fs/fscache/page.c

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