implement POSIX asynchronous io
authorRich Felker <dalias@aerifal.cx>
Fri, 9 Sep 2011 05:07:38 +0000 (01:07 -0400)
committerRich Felker <dalias@aerifal.cx>
Fri, 9 Sep 2011 05:07:38 +0000 (01:07 -0400)
some features are not yet supported, and only minimal testing has been
performed. should be considered experimental at this point.

include/aio.h [new file with mode: 0644]
src/aio/aio_cancel.c [new file with mode: 0644]
src/aio/aio_error.c [new file with mode: 0644]
src/aio/aio_fsync.c [new file with mode: 0644]
src/aio/aio_readwrite.c [new file with mode: 0644]
src/aio/aio_return.c [new file with mode: 0644]
src/aio/aio_suspend.c [new file with mode: 0644]
src/aio/lio_listio.c [new file with mode: 0644]

diff --git a/include/aio.h b/include/aio.h
new file mode 100644 (file)
index 0000000..893c0e9
--- /dev/null
@@ -0,0 +1,60 @@
+#ifndef _AIO_H
+#define _AIO_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(_POSIX_SOURCE) || defined(_POSIX_C_SOURCE) \
+ || defined(_XOPEN_SOURCE) || defined(_GNU_SOURCE)
+
+#include <signal.h>
+#include <time.h>
+
+#define __NEED_ssize_t
+#define __NEED_off_t
+
+#include <bits/alltypes.h>
+
+struct aiocb {
+       int aio_filedes, aio_lio_opcode, aio_reqprio;
+       volatile void *aio_buf;
+       size_t aio_nbytes;
+       struct sigevent aio_sigevent;
+       void *__td;
+       int __lock[2];
+       int __err;
+       ssize_t __ret;
+       off_t aio_offset;
+       void *__next, *__prev;
+       char __dummy4[32-2*sizeof(void *)];
+};
+
+#define AIO_CANCELED 0
+#define AIO_NOTCANCELED 1
+#define AIO_ALLDONE 2
+
+#define LIO_READ 0
+#define LIO_WRITE 1
+#define LIO_NOP 2
+
+#define LIO_WAIT 0
+#define LIO_NOWAIT 1
+
+int aio_read(struct aiocb *);
+int aio_write(struct aiocb *);
+int aio_error(struct aiocb *);
+ssize_t aio_return(struct aiocb *);
+int aio_cancel(int, struct aiocb *);
+int aio_suspend(struct aiocb *const [], int, const struct timespec *);
+int aio_fsync(int, struct aiocb *);
+
+int lio_listio(int, struct aiocb *const [], int, struct sigevent *);
+
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/src/aio/aio_cancel.c b/src/aio/aio_cancel.c
new file mode 100644 (file)
index 0000000..5a753b1
--- /dev/null
@@ -0,0 +1,16 @@
+#include <aio.h>
+#include <pthread.h>
+#include <errno.h>
+
+int aio_cancel(int fd, struct aiocb *cb)
+{
+       if (!cb) {
+               /* FIXME: for correctness, we should return AIO_ALLDONE
+                * if there are no outstanding aio operations on this
+                * file descriptor, but that would require making aio
+                * much slower, and seems to have little advantage since
+                * we don't support cancellation anyway. */
+               return AIO_NOTCANCELED;
+       }
+       return cb->__err==EINPROGRESS ? AIO_NOTCANCELED : AIO_ALLDONE;
+}
diff --git a/src/aio/aio_error.c b/src/aio/aio_error.c
new file mode 100644 (file)
index 0000000..169a9a3
--- /dev/null
@@ -0,0 +1,6 @@
+#include <aio.h>
+
+int aio_error(struct aiocb *cb)
+{
+       return cb->__err;
+}
diff --git a/src/aio/aio_fsync.c b/src/aio/aio_fsync.c
new file mode 100644 (file)
index 0000000..0ac6ea8
--- /dev/null
@@ -0,0 +1,9 @@
+#include <aio.h>
+#include <errno.h>
+
+int aio_fsync(int op, struct aiocb *cb)
+{
+       /* FIXME: unsupported */
+       errno = EINVAL;
+       return -1;
+}
diff --git a/src/aio/aio_readwrite.c b/src/aio/aio_readwrite.c
new file mode 100644 (file)
index 0000000..27168f2
--- /dev/null
@@ -0,0 +1,104 @@
+#include <aio.h>
+#include <fcntl.h>
+#include "pthread_impl.h"
+
+static void dummy(void)
+{
+}
+
+weak_alias(dummy, __aio_wake);
+
+static void notify_signal(struct sigevent *sev)
+{
+       siginfo_t si = {
+               .si_signo = sev->sigev_signo,
+               .si_value = sev->sigev_value,
+               .si_code = SI_ASYNCIO,
+               .si_pid = __pthread_self()->pid,
+               .si_uid = getuid()
+       };
+       __syscall(SYS_rt_sigqueueinfo, si.si_pid, si.si_signo, &si);
+}
+
+static void *io_thread(void *p)
+{
+       struct aiocb *cb = p;
+       int fd = cb->aio_filedes;
+       void *buf = (void *)cb->aio_buf;
+       size_t len = cb->aio_nbytes;
+       off_t off = cb->aio_offset;
+       int op = cb->aio_lio_opcode;
+       struct sigevent sev = cb->aio_sigevent;
+       ssize_t ret;
+
+       if (op == LIO_WRITE) {
+               if (  (fcntl(fd, F_GETFL) & O_APPEND)
+                   ||((ret = pwrite(fd, buf, len, off))<0 && errno==ESPIPE) )
+                       ret = write(fd, buf, len);
+       } else if (op == LIO_READ) {
+               if ( (ret = pread(fd, buf, len, off))<0 && errno==ESPIPE )
+                       ret = read(fd, buf, len);
+       } else {
+               ret = 0;
+       }
+       cb->__ret = ret;
+
+       if (ret < 0) a_store(&cb->__err, errno);
+       else a_store(&cb->__err, 0);
+
+       __aio_wake();
+
+       switch (cb->aio_sigevent.sigev_notify) {
+       case SIGEV_SIGNAL:
+               notify_signal(&sev);
+               break;
+       case SIGEV_THREAD:
+               sev.sigev_notify_function(sev.sigev_value);
+               break;
+       }
+
+       return 0;
+}
+
+static int new_req(struct aiocb *cb)
+{
+       int ret = 0;
+       pthread_attr_t a;
+       sigset_t set;
+       pthread_t td;
+
+       if (cb->aio_sigevent.sigev_notify == SIGEV_THREAD) {
+               if (cb->aio_sigevent.sigev_notify_attributes)
+                       a = *cb->aio_sigevent.sigev_notify_attributes;
+               else
+                       pthread_attr_init(&a);
+       } else {
+               pthread_attr_init(&a);
+               pthread_attr_setstacksize(&a, PAGE_SIZE);
+               pthread_attr_setguardsize(&a, 0);
+       }
+       pthread_attr_setdetachstate(&a, PTHREAD_CREATE_DETACHED);
+       sigfillset(&set);
+       pthread_sigmask(SIG_BLOCK, &set, &set);
+       cb->__err = EINPROGRESS;
+       if (pthread_create(&td, &a, io_thread, cb)) {
+               errno = EAGAIN;
+               ret = -1;
+       }
+       pthread_sigmask(SIG_SETMASK, &set, 0);
+       cb->__td = td;
+
+       return ret;
+}
+
+ssize_t aio_read(struct aiocb *cb)
+{
+       cb->aio_lio_opcode = LIO_READ;
+       return new_req(cb);
+}
+
+ssize_t aio_write(struct aiocb *cb)
+{
+       cb->aio_lio_opcode = LIO_WRITE;
+       return new_req(cb);
+}
diff --git a/src/aio/aio_return.c b/src/aio/aio_return.c
new file mode 100644 (file)
index 0000000..df10bdb
--- /dev/null
@@ -0,0 +1,6 @@
+#include <aio.h>
+
+ssize_t aio_return(struct aiocb *cb)
+{
+       return cb->__ret;
+}
diff --git a/src/aio/aio_suspend.c b/src/aio/aio_suspend.c
new file mode 100644 (file)
index 0000000..cb2539e
--- /dev/null
@@ -0,0 +1,57 @@
+#include <aio.h>
+#include <errno.h>
+#include "pthread_impl.h"
+
+/* Due to the requirement that aio_suspend be async-signal-safe, we cannot
+ * use any locks, wait queues, etc. that would make it more efficient. The
+ * only obviously-correct algorithm is to generate a wakeup every time any
+ * aio operation finishes and have aio_suspend re-evaluate the completion
+ * status of each aiocb it was waiting on. */
+
+static volatile int seq;
+
+void __aio_wake(void)
+{
+       a_inc(&seq);
+       __wake(&seq, -1, 1);
+}
+
+int aio_suspend(struct aiocb *const cbs[], int cnt, const struct timespec *ts)
+{
+       int i, last, first=1, ret=0;
+       struct timespec at;
+
+       if (cnt<0) {
+               errno = EINVAL;
+               return -1;
+       }
+
+       for (;;) {
+               last = seq;
+
+               for (i=0; i<cnt; i++) {
+                       if (cbs[i] && cbs[i]->__err != EINPROGRESS)
+                               return 0;
+               }
+
+               if (first && ts) {
+                       clock_gettime(CLOCK_MONOTONIC, &at);
+                       at.tv_sec += ts->tv_sec;
+                       if ((at.tv_nsec += ts->tv_nsec) >= 1000000000) {
+                               at.tv_nsec -= 1000000000;
+                               at.tv_sec++;
+                       }
+                       first = 0;
+               }
+
+               ret = __timedwait(&seq, last, CLOCK_MONOTONIC,
+                       ts ? &at : 0, 0, 0, 1);
+
+               if (ret == ETIMEDOUT) ret = EAGAIN;
+
+               if (ret) {
+                       errno = ret;
+                       return -1;
+               }
+       }
+}
diff --git a/src/aio/lio_listio.c b/src/aio/lio_listio.c
new file mode 100644 (file)
index 0000000..8865029
--- /dev/null
@@ -0,0 +1,140 @@
+#include <aio.h>
+#include <errno.h>
+#include "pthread_impl.h"
+
+struct lio_state {
+       struct sigevent *sev;
+       int cnt;
+       struct aiocb *cbs[];
+};
+
+static int lio_wait(struct lio_state *st)
+{
+       int i, err, got_err;
+       int cnt = st->cnt;
+       struct aiocb **cbs = st->cbs;
+
+       for (;;) {
+               for (i=0; i<cnt; i++) {
+                       if (!cbs[i]) continue;
+                       err = aio_error(cbs[i]);
+                       if (err==EINPROGRESS)
+                               break;
+                       if (err) got_err=1;
+                       cbs[i] = 0;
+               }
+               if (i==cnt) {
+                       if (got_err) {
+                               errno = EIO;
+                               return -1;
+                       }
+                       return 0;
+               }
+               if (aio_suspend(cbs, cnt, 0))
+                       return -1;
+       }
+}
+
+static void notify_signal(struct sigevent *sev)
+{
+       siginfo_t si = {
+               .si_signo = sev->sigev_signo,
+               .si_value = sev->sigev_value,
+               .si_code = SI_ASYNCIO,
+               .si_pid = __pthread_self()->pid,
+               .si_uid = getuid()
+       };
+       __syscall(SYS_rt_sigqueueinfo, si.si_pid, si.si_signo, &si);
+}
+
+static void *wait_thread(void *p)
+{
+       struct lio_state *st = p;
+       struct sigevent *sev = st->sev;
+       lio_wait(st);
+       free(st);
+       switch (sev->sigev_notify) {
+       case SIGEV_SIGNAL:
+               notify_signal(sev);
+               break;
+       case SIGEV_THREAD:
+               sev->sigev_notify_function(sev->sigev_value);
+               break;
+       }
+       return 0;
+}
+
+int lio_listio(int mode, struct aiocb *const cbs[], int cnt, struct sigevent *sev)
+{
+       int i, ret;
+       struct lio_state *st=0;
+
+       if (cnt < 0) {
+               errno = EINVAL;
+               return -1;
+       }
+
+       if (mode == LIO_WAIT || (sev && sev->sigev_notify != SIGEV_NONE)) {
+               if (!(st = malloc(sizeof *st + cnt*sizeof *cbs))) {
+                       errno = EAGAIN;
+                       return -1;
+               }
+               st->cnt = cnt;
+               st->sev = sev;
+               memcpy(st->cbs, cbs, cnt*sizeof *cbs);
+       }
+
+       for (i=0; i<cnt; i++) {
+               if (!cbs[i]) continue;
+               switch (cbs[i]->aio_lio_opcode) {
+               case LIO_READ:
+                       ret = aio_read(cbs[i]);
+                       break;
+               case LIO_WRITE:
+                       ret = aio_write(cbs[i]);
+                       break;
+               default:
+                       continue;
+               }
+               if (ret) {
+                       free(st);
+                       errno = EAGAIN;
+                       return -1;
+               }
+       }
+
+       if (mode == LIO_WAIT) {
+               ret = lio_wait(st);
+               free(st);
+               return 0;
+       }
+
+       if (st) {
+               pthread_attr_t a;
+               sigset_t set;
+               pthread_t td;
+
+               if (sev->sigev_notify == SIGEV_THREAD) {
+                       if (sev->sigev_notify_attributes)
+                               a = *sev->sigev_notify_attributes;
+                       else
+                               pthread_attr_init(&a);
+               } else {
+                       pthread_attr_init(&a);
+                       pthread_attr_setstacksize(&a, PAGE_SIZE);
+                       pthread_attr_setguardsize(&a, 0);
+               }
+               pthread_attr_setdetachstate(&a, PTHREAD_CREATE_DETACHED);
+               sigfillset(&set);
+               pthread_sigmask(SIG_BLOCK, &set, &set);
+               if (pthread_create(&td, &a, wait_thread, st)) {
+                       free(st);
+                       errno = EAGAIN;
+                       return -1;
+               }
+               pthread_sigmask(SIG_SETMASK, &set, 0);
+       }
+
+       return 0;
+}
+