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