| 1 | /* |
| 2 | * QEMU aio implementation |
| 3 | * |
| 4 | * Copyright IBM, Corp. 2008 |
| 5 | * |
| 6 | * Authors: |
| 7 | * Anthony Liguori <aliguori@us.ibm.com> |
| 8 | * |
| 9 | * This work is licensed under the terms of the GNU GPL, version 2. See |
| 10 | * the COPYING file in the top-level directory. |
| 11 | * |
| 12 | */ |
| 13 | |
| 14 | #ifndef QEMU_AIO_H |
| 15 | #define QEMU_AIO_H |
| 16 | |
| 17 | #ifdef CONFIG_LINUX_IO_URING |
| 18 | #include <liburing.h> |
| 19 | #endif |
| 20 | #include "qemu/coroutine-core.h" |
| 21 | #include "qemu/queue.h" |
| 22 | #include "qemu/event_notifier.h" |
| 23 | #include "qemu/lockcnt.h" |
| 24 | #include "qemu/thread.h" |
| 25 | #include "qemu/timer.h" |
| 26 | |
| 27 | struct MemReentrancyGuard; |
| 28 | |
| 29 | typedef struct AioHandler AioHandler; |
| 30 | typedef QLIST_HEAD(, AioHandler) AioHandlerList; |
| 31 | typedef void QEMUBHFunc(void *opaque); |
| 32 | typedef bool AioPollFn(void *opaque); |
| 33 | typedef void IOHandler(void *opaque); |
| 34 | |
| 35 | struct ThreadPoolAio; |
| 36 | struct LinuxAioState; |
| 37 | typedef struct LuringState LuringState; |
| 38 | |
| 39 | /* Is polling disabled? */ |
| 40 | bool aio_poll_disabled(AioContext *ctx); |
| 41 | |
| 42 | #ifdef CONFIG_LINUX_IO_URING |
| 43 | /* |
| 44 | * Each io_uring request must have a unique CqeHandler that processes the cqe. |
| 45 | * The lifetime of a CqeHandler must be at least from aio_add_sqe() until |
| 46 | * ->cb() invocation. |
| 47 | */ |
| 48 | typedef struct CqeHandler CqeHandler; |
| 49 | struct CqeHandler { |
| 50 | /* Called by the AioContext when the request has completed */ |
| 51 | void (*cb)(CqeHandler *handler); |
| 52 | |
| 53 | /* Used internally, do not access this */ |
| 54 | QSIMPLEQ_ENTRY(CqeHandler) next; |
| 55 | |
| 56 | /* This field is filled in before ->cb() is called */ |
| 57 | struct io_uring_cqe cqe; |
| 58 | }; |
| 59 | |
| 60 | typedef QSIMPLEQ_HEAD(, CqeHandler) CqeHandlerSimpleQ; |
| 61 | #endif /* CONFIG_LINUX_IO_URING */ |
| 62 | |
| 63 | /* Callbacks for file descriptor monitoring implementations */ |
| 64 | typedef struct { |
| 65 | /* |
| 66 | * update: |
| 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 |
| 71 | * removed |
| 72 | * |
| 73 | * Add/remove/modify a monitored file descriptor. |
| 74 | * |
| 75 | * Called with ctx->list_lock acquired. |
| 76 | */ |
| 77 | void (*update)(AioContext *ctx, AioHandler *old_node, AioHandler *new_node); |
| 78 | |
| 79 | /* |
| 80 | * wait: |
| 81 | * @ctx: the AioContext |
| 82 | * @ready_list: list for handlers that become ready |
| 83 | * @timeout: maximum duration to wait, in nanoseconds |
| 84 | * |
| 85 | * Wait for file descriptors to become ready and place them on ready_list. |
| 86 | * |
| 87 | * Called with ctx->list_lock incremented but not locked. |
| 88 | * |
| 89 | * Returns: number of ready file descriptors. |
| 90 | */ |
| 91 | int (*wait)(AioContext *ctx, AioHandlerList *ready_list, int64_t timeout); |
| 92 | |
| 93 | /* |
| 94 | * need_wait: |
| 95 | * @ctx: the AioContext |
| 96 | * |
| 97 | * Tell aio_poll() when to stop userspace polling early because ->wait() |
| 98 | * has fds ready. |
| 99 | * |
| 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. |
| 104 | * |
| 105 | * Returns: true if ->wait() should be called, false otherwise. |
| 106 | */ |
| 107 | bool (*need_wait)(AioContext *ctx); |
| 108 | |
| 109 | /* |
| 110 | * dispatch: |
| 111 | * @ctx: the AioContext |
| 112 | * |
| 113 | * Dispatch any work that is specific to this file descriptor monitoring |
| 114 | * implementation. Usually the event loop's generic file descriptor |
| 115 | * monitoring, BH, and timer dispatching code is sufficient, but file |
| 116 | * descriptor monitoring implementations offering additional functionality |
| 117 | * may need to implement this function for custom behavior. Called at a |
| 118 | * point in the event loop when it is safe to invoke user-defined |
| 119 | * callbacks. |
| 120 | * |
| 121 | * This function is optional and may be NULL. |
| 122 | * |
| 123 | * Returns: true if progress was made (see aio_poll()'s return value), |
| 124 | * false otherwise. |
| 125 | */ |
| 126 | bool (*dispatch)(AioContext *ctx); |
| 127 | |
| 128 | /* |
| 129 | * gsource_prepare: |
| 130 | * @ctx: the AioContext |
| 131 | * |
| 132 | * Prepare for the glib event loop to wait for events instead of the usual |
| 133 | * ->wait() call. See glib's GSourceFuncs->prepare(). |
| 134 | */ |
| 135 | void (*gsource_prepare)(AioContext *ctx); |
| 136 | |
| 137 | /* |
| 138 | * gsource_check: |
| 139 | * @ctx: the AioContext |
| 140 | * |
| 141 | * Called by the glib event loop from glib's GSourceFuncs->check() after |
| 142 | * waiting for events. |
| 143 | * |
| 144 | * Returns: true when ready to be dispatched. |
| 145 | */ |
| 146 | bool (*gsource_check)(AioContext *ctx); |
| 147 | |
| 148 | /* |
| 149 | * gsource_dispatch: |
| 150 | * @ctx: the AioContext |
| 151 | * @ready_list: list for handlers that become ready |
| 152 | * |
| 153 | * Place ready AioHandlers on ready_list. Called as part of the glib event |
| 154 | * loop from glib's GSourceFuncs->dispatch(). |
| 155 | * |
| 156 | * Called with list_lock incremented. |
| 157 | */ |
| 158 | void (*gsource_dispatch)(AioContext *ctx, AioHandlerList *ready_list); |
| 159 | |
| 160 | #ifdef CONFIG_LINUX_IO_URING |
| 161 | /** |
| 162 | * add_sqe: Add an io_uring sqe for submission. |
| 163 | * @prep_sqe: invoked with an sqe that should be prepared for submission |
| 164 | * @opaque: user-defined argument to @prep_sqe() |
| 165 | * @cqe_handler: the unique cqe handler associated with this request |
| 166 | * |
| 167 | * The caller's @prep_sqe() function is invoked to fill in the details of |
| 168 | * the sqe. Do not call io_uring_sqe_set_data() on this sqe. |
| 169 | * |
| 170 | * The kernel may see the sqe as soon as @prep_sqe() returns or it may take |
| 171 | * until the next event loop iteration. |
| 172 | * |
| 173 | * This function is called from the current AioContext and is not |
| 174 | * thread-safe. |
| 175 | */ |
| 176 | void (*add_sqe)(AioContext *ctx, |
| 177 | void (*prep_sqe)(struct io_uring_sqe *sqe, void *opaque), |
| 178 | void *opaque, CqeHandler *cqe_handler); |
| 179 | #endif /* CONFIG_LINUX_IO_URING */ |
| 180 | } FDMonOps; |
| 181 | |
| 182 | /* |
| 183 | * Each aio_bh_poll() call carves off a slice of the BH list, so that newly |
| 184 | * scheduled BHs are not processed until the next aio_bh_poll() call. All |
| 185 | * active aio_bh_poll() calls chain their slices together in a list, so that |
| 186 | * nested aio_bh_poll() calls process all scheduled bottom halves. |
| 187 | */ |
| 188 | typedef QSLIST_HEAD(, QEMUBH) BHList; |
| 189 | typedef struct BHListSlice BHListSlice; |
| 190 | struct BHListSlice { |
| 191 | BHList bh_list; |
| 192 | QSIMPLEQ_ENTRY(BHListSlice) next; |
| 193 | }; |
| 194 | |
| 195 | typedef QSLIST_HEAD(, AioHandler) AioHandlerSList; |
| 196 | |
| 197 | typedef struct AioPolledEvent { |
| 198 | int64_t ns; /* estimated block time in nanoseconds */ |
| 199 | } AioPolledEvent; |
| 200 | |
| 201 | struct AioContext { |
| 202 | GSource source; |
| 203 | |
| 204 | /* Used by AioContext users to protect from multi-threaded access. */ |
| 205 | QemuRecMutex lock; |
| 206 | |
| 207 | /* |
| 208 | * Keep track of readers and writers of the block layer graph. |
| 209 | * This is essential to avoid performing additions and removal |
| 210 | * of nodes and edges from block graph while some |
| 211 | * other thread is traversing it. |
| 212 | */ |
| 213 | struct BdrvGraphRWlock *bdrv_graph; |
| 214 | |
| 215 | /* The list of registered AIO handlers. Protected by ctx->list_lock. */ |
| 216 | AioHandlerList aio_handlers; |
| 217 | |
| 218 | /* The list of AIO handlers to be deleted. Protected by ctx->list_lock. */ |
| 219 | AioHandlerList deleted_aio_handlers; |
| 220 | |
| 221 | /* Used to avoid unnecessary event_notifier_set calls in aio_notify; |
| 222 | * only written from the AioContext home thread, or under the BQL in |
| 223 | * the case of the main AioContext. However, it is read from any |
| 224 | * thread so it is still accessed with atomic primitives. |
| 225 | * |
| 226 | * If this field is 0, everything (file descriptors, bottom halves, |
| 227 | * timers) will be re-evaluated before the next blocking poll() or |
| 228 | * io_uring wait; therefore, the event_notifier_set call can be |
| 229 | * skipped. If it is non-zero, you may need to wake up a concurrent |
| 230 | * aio_poll or the glib main event loop, making event_notifier_set |
| 231 | * necessary. |
| 232 | * |
| 233 | * Bit 0 is reserved for GSource usage of the AioContext, and is 1 |
| 234 | * between a call to aio_ctx_prepare and the next call to aio_ctx_check. |
| 235 | * Bits 1-31 simply count the number of active calls to aio_poll |
| 236 | * that are in the prepare or poll phase. |
| 237 | * |
| 238 | * The GSource and aio_poll must use a different mechanism because |
| 239 | * there is no certainty that a call to GSource's prepare callback |
| 240 | * (via g_main_context_prepare) is indeed followed by check and |
| 241 | * dispatch. It's not clear whether this would be a bug, but let's |
| 242 | * play safe and allow it---it will just cause extra calls to |
| 243 | * event_notifier_set until the next call to dispatch. |
| 244 | * |
| 245 | * Instead, the aio_poll calls include both the prepare and the |
| 246 | * dispatch phase, hence a simple counter is enough for them. |
| 247 | */ |
| 248 | uint32_t notify_me; |
| 249 | |
| 250 | /* A lock to protect between QEMUBH and AioHandler adders and deleter, |
| 251 | * and to ensure that no callbacks are removed while we're walking and |
| 252 | * dispatching them. |
| 253 | */ |
| 254 | QemuLockCnt list_lock; |
| 255 | |
| 256 | /* Bottom Halves pending aio_bh_poll() processing */ |
| 257 | BHList bh_list; |
| 258 | |
| 259 | /* Chained BH list slices for each nested aio_bh_poll() call */ |
| 260 | QSIMPLEQ_HEAD(, BHListSlice) bh_slice_list; |
| 261 | |
| 262 | /* Used by aio_notify. |
| 263 | * |
| 264 | * "notified" is used to avoid expensive event_notifier_test_and_clear |
| 265 | * calls. When it is clear, the EventNotifier is clear, or one thread |
| 266 | * is going to clear "notified" before processing more events. False |
| 267 | * positives are possible, i.e. "notified" could be set even though the |
| 268 | * EventNotifier is clear. |
| 269 | * |
| 270 | * Note that event_notifier_set *cannot* be optimized the same way. For |
| 271 | * more information on the problem that would result, see "#ifdef BUG2" |
| 272 | * in the docs/aio_notify_accept.promela formal model. |
| 273 | */ |
| 274 | bool notified; |
| 275 | EventNotifier notifier; |
| 276 | |
| 277 | QSLIST_HEAD(, Coroutine) scheduled_coroutines; |
| 278 | QEMUBH *co_schedule_bh; |
| 279 | |
| 280 | int thread_pool_min; |
| 281 | int thread_pool_max; |
| 282 | /* Thread pool for performing work and receiving completion callbacks. |
| 283 | * Has its own locking. |
| 284 | */ |
| 285 | struct ThreadPoolAio *thread_pool; |
| 286 | |
| 287 | #ifdef CONFIG_LINUX_AIO |
| 288 | struct LinuxAioState *linux_aio; |
| 289 | #endif |
| 290 | #ifdef CONFIG_LINUX_IO_URING |
| 291 | /* State for file descriptor monitoring using Linux io_uring */ |
| 292 | struct io_uring fdmon_io_uring; |
| 293 | AioHandlerSList submit_list; |
| 294 | void *io_uring_fd_tag; |
| 295 | |
| 296 | /* Pending callback state for cqe handlers */ |
| 297 | CqeHandlerSimpleQ cqe_handler_ready_list; |
| 298 | #endif /* CONFIG_LINUX_IO_URING */ |
| 299 | |
| 300 | /* TimerLists for calling timers - one per clock type. Has its own |
| 301 | * locking. |
| 302 | */ |
| 303 | QEMUTimerListGroup tlg; |
| 304 | |
| 305 | /* Number of AioHandlers without .io_poll() */ |
| 306 | int poll_disable_cnt; |
| 307 | |
| 308 | /* Polling mode parameters */ |
| 309 | int64_t poll_ns; /* current polling time in nanoseconds */ |
| 310 | int64_t poll_max_ns; /* maximum polling time in nanoseconds */ |
| 311 | int64_t poll_grow; /* polling time growth factor */ |
| 312 | int64_t poll_shrink; /* polling time shrink factor */ |
| 313 | int64_t poll_weight; /* weight of current interval in calculation */ |
| 314 | |
| 315 | /* AIO engine parameters */ |
| 316 | int64_t aio_max_batch; /* maximum number of requests in a batch */ |
| 317 | |
| 318 | /* |
| 319 | * List of handlers participating in userspace polling. Protected by |
| 320 | * ctx->list_lock. Iterated and modified mostly by the event loop thread |
| 321 | * from aio_poll() with ctx->list_lock incremented. aio_set_fd_handler() |
| 322 | * only touches the list to delete nodes if ctx->list_lock's count is zero. |
| 323 | */ |
| 324 | AioHandlerList poll_aio_handlers; |
| 325 | |
| 326 | /* Are we in polling mode or monitoring file descriptors? */ |
| 327 | bool poll_started; |
| 328 | |
| 329 | /* epoll(7) state used when built with CONFIG_EPOLL */ |
| 330 | int epollfd; |
| 331 | |
| 332 | /* The GSource unix fd tag for epollfd */ |
| 333 | void *epollfd_tag; |
| 334 | |
| 335 | const FDMonOps *fdmon_ops; |
| 336 | |
| 337 | /* Was aio_context_new() successful? */ |
| 338 | bool initialized; |
| 339 | }; |
| 340 | |
| 341 | /** |
| 342 | * aio_context_new: Allocate a new AioContext. |
| 343 | * |
| 344 | * AioContext provide a mini event-loop that can be waited on synchronously. |
| 345 | * They also provide bottom halves, a service to execute a piece of code |
| 346 | * as soon as possible. |
| 347 | */ |
| 348 | AioContext *aio_context_new(Error **errp); |
| 349 | |
| 350 | /** |
| 351 | * aio_context_ref: |
| 352 | * @ctx: The AioContext to operate on. |
| 353 | * |
| 354 | * Add a reference to an AioContext. |
| 355 | */ |
| 356 | void aio_context_ref(AioContext *ctx); |
| 357 | |
| 358 | /** |
| 359 | * aio_context_unref: |
| 360 | * @ctx: The AioContext to operate on. |
| 361 | * |
| 362 | * Drop a reference to an AioContext. |
| 363 | */ |
| 364 | void aio_context_unref(AioContext *ctx); |
| 365 | |
| 366 | /** |
| 367 | * aio_bh_schedule_oneshot_full: Allocate a new bottom half structure that will |
| 368 | * run only once and as soon as possible. |
| 369 | * |
| 370 | * @name: A human-readable identifier for debugging purposes. |
| 371 | */ |
| 372 | void aio_bh_schedule_oneshot_full(AioContext *ctx, QEMUBHFunc *cb, void *opaque, |
| 373 | const char *name); |
| 374 | |
| 375 | /** |
| 376 | * aio_bh_schedule_oneshot: Allocate a new bottom half structure that will run |
| 377 | * only once and as soon as possible. |
| 378 | * |
| 379 | * A convenience wrapper for aio_bh_schedule_oneshot_full() that uses cb as the |
| 380 | * name string. |
| 381 | */ |
| 382 | #define aio_bh_schedule_oneshot(ctx, cb, opaque) \ |
| 383 | aio_bh_schedule_oneshot_full((ctx), (cb), (opaque), (stringify(cb))) |
| 384 | |
| 385 | /** |
| 386 | * aio_bh_new_full: Allocate a new bottom half structure. |
| 387 | * |
| 388 | * Bottom halves are lightweight callbacks whose invocation is guaranteed |
| 389 | * to be wait-free, thread-safe and signal-safe. The #QEMUBH structure |
| 390 | * is opaque and must be allocated prior to its use. |
| 391 | * |
| 392 | * @name: A human-readable identifier for debugging purposes. |
| 393 | * @reentrancy_guard: A guard set when entering a cb to prevent |
| 394 | * device-reentrancy issues |
| 395 | */ |
| 396 | QEMUBH *aio_bh_new_full(AioContext *ctx, QEMUBHFunc *cb, void *opaque, |
| 397 | const char *name, struct MemReentrancyGuard *reentrancy_guard); |
| 398 | |
| 399 | /** |
| 400 | * aio_bh_new: Allocate a new bottom half structure |
| 401 | * |
| 402 | * A convenience wrapper for aio_bh_new_full() that uses the cb as the name |
| 403 | * string. |
| 404 | */ |
| 405 | #define aio_bh_new(ctx, cb, opaque) \ |
| 406 | aio_bh_new_full((ctx), (cb), (opaque), (stringify(cb)), NULL) |
| 407 | |
| 408 | /** |
| 409 | * aio_bh_new_guarded: Allocate a new bottom half structure with a |
| 410 | * reentrancy_guard |
| 411 | * |
| 412 | * A convenience wrapper for aio_bh_new_full() that uses the cb as the name |
| 413 | * string. |
| 414 | */ |
| 415 | #define aio_bh_new_guarded(ctx, cb, opaque, guard) \ |
| 416 | aio_bh_new_full((ctx), (cb), (opaque), (stringify(cb)), guard) |
| 417 | |
| 418 | /** |
| 419 | * aio_notify: Force processing of pending events. |
| 420 | * |
| 421 | * Similar to signaling a condition variable, aio_notify forces |
| 422 | * aio_poll to exit, so that the next call will re-examine pending events. |
| 423 | * The caller of aio_notify will usually call aio_poll again very soon, |
| 424 | * or go through another iteration of the GLib main loop. Hence, aio_notify |
| 425 | * also has the side effect of recalculating the sets of file descriptors |
| 426 | * that the main loop waits for. |
| 427 | * |
| 428 | * Calling aio_notify is rarely necessary, because for example scheduling |
| 429 | * a bottom half calls it already. |
| 430 | */ |
| 431 | void aio_notify(AioContext *ctx); |
| 432 | |
| 433 | /** |
| 434 | * aio_notify_accept: Acknowledge receiving an aio_notify. |
| 435 | * |
| 436 | * aio_notify() uses an EventNotifier in order to wake up a sleeping |
| 437 | * aio_poll() or g_main_context_iteration(). Calls to aio_notify() are |
| 438 | * usually rare, but the AioContext has to clear the EventNotifier on |
| 439 | * every aio_poll() or g_main_context_iteration() in order to avoid |
| 440 | * busy waiting. This event_notifier_test_and_clear() cannot be done |
| 441 | * using the usual aio_context_set_event_notifier(), because it must |
| 442 | * be done before processing all events (file descriptors, bottom halves, |
| 443 | * timers). |
| 444 | * |
| 445 | * aio_notify_accept() is an optimized event_notifier_test_and_clear() |
| 446 | * that is specific to an AioContext's notifier; it is used internally |
| 447 | * to clear the EventNotifier only if aio_notify() had been called. |
| 448 | */ |
| 449 | void aio_notify_accept(AioContext *ctx); |
| 450 | |
| 451 | /** |
| 452 | * aio_bh_call: Executes callback function of the specified BH. |
| 453 | */ |
| 454 | void aio_bh_call(QEMUBH *bh); |
| 455 | |
| 456 | /** |
| 457 | * aio_bh_poll: Poll bottom halves for an AioContext. |
| 458 | * |
| 459 | * These are internal functions used by the QEMU main loop. |
| 460 | * And notice that multiple occurrences of aio_bh_poll cannot |
| 461 | * be called concurrently |
| 462 | */ |
| 463 | int aio_bh_poll(AioContext *ctx); |
| 464 | |
| 465 | /** |
| 466 | * qemu_bh_schedule: Schedule a bottom half. |
| 467 | * |
| 468 | * Scheduling a bottom half interrupts the main loop and causes the |
| 469 | * execution of the callback that was passed to qemu_bh_new. |
| 470 | * |
| 471 | * Bottom halves that are scheduled from a bottom half handler are instantly |
| 472 | * invoked. This can create an infinite loop if a bottom half handler |
| 473 | * schedules itself. |
| 474 | * |
| 475 | * @bh: The bottom half to be scheduled. |
| 476 | */ |
| 477 | void qemu_bh_schedule(QEMUBH *bh); |
| 478 | |
| 479 | /** |
| 480 | * qemu_bh_cancel: Cancel execution of a bottom half. |
| 481 | * |
| 482 | * Canceling execution of a bottom half undoes the effect of calls to |
| 483 | * qemu_bh_schedule without freeing its resources yet. While cancellation |
| 484 | * itself is also wait-free and thread-safe, it can of course race with the |
| 485 | * loop that executes bottom halves unless you are holding the iothread |
| 486 | * mutex. This makes it mostly useless if you are not holding the mutex. |
| 487 | * |
| 488 | * @bh: The bottom half to be canceled. |
| 489 | */ |
| 490 | void qemu_bh_cancel(QEMUBH *bh); |
| 491 | |
| 492 | /** |
| 493 | *qemu_bh_delete: Cancel execution of a bottom half and free its resources. |
| 494 | * |
| 495 | * Deleting a bottom half frees the memory that was allocated for it by |
| 496 | * qemu_bh_new. It also implies canceling the bottom half if it was |
| 497 | * scheduled. |
| 498 | * This func is async. The bottom half will do the delete action at the finial |
| 499 | * end. |
| 500 | * |
| 501 | * @bh: The bottom half to be deleted. |
| 502 | */ |
| 503 | void qemu_bh_delete(QEMUBH *bh); |
| 504 | |
| 505 | /* Return whether there are any pending callbacks from the GSource |
| 506 | * attached to the AioContext, before g_poll is invoked. |
| 507 | * |
| 508 | * This is used internally in the implementation of the GSource. |
| 509 | */ |
| 510 | bool aio_prepare(AioContext *ctx); |
| 511 | |
| 512 | /* Return whether there are any pending callbacks from the GSource |
| 513 | * attached to the AioContext, after g_poll is invoked. |
| 514 | * |
| 515 | * This is used internally in the implementation of the GSource. |
| 516 | */ |
| 517 | bool aio_pending(AioContext *ctx); |
| 518 | |
| 519 | /* Dispatch any pending callbacks from the GSource attached to the AioContext. |
| 520 | * |
| 521 | * This is used internally in the implementation of the GSource. |
| 522 | */ |
| 523 | void aio_dispatch(AioContext *ctx); |
| 524 | |
| 525 | /* Progress in completing AIO work to occur. This can issue new pending |
| 526 | * aio as a result of executing I/O completion or bh callbacks. |
| 527 | * |
| 528 | * Return whether any progress was made by executing AIO or bottom half |
| 529 | * handlers. If @blocking == true, this should always be true except |
| 530 | * if someone called aio_notify. |
| 531 | * |
| 532 | * If there are no pending bottom halves, but there are pending AIO |
| 533 | * operations, it may not be possible to make any progress without |
| 534 | * blocking. If @blocking is true, this function will wait until one |
| 535 | * or more AIO events have completed, to ensure something has moved |
| 536 | * before returning. |
| 537 | */ |
| 538 | bool no_coroutine_fn aio_poll(AioContext *ctx, bool blocking); |
| 539 | |
| 540 | /* Register a file descriptor and associated callbacks. Behaves very similarly |
| 541 | * to qemu_set_fd_handler. Unlike qemu_set_fd_handler, these callbacks will |
| 542 | * be invoked when using aio_poll(). |
| 543 | * |
| 544 | * Code that invokes AIO completion functions should rely on this function |
| 545 | * instead of qemu_set_fd_handler[2]. |
| 546 | */ |
| 547 | void aio_set_fd_handler(AioContext *ctx, |
| 548 | int fd, |
| 549 | IOHandler *io_read, |
| 550 | IOHandler *io_write, |
| 551 | AioPollFn *io_poll, |
| 552 | IOHandler *io_poll_ready, |
| 553 | void *opaque); |
| 554 | |
| 555 | /* Register an event notifier and associated callbacks. Behaves very similarly |
| 556 | * to event_notifier_set_handler. Unlike event_notifier_set_handler, these callbacks |
| 557 | * will be invoked when using aio_poll(). |
| 558 | * |
| 559 | * Code that invokes AIO completion functions should rely on this function |
| 560 | * instead of event_notifier_set_handler. |
| 561 | */ |
| 562 | void aio_set_event_notifier(AioContext *ctx, |
| 563 | EventNotifier *notifier, |
| 564 | EventNotifierHandler *io_read, |
| 565 | AioPollFn *io_poll, |
| 566 | EventNotifierHandler *io_poll_ready); |
| 567 | |
| 568 | /* |
| 569 | * Set polling begin/end callbacks for an event notifier that has already been |
| 570 | * registered with aio_set_event_notifier. Do nothing if the event notifier is |
| 571 | * not registered. |
| 572 | * |
| 573 | * Note that if the io_poll_end() callback (or the entire notifier) is removed |
| 574 | * during polling, it will not be called, so an io_poll_begin() is not |
| 575 | * necessarily always followed by an io_poll_end(). |
| 576 | */ |
| 577 | void aio_set_event_notifier_poll(AioContext *ctx, |
| 578 | EventNotifier *notifier, |
| 579 | EventNotifierHandler *io_poll_begin, |
| 580 | EventNotifierHandler *io_poll_end); |
| 581 | |
| 582 | /* Return a GSource that lets the main loop poll the file descriptors attached |
| 583 | * to this AioContext. |
| 584 | */ |
| 585 | GSource *aio_get_g_source(AioContext *ctx); |
| 586 | |
| 587 | /* Return the ThreadPoolAio bound to this AioContext */ |
| 588 | struct ThreadPoolAio *aio_get_thread_pool(AioContext *ctx); |
| 589 | |
| 590 | /* Setup the LinuxAioState bound to this AioContext */ |
| 591 | struct LinuxAioState *aio_setup_linux_aio(AioContext *ctx, Error **errp); |
| 592 | |
| 593 | /* Return the LinuxAioState bound to this AioContext */ |
| 594 | struct LinuxAioState *aio_get_linux_aio(AioContext *ctx); |
| 595 | |
| 596 | /** |
| 597 | * aio_timer_new_with_attrs: |
| 598 | * @ctx: the aio context |
| 599 | * @type: the clock type |
| 600 | * @scale: the scale |
| 601 | * @attributes: 0, or one to multiple OR'ed QEMU_TIMER_ATTR_<id> values |
| 602 | * to assign |
| 603 | * @cb: the callback to call on timer expiry |
| 604 | * @opaque: the opaque pointer to pass to the callback |
| 605 | * |
| 606 | * Allocate a new timer (with attributes) attached to the context @ctx. |
| 607 | * The function is responsible for memory allocation. |
| 608 | * |
| 609 | * The preferred interface is aio_timer_init or aio_timer_init_with_attrs. |
| 610 | * Use that unless you really need dynamic memory allocation. |
| 611 | * |
| 612 | * Returns: a pointer to the new timer |
| 613 | */ |
| 614 | static inline QEMUTimer *aio_timer_new_with_attrs(AioContext *ctx, |
| 615 | QEMUClockType type, |
| 616 | int scale, int attributes, |
| 617 | QEMUTimerCB *cb, void *opaque) |
| 618 | { |
| 619 | return timer_new_full(&ctx->tlg, type, scale, attributes, cb, opaque); |
| 620 | } |
| 621 | |
| 622 | /** |
| 623 | * aio_timer_new: |
| 624 | * @ctx: the aio context |
| 625 | * @type: the clock type |
| 626 | * @scale: the scale |
| 627 | * @cb: the callback to call on timer expiry |
| 628 | * @opaque: the opaque pointer to pass to the callback |
| 629 | * |
| 630 | * Allocate a new timer attached to the context @ctx. |
| 631 | * See aio_timer_new_with_attrs for details. |
| 632 | * |
| 633 | * Returns: a pointer to the new timer |
| 634 | */ |
| 635 | static inline QEMUTimer *aio_timer_new(AioContext *ctx, QEMUClockType type, |
| 636 | int scale, |
| 637 | QEMUTimerCB *cb, void *opaque) |
| 638 | { |
| 639 | return timer_new_full(&ctx->tlg, type, scale, 0, cb, opaque); |
| 640 | } |
| 641 | |
| 642 | /** |
| 643 | * aio_timer_init_with_attrs: |
| 644 | * @ctx: the aio context |
| 645 | * @ts: the timer |
| 646 | * @type: the clock type |
| 647 | * @scale: the scale |
| 648 | * @attributes: 0, or one to multiple OR'ed QEMU_TIMER_ATTR_<id> values |
| 649 | * to assign |
| 650 | * @cb: the callback to call on timer expiry |
| 651 | * @opaque: the opaque pointer to pass to the callback |
| 652 | * |
| 653 | * Initialise a new timer (with attributes) attached to the context @ctx. |
| 654 | * The caller is responsible for memory allocation. |
| 655 | */ |
| 656 | static inline void aio_timer_init_with_attrs(AioContext *ctx, |
| 657 | QEMUTimer *ts, QEMUClockType type, |
| 658 | int scale, int attributes, |
| 659 | QEMUTimerCB *cb, void *opaque) |
| 660 | { |
| 661 | timer_init_full(ts, &ctx->tlg, type, scale, attributes, cb, opaque); |
| 662 | } |
| 663 | |
| 664 | /** |
| 665 | * aio_timer_init: |
| 666 | * @ctx: the aio context |
| 667 | * @ts: the timer |
| 668 | * @type: the clock type |
| 669 | * @scale: the scale |
| 670 | * @cb: the callback to call on timer expiry |
| 671 | * @opaque: the opaque pointer to pass to the callback |
| 672 | * |
| 673 | * Initialise a new timer attached to the context @ctx. |
| 674 | * See aio_timer_init_with_attrs for details. |
| 675 | */ |
| 676 | static inline void aio_timer_init(AioContext *ctx, |
| 677 | QEMUTimer *ts, QEMUClockType type, |
| 678 | int scale, |
| 679 | QEMUTimerCB *cb, void *opaque) |
| 680 | { |
| 681 | timer_init_full(ts, &ctx->tlg, type, scale, 0, cb, opaque); |
| 682 | } |
| 683 | |
| 684 | /** |
| 685 | * aio_compute_timeout: |
| 686 | * @ctx: the aio context |
| 687 | * |
| 688 | * Compute the timeout that a blocking aio_poll should use. |
| 689 | */ |
| 690 | int64_t aio_compute_timeout(AioContext *ctx); |
| 691 | |
| 692 | /** |
| 693 | * aio_co_schedule: |
| 694 | * @ctx: the aio context |
| 695 | * @co: the coroutine |
| 696 | * |
| 697 | * Start a coroutine on a remote AioContext. |
| 698 | * |
| 699 | * The coroutine must not be entered by anyone else while aio_co_schedule() |
| 700 | * is active. In addition the coroutine must have yielded unless ctx |
| 701 | * is the context in which the coroutine is running (i.e. the value of |
| 702 | * qemu_get_current_aio_context() from the coroutine itself). |
| 703 | */ |
| 704 | void aio_co_schedule(AioContext *ctx, Coroutine *co); |
| 705 | |
| 706 | /** |
| 707 | * aio_co_reschedule_self: |
| 708 | * @new_ctx: the new context |
| 709 | * |
| 710 | * Move the currently running coroutine to new_ctx. If the coroutine is already |
| 711 | * running in new_ctx, do nothing. |
| 712 | * |
| 713 | * Note that this function cannot reschedule from iohandler_ctx to |
| 714 | * qemu_aio_context. |
| 715 | */ |
| 716 | void coroutine_fn aio_co_reschedule_self(AioContext *new_ctx); |
| 717 | |
| 718 | /** |
| 719 | * aio_co_wake: |
| 720 | * @co: the coroutine |
| 721 | * |
| 722 | * Restart a coroutine on the AioContext where it was running last, thus |
| 723 | * preventing coroutines from jumping from one context to another when they |
| 724 | * go to sleep. |
| 725 | * |
| 726 | * aio_co_wake may be executed either in coroutine or non-coroutine |
| 727 | * context. The coroutine must not be entered by anyone else while |
| 728 | * aio_co_wake() is active. |
| 729 | * |
| 730 | * If `co`'s AioContext differs from the current AioContext, this will call |
| 731 | * aio_co_schedule(), which makes this safe to use even when `co` has not |
| 732 | * yielded yet. In such a case, it will be entered once it yields. |
| 733 | * |
| 734 | * In contrast, if `co`'s AioContext is equal to the current one, it is |
| 735 | * required for `co` to currently be yielding. This is generally the case |
| 736 | * if the caller is not in `co` (i.e. invoked by `co`), because the only |
| 737 | * other way for the caller to be running then is for `co` to currently be |
| 738 | * yielding. |
| 739 | * |
| 740 | * Therefore, if there is no way for the caller to be invoked/entered by |
| 741 | * `co`, it is generally safe to call this regardless of whether `co` is |
| 742 | * known to already be yielding or not -- it only has to yield at some |
| 743 | * point. |
| 744 | */ |
| 745 | void aio_co_wake(Coroutine *co); |
| 746 | |
| 747 | /** |
| 748 | * aio_co_enter: |
| 749 | * @ctx: the context to run the coroutine |
| 750 | * @co: the coroutine to run |
| 751 | * |
| 752 | * Enter a coroutine in the specified AioContext. |
| 753 | */ |
| 754 | void aio_co_enter(AioContext *ctx, Coroutine *co); |
| 755 | |
| 756 | /** |
| 757 | * Return the AioContext whose event loop runs in the current thread. |
| 758 | * |
| 759 | * If called from an IOThread this will be the IOThread's AioContext. If |
| 760 | * called from the main thread or with the "big QEMU lock" taken it |
| 761 | * will be the main loop AioContext. |
| 762 | * |
| 763 | * Note that the return value is never the main loop's iohandler_ctx and the |
| 764 | * return value is the main loop AioContext instead. |
| 765 | */ |
| 766 | AioContext *qemu_get_current_aio_context(void); |
| 767 | |
| 768 | void qemu_set_current_aio_context(AioContext *ctx); |
| 769 | |
| 770 | /** |
| 771 | * aio_context_setup: |
| 772 | * @ctx: the aio context |
| 773 | * @errp: error pointer |
| 774 | * |
| 775 | * Initialize the aio context. |
| 776 | * |
| 777 | * Returns: true on success, false otherwise |
| 778 | */ |
| 779 | bool aio_context_setup(AioContext *ctx, Error **errp); |
| 780 | |
| 781 | /** |
| 782 | * aio_context_destroy: |
| 783 | * @ctx: the aio context |
| 784 | * |
| 785 | * Destroy the aio context. |
| 786 | */ |
| 787 | void aio_context_destroy(AioContext *ctx); |
| 788 | |
| 789 | /** |
| 790 | * aio_context_set_poll_params: |
| 791 | * @ctx: the aio context |
| 792 | * @max_ns: how long to busy poll for, in nanoseconds |
| 793 | * @grow: polling time growth factor |
| 794 | * @shrink: polling time shrink factor |
| 795 | * @weight: weight factor applied to the current polling interval |
| 796 | * |
| 797 | * Poll mode can be disabled by setting poll_max_ns to 0. |
| 798 | */ |
| 799 | void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns, |
| 800 | int64_t grow, int64_t shrink, |
| 801 | int64_t weight, Error **errp); |
| 802 | |
| 803 | /** |
| 804 | * aio_context_set_aio_params: |
| 805 | * @ctx: the aio context |
| 806 | * @max_batch: maximum number of requests in a batch, 0 means that the |
| 807 | * engine will use its default |
| 808 | */ |
| 809 | void aio_context_set_aio_params(AioContext *ctx, int64_t max_batch); |
| 810 | |
| 811 | /** |
| 812 | * aio_context_set_thread_pool_params: |
| 813 | * @ctx: the aio context |
| 814 | * @min: min number of threads to have readily available in the thread pool |
| 815 | * @min: max number of threads the thread pool can contain |
| 816 | */ |
| 817 | void aio_context_set_thread_pool_params(AioContext *ctx, int64_t min, |
| 818 | int64_t max, Error **errp); |
| 819 | |
| 820 | #ifdef CONFIG_LINUX_IO_URING |
| 821 | /** |
| 822 | * aio_has_io_uring: Return whether io_uring is available. |
| 823 | * |
| 824 | * io_uring is either available in all AioContexts or in none, so this only |
| 825 | * needs to be called once from within any thread's AioContext. |
| 826 | */ |
| 827 | static inline bool aio_has_io_uring(void) |
| 828 | { |
| 829 | AioContext *ctx = qemu_get_current_aio_context(); |
| 830 | return ctx->fdmon_ops->add_sqe; |
| 831 | } |
| 832 | |
| 833 | /** |
| 834 | * aio_add_sqe: Add an io_uring sqe for submission. |
| 835 | * @prep_sqe: invoked with an sqe that should be prepared for submission |
| 836 | * @opaque: user-defined argument to @prep_sqe() |
| 837 | * @cqe_handler: the unique cqe handler associated with this request |
| 838 | * |
| 839 | * The caller's @prep_sqe() function is invoked to fill in the details of the |
| 840 | * sqe. Do not call io_uring_sqe_set_data() on this sqe. |
| 841 | * |
| 842 | * The sqe is submitted by the current AioContext. The kernel may see the sqe |
| 843 | * as soon as @prep_sqe() returns or it may take until the next event loop |
| 844 | * iteration. |
| 845 | * |
| 846 | * When the AioContext is destroyed, pending sqes are ignored and their |
| 847 | * CqeHandlers are not invoked. |
| 848 | * |
| 849 | * This function must be called only when aio_has_io_uring() returns true. |
| 850 | */ |
| 851 | void aio_add_sqe(void (*prep_sqe)(struct io_uring_sqe *sqe, void *opaque), |
| 852 | void *opaque, CqeHandler *cqe_handler); |
| 853 | #endif /* CONFIG_LINUX_IO_URING */ |
| 854 | |
| 855 | #endif |