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>
25 #include "qemu-queue.h"
27 #include "qemu-common.h"
28 #include "block_int.h"
30 #include "block/raw-posix-aio.h"
34 BlockDriverAIOCB common;
37 struct iovec *aio_iov;
42 #define aio_ioctl_cmd aio_nbytes /* for QEMU_AIO_IOCTL */
46 QTAILQ_ENTRY(qemu_paiocb) node;
50 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 QTAILQ_HEAD(, qemu_paiocb) request_list;
71 static int preadv_present = 1;
73 static int preadv_present = 0;
76 static void die2(int err, const char *what)
78 fprintf(stderr, "%s failed: %s\n", what, strerror(err));
82 static void die(const char *what)
87 static void mutex_lock(pthread_mutex_t *mutex)
89 int ret = pthread_mutex_lock(mutex);
90 if (ret) die2(ret, "pthread_mutex_lock");
93 static void mutex_unlock(pthread_mutex_t *mutex)
95 int ret = pthread_mutex_unlock(mutex);
96 if (ret) die2(ret, "pthread_mutex_unlock");
99 static int cond_timedwait(pthread_cond_t *cond, pthread_mutex_t *mutex,
102 int ret = pthread_cond_timedwait(cond, mutex, ts);
103 if (ret && ret != ETIMEDOUT) die2(ret, "pthread_cond_timedwait");
107 static void cond_signal(pthread_cond_t *cond)
109 int ret = pthread_cond_signal(cond);
110 if (ret) die2(ret, "pthread_cond_signal");
113 static void thread_create(pthread_t *thread, pthread_attr_t *attr,
114 void *(*start_routine)(void*), void *arg)
116 int ret = pthread_create(thread, attr, start_routine, arg);
117 if (ret) die2(ret, "pthread_create");
120 static ssize_t handle_aiocb_ioctl(struct qemu_paiocb *aiocb)
124 ret = ioctl(aiocb->aio_fildes, aiocb->aio_ioctl_cmd, aiocb->aio_ioctl_buf);
129 * This looks weird, but the aio code only consideres a request
130 * successfull if it has written the number full number of bytes.
132 * Now we overload aio_nbytes as aio_ioctl_cmd for the ioctl command,
133 * so in fact we return the ioctl command here to make posix_aio_read()
136 return aiocb->aio_nbytes;
139 static ssize_t handle_aiocb_flush(struct qemu_paiocb *aiocb)
143 ret = qemu_fdatasync(aiocb->aio_fildes);
152 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
154 return preadv(fd, iov, nr_iov, offset);
158 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
160 return pwritev(fd, iov, nr_iov, offset);
166 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
172 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
179 static ssize_t handle_aiocb_rw_vector(struct qemu_paiocb *aiocb)
185 if (aiocb->aio_type & QEMU_AIO_WRITE)
186 len = qemu_pwritev(aiocb->aio_fildes,
189 aiocb->aio_offset + offset);
191 len = qemu_preadv(aiocb->aio_fildes,
194 aiocb->aio_offset + offset);
195 } while (len == -1 && errno == EINTR);
202 static ssize_t handle_aiocb_rw_linear(struct qemu_paiocb *aiocb, char *buf)
207 while (offset < aiocb->aio_nbytes) {
208 if (aiocb->aio_type & QEMU_AIO_WRITE)
209 len = pwrite(aiocb->aio_fildes,
210 (const char *)buf + offset,
211 aiocb->aio_nbytes - offset,
212 aiocb->aio_offset + offset);
214 len = pread(aiocb->aio_fildes,
216 aiocb->aio_nbytes - offset,
217 aiocb->aio_offset + offset);
219 if (len == -1 && errno == EINTR)
221 else if (len == -1) {
233 static ssize_t handle_aiocb_rw(struct qemu_paiocb *aiocb)
238 if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
240 * If there is just a single buffer, and it is properly aligned
241 * we can just use plain pread/pwrite without any problems.
243 if (aiocb->aio_niov == 1)
244 return handle_aiocb_rw_linear(aiocb, aiocb->aio_iov->iov_base);
247 * We have more than one iovec, and all are properly aligned.
249 * Try preadv/pwritev first and fall back to linearizing the
250 * buffer if it's not supported.
252 if (preadv_present) {
253 nbytes = handle_aiocb_rw_vector(aiocb);
254 if (nbytes == aiocb->aio_nbytes)
256 if (nbytes < 0 && nbytes != -ENOSYS)
262 * XXX(hch): short read/write. no easy way to handle the reminder
263 * using these interfaces. For now retry using plain
269 * Ok, we have to do it the hard way, copy all segments into
270 * a single aligned buffer.
272 buf = qemu_memalign(512, aiocb->aio_nbytes);
273 if (aiocb->aio_type & QEMU_AIO_WRITE) {
277 for (i = 0; i < aiocb->aio_niov; ++i) {
278 memcpy(p, aiocb->aio_iov[i].iov_base, aiocb->aio_iov[i].iov_len);
279 p += aiocb->aio_iov[i].iov_len;
283 nbytes = handle_aiocb_rw_linear(aiocb, buf);
284 if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
286 size_t count = aiocb->aio_nbytes, copy;
289 for (i = 0; i < aiocb->aio_niov && count; ++i) {
291 if (copy > aiocb->aio_iov[i].iov_len)
292 copy = aiocb->aio_iov[i].iov_len;
293 memcpy(aiocb->aio_iov[i].iov_base, p, copy);
303 static void *aio_thread(void *unused)
310 struct qemu_paiocb *aiocb;
315 qemu_gettimeofday(&tv);
316 ts.tv_sec = tv.tv_sec + 10;
321 while (QTAILQ_EMPTY(&request_list) &&
322 !(ret == ETIMEDOUT)) {
323 ret = cond_timedwait(&cond, &lock, &ts);
326 if (QTAILQ_EMPTY(&request_list))
329 aiocb = QTAILQ_FIRST(&request_list);
330 QTAILQ_REMOVE(&request_list, aiocb, node);
335 switch (aiocb->aio_type & QEMU_AIO_TYPE_MASK) {
338 ret = handle_aiocb_rw(aiocb);
341 ret = handle_aiocb_flush(aiocb);
344 ret = handle_aiocb_ioctl(aiocb);
347 fprintf(stderr, "invalid aio request (0x%x)\n", aiocb->aio_type);
357 if (kill(pid, aiocb->ev_signo)) die("kill failed");
367 static void spawn_thread(void)
369 sigset_t set, oldset;
374 /* block all signals */
375 if (sigfillset(&set)) die("sigfillset");
376 if (sigprocmask(SIG_SETMASK, &set, &oldset)) die("sigprocmask");
378 thread_create(&thread_id, &attr, aio_thread, NULL);
380 if (sigprocmask(SIG_SETMASK, &oldset, NULL)) die("sigprocmask restore");
383 static void qemu_paio_submit(struct qemu_paiocb *aiocb)
385 aiocb->ret = -EINPROGRESS;
388 if (idle_threads == 0 && cur_threads < max_threads)
390 QTAILQ_INSERT_TAIL(&request_list, aiocb, node);
395 static ssize_t qemu_paio_return(struct qemu_paiocb *aiocb)
406 static int qemu_paio_error(struct qemu_paiocb *aiocb)
408 ssize_t ret = qemu_paio_return(aiocb);
418 static int posix_aio_process_queue(void *opaque)
420 PosixAioState *s = opaque;
421 struct qemu_paiocb *acb, **pacb;
424 int async_context_id = get_async_context_id();
427 pacb = &s->first_aio;
433 /* we're only interested in requests in the right context */
434 if (acb->async_context_id != async_context_id) {
439 ret = qemu_paio_error(acb);
440 if (ret == ECANCELED) {
441 /* remove the request */
443 qemu_aio_release(acb);
445 } else if (ret != EINPROGRESS) {
448 ret = qemu_paio_return(acb);
449 if (ret == acb->aio_nbytes)
456 /* remove the request */
458 /* call the callback */
459 acb->common.cb(acb->common.opaque, ret);
460 qemu_aio_release(acb);
472 static void posix_aio_read(void *opaque)
474 PosixAioState *s = opaque;
477 /* read all bytes from signal pipe */
481 len = read(s->rfd, bytes, sizeof(bytes));
482 if (len == -1 && errno == EINTR)
483 continue; /* try again */
484 if (len == sizeof(bytes))
485 continue; /* more to read */
489 posix_aio_process_queue(s);
492 static int posix_aio_flush(void *opaque)
494 PosixAioState *s = opaque;
495 return !!s->first_aio;
498 static PosixAioState *posix_aio_state;
500 static void aio_signal_handler(int signum)
502 if (posix_aio_state) {
506 ret = write(posix_aio_state->wfd, &byte, sizeof(byte));
507 if (ret < 0 && errno != EAGAIN)
514 static void paio_remove(struct qemu_paiocb *acb)
516 struct qemu_paiocb **pacb;
518 /* remove the callback from the queue */
519 pacb = &posix_aio_state->first_aio;
522 fprintf(stderr, "paio_remove: aio request not found!\n");
524 } else if (*pacb == acb) {
526 qemu_aio_release(acb);
529 pacb = &(*pacb)->next;
533 static void paio_cancel(BlockDriverAIOCB *blockacb)
535 struct qemu_paiocb *acb = (struct qemu_paiocb *)blockacb;
540 QTAILQ_REMOVE(&request_list, acb, node);
541 acb->ret = -ECANCELED;
542 } else if (acb->ret == -EINPROGRESS) {
548 /* fail safe: if the aio could not be canceled, we wait for
550 while (qemu_paio_error(acb) == EINPROGRESS)
557 static AIOPool raw_aio_pool = {
558 .aiocb_size = sizeof(struct qemu_paiocb),
559 .cancel = paio_cancel,
562 BlockDriverAIOCB *paio_submit(BlockDriverState *bs, int fd,
563 int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
564 BlockDriverCompletionFunc *cb, void *opaque, int type)
566 struct qemu_paiocb *acb;
568 acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
571 acb->aio_type = type;
572 acb->aio_fildes = fd;
573 acb->ev_signo = SIGUSR2;
574 acb->async_context_id = get_async_context_id();
577 acb->aio_iov = qiov->iov;
578 acb->aio_niov = qiov->niov;
580 acb->aio_nbytes = nb_sectors * 512;
581 acb->aio_offset = sector_num * 512;
583 acb->next = posix_aio_state->first_aio;
584 posix_aio_state->first_aio = acb;
586 qemu_paio_submit(acb);
590 BlockDriverAIOCB *paio_ioctl(BlockDriverState *bs, int fd,
591 unsigned long int req, void *buf,
592 BlockDriverCompletionFunc *cb, void *opaque)
594 struct qemu_paiocb *acb;
596 acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
599 acb->aio_type = QEMU_AIO_IOCTL;
600 acb->aio_fildes = fd;
601 acb->ev_signo = SIGUSR2;
603 acb->aio_ioctl_buf = buf;
604 acb->aio_ioctl_cmd = req;
606 acb->next = posix_aio_state->first_aio;
607 posix_aio_state->first_aio = acb;
609 qemu_paio_submit(acb);
615 struct sigaction act;
623 s = qemu_malloc(sizeof(PosixAioState));
625 sigfillset(&act.sa_mask);
626 act.sa_flags = 0; /* do not restart syscalls to interrupt select() */
627 act.sa_handler = aio_signal_handler;
628 sigaction(SIGUSR2, &act, NULL);
631 if (qemu_pipe(fds) == -1) {
632 fprintf(stderr, "failed to create pipe\n");
639 fcntl(s->rfd, F_SETFL, O_NONBLOCK);
640 fcntl(s->wfd, F_SETFL, O_NONBLOCK);
642 qemu_aio_set_fd_handler(s->rfd, posix_aio_read, NULL, posix_aio_flush,
643 posix_aio_process_queue, s);
645 ret = pthread_attr_init(&attr);
647 die2(ret, "pthread_attr_init");
649 ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
651 die2(ret, "pthread_attr_setdetachstate");
653 QTAILQ_INIT(&request_list);