xref: /openbmc/qemu/block/linux-aio.c (revision d525f73f)
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 
19 #include <libaio.h>
20 
21 /*
22  * Queue size (per-device).
23  *
24  * XXX: eventually we need to communicate this to the guest and/or make it
25  *      tunable by the guest.  If we get more outstanding requests at a time
26  *      than this we will get EAGAIN from io_submit which is communicated to
27  *      the guest as an I/O error.
28  */
29 #define MAX_EVENTS 1024
30 
31 /* Maximum number of requests in a batch. (default value) */
32 #define DEFAULT_MAX_BATCH 32
33 
34 struct qemu_laiocb {
35     Coroutine *co;
36     LinuxAioState *ctx;
37     struct iocb iocb;
38     ssize_t ret;
39     size_t nbytes;
40     QEMUIOVector *qiov;
41     bool is_read;
42     QSIMPLEQ_ENTRY(qemu_laiocb) next;
43 };
44 
45 typedef struct {
46     int plugged;
47     unsigned int in_queue;
48     unsigned int in_flight;
49     bool blocked;
50     QSIMPLEQ_HEAD(, qemu_laiocb) pending;
51 } LaioQueue;
52 
53 struct LinuxAioState {
54     AioContext *aio_context;
55 
56     io_context_t ctx;
57     EventNotifier e;
58 
59     /* io queue for submit at batch.  Protected by AioContext lock. */
60     LaioQueue io_q;
61 
62     /* I/O completion processing.  Only runs in I/O thread.  */
63     QEMUBH *completion_bh;
64     int event_idx;
65     int event_max;
66 };
67 
68 static void ioq_submit(LinuxAioState *s);
69 
70 static inline ssize_t io_event_ret(struct io_event *ev)
71 {
72     return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
73 }
74 
75 /*
76  * Completes an AIO request.
77  */
78 static void qemu_laio_process_completion(struct qemu_laiocb *laiocb)
79 {
80     int ret;
81 
82     ret = laiocb->ret;
83     if (ret != -ECANCELED) {
84         if (ret == laiocb->nbytes) {
85             ret = 0;
86         } else if (ret >= 0) {
87             /* Short reads mean EOF, pad with zeros. */
88             if (laiocb->is_read) {
89                 qemu_iovec_memset(laiocb->qiov, ret, 0,
90                     laiocb->qiov->size - ret);
91             } else {
92                 ret = -ENOSPC;
93             }
94         }
95     }
96 
97     laiocb->ret = ret;
98 
99     /*
100      * If the coroutine is already entered it must be in ioq_submit() and
101      * will notice laio->ret has been filled in when it eventually runs
102      * later.  Coroutines cannot be entered recursively so avoid doing
103      * that!
104      */
105     if (!qemu_coroutine_entered(laiocb->co)) {
106         aio_co_wake(laiocb->co);
107     }
108 }
109 
110 /**
111  * aio_ring buffer which is shared between userspace and kernel.
112  *
113  * This copied from linux/fs/aio.c, common header does not exist
114  * but AIO exists for ages so we assume ABI is stable.
115  */
116 struct aio_ring {
117     unsigned    id;    /* kernel internal index number */
118     unsigned    nr;    /* number of io_events */
119     unsigned    head;  /* Written to by userland or by kernel. */
120     unsigned    tail;
121 
122     unsigned    magic;
123     unsigned    compat_features;
124     unsigned    incompat_features;
125     unsigned    header_length;  /* size of aio_ring */
126 
127     struct io_event io_events[];
128 };
129 
130 /**
131  * io_getevents_peek:
132  * @ctx: AIO context
133  * @events: pointer on events array, output value
134 
135  * Returns the number of completed events and sets a pointer
136  * on events array.  This function does not update the internal
137  * ring buffer, only reads head and tail.  When @events has been
138  * processed io_getevents_commit() must be called.
139  */
140 static inline unsigned int io_getevents_peek(io_context_t ctx,
141                                              struct io_event **events)
142 {
143     struct aio_ring *ring = (struct aio_ring *)ctx;
144     unsigned int head = ring->head, tail = ring->tail;
145     unsigned int nr;
146 
147     nr = tail >= head ? tail - head : ring->nr - head;
148     *events = ring->io_events + head;
149     /* To avoid speculative loads of s->events[i] before observing tail.
150        Paired with smp_wmb() inside linux/fs/aio.c: aio_complete(). */
151     smp_rmb();
152 
153     return nr;
154 }
155 
156 /**
157  * io_getevents_commit:
158  * @ctx: AIO context
159  * @nr: the number of events on which head should be advanced
160  *
161  * Advances head of a ring buffer.
162  */
163 static inline void io_getevents_commit(io_context_t ctx, unsigned int nr)
164 {
165     struct aio_ring *ring = (struct aio_ring *)ctx;
166 
167     if (nr) {
168         ring->head = (ring->head + nr) % ring->nr;
169     }
170 }
171 
172 /**
173  * io_getevents_advance_and_peek:
174  * @ctx: AIO context
175  * @events: pointer on events array, output value
176  * @nr: the number of events on which head should be advanced
177  *
178  * Advances head of a ring buffer and returns number of elements left.
179  */
180 static inline unsigned int
181 io_getevents_advance_and_peek(io_context_t ctx,
182                               struct io_event **events,
183                               unsigned int nr)
184 {
185     io_getevents_commit(ctx, nr);
186     return io_getevents_peek(ctx, events);
187 }
188 
189 /**
190  * qemu_laio_process_completions:
191  * @s: AIO state
192  *
193  * Fetches completed I/O requests and invokes their callbacks.
194  *
195  * The function is somewhat tricky because it supports nested event loops, for
196  * example when a request callback invokes aio_poll().  In order to do this,
197  * indices are kept in LinuxAioState.  Function schedules BH completion so it
198  * can be called again in a nested event loop.  When there are no events left
199  * to complete the BH is being canceled.
200  */
201 static void qemu_laio_process_completions(LinuxAioState *s)
202 {
203     struct io_event *events;
204 
205     /* Reschedule so nested event loops see currently pending completions */
206     qemu_bh_schedule(s->completion_bh);
207 
208     while ((s->event_max = io_getevents_advance_and_peek(s->ctx, &events,
209                                                          s->event_idx))) {
210         for (s->event_idx = 0; s->event_idx < s->event_max; ) {
211             struct iocb *iocb = events[s->event_idx].obj;
212             struct qemu_laiocb *laiocb =
213                 container_of(iocb, struct qemu_laiocb, iocb);
214 
215             laiocb->ret = io_event_ret(&events[s->event_idx]);
216 
217             /* Change counters one-by-one because we can be nested. */
218             s->io_q.in_flight--;
219             s->event_idx++;
220             qemu_laio_process_completion(laiocb);
221         }
222     }
223 
224     qemu_bh_cancel(s->completion_bh);
225 
226     /* If we are nested we have to notify the level above that we are done
227      * by setting event_max to zero, upper level will then jump out of it's
228      * own `for` loop.  If we are the last all counters droped to zero. */
229     s->event_max = 0;
230     s->event_idx = 0;
231 }
232 
233 static void qemu_laio_process_completions_and_submit(LinuxAioState *s)
234 {
235     aio_context_acquire(s->aio_context);
236     qemu_laio_process_completions(s);
237 
238     if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
239         ioq_submit(s);
240     }
241     aio_context_release(s->aio_context);
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->plugged = 0;
281     io_q->in_queue = 0;
282     io_q->in_flight = 0;
283     io_q->blocked = false;
284 }
285 
286 static void ioq_submit(LinuxAioState *s)
287 {
288     int ret, len;
289     struct qemu_laiocb *aiocb;
290     struct iocb *iocbs[MAX_EVENTS];
291     QSIMPLEQ_HEAD(, qemu_laiocb) completed;
292 
293     do {
294         if (s->io_q.in_flight >= MAX_EVENTS) {
295             break;
296         }
297         len = 0;
298         QSIMPLEQ_FOREACH(aiocb, &s->io_q.pending, next) {
299             iocbs[len++] = &aiocb->iocb;
300             if (s->io_q.in_flight + len >= MAX_EVENTS) {
301                 break;
302             }
303         }
304 
305         ret = io_submit(s->ctx, len, iocbs);
306         if (ret == -EAGAIN) {
307             break;
308         }
309         if (ret < 0) {
310             /* Fail the first request, retry the rest */
311             aiocb = QSIMPLEQ_FIRST(&s->io_q.pending);
312             QSIMPLEQ_REMOVE_HEAD(&s->io_q.pending, next);
313             s->io_q.in_queue--;
314             aiocb->ret = ret;
315             qemu_laio_process_completion(aiocb);
316             continue;
317         }
318 
319         s->io_q.in_flight += ret;
320         s->io_q.in_queue  -= ret;
321         aiocb = container_of(iocbs[ret - 1], struct qemu_laiocb, iocb);
322         QSIMPLEQ_SPLIT_AFTER(&s->io_q.pending, aiocb, next, &completed);
323     } while (ret == len && !QSIMPLEQ_EMPTY(&s->io_q.pending));
324     s->io_q.blocked = (s->io_q.in_queue > 0);
325 
326     if (s->io_q.in_flight) {
327         /* We can try to complete something just right away if there are
328          * still requests in-flight. */
329         qemu_laio_process_completions(s);
330         /*
331          * Even we have completed everything (in_flight == 0), the queue can
332          * have still pended requests (in_queue > 0).  We do not attempt to
333          * repeat submission to avoid IO hang.  The reason is simple: s->e is
334          * still set and completion callback will be called shortly and all
335          * pended requests will be submitted from there.
336          */
337     }
338 }
339 
340 static uint64_t laio_max_batch(LinuxAioState *s, uint64_t dev_max_batch)
341 {
342     uint64_t max_batch = s->aio_context->aio_max_batch ?: DEFAULT_MAX_BATCH;
343 
344     /*
345      * AIO context can be shared between multiple block devices, so
346      * `dev_max_batch` allows reducing the batch size for latency-sensitive
347      * devices.
348      */
349     max_batch = MIN_NON_ZERO(dev_max_batch, max_batch);
350 
351     /* limit the batch with the number of available events */
352     max_batch = MIN_NON_ZERO(MAX_EVENTS - s->io_q.in_flight, max_batch);
353 
354     return max_batch;
355 }
356 
357 void laio_io_plug(BlockDriverState *bs, LinuxAioState *s)
358 {
359     s->io_q.plugged++;
360 }
361 
362 void laio_io_unplug(BlockDriverState *bs, LinuxAioState *s,
363                     uint64_t dev_max_batch)
364 {
365     assert(s->io_q.plugged);
366     s->io_q.plugged--;
367 
368     /*
369      * Why max batch checking is performed here:
370      * Another BDS may have queued requests with a higher dev_max_batch and
371      * therefore in_queue could now exceed our dev_max_batch. Re-check the max
372      * batch so we can honor our device's dev_max_batch.
373      */
374     if (s->io_q.in_queue >= laio_max_batch(s, dev_max_batch) ||
375         (!s->io_q.plugged &&
376          !s->io_q.blocked && !QSIMPLEQ_EMPTY(&s->io_q.pending))) {
377         ioq_submit(s);
378     }
379 }
380 
381 static int laio_do_submit(int fd, struct qemu_laiocb *laiocb, off_t offset,
382                           int type, uint64_t dev_max_batch)
383 {
384     LinuxAioState *s = laiocb->ctx;
385     struct iocb *iocbs = &laiocb->iocb;
386     QEMUIOVector *qiov = laiocb->qiov;
387 
388     switch (type) {
389     case QEMU_AIO_WRITE:
390         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
391         break;
392     case QEMU_AIO_READ:
393         io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
394         break;
395     /* Currently Linux kernel does not support other operations */
396     default:
397         fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
398                         __func__, type);
399         return -EIO;
400     }
401     io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
402 
403     QSIMPLEQ_INSERT_TAIL(&s->io_q.pending, laiocb, next);
404     s->io_q.in_queue++;
405     if (!s->io_q.blocked &&
406         (!s->io_q.plugged ||
407          s->io_q.in_queue >= laio_max_batch(s, dev_max_batch))) {
408         ioq_submit(s);
409     }
410 
411     return 0;
412 }
413 
414 int coroutine_fn laio_co_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
415                                 uint64_t offset, QEMUIOVector *qiov, int type,
416                                 uint64_t dev_max_batch)
417 {
418     int ret;
419     struct qemu_laiocb laiocb = {
420         .co         = qemu_coroutine_self(),
421         .nbytes     = qiov->size,
422         .ctx        = s,
423         .ret        = -EINPROGRESS,
424         .is_read    = (type == QEMU_AIO_READ),
425         .qiov       = qiov,
426     };
427 
428     ret = laio_do_submit(fd, &laiocb, offset, type, dev_max_batch);
429     if (ret < 0) {
430         return ret;
431     }
432 
433     if (laiocb.ret == -EINPROGRESS) {
434         qemu_coroutine_yield();
435     }
436     return laiocb.ret;
437 }
438 
439 void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context)
440 {
441     aio_set_event_notifier(old_context, &s->e, false, NULL, NULL, NULL);
442     qemu_bh_delete(s->completion_bh);
443     s->aio_context = NULL;
444 }
445 
446 void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context)
447 {
448     s->aio_context = new_context;
449     s->completion_bh = aio_bh_new(new_context, qemu_laio_completion_bh, s);
450     aio_set_event_notifier(new_context, &s->e, false,
451                            qemu_laio_completion_cb,
452                            qemu_laio_poll_cb,
453                            qemu_laio_poll_ready);
454 }
455 
456 LinuxAioState *laio_init(Error **errp)
457 {
458     int rc;
459     LinuxAioState *s;
460 
461     s = g_malloc0(sizeof(*s));
462     rc = event_notifier_init(&s->e, false);
463     if (rc < 0) {
464         error_setg_errno(errp, -rc, "failed to initialize event notifier");
465         goto out_free_state;
466     }
467 
468     rc = io_setup(MAX_EVENTS, &s->ctx);
469     if (rc < 0) {
470         error_setg_errno(errp, -rc, "failed to create linux AIO context");
471         goto out_close_efd;
472     }
473 
474     ioq_init(&s->io_q);
475 
476     return s;
477 
478 out_close_efd:
479     event_notifier_cleanup(&s->e);
480 out_free_state:
481     g_free(s);
482     return NULL;
483 }
484 
485 void laio_cleanup(LinuxAioState *s)
486 {
487     event_notifier_cleanup(&s->e);
488 
489     if (io_destroy(s->ctx) != 0) {
490         fprintf(stderr, "%s: destroy AIO context %p failed\n",
491                         __func__, &s->ctx);
492     }
493     g_free(s);
494 }
495