fs/fscache/operation.c

   1 /* FS-Cache worker operation management routines
   2  *
   3  * Copyright (C) 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  * See Documentation/filesystems/caching/operations.txt
  12  */
  13
  14 #define FSCACHE_DEBUG_LEVEL OPERATION
  15 #include <linux/module.h>
  16 #include <linux/seq_file.h>
  17 #include <linux/slab.h>
  18 #include "internal.h"
  19
  20 atomic_t fscache_op_debug_id;
  21 EXPORT_SYMBOL(fscache_op_debug_id);
  22
  23 /**
  24  * fscache_operation_init - Do basic initialisation of an operation
  25  * @op: The operation to initialise
  26  * @release: The release function to assign
  27  *
  28  * Do basic initialisation of an operation.  The caller must still set flags,
  29  * object and processor if needed.
  30  */
  31 void fscache_operation_init(struct fscache_operation *op,
  32                             fscache_operation_processor_t processor,
  33                             fscache_operation_release_t release)
  34 {
  35         INIT_WORK(&op->work, fscache_op_work_func);
  36         atomic_set(&op->usage, 1);
  37         op->state = FSCACHE_OP_ST_INITIALISED;
  38         op->debug_id = atomic_inc_return(&fscache_op_debug_id);
  39         op->processor = processor;
  40         op->release = release;
  41         INIT_LIST_HEAD(&op->pend_link);
  42         fscache_stat(&fscache_n_op_initialised);
  43 }
  44 EXPORT_SYMBOL(fscache_operation_init);
  45
  46 /**
  47  * fscache_enqueue_operation - Enqueue an operation for processing
  48  * @op: The operation to enqueue
  49  *
  50  * Enqueue an operation for processing by the FS-Cache thread pool.
  51  *
  52  * This will get its own ref on the object.
  53  */
  54 void fscache_enqueue_operation(struct fscache_operation *op)
  55 {
  56         _enter("{OBJ%x OP%x,%u}",
  57                op->object->debug_id, op->debug_id, atomic_read(&op->usage));
  58
  59         ASSERT(list_empty(&op->pend_link));
  60         ASSERT(op->processor != NULL);
  61         ASSERT(fscache_object_is_available(op->object));
  62         ASSERTCMP(atomic_read(&op->usage), >, 0);
  63         ASSERTCMP(op->state, ==, FSCACHE_OP_ST_IN_PROGRESS);
  64
  65         fscache_stat(&fscache_n_op_enqueue);
  66         switch (op->flags & FSCACHE_OP_TYPE) {
  67         case FSCACHE_OP_ASYNC:
  68                 _debug("queue async");
  69                 atomic_inc(&op->usage);
  70                 if (!queue_work(fscache_op_wq, &op->work))
  71                         fscache_put_operation(op);
  72                 break;
  73         case FSCACHE_OP_MYTHREAD:
  74                 _debug("queue for caller's attention");
  75                 break;
  76         default:
  77                 pr_err("Unexpected op type %lx", op->flags);
  78                 BUG();
  79                 break;
  80         }
  81 }
  82 EXPORT_SYMBOL(fscache_enqueue_operation);
  83
  84 /*
  85  * start an op running
  86  */
  87 static void fscache_run_op(struct fscache_object *object,
  88                            struct fscache_operation *op)
  89 {
  90         ASSERTCMP(op->state, ==, FSCACHE_OP_ST_PENDING);
  91
  92         op->state = FSCACHE_OP_ST_IN_PROGRESS;
  93         object->n_in_progress++;
  94         if (test_and_clear_bit(FSCACHE_OP_WAITING, &op->flags))
  95                 wake_up_bit(&op->flags, FSCACHE_OP_WAITING);
  96         if (op->processor)
  97                 fscache_enqueue_operation(op);
  98         fscache_stat(&fscache_n_op_run);
  99 }
 100
 101 /*
 102  * report an unexpected submission
 103  */
 104 static void fscache_report_unexpected_submission(struct fscache_object *object,
 105                                                  struct fscache_operation *op,
 106                                                  const struct fscache_state *ostate)
 107 {
 108         static bool once_only;
 109         struct fscache_operation *p;
 110         unsigned n;
 111
 112         if (once_only)
 113                 return;
 114         once_only = true;
 115
 116         kdebug("unexpected submission OP%x [OBJ%x %s]",
 117                op->debug_id, object->debug_id, object->state->name);
 118         kdebug("objstate=%s [%s]", object->state->name, ostate->name);
 119         kdebug("objflags=%lx", object->flags);
 120         kdebug("objevent=%lx [%lx]", object->events, object->event_mask);
 121         kdebug("ops=%u inp=%u exc=%u",
 122                object->n_ops, object->n_in_progress, object->n_exclusive);
 123
 124         if (!list_empty(&object->pending_ops)) {
 125                 n = 0;
 126                 list_for_each_entry(p, &object->pending_ops, pend_link) {
 127                         ASSERTCMP(p->object, ==, object);
 128                         kdebug("%p %p", op->processor, op->release);
 129                         n++;
 130                 }
 131
 132                 kdebug("n=%u", n);
 133         }
 134
 135         dump_stack();
 136 }
 137
 138 /*
 139  * submit an exclusive operation for an object
 140  * - other ops are excluded from running simultaneously with this one
 141  * - this gets any extra refs it needs on an op
 142  */
 143 int fscache_submit_exclusive_op(struct fscache_object *object,
 144                                 struct fscache_operation *op)
 145 {
 146         const struct fscache_state *ostate;
 147         unsigned long flags;
 148         int ret;
 149
 150         _enter("{OBJ%x OP%x},", object->debug_id, op->debug_id);
 151
 152         ASSERTCMP(op->state, ==, FSCACHE_OP_ST_INITIALISED);
 153         ASSERTCMP(atomic_read(&op->usage), >, 0);
 154
 155         spin_lock(&object->lock);
 156         ASSERTCMP(object->n_ops, >=, object->n_in_progress);
 157         ASSERTCMP(object->n_ops, >=, object->n_exclusive);
 158         ASSERT(list_empty(&op->pend_link));
 159
 160         ostate = object->state;
 161         smp_rmb();
 162
 163         op->state = FSCACHE_OP_ST_PENDING;
 164         flags = READ_ONCE(object->flags);
 165         if (unlikely(!(flags & BIT(FSCACHE_OBJECT_IS_LIVE)))) {
 166                 fscache_stat(&fscache_n_op_rejected);
 167                 op->state = FSCACHE_OP_ST_CANCELLED;
 168                 ret = -ENOBUFS;
 169         } else if (unlikely(fscache_cache_is_broken(object))) {
 170                 op->state = FSCACHE_OP_ST_CANCELLED;
 171                 ret = -EIO;
 172         } else if (flags & BIT(FSCACHE_OBJECT_IS_AVAILABLE)) {
 173                 op->object = object;
 174                 object->n_ops++;
 175                 object->n_exclusive++;  /* reads and writes must wait */
 176
 177                 if (object->n_in_progress > 0) {
 178                         atomic_inc(&op->usage);
 179                         list_add_tail(&op->pend_link, &object->pending_ops);
 180                         fscache_stat(&fscache_n_op_pend);
 181                 } else if (!list_empty(&object->pending_ops)) {
 182                         atomic_inc(&op->usage);
 183                         list_add_tail(&op->pend_link, &object->pending_ops);
 184                         fscache_stat(&fscache_n_op_pend);
 185                         fscache_start_operations(object);
 186                 } else {
 187                         ASSERTCMP(object->n_in_progress, ==, 0);
 188                         fscache_run_op(object, op);
 189                 }
 190
 191                 /* need to issue a new write op after this */
 192                 clear_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags);
 193                 ret = 0;
 194         } else if (flags & BIT(FSCACHE_OBJECT_IS_LOOKED_UP)) {
 195                 op->object = object;
 196                 object->n_ops++;
 197                 object->n_exclusive++;  /* reads and writes must wait */
 198                 atomic_inc(&op->usage);
 199                 list_add_tail(&op->pend_link, &object->pending_ops);
 200                 fscache_stat(&fscache_n_op_pend);
 201                 ret = 0;
 202         } else if (flags & BIT(FSCACHE_OBJECT_KILLED_BY_CACHE)) {
 203                 op->state = FSCACHE_OP_ST_CANCELLED;
 204                 ret = -ENOBUFS;
 205         } else {
 206                 fscache_report_unexpected_submission(object, op, ostate);
 207                 op->state = FSCACHE_OP_ST_CANCELLED;
 208                 ret = -ENOBUFS;
 209         }
 210
 211         spin_unlock(&object->lock);
 212         return ret;
 213 }
 214
 215 /*
 216  * submit an operation for an object
 217  * - objects may be submitted only in the following states:
 218  *   - during object creation (write ops may be submitted)
 219  *   - whilst the object is active
 220  *   - after an I/O error incurred in one of the two above states (op rejected)
 221  * - this gets any extra refs it needs on an op
 222  */
 223 int fscache_submit_op(struct fscache_object *object,
 224                       struct fscache_operation *op)
 225 {
 226         const struct fscache_state *ostate;
 227         unsigned long flags;
 228         int ret;
 229
 230         _enter("{OBJ%x OP%x},{%u}",
 231                object->debug_id, op->debug_id, atomic_read(&op->usage));
 232
 233         ASSERTCMP(op->state, ==, FSCACHE_OP_ST_INITIALISED);
 234         ASSERTCMP(atomic_read(&op->usage), >, 0);
 235
 236         spin_lock(&object->lock);
 237         ASSERTCMP(object->n_ops, >=, object->n_in_progress);
 238         ASSERTCMP(object->n_ops, >=, object->n_exclusive);
 239         ASSERT(list_empty(&op->pend_link));
 240
 241         ostate = object->state;
 242         smp_rmb();
 243
 244         op->state = FSCACHE_OP_ST_PENDING;
 245         flags = READ_ONCE(object->flags);
 246         if (unlikely(!(flags & BIT(FSCACHE_OBJECT_IS_LIVE)))) {
 247                 fscache_stat(&fscache_n_op_rejected);
 248                 op->state = FSCACHE_OP_ST_CANCELLED;
 249                 ret = -ENOBUFS;
 250         } else if (unlikely(fscache_cache_is_broken(object))) {
 251                 op->state = FSCACHE_OP_ST_CANCELLED;
 252                 ret = -EIO;
 253         } else if (flags & BIT(FSCACHE_OBJECT_IS_AVAILABLE)) {
 254                 op->object = object;
 255                 object->n_ops++;
 256
 257                 if (object->n_exclusive > 0) {
 258                         atomic_inc(&op->usage);
 259                         list_add_tail(&op->pend_link, &object->pending_ops);
 260                         fscache_stat(&fscache_n_op_pend);
 261                 } else if (!list_empty(&object->pending_ops)) {
 262                         atomic_inc(&op->usage);
 263                         list_add_tail(&op->pend_link, &object->pending_ops);
 264                         fscache_stat(&fscache_n_op_pend);
 265                         fscache_start_operations(object);
 266                 } else {
 267                         ASSERTCMP(object->n_exclusive, ==, 0);
 268                         fscache_run_op(object, op);
 269                 }
 270                 ret = 0;
 271         } else if (flags & BIT(FSCACHE_OBJECT_IS_LOOKED_UP)) {
 272                 op->object = object;
 273                 object->n_ops++;
 274                 atomic_inc(&op->usage);
 275                 list_add_tail(&op->pend_link, &object->pending_ops);
 276                 fscache_stat(&fscache_n_op_pend);
 277                 ret = 0;
 278         } else if (flags & BIT(FSCACHE_OBJECT_KILLED_BY_CACHE)) {
 279                 op->state = FSCACHE_OP_ST_CANCELLED;
 280                 ret = -ENOBUFS;
 281         } else {
 282                 fscache_report_unexpected_submission(object, op, ostate);
 283                 ASSERT(!fscache_object_is_active(object));
 284                 op->state = FSCACHE_OP_ST_CANCELLED;
 285                 ret = -ENOBUFS;
 286         }
 287
 288         spin_unlock(&object->lock);
 289         return ret;
 290 }
 291
 292 /*
 293  * queue an object for withdrawal on error, aborting all following asynchronous
 294  * operations
 295  */
 296 void fscache_abort_object(struct fscache_object *object)
 297 {
 298         _enter("{OBJ%x}", object->debug_id);
 299
 300         fscache_raise_event(object, FSCACHE_OBJECT_EV_ERROR);
 301 }
 302
 303 /*
 304  * Jump start the operation processing on an object.  The caller must hold
 305  * object->lock.
 306  */
 307 void fscache_start_operations(struct fscache_object *object)
 308 {
 309         struct fscache_operation *op;
 310         bool stop = false;
 311
 312         while (!list_empty(&object->pending_ops) && !stop) {
 313                 op = list_entry(object->pending_ops.next,
 314                                 struct fscache_operation, pend_link);
 315
 316                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags)) {
 317                         if (object->n_in_progress > 0)
 318                                 break;
 319                         stop = true;
 320                 }
 321                 list_del_init(&op->pend_link);
 322                 fscache_run_op(object, op);
 323
 324                 /* the pending queue was holding a ref on the object */
 325                 fscache_put_operation(op);
 326         }
 327
 328         ASSERTCMP(object->n_in_progress, <=, object->n_ops);
 329
 330         _debug("woke %d ops on OBJ%x",
 331                object->n_in_progress, object->debug_id);
 332 }
 333
 334 /*
 335  * cancel an operation that's pending on an object
 336  */
 337 int fscache_cancel_op(struct fscache_operation *op,
 338                       void (*do_cancel)(struct fscache_operation *),
 339                       bool cancel_in_progress_op)
 340 {
 341         struct fscache_object *object = op->object;
 342         bool put = false;
 343         int ret;
 344
 345         _enter("OBJ%x OP%x}", op->object->debug_id, op->debug_id);
 346
 347         ASSERTCMP(op->state, >=, FSCACHE_OP_ST_PENDING);
 348         ASSERTCMP(op->state, !=, FSCACHE_OP_ST_CANCELLED);
 349         ASSERTCMP(atomic_read(&op->usage), >, 0);
 350
 351         spin_lock(&object->lock);
 352
 353         ret = -EBUSY;
 354         if (op->state == FSCACHE_OP_ST_PENDING) {
 355                 ASSERT(!list_empty(&op->pend_link));
 356                 list_del_init(&op->pend_link);
 357                 put = true;
 358                 fscache_stat(&fscache_n_op_cancelled);
 359                 if (do_cancel)
 360                         do_cancel(op);
 361                 op->state = FSCACHE_OP_ST_CANCELLED;
 362                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 363                         object->n_exclusive--;
 364                 if (test_and_clear_bit(FSCACHE_OP_WAITING, &op->flags))
 365                         wake_up_bit(&op->flags, FSCACHE_OP_WAITING);
 366                 ret = 0;
 367         } else if (op->state == FSCACHE_OP_ST_IN_PROGRESS && cancel_in_progress_op) {
 368                 fscache_stat(&fscache_n_op_cancelled);
 369                 if (do_cancel)
 370                         do_cancel(op);
 371                 op->state = FSCACHE_OP_ST_CANCELLED;
 372                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 373                         object->n_exclusive--;
 374                 if (test_and_clear_bit(FSCACHE_OP_WAITING, &op->flags))
 375                         wake_up_bit(&op->flags, FSCACHE_OP_WAITING);
 376                 ret = 0;
 377         }
 378
 379         if (put)
 380                 fscache_put_operation(op);
 381         spin_unlock(&object->lock);
 382         _leave(" = %d", ret);
 383         return ret;
 384 }
 385
 386 /*
 387  * Cancel all pending operations on an object
 388  */
 389 void fscache_cancel_all_ops(struct fscache_object *object)
 390 {
 391         struct fscache_operation *op;
 392
 393         _enter("OBJ%x", object->debug_id);
 394
 395         spin_lock(&object->lock);
 396
 397         while (!list_empty(&object->pending_ops)) {
 398                 op = list_entry(object->pending_ops.next,
 399                                 struct fscache_operation, pend_link);
 400                 fscache_stat(&fscache_n_op_cancelled);
 401                 list_del_init(&op->pend_link);
 402
 403                 ASSERTCMP(op->state, ==, FSCACHE_OP_ST_PENDING);
 404                 op->state = FSCACHE_OP_ST_CANCELLED;
 405
 406                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 407                         object->n_exclusive--;
 408                 if (test_and_clear_bit(FSCACHE_OP_WAITING, &op->flags))
 409                         wake_up_bit(&op->flags, FSCACHE_OP_WAITING);
 410                 fscache_put_operation(op);
 411                 cond_resched_lock(&object->lock);
 412         }
 413
 414         spin_unlock(&object->lock);
 415         _leave("");
 416 }
 417
 418 /*
 419  * Record the completion or cancellation of an in-progress operation.
 420  */
 421 void fscache_op_complete(struct fscache_operation *op, bool cancelled)
 422 {
 423         struct fscache_object *object = op->object;
 424
 425         _enter("OBJ%x", object->debug_id);
 426
 427         ASSERTCMP(op->state, ==, FSCACHE_OP_ST_IN_PROGRESS);
 428         ASSERTCMP(object->n_in_progress, >, 0);
 429         ASSERTIFCMP(test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags),
 430                     object->n_exclusive, >, 0);
 431         ASSERTIFCMP(test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags),
 432                     object->n_in_progress, ==, 1);
 433
 434         spin_lock(&object->lock);
 435
 436         op->state = cancelled ?
 437                 FSCACHE_OP_ST_CANCELLED : FSCACHE_OP_ST_COMPLETE;
 438
 439         if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 440                 object->n_exclusive--;
 441         object->n_in_progress--;
 442         if (object->n_in_progress == 0)
 443                 fscache_start_operations(object);
 444
 445         spin_unlock(&object->lock);
 446         _leave("");
 447 }
 448 EXPORT_SYMBOL(fscache_op_complete);
 449
 450 /*
 451  * release an operation
 452  * - queues pending ops if this is the last in-progress op
 453  */
 454 void fscache_put_operation(struct fscache_operation *op)
 455 {
 456         struct fscache_object *object;
 457         struct fscache_cache *cache;
 458
 459         _enter("{OBJ%x OP%x,%d}",
 460                op->object->debug_id, op->debug_id, atomic_read(&op->usage));
 461
 462         ASSERTCMP(atomic_read(&op->usage), >, 0);
 463
 464         if (!atomic_dec_and_test(&op->usage))
 465                 return;
 466
 467         _debug("PUT OP");
 468         ASSERTIFCMP(op->state != FSCACHE_OP_ST_COMPLETE,
 469                     op->state, ==, FSCACHE_OP_ST_CANCELLED);
 470         op->state = FSCACHE_OP_ST_DEAD;
 471
 472         fscache_stat(&fscache_n_op_release);
 473
 474         if (op->release) {
 475                 op->release(op);
 476                 op->release = NULL;
 477         }
 478
 479         object = op->object;
 480
 481         if (test_bit(FSCACHE_OP_DEC_READ_CNT, &op->flags))
 482                 atomic_dec(&object->n_reads);
 483         if (test_bit(FSCACHE_OP_UNUSE_COOKIE, &op->flags))
 484                 fscache_unuse_cookie(object);
 485
 486         /* now... we may get called with the object spinlock held, so we
 487          * complete the cleanup here only if we can immediately acquire the
 488          * lock, and defer it otherwise */
 489         if (!spin_trylock(&object->lock)) {
 490                 _debug("defer put");
 491                 fscache_stat(&fscache_n_op_deferred_release);
 492
 493                 cache = object->cache;
 494                 spin_lock(&cache->op_gc_list_lock);
 495                 list_add_tail(&op->pend_link, &cache->op_gc_list);
 496                 spin_unlock(&cache->op_gc_list_lock);
 497                 schedule_work(&cache->op_gc);
 498                 _leave(" [defer]");
 499                 return;
 500         }
 501
 502         ASSERTCMP(object->n_ops, >, 0);
 503         object->n_ops--;
 504         if (object->n_ops == 0)
 505                 fscache_raise_event(object, FSCACHE_OBJECT_EV_CLEARED);
 506
 507         spin_unlock(&object->lock);
 508
 509         kfree(op);
 510         _leave(" [done]");
 511 }
 512 EXPORT_SYMBOL(fscache_put_operation);
 513
 514 /*
 515  * garbage collect operations that have had their release deferred
 516  */
 517 void fscache_operation_gc(struct work_struct *work)
 518 {
 519         struct fscache_operation *op;
 520         struct fscache_object *object;
 521         struct fscache_cache *cache =
 522                 container_of(work, struct fscache_cache, op_gc);
 523         int count = 0;
 524
 525         _enter("");
 526
 527         do {
 528                 spin_lock(&cache->op_gc_list_lock);
 529                 if (list_empty(&cache->op_gc_list)) {
 530                         spin_unlock(&cache->op_gc_list_lock);
 531                         break;
 532                 }
 533
 534                 op = list_entry(cache->op_gc_list.next,
 535                                 struct fscache_operation, pend_link);
 536                 list_del(&op->pend_link);
 537                 spin_unlock(&cache->op_gc_list_lock);
 538
 539                 object = op->object;
 540                 spin_lock(&object->lock);
 541
 542                 _debug("GC DEFERRED REL OBJ%x OP%x",
 543                        object->debug_id, op->debug_id);
 544                 fscache_stat(&fscache_n_op_gc);
 545
 546                 ASSERTCMP(atomic_read(&op->usage), ==, 0);
 547                 ASSERTCMP(op->state, ==, FSCACHE_OP_ST_DEAD);
 548
 549                 ASSERTCMP(object->n_ops, >, 0);
 550                 object->n_ops--;
 551                 if (object->n_ops == 0)
 552                         fscache_raise_event(object, FSCACHE_OBJECT_EV_CLEARED);
 553
 554                 spin_unlock(&object->lock);
 555                 kfree(op);
 556
 557         } while (count++ < 20);
 558
 559         if (!list_empty(&cache->op_gc_list))
 560                 schedule_work(&cache->op_gc);
 561
 562         _leave("");
 563 }
 564
 565 /*
 566  * execute an operation using fs_op_wq to provide processing context -
 567  * the caller holds a ref to this object, so we don't need to hold one
 568  */
 569 void fscache_op_work_func(struct work_struct *work)
 570 {
 571         struct fscache_operation *op =
 572                 container_of(work, struct fscache_operation, work);
 573         unsigned long start;
 574
 575         _enter("{OBJ%x OP%x,%d}",
 576                op->object->debug_id, op->debug_id, atomic_read(&op->usage));
 577
 578         ASSERT(op->processor != NULL);
 579         start = jiffies;
 580         op->processor(op);
 581         fscache_hist(fscache_ops_histogram, start);
 582         fscache_put_operation(op);
 583
 584         _leave("");
 585 }