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/slab.h>
  10
  11 #include "blk.h"
  12
  13 /*
  14  * For io context allocations
  15  */
  16 static struct kmem_cache *iocontext_cachep;
  17
  18 /**
  19  * get_io_context - increment reference count to io_context
  20  * @ioc: io_context to get
  21  *
  22  * Increment reference count to @ioc.
  23  */
  24 void get_io_context(struct io_context *ioc)
  25 {
  26         BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
  27         atomic_long_inc(&ioc->refcount);
  28 }
  29 EXPORT_SYMBOL(get_io_context);
  30
  31 static void icq_free_icq_rcu(struct rcu_head *head)
  32 {
  33         struct io_cq *icq = container_of(head, struct io_cq, __rcu_head);
  34
  35         kmem_cache_free(icq->__rcu_icq_cache, icq);
  36 }
  37
  38 /*
  39  * Exit an icq. Called with both ioc and q locked for sq, only ioc locked for
  40  * mq.
  41  */
  42 static void ioc_exit_icq(struct io_cq *icq)
  43 {
  44         struct elevator_type *et = icq->q->elevator->type;
  45
  46         if (icq->flags & ICQ_EXITED)
  47                 return;
  48
  49         if (et->uses_mq && et->ops.mq.exit_icq)
  50                 et->ops.mq.exit_icq(icq);
  51         else if (!et->uses_mq && et->ops.sq.elevator_exit_icq_fn)
  52                 et->ops.sq.elevator_exit_icq_fn(icq);
  53
  54         icq->flags |= ICQ_EXITED;
  55 }
  56
  57 /* Release an icq.  Called with both ioc and q locked. */
  58 static void ioc_destroy_icq(struct io_cq *icq)
  59 {
  60         struct io_context *ioc = icq->ioc;
  61         struct request_queue *q = icq->q;
  62         struct elevator_type *et = q->elevator->type;
  63
  64         lockdep_assert_held(&ioc->lock);
  65         lockdep_assert_held(q->queue_lock);
  66
  67         radix_tree_delete(&ioc->icq_tree, icq->q->id);
  68         hlist_del_init(&icq->ioc_node);
  69         list_del_init(&icq->q_node);
  70
  71         /*
  72          * Both setting lookup hint to and clearing it from @icq are done
  73          * under queue_lock.  If it's not pointing to @icq now, it never
  74          * will.  Hint assignment itself can race safely.
  75          */
  76         if (rcu_access_pointer(ioc->icq_hint) == icq)
  77                 rcu_assign_pointer(ioc->icq_hint, NULL);
  78
  79         ioc_exit_icq(icq);
  80
  81         /*
  82          * @icq->q might have gone away by the time RCU callback runs
  83          * making it impossible to determine icq_cache.  Record it in @icq.
  84          */
  85         icq->__rcu_icq_cache = et->icq_cache;
  86         call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
  87 }
  88
  89 /*
  90  * Slow path for ioc release in put_io_context().  Performs double-lock
  91  * dancing to unlink all icq's and then frees ioc.
  92  */
  93 static void ioc_release_fn(struct work_struct *work)
  94 {
  95         struct io_context *ioc = container_of(work, struct io_context,
  96                                               release_work);
  97         unsigned long flags;
  98
  99         /*
 100          * Exiting icq may call into put_io_context() through elevator
 101          * which will trigger lockdep warning.  The ioc's are guaranteed to
 102          * be different, use a different locking subclass here.  Use
 103          * irqsave variant as there's no spin_lock_irq_nested().
 104          */
 105         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 106
 107         while (!hlist_empty(&ioc->icq_list)) {
 108                 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
 109                                                 struct io_cq, ioc_node);
 110                 struct request_queue *q = icq->q;
 111
 112                 if (spin_trylock(q->queue_lock)) {
 113                         ioc_destroy_icq(icq);
 114                         spin_unlock(q->queue_lock);
 115                 } else {
 116                         spin_unlock_irqrestore(&ioc->lock, flags);
 117                         cpu_relax();
 118                         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 119                 }
 120         }
 121
 122         spin_unlock_irqrestore(&ioc->lock, flags);
 123
 124         kmem_cache_free(iocontext_cachep, ioc);
 125 }
 126
 127 /**
 128  * put_io_context - put a reference of io_context
 129  * @ioc: io_context to put
 130  *
 131  * Decrement reference count of @ioc and release it if the count reaches
 132  * zero.
 133  */
 134 void put_io_context(struct io_context *ioc)
 135 {
 136         unsigned long flags;
 137         bool free_ioc = false;
 138
 139         if (ioc == NULL)
 140                 return;
 141
 142         BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
 143
 144         /*
 145          * Releasing ioc requires reverse order double locking and we may
 146          * already be holding a queue_lock.  Do it asynchronously from wq.
 147          */
 148         if (atomic_long_dec_and_test(&ioc->refcount)) {
 149                 spin_lock_irqsave(&ioc->lock, flags);
 150                 if (!hlist_empty(&ioc->icq_list))
 151                         queue_work(system_power_efficient_wq,
 152                                         &ioc->release_work);
 153                 else
 154                         free_ioc = true;
 155                 spin_unlock_irqrestore(&ioc->lock, flags);
 156         }
 157
 158         if (free_ioc)
 159                 kmem_cache_free(iocontext_cachep, ioc);
 160 }
 161 EXPORT_SYMBOL(put_io_context);
 162
 163 /**
 164  * put_io_context_active - put active reference on ioc
 165  * @ioc: ioc of interest
 166  *
 167  * Undo get_io_context_active().  If active reference reaches zero after
 168  * put, @ioc can never issue further IOs and ioscheds are notified.
 169  */
 170 void put_io_context_active(struct io_context *ioc)
 171 {
 172         struct elevator_type *et;
 173         unsigned long flags;
 174         struct io_cq *icq;
 175
 176         if (!atomic_dec_and_test(&ioc->active_ref)) {
 177                 put_io_context(ioc);
 178                 return;
 179         }
 180
 181         /*
 182          * Need ioc lock to walk icq_list and q lock to exit icq.  Perform
 183          * reverse double locking.  Read comment in ioc_release_fn() for
 184          * explanation on the nested locking annotation.
 185          */
 186 retry:
 187         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 188         hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) {
 189                 if (icq->flags & ICQ_EXITED)
 190                         continue;
 191
 192                 et = icq->q->elevator->type;
 193                 if (et->uses_mq) {
 194                         ioc_exit_icq(icq);
 195                 } else {
 196                         if (spin_trylock(icq->q->queue_lock)) {
 197                                 ioc_exit_icq(icq);
 198                                 spin_unlock(icq->q->queue_lock);
 199                         } else {
 200                                 spin_unlock_irqrestore(&ioc->lock, flags);
 201                                 cpu_relax();
 202                                 goto retry;
 203                         }
 204                 }
 205         }
 206         spin_unlock_irqrestore(&ioc->lock, flags);
 207
 208         put_io_context(ioc);
 209 }
 210
 211 /* Called by the exiting task */
 212 void exit_io_context(struct task_struct *task)
 213 {
 214         struct io_context *ioc;
 215
 216         task_lock(task);
 217         ioc = task->io_context;
 218         task->io_context = NULL;
 219         task_unlock(task);
 220
 221         atomic_dec(&ioc->nr_tasks);
 222         put_io_context_active(ioc);
 223 }
 224
 225 /**
 226  * ioc_clear_queue - break any ioc association with the specified queue
 227  * @q: request_queue being cleared
 228  *
 229  * Walk @q->icq_list and exit all io_cq's.  Must be called with @q locked.
 230  */
 231 void ioc_clear_queue(struct request_queue *q)
 232 {
 233         lockdep_assert_held(q->queue_lock);
 234
 235         while (!list_empty(&q->icq_list)) {
 236                 struct io_cq *icq = list_entry(q->icq_list.next,
 237                                                struct io_cq, q_node);
 238                 struct io_context *ioc = icq->ioc;
 239
 240                 spin_lock(&ioc->lock);
 241                 ioc_destroy_icq(icq);
 242                 spin_unlock(&ioc->lock);
 243         }
 244 }
 245
 246 int create_task_io_context(struct task_struct *task, gfp_t gfp_flags, int node)
 247 {
 248         struct io_context *ioc;
 249         int ret;
 250
 251         ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
 252                                     node);
 253         if (unlikely(!ioc))
 254                 return -ENOMEM;
 255
 256         /* initialize */
 257         atomic_long_set(&ioc->refcount, 1);
 258         atomic_set(&ioc->nr_tasks, 1);
 259         atomic_set(&ioc->active_ref, 1);
 260         spin_lock_init(&ioc->lock);
 261         INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
 262         INIT_HLIST_HEAD(&ioc->icq_list);
 263         INIT_WORK(&ioc->release_work, ioc_release_fn);
 264
 265         /*
 266          * Try to install.  ioc shouldn't be installed if someone else
 267          * already did or @task, which isn't %current, is exiting.  Note
 268          * that we need to allow ioc creation on exiting %current as exit
 269          * path may issue IOs from e.g. exit_files().  The exit path is
 270          * responsible for not issuing IO after exit_io_context().
 271          */
 272         task_lock(task);
 273         if (!task->io_context &&
 274             (task == current || !(task->flags & PF_EXITING)))
 275                 task->io_context = ioc;
 276         else
 277                 kmem_cache_free(iocontext_cachep, ioc);
 278
 279         ret = task->io_context ? 0 : -EBUSY;
 280
 281         task_unlock(task);
 282
 283         return ret;
 284 }
 285
 286 /**
 287  * get_task_io_context - get io_context of a task
 288  * @task: task of interest
 289  * @gfp_flags: allocation flags, used if allocation is necessary
 290  * @node: allocation node, used if allocation is necessary
 291  *
 292  * Return io_context of @task.  If it doesn't exist, it is created with
 293  * @gfp_flags and @node.  The returned io_context has its reference count
 294  * incremented.
 295  *
 296  * This function always goes through task_lock() and it's better to use
 297  * %current->io_context + get_io_context() for %current.
 298  */
 299 struct io_context *get_task_io_context(struct task_struct *task,
 300                                        gfp_t gfp_flags, int node)
 301 {
 302         struct io_context *ioc;
 303
 304         might_sleep_if(gfpflags_allow_blocking(gfp_flags));
 305
 306         do {
 307                 task_lock(task);
 308                 ioc = task->io_context;
 309                 if (likely(ioc)) {
 310                         get_io_context(ioc);
 311                         task_unlock(task);
 312                         return ioc;
 313                 }
 314                 task_unlock(task);
 315         } while (!create_task_io_context(task, gfp_flags, node));
 316
 317         return NULL;
 318 }
 319 EXPORT_SYMBOL(get_task_io_context);
 320
 321 /**
 322  * ioc_lookup_icq - lookup io_cq from ioc
 323  * @ioc: the associated io_context
 324  * @q: the associated request_queue
 325  *
 326  * Look up io_cq associated with @ioc - @q pair from @ioc.  Must be called
 327  * with @q->queue_lock held.
 328  */
 329 struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
 330 {
 331         struct io_cq *icq;
 332
 333         lockdep_assert_held(q->queue_lock);
 334
 335         /*
 336          * icq's are indexed from @ioc using radix tree and hint pointer,
 337          * both of which are protected with RCU.  All removals are done
 338          * holding both q and ioc locks, and we're holding q lock - if we
 339          * find a icq which points to us, it's guaranteed to be valid.
 340          */
 341         rcu_read_lock();
 342         icq = rcu_dereference(ioc->icq_hint);
 343         if (icq && icq->q == q)
 344                 goto out;
 345
 346         icq = radix_tree_lookup(&ioc->icq_tree, q->id);
 347         if (icq && icq->q == q)
 348                 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
 349         else
 350                 icq = NULL;
 351 out:
 352         rcu_read_unlock();
 353         return icq;
 354 }
 355 EXPORT_SYMBOL(ioc_lookup_icq);
 356
 357 /**
 358  * ioc_create_icq - create and link io_cq
 359  * @ioc: io_context of interest
 360  * @q: request_queue of interest
 361  * @gfp_mask: allocation mask
 362  *
 363  * Make sure io_cq linking @ioc and @q exists.  If icq doesn't exist, they
 364  * will be created using @gfp_mask.
 365  *
 366  * The caller is responsible for ensuring @ioc won't go away and @q is
 367  * alive and will stay alive until this function returns.
 368  */
 369 struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q,
 370                              gfp_t gfp_mask)
 371 {
 372         struct elevator_type *et = q->elevator->type;
 373         struct io_cq *icq;
 374
 375         /* allocate stuff */
 376         icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
 377                                     q->node);
 378         if (!icq)
 379                 return NULL;
 380
 381         if (radix_tree_maybe_preload(gfp_mask) < 0) {
 382                 kmem_cache_free(et->icq_cache, icq);
 383                 return NULL;
 384         }
 385
 386         icq->ioc = ioc;
 387         icq->q = q;
 388         INIT_LIST_HEAD(&icq->q_node);
 389         INIT_HLIST_NODE(&icq->ioc_node);
 390
 391         /* lock both q and ioc and try to link @icq */
 392         spin_lock_irq(q->queue_lock);
 393         spin_lock(&ioc->lock);
 394
 395         if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
 396                 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
 397                 list_add(&icq->q_node, &q->icq_list);
 398                 if (et->uses_mq && et->ops.mq.init_icq)
 399                         et->ops.mq.init_icq(icq);
 400                 else if (!et->uses_mq && et->ops.sq.elevator_init_icq_fn)
 401                         et->ops.sq.elevator_init_icq_fn(icq);
 402         } else {
 403                 kmem_cache_free(et->icq_cache, icq);
 404                 icq = ioc_lookup_icq(ioc, q);
 405                 if (!icq)
 406                         printk(KERN_ERR "cfq: icq link failed!\n");
 407         }
 408
 409         spin_unlock(&ioc->lock);
 410         spin_unlock_irq(q->queue_lock);
 411         radix_tree_preload_end();
 412         return icq;
 413 }
 414
 415 static int __init blk_ioc_init(void)
 416 {
 417         iocontext_cachep = kmem_cache_create("blkdev_ioc",
 418                         sizeof(struct io_context), 0, SLAB_PANIC, NULL);
 419         return 0;
 420 }
 421 subsys_initcall(blk_ioc_init);