]> git.sesse.net Git - bcachefs-tools-debian/commitdiff
Fall back to synchronous io when kernel doesn't support aio.
authorBrett Holman <bholman.devel@gmail.com>
Sun, 10 Oct 2021 20:13:40 +0000 (14:13 -0600)
committerBrett Holman <bholman.devel@gmail.com>
Sun, 10 Oct 2021 21:23:57 +0000 (15:23 -0600)
Signed-off-by: Brett Holman <bholman.devel@gmail.com>
linux/blkdev.c

index 898fd53ad4fd6d4e3185120516c1151ba2adf6c1..6c940f79fcaf8bce58f959ad4e80c8eac9adfa9a 100644 (file)
 
 #include "tools-util.h"
 
+struct fops {
+       void (*init)(void);
+       void (*cleanup)(void);
+       void (*read)(struct bio *bio, struct iovec * iov, unsigned i);
+       void (*write)(struct bio *bio, struct iovec * iov, unsigned i);
+};
+
+static struct fops *fops;
 static io_context_t aio_ctx;
 static atomic_t running_requests;
 
@@ -66,35 +74,12 @@ void generic_make_request(struct bio *bio)
 #endif
        }
 
-       struct iocb iocb = {
-               .data           = bio,
-               .aio_fildes     = bio->bi_opf & REQ_FUA
-                       ? bio->bi_bdev->bd_sync_fd
-                       : bio->bi_bdev->bd_fd,
-       }, *iocbp = &iocb;
-
        switch (bio_op(bio)) {
        case REQ_OP_READ:
-               iocb.aio_lio_opcode     = IO_CMD_PREADV;
-               iocb.u.c.buf            = iov;
-               iocb.u.c.nbytes         = i;
-               iocb.u.c.offset         = bio->bi_iter.bi_sector << 9;
-
-               atomic_inc(&running_requests);
-               ret = io_submit(aio_ctx, 1, &iocbp);
-               if (ret != 1)
-                       die("io_submit err: %s", strerror(-ret));
+               fops->read(bio, iov, i);
                break;
        case REQ_OP_WRITE:
-               iocb.aio_lio_opcode     = IO_CMD_PWRITEV;
-               iocb.u.c.buf            = iov;
-               iocb.u.c.nbytes         = i;
-               iocb.u.c.offset         = bio->bi_iter.bi_sector << 9;
-
-               atomic_inc(&running_requests);
-               ret = io_submit(aio_ctx, 1, &iocbp);
-               if (ret != 1)
-                       die("io_submit err: %s", strerror(-ret));
+               fops->write(bio, iov, i);
                break;
        case REQ_OP_FLUSH:
                ret = fsync(bio->bi_bdev->bd_fd);
@@ -236,6 +221,55 @@ int lookup_bdev(const char *path, dev_t *dev)
        return -EINVAL;
 }
 
+static void io_fallback(void)
+{
+       fops++;
+       if (fops->init == NULL)
+               die("no fallback possible, something is very wrong");
+       fops->init();
+}
+
+static void sync_check(struct bio *bio, int ret)
+{
+       if (ret != bio->bi_iter.bi_size) {
+               die("IO error: %s\n", strerror(-ret));
+       }
+
+       if (bio->bi_opf & REQ_FUA) {
+               ret = fdatasync(bio->bi_bdev->bd_fd);
+               if (ret)
+                       die("fsync error: %s\n", strerror(-ret));
+       }
+       bio_endio(bio);
+}
+
+static void sync_init(void) {}
+
+static void sync_cleanup(void)
+{
+       /* not necessary? */
+       sync();
+}
+
+static void sync_read(struct bio *bio, struct iovec * iov, unsigned i)
+{
+
+       int fd = bio->bi_opf & REQ_FUA
+                       ? bio->bi_bdev->bd_sync_fd
+                       : bio->bi_bdev->bd_fd;
+       ssize_t ret = preadv(fd, iov, i, bio->bi_iter.bi_sector << 9);
+       sync_check(bio, ret);
+}
+
+static void sync_write(struct bio *bio, struct iovec * iov, unsigned i)
+{
+       int fd = bio->bi_opf & REQ_FUA
+                       ? bio->bi_bdev->bd_sync_fd
+                       : bio->bi_bdev->bd_fd;
+       ssize_t ret = pwritev(fd, iov, i, bio->bi_iter.bi_sector << 9);
+       sync_check(bio, ret);
+}
+
 static int aio_completion_thread(void *arg)
 {
        struct io_event events[8], *ev;
@@ -274,23 +308,24 @@ static int aio_completion_thread(void *arg)
 
 static struct task_struct *aio_task = NULL;
 
-__attribute__((constructor(102)))
-static void blkdev_init(void)
+static void aio_init(void)
 {
        struct task_struct *p;
        long err = io_setup(256, &aio_ctx);
+       if (!err) {
+               p = kthread_run(aio_completion_thread, NULL, "aio_completion");
+               BUG_ON(IS_ERR(p));
 
-       if (err)
-               die("io_setup() error: %s", strerror(-err));
-
-       p = kthread_run(aio_completion_thread, NULL, "aio_completion");
-       BUG_ON(IS_ERR(p));
+               aio_task = p;
 
-       aio_task = p;
+       } else if (err == -ENOSYS) {
+               io_fallback();
+       } else {
+               die("io_setup() error: %s", strerror(err));
+       }
 }
 
-__attribute__((destructor(102)))
-static void blkdev_cleanup(void)
+static void aio_cleanup(void)
 {
        struct task_struct *p = NULL;
        swap(aio_task, p);
@@ -323,3 +358,71 @@ static void blkdev_cleanup(void)
        close(fds[0]);
        close(fds[1]);
 }
+
+static void aio_op(struct bio *bio, struct iovec *iov, unsigned i, int opcode)
+{
+       ssize_t ret;
+       struct iocb iocb = {
+               .data           = bio,
+               .aio_fildes     = bio->bi_opf & REQ_FUA
+                       ? bio->bi_bdev->bd_sync_fd
+                       : bio->bi_bdev->bd_fd,
+               .aio_lio_opcode = opcode,
+               .u.v.vec        = iov,
+               .u.v.nr         = i,
+               .u.v.offset     = bio->bi_iter.bi_sector << 9,
+
+       }, *iocbp = &iocb;
+
+       atomic_inc(&running_requests);
+       ret = io_submit(aio_ctx, 1, &iocbp);
+       if (ret != 1)
+               die("io_submit err: %s", strerror(-ret));
+}
+
+static void aio_read(struct bio *bio, struct iovec *iov, unsigned i)
+{
+       aio_op(bio, iov, i, IO_CMD_PREADV);
+}
+
+static void aio_write(struct bio *bio, struct iovec * iov, unsigned i)
+{
+       aio_op(bio, iov, i, IO_CMD_PWRITEV);
+}
+
+
+/* not implemented */
+static void uring_init(void) {
+       io_fallback();
+}
+
+struct fops fops_list[] = {
+       {
+               .init           = uring_init,
+       }, {
+               .init           = aio_init,
+               .cleanup        = aio_cleanup,
+               .read           = aio_read,
+               .write          = aio_write,
+       }, {
+               .init           = sync_init,
+               .cleanup        = sync_cleanup,
+               .read           = sync_read,
+               .write          = sync_write,
+       }, {
+               /* NULL */
+       }
+};
+
+__attribute__((constructor(102)))
+static void blkdev_init(void)
+{
+       fops = fops_list;
+       fops->init();
+}
+
+__attribute__((destructor(102)))
+static void blkdev_cleanup(void)
+{
+       fops->cleanup();
+}