On Fri, Dec 14, 2012 at 07:20:41PM +0400, Maxim V. Patlasov wrote:
> The patch implements a framework to process an IO request asynchronously. The
> idea is to associate several fuse requests with a single kiocb by means of
> fuse_io_priv structure. The structure plays the same role for FUSE as 'struct
> dio' for direct-io.c.
>
> The framework is supposed to be used like this:
> - someone (who wants to process an IO asynchronously) allocates fuse_io_priv
> and initializes it setting 'async' field to non-zero value.
> - as soon as fuse request is filled, it can be submitted (in non-blocking way)
> by fuse_async_req_send()
> - when all submitted requests are ACKed by userspace, io->reqs drops to zero
> triggering aio_complete()
>
> In case of IO initiated by libaio, aio_complete() will finish processing the
> same way as in case of dio_complete() calling aio_complete(). But the
> framework may be also used for internal FUSE use when initial IO request
> was synchronous (from user perspective), but it's beneficial to process it
> asynchronously. Then the caller should wait on kiocb explicitly and
> aio_complete() will wake the caller up.
>
> Signed-off-by: Maxim Patlasov <[email protected]>
> ---
> fs/fuse/file.c | 92 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
> fs/fuse/fuse_i.h | 17 ++++++++++
> 2 files changed, 109 insertions(+), 0 deletions(-)
>
> diff --git a/fs/fuse/file.c b/fs/fuse/file.c
> index 6685cb0..8dd931f 100644
> --- a/fs/fuse/file.c
> +++ b/fs/fuse/file.c
> @@ -503,6 +503,98 @@ static void fuse_release_user_pages(struct fuse_req *req, int write)
> }
> }
>
> +/**
> + * In case of short read, the caller sets 'pos' to the position of
> + * actual end of fuse request in IO request. Otherwise, if bytes_requested
> + * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
> + *
> + * An example:
> + * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
> + * both submitted asynchronously. The first of them was ACKed by userspace as
> + * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
> + * second request was ACKed as short, e.g. only 1K was read, resulting in
> + * pos == 33K.
> + *
> + * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
> + * will be equal to the length of the longest contiguous fragment of
> + * transferred data starting from the beginning of IO request.
> + */
> +static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
> +{
> + int left;
> +
> + spin_lock(&io->lock);
> + if (err)
> + io->err = io->err ? : err;
> + else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
> + io->bytes = pos;
> +
> + left = --io->reqs;
> + spin_unlock(&io->lock);
> +
> + if (!left) {
> + long res;
> +
> + if (io->err)
> + res = io->err;
> + else if (io->bytes >= 0 && io->write)
> + res = -EIO;
> + else {
> + res = io->bytes < 0 ? io->size : io->bytes;
> +
> + if (!is_sync_kiocb(io->iocb)) {
> + struct path *path = &io->iocb->ki_filp->f_path;
> + struct inode *inode = path->dentry->d_inode;
> + struct fuse_conn *fc = get_fuse_conn(inode);
> + struct fuse_inode *fi = get_fuse_inode(inode);
> +
> + spin_lock(&fc->lock);
> + fi->attr_version = ++fc->attr_version;
> + spin_unlock(&fc->lock);
Hmm, what is this? Incrementing the attr version without setting any attributes
doesn't make sense.
Thanks,
Miklos
> + }
> + }
> +
> + aio_complete(io->iocb, res, 0);
> + kfree(io);
> + }
> +}
> +
> +static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
> +{
> + struct fuse_io_priv *io = req->io;
> + ssize_t pos = -1;
> +
> + fuse_release_user_pages(req, !io->write);
> +
> + if (io->write) {
> + if (req->misc.write.in.size != req->misc.write.out.size)
> + pos = req->misc.write.in.offset - io->offset +
> + req->misc.write.out.size;
> + } else {
> + if (req->misc.read.in.size != req->out.args[0].size)
> + pos = req->misc.read.in.offset - io->offset +
> + req->out.args[0].size;
> + }
> +
> + fuse_aio_complete(io, req->out.h.error, pos);
> +}
> +
> +static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
> + size_t num_bytes, struct fuse_io_priv *io)
> +{
> + spin_lock(&io->lock);
> + io->size += num_bytes;
> + io->reqs++;
> + spin_unlock(&io->lock);
> +
> + req->io = io;
> + req->end = fuse_aio_complete_req;
> +
> + fuse_request_send_background(fc, req);
> +
> + return num_bytes;
> +}
> +
> static size_t fuse_send_read(struct fuse_req *req, struct file *file,
> loff_t pos, size_t count, fl_owner_t owner)
> {
> diff --git a/fs/fuse/fuse_i.h b/fs/fuse/fuse_i.h
> index e4f70ea..e0a5b65 100644
> --- a/fs/fuse/fuse_i.h
> +++ b/fs/fuse/fuse_i.h
> @@ -219,6 +219,20 @@ enum fuse_req_state {
> FUSE_REQ_FINISHED
> };
>
> +/** The request IO state (for asynchronous processing) */
> +struct fuse_io_priv {
> + int async;
> + spinlock_t lock;
> + unsigned reqs;
> + ssize_t bytes;
> + size_t size;
> + __u64 offset;
> + bool write;
> + int err;
> + struct kiocb *iocb;
> + struct file *file;
> +};
> +
> /**
> * A request to the client
> */
> @@ -323,6 +337,9 @@ struct fuse_req {
> /** Inode used in the request or NULL */
> struct inode *inode;
>
> + /** AIO control block */
> + struct fuse_io_priv *io;
> +
> /** Link on fi->writepages */
> struct list_head writepages_entry;
>
>
Hi Miklos,
04/22/2013 08:34 PM, Miklos Szeredi пишет:
> On Fri, Dec 14, 2012 at 07:20:41PM +0400, Maxim V. Patlasov wrote:
>> The patch implements a framework to process an IO request asynchronously. The
>> idea is to associate several fuse requests with a single kiocb by means of
>> fuse_io_priv structure. The structure plays the same role for FUSE as 'struct
>> dio' for direct-io.c.
>>
>> The framework is supposed to be used like this:
>> - someone (who wants to process an IO asynchronously) allocates fuse_io_priv
>> and initializes it setting 'async' field to non-zero value.
>> - as soon as fuse request is filled, it can be submitted (in non-blocking way)
>> by fuse_async_req_send()
>> - when all submitted requests are ACKed by userspace, io->reqs drops to zero
>> triggering aio_complete()
>>
>> In case of IO initiated by libaio, aio_complete() will finish processing the
>> same way as in case of dio_complete() calling aio_complete(). But the
>> framework may be also used for internal FUSE use when initial IO request
>> was synchronous (from user perspective), but it's beneficial to process it
>> asynchronously. Then the caller should wait on kiocb explicitly and
>> aio_complete() will wake the caller up.
>>
>> Signed-off-by: Maxim Patlasov <[email protected]>
>> ---
>> fs/fuse/file.c | 92 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
>> fs/fuse/fuse_i.h | 17 ++++++++++
>> 2 files changed, 109 insertions(+), 0 deletions(-)
>>
>> diff --git a/fs/fuse/file.c b/fs/fuse/file.c
>> index 6685cb0..8dd931f 100644
>> --- a/fs/fuse/file.c
>> +++ b/fs/fuse/file.c
>> @@ -503,6 +503,98 @@ static void fuse_release_user_pages(struct fuse_req *req, int write)
>> }
>> }
>>
>> +/**
>> + * In case of short read, the caller sets 'pos' to the position of
>> + * actual end of fuse request in IO request. Otherwise, if bytes_requested
>> + * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
>> + *
>> + * An example:
>> + * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
>> + * both submitted asynchronously. The first of them was ACKed by userspace as
>> + * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
>> + * second request was ACKed as short, e.g. only 1K was read, resulting in
>> + * pos == 33K.
>> + *
>> + * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
>> + * will be equal to the length of the longest contiguous fragment of
>> + * transferred data starting from the beginning of IO request.
>> + */
>> +static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
>> +{
>> + int left;
>> +
>> + spin_lock(&io->lock);
>> + if (err)
>> + io->err = io->err ? : err;
>> + else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
>> + io->bytes = pos;
>> +
>> + left = --io->reqs;
>> + spin_unlock(&io->lock);
>> +
>> + if (!left) {
>> + long res;
>> +
>> + if (io->err)
>> + res = io->err;
>> + else if (io->bytes >= 0 && io->write)
>> + res = -EIO;
>> + else {
>> + res = io->bytes < 0 ? io->size : io->bytes;
>> +
>> + if (!is_sync_kiocb(io->iocb)) {
>> + struct path *path = &io->iocb->ki_filp->f_path;
>> + struct inode *inode = path->dentry->d_inode;
>> + struct fuse_conn *fc = get_fuse_conn(inode);
>> + struct fuse_inode *fi = get_fuse_inode(inode);
>> +
>> + spin_lock(&fc->lock);
>> + fi->attr_version = ++fc->attr_version;
>> + spin_unlock(&fc->lock);
> Hmm, what is this? Incrementing the attr version without setting any attributes
> doesn't make sense.
It makes sense at least for writes. __fuse_direct_write() always called
fuse_write_update_size() and the latter always incremented attr_version,
even if *ppos <= inode->i_size. I believed it was implemented in this
way intentionally: if write succeeded, the file is changed on server,
hence attrs requested from server early should be regarded as stale.
Adding async IO support to fuse, a case emerges when
fuse_write_update_size() won't be called: incoming direct IO write is
asynchronous (e.g. it came from libaio), it's not extending write, so
it's allowable to process it by submitting fuse requests to background
and return -EIOCBQUEUED without waiting for completions (see 4th patch
of this patch-set). But in this case the file on server will be changed
anyway. That's why I bump attr_version in fuse_aio_complete() -- to be
consistent with the model we had before this patch-set.
The fact that I did the trick both for writes and reads was probably
overlook. I'd suggest to fix it like this:
> - if (!is_sync_kiocb(io->iocb)) {
> + if (!is_sync_kiocb(io->iocb) && io->write) {
Thanks,
Maxim