2 * QEMU aio implementation
4 * Copyright IBM, Corp. 2008
7 * Anthony Liguori <aliguori@us.ibm.com>
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
17 #ifdef CONFIG_LINUX_IO_URING
20 #include "qemu/coroutine-core.h"
21 #include "qemu/queue.h"
22 #include "qemu/event_notifier.h"
23 #include "qemu/thread.h"
24 #include "qemu/timer.h"
25 #include "block/graph-lock.h"
26 #include "hw/qdev-core.h"
29 typedef struct BlockAIOCB BlockAIOCB
;
30 typedef void BlockCompletionFunc(void *opaque
, int ret
);
32 typedef struct AIOCBInfo
{
33 void (*cancel_async
)(BlockAIOCB
*acb
);
38 const AIOCBInfo
*aiocb_info
;
40 BlockCompletionFunc
*cb
;
45 void *qemu_aio_get(const AIOCBInfo
*aiocb_info
, BlockDriverState
*bs
,
46 BlockCompletionFunc
*cb
, void *opaque
);
47 void qemu_aio_unref(void *p
);
48 void qemu_aio_ref(void *p
);
50 typedef struct AioHandler AioHandler
;
51 typedef QLIST_HEAD(, AioHandler
) AioHandlerList
;
52 typedef void QEMUBHFunc(void *opaque
);
53 typedef bool AioPollFn(void *opaque
);
54 typedef void IOHandler(void *opaque
);
60 /* Is polling disabled? */
61 bool aio_poll_disabled(AioContext
*ctx
);
63 /* Callbacks for file descriptor monitoring implementations */
67 * @ctx: the AioContext
68 * @old_node: the existing handler or NULL if this file descriptor is being
69 * monitored for the first time
70 * @new_node: the new handler or NULL if this file descriptor is being
73 * Add/remove/modify a monitored file descriptor.
75 * Called with ctx->list_lock acquired.
77 void (*update
)(AioContext
*ctx
, AioHandler
*old_node
, AioHandler
*new_node
);
81 * @ctx: the AioContext
82 * @ready_list: list for handlers that become ready
83 * @timeout: maximum duration to wait, in nanoseconds
85 * Wait for file descriptors to become ready and place them on ready_list.
87 * Called with ctx->list_lock incremented but not locked.
89 * Returns: number of ready file descriptors.
91 int (*wait
)(AioContext
*ctx
, AioHandlerList
*ready_list
, int64_t timeout
);
95 * @ctx: the AioContext
97 * Tell aio_poll() when to stop userspace polling early because ->wait()
100 * File descriptor monitoring implementations that cannot poll fd readiness
101 * from userspace should use aio_poll_disabled() here. This ensures that
102 * file descriptors are not starved by handlers that frequently make
103 * progress via userspace polling.
105 * Returns: true if ->wait() should be called, false otherwise.
107 bool (*need_wait
)(AioContext
*ctx
);
111 * Each aio_bh_poll() call carves off a slice of the BH list, so that newly
112 * scheduled BHs are not processed until the next aio_bh_poll() call. All
113 * active aio_bh_poll() calls chain their slices together in a list, so that
114 * nested aio_bh_poll() calls process all scheduled bottom halves.
116 typedef QSLIST_HEAD(, QEMUBH
) BHList
;
117 typedef struct BHListSlice BHListSlice
;
120 QSIMPLEQ_ENTRY(BHListSlice
) next
;
123 typedef QSLIST_HEAD(, AioHandler
) AioHandlerSList
;
128 /* Used by AioContext users to protect from multi-threaded access. */
132 * Keep track of readers and writers of the block layer graph.
133 * This is essential to avoid performing additions and removal
134 * of nodes and edges from block graph while some
135 * other thread is traversing it.
137 BdrvGraphRWlock
*bdrv_graph
;
139 /* The list of registered AIO handlers. Protected by ctx->list_lock. */
140 AioHandlerList aio_handlers
;
142 /* The list of AIO handlers to be deleted. Protected by ctx->list_lock. */
143 AioHandlerList deleted_aio_handlers
;
145 /* Used to avoid unnecessary event_notifier_set calls in aio_notify;
146 * only written from the AioContext home thread, or under the BQL in
147 * the case of the main AioContext. However, it is read from any
148 * thread so it is still accessed with atomic primitives.
150 * If this field is 0, everything (file descriptors, bottom halves,
151 * timers) will be re-evaluated before the next blocking poll() or
152 * io_uring wait; therefore, the event_notifier_set call can be
153 * skipped. If it is non-zero, you may need to wake up a concurrent
154 * aio_poll or the glib main event loop, making event_notifier_set
157 * Bit 0 is reserved for GSource usage of the AioContext, and is 1
158 * between a call to aio_ctx_prepare and the next call to aio_ctx_check.
159 * Bits 1-31 simply count the number of active calls to aio_poll
160 * that are in the prepare or poll phase.
162 * The GSource and aio_poll must use a different mechanism because
163 * there is no certainty that a call to GSource's prepare callback
164 * (via g_main_context_prepare) is indeed followed by check and
165 * dispatch. It's not clear whether this would be a bug, but let's
166 * play safe and allow it---it will just cause extra calls to
167 * event_notifier_set until the next call to dispatch.
169 * Instead, the aio_poll calls include both the prepare and the
170 * dispatch phase, hence a simple counter is enough for them.
174 /* A lock to protect between QEMUBH and AioHandler adders and deleter,
175 * and to ensure that no callbacks are removed while we're walking and
178 QemuLockCnt list_lock
;
180 /* Bottom Halves pending aio_bh_poll() processing */
183 /* Chained BH list slices for each nested aio_bh_poll() call */
184 QSIMPLEQ_HEAD(, BHListSlice
) bh_slice_list
;
186 /* Used by aio_notify.
188 * "notified" is used to avoid expensive event_notifier_test_and_clear
189 * calls. When it is clear, the EventNotifier is clear, or one thread
190 * is going to clear "notified" before processing more events. False
191 * positives are possible, i.e. "notified" could be set even though the
192 * EventNotifier is clear.
194 * Note that event_notifier_set *cannot* be optimized the same way. For
195 * more information on the problem that would result, see "#ifdef BUG2"
196 * in the docs/aio_notify_accept.promela formal model.
199 EventNotifier notifier
;
201 QSLIST_HEAD(, Coroutine
) scheduled_coroutines
;
202 QEMUBH
*co_schedule_bh
;
206 /* Thread pool for performing work and receiving completion callbacks.
207 * Has its own locking.
209 struct ThreadPool
*thread_pool
;
211 #ifdef CONFIG_LINUX_AIO
212 struct LinuxAioState
*linux_aio
;
214 #ifdef CONFIG_LINUX_IO_URING
215 struct LuringState
*linux_io_uring
;
217 /* State for file descriptor monitoring using Linux io_uring */
218 struct io_uring fdmon_io_uring
;
219 AioHandlerSList submit_list
;
222 /* TimerLists for calling timers - one per clock type. Has its own
225 QEMUTimerListGroup tlg
;
227 /* Number of AioHandlers without .io_poll() */
228 int poll_disable_cnt
;
230 /* Polling mode parameters */
231 int64_t poll_ns
; /* current polling time in nanoseconds */
232 int64_t poll_max_ns
; /* maximum polling time in nanoseconds */
233 int64_t poll_grow
; /* polling time growth factor */
234 int64_t poll_shrink
; /* polling time shrink factor */
236 /* AIO engine parameters */
237 int64_t aio_max_batch
; /* maximum number of requests in a batch */
240 * List of handlers participating in userspace polling. Protected by
241 * ctx->list_lock. Iterated and modified mostly by the event loop thread
242 * from aio_poll() with ctx->list_lock incremented. aio_set_fd_handler()
243 * only touches the list to delete nodes if ctx->list_lock's count is zero.
245 AioHandlerList poll_aio_handlers
;
247 /* Are we in polling mode or monitoring file descriptors? */
250 /* epoll(7) state used when built with CONFIG_EPOLL */
253 const FDMonOps
*fdmon_ops
;
257 * aio_context_new: Allocate a new AioContext.
259 * AioContext provide a mini event-loop that can be waited on synchronously.
260 * They also provide bottom halves, a service to execute a piece of code
261 * as soon as possible.
263 AioContext
*aio_context_new(Error
**errp
);
267 * @ctx: The AioContext to operate on.
269 * Add a reference to an AioContext.
271 void aio_context_ref(AioContext
*ctx
);
275 * @ctx: The AioContext to operate on.
277 * Drop a reference to an AioContext.
279 void aio_context_unref(AioContext
*ctx
);
281 /* Take ownership of the AioContext. If the AioContext will be shared between
282 * threads, and a thread does not want to be interrupted, it will have to
283 * take ownership around calls to aio_poll(). Otherwise, aio_poll()
284 * automatically takes care of calling aio_context_acquire and
285 * aio_context_release.
287 * Note that this is separate from bdrv_drained_begin/bdrv_drained_end. A
288 * thread still has to call those to avoid being interrupted by the guest.
290 * Bottom halves, timers and callbacks can be created or removed without
291 * acquiring the AioContext.
293 void aio_context_acquire(AioContext
*ctx
);
295 /* Relinquish ownership of the AioContext. */
296 void aio_context_release(AioContext
*ctx
);
299 * aio_bh_schedule_oneshot_full: Allocate a new bottom half structure that will
300 * run only once and as soon as possible.
302 * @name: A human-readable identifier for debugging purposes.
304 void aio_bh_schedule_oneshot_full(AioContext
*ctx
, QEMUBHFunc
*cb
, void *opaque
,
308 * aio_bh_schedule_oneshot: Allocate a new bottom half structure that will run
309 * only once and as soon as possible.
311 * A convenience wrapper for aio_bh_schedule_oneshot_full() that uses cb as the
314 #define aio_bh_schedule_oneshot(ctx, cb, opaque) \
315 aio_bh_schedule_oneshot_full((ctx), (cb), (opaque), (stringify(cb)))
318 * aio_bh_new_full: Allocate a new bottom half structure.
320 * Bottom halves are lightweight callbacks whose invocation is guaranteed
321 * to be wait-free, thread-safe and signal-safe. The #QEMUBH structure
322 * is opaque and must be allocated prior to its use.
324 * @name: A human-readable identifier for debugging purposes.
325 * @reentrancy_guard: A guard set when entering a cb to prevent
326 * device-reentrancy issues
328 QEMUBH
*aio_bh_new_full(AioContext
*ctx
, QEMUBHFunc
*cb
, void *opaque
,
329 const char *name
, MemReentrancyGuard
*reentrancy_guard
);
332 * aio_bh_new: Allocate a new bottom half structure
334 * A convenience wrapper for aio_bh_new_full() that uses the cb as the name
337 #define aio_bh_new(ctx, cb, opaque) \
338 aio_bh_new_full((ctx), (cb), (opaque), (stringify(cb)), NULL)
341 * aio_bh_new_guarded: Allocate a new bottom half structure with a
344 * A convenience wrapper for aio_bh_new_full() that uses the cb as the name
347 #define aio_bh_new_guarded(ctx, cb, opaque, guard) \
348 aio_bh_new_full((ctx), (cb), (opaque), (stringify(cb)), guard)
351 * aio_notify: Force processing of pending events.
353 * Similar to signaling a condition variable, aio_notify forces
354 * aio_poll to exit, so that the next call will re-examine pending events.
355 * The caller of aio_notify will usually call aio_poll again very soon,
356 * or go through another iteration of the GLib main loop. Hence, aio_notify
357 * also has the side effect of recalculating the sets of file descriptors
358 * that the main loop waits for.
360 * Calling aio_notify is rarely necessary, because for example scheduling
361 * a bottom half calls it already.
363 void aio_notify(AioContext
*ctx
);
366 * aio_notify_accept: Acknowledge receiving an aio_notify.
368 * aio_notify() uses an EventNotifier in order to wake up a sleeping
369 * aio_poll() or g_main_context_iteration(). Calls to aio_notify() are
370 * usually rare, but the AioContext has to clear the EventNotifier on
371 * every aio_poll() or g_main_context_iteration() in order to avoid
372 * busy waiting. This event_notifier_test_and_clear() cannot be done
373 * using the usual aio_context_set_event_notifier(), because it must
374 * be done before processing all events (file descriptors, bottom halves,
377 * aio_notify_accept() is an optimized event_notifier_test_and_clear()
378 * that is specific to an AioContext's notifier; it is used internally
379 * to clear the EventNotifier only if aio_notify() had been called.
381 void aio_notify_accept(AioContext
*ctx
);
384 * aio_bh_call: Executes callback function of the specified BH.
386 void aio_bh_call(QEMUBH
*bh
);
389 * aio_bh_poll: Poll bottom halves for an AioContext.
391 * These are internal functions used by the QEMU main loop.
392 * And notice that multiple occurrences of aio_bh_poll cannot
393 * be called concurrently
395 int aio_bh_poll(AioContext
*ctx
);
398 * qemu_bh_schedule: Schedule a bottom half.
400 * Scheduling a bottom half interrupts the main loop and causes the
401 * execution of the callback that was passed to qemu_bh_new.
403 * Bottom halves that are scheduled from a bottom half handler are instantly
404 * invoked. This can create an infinite loop if a bottom half handler
407 * @bh: The bottom half to be scheduled.
409 void qemu_bh_schedule(QEMUBH
*bh
);
412 * qemu_bh_cancel: Cancel execution of a bottom half.
414 * Canceling execution of a bottom half undoes the effect of calls to
415 * qemu_bh_schedule without freeing its resources yet. While cancellation
416 * itself is also wait-free and thread-safe, it can of course race with the
417 * loop that executes bottom halves unless you are holding the iothread
418 * mutex. This makes it mostly useless if you are not holding the mutex.
420 * @bh: The bottom half to be canceled.
422 void qemu_bh_cancel(QEMUBH
*bh
);
425 *qemu_bh_delete: Cancel execution of a bottom half and free its resources.
427 * Deleting a bottom half frees the memory that was allocated for it by
428 * qemu_bh_new. It also implies canceling the bottom half if it was
430 * This func is async. The bottom half will do the delete action at the finial
433 * @bh: The bottom half to be deleted.
435 void qemu_bh_delete(QEMUBH
*bh
);
437 /* Return whether there are any pending callbacks from the GSource
438 * attached to the AioContext, before g_poll is invoked.
440 * This is used internally in the implementation of the GSource.
442 bool aio_prepare(AioContext
*ctx
);
444 /* Return whether there are any pending callbacks from the GSource
445 * attached to the AioContext, after g_poll is invoked.
447 * This is used internally in the implementation of the GSource.
449 bool aio_pending(AioContext
*ctx
);
451 /* Dispatch any pending callbacks from the GSource attached to the AioContext.
453 * This is used internally in the implementation of the GSource.
455 void aio_dispatch(AioContext
*ctx
);
457 /* Progress in completing AIO work to occur. This can issue new pending
458 * aio as a result of executing I/O completion or bh callbacks.
460 * Return whether any progress was made by executing AIO or bottom half
461 * handlers. If @blocking == true, this should always be true except
462 * if someone called aio_notify.
464 * If there are no pending bottom halves, but there are pending AIO
465 * operations, it may not be possible to make any progress without
466 * blocking. If @blocking is true, this function will wait until one
467 * or more AIO events have completed, to ensure something has moved
470 bool no_coroutine_fn
aio_poll(AioContext
*ctx
, bool blocking
);
472 /* Register a file descriptor and associated callbacks. Behaves very similarly
473 * to qemu_set_fd_handler. Unlike qemu_set_fd_handler, these callbacks will
474 * be invoked when using aio_poll().
476 * Code that invokes AIO completion functions should rely on this function
477 * instead of qemu_set_fd_handler[2].
479 void aio_set_fd_handler(AioContext
*ctx
,
484 IOHandler
*io_poll_ready
,
487 /* Register an event notifier and associated callbacks. Behaves very similarly
488 * to event_notifier_set_handler. Unlike event_notifier_set_handler, these callbacks
489 * will be invoked when using aio_poll().
491 * Code that invokes AIO completion functions should rely on this function
492 * instead of event_notifier_set_handler.
494 void aio_set_event_notifier(AioContext
*ctx
,
495 EventNotifier
*notifier
,
496 EventNotifierHandler
*io_read
,
498 EventNotifierHandler
*io_poll_ready
);
500 /* Set polling begin/end callbacks for an event notifier that has already been
501 * registered with aio_set_event_notifier. Do nothing if the event notifier is
504 void aio_set_event_notifier_poll(AioContext
*ctx
,
505 EventNotifier
*notifier
,
506 EventNotifierHandler
*io_poll_begin
,
507 EventNotifierHandler
*io_poll_end
);
509 /* Return a GSource that lets the main loop poll the file descriptors attached
510 * to this AioContext.
512 GSource
*aio_get_g_source(AioContext
*ctx
);
514 /* Return the ThreadPool bound to this AioContext */
515 struct ThreadPool
*aio_get_thread_pool(AioContext
*ctx
);
517 /* Setup the LinuxAioState bound to this AioContext */
518 struct LinuxAioState
*aio_setup_linux_aio(AioContext
*ctx
, Error
**errp
);
520 /* Return the LinuxAioState bound to this AioContext */
521 struct LinuxAioState
*aio_get_linux_aio(AioContext
*ctx
);
523 /* Setup the LuringState bound to this AioContext */
524 struct LuringState
*aio_setup_linux_io_uring(AioContext
*ctx
, Error
**errp
);
526 /* Return the LuringState bound to this AioContext */
527 struct LuringState
*aio_get_linux_io_uring(AioContext
*ctx
);
529 * aio_timer_new_with_attrs:
530 * @ctx: the aio context
531 * @type: the clock type
533 * @attributes: 0, or one to multiple OR'ed QEMU_TIMER_ATTR_<id> values
535 * @cb: the callback to call on timer expiry
536 * @opaque: the opaque pointer to pass to the callback
538 * Allocate a new timer (with attributes) attached to the context @ctx.
539 * The function is responsible for memory allocation.
541 * The preferred interface is aio_timer_init or aio_timer_init_with_attrs.
542 * Use that unless you really need dynamic memory allocation.
544 * Returns: a pointer to the new timer
546 static inline QEMUTimer
*aio_timer_new_with_attrs(AioContext
*ctx
,
548 int scale
, int attributes
,
549 QEMUTimerCB
*cb
, void *opaque
)
551 return timer_new_full(&ctx
->tlg
, type
, scale
, attributes
, cb
, opaque
);
556 * @ctx: the aio context
557 * @type: the clock type
559 * @cb: the callback to call on timer expiry
560 * @opaque: the opaque pointer to pass to the callback
562 * Allocate a new timer attached to the context @ctx.
563 * See aio_timer_new_with_attrs for details.
565 * Returns: a pointer to the new timer
567 static inline QEMUTimer
*aio_timer_new(AioContext
*ctx
, QEMUClockType type
,
569 QEMUTimerCB
*cb
, void *opaque
)
571 return timer_new_full(&ctx
->tlg
, type
, scale
, 0, cb
, opaque
);
575 * aio_timer_init_with_attrs:
576 * @ctx: the aio context
578 * @type: the clock type
580 * @attributes: 0, or one to multiple OR'ed QEMU_TIMER_ATTR_<id> values
582 * @cb: the callback to call on timer expiry
583 * @opaque: the opaque pointer to pass to the callback
585 * Initialise a new timer (with attributes) attached to the context @ctx.
586 * The caller is responsible for memory allocation.
588 static inline void aio_timer_init_with_attrs(AioContext
*ctx
,
589 QEMUTimer
*ts
, QEMUClockType type
,
590 int scale
, int attributes
,
591 QEMUTimerCB
*cb
, void *opaque
)
593 timer_init_full(ts
, &ctx
->tlg
, type
, scale
, attributes
, cb
, opaque
);
598 * @ctx: the aio context
600 * @type: the clock type
602 * @cb: the callback to call on timer expiry
603 * @opaque: the opaque pointer to pass to the callback
605 * Initialise a new timer attached to the context @ctx.
606 * See aio_timer_init_with_attrs for details.
608 static inline void aio_timer_init(AioContext
*ctx
,
609 QEMUTimer
*ts
, QEMUClockType type
,
611 QEMUTimerCB
*cb
, void *opaque
)
613 timer_init_full(ts
, &ctx
->tlg
, type
, scale
, 0, cb
, opaque
);
617 * aio_compute_timeout:
618 * @ctx: the aio context
620 * Compute the timeout that a blocking aio_poll should use.
622 int64_t aio_compute_timeout(AioContext
*ctx
);
626 * @ctx: the aio context
629 * Start a coroutine on a remote AioContext.
631 * The coroutine must not be entered by anyone else while aio_co_schedule()
632 * is active. In addition the coroutine must have yielded unless ctx
633 * is the context in which the coroutine is running (i.e. the value of
634 * qemu_get_current_aio_context() from the coroutine itself).
636 void aio_co_schedule(AioContext
*ctx
, Coroutine
*co
);
639 * aio_co_reschedule_self:
640 * @new_ctx: the new context
642 * Move the currently running coroutine to new_ctx. If the coroutine is already
643 * running in new_ctx, do nothing.
645 void coroutine_fn
aio_co_reschedule_self(AioContext
*new_ctx
);
651 * Restart a coroutine on the AioContext where it was running last, thus
652 * preventing coroutines from jumping from one context to another when they
655 * aio_co_wake may be executed either in coroutine or non-coroutine
656 * context. The coroutine must not be entered by anyone else while
657 * aio_co_wake() is active.
659 void aio_co_wake(Coroutine
*co
);
663 * @ctx: the context to run the coroutine
664 * @co: the coroutine to run
666 * Enter a coroutine in the specified AioContext.
668 void aio_co_enter(AioContext
*ctx
, Coroutine
*co
);
671 * Return the AioContext whose event loop runs in the current thread.
673 * If called from an IOThread this will be the IOThread's AioContext. If
674 * called from the main thread or with the "big QEMU lock" taken it
675 * will be the main loop AioContext.
677 AioContext
*qemu_get_current_aio_context(void);
679 void qemu_set_current_aio_context(AioContext
*ctx
);
683 * @ctx: the aio context
685 * Initialize the aio context.
687 void aio_context_setup(AioContext
*ctx
);
690 * aio_context_destroy:
691 * @ctx: the aio context
693 * Destroy the aio context.
695 void aio_context_destroy(AioContext
*ctx
);
697 /* Used internally, do not call outside AioContext code */
698 void aio_context_use_g_source(AioContext
*ctx
);
701 * aio_context_set_poll_params:
702 * @ctx: the aio context
703 * @max_ns: how long to busy poll for, in nanoseconds
704 * @grow: polling time growth factor
705 * @shrink: polling time shrink factor
707 * Poll mode can be disabled by setting poll_max_ns to 0.
709 void aio_context_set_poll_params(AioContext
*ctx
, int64_t max_ns
,
710 int64_t grow
, int64_t shrink
,
714 * aio_context_set_aio_params:
715 * @ctx: the aio context
716 * @max_batch: maximum number of requests in a batch, 0 means that the
717 * engine will use its default
719 void aio_context_set_aio_params(AioContext
*ctx
, int64_t max_batch
,
723 * aio_context_set_thread_pool_params:
724 * @ctx: the aio context
725 * @min: min number of threads to have readily available in the thread pool
726 * @min: max number of threads the thread pool can contain
728 void aio_context_set_thread_pool_params(AioContext
*ctx
, int64_t min
,
729 int64_t max
, Error
**errp
);