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