)) { 1806 goto again; 1807 } 1808 #endif 1809 1810 /* 1811 * There is at least one other VA mapping this page. 1812 * Place this entry after the header. 1813 * 1814 * Note: the entry may already be in the table if 1815 * we are only changing the protection bits. 1816 */ 1817 1818 #ifdef PARANOIADIAG 1819 const paddr_t pa = VM_PAGE_TO_PHYS(pg); 1820 #endif 1821 for (npv = pv; npv; npv = npv->pv_next) { 1822 if (pmap == npv->pv_pmap 1823 && va == trunc_page(npv->pv_va)) { 1824 #ifdef PARANOIADIAG 1825 pt_entry_t *ptep = pmap_pte_lookup(pmap, va); 1826 pt_entry_t pte = (ptep != NULL) ? *ptep : 0; 1827 if (!pte_valid_p(pte) || pte_to_paddr(pte) != pa) 1828 printf("%s: found va %#"PRIxVADDR 1829 " pa %#"PRIxPADDR 1830 " in pv_table but != %#"PRIxPTE"\n", 1831 __func__, va, pa, pte_value(pte)); 1832 #endif 1833 PMAP_COUNT(remappings); 1834 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1835 if (__predict_false(apv != NULL)) 1836 pmap_pv_free(apv); 1837 1838 UVMHIST_LOG(pmaphist, " <-- done pv=%p (reused)", 1839 pv, 0, 0, 0); 1840 return; 1841 } 1842 } 1843 if (__predict_true(apv == NULL)) { 1844 /* 1845 * To allocate a PV, we have to release the PVLIST lock 1846 * so get the page generation. We allocate the PV, and 1847 * then reacquire the lock. 1848 */ 1849 pmap_pvlist_check(mdpg); 1850 const uintptr_t gen = VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1851 1852 apv = (pv_entry_t)pmap_pv_alloc(); 1853 if (apv == NULL) 1854 panic("pmap_enter_pv: pmap_pv_alloc() failed"); 1855 1856 /* 1857 * If the generation has changed, then someone else 1858 * tinkered with this page so we should start over. 1859 */ 1860 if (gen != VM_PAGEMD_PVLIST_LOCK(mdpg)) 1861 goto again; 1862 } 1863 npv = apv; 1864 apv = NULL; 1865 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1866 /* 1867 * If need to deal with virtual cache aliases, keep mappings 1868 * in the kernel pmap at the head of the list. This allows 1869 * the VCA code to easily use them for cache operations if 1870 * present. 1871 */ 1872 pmap_t kpmap = pmap_kernel(); 1873 if (pmap != kpmap) { 1874 while (pv->pv_pmap == kpmap && pv->pv_next != NULL) { 1875 pv = pv->pv_next; 1876 } 1877 } 1878 #endif 1879 npv->pv_va = va | flags; 1880 npv->pv_pmap = pmap; 1881 npv->pv_next = pv->pv_next; 1882 pv->pv_next = npv; 1883 PMAP_COUNT(mappings); 1884 } 1885 pmap_pvlist_check(mdpg); 1886 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1887 if (__predict_false(apv != NULL)) 1888 pmap_pv_free(apv); 1889 1890 UVMHIST_LOG(pmaphist, " <-- done pv=%p (first %u)", pv, first, 0, 0); 1891 } 1892 1893 /* 1894 * Remove a physical to virtual address translation. 1895 * If cache was inhibited on this page, and there are no more cache 1896 * conflicts, restore caching. 1897 * Flush the cache if the last page is removed (should always be cached 1898 * at this point). 1899 */ 1900 void 1901 pmap_remove_pv(pmap_t pmap, vaddr_t va, struct vm_page *pg, bool dirty) 1902 { 1903 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1904 pv_entry_t pv, npv; 1905 bool last; 1906 1907 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1908 UVMHIST_LOG(pmaphist, 1909 "(pmap=%p, va=%#"PRIxVADDR", pg=%p (pa %#"PRIxPADDR")", 1910 pmap, va, pg, VM_PAGE_TO_PHYS(pg)); 1911 UVMHIST_LOG(pmaphist, "dirty=%u)", dirty, 0, 0, 0); 1912 1913 KASSERT(kpreempt_disabled()); 1914 KASSERT((va & PAGE_MASK) == 0); 1915 pv = &mdpg->mdpg_first; 1916 1917 VM_PAGEMD_PVLIST_LOCK(mdpg); 1918 pmap_pvlist_check(mdpg); 1919 1920 /* 1921 * If it is the first entry on the list, it is actually 1922 * in the header and we must copy the following entry up 1923 * to the header. Otherwise we must search the list for 1924 * the entry. In either case we free the now unused entry. 1925 */ 1926 1927 last = false; 1928 if (pmap == pv->pv_pmap && va == trunc_page(pv->pv_va)) { 1929 npv = pv->pv_next; 1930 if (npv) { 1931 *pv = *npv; 1932 KASSERT(pv->pv_pmap != NULL); 1933 } else { 1934 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1935 pmap_page_clear_attributes(mdpg, VM_PAGEMD_UNCACHED); 1936 #endif 1937 pv->pv_pmap = NULL; 1938 last = true; /* Last mapping removed */ 1939 } 1940 PMAP_COUNT(remove_pvfirst); 1941 } else { 1942 for (npv = pv->pv_next; npv; pv = npv, npv = npv->pv_next) { 1943 PMAP_COUNT(remove_pvsearch); 1944 if (pmap == npv->pv_pmap && va == trunc_page(npv->pv_va)) 1945 break; 1946 } 1947 if (npv) { 1948 pv->pv_next = npv->pv_next; 1949 } 1950 } 1951 1952 pmap_pvlist_check(mdpg); 1953 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1954 1955 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1956 pmap_md_vca_remove(pg, va, dirty, last); 1957 #endif 1958 1959 /* 1960 * Free the pv_entry if needed. 1961 */ 1962 if (npv) 1963 pmap_pv_free(npv); 1964 if (VM_PAGEMD_EXECPAGE_P(mdpg) && dirty) { 1965 if (last) { 1966 /* 1967 * If this was the page's last mapping, we no longer 1968 * care about its execness. 1969 */ 1970 UVMHIST_LOG(pmapexechist, 1971 "pg %p (pa %#"PRIxPADDR")last %u: execpage cleared", 1972 pg, VM_PAGE_TO_PHYS(pg), last, 0); 1973 pmap_page_clear_attributes(mdpg, VM_PAGEMD_EXECPAGE); 1974 PMAP_COUNT(exec_uncached_remove); 1975 } else { 1976 /* 1977 * Someone still has it mapped as an executable page 1978 * so we must sync it. 1979 */ 1980 UVMHIST_LOG(pmapexechist, 1981 "pg %p (pa %#"PRIxPADDR")last %u: performed syncicache", 1982 pg, VM_PAGE_TO_PHYS(pg), last, 0); 1983 pmap_page_syncicache(pg); 1984 PMAP_COUNT(exec_synced_remove); 1985 } 1986 } 1987 1988 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1989 } 1990 1991 #if defined(MULTIPROCESSOR) 1992 struct pmap_pvlist_info { 1993 kmutex_t *pli_locks[PAGE_SIZE / 32]; 1994 volatile u_int pli_lock_refs[PAGE_SIZE / 32]; 1995 volatile u_int pli_lock_index; 1996 u_int pli_lock_mask; 1997 } pmap_pvlist_info; 1998 1999 void 2000 pmap_pvlist_lock_init(size_t cache_line_size) 2001 { 2002 struct pmap_pvlist_info * const pli = &pmap_pvlist_info; 2003 const vaddr_t lock_page = uvm_pageboot_alloc(PAGE_SIZE); 2004 vaddr_t lock_va = lock_page; 2005 if (sizeof(kmutex_t) > cache_line_size) { 2006 cache_line_size = roundup2(sizeof(kmutex_t), cache_line_size); 2007 } 2008 const size_t nlocks = PAGE_SIZE / cache_line_size; 2009 KASSERT((nlocks & (nlocks - 1)) == 0); 2010 /* 2011 * Now divide the page into a number of mutexes, one per cacheline. 2012 */ 2013 for (size_t i = 0; i < nlocks; lock_va += cache_line_size, i++) { 2014 kmutex_t * const lock = (kmutex_t *)lock_va; 2015 mutex_init(lock, MUTEX_DEFAULT, IPL_HIGH); 2016 pli->pli_locks[i] = lock; 2017 } 2018 pli->pli_lock_mask = nlocks - 1; 2019 } 2020 2021 kmutex_t * 2022 pmap_pvlist_lock_addr(struct vm_page_md *mdpg) 2023 { 2024 struct pmap_pvlist_info * const pli = &pmap_pvlist_info; 2025 kmutex_t *lock = mdpg->mdpg_lock; 2026 2027 /* 2028 * Allocate a lock on an as-needed basis. This will hopefully give us 2029 * semi-random distribution not based on page color. 2030 */ 2031 if (__predict_false(lock == NULL)) { 2032 size_t locknum = atomic_add_int_nv(&pli->pli_lock_index, 37); 2033 size_t lockid = locknum & pli->pli_lock_mask; 2034 kmutex_t * const new_lock = pli->pli_locks[lockid]; 2035 /* 2036 * Set the lock. If some other thread already did, just use 2037 * the one they assigned. 2038 */ 2039 lock = atomic_cas_ptr(&mdpg->mdpg_lock, NULL, new_lock); 2040 if (lock == NULL) { 2041 lock = new_lock; 2042 atomic_inc_uint(&pli->pli_lock_refs[lockid]); 2043 } 2044 } 2045 2046 /* 2047 * Now finally provide the lock. 2048 */ 2049 return lock; 2050 } 2051 #else /* !MULTIPROCESSOR */ 2052 void 2053 pmap_pvlist_lock_init(size_t cache_line_size) 2054 { 2055 mutex_init(&pmap_pvlist_mutex, MUTEX_DEFAULT, IPL_HIGH); 2056 } 2057 2058 #ifdef MODULAR 2059 kmutex_t * 2060 pmap_pvlist_lock_addr(struct vm_page_md *mdpg) 2061 { 2062 /* 2063 * We just use a global lock. 2064 */ 2065 if (__predict_false(mdpg->mdpg_lock == NULL)) { 2066 mdpg->mdpg_lock = &pmap_pvlist_mutex; 2067 } 2068 2069 /* 2070 * Now finally provide the lock. 2071 */ 2072 return mdpg->mdpg_lock; 2073 } 2074 #endif /* MODULAR */ 2075 #endif /* !MULTIPROCESSOR */ 2076 2077 /* 2078 * pmap_pv_page_alloc: 2079 * 2080 * Allocate a page for the pv_entry pool. 2081 */ 2082 void * 2083 pmap_pv_page_alloc(struct pool *pp, int flags) 2084 { 2085 struct vm_page * const pg = PMAP_ALLOC_POOLPAGE(UVM_PGA_USERESERVE); 2086 if (pg == NULL) 2087 return NULL; 2088 2089 return (void *)pmap_map_poolpage(VM_PAGE_TO_PHYS(pg)); 2090 } 2091 2092 /* 2093 * pmap_pv_page_free: 2094 * 2095 * Free a pv_entry pool page. 2096 */ 2097 void 2098 pmap_pv_page_free(struct pool *pp, void *v) 2099 { 2100 vaddr_t va = (vaddr_t)v; 2101 2102 KASSERT(pmap_md_direct_mapped_vaddr_p(va)); 2103 const paddr_t pa = pmap_md_direct_mapped_vaddr_to_paddr(va); 2104 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 2105 KASSERT(pg != NULL); 2106 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 2107 kpreempt_disable(); 2108 pmap_md_vca_remove(pg, va, true, true); 2109 kpreempt_enable(); 2110 #endif 2111 pmap_page_clear_attributes(VM_PAGE_TO_MD(pg), VM_PAGEMD_POOLPAGE); 2112 KASSERT(!VM_PAGEMD_EXECPAGE_P(VM_PAGE_TO_MD(pg))); 2113 uvm_pagefree(pg); 2114 } 2115 2116 #ifdef PMAP_PREFER 2117 /* 2118 * Find first virtual address >= *vap that doesn't cause 2119 * a cache alias conflict. 2120 */ 2121 void 2122 pmap_prefer(vaddr_t foff, vaddr_t *vap, vsize_t sz, int td) 2123 { 2124 vsize_t prefer_mask = ptoa(uvmexp.colormask); 2125 2126 PMAP_COUNT(prefer_requests); 2127 2128 prefer_mask |= pmap_md_cache_prefer_mask(); 2129 2130 if (prefer_mask) { 2131 vaddr_t va = *vap; 2132 vsize_t d = (foff - va) & prefer_mask; 2133 if (d) { 2134 if (td) 2135 *vap = trunc_page(va - ((-d) & prefer_mask)); 2136 else 2137 *vap = round_page(va + d); 2138 PMAP_COUNT(prefer_adjustments); 2139 } 2140 } 2141 } 2142 #endif /* PMAP_PREFER */ 2143 2144 #ifdef PMAP_MAP_POOLPAGE 2145 vaddr_t 2146 pmap_map_poolpage(paddr_t pa) 2147 { 2148 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 2149 KASSERT(pg); 2150 2151 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 2152 KASSERT(!VM_PAGEMD_EXECPAGE_P(mdpg)); 2153 2154 pmap_page_set_attributes(mdpg, VM_PAGEMD_POOLPAGE); 2155 2156 return pmap_md_map_poolpage(pa, NBPG); 2157 } 2158 2159 paddr_t 2160 pmap_unmap_poolpage(vaddr_t va) 2161 { 2162 KASSERT(pmap_md_direct_mapped_vaddr_p(va)); 2163 paddr_t pa = pmap_md_direct_mapped_vaddr_to_paddr(va); 2164 2165 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 2166 KASSERT(pg != NULL); 2167 KASSERT(!VM_PAGEMD_EXECPAGE_P(VM_PAGE_TO_MD(pg))); 2168 2169 pmap_page_clear_attributes(VM_PAGE_TO_MD(pg), VM_PAGEMD_POOLPAGE); 2170 pmap_md_unmap_poolpage(va, NBPG); 2171 2172 return pa; 2173 } 2174 #endif /* PMAP_MAP_POOLPAGE */ 2175
Indexes created Fri Sep 26 20:09:58 GMT 2025