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.h"
21 #include "qemu/queue.h"
22 #include "qemu/event_notifier.h"
23 #include "qemu/thread.h"
24 #include "qemu/timer.h"
26 typedef struct BlockAIOCB BlockAIOCB
;
27 typedef void BlockCompletionFunc(void *opaque
, int ret
);
29 typedef struct AIOCBInfo
{
30 void (*cancel_async
)(BlockAIOCB
*acb
);
31 AioContext
*(*get_aio_context
)(BlockAIOCB
*acb
);
36 const AIOCBInfo
*aiocb_info
;
38 BlockCompletionFunc
*cb
;
43 void *qemu_aio_get(const AIOCBInfo
*aiocb_info
, BlockDriverState
*bs
,
44 BlockCompletionFunc
*cb
, void *opaque
);
45 void qemu_aio_unref(void *p
);
46 void qemu_aio_ref(void *p
);
48 typedef struct AioHandler AioHandler
;
49 typedef QLIST_HEAD(, AioHandler
) AioHandlerList
;
50 typedef void QEMUBHFunc(void *opaque
);
51 typedef bool AioPollFn(void *opaque
);
52 typedef void IOHandler(void *opaque
);
59 /* Is polling disabled? */
60 bool aio_poll_disabled(AioContext
*ctx
);
62 /* Callbacks for file descriptor monitoring implementations */
66 * @ctx: the AioContext
67 * @old_node: the existing handler or NULL if this file descriptor is being
68 * monitored for the first time
69 * @new_node: the new handler or NULL if this file descriptor is being
72 * Add/remove/modify a monitored file descriptor.
74 * Called with ctx->list_lock acquired.
76 void (*update
)(AioContext
*ctx
, AioHandler
*old_node
, AioHandler
*new_node
);
80 * @ctx: the AioContext
81 * @ready_list: list for handlers that become ready
82 * @timeout: maximum duration to wait, in nanoseconds
84 * Wait for file descriptors to become ready and place them on ready_list.
86 * Called with ctx->list_lock incremented but not locked.
88 * Returns: number of ready file descriptors.
90 int (*wait
)(AioContext
*ctx
, AioHandlerList
*ready_list
, int64_t timeout
);
94 * @ctx: the AioContext
96 * Tell aio_poll() when to stop userspace polling early because ->wait()
99 * File descriptor monitoring implementations that cannot poll fd readiness
100 * from userspace should use aio_poll_disabled() here. This ensures that
101 * file descriptors are not starved by handlers that frequently make
102 * progress via userspace polling.
104 * Returns: true if ->wait() should be called, false otherwise.
106 bool (*need_wait
)(AioContext
*ctx
);
110 * Each aio_bh_poll() call carves off a slice of the BH list, so that newly
111 * scheduled BHs are not processed until the next aio_bh_poll() call. All
112 * active aio_bh_poll() calls chain their slices together in a list, so that
113 * nested aio_bh_poll() calls process all scheduled bottom halves.
115 typedef QSLIST_HEAD(, QEMUBH
) BHList
;
116 typedef struct BHListSlice BHListSlice
;
119 QSIMPLEQ_ENTRY(BHListSlice
) next
;
122 typedef QSLIST_HEAD(, AioHandler
) AioHandlerSList
;
127 /* Used by AioContext users to protect from multi-threaded access. */
130 /* The list of registered AIO handlers. Protected by ctx->list_lock. */
131 AioHandlerList aio_handlers
;
133 /* The list of AIO handlers to be deleted. Protected by ctx->list_lock. */
134 AioHandlerList deleted_aio_handlers
;
136 /* Used to avoid unnecessary event_notifier_set calls in aio_notify;
137 * only written from the AioContext home thread, or under the BQL in
138 * the case of the main AioContext. However, it is read from any
139 * thread so it is still accessed with atomic primitives.
141 * If this field is 0, everything (file descriptors, bottom halves,
142 * timers) will be re-evaluated before the next blocking poll() or
143 * io_uring wait; therefore, the event_notifier_set call can be
144 * skipped. If it is non-zero, you may need to wake up a concurrent
145 * aio_poll or the glib main event loop, making event_notifier_set
148 * Bit 0 is reserved for GSource usage of the AioContext, and is 1
149 * between a call to aio_ctx_prepare and the next call to aio_ctx_check.
150 * Bits 1-31 simply count the number of active calls to aio_poll
151 * that are in the prepare or poll phase.
153 * The GSource and aio_poll must use a different mechanism because
154 * there is no certainty that a call to GSource's prepare callback
155 * (via g_main_context_prepare) is indeed followed by check and
156 * dispatch. It's not clear whether this would be a bug, but let's
157 * play safe and allow it---it will just cause extra calls to
158 * event_notifier_set until the next call to dispatch.
160 * Instead, the aio_poll calls include both the prepare and the
161 * dispatch phase, hence a simple counter is enough for them.
165 /* A lock to protect between QEMUBH and AioHandler adders and deleter,
166 * and to ensure that no callbacks are removed while we're walking and
169 QemuLockCnt list_lock
;
171 /* Bottom Halves pending aio_bh_poll() processing */
174 /* Chained BH list slices for each nested aio_bh_poll() call */
175 QSIMPLEQ_HEAD(, BHListSlice
) bh_slice_list
;
177 /* Used by aio_notify.
179 * "notified" is used to avoid expensive event_notifier_test_and_clear
180 * calls. When it is clear, the EventNotifier is clear, or one thread
181 * is going to clear "notified" before processing more events. False
182 * positives are possible, i.e. "notified" could be set even though the
183 * EventNotifier is clear.
185 * Note that event_notifier_set *cannot* be optimized the same way. For
186 * more information on the problem that would result, see "#ifdef BUG2"
187 * in the docs/aio_notify_accept.promela formal model.
190 EventNotifier notifier
;
192 QSLIST_HEAD(, Coroutine
) scheduled_coroutines
;
193 QEMUBH
*co_schedule_bh
;
195 /* Thread pool for performing work and receiving completion callbacks.
196 * Has its own locking.
198 struct ThreadPool
*thread_pool
;
200 #ifdef CONFIG_LINUX_AIO
202 * State for native Linux AIO. Uses aio_context_acquire/release for
205 struct LinuxAioState
*linux_aio
;
207 #ifdef CONFIG_LINUX_IO_URING
209 * State for Linux io_uring. Uses aio_context_acquire/release for
212 struct LuringState
*linux_io_uring
;
214 /* State for file descriptor monitoring using Linux io_uring */
215 struct io_uring fdmon_io_uring
;
216 AioHandlerSList submit_list
;
219 /* TimerLists for calling timers - one per clock type. Has its own
222 QEMUTimerListGroup tlg
;
224 int external_disable_cnt
;
226 /* Number of AioHandlers without .io_poll() */
227 int poll_disable_cnt
;
229 /* Polling mode parameters */
230 int64_t poll_ns
; /* current polling time in nanoseconds */
231 int64_t poll_max_ns
; /* maximum polling time in nanoseconds */
232 int64_t poll_grow
; /* polling time growth factor */
233 int64_t poll_shrink
; /* polling time shrink factor */
235 /* AIO engine parameters */
236 int64_t aio_max_batch
; /* maximum number of requests in a batch */
239 * List of handlers participating in userspace polling. Protected by
240 * ctx->list_lock. Iterated and modified mostly by the event loop thread
241 * from aio_poll() with ctx->list_lock incremented. aio_set_fd_handler()
242 * only touches the list to delete nodes if ctx->list_lock's count is zero.
244 AioHandlerList poll_aio_handlers
;
246 /* Are we in polling mode or monitoring file descriptors? */
249 /* epoll(7) state used when built with CONFIG_EPOLL */
252 const FDMonOps
*fdmon_ops
;
256 * aio_context_new: Allocate a new AioContext.
258 * AioContext provide a mini event-loop that can be waited on synchronously.
259 * They also provide bottom halves, a service to execute a piece of code
260 * as soon as possible.
262 AioContext
*aio_context_new(Error
**errp
);
266 * @ctx: The AioContext to operate on.
268 * Add a reference to an AioContext.
270 void aio_context_ref(AioContext
*ctx
);
274 * @ctx: The AioContext to operate on.
276 * Drop a reference to an AioContext.
278 void aio_context_unref(AioContext
*ctx
);
280 /* Take ownership of the AioContext. If the AioContext will be shared between
281 * threads, and a thread does not want to be interrupted, it will have to
282 * take ownership around calls to aio_poll(). Otherwise, aio_poll()
283 * automatically takes care of calling aio_context_acquire and
284 * aio_context_release.
286 * Note that this is separate from bdrv_drained_begin/bdrv_drained_end. A
287 * thread still has to call those to avoid being interrupted by the guest.
289 * Bottom halves, timers and callbacks can be created or removed without
290 * acquiring the AioContext.
292 void aio_context_acquire(AioContext
*ctx
);
294 /* Relinquish ownership of the AioContext. */
295 void aio_context_release(AioContext
*ctx
);
298 * aio_bh_schedule_oneshot_full: Allocate a new bottom half structure that will
299 * run only once and as soon as possible.
301 * @name: A human-readable identifier for debugging purposes.
303 void aio_bh_schedule_oneshot_full(AioContext
*ctx
, QEMUBHFunc
*cb
, void *opaque
,
307 * aio_bh_schedule_oneshot: Allocate a new bottom half structure that will run
308 * only once and as soon as possible.
310 * A convenience wrapper for aio_bh_schedule_oneshot_full() that uses cb as the
313 #define aio_bh_schedule_oneshot(ctx, cb, opaque) \
314 aio_bh_schedule_oneshot_full((ctx), (cb), (opaque), (stringify(cb)))
317 * aio_bh_new_full: Allocate a new bottom half structure.
319 * Bottom halves are lightweight callbacks whose invocation is guaranteed
320 * to be wait-free, thread-safe and signal-safe. The #QEMUBH structure
321 * is opaque and must be allocated prior to its use.
323 * @name: A human-readable identifier for debugging purposes.
325 QEMUBH
*aio_bh_new_full(AioContext
*ctx
, QEMUBHFunc
*cb
, void *opaque
,
329 * aio_bh_new: Allocate a new bottom half structure
331 * A convenience wrapper for aio_bh_new_full() that uses the cb as the name
334 #define aio_bh_new(ctx, cb, opaque) \
335 aio_bh_new_full((ctx), (cb), (opaque), (stringify(cb)))
338 * aio_notify: Force processing of pending events.
340 * Similar to signaling a condition variable, aio_notify forces
341 * aio_poll to exit, so that the next call will re-examine pending events.
342 * The caller of aio_notify will usually call aio_poll again very soon,
343 * or go through another iteration of the GLib main loop. Hence, aio_notify
344 * also has the side effect of recalculating the sets of file descriptors
345 * that the main loop waits for.
347 * Calling aio_notify is rarely necessary, because for example scheduling
348 * a bottom half calls it already.
350 void aio_notify(AioContext
*ctx
);
353 * aio_notify_accept: Acknowledge receiving an aio_notify.
355 * aio_notify() uses an EventNotifier in order to wake up a sleeping
356 * aio_poll() or g_main_context_iteration(). Calls to aio_notify() are
357 * usually rare, but the AioContext has to clear the EventNotifier on
358 * every aio_poll() or g_main_context_iteration() in order to avoid
359 * busy waiting. This event_notifier_test_and_clear() cannot be done
360 * using the usual aio_context_set_event_notifier(), because it must
361 * be done before processing all events (file descriptors, bottom halves,
364 * aio_notify_accept() is an optimized event_notifier_test_and_clear()
365 * that is specific to an AioContext's notifier; it is used internally
366 * to clear the EventNotifier only if aio_notify() had been called.
368 void aio_notify_accept(AioContext
*ctx
);
371 * aio_bh_call: Executes callback function of the specified BH.
373 void aio_bh_call(QEMUBH
*bh
);
376 * aio_bh_poll: Poll bottom halves for an AioContext.
378 * These are internal functions used by the QEMU main loop.
379 * And notice that multiple occurrences of aio_bh_poll cannot
380 * be called concurrently
382 int aio_bh_poll(AioContext
*ctx
);
385 * qemu_bh_schedule: Schedule a bottom half.
387 * Scheduling a bottom half interrupts the main loop and causes the
388 * execution of the callback that was passed to qemu_bh_new.
390 * Bottom halves that are scheduled from a bottom half handler are instantly
391 * invoked. This can create an infinite loop if a bottom half handler
394 * @bh: The bottom half to be scheduled.
396 void qemu_bh_schedule(QEMUBH
*bh
);
399 * qemu_bh_cancel: Cancel execution of a bottom half.
401 * Canceling execution of a bottom half undoes the effect of calls to
402 * qemu_bh_schedule without freeing its resources yet. While cancellation
403 * itself is also wait-free and thread-safe, it can of course race with the
404 * loop that executes bottom halves unless you are holding the iothread
405 * mutex. This makes it mostly useless if you are not holding the mutex.
407 * @bh: The bottom half to be canceled.
409 void qemu_bh_cancel(QEMUBH
*bh
);
412 *qemu_bh_delete: Cancel execution of a bottom half and free its resources.
414 * Deleting a bottom half frees the memory that was allocated for it by
415 * qemu_bh_new. It also implies canceling the bottom half if it was
417 * This func is async. The bottom half will do the delete action at the finial
420 * @bh: The bottom half to be deleted.
422 void qemu_bh_delete(QEMUBH
*bh
);
424 /* Return whether there are any pending callbacks from the GSource
425 * attached to the AioContext, before g_poll is invoked.
427 * This is used internally in the implementation of the GSource.
429 bool aio_prepare(AioContext
*ctx
);
431 /* Return whether there are any pending callbacks from the GSource
432 * attached to the AioContext, after g_poll is invoked.
434 * This is used internally in the implementation of the GSource.
436 bool aio_pending(AioContext
*ctx
);
438 /* Dispatch any pending callbacks from the GSource attached to the AioContext.
440 * This is used internally in the implementation of the GSource.
442 void aio_dispatch(AioContext
*ctx
);
444 /* Progress in completing AIO work to occur. This can issue new pending
445 * aio as a result of executing I/O completion or bh callbacks.
447 * Return whether any progress was made by executing AIO or bottom half
448 * handlers. If @blocking == true, this should always be true except
449 * if someone called aio_notify.
451 * If there are no pending bottom halves, but there are pending AIO
452 * operations, it may not be possible to make any progress without
453 * blocking. If @blocking is true, this function will wait until one
454 * or more AIO events have completed, to ensure something has moved
457 bool aio_poll(AioContext
*ctx
, bool blocking
);
459 /* Register a file descriptor and associated callbacks. Behaves very similarly
460 * to qemu_set_fd_handler. Unlike qemu_set_fd_handler, these callbacks will
461 * be invoked when using aio_poll().
463 * Code that invokes AIO completion functions should rely on this function
464 * instead of qemu_set_fd_handler[2].
466 void aio_set_fd_handler(AioContext
*ctx
,
474 /* Set polling begin/end callbacks for a file descriptor that has already been
475 * registered with aio_set_fd_handler. Do nothing if the file descriptor is
478 void aio_set_fd_poll(AioContext
*ctx
, int fd
,
479 IOHandler
*io_poll_begin
,
480 IOHandler
*io_poll_end
);
482 /* Register an event notifier and associated callbacks. Behaves very similarly
483 * to event_notifier_set_handler. Unlike event_notifier_set_handler, these callbacks
484 * will be invoked when using aio_poll().
486 * Code that invokes AIO completion functions should rely on this function
487 * instead of event_notifier_set_handler.
489 void aio_set_event_notifier(AioContext
*ctx
,
490 EventNotifier
*notifier
,
492 EventNotifierHandler
*io_read
,
495 /* Set polling begin/end callbacks for an event notifier that has already been
496 * registered with aio_set_event_notifier. Do nothing if the event notifier is
499 void aio_set_event_notifier_poll(AioContext
*ctx
,
500 EventNotifier
*notifier
,
501 EventNotifierHandler
*io_poll_begin
,
502 EventNotifierHandler
*io_poll_end
);
504 /* Return a GSource that lets the main loop poll the file descriptors attached
505 * to this AioContext.
507 GSource
*aio_get_g_source(AioContext
*ctx
);
509 /* Return the ThreadPool bound to this AioContext */
510 struct ThreadPool
*aio_get_thread_pool(AioContext
*ctx
);
512 /* Setup the LinuxAioState bound to this AioContext */
513 struct LinuxAioState
*aio_setup_linux_aio(AioContext
*ctx
, Error
**errp
);
515 /* Return the LinuxAioState bound to this AioContext */
516 struct LinuxAioState
*aio_get_linux_aio(AioContext
*ctx
);
518 /* Setup the LuringState bound to this AioContext */
519 struct LuringState
*aio_setup_linux_io_uring(AioContext
*ctx
, Error
**errp
);
521 /* Return the LuringState bound to this AioContext */
522 struct LuringState
*aio_get_linux_io_uring(AioContext
*ctx
);
524 * aio_timer_new_with_attrs:
525 * @ctx: the aio context
526 * @type: the clock type
528 * @attributes: 0, or one to multiple OR'ed QEMU_TIMER_ATTR_<id> values
530 * @cb: the callback to call on timer expiry
531 * @opaque: the opaque pointer to pass to the callback
533 * Allocate a new timer (with attributes) attached to the context @ctx.
534 * The function is responsible for memory allocation.
536 * The preferred interface is aio_timer_init or aio_timer_init_with_attrs.
537 * Use that unless you really need dynamic memory allocation.
539 * Returns: a pointer to the new timer
541 static inline QEMUTimer
*aio_timer_new_with_attrs(AioContext
*ctx
,
543 int scale
, int attributes
,
544 QEMUTimerCB
*cb
, void *opaque
)
546 return timer_new_full(&ctx
->tlg
, type
, scale
, attributes
, cb
, opaque
);
551 * @ctx: the aio context
552 * @type: the clock type
554 * @cb: the callback to call on timer expiry
555 * @opaque: the opaque pointer to pass to the callback
557 * Allocate a new timer attached to the context @ctx.
558 * See aio_timer_new_with_attrs for details.
560 * Returns: a pointer to the new timer
562 static inline QEMUTimer
*aio_timer_new(AioContext
*ctx
, QEMUClockType type
,
564 QEMUTimerCB
*cb
, void *opaque
)
566 return timer_new_full(&ctx
->tlg
, type
, scale
, 0, cb
, opaque
);
570 * aio_timer_init_with_attrs:
571 * @ctx: the aio context
573 * @type: the clock type
575 * @attributes: 0, or one to multiple OR'ed QEMU_TIMER_ATTR_<id> values
577 * @cb: the callback to call on timer expiry
578 * @opaque: the opaque pointer to pass to the callback
580 * Initialise a new timer (with attributes) attached to the context @ctx.
581 * The caller is responsible for memory allocation.
583 static inline void aio_timer_init_with_attrs(AioContext
*ctx
,
584 QEMUTimer
*ts
, QEMUClockType type
,
585 int scale
, int attributes
,
586 QEMUTimerCB
*cb
, void *opaque
)
588 timer_init_full(ts
, &ctx
->tlg
, type
, scale
, attributes
, cb
, opaque
);
593 * @ctx: the aio context
595 * @type: the clock type
597 * @cb: the callback to call on timer expiry
598 * @opaque: the opaque pointer to pass to the callback
600 * Initialise a new timer attached to the context @ctx.
601 * See aio_timer_init_with_attrs for details.
603 static inline void aio_timer_init(AioContext
*ctx
,
604 QEMUTimer
*ts
, QEMUClockType type
,
606 QEMUTimerCB
*cb
, void *opaque
)
608 timer_init_full(ts
, &ctx
->tlg
, type
, scale
, 0, cb
, opaque
);
612 * aio_compute_timeout:
613 * @ctx: the aio context
615 * Compute the timeout that a blocking aio_poll should use.
617 int64_t aio_compute_timeout(AioContext
*ctx
);
620 * aio_disable_external:
621 * @ctx: the aio context
623 * Disable the further processing of external clients.
625 static inline void aio_disable_external(AioContext
*ctx
)
627 qatomic_inc(&ctx
->external_disable_cnt
);
631 * aio_enable_external:
632 * @ctx: the aio context
634 * Enable the processing of external clients.
636 static inline void aio_enable_external(AioContext
*ctx
)
640 old
= qatomic_fetch_dec(&ctx
->external_disable_cnt
);
643 /* Kick event loop so it re-arms file descriptors */
649 * aio_external_disabled:
650 * @ctx: the aio context
652 * Return true if the external clients are disabled.
654 static inline bool aio_external_disabled(AioContext
*ctx
)
656 return qatomic_read(&ctx
->external_disable_cnt
);
661 * @ctx: the aio context
662 * @is_external: Whether or not the checked node is an external event source.
664 * Check if the node's is_external flag is okay to be polled by the ctx at this
665 * moment. True means green light.
667 static inline bool aio_node_check(AioContext
*ctx
, bool is_external
)
669 return !is_external
|| !qatomic_read(&ctx
->external_disable_cnt
);
674 * @ctx: the aio context
677 * Start a coroutine on a remote AioContext.
679 * The coroutine must not be entered by anyone else while aio_co_schedule()
680 * is active. In addition the coroutine must have yielded unless ctx
681 * is the context in which the coroutine is running (i.e. the value of
682 * qemu_get_current_aio_context() from the coroutine itself).
684 void aio_co_schedule(AioContext
*ctx
, struct Coroutine
*co
);
687 * aio_co_reschedule_self:
688 * @new_ctx: the new context
690 * Move the currently running coroutine to new_ctx. If the coroutine is already
691 * running in new_ctx, do nothing.
693 void coroutine_fn
aio_co_reschedule_self(AioContext
*new_ctx
);
699 * Restart a coroutine on the AioContext where it was running last, thus
700 * preventing coroutines from jumping from one context to another when they
703 * aio_co_wake may be executed either in coroutine or non-coroutine
704 * context. The coroutine must not be entered by anyone else while
705 * aio_co_wake() is active.
707 void aio_co_wake(struct Coroutine
*co
);
711 * @ctx: the context to run the coroutine
712 * @co: the coroutine to run
714 * Enter a coroutine in the specified AioContext.
716 void aio_co_enter(AioContext
*ctx
, struct Coroutine
*co
);
719 * Return the AioContext whose event loop runs in the current thread.
721 * If called from an IOThread this will be the IOThread's AioContext. If
722 * called from the main thread or with the "big QEMU lock" taken it
723 * will be the main loop AioContext.
725 AioContext
*qemu_get_current_aio_context(void);
727 void qemu_set_current_aio_context(AioContext
*ctx
);
731 * @ctx: the aio context
733 * Initialize the aio context.
735 void aio_context_setup(AioContext
*ctx
);
738 * aio_context_destroy:
739 * @ctx: the aio context
741 * Destroy the aio context.
743 void aio_context_destroy(AioContext
*ctx
);
745 /* Used internally, do not call outside AioContext code */
746 void aio_context_use_g_source(AioContext
*ctx
);
749 * aio_context_set_poll_params:
750 * @ctx: the aio context
751 * @max_ns: how long to busy poll for, in nanoseconds
752 * @grow: polling time growth factor
753 * @shrink: polling time shrink factor
755 * Poll mode can be disabled by setting poll_max_ns to 0.
757 void aio_context_set_poll_params(AioContext
*ctx
, int64_t max_ns
,
758 int64_t grow
, int64_t shrink
,
762 * aio_context_set_aio_params:
763 * @ctx: the aio context
764 * @max_batch: maximum number of requests in a batch, 0 means that the
765 * engine will use its default
767 void aio_context_set_aio_params(AioContext
*ctx
, int64_t max_batch
,