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 static void fscache_operation_dummy_cancel(struct fscache_operation *op)
  24 {
  25 }
  26
  27 /**
  28  * fscache_operation_init - Do basic initialisation of an operation
  29  * @op: The operation to initialise
  30  * @release: The release function to assign
  31  *
  32  * Do basic initialisation of an operation.  The caller must still set flags,
  33  * object and processor if needed.
  34  */
  35 void fscache_operation_init(struct fscache_operation *op,
  36                             fscache_operation_processor_t processor,
  37                             fscache_operation_cancel_t cancel,
  38                             fscache_operation_release_t release)
  39 {
  40         INIT_WORK(&op->work, fscache_op_work_func);
  41         atomic_set(&op->usage, 1);
  42         op->state = FSCACHE_OP_ST_INITIALISED;
  43         op->debug_id = atomic_inc_return(&fscache_op_debug_id);
  44         op->processor = processor;
  45         op->cancel = cancel ?: fscache_operation_dummy_cancel;
  46         op->release = release;
  47         INIT_LIST_HEAD(&op->pend_link);
  48         fscache_stat(&fscache_n_op_initialised);
  49 }
  50 EXPORT_SYMBOL(fscache_operation_init);
  51
  52 /**
  53  * fscache_enqueue_operation - Enqueue an operation for processing
  54  * @op: The operation to enqueue
  55  *
  56  * Enqueue an operation for processing by the FS-Cache thread pool.
  57  *
  58  * This will get its own ref on the object.
  59  */
  60 void fscache_enqueue_operation(struct fscache_operation *op)
  61 {
  62         _enter("{OBJ%x OP%x,%u}",
  63                op->object->debug_id, op->debug_id, atomic_read(&op->usage));
  64
  65         ASSERT(list_empty(&op->pend_link));
  66         ASSERT(op->processor != NULL);
  67         ASSERT(fscache_object_is_available(op->object));
  68         ASSERTCMP(atomic_read(&op->usage), >, 0);
  69         ASSERTIFCMP(op->state != FSCACHE_OP_ST_IN_PROGRESS,
  70                     op->state, ==,  FSCACHE_OP_ST_CANCELLED);
  71
  72         fscache_stat(&fscache_n_op_enqueue);
  73         switch (op->flags & FSCACHE_OP_TYPE) {
  74         case FSCACHE_OP_ASYNC:
  75                 _debug("queue async");
  76                 atomic_inc(&op->usage);
  77                 if (!queue_work(fscache_op_wq, &op->work))
  78                         fscache_put_operation(op);
  79                 break;
  80         case FSCACHE_OP_MYTHREAD:
  81                 _debug("queue for caller's attention");
  82                 break;
  83         default:
  84                 pr_err("Unexpected op type %lx", op->flags);
  85                 BUG();
  86                 break;
  87         }
  88 }
  89 EXPORT_SYMBOL(fscache_enqueue_operation);
  90
  91 /*
  92  * start an op running
  93  */
  94 static void fscache_run_op(struct fscache_object *object,
  95                            struct fscache_operation *op)
  96 {
  97         ASSERTCMP(op->state, ==, FSCACHE_OP_ST_PENDING);
  98
  99         op->state = FSCACHE_OP_ST_IN_PROGRESS;
 100         object->n_in_progress++;
 101         if (test_and_clear_bit(FSCACHE_OP_WAITING, &op->flags))
 102                 wake_up_bit(&op->flags, FSCACHE_OP_WAITING);
 103         if (op->processor)
 104                 fscache_enqueue_operation(op);
 105         fscache_stat(&fscache_n_op_run);
 106 }
 107
 108 /*
 109  * report an unexpected submission
 110  */
 111 static void fscache_report_unexpected_submission(struct fscache_object *object,
 112                                                  struct fscache_operation *op,
 113                                                  const struct fscache_state *ostate)
 114 {
 115         static bool once_only;
 116         struct fscache_operation *p;
 117         unsigned n;
 118
 119         if (once_only)
 120                 return;
 121         once_only = true;
 122
 123         kdebug("unexpected submission OP%x [OBJ%x %s]",
 124                op->debug_id, object->debug_id, object->state->name);
 125         kdebug("objstate=%s [%s]", object->state->name, ostate->name);
 126         kdebug("objflags=%lx", object->flags);
 127         kdebug("objevent=%lx [%lx]", object->events, object->event_mask);
 128         kdebug("ops=%u inp=%u exc=%u",
 129                object->n_ops, object->n_in_progress, object->n_exclusive);
 130
 131         if (!list_empty(&object->pending_ops)) {
 132                 n = 0;
 133                 list_for_each_entry(p, &object->pending_ops, pend_link) {
 134                         ASSERTCMP(p->object, ==, object);
 135                         kdebug("%p %p", op->processor, op->release);
 136                         n++;
 137                 }
 138
 139                 kdebug("n=%u", n);
 140         }
 141
 142         dump_stack();
 143 }
 144
 145 /*
 146  * submit an exclusive operation for an object
 147  * - other ops are excluded from running simultaneously with this one
 148  * - this gets any extra refs it needs on an op
 149  */
 150 int fscache_submit_exclusive_op(struct fscache_object *object,
 151                                 struct fscache_operation *op)
 152 {
 153         const struct fscache_state *ostate;
 154         unsigned long flags;
 155         int ret;
 156
 157         _enter("{OBJ%x OP%x},", object->debug_id, op->debug_id);
 158
 159         ASSERTCMP(op->state, ==, FSCACHE_OP_ST_INITIALISED);
 160         ASSERTCMP(atomic_read(&op->usage), >, 0);
 161
 162         spin_lock(&object->lock);
 163         ASSERTCMP(object->n_ops, >=, object->n_in_progress);
 164         ASSERTCMP(object->n_ops, >=, object->n_exclusive);
 165         ASSERT(list_empty(&op->pend_link));
 166
 167         ostate = object->state;
 168         smp_rmb();
 169
 170         op->state = FSCACHE_OP_ST_PENDING;
 171         flags = READ_ONCE(object->flags);
 172         if (unlikely(!(flags & BIT(FSCACHE_OBJECT_IS_LIVE)))) {
 173                 fscache_stat(&fscache_n_op_rejected);
 174                 op->cancel(op);
 175                 op->state = FSCACHE_OP_ST_CANCELLED;
 176                 ret = -ENOBUFS;
 177         } else if (unlikely(fscache_cache_is_broken(object))) {
 178                 op->cancel(op);
 179                 op->state = FSCACHE_OP_ST_CANCELLED;
 180                 ret = -EIO;
 181         } else if (flags & BIT(FSCACHE_OBJECT_IS_AVAILABLE)) {
 182                 op->object = object;
 183                 object->n_ops++;
 184                 object->n_exclusive++;  /* reads and writes must wait */
 185
 186                 if (object->n_in_progress > 0) {
 187                         atomic_inc(&op->usage);
 188                         list_add_tail(&op->pend_link, &object->pending_ops);
 189                         fscache_stat(&fscache_n_op_pend);
 190                 } else if (!list_empty(&object->pending_ops)) {
 191                         atomic_inc(&op->usage);
 192                         list_add_tail(&op->pend_link, &object->pending_ops);
 193                         fscache_stat(&fscache_n_op_pend);
 194                         fscache_start_operations(object);
 195                 } else {
 196                         ASSERTCMP(object->n_in_progress, ==, 0);
 197                         fscache_run_op(object, op);
 198                 }
 199
 200                 /* need to issue a new write op after this */
 201                 clear_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags);
 202                 ret = 0;
 203         } else if (flags & BIT(FSCACHE_OBJECT_IS_LOOKED_UP)) {
 204                 op->object = object;
 205                 object->n_ops++;
 206                 object->n_exclusive++;  /* reads and writes must wait */
 207                 atomic_inc(&op->usage);
 208                 list_add_tail(&op->pend_link, &object->pending_ops);
 209                 fscache_stat(&fscache_n_op_pend);
 210                 ret = 0;
 211         } else if (flags & BIT(FSCACHE_OBJECT_KILLED_BY_CACHE)) {
 212                 op->cancel(op);
 213                 op->state = FSCACHE_OP_ST_CANCELLED;
 214                 ret = -ENOBUFS;
 215         } else {
 216                 fscache_report_unexpected_submission(object, op, ostate);
 217                 op->cancel(op);
 218                 op->state = FSCACHE_OP_ST_CANCELLED;
 219                 ret = -ENOBUFS;
 220         }
 221
 222         spin_unlock(&object->lock);
 223         return ret;
 224 }
 225
 226 /*
 227  * submit an operation for an object
 228  * - objects may be submitted only in the following states:
 229  *   - during object creation (write ops may be submitted)
 230  *   - whilst the object is active
 231  *   - after an I/O error incurred in one of the two above states (op rejected)
 232  * - this gets any extra refs it needs on an op
 233  */
 234 int fscache_submit_op(struct fscache_object *object,
 235                       struct fscache_operation *op)
 236 {
 237         const struct fscache_state *ostate;
 238         unsigned long flags;
 239         int ret;
 240
 241         _enter("{OBJ%x OP%x},{%u}",
 242                object->debug_id, op->debug_id, atomic_read(&op->usage));
 243
 244         ASSERTCMP(op->state, ==, FSCACHE_OP_ST_INITIALISED);
 245         ASSERTCMP(atomic_read(&op->usage), >, 0);
 246
 247         spin_lock(&object->lock);
 248         ASSERTCMP(object->n_ops, >=, object->n_in_progress);
 249         ASSERTCMP(object->n_ops, >=, object->n_exclusive);
 250         ASSERT(list_empty(&op->pend_link));
 251
 252         ostate = object->state;
 253         smp_rmb();
 254
 255         op->state = FSCACHE_OP_ST_PENDING;
 256         flags = READ_ONCE(object->flags);
 257         if (unlikely(!(flags & BIT(FSCACHE_OBJECT_IS_LIVE)))) {
 258                 fscache_stat(&fscache_n_op_rejected);
 259                 op->cancel(op);
 260                 op->state = FSCACHE_OP_ST_CANCELLED;
 261                 ret = -ENOBUFS;
 262         } else if (unlikely(fscache_cache_is_broken(object))) {
 263                 op->cancel(op);
 264                 op->state = FSCACHE_OP_ST_CANCELLED;
 265                 ret = -EIO;
 266         } else if (flags & BIT(FSCACHE_OBJECT_IS_AVAILABLE)) {
 267                 op->object = object;
 268                 object->n_ops++;
 269
 270                 if (object->n_exclusive > 0) {
 271                         atomic_inc(&op->usage);
 272                         list_add_tail(&op->pend_link, &object->pending_ops);
 273                         fscache_stat(&fscache_n_op_pend);
 274                 } else if (!list_empty(&object->pending_ops)) {
 275                         atomic_inc(&op->usage);
 276                         list_add_tail(&op->pend_link, &object->pending_ops);
 277                         fscache_stat(&fscache_n_op_pend);
 278                         fscache_start_operations(object);
 279                 } else {
 280                         ASSERTCMP(object->n_exclusive, ==, 0);
 281                         fscache_run_op(object, op);
 282                 }
 283                 ret = 0;
 284         } else if (flags & BIT(FSCACHE_OBJECT_IS_LOOKED_UP)) {
 285                 op->object = object;
 286                 object->n_ops++;
 287                 atomic_inc(&op->usage);
 288                 list_add_tail(&op->pend_link, &object->pending_ops);
 289                 fscache_stat(&fscache_n_op_pend);
 290                 ret = 0;
 291         } else if (flags & BIT(FSCACHE_OBJECT_KILLED_BY_CACHE)) {
 292                 op->cancel(op);
 293                 op->state = FSCACHE_OP_ST_CANCELLED;
 294                 ret = -ENOBUFS;
 295         } else {
 296                 fscache_report_unexpected_submission(object, op, ostate);
 297                 ASSERT(!fscache_object_is_active(object));
 298                 op->cancel(op);
 299                 op->state = FSCACHE_OP_ST_CANCELLED;
 300                 ret = -ENOBUFS;
 301         }
 302
 303         spin_unlock(&object->lock);
 304         return ret;
 305 }
 306
 307 /*
 308  * queue an object for withdrawal on error, aborting all following asynchronous
 309  * operations
 310  */
 311 void fscache_abort_object(struct fscache_object *object)
 312 {
 313         _enter("{OBJ%x}", object->debug_id);
 314
 315         fscache_raise_event(object, FSCACHE_OBJECT_EV_ERROR);
 316 }
 317
 318 /*
 319  * Jump start the operation processing on an object.  The caller must hold
 320  * object->lock.
 321  */
 322 void fscache_start_operations(struct fscache_object *object)
 323 {
 324         struct fscache_operation *op;
 325         bool stop = false;
 326
 327         while (!list_empty(&object->pending_ops) && !stop) {
 328                 op = list_entry(object->pending_ops.next,
 329                                 struct fscache_operation, pend_link);
 330
 331                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags)) {
 332                         if (object->n_in_progress > 0)
 333                                 break;
 334                         stop = true;
 335                 }
 336                 list_del_init(&op->pend_link);
 337                 fscache_run_op(object, op);
 338
 339                 /* the pending queue was holding a ref on the object */
 340                 fscache_put_operation(op);
 341         }
 342
 343         ASSERTCMP(object->n_in_progress, <=, object->n_ops);
 344
 345         _debug("woke %d ops on OBJ%x",
 346                object->n_in_progress, object->debug_id);
 347 }
 348
 349 /*
 350  * cancel an operation that's pending on an object
 351  */
 352 int fscache_cancel_op(struct fscache_operation *op,
 353                       bool cancel_in_progress_op)
 354 {
 355         struct fscache_object *object = op->object;
 356         bool put = false;
 357         int ret;
 358
 359         _enter("OBJ%x OP%x}", op->object->debug_id, op->debug_id);
 360
 361         ASSERTCMP(op->state, >=, FSCACHE_OP_ST_PENDING);
 362         ASSERTCMP(op->state, !=, FSCACHE_OP_ST_CANCELLED);
 363         ASSERTCMP(atomic_read(&op->usage), >, 0);
 364
 365         spin_lock(&object->lock);
 366
 367         ret = -EBUSY;
 368         if (op->state == FSCACHE_OP_ST_PENDING) {
 369                 ASSERT(!list_empty(&op->pend_link));
 370                 list_del_init(&op->pend_link);
 371                 put = true;
 372
 373                 fscache_stat(&fscache_n_op_cancelled);
 374                 op->cancel(op);
 375                 op->state = FSCACHE_OP_ST_CANCELLED;
 376                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 377                         object->n_exclusive--;
 378                 if (test_and_clear_bit(FSCACHE_OP_WAITING, &op->flags))
 379                         wake_up_bit(&op->flags, FSCACHE_OP_WAITING);
 380                 ret = 0;
 381         } else if (op->state == FSCACHE_OP_ST_IN_PROGRESS && cancel_in_progress_op) {
 382                 ASSERTCMP(object->n_in_progress, >, 0);
 383                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 384                         object->n_exclusive--;
 385                 object->n_in_progress--;
 386                 if (object->n_in_progress == 0)
 387                         fscache_start_operations(object);
 388
 389                 fscache_stat(&fscache_n_op_cancelled);
 390                 op->cancel(op);
 391                 op->state = FSCACHE_OP_ST_CANCELLED;
 392                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 393                         object->n_exclusive--;
 394                 if (test_and_clear_bit(FSCACHE_OP_WAITING, &op->flags))
 395                         wake_up_bit(&op->flags, FSCACHE_OP_WAITING);
 396                 ret = 0;
 397         }
 398
 399         if (put)
 400                 fscache_put_operation(op);
 401         spin_unlock(&object->lock);
 402         _leave(" = %d", ret);
 403         return ret;
 404 }
 405
 406 /*
 407  * Cancel all pending operations on an object
 408  */
 409 void fscache_cancel_all_ops(struct fscache_object *object)
 410 {
 411         struct fscache_operation *op;
 412
 413         _enter("OBJ%x", object->debug_id);
 414
 415         spin_lock(&object->lock);
 416
 417         while (!list_empty(&object->pending_ops)) {
 418                 op = list_entry(object->pending_ops.next,
 419                                 struct fscache_operation, pend_link);
 420                 fscache_stat(&fscache_n_op_cancelled);
 421                 list_del_init(&op->pend_link);
 422
 423                 ASSERTCMP(op->state, ==, FSCACHE_OP_ST_PENDING);
 424                 op->cancel(op);
 425                 op->state = FSCACHE_OP_ST_CANCELLED;
 426
 427                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 428                         object->n_exclusive--;
 429                 if (test_and_clear_bit(FSCACHE_OP_WAITING, &op->flags))
 430                         wake_up_bit(&op->flags, FSCACHE_OP_WAITING);
 431                 fscache_put_operation(op);
 432                 cond_resched_lock(&object->lock);
 433         }
 434
 435         spin_unlock(&object->lock);
 436         _leave("");
 437 }
 438
 439 /*
 440  * Record the completion or cancellation of an in-progress operation.
 441  */
 442 void fscache_op_complete(struct fscache_operation *op, bool cancelled)
 443 {
 444         struct fscache_object *object = op->object;
 445
 446         _enter("OBJ%x", object->debug_id);
 447
 448         ASSERTCMP(op->state, ==, FSCACHE_OP_ST_IN_PROGRESS);
 449         ASSERTCMP(object->n_in_progress, >, 0);
 450         ASSERTIFCMP(test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags),
 451                     object->n_exclusive, >, 0);
 452         ASSERTIFCMP(test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags),
 453                     object->n_in_progress, ==, 1);
 454
 455         spin_lock(&object->lock);
 456
 457         if (!cancelled) {
 458                 op->state = FSCACHE_OP_ST_COMPLETE;
 459         } else {
 460                 op->cancel(op);
 461                 op->state = FSCACHE_OP_ST_CANCELLED;
 462         }
 463
 464         if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 465                 object->n_exclusive--;
 466         object->n_in_progress--;
 467         if (object->n_in_progress == 0)
 468                 fscache_start_operations(object);
 469
 470         spin_unlock(&object->lock);
 471         _leave("");
 472 }
 473 EXPORT_SYMBOL(fscache_op_complete);
 474
 475 /*
 476  * release an operation
 477  * - queues pending ops if this is the last in-progress op
 478  */
 479 void fscache_put_operation(struct fscache_operation *op)
 480 {
 481         struct fscache_object *object;
 482         struct fscache_cache *cache;
 483
 484         _enter("{OBJ%x OP%x,%d}",
 485                op->object ? op->object->debug_id : 0,
 486                op->debug_id, atomic_read(&op->usage));
 487
 488         ASSERTCMP(atomic_read(&op->usage), >, 0);
 489
 490         if (!atomic_dec_and_test(&op->usage))
 491                 return;
 492
 493         _debug("PUT OP");
 494         ASSERTIFCMP(op->state != FSCACHE_OP_ST_INITIALISED &&
 495                     op->state != FSCACHE_OP_ST_COMPLETE,
 496                     op->state, ==, FSCACHE_OP_ST_CANCELLED);
 497
 498         fscache_stat(&fscache_n_op_release);
 499
 500         if (op->release) {
 501                 op->release(op);
 502                 op->release = NULL;
 503         }
 504         op->state = FSCACHE_OP_ST_DEAD;
 505
 506         object = op->object;
 507         if (likely(object)) {
 508                 if (test_bit(FSCACHE_OP_DEC_READ_CNT, &op->flags))
 509                         atomic_dec(&object->n_reads);
 510                 if (test_bit(FSCACHE_OP_UNUSE_COOKIE, &op->flags))
 511                         fscache_unuse_cookie(object);
 512
 513                 /* now... we may get called with the object spinlock held, so we
 514                  * complete the cleanup here only if we can immediately acquire the
 515                  * lock, and defer it otherwise */
 516                 if (!spin_trylock(&object->lock)) {
 517                         _debug("defer put");
 518                         fscache_stat(&fscache_n_op_deferred_release);
 519
 520                         cache = object->cache;
 521                         spin_lock(&cache->op_gc_list_lock);
 522                         list_add_tail(&op->pend_link, &cache->op_gc_list);
 523                         spin_unlock(&cache->op_gc_list_lock);
 524                         schedule_work(&cache->op_gc);
 525                         _leave(" [defer]");
 526                         return;
 527                 }
 528
 529                 ASSERTCMP(object->n_ops, >, 0);
 530                 object->n_ops--;
 531                 if (object->n_ops == 0)
 532                         fscache_raise_event(object, FSCACHE_OBJECT_EV_CLEARED);
 533
 534                 spin_unlock(&object->lock);
 535         }
 536
 537         kfree(op);
 538         _leave(" [done]");
 539 }
 540 EXPORT_SYMBOL(fscache_put_operation);
 541
 542 /*
 543  * garbage collect operations that have had their release deferred
 544  */
 545 void fscache_operation_gc(struct work_struct *work)
 546 {
 547         struct fscache_operation *op;
 548         struct fscache_object *object;
 549         struct fscache_cache *cache =
 550                 container_of(work, struct fscache_cache, op_gc);
 551         int count = 0;
 552
 553         _enter("");
 554
 555         do {
 556                 spin_lock(&cache->op_gc_list_lock);
 557                 if (list_empty(&cache->op_gc_list)) {
 558                         spin_unlock(&cache->op_gc_list_lock);
 559                         break;
 560                 }
 561
 562                 op = list_entry(cache->op_gc_list.next,
 563                                 struct fscache_operation, pend_link);
 564                 list_del(&op->pend_link);
 565                 spin_unlock(&cache->op_gc_list_lock);
 566
 567                 object = op->object;
 568                 spin_lock(&object->lock);
 569
 570                 _debug("GC DEFERRED REL OBJ%x OP%x",
 571                        object->debug_id, op->debug_id);
 572                 fscache_stat(&fscache_n_op_gc);
 573
 574                 ASSERTCMP(atomic_read(&op->usage), ==, 0);
 575                 ASSERTCMP(op->state, ==, FSCACHE_OP_ST_DEAD);
 576
 577                 ASSERTCMP(object->n_ops, >, 0);
 578                 object->n_ops--;
 579                 if (object->n_ops == 0)
 580                         fscache_raise_event(object, FSCACHE_OBJECT_EV_CLEARED);
 581
 582                 spin_unlock(&object->lock);
 583                 kfree(op);
 584
 585         } while (count++ < 20);
 586
 587         if (!list_empty(&cache->op_gc_list))
 588                 schedule_work(&cache->op_gc);
 589
 590         _leave("");
 591 }
 592
 593 /*
 594  * execute an operation using fs_op_wq to provide processing context -
 595  * the caller holds a ref to this object, so we don't need to hold one
 596  */
 597 void fscache_op_work_func(struct work_struct *work)
 598 {
 599         struct fscache_operation *op =
 600                 container_of(work, struct fscache_operation, work);
 601         unsigned long start;
 602
 603         _enter("{OBJ%x OP%x,%d}",
 604                op->object->debug_id, op->debug_id, atomic_read(&op->usage));
 605
 606         ASSERT(op->processor != NULL);
 607         start = jiffies;
 608         op->processor(op);
 609         fscache_hist(fscache_ops_histogram, start);
 610         fscache_put_operation(op);
 611
 612         _leave("");
 613 }