1 /* $NetBSD: uvm_aobj.c,v 1.156 2022/05/31 08:43:16 andvar Exp $ */ 2 3 /* 4 * Copyright (c) 1998 Chuck Silvers, Charles D. Cranor and 5 * Washington University. 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 * 28 * from: Id: uvm_aobj.c,v 1.1.2.5 1998/02/06 05:14:38 chs Exp 29 */ 30 31 /* 32 * uvm_aobj.c: anonymous memory uvm_object pager 33 * 34 * author: Chuck Silvers <chuq@chuq.com> 35 * started: Jan-1998 36 * 37 * - design mostly from Chuck Cranor 38 */ 39 40 #include <sys/cdefs.h> 41 __KERNEL_RCSID(0, "$NetBSD: uvm_aobj.c,v 1.156 2022/05/31 08:43:16 andvar Exp $"); 42 43 #ifdef _KERNEL_OPT 44 #include "opt_uvmhist.h" 45 #endif 46 47 #include <sys/param.h> 48 #include <sys/systm.h> 49 #include <sys/kernel.h> 50 #include <sys/kmem.h> 51 #include <sys/pool.h> 52 #include <sys/atomic.h> 53 54 #include <uvm/uvm.h> 55 #include <uvm/uvm_page_array.h> 56 57 /* 58 * An anonymous UVM object (aobj) manages anonymous-memory. In addition to 59 * keeping the list of resident pages, it may also keep a list of allocated 60 * swap blocks. Depending on the size of the object, this list is either 61 * stored in an array (small objects) or in a hash table (large objects). 62 * 63 * Lock order 64 * 65 * uao_list_lock -> 66 * uvm_object::vmobjlock 67 */ 68 69 /* 70 * Note: for hash tables, we break the address space of the aobj into blocks 71 * of UAO_SWHASH_CLUSTER_SIZE pages, which shall be a power of two. 72 */ 73 74 #define UAO_SWHASH_CLUSTER_SHIFT 4 75 #define UAO_SWHASH_CLUSTER_SIZE (1 << UAO_SWHASH_CLUSTER_SHIFT) 76 77 /* Get the "tag" for this page index. */ 78 #define UAO_SWHASH_ELT_TAG(idx) ((idx) >> UAO_SWHASH_CLUSTER_SHIFT) 79 #define UAO_SWHASH_ELT_PAGESLOT_IDX(idx) \ 80 ((idx) & (UAO_SWHASH_CLUSTER_SIZE - 1)) 81 82 /* Given an ELT and a page index, find the swap slot. */ 83 #define UAO_SWHASH_ELT_PAGESLOT(elt, idx) \ 84 ((elt)->slots[UAO_SWHASH_ELT_PAGESLOT_IDX(idx)]) 85 86 /* Given an ELT, return its pageidx base. */ 87 #define UAO_SWHASH_ELT_PAGEIDX_BASE(ELT) \ 88 ((elt)->tag << UAO_SWHASH_CLUSTER_SHIFT) 89 90 /* The hash function. */ 91 #define UAO_SWHASH_HASH(aobj, idx) \ 92 (&(aobj)->u_swhash[(((idx) >> UAO_SWHASH_CLUSTER_SHIFT) \ 93 & (aobj)->u_swhashmask)]) 94 95 /* 96 * The threshold which determines whether we will use an array or a 97 * hash table to store the list of allocated swap blocks. 98 */ 99 #define UAO_SWHASH_THRESHOLD (UAO_SWHASH_CLUSTER_SIZE * 4) 100 #define UAO_USES_SWHASH(aobj) \ 101 ((aobj)->u_pages > UAO_SWHASH_THRESHOLD) 102 103 /* The number of buckets in a hash, with an upper bound. */ 104 #define UAO_SWHASH_MAXBUCKETS 256 105 #define UAO_SWHASH_BUCKETS(aobj) \ 106 (MIN((aobj)->u_pages >> UAO_SWHASH_CLUSTER_SHIFT, UAO_SWHASH_MAXBUCKETS)) 107 108 /* 109 * uao_swhash_elt: when a hash table is being used, this structure defines 110 * the format of an entry in the bucket list. 111 */ 112 113 struct uao_swhash_elt { 114 LIST_ENTRY(uao_swhash_elt) list; /* the hash list */ 115 voff_t tag; /* our 'tag' */ 116 int count; /* our number of active slots */ 117 int slots[UAO_SWHASH_CLUSTER_SIZE]; /* the slots */ 118 }; 119 120 /* 121 * uao_swhash: the swap hash table structure 122 */ 123 124 LIST_HEAD(uao_swhash, uao_swhash_elt); 125 126 /* 127 * uao_swhash_elt_pool: pool of uao_swhash_elt structures. 128 * Note: pages for this pool must not come from a pageable kernel map. 129 */ 130 static struct pool uao_swhash_elt_pool __cacheline_aligned; 131 132 /* 133 * uvm_aobj: the actual anon-backed uvm_object 134 * 135 * => the uvm_object is at the top of the structure, this allows 136 * (struct uvm_aobj *) == (struct uvm_object *) 137 * => only one of u_swslots and u_swhash is used in any given aobj 138 */ 139 140 struct uvm_aobj { 141 struct uvm_object u_obj; /* has: lock, pgops, #pages, #refs */ 142 pgoff_t u_pages; /* number of pages in entire object */ 143 int u_flags; /* the flags (see uvm_aobj.h) */ 144 int *u_swslots; /* array of offset->swapslot mappings */ 145 /* 146 * hashtable of offset->swapslot mappings 147 * (u_swhash is an array of bucket heads) 148 */ 149 struct uao_swhash *u_swhash; 150 u_long u_swhashmask; /* mask for hashtable */ 151 LIST_ENTRY(uvm_aobj) u_list; /* global list of aobjs */ 152 int u_freelist; /* freelist to allocate pages from */ 153 }; 154 155 static void uao_free(struct uvm_aobj *); 156 static int uao_get(struct uvm_object *, voff_t, struct vm_page **, 157 int *, int, vm_prot_t, int, int); 158 static int uao_put(struct uvm_object *, voff_t, voff_t, int); 159 160 #if defined(VMSWAP) 161 static struct uao_swhash_elt *uao_find_swhash_elt 162 (struct uvm_aobj *, int, bool); 163 164 static bool uao_pagein(struct uvm_aobj *, int, int); 165 static bool uao_pagein_page(struct uvm_aobj *, int); 166 #endif /* defined(VMSWAP) */ 167 168 static struct vm_page *uao_pagealloc(struct uvm_object *, voff_t, int); 169 170 /* 171 * aobj_pager 172 * 173 * note that some functions (e.g. put) are handled elsewhere 174 */ 175 176 const struct uvm_pagerops aobj_pager = { 177 .pgo_reference = uao_reference, 178 .pgo_detach = uao_detach, 179 .pgo_get = uao_get, 180 .pgo_put = uao_put, 181 }; 182 183 /* 184 * uao_list: global list of active aobjs, locked by uao_list_lock 185 */ 186 187 static LIST_HEAD(aobjlist, uvm_aobj) uao_list __cacheline_aligned; 188 static kmutex_t uao_list_lock __cacheline_aligned; 189 190 /* 191 * hash table/array related functions 192 */ 193 194 #if defined(VMSWAP) 195 196 /* 197 * uao_find_swhash_elt: find (or create) a hash table entry for a page 198 * offset. 199 * 200 * => the object should be locked by the caller 201 */ 202 203 static struct uao_swhash_elt * 204 uao_find_swhash_elt(struct uvm_aobj *aobj, int pageidx, bool create) 205 { 206 struct uao_swhash *swhash; 207 struct uao_swhash_elt *elt; 208 voff_t page_tag; 209 210 swhash = UAO_SWHASH_HASH(aobj, pageidx); 211 page_tag = UAO_SWHASH_ELT_TAG(pageidx); 212 213 /* 214 * now search the bucket for the requested tag 215 */ 216 217 LIST_FOREACH(elt, swhash, list) { 218 if (elt->tag == page_tag) { 219 return elt; 220 } 221 } 222 if (!create) { 223 return NULL; 224 } 225 226 /* 227 * allocate a new entry for the bucket and init/insert it in 228 */ 229 230 elt = pool_get(&uao_swhash_elt_pool, PR_NOWAIT); 231 if (elt == NULL) { 232 return NULL; 233 } 234 LIST_INSERT_HEAD(swhash, elt, list); 235 elt->tag = page_tag; 236 elt->count = 0; 237 memset(elt->slots, 0, sizeof(elt->slots)); 238 return elt; 239 } 240 241 /* 242 * uao_find_swslot: find the swap slot number for an aobj/pageidx 243 * 244 * => object must be locked by caller 245 */ 246 247 int 248 uao_find_swslot(struct uvm_object *uobj, int pageidx) 249 { 250 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 251 struct uao_swhash_elt *elt; 252 253 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 254 255 /* 256 * if noswap flag is set, then we never return a slot 257 */ 258 259 if (aobj->u_flags & UAO_FLAG_NOSWAP) 260 return 0; 261 262 /* 263 * if hashing, look in hash table. 264 */ 265 266 if (UAO_USES_SWHASH(aobj)) { 267 elt = uao_find_swhash_elt(aobj, pageidx, false); 268 return elt ? UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) : 0; 269 } 270 271 /* 272 * otherwise, look in the array 273 */ 274 275 return aobj->u_swslots[pageidx]; 276 } 277 278 /* 279 * uao_set_swslot: set the swap slot for a page in an aobj. 280 * 281 * => setting a slot to zero frees the slot 282 * => object must be locked by caller 283 * => we return the old slot number, or -1 if we failed to allocate 284 * memory to record the new slot number 285 */ 286 287 int 288 uao_set_swslot(struct uvm_object *uobj, int pageidx, int slot) 289 { 290 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 291 struct uao_swhash_elt *elt; 292 int oldslot; 293 UVMHIST_FUNC(__func__); 294 UVMHIST_CALLARGS(pdhist, "aobj %#jx pageidx %jd slot %jd", 295 (uintptr_t)aobj, pageidx, slot, 0); 296 297 KASSERT(rw_write_held(uobj->vmobjlock) || uobj->uo_refs == 0); 298 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 299 300 /* 301 * if noswap flag is set, then we can't set a non-zero slot. 302 */ 303 304 if (aobj->u_flags & UAO_FLAG_NOSWAP) { 305 KASSERTMSG(slot == 0, "uao_set_swslot: no swap object"); 306 return 0; 307 } 308 309 /* 310 * are we using a hash table? if so, add it in the hash. 311 */ 312 313 if (UAO_USES_SWHASH(aobj)) { 314 315 /* 316 * Avoid allocating an entry just to free it again if 317 * the page had not swap slot in the first place, and 318 * we are freeing. 319 */ 320 321 elt = uao_find_swhash_elt(aobj, pageidx, slot != 0); 322 if (elt == NULL) { 323 return slot ? -1 : 0; 324 } 325 326 oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx); 327 UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot; 328 329 /* 330 * now adjust the elt's reference counter and free it if we've 331 * dropped it to zero. 332 */ 333 334 if (slot) { 335 if (oldslot == 0) 336 elt->count++; 337 } else { 338 if (oldslot) 339 elt->count--; 340 341 if (elt->count == 0) { 342 LIST_REMOVE(elt, list); 343 pool_put(&uao_swhash_elt_pool, elt); 344 } 345 } 346 } else { 347 /* we are using an array */ 348 oldslot = aobj->u_swslots[pageidx]; 349 aobj->u_swslots[pageidx] = slot; 350 } 351 return oldslot; 352 } 353 354 #endif /* defined(VMSWAP) */ 355 356 /* 357 * end of hash/array functions 358 */ 359 360 /* 361 * uao_free: free all resources held by an aobj, and then free the aobj 362 * 363 * => the aobj should be dead 364 */ 365 366 static void 367 uao_free(struct uvm_aobj *aobj) 368 { 369 struct uvm_object *uobj = &aobj->u_obj; 370 371 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 372 KASSERT(rw_write_held(uobj->vmobjlock)); 373 uao_dropswap_range(uobj, 0, 0); 374 rw_exit(uobj->vmobjlock); 375 376 #if defined(VMSWAP) 377 if (UAO_USES_SWHASH(aobj)) { 378 379 /* 380 * free the hash table itself. 381 */ 382 383 hashdone(aobj->u_swhash, HASH_LIST, aobj->u_swhashmask); 384 } else { 385 386 /* 387 * free the array itself. 388 */ 389 390 kmem_free(aobj->u_swslots, aobj->u_pages * sizeof(int)); 391 } 392 #endif /* defined(VMSWAP) */ 393 394 /* 395 * finally free the aobj itself 396 */ 397 398 uvm_obj_destroy(uobj, true); 399 kmem_free(aobj, sizeof(struct uvm_aobj)); 400 } 401 402 /* 403 * pager functions 404 */ 405 406 /* 407 * uao_create: create an aobj of the given size and return its uvm_object. 408 * 409 * => for normal use, flags are always zero 410 * => for the kernel object, the flags are: 411 * UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once) 412 * UAO_FLAG_KERNSWAP - enable swapping of kernel object (" ") 413 */ 414 415 struct uvm_object * 416 uao_create(voff_t size, int flags) 417 { 418 static struct uvm_aobj kernel_object_store; 419 static krwlock_t bootstrap_kernel_object_lock; 420 static int kobj_alloced __diagused = 0; 421 pgoff_t pages = round_page((uint64_t)size) >> PAGE_SHIFT; 422 struct uvm_aobj *aobj; 423 int refs; 424 425 /* 426 * Allocate a new aobj, unless kernel object is requested. 427 */ 428 429 if (flags & UAO_FLAG_KERNOBJ) { 430 KASSERT(!kobj_alloced); 431 aobj = &kernel_object_store; 432 aobj->u_pages = pages; 433 aobj->u_flags = UAO_FLAG_NOSWAP; 434 refs = UVM_OBJ_KERN; 435 kobj_alloced = UAO_FLAG_KERNOBJ; 436 } else if (flags & UAO_FLAG_KERNSWAP) { 437 KASSERT(kobj_alloced == UAO_FLAG_KERNOBJ); 438 aobj = &kernel_object_store; 439 kobj_alloced = UAO_FLAG_KERNSWAP; 440 refs = 0xdeadbeaf; /* XXX: gcc */ 441 } else { 442 aobj = kmem_alloc(sizeof(struct uvm_aobj), KM_SLEEP); 443 aobj->u_pages = pages; 444 aobj->u_flags = 0; 445 refs = 1; 446 } 447 448 /* 449 * no freelist by default 450 */ 451 452 aobj->u_freelist = VM_NFREELIST; 453 454 /* 455 * allocate hash/array if necessary 456 * 457 * note: in the KERNSWAP case no need to worry about locking since 458 * we are still booting we should be the only thread around. 459 */ 460 461 const int kernswap = (flags & UAO_FLAG_KERNSWAP) != 0; 462 if (flags == 0 || kernswap) { 463 #if defined(VMSWAP) 464 465 /* allocate hash table or array depending on object size */ 466 if (UAO_USES_SWHASH(aobj)) { 467 aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(aobj), 468 HASH_LIST, true, &aobj->u_swhashmask); 469 } else { 470 aobj->u_swslots = kmem_zalloc(pages * sizeof(int), 471 KM_SLEEP); 472 } 473 #endif /* defined(VMSWAP) */ 474 475 /* 476 * Replace kernel_object's temporary static lock with 477 * a regular rw_obj. We cannot use uvm_obj_setlock() 478 * because that would try to free the old lock. 479 */ 480 481 if (kernswap) { 482 aobj->u_obj.vmobjlock = rw_obj_alloc(); 483 rw_destroy(&bootstrap_kernel_object_lock); 484 } 485 if (flags) { 486 aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */ 487 return &aobj->u_obj; 488 } 489 } 490 491 /* 492 * Initialise UVM object. 493 */ 494 495 const bool kernobj = (flags & UAO_FLAG_KERNOBJ) != 0; 496 uvm_obj_init(&aobj->u_obj, &aobj_pager, !kernobj, refs); 497 if (__predict_false(kernobj)) { 498 /* Use a temporary static lock for kernel_object. */ 499 rw_init(&bootstrap_kernel_object_lock); 500 uvm_obj_setlock(&aobj->u_obj, &bootstrap_kernel_object_lock); 501 } 502 503 /* 504 * now that aobj is ready, add it to the global list 505 */ 506 507 mutex_enter(&uao_list_lock); 508 LIST_INSERT_HEAD(&uao_list, aobj, u_list); 509 mutex_exit(&uao_list_lock); 510 return(&aobj->u_obj); 511 } 512 513 /* 514 * uao_set_pgfl: allocate pages only from the specified freelist. 515 * 516 * => must be called before any pages are allocated for the object. 517 * => reset by setting it to VM_NFREELIST, meaning any freelist. 518 */ 519 520 void 521 uao_set_pgfl(struct uvm_object *uobj, int freelist) 522 { 523 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 524 525 KASSERTMSG((0 <= freelist), "invalid freelist %d", freelist); 526 KASSERTMSG((freelist <= VM_NFREELIST), "invalid freelist %d", 527 freelist); 528 529 aobj->u_freelist = freelist; 530 } 531 532 /* 533 * uao_pagealloc: allocate a page for aobj. 534 */ 535 536 static inline struct vm_page * 537 uao_pagealloc(struct uvm_object *uobj, voff_t offset, int flags) 538 { 539 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 540 541 if (__predict_true(aobj->u_freelist == VM_NFREELIST)) 542 return uvm_pagealloc(uobj, offset, NULL, flags); 543 else 544 return uvm_pagealloc_strat(uobj, offset, NULL, flags, 545 UVM_PGA_STRAT_ONLY, aobj->u_freelist); 546 } 547 548 /* 549 * uao_init: set up aobj pager subsystem 550 * 551 * => called at boot time from uvm_pager_init() 552 */ 553 554 void 555 uao_init(void) 556 { 557 static int uao_initialized; 558 559 if (uao_initialized) 560 return; 561 uao_initialized = true; 562 LIST_INIT(&uao_list); 563 mutex_init(&uao_list_lock, MUTEX_DEFAULT, IPL_NONE); 564 pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 565 0, 0, 0, "uaoeltpl", NULL, IPL_VM); 566 } 567 568 /* 569 * uao_reference: hold a reference to an anonymous UVM object. 570 */ 571 void 572 uao_reference(struct uvm_object *uobj) 573 { 574 /* Kernel object is persistent. */ 575 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) { 576 return; 577 } 578 atomic_inc_uint(&uobj->uo_refs); 579 } 580 581 /* 582 * uao_detach: drop a reference to an anonymous UVM object. 583 */ 584 void 585 uao_detach(struct uvm_object *uobj) 586 { 587 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 588 struct uvm_page_array a; 589 struct vm_page *pg; 590 591 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 592 593 /* 594 * Detaching from kernel object is a NOP. 595 */ 596 597 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 598 return; 599 600 /* 601 * Drop the reference. If it was the last one, destroy the object. 602 */ 603 604 KASSERT(uobj->uo_refs > 0); 605 UVMHIST_LOG(maphist," (uobj=%#jx) ref=%jd", 606 (uintptr_t)uobj, uobj->uo_refs, 0, 0); 607 #ifndef __HAVE_ATOMIC_AS_MEMBAR 608 membar_release(); 609 #endif 610 if (atomic_dec_uint_nv(&uobj->uo_refs) > 0) { 611 UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0); 612 return; 613 } 614 #ifndef __HAVE_ATOMIC_AS_MEMBAR 615 membar_acquire(); 616 #endif 617 618 /* 619 * Remove the aobj from the global list. 620 */ 621 622 mutex_enter(&uao_list_lock); 623 LIST_REMOVE(aobj, u_list); 624 mutex_exit(&uao_list_lock); 625 626 /* 627 * Free all the pages left in the aobj. For each page, when the 628 * page is no longer busy (and thus after any disk I/O that it is 629 * involved in is complete), release any swap resources and free 630 * the page itself. 631 */ 632 uvm_page_array_init(&a, uobj, 0); 633 rw_enter(uobj->vmobjlock, RW_WRITER); 634 while ((pg = uvm_page_array_fill_and_peek(&a, 0, 0)) != NULL) { 635 uvm_page_array_advance(&a); 636 pmap_page_protect(pg, VM_PROT_NONE); 637 if (pg->flags & PG_BUSY) { 638 uvm_pagewait(pg, uobj->vmobjlock, "uao_det"); 639 uvm_page_array_clear(&a); 640 rw_enter(uobj->vmobjlock, RW_WRITER); 641 continue; 642 } 643 uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT); 644 uvm_pagefree(pg); 645 } 646 uvm_page_array_fini(&a); 647 648 /* 649 * Finally, free the anonymous UVM object itself. 650 */ 651 652 uao_free(aobj); 653 } 654 655 /* 656 * uao_put: flush pages out of a uvm object 657 * 658 * => object should be locked by caller. we may _unlock_ the object 659 * if (and only if) we need to clean a page (PGO_CLEANIT). 660 * XXXJRT Currently, however, we don't. In the case of cleaning 661 * XXXJRT a page, we simply just deactivate it. Should probably 662 * XXXJRT handle this better, in the future (although "flushing" 663 * XXXJRT anonymous memory isn't terribly important). 664 * => if PGO_CLEANIT is not set, then we will neither unlock the object 665 * or block. 666 * => if PGO_ALLPAGE is set, then all pages in the object are valid targets 667 * for flushing. 668 * => we return 0 unless we encountered some sort of I/O error 669 * XXXJRT currently never happens, as we never directly initiate 670 * XXXJRT I/O 671 */ 672 673 static int 674 uao_put(struct uvm_object *uobj, voff_t start, voff_t stop, int flags) 675 { 676 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 677 struct uvm_page_array a; 678 struct vm_page *pg; 679 voff_t curoff; 680 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 681 682 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 683 KASSERT(rw_write_held(uobj->vmobjlock)); 684 685 if (flags & PGO_ALLPAGES) { 686 start = 0; 687 stop = aobj->u_pages << PAGE_SHIFT; 688 } else { 689 start = trunc_page(start); 690 if (stop == 0) { 691 stop = aobj->u_pages << PAGE_SHIFT; 692 } else { 693 stop = round_page(stop); 694 } 695 if (stop > (uint64_t)(aobj->u_pages << PAGE_SHIFT)) { 696 printf("uao_put: strange, got an out of range " 697 "flush %#jx > %#jx (fixed)\n", 698 (uintmax_t)stop, 699 (uintmax_t)(aobj->u_pages << PAGE_SHIFT)); 700 stop = aobj->u_pages << PAGE_SHIFT; 701 } 702 } 703 UVMHIST_LOG(maphist, 704 " flush start=%#jx, stop=%#jx, flags=%#jx", 705 start, stop, flags, 0); 706 707 /* 708 * Don't need to do any work here if we're not freeing 709 * or deactivating pages. 710 */ 711 712 if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) { 713 rw_exit(uobj->vmobjlock); 714 return 0; 715 } 716 717 /* locked: uobj */ 718 uvm_page_array_init(&a, uobj, 0); 719 curoff = start; 720 while ((pg = uvm_page_array_fill_and_peek(&a, curoff, 0)) != NULL) { 721 if (pg->offset >= stop) { 722 break; 723 } 724 725 /* 726 * wait and try again if the page is busy. 727 */ 728 729 if (pg->flags & PG_BUSY) { 730 uvm_pagewait(pg, uobj->vmobjlock, "uao_put"); 731 uvm_page_array_clear(&a); 732 rw_enter(uobj->vmobjlock, RW_WRITER); 733 continue; 734 } 735 uvm_page_array_advance(&a); 736 curoff = pg->offset + PAGE_SIZE; 737 738 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 739 740 /* 741 * XXX In these first 3 cases, we always just 742 * XXX deactivate the page. We may want to 743 * XXX handle the different cases more specifically 744 * XXX in the future. 745 */ 746 747 case PGO_CLEANIT|PGO_FREE: 748 case PGO_CLEANIT|PGO_DEACTIVATE: 749 case PGO_DEACTIVATE: 750 deactivate_it: 751 uvm_pagelock(pg); 752 uvm_pagedeactivate(pg); 753 uvm_pageunlock(pg); 754 break; 755 756 case PGO_FREE: 757 /* 758 * If there are multiple references to 759 * the object, just deactivate the page. 760 */ 761 762 if (uobj->uo_refs > 1) 763 goto deactivate_it; 764 765 /* 766 * free the swap slot and the page. 767 */ 768 769 pmap_page_protect(pg, VM_PROT_NONE); 770 771 /* 772 * freeing swapslot here is not strictly necessary. 773 * however, leaving it here doesn't save much 774 * because we need to update swap accounting anyway. 775 */ 776 777 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 778 uvm_pagefree(pg); 779 break; 780 781 default: 782 panic("%s: impossible", __func__); 783 } 784 } 785 rw_exit(uobj->vmobjlock); 786 uvm_page_array_fini(&a); 787 return 0; 788 } 789 790 /* 791 * uao_get: fetch me a page 792 * 793 * we have three cases: 794 * 1: page is resident -> just return the page. 795 * 2: page is zero-fill -> allocate a new page and zero it. 796 * 3: page is swapped out -> fetch the page from swap. 797 * 798 * case 1 can be handled with PGO_LOCKED, cases 2 and 3 cannot. 799 * so, if the "center" page hits case 2/3 then we will need to return EBUSY. 800 * 801 * => prefer map unlocked (not required) 802 * => object must be locked! we will _unlock_ it before starting any I/O. 803 * => flags: PGO_LOCKED: fault data structures are locked 804 * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx] 805 * => NOTE: caller must check for released pages!! 806 */ 807 808 static int 809 uao_get(struct uvm_object *uobj, voff_t offset, struct vm_page **pps, 810 int *npagesp, int centeridx, vm_prot_t access_type, int advice, int flags) 811 { 812 voff_t current_offset; 813 struct vm_page *ptmp; 814 int lcv, gotpages, maxpages, swslot, pageidx; 815 bool overwrite = ((flags & PGO_OVERWRITE) != 0); 816 struct uvm_page_array a; 817 818 UVMHIST_FUNC(__func__); 819 UVMHIST_CALLARGS(pdhist, "aobj=%#jx offset=%jd, flags=%#jx", 820 (uintptr_t)uobj, offset, flags,0); 821 822 /* 823 * the object must be locked. it can only be a read lock when 824 * processing a read fault with PGO_LOCKED. 825 */ 826 827 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 828 KASSERT(rw_lock_held(uobj->vmobjlock)); 829 KASSERT(rw_write_held(uobj->vmobjlock) || 830 ((flags & PGO_LOCKED) != 0 && (access_type & VM_PROT_WRITE) == 0)); 831 832 /* 833 * get number of pages 834 */ 835 836 maxpages = *npagesp; 837 838 /* 839 * step 1: handled the case where fault data structures are locked. 840 */ 841 842 if (flags & PGO_LOCKED) { 843 844 /* 845 * step 1a: get pages that are already resident. only do 846 * this if the data structures are locked (i.e. the first 847 * time through). 848 */ 849 850 uvm_page_array_init(&a, uobj, 0); 851 gotpages = 0; /* # of pages we got so far */ 852 for (lcv = 0; lcv < maxpages; lcv++) { 853 ptmp = uvm_page_array_fill_and_peek(&a, 854 offset + (lcv << PAGE_SHIFT), maxpages); 855 if (ptmp == NULL) { 856 break; 857 } 858 KASSERT(ptmp->offset >= offset); 859 lcv = (ptmp->offset - offset) >> PAGE_SHIFT; 860 if (lcv >= maxpages) { 861 break; 862 } 863 uvm_page_array_advance(&a); 864 865 /* 866 * to be useful must get a non-busy page 867 */ 868 869 if ((ptmp->flags & PG_BUSY) != 0) { 870 continue; 871 } 872 873 /* 874 * useful page: plug it in our result array 875 */ 876 877 KASSERT(uvm_pagegetdirty(ptmp) != 878 UVM_PAGE_STATUS_CLEAN); 879 pps[lcv] = ptmp; 880 gotpages++; 881 } 882 uvm_page_array_fini(&a); 883 884 /* 885 * step 1b: now we've either done everything needed or we 886 * to unlock and do some waiting or I/O. 887 */ 888 889 UVMHIST_LOG(pdhist, "<- done (done=%jd)", 890 (pps[centeridx] != NULL), 0,0,0); 891 *npagesp = gotpages; 892 return pps[centeridx] != NULL ? 0 : EBUSY; 893 } 894 895 /* 896 * step 2: get non-resident or busy pages. 897 * object is locked. data structures are unlocked. 898 */ 899 900 if ((flags & PGO_SYNCIO) == 0) { 901 goto done; 902 } 903 904 uvm_page_array_init(&a, uobj, 0); 905 for (lcv = 0, current_offset = offset ; lcv < maxpages ;) { 906 907 /* 908 * we have yet to locate the current page (pps[lcv]). we 909 * first look for a page that is already at the current offset. 910 * if we find a page, we check to see if it is busy or 911 * released. if that is the case, then we sleep on the page 912 * until it is no longer busy or released and repeat the lookup. 913 * if the page we found is neither busy nor released, then we 914 * busy it (so we own it) and plug it into pps[lcv]. we are 915 * ready to move on to the next page. 916 */ 917 918 ptmp = uvm_page_array_fill_and_peek(&a, current_offset, 919 maxpages - lcv); 920 921 if (ptmp != NULL && ptmp->offset == current_offset) { 922 /* page is there, see if we need to wait on it */ 923 if ((ptmp->flags & PG_BUSY) != 0) { 924 UVMHIST_LOG(pdhist, 925 "sleeping, ptmp->flags %#jx\n", 926 ptmp->flags,0,0,0); 927 uvm_pagewait(ptmp, uobj->vmobjlock, "uao_get"); 928 rw_enter(uobj->vmobjlock, RW_WRITER); 929 uvm_page_array_clear(&a); 930 continue; 931 } 932 933 /* 934 * if we get here then the page is resident and 935 * unbusy. we busy it now (so we own it). if 936 * overwriting, mark the page dirty up front as 937 * it will be zapped via an unmanaged mapping. 938 */ 939 940 KASSERT(uvm_pagegetdirty(ptmp) != 941 UVM_PAGE_STATUS_CLEAN); 942 if (overwrite) { 943 uvm_pagemarkdirty(ptmp, UVM_PAGE_STATUS_DIRTY); 944 } 945 /* we own it, caller must un-busy */ 946 ptmp->flags |= PG_BUSY; 947 UVM_PAGE_OWN(ptmp, "uao_get2"); 948 pps[lcv++] = ptmp; 949 current_offset += PAGE_SIZE; 950 uvm_page_array_advance(&a); 951 continue; 952 } else { 953 KASSERT(ptmp == NULL || ptmp->offset > current_offset); 954 } 955 956 /* 957 * not resident. allocate a new busy/fake/clean page in the 958 * object. if it's in swap we need to do I/O to fill in the 959 * data, otherwise the page needs to be cleared: if it's not 960 * destined to be overwritten, then zero it here and now. 961 */ 962 963 pageidx = current_offset >> PAGE_SHIFT; 964 swslot = uao_find_swslot(uobj, pageidx); 965 ptmp = uao_pagealloc(uobj, current_offset, 966 swslot != 0 || overwrite ? 0 : UVM_PGA_ZERO); 967 968 /* out of RAM? */ 969 if (ptmp == NULL) { 970 rw_exit(uobj->vmobjlock); 971 UVMHIST_LOG(pdhist, "sleeping, ptmp == NULL",0,0,0,0); 972 uvm_wait("uao_getpage"); 973 rw_enter(uobj->vmobjlock, RW_WRITER); 974 uvm_page_array_clear(&a); 975 continue; 976 } 977 978 /* 979 * if swslot == 0, page hasn't existed before and is zeroed. 980 * otherwise we have a "fake/busy/clean" page that we just 981 * allocated. do the needed "i/o", reading from swap. 982 */ 983 984 if (swslot != 0) { 985 #if defined(VMSWAP) 986 int error; 987 988 UVMHIST_LOG(pdhist, "pagein from swslot %jd", 989 swslot, 0,0,0); 990 991 /* 992 * page in the swapped-out page. 993 * unlock object for i/o, relock when done. 994 */ 995 996 uvm_page_array_clear(&a); 997 rw_exit(uobj->vmobjlock); 998 error = uvm_swap_get(ptmp, swslot, PGO_SYNCIO); 999 rw_enter(uobj->vmobjlock, RW_WRITER); 1000 1001 /* 1002 * I/O done. check for errors. 1003 */ 1004 1005 if (error != 0) { 1006 UVMHIST_LOG(pdhist, "<- done (error=%jd)", 1007 error,0,0,0); 1008 1009 /* 1010 * remove the swap slot from the aobj 1011 * and mark the aobj as having no real slot. 1012 * don't free the swap slot, thus preventing 1013 * it from being used again. 1014 */ 1015 1016 swslot = uao_set_swslot(uobj, pageidx, 1017 SWSLOT_BAD); 1018 if (swslot > 0) { 1019 uvm_swap_markbad(swslot, 1); 1020 } 1021 1022 uvm_pagefree(ptmp); 1023 rw_exit(uobj->vmobjlock); 1024 UVMHIST_LOG(pdhist, "<- done (error)", 1025 error,lcv,0,0); 1026 if (lcv != 0) { 1027 uvm_page_unbusy(pps, lcv); 1028 } 1029 memset(pps, 0, maxpages * sizeof(pps[0])); 1030 uvm_page_array_fini(&a); 1031 return error; 1032 } 1033 #else /* defined(VMSWAP) */ 1034 panic("%s: pagein", __func__); 1035 #endif /* defined(VMSWAP) */ 1036 } 1037 1038 /* 1039 * note that we will allow the page being writably-mapped 1040 * (!PG_RDONLY) regardless of access_type. if overwrite, 1041 * the page can be modified through an unmanaged mapping 1042 * so mark it dirty up front. 1043 */ 1044 if (overwrite) { 1045 uvm_pagemarkdirty(ptmp, UVM_PAGE_STATUS_DIRTY); 1046 } else { 1047 uvm_pagemarkdirty(ptmp, UVM_PAGE_STATUS_UNKNOWN); 1048 } 1049 1050 /* 1051 * we got the page! clear the fake flag (indicates valid 1052 * data now in page) and plug into our result array. note 1053 * that page is still busy. 1054 * 1055 * it is the callers job to: 1056 * => check if the page is released 1057 * => unbusy the page 1058 * => activate the page 1059 */ 1060 KASSERT(uvm_pagegetdirty(ptmp) != UVM_PAGE_STATUS_CLEAN); 1061 KASSERT((ptmp->flags & PG_FAKE) != 0); 1062 KASSERT(ptmp->offset == current_offset); 1063 ptmp->flags &= ~PG_FAKE; 1064 pps[lcv++] = ptmp; 1065 current_offset += PAGE_SIZE; 1066 } 1067 uvm_page_array_fini(&a); 1068 1069 /* 1070 * finally, unlock object and return. 1071 */ 1072 1073 done: 1074 rw_exit(uobj->vmobjlock); 1075 UVMHIST_LOG(pdhist, "<- done (OK)",0,0,0,0); 1076 return 0; 1077 } 1078 1079 #if defined(VMSWAP) 1080 1081 /* 1082 * uao_dropswap: release any swap resources from this aobj page. 1083 * 1084 * => aobj must be locked or have a reference count of 0. 1085 */ 1086 1087 void 1088 uao_dropswap(struct uvm_object *uobj, int pageidx) 1089 { 1090 int slot; 1091 1092 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 1093 1094 slot = uao_set_swslot(uobj, pageidx, 0); 1095 if (slot) { 1096 uvm_swap_free(slot, 1); 1097 } 1098 } 1099 1100 /* 1101 * page in every page in every aobj that is paged-out to a range of swslots. 1102 * 1103 * => nothing should be locked. 1104 * => returns true if pagein was aborted due to lack of memory. 1105 */ 1106 1107 bool 1108 uao_swap_off(int startslot, int endslot) 1109 { 1110 struct uvm_aobj *aobj; 1111 1112 /* 1113 * Walk the list of all anonymous UVM objects. Grab the first. 1114 */ 1115 mutex_enter(&uao_list_lock); 1116 if ((aobj = LIST_FIRST(&uao_list)) == NULL) { 1117 mutex_exit(&uao_list_lock); 1118 return false; 1119 } 1120 uao_reference(&aobj->u_obj); 1121 1122 do { 1123 struct uvm_aobj *nextaobj; 1124 bool rv; 1125 1126 /* 1127 * Prefetch the next object and immediately hold a reference 1128 * on it, so neither the current nor the next entry could 1129 * disappear while we are iterating. 1130 */ 1131 if ((nextaobj = LIST_NEXT(aobj, u_list)) != NULL) { 1132 uao_reference(&nextaobj->u_obj); 1133 } 1134 mutex_exit(&uao_list_lock); 1135 1136 /* 1137 * Page in all pages in the swap slot range. 1138 */ 1139 rw_enter(aobj->u_obj.vmobjlock, RW_WRITER); 1140 rv = uao_pagein(aobj, startslot, endslot); 1141 rw_exit(aobj->u_obj.vmobjlock); 1142 1143 /* Drop the reference of the current object. */ 1144 uao_detach(&aobj->u_obj); 1145 if (rv) { 1146 if (nextaobj) { 1147 uao_detach(&nextaobj->u_obj); 1148 } 1149 return rv; 1150 } 1151 1152 aobj = nextaobj; 1153 mutex_enter(&uao_list_lock); 1154 } while (aobj); 1155 1156 mutex_exit(&uao_list_lock); 1157 return false; 1158 } 1159 1160 /* 1161 * page in any pages from aobj in the given range. 1162 * 1163 * => aobj must be locked and is returned locked. 1164 * => returns true if pagein was aborted due to lack of memory. 1165 */ 1166 static bool 1167 uao_pagein(struct uvm_aobj *aobj, int startslot, int endslot) 1168 { 1169 bool rv; 1170 1171 if (UAO_USES_SWHASH(aobj)) { 1172 struct uao_swhash_elt *elt; 1173 int buck; 1174 1175 restart: 1176 for (buck = aobj->u_swhashmask; buck >= 0; buck--) { 1177 for (elt = LIST_FIRST(&aobj->u_swhash[buck]); 1178 elt != NULL; 1179 elt = LIST_NEXT(elt, list)) { 1180 int i; 1181 1182 for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) { 1183 int slot = elt->slots[i]; 1184 1185 /* 1186 * if the slot isn't in range, skip it. 1187 */ 1188 1189 if (slot < startslot || 1190 slot >= endslot) { 1191 continue; 1192 } 1193 1194 /* 1195 * process the page, 1196 * the start over on this object 1197 * since the swhash elt 1198 * may have been freed. 1199 */ 1200 1201 rv = uao_pagein_page(aobj, 1202 UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i); 1203 if (rv) { 1204 return rv; 1205 } 1206 goto restart; 1207 } 1208 } 1209 } 1210 } else { 1211 int i; 1212 1213 for (i = 0; i < aobj->u_pages; i++) { 1214 int slot = aobj->u_swslots[i]; 1215 1216 /* 1217 * if the slot isn't in range, skip it 1218 */ 1219 1220 if (slot < startslot || slot >= endslot) { 1221 continue; 1222 } 1223 1224 /* 1225 * process the page. 1226 */ 1227 1228 rv = uao_pagein_page(aobj, i); 1229 if (rv) { 1230 return rv; 1231 } 1232 } 1233 } 1234 1235 return false; 1236 } 1237 1238 /* 1239 * uao_pagein_page: page in a single page from an anonymous UVM object. 1240 * 1241 * => Returns true if pagein was aborted due to lack of memory. 1242 * => Object must be locked and is returned locked. 1243 */ 1244 1245 static bool 1246 uao_pagein_page(struct uvm_aobj *aobj, int pageidx) 1247 { 1248 struct uvm_object *uobj = &aobj->u_obj; 1249 struct vm_page *pg; 1250 int rv, npages; 1251 1252 pg = NULL; 1253 npages = 1; 1254 1255 KASSERT(rw_write_held(uobj->vmobjlock)); 1256 rv = uao_get(uobj, (voff_t)pageidx << PAGE_SHIFT, &pg, &npages, 1257 0, VM_PROT_READ | VM_PROT_WRITE, 0, PGO_SYNCIO); 1258 1259 /* 1260 * relock and finish up. 1261 */ 1262 1263 rw_enter(uobj->vmobjlock, RW_WRITER); 1264 switch (rv) { 1265 case 0: 1266 break; 1267 1268 case EIO: 1269 case ERESTART: 1270 1271 /* 1272 * nothing more to do on errors. 1273 * ERESTART can only mean that the anon was freed, 1274 * so again there's nothing to do. 1275 */ 1276 1277 return false; 1278 1279 default: 1280 return true; 1281 } 1282 1283 /* 1284 * ok, we've got the page now. 1285 * mark it as dirty, clear its swslot and un-busy it. 1286 */ 1287 uao_dropswap(&aobj->u_obj, pageidx); 1288 1289 /* 1290 * make sure it's on a page queue. 1291 */ 1292 uvm_pagelock(pg); 1293 uvm_pageenqueue(pg); 1294 uvm_pagewakeup(pg); 1295 uvm_pageunlock(pg); 1296 1297 pg->flags &= ~(PG_BUSY|PG_FAKE); 1298 uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY); 1299 UVM_PAGE_OWN(pg, NULL); 1300 1301 return false; 1302 } 1303 1304 /* 1305 * uao_dropswap_range: drop swapslots in the range. 1306 * 1307 * => aobj must be locked and is returned locked. 1308 * => start is inclusive. end is exclusive. 1309 */ 1310 1311 void 1312 uao_dropswap_range(struct uvm_object *uobj, voff_t start, voff_t end) 1313 { 1314 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 1315 int swpgonlydelta = 0; 1316 1317 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 1318 KASSERT(rw_write_held(uobj->vmobjlock)); 1319 1320 if (end == 0) { 1321 end = INT64_MAX; 1322 } 1323 1324 if (UAO_USES_SWHASH(aobj)) { 1325 int i, hashbuckets = aobj->u_swhashmask + 1; 1326 voff_t taghi; 1327 voff_t taglo; 1328 1329 taglo = UAO_SWHASH_ELT_TAG(start); 1330 taghi = UAO_SWHASH_ELT_TAG(end); 1331 1332 for (i = 0; i < hashbuckets; i++) { 1333 struct uao_swhash_elt *elt, *next; 1334 1335 for (elt = LIST_FIRST(&aobj->u_swhash[i]); 1336 elt != NULL; 1337 elt = next) { 1338 int startidx, endidx; 1339 int j; 1340 1341 next = LIST_NEXT(elt, list); 1342 1343 if (elt->tag < taglo || taghi < elt->tag) { 1344 continue; 1345 } 1346 1347 if (elt->tag == taglo) { 1348 startidx = 1349 UAO_SWHASH_ELT_PAGESLOT_IDX(start); 1350 } else { 1351 startidx = 0; 1352 } 1353 1354 if (elt->tag == taghi) { 1355 endidx = 1356 UAO_SWHASH_ELT_PAGESLOT_IDX(end); 1357 } else { 1358 endidx = UAO_SWHASH_CLUSTER_SIZE; 1359 } 1360 1361 for (j = startidx; j < endidx; j++) { 1362 int slot = elt->slots[j]; 1363 1364 KASSERT(uvm_pagelookup(&aobj->u_obj, 1365 (UAO_SWHASH_ELT_PAGEIDX_BASE(elt) 1366 + j) << PAGE_SHIFT) == NULL); 1367 if (slot > 0) { 1368 uvm_swap_free(slot, 1); 1369 swpgonlydelta++; 1370 KASSERT(elt->count > 0); 1371 elt->slots[j] = 0; 1372 elt->count--; 1373 } 1374 } 1375 1376 if (elt->count == 0) { 1377 LIST_REMOVE(elt, list); 1378 pool_put(&uao_swhash_elt_pool, elt); 1379 } 1380 } 1381 } 1382 } else { 1383 int i; 1384 1385 if (aobj->u_pages < end) { 1386 end = aobj->u_pages; 1387 } 1388 for (i = start; i < end; i++) { 1389 int slot = aobj->u_swslots[i]; 1390 1391 if (slot > 0) { 1392 uvm_swap_free(slot, 1); 1393 swpgonlydelta++; 1394 } 1395 } 1396 } 1397 1398 /* 1399 * adjust the counter of pages only in swap for all 1400 * the swap slots we've freed. 1401 */ 1402 1403 if (swpgonlydelta > 0) { 1404 KASSERT(uvmexp.swpgonly >= swpgonlydelta); 1405 atomic_add_int(&uvmexp.swpgonly, -swpgonlydelta); 1406 } 1407 } 1408 1409 #endif /* defined(VMSWAP) */ 1410