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         icq->flags |= ICQ_DESTROYED;
  91         call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
  92 }
  93
  94 /*
  95  * Slow path for ioc release in put_io_context().  Performs double-lock
  96  * dancing to unlink all icq's and then frees ioc.
  97  */
  98 static void ioc_release_fn(struct work_struct *work)
  99 {
 100         struct io_context *ioc = container_of(work, struct io_context,
 101                                               release_work);
 102         unsigned long flags;
 103
 104         /*
 105          * Exiting icq may call into put_io_context() through elevator
 106          * which will trigger lockdep warning.  The ioc's are guaranteed to
 107          * be different, use a different locking subclass here.  Use
 108          * irqsave variant as there's no spin_lock_irq_nested().
 109          */
 110         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 111
 112         while (!hlist_empty(&ioc->icq_list)) {
 113                 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
 114                                                 struct io_cq, ioc_node);
 115                 struct request_queue *q = icq->q;
 116
 117                 if (spin_trylock(q->queue_lock)) {
 118                         ioc_destroy_icq(icq);
 119                         spin_unlock(q->queue_lock);
 120                 } else {
 121                         spin_unlock_irqrestore(&ioc->lock, flags);
 122                         cpu_relax();
 123                         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 124                 }
 125         }
 126
 127         spin_unlock_irqrestore(&ioc->lock, flags);
 128
 129         kmem_cache_free(iocontext_cachep, ioc);
 130 }
 131
 132 /**
 133  * put_io_context - put a reference of io_context
 134  * @ioc: io_context to put
 135  *
 136  * Decrement reference count of @ioc and release it if the count reaches
 137  * zero.
 138  */
 139 void put_io_context(struct io_context *ioc)
 140 {
 141         unsigned long flags;
 142         bool free_ioc = false;
 143
 144         if (ioc == NULL)
 145                 return;
 146
 147         BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
 148
 149         /*
 150          * Releasing ioc requires reverse order double locking and we may
 151          * already be holding a queue_lock.  Do it asynchronously from wq.
 152          */
 153         if (atomic_long_dec_and_test(&ioc->refcount)) {
 154                 spin_lock_irqsave(&ioc->lock, flags);
 155                 if (!hlist_empty(&ioc->icq_list))
 156                         queue_work(system_power_efficient_wq,
 157                                         &ioc->release_work);
 158                 else
 159                         free_ioc = true;
 160                 spin_unlock_irqrestore(&ioc->lock, flags);
 161         }
 162
 163         if (free_ioc)
 164                 kmem_cache_free(iocontext_cachep, ioc);
 165 }
 166 EXPORT_SYMBOL(put_io_context);
 167
 168 /**
 169  * put_io_context_active - put active reference on ioc
 170  * @ioc: ioc of interest
 171  *
 172  * Undo get_io_context_active().  If active reference reaches zero after
 173  * put, @ioc can never issue further IOs and ioscheds are notified.
 174  */
 175 void put_io_context_active(struct io_context *ioc)
 176 {
 177         struct elevator_type *et;
 178         unsigned long flags;
 179         struct io_cq *icq;
 180
 181         if (!atomic_dec_and_test(&ioc->active_ref)) {
 182                 put_io_context(ioc);
 183                 return;
 184         }
 185
 186         /*
 187          * Need ioc lock to walk icq_list and q lock to exit icq.  Perform
 188          * reverse double locking.  Read comment in ioc_release_fn() for
 189          * explanation on the nested locking annotation.
 190          */
 191 retry:
 192         spin_lock_irqsave_nested(&ioc->lock, flags, 1);
 193         hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) {
 194                 if (icq->flags & ICQ_EXITED)
 195                         continue;
 196
 197                 et = icq->q->elevator->type;
 198                 if (et->uses_mq) {
 199                         ioc_exit_icq(icq);
 200                 } else {
 201                         if (spin_trylock(icq->q->queue_lock)) {
 202                                 ioc_exit_icq(icq);
 203                                 spin_unlock(icq->q->queue_lock);
 204                         } else {
 205                                 spin_unlock_irqrestore(&ioc->lock, flags);
 206                                 cpu_relax();
 207                                 goto retry;
 208                         }
 209                 }
 210         }
 211         spin_unlock_irqrestore(&ioc->lock, flags);
 212
 213         put_io_context(ioc);
 214 }
 215
 216 /* Called by the exiting task */
 217 void exit_io_context(struct task_struct *task)
 218 {
 219         struct io_context *ioc;
 220
 221         task_lock(task);
 222         ioc = task->io_context;
 223         task->io_context = NULL;
 224         task_unlock(task);
 225
 226         atomic_dec(&ioc->nr_tasks);
 227         put_io_context_active(ioc);
 228 }
 229
 230 static void __ioc_clear_queue(struct list_head *icq_list)
 231 {
 232         unsigned long flags;
 233
 234         rcu_read_lock();
 235         while (!list_empty(icq_list)) {
 236                 struct io_cq *icq = list_entry(icq_list->next,
 237                                                struct io_cq, q_node);
 238                 struct io_context *ioc = icq->ioc;
 239
 240                 spin_lock_irqsave(&ioc->lock, flags);
 241                 if (icq->flags & ICQ_DESTROYED) {
 242                         spin_unlock_irqrestore(&ioc->lock, flags);
 243                         continue;
 244                 }
 245                 ioc_destroy_icq(icq);
 246                 spin_unlock_irqrestore(&ioc->lock, flags);
 247         }
 248         rcu_read_unlock();
 249 }
 250
 251 /**
 252  * ioc_clear_queue - break any ioc association with the specified queue
 253  * @q: request_queue being cleared
 254  *
 255  * Walk @q->icq_list and exit all io_cq's.
 256  */
 257 void ioc_clear_queue(struct request_queue *q)
 258 {
 259         LIST_HEAD(icq_list);
 260
 261         spin_lock_irq(q->queue_lock);
 262         list_splice_init(&q->icq_list, &icq_list);
 263
 264         if (q->mq_ops) {
 265                 spin_unlock_irq(q->queue_lock);
 266                 __ioc_clear_queue(&icq_list);
 267         } else {
 268                 __ioc_clear_queue(&icq_list);
 269                 spin_unlock_irq(q->queue_lock);
 270         }
 271 }
 272
 273 int create_task_io_context(struct task_struct *task, gfp_t gfp_flags, int node)
 274 {
 275         struct io_context *ioc;
 276         int ret;
 277
 278         ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
 279                                     node);
 280         if (unlikely(!ioc))
 281                 return -ENOMEM;
 282
 283         /* initialize */
 284         atomic_long_set(&ioc->refcount, 1);
 285         atomic_set(&ioc->nr_tasks, 1);
 286         atomic_set(&ioc->active_ref, 1);
 287         spin_lock_init(&ioc->lock);
 288         INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
 289         INIT_HLIST_HEAD(&ioc->icq_list);
 290         INIT_WORK(&ioc->release_work, ioc_release_fn);
 291
 292         /*
 293          * Try to install.  ioc shouldn't be installed if someone else
 294          * already did or @task, which isn't %current, is exiting.  Note
 295          * that we need to allow ioc creation on exiting %current as exit
 296          * path may issue IOs from e.g. exit_files().  The exit path is
 297          * responsible for not issuing IO after exit_io_context().
 298          */
 299         task_lock(task);
 300         if (!task->io_context &&
 301             (task == current || !(task->flags & PF_EXITING)))
 302                 task->io_context = ioc;
 303         else
 304                 kmem_cache_free(iocontext_cachep, ioc);
 305
 306         ret = task->io_context ? 0 : -EBUSY;
 307
 308         task_unlock(task);
 309
 310         return ret;
 311 }
 312
 313 /**
 314  * get_task_io_context - get io_context of a task
 315  * @task: task of interest
 316  * @gfp_flags: allocation flags, used if allocation is necessary
 317  * @node: allocation node, used if allocation is necessary
 318  *
 319  * Return io_context of @task.  If it doesn't exist, it is created with
 320  * @gfp_flags and @node.  The returned io_context has its reference count
 321  * incremented.
 322  *
 323  * This function always goes through task_lock() and it's better to use
 324  * %current->io_context + get_io_context() for %current.
 325  */
 326 struct io_context *get_task_io_context(struct task_struct *task,
 327                                        gfp_t gfp_flags, int node)
 328 {
 329         struct io_context *ioc;
 330
 331         might_sleep_if(gfpflags_allow_blocking(gfp_flags));
 332
 333         do {
 334                 task_lock(task);
 335                 ioc = task->io_context;
 336                 if (likely(ioc)) {
 337                         get_io_context(ioc);
 338                         task_unlock(task);
 339                         return ioc;
 340                 }
 341                 task_unlock(task);
 342         } while (!create_task_io_context(task, gfp_flags, node));
 343
 344         return NULL;
 345 }
 346 EXPORT_SYMBOL(get_task_io_context);
 347
 348 /**
 349  * ioc_lookup_icq - lookup io_cq from ioc
 350  * @ioc: the associated io_context
 351  * @q: the associated request_queue
 352  *
 353  * Look up io_cq associated with @ioc - @q pair from @ioc.  Must be called
 354  * with @q->queue_lock held.
 355  */
 356 struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
 357 {
 358         struct io_cq *icq;
 359
 360         lockdep_assert_held(q->queue_lock);
 361
 362         /*
 363          * icq's are indexed from @ioc using radix tree and hint pointer,
 364          * both of which are protected with RCU.  All removals are done
 365          * holding both q and ioc locks, and we're holding q lock - if we
 366          * find a icq which points to us, it's guaranteed to be valid.
 367          */
 368         rcu_read_lock();
 369         icq = rcu_dereference(ioc->icq_hint);
 370         if (icq && icq->q == q)
 371                 goto out;
 372
 373         icq = radix_tree_lookup(&ioc->icq_tree, q->id);
 374         if (icq && icq->q == q)
 375                 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
 376         else
 377                 icq = NULL;
 378 out:
 379         rcu_read_unlock();
 380         return icq;
 381 }
 382 EXPORT_SYMBOL(ioc_lookup_icq);
 383
 384 /**
 385  * ioc_create_icq - create and link io_cq
 386  * @ioc: io_context of interest
 387  * @q: request_queue of interest
 388  * @gfp_mask: allocation mask
 389  *
 390  * Make sure io_cq linking @ioc and @q exists.  If icq doesn't exist, they
 391  * will be created using @gfp_mask.
 392  *
 393  * The caller is responsible for ensuring @ioc won't go away and @q is
 394  * alive and will stay alive until this function returns.
 395  */
 396 struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q,
 397                              gfp_t gfp_mask)
 398 {
 399         struct elevator_type *et = q->elevator->type;
 400         struct io_cq *icq;
 401
 402         /* allocate stuff */
 403         icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
 404                                     q->node);
 405         if (!icq)
 406                 return NULL;
 407
 408         if (radix_tree_maybe_preload(gfp_mask) < 0) {
 409                 kmem_cache_free(et->icq_cache, icq);
 410                 return NULL;
 411         }
 412
 413         icq->ioc = ioc;
 414         icq->q = q;
 415         INIT_LIST_HEAD(&icq->q_node);
 416         INIT_HLIST_NODE(&icq->ioc_node);
 417
 418         /* lock both q and ioc and try to link @icq */
 419         spin_lock_irq(q->queue_lock);
 420         spin_lock(&ioc->lock);
 421
 422         if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
 423                 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
 424                 list_add(&icq->q_node, &q->icq_list);
 425                 if (et->uses_mq && et->ops.mq.init_icq)
 426                         et->ops.mq.init_icq(icq);
 427                 else if (!et->uses_mq && et->ops.sq.elevator_init_icq_fn)
 428                         et->ops.sq.elevator_init_icq_fn(icq);
 429         } else {
 430                 kmem_cache_free(et->icq_cache, icq);
 431                 icq = ioc_lookup_icq(ioc, q);
 432                 if (!icq)
 433                         printk(KERN_ERR "cfq: icq link failed!\n");
 434         }
 435
 436         spin_unlock(&ioc->lock);
 437         spin_unlock_irq(q->queue_lock);
 438         radix_tree_preload_end();
 439         return icq;
 440 }
 441
 442 static int __init blk_ioc_init(void)
 443 {
 444         iocontext_cachep = kmem_cache_create("blkdev_ioc",
 445                         sizeof(struct io_context), 0, SLAB_PANIC, NULL);
 446         return 0;
 447 }
 448 subsys_initcall(blk_ioc_init);