block/blk-ioc.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Functions related to io context handling
   4  */
   5 #include <linux/kernel.h>
   6 #include <linux/module.h>
   7 #include <linux/init.h>
   8 #include <linux/bio.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/slab.h>
  11 #include <linux/sched/task.h>
  12
  13 #include "blk.h"
  14
  15 /*
  16  * For io context allocations
  17  */
  18 static struct kmem_cache *iocontext_cachep;
  19
  20 /**
  21  * get_io_context - increment reference count to io_context
  22  * @ioc: io_context to get
  23  *
  24  * Increment reference count to @ioc.
  25  */
  26 void get_io_context(struct io_context *ioc)
  27 {
  28         BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
  29         atomic_long_inc(&ioc->refcount);
  30 }
  31 EXPORT_SYMBOL(get_io_context);
  32
  33 static void icq_free_icq_rcu(struct rcu_head *head)
  34 {
  35         struct io_cq *icq = container_of(head, struct io_cq, __rcu_head);
  36
  37         kmem_cache_free(icq->__rcu_icq_cache, icq);
  38 }
  39
  40 /*
  41  * Exit an icq. Called with ioc locked for blk-mq, and with both ioc
  42  * and queue locked for legacy.
  43  */
  44 static void ioc_exit_icq(struct io_cq *icq)
  45 {
  46         struct elevator_type *et = icq->q->elevator->type;
  47
  48         if (icq->flags & ICQ_EXITED)
  49                 return;
  50
  51         if (et->uses_mq && et->ops.mq.exit_icq)
  52                 et->ops.mq.exit_icq(icq);
  53         else if (!et->uses_mq && et->ops.sq.elevator_exit_icq_fn)
  54                 et->ops.sq.elevator_exit_icq_fn(icq);
  55
  56         icq->flags |= ICQ_EXITED;
  57 }
  58
  59 /*
  60  * Release an icq. Called with ioc locked for blk-mq, and with both ioc
  61  * and queue locked for legacy.
  62  */
  63 static void ioc_destroy_icq(struct io_cq *icq)
  64 {
  65         struct io_context *ioc = icq->ioc;
  66         struct request_queue *q = icq->q;
  67         struct elevator_type *et = q->elevator->type;
  68
  69         lockdep_assert_held(&ioc->lock);
  70
  71         radix_tree_delete(&ioc->icq_tree, icq->q->id);
  72         hlist_del_init(&icq->ioc_node);
  73         list_del_init(&icq->q_node);
  74
  75         /*
  76          * Both setting lookup hint to and clearing it from @icq are done
  77          * under queue_lock.  If it's not pointing to @icq now, it never
  78          * will.  Hint assignment itself can race safely.
  79          */
  80         if (rcu_access_pointer(ioc->icq_hint) == icq)
  81                 rcu_assign_pointer(ioc->icq_hint, NULL);
  82
  83         ioc_exit_icq(icq);
  84
  85         /*
  86          * @icq->q might have gone away by the time RCU callback runs
  87          * making it impossible to determine icq_cache.  Record it in @icq.
  88          */
  89         icq->__rcu_icq_cache = et->icq_cache;
  90         call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
  91 }
  92
  93 /*
  94  * Slow path for ioc release in put_io_context().  Performs double-lock
  95  * dancing to unlink all icq's and then frees ioc.
  96  */
  97 static void ioc_release_fn(struct work_struct *work)
  98 {
  99         struct io_context *ioc = container_of(work, struct io_context,
 100                                               release_work);
 101         unsigned long flags;
 102
 103         /*
 104          * Exiting icq may call into put_io_context() through elevator
 105          * which will trigger lockdep warning.  The ioc's are guaranteed to
 106          * be different, use a different locking subclass here.  Use
 107          * irqsave variant as there's no spin_lock_irq_nested().
 108          */
 109         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 110
 111         while (!hlist_empty(&ioc->icq_list)) {
 112                 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
 113                                                 struct io_cq, ioc_node);
 114                 struct request_queue *q = icq->q;
 115
 116                 if (spin_trylock(q->queue_lock)) {
 117                         ioc_destroy_icq(icq);
 118                         spin_unlock(q->queue_lock);
 119                 } else {
 120                         spin_unlock_irqrestore(&ioc->lock, flags);
 121                         cpu_relax();
 122                         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 123                 }
 124         }
 125
 126         spin_unlock_irqrestore(&ioc->lock, flags);
 127
 128         kmem_cache_free(iocontext_cachep, ioc);
 129 }
 130
 131 /**
 132  * put_io_context - put a reference of io_context
 133  * @ioc: io_context to put
 134  *
 135  * Decrement reference count of @ioc and release it if the count reaches
 136  * zero.
 137  */
 138 void put_io_context(struct io_context *ioc)
 139 {
 140         unsigned long flags;
 141         bool free_ioc = false;
 142
 143         if (ioc == NULL)
 144                 return;
 145
 146         BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
 147
 148         /*
 149          * Releasing ioc requires reverse order double locking and we may
 150          * already be holding a queue_lock.  Do it asynchronously from wq.
 151          */
 152         if (atomic_long_dec_and_test(&ioc->refcount)) {
 153                 spin_lock_irqsave(&ioc->lock, flags);
 154                 if (!hlist_empty(&ioc->icq_list))
 155                         queue_work(system_power_efficient_wq,
 156                                         &ioc->release_work);
 157                 else
 158                         free_ioc = true;
 159                 spin_unlock_irqrestore(&ioc->lock, flags);
 160         }
 161
 162         if (free_ioc)
 163                 kmem_cache_free(iocontext_cachep, ioc);
 164 }
 165 EXPORT_SYMBOL(put_io_context);
 166
 167 /**
 168  * put_io_context_active - put active reference on ioc
 169  * @ioc: ioc of interest
 170  *
 171  * Undo get_io_context_active().  If active reference reaches zero after
 172  * put, @ioc can never issue further IOs and ioscheds are notified.
 173  */
 174 void put_io_context_active(struct io_context *ioc)
 175 {
 176         struct elevator_type *et;
 177         unsigned long flags;
 178         struct io_cq *icq;
 179
 180         if (!atomic_dec_and_test(&ioc->active_ref)) {
 181                 put_io_context(ioc);
 182                 return;
 183         }
 184
 185         /*
 186          * Need ioc lock to walk icq_list and q lock to exit icq.  Perform
 187          * reverse double locking.  Read comment in ioc_release_fn() for
 188          * explanation on the nested locking annotation.
 189          */
 190 retry:
 191         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 192         hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) {
 193                 if (icq->flags & ICQ_EXITED)
 194                         continue;
 195
 196                 et = icq->q->elevator->type;
 197                 if (et->uses_mq) {
 198                         ioc_exit_icq(icq);
 199                 } else {
 200                         if (spin_trylock(icq->q->queue_lock)) {
 201                                 ioc_exit_icq(icq);
 202                                 spin_unlock(icq->q->queue_lock);
 203                         } else {
 204                                 spin_unlock_irqrestore(&ioc->lock, flags);
 205                                 cpu_relax();
 206                                 goto retry;
 207                         }
 208                 }
 209         }
 210         spin_unlock_irqrestore(&ioc->lock, flags);
 211
 212         put_io_context(ioc);
 213 }
 214
 215 /* Called by the exiting task */
 216 void exit_io_context(struct task_struct *task)
 217 {
 218         struct io_context *ioc;
 219
 220         task_lock(task);
 221         ioc = task->io_context;
 222         task->io_context = NULL;
 223         task_unlock(task);
 224
 225         atomic_dec(&ioc->nr_tasks);
 226         put_io_context_active(ioc);
 227 }
 228
 229 static void __ioc_clear_queue(struct list_head *icq_list)
 230 {
 231         unsigned long flags;
 232
 233         while (!list_empty(icq_list)) {
 234                 struct io_cq *icq = list_entry(icq_list->next,
 235                                                struct io_cq, q_node);
 236                 struct io_context *ioc = icq->ioc;
 237
 238                 spin_lock_irqsave(&ioc->lock, flags);
 239                 ioc_destroy_icq(icq);
 240                 spin_unlock_irqrestore(&ioc->lock, flags);
 241         }
 242 }
 243
 244 /**
 245  * ioc_clear_queue - break any ioc association with the specified queue
 246  * @q: request_queue being cleared
 247  *
 248  * Walk @q->icq_list and exit all io_cq's.
 249  */
 250 void ioc_clear_queue(struct request_queue *q)
 251 {
 252         LIST_HEAD(icq_list);
 253
 254         spin_lock_irq(q->queue_lock);
 255         list_splice_init(&q->icq_list, &icq_list);
 256
 257         if (q->mq_ops) {
 258                 spin_unlock_irq(q->queue_lock);
 259                 __ioc_clear_queue(&icq_list);
 260         } else {
 261                 __ioc_clear_queue(&icq_list);
 262                 spin_unlock_irq(q->queue_lock);
 263         }
 264 }
 265
 266 int create_task_io_context(struct task_struct *task, gfp_t gfp_flags, int node)
 267 {
 268         struct io_context *ioc;
 269         int ret;
 270
 271         ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
 272                                     node);
 273         if (unlikely(!ioc))
 274                 return -ENOMEM;
 275
 276         /* initialize */
 277         atomic_long_set(&ioc->refcount, 1);
 278         atomic_set(&ioc->nr_tasks, 1);
 279         atomic_set(&ioc->active_ref, 1);
 280         spin_lock_init(&ioc->lock);
 281         INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC);
 282         INIT_HLIST_HEAD(&ioc->icq_list);
 283         INIT_WORK(&ioc->release_work, ioc_release_fn);
 284
 285         /*
 286          * Try to install.  ioc shouldn't be installed if someone else
 287          * already did or @task, which isn't %current, is exiting.  Note
 288          * that we need to allow ioc creation on exiting %current as exit
 289          * path may issue IOs from e.g. exit_files().  The exit path is
 290          * responsible for not issuing IO after exit_io_context().
 291          */
 292         task_lock(task);
 293         if (!task->io_context &&
 294             (task == current || !(task->flags & PF_EXITING)))
 295                 task->io_context = ioc;
 296         else
 297                 kmem_cache_free(iocontext_cachep, ioc);
 298
 299         ret = task->io_context ? 0 : -EBUSY;
 300
 301         task_unlock(task);
 302
 303         return ret;
 304 }
 305
 306 /**
 307  * get_task_io_context - get io_context of a task
 308  * @task: task of interest
 309  * @gfp_flags: allocation flags, used if allocation is necessary
 310  * @node: allocation node, used if allocation is necessary
 311  *
 312  * Return io_context of @task.  If it doesn't exist, it is created with
 313  * @gfp_flags and @node.  The returned io_context has its reference count
 314  * incremented.
 315  *
 316  * This function always goes through task_lock() and it's better to use
 317  * %current->io_context + get_io_context() for %current.
 318  */
 319 struct io_context *get_task_io_context(struct task_struct *task,
 320                                        gfp_t gfp_flags, int node)
 321 {
 322         struct io_context *ioc;
 323
 324         might_sleep_if(gfpflags_allow_blocking(gfp_flags));
 325
 326         do {
 327                 task_lock(task);
 328                 ioc = task->io_context;
 329                 if (likely(ioc)) {
 330                         get_io_context(ioc);
 331                         task_unlock(task);
 332                         return ioc;
 333                 }
 334                 task_unlock(task);
 335         } while (!create_task_io_context(task, gfp_flags, node));
 336
 337         return NULL;
 338 }
 339 EXPORT_SYMBOL(get_task_io_context);
 340
 341 /**
 342  * ioc_lookup_icq - lookup io_cq from ioc
 343  * @ioc: the associated io_context
 344  * @q: the associated request_queue
 345  *
 346  * Look up io_cq associated with @ioc - @q pair from @ioc.  Must be called
 347  * with @q->queue_lock held.
 348  */
 349 struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
 350 {
 351         struct io_cq *icq;
 352
 353         lockdep_assert_held(q->queue_lock);
 354
 355         /*
 356          * icq's are indexed from @ioc using radix tree and hint pointer,
 357          * both of which are protected with RCU.  All removals are done
 358          * holding both q and ioc locks, and we're holding q lock - if we
 359          * find a icq which points to us, it's guaranteed to be valid.
 360          */
 361         rcu_read_lock();
 362         icq = rcu_dereference(ioc->icq_hint);
 363         if (icq && icq->q == q)
 364                 goto out;
 365
 366         icq = radix_tree_lookup(&ioc->icq_tree, q->id);
 367         if (icq && icq->q == q)
 368                 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
 369         else
 370                 icq = NULL;
 371 out:
 372         rcu_read_unlock();
 373         return icq;
 374 }
 375 EXPORT_SYMBOL(ioc_lookup_icq);
 376
 377 /**
 378  * ioc_create_icq - create and link io_cq
 379  * @ioc: io_context of interest
 380  * @q: request_queue of interest
 381  * @gfp_mask: allocation mask
 382  *
 383  * Make sure io_cq linking @ioc and @q exists.  If icq doesn't exist, they
 384  * will be created using @gfp_mask.
 385  *
 386  * The caller is responsible for ensuring @ioc won't go away and @q is
 387  * alive and will stay alive until this function returns.
 388  */
 389 struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q,
 390                              gfp_t gfp_mask)
 391 {
 392         struct elevator_type *et = q->elevator->type;
 393         struct io_cq *icq;
 394
 395         /* allocate stuff */
 396         icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
 397                                     q->node);
 398         if (!icq)
 399                 return NULL;
 400
 401         if (radix_tree_maybe_preload(gfp_mask) < 0) {
 402                 kmem_cache_free(et->icq_cache, icq);
 403                 return NULL;
 404         }
 405
 406         icq->ioc = ioc;
 407         icq->q = q;
 408         INIT_LIST_HEAD(&icq->q_node);
 409         INIT_HLIST_NODE(&icq->ioc_node);
 410
 411         /* lock both q and ioc and try to link @icq */
 412         spin_lock_irq(q->queue_lock);
 413         spin_lock(&ioc->lock);
 414
 415         if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
 416                 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
 417                 list_add(&icq->q_node, &q->icq_list);
 418                 if (et->uses_mq && et->ops.mq.init_icq)
 419                         et->ops.mq.init_icq(icq);
 420                 else if (!et->uses_mq && et->ops.sq.elevator_init_icq_fn)
 421                         et->ops.sq.elevator_init_icq_fn(icq);
 422         } else {
 423                 kmem_cache_free(et->icq_cache, icq);
 424                 icq = ioc_lookup_icq(ioc, q);
 425                 if (!icq)
 426                         printk(KERN_ERR "cfq: icq link failed!\n");
 427         }
 428
 429         spin_unlock(&ioc->lock);
 430         spin_unlock_irq(q->queue_lock);
 431         radix_tree_preload_end();
 432         return icq;
 433 }
 434
 435 static int __init blk_ioc_init(void)
 436 {
 437         iocontext_cachep = kmem_cache_create("blkdev_ioc",
 438                         sizeof(struct io_context), 0, SLAB_PANIC, NULL);
 439         return 0;
 440 }
 441 subsys_initcall(blk_ioc_init);