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
 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                         schedule_work(&ioc->release_work);
 147                 spin_unlock_irqrestore(&ioc->lock, flags);
 148         }
 149 }
 150 EXPORT_SYMBOL(put_io_context);
 151
 152 /* Called by the exiting task */
 153 void exit_io_context(struct task_struct *task)
 154 {
 155         struct io_context *ioc;
 156         struct io_cq *icq;
 157         struct hlist_node *n;
 158         unsigned long flags;
 159
 160         task_lock(task);
 161         ioc = task->io_context;
 162         task->io_context = NULL;
 163         task_unlock(task);
 164
 165         if (!atomic_dec_and_test(&ioc->nr_tasks)) {
 166                 put_io_context(ioc);
 167                 return;
 168         }
 169
 170         /*
 171          * Need ioc lock to walk icq_list and q lock to exit icq.  Perform
 172          * reverse double locking.  Read comment in ioc_release_fn() for
 173          * explanation on the nested locking annotation.
 174          */
 175 retry:
 176         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 177         hlist_for_each_entry(icq, n, &ioc->icq_list, ioc_node) {
 178                 if (icq->flags & ICQ_EXITED)
 179                         continue;
 180                 if (spin_trylock(icq->q->queue_lock)) {
 181                         ioc_exit_icq(icq);
 182                         spin_unlock(icq->q->queue_lock);
 183                 } else {
 184                         spin_unlock_irqrestore(&ioc->lock, flags);
 185                         cpu_relax();
 186                         goto retry;
 187                 }
 188         }
 189         spin_unlock_irqrestore(&ioc->lock, flags);
 190
 191         put_io_context(ioc);
 192 }
 193
 194 /**
 195  * ioc_clear_queue - break any ioc association with the specified queue
 196  * @q: request_queue being cleared
 197  *
 198  * Walk @q->icq_list and exit all io_cq's.  Must be called with @q locked.
 199  */
 200 void ioc_clear_queue(struct request_queue *q)
 201 {
 202         lockdep_assert_held(q->queue_lock);
 203
 204         while (!list_empty(&q->icq_list)) {
 205                 struct io_cq *icq = list_entry(q->icq_list.next,
 206                                                struct io_cq, q_node);
 207                 struct io_context *ioc = icq->ioc;
 208
 209                 spin_lock(&ioc->lock);
 210                 ioc_destroy_icq(icq);
 211                 spin_unlock(&ioc->lock);
 212         }
 213 }
 214
 215 void create_io_context_slowpath(struct task_struct *task, gfp_t gfp_flags,
 216                                 int node)
 217 {
 218         struct io_context *ioc;
 219
 220         ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
 221                                     node);
 222         if (unlikely(!ioc))
 223                 return;
 224
 225         /* initialize */
 226         atomic_long_set(&ioc->refcount, 1);
 227         atomic_set(&ioc->nr_tasks, 1);
 228         spin_lock_init(&ioc->lock);
 229         INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
 230         INIT_HLIST_HEAD(&ioc->icq_list);
 231         INIT_WORK(&ioc->release_work, ioc_release_fn);
 232
 233         /*
 234          * Try to install.  ioc shouldn't be installed if someone else
 235          * already did or @task, which isn't %current, is exiting.  Note
 236          * that we need to allow ioc creation on exiting %current as exit
 237          * path may issue IOs from e.g. exit_files().  The exit path is
 238          * responsible for not issuing IO after exit_io_context().
 239          */
 240         task_lock(task);
 241         if (!task->io_context &&
 242             (task == current || !(task->flags & PF_EXITING)))
 243                 task->io_context = ioc;
 244         else
 245                 kmem_cache_free(iocontext_cachep, ioc);
 246         task_unlock(task);
 247 }
 248
 249 /**
 250  * get_task_io_context - get io_context of a task
 251  * @task: task of interest
 252  * @gfp_flags: allocation flags, used if allocation is necessary
 253  * @node: allocation node, used if allocation is necessary
 254  *
 255  * Return io_context of @task.  If it doesn't exist, it is created with
 256  * @gfp_flags and @node.  The returned io_context has its reference count
 257  * incremented.
 258  *
 259  * This function always goes through task_lock() and it's better to use
 260  * %current->io_context + get_io_context() for %current.
 261  */
 262 struct io_context *get_task_io_context(struct task_struct *task,
 263                                        gfp_t gfp_flags, int node)
 264 {
 265         struct io_context *ioc;
 266
 267         might_sleep_if(gfp_flags & __GFP_WAIT);
 268
 269         do {
 270                 task_lock(task);
 271                 ioc = task->io_context;
 272                 if (likely(ioc)) {
 273                         get_io_context(ioc);
 274                         task_unlock(task);
 275                         return ioc;
 276                 }
 277                 task_unlock(task);
 278         } while (create_io_context(task, gfp_flags, node));
 279
 280         return NULL;
 281 }
 282 EXPORT_SYMBOL(get_task_io_context);
 283
 284 /**
 285  * ioc_lookup_icq - lookup io_cq from ioc
 286  * @ioc: the associated io_context
 287  * @q: the associated request_queue
 288  *
 289  * Look up io_cq associated with @ioc - @q pair from @ioc.  Must be called
 290  * with @q->queue_lock held.
 291  */
 292 struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
 293 {
 294         struct io_cq *icq;
 295
 296         lockdep_assert_held(q->queue_lock);
 297
 298         /*
 299          * icq's are indexed from @ioc using radix tree and hint pointer,
 300          * both of which are protected with RCU.  All removals are done
 301          * holding both q and ioc locks, and we're holding q lock - if we
 302          * find a icq which points to us, it's guaranteed to be valid.
 303          */
 304         rcu_read_lock();
 305         icq = rcu_dereference(ioc->icq_hint);
 306         if (icq && icq->q == q)
 307                 goto out;
 308
 309         icq = radix_tree_lookup(&ioc->icq_tree, q->id);
 310         if (icq && icq->q == q)
 311                 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
 312         else
 313                 icq = NULL;
 314 out:
 315         rcu_read_unlock();
 316         return icq;
 317 }
 318 EXPORT_SYMBOL(ioc_lookup_icq);
 319
 320 /**
 321  * ioc_create_icq - create and link io_cq
 322  * @q: request_queue of interest
 323  * @gfp_mask: allocation mask
 324  *
 325  * Make sure io_cq linking %current->io_context and @q exists.  If either
 326  * io_context and/or icq don't exist, they will be created using @gfp_mask.
 327  *
 328  * The caller is responsible for ensuring @ioc won't go away and @q is
 329  * alive and will stay alive until this function returns.
 330  */
 331 struct io_cq *ioc_create_icq(struct request_queue *q, gfp_t gfp_mask)
 332 {
 333         struct elevator_type *et = q->elevator->type;
 334         struct io_context *ioc;
 335         struct io_cq *icq;
 336
 337         /* allocate stuff */
 338         ioc = create_io_context(current, gfp_mask, q->node);
 339         if (!ioc)
 340                 return NULL;
 341
 342         icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
 343                                     q->node);
 344         if (!icq)
 345                 return NULL;
 346
 347         if (radix_tree_preload(gfp_mask) < 0) {
 348                 kmem_cache_free(et->icq_cache, icq);
 349                 return NULL;
 350         }
 351
 352         icq->ioc = ioc;
 353         icq->q = q;
 354         INIT_LIST_HEAD(&icq->q_node);
 355         INIT_HLIST_NODE(&icq->ioc_node);
 356
 357         /* lock both q and ioc and try to link @icq */
 358         spin_lock_irq(q->queue_lock);
 359         spin_lock(&ioc->lock);
 360
 361         if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
 362                 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
 363                 list_add(&icq->q_node, &q->icq_list);
 364                 if (et->ops.elevator_init_icq_fn)
 365                         et->ops.elevator_init_icq_fn(icq);
 366         } else {
 367                 kmem_cache_free(et->icq_cache, icq);
 368                 icq = ioc_lookup_icq(ioc, q);
 369                 if (!icq)
 370                         printk(KERN_ERR "cfq: icq link failed!\n");
 371         }
 372
 373         spin_unlock(&ioc->lock);
 374         spin_unlock_irq(q->queue_lock);
 375         radix_tree_preload_end();
 376         return icq;
 377 }
 378
 379 void ioc_set_icq_flags(struct io_context *ioc, unsigned int flags)
 380 {
 381         struct io_cq *icq;
 382         struct hlist_node *n;
 383
 384         hlist_for_each_entry(icq, n, &ioc->icq_list, ioc_node)
 385                 icq->flags |= flags;
 386 }
 387
 388 /**
 389  * ioc_ioprio_changed - notify ioprio change
 390  * @ioc: io_context of interest
 391  * @ioprio: new ioprio
 392  *
 393  * @ioc's ioprio has changed to @ioprio.  Set %ICQ_IOPRIO_CHANGED for all
 394  * icq's.  iosched is responsible for checking the bit and applying it on
 395  * request issue path.
 396  */
 397 void ioc_ioprio_changed(struct io_context *ioc, int ioprio)
 398 {
 399         unsigned long flags;
 400
 401         spin_lock_irqsave(&ioc->lock, flags);
 402         ioc->ioprio = ioprio;
 403         ioc_set_icq_flags(ioc, ICQ_IOPRIO_CHANGED);
 404         spin_unlock_irqrestore(&ioc->lock, flags);
 405 }
 406
 407 /**
 408  * ioc_cgroup_changed - notify cgroup change
 409  * @ioc: io_context of interest
 410  *
 411  * @ioc's cgroup has changed.  Set %ICQ_CGROUP_CHANGED for all icq's.
 412  * iosched is responsible for checking the bit and applying it on request
 413  * issue path.
 414  */
 415 void ioc_cgroup_changed(struct io_context *ioc)
 416 {
 417         unsigned long flags;
 418
 419         spin_lock_irqsave(&ioc->lock, flags);
 420         ioc_set_icq_flags(ioc, ICQ_CGROUP_CHANGED);
 421         spin_unlock_irqrestore(&ioc->lock, flags);
 422 }
 423 EXPORT_SYMBOL(ioc_cgroup_changed);
 424
 425 /**
 426  * icq_get_changed - fetch and clear icq changed mask
 427  * @icq: icq of interest
 428  *
 429  * Fetch and clear ICQ_*_CHANGED bits from @icq.  Grabs and releases
 430  * @icq->ioc->lock.
 431  */
 432 unsigned icq_get_changed(struct io_cq *icq)
 433 {
 434         unsigned int changed = 0;
 435         unsigned long flags;
 436
 437         if (unlikely(icq->flags & ICQ_CHANGED_MASK)) {
 438                 spin_lock_irqsave(&icq->ioc->lock, flags);
 439                 changed = icq->flags & ICQ_CHANGED_MASK;
 440                 icq->flags &= ~ICQ_CHANGED_MASK;
 441                 spin_unlock_irqrestore(&icq->ioc->lock, flags);
 442         }
 443         return changed;
 444 }
 445 EXPORT_SYMBOL(icq_get_changed);
 446
 447 static int __init blk_ioc_init(void)
 448 {
 449         iocontext_cachep = kmem_cache_create("blkdev_ioc",
 450                         sizeof(struct io_context), 0, SLAB_PANIC, NULL);
 451         return 0;
 452 }
 453 subsys_initcall(blk_ioc_init);