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 #include "qemu-common.h"
18 #include "qemu/queue.h"
19 #include "qemu/event_notifier.h"
20 #include "qemu/thread.h"
21 #include "qemu/rfifolock.h"
22 #include "qemu/timer.h"
24 typedef struct BlockAIOCB BlockAIOCB
;
25 typedef void BlockCompletionFunc(void *opaque
, int ret
);
27 typedef struct AIOCBInfo
{
28 void (*cancel_async
)(BlockAIOCB
*acb
);
29 AioContext
*(*get_aio_context
)(BlockAIOCB
*acb
);
34 const AIOCBInfo
*aiocb_info
;
36 BlockCompletionFunc
*cb
;
41 void *qemu_aio_get(const AIOCBInfo
*aiocb_info
, BlockDriverState
*bs
,
42 BlockCompletionFunc
*cb
, void *opaque
);
43 void qemu_aio_unref(void *p
);
44 void qemu_aio_ref(void *p
);
46 typedef struct AioHandler AioHandler
;
47 typedef void QEMUBHFunc(void *opaque
);
48 typedef void IOHandler(void *opaque
);
56 /* Protects all fields from multi-threaded access */
59 /* The list of registered AIO handlers */
60 QLIST_HEAD(, AioHandler
) aio_handlers
;
62 /* This is a simple lock used to protect the aio_handlers list.
63 * Specifically, it's used to ensure that no callbacks are removed while
64 * we're walking and dispatching callbacks.
68 /* Used to avoid unnecessary event_notifier_set calls in aio_notify;
69 * accessed with atomic primitives. If this field is 0, everything
70 * (file descriptors, bottom halves, timers) will be re-evaluated
71 * before the next blocking poll(), thus the event_notifier_set call
72 * can be skipped. If it is non-zero, you may need to wake up a
73 * concurrent aio_poll or the glib main event loop, making
74 * event_notifier_set necessary.
76 * Bit 0 is reserved for GSource usage of the AioContext, and is 1
77 * between a call to aio_ctx_prepare and the next call to aio_ctx_check.
78 * Bits 1-31 simply count the number of active calls to aio_poll
79 * that are in the prepare or poll phase.
81 * The GSource and aio_poll must use a different mechanism because
82 * there is no certainty that a call to GSource's prepare callback
83 * (via g_main_context_prepare) is indeed followed by check and
84 * dispatch. It's not clear whether this would be a bug, but let's
85 * play safe and allow it---it will just cause extra calls to
86 * event_notifier_set until the next call to dispatch.
88 * Instead, the aio_poll calls include both the prepare and the
89 * dispatch phase, hence a simple counter is enough for them.
93 /* lock to protect between bh's adders and deleter */
96 /* Anchor of the list of Bottom Halves belonging to the context */
97 struct QEMUBH
*first_bh
;
99 /* A simple lock used to protect the first_bh list, and ensure that
100 * no callbacks are removed while we're walking and dispatching callbacks.
104 /* Used by aio_notify.
106 * "notified" is used to avoid expensive event_notifier_test_and_clear
107 * calls. When it is clear, the EventNotifier is clear, or one thread
108 * is going to clear "notified" before processing more events. False
109 * positives are possible, i.e. "notified" could be set even though the
110 * EventNotifier is clear.
112 * Note that event_notifier_set *cannot* be optimized the same way. For
113 * more information on the problem that would result, see "#ifdef BUG2"
114 * in the docs/aio_notify_accept.promela formal model.
117 EventNotifier notifier
;
119 /* Scheduling this BH forces the event loop it iterate */
120 QEMUBH
*notify_dummy_bh
;
122 /* Thread pool for performing work and receiving completion callbacks */
123 struct ThreadPool
*thread_pool
;
125 #ifdef CONFIG_LINUX_AIO
126 /* State for native Linux AIO. Uses aio_context_acquire/release for
129 struct LinuxAioState
*linux_aio
;
132 /* TimerLists for calling timers - one per clock type */
133 QEMUTimerListGroup tlg
;
135 int external_disable_cnt
;
137 /* epoll(7) state used when built with CONFIG_EPOLL */
140 bool epoll_available
;
144 * aio_context_new: Allocate a new AioContext.
146 * AioContext provide a mini event-loop that can be waited on synchronously.
147 * They also provide bottom halves, a service to execute a piece of code
148 * as soon as possible.
150 AioContext
*aio_context_new(Error
**errp
);
154 * @ctx: The AioContext to operate on.
156 * Add a reference to an AioContext.
158 void aio_context_ref(AioContext
*ctx
);
162 * @ctx: The AioContext to operate on.
164 * Drop a reference to an AioContext.
166 void aio_context_unref(AioContext
*ctx
);
168 /* Take ownership of the AioContext. If the AioContext will be shared between
169 * threads, and a thread does not want to be interrupted, it will have to
170 * take ownership around calls to aio_poll(). Otherwise, aio_poll()
171 * automatically takes care of calling aio_context_acquire and
172 * aio_context_release.
174 * Access to timers and BHs from a thread that has not acquired AioContext
175 * is possible. Access to callbacks for now must be done while the AioContext
176 * is owned by the thread (FIXME).
178 void aio_context_acquire(AioContext
*ctx
);
180 /* Relinquish ownership of the AioContext. */
181 void aio_context_release(AioContext
*ctx
);
184 * aio_bh_schedule_oneshot: Allocate a new bottom half structure that will run
185 * only once and as soon as possible.
187 void aio_bh_schedule_oneshot(AioContext
*ctx
, QEMUBHFunc
*cb
, void *opaque
);
190 * aio_bh_new: Allocate a new bottom half structure.
192 * Bottom halves are lightweight callbacks whose invocation is guaranteed
193 * to be wait-free, thread-safe and signal-safe. The #QEMUBH structure
194 * is opaque and must be allocated prior to its use.
196 QEMUBH
*aio_bh_new(AioContext
*ctx
, QEMUBHFunc
*cb
, void *opaque
);
199 * aio_notify: Force processing of pending events.
201 * Similar to signaling a condition variable, aio_notify forces
202 * aio_wait to exit, so that the next call will re-examine pending events.
203 * The caller of aio_notify will usually call aio_wait again very soon,
204 * or go through another iteration of the GLib main loop. Hence, aio_notify
205 * also has the side effect of recalculating the sets of file descriptors
206 * that the main loop waits for.
208 * Calling aio_notify is rarely necessary, because for example scheduling
209 * a bottom half calls it already.
211 void aio_notify(AioContext
*ctx
);
214 * aio_notify_accept: Acknowledge receiving an aio_notify.
216 * aio_notify() uses an EventNotifier in order to wake up a sleeping
217 * aio_poll() or g_main_context_iteration(). Calls to aio_notify() are
218 * usually rare, but the AioContext has to clear the EventNotifier on
219 * every aio_poll() or g_main_context_iteration() in order to avoid
220 * busy waiting. This event_notifier_test_and_clear() cannot be done
221 * using the usual aio_context_set_event_notifier(), because it must
222 * be done before processing all events (file descriptors, bottom halves,
225 * aio_notify_accept() is an optimized event_notifier_test_and_clear()
226 * that is specific to an AioContext's notifier; it is used internally
227 * to clear the EventNotifier only if aio_notify() had been called.
229 void aio_notify_accept(AioContext
*ctx
);
232 * aio_bh_call: Executes callback function of the specified BH.
234 void aio_bh_call(QEMUBH
*bh
);
237 * aio_bh_poll: Poll bottom halves for an AioContext.
239 * These are internal functions used by the QEMU main loop.
240 * And notice that multiple occurrences of aio_bh_poll cannot
241 * be called concurrently
243 int aio_bh_poll(AioContext
*ctx
);
246 * qemu_bh_schedule: Schedule a bottom half.
248 * Scheduling a bottom half interrupts the main loop and causes the
249 * execution of the callback that was passed to qemu_bh_new.
251 * Bottom halves that are scheduled from a bottom half handler are instantly
252 * invoked. This can create an infinite loop if a bottom half handler
255 * @bh: The bottom half to be scheduled.
257 void qemu_bh_schedule(QEMUBH
*bh
);
260 * qemu_bh_cancel: Cancel execution of a bottom half.
262 * Canceling execution of a bottom half undoes the effect of calls to
263 * qemu_bh_schedule without freeing its resources yet. While cancellation
264 * itself is also wait-free and thread-safe, it can of course race with the
265 * loop that executes bottom halves unless you are holding the iothread
266 * mutex. This makes it mostly useless if you are not holding the mutex.
268 * @bh: The bottom half to be canceled.
270 void qemu_bh_cancel(QEMUBH
*bh
);
273 *qemu_bh_delete: Cancel execution of a bottom half and free its resources.
275 * Deleting a bottom half frees the memory that was allocated for it by
276 * qemu_bh_new. It also implies canceling the bottom half if it was
278 * This func is async. The bottom half will do the delete action at the finial
281 * @bh: The bottom half to be deleted.
283 void qemu_bh_delete(QEMUBH
*bh
);
285 /* Return whether there are any pending callbacks from the GSource
286 * attached to the AioContext, before g_poll is invoked.
288 * This is used internally in the implementation of the GSource.
290 bool aio_prepare(AioContext
*ctx
);
292 /* Return whether there are any pending callbacks from the GSource
293 * attached to the AioContext, after g_poll is invoked.
295 * This is used internally in the implementation of the GSource.
297 bool aio_pending(AioContext
*ctx
);
299 /* Dispatch any pending callbacks from the GSource attached to the AioContext.
301 * This is used internally in the implementation of the GSource.
303 bool aio_dispatch(AioContext
*ctx
);
305 /* Progress in completing AIO work to occur. This can issue new pending
306 * aio as a result of executing I/O completion or bh callbacks.
308 * Return whether any progress was made by executing AIO or bottom half
309 * handlers. If @blocking == true, this should always be true except
310 * if someone called aio_notify.
312 * If there are no pending bottom halves, but there are pending AIO
313 * operations, it may not be possible to make any progress without
314 * blocking. If @blocking is true, this function will wait until one
315 * or more AIO events have completed, to ensure something has moved
318 bool aio_poll(AioContext
*ctx
, bool blocking
);
320 /* Register a file descriptor and associated callbacks. Behaves very similarly
321 * to qemu_set_fd_handler. Unlike qemu_set_fd_handler, these callbacks will
322 * be invoked when using aio_poll().
324 * Code that invokes AIO completion functions should rely on this function
325 * instead of qemu_set_fd_handler[2].
327 void aio_set_fd_handler(AioContext
*ctx
,
334 /* Register an event notifier and associated callbacks. Behaves very similarly
335 * to event_notifier_set_handler. Unlike event_notifier_set_handler, these callbacks
336 * will be invoked when using aio_poll().
338 * Code that invokes AIO completion functions should rely on this function
339 * instead of event_notifier_set_handler.
341 void aio_set_event_notifier(AioContext
*ctx
,
342 EventNotifier
*notifier
,
344 EventNotifierHandler
*io_read
);
346 /* Return a GSource that lets the main loop poll the file descriptors attached
347 * to this AioContext.
349 GSource
*aio_get_g_source(AioContext
*ctx
);
351 /* Return the ThreadPool bound to this AioContext */
352 struct ThreadPool
*aio_get_thread_pool(AioContext
*ctx
);
354 /* Return the LinuxAioState bound to this AioContext */
355 struct LinuxAioState
*aio_get_linux_aio(AioContext
*ctx
);
359 * @ctx: the aio context
360 * @type: the clock type
362 * @cb: the callback to call on timer expiry
363 * @opaque: the opaque pointer to pass to the callback
365 * Allocate a new timer attached to the context @ctx.
366 * The function is responsible for memory allocation.
368 * The preferred interface is aio_timer_init. Use that
369 * unless you really need dynamic memory allocation.
371 * Returns: a pointer to the new timer
373 static inline QEMUTimer
*aio_timer_new(AioContext
*ctx
, QEMUClockType type
,
375 QEMUTimerCB
*cb
, void *opaque
)
377 return timer_new_tl(ctx
->tlg
.tl
[type
], scale
, cb
, opaque
);
382 * @ctx: the aio context
384 * @type: the clock type
386 * @cb: the callback to call on timer expiry
387 * @opaque: the opaque pointer to pass to the callback
389 * Initialise a new timer attached to the context @ctx.
390 * The caller is responsible for memory allocation.
392 static inline void aio_timer_init(AioContext
*ctx
,
393 QEMUTimer
*ts
, QEMUClockType type
,
395 QEMUTimerCB
*cb
, void *opaque
)
397 timer_init_tl(ts
, ctx
->tlg
.tl
[type
], scale
, cb
, opaque
);
401 * aio_compute_timeout:
402 * @ctx: the aio context
404 * Compute the timeout that a blocking aio_poll should use.
406 int64_t aio_compute_timeout(AioContext
*ctx
);
409 * aio_disable_external:
410 * @ctx: the aio context
412 * Disable the further processing of external clients.
414 static inline void aio_disable_external(AioContext
*ctx
)
416 atomic_inc(&ctx
->external_disable_cnt
);
420 * aio_enable_external:
421 * @ctx: the aio context
423 * Enable the processing of external clients.
425 static inline void aio_enable_external(AioContext
*ctx
)
427 assert(ctx
->external_disable_cnt
> 0);
428 atomic_dec(&ctx
->external_disable_cnt
);
432 * aio_external_disabled:
433 * @ctx: the aio context
435 * Return true if the external clients are disabled.
437 static inline bool aio_external_disabled(AioContext
*ctx
)
439 return atomic_read(&ctx
->external_disable_cnt
);
444 * @ctx: the aio context
445 * @is_external: Whether or not the checked node is an external event source.
447 * Check if the node's is_external flag is okay to be polled by the ctx at this
448 * moment. True means green light.
450 static inline bool aio_node_check(AioContext
*ctx
, bool is_external
)
452 return !is_external
|| !atomic_read(&ctx
->external_disable_cnt
);
457 * @ctx: the aio context
459 * Initialize the aio context.
461 void aio_context_setup(AioContext
*ctx
);