2 * QEMU posix-aio emulation
4 * Copyright IBM, Corp. 2008
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
14 #include <sys/ioctl.h>
15 #include <sys/types.h>
24 #include "qemu-queue.h"
27 #include "qemu-common.h"
29 #include "block_int.h"
31 #include "block/raw-posix-aio.h"
33 static void do_spawn_thread(void);
36 BlockDriverAIOCB common;
39 struct iovec *aio_iov;
44 #define aio_ioctl_cmd aio_nbytes /* for QEMU_AIO_IOCTL */
48 QTAILQ_ENTRY(qemu_paiocb) node;
52 struct qemu_paiocb *next;
55 typedef struct PosixAioState {
57 struct qemu_paiocb *first_aio;
61 static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
62 static pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
63 static pthread_t thread_id;
64 static pthread_attr_t attr;
65 static int max_threads = 64;
66 static int cur_threads = 0;
67 static int idle_threads = 0;
68 static int new_threads = 0; /* backlog of threads we need to create */
69 static int pending_threads = 0; /* threads created but not running yet */
70 static QEMUBH *new_thread_bh;
71 static QTAILQ_HEAD(, qemu_paiocb) request_list;
74 static int preadv_present = 1;
76 static int preadv_present = 0;
79 static void die2(int err, const char *what)
81 fprintf(stderr, "%s failed: %s\n", what, strerror(err));
85 static void die(const char *what)
90 static void mutex_lock(pthread_mutex_t *mutex)
92 int ret = pthread_mutex_lock(mutex);
93 if (ret) die2(ret, "pthread_mutex_lock");
96 static void mutex_unlock(pthread_mutex_t *mutex)
98 int ret = pthread_mutex_unlock(mutex);
99 if (ret) die2(ret, "pthread_mutex_unlock");
102 static int cond_timedwait(pthread_cond_t *cond, pthread_mutex_t *mutex,
105 int ret = pthread_cond_timedwait(cond, mutex, ts);
106 if (ret && ret != ETIMEDOUT) die2(ret, "pthread_cond_timedwait");
110 static void cond_signal(pthread_cond_t *cond)
112 int ret = pthread_cond_signal(cond);
113 if (ret) die2(ret, "pthread_cond_signal");
116 static void thread_create(pthread_t *thread, pthread_attr_t *attr,
117 void *(*start_routine)(void*), void *arg)
119 int ret = pthread_create(thread, attr, start_routine, arg);
120 if (ret) die2(ret, "pthread_create");
123 static ssize_t handle_aiocb_ioctl(struct qemu_paiocb *aiocb)
127 ret = ioctl(aiocb->aio_fildes, aiocb->aio_ioctl_cmd, aiocb->aio_ioctl_buf);
132 * This looks weird, but the aio code only consideres a request
133 * successful if it has written the number full number of bytes.
135 * Now we overload aio_nbytes as aio_ioctl_cmd for the ioctl command,
136 * so in fact we return the ioctl command here to make posix_aio_read()
139 return aiocb->aio_nbytes;
142 static ssize_t handle_aiocb_flush(struct qemu_paiocb *aiocb)
146 ret = qemu_fdatasync(aiocb->aio_fildes);
155 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
157 return preadv(fd, iov, nr_iov, offset);
161 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
163 return pwritev(fd, iov, nr_iov, offset);
169 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
175 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
182 static ssize_t handle_aiocb_rw_vector(struct qemu_paiocb *aiocb)
188 if (aiocb->aio_type & QEMU_AIO_WRITE)
189 len = qemu_pwritev(aiocb->aio_fildes,
192 aiocb->aio_offset + offset);
194 len = qemu_preadv(aiocb->aio_fildes,
197 aiocb->aio_offset + offset);
198 } while (len == -1 && errno == EINTR);
206 * Read/writes the data to/from a given linear buffer.
208 * Returns the number of bytes handles or -errno in case of an error. Short
209 * reads are only returned if the end of the file is reached.
211 static ssize_t handle_aiocb_rw_linear(struct qemu_paiocb *aiocb, char *buf)
216 while (offset < aiocb->aio_nbytes) {
217 if (aiocb->aio_type & QEMU_AIO_WRITE)
218 len = pwrite(aiocb->aio_fildes,
219 (const char *)buf + offset,
220 aiocb->aio_nbytes - offset,
221 aiocb->aio_offset + offset);
223 len = pread(aiocb->aio_fildes,
225 aiocb->aio_nbytes - offset,
226 aiocb->aio_offset + offset);
228 if (len == -1 && errno == EINTR)
230 else if (len == -1) {
242 static ssize_t handle_aiocb_rw(struct qemu_paiocb *aiocb)
247 if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
249 * If there is just a single buffer, and it is properly aligned
250 * we can just use plain pread/pwrite without any problems.
252 if (aiocb->aio_niov == 1)
253 return handle_aiocb_rw_linear(aiocb, aiocb->aio_iov->iov_base);
256 * We have more than one iovec, and all are properly aligned.
258 * Try preadv/pwritev first and fall back to linearizing the
259 * buffer if it's not supported.
261 if (preadv_present) {
262 nbytes = handle_aiocb_rw_vector(aiocb);
263 if (nbytes == aiocb->aio_nbytes)
265 if (nbytes < 0 && nbytes != -ENOSYS)
271 * XXX(hch): short read/write. no easy way to handle the reminder
272 * using these interfaces. For now retry using plain
278 * Ok, we have to do it the hard way, copy all segments into
279 * a single aligned buffer.
281 buf = qemu_blockalign(aiocb->common.bs, aiocb->aio_nbytes);
282 if (aiocb->aio_type & QEMU_AIO_WRITE) {
286 for (i = 0; i < aiocb->aio_niov; ++i) {
287 memcpy(p, aiocb->aio_iov[i].iov_base, aiocb->aio_iov[i].iov_len);
288 p += aiocb->aio_iov[i].iov_len;
292 nbytes = handle_aiocb_rw_linear(aiocb, buf);
293 if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
295 size_t count = aiocb->aio_nbytes, copy;
298 for (i = 0; i < aiocb->aio_niov && count; ++i) {
300 if (copy > aiocb->aio_iov[i].iov_len)
301 copy = aiocb->aio_iov[i].iov_len;
302 memcpy(aiocb->aio_iov[i].iov_base, p, copy);
312 static void *aio_thread(void *unused)
324 struct qemu_paiocb *aiocb;
329 qemu_gettimeofday(&tv);
330 ts.tv_sec = tv.tv_sec + 10;
335 while (QTAILQ_EMPTY(&request_list) &&
336 !(ret == ETIMEDOUT)) {
338 ret = cond_timedwait(&cond, &lock, &ts);
342 if (QTAILQ_EMPTY(&request_list))
345 aiocb = QTAILQ_FIRST(&request_list);
346 QTAILQ_REMOVE(&request_list, aiocb, node);
350 switch (aiocb->aio_type & QEMU_AIO_TYPE_MASK) {
352 ret = handle_aiocb_rw(aiocb);
353 if (ret >= 0 && ret < aiocb->aio_nbytes && aiocb->common.bs->growable) {
354 /* A short read means that we have reached EOF. Pad the buffer
355 * with zeros for bytes after EOF. */
358 qemu_iovec_init_external(&qiov, aiocb->aio_iov,
360 qemu_iovec_memset_skip(&qiov, 0, aiocb->aio_nbytes - ret, ret);
362 ret = aiocb->aio_nbytes;
366 ret = handle_aiocb_rw(aiocb);
369 ret = handle_aiocb_flush(aiocb);
372 ret = handle_aiocb_ioctl(aiocb);
375 fprintf(stderr, "invalid aio request (0x%x)\n", aiocb->aio_type);
384 if (kill(pid, aiocb->ev_signo)) die("kill failed");
393 static void do_spawn_thread(void)
395 sigset_t set, oldset;
408 /* block all signals */
409 if (sigfillset(&set)) die("sigfillset");
410 if (sigprocmask(SIG_SETMASK, &set, &oldset)) die("sigprocmask");
412 thread_create(&thread_id, &attr, aio_thread, NULL);
414 if (sigprocmask(SIG_SETMASK, &oldset, NULL)) die("sigprocmask restore");
417 static void spawn_thread_bh_fn(void *opaque)
422 static void spawn_thread(void)
426 /* If there are threads being created, they will spawn new workers, so
427 * we don't spend time creating many threads in a loop holding a mutex or
428 * starving the current vcpu.
430 * If there are no idle threads, ask the main thread to create one, so we
431 * inherit the correct affinity instead of the vcpu affinity.
433 if (!pending_threads) {
434 qemu_bh_schedule(new_thread_bh);
438 static void qemu_paio_submit(struct qemu_paiocb *aiocb)
440 aiocb->ret = -EINPROGRESS;
443 if (idle_threads == 0 && cur_threads < max_threads)
445 QTAILQ_INSERT_TAIL(&request_list, aiocb, node);
450 static ssize_t qemu_paio_return(struct qemu_paiocb *aiocb)
461 static int qemu_paio_error(struct qemu_paiocb *aiocb)
463 ssize_t ret = qemu_paio_return(aiocb);
473 static int posix_aio_process_queue(void *opaque)
475 PosixAioState *s = opaque;
476 struct qemu_paiocb *acb, **pacb;
481 pacb = &s->first_aio;
487 ret = qemu_paio_error(acb);
488 if (ret == ECANCELED) {
489 /* remove the request */
491 qemu_aio_release(acb);
493 } else if (ret != EINPROGRESS) {
496 ret = qemu_paio_return(acb);
497 if (ret == acb->aio_nbytes)
505 trace_paio_complete(acb, acb->common.opaque, ret);
507 /* remove the request */
509 /* call the callback */
510 acb->common.cb(acb->common.opaque, ret);
511 qemu_aio_release(acb);
523 static void posix_aio_read(void *opaque)
525 PosixAioState *s = opaque;
528 /* read all bytes from signal pipe */
532 len = read(s->rfd, bytes, sizeof(bytes));
533 if (len == -1 && errno == EINTR)
534 continue; /* try again */
535 if (len == sizeof(bytes))
536 continue; /* more to read */
540 posix_aio_process_queue(s);
543 static int posix_aio_flush(void *opaque)
545 PosixAioState *s = opaque;
546 return !!s->first_aio;
549 static PosixAioState *posix_aio_state;
551 static void aio_signal_handler(int signum)
553 if (posix_aio_state) {
557 ret = write(posix_aio_state->wfd, &byte, sizeof(byte));
558 if (ret < 0 && errno != EAGAIN)
565 static void paio_remove(struct qemu_paiocb *acb)
567 struct qemu_paiocb **pacb;
569 /* remove the callback from the queue */
570 pacb = &posix_aio_state->first_aio;
573 fprintf(stderr, "paio_remove: aio request not found!\n");
575 } else if (*pacb == acb) {
577 qemu_aio_release(acb);
580 pacb = &(*pacb)->next;
584 static void paio_cancel(BlockDriverAIOCB *blockacb)
586 struct qemu_paiocb *acb = (struct qemu_paiocb *)blockacb;
589 trace_paio_cancel(acb, acb->common.opaque);
593 QTAILQ_REMOVE(&request_list, acb, node);
594 acb->ret = -ECANCELED;
595 } else if (acb->ret == -EINPROGRESS) {
601 /* fail safe: if the aio could not be canceled, we wait for
603 while (qemu_paio_error(acb) == EINPROGRESS)
610 static AIOPool raw_aio_pool = {
611 .aiocb_size = sizeof(struct qemu_paiocb),
612 .cancel = paio_cancel,
615 BlockDriverAIOCB *paio_submit(BlockDriverState *bs, int fd,
616 int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
617 BlockDriverCompletionFunc *cb, void *opaque, int type)
619 struct qemu_paiocb *acb;
621 acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
624 acb->aio_type = type;
625 acb->aio_fildes = fd;
626 acb->ev_signo = SIGUSR2;
629 acb->aio_iov = qiov->iov;
630 acb->aio_niov = qiov->niov;
632 acb->aio_nbytes = nb_sectors * 512;
633 acb->aio_offset = sector_num * 512;
635 acb->next = posix_aio_state->first_aio;
636 posix_aio_state->first_aio = acb;
638 trace_paio_submit(acb, opaque, sector_num, nb_sectors, type);
639 qemu_paio_submit(acb);
643 BlockDriverAIOCB *paio_ioctl(BlockDriverState *bs, int fd,
644 unsigned long int req, void *buf,
645 BlockDriverCompletionFunc *cb, void *opaque)
647 struct qemu_paiocb *acb;
649 acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
652 acb->aio_type = QEMU_AIO_IOCTL;
653 acb->aio_fildes = fd;
654 acb->ev_signo = SIGUSR2;
656 acb->aio_ioctl_buf = buf;
657 acb->aio_ioctl_cmd = req;
659 acb->next = posix_aio_state->first_aio;
660 posix_aio_state->first_aio = acb;
662 qemu_paio_submit(acb);
668 struct sigaction act;
676 s = g_malloc(sizeof(PosixAioState));
678 sigfillset(&act.sa_mask);
679 act.sa_flags = 0; /* do not restart syscalls to interrupt select() */
680 act.sa_handler = aio_signal_handler;
681 sigaction(SIGUSR2, &act, NULL);
684 if (qemu_pipe(fds) == -1) {
685 fprintf(stderr, "failed to create pipe\n");
692 fcntl(s->rfd, F_SETFL, O_NONBLOCK);
693 fcntl(s->wfd, F_SETFL, O_NONBLOCK);
695 qemu_aio_set_fd_handler(s->rfd, posix_aio_read, NULL, posix_aio_flush,
696 posix_aio_process_queue, s);
698 ret = pthread_attr_init(&attr);
700 die2(ret, "pthread_attr_init");
702 ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
704 die2(ret, "pthread_attr_setdetachstate");
706 QTAILQ_INIT(&request_list);
707 new_thread_bh = qemu_bh_new(spawn_thread_bh_fn, NULL);