block/blk-ioc.c

   1 /*
   2  * Functions related to io context handling
   3  */
   4 #include <linux/kernel.h>
   5 #include <linux/module.h>
   6 #include <linux/init.h>
   7 #include <linux/bio.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/bootmem.h>      /* for max_pfn/max_low_pfn */
  10 #include <linux/slab.h>
  11
  12 #include "blk.h"
  13
  14 /*
  15  * For io context allocations
  16  */
  17 static struct kmem_cache *iocontext_cachep;
  18
  19 /**
  20  * get_io_context - increment reference count to io_context
  21  * @ioc: io_context to get
  22  *
  23  * Increment reference count to @ioc.
  24  */
  25 void get_io_context(struct io_context *ioc)
  26 {
  27         BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
  28         atomic_long_inc(&ioc->refcount);
  29 }
  30 EXPORT_SYMBOL(get_io_context);
  31
  32 /*
  33  * Releasing ioc may nest into another put_io_context() leading to nested
  34  * fast path release.  As the ioc's can't be the same, this is okay but
  35  * makes lockdep whine.  Keep track of nesting and use it as subclass.
  36  */
  37 #ifdef CONFIG_LOCKDEP
  38 #define ioc_release_depth(q)            ((q) ? (q)->ioc_release_depth : 0)
  39 #define ioc_release_depth_inc(q)        (q)->ioc_release_depth++
  40 #define ioc_release_depth_dec(q)        (q)->ioc_release_depth--
  41 #else
  42 #define ioc_release_depth(q)            0
  43 #define ioc_release_depth_inc(q)        do { } while (0)
  44 #define ioc_release_depth_dec(q)        do { } while (0)
  45 #endif
  46
  47 static void icq_free_icq_rcu(struct rcu_head *head)
  48 {
  49         struct io_cq *icq = container_of(head, struct io_cq, __rcu_head);
  50
  51         kmem_cache_free(icq->__rcu_icq_cache, icq);
  52 }
  53
  54 /*
  55  * Exit and free an icq.  Called with both ioc and q locked.
  56  */
  57 static void ioc_exit_icq(struct io_cq *icq)
  58 {
  59         struct io_context *ioc = icq->ioc;
  60         struct request_queue *q = icq->q;
  61         struct elevator_type *et = q->elevator->type;
  62
  63         lockdep_assert_held(&ioc->lock);
  64         lockdep_assert_held(q->queue_lock);
  65
  66         radix_tree_delete(&ioc->icq_tree, icq->q->id);
  67         hlist_del_init(&icq->ioc_node);
  68         list_del_init(&icq->q_node);
  69
  70         /*
  71          * Both setting lookup hint to and clearing it from @icq are done
  72          * under queue_lock.  If it's not pointing to @icq now, it never
  73          * will.  Hint assignment itself can race safely.
  74          */
  75         if (rcu_dereference_raw(ioc->icq_hint) == icq)
  76                 rcu_assign_pointer(ioc->icq_hint, NULL);
  77
  78         if (et->ops.elevator_exit_icq_fn) {
  79                 ioc_release_depth_inc(q);
  80                 et->ops.elevator_exit_icq_fn(icq);
  81                 ioc_release_depth_dec(q);
  82         }
  83
  84         /*
  85          * @icq->q might have gone away by the time RCU callback runs
  86          * making it impossible to determine icq_cache.  Record it in @icq.
  87          */
  88         icq->__rcu_icq_cache = et->icq_cache;
  89         call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
  90 }
  91
  92 /*
  93  * Slow path for ioc release in put_io_context().  Performs double-lock
  94  * dancing to unlink all icq's and then frees ioc.
  95  */
  96 static void ioc_release_fn(struct work_struct *work)
  97 {
  98         struct io_context *ioc = container_of(work, struct io_context,
  99                                               release_work);
 100         struct request_queue *last_q = NULL;
 101
 102         spin_lock_irq(&ioc->lock);
 103
 104         while (!hlist_empty(&ioc->icq_list)) {
 105                 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
 106                                                 struct io_cq, ioc_node);
 107                 struct request_queue *this_q = icq->q;
 108
 109                 if (this_q != last_q) {
 110                         /*
 111                          * Need to switch to @this_q.  Once we release
 112                          * @ioc->lock, it can go away along with @cic.
 113                          * Hold on to it.
 114                          */
 115                         __blk_get_queue(this_q);
 116
 117                         /*
 118                          * blk_put_queue() might sleep thanks to kobject
 119                          * idiocy.  Always release both locks, put and
 120                          * restart.
 121                          */
 122                         if (last_q) {
 123                                 spin_unlock(last_q->queue_lock);
 124                                 spin_unlock_irq(&ioc->lock);
 125                                 blk_put_queue(last_q);
 126                         } else {
 127                                 spin_unlock_irq(&ioc->lock);
 128                         }
 129
 130                         last_q = this_q;
 131                         spin_lock_irq(this_q->queue_lock);
 132                         spin_lock(&ioc->lock);
 133                         continue;
 134                 }
 135                 ioc_exit_icq(icq);
 136         }
 137
 138         if (last_q) {
 139                 spin_unlock(last_q->queue_lock);
 140                 spin_unlock_irq(&ioc->lock);
 141                 blk_put_queue(last_q);
 142         } else {
 143                 spin_unlock_irq(&ioc->lock);
 144         }
 145
 146         kmem_cache_free(iocontext_cachep, ioc);
 147 }
 148
 149 /**
 150  * put_io_context - put a reference of io_context
 151  * @ioc: io_context to put
 152  * @locked_q: request_queue the caller is holding queue_lock of (hint)
 153  *
 154  * Decrement reference count of @ioc and release it if the count reaches
 155  * zero.  If the caller is holding queue_lock of a queue, it can indicate
 156  * that with @locked_q.  This is an optimization hint and the caller is
 157  * allowed to pass in %NULL even when it's holding a queue_lock.
 158  */
 159 void put_io_context(struct io_context *ioc, struct request_queue *locked_q)
 160 {
 161         struct request_queue *last_q = locked_q;
 162         unsigned long flags;
 163
 164         if (ioc == NULL)
 165                 return;
 166
 167         BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
 168         if (locked_q)
 169                 lockdep_assert_held(locked_q->queue_lock);
 170
 171         if (!atomic_long_dec_and_test(&ioc->refcount))
 172                 return;
 173
 174         /*
 175          * Destroy @ioc.  This is a bit messy because icq's are chained
 176          * from both ioc and queue, and ioc->lock nests inside queue_lock.
 177          * The inner ioc->lock should be held to walk our icq_list and then
 178          * for each icq the outer matching queue_lock should be grabbed.
 179          * ie. We need to do reverse-order double lock dancing.
 180          *
 181          * Another twist is that we are often called with one of the
 182          * matching queue_locks held as indicated by @locked_q, which
 183          * prevents performing double-lock dance for other queues.
 184          *
 185          * So, we do it in two stages.  The fast path uses the queue_lock
 186          * the caller is holding and, if other queues need to be accessed,
 187          * uses trylock to avoid introducing locking dependency.  This can
 188          * handle most cases, especially if @ioc was performing IO on only
 189          * single device.
 190          *
 191          * If trylock doesn't cut it, we defer to @ioc->release_work which
 192          * can do all the double-locking dancing.
 193          */
 194         spin_lock_irqsave_nested(&ioc->lock, flags,
 195                                  ioc_release_depth(locked_q));
 196
 197         while (!hlist_empty(&ioc->icq_list)) {
 198                 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
 199                                                 struct io_cq, ioc_node);
 200                 struct request_queue *this_q = icq->q;
 201
 202                 if (this_q != last_q) {
 203                         if (last_q && last_q != locked_q)
 204                                 spin_unlock(last_q->queue_lock);
 205                         last_q = NULL;
 206
 207                         if (!spin_trylock(this_q->queue_lock))
 208                                 break;
 209                         last_q = this_q;
 210                         continue;
 211                 }
 212                 ioc_exit_icq(icq);
 213         }
 214
 215         if (last_q && last_q != locked_q)
 216                 spin_unlock(last_q->queue_lock);
 217
 218         spin_unlock_irqrestore(&ioc->lock, flags);
 219
 220         /* if no icq is left, we're done; otherwise, kick release_work */
 221         if (hlist_empty(&ioc->icq_list))
 222                 kmem_cache_free(iocontext_cachep, ioc);
 223         else
 224                 schedule_work(&ioc->release_work);
 225 }
 226 EXPORT_SYMBOL(put_io_context);
 227
 228 /* Called by the exiting task */
 229 void exit_io_context(struct task_struct *task)
 230 {
 231         struct io_context *ioc;
 232
 233         task_lock(task);
 234         ioc = task->io_context;
 235         task->io_context = NULL;
 236         task_unlock(task);
 237
 238         atomic_dec(&ioc->nr_tasks);
 239         put_io_context(ioc, NULL);
 240 }
 241
 242 /**
 243  * ioc_clear_queue - break any ioc association with the specified queue
 244  * @q: request_queue being cleared
 245  *
 246  * Walk @q->icq_list and exit all io_cq's.  Must be called with @q locked.
 247  */
 248 void ioc_clear_queue(struct request_queue *q)
 249 {
 250         lockdep_assert_held(q->queue_lock);
 251
 252         while (!list_empty(&q->icq_list)) {
 253                 struct io_cq *icq = list_entry(q->icq_list.next,
 254                                                struct io_cq, q_node);
 255                 struct io_context *ioc = icq->ioc;
 256
 257                 spin_lock(&ioc->lock);
 258                 ioc_exit_icq(icq);
 259                 spin_unlock(&ioc->lock);
 260         }
 261 }
 262
 263 void create_io_context_slowpath(struct task_struct *task, gfp_t gfp_flags,
 264                                 int node)
 265 {
 266         struct io_context *ioc;
 267
 268         ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
 269                                     node);
 270         if (unlikely(!ioc))
 271                 return;
 272
 273         /* initialize */
 274         atomic_long_set(&ioc->refcount, 1);
 275         atomic_set(&ioc->nr_tasks, 1);
 276         spin_lock_init(&ioc->lock);
 277         INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
 278         INIT_HLIST_HEAD(&ioc->icq_list);
 279         INIT_WORK(&ioc->release_work, ioc_release_fn);
 280
 281         /*
 282          * Try to install.  ioc shouldn't be installed if someone else
 283          * already did or @task, which isn't %current, is exiting.  Note
 284          * that we need to allow ioc creation on exiting %current as exit
 285          * path may issue IOs from e.g. exit_files().  The exit path is
 286          * responsible for not issuing IO after exit_io_context().
 287          */
 288         task_lock(task);
 289         if (!task->io_context &&
 290             (task == current || !(task->flags & PF_EXITING)))
 291                 task->io_context = ioc;
 292         else
 293                 kmem_cache_free(iocontext_cachep, ioc);
 294         task_unlock(task);
 295 }
 296
 297 /**
 298  * get_task_io_context - get io_context of a task
 299  * @task: task of interest
 300  * @gfp_flags: allocation flags, used if allocation is necessary
 301  * @node: allocation node, used if allocation is necessary
 302  *
 303  * Return io_context of @task.  If it doesn't exist, it is created with
 304  * @gfp_flags and @node.  The returned io_context has its reference count
 305  * incremented.
 306  *
 307  * This function always goes through task_lock() and it's better to use
 308  * %current->io_context + get_io_context() for %current.
 309  */
 310 struct io_context *get_task_io_context(struct task_struct *task,
 311                                        gfp_t gfp_flags, int node)
 312 {
 313         struct io_context *ioc;
 314
 315         might_sleep_if(gfp_flags & __GFP_WAIT);
 316
 317         do {
 318                 task_lock(task);
 319                 ioc = task->io_context;
 320                 if (likely(ioc)) {
 321                         get_io_context(ioc);
 322                         task_unlock(task);
 323                         return ioc;
 324                 }
 325                 task_unlock(task);
 326         } while (create_io_context(task, gfp_flags, node));
 327
 328         return NULL;
 329 }
 330 EXPORT_SYMBOL(get_task_io_context);
 331
 332 /**
 333  * ioc_lookup_icq - lookup io_cq from ioc
 334  * @ioc: the associated io_context
 335  * @q: the associated request_queue
 336  *
 337  * Look up io_cq associated with @ioc - @q pair from @ioc.  Must be called
 338  * with @q->queue_lock held.
 339  */
 340 struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
 341 {
 342         struct io_cq *icq;
 343
 344         lockdep_assert_held(q->queue_lock);
 345
 346         /*
 347          * icq's are indexed from @ioc using radix tree and hint pointer,
 348          * both of which are protected with RCU.  All removals are done
 349          * holding both q and ioc locks, and we're holding q lock - if we
 350          * find a icq which points to us, it's guaranteed to be valid.
 351          */
 352         rcu_read_lock();
 353         icq = rcu_dereference(ioc->icq_hint);
 354         if (icq && icq->q == q)
 355                 goto out;
 356
 357         icq = radix_tree_lookup(&ioc->icq_tree, q->id);
 358         if (icq && icq->q == q)
 359                 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
 360         else
 361                 icq = NULL;
 362 out:
 363         rcu_read_unlock();
 364         return icq;
 365 }
 366 EXPORT_SYMBOL(ioc_lookup_icq);
 367
 368 /**
 369  * ioc_create_icq - create and link io_cq
 370  * @q: request_queue of interest
 371  * @gfp_mask: allocation mask
 372  *
 373  * Make sure io_cq linking %current->io_context and @q exists.  If either
 374  * io_context and/or icq don't exist, they will be created using @gfp_mask.
 375  *
 376  * The caller is responsible for ensuring @ioc won't go away and @q is
 377  * alive and will stay alive until this function returns.
 378  */
 379 struct io_cq *ioc_create_icq(struct request_queue *q, gfp_t gfp_mask)
 380 {
 381         struct elevator_type *et = q->elevator->type;
 382         struct io_context *ioc;
 383         struct io_cq *icq;
 384
 385         /* allocate stuff */
 386         ioc = create_io_context(current, gfp_mask, q->node);
 387         if (!ioc)
 388                 return NULL;
 389
 390         icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
 391                                     q->node);
 392         if (!icq)
 393                 return NULL;
 394
 395         if (radix_tree_preload(gfp_mask) < 0) {
 396                 kmem_cache_free(et->icq_cache, icq);
 397                 return NULL;
 398         }
 399
 400         icq->ioc = ioc;
 401         icq->q = q;
 402         INIT_LIST_HEAD(&icq->q_node);
 403         INIT_HLIST_NODE(&icq->ioc_node);
 404
 405         /* lock both q and ioc and try to link @icq */
 406         spin_lock_irq(q->queue_lock);
 407         spin_lock(&ioc->lock);
 408
 409         if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
 410                 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
 411                 list_add(&icq->q_node, &q->icq_list);
 412                 if (et->ops.elevator_init_icq_fn)
 413                         et->ops.elevator_init_icq_fn(icq);
 414         } else {
 415                 kmem_cache_free(et->icq_cache, icq);
 416                 icq = ioc_lookup_icq(ioc, q);
 417                 if (!icq)
 418                         printk(KERN_ERR "cfq: icq link failed!\n");
 419         }
 420
 421         spin_unlock(&ioc->lock);
 422         spin_unlock_irq(q->queue_lock);
 423         radix_tree_preload_end();
 424         return icq;
 425 }
 426
 427 void ioc_set_changed(struct io_context *ioc, int which)
 428 {
 429         struct io_cq *icq;
 430         struct hlist_node *n;
 431
 432         hlist_for_each_entry(icq, n, &ioc->icq_list, ioc_node)
 433                 set_bit(which, &icq->changed);
 434 }
 435
 436 /**
 437  * ioc_ioprio_changed - notify ioprio change
 438  * @ioc: io_context of interest
 439  * @ioprio: new ioprio
 440  *
 441  * @ioc's ioprio has changed to @ioprio.  Set %ICQ_IOPRIO_CHANGED for all
 442  * icq's.  iosched is responsible for checking the bit and applying it on
 443  * request issue path.
 444  */
 445 void ioc_ioprio_changed(struct io_context *ioc, int ioprio)
 446 {
 447         unsigned long flags;
 448
 449         spin_lock_irqsave(&ioc->lock, flags);
 450         ioc->ioprio = ioprio;
 451         ioc_set_changed(ioc, ICQ_IOPRIO_CHANGED);
 452         spin_unlock_irqrestore(&ioc->lock, flags);
 453 }
 454
 455 /**
 456  * ioc_cgroup_changed - notify cgroup change
 457  * @ioc: io_context of interest
 458  *
 459  * @ioc's cgroup has changed.  Set %ICQ_CGROUP_CHANGED for all icq's.
 460  * iosched is responsible for checking the bit and applying it on request
 461  * issue path.
 462  */
 463 void ioc_cgroup_changed(struct io_context *ioc)
 464 {
 465         unsigned long flags;
 466
 467         spin_lock_irqsave(&ioc->lock, flags);
 468         ioc_set_changed(ioc, ICQ_CGROUP_CHANGED);
 469         spin_unlock_irqrestore(&ioc->lock, flags);
 470 }
 471 EXPORT_SYMBOL(ioc_cgroup_changed);
 472
 473 static int __init blk_ioc_init(void)
 474 {
 475         iocontext_cachep = kmem_cache_create("blkdev_ioc",
 476                         sizeof(struct io_context), 0, SLAB_PANIC, NULL);
 477         return 0;
 478 }
 479 subsys_initcall(blk_ioc_init);