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 /* Called by the exiting task */
 159 void exit_io_context(struct task_struct *task)
 160 {
 161         struct io_context *ioc;
 162         struct io_cq *icq;
 163         struct hlist_node *n;
 164         unsigned long flags;
 165
 166         task_lock(task);
 167         ioc = task->io_context;
 168         task->io_context = NULL;
 169         task_unlock(task);
 170
 171         if (!atomic_dec_and_test(&ioc->nr_tasks)) {
 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 /**
 201  * ioc_clear_queue - break any ioc association with the specified queue
 202  * @q: request_queue being cleared
 203  *
 204  * Walk @q->icq_list and exit all io_cq's.  Must be called with @q locked.
 205  */
 206 void ioc_clear_queue(struct request_queue *q)
 207 {
 208         lockdep_assert_held(q->queue_lock);
 209
 210         while (!list_empty(&q->icq_list)) {
 211                 struct io_cq *icq = list_entry(q->icq_list.next,
 212                                                struct io_cq, q_node);
 213                 struct io_context *ioc = icq->ioc;
 214
 215                 spin_lock(&ioc->lock);
 216                 ioc_destroy_icq(icq);
 217                 spin_unlock(&ioc->lock);
 218         }
 219 }
 220
 221 void create_io_context_slowpath(struct task_struct *task, gfp_t gfp_flags,
 222                                 int node)
 223 {
 224         struct io_context *ioc;
 225
 226         ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
 227                                     node);
 228         if (unlikely(!ioc))
 229                 return;
 230
 231         /* initialize */
 232         atomic_long_set(&ioc->refcount, 1);
 233         atomic_set(&ioc->nr_tasks, 1);
 234         spin_lock_init(&ioc->lock);
 235         INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
 236         INIT_HLIST_HEAD(&ioc->icq_list);
 237         INIT_WORK(&ioc->release_work, ioc_release_fn);
 238
 239         /*
 240          * Try to install.  ioc shouldn't be installed if someone else
 241          * already did or @task, which isn't %current, is exiting.  Note
 242          * that we need to allow ioc creation on exiting %current as exit
 243          * path may issue IOs from e.g. exit_files().  The exit path is
 244          * responsible for not issuing IO after exit_io_context().
 245          */
 246         task_lock(task);
 247         if (!task->io_context &&
 248             (task == current || !(task->flags & PF_EXITING)))
 249                 task->io_context = ioc;
 250         else
 251                 kmem_cache_free(iocontext_cachep, ioc);
 252         task_unlock(task);
 253 }
 254
 255 /**
 256  * get_task_io_context - get io_context of a task
 257  * @task: task of interest
 258  * @gfp_flags: allocation flags, used if allocation is necessary
 259  * @node: allocation node, used if allocation is necessary
 260  *
 261  * Return io_context of @task.  If it doesn't exist, it is created with
 262  * @gfp_flags and @node.  The returned io_context has its reference count
 263  * incremented.
 264  *
 265  * This function always goes through task_lock() and it's better to use
 266  * %current->io_context + get_io_context() for %current.
 267  */
 268 struct io_context *get_task_io_context(struct task_struct *task,
 269                                        gfp_t gfp_flags, int node)
 270 {
 271         struct io_context *ioc;
 272
 273         might_sleep_if(gfp_flags & __GFP_WAIT);
 274
 275         do {
 276                 task_lock(task);
 277                 ioc = task->io_context;
 278                 if (likely(ioc)) {
 279                         get_io_context(ioc);
 280                         task_unlock(task);
 281                         return ioc;
 282                 }
 283                 task_unlock(task);
 284         } while (create_io_context(task, gfp_flags, node));
 285
 286         return NULL;
 287 }
 288 EXPORT_SYMBOL(get_task_io_context);
 289
 290 /**
 291  * ioc_lookup_icq - lookup io_cq from ioc
 292  * @ioc: the associated io_context
 293  * @q: the associated request_queue
 294  *
 295  * Look up io_cq associated with @ioc - @q pair from @ioc.  Must be called
 296  * with @q->queue_lock held.
 297  */
 298 struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
 299 {
 300         struct io_cq *icq;
 301
 302         lockdep_assert_held(q->queue_lock);
 303
 304         /*
 305          * icq's are indexed from @ioc using radix tree and hint pointer,
 306          * both of which are protected with RCU.  All removals are done
 307          * holding both q and ioc locks, and we're holding q lock - if we
 308          * find a icq which points to us, it's guaranteed to be valid.
 309          */
 310         rcu_read_lock();
 311         icq = rcu_dereference(ioc->icq_hint);
 312         if (icq && icq->q == q)
 313                 goto out;
 314
 315         icq = radix_tree_lookup(&ioc->icq_tree, q->id);
 316         if (icq && icq->q == q)
 317                 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
 318         else
 319                 icq = NULL;
 320 out:
 321         rcu_read_unlock();
 322         return icq;
 323 }
 324 EXPORT_SYMBOL(ioc_lookup_icq);
 325
 326 /**
 327  * ioc_create_icq - create and link io_cq
 328  * @q: request_queue of interest
 329  * @gfp_mask: allocation mask
 330  *
 331  * Make sure io_cq linking %current->io_context and @q exists.  If either
 332  * io_context and/or icq don't exist, they will be created using @gfp_mask.
 333  *
 334  * The caller is responsible for ensuring @ioc won't go away and @q is
 335  * alive and will stay alive until this function returns.
 336  */
 337 struct io_cq *ioc_create_icq(struct request_queue *q, gfp_t gfp_mask)
 338 {
 339         struct elevator_type *et = q->elevator->type;
 340         struct io_context *ioc;
 341         struct io_cq *icq;
 342
 343         /* allocate stuff */
 344         ioc = create_io_context(current, gfp_mask, q->node);
 345         if (!ioc)
 346                 return NULL;
 347
 348         icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
 349                                     q->node);
 350         if (!icq)
 351                 return NULL;
 352
 353         if (radix_tree_preload(gfp_mask) < 0) {
 354                 kmem_cache_free(et->icq_cache, icq);
 355                 return NULL;
 356         }
 357
 358         icq->ioc = ioc;
 359         icq->q = q;
 360         INIT_LIST_HEAD(&icq->q_node);
 361         INIT_HLIST_NODE(&icq->ioc_node);
 362
 363         /* lock both q and ioc and try to link @icq */
 364         spin_lock_irq(q->queue_lock);
 365         spin_lock(&ioc->lock);
 366
 367         if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
 368                 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
 369                 list_add(&icq->q_node, &q->icq_list);
 370                 if (et->ops.elevator_init_icq_fn)
 371                         et->ops.elevator_init_icq_fn(icq);
 372         } else {
 373                 kmem_cache_free(et->icq_cache, icq);
 374                 icq = ioc_lookup_icq(ioc, q);
 375                 if (!icq)
 376                         printk(KERN_ERR "cfq: icq link failed!\n");
 377         }
 378
 379         spin_unlock(&ioc->lock);
 380         spin_unlock_irq(q->queue_lock);
 381         radix_tree_preload_end();
 382         return icq;
 383 }
 384
 385 void ioc_set_icq_flags(struct io_context *ioc, unsigned int flags)
 386 {
 387         struct io_cq *icq;
 388         struct hlist_node *n;
 389
 390         hlist_for_each_entry(icq, n, &ioc->icq_list, ioc_node)
 391                 icq->flags |= flags;
 392 }
 393
 394 /**
 395  * ioc_ioprio_changed - notify ioprio change
 396  * @ioc: io_context of interest
 397  * @ioprio: new ioprio
 398  *
 399  * @ioc's ioprio has changed to @ioprio.  Set %ICQ_IOPRIO_CHANGED for all
 400  * icq's.  iosched is responsible for checking the bit and applying it on
 401  * request issue path.
 402  */
 403 void ioc_ioprio_changed(struct io_context *ioc, int ioprio)
 404 {
 405         unsigned long flags;
 406
 407         spin_lock_irqsave(&ioc->lock, flags);
 408         ioc->ioprio = ioprio;
 409         ioc_set_icq_flags(ioc, ICQ_IOPRIO_CHANGED);
 410         spin_unlock_irqrestore(&ioc->lock, flags);
 411 }
 412
 413 /**
 414  * ioc_cgroup_changed - notify cgroup change
 415  * @ioc: io_context of interest
 416  *
 417  * @ioc's cgroup has changed.  Set %ICQ_CGROUP_CHANGED for all icq's.
 418  * iosched is responsible for checking the bit and applying it on request
 419  * issue path.
 420  */
 421 void ioc_cgroup_changed(struct io_context *ioc)
 422 {
 423         unsigned long flags;
 424
 425         spin_lock_irqsave(&ioc->lock, flags);
 426         ioc_set_icq_flags(ioc, ICQ_CGROUP_CHANGED);
 427         spin_unlock_irqrestore(&ioc->lock, flags);
 428 }
 429 EXPORT_SYMBOL(ioc_cgroup_changed);
 430
 431 /**
 432  * icq_get_changed - fetch and clear icq changed mask
 433  * @icq: icq of interest
 434  *
 435  * Fetch and clear ICQ_*_CHANGED bits from @icq.  Grabs and releases
 436  * @icq->ioc->lock.
 437  */
 438 unsigned icq_get_changed(struct io_cq *icq)
 439 {
 440         unsigned int changed = 0;
 441         unsigned long flags;
 442
 443         if (unlikely(icq->flags & ICQ_CHANGED_MASK)) {
 444                 spin_lock_irqsave(&icq->ioc->lock, flags);
 445                 changed = icq->flags & ICQ_CHANGED_MASK;
 446                 icq->flags &= ~ICQ_CHANGED_MASK;
 447                 spin_unlock_irqrestore(&icq->ioc->lock, flags);
 448         }
 449         return changed;
 450 }
 451 EXPORT_SYMBOL(icq_get_changed);
 452
 453 static int __init blk_ioc_init(void)
 454 {
 455         iocontext_cachep = kmem_cache_create("blkdev_ioc",
 456                         sizeof(struct io_context), 0, SLAB_PANIC, NULL);
 457         return 0;
 458 }
 459 subsys_initcall(blk_ioc_init);