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