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 ioc locked for blk-mq, and with both ioc
  40  * and queue locked for legacy.
  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 /*
  58  * Release an icq. Called with ioc locked for blk-mq, and with both ioc
  59  * and queue locked for legacy.
  60  */
  61 static void ioc_destroy_icq(struct io_cq *icq)
  62 {
  63         struct io_context *ioc = icq->ioc;
  64         struct request_queue *q = icq->q;
  65         struct elevator_type *et = q->elevator->type;
  66
  67         lockdep_assert_held(&ioc->lock);
  68
  69         radix_tree_delete(&ioc->icq_tree, icq->q->id);
  70         hlist_del_init(&icq->ioc_node);
  71         list_del_init(&icq->q_node);
  72
  73         /*
  74          * Both setting lookup hint to and clearing it from @icq are done
  75          * under queue_lock.  If it's not pointing to @icq now, it never
  76          * will.  Hint assignment itself can race safely.
  77          */
  78         if (rcu_access_pointer(ioc->icq_hint) == icq)
  79                 rcu_assign_pointer(ioc->icq_hint, NULL);
  80
  81         ioc_exit_icq(icq);
  82
  83         /*
  84          * @icq->q might have gone away by the time RCU callback runs
  85          * making it impossible to determine icq_cache.  Record it in @icq.
  86          */
  87         icq->__rcu_icq_cache = et->icq_cache;
  88         call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
  89 }
  90
  91 /*
  92  * Slow path for ioc release in put_io_context().  Performs double-lock
  93  * dancing to unlink all icq's and then frees ioc.
  94  */
  95 static void ioc_release_fn(struct work_struct *work)
  96 {
  97         struct io_context *ioc = container_of(work, struct io_context,
  98                                               release_work);
  99         unsigned long flags;
 100
 101         /*
 102          * Exiting icq may call into put_io_context() through elevator
 103          * which will trigger lockdep warning.  The ioc's are guaranteed to
 104          * be different, use a different locking subclass here.  Use
 105          * irqsave variant as there's no spin_lock_irq_nested().
 106          */
 107         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 108
 109         while (!hlist_empty(&ioc->icq_list)) {
 110                 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
 111                                                 struct io_cq, ioc_node);
 112                 struct request_queue *q = icq->q;
 113
 114                 if (spin_trylock(q->queue_lock)) {
 115                         ioc_destroy_icq(icq);
 116                         spin_unlock(q->queue_lock);
 117                 } else {
 118                         spin_unlock_irqrestore(&ioc->lock, flags);
 119                         cpu_relax();
 120                         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 121                 }
 122         }
 123
 124         spin_unlock_irqrestore(&ioc->lock, flags);
 125
 126         kmem_cache_free(iocontext_cachep, ioc);
 127 }
 128
 129 /**
 130  * put_io_context - put a reference of io_context
 131  * @ioc: io_context to put
 132  *
 133  * Decrement reference count of @ioc and release it if the count reaches
 134  * zero.
 135  */
 136 void put_io_context(struct io_context *ioc)
 137 {
 138         unsigned long flags;
 139         bool free_ioc = false;
 140
 141         if (ioc == NULL)
 142                 return;
 143
 144         BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
 145
 146         /*
 147          * Releasing ioc requires reverse order double locking and we may
 148          * already be holding a queue_lock.  Do it asynchronously from wq.
 149          */
 150         if (atomic_long_dec_and_test(&ioc->refcount)) {
 151                 spin_lock_irqsave(&ioc->lock, flags);
 152                 if (!hlist_empty(&ioc->icq_list))
 153                         queue_work(system_power_efficient_wq,
 154                                         &ioc->release_work);
 155                 else
 156                         free_ioc = true;
 157                 spin_unlock_irqrestore(&ioc->lock, flags);
 158         }
 159
 160         if (free_ioc)
 161                 kmem_cache_free(iocontext_cachep, ioc);
 162 }
 163 EXPORT_SYMBOL(put_io_context);
 164
 165 /**
 166  * put_io_context_active - put active reference on ioc
 167  * @ioc: ioc of interest
 168  *
 169  * Undo get_io_context_active().  If active reference reaches zero after
 170  * put, @ioc can never issue further IOs and ioscheds are notified.
 171  */
 172 void put_io_context_active(struct io_context *ioc)
 173 {
 174         struct elevator_type *et;
 175         unsigned long flags;
 176         struct io_cq *icq;
 177
 178         if (!atomic_dec_and_test(&ioc->active_ref)) {
 179                 put_io_context(ioc);
 180                 return;
 181         }
 182
 183         /*
 184          * Need ioc lock to walk icq_list and q lock to exit icq.  Perform
 185          * reverse double locking.  Read comment in ioc_release_fn() for
 186          * explanation on the nested locking annotation.
 187          */
 188 retry:
 189         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 190         hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) {
 191                 if (icq->flags & ICQ_EXITED)
 192                         continue;
 193
 194                 et = icq->q->elevator->type;
 195                 if (et->uses_mq) {
 196                         ioc_exit_icq(icq);
 197                 } else {
 198                         if (spin_trylock(icq->q->queue_lock)) {
 199                                 ioc_exit_icq(icq);
 200                                 spin_unlock(icq->q->queue_lock);
 201                         } else {
 202                                 spin_unlock_irqrestore(&ioc->lock, flags);
 203                                 cpu_relax();
 204                                 goto retry;
 205                         }
 206                 }
 207         }
 208         spin_unlock_irqrestore(&ioc->lock, flags);
 209
 210         put_io_context(ioc);
 211 }
 212
 213 /* Called by the exiting task */
 214 void exit_io_context(struct task_struct *task)
 215 {
 216         struct io_context *ioc;
 217
 218         task_lock(task);
 219         ioc = task->io_context;
 220         task->io_context = NULL;
 221         task_unlock(task);
 222
 223         atomic_dec(&ioc->nr_tasks);
 224         put_io_context_active(ioc);
 225 }
 226
 227 static void __ioc_clear_queue(struct list_head *icq_list)
 228 {
 229         unsigned long flags;
 230
 231         while (!list_empty(icq_list)) {
 232                 struct io_cq *icq = list_entry(icq_list->next,
 233                                                struct io_cq, q_node);
 234                 struct io_context *ioc = icq->ioc;
 235
 236                 spin_lock_irqsave(&ioc->lock, flags);
 237                 ioc_destroy_icq(icq);
 238                 spin_unlock_irqrestore(&ioc->lock, flags);
 239         }
 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.
 247  */
 248 void ioc_clear_queue(struct request_queue *q)
 249 {
 250         LIST_HEAD(icq_list);
 251
 252         spin_lock_irq(q->queue_lock);
 253         list_splice_init(&q->icq_list, &icq_list);
 254
 255         if (q->mq_ops) {
 256                 spin_unlock_irq(q->queue_lock);
 257                 __ioc_clear_queue(&icq_list);
 258         } else {
 259                 __ioc_clear_queue(&icq_list);
 260                 spin_unlock_irq(q->queue_lock);
 261         }
 262 }
 263
 264 int create_task_io_context(struct task_struct *task, gfp_t gfp_flags, int node)
 265 {
 266         struct io_context *ioc;
 267         int ret;
 268
 269         ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
 270                                     node);
 271         if (unlikely(!ioc))
 272                 return -ENOMEM;
 273
 274         /* initialize */
 275         atomic_long_set(&ioc->refcount, 1);
 276         atomic_set(&ioc->nr_tasks, 1);
 277         atomic_set(&ioc->active_ref, 1);
 278         spin_lock_init(&ioc->lock);
 279         INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
 280         INIT_HLIST_HEAD(&ioc->icq_list);
 281         INIT_WORK(&ioc->release_work, ioc_release_fn);
 282
 283         /*
 284          * Try to install.  ioc shouldn't be installed if someone else
 285          * already did or @task, which isn't %current, is exiting.  Note
 286          * that we need to allow ioc creation on exiting %current as exit
 287          * path may issue IOs from e.g. exit_files().  The exit path is
 288          * responsible for not issuing IO after exit_io_context().
 289          */
 290         task_lock(task);
 291         if (!task->io_context &&
 292             (task == current || !(task->flags & PF_EXITING)))
 293                 task->io_context = ioc;
 294         else
 295                 kmem_cache_free(iocontext_cachep, ioc);
 296
 297         ret = task->io_context ? 0 : -EBUSY;
 298
 299         task_unlock(task);
 300
 301         return ret;
 302 }
 303
 304 /**
 305  * get_task_io_context - get io_context of a task
 306  * @task: task of interest
 307  * @gfp_flags: allocation flags, used if allocation is necessary
 308  * @node: allocation node, used if allocation is necessary
 309  *
 310  * Return io_context of @task.  If it doesn't exist, it is created with
 311  * @gfp_flags and @node.  The returned io_context has its reference count
 312  * incremented.
 313  *
 314  * This function always goes through task_lock() and it's better to use
 315  * %current->io_context + get_io_context() for %current.
 316  */
 317 struct io_context *get_task_io_context(struct task_struct *task,
 318                                        gfp_t gfp_flags, int node)
 319 {
 320         struct io_context *ioc;
 321
 322         might_sleep_if(gfpflags_allow_blocking(gfp_flags));
 323
 324         do {
 325                 task_lock(task);
 326                 ioc = task->io_context;
 327                 if (likely(ioc)) {
 328                         get_io_context(ioc);
 329                         task_unlock(task);
 330                         return ioc;
 331                 }
 332                 task_unlock(task);
 333         } while (!create_task_io_context(task, gfp_flags, node));
 334
 335         return NULL;
 336 }
 337 EXPORT_SYMBOL(get_task_io_context);
 338
 339 /**
 340  * ioc_lookup_icq - lookup io_cq from ioc
 341  * @ioc: the associated io_context
 342  * @q: the associated request_queue
 343  *
 344  * Look up io_cq associated with @ioc - @q pair from @ioc.  Must be called
 345  * with @q->queue_lock held.
 346  */
 347 struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
 348 {
 349         struct io_cq *icq;
 350
 351         lockdep_assert_held(q->queue_lock);
 352
 353         /*
 354          * icq's are indexed from @ioc using radix tree and hint pointer,
 355          * both of which are protected with RCU.  All removals are done
 356          * holding both q and ioc locks, and we're holding q lock - if we
 357          * find a icq which points to us, it's guaranteed to be valid.
 358          */
 359         rcu_read_lock();
 360         icq = rcu_dereference(ioc->icq_hint);
 361         if (icq && icq->q == q)
 362                 goto out;
 363
 364         icq = radix_tree_lookup(&ioc->icq_tree, q->id);
 365         if (icq && icq->q == q)
 366                 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
 367         else
 368                 icq = NULL;
 369 out:
 370         rcu_read_unlock();
 371         return icq;
 372 }
 373 EXPORT_SYMBOL(ioc_lookup_icq);
 374
 375 /**
 376  * ioc_create_icq - create and link io_cq
 377  * @ioc: io_context of interest
 378  * @q: request_queue of interest
 379  * @gfp_mask: allocation mask
 380  *
 381  * Make sure io_cq linking @ioc and @q exists.  If icq doesn't exist, they
 382  * will be created using @gfp_mask.
 383  *
 384  * The caller is responsible for ensuring @ioc won't go away and @q is
 385  * alive and will stay alive until this function returns.
 386  */
 387 struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q,
 388                              gfp_t gfp_mask)
 389 {
 390         struct elevator_type *et = q->elevator->type;
 391         struct io_cq *icq;
 392
 393         /* allocate stuff */
 394         icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
 395                                     q->node);
 396         if (!icq)
 397                 return NULL;
 398
 399         if (radix_tree_maybe_preload(gfp_mask) < 0) {
 400                 kmem_cache_free(et->icq_cache, icq);
 401                 return NULL;
 402         }
 403
 404         icq->ioc = ioc;
 405         icq->q = q;
 406         INIT_LIST_HEAD(&icq->q_node);
 407         INIT_HLIST_NODE(&icq->ioc_node);
 408
 409         /* lock both q and ioc and try to link @icq */
 410         spin_lock_irq(q->queue_lock);
 411         spin_lock(&ioc->lock);
 412
 413         if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
 414                 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
 415                 list_add(&icq->q_node, &q->icq_list);
 416                 if (et->uses_mq && et->ops.mq.init_icq)
 417                         et->ops.mq.init_icq(icq);
 418                 else if (!et->uses_mq && et->ops.sq.elevator_init_icq_fn)
 419                         et->ops.sq.elevator_init_icq_fn(icq);
 420         } else {
 421                 kmem_cache_free(et->icq_cache, icq);
 422                 icq = ioc_lookup_icq(ioc, q);
 423                 if (!icq)
 424                         printk(KERN_ERR "cfq: icq link failed!\n");
 425         }
 426
 427         spin_unlock(&ioc->lock);
 428         spin_unlock_irq(q->queue_lock);
 429         radix_tree_preload_end();
 430         return icq;
 431 }
 432
 433 static int __init blk_ioc_init(void)
 434 {
 435         iocontext_cachep = kmem_cache_create("blkdev_ioc",
 436                         sizeof(struct io_context), 0, SLAB_PANIC, NULL);
 437         return 0;
 438 }
 439 subsys_initcall(blk_ioc_init);