1 /* $NetBSD: pmap.c,v 1.27 2016/12/23 09:16:46 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 1.27 2016/12/23 09:16:46 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 #define LNAME(l) \ 668 ((l)->l_name ? (l)->l_name : (l)->l_proc->p_comm) 669 670 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 671 UVMHIST_LOG(pmaphist, "(l=%p pmap=%p (%s))", l, pmap, LNAME(l), 0); 672 PMAP_COUNT(activate); 673 674 kpreempt_disable(); 675 pmap_md_tlb_miss_lock_enter(); 676 pmap_tlb_asid_acquire(pmap, l); 677 if (l == curlwp) { 678 pmap_segtab_activate(pmap, l); 679 } 680 pmap_md_tlb_miss_lock_exit(); 681 kpreempt_enable(); 682 683 UVMHIST_LOG(pmaphist, " <-- done (%u:%u)", l->l_proc->p_pid, l->l_lid, 0, 0); 684 } 685 686 /* 687 * Remove this page from all physical maps in which it resides. 688 * Reflects back modify bits to the pager. 689 */ 690 void 691 pmap_page_remove(struct vm_page *pg) 692 { 693 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 694 695 kpreempt_disable(); 696 VM_PAGEMD_PVLIST_LOCK(mdpg); 697 pmap_pvlist_check(mdpg); 698 699 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 700 701 UVMHIST_LOG(pmapexechist, "pg %p (pa %#"PRIxPADDR")%s: %s", 702 pg, VM_PAGE_TO_PHYS(pg), " [page removed]", "execpage cleared"); 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 *ptep = npte; 777 /* 778 * Flush the TLB for the given address. 779 */ 780 pmap_tlb_invalidate_addr(pmap, va); 781 pmap_md_tlb_miss_lock_exit(); 782 783 /* 784 * non-null means this is a non-pvh_first pv, so we should 785 * free it. 786 */ 787 if (pvp) { 788 KASSERT(pvp->pv_pmap == pmap_kernel()); 789 KASSERT(pvp->pv_next == NULL); 790 pmap_pv_free(pv); 791 } else { 792 pv->pv_pmap = NULL; 793 pv->pv_next = NULL; 794 } 795 } 796 797 pmap_pvlist_check(mdpg); 798 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 799 kpreempt_enable(); 800 801 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 802 } 803 804 805 /* 806 * Make a previously active pmap (vmspace) inactive. 807 */ 808 void 809 pmap_deactivate(struct lwp *l) 810 { 811 pmap_t pmap = l->l_proc->p_vmspace->vm_map.pmap; 812 813 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 814 UVMHIST_LOG(pmaphist, "(l=%p pmap=%p (%s))", l, pmap, LNAME(l), 0); 815 PMAP_COUNT(deactivate); 816 817 kpreempt_disable(); 818 KASSERT(l == curlwp || l->l_cpu == curlwp->l_cpu); 819 pmap_md_tlb_miss_lock_enter(); 820 curcpu()->ci_pmap_user_segtab = PMAP_INVALID_SEGTAB_ADDRESS; 821 #ifdef _LP64 822 curcpu()->ci_pmap_user_seg0tab = NULL; 823 #endif 824 pmap_tlb_asid_deactivate(pmap); 825 pmap_md_tlb_miss_lock_exit(); 826 kpreempt_enable(); 827 828 UVMHIST_LOG(pmaphist, " <-- done (%u:%u)", l->l_proc->p_pid, l->l_lid, 0, 0); 829 } 830 831 void 832 pmap_update(struct pmap *pmap) 833 { 834 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 835 UVMHIST_LOG(pmaphist, "(pmap=%p)", pmap, 0, 0, 0); 836 PMAP_COUNT(update); 837 838 kpreempt_disable(); 839 #if defined(MULTIPROCESSOR) && defined(PMAP_TLB_NEED_SHOOTDOWN) 840 u_int pending = atomic_swap_uint(&pmap->pm_shootdown_pending, 0); 841 if (pending && pmap_tlb_shootdown_bystanders(pmap)) 842 PMAP_COUNT(shootdown_ipis); 843 #endif 844 pmap_md_tlb_miss_lock_enter(); 845 #if defined(DEBUG) && !defined(MULTIPROCESSOR) 846 pmap_tlb_check(pmap, pmap_md_tlb_check_entry); 847 #endif /* DEBUG */ 848 849 /* 850 * If pmap_remove_all was called, we deactivated ourselves and nuked 851 * our ASID. Now we have to reactivate ourselves. 852 */ 853 if (__predict_false(pmap->pm_flags & PMAP_DEFERRED_ACTIVATE)) { 854 pmap->pm_flags ^= PMAP_DEFERRED_ACTIVATE; 855 pmap_tlb_asid_acquire(pmap, curlwp); 856 pmap_segtab_activate(pmap, curlwp); 857 } 858 pmap_md_tlb_miss_lock_exit(); 859 kpreempt_enable(); 860 861 UVMHIST_LOG(pmaphist, " <-- done%s", 862 (pmap == pmap_kernel()) ? " (kernel)" : "", 0, 0, 0); 863 } 864 865 /* 866 * Remove the given range of addresses from the specified map. 867 * 868 * It is assumed that the start and end are properly 869 * rounded to the page size. 870 */ 871 872 static bool 873 pmap_pte_remove(pmap_t pmap, vaddr_t sva, vaddr_t eva, pt_entry_t *ptep, 874 uintptr_t flags) 875 { 876 const pt_entry_t npte = flags; 877 const bool is_kernel_pmap_p = (pmap == pmap_kernel()); 878 879 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 880 UVMHIST_LOG(pmaphist, "(pmap=%p %sva=%#"PRIxVADDR"..%#"PRIxVADDR, 881 pmap, (is_kernel_pmap_p ? "(kernel) " : ""), sva, eva); 882 UVMHIST_LOG(pmaphist, "ptep=%p, flags(npte)=%#"PRIxPTR")", 883 ptep, flags, 0, 0); 884 885 KASSERT(kpreempt_disabled()); 886 887 for (; sva < eva; sva += NBPG, ptep++) { 888 const pt_entry_t pte = *ptep; 889 if (!pte_valid_p(pte)) 890 continue; 891 if (is_kernel_pmap_p) { 892 PMAP_COUNT(remove_kernel_pages); 893 } else { 894 PMAP_COUNT(remove_user_pages); 895 } 896 if (pte_wired_p(pte)) 897 pmap->pm_stats.wired_count--; 898 pmap->pm_stats.resident_count--; 899 struct vm_page * const pg = PHYS_TO_VM_PAGE(pte_to_paddr(pte)); 900 if (__predict_true(pg != NULL)) { 901 pmap_remove_pv(pmap, sva, pg, pte_modified_p(pte)); 902 } 903 pmap_md_tlb_miss_lock_enter(); 904 *ptep = npte; 905 /* 906 * Flush the TLB for the given address. 907 */ 908 pmap_tlb_invalidate_addr(pmap, sva); 909 pmap_md_tlb_miss_lock_exit(); 910 } 911 912 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 913 914 return false; 915 } 916 917 void 918 pmap_remove(pmap_t pmap, vaddr_t sva, vaddr_t eva) 919 { 920 const bool is_kernel_pmap_p = (pmap == pmap_kernel()); 921 const pt_entry_t npte = pte_nv_entry(is_kernel_pmap_p); 922 923 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 924 UVMHIST_LOG(pmaphist, "(pmap=%p, va=%#"PRIxVADDR"..%#"PRIxVADDR")", 925 pmap, sva, eva, 0); 926 927 if (is_kernel_pmap_p) { 928 PMAP_COUNT(remove_kernel_calls); 929 } else { 930 PMAP_COUNT(remove_user_calls); 931 } 932 #ifdef PMAP_FAULTINFO 933 curpcb->pcb_faultinfo.pfi_faultaddr = 0; 934 curpcb->pcb_faultinfo.pfi_repeats = 0; 935 curpcb->pcb_faultinfo.pfi_faultpte = NULL; 936 #endif 937 kpreempt_disable(); 938 pmap_addr_range_check(pmap, sva, eva, __func__); 939 pmap_pte_process(pmap, sva, eva, pmap_pte_remove, npte); 940 kpreempt_enable(); 941 942 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 943 } 944 945 /* 946 * pmap_page_protect: 947 * 948 * Lower the permission for all mappings to a given page. 949 */ 950 void 951 pmap_page_protect(struct vm_page *pg, vm_prot_t prot) 952 { 953 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 954 pv_entry_t pv; 955 vaddr_t va; 956 957 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 958 UVMHIST_LOG(pmaphist, "(pg=%p (pa %#"PRIxPADDR") prot=%#x)", 959 pg, VM_PAGE_TO_PHYS(pg), prot, 0); 960 PMAP_COUNT(page_protect); 961 962 switch (prot) { 963 case VM_PROT_READ|VM_PROT_WRITE: 964 case VM_PROT_ALL: 965 break; 966 967 /* copy_on_write */ 968 case VM_PROT_READ: 969 case VM_PROT_READ|VM_PROT_EXECUTE: 970 pv = &mdpg->mdpg_first; 971 kpreempt_disable(); 972 VM_PAGEMD_PVLIST_READLOCK(mdpg); 973 pmap_pvlist_check(mdpg); 974 /* 975 * Loop over all current mappings setting/clearing as apropos. 976 */ 977 if (pv->pv_pmap != NULL) { 978 while (pv != NULL) { 979 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 980 if (pv->pv_va & PV_KENTER) { 981 pv = pv->pv_next; 982 continue; 983 } 984 #endif 985 const pmap_t pmap = pv->pv_pmap; 986 va = trunc_page(pv->pv_va); 987 const uintptr_t gen = 988 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 989 pmap_protect(pmap, va, va + PAGE_SIZE, prot); 990 KASSERT(pv->pv_pmap == pmap); 991 pmap_update(pmap); 992 if (gen != VM_PAGEMD_PVLIST_READLOCK(mdpg)) { 993 pv = &mdpg->mdpg_first; 994 } else { 995 pv = pv->pv_next; 996 } 997 pmap_pvlist_check(mdpg); 998 } 999 } 1000 pmap_pvlist_check(mdpg); 1001 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1002 kpreempt_enable(); 1003 break; 1004 1005 /* remove_all */ 1006 default: 1007 pmap_page_remove(pg); 1008 } 1009 1010 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1011 } 1012 1013 static bool 1014 pmap_pte_protect(pmap_t pmap, vaddr_t sva, vaddr_t eva, pt_entry_t *ptep, 1015 uintptr_t flags) 1016 { 1017 const vm_prot_t prot = (flags & VM_PROT_ALL); 1018 1019 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1020 UVMHIST_LOG(pmaphist, "(pmap=%p %sva=%#"PRIxVADDR"..%#"PRIxVADDR, 1021 pmap, (pmap == pmap_kernel() ? "(kernel) " : ""), sva, eva); 1022 UVMHIST_LOG(pmaphist, "ptep=%p, flags(npte)=%#"PRIxPTR")", 1023 ptep, flags, 0, 0); 1024 1025 KASSERT(kpreempt_disabled()); 1026 /* 1027 * Change protection on every valid mapping within this segment. 1028 */ 1029 for (; sva < eva; sva += NBPG, ptep++) { 1030 pt_entry_t pte = *ptep; 1031 if (!pte_valid_p(pte)) 1032 continue; 1033 struct vm_page * const pg = PHYS_TO_VM_PAGE(pte_to_paddr(pte)); 1034 if (pg != NULL && pte_modified_p(pte)) { 1035 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1036 if (VM_PAGEMD_EXECPAGE_P(mdpg)) { 1037 KASSERT(mdpg->mdpg_first.pv_pmap != NULL); 1038 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1039 if (VM_PAGEMD_CACHED_P(mdpg)) { 1040 #endif 1041 UVMHIST_LOG(pmapexechist, 1042 "pg %p (pa %#"PRIxPADDR"): %s", 1043 pg, VM_PAGE_TO_PHYS(pg), 1044 "syncicached performed", 0); 1045 pmap_page_syncicache(pg); 1046 PMAP_COUNT(exec_synced_protect); 1047 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1048 } 1049 #endif 1050 } 1051 } 1052 pte = pte_prot_downgrade(pte, prot); 1053 if (*ptep != pte) { 1054 pmap_md_tlb_miss_lock_enter(); 1055 *ptep = pte; 1056 /* 1057 * Update the TLB if needed. 1058 */ 1059 pmap_tlb_update_addr(pmap, sva, pte, PMAP_TLB_NEED_IPI); 1060 pmap_md_tlb_miss_lock_exit(); 1061 } 1062 } 1063 1064 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1065 1066 return false; 1067 } 1068 1069 /* 1070 * Set the physical protection on the 1071 * specified range of this map as requested. 1072 */ 1073 void 1074 pmap_protect(pmap_t pmap, vaddr_t sva, vaddr_t eva, vm_prot_t prot) 1075 { 1076 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1077 UVMHIST_LOG(pmaphist, 1078 "(pmap=%p, va=%#"PRIxVADDR"..%#"PRIxVADDR", prot=%u)", 1079 pmap, sva, eva, prot); 1080 PMAP_COUNT(protect); 1081 1082 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1083 pmap_remove(pmap, sva, eva); 1084 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1085 return; 1086 } 1087 1088 /* 1089 * Change protection on every valid mapping within this segment. 1090 */ 1091 kpreempt_disable(); 1092 pmap_addr_range_check(pmap, sva, eva, __func__); 1093 pmap_pte_process(pmap, sva, eva, pmap_pte_protect, prot); 1094 kpreempt_enable(); 1095 1096 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1097 } 1098 1099 #if defined(PMAP_VIRTUAL_CACHE_ALIASES) && !defined(PMAP_NO_PV_UNCACHED) 1100 /* 1101 * pmap_page_cache: 1102 * 1103 * Change all mappings of a managed page to cached/uncached. 1104 */ 1105 void 1106 pmap_page_cache(struct vm_page *pg, bool cached) 1107 { 1108 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1109 1110 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1111 UVMHIST_LOG(pmaphist, "(pg=%p (pa %#"PRIxPADDR") cached=%s)", 1112 pg, VM_PAGE_TO_PHYS(pg), cached ? "true" : "false", 0); 1113 1114 KASSERT(kpreempt_disabled()); 1115 KASSERT(VM_PAGEMD_PVLIST_LOCKED_P(mdpg)); 1116 1117 if (cached) { 1118 pmap_page_clear_attributes(mdpg, VM_PAGEMD_UNCACHED); 1119 PMAP_COUNT(page_cache_restorations); 1120 } else { 1121 pmap_page_set_attributes(mdpg, VM_PAGEMD_UNCACHED); 1122 PMAP_COUNT(page_cache_evictions); 1123 } 1124 1125 for (pv_entry_t pv = &mdpg->mdpg_first; pv != NULL; pv = pv->pv_next) { 1126 pmap_t pmap = pv->pv_pmap; 1127 vaddr_t va = trunc_page(pv->pv_va); 1128 1129 KASSERT(pmap != NULL); 1130 KASSERT(pmap != pmap_kernel() || !pmap_md_direct_mapped_vaddr_p(va)); 1131 pt_entry_t * const ptep = pmap_pte_lookup(pmap, va); 1132 if (ptep == NULL) 1133 continue; 1134 pt_entry_t pte = *ptep; 1135 if (pte_valid_p(pte)) { 1136 pte = pte_cached_change(pte, cached); 1137 pmap_md_tlb_miss_lock_enter(); 1138 *ptep = pte; 1139 pmap_tlb_update_addr(pmap, va, pte, PMAP_TLB_NEED_IPI); 1140 pmap_md_tlb_miss_lock_exit(); 1141 } 1142 } 1143 1144 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1145 } 1146 #endif /* PMAP_VIRTUAL_CACHE_ALIASES && !PMAP_NO_PV_UNCACHED */ 1147 1148 /* 1149 * Insert the given physical page (p) at 1150 * the specified virtual address (v) in the 1151 * target physical map with the protection requested. 1152 * 1153 * If specified, the page will be wired down, meaning 1154 * that the related pte can not be reclaimed. 1155 * 1156 * NB: This is the only routine which MAY NOT lazy-evaluate 1157 * or lose information. That is, this routine must actually 1158 * insert this page into the given map NOW. 1159 */ 1160 int 1161 pmap_enter(pmap_t pmap, vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags) 1162 { 1163 const bool wired = (flags & PMAP_WIRED) != 0; 1164 const bool is_kernel_pmap_p = (pmap == pmap_kernel()); 1165 u_int update_flags = (flags & VM_PROT_ALL) != 0 ? PMAP_TLB_INSERT : 0; 1166 #ifdef UVMHIST 1167 struct kern_history * const histp = 1168 ((prot & VM_PROT_EXECUTE) ? &pmapexechist : &pmaphist); 1169 #endif 1170 1171 UVMHIST_FUNC(__func__); UVMHIST_CALLED(*histp); 1172 #define VM_PROT_STRING(prot) \ 1173 &"\0 " \ 1174 "(R)\0 " \ 1175 "(W)\0 " \ 1176 "(RW)\0 " \ 1177 "(X)\0 " \ 1178 "(RX)\0 " \ 1179 "(WX)\0 " \ 1180 "(RWX)\0"[UVM_PROTECTION(prot)*6] 1181 UVMHIST_LOG(*histp, "(pmap=%p, va=%#"PRIxVADDR", pa=%#"PRIxPADDR, 1182 pmap, va, pa, 0); 1183 UVMHIST_LOG(*histp, "prot=%#x%s flags=%#x%s)", 1184 prot, VM_PROT_STRING(prot), flags, VM_PROT_STRING(flags)); 1185 1186 const bool good_color = PMAP_PAGE_COLOROK_P(pa, va); 1187 if (is_kernel_pmap_p) { 1188 PMAP_COUNT(kernel_mappings); 1189 if (!good_color) 1190 PMAP_COUNT(kernel_mappings_bad); 1191 } else { 1192 PMAP_COUNT(user_mappings); 1193 if (!good_color) 1194 PMAP_COUNT(user_mappings_bad); 1195 } 1196 pmap_addr_range_check(pmap, va, va, __func__); 1197 1198 KASSERTMSG(prot & VM_PROT_READ, "no READ (%#x) in prot %#x", 1199 VM_PROT_READ, prot); 1200 1201 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 1202 struct vm_page_md * const mdpg = (pg ? VM_PAGE_TO_MD(pg) : NULL); 1203 1204 if (pg) { 1205 /* Set page referenced/modified status based on flags */ 1206 if (flags & VM_PROT_WRITE) { 1207 pmap_page_set_attributes(mdpg, VM_PAGEMD_MODIFIED|VM_PAGEMD_REFERENCED); 1208 } else if (flags & VM_PROT_ALL) { 1209 pmap_page_set_attributes(mdpg, VM_PAGEMD_REFERENCED); 1210 } 1211 1212 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1213 if (!VM_PAGEMD_CACHED_P(mdpg)) { 1214 flags |= PMAP_NOCACHE; 1215 PMAP_COUNT(uncached_mappings); 1216 } 1217 #endif 1218 1219 PMAP_COUNT(managed_mappings); 1220 } else { 1221 /* 1222 * Assumption: if it is not part of our managed memory 1223 * then it must be device memory which may be volatile. 1224 */ 1225 if ((flags & PMAP_CACHE_MASK) == 0) 1226 flags |= PMAP_NOCACHE; 1227 PMAP_COUNT(unmanaged_mappings); 1228 } 1229 1230 pt_entry_t npte = pte_make_enter(pa, mdpg, prot, flags, 1231 is_kernel_pmap_p); 1232 1233 kpreempt_disable(); 1234 1235 pt_entry_t * const ptep = pmap_pte_reserve(pmap, va, flags); 1236 if (__predict_false(ptep == NULL)) { 1237 kpreempt_enable(); 1238 UVMHIST_LOG(*histp, " <-- ENOMEM", 0, 0, 0, 0); 1239 return ENOMEM; 1240 } 1241 const pt_entry_t opte = *ptep; 1242 const bool resident = pte_valid_p(opte); 1243 bool remap = false; 1244 if (resident) { 1245 if (pte_to_paddr(opte) != pa) { 1246 KASSERT(!is_kernel_pmap_p); 1247 const pt_entry_t rpte = pte_nv_entry(false); 1248 1249 pmap_addr_range_check(pmap, va, va + NBPG, __func__); 1250 pmap_pte_process(pmap, va, va + NBPG, pmap_pte_remove, 1251 rpte); 1252 PMAP_COUNT(user_mappings_changed); 1253 remap = true; 1254 } 1255 update_flags |= PMAP_TLB_NEED_IPI; 1256 } 1257 1258 if (!resident || remap) { 1259 pmap->pm_stats.resident_count++; 1260 } 1261 1262 /* Done after case that may sleep/return. */ 1263 if (pg) 1264 pmap_enter_pv(pmap, va, pg, &npte, 0); 1265 1266 /* 1267 * Now validate mapping with desired protection/wiring. 1268 * Assume uniform modified and referenced status for all 1269 * MIPS pages in a MACH page. 1270 */ 1271 if (wired) { 1272 pmap->pm_stats.wired_count++; 1273 npte = pte_wire_entry(npte); 1274 } 1275 1276 UVMHIST_LOG(*histp, "new pte %#"PRIxPTE" (pa %#"PRIxPADDR")", 1277 pte_value(npte), pa, 0, 0); 1278 1279 KASSERT(pte_valid_p(npte)); 1280 1281 pmap_md_tlb_miss_lock_enter(); 1282 *ptep = npte; 1283 pmap_tlb_update_addr(pmap, va, npte, update_flags); 1284 pmap_md_tlb_miss_lock_exit(); 1285 kpreempt_enable(); 1286 1287 if (pg != NULL && (prot == (VM_PROT_READ | VM_PROT_EXECUTE))) { 1288 KASSERT(mdpg != NULL); 1289 PMAP_COUNT(exec_mappings); 1290 if (!VM_PAGEMD_EXECPAGE_P(mdpg) && pte_cached_p(npte)) { 1291 if (!pte_deferred_exec_p(npte)) { 1292 UVMHIST_LOG(*histp, 1293 "va=%#"PRIxVADDR" pg %p: %s syncicache%s", 1294 va, pg, "immediate", ""); 1295 pmap_page_syncicache(pg); 1296 pmap_page_set_attributes(mdpg, 1297 VM_PAGEMD_EXECPAGE); 1298 PMAP_COUNT(exec_synced_mappings); 1299 } else { 1300 UVMHIST_LOG(*histp, "va=%#"PRIxVADDR 1301 " pg %p: %s syncicache: pte %#x", 1302 va, pg, "defer", npte); 1303 } 1304 } else { 1305 UVMHIST_LOG(*histp, 1306 "va=%#"PRIxVADDR" pg %p: %s syncicache%s", 1307 va, pg, "no", 1308 (pte_cached_p(npte) 1309 ? " (already exec)" 1310 : " (uncached)")); 1311 } 1312 } else if (pg != NULL && (prot & VM_PROT_EXECUTE)) { 1313 KASSERT(mdpg != NULL); 1314 KASSERT(prot & VM_PROT_WRITE); 1315 PMAP_COUNT(exec_mappings); 1316 pmap_page_syncicache(pg); 1317 pmap_page_clear_attributes(mdpg, VM_PAGEMD_EXECPAGE); 1318 UVMHIST_LOG(*histp, 1319 "va=%#"PRIxVADDR" pg %p: %s syncicache%s", 1320 va, pg, "immediate", " (writeable)"); 1321 } 1322 1323 UVMHIST_LOG(*histp, " <-- 0 (OK)", 0, 0, 0, 0); 1324 return 0; 1325 } 1326 1327 void 1328 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags) 1329 { 1330 pmap_t pmap = pmap_kernel(); 1331 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 1332 struct vm_page_md * const mdpg = (pg ? VM_PAGE_TO_MD(pg) : NULL); 1333 1334 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1335 UVMHIST_LOG(pmaphist, 1336 "(va=%#"PRIxVADDR", pa=%#"PRIxPADDR", prot=%u, flags=%#x)", 1337 va, pa, prot, flags); 1338 PMAP_COUNT(kenter_pa); 1339 1340 if (mdpg == NULL) { 1341 PMAP_COUNT(kenter_pa_unmanaged); 1342 if ((flags & PMAP_CACHE_MASK) == 0) 1343 flags |= PMAP_NOCACHE; 1344 } else { 1345 if ((flags & PMAP_NOCACHE) == 0 && !PMAP_PAGE_COLOROK_P(pa, va)) 1346 PMAP_COUNT(kenter_pa_bad); 1347 } 1348 1349 pt_entry_t npte = pte_make_kenter_pa(pa, mdpg, prot, flags); 1350 kpreempt_disable(); 1351 pt_entry_t * const ptep = pmap_pte_lookup(pmap, va); 1352 KASSERTMSG(ptep != NULL, "%#"PRIxVADDR " %#"PRIxVADDR, va, 1353 pmap_limits.virtual_end); 1354 KASSERT(!pte_valid_p(*ptep)); 1355 1356 /* 1357 * No need to track non-managed pages or PMAP_KMPAGEs pages for aliases 1358 */ 1359 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1360 if (pg != NULL && (flags & PMAP_KMPAGE) == 0 1361 && pmap_md_virtual_cache_aliasing_p()) { 1362 pmap_enter_pv(pmap, va, pg, &npte, PV_KENTER); 1363 } 1364 #endif 1365 1366 /* 1367 * We have the option to force this mapping into the TLB but we 1368 * don't. Instead let the next reference to the page do it. 1369 */ 1370 pmap_md_tlb_miss_lock_enter(); 1371 *ptep = npte; 1372 pmap_tlb_update_addr(pmap_kernel(), va, npte, 0); 1373 pmap_md_tlb_miss_lock_exit(); 1374 kpreempt_enable(); 1375 #if DEBUG > 1 1376 for (u_int i = 0; i < PAGE_SIZE / sizeof(long); i++) { 1377 if (((long *)va)[i] != ((long *)pa)[i]) 1378 panic("%s: contents (%lx) of va %#"PRIxVADDR 1379 " != contents (%lx) of pa %#"PRIxPADDR, __func__, 1380 ((long *)va)[i], va, ((long *)pa)[i], pa); 1381 } 1382 #endif 1383 1384 UVMHIST_LOG(pmaphist, " <-- done (ptep=%p)", ptep, 0, 0, 0); 1385 } 1386 1387 /* 1388 * Remove the given range of addresses from the kernel map. 1389 * 1390 * It is assumed that the start and end are properly 1391 * rounded to the page size. 1392 */ 1393 1394 static bool 1395 pmap_pte_kremove(pmap_t pmap, vaddr_t sva, vaddr_t eva, pt_entry_t *ptep, 1396 uintptr_t flags) 1397 { 1398 const pt_entry_t new_pte = pte_nv_entry(true); 1399 1400 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1401 UVMHIST_LOG(pmaphist, 1402 "(pmap=%p, sva=%#"PRIxVADDR", eva=%#"PRIxVADDR", ptep=%p)", 1403 pmap, sva, eva, ptep); 1404 1405 KASSERT(kpreempt_disabled()); 1406 1407 for (; sva < eva; sva += NBPG, ptep++) { 1408 pt_entry_t pte = *ptep; 1409 if (!pte_valid_p(pte)) 1410 continue; 1411 1412 PMAP_COUNT(kremove_pages); 1413 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1414 struct vm_page * const pg = PHYS_TO_VM_PAGE(pte_to_paddr(pte)); 1415 if (pg != NULL && pmap_md_virtual_cache_aliasing_p()) { 1416 pmap_remove_pv(pmap, sva, pg, !pte_readonly_p(pte)); 1417 } 1418 #endif 1419 1420 pmap_md_tlb_miss_lock_enter(); 1421 *ptep = new_pte; 1422 pmap_tlb_invalidate_addr(pmap, sva); 1423 pmap_md_tlb_miss_lock_exit(); 1424 } 1425 1426 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1427 1428 return false; 1429 } 1430 1431 void 1432 pmap_kremove(vaddr_t va, vsize_t len) 1433 { 1434 const vaddr_t sva = trunc_page(va); 1435 const vaddr_t eva = round_page(va + len); 1436 1437 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1438 UVMHIST_LOG(pmaphist, "(va=%#"PRIxVADDR", len=%#"PRIxVSIZE")", 1439 va, len, 0, 0); 1440 1441 kpreempt_disable(); 1442 pmap_pte_process(pmap_kernel(), sva, eva, pmap_pte_kremove, 0); 1443 kpreempt_enable(); 1444 1445 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1446 } 1447 1448 void 1449 pmap_remove_all(struct pmap *pmap) 1450 { 1451 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1452 UVMHIST_LOG(pmaphist, "(pm=%p)", pmap, 0, 0, 0); 1453 1454 KASSERT(pmap != pmap_kernel()); 1455 1456 kpreempt_disable(); 1457 /* 1458 * Free all of our ASIDs which means we can skip doing all the 1459 * tlb_invalidate_addrs(). 1460 */ 1461 pmap_md_tlb_miss_lock_enter(); 1462 #ifdef MULTIPROCESSOR 1463 // This should be the last CPU with this pmap onproc 1464 KASSERT(!kcpuset_isotherset(pmap->pm_onproc, cpu_index(curcpu()))); 1465 if (kcpuset_isset(pmap->pm_onproc, cpu_index(curcpu()))) 1466 #endif 1467 pmap_tlb_asid_deactivate(pmap); 1468 #ifdef MULTIPROCESSOR 1469 KASSERT(kcpuset_iszero(pmap->pm_onproc)); 1470 #endif 1471 pmap_tlb_asid_release_all(pmap); 1472 pmap_md_tlb_miss_lock_exit(); 1473 pmap->pm_flags |= PMAP_DEFERRED_ACTIVATE; 1474 1475 #ifdef PMAP_FAULTINFO 1476 curpcb->pcb_faultinfo.pfi_faultaddr = 0; 1477 curpcb->pcb_faultinfo.pfi_repeats = 0; 1478 curpcb->pcb_faultinfo.pfi_faultpte = NULL; 1479 #endif 1480 kpreempt_enable(); 1481 1482 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1483 } 1484 1485 /* 1486 * Routine: pmap_unwire 1487 * Function: Clear the wired attribute for a map/virtual-address 1488 * pair. 1489 * In/out conditions: 1490 * The mapping must already exist in the pmap. 1491 */ 1492 void 1493 pmap_unwire(pmap_t pmap, vaddr_t va) 1494 { 1495 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1496 UVMHIST_LOG(pmaphist, "(pmap=%p, va=%#"PRIxVADDR")", pmap, va, 0, 0); 1497 PMAP_COUNT(unwire); 1498 1499 /* 1500 * Don't need to flush the TLB since PG_WIRED is only in software. 1501 */ 1502 kpreempt_disable(); 1503 pmap_addr_range_check(pmap, va, va, __func__); 1504 pt_entry_t * const ptep = pmap_pte_lookup(pmap, va); 1505 KASSERTMSG(ptep != NULL, "pmap %p va %#"PRIxVADDR" invalid STE", 1506 pmap, va); 1507 pt_entry_t pte = *ptep; 1508 KASSERTMSG(pte_valid_p(pte), 1509 "pmap %p va %#"PRIxVADDR" invalid PTE %#"PRIxPTE" @ %p", 1510 pmap, va, pte_value(pte), ptep); 1511 1512 if (pte_wired_p(pte)) { 1513 pmap_md_tlb_miss_lock_enter(); 1514 *ptep = pte_unwire_entry(pte); 1515 pmap_md_tlb_miss_lock_exit(); 1516 pmap->pm_stats.wired_count--; 1517 } 1518 #ifdef DIAGNOSTIC 1519 else { 1520 printf("%s: wiring for pmap %p va %#"PRIxVADDR" unchanged!\n", 1521 __func__, pmap, va); 1522 } 1523 #endif 1524 kpreempt_enable(); 1525 1526 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1527 } 1528 1529 /* 1530 * Routine: pmap_extract 1531 * Function: 1532 * Extract the physical page address associated 1533 * with the given map/virtual_address pair. 1534 */ 1535 bool 1536 pmap_extract(pmap_t pmap, vaddr_t va, paddr_t *pap) 1537 { 1538 paddr_t pa; 1539 1540 if (pmap == pmap_kernel()) { 1541 if (pmap_md_direct_mapped_vaddr_p(va)) { 1542 pa = pmap_md_direct_mapped_vaddr_to_paddr(va); 1543 goto done; 1544 } 1545 if (pmap_md_io_vaddr_p(va)) 1546 panic("pmap_extract: io address %#"PRIxVADDR"", va); 1547 1548 if (va >= pmap_limits.virtual_end) 1549 panic("%s: illegal kernel mapped address %#"PRIxVADDR, 1550 __func__, va); 1551 } 1552 kpreempt_disable(); 1553 const pt_entry_t * const ptep = pmap_pte_lookup(pmap, va); 1554 if (ptep == NULL || !pte_valid_p(*ptep)) { 1555 kpreempt_enable(); 1556 return false; 1557 } 1558 pa = pte_to_paddr(*ptep) | (va & PGOFSET); 1559 kpreempt_enable(); 1560 done: 1561 if (pap != NULL) { 1562 *pap = pa; 1563 } 1564 return true; 1565 } 1566 1567 /* 1568 * Copy the range specified by src_addr/len 1569 * from the source map to the range dst_addr/len 1570 * in the destination map. 1571 * 1572 * This routine is only advisory and need not do anything. 1573 */ 1574 void 1575 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vaddr_t dst_addr, vsize_t len, 1576 vaddr_t src_addr) 1577 { 1578 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1579 PMAP_COUNT(copy); 1580 } 1581 1582 /* 1583 * pmap_clear_reference: 1584 * 1585 * Clear the reference bit on the specified physical page. 1586 */ 1587 bool 1588 pmap_clear_reference(struct vm_page *pg) 1589 { 1590 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1591 1592 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1593 UVMHIST_LOG(pmaphist, "(pg=%p (pa %#"PRIxPADDR"))", 1594 pg, VM_PAGE_TO_PHYS(pg), 0,0); 1595 1596 bool rv = pmap_page_clear_attributes(mdpg, VM_PAGEMD_REFERENCED); 1597 1598 UVMHIST_LOG(pmaphist, " <-- %s", rv ? "true" : "false", 0, 0, 0); 1599 1600 return rv; 1601 } 1602 1603 /* 1604 * pmap_is_referenced: 1605 * 1606 * Return whether or not the specified physical page is referenced 1607 * by any physical maps. 1608 */ 1609 bool 1610 pmap_is_referenced(struct vm_page *pg) 1611 { 1612 return VM_PAGEMD_REFERENCED_P(VM_PAGE_TO_MD(pg)); 1613 } 1614 1615 /* 1616 * Clear the modify bits on the specified physical page. 1617 */ 1618 bool 1619 pmap_clear_modify(struct vm_page *pg) 1620 { 1621 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1622 pv_entry_t pv = &mdpg->mdpg_first; 1623 pv_entry_t pv_next; 1624 1625 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1626 UVMHIST_LOG(pmaphist, "(pg=%p (%#"PRIxPADDR"))", 1627 pg, VM_PAGE_TO_PHYS(pg), 0,0); 1628 PMAP_COUNT(clear_modify); 1629 1630 if (VM_PAGEMD_EXECPAGE_P(mdpg)) { 1631 if (pv->pv_pmap == NULL) { 1632 UVMHIST_LOG(pmapexechist, 1633 "pg %p (pa %#"PRIxPADDR"): %s", 1634 pg, VM_PAGE_TO_PHYS(pg), "execpage cleared", 0); 1635 pmap_page_clear_attributes(mdpg, VM_PAGEMD_EXECPAGE); 1636 PMAP_COUNT(exec_uncached_clear_modify); 1637 } else { 1638 UVMHIST_LOG(pmapexechist, 1639 "pg %p (pa %#"PRIxPADDR"): %s", 1640 pg, VM_PAGE_TO_PHYS(pg), "syncicache performed", 0); 1641 pmap_page_syncicache(pg); 1642 PMAP_COUNT(exec_synced_clear_modify); 1643 } 1644 } 1645 if (!pmap_page_clear_attributes(mdpg, VM_PAGEMD_MODIFIED)) { 1646 UVMHIST_LOG(pmaphist, " <-- false", 0, 0, 0, 0); 1647 return false; 1648 } 1649 if (pv->pv_pmap == NULL) { 1650 UVMHIST_LOG(pmaphist, " <-- true (no mappings)", 0, 0, 0, 0); 1651 return true; 1652 } 1653 1654 /* 1655 * remove write access from any pages that are dirty 1656 * so we can tell if they are written to again later. 1657 * flush the VAC first if there is one. 1658 */ 1659 kpreempt_disable(); 1660 KASSERT(!VM_PAGEMD_PVLIST_LOCKED_P(mdpg)); 1661 VM_PAGEMD_PVLIST_READLOCK(mdpg); 1662 pmap_pvlist_check(mdpg); 1663 for (; pv != NULL; pv = pv_next) { 1664 pmap_t pmap = pv->pv_pmap; 1665 vaddr_t va = trunc_page(pv->pv_va); 1666 1667 pv_next = pv->pv_next; 1668 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1669 if (pv->pv_va & PV_KENTER) 1670 continue; 1671 #endif 1672 pt_entry_t * const ptep = pmap_pte_lookup(pmap, va); 1673 KASSERT(ptep); 1674 pt_entry_t pte = pte_prot_nowrite(*ptep); 1675 if (*ptep == pte) { 1676 continue; 1677 } 1678 KASSERT(pte_valid_p(pte)); 1679 const uintptr_t gen = VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1680 pmap_md_tlb_miss_lock_enter(); 1681 *ptep = pte; 1682 pmap_tlb_invalidate_addr(pmap, va); 1683 pmap_md_tlb_miss_lock_exit(); 1684 pmap_update(pmap); 1685 if (__predict_false(gen != VM_PAGEMD_PVLIST_READLOCK(mdpg))) { 1686 /* 1687 * The list changed! So restart from the beginning. 1688 */ 1689 pv_next = &mdpg->mdpg_first; 1690 pmap_pvlist_check(mdpg); 1691 } 1692 } 1693 pmap_pvlist_check(mdpg); 1694 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1695 kpreempt_enable(); 1696 1697 UVMHIST_LOG(pmaphist, " <-- true (mappings changed)", 0, 0, 0, 0); 1698 return true; 1699 } 1700 1701 /* 1702 * pmap_is_modified: 1703 * 1704 * Return whether or not the specified physical page is modified 1705 * by any physical maps. 1706 */ 1707 bool 1708 pmap_is_modified(struct vm_page *pg) 1709 { 1710 return VM_PAGEMD_MODIFIED_P(VM_PAGE_TO_MD(pg)); 1711 } 1712 1713 /* 1714 * pmap_set_modified: 1715 * 1716 * Sets the page modified reference bit for the specified page. 1717 */ 1718 void 1719 pmap_set_modified(paddr_t pa) 1720 { 1721 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 1722 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1723 pmap_page_set_attributes(mdpg, VM_PAGEMD_MODIFIED|VM_PAGEMD_REFERENCED); 1724 } 1725 1726 /******************** pv_entry management ********************/ 1727 1728 static void 1729 pmap_pvlist_check(struct vm_page_md *mdpg) 1730 { 1731 #ifdef DEBUG 1732 pv_entry_t pv = &mdpg->mdpg_first; 1733 if (pv->pv_pmap != NULL) { 1734 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1735 const u_int colormask = uvmexp.colormask; 1736 u_int colors = 0; 1737 #endif 1738 for (; pv != NULL; pv = pv->pv_next) { 1739 KASSERT(pv->pv_pmap != pmap_kernel() || !pmap_md_direct_mapped_vaddr_p(pv->pv_va)); 1740 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1741 colors |= __BIT(atop(pv->pv_va) & colormask); 1742 #endif 1743 } 1744 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1745 // Assert there if there more than 1 color mapped, that they 1746 // are uncached. 1747 KASSERTMSG(!pmap_md_virtual_cache_aliasing_p() 1748 || colors == 0 || (colors & (colors-1)) == 0 1749 || VM_PAGEMD_UNCACHED_P(mdpg), "colors=%#x uncached=%u", 1750 colors, VM_PAGEMD_UNCACHED_P(mdpg)); 1751 #endif 1752 } 1753 #endif /* DEBUG */ 1754 } 1755 1756 /* 1757 * Enter the pmap and virtual address into the 1758 * physical to virtual map table. 1759 */ 1760 void 1761 pmap_enter_pv(pmap_t pmap, vaddr_t va, struct vm_page *pg, pt_entry_t *nptep, 1762 u_int flags) 1763 { 1764 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1765 pv_entry_t pv, npv, apv; 1766 #ifdef UVMHIST 1767 bool first = false; 1768 #endif 1769 1770 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1771 UVMHIST_LOG(pmaphist, 1772 "(pmap=%p va=%#"PRIxVADDR" pg=%p (%#"PRIxPADDR")", 1773 pmap, va, pg, VM_PAGE_TO_PHYS(pg)); 1774 UVMHIST_LOG(pmaphist, "nptep=%p (%#"PRIxPTE"))", 1775 nptep, pte_value(*nptep), 0, 0); 1776 1777 KASSERT(kpreempt_disabled()); 1778 KASSERT(pmap != pmap_kernel() || !pmap_md_direct_mapped_vaddr_p(va)); 1779 KASSERTMSG(pmap != pmap_kernel() || !pmap_md_io_vaddr_p(va), 1780 "va %#"PRIxVADDR, va); 1781 1782 apv = NULL; 1783 VM_PAGEMD_PVLIST_LOCK(mdpg); 1784 again: 1785 pv = &mdpg->mdpg_first; 1786 pmap_pvlist_check(mdpg); 1787 if (pv->pv_pmap == NULL) { 1788 KASSERT(pv->pv_next == NULL); 1789 /* 1790 * No entries yet, use header as the first entry 1791 */ 1792 PMAP_COUNT(primary_mappings); 1793 PMAP_COUNT(mappings); 1794 #ifdef UVMHIST 1795 first = true; 1796 #endif 1797 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1798 KASSERT(VM_PAGEMD_CACHED_P(mdpg)); 1799 // If the new mapping has an incompatible color the last 1800 // mapping of this page, clean the page before using it. 1801 if (!PMAP_PAGE_COLOROK_P(va, pv->pv_va)) { 1802 pmap_md_vca_clean(pg, PMAP_WBINV); 1803 } 1804 #endif 1805 pv->pv_pmap = pmap; 1806 pv->pv_va = va | flags; 1807 } else { 1808 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1809 if (pmap_md_vca_add(pg, va, nptep)) { 1810 goto again; 1811 } 1812 #endif 1813 1814 /* 1815 * There is at least one other VA mapping this page. 1816 * Place this entry after the header. 1817 * 1818 * Note: the entry may already be in the table if 1819 * we are only changing the protection bits. 1820 */ 1821 1822 #ifdef PARANOIADIAG 1823 const paddr_t pa = VM_PAGE_TO_PHYS(pg); 1824 #endif 1825 for (npv = pv; npv; npv = npv->pv_next) { 1826 if (pmap == npv->pv_pmap 1827 && va == trunc_page(npv->pv_va)) { 1828 #ifdef PARANOIADIAG 1829 pt_entry_t *ptep = pmap_pte_lookup(pmap, va); 1830 pt_entry_t pte = (ptep != NULL) ? *ptep : 0; 1831 if (!pte_valid_p(pte) || pte_to_paddr(pte) != pa) 1832 printf("%s: found va %#"PRIxVADDR 1833 " pa %#"PRIxPADDR 1834 " in pv_table but != %#"PRIxPTE"\n", 1835 __func__, va, pa, pte_value(pte)); 1836 #endif 1837 PMAP_COUNT(remappings); 1838 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1839 if (__predict_false(apv != NULL)) 1840 pmap_pv_free(apv); 1841 1842 UVMHIST_LOG(pmaphist, " <-- done pv=%p%s", 1843 pv, " (reused)", 0, 0); 1844 return; 1845 } 1846 } 1847 if (__predict_true(apv == NULL)) { 1848 /* 1849 * To allocate a PV, we have to release the PVLIST lock 1850 * so get the page generation. We allocate the PV, and 1851 * then reacquire the lock. 1852 */ 1853 pmap_pvlist_check(mdpg); 1854 const uintptr_t gen = VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1855 1856 apv = (pv_entry_t)pmap_pv_alloc(); 1857 if (apv == NULL) 1858 panic("pmap_enter_pv: pmap_pv_alloc() failed"); 1859 1860 /* 1861 * If the generation has changed, then someone else 1862 * tinkered with this page so we should start over. 1863 */ 1864 if (gen != VM_PAGEMD_PVLIST_LOCK(mdpg)) 1865 goto again; 1866 } 1867 npv = apv; 1868 apv = NULL; 1869 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1870 /* 1871 * If need to deal with virtual cache aliases, keep mappings 1872 * in the kernel pmap at the head of the list. This allows 1873 * the VCA code to easily use them for cache operations if 1874 * present. 1875 */ 1876 pmap_t kpmap = pmap_kernel(); 1877 if (pmap != kpmap) { 1878 while (pv->pv_pmap == kpmap && pv->pv_next != NULL) { 1879 pv = pv->pv_next; 1880 } 1881 } 1882 #endif 1883 npv->pv_va = va | flags; 1884 npv->pv_pmap = pmap; 1885 npv->pv_next = pv->pv_next; 1886 pv->pv_next = npv; 1887 PMAP_COUNT(mappings); 1888 } 1889 pmap_pvlist_check(mdpg); 1890 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1891 if (__predict_false(apv != NULL)) 1892 pmap_pv_free(apv); 1893 1894 UVMHIST_LOG(pmaphist, " <-- done pv=%p%s", 1895 pv, first ? " (first pv)" : "",0,0); 1896 } 1897 1898 /* 1899 * Remove a physical to virtual address translation. 1900 * If cache was inhibited on this page, and there are no more cache 1901 * conflicts, restore caching. 1902 * Flush the cache if the last page is removed (should always be cached 1903 * at this point). 1904 */ 1905 void 1906 pmap_remove_pv(pmap_t pmap, vaddr_t va, struct vm_page *pg, bool dirty) 1907 { 1908 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 1909 pv_entry_t pv, npv; 1910 bool last; 1911 1912 UVMHIST_FUNC(__func__); UVMHIST_CALLED(pmaphist); 1913 UVMHIST_LOG(pmaphist, 1914 "(pmap=%p, va=%#"PRIxVADDR", pg=%p (pa %#"PRIxPADDR")", 1915 pmap, va, pg, VM_PAGE_TO_PHYS(pg)); 1916 UVMHIST_LOG(pmaphist, "dirty=%s)", dirty ? "true" : "false", 0, 0, 0); 1917 1918 KASSERT(kpreempt_disabled()); 1919 KASSERT((va & PAGE_MASK) == 0); 1920 pv = &mdpg->mdpg_first; 1921 1922 VM_PAGEMD_PVLIST_LOCK(mdpg); 1923 pmap_pvlist_check(mdpg); 1924 1925 /* 1926 * If it is the first entry on the list, it is actually 1927 * in the header and we must copy the following entry up 1928 * to the header. Otherwise we must search the list for 1929 * the entry. In either case we free the now unused entry. 1930 */ 1931 1932 last = false; 1933 if (pmap == pv->pv_pmap && va == trunc_page(pv->pv_va)) { 1934 npv = pv->pv_next; 1935 if (npv) { 1936 *pv = *npv; 1937 KASSERT(pv->pv_pmap != NULL); 1938 } else { 1939 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1940 pmap_page_clear_attributes(mdpg, VM_PAGEMD_UNCACHED); 1941 #endif 1942 pv->pv_pmap = NULL; 1943 last = true; /* Last mapping removed */ 1944 } 1945 PMAP_COUNT(remove_pvfirst); 1946 } else { 1947 for (npv = pv->pv_next; npv; pv = npv, npv = npv->pv_next) { 1948 PMAP_COUNT(remove_pvsearch); 1949 if (pmap == npv->pv_pmap && va == trunc_page(npv->pv_va)) 1950 break; 1951 } 1952 if (npv) { 1953 pv->pv_next = npv->pv_next; 1954 } 1955 } 1956 1957 pmap_pvlist_check(mdpg); 1958 VM_PAGEMD_PVLIST_UNLOCK(mdpg); 1959 1960 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 1961 pmap_md_vca_remove(pg, va, dirty, last); 1962 #endif 1963 1964 /* 1965 * Free the pv_entry if needed. 1966 */ 1967 if (npv) 1968 pmap_pv_free(npv); 1969 if (VM_PAGEMD_EXECPAGE_P(mdpg) && dirty) { 1970 if (last) { 1971 /* 1972 * If this was the page's last mapping, we no longer 1973 * care about its execness. 1974 */ 1975 UVMHIST_LOG(pmapexechist, 1976 "pg %p (pa %#"PRIxPADDR")%s: %s", 1977 pg, VM_PAGE_TO_PHYS(pg), 1978 last ? " [last mapping]" : "", 1979 "execpage cleared"); 1980 pmap_page_clear_attributes(mdpg, VM_PAGEMD_EXECPAGE); 1981 PMAP_COUNT(exec_uncached_remove); 1982 } else { 1983 /* 1984 * Someone still has it mapped as an executable page 1985 * so we must sync it. 1986 */ 1987 UVMHIST_LOG(pmapexechist, 1988 "pg %p (pa %#"PRIxPADDR")%s: %s", 1989 pg, VM_PAGE_TO_PHYS(pg), 1990 last ? " [last mapping]" : "", 1991 "performed syncicache"); 1992 pmap_page_syncicache(pg); 1993 PMAP_COUNT(exec_synced_remove); 1994 } 1995 } 1996 1997 UVMHIST_LOG(pmaphist, " <-- done", 0, 0, 0, 0); 1998 } 1999 2000 #if defined(MULTIPROCESSOR) 2001 struct pmap_pvlist_info { 2002 kmutex_t *pli_locks[PAGE_SIZE / 32]; 2003 volatile u_int pli_lock_refs[PAGE_SIZE / 32]; 2004 volatile u_int pli_lock_index; 2005 u_int pli_lock_mask; 2006 } pmap_pvlist_info; 2007 2008 void 2009 pmap_pvlist_lock_init(size_t cache_line_size) 2010 { 2011 struct pmap_pvlist_info * const pli = &pmap_pvlist_info; 2012 const vaddr_t lock_page = uvm_pageboot_alloc(PAGE_SIZE); 2013 vaddr_t lock_va = lock_page; 2014 if (sizeof(kmutex_t) > cache_line_size) { 2015 cache_line_size = roundup2(sizeof(kmutex_t), cache_line_size); 2016 } 2017 const size_t nlocks = PAGE_SIZE / cache_line_size; 2018 KASSERT((nlocks & (nlocks - 1)) == 0); 2019 /* 2020 * Now divide the page into a number of mutexes, one per cacheline. 2021 */ 2022 for (size_t i = 0; i < nlocks; lock_va += cache_line_size, i++) { 2023 kmutex_t * const lock = (kmutex_t *)lock_va; 2024 mutex_init(lock, MUTEX_DEFAULT, IPL_HIGH); 2025 pli->pli_locks[i] = lock; 2026 } 2027 pli->pli_lock_mask = nlocks - 1; 2028 } 2029 2030 kmutex_t * 2031 pmap_pvlist_lock_addr(struct vm_page_md *mdpg) 2032 { 2033 struct pmap_pvlist_info * const pli = &pmap_pvlist_info; 2034 kmutex_t *lock = mdpg->mdpg_lock; 2035 2036 /* 2037 * Allocate a lock on an as-needed basis. This will hopefully give us 2038 * semi-random distribution not based on page color. 2039 */ 2040 if (__predict_false(lock == NULL)) { 2041 size_t locknum = atomic_add_int_nv(&pli->pli_lock_index, 37); 2042 size_t lockid = locknum & pli->pli_lock_mask; 2043 kmutex_t * const new_lock = pli->pli_locks[lockid]; 2044 /* 2045 * Set the lock. If some other thread already did, just use 2046 * the one they assigned. 2047 */ 2048 lock = atomic_cas_ptr(&mdpg->mdpg_lock, NULL, new_lock); 2049 if (lock == NULL) { 2050 lock = new_lock; 2051 atomic_inc_uint(&pli->pli_lock_refs[lockid]); 2052 } 2053 } 2054 2055 /* 2056 * Now finally provide the lock. 2057 */ 2058 return lock; 2059 } 2060 #else /* !MULTIPROCESSOR */ 2061 void 2062 pmap_pvlist_lock_init(size_t cache_line_size) 2063 { 2064 mutex_init(&pmap_pvlist_mutex, MUTEX_DEFAULT, IPL_HIGH); 2065 } 2066 2067 #ifdef MODULAR 2068 kmutex_t * 2069 pmap_pvlist_lock_addr(struct vm_page_md *mdpg) 2070 { 2071 /* 2072 * We just use a global lock. 2073 */ 2074 if (__predict_false(mdpg->mdpg_lock == NULL)) { 2075 mdpg->mdpg_lock = &pmap_pvlist_mutex; 2076 } 2077 2078 /* 2079 * Now finally provide the lock. 2080 */ 2081 return mdpg->mdpg_lock; 2082 } 2083 #endif /* MODULAR */ 2084 #endif /* !MULTIPROCESSOR */ 2085 2086 /* 2087 * pmap_pv_page_alloc: 2088 * 2089 * Allocate a page for the pv_entry pool. 2090 */ 2091 void * 2092 pmap_pv_page_alloc(struct pool *pp, int flags) 2093 { 2094 struct vm_page * const pg = PMAP_ALLOC_POOLPAGE(UVM_PGA_USERESERVE); 2095 if (pg == NULL) 2096 return NULL; 2097 2098 return (void *)pmap_map_poolpage(VM_PAGE_TO_PHYS(pg)); 2099 } 2100 2101 /* 2102 * pmap_pv_page_free: 2103 * 2104 * Free a pv_entry pool page. 2105 */ 2106 void 2107 pmap_pv_page_free(struct pool *pp, void *v) 2108 { 2109 vaddr_t va = (vaddr_t)v; 2110 2111 KASSERT(pmap_md_direct_mapped_vaddr_p(va)); 2112 const paddr_t pa = pmap_md_direct_mapped_vaddr_to_paddr(va); 2113 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 2114 KASSERT(pg != NULL); 2115 #ifdef PMAP_VIRTUAL_CACHE_ALIASES 2116 kpreempt_disable(); 2117 pmap_md_vca_remove(pg, va, true, true); 2118 kpreempt_enable(); 2119 #endif 2120 pmap_page_clear_attributes(VM_PAGE_TO_MD(pg), VM_PAGEMD_POOLPAGE); 2121 uvm_pagefree(pg); 2122 } 2123 2124 #ifdef PMAP_PREFER 2125 /* 2126 * Find first virtual address >= *vap that doesn't cause 2127 * a cache alias conflict. 2128 */ 2129 void 2130 pmap_prefer(vaddr_t foff, vaddr_t *vap, vsize_t sz, int td) 2131 { 2132 vsize_t prefer_mask = ptoa(uvmexp.colormask); 2133 2134 PMAP_COUNT(prefer_requests); 2135 2136 prefer_mask |= pmap_md_cache_prefer_mask(); 2137 2138 if (prefer_mask) { 2139 vaddr_t va = *vap; 2140 vsize_t d = (foff - va) & prefer_mask; 2141 if (d) { 2142 if (td) 2143 *vap = trunc_page(va - ((-d) & prefer_mask)); 2144 else 2145 *vap = round_page(va + d); 2146 PMAP_COUNT(prefer_adjustments); 2147 } 2148 } 2149 } 2150 #endif /* PMAP_PREFER */ 2151 2152 #ifdef PMAP_MAP_POOLPAGE 2153 vaddr_t 2154 pmap_map_poolpage(paddr_t pa) 2155 { 2156 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 2157 KASSERT(pg); 2158 struct vm_page_md * const mdpg = VM_PAGE_TO_MD(pg); 2159 pmap_page_set_attributes(mdpg, VM_PAGEMD_POOLPAGE); 2160 2161 return pmap_md_map_poolpage(pa, NBPG); 2162 } 2163 2164 paddr_t 2165 pmap_unmap_poolpage(vaddr_t va) 2166 { 2167 KASSERT(pmap_md_direct_mapped_vaddr_p(va)); 2168 paddr_t pa = pmap_md_direct_mapped_vaddr_to_paddr(va); 2169 2170 struct vm_page * const pg = PHYS_TO_VM_PAGE(pa); 2171 KASSERT(pg != NULL); 2172 pmap_page_clear_attributes(VM_PAGE_TO_MD(pg), VM_PAGEMD_POOLPAGE); 2173 pmap_md_unmap_poolpage(va, NBPG); 2174 2175 return pa; 2176 } 2177 #endif /* PMAP_MAP_POOLPAGE */ 2178