1 /* $NetBSD: uvm_aobj.c,v 1.128 2019/07/28 05:28:53 msaitoh 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 1.128 2019/07/28 05:28:53 msaitoh 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 56 /* 57 * An anonymous UVM object (aobj) manages anonymous-memory. In addition to 58 * keeping the list of resident pages, it may also keep a list of allocated 59 * swap blocks. Depending on the size of the object, this list is either 60 * stored in an array (small objects) or in a hash table (large objects). 61 * 62 * Lock order 63 * 64 * uao_list_lock -> 65 * uvm_object::vmobjlock 66 */ 67 68 /* 69 * Note: for hash tables, we break the address space of the aobj into blocks 70 * of UAO_SWHASH_CLUSTER_SIZE pages, which shall be a power of two. 71 */ 72 73 #define UAO_SWHASH_CLUSTER_SHIFT 4 74 #define UAO_SWHASH_CLUSTER_SIZE (1 << UAO_SWHASH_CLUSTER_SHIFT) 75 76 /* Get the "tag" for this page index. */ 77 #define UAO_SWHASH_ELT_TAG(idx) ((idx) >> UAO_SWHASH_CLUSTER_SHIFT) 78 #define UAO_SWHASH_ELT_PAGESLOT_IDX(idx) \ 79 ((idx) & (UAO_SWHASH_CLUSTER_SIZE - 1)) 80 81 /* Given an ELT and a page index, find the swap slot. */ 82 #define UAO_SWHASH_ELT_PAGESLOT(elt, idx) \ 83 ((elt)->slots[UAO_SWHASH_ELT_PAGESLOT_IDX(idx)]) 84 85 /* Given an ELT, return its pageidx base. */ 86 #define UAO_SWHASH_ELT_PAGEIDX_BASE(ELT) \ 87 ((elt)->tag << UAO_SWHASH_CLUSTER_SHIFT) 88 89 /* The hash function. */ 90 #define UAO_SWHASH_HASH(aobj, idx) \ 91 (&(aobj)->u_swhash[(((idx) >> UAO_SWHASH_CLUSTER_SHIFT) \ 92 & (aobj)->u_swhashmask)]) 93 94 /* 95 * The threshold which determines whether we will use an array or a 96 * hash table to store the list of allocated swap blocks. 97 */ 98 #define UAO_SWHASH_THRESHOLD (UAO_SWHASH_CLUSTER_SIZE * 4) 99 #define UAO_USES_SWHASH(aobj) \ 100 ((aobj)->u_pages > UAO_SWHASH_THRESHOLD) 101 102 /* The number of buckets in a hash, with an upper bound. */ 103 #define UAO_SWHASH_MAXBUCKETS 256 104 #define UAO_SWHASH_BUCKETS(aobj) \ 105 (MIN((aobj)->u_pages >> UAO_SWHASH_CLUSTER_SHIFT, UAO_SWHASH_MAXBUCKETS)) 106 107 /* 108 * uao_swhash_elt: when a hash table is being used, this structure defines 109 * the format of an entry in the bucket list. 110 */ 111 112 struct uao_swhash_elt { 113 LIST_ENTRY(uao_swhash_elt) list; /* the hash list */ 114 voff_t tag; /* our 'tag' */ 115 int count; /* our number of active slots */ 116 int slots[UAO_SWHASH_CLUSTER_SIZE]; /* the slots */ 117 }; 118 119 /* 120 * uao_swhash: the swap hash table structure 121 */ 122 123 LIST_HEAD(uao_swhash, uao_swhash_elt); 124 125 /* 126 * uao_swhash_elt_pool: pool of uao_swhash_elt structures. 127 * Note: pages for this pool must not come from a pageable kernel map. 128 */ 129 static struct pool uao_swhash_elt_pool __cacheline_aligned; 130 131 /* 132 * uvm_aobj: the actual anon-backed uvm_object 133 * 134 * => the uvm_object is at the top of the structure, this allows 135 * (struct uvm_aobj *) == (struct uvm_object *) 136 * => only one of u_swslots and u_swhash is used in any given aobj 137 */ 138 139 struct uvm_aobj { 140 struct uvm_object u_obj; /* has: lock, pgops, memq, #pages, #refs */ 141 pgoff_t u_pages; /* number of pages in entire object */ 142 int u_flags; /* the flags (see uvm_aobj.h) */ 143 int *u_swslots; /* array of offset->swapslot mappings */ 144 /* 145 * hashtable of offset->swapslot mappings 146 * (u_swhash is an array of bucket heads) 147 */ 148 struct uao_swhash *u_swhash; 149 u_long u_swhashmask; /* mask for hashtable */ 150 LIST_ENTRY(uvm_aobj) u_list; /* global list of aobjs */ 151 int u_freelist; /* freelist to allocate pages from */ 152 }; 153 154 static void uao_free(struct uvm_aobj *); 155 static int uao_get(struct uvm_object *, voff_t, struct vm_page **, 156 int *, int, vm_prot_t, int, int); 157 static int uao_put(struct uvm_object *, voff_t, voff_t, int); 158 159 #if defined(VMSWAP) 160 static struct uao_swhash_elt *uao_find_swhash_elt 161 (struct uvm_aobj *, int, bool); 162 163 static bool uao_pagein(struct uvm_aobj *, int, int); 164 static bool uao_pagein_page(struct uvm_aobj *, int); 165 #endif /* defined(VMSWAP) */ 166 167 static struct vm_page *uao_pagealloc(struct uvm_object *, voff_t, int); 168 169 /* 170 * aobj_pager 171 * 172 * note that some functions (e.g. put) are handled elsewhere 173 */ 174 175 const struct uvm_pagerops aobj_pager = { 176 .pgo_reference = uao_reference, 177 .pgo_detach = uao_detach, 178 .pgo_get = uao_get, 179 .pgo_put = uao_put, 180 }; 181 182 /* 183 * uao_list: global list of active aobjs, locked by uao_list_lock 184 */ 185 186 static LIST_HEAD(aobjlist, uvm_aobj) uao_list __cacheline_aligned; 187 static kmutex_t uao_list_lock __cacheline_aligned; 188 189 /* 190 * hash table/array related functions 191 */ 192 193 #if defined(VMSWAP) 194 195 /* 196 * uao_find_swhash_elt: find (or create) a hash table entry for a page 197 * offset. 198 * 199 * => the object should be locked by the caller 200 */ 201 202 static struct uao_swhash_elt * 203 uao_find_swhash_elt(struct uvm_aobj *aobj, int pageidx, bool create) 204 { 205 struct uao_swhash *swhash; 206 struct uao_swhash_elt *elt; 207 voff_t page_tag; 208 209 swhash = UAO_SWHASH_HASH(aobj, pageidx); 210 page_tag = UAO_SWHASH_ELT_TAG(pageidx); 211 212 /* 213 * now search the bucket for the requested tag 214 */ 215 216 LIST_FOREACH(elt, swhash, list) { 217 if (elt->tag == page_tag) { 218 return elt; 219 } 220 } 221 if (!create) { 222 return NULL; 223 } 224 225 /* 226 * allocate a new entry for the bucket and init/insert it in 227 */ 228 229 elt = pool_get(&uao_swhash_elt_pool, PR_NOWAIT); 230 if (elt == NULL) { 231 return NULL; 232 } 233 LIST_INSERT_HEAD(swhash, elt, list); 234 elt->tag = page_tag; 235 elt->count = 0; 236 memset(elt->slots, 0, sizeof(elt->slots)); 237 return elt; 238 } 239 240 /* 241 * uao_find_swslot: find the swap slot number for an aobj/pageidx 242 * 243 * => object must be locked by caller 244 */ 245 246 int 247 uao_find_swslot(struct uvm_object *uobj, int pageidx) 248 { 249 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 250 struct uao_swhash_elt *elt; 251 252 /* 253 * if noswap flag is set, then we never return a slot 254 */ 255 256 if (aobj->u_flags & UAO_FLAG_NOSWAP) 257 return 0; 258 259 /* 260 * if hashing, look in hash table. 261 */ 262 263 if (UAO_USES_SWHASH(aobj)) { 264 elt = uao_find_swhash_elt(aobj, pageidx, false); 265 return elt ? UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) : 0; 266 } 267 268 /* 269 * otherwise, look in the array 270 */ 271 272 return aobj->u_swslots[pageidx]; 273 } 274 275 /* 276 * uao_set_swslot: set the swap slot for a page in an aobj. 277 * 278 * => setting a slot to zero frees the slot 279 * => object must be locked by caller 280 * => we return the old slot number, or -1 if we failed to allocate 281 * memory to record the new slot number 282 */ 283 284 int 285 uao_set_swslot(struct uvm_object *uobj, int pageidx, int slot) 286 { 287 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 288 struct uao_swhash_elt *elt; 289 int oldslot; 290 UVMHIST_FUNC("uao_set_swslot"); UVMHIST_CALLED(pdhist); 291 UVMHIST_LOG(pdhist, "aobj %#jx pageidx %jd slot %jd", 292 (uintptr_t)aobj, pageidx, slot, 0); 293 294 KASSERT(mutex_owned(uobj->vmobjlock) || uobj->uo_refs == 0); 295 296 /* 297 * if noswap flag is set, then we can't set a non-zero slot. 298 */ 299 300 if (aobj->u_flags & UAO_FLAG_NOSWAP) { 301 KASSERTMSG(slot == 0, "uao_set_swslot: no swap object"); 302 return 0; 303 } 304 305 /* 306 * are we using a hash table? if so, add it in the hash. 307 */ 308 309 if (UAO_USES_SWHASH(aobj)) { 310 311 /* 312 * Avoid allocating an entry just to free it again if 313 * the page had not swap slot in the first place, and 314 * we are freeing. 315 */ 316 317 elt = uao_find_swhash_elt(aobj, pageidx, slot != 0); 318 if (elt == NULL) { 319 return slot ? -1 : 0; 320 } 321 322 oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx); 323 UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot; 324 325 /* 326 * now adjust the elt's reference counter and free it if we've 327 * dropped it to zero. 328 */ 329 330 if (slot) { 331 if (oldslot == 0) 332 elt->count++; 333 } else { 334 if (oldslot) 335 elt->count--; 336 337 if (elt->count == 0) { 338 LIST_REMOVE(elt, list); 339 pool_put(&uao_swhash_elt_pool, elt); 340 } 341 } 342 } else { 343 /* we are using an array */ 344 oldslot = aobj->u_swslots[pageidx]; 345 aobj->u_swslots[pageidx] = slot; 346 } 347 return oldslot; 348 } 349 350 #endif /* defined(VMSWAP) */ 351 352 /* 353 * end of hash/array functions 354 */ 355 356 /* 357 * uao_free: free all resources held by an aobj, and then free the aobj 358 * 359 * => the aobj should be dead 360 */ 361 362 static void 363 uao_free(struct uvm_aobj *aobj) 364 { 365 struct uvm_object *uobj = &aobj->u_obj; 366 367 KASSERT(mutex_owned(uobj->vmobjlock)); 368 uao_dropswap_range(uobj, 0, 0); 369 mutex_exit(uobj->vmobjlock); 370 371 #if defined(VMSWAP) 372 if (UAO_USES_SWHASH(aobj)) { 373 374 /* 375 * free the hash table itself. 376 */ 377 378 hashdone(aobj->u_swhash, HASH_LIST, aobj->u_swhashmask); 379 } else { 380 381 /* 382 * free the array itsself. 383 */ 384 385 kmem_free(aobj->u_swslots, aobj->u_pages * sizeof(int)); 386 } 387 #endif /* defined(VMSWAP) */ 388 389 /* 390 * finally free the aobj itself 391 */ 392 393 uvm_obj_destroy(uobj, true); 394 kmem_free(aobj, sizeof(struct uvm_aobj)); 395 } 396 397 /* 398 * pager functions 399 */ 400 401 /* 402 * uao_create: create an aobj of the given size and return its uvm_object. 403 * 404 * => for normal use, flags are always zero 405 * => for the kernel object, the flags are: 406 * UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once) 407 * UAO_FLAG_KERNSWAP - enable swapping of kernel object (" ") 408 */ 409 410 struct uvm_object * 411 uao_create(voff_t size, int flags) 412 { 413 static struct uvm_aobj kernel_object_store; 414 static kmutex_t kernel_object_lock; 415 static int kobj_alloced __diagused = 0; 416 pgoff_t pages = round_page((uint64_t)size) >> PAGE_SHIFT; 417 struct uvm_aobj *aobj; 418 int refs; 419 420 /* 421 * Allocate a new aobj, unless kernel object is requested. 422 */ 423 424 if (flags & UAO_FLAG_KERNOBJ) { 425 KASSERT(!kobj_alloced); 426 aobj = &kernel_object_store; 427 aobj->u_pages = pages; 428 aobj->u_flags = UAO_FLAG_NOSWAP; 429 refs = UVM_OBJ_KERN; 430 kobj_alloced = UAO_FLAG_KERNOBJ; 431 } else if (flags & UAO_FLAG_KERNSWAP) { 432 KASSERT(kobj_alloced == UAO_FLAG_KERNOBJ); 433 aobj = &kernel_object_store; 434 kobj_alloced = UAO_FLAG_KERNSWAP; 435 refs = 0xdeadbeaf; /* XXX: gcc */ 436 } else { 437 aobj = kmem_alloc(sizeof(struct uvm_aobj), KM_SLEEP); 438 aobj->u_pages = pages; 439 aobj->u_flags = 0; 440 refs = 1; 441 } 442 443 /* 444 * no freelist by default 445 */ 446 447 aobj->u_freelist = VM_NFREELIST; 448 449 /* 450 * allocate hash/array if necessary 451 * 452 * note: in the KERNSWAP case no need to worry about locking since 453 * we are still booting we should be the only thread around. 454 */ 455 456 if (flags == 0 || (flags & UAO_FLAG_KERNSWAP) != 0) { 457 #if defined(VMSWAP) 458 const int kernswap = (flags & UAO_FLAG_KERNSWAP) != 0; 459 460 /* allocate hash table or array depending on object size */ 461 if (UAO_USES_SWHASH(aobj)) { 462 aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(aobj), 463 HASH_LIST, kernswap ? false : true, 464 &aobj->u_swhashmask); 465 if (aobj->u_swhash == NULL) 466 panic("uao_create: hashinit swhash failed"); 467 } else { 468 aobj->u_swslots = kmem_zalloc(pages * sizeof(int), 469 kernswap ? KM_NOSLEEP : KM_SLEEP); 470 if (aobj->u_swslots == NULL) 471 panic("uao_create: swslots allocation failed"); 472 } 473 #endif /* defined(VMSWAP) */ 474 475 if (flags) { 476 aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */ 477 return &aobj->u_obj; 478 } 479 } 480 481 /* 482 * Initialise UVM object. 483 */ 484 485 const bool kernobj = (flags & UAO_FLAG_KERNOBJ) != 0; 486 uvm_obj_init(&aobj->u_obj, &aobj_pager, !kernobj, refs); 487 if (__predict_false(kernobj)) { 488 /* Initialisation only once, for UAO_FLAG_KERNOBJ. */ 489 mutex_init(&kernel_object_lock, MUTEX_DEFAULT, IPL_NONE); 490 uvm_obj_setlock(&aobj->u_obj, &kernel_object_lock); 491 } 492 493 /* 494 * now that aobj is ready, add it to the global list 495 */ 496 497 mutex_enter(&uao_list_lock); 498 LIST_INSERT_HEAD(&uao_list, aobj, u_list); 499 mutex_exit(&uao_list_lock); 500 return(&aobj->u_obj); 501 } 502 503 /* 504 * uao_set_pgfl: allocate pages only from the specified freelist. 505 * 506 * => must be called before any pages are allocated for the object. 507 * => reset by setting it to VM_NFREELIST, meaning any freelist. 508 */ 509 510 void 511 uao_set_pgfl(struct uvm_object *uobj, int freelist) 512 { 513 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 514 515 KASSERTMSG((0 <= freelist), "invalid freelist %d", freelist); 516 KASSERTMSG((freelist <= VM_NFREELIST), "invalid freelist %d", 517 freelist); 518 519 aobj->u_freelist = freelist; 520 } 521 522 /* 523 * uao_pagealloc: allocate a page for aobj. 524 */ 525 526 static inline struct vm_page * 527 uao_pagealloc(struct uvm_object *uobj, voff_t offset, int flags) 528 { 529 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 530 531 if (__predict_true(aobj->u_freelist == VM_NFREELIST)) 532 return uvm_pagealloc(uobj, offset, NULL, flags); 533 else 534 return uvm_pagealloc_strat(uobj, offset, NULL, flags, 535 UVM_PGA_STRAT_ONLY, aobj->u_freelist); 536 } 537 538 /* 539 * uao_init: set up aobj pager subsystem 540 * 541 * => called at boot time from uvm_pager_init() 542 */ 543 544 void 545 uao_init(void) 546 { 547 static int uao_initialized; 548 549 if (uao_initialized) 550 return; 551 uao_initialized = true; 552 LIST_INIT(&uao_list); 553 mutex_init(&uao_list_lock, MUTEX_DEFAULT, IPL_NONE); 554 pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 555 0, 0, 0, "uaoeltpl", NULL, IPL_VM); 556 } 557 558 /* 559 * uao_reference: hold a reference to an anonymous UVM object. 560 */ 561 void 562 uao_reference(struct uvm_object *uobj) 563 { 564 /* Kernel object is persistent. */ 565 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) { 566 return; 567 } 568 atomic_inc_uint(&uobj->uo_refs); 569 } 570 571 /* 572 * uao_detach: drop a reference to an anonymous UVM object. 573 */ 574 void 575 uao_detach(struct uvm_object *uobj) 576 { 577 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 578 struct vm_page *pg; 579 580 UVMHIST_FUNC("uao_detach"); UVMHIST_CALLED(maphist); 581 582 /* 583 * Detaching from kernel object is a NOP. 584 */ 585 586 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 587 return; 588 589 /* 590 * Drop the reference. If it was the last one, destroy the object. 591 */ 592 593 KASSERT(uobj->uo_refs > 0); 594 UVMHIST_LOG(maphist," (uobj=0x%#jx) ref=%jd", 595 (uintptr_t)uobj, uobj->uo_refs, 0, 0); 596 if (atomic_dec_uint_nv(&uobj->uo_refs) > 0) { 597 UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0); 598 return; 599 } 600 601 /* 602 * Remove the aobj from the global list. 603 */ 604 605 mutex_enter(&uao_list_lock); 606 LIST_REMOVE(aobj, u_list); 607 mutex_exit(&uao_list_lock); 608 609 /* 610 * Free all the pages left in the aobj. For each page, when the 611 * page is no longer busy (and thus after any disk I/O that it is 612 * involved in is complete), release any swap resources and free 613 * the page itself. 614 */ 615 616 mutex_enter(uobj->vmobjlock); 617 mutex_enter(&uvm_pageqlock); 618 while ((pg = TAILQ_FIRST(&uobj->memq)) != NULL) { 619 pmap_page_protect(pg, VM_PROT_NONE); 620 if (pg->flags & PG_BUSY) { 621 pg->flags |= PG_WANTED; 622 mutex_exit(&uvm_pageqlock); 623 UVM_UNLOCK_AND_WAIT(pg, uobj->vmobjlock, false, 624 "uao_det", 0); 625 mutex_enter(uobj->vmobjlock); 626 mutex_enter(&uvm_pageqlock); 627 continue; 628 } 629 uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT); 630 uvm_pagefree(pg); 631 } 632 mutex_exit(&uvm_pageqlock); 633 634 /* 635 * Finally, free the anonymous UVM object itself. 636 */ 637 638 uao_free(aobj); 639 } 640 641 /* 642 * uao_put: flush pages out of a uvm object 643 * 644 * => object should be locked by caller. we may _unlock_ the object 645 * if (and only if) we need to clean a page (PGO_CLEANIT). 646 * XXXJRT Currently, however, we don't. In the case of cleaning 647 * XXXJRT a page, we simply just deactivate it. Should probably 648 * XXXJRT handle this better, in the future (although "flushing" 649 * XXXJRT anonymous memory isn't terribly important). 650 * => if PGO_CLEANIT is not set, then we will neither unlock the object 651 * or block. 652 * => if PGO_ALLPAGE is set, then all pages in the object are valid targets 653 * for flushing. 654 * => NOTE: we rely on the fact that the object's memq is a TAILQ and 655 * that new pages are inserted on the tail end of the list. thus, 656 * we can make a complete pass through the object in one go by starting 657 * at the head and working towards the tail (new pages are put in 658 * front of us). 659 * => NOTE: we are allowed to lock the page queues, so the caller 660 * must not be holding the lock on them [e.g. pagedaemon had 661 * better not call us with the queues locked] 662 * => we return 0 unless we encountered some sort of I/O error 663 * XXXJRT currently never happens, as we never directly initiate 664 * XXXJRT I/O 665 * 666 * note on page traversal: 667 * we can traverse the pages in an object either by going down the 668 * linked list in "uobj->memq", or we can go over the address range 669 * by page doing hash table lookups for each address. depending 670 * on how many pages are in the object it may be cheaper to do one 671 * or the other. we set "by_list" to true if we are using memq. 672 * if the cost of a hash lookup was equal to the cost of the list 673 * traversal we could compare the number of pages in the start->stop 674 * range to the total number of pages in the object. however, it 675 * seems that a hash table lookup is more expensive than the linked 676 * list traversal, so we multiply the number of pages in the 677 * start->stop range by a penalty which we define below. 678 */ 679 680 static int 681 uao_put(struct uvm_object *uobj, voff_t start, voff_t stop, int flags) 682 { 683 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 684 struct vm_page *pg, *nextpg, curmp, endmp; 685 bool by_list; 686 voff_t curoff; 687 UVMHIST_FUNC("uao_put"); UVMHIST_CALLED(maphist); 688 689 KASSERT(mutex_owned(uobj->vmobjlock)); 690 691 curoff = 0; 692 if (flags & PGO_ALLPAGES) { 693 start = 0; 694 stop = aobj->u_pages << PAGE_SHIFT; 695 by_list = true; /* always go by the list */ 696 } else { 697 start = trunc_page(start); 698 if (stop == 0) { 699 stop = aobj->u_pages << PAGE_SHIFT; 700 } else { 701 stop = round_page(stop); 702 } 703 if (stop > (uint64_t)(aobj->u_pages << PAGE_SHIFT)) { 704 printf("uao_put: strange, got an out of range " 705 "flush 0x%jx > 0x%jx (fixed)\n", 706 (uintmax_t)stop, 707 (uintmax_t)(aobj->u_pages << PAGE_SHIFT)); 708 stop = aobj->u_pages << PAGE_SHIFT; 709 } 710 by_list = (uobj->uo_npages <= 711 ((stop - start) >> PAGE_SHIFT) * UVM_PAGE_TREE_PENALTY); 712 } 713 UVMHIST_LOG(maphist, 714 " flush start=0x%jx, stop=0x%jx, by_list=%jd, flags=0x%jx", 715 start, stop, by_list, flags); 716 717 /* 718 * Don't need to do any work here if we're not freeing 719 * or deactivating pages. 720 */ 721 722 if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) { 723 mutex_exit(uobj->vmobjlock); 724 return 0; 725 } 726 727 /* 728 * Initialize the marker pages. See the comment in 729 * genfs_putpages() also. 730 */ 731 732 curmp.flags = PG_MARKER; 733 endmp.flags = PG_MARKER; 734 735 /* 736 * now do it. note: we must update nextpg in the body of loop or we 737 * will get stuck. we need to use nextpg if we'll traverse the list 738 * because we may free "pg" before doing the next loop. 739 */ 740 741 if (by_list) { 742 TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq.queue); 743 nextpg = TAILQ_FIRST(&uobj->memq); 744 } else { 745 curoff = start; 746 nextpg = NULL; /* Quell compiler warning */ 747 } 748 749 /* locked: uobj */ 750 for (;;) { 751 if (by_list) { 752 pg = nextpg; 753 if (pg == &endmp) 754 break; 755 nextpg = TAILQ_NEXT(pg, listq.queue); 756 if (pg->flags & PG_MARKER) 757 continue; 758 if (pg->offset < start || pg->offset >= stop) 759 continue; 760 } else { 761 if (curoff < stop) { 762 pg = uvm_pagelookup(uobj, curoff); 763 curoff += PAGE_SIZE; 764 } else 765 break; 766 if (pg == NULL) 767 continue; 768 } 769 770 /* 771 * wait and try again if the page is busy. 772 */ 773 774 if (pg->flags & PG_BUSY) { 775 if (by_list) { 776 TAILQ_INSERT_BEFORE(pg, &curmp, listq.queue); 777 } 778 pg->flags |= PG_WANTED; 779 UVM_UNLOCK_AND_WAIT(pg, uobj->vmobjlock, 0, 780 "uao_put", 0); 781 mutex_enter(uobj->vmobjlock); 782 if (by_list) { 783 nextpg = TAILQ_NEXT(&curmp, listq.queue); 784 TAILQ_REMOVE(&uobj->memq, &curmp, 785 listq.queue); 786 } else 787 curoff -= PAGE_SIZE; 788 continue; 789 } 790 791 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 792 793 /* 794 * XXX In these first 3 cases, we always just 795 * XXX deactivate the page. We may want to 796 * XXX handle the different cases more specifically 797 * XXX in the future. 798 */ 799 800 case PGO_CLEANIT|PGO_FREE: 801 case PGO_CLEANIT|PGO_DEACTIVATE: 802 case PGO_DEACTIVATE: 803 deactivate_it: 804 mutex_enter(&uvm_pageqlock); 805 /* skip the page if it's wired */ 806 if (pg->wire_count == 0) { 807 uvm_pagedeactivate(pg); 808 } 809 mutex_exit(&uvm_pageqlock); 810 break; 811 812 case PGO_FREE: 813 /* 814 * If there are multiple references to 815 * the object, just deactivate the page. 816 */ 817 818 if (uobj->uo_refs > 1) 819 goto deactivate_it; 820 821 /* 822 * free the swap slot and the page. 823 */ 824 825 pmap_page_protect(pg, VM_PROT_NONE); 826 827 /* 828 * freeing swapslot here is not strictly necessary. 829 * however, leaving it here doesn't save much 830 * because we need to update swap accounting anyway. 831 */ 832 833 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 834 mutex_enter(&uvm_pageqlock); 835 uvm_pagefree(pg); 836 mutex_exit(&uvm_pageqlock); 837 break; 838 839 default: 840 panic("%s: impossible", __func__); 841 } 842 } 843 if (by_list) { 844 TAILQ_REMOVE(&uobj->memq, &endmp, listq.queue); 845 } 846 mutex_exit(uobj->vmobjlock); 847 return 0; 848 } 849 850 /* 851 * uao_get: fetch me a page 852 * 853 * we have three cases: 854 * 1: page is resident -> just return the page. 855 * 2: page is zero-fill -> allocate a new page and zero it. 856 * 3: page is swapped out -> fetch the page from swap. 857 * 858 * cases 1 and 2 can be handled with PGO_LOCKED, case 3 cannot. 859 * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES), 860 * then we will need to return EBUSY. 861 * 862 * => prefer map unlocked (not required) 863 * => object must be locked! we will _unlock_ it before starting any I/O. 864 * => flags: PGO_ALLPAGES: get all of the pages 865 * PGO_LOCKED: fault data structures are locked 866 * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx] 867 * => NOTE: caller must check for released pages!! 868 */ 869 870 static int 871 uao_get(struct uvm_object *uobj, voff_t offset, struct vm_page **pps, 872 int *npagesp, int centeridx, vm_prot_t access_type, int advice, int flags) 873 { 874 voff_t current_offset; 875 struct vm_page *ptmp = NULL; /* Quell compiler warning */ 876 int lcv, gotpages, maxpages, swslot, pageidx; 877 bool done; 878 UVMHIST_FUNC("uao_get"); UVMHIST_CALLED(pdhist); 879 880 UVMHIST_LOG(pdhist, "aobj=%#jx offset=%jd, flags=%jd", 881 (uintptr_t)uobj, offset, flags,0); 882 883 /* 884 * get number of pages 885 */ 886 887 maxpages = *npagesp; 888 889 /* 890 * step 1: handled the case where fault data structures are locked. 891 */ 892 893 if (flags & PGO_LOCKED) { 894 895 /* 896 * step 1a: get pages that are already resident. only do 897 * this if the data structures are locked (i.e. the first 898 * time through). 899 */ 900 901 done = true; /* be optimistic */ 902 gotpages = 0; /* # of pages we got so far */ 903 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 904 lcv++, current_offset += PAGE_SIZE) { 905 /* do we care about this page? if not, skip it */ 906 if (pps[lcv] == PGO_DONTCARE) 907 continue; 908 ptmp = uvm_pagelookup(uobj, current_offset); 909 910 /* 911 * if page is new, attempt to allocate the page, 912 * zero-fill'd. 913 */ 914 915 if (ptmp == NULL && uao_find_swslot(uobj, 916 current_offset >> PAGE_SHIFT) == 0) { 917 ptmp = uao_pagealloc(uobj, current_offset, 918 UVM_FLAG_COLORMATCH|UVM_PGA_ZERO); 919 if (ptmp) { 920 /* new page */ 921 ptmp->flags &= ~(PG_FAKE); 922 ptmp->pqflags |= PQ_AOBJ; 923 goto gotpage; 924 } 925 } 926 927 /* 928 * to be useful must get a non-busy page 929 */ 930 931 if (ptmp == NULL || (ptmp->flags & PG_BUSY) != 0) { 932 if (lcv == centeridx || 933 (flags & PGO_ALLPAGES) != 0) 934 /* need to do a wait or I/O! */ 935 done = false; 936 continue; 937 } 938 939 /* 940 * useful page: busy/lock it and plug it in our 941 * result array 942 */ 943 944 /* caller must un-busy this page */ 945 ptmp->flags |= PG_BUSY; 946 UVM_PAGE_OWN(ptmp, "uao_get1"); 947 gotpage: 948 pps[lcv] = ptmp; 949 gotpages++; 950 } 951 952 /* 953 * step 1b: now we've either done everything needed or we 954 * to unlock and do some waiting or I/O. 955 */ 956 957 UVMHIST_LOG(pdhist, "<- done (done=%jd)", done, 0,0,0); 958 *npagesp = gotpages; 959 if (done) 960 return 0; 961 else 962 return EBUSY; 963 } 964 965 /* 966 * step 2: get non-resident or busy pages. 967 * object is locked. data structures are unlocked. 968 */ 969 970 if ((flags & PGO_SYNCIO) == 0) { 971 goto done; 972 } 973 974 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 975 lcv++, current_offset += PAGE_SIZE) { 976 977 /* 978 * - skip over pages we've already gotten or don't want 979 * - skip over pages we don't _have_ to get 980 */ 981 982 if (pps[lcv] != NULL || 983 (lcv != centeridx && (flags & PGO_ALLPAGES) == 0)) 984 continue; 985 986 pageidx = current_offset >> PAGE_SHIFT; 987 988 /* 989 * we have yet to locate the current page (pps[lcv]). we 990 * first look for a page that is already at the current offset. 991 * if we find a page, we check to see if it is busy or 992 * released. if that is the case, then we sleep on the page 993 * until it is no longer busy or released and repeat the lookup. 994 * if the page we found is neither busy nor released, then we 995 * busy it (so we own it) and plug it into pps[lcv]. this 996 * 'break's the following while loop and indicates we are 997 * ready to move on to the next page in the "lcv" loop above. 998 * 999 * if we exit the while loop with pps[lcv] still set to NULL, 1000 * then it means that we allocated a new busy/fake/clean page 1001 * ptmp in the object and we need to do I/O to fill in the data. 1002 */ 1003 1004 /* top of "pps" while loop */ 1005 while (pps[lcv] == NULL) { 1006 /* look for a resident page */ 1007 ptmp = uvm_pagelookup(uobj, current_offset); 1008 1009 /* not resident? allocate one now (if we can) */ 1010 if (ptmp == NULL) { 1011 1012 ptmp = uao_pagealloc(uobj, current_offset, 0); 1013 1014 /* out of RAM? */ 1015 if (ptmp == NULL) { 1016 mutex_exit(uobj->vmobjlock); 1017 UVMHIST_LOG(pdhist, 1018 "sleeping, ptmp == NULL\n",0,0,0,0); 1019 uvm_wait("uao_getpage"); 1020 mutex_enter(uobj->vmobjlock); 1021 continue; 1022 } 1023 1024 /* 1025 * safe with PQ's unlocked: because we just 1026 * alloc'd the page 1027 */ 1028 1029 ptmp->pqflags |= PQ_AOBJ; 1030 1031 /* 1032 * got new page ready for I/O. break pps while 1033 * loop. pps[lcv] is still NULL. 1034 */ 1035 1036 break; 1037 } 1038 1039 /* page is there, see if we need to wait on it */ 1040 if ((ptmp->flags & PG_BUSY) != 0) { 1041 ptmp->flags |= PG_WANTED; 1042 UVMHIST_LOG(pdhist, 1043 "sleeping, ptmp->flags 0x%jx\n", 1044 ptmp->flags,0,0,0); 1045 UVM_UNLOCK_AND_WAIT(ptmp, uobj->vmobjlock, 1046 false, "uao_get", 0); 1047 mutex_enter(uobj->vmobjlock); 1048 continue; 1049 } 1050 1051 /* 1052 * if we get here then the page has become resident and 1053 * unbusy between steps 1 and 2. we busy it now (so we 1054 * own it) and set pps[lcv] (so that we exit the while 1055 * loop). 1056 */ 1057 1058 /* we own it, caller must un-busy */ 1059 ptmp->flags |= PG_BUSY; 1060 UVM_PAGE_OWN(ptmp, "uao_get2"); 1061 pps[lcv] = ptmp; 1062 } 1063 1064 /* 1065 * if we own the valid page at the correct offset, pps[lcv] will 1066 * point to it. nothing more to do except go to the next page. 1067 */ 1068 1069 if (pps[lcv]) 1070 continue; /* next lcv */ 1071 1072 /* 1073 * we have a "fake/busy/clean" page that we just allocated. 1074 * do the needed "i/o", either reading from swap or zeroing. 1075 */ 1076 1077 swslot = uao_find_swslot(uobj, pageidx); 1078 1079 /* 1080 * just zero the page if there's nothing in swap. 1081 */ 1082 1083 if (swslot == 0) { 1084 1085 /* 1086 * page hasn't existed before, just zero it. 1087 */ 1088 1089 uvm_pagezero(ptmp); 1090 } else { 1091 #if defined(VMSWAP) 1092 int error; 1093 1094 UVMHIST_LOG(pdhist, "pagein from swslot %jd", 1095 swslot, 0,0,0); 1096 1097 /* 1098 * page in the swapped-out page. 1099 * unlock object for i/o, relock when done. 1100 */ 1101 1102 mutex_exit(uobj->vmobjlock); 1103 error = uvm_swap_get(ptmp, swslot, PGO_SYNCIO); 1104 mutex_enter(uobj->vmobjlock); 1105 1106 /* 1107 * I/O done. check for errors. 1108 */ 1109 1110 if (error != 0) { 1111 UVMHIST_LOG(pdhist, "<- done (error=%jd)", 1112 error,0,0,0); 1113 if (ptmp->flags & PG_WANTED) 1114 wakeup(ptmp); 1115 1116 /* 1117 * remove the swap slot from the aobj 1118 * and mark the aobj as having no real slot. 1119 * don't free the swap slot, thus preventing 1120 * it from being used again. 1121 */ 1122 1123 swslot = uao_set_swslot(uobj, pageidx, 1124 SWSLOT_BAD); 1125 if (swslot > 0) { 1126 uvm_swap_markbad(swslot, 1); 1127 } 1128 1129 mutex_enter(&uvm_pageqlock); 1130 uvm_pagefree(ptmp); 1131 mutex_exit(&uvm_pageqlock); 1132 mutex_exit(uobj->vmobjlock); 1133 return error; 1134 } 1135 #else /* defined(VMSWAP) */ 1136 panic("%s: pagein", __func__); 1137 #endif /* defined(VMSWAP) */ 1138 } 1139 1140 if ((access_type & VM_PROT_WRITE) == 0) { 1141 ptmp->flags |= PG_CLEAN; 1142 pmap_clear_modify(ptmp); 1143 } 1144 1145 /* 1146 * we got the page! clear the fake flag (indicates valid 1147 * data now in page) and plug into our result array. note 1148 * that page is still busy. 1149 * 1150 * it is the callers job to: 1151 * => check if the page is released 1152 * => unbusy the page 1153 * => activate the page 1154 */ 1155 1156 ptmp->flags &= ~PG_FAKE; 1157 pps[lcv] = ptmp; 1158 } 1159 1160 /* 1161 * finally, unlock object and return. 1162 */ 1163 1164 done: 1165 mutex_exit(uobj->vmobjlock); 1166 UVMHIST_LOG(pdhist, "<- done (OK)",0,0,0,0); 1167 return 0; 1168 } 1169 1170 #if defined(VMSWAP) 1171 1172 /* 1173 * uao_dropswap: release any swap resources from this aobj page. 1174 * 1175 * => aobj must be locked or have a reference count of 0. 1176 */ 1177 1178 void 1179 uao_dropswap(struct uvm_object *uobj, int pageidx) 1180 { 1181 int slot; 1182 1183 slot = uao_set_swslot(uobj, pageidx, 0); 1184 if (slot) { 1185 uvm_swap_free(slot, 1); 1186 } 1187 } 1188 1189 /* 1190 * page in every page in every aobj that is paged-out to a range of swslots. 1191 * 1192 * => nothing should be locked. 1193 * => returns true if pagein was aborted due to lack of memory. 1194 */ 1195 1196 bool 1197 uao_swap_off(int startslot, int endslot) 1198 { 1199 struct uvm_aobj *aobj; 1200 1201 /* 1202 * Walk the list of all anonymous UVM objects. Grab the first. 1203 */ 1204 mutex_enter(&uao_list_lock); 1205 if ((aobj = LIST_FIRST(&uao_list)) == NULL) { 1206 mutex_exit(&uao_list_lock); 1207 return false; 1208 } 1209 uao_reference(&aobj->u_obj); 1210 1211 do { 1212 struct uvm_aobj *nextaobj; 1213 bool rv; 1214 1215 /* 1216 * Prefetch the next object and immediately hold a reference 1217 * on it, so neither the current nor the next entry could 1218 * disappear while we are iterating. 1219 */ 1220 if ((nextaobj = LIST_NEXT(aobj, u_list)) != NULL) { 1221 uao_reference(&nextaobj->u_obj); 1222 } 1223 mutex_exit(&uao_list_lock); 1224 1225 /* 1226 * Page in all pages in the swap slot range. 1227 */ 1228 mutex_enter(aobj->u_obj.vmobjlock); 1229 rv = uao_pagein(aobj, startslot, endslot); 1230 mutex_exit(aobj->u_obj.vmobjlock); 1231 1232 /* Drop the reference of the current object. */ 1233 uao_detach(&aobj->u_obj); 1234 if (rv) { 1235 if (nextaobj) { 1236 uao_detach(&nextaobj->u_obj); 1237 } 1238 return rv; 1239 } 1240 1241 aobj = nextaobj; 1242 mutex_enter(&uao_list_lock); 1243 } while (aobj); 1244 1245 mutex_exit(&uao_list_lock); 1246 return false; 1247 } 1248 1249 /* 1250 * page in any pages from aobj in the given range. 1251 * 1252 * => aobj must be locked and is returned locked. 1253 * => returns true if pagein was aborted due to lack of memory. 1254 */ 1255 static bool 1256 uao_pagein(struct uvm_aobj *aobj, int startslot, int endslot) 1257 { 1258 bool rv; 1259 1260 if (UAO_USES_SWHASH(aobj)) { 1261 struct uao_swhash_elt *elt; 1262 int buck; 1263 1264 restart: 1265 for (buck = aobj->u_swhashmask; buck >= 0; buck--) { 1266 for (elt = LIST_FIRST(&aobj->u_swhash[buck]); 1267 elt != NULL; 1268 elt = LIST_NEXT(elt, list)) { 1269 int i; 1270 1271 for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) { 1272 int slot = elt->slots[i]; 1273 1274 /* 1275 * if the slot isn't in range, skip it. 1276 */ 1277 1278 if (slot < startslot || 1279 slot >= endslot) { 1280 continue; 1281 } 1282 1283 /* 1284 * process the page, 1285 * the start over on this object 1286 * since the swhash elt 1287 * may have been freed. 1288 */ 1289 1290 rv = uao_pagein_page(aobj, 1291 UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i); 1292 if (rv) { 1293 return rv; 1294 } 1295 goto restart; 1296 } 1297 } 1298 } 1299 } else { 1300 int i; 1301 1302 for (i = 0; i < aobj->u_pages; i++) { 1303 int slot = aobj->u_swslots[i]; 1304 1305 /* 1306 * if the slot isn't in range, skip it 1307 */ 1308 1309 if (slot < startslot || slot >= endslot) { 1310 continue; 1311 } 1312 1313 /* 1314 * process the page. 1315 */ 1316 1317 rv = uao_pagein_page(aobj, i); 1318 if (rv) { 1319 return rv; 1320 } 1321 } 1322 } 1323 1324 return false; 1325 } 1326 1327 /* 1328 * uao_pagein_page: page in a single page from an anonymous UVM object. 1329 * 1330 * => Returns true if pagein was aborted due to lack of memory. 1331 * => Object must be locked and is returned locked. 1332 */ 1333 1334 static bool 1335 uao_pagein_page(struct uvm_aobj *aobj, int pageidx) 1336 { 1337 struct uvm_object *uobj = &aobj->u_obj; 1338 struct vm_page *pg; 1339 int rv, npages; 1340 1341 pg = NULL; 1342 npages = 1; 1343 1344 KASSERT(mutex_owned(uobj->vmobjlock)); 1345 rv = uao_get(uobj, (voff_t)pageidx << PAGE_SHIFT, &pg, &npages, 1346 0, VM_PROT_READ | VM_PROT_WRITE, 0, PGO_SYNCIO); 1347 1348 /* 1349 * relock and finish up. 1350 */ 1351 1352 mutex_enter(uobj->vmobjlock); 1353 switch (rv) { 1354 case 0: 1355 break; 1356 1357 case EIO: 1358 case ERESTART: 1359 1360 /* 1361 * nothing more to do on errors. 1362 * ERESTART can only mean that the anon was freed, 1363 * so again there's nothing to do. 1364 */ 1365 1366 return false; 1367 1368 default: 1369 return true; 1370 } 1371 1372 /* 1373 * ok, we've got the page now. 1374 * mark it as dirty, clear its swslot and un-busy it. 1375 */ 1376 uao_dropswap(&aobj->u_obj, pageidx); 1377 1378 /* 1379 * make sure it's on a page queue. 1380 */ 1381 mutex_enter(&uvm_pageqlock); 1382 if (pg->wire_count == 0) 1383 uvm_pageenqueue(pg); 1384 mutex_exit(&uvm_pageqlock); 1385 1386 if (pg->flags & PG_WANTED) { 1387 wakeup(pg); 1388 } 1389 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_CLEAN|PG_FAKE); 1390 UVM_PAGE_OWN(pg, NULL); 1391 1392 return false; 1393 } 1394 1395 /* 1396 * uao_dropswap_range: drop swapslots in the range. 1397 * 1398 * => aobj must be locked and is returned locked. 1399 * => start is inclusive. end is exclusive. 1400 */ 1401 1402 void 1403 uao_dropswap_range(struct uvm_object *uobj, voff_t start, voff_t end) 1404 { 1405 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 1406 int swpgonlydelta = 0; 1407 1408 KASSERT(mutex_owned(uobj->vmobjlock)); 1409 1410 if (end == 0) { 1411 end = INT64_MAX; 1412 } 1413 1414 if (UAO_USES_SWHASH(aobj)) { 1415 int i, hashbuckets = aobj->u_swhashmask + 1; 1416 voff_t taghi; 1417 voff_t taglo; 1418 1419 taglo = UAO_SWHASH_ELT_TAG(start); 1420 taghi = UAO_SWHASH_ELT_TAG(end); 1421 1422 for (i = 0; i < hashbuckets; i++) { 1423 struct uao_swhash_elt *elt, *next; 1424 1425 for (elt = LIST_FIRST(&aobj->u_swhash[i]); 1426 elt != NULL; 1427 elt = next) { 1428 int startidx, endidx; 1429 int j; 1430 1431 next = LIST_NEXT(elt, list); 1432 1433 if (elt->tag < taglo || taghi < elt->tag) { 1434 continue; 1435 } 1436 1437 if (elt->tag == taglo) { 1438 startidx = 1439 UAO_SWHASH_ELT_PAGESLOT_IDX(start); 1440 } else { 1441 startidx = 0; 1442 } 1443 1444 if (elt->tag == taghi) { 1445 endidx = 1446 UAO_SWHASH_ELT_PAGESLOT_IDX(end); 1447 } else { 1448 endidx = UAO_SWHASH_CLUSTER_SIZE; 1449 } 1450 1451 for (j = startidx; j < endidx; j++) { 1452 int slot = elt->slots[j]; 1453 1454 KASSERT(uvm_pagelookup(&aobj->u_obj, 1455 (UAO_SWHASH_ELT_PAGEIDX_BASE(elt) 1456 + j) << PAGE_SHIFT) == NULL); 1457 if (slot > 0) { 1458 uvm_swap_free(slot, 1); 1459 swpgonlydelta++; 1460 KASSERT(elt->count > 0); 1461 elt->slots[j] = 0; 1462 elt->count--; 1463 } 1464 } 1465 1466 if (elt->count == 0) { 1467 LIST_REMOVE(elt, list); 1468 pool_put(&uao_swhash_elt_pool, elt); 1469 } 1470 } 1471 } 1472 } else { 1473 int i; 1474 1475 if (aobj->u_pages < end) { 1476 end = aobj->u_pages; 1477 } 1478 for (i = start; i < end; i++) { 1479 int slot = aobj->u_swslots[i]; 1480 1481 if (slot > 0) { 1482 uvm_swap_free(slot, 1); 1483 swpgonlydelta++; 1484 } 1485 } 1486 } 1487 1488 /* 1489 * adjust the counter of pages only in swap for all 1490 * the swap slots we've freed. 1491 */ 1492 1493 if (swpgonlydelta > 0) { 1494 mutex_enter(&uvm_swap_data_lock); 1495 KASSERT(uvmexp.swpgonly >= swpgonlydelta); 1496 uvmexp.swpgonly -= swpgonlydelta; 1497 mutex_exit(&uvm_swap_data_lock); 1498 } 1499 } 1500 1501 #endif /* defined(VMSWAP) */ 1502