xref: /qemu/block/linux-aio.c (revision 8b7b9c5c)
1 /*
2  * Linux native AIO support.
3  *
4  * Copyright (C) 2009 IBM, Corp.
5  * Copyright (C) 2009 Red Hat, Inc.
6  *
7  * This work is licensed under the terms of the GNU GPL, version 2 or later.
8  * See the COPYING file in the top-level directory.
9  */
10 #include "qemu/osdep.h"
11 #include "block/aio.h"
12 #include "qemu/queue.h"
13 #include "block/block.h"
14 #include "block/raw-aio.h"
15 #include "qemu/event_notifier.h"
16 #include "qemu/coroutine.h"
17 #include "qapi/error.h"
18 #include "sysemu/block-backend.h"
19 
20 /* Only used for assertions.  */
21 #include "qemu/coroutine_int.h"
22 
23 #include <libaio.h>
24 
25 /*
26  * Queue size (per-device).
27  *
28  * XXX: eventually we need to communicate this to the guest and/or make it
29  *      tunable by the guest.  If we get more outstanding requests at a time
30  *      than this we will get EAGAIN from io_submit which is communicated to
31  *      the guest as an I/O error.
32  */
33 #define MAX_EVENTS 1024
34 
35 /* Maximum number of requests in a batch. (default value) */
36 #define DEFAULT_MAX_BATCH 32
37 
38 struct qemu_laiocb {
39     Coroutine *co;
40     LinuxAioState *ctx;
41     struct iocb iocb;
42     ssize_t ret;
43     size_t nbytes;
44     QEMUIOVector *qiov;
45     bool is_read;
46     QSIMPLEQ_ENTRY(qemu_laiocb) next;
47 };
48 
49 typedef struct {
50     unsigned int in_queue;
51     unsigned int in_flight;
52     bool blocked;
53     QSIMPLEQ_HEAD(, qemu_laiocb) pending;
54 } LaioQueue;
55 
56 struct LinuxAioState {
57     AioContext *aio_context;
58 
59     io_context_t ctx;
60     EventNotifier e;
61 
62     /* No locking required, only accessed from AioContext home thread */
63     LaioQueue io_q;
64     QEMUBH *completion_bh;
65     int event_idx;
66     int event_max;
67 };
68 
69 static void ioq_submit(LinuxAioState *s);
70 
71 static inline ssize_t io_event_ret(struct io_event *ev)
72 {
73     return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
74 }
75 
76 /*
77  * Completes an AIO request.
78  */
79 static void qemu_laio_process_completion(struct qemu_laiocb *laiocb)
80 {
81     int ret;
82 
83     ret = laiocb->ret;
84     if (ret != -ECANCELED) {
85         if (ret == laiocb->nbytes) {
86             ret = 0;
87         } else if (ret >= 0) {
88             /* Short reads mean EOF, pad with zeros. */
89             if (laiocb->is_read) {
90                 qemu_iovec_memset(laiocb->qiov, ret, 0,
91                     laiocb->qiov->size - ret);
92             } else {
93                 ret = -ENOSPC;
94             }
95         }
96     }
97 
98     laiocb->ret = ret;
99 
100     /*
101      * If the coroutine is already entered it must be in ioq_submit() and
102      * will notice laio->ret has been filled in when it eventually runs
103      * later.  Coroutines cannot be entered recursively so avoid doing
104      * that!
105      */
106     assert(laiocb->co->ctx == laiocb->ctx->aio_context);
107     if (!qemu_coroutine_entered(laiocb->co)) {
108         aio_co_wake(laiocb->co);
109     }
110 }
111 
112 /**
113  * aio_ring buffer which is shared between userspace and kernel.
114  *
115  * This copied from linux/fs/aio.c, common header does not exist
116  * but AIO exists for ages so we assume ABI is stable.
117  */
118 struct aio_ring {
119     unsigned    id;    /* kernel internal index number */
120     unsigned    nr;    /* number of io_events */
121     unsigned    head;  /* Written to by userland or by kernel. */
122     unsigned    tail;
123 
124     unsigned    magic;
125     unsigned    compat_features;
126     unsigned    incompat_features;
127     unsigned    header_length;  /* size of aio_ring */
128 
129     struct io_event io_events[];
130 };
131 
132 /**
133  * io_getevents_peek:
134  * @ctx: AIO context
135  * @events: pointer on events array, output value
136 
137  * Returns the number of completed events and sets a pointer
138  * on events array.  This function does not update the internal
139  * ring buffer, only reads head and tail.  When @events has been
140  * processed io_getevents_commit() must be called.
141  */
142 static inline unsigned int io_getevents_peek(io_context_t ctx,
143                                              struct io_event **events)
144 {
145     struct aio_ring *ring = (struct aio_ring *)ctx;
146     unsigned int head = ring->head, tail = ring->tail;
147     unsigned int nr;
148 
149     nr = tail >= head ? tail - head : ring->nr - head;
150     *events = ring->io_events + head;
151     /* To avoid speculative loads of s->events[i] before observing tail.
152        Paired with smp_wmb() inside linux/fs/aio.c: aio_complete(). */
153     smp_rmb();
154 
155     return nr;
156 }
157 
158 /**
159  * io_getevents_commit:
160  * @ctx: AIO context
161  * @nr: the number of events on which head should be advanced
162  *
163  * Advances head of a ring buffer.
164  */
165 static inline void io_getevents_commit(io_context_t ctx, unsigned int nr)
166 {
167     struct aio_ring *ring = (struct aio_ring *)ctx;
168 
169     if (nr) {
170         ring->head = (ring->head + nr) % ring->nr;
171     }
172 }
173 
174 /**
175  * io_getevents_advance_and_peek:
176  * @ctx: AIO context
177  * @events: pointer on events array, output value
178  * @nr: the number of events on which head should be advanced
179  *
180  * Advances head of a ring buffer and returns number of elements left.
181  */
182 static inline unsigned int
183 io_getevents_advance_and_peek(io_context_t ctx,
184                               struct io_event **events,
185                               unsigned int nr)
186 {
187     io_getevents_commit(ctx, nr);
188     return io_getevents_peek(ctx, events);
189 }
190 
191 /**
192  * qemu_laio_process_completions:
193  * @s: AIO state
194  *
195  * Fetches completed I/O requests and invokes their callbacks.
196  *
197  * The function is somewhat tricky because it supports nested event loops, for
198  * example when a request callback invokes aio_poll().  In order to do this,
199  * indices are kept in LinuxAioState.  Function schedules BH completion so it
200  * can be called again in a nested event loop.  When there are no events left
201  * to complete the BH is being canceled.
202  */
203 static void qemu_laio_process_completions(LinuxAioState *s)
204 {
205     struct io_event *events;
206 
207     /* Reschedule so nested event loops see currently pending completions */
208     qemu_bh_schedule(s->completion_bh);
209 
210     while ((s->event_max = io_getevents_advance_and_peek(s->ctx, &events,
211                                                          s->event_idx))) {
212         for (s->event_idx = 0; s->event_idx < s->event_max; ) {
213             struct iocb *iocb = events[s->event_idx].obj;
214             struct qemu_laiocb *laiocb =
215                 container_of(iocb, struct qemu_laiocb, iocb);
216 
217             laiocb->ret = io_event_ret(&events[s->event_idx]);
218 
219             /* Change counters one-by-one because we can be nested. */
220             s->io_q.in_flight--;
221             s->event_idx++;
222             qemu_laio_process_completion(laiocb);
223         }
224     }
225 
226     qemu_bh_cancel(s->completion_bh);
227 
228     /* If we are nested we have to notify the level above that we are done
229      * by setting event_max to zero, upper level will then jump out of it's
230      * own `for` loop.  If we are the last all counters dropped to zero. */
231     s->event_max = 0;
232     s->event_idx = 0;
233 }
234 
235 static void qemu_laio_process_completions_and_submit(LinuxAioState *s)
236 {
237     qemu_laio_process_completions(s);
238 
239     if (!QSIMPLEQ_EMPTY(&s->io_q.pending)) {
240         ioq_submit(s);
241     }
242 }
243 
244 static void qemu_laio_completion_bh(void *opaque)
245 {
246     LinuxAioState *s = opaque;
247 
248     qemu_laio_process_completions_and_submit(s);
249 }
250 
251 static void qemu_laio_completion_cb(EventNotifier *e)
252 {
253     LinuxAioState *s = container_of(e, LinuxAioState, e);
254 
255     if (event_notifier_test_and_clear(&s->e)) {
256         qemu_laio_process_completions_and_submit(s);
257     }
258 }
259 
260 static bool qemu_laio_poll_cb(void *opaque)
261 {
262     EventNotifier *e = opaque;
263     LinuxAioState *s = container_of(e, LinuxAioState, e);
264     struct io_event *events;
265 
266     return io_getevents_peek(s->ctx, &events);
267 }
268 
269 static void qemu_laio_poll_ready(EventNotifier *opaque)
270 {
271     EventNotifier *e = opaque;
272     LinuxAioState *s = container_of(e, LinuxAioState, e);
273 
274     qemu_laio_process_completions_and_submit(s);
275 }
276 
277 static void ioq_init(LaioQueue *io_q)
278 {
279     QSIMPLEQ_INIT(&io_q->pending);
280     io_q->in_queue = 0;
281     io_q->in_flight = 0;
282     io_q->blocked = false;
283 }
284 
285 static void ioq_submit(LinuxAioState *s)
286 {
287     int ret, len;
288     struct qemu_laiocb *aiocb;
289     struct iocb *iocbs[MAX_EVENTS];
290     QSIMPLEQ_HEAD(, qemu_laiocb) completed;
291 
292     do {
293         if (s->io_q.in_flight >= MAX_EVENTS) {
294             break;
295         }
296         len = 0;
297         QSIMPLEQ_FOREACH(aiocb, &s->io_q.pending, next) {
298             iocbs[len++] = &aiocb->iocb;
299             if (s->io_q.in_flight + len >= MAX_EVENTS) {
300                 break;
301             }
302         }
303 
304         ret = io_submit(s->ctx, len, iocbs);
305         if (ret == -EAGAIN) {
306             break;
307         }
308         if (ret < 0) {
309             /* Fail the first request, retry the rest */
310             aiocb = QSIMPLEQ_FIRST(&s->io_q.pending);
311             QSIMPLEQ_REMOVE_HEAD(&s->io_q.pending, next);
312             s->io_q.in_queue--;
313             aiocb->ret = ret;
314             qemu_laio_process_completion(aiocb);
315             continue;
316         }
317 
318         s->io_q.in_flight += ret;
319         s->io_q.in_queue  -= ret;
320         aiocb = container_of(iocbs[ret - 1], struct qemu_laiocb, iocb);
321         QSIMPLEQ_SPLIT_AFTER(&s->io_q.pending, aiocb, next, &completed);
322     } while (ret == len && !QSIMPLEQ_EMPTY(&s->io_q.pending));
323     s->io_q.blocked = (s->io_q.in_queue > 0);
324 
325     if (s->io_q.in_flight) {
326         /* We can try to complete something just right away if there are
327          * still requests in-flight. */
328         qemu_laio_process_completions(s);
329         /*
330          * Even we have completed everything (in_flight == 0), the queue can
331          * have still pended requests (in_queue > 0).  We do not attempt to
332          * repeat submission to avoid IO hang.  The reason is simple: s->e is
333          * still set and completion callback will be called shortly and all
334          * pended requests will be submitted from there.
335          */
336     }
337 }
338 
339 static uint64_t laio_max_batch(LinuxAioState *s, uint64_t dev_max_batch)
340 {
341     uint64_t max_batch = s->aio_context->aio_max_batch ?: DEFAULT_MAX_BATCH;
342 
343     /*
344      * AIO context can be shared between multiple block devices, so
345      * `dev_max_batch` allows reducing the batch size for latency-sensitive
346      * devices.
347      */
348     max_batch = MIN_NON_ZERO(dev_max_batch, max_batch);
349 
350     /* limit the batch with the number of available events */
351     max_batch = MIN_NON_ZERO(MAX_EVENTS - s->io_q.in_flight, max_batch);
352 
353     return max_batch;
354 }
355 
356 static void laio_unplug_fn(void *opaque)
357 {
358     LinuxAioState *s = opaque;
359 
360     if (!s->io_q.blocked && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
361         ioq_submit(s);
362     }
363 }
364 
365 static int laio_do_submit(int fd, struct qemu_laiocb *laiocb, off_t offset,
366                           int type, uint64_t dev_max_batch)
367 {
368     LinuxAioState *s = laiocb->ctx;
369     struct iocb *iocbs = &laiocb->iocb;
370     QEMUIOVector *qiov = laiocb->qiov;
371 
372     switch (type) {
373     case QEMU_AIO_WRITE:
374         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
375         break;
376     case QEMU_AIO_ZONE_APPEND:
377         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
378         break;
379     case QEMU_AIO_READ:
380         io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
381         break;
382     /* Currently Linux kernel does not support other operations */
383     default:
384         fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
385                         __func__, type);
386         return -EIO;
387     }
388     io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
389 
390     QSIMPLEQ_INSERT_TAIL(&s->io_q.pending, laiocb, next);
391     s->io_q.in_queue++;
392     if (!s->io_q.blocked) {
393         if (s->io_q.in_queue >= laio_max_batch(s, dev_max_batch)) {
394             ioq_submit(s);
395         } else {
396             blk_io_plug_call(laio_unplug_fn, s);
397         }
398     }
399 
400     return 0;
401 }
402 
403 int coroutine_fn laio_co_submit(int fd, uint64_t offset, QEMUIOVector *qiov,
404                                 int type, uint64_t dev_max_batch)
405 {
406     int ret;
407     AioContext *ctx = qemu_get_current_aio_context();
408     struct qemu_laiocb laiocb = {
409         .co         = qemu_coroutine_self(),
410         .nbytes     = qiov->size,
411         .ctx        = aio_get_linux_aio(ctx),
412         .ret        = -EINPROGRESS,
413         .is_read    = (type == QEMU_AIO_READ),
414         .qiov       = qiov,
415     };
416 
417     ret = laio_do_submit(fd, &laiocb, offset, type, dev_max_batch);
418     if (ret < 0) {
419         return ret;
420     }
421 
422     if (laiocb.ret == -EINPROGRESS) {
423         qemu_coroutine_yield();
424     }
425     return laiocb.ret;
426 }
427 
428 void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context)
429 {
430     aio_set_event_notifier(old_context, &s->e, NULL, NULL, NULL);
431     qemu_bh_delete(s->completion_bh);
432     s->aio_context = NULL;
433 }
434 
435 void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context)
436 {
437     s->aio_context = new_context;
438     s->completion_bh = aio_bh_new(new_context, qemu_laio_completion_bh, s);
439     aio_set_event_notifier(new_context, &s->e,
440                            qemu_laio_completion_cb,
441                            qemu_laio_poll_cb,
442                            qemu_laio_poll_ready);
443 }
444 
445 LinuxAioState *laio_init(Error **errp)
446 {
447     int rc;
448     LinuxAioState *s;
449 
450     s = g_malloc0(sizeof(*s));
451     rc = event_notifier_init(&s->e, false);
452     if (rc < 0) {
453         error_setg_errno(errp, -rc, "failed to initialize event notifier");
454         goto out_free_state;
455     }
456 
457     rc = io_setup(MAX_EVENTS, &s->ctx);
458     if (rc < 0) {
459         error_setg_errno(errp, -rc, "failed to create linux AIO context");
460         goto out_close_efd;
461     }
462 
463     ioq_init(&s->io_q);
464 
465     return s;
466 
467 out_close_efd:
468     event_notifier_cleanup(&s->e);
469 out_free_state:
470     g_free(s);
471     return NULL;
472 }
473 
474 void laio_cleanup(LinuxAioState *s)
475 {
476     event_notifier_cleanup(&s->e);
477 
478     if (io_destroy(s->ctx) != 0) {
479         fprintf(stderr, "%s: destroy AIO context %p failed\n",
480                         __func__, &s->ctx);
481     }
482     g_free(s);
483 }
484