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