1 /* $OpenBSD: uvm_aobj.c,v 1.103 2021/12/29 20:22:06 mpi Exp $ */ 2 /* $NetBSD: uvm_aobj.c,v 1.39 2001/02/18 21:19:08 chs Exp $ */ 3 4 /* 5 * Copyright (c) 1998 Chuck Silvers, Charles D. Cranor and 6 * Washington University. 7 * All rights reserved. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 19 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 20 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 21 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 22 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 23 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 24 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 27 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 28 * 29 * from: Id: uvm_aobj.c,v 1.1.2.5 1998/02/06 05:14:38 chs Exp 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/param.h> 41 #include <sys/systm.h> 42 #include <sys/malloc.h> 43 #include <sys/kernel.h> 44 #include <sys/pool.h> 45 #include <sys/stdint.h> 46 #include <sys/atomic.h> 47 48 #include <uvm/uvm.h> 49 50 /* 51 * An anonymous UVM object (aobj) manages anonymous-memory. In addition to 52 * keeping the list of resident pages, it may also keep a list of allocated 53 * swap blocks. Depending on the size of the object, this list is either 54 * stored in an array (small objects) or in a hash table (large objects). 55 */ 56 57 /* 58 * Note: for hash tables, we break the address space of the aobj into blocks 59 * of UAO_SWHASH_CLUSTER_SIZE pages, which shall be a power of two. 60 */ 61 #define UAO_SWHASH_CLUSTER_SHIFT 4 62 #define UAO_SWHASH_CLUSTER_SIZE (1 << UAO_SWHASH_CLUSTER_SHIFT) 63 64 /* Get the "tag" for this page index. */ 65 #define UAO_SWHASH_ELT_TAG(idx) ((idx) >> UAO_SWHASH_CLUSTER_SHIFT) 66 #define UAO_SWHASH_ELT_PAGESLOT_IDX(idx) \ 67 ((idx) & (UAO_SWHASH_CLUSTER_SIZE - 1)) 68 69 /* Given an ELT and a page index, find the swap slot. */ 70 #define UAO_SWHASH_ELT_PAGESLOT(elt, idx) \ 71 ((elt)->slots[UAO_SWHASH_ELT_PAGESLOT_IDX(idx)]) 72 73 /* Given an ELT, return its pageidx base. */ 74 #define UAO_SWHASH_ELT_PAGEIDX_BASE(elt) \ 75 ((elt)->tag << UAO_SWHASH_CLUSTER_SHIFT) 76 77 /* The hash function. */ 78 #define UAO_SWHASH_HASH(aobj, idx) \ 79 (&(aobj)->u_swhash[(((idx) >> UAO_SWHASH_CLUSTER_SHIFT) \ 80 & (aobj)->u_swhashmask)]) 81 82 /* 83 * The threshold which determines whether we will use an array or a 84 * hash table to store the list of allocated swap blocks. 85 */ 86 #define UAO_SWHASH_THRESHOLD (UAO_SWHASH_CLUSTER_SIZE * 4) 87 #define UAO_USES_SWHASH(aobj) \ 88 ((aobj)->u_pages > UAO_SWHASH_THRESHOLD) 89 90 /* The number of buckets in a hash, with an upper bound. */ 91 #define UAO_SWHASH_MAXBUCKETS 256 92 #define UAO_SWHASH_BUCKETS(pages) \ 93 (min((pages) >> UAO_SWHASH_CLUSTER_SHIFT, UAO_SWHASH_MAXBUCKETS)) 94 95 96 /* 97 * uao_swhash_elt: when a hash table is being used, this structure defines 98 * the format of an entry in the bucket list. 99 */ 100 struct uao_swhash_elt { 101 LIST_ENTRY(uao_swhash_elt) list; /* the hash list */ 102 voff_t tag; /* our 'tag' */ 103 int count; /* our number of active slots */ 104 int slots[UAO_SWHASH_CLUSTER_SIZE]; /* the slots */ 105 }; 106 107 /* 108 * uao_swhash: the swap hash table structure 109 */ 110 LIST_HEAD(uao_swhash, uao_swhash_elt); 111 112 /* 113 * uao_swhash_elt_pool: pool of uao_swhash_elt structures 114 */ 115 struct pool uao_swhash_elt_pool; 116 117 /* 118 * uvm_aobj: the actual anon-backed uvm_object 119 * 120 * => the uvm_object is at the top of the structure, this allows 121 * (struct uvm_aobj *) == (struct uvm_object *) 122 * => only one of u_swslots and u_swhash is used in any given aobj 123 */ 124 struct uvm_aobj { 125 struct uvm_object u_obj; /* has: pgops, memt, #pages, #refs */ 126 int u_pages; /* number of pages in entire object */ 127 int u_flags; /* the flags (see uvm_aobj.h) */ 128 /* 129 * Either an array or hashtable (array of bucket heads) of 130 * offset -> swapslot mappings for the aobj. 131 */ 132 #define u_swslots u_swap.slot_array 133 #define u_swhash u_swap.slot_hash 134 union swslots { 135 int *slot_array; 136 struct uao_swhash *slot_hash; 137 } u_swap; 138 u_long u_swhashmask; /* mask for hashtable */ 139 LIST_ENTRY(uvm_aobj) u_list; /* global list of aobjs */ 140 }; 141 142 struct pool uvm_aobj_pool; 143 144 static struct uao_swhash_elt *uao_find_swhash_elt(struct uvm_aobj *, int, 145 boolean_t); 146 static int uao_find_swslot(struct uvm_object *, int); 147 static boolean_t uao_flush(struct uvm_object *, voff_t, 148 voff_t, int); 149 static void uao_free(struct uvm_aobj *); 150 static int uao_get(struct uvm_object *, voff_t, 151 vm_page_t *, int *, int, vm_prot_t, 152 int, int); 153 static boolean_t uao_pagein(struct uvm_aobj *, int, int); 154 static boolean_t uao_pagein_page(struct uvm_aobj *, int); 155 156 void uao_dropswap_range(struct uvm_object *, voff_t, voff_t); 157 void uao_shrink_flush(struct uvm_object *, int, int); 158 int uao_shrink_hash(struct uvm_object *, int); 159 int uao_shrink_array(struct uvm_object *, int); 160 int uao_shrink_convert(struct uvm_object *, int); 161 162 int uao_grow_hash(struct uvm_object *, int); 163 int uao_grow_array(struct uvm_object *, int); 164 int uao_grow_convert(struct uvm_object *, int); 165 166 /* 167 * aobj_pager 168 * 169 * note that some functions (e.g. put) are handled elsewhere 170 */ 171 const struct uvm_pagerops aobj_pager = { 172 .pgo_reference = uao_reference, 173 .pgo_detach = uao_detach, 174 .pgo_flush = uao_flush, 175 .pgo_get = uao_get, 176 }; 177 178 /* 179 * uao_list: global list of active aobjs, locked by uao_list_lock 180 * 181 * Lock ordering: generally the locking order is object lock, then list lock. 182 * in the case of swap off we have to iterate over the list, and thus the 183 * ordering is reversed. In that case we must use trylocking to prevent 184 * deadlock. 185 */ 186 static LIST_HEAD(aobjlist, uvm_aobj) uao_list = LIST_HEAD_INITIALIZER(uao_list); 187 static struct mutex uao_list_lock = MUTEX_INITIALIZER(IPL_MPFLOOR); 188 189 190 /* 191 * functions 192 */ 193 /* 194 * hash table/array related functions 195 */ 196 /* 197 * uao_find_swhash_elt: find (or create) a hash table entry for a page 198 * offset. 199 */ 200 static struct uao_swhash_elt * 201 uao_find_swhash_elt(struct uvm_aobj *aobj, int pageidx, boolean_t create) 202 { 203 struct uao_swhash *swhash; 204 struct uao_swhash_elt *elt; 205 voff_t page_tag; 206 207 swhash = UAO_SWHASH_HASH(aobj, pageidx); /* first hash to get bucket */ 208 page_tag = UAO_SWHASH_ELT_TAG(pageidx); /* tag to search for */ 209 210 /* 211 * now search the bucket for the requested tag 212 */ 213 LIST_FOREACH(elt, swhash, list) { 214 if (elt->tag == page_tag) 215 return elt; 216 } 217 218 if (!create) 219 return NULL; 220 221 /* 222 * allocate a new entry for the bucket and init/insert it in 223 */ 224 elt = pool_get(&uao_swhash_elt_pool, PR_NOWAIT | PR_ZERO); 225 /* 226 * XXX We cannot sleep here as the hash table might disappear 227 * from under our feet. And we run the risk of deadlocking 228 * the pagedeamon. In fact this code will only be called by 229 * the pagedaemon and allocation will only fail if we 230 * exhausted the pagedeamon reserve. In that case we're 231 * doomed anyway, so panic. 232 */ 233 if (elt == NULL) 234 panic("%s: can't allocate entry", __func__); 235 LIST_INSERT_HEAD(swhash, elt, list); 236 elt->tag = page_tag; 237 238 return elt; 239 } 240 241 /* 242 * uao_find_swslot: find the swap slot number for an aobj/pageidx 243 */ 244 inline static int 245 uao_find_swslot(struct uvm_object *uobj, int pageidx) 246 { 247 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 248 249 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 250 251 /* 252 * if noswap flag is set, then we never return a slot 253 */ 254 if (aobj->u_flags & UAO_FLAG_NOSWAP) 255 return 0; 256 257 /* 258 * if hashing, look in hash table. 259 */ 260 if (UAO_USES_SWHASH(aobj)) { 261 struct uao_swhash_elt *elt = 262 uao_find_swhash_elt(aobj, pageidx, FALSE); 263 264 if (elt) 265 return UAO_SWHASH_ELT_PAGESLOT(elt, pageidx); 266 else 267 return 0; 268 } 269 270 /* 271 * otherwise, look in the array 272 */ 273 return aobj->u_swslots[pageidx]; 274 } 275 276 /* 277 * uao_set_swslot: set the swap slot for a page in an aobj. 278 * 279 * => setting a slot to zero frees the slot 280 * => object must be locked by caller 281 * => we return the old slot number, or -1 if we failed to allocate 282 * memory to record the new slot number 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 int oldslot; 289 290 KASSERT(rw_write_held(uobj->vmobjlock) || uobj->uo_refs == 0); 291 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 292 293 /* 294 * if noswap flag is set, then we can't set a slot 295 */ 296 if (aobj->u_flags & UAO_FLAG_NOSWAP) { 297 if (slot == 0) 298 return 0; /* a clear is ok */ 299 300 /* but a set is not */ 301 printf("uao_set_swslot: uobj = %p\n", uobj); 302 panic("uao_set_swslot: attempt to set a slot on a NOSWAP object"); 303 } 304 305 /* 306 * are we using a hash table? if so, add it in the hash. 307 */ 308 if (UAO_USES_SWHASH(aobj)) { 309 /* 310 * Avoid allocating an entry just to free it again if 311 * the page had not swap slot in the first place, and 312 * we are freeing. 313 */ 314 struct uao_swhash_elt *elt = 315 uao_find_swhash_elt(aobj, pageidx, slot ? TRUE : FALSE); 316 if (elt == NULL) { 317 KASSERT(slot == 0); 318 return 0; 319 } 320 321 oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx); 322 UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot; 323 324 /* 325 * now adjust the elt's reference counter and free it if we've 326 * dropped it to zero. 327 */ 328 if (slot) { 329 if (oldslot == 0) 330 elt->count++; 331 } else { 332 if (oldslot) 333 elt->count--; 334 335 if (elt->count == 0) { 336 LIST_REMOVE(elt, list); 337 pool_put(&uao_swhash_elt_pool, elt); 338 } 339 } 340 } else { 341 /* we are using an array */ 342 oldslot = aobj->u_swslots[pageidx]; 343 aobj->u_swslots[pageidx] = slot; 344 } 345 return oldslot; 346 } 347 /* 348 * end of hash/array functions 349 */ 350 351 /* 352 * uao_free: free all resources held by an aobj, and then free the aobj 353 * 354 * => the aobj should be dead 355 */ 356 static void 357 uao_free(struct uvm_aobj *aobj) 358 { 359 struct uvm_object *uobj = &aobj->u_obj; 360 361 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 362 KASSERT(rw_write_held(uobj->vmobjlock)); 363 uao_dropswap_range(uobj, 0, 0); 364 rw_exit(uobj->vmobjlock); 365 366 if (UAO_USES_SWHASH(aobj)) { 367 /* 368 * free the hash table itself. 369 */ 370 hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ); 371 } else { 372 free(aobj->u_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int)); 373 } 374 375 /* 376 * finally free the aobj itself 377 */ 378 uvm_obj_destroy(uobj); 379 pool_put(&uvm_aobj_pool, aobj); 380 } 381 382 /* 383 * pager functions 384 */ 385 386 #ifdef TMPFS 387 /* 388 * Shrink an aobj to a given number of pages. The procedure is always the same: 389 * assess the necessity of data structure conversion (hash to array), secure 390 * resources, flush pages and drop swap slots. 391 * 392 */ 393 394 void 395 uao_shrink_flush(struct uvm_object *uobj, int startpg, int endpg) 396 { 397 KASSERT(startpg < endpg); 398 KASSERT(uobj->uo_refs == 1); 399 uao_flush(uobj, (voff_t)startpg << PAGE_SHIFT, 400 (voff_t)endpg << PAGE_SHIFT, PGO_FREE); 401 uao_dropswap_range(uobj, startpg, endpg); 402 } 403 404 int 405 uao_shrink_hash(struct uvm_object *uobj, int pages) 406 { 407 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 408 struct uao_swhash *new_swhash; 409 struct uao_swhash_elt *elt; 410 unsigned long new_hashmask; 411 int i; 412 413 KASSERT(UAO_USES_SWHASH(aobj)); 414 415 /* 416 * If the size of the hash table doesn't change, all we need to do is 417 * to adjust the page count. 418 */ 419 if (UAO_SWHASH_BUCKETS(aobj->u_pages) == UAO_SWHASH_BUCKETS(pages)) { 420 uao_shrink_flush(uobj, pages, aobj->u_pages); 421 aobj->u_pages = pages; 422 return 0; 423 } 424 425 new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ, 426 M_WAITOK | M_CANFAIL, &new_hashmask); 427 if (new_swhash == NULL) 428 return ENOMEM; 429 430 uao_shrink_flush(uobj, pages, aobj->u_pages); 431 432 /* 433 * Even though the hash table size is changing, the hash of the buckets 434 * we are interested in copying should not change. 435 */ 436 for (i = 0; i < UAO_SWHASH_BUCKETS(aobj->u_pages); i++) { 437 while (LIST_EMPTY(&aobj->u_swhash[i]) == 0) { 438 elt = LIST_FIRST(&aobj->u_swhash[i]); 439 LIST_REMOVE(elt, list); 440 LIST_INSERT_HEAD(&new_swhash[i], elt, list); 441 } 442 } 443 444 hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ); 445 446 aobj->u_swhash = new_swhash; 447 aobj->u_pages = pages; 448 aobj->u_swhashmask = new_hashmask; 449 450 return 0; 451 } 452 453 int 454 uao_shrink_convert(struct uvm_object *uobj, int pages) 455 { 456 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 457 struct uao_swhash_elt *elt; 458 int i, *new_swslots; 459 460 new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ, 461 M_WAITOK | M_CANFAIL | M_ZERO); 462 if (new_swslots == NULL) 463 return ENOMEM; 464 465 uao_shrink_flush(uobj, pages, aobj->u_pages); 466 467 /* Convert swap slots from hash to array. */ 468 for (i = 0; i < pages; i++) { 469 elt = uao_find_swhash_elt(aobj, i, FALSE); 470 if (elt != NULL) { 471 new_swslots[i] = UAO_SWHASH_ELT_PAGESLOT(elt, i); 472 if (new_swslots[i] != 0) 473 elt->count--; 474 if (elt->count == 0) { 475 LIST_REMOVE(elt, list); 476 pool_put(&uao_swhash_elt_pool, elt); 477 } 478 } 479 } 480 481 hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ); 482 483 aobj->u_swslots = new_swslots; 484 aobj->u_pages = pages; 485 486 return 0; 487 } 488 489 int 490 uao_shrink_array(struct uvm_object *uobj, int pages) 491 { 492 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 493 int i, *new_swslots; 494 495 new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ, 496 M_WAITOK | M_CANFAIL | M_ZERO); 497 if (new_swslots == NULL) 498 return ENOMEM; 499 500 uao_shrink_flush(uobj, pages, aobj->u_pages); 501 502 for (i = 0; i < pages; i++) 503 new_swslots[i] = aobj->u_swslots[i]; 504 505 free(aobj->u_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int)); 506 507 aobj->u_swslots = new_swslots; 508 aobj->u_pages = pages; 509 510 return 0; 511 } 512 513 int 514 uao_shrink(struct uvm_object *uobj, int pages) 515 { 516 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 517 518 KASSERT(pages < aobj->u_pages); 519 520 /* 521 * Distinguish between three possible cases: 522 * 1. aobj uses hash and must be converted to array. 523 * 2. aobj uses array and array size needs to be adjusted. 524 * 3. aobj uses hash and hash size needs to be adjusted. 525 */ 526 if (pages > UAO_SWHASH_THRESHOLD) 527 return uao_shrink_hash(uobj, pages); /* case 3 */ 528 else if (aobj->u_pages > UAO_SWHASH_THRESHOLD) 529 return uao_shrink_convert(uobj, pages); /* case 1 */ 530 else 531 return uao_shrink_array(uobj, pages); /* case 2 */ 532 } 533 534 /* 535 * Grow an aobj to a given number of pages. Right now we only adjust the swap 536 * slots. We could additionally handle page allocation directly, so that they 537 * don't happen through uvm_fault(). That would allow us to use another 538 * mechanism for the swap slots other than malloc(). It is thus mandatory that 539 * the caller of these functions does not allow faults to happen in case of 540 * growth error. 541 */ 542 int 543 uao_grow_array(struct uvm_object *uobj, int pages) 544 { 545 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 546 int i, *new_swslots; 547 548 KASSERT(aobj->u_pages <= UAO_SWHASH_THRESHOLD); 549 550 new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ, 551 M_WAITOK | M_CANFAIL | M_ZERO); 552 if (new_swslots == NULL) 553 return ENOMEM; 554 555 for (i = 0; i < aobj->u_pages; i++) 556 new_swslots[i] = aobj->u_swslots[i]; 557 558 free(aobj->u_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int)); 559 560 aobj->u_swslots = new_swslots; 561 aobj->u_pages = pages; 562 563 return 0; 564 } 565 566 int 567 uao_grow_hash(struct uvm_object *uobj, int pages) 568 { 569 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 570 struct uao_swhash *new_swhash; 571 struct uao_swhash_elt *elt; 572 unsigned long new_hashmask; 573 int i; 574 575 KASSERT(pages > UAO_SWHASH_THRESHOLD); 576 577 /* 578 * If the size of the hash table doesn't change, all we need to do is 579 * to adjust the page count. 580 */ 581 if (UAO_SWHASH_BUCKETS(aobj->u_pages) == UAO_SWHASH_BUCKETS(pages)) { 582 aobj->u_pages = pages; 583 return 0; 584 } 585 586 KASSERT(UAO_SWHASH_BUCKETS(aobj->u_pages) < UAO_SWHASH_BUCKETS(pages)); 587 588 new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ, 589 M_WAITOK | M_CANFAIL, &new_hashmask); 590 if (new_swhash == NULL) 591 return ENOMEM; 592 593 for (i = 0; i < UAO_SWHASH_BUCKETS(aobj->u_pages); i++) { 594 while (LIST_EMPTY(&aobj->u_swhash[i]) == 0) { 595 elt = LIST_FIRST(&aobj->u_swhash[i]); 596 LIST_REMOVE(elt, list); 597 LIST_INSERT_HEAD(&new_swhash[i], elt, list); 598 } 599 } 600 601 hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ); 602 603 aobj->u_swhash = new_swhash; 604 aobj->u_pages = pages; 605 aobj->u_swhashmask = new_hashmask; 606 607 return 0; 608 } 609 610 int 611 uao_grow_convert(struct uvm_object *uobj, int pages) 612 { 613 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 614 struct uao_swhash *new_swhash; 615 struct uao_swhash_elt *elt; 616 unsigned long new_hashmask; 617 int i, *old_swslots; 618 619 new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ, 620 M_WAITOK | M_CANFAIL, &new_hashmask); 621 if (new_swhash == NULL) 622 return ENOMEM; 623 624 /* Set these now, so we can use uao_find_swhash_elt(). */ 625 old_swslots = aobj->u_swslots; 626 aobj->u_swhash = new_swhash; 627 aobj->u_swhashmask = new_hashmask; 628 629 for (i = 0; i < aobj->u_pages; i++) { 630 if (old_swslots[i] != 0) { 631 elt = uao_find_swhash_elt(aobj, i, TRUE); 632 elt->count++; 633 UAO_SWHASH_ELT_PAGESLOT(elt, i) = old_swslots[i]; 634 } 635 } 636 637 free(old_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int)); 638 aobj->u_pages = pages; 639 640 return 0; 641 } 642 643 int 644 uao_grow(struct uvm_object *uobj, int pages) 645 { 646 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 647 648 KASSERT(pages > aobj->u_pages); 649 650 /* 651 * Distinguish between three possible cases: 652 * 1. aobj uses hash and hash size needs to be adjusted. 653 * 2. aobj uses array and array size needs to be adjusted. 654 * 3. aobj uses array and must be converted to hash. 655 */ 656 if (pages <= UAO_SWHASH_THRESHOLD) 657 return uao_grow_array(uobj, pages); /* case 2 */ 658 else if (aobj->u_pages > UAO_SWHASH_THRESHOLD) 659 return uao_grow_hash(uobj, pages); /* case 1 */ 660 else 661 return uao_grow_convert(uobj, pages); 662 } 663 #endif /* TMPFS */ 664 665 /* 666 * uao_create: create an aobj of the given size and return its uvm_object. 667 * 668 * => for normal use, flags are zero or UAO_FLAG_CANFAIL. 669 * => for the kernel object, the flags are: 670 * UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once) 671 * UAO_FLAG_KERNSWAP - enable swapping of kernel object (" ") 672 */ 673 struct uvm_object * 674 uao_create(vsize_t size, int flags) 675 { 676 static struct uvm_aobj kernel_object_store; 677 static struct rwlock bootstrap_kernel_object_lock; 678 static int kobj_alloced = 0; 679 int pages = round_page(size) >> PAGE_SHIFT; 680 struct uvm_aobj *aobj; 681 int refs; 682 683 /* 684 * Allocate a new aobj, unless kernel object is requested. 685 */ 686 if (flags & UAO_FLAG_KERNOBJ) { 687 KASSERT(!kobj_alloced); 688 aobj = &kernel_object_store; 689 aobj->u_pages = pages; 690 aobj->u_flags = UAO_FLAG_NOSWAP; 691 refs = UVM_OBJ_KERN; 692 kobj_alloced = UAO_FLAG_KERNOBJ; 693 } else if (flags & UAO_FLAG_KERNSWAP) { 694 KASSERT(kobj_alloced == UAO_FLAG_KERNOBJ); 695 aobj = &kernel_object_store; 696 kobj_alloced = UAO_FLAG_KERNSWAP; 697 } else { 698 aobj = pool_get(&uvm_aobj_pool, PR_WAITOK); 699 aobj->u_pages = pages; 700 aobj->u_flags = 0; 701 refs = 1; 702 } 703 704 /* 705 * allocate hash/array if necessary 706 */ 707 if (flags == 0 || (flags & (UAO_FLAG_KERNSWAP | UAO_FLAG_CANFAIL))) { 708 int mflags; 709 710 if (flags) 711 mflags = M_NOWAIT; 712 else 713 mflags = M_WAITOK; 714 715 /* allocate hash table or array depending on object size */ 716 if (UAO_USES_SWHASH(aobj)) { 717 aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), 718 M_UVMAOBJ, mflags, &aobj->u_swhashmask); 719 if (aobj->u_swhash == NULL) { 720 if (flags & UAO_FLAG_CANFAIL) { 721 pool_put(&uvm_aobj_pool, aobj); 722 return NULL; 723 } 724 panic("uao_create: hashinit swhash failed"); 725 } 726 } else { 727 aobj->u_swslots = mallocarray(pages, sizeof(int), 728 M_UVMAOBJ, mflags|M_ZERO); 729 if (aobj->u_swslots == NULL) { 730 if (flags & UAO_FLAG_CANFAIL) { 731 pool_put(&uvm_aobj_pool, aobj); 732 return NULL; 733 } 734 panic("uao_create: malloc swslots failed"); 735 } 736 } 737 738 if (flags & UAO_FLAG_KERNSWAP) { 739 aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */ 740 return &aobj->u_obj; 741 /* done! */ 742 } 743 } 744 745 /* 746 * Initialise UVM object. 747 */ 748 uvm_obj_init(&aobj->u_obj, &aobj_pager, refs); 749 if (flags & UAO_FLAG_KERNOBJ) { 750 /* Use a temporary static lock for kernel_object. */ 751 rw_init(&bootstrap_kernel_object_lock, "kobjlk"); 752 uvm_obj_setlock(&aobj->u_obj, &bootstrap_kernel_object_lock); 753 } 754 755 /* 756 * now that aobj is ready, add it to the global list 757 */ 758 mtx_enter(&uao_list_lock); 759 LIST_INSERT_HEAD(&uao_list, aobj, u_list); 760 mtx_leave(&uao_list_lock); 761 762 return &aobj->u_obj; 763 } 764 765 766 767 /* 768 * uao_init: set up aobj pager subsystem 769 * 770 * => called at boot time from uvm_pager_init() 771 */ 772 void 773 uao_init(void) 774 { 775 /* 776 * NOTE: Pages for this pool must not come from a pageable 777 * kernel map! 778 */ 779 pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 0, 780 IPL_NONE, PR_WAITOK, "uaoeltpl", NULL); 781 pool_init(&uvm_aobj_pool, sizeof(struct uvm_aobj), 0, 782 IPL_NONE, PR_WAITOK, "aobjpl", NULL); 783 } 784 785 /* 786 * uao_reference: hold a reference to an anonymous UVM object. 787 */ 788 void 789 uao_reference(struct uvm_object *uobj) 790 { 791 /* Kernel object is persistent. */ 792 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 793 return; 794 795 atomic_inc_int(&uobj->uo_refs); 796 } 797 798 799 /* 800 * uao_detach: drop a reference to an anonymous UVM object. 801 */ 802 void 803 uao_detach(struct uvm_object *uobj) 804 { 805 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 806 struct vm_page *pg; 807 808 /* 809 * Detaching from kernel_object is a NOP. 810 */ 811 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 812 return; 813 814 /* 815 * Drop the reference. If it was the last one, destroy the object. 816 */ 817 if (atomic_dec_int_nv(&uobj->uo_refs) > 0) { 818 return; 819 } 820 821 /* 822 * Remove the aobj from the global list. 823 */ 824 mtx_enter(&uao_list_lock); 825 LIST_REMOVE(aobj, u_list); 826 mtx_leave(&uao_list_lock); 827 828 /* 829 * Free all the pages left in the aobj. For each page, when the 830 * page is no longer busy (and thus after any disk I/O that it is 831 * involved in is complete), release any swap resources and free 832 * the page itself. 833 */ 834 rw_enter(uobj->vmobjlock, RW_WRITE); 835 while ((pg = RBT_ROOT(uvm_objtree, &uobj->memt)) != NULL) { 836 pmap_page_protect(pg, PROT_NONE); 837 if (pg->pg_flags & PG_BUSY) { 838 atomic_setbits_int(&pg->pg_flags, PG_WANTED); 839 rwsleep_nsec(pg, uobj->vmobjlock, PVM, "uao_det", 840 INFSLP); 841 continue; 842 } 843 uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT); 844 uvm_lock_pageq(); 845 uvm_pagefree(pg); 846 uvm_unlock_pageq(); 847 } 848 849 /* 850 * Finally, free the anonymous UVM object itself. 851 */ 852 uao_free(aobj); 853 } 854 855 /* 856 * uao_flush: flush pages out of a uvm object 857 * 858 * => if PGO_CLEANIT is not set, then we will not block. 859 * => if PGO_ALLPAGE is set, then all pages in the object are valid targets 860 * for flushing. 861 * => NOTE: we are allowed to lock the page queues, so the caller 862 * must not be holding the lock on them [e.g. pagedaemon had 863 * better not call us with the queues locked] 864 * => we return TRUE unless we encountered some sort of I/O error 865 * XXXJRT currently never happens, as we never directly initiate 866 * XXXJRT I/O 867 */ 868 boolean_t 869 uao_flush(struct uvm_object *uobj, voff_t start, voff_t stop, int flags) 870 { 871 struct uvm_aobj *aobj = (struct uvm_aobj *) uobj; 872 struct vm_page *pg; 873 voff_t curoff; 874 875 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 876 KASSERT(rw_write_held(uobj->vmobjlock)); 877 878 if (flags & PGO_ALLPAGES) { 879 start = 0; 880 stop = (voff_t)aobj->u_pages << PAGE_SHIFT; 881 } else { 882 start = trunc_page(start); 883 stop = round_page(stop); 884 if (stop > ((voff_t)aobj->u_pages << PAGE_SHIFT)) { 885 printf("uao_flush: strange, got an out of range " 886 "flush (fixed)\n"); 887 stop = (voff_t)aobj->u_pages << PAGE_SHIFT; 888 } 889 } 890 891 /* 892 * Don't need to do any work here if we're not freeing 893 * or deactivating pages. 894 */ 895 if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) { 896 return TRUE; 897 } 898 899 curoff = start; 900 for (;;) { 901 if (curoff < stop) { 902 pg = uvm_pagelookup(uobj, curoff); 903 curoff += PAGE_SIZE; 904 if (pg == NULL) 905 continue; 906 } else { 907 break; 908 } 909 910 /* Make sure page is unbusy, else wait for it. */ 911 if (pg->pg_flags & PG_BUSY) { 912 atomic_setbits_int(&pg->pg_flags, PG_WANTED); 913 rwsleep_nsec(pg, uobj->vmobjlock, PVM, "uaoflsh", 914 INFSLP); 915 curoff -= PAGE_SIZE; 916 continue; 917 } 918 919 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 920 /* 921 * XXX In these first 3 cases, we always just 922 * XXX deactivate the page. We may want to 923 * XXX handle the different cases more specifically 924 * XXX in the future. 925 */ 926 case PGO_CLEANIT|PGO_FREE: 927 /* FALLTHROUGH */ 928 case PGO_CLEANIT|PGO_DEACTIVATE: 929 /* FALLTHROUGH */ 930 case PGO_DEACTIVATE: 931 deactivate_it: 932 if (pg->wire_count != 0) 933 continue; 934 935 uvm_lock_pageq(); 936 pmap_page_protect(pg, PROT_NONE); 937 uvm_pagedeactivate(pg); 938 uvm_unlock_pageq(); 939 940 continue; 941 case PGO_FREE: 942 /* 943 * If there are multiple references to 944 * the object, just deactivate the page. 945 */ 946 if (uobj->uo_refs > 1) 947 goto deactivate_it; 948 949 /* XXX skip the page if it's wired */ 950 if (pg->wire_count != 0) 951 continue; 952 953 /* 954 * free the swap slot and the page. 955 */ 956 pmap_page_protect(pg, PROT_NONE); 957 958 /* 959 * freeing swapslot here is not strictly necessary. 960 * however, leaving it here doesn't save much 961 * because we need to update swap accounting anyway. 962 */ 963 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 964 uvm_lock_pageq(); 965 uvm_pagefree(pg); 966 uvm_unlock_pageq(); 967 968 continue; 969 default: 970 panic("uao_flush: weird flags"); 971 } 972 } 973 974 return TRUE; 975 } 976 977 /* 978 * uao_get: fetch me a page 979 * 980 * we have three cases: 981 * 1: page is resident -> just return the page. 982 * 2: page is zero-fill -> allocate a new page and zero it. 983 * 3: page is swapped out -> fetch the page from swap. 984 * 985 * cases 1 can be handled with PGO_LOCKED, cases 2 and 3 cannot. 986 * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES), 987 * then we will need to return VM_PAGER_UNLOCK. 988 * 989 * => flags: PGO_ALLPAGES: get all of the pages 990 * PGO_LOCKED: fault data structures are locked 991 * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx] 992 * => NOTE: caller must check for released pages!! 993 */ 994 static int 995 uao_get(struct uvm_object *uobj, voff_t offset, struct vm_page **pps, 996 int *npagesp, int centeridx, vm_prot_t access_type, int advice, int flags) 997 { 998 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 999 voff_t current_offset; 1000 vm_page_t ptmp; 1001 int lcv, gotpages, maxpages, swslot, rv, pageidx; 1002 boolean_t done; 1003 1004 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 1005 KASSERT(rw_write_held(uobj->vmobjlock)); 1006 1007 /* 1008 * get number of pages 1009 */ 1010 maxpages = *npagesp; 1011 1012 if (flags & PGO_LOCKED) { 1013 /* 1014 * step 1a: get pages that are already resident. only do 1015 * this if the data structures are locked (i.e. the first 1016 * time through). 1017 */ 1018 1019 done = TRUE; /* be optimistic */ 1020 gotpages = 0; /* # of pages we got so far */ 1021 1022 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 1023 lcv++, current_offset += PAGE_SIZE) { 1024 /* do we care about this page? if not, skip it */ 1025 if (pps[lcv] == PGO_DONTCARE) 1026 continue; 1027 1028 ptmp = uvm_pagelookup(uobj, current_offset); 1029 1030 /* 1031 * if page is new, attempt to allocate the page, 1032 * zero-fill'd. 1033 */ 1034 if (ptmp == NULL && uao_find_swslot(uobj, 1035 current_offset >> PAGE_SHIFT) == 0) { 1036 ptmp = uvm_pagealloc(uobj, current_offset, 1037 NULL, UVM_PGA_ZERO); 1038 if (ptmp) { 1039 /* new page */ 1040 atomic_clearbits_int(&ptmp->pg_flags, 1041 PG_BUSY|PG_FAKE); 1042 atomic_setbits_int(&ptmp->pg_flags, 1043 PQ_AOBJ); 1044 UVM_PAGE_OWN(ptmp, NULL); 1045 } 1046 } 1047 1048 /* 1049 * to be useful must get a non-busy page 1050 */ 1051 if (ptmp == NULL || 1052 (ptmp->pg_flags & PG_BUSY) != 0) { 1053 if (lcv == centeridx || 1054 (flags & PGO_ALLPAGES) != 0) 1055 /* need to do a wait or I/O! */ 1056 done = FALSE; 1057 continue; 1058 } 1059 1060 /* 1061 * useful page: plug it in our result array 1062 */ 1063 atomic_setbits_int(&ptmp->pg_flags, PG_BUSY); 1064 UVM_PAGE_OWN(ptmp, "uao_get1"); 1065 pps[lcv] = ptmp; 1066 gotpages++; 1067 1068 } 1069 1070 /* 1071 * step 1b: now we've either done everything needed or we 1072 * to unlock and do some waiting or I/O. 1073 */ 1074 *npagesp = gotpages; 1075 if (done) 1076 /* bingo! */ 1077 return VM_PAGER_OK; 1078 else 1079 /* EEK! Need to unlock and I/O */ 1080 return VM_PAGER_UNLOCK; 1081 } 1082 1083 /* 1084 * step 2: get non-resident or busy pages. 1085 * data structures are unlocked. 1086 */ 1087 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 1088 lcv++, current_offset += PAGE_SIZE) { 1089 /* 1090 * - skip over pages we've already gotten or don't want 1091 * - skip over pages we don't _have_ to get 1092 */ 1093 if (pps[lcv] != NULL || 1094 (lcv != centeridx && (flags & PGO_ALLPAGES) == 0)) 1095 continue; 1096 1097 pageidx = current_offset >> PAGE_SHIFT; 1098 1099 /* 1100 * we have yet to locate the current page (pps[lcv]). we 1101 * first look for a page that is already at the current offset. 1102 * if we find a page, we check to see if it is busy or 1103 * released. if that is the case, then we sleep on the page 1104 * until it is no longer busy or released and repeat the lookup. 1105 * if the page we found is neither busy nor released, then we 1106 * busy it (so we own it) and plug it into pps[lcv]. this 1107 * 'break's the following while loop and indicates we are 1108 * ready to move on to the next page in the "lcv" loop above. 1109 * 1110 * if we exit the while loop with pps[lcv] still set to NULL, 1111 * then it means that we allocated a new busy/fake/clean page 1112 * ptmp in the object and we need to do I/O to fill in the data. 1113 */ 1114 1115 /* top of "pps" while loop */ 1116 while (pps[lcv] == NULL) { 1117 /* look for a resident page */ 1118 ptmp = uvm_pagelookup(uobj, current_offset); 1119 1120 /* not resident? allocate one now (if we can) */ 1121 if (ptmp == NULL) { 1122 1123 ptmp = uvm_pagealloc(uobj, current_offset, 1124 NULL, 0); 1125 1126 /* out of RAM? */ 1127 if (ptmp == NULL) { 1128 rw_exit(uobj->vmobjlock); 1129 uvm_wait("uao_getpage"); 1130 rw_enter(uobj->vmobjlock, RW_WRITE); 1131 /* goto top of pps while loop */ 1132 continue; 1133 } 1134 1135 /* 1136 * safe with PQ's unlocked: because we just 1137 * alloc'd the page 1138 */ 1139 atomic_setbits_int(&ptmp->pg_flags, PQ_AOBJ); 1140 1141 /* 1142 * got new page ready for I/O. break pps while 1143 * loop. pps[lcv] is still NULL. 1144 */ 1145 break; 1146 } 1147 1148 /* page is there, see if we need to wait on it */ 1149 if ((ptmp->pg_flags & PG_BUSY) != 0) { 1150 atomic_setbits_int(&ptmp->pg_flags, PG_WANTED); 1151 rwsleep_nsec(ptmp, uobj->vmobjlock, PVM, 1152 "uao_get", INFSLP); 1153 continue; /* goto top of pps while loop */ 1154 } 1155 1156 /* 1157 * if we get here then the page is resident and 1158 * unbusy. we busy it now (so we own it). 1159 */ 1160 /* we own it, caller must un-busy */ 1161 atomic_setbits_int(&ptmp->pg_flags, PG_BUSY); 1162 UVM_PAGE_OWN(ptmp, "uao_get2"); 1163 pps[lcv] = ptmp; 1164 } 1165 1166 /* 1167 * if we own the valid page at the correct offset, pps[lcv] will 1168 * point to it. nothing more to do except go to the next page. 1169 */ 1170 if (pps[lcv]) 1171 continue; /* next lcv */ 1172 1173 /* 1174 * we have a "fake/busy/clean" page that we just allocated. 1175 * do the needed "i/o", either reading from swap or zeroing. 1176 */ 1177 swslot = uao_find_swslot(uobj, pageidx); 1178 1179 /* just zero the page if there's nothing in swap. */ 1180 if (swslot == 0) { 1181 /* page hasn't existed before, just zero it. */ 1182 uvm_pagezero(ptmp); 1183 } else { 1184 /* 1185 * page in the swapped-out page. 1186 * unlock object for i/o, relock when done. 1187 */ 1188 1189 rw_exit(uobj->vmobjlock); 1190 rv = uvm_swap_get(ptmp, swslot, PGO_SYNCIO); 1191 rw_enter(uobj->vmobjlock, RW_WRITE); 1192 1193 /* 1194 * I/O done. check for errors. 1195 */ 1196 if (rv != VM_PAGER_OK) { 1197 /* 1198 * remove the swap slot from the aobj 1199 * and mark the aobj as having no real slot. 1200 * don't free the swap slot, thus preventing 1201 * it from being used again. 1202 */ 1203 swslot = uao_set_swslot(&aobj->u_obj, pageidx, 1204 SWSLOT_BAD); 1205 uvm_swap_markbad(swslot, 1); 1206 1207 if (ptmp->pg_flags & PG_WANTED) 1208 wakeup(ptmp); 1209 atomic_clearbits_int(&ptmp->pg_flags, 1210 PG_WANTED|PG_BUSY); 1211 UVM_PAGE_OWN(ptmp, NULL); 1212 uvm_lock_pageq(); 1213 uvm_pagefree(ptmp); 1214 uvm_unlock_pageq(); 1215 rw_exit(uobj->vmobjlock); 1216 1217 return rv; 1218 } 1219 } 1220 1221 /* 1222 * we got the page! clear the fake flag (indicates valid 1223 * data now in page) and plug into our result array. note 1224 * that page is still busy. 1225 * 1226 * it is the callers job to: 1227 * => check if the page is released 1228 * => unbusy the page 1229 * => activate the page 1230 */ 1231 atomic_clearbits_int(&ptmp->pg_flags, PG_FAKE); 1232 pmap_clear_modify(ptmp); /* ... and clean */ 1233 pps[lcv] = ptmp; 1234 1235 } /* lcv loop */ 1236 1237 rw_exit(uobj->vmobjlock); 1238 return VM_PAGER_OK; 1239 } 1240 1241 /* 1242 * uao_dropswap: release any swap resources from this aobj page. 1243 * 1244 * => aobj must be locked or have a reference count of 0. 1245 */ 1246 int 1247 uao_dropswap(struct uvm_object *uobj, int pageidx) 1248 { 1249 int slot; 1250 1251 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 1252 1253 slot = uao_set_swslot(uobj, pageidx, 0); 1254 if (slot) { 1255 uvm_swap_free(slot, 1); 1256 } 1257 return slot; 1258 } 1259 1260 /* 1261 * page in every page in every aobj that is paged-out to a range of swslots. 1262 * 1263 * => aobj must be locked and is returned locked. 1264 * => returns TRUE if pagein was aborted due to lack of memory. 1265 */ 1266 boolean_t 1267 uao_swap_off(int startslot, int endslot) 1268 { 1269 struct uvm_aobj *aobj; 1270 1271 /* 1272 * Walk the list of all anonymous UVM objects. Grab the first. 1273 */ 1274 mtx_enter(&uao_list_lock); 1275 if ((aobj = LIST_FIRST(&uao_list)) == NULL) { 1276 mtx_leave(&uao_list_lock); 1277 return FALSE; 1278 } 1279 uao_reference(&aobj->u_obj); 1280 1281 do { 1282 struct uvm_aobj *nextaobj; 1283 boolean_t rv; 1284 1285 /* 1286 * Prefetch the next object and immediately hold a reference 1287 * on it, so neither the current nor the next entry could 1288 * disappear while we are iterating. 1289 */ 1290 if ((nextaobj = LIST_NEXT(aobj, u_list)) != NULL) { 1291 uao_reference(&nextaobj->u_obj); 1292 } 1293 mtx_leave(&uao_list_lock); 1294 1295 /* 1296 * Page in all pages in the swap slot range. 1297 */ 1298 rw_enter(aobj->u_obj.vmobjlock, RW_WRITE); 1299 rv = uao_pagein(aobj, startslot, endslot); 1300 rw_exit(aobj->u_obj.vmobjlock); 1301 1302 /* Drop the reference of the current object. */ 1303 uao_detach(&aobj->u_obj); 1304 if (rv) { 1305 if (nextaobj) { 1306 uao_detach(&nextaobj->u_obj); 1307 } 1308 return rv; 1309 } 1310 1311 aobj = nextaobj; 1312 mtx_enter(&uao_list_lock); 1313 } while (aobj); 1314 1315 /* 1316 * done with traversal, unlock the list 1317 */ 1318 mtx_leave(&uao_list_lock); 1319 return FALSE; 1320 } 1321 1322 /* 1323 * page in any pages from aobj in the given range. 1324 * 1325 * => returns TRUE if pagein was aborted due to lack of memory. 1326 */ 1327 static boolean_t 1328 uao_pagein(struct uvm_aobj *aobj, int startslot, int endslot) 1329 { 1330 boolean_t rv; 1331 1332 if (UAO_USES_SWHASH(aobj)) { 1333 struct uao_swhash_elt *elt; 1334 int bucket; 1335 1336 restart: 1337 for (bucket = aobj->u_swhashmask; bucket >= 0; bucket--) { 1338 for (elt = LIST_FIRST(&aobj->u_swhash[bucket]); 1339 elt != NULL; 1340 elt = LIST_NEXT(elt, list)) { 1341 int i; 1342 1343 for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) { 1344 int slot = elt->slots[i]; 1345 1346 /* 1347 * if the slot isn't in range, skip it. 1348 */ 1349 if (slot < startslot || 1350 slot >= endslot) { 1351 continue; 1352 } 1353 1354 /* 1355 * process the page, 1356 * the start over on this object 1357 * since the swhash elt 1358 * may have been freed. 1359 */ 1360 rv = uao_pagein_page(aobj, 1361 UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i); 1362 if (rv) { 1363 return rv; 1364 } 1365 goto restart; 1366 } 1367 } 1368 } 1369 } else { 1370 int i; 1371 1372 for (i = 0; i < aobj->u_pages; i++) { 1373 int slot = aobj->u_swslots[i]; 1374 1375 /* 1376 * if the slot isn't in range, skip it 1377 */ 1378 if (slot < startslot || slot >= endslot) { 1379 continue; 1380 } 1381 1382 /* 1383 * process the page. 1384 */ 1385 rv = uao_pagein_page(aobj, i); 1386 if (rv) { 1387 return rv; 1388 } 1389 } 1390 } 1391 1392 return FALSE; 1393 } 1394 1395 /* 1396 * uao_pagein_page: page in a single page from an anonymous UVM object. 1397 * 1398 * => Returns TRUE if pagein was aborted due to lack of memory. 1399 */ 1400 static boolean_t 1401 uao_pagein_page(struct uvm_aobj *aobj, int pageidx) 1402 { 1403 struct uvm_object *uobj = &aobj->u_obj; 1404 struct vm_page *pg; 1405 int rv, slot, npages; 1406 1407 pg = NULL; 1408 npages = 1; 1409 1410 KASSERT(rw_write_held(uobj->vmobjlock)); 1411 rv = uao_get(&aobj->u_obj, (voff_t)pageidx << PAGE_SHIFT, 1412 &pg, &npages, 0, PROT_READ | PROT_WRITE, 0, 0); 1413 1414 /* 1415 * relock and finish up. 1416 */ 1417 rw_enter(uobj->vmobjlock, RW_WRITE); 1418 switch (rv) { 1419 case VM_PAGER_OK: 1420 break; 1421 1422 case VM_PAGER_ERROR: 1423 case VM_PAGER_REFAULT: 1424 /* 1425 * nothing more to do on errors. 1426 * VM_PAGER_REFAULT can only mean that the anon was freed, 1427 * so again there's nothing to do. 1428 */ 1429 return FALSE; 1430 } 1431 1432 /* 1433 * ok, we've got the page now. 1434 * mark it as dirty, clear its swslot and un-busy it. 1435 */ 1436 slot = uao_set_swslot(&aobj->u_obj, pageidx, 0); 1437 uvm_swap_free(slot, 1); 1438 atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_CLEAN|PG_FAKE); 1439 UVM_PAGE_OWN(pg, NULL); 1440 1441 /* 1442 * deactivate the page (to put it on a page queue). 1443 */ 1444 pmap_clear_reference(pg); 1445 uvm_lock_pageq(); 1446 uvm_pagedeactivate(pg); 1447 uvm_unlock_pageq(); 1448 1449 return FALSE; 1450 } 1451 1452 /* 1453 * uao_dropswap_range: drop swapslots in the range. 1454 * 1455 * => aobj must be locked and is returned locked. 1456 * => start is inclusive. end is exclusive. 1457 */ 1458 void 1459 uao_dropswap_range(struct uvm_object *uobj, voff_t start, voff_t end) 1460 { 1461 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 1462 int swpgonlydelta = 0; 1463 1464 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 1465 KASSERT(rw_write_held(uobj->vmobjlock)); 1466 1467 if (end == 0) { 1468 end = INT64_MAX; 1469 } 1470 1471 if (UAO_USES_SWHASH(aobj)) { 1472 int i, hashbuckets = aobj->u_swhashmask + 1; 1473 voff_t taghi; 1474 voff_t taglo; 1475 1476 taglo = UAO_SWHASH_ELT_TAG(start); 1477 taghi = UAO_SWHASH_ELT_TAG(end); 1478 1479 for (i = 0; i < hashbuckets; i++) { 1480 struct uao_swhash_elt *elt, *next; 1481 1482 for (elt = LIST_FIRST(&aobj->u_swhash[i]); 1483 elt != NULL; 1484 elt = next) { 1485 int startidx, endidx; 1486 int j; 1487 1488 next = LIST_NEXT(elt, list); 1489 1490 if (elt->tag < taglo || taghi < elt->tag) { 1491 continue; 1492 } 1493 1494 if (elt->tag == taglo) { 1495 startidx = 1496 UAO_SWHASH_ELT_PAGESLOT_IDX(start); 1497 } else { 1498 startidx = 0; 1499 } 1500 1501 if (elt->tag == taghi) { 1502 endidx = 1503 UAO_SWHASH_ELT_PAGESLOT_IDX(end); 1504 } else { 1505 endidx = UAO_SWHASH_CLUSTER_SIZE; 1506 } 1507 1508 for (j = startidx; j < endidx; j++) { 1509 int slot = elt->slots[j]; 1510 1511 KASSERT(uvm_pagelookup(&aobj->u_obj, 1512 (voff_t)(UAO_SWHASH_ELT_PAGEIDX_BASE(elt) 1513 + j) << PAGE_SHIFT) == NULL); 1514 1515 if (slot > 0) { 1516 uvm_swap_free(slot, 1); 1517 swpgonlydelta++; 1518 KASSERT(elt->count > 0); 1519 elt->slots[j] = 0; 1520 elt->count--; 1521 } 1522 } 1523 1524 if (elt->count == 0) { 1525 LIST_REMOVE(elt, list); 1526 pool_put(&uao_swhash_elt_pool, elt); 1527 } 1528 } 1529 } 1530 } else { 1531 int i; 1532 1533 if (aobj->u_pages < end) { 1534 end = aobj->u_pages; 1535 } 1536 for (i = start; i < end; i++) { 1537 int slot = aobj->u_swslots[i]; 1538 1539 if (slot > 0) { 1540 uvm_swap_free(slot, 1); 1541 swpgonlydelta++; 1542 } 1543 } 1544 } 1545 1546 /* 1547 * adjust the counter of pages only in swap for all 1548 * the swap slots we've freed. 1549 */ 1550 if (swpgonlydelta > 0) { 1551 KASSERT(uvmexp.swpgonly >= swpgonlydelta); 1552 atomic_add_int(&uvmexp.swpgonly, -swpgonlydelta); 1553 } 1554 } 1555