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
);
34 AioContext
*(*get_aio_context
)(BlockAIOCB
*acb
);
39 const AIOCBInfo
*aiocb_info
;
41 BlockCompletionFunc
*cb
;
46 void *qemu_aio_get(const AIOCBInfo
*aiocb_info
, BlockDriverState
*bs
,
47 BlockCompletionFunc
*cb
, void *opaque
);
48 void qemu_aio_unref(void *p
);
49 void qemu_aio_ref(void *p
);
51 typedef struct AioHandler AioHandler
;
52 typedef QLIST_HEAD(, AioHandler
) AioHandlerList
;
53 typedef void QEMUBHFunc(void *opaque
);
54 typedef bool AioPollFn(void *opaque
);
55 typedef void IOHandler(void *opaque
);
61 /* Is polling disabled? */
62 bool aio_poll_disabled(AioContext
*ctx
);
64 /* Callbacks for file descriptor monitoring implementations */
68 * @ctx: the AioContext
69 * @old_node: the existing handler or NULL if this file descriptor is being
70 * monitored for the first time
71 * @new_node: the new handler or NULL if this file descriptor is being
74 * Add/remove/modify a monitored file descriptor.
76 * Called with ctx->list_lock acquired.
78 void (*update
)(AioContext
*ctx
, AioHandler
*old_node
, AioHandler
*new_node
);
82 * @ctx: the AioContext
83 * @ready_list: list for handlers that become ready
84 * @timeout: maximum duration to wait, in nanoseconds
86 * Wait for file descriptors to become ready and place them on ready_list.
88 * Called with ctx->list_lock incremented but not locked.
90 * Returns: number of ready file descriptors.
92 int (*wait
)(AioContext
*ctx
, AioHandlerList
*ready_list
, int64_t timeout
);
96 * @ctx: the AioContext
98 * Tell aio_poll() when to stop userspace polling early because ->wait()
101 * File descriptor monitoring implementations that cannot poll fd readiness
102 * from userspace should use aio_poll_disabled() here. This ensures that
103 * file descriptors are not starved by handlers that frequently make
104 * progress via userspace polling.
106 * Returns: true if ->wait() should be called, false otherwise.
108 bool (*need_wait
)(AioContext
*ctx
);
112 * Each aio_bh_poll() call carves off a slice of the BH list, so that newly
113 * scheduled BHs are not processed until the next aio_bh_poll() call. All
114 * active aio_bh_poll() calls chain their slices together in a list, so that
115 * nested aio_bh_poll() calls process all scheduled bottom halves.
117 typedef QSLIST_HEAD(, QEMUBH
) BHList
;
118 typedef struct BHListSlice BHListSlice
;
121 QSIMPLEQ_ENTRY(BHListSlice
) next
;
124 typedef QSLIST_HEAD(, AioHandler
) AioHandlerSList
;
129 /* Used by AioContext users to protect from multi-threaded access. */
133 * Keep track of readers and writers of the block layer graph.
134 * This is essential to avoid performing additions and removal
135 * of nodes and edges from block graph while some
136 * other thread is traversing it.
138 BdrvGraphRWlock
*bdrv_graph
;
140 /* The list of registered AIO handlers. Protected by ctx->list_lock. */
141 AioHandlerList aio_handlers
;
143 /* The list of AIO handlers to be deleted. Protected by ctx->list_lock. */
144 AioHandlerList deleted_aio_handlers
;
146 /* Used to avoid unnecessary event_notifier_set calls in aio_notify;
147 * only written from the AioContext home thread, or under the BQL in
148 * the case of the main AioContext. However, it is read from any
149 * thread so it is still accessed with atomic primitives.
151 * If this field is 0, everything (file descriptors, bottom halves,
152 * timers) will be re-evaluated before the next blocking poll() or
153 * io_uring wait; therefore, the event_notifier_set call can be
154 * skipped. If it is non-zero, you may need to wake up a concurrent
155 * aio_poll or the glib main event loop, making event_notifier_set
158 * Bit 0 is reserved for GSource usage of the AioContext, and is 1
159 * between a call to aio_ctx_prepare and the next call to aio_ctx_check.
160 * Bits 1-31 simply count the number of active calls to aio_poll
161 * that are in the prepare or poll phase.
163 * The GSource and aio_poll must use a different mechanism because
164 * there is no certainty that a call to GSource's prepare callback
165 * (via g_main_context_prepare) is indeed followed by check and
166 * dispatch. It's not clear whether this would be a bug, but let's
167 * play safe and allow it---it will just cause extra calls to
168 * event_notifier_set until the next call to dispatch.
170 * Instead, the aio_poll calls include both the prepare and the
171 * dispatch phase, hence a simple counter is enough for them.
175 /* A lock to protect between QEMUBH and AioHandler adders and deleter,
176 * and to ensure that no callbacks are removed while we're walking and
179 QemuLockCnt list_lock
;
181 /* Bottom Halves pending aio_bh_poll() processing */
184 /* Chained BH list slices for each nested aio_bh_poll() call */
185 QSIMPLEQ_HEAD(, BHListSlice
) bh_slice_list
;
187 /* Used by aio_notify.
189 * "notified" is used to avoid expensive event_notifier_test_and_clear
190 * calls. When it is clear, the EventNotifier is clear, or one thread
191 * is going to clear "notified" before processing more events. False
192 * positives are possible, i.e. "notified" could be set even though the
193 * EventNotifier is clear.
195 * Note that event_notifier_set *cannot* be optimized the same way. For
196 * more information on the problem that would result, see "#ifdef BUG2"
197 * in the docs/aio_notify_accept.promela formal model.
200 EventNotifier notifier
;
202 QSLIST_HEAD(, Coroutine
) scheduled_coroutines
;
203 QEMUBH
*co_schedule_bh
;
207 /* Thread pool for performing work and receiving completion callbacks.
208 * Has its own locking.
210 struct ThreadPool
*thread_pool
;
212 #ifdef CONFIG_LINUX_AIO
213 struct LinuxAioState
*linux_aio
;
215 #ifdef CONFIG_LINUX_IO_URING
216 struct LuringState
*linux_io_uring
;
218 /* State for file descriptor monitoring using Linux io_uring */
219 struct io_uring fdmon_io_uring
;
220 AioHandlerSList submit_list
;
223 /* TimerLists for calling timers - one per clock type. Has its own
226 QEMUTimerListGroup tlg
;
228 /* Number of AioHandlers without .io_poll() */
229 int poll_disable_cnt
;
231 /* Polling mode parameters */
232 int64_t poll_ns
; /* current polling time in nanoseconds */
233 int64_t poll_max_ns
; /* maximum polling time in nanoseconds */
234 int64_t poll_grow
; /* polling time growth factor */
235 int64_t poll_shrink
; /* polling time shrink factor */
237 /* AIO engine parameters */
238 int64_t aio_max_batch
; /* maximum number of requests in a batch */
241 * List of handlers participating in userspace polling. Protected by
242 * ctx->list_lock. Iterated and modified mostly by the event loop thread
243 * from aio_poll() with ctx->list_lock incremented. aio_set_fd_handler()
244 * only touches the list to delete nodes if ctx->list_lock's count is zero.
246 AioHandlerList poll_aio_handlers
;
248 /* Are we in polling mode or monitoring file descriptors? */
251 /* epoll(7) state used when built with CONFIG_EPOLL */
254 const FDMonOps
*fdmon_ops
;
258 * aio_context_new: Allocate a new AioContext.
260 * AioContext provide a mini event-loop that can be waited on synchronously.
261 * They also provide bottom halves, a service to execute a piece of code
262 * as soon as possible.
264 AioContext
*aio_context_new(Error
**errp
);
268 * @ctx: The AioContext to operate on.
270 * Add a reference to an AioContext.
272 void aio_context_ref(AioContext
*ctx
);
276 * @ctx: The AioContext to operate on.
278 * Drop a reference to an AioContext.
280 void aio_context_unref(AioContext
*ctx
);
282 /* Take ownership of the AioContext. If the AioContext will be shared between
283 * threads, and a thread does not want to be interrupted, it will have to
284 * take ownership around calls to aio_poll(). Otherwise, aio_poll()
285 * automatically takes care of calling aio_context_acquire and
286 * aio_context_release.
288 * Note that this is separate from bdrv_drained_begin/bdrv_drained_end. A
289 * thread still has to call those to avoid being interrupted by the guest.
291 * Bottom halves, timers and callbacks can be created or removed without
292 * acquiring the AioContext.
294 void aio_context_acquire(AioContext
*ctx
);
296 /* Relinquish ownership of the AioContext. */
297 void aio_context_release(AioContext
*ctx
);
300 * aio_bh_schedule_oneshot_full: Allocate a new bottom half structure that will
301 * run only once and as soon as possible.
303 * @name: A human-readable identifier for debugging purposes.
305 void aio_bh_schedule_oneshot_full(AioContext
*ctx
, QEMUBHFunc
*cb
, void *opaque
,
309 * aio_bh_schedule_oneshot: Allocate a new bottom half structure that will run
310 * only once and as soon as possible.
312 * A convenience wrapper for aio_bh_schedule_oneshot_full() that uses cb as the
315 #define aio_bh_schedule_oneshot(ctx, cb, opaque) \
316 aio_bh_schedule_oneshot_full((ctx), (cb), (opaque), (stringify(cb)))
319 * aio_bh_new_full: Allocate a new bottom half structure.
321 * Bottom halves are lightweight callbacks whose invocation is guaranteed
322 * to be wait-free, thread-safe and signal-safe. The #QEMUBH structure
323 * is opaque and must be allocated prior to its use.
325 * @name: A human-readable identifier for debugging purposes.
326 * @reentrancy_guard: A guard set when entering a cb to prevent
327 * device-reentrancy issues
329 QEMUBH
*aio_bh_new_full(AioContext
*ctx
, QEMUBHFunc
*cb
, void *opaque
,
330 const char *name
, MemReentrancyGuard
*reentrancy_guard
);
333 * aio_bh_new: Allocate a new bottom half structure
335 * A convenience wrapper for aio_bh_new_full() that uses the cb as the name
338 #define aio_bh_new(ctx, cb, opaque) \
339 aio_bh_new_full((ctx), (cb), (opaque), (stringify(cb)), NULL)
342 * aio_bh_new_guarded: Allocate a new bottom half structure with a
345 * A convenience wrapper for aio_bh_new_full() that uses the cb as the name
348 #define aio_bh_new_guarded(ctx, cb, opaque, guard) \
349 aio_bh_new_full((ctx), (cb), (opaque), (stringify(cb)), guard)
352 * aio_notify: Force processing of pending events.
354 * Similar to signaling a condition variable, aio_notify forces
355 * aio_poll to exit, so that the next call will re-examine pending events.
356 * The caller of aio_notify will usually call aio_poll again very soon,
357 * or go through another iteration of the GLib main loop. Hence, aio_notify
358 * also has the side effect of recalculating the sets of file descriptors
359 * that the main loop waits for.
361 * Calling aio_notify is rarely necessary, because for example scheduling
362 * a bottom half calls it already.
364 void aio_notify(AioContext
*ctx
);
367 * aio_notify_accept: Acknowledge receiving an aio_notify.
369 * aio_notify() uses an EventNotifier in order to wake up a sleeping
370 * aio_poll() or g_main_context_iteration(). Calls to aio_notify() are
371 * usually rare, but the AioContext has to clear the EventNotifier on
372 * every aio_poll() or g_main_context_iteration() in order to avoid
373 * busy waiting. This event_notifier_test_and_clear() cannot be done
374 * using the usual aio_context_set_event_notifier(), because it must
375 * be done before processing all events (file descriptors, bottom halves,
378 * aio_notify_accept() is an optimized event_notifier_test_and_clear()
379 * that is specific to an AioContext's notifier; it is used internally
380 * to clear the EventNotifier only if aio_notify() had been called.
382 void aio_notify_accept(AioContext
*ctx
);
385 * aio_bh_call: Executes callback function of the specified BH.
387 void aio_bh_call(QEMUBH
*bh
);
390 * aio_bh_poll: Poll bottom halves for an AioContext.
392 * These are internal functions used by the QEMU main loop.
393 * And notice that multiple occurrences of aio_bh_poll cannot
394 * be called concurrently
396 int aio_bh_poll(AioContext
*ctx
);
399 * qemu_bh_schedule: Schedule a bottom half.
401 * Scheduling a bottom half interrupts the main loop and causes the
402 * execution of the callback that was passed to qemu_bh_new.
404 * Bottom halves that are scheduled from a bottom half handler are instantly
405 * invoked. This can create an infinite loop if a bottom half handler
408 * @bh: The bottom half to be scheduled.
410 void qemu_bh_schedule(QEMUBH
*bh
);
413 * qemu_bh_cancel: Cancel execution of a bottom half.
415 * Canceling execution of a bottom half undoes the effect of calls to
416 * qemu_bh_schedule without freeing its resources yet. While cancellation
417 * itself is also wait-free and thread-safe, it can of course race with the
418 * loop that executes bottom halves unless you are holding the iothread
419 * mutex. This makes it mostly useless if you are not holding the mutex.
421 * @bh: The bottom half to be canceled.
423 void qemu_bh_cancel(QEMUBH
*bh
);
426 *qemu_bh_delete: Cancel execution of a bottom half and free its resources.
428 * Deleting a bottom half frees the memory that was allocated for it by
429 * qemu_bh_new. It also implies canceling the bottom half if it was
431 * This func is async. The bottom half will do the delete action at the finial
434 * @bh: The bottom half to be deleted.
436 void qemu_bh_delete(QEMUBH
*bh
);
438 /* Return whether there are any pending callbacks from the GSource
439 * attached to the AioContext, before g_poll is invoked.
441 * This is used internally in the implementation of the GSource.
443 bool aio_prepare(AioContext
*ctx
);
445 /* Return whether there are any pending callbacks from the GSource
446 * attached to the AioContext, after g_poll is invoked.
448 * This is used internally in the implementation of the GSource.
450 bool aio_pending(AioContext
*ctx
);
452 /* Dispatch any pending callbacks from the GSource attached to the AioContext.
454 * This is used internally in the implementation of the GSource.
456 void aio_dispatch(AioContext
*ctx
);
458 /* Progress in completing AIO work to occur. This can issue new pending
459 * aio as a result of executing I/O completion or bh callbacks.
461 * Return whether any progress was made by executing AIO or bottom half
462 * handlers. If @blocking == true, this should always be true except
463 * if someone called aio_notify.
465 * If there are no pending bottom halves, but there are pending AIO
466 * operations, it may not be possible to make any progress without
467 * blocking. If @blocking is true, this function will wait until one
468 * or more AIO events have completed, to ensure something has moved
471 bool aio_poll(AioContext
*ctx
, bool blocking
);
473 /* Register a file descriptor and associated callbacks. Behaves very similarly
474 * to qemu_set_fd_handler. Unlike qemu_set_fd_handler, these callbacks will
475 * be invoked when using aio_poll().
477 * Code that invokes AIO completion functions should rely on this function
478 * instead of qemu_set_fd_handler[2].
480 void aio_set_fd_handler(AioContext
*ctx
,
485 IOHandler
*io_poll_ready
,
488 /* Register an event notifier and associated callbacks. Behaves very similarly
489 * to event_notifier_set_handler. Unlike event_notifier_set_handler, these callbacks
490 * will be invoked when using aio_poll().
492 * Code that invokes AIO completion functions should rely on this function
493 * instead of event_notifier_set_handler.
495 void aio_set_event_notifier(AioContext
*ctx
,
496 EventNotifier
*notifier
,
497 EventNotifierHandler
*io_read
,
499 EventNotifierHandler
*io_poll_ready
);
501 /* Set polling begin/end callbacks for an event notifier that has already been
502 * registered with aio_set_event_notifier. Do nothing if the event notifier is
505 void aio_set_event_notifier_poll(AioContext
*ctx
,
506 EventNotifier
*notifier
,
507 EventNotifierHandler
*io_poll_begin
,
508 EventNotifierHandler
*io_poll_end
);
510 /* Return a GSource that lets the main loop poll the file descriptors attached
511 * to this AioContext.
513 GSource
*aio_get_g_source(AioContext
*ctx
);
515 /* Return the ThreadPool bound to this AioContext */
516 struct ThreadPool
*aio_get_thread_pool(AioContext
*ctx
);
518 /* Setup the LinuxAioState bound to this AioContext */
519 struct LinuxAioState
*aio_setup_linux_aio(AioContext
*ctx
, Error
**errp
);
521 /* Return the LinuxAioState bound to this AioContext */
522 struct LinuxAioState
*aio_get_linux_aio(AioContext
*ctx
);
524 /* Setup the LuringState bound to this AioContext */
525 struct LuringState
*aio_setup_linux_io_uring(AioContext
*ctx
, Error
**errp
);
527 /* Return the LuringState bound to this AioContext */
528 struct LuringState
*aio_get_linux_io_uring(AioContext
*ctx
);
530 * aio_timer_new_with_attrs:
531 * @ctx: the aio context
532 * @type: the clock type
534 * @attributes: 0, or one to multiple OR'ed QEMU_TIMER_ATTR_<id> values
536 * @cb: the callback to call on timer expiry
537 * @opaque: the opaque pointer to pass to the callback
539 * Allocate a new timer (with attributes) attached to the context @ctx.
540 * The function is responsible for memory allocation.
542 * The preferred interface is aio_timer_init or aio_timer_init_with_attrs.
543 * Use that unless you really need dynamic memory allocation.
545 * Returns: a pointer to the new timer
547 static inline QEMUTimer
*aio_timer_new_with_attrs(AioContext
*ctx
,
549 int scale
, int attributes
,
550 QEMUTimerCB
*cb
, void *opaque
)
552 return timer_new_full(&ctx
->tlg
, type
, scale
, attributes
, cb
, opaque
);
557 * @ctx: the aio context
558 * @type: the clock type
560 * @cb: the callback to call on timer expiry
561 * @opaque: the opaque pointer to pass to the callback
563 * Allocate a new timer attached to the context @ctx.
564 * See aio_timer_new_with_attrs for details.
566 * Returns: a pointer to the new timer
568 static inline QEMUTimer
*aio_timer_new(AioContext
*ctx
, QEMUClockType type
,
570 QEMUTimerCB
*cb
, void *opaque
)
572 return timer_new_full(&ctx
->tlg
, type
, scale
, 0, cb
, opaque
);
576 * aio_timer_init_with_attrs:
577 * @ctx: the aio context
579 * @type: the clock type
581 * @attributes: 0, or one to multiple OR'ed QEMU_TIMER_ATTR_<id> values
583 * @cb: the callback to call on timer expiry
584 * @opaque: the opaque pointer to pass to the callback
586 * Initialise a new timer (with attributes) attached to the context @ctx.
587 * The caller is responsible for memory allocation.
589 static inline void aio_timer_init_with_attrs(AioContext
*ctx
,
590 QEMUTimer
*ts
, QEMUClockType type
,
591 int scale
, int attributes
,
592 QEMUTimerCB
*cb
, void *opaque
)
594 timer_init_full(ts
, &ctx
->tlg
, type
, scale
, attributes
, cb
, opaque
);
599 * @ctx: the aio context
601 * @type: the clock type
603 * @cb: the callback to call on timer expiry
604 * @opaque: the opaque pointer to pass to the callback
606 * Initialise a new timer attached to the context @ctx.
607 * See aio_timer_init_with_attrs for details.
609 static inline void aio_timer_init(AioContext
*ctx
,
610 QEMUTimer
*ts
, QEMUClockType type
,
612 QEMUTimerCB
*cb
, void *opaque
)
614 timer_init_full(ts
, &ctx
->tlg
, type
, scale
, 0, cb
, opaque
);
618 * aio_compute_timeout:
619 * @ctx: the aio context
621 * Compute the timeout that a blocking aio_poll should use.
623 int64_t aio_compute_timeout(AioContext
*ctx
);
627 * @ctx: the aio context
630 * Start a coroutine on a remote AioContext.
632 * The coroutine must not be entered by anyone else while aio_co_schedule()
633 * is active. In addition the coroutine must have yielded unless ctx
634 * is the context in which the coroutine is running (i.e. the value of
635 * qemu_get_current_aio_context() from the coroutine itself).
637 void aio_co_schedule(AioContext
*ctx
, Coroutine
*co
);
640 * aio_co_reschedule_self:
641 * @new_ctx: the new context
643 * Move the currently running coroutine to new_ctx. If the coroutine is already
644 * running in new_ctx, do nothing.
646 void coroutine_fn
aio_co_reschedule_self(AioContext
*new_ctx
);
652 * Restart a coroutine on the AioContext where it was running last, thus
653 * preventing coroutines from jumping from one context to another when they
656 * aio_co_wake may be executed either in coroutine or non-coroutine
657 * context. The coroutine must not be entered by anyone else while
658 * aio_co_wake() is active.
660 void aio_co_wake(Coroutine
*co
);
664 * @ctx: the context to run the coroutine
665 * @co: the coroutine to run
667 * Enter a coroutine in the specified AioContext.
669 void aio_co_enter(AioContext
*ctx
, Coroutine
*co
);
672 * Return the AioContext whose event loop runs in the current thread.
674 * If called from an IOThread this will be the IOThread's AioContext. If
675 * called from the main thread or with the "big QEMU lock" taken it
676 * will be the main loop AioContext.
678 AioContext
*qemu_get_current_aio_context(void);
680 void qemu_set_current_aio_context(AioContext
*ctx
);
684 * @ctx: the aio context
686 * Initialize the aio context.
688 void aio_context_setup(AioContext
*ctx
);
691 * aio_context_destroy:
692 * @ctx: the aio context
694 * Destroy the aio context.
696 void aio_context_destroy(AioContext
*ctx
);
698 /* Used internally, do not call outside AioContext code */
699 void aio_context_use_g_source(AioContext
*ctx
);
702 * aio_context_set_poll_params:
703 * @ctx: the aio context
704 * @max_ns: how long to busy poll for, in nanoseconds
705 * @grow: polling time growth factor
706 * @shrink: polling time shrink factor
708 * Poll mode can be disabled by setting poll_max_ns to 0.
710 void aio_context_set_poll_params(AioContext
*ctx
, int64_t max_ns
,
711 int64_t grow
, int64_t shrink
,
715 * aio_context_set_aio_params:
716 * @ctx: the aio context
717 * @max_batch: maximum number of requests in a batch, 0 means that the
718 * engine will use its default
720 void aio_context_set_aio_params(AioContext
*ctx
, int64_t max_batch
,
724 * aio_context_set_thread_pool_params:
725 * @ctx: the aio context
726 * @min: min number of threads to have readily available in the thread pool
727 * @min: max number of threads the thread pool can contain
729 void aio_context_set_thread_pool_params(AioContext
*ctx
, int64_t min
,
730 int64_t max
, Error
**errp
);