fs/fscache/page.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* Cache page management and data I/O routines
   3  *
   4  * Copyright (C) 2004-2008 Red Hat, Inc. All Rights Reserved.
   5  * Written by David Howells (dhowells@redhat.com)
   6  */
   7
   8 #define FSCACHE_DEBUG_LEVEL PAGE
   9 #include <linux/module.h>
  10 #include <linux/fscache-cache.h>
  11 #include <linux/buffer_head.h>
  12 #include <linux/pagevec.h>
  13 #include <linux/slab.h>
  14 #include "internal.h"
  15
  16 /*
  17  * check to see if a page is being written to the cache
  18  */
  19 bool __fscache_check_page_write(struct fscache_cookie *cookie, struct page *page)
  20 {
  21         void *val;
  22
  23         rcu_read_lock();
  24         val = radix_tree_lookup(&cookie->stores, page->index);
  25         rcu_read_unlock();
  26         trace_fscache_check_page(cookie, page, val, 0);
  27
  28         return val != NULL;
  29 }
  30 EXPORT_SYMBOL(__fscache_check_page_write);
  31
  32 /*
  33  * wait for a page to finish being written to the cache
  34  */
  35 void __fscache_wait_on_page_write(struct fscache_cookie *cookie, struct page *page)
  36 {
  37         wait_queue_head_t *wq = bit_waitqueue(&cookie->flags, 0);
  38
  39         trace_fscache_page(cookie, page, fscache_page_write_wait);
  40
  41         wait_event(*wq, !__fscache_check_page_write(cookie, page));
  42 }
  43 EXPORT_SYMBOL(__fscache_wait_on_page_write);
  44
  45 /*
  46  * wait for a page to finish being written to the cache. Put a timeout here
  47  * since we might be called recursively via parent fs.
  48  */
  49 static
  50 bool release_page_wait_timeout(struct fscache_cookie *cookie, struct page *page)
  51 {
  52         wait_queue_head_t *wq = bit_waitqueue(&cookie->flags, 0);
  53
  54         return wait_event_timeout(*wq, !__fscache_check_page_write(cookie, page),
  55                                   HZ);
  56 }
  57
  58 /*
  59  * decide whether a page can be released, possibly by cancelling a store to it
  60  * - we're allowed to sleep if __GFP_DIRECT_RECLAIM is flagged
  61  */
  62 bool __fscache_maybe_release_page(struct fscache_cookie *cookie,
  63                                   struct page *page,
  64                                   gfp_t gfp)
  65 {
  66         struct page *xpage;
  67         void *val;
  68
  69         _enter("%p,%p,%x", cookie, page, gfp);
  70
  71         trace_fscache_page(cookie, page, fscache_page_maybe_release);
  72
  73 try_again:
  74         rcu_read_lock();
  75         val = radix_tree_lookup(&cookie->stores, page->index);
  76         if (!val) {
  77                 rcu_read_unlock();
  78                 fscache_stat(&fscache_n_store_vmscan_not_storing);
  79                 __fscache_uncache_page(cookie, page);
  80                 return true;
  81         }
  82
  83         /* see if the page is actually undergoing storage - if so we can't get
  84          * rid of it till the cache has finished with it */
  85         if (radix_tree_tag_get(&cookie->stores, page->index,
  86                                FSCACHE_COOKIE_STORING_TAG)) {
  87                 rcu_read_unlock();
  88                 goto page_busy;
  89         }
  90
  91         /* the page is pending storage, so we attempt to cancel the store and
  92          * discard the store request so that the page can be reclaimed */
  93         spin_lock(&cookie->stores_lock);
  94         rcu_read_unlock();
  95
  96         if (radix_tree_tag_get(&cookie->stores, page->index,
  97                                FSCACHE_COOKIE_STORING_TAG)) {
  98                 /* the page started to undergo storage whilst we were looking,
  99                  * so now we can only wait or return */
 100                 spin_unlock(&cookie->stores_lock);
 101                 goto page_busy;
 102         }
 103
 104         xpage = radix_tree_delete(&cookie->stores, page->index);
 105         trace_fscache_page(cookie, page, fscache_page_radix_delete);
 106         spin_unlock(&cookie->stores_lock);
 107
 108         if (xpage) {
 109                 fscache_stat(&fscache_n_store_vmscan_cancelled);
 110                 fscache_stat(&fscache_n_store_radix_deletes);
 111                 ASSERTCMP(xpage, ==, page);
 112         } else {
 113                 fscache_stat(&fscache_n_store_vmscan_gone);
 114         }
 115
 116         wake_up_bit(&cookie->flags, 0);
 117         trace_fscache_wake_cookie(cookie);
 118         if (xpage)
 119                 put_page(xpage);
 120         __fscache_uncache_page(cookie, page);
 121         return true;
 122
 123 page_busy:
 124         /* We will wait here if we're allowed to, but that could deadlock the
 125          * allocator as the work threads writing to the cache may all end up
 126          * sleeping on memory allocation, so we may need to impose a timeout
 127          * too. */
 128         if (!(gfp & __GFP_DIRECT_RECLAIM) || !(gfp & __GFP_FS)) {
 129                 fscache_stat(&fscache_n_store_vmscan_busy);
 130                 return false;
 131         }
 132
 133         fscache_stat(&fscache_n_store_vmscan_wait);
 134         if (!release_page_wait_timeout(cookie, page))
 135                 _debug("fscache writeout timeout page: %p{%lx}",
 136                         page, page->index);
 137
 138         gfp &= ~__GFP_DIRECT_RECLAIM;
 139         goto try_again;
 140 }
 141 EXPORT_SYMBOL(__fscache_maybe_release_page);
 142
 143 /*
 144  * note that a page has finished being written to the cache
 145  */
 146 static void fscache_end_page_write(struct fscache_object *object,
 147                                    struct page *page)
 148 {
 149         struct fscache_cookie *cookie;
 150         struct page *xpage = NULL, *val;
 151
 152         spin_lock(&object->lock);
 153         cookie = object->cookie;
 154         if (cookie) {
 155                 /* delete the page from the tree if it is now no longer
 156                  * pending */
 157                 spin_lock(&cookie->stores_lock);
 158                 radix_tree_tag_clear(&cookie->stores, page->index,
 159                                      FSCACHE_COOKIE_STORING_TAG);
 160                 trace_fscache_page(cookie, page, fscache_page_radix_clear_store);
 161                 if (!radix_tree_tag_get(&cookie->stores, page->index,
 162                                         FSCACHE_COOKIE_PENDING_TAG)) {
 163                         fscache_stat(&fscache_n_store_radix_deletes);
 164                         xpage = radix_tree_delete(&cookie->stores, page->index);
 165                         trace_fscache_page(cookie, page, fscache_page_radix_delete);
 166                         trace_fscache_page(cookie, page, fscache_page_write_end);
 167
 168                         val = radix_tree_lookup(&cookie->stores, page->index);
 169                         trace_fscache_check_page(cookie, page, val, 1);
 170                 } else {
 171                         trace_fscache_page(cookie, page, fscache_page_write_end_pend);
 172                 }
 173                 spin_unlock(&cookie->stores_lock);
 174                 wake_up_bit(&cookie->flags, 0);
 175                 trace_fscache_wake_cookie(cookie);
 176         } else {
 177                 trace_fscache_page(cookie, page, fscache_page_write_end_noc);
 178         }
 179         spin_unlock(&object->lock);
 180         if (xpage)
 181                 put_page(xpage);
 182 }
 183
 184 /*
 185  * actually apply the changed attributes to a cache object
 186  */
 187 static void fscache_attr_changed_op(struct fscache_operation *op)
 188 {
 189         struct fscache_object *object = op->object;
 190         int ret;
 191
 192         _enter("{OBJ%x OP%x}", object->debug_id, op->debug_id);
 193
 194         fscache_stat(&fscache_n_attr_changed_calls);
 195
 196         if (fscache_object_is_active(object)) {
 197                 fscache_stat(&fscache_n_cop_attr_changed);
 198                 ret = object->cache->ops->attr_changed(object);
 199                 fscache_stat_d(&fscache_n_cop_attr_changed);
 200                 if (ret < 0)
 201                         fscache_abort_object(object);
 202                 fscache_op_complete(op, ret < 0);
 203         } else {
 204                 fscache_op_complete(op, true);
 205         }
 206
 207         _leave("");
 208 }
 209
 210 /*
 211  * notification that the attributes on an object have changed
 212  */
 213 int __fscache_attr_changed(struct fscache_cookie *cookie)
 214 {
 215         struct fscache_operation *op;
 216         struct fscache_object *object;
 217         bool wake_cookie = false;
 218
 219         _enter("%p", cookie);
 220
 221         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
 222
 223         fscache_stat(&fscache_n_attr_changed);
 224
 225         op = kzalloc(sizeof(*op), GFP_KERNEL);
 226         if (!op) {
 227                 fscache_stat(&fscache_n_attr_changed_nomem);
 228                 _leave(" = -ENOMEM");
 229                 return -ENOMEM;
 230         }
 231
 232         fscache_operation_init(cookie, op, fscache_attr_changed_op, NULL, NULL);
 233         trace_fscache_page_op(cookie, NULL, op, fscache_page_op_attr_changed);
 234         op->flags = FSCACHE_OP_ASYNC |
 235                 (1 << FSCACHE_OP_EXCLUSIVE) |
 236                 (1 << FSCACHE_OP_UNUSE_COOKIE);
 237
 238         spin_lock(&cookie->lock);
 239
 240         if (!fscache_cookie_enabled(cookie) ||
 241             hlist_empty(&cookie->backing_objects))
 242                 goto nobufs;
 243         object = hlist_entry(cookie->backing_objects.first,
 244                              struct fscache_object, cookie_link);
 245
 246         __fscache_use_cookie(cookie);
 247         if (fscache_submit_exclusive_op(object, op) < 0)
 248                 goto nobufs_dec;
 249         spin_unlock(&cookie->lock);
 250         fscache_stat(&fscache_n_attr_changed_ok);
 251         fscache_put_operation(op);
 252         _leave(" = 0");
 253         return 0;
 254
 255 nobufs_dec:
 256         wake_cookie = __fscache_unuse_cookie(cookie);
 257 nobufs:
 258         spin_unlock(&cookie->lock);
 259         fscache_put_operation(op);
 260         if (wake_cookie)
 261                 __fscache_wake_unused_cookie(cookie);
 262         fscache_stat(&fscache_n_attr_changed_nobufs);
 263         _leave(" = %d", -ENOBUFS);
 264         return -ENOBUFS;
 265 }
 266 EXPORT_SYMBOL(__fscache_attr_changed);
 267
 268 /*
 269  * Handle cancellation of a pending retrieval op
 270  */
 271 static void fscache_do_cancel_retrieval(struct fscache_operation *_op)
 272 {
 273         struct fscache_retrieval *op =
 274                 container_of(_op, struct fscache_retrieval, op);
 275
 276         atomic_set(&op->n_pages, 0);
 277 }
 278
 279 /*
 280  * release a retrieval op reference
 281  */
 282 static void fscache_release_retrieval_op(struct fscache_operation *_op)
 283 {
 284         struct fscache_retrieval *op =
 285                 container_of(_op, struct fscache_retrieval, op);
 286
 287         _enter("{OP%x}", op->op.debug_id);
 288
 289         ASSERTIFCMP(op->op.state != FSCACHE_OP_ST_INITIALISED,
 290                     atomic_read(&op->n_pages), ==, 0);
 291
 292         fscache_hist(fscache_retrieval_histogram, op->start_time);
 293         if (op->context)
 294                 fscache_put_context(op->cookie, op->context);
 295
 296         _leave("");
 297 }
 298
 299 /*
 300  * allocate a retrieval op
 301  */
 302 static struct fscache_retrieval *fscache_alloc_retrieval(
 303         struct fscache_cookie *cookie,
 304         struct address_space *mapping,
 305         fscache_rw_complete_t end_io_func,
 306         void *context)
 307 {
 308         struct fscache_retrieval *op;
 309
 310         /* allocate a retrieval operation and attempt to submit it */
 311         op = kzalloc(sizeof(*op), GFP_NOIO);
 312         if (!op) {
 313                 fscache_stat(&fscache_n_retrievals_nomem);
 314                 return NULL;
 315         }
 316
 317         fscache_operation_init(cookie, &op->op, NULL,
 318                                fscache_do_cancel_retrieval,
 319                                fscache_release_retrieval_op);
 320         op->op.flags    = FSCACHE_OP_MYTHREAD |
 321                 (1UL << FSCACHE_OP_WAITING) |
 322                 (1UL << FSCACHE_OP_UNUSE_COOKIE);
 323         op->cookie      = cookie;
 324         op->mapping     = mapping;
 325         op->end_io_func = end_io_func;
 326         op->context     = context;
 327         op->start_time  = jiffies;
 328         INIT_LIST_HEAD(&op->to_do);
 329
 330         /* Pin the netfs read context in case we need to do the actual netfs
 331          * read because we've encountered a cache read failure.
 332          */
 333         if (context)
 334                 fscache_get_context(op->cookie, context);
 335         return op;
 336 }
 337
 338 /*
 339  * wait for a deferred lookup to complete
 340  */
 341 int fscache_wait_for_deferred_lookup(struct fscache_cookie *cookie)
 342 {
 343         unsigned long jif;
 344
 345         _enter("");
 346
 347         if (!test_bit(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags)) {
 348                 _leave(" = 0 [imm]");
 349                 return 0;
 350         }
 351
 352         fscache_stat(&fscache_n_retrievals_wait);
 353
 354         jif = jiffies;
 355         if (wait_on_bit(&cookie->flags, FSCACHE_COOKIE_LOOKING_UP,
 356                         TASK_INTERRUPTIBLE) != 0) {
 357                 fscache_stat(&fscache_n_retrievals_intr);
 358                 _leave(" = -ERESTARTSYS");
 359                 return -ERESTARTSYS;
 360         }
 361
 362         ASSERT(!test_bit(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags));
 363
 364         smp_rmb();
 365         fscache_hist(fscache_retrieval_delay_histogram, jif);
 366         _leave(" = 0 [dly]");
 367         return 0;
 368 }
 369
 370 /*
 371  * wait for an object to become active (or dead)
 372  */
 373 int fscache_wait_for_operation_activation(struct fscache_object *object,
 374                                           struct fscache_operation *op,
 375                                           atomic_t *stat_op_waits,
 376                                           atomic_t *stat_object_dead)
 377 {
 378         int ret;
 379
 380         if (!test_bit(FSCACHE_OP_WAITING, &op->flags))
 381                 goto check_if_dead;
 382
 383         _debug(">>> WT");
 384         if (stat_op_waits)
 385                 fscache_stat(stat_op_waits);
 386         if (wait_on_bit(&op->flags, FSCACHE_OP_WAITING,
 387                         TASK_INTERRUPTIBLE) != 0) {
 388                 trace_fscache_op(object->cookie, op, fscache_op_signal);
 389                 ret = fscache_cancel_op(op, false);
 390                 if (ret == 0)
 391                         return -ERESTARTSYS;
 392
 393                 /* it's been removed from the pending queue by another party,
 394                  * so we should get to run shortly */
 395                 wait_on_bit(&op->flags, FSCACHE_OP_WAITING,
 396                             TASK_UNINTERRUPTIBLE);
 397         }
 398         _debug("<<< GO");
 399
 400 check_if_dead:
 401         if (op->state == FSCACHE_OP_ST_CANCELLED) {
 402                 if (stat_object_dead)
 403                         fscache_stat(stat_object_dead);
 404                 _leave(" = -ENOBUFS [cancelled]");
 405                 return -ENOBUFS;
 406         }
 407         if (unlikely(fscache_object_is_dying(object) ||
 408                      fscache_cache_is_broken(object))) {
 409                 enum fscache_operation_state state = op->state;
 410                 trace_fscache_op(object->cookie, op, fscache_op_signal);
 411                 fscache_cancel_op(op, true);
 412                 if (stat_object_dead)
 413                         fscache_stat(stat_object_dead);
 414                 _leave(" = -ENOBUFS [obj dead %d]", state);
 415                 return -ENOBUFS;
 416         }
 417         return 0;
 418 }
 419
 420 /*
 421  * read a page from the cache or allocate a block in which to store it
 422  * - we return:
 423  *   -ENOMEM    - out of memory, nothing done
 424  *   -ERESTARTSYS - interrupted
 425  *   -ENOBUFS   - no backing object available in which to cache the block
 426  *   -ENODATA   - no data available in the backing object for this block
 427  *   0          - dispatched a read - it'll call end_io_func() when finished
 428  */
 429 int __fscache_read_or_alloc_page(struct fscache_cookie *cookie,
 430                                  struct page *page,
 431                                  fscache_rw_complete_t end_io_func,
 432                                  void *context,
 433                                  gfp_t gfp)
 434 {
 435         struct fscache_retrieval *op;
 436         struct fscache_object *object;
 437         bool wake_cookie = false;
 438         int ret;
 439
 440         _enter("%p,%p,,,", cookie, page);
 441
 442         fscache_stat(&fscache_n_retrievals);
 443
 444         if (hlist_empty(&cookie->backing_objects))
 445                 goto nobufs;
 446
 447         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
 448                 _leave(" = -ENOBUFS [invalidating]");
 449                 return -ENOBUFS;
 450         }
 451
 452         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
 453         ASSERTCMP(page, !=, NULL);
 454
 455         if (fscache_wait_for_deferred_lookup(cookie) < 0)
 456                 return -ERESTARTSYS;
 457
 458         op = fscache_alloc_retrieval(cookie, page->mapping,
 459                                      end_io_func, context);
 460         if (!op) {
 461                 _leave(" = -ENOMEM");
 462                 return -ENOMEM;
 463         }
 464         atomic_set(&op->n_pages, 1);
 465         trace_fscache_page_op(cookie, page, &op->op, fscache_page_op_retr_one);
 466
 467         spin_lock(&cookie->lock);
 468
 469         if (!fscache_cookie_enabled(cookie) ||
 470             hlist_empty(&cookie->backing_objects))
 471                 goto nobufs_unlock;
 472         object = hlist_entry(cookie->backing_objects.first,
 473                              struct fscache_object, cookie_link);
 474
 475         ASSERT(test_bit(FSCACHE_OBJECT_IS_LOOKED_UP, &object->flags));
 476
 477         __fscache_use_cookie(cookie);
 478         atomic_inc(&object->n_reads);
 479         __set_bit(FSCACHE_OP_DEC_READ_CNT, &op->op.flags);
 480
 481         if (fscache_submit_op(object, &op->op) < 0)
 482                 goto nobufs_unlock_dec;
 483         spin_unlock(&cookie->lock);
 484
 485         fscache_stat(&fscache_n_retrieval_ops);
 486
 487         /* we wait for the operation to become active, and then process it
 488          * *here*, in this thread, and not in the thread pool */
 489         ret = fscache_wait_for_operation_activation(
 490                 object, &op->op,
 491                 __fscache_stat(&fscache_n_retrieval_op_waits),
 492                 __fscache_stat(&fscache_n_retrievals_object_dead));
 493         if (ret < 0)
 494                 goto error;
 495
 496         /* ask the cache to honour the operation */
 497         if (test_bit(FSCACHE_COOKIE_NO_DATA_YET, &object->cookie->flags)) {
 498                 fscache_stat(&fscache_n_cop_allocate_page);
 499                 ret = object->cache->ops->allocate_page(op, page, gfp);
 500                 fscache_stat_d(&fscache_n_cop_allocate_page);
 501                 if (ret == 0)
 502                         ret = -ENODATA;
 503         } else {
 504                 fscache_stat(&fscache_n_cop_read_or_alloc_page);
 505                 ret = object->cache->ops->read_or_alloc_page(op, page, gfp);
 506                 fscache_stat_d(&fscache_n_cop_read_or_alloc_page);
 507         }
 508
 509 error:
 510         if (ret == -ENOMEM)
 511                 fscache_stat(&fscache_n_retrievals_nomem);
 512         else if (ret == -ERESTARTSYS)
 513                 fscache_stat(&fscache_n_retrievals_intr);
 514         else if (ret == -ENODATA)
 515                 fscache_stat(&fscache_n_retrievals_nodata);
 516         else if (ret < 0)
 517                 fscache_stat(&fscache_n_retrievals_nobufs);
 518         else
 519                 fscache_stat(&fscache_n_retrievals_ok);
 520
 521         fscache_put_retrieval(op);
 522         _leave(" = %d", ret);
 523         return ret;
 524
 525 nobufs_unlock_dec:
 526         atomic_dec(&object->n_reads);
 527         wake_cookie = __fscache_unuse_cookie(cookie);
 528 nobufs_unlock:
 529         spin_unlock(&cookie->lock);
 530         if (wake_cookie)
 531                 __fscache_wake_unused_cookie(cookie);
 532         fscache_put_retrieval(op);
 533 nobufs:
 534         fscache_stat(&fscache_n_retrievals_nobufs);
 535         _leave(" = -ENOBUFS");
 536         return -ENOBUFS;
 537 }
 538 EXPORT_SYMBOL(__fscache_read_or_alloc_page);
 539
 540 /*
 541  * read a list of page from the cache or allocate a block in which to store
 542  * them
 543  * - we return:
 544  *   -ENOMEM    - out of memory, some pages may be being read
 545  *   -ERESTARTSYS - interrupted, some pages may be being read
 546  *   -ENOBUFS   - no backing object or space available in which to cache any
 547  *                pages not being read
 548  *   -ENODATA   - no data available in the backing object for some or all of
 549  *                the pages
 550  *   0          - dispatched a read on all pages
 551  *
 552  * end_io_func() will be called for each page read from the cache as it is
 553  * finishes being read
 554  *
 555  * any pages for which a read is dispatched will be removed from pages and
 556  * nr_pages
 557  */
 558 int __fscache_read_or_alloc_pages(struct fscache_cookie *cookie,
 559                                   struct address_space *mapping,
 560                                   struct list_head *pages,
 561                                   unsigned *nr_pages,
 562                                   fscache_rw_complete_t end_io_func,
 563                                   void *context,
 564                                   gfp_t gfp)
 565 {
 566         struct fscache_retrieval *op;
 567         struct fscache_object *object;
 568         bool wake_cookie = false;
 569         int ret;
 570
 571         _enter("%p,,%d,,,", cookie, *nr_pages);
 572
 573         fscache_stat(&fscache_n_retrievals);
 574
 575         if (hlist_empty(&cookie->backing_objects))
 576                 goto nobufs;
 577
 578         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
 579                 _leave(" = -ENOBUFS [invalidating]");
 580                 return -ENOBUFS;
 581         }
 582
 583         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
 584         ASSERTCMP(*nr_pages, >, 0);
 585         ASSERT(!list_empty(pages));
 586
 587         if (fscache_wait_for_deferred_lookup(cookie) < 0)
 588                 return -ERESTARTSYS;
 589
 590         op = fscache_alloc_retrieval(cookie, mapping, end_io_func, context);
 591         if (!op)
 592                 return -ENOMEM;
 593         atomic_set(&op->n_pages, *nr_pages);
 594         trace_fscache_page_op(cookie, NULL, &op->op, fscache_page_op_retr_multi);
 595
 596         spin_lock(&cookie->lock);
 597
 598         if (!fscache_cookie_enabled(cookie) ||
 599             hlist_empty(&cookie->backing_objects))
 600                 goto nobufs_unlock;
 601         object = hlist_entry(cookie->backing_objects.first,
 602                              struct fscache_object, cookie_link);
 603
 604         __fscache_use_cookie(cookie);
 605         atomic_inc(&object->n_reads);
 606         __set_bit(FSCACHE_OP_DEC_READ_CNT, &op->op.flags);
 607
 608         if (fscache_submit_op(object, &op->op) < 0)
 609                 goto nobufs_unlock_dec;
 610         spin_unlock(&cookie->lock);
 611
 612         fscache_stat(&fscache_n_retrieval_ops);
 613
 614         /* we wait for the operation to become active, and then process it
 615          * *here*, in this thread, and not in the thread pool */
 616         ret = fscache_wait_for_operation_activation(
 617                 object, &op->op,
 618                 __fscache_stat(&fscache_n_retrieval_op_waits),
 619                 __fscache_stat(&fscache_n_retrievals_object_dead));
 620         if (ret < 0)
 621                 goto error;
 622
 623         /* ask the cache to honour the operation */
 624         if (test_bit(FSCACHE_COOKIE_NO_DATA_YET, &object->cookie->flags)) {
 625                 fscache_stat(&fscache_n_cop_allocate_pages);
 626                 ret = object->cache->ops->allocate_pages(
 627                         op, pages, nr_pages, gfp);
 628                 fscache_stat_d(&fscache_n_cop_allocate_pages);
 629         } else {
 630                 fscache_stat(&fscache_n_cop_read_or_alloc_pages);
 631                 ret = object->cache->ops->read_or_alloc_pages(
 632                         op, pages, nr_pages, gfp);
 633                 fscache_stat_d(&fscache_n_cop_read_or_alloc_pages);
 634         }
 635
 636 error:
 637         if (ret == -ENOMEM)
 638                 fscache_stat(&fscache_n_retrievals_nomem);
 639         else if (ret == -ERESTARTSYS)
 640                 fscache_stat(&fscache_n_retrievals_intr);
 641         else if (ret == -ENODATA)
 642                 fscache_stat(&fscache_n_retrievals_nodata);
 643         else if (ret < 0)
 644                 fscache_stat(&fscache_n_retrievals_nobufs);
 645         else
 646                 fscache_stat(&fscache_n_retrievals_ok);
 647
 648         fscache_put_retrieval(op);
 649         _leave(" = %d", ret);
 650         return ret;
 651
 652 nobufs_unlock_dec:
 653         atomic_dec(&object->n_reads);
 654         wake_cookie = __fscache_unuse_cookie(cookie);
 655 nobufs_unlock:
 656         spin_unlock(&cookie->lock);
 657         fscache_put_retrieval(op);
 658         if (wake_cookie)
 659                 __fscache_wake_unused_cookie(cookie);
 660 nobufs:
 661         fscache_stat(&fscache_n_retrievals_nobufs);
 662         _leave(" = -ENOBUFS");
 663         return -ENOBUFS;
 664 }
 665 EXPORT_SYMBOL(__fscache_read_or_alloc_pages);
 666
 667 /*
 668  * allocate a block in the cache on which to store a page
 669  * - we return:
 670  *   -ENOMEM    - out of memory, nothing done
 671  *   -ERESTARTSYS - interrupted
 672  *   -ENOBUFS   - no backing object available in which to cache the block
 673  *   0          - block allocated
 674  */
 675 int __fscache_alloc_page(struct fscache_cookie *cookie,
 676                          struct page *page,
 677                          gfp_t gfp)
 678 {
 679         struct fscache_retrieval *op;
 680         struct fscache_object *object;
 681         bool wake_cookie = false;
 682         int ret;
 683
 684         _enter("%p,%p,,,", cookie, page);
 685
 686         fscache_stat(&fscache_n_allocs);
 687
 688         if (hlist_empty(&cookie->backing_objects))
 689                 goto nobufs;
 690
 691         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
 692         ASSERTCMP(page, !=, NULL);
 693
 694         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
 695                 _leave(" = -ENOBUFS [invalidating]");
 696                 return -ENOBUFS;
 697         }
 698
 699         if (fscache_wait_for_deferred_lookup(cookie) < 0)
 700                 return -ERESTARTSYS;
 701
 702         op = fscache_alloc_retrieval(cookie, page->mapping, NULL, NULL);
 703         if (!op)
 704                 return -ENOMEM;
 705         atomic_set(&op->n_pages, 1);
 706         trace_fscache_page_op(cookie, page, &op->op, fscache_page_op_alloc_one);
 707
 708         spin_lock(&cookie->lock);
 709
 710         if (!fscache_cookie_enabled(cookie) ||
 711             hlist_empty(&cookie->backing_objects))
 712                 goto nobufs_unlock;
 713         object = hlist_entry(cookie->backing_objects.first,
 714                              struct fscache_object, cookie_link);
 715
 716         __fscache_use_cookie(cookie);
 717         if (fscache_submit_op(object, &op->op) < 0)
 718                 goto nobufs_unlock_dec;
 719         spin_unlock(&cookie->lock);
 720
 721         fscache_stat(&fscache_n_alloc_ops);
 722
 723         ret = fscache_wait_for_operation_activation(
 724                 object, &op->op,
 725                 __fscache_stat(&fscache_n_alloc_op_waits),
 726                 __fscache_stat(&fscache_n_allocs_object_dead));
 727         if (ret < 0)
 728                 goto error;
 729
 730         /* ask the cache to honour the operation */
 731         fscache_stat(&fscache_n_cop_allocate_page);
 732         ret = object->cache->ops->allocate_page(op, page, gfp);
 733         fscache_stat_d(&fscache_n_cop_allocate_page);
 734
 735 error:
 736         if (ret == -ERESTARTSYS)
 737                 fscache_stat(&fscache_n_allocs_intr);
 738         else if (ret < 0)
 739                 fscache_stat(&fscache_n_allocs_nobufs);
 740         else
 741                 fscache_stat(&fscache_n_allocs_ok);
 742
 743         fscache_put_retrieval(op);
 744         _leave(" = %d", ret);
 745         return ret;
 746
 747 nobufs_unlock_dec:
 748         wake_cookie = __fscache_unuse_cookie(cookie);
 749 nobufs_unlock:
 750         spin_unlock(&cookie->lock);
 751         fscache_put_retrieval(op);
 752         if (wake_cookie)
 753                 __fscache_wake_unused_cookie(cookie);
 754 nobufs:
 755         fscache_stat(&fscache_n_allocs_nobufs);
 756         _leave(" = -ENOBUFS");
 757         return -ENOBUFS;
 758 }
 759 EXPORT_SYMBOL(__fscache_alloc_page);
 760
 761 /*
 762  * Unmark pages allocate in the readahead code path (via:
 763  * fscache_readpages_or_alloc) after delegating to the base filesystem
 764  */
 765 void __fscache_readpages_cancel(struct fscache_cookie *cookie,
 766                                 struct list_head *pages)
 767 {
 768         struct page *page;
 769
 770         list_for_each_entry(page, pages, lru) {
 771                 if (PageFsCache(page))
 772                         __fscache_uncache_page(cookie, page);
 773         }
 774 }
 775 EXPORT_SYMBOL(__fscache_readpages_cancel);
 776
 777 /*
 778  * release a write op reference
 779  */
 780 static void fscache_release_write_op(struct fscache_operation *_op)
 781 {
 782         _enter("{OP%x}", _op->debug_id);
 783 }
 784
 785 /*
 786  * perform the background storage of a page into the cache
 787  */
 788 static void fscache_write_op(struct fscache_operation *_op)
 789 {
 790         struct fscache_storage *op =
 791                 container_of(_op, struct fscache_storage, op);
 792         struct fscache_object *object = op->op.object;
 793         struct fscache_cookie *cookie;
 794         struct page *page;
 795         unsigned n;
 796         void *results[1];
 797         int ret;
 798
 799         _enter("{OP%x,%d}", op->op.debug_id, atomic_read(&op->op.usage));
 800
 801 again:
 802         spin_lock(&object->lock);
 803         cookie = object->cookie;
 804
 805         if (!fscache_object_is_active(object)) {
 806                 /* If we get here, then the on-disk cache object likely no
 807                  * longer exists, so we should just cancel this write
 808                  * operation.
 809                  */
 810                 spin_unlock(&object->lock);
 811                 fscache_op_complete(&op->op, true);
 812                 _leave(" [inactive]");
 813                 return;
 814         }
 815
 816         if (!cookie) {
 817                 /* If we get here, then the cookie belonging to the object was
 818                  * detached, probably by the cookie being withdrawn due to
 819                  * memory pressure, which means that the pages we might write
 820                  * to the cache from no longer exist - therefore, we can just
 821                  * cancel this write operation.
 822                  */
 823                 spin_unlock(&object->lock);
 824                 fscache_op_complete(&op->op, true);
 825                 _leave(" [cancel] op{f=%lx s=%u} obj{s=%s f=%lx}",
 826                        _op->flags, _op->state, object->state->short_name,
 827                        object->flags);
 828                 return;
 829         }
 830
 831         spin_lock(&cookie->stores_lock);
 832
 833         fscache_stat(&fscache_n_store_calls);
 834
 835         /* find a page to store */
 836         results[0] = NULL;
 837         page = NULL;
 838         n = radix_tree_gang_lookup_tag(&cookie->stores, results, 0, 1,
 839                                        FSCACHE_COOKIE_PENDING_TAG);
 840         trace_fscache_gang_lookup(cookie, &op->op, results, n, op->store_limit);
 841         if (n != 1)
 842                 goto superseded;
 843         page = results[0];
 844         _debug("gang %d [%lx]", n, page->index);
 845
 846         radix_tree_tag_set(&cookie->stores, page->index,
 847                            FSCACHE_COOKIE_STORING_TAG);
 848         radix_tree_tag_clear(&cookie->stores, page->index,
 849                              FSCACHE_COOKIE_PENDING_TAG);
 850         trace_fscache_page(cookie, page, fscache_page_radix_pend2store);
 851
 852         spin_unlock(&cookie->stores_lock);
 853         spin_unlock(&object->lock);
 854
 855         if (page->index >= op->store_limit)
 856                 goto discard_page;
 857
 858         fscache_stat(&fscache_n_store_pages);
 859         fscache_stat(&fscache_n_cop_write_page);
 860         ret = object->cache->ops->write_page(op, page);
 861         fscache_stat_d(&fscache_n_cop_write_page);
 862         trace_fscache_wrote_page(cookie, page, &op->op, ret);
 863         fscache_end_page_write(object, page);
 864         if (ret < 0) {
 865                 fscache_abort_object(object);
 866                 fscache_op_complete(&op->op, true);
 867         } else {
 868                 fscache_enqueue_operation(&op->op);
 869         }
 870
 871         _leave("");
 872         return;
 873
 874 discard_page:
 875         fscache_stat(&fscache_n_store_pages_over_limit);
 876         trace_fscache_wrote_page(cookie, page, &op->op, -ENOBUFS);
 877         fscache_end_page_write(object, page);
 878         goto again;
 879
 880 superseded:
 881         /* this writer is going away and there aren't any more things to
 882          * write */
 883         _debug("cease");
 884         spin_unlock(&cookie->stores_lock);
 885         clear_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags);
 886         spin_unlock(&object->lock);
 887         fscache_op_complete(&op->op, false);
 888         _leave("");
 889 }
 890
 891 /*
 892  * Clear the pages pending writing for invalidation
 893  */
 894 void fscache_invalidate_writes(struct fscache_cookie *cookie)
 895 {
 896         struct page *page;
 897         void *results[16];
 898         int n, i;
 899
 900         _enter("");
 901
 902         for (;;) {
 903                 spin_lock(&cookie->stores_lock);
 904                 n = radix_tree_gang_lookup_tag(&cookie->stores, results, 0,
 905                                                ARRAY_SIZE(results),
 906                                                FSCACHE_COOKIE_PENDING_TAG);
 907                 if (n == 0) {
 908                         spin_unlock(&cookie->stores_lock);
 909                         break;
 910                 }
 911
 912                 for (i = n - 1; i >= 0; i--) {
 913                         page = results[i];
 914                         radix_tree_delete(&cookie->stores, page->index);
 915                         trace_fscache_page(cookie, page, fscache_page_radix_delete);
 916                         trace_fscache_page(cookie, page, fscache_page_inval);
 917                 }
 918
 919                 spin_unlock(&cookie->stores_lock);
 920
 921                 for (i = n - 1; i >= 0; i--)
 922                         put_page(results[i]);
 923         }
 924
 925         wake_up_bit(&cookie->flags, 0);
 926         trace_fscache_wake_cookie(cookie);
 927
 928         _leave("");
 929 }
 930
 931 /*
 932  * request a page be stored in the cache
 933  * - returns:
 934  *   -ENOMEM    - out of memory, nothing done
 935  *   -ENOBUFS   - no backing object available in which to cache the page
 936  *   0          - dispatched a write - it'll call end_io_func() when finished
 937  *
 938  * if the cookie still has a backing object at this point, that object can be
 939  * in one of a few states with respect to storage processing:
 940  *
 941  *  (1) negative lookup, object not yet created (FSCACHE_COOKIE_CREATING is
 942  *      set)
 943  *
 944  *      (a) no writes yet
 945  *
 946  *      (b) writes deferred till post-creation (mark page for writing and
 947  *          return immediately)
 948  *
 949  *  (2) negative lookup, object created, initial fill being made from netfs
 950  *
 951  *      (a) fill point not yet reached this page (mark page for writing and
 952  *          return)
 953  *
 954  *      (b) fill point passed this page (queue op to store this page)
 955  *
 956  *  (3) object extant (queue op to store this page)
 957  *
 958  * any other state is invalid
 959  */
 960 int __fscache_write_page(struct fscache_cookie *cookie,
 961                          struct page *page,
 962                          loff_t object_size,
 963                          gfp_t gfp)
 964 {
 965         struct fscache_storage *op;
 966         struct fscache_object *object;
 967         bool wake_cookie = false;
 968         int ret;
 969
 970         _enter("%p,%x,", cookie, (u32) page->flags);
 971
 972         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
 973         ASSERT(PageFsCache(page));
 974
 975         fscache_stat(&fscache_n_stores);
 976
 977         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
 978                 _leave(" = -ENOBUFS [invalidating]");
 979                 return -ENOBUFS;
 980         }
 981
 982         op = kzalloc(sizeof(*op), GFP_NOIO | __GFP_NOMEMALLOC | __GFP_NORETRY);
 983         if (!op)
 984                 goto nomem;
 985
 986         fscache_operation_init(cookie, &op->op, fscache_write_op, NULL,
 987                                fscache_release_write_op);
 988         op->op.flags = FSCACHE_OP_ASYNC |
 989                 (1 << FSCACHE_OP_WAITING) |
 990                 (1 << FSCACHE_OP_UNUSE_COOKIE);
 991
 992         ret = radix_tree_maybe_preload(gfp & ~__GFP_HIGHMEM);
 993         if (ret < 0)
 994                 goto nomem_free;
 995
 996         trace_fscache_page_op(cookie, page, &op->op, fscache_page_op_write_one);
 997
 998         ret = -ENOBUFS;
 999         spin_lock(&cookie->lock);
1000
1001         if (!fscache_cookie_enabled(cookie) ||
1002             hlist_empty(&cookie->backing_objects))
1003                 goto nobufs;
1004         object = hlist_entry(cookie->backing_objects.first,
1005                              struct fscache_object, cookie_link);
1006         if (test_bit(FSCACHE_IOERROR, &object->cache->flags))
1007                 goto nobufs;
1008
1009         trace_fscache_page(cookie, page, fscache_page_write);
1010
1011         /* add the page to the pending-storage radix tree on the backing
1012          * object */
1013         spin_lock(&object->lock);
1014
1015         if (object->store_limit_l != object_size)
1016                 fscache_set_store_limit(object, object_size);
1017
1018         spin_lock(&cookie->stores_lock);
1019
1020         _debug("store limit %llx", (unsigned long long) object->store_limit);
1021
1022         ret = radix_tree_insert(&cookie->stores, page->index, page);
1023         if (ret < 0) {
1024                 if (ret == -EEXIST)
1025                         goto already_queued;
1026                 _debug("insert failed %d", ret);
1027                 goto nobufs_unlock_obj;
1028         }
1029
1030         trace_fscache_page(cookie, page, fscache_page_radix_insert);
1031         radix_tree_tag_set(&cookie->stores, page->index,
1032                            FSCACHE_COOKIE_PENDING_TAG);
1033         trace_fscache_page(cookie, page, fscache_page_radix_set_pend);
1034         get_page(page);
1035
1036         /* we only want one writer at a time, but we do need to queue new
1037          * writers after exclusive ops */
1038         if (test_and_set_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags))
1039                 goto already_pending;
1040
1041         spin_unlock(&cookie->stores_lock);
1042         spin_unlock(&object->lock);
1043
1044         op->op.debug_id = atomic_inc_return(&fscache_op_debug_id);
1045         op->store_limit = object->store_limit;
1046
1047         __fscache_use_cookie(cookie);
1048         if (fscache_submit_op(object, &op->op) < 0)
1049                 goto submit_failed;
1050
1051         spin_unlock(&cookie->lock);
1052         radix_tree_preload_end();
1053         fscache_stat(&fscache_n_store_ops);
1054         fscache_stat(&fscache_n_stores_ok);
1055
1056         /* the work queue now carries its own ref on the object */
1057         fscache_put_operation(&op->op);
1058         _leave(" = 0");
1059         return 0;
1060
1061 already_queued:
1062         fscache_stat(&fscache_n_stores_again);
1063 already_pending:
1064         spin_unlock(&cookie->stores_lock);
1065         spin_unlock(&object->lock);
1066         spin_unlock(&cookie->lock);
1067         radix_tree_preload_end();
1068         fscache_put_operation(&op->op);
1069         fscache_stat(&fscache_n_stores_ok);
1070         _leave(" = 0");
1071         return 0;
1072
1073 submit_failed:
1074         spin_lock(&cookie->stores_lock);
1075         radix_tree_delete(&cookie->stores, page->index);
1076         trace_fscache_page(cookie, page, fscache_page_radix_delete);
1077         spin_unlock(&cookie->stores_lock);
1078         wake_cookie = __fscache_unuse_cookie(cookie);
1079         put_page(page);
1080         ret = -ENOBUFS;
1081         goto nobufs;
1082
1083 nobufs_unlock_obj:
1084         spin_unlock(&cookie->stores_lock);
1085         spin_unlock(&object->lock);
1086 nobufs:
1087         spin_unlock(&cookie->lock);
1088         radix_tree_preload_end();
1089         fscache_put_operation(&op->op);
1090         if (wake_cookie)
1091                 __fscache_wake_unused_cookie(cookie);
1092         fscache_stat(&fscache_n_stores_nobufs);
1093         _leave(" = -ENOBUFS");
1094         return -ENOBUFS;
1095
1096 nomem_free:
1097         fscache_put_operation(&op->op);
1098 nomem:
1099         fscache_stat(&fscache_n_stores_oom);
1100         _leave(" = -ENOMEM");
1101         return -ENOMEM;
1102 }
1103 EXPORT_SYMBOL(__fscache_write_page);
1104
1105 /*
1106  * remove a page from the cache
1107  */
1108 void __fscache_uncache_page(struct fscache_cookie *cookie, struct page *page)
1109 {
1110         struct fscache_object *object;
1111
1112         _enter(",%p", page);
1113
1114         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
1115         ASSERTCMP(page, !=, NULL);
1116
1117         fscache_stat(&fscache_n_uncaches);
1118
1119         /* cache withdrawal may beat us to it */
1120         if (!PageFsCache(page))
1121                 goto done;
1122
1123         trace_fscache_page(cookie, page, fscache_page_uncache);
1124
1125         /* get the object */
1126         spin_lock(&cookie->lock);
1127
1128         if (hlist_empty(&cookie->backing_objects)) {
1129                 ClearPageFsCache(page);
1130                 goto done_unlock;
1131         }
1132
1133         object = hlist_entry(cookie->backing_objects.first,
1134                              struct fscache_object, cookie_link);
1135
1136         /* there might now be stuff on disk we could read */
1137         clear_bit(FSCACHE_COOKIE_NO_DATA_YET, &cookie->flags);
1138
1139         /* only invoke the cache backend if we managed to mark the page
1140          * uncached here; this deals with synchronisation vs withdrawal */
1141         if (TestClearPageFsCache(page) &&
1142             object->cache->ops->uncache_page) {
1143                 /* the cache backend releases the cookie lock */
1144                 fscache_stat(&fscache_n_cop_uncache_page);
1145                 object->cache->ops->uncache_page(object, page);
1146                 fscache_stat_d(&fscache_n_cop_uncache_page);
1147                 goto done;
1148         }
1149
1150 done_unlock:
1151         spin_unlock(&cookie->lock);
1152 done:
1153         _leave("");
1154 }
1155 EXPORT_SYMBOL(__fscache_uncache_page);
1156
1157 /**
1158  * fscache_mark_page_cached - Mark a page as being cached
1159  * @op: The retrieval op pages are being marked for
1160  * @page: The page to be marked
1161  *
1162  * Mark a netfs page as being cached.  After this is called, the netfs
1163  * must call fscache_uncache_page() to remove the mark.
1164  */
1165 void fscache_mark_page_cached(struct fscache_retrieval *op, struct page *page)
1166 {
1167         struct fscache_cookie *cookie = op->op.object->cookie;
1168
1169 #ifdef CONFIG_FSCACHE_STATS
1170         atomic_inc(&fscache_n_marks);
1171 #endif
1172
1173         trace_fscache_page(cookie, page, fscache_page_cached);
1174
1175         _debug("- mark %p{%lx}", page, page->index);
1176         if (TestSetPageFsCache(page)) {
1177                 static bool once_only;
1178                 if (!once_only) {
1179                         once_only = true;
1180                         pr_warn("Cookie type %s marked page %lx multiple times\n",
1181                                 cookie->def->name, page->index);
1182                 }
1183         }
1184
1185         if (cookie->def->mark_page_cached)
1186                 cookie->def->mark_page_cached(cookie->netfs_data,
1187                                               op->mapping, page);
1188 }
1189 EXPORT_SYMBOL(fscache_mark_page_cached);
1190
1191 /**
1192  * fscache_mark_pages_cached - Mark pages as being cached
1193  * @op: The retrieval op pages are being marked for
1194  * @pagevec: The pages to be marked
1195  *
1196  * Mark a bunch of netfs pages as being cached.  After this is called,
1197  * the netfs must call fscache_uncache_page() to remove the mark.
1198  */
1199 void fscache_mark_pages_cached(struct fscache_retrieval *op,
1200                                struct pagevec *pagevec)
1201 {
1202         unsigned long loop;
1203
1204         for (loop = 0; loop < pagevec->nr; loop++)
1205                 fscache_mark_page_cached(op, pagevec->pages[loop]);
1206
1207         pagevec_reinit(pagevec);
1208 }
1209 EXPORT_SYMBOL(fscache_mark_pages_cached);
1210
1211 /*
1212  * Uncache all the pages in an inode that are marked PG_fscache, assuming them
1213  * to be associated with the given cookie.
1214  */
1215 void __fscache_uncache_all_inode_pages(struct fscache_cookie *cookie,
1216                                        struct inode *inode)
1217 {
1218         struct address_space *mapping = inode->i_mapping;
1219         struct pagevec pvec;
1220         pgoff_t next;
1221         int i;
1222
1223         _enter("%p,%p", cookie, inode);
1224
1225         if (!mapping || mapping->nrpages == 0) {
1226                 _leave(" [no pages]");
1227                 return;
1228         }
1229
1230         pagevec_init(&pvec);
1231         next = 0;
1232         do {
1233                 if (!pagevec_lookup(&pvec, mapping, &next))
1234                         break;
1235                 for (i = 0; i < pagevec_count(&pvec); i++) {
1236                         struct page *page = pvec.pages[i];
1237                         if (PageFsCache(page)) {
1238                                 __fscache_wait_on_page_write(cookie, page);
1239                                 __fscache_uncache_page(cookie, page);
1240                         }
1241                 }
1242                 pagevec_release(&pvec);
1243                 cond_resched();
1244         } while (next);
1245
1246         _leave("");
1247 }
1248 EXPORT_SYMBOL(__fscache_uncache_all_inode_pages);