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                       bool cancel_in_progress_op)
 317 {
 318         struct fscache_object *object = op->object;
 319         bool put = false;
 320         int ret;
 321
 322         _enter("OBJ%x OP%x}", op->object->debug_id, op->debug_id);
 323
 324         ASSERTCMP(op->state, >=, FSCACHE_OP_ST_PENDING);
 325         ASSERTCMP(op->state, !=, FSCACHE_OP_ST_CANCELLED);
 326         ASSERTCMP(atomic_read(&op->usage), >, 0);
 327
 328         spin_lock(&object->lock);
 329
 330         ret = -EBUSY;
 331         if (op->state == FSCACHE_OP_ST_PENDING) {
 332                 ASSERT(!list_empty(&op->pend_link));
 333                 list_del_init(&op->pend_link);
 334                 put = true;
 335                 fscache_stat(&fscache_n_op_cancelled);
 336                 if (do_cancel)
 337                         do_cancel(op);
 338                 op->state = FSCACHE_OP_ST_CANCELLED;
 339                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 340                         object->n_exclusive--;
 341                 if (test_and_clear_bit(FSCACHE_OP_WAITING, &op->flags))
 342                         wake_up_bit(&op->flags, FSCACHE_OP_WAITING);
 343                 ret = 0;
 344         } else if (op->state == FSCACHE_OP_ST_IN_PROGRESS && cancel_in_progress_op) {
 345                 fscache_stat(&fscache_n_op_cancelled);
 346                 if (do_cancel)
 347                         do_cancel(op);
 348                 op->state = FSCACHE_OP_ST_CANCELLED;
 349                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 350                         object->n_exclusive--;
 351                 if (test_and_clear_bit(FSCACHE_OP_WAITING, &op->flags))
 352                         wake_up_bit(&op->flags, FSCACHE_OP_WAITING);
 353                 ret = 0;
 354         }
 355
 356         if (put)
 357                 fscache_put_operation(op);
 358         spin_unlock(&object->lock);
 359         _leave(" = %d", ret);
 360         return ret;
 361 }
 362
 363 /*
 364  * Cancel all pending operations on an object
 365  */
 366 void fscache_cancel_all_ops(struct fscache_object *object)
 367 {
 368         struct fscache_operation *op;
 369
 370         _enter("OBJ%x", object->debug_id);
 371
 372         spin_lock(&object->lock);
 373
 374         while (!list_empty(&object->pending_ops)) {
 375                 op = list_entry(object->pending_ops.next,
 376                                 struct fscache_operation, pend_link);
 377                 fscache_stat(&fscache_n_op_cancelled);
 378                 list_del_init(&op->pend_link);
 379
 380                 ASSERTCMP(op->state, ==, FSCACHE_OP_ST_PENDING);
 381                 op->state = FSCACHE_OP_ST_CANCELLED;
 382
 383                 if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 384                         object->n_exclusive--;
 385                 if (test_and_clear_bit(FSCACHE_OP_WAITING, &op->flags))
 386                         wake_up_bit(&op->flags, FSCACHE_OP_WAITING);
 387                 fscache_put_operation(op);
 388                 cond_resched_lock(&object->lock);
 389         }
 390
 391         spin_unlock(&object->lock);
 392         _leave("");
 393 }
 394
 395 /*
 396  * Record the completion or cancellation of an in-progress operation.
 397  */
 398 void fscache_op_complete(struct fscache_operation *op, bool cancelled)
 399 {
 400         struct fscache_object *object = op->object;
 401
 402         _enter("OBJ%x", object->debug_id);
 403
 404         ASSERTCMP(op->state, ==, FSCACHE_OP_ST_IN_PROGRESS);
 405         ASSERTCMP(object->n_in_progress, >, 0);
 406         ASSERTIFCMP(test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags),
 407                     object->n_exclusive, >, 0);
 408         ASSERTIFCMP(test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags),
 409                     object->n_in_progress, ==, 1);
 410
 411         spin_lock(&object->lock);
 412
 413         op->state = cancelled ?
 414                 FSCACHE_OP_ST_CANCELLED : FSCACHE_OP_ST_COMPLETE;
 415
 416         if (test_bit(FSCACHE_OP_EXCLUSIVE, &op->flags))
 417                 object->n_exclusive--;
 418         object->n_in_progress--;
 419         if (object->n_in_progress == 0)
 420                 fscache_start_operations(object);
 421
 422         spin_unlock(&object->lock);
 423         _leave("");
 424 }
 425 EXPORT_SYMBOL(fscache_op_complete);
 426
 427 /*
 428  * release an operation
 429  * - queues pending ops if this is the last in-progress op
 430  */
 431 void fscache_put_operation(struct fscache_operation *op)
 432 {
 433         struct fscache_object *object;
 434         struct fscache_cache *cache;
 435
 436         _enter("{OBJ%x OP%x,%d}",
 437                op->object->debug_id, op->debug_id, atomic_read(&op->usage));
 438
 439         ASSERTCMP(atomic_read(&op->usage), >, 0);
 440
 441         if (!atomic_dec_and_test(&op->usage))
 442                 return;
 443
 444         _debug("PUT OP");
 445         ASSERTIFCMP(op->state != FSCACHE_OP_ST_COMPLETE,
 446                     op->state, ==, FSCACHE_OP_ST_CANCELLED);
 447         op->state = FSCACHE_OP_ST_DEAD;
 448
 449         fscache_stat(&fscache_n_op_release);
 450
 451         if (op->release) {
 452                 op->release(op);
 453                 op->release = NULL;
 454         }
 455
 456         object = op->object;
 457
 458         if (test_bit(FSCACHE_OP_DEC_READ_CNT, &op->flags))
 459                 atomic_dec(&object->n_reads);
 460         if (test_bit(FSCACHE_OP_UNUSE_COOKIE, &op->flags))
 461                 fscache_unuse_cookie(object);
 462
 463         /* now... we may get called with the object spinlock held, so we
 464          * complete the cleanup here only if we can immediately acquire the
 465          * lock, and defer it otherwise */
 466         if (!spin_trylock(&object->lock)) {
 467                 _debug("defer put");
 468                 fscache_stat(&fscache_n_op_deferred_release);
 469
 470                 cache = object->cache;
 471                 spin_lock(&cache->op_gc_list_lock);
 472                 list_add_tail(&op->pend_link, &cache->op_gc_list);
 473                 spin_unlock(&cache->op_gc_list_lock);
 474                 schedule_work(&cache->op_gc);
 475                 _leave(" [defer]");
 476                 return;
 477         }
 478
 479         ASSERTCMP(object->n_ops, >, 0);
 480         object->n_ops--;
 481         if (object->n_ops == 0)
 482                 fscache_raise_event(object, FSCACHE_OBJECT_EV_CLEARED);
 483
 484         spin_unlock(&object->lock);
 485
 486         kfree(op);
 487         _leave(" [done]");
 488 }
 489 EXPORT_SYMBOL(fscache_put_operation);
 490
 491 /*
 492  * garbage collect operations that have had their release deferred
 493  */
 494 void fscache_operation_gc(struct work_struct *work)
 495 {
 496         struct fscache_operation *op;
 497         struct fscache_object *object;
 498         struct fscache_cache *cache =
 499                 container_of(work, struct fscache_cache, op_gc);
 500         int count = 0;
 501
 502         _enter("");
 503
 504         do {
 505                 spin_lock(&cache->op_gc_list_lock);
 506                 if (list_empty(&cache->op_gc_list)) {
 507                         spin_unlock(&cache->op_gc_list_lock);
 508                         break;
 509                 }
 510
 511                 op = list_entry(cache->op_gc_list.next,
 512                                 struct fscache_operation, pend_link);
 513                 list_del(&op->pend_link);
 514                 spin_unlock(&cache->op_gc_list_lock);
 515
 516                 object = op->object;
 517                 spin_lock(&object->lock);
 518
 519                 _debug("GC DEFERRED REL OBJ%x OP%x",
 520                        object->debug_id, op->debug_id);
 521                 fscache_stat(&fscache_n_op_gc);
 522
 523                 ASSERTCMP(atomic_read(&op->usage), ==, 0);
 524                 ASSERTCMP(op->state, ==, FSCACHE_OP_ST_DEAD);
 525
 526                 ASSERTCMP(object->n_ops, >, 0);
 527                 object->n_ops--;
 528                 if (object->n_ops == 0)
 529                         fscache_raise_event(object, FSCACHE_OBJECT_EV_CLEARED);
 530
 531                 spin_unlock(&object->lock);
 532                 kfree(op);
 533
 534         } while (count++ < 20);
 535
 536         if (!list_empty(&cache->op_gc_list))
 537                 schedule_work(&cache->op_gc);
 538
 539         _leave("");
 540 }
 541
 542 /*
 543  * execute an operation using fs_op_wq to provide processing context -
 544  * the caller holds a ref to this object, so we don't need to hold one
 545  */
 546 void fscache_op_work_func(struct work_struct *work)
 547 {
 548         struct fscache_operation *op =
 549                 container_of(work, struct fscache_operation, work);
 550         unsigned long start;
 551
 552         _enter("{OBJ%x OP%x,%d}",
 553                op->object->debug_id, op->debug_id, atomic_read(&op->usage));
 554
 555         ASSERT(op->processor != NULL);
 556         start = jiffies;
 557         op->processor(op);
 558         fscache_hist(fscache_ops_histogram, start);
 559         fscache_put_operation(op);
 560
 561         _leave("");
 562 }