1 // SPDX-License-Identifier: GPL-2.0
3 * Functions related to io context handling
5 #include <linux/kernel.h>
6 #include <linux/module.h>
7 #include <linux/init.h>
9 #include <linux/blkdev.h>
10 #include <linux/slab.h>
11 #include <linux/security.h>
12 #include <linux/sched/task.h>
15 #include "blk-mq-sched.h"
18 * For io context allocations
20 static struct kmem_cache *iocontext_cachep;
24 * get_io_context - increment reference count to io_context
25 * @ioc: io_context to get
27 * Increment reference count to @ioc.
29 static void get_io_context(struct io_context *ioc)
31 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
32 atomic_long_inc(&ioc->refcount);
35 static void icq_free_icq_rcu(struct rcu_head *head)
37 struct io_cq *icq = container_of(head, struct io_cq, __rcu_head);
39 kmem_cache_free(icq->__rcu_icq_cache, icq);
43 * Exit an icq. Called with ioc locked for blk-mq, and with both ioc
44 * and queue locked for legacy.
46 static void ioc_exit_icq(struct io_cq *icq)
48 struct elevator_type *et = icq->q->elevator->type;
50 if (icq->flags & ICQ_EXITED)
54 et->ops.exit_icq(icq);
56 icq->flags |= ICQ_EXITED;
59 static void ioc_exit_icqs(struct io_context *ioc)
63 spin_lock_irq(&ioc->lock);
64 hlist_for_each_entry(icq, &ioc->icq_list, ioc_node)
66 spin_unlock_irq(&ioc->lock);
70 * Release an icq. Called with ioc locked for blk-mq, and with both ioc
71 * and queue locked for legacy.
73 static void ioc_destroy_icq(struct io_cq *icq)
75 struct io_context *ioc = icq->ioc;
76 struct request_queue *q = icq->q;
77 struct elevator_type *et = q->elevator->type;
79 lockdep_assert_held(&ioc->lock);
80 lockdep_assert_held(&q->queue_lock);
82 if (icq->flags & ICQ_DESTROYED)
85 radix_tree_delete(&ioc->icq_tree, icq->q->id);
86 hlist_del_init(&icq->ioc_node);
87 list_del_init(&icq->q_node);
90 * Both setting lookup hint to and clearing it from @icq are done
91 * under queue_lock. If it's not pointing to @icq now, it never
92 * will. Hint assignment itself can race safely.
94 if (rcu_access_pointer(ioc->icq_hint) == icq)
95 rcu_assign_pointer(ioc->icq_hint, NULL);
100 * @icq->q might have gone away by the time RCU callback runs
101 * making it impossible to determine icq_cache. Record it in @icq.
103 icq->__rcu_icq_cache = et->icq_cache;
104 icq->flags |= ICQ_DESTROYED;
105 call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
109 * Slow path for ioc release in put_io_context(). Performs double-lock
110 * dancing to unlink all icq's and then frees ioc.
112 static void ioc_release_fn(struct work_struct *work)
114 struct io_context *ioc = container_of(work, struct io_context,
116 spin_lock_irq(&ioc->lock);
118 while (!hlist_empty(&ioc->icq_list)) {
119 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
120 struct io_cq, ioc_node);
121 struct request_queue *q = icq->q;
123 if (spin_trylock(&q->queue_lock)) {
124 ioc_destroy_icq(icq);
125 spin_unlock(&q->queue_lock);
127 /* Make sure q and icq cannot be freed. */
130 /* Re-acquire the locks in the correct order. */
131 spin_unlock(&ioc->lock);
132 spin_lock(&q->queue_lock);
133 spin_lock(&ioc->lock);
135 ioc_destroy_icq(icq);
137 spin_unlock(&q->queue_lock);
142 spin_unlock_irq(&ioc->lock);
144 kmem_cache_free(iocontext_cachep, ioc);
148 * Releasing icqs requires reverse order double locking and we may already be
149 * holding a queue_lock. Do it asynchronously from a workqueue.
151 static bool ioc_delay_free(struct io_context *ioc)
155 spin_lock_irqsave(&ioc->lock, flags);
156 if (!hlist_empty(&ioc->icq_list)) {
157 queue_work(system_power_efficient_wq, &ioc->release_work);
158 spin_unlock_irqrestore(&ioc->lock, flags);
161 spin_unlock_irqrestore(&ioc->lock, flags);
166 * ioc_clear_queue - break any ioc association with the specified queue
167 * @q: request_queue being cleared
169 * Walk @q->icq_list and exit all io_cq's.
171 void ioc_clear_queue(struct request_queue *q)
173 spin_lock_irq(&q->queue_lock);
174 while (!list_empty(&q->icq_list)) {
176 list_first_entry(&q->icq_list, struct io_cq, q_node);
179 * Other context won't hold ioc lock to wait for queue_lock, see
180 * details in ioc_release_fn().
182 spin_lock(&icq->ioc->lock);
183 ioc_destroy_icq(icq);
184 spin_unlock(&icq->ioc->lock);
186 spin_unlock_irq(&q->queue_lock);
188 #else /* CONFIG_BLK_ICQ */
189 static inline void ioc_exit_icqs(struct io_context *ioc)
192 static inline bool ioc_delay_free(struct io_context *ioc)
196 #endif /* CONFIG_BLK_ICQ */
199 * put_io_context - put a reference of io_context
200 * @ioc: io_context to put
202 * Decrement reference count of @ioc and release it if the count reaches
205 void put_io_context(struct io_context *ioc)
207 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
208 if (atomic_long_dec_and_test(&ioc->refcount) && !ioc_delay_free(ioc))
209 kmem_cache_free(iocontext_cachep, ioc);
211 EXPORT_SYMBOL_GPL(put_io_context);
213 /* Called by the exiting task */
214 void exit_io_context(struct task_struct *task)
216 struct io_context *ioc;
219 ioc = task->io_context;
220 task->io_context = NULL;
223 if (atomic_dec_and_test(&ioc->active_ref)) {
229 static struct io_context *alloc_io_context(gfp_t gfp_flags, int node)
231 struct io_context *ioc;
233 ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
238 atomic_long_set(&ioc->refcount, 1);
239 atomic_set(&ioc->active_ref, 1);
240 #ifdef CONFIG_BLK_ICQ
241 spin_lock_init(&ioc->lock);
242 INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC);
243 INIT_HLIST_HEAD(&ioc->icq_list);
244 INIT_WORK(&ioc->release_work, ioc_release_fn);
246 ioc->ioprio = IOPRIO_DEFAULT;
251 int set_task_ioprio(struct task_struct *task, int ioprio)
254 const struct cred *cred = current_cred(), *tcred;
257 tcred = __task_cred(task);
258 if (!uid_eq(tcred->uid, cred->euid) &&
259 !uid_eq(tcred->uid, cred->uid) && !capable(CAP_SYS_NICE)) {
265 err = security_task_setioprio(task, ioprio);
270 if (unlikely(!task->io_context)) {
271 struct io_context *ioc;
275 ioc = alloc_io_context(GFP_ATOMIC, NUMA_NO_NODE);
280 if (task->flags & PF_EXITING) {
281 kmem_cache_free(iocontext_cachep, ioc);
284 if (task->io_context)
285 kmem_cache_free(iocontext_cachep, ioc);
287 task->io_context = ioc;
289 task->io_context->ioprio = ioprio;
294 EXPORT_SYMBOL_GPL(set_task_ioprio);
296 int __copy_io(unsigned long clone_flags, struct task_struct *tsk)
298 struct io_context *ioc = current->io_context;
301 * Share io context with parent, if CLONE_IO is set
303 if (clone_flags & CLONE_IO) {
304 atomic_inc(&ioc->active_ref);
305 tsk->io_context = ioc;
306 } else if (ioprio_valid(ioc->ioprio)) {
307 tsk->io_context = alloc_io_context(GFP_KERNEL, NUMA_NO_NODE);
308 if (!tsk->io_context)
310 tsk->io_context->ioprio = ioc->ioprio;
316 #ifdef CONFIG_BLK_ICQ
318 * ioc_lookup_icq - lookup io_cq from ioc
319 * @q: the associated request_queue
321 * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called
322 * with @q->queue_lock held.
324 struct io_cq *ioc_lookup_icq(struct request_queue *q)
326 struct io_context *ioc = current->io_context;
329 lockdep_assert_held(&q->queue_lock);
332 * icq's are indexed from @ioc using radix tree and hint pointer,
333 * both of which are protected with RCU. All removals are done
334 * holding both q and ioc locks, and we're holding q lock - if we
335 * find a icq which points to us, it's guaranteed to be valid.
338 icq = rcu_dereference(ioc->icq_hint);
339 if (icq && icq->q == q)
342 icq = radix_tree_lookup(&ioc->icq_tree, q->id);
343 if (icq && icq->q == q)
344 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
351 EXPORT_SYMBOL(ioc_lookup_icq);
354 * ioc_create_icq - create and link io_cq
355 * @q: request_queue of interest
357 * Make sure io_cq linking @ioc and @q exists. If icq doesn't exist, they
358 * will be created using @gfp_mask.
360 * The caller is responsible for ensuring @ioc won't go away and @q is
361 * alive and will stay alive until this function returns.
363 static struct io_cq *ioc_create_icq(struct request_queue *q)
365 struct io_context *ioc = current->io_context;
366 struct elevator_type *et = q->elevator->type;
370 icq = kmem_cache_alloc_node(et->icq_cache, GFP_ATOMIC | __GFP_ZERO,
375 if (radix_tree_maybe_preload(GFP_ATOMIC) < 0) {
376 kmem_cache_free(et->icq_cache, icq);
382 INIT_LIST_HEAD(&icq->q_node);
383 INIT_HLIST_NODE(&icq->ioc_node);
385 /* lock both q and ioc and try to link @icq */
386 spin_lock_irq(&q->queue_lock);
387 spin_lock(&ioc->lock);
389 if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
390 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
391 list_add(&icq->q_node, &q->icq_list);
392 if (et->ops.init_icq)
393 et->ops.init_icq(icq);
395 kmem_cache_free(et->icq_cache, icq);
396 icq = ioc_lookup_icq(q);
398 printk(KERN_ERR "cfq: icq link failed!\n");
401 spin_unlock(&ioc->lock);
402 spin_unlock_irq(&q->queue_lock);
403 radix_tree_preload_end();
407 struct io_cq *ioc_find_get_icq(struct request_queue *q)
409 struct io_context *ioc = current->io_context;
410 struct io_cq *icq = NULL;
412 if (unlikely(!ioc)) {
413 ioc = alloc_io_context(GFP_ATOMIC, q->node);
418 if (current->io_context) {
419 kmem_cache_free(iocontext_cachep, ioc);
420 ioc = current->io_context;
422 current->io_context = ioc;
426 task_unlock(current);
430 spin_lock_irq(&q->queue_lock);
431 icq = ioc_lookup_icq(q);
432 spin_unlock_irq(&q->queue_lock);
436 icq = ioc_create_icq(q);
444 EXPORT_SYMBOL_GPL(ioc_find_get_icq);
445 #endif /* CONFIG_BLK_ICQ */
447 static int __init blk_ioc_init(void)
449 iocontext_cachep = kmem_cache_create("blkdev_ioc",
450 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
453 subsys_initcall(blk_ioc_init);