On Sat, Aug 16, 2025 at 07:13:53PM -0400, Brian Song wrote:
> 
> 
> On 8/14/25 11:46 PM, Brian Song wrote:
> > From: Brian Song <hibrians...@gmail.com>
> > 
> > This patch adds a new export option for storage-export-daemon to enable
> > or disable FUSE-over-io_uring via the switch io-uring=on|off (disable
> > by default). It also implements the protocol handshake with the Linux
> > kernel during the FUSE-over-io_uring initialization phase.
> > 
> > See: https://docs.kernel.org/filesystems/fuse-io-uring.html
> > 
> > The kernel documentation describes in detail how FUSE-over-io_uring
> > works. This patch implements the Initial SQE stage shown in thediagram:
> > it initializes one queue per IOThread, each currently supporting a
> > single submission queue entry (SQE). When the FUSE driver sends the
> > first FUSE request (FUSE_INIT), storage-export-daemon calls
> > fuse_uring_start() to complete initialization, ultimately submitting
> > the SQE with the FUSE_IO_URING_CMD_REGISTER command to confirm
> > successful initialization with the kernel.
> > 
> > Suggested-by: Kevin Wolf <kw...@redhat.com>
> > Suggested-by: Stefan Hajnoczi <stefa...@redhat.com>
> > Signed-off-by: Brian Song <hibrians...@gmail.com>
> > ---
> >   block/export/fuse.c                  | 161 ++++++++++++++++++++++++---
> >   docs/tools/qemu-storage-daemon.rst   |  11 +-
> >   qapi/block-export.json               |   5 +-
> >   storage-daemon/qemu-storage-daemon.c |   1 +
> >   util/fdmon-io_uring.c                |   5 +-
> >   5 files changed, 159 insertions(+), 24 deletions(-)
> > 
> > diff --git a/block/export/fuse.c b/block/export/fuse.c
> > index c0ad4696ce..59fa79f486 100644
> > --- a/block/export/fuse.c
> > +++ b/block/export/fuse.c
> > @@ -48,6 +48,11 @@
> >   #include <linux/fs.h>
> >   #endif
> > 
> > +#define FUSE_DEFAULT_MAX_PAGES_PER_REQ 32
> > +
> > +/* room needed in buffer to accommodate header */
> > +#define FUSE_BUFFER_HEADER_SIZE 0x1000
> > +
> >   /* Prevent overly long bounce buffer allocations */
> >   #define FUSE_MAX_READ_BYTES (MIN(BDRV_REQUEST_MAX_BYTES, 1 * 1024 * 1024))
> >   /*
> > @@ -63,12 +68,31 @@
> >       (FUSE_MAX_WRITE_BYTES - FUSE_IN_PLACE_WRITE_BYTES)
> > 
> >   typedef struct FuseExport FuseExport;
> > +typedef struct FuseQueue FuseQueue;
> > +
> > +typedef struct FuseRingEnt {
> > +    /* back pointer */
> > +    FuseQueue *q;
> > +
> > +    /* commit id of a fuse request */
> > +    uint64_t req_commit_id;
> > +
> > +    /* fuse request header and payload */
> > +    struct fuse_uring_req_header req_header;
> > +    void *op_payload;
> > +    size_t req_payload_sz;
> > +
> > +    /* The vector passed to the kernel */
> > +    struct iovec iov[2];
> > +
> > +    CqeHandler fuse_cqe_handler;
> > +} FuseRingEnt;
> > 
> >   /*
> >    * One FUSE "queue", representing one FUSE FD from which requests are 
> > fetched
> >    * and processed.  Each queue is tied to an AioContext.
> >    */
> > -typedef struct FuseQueue {
> > +struct FuseQueue {
> >       FuseExport *exp;
> > 
> >       AioContext *ctx;
> > @@ -109,7 +133,12 @@ typedef struct FuseQueue {
> >        * Free this buffer with qemu_vfree().
> >        */
> >       void *spillover_buf;
> > -} FuseQueue;
> > +
> > +#ifdef CONFIG_LINUX_IO_URING
> > +    int qid;
> > +    FuseRingEnt ent;
> > +#endif
> > +};
> > 
> >   /*
> >    * Verify that FuseQueue.request_buf plus the spill-over buffer together
> > @@ -148,6 +177,7 @@ struct FuseExport {
> >       bool growable;
> >       /* Whether allow_other was used as a mount option or not */
> >       bool allow_other;
> > +    bool is_uring;
> > 
> >       mode_t st_mode;
> >       uid_t st_uid;
> > @@ -257,6 +287,93 @@ static const BlockDevOps fuse_export_blk_dev_ops = {
> >       .drained_poll  = fuse_export_drained_poll,
> >   };
> > 
> > +#ifdef CONFIG_LINUX_IO_URING
> > +
> > +static void fuse_uring_sqe_set_req_data(struct fuse_uring_cmd_req *req,
> > +                    const unsigned int qid,
> > +                    const unsigned int commit_id)
> > +{
> > +    req->qid = qid;
> > +    req->commit_id = commit_id;
> > +    req->flags = 0;
> > +}
> > +
> > +static void fuse_uring_sqe_prepare(struct io_uring_sqe *sqe, FuseQueue *q,
> > +               __u32 cmd_op)
> > +{
> > +    sqe->opcode = IORING_OP_URING_CMD;
> > +
> > +    sqe->fd = q->fuse_fd;
> > +    sqe->rw_flags = 0;
> > +    sqe->ioprio = 0;
> > +    sqe->off = 0;
> > +
> > +    sqe->cmd_op = cmd_op;
> > +    sqe->__pad1 = 0;
> > +}
> > +
> > +static void fuse_uring_prep_sqe_register(struct io_uring_sqe *sqe, void 
> > *opaque)
> > +{
> > +    FuseQueue *q = opaque;
> > +    struct fuse_uring_cmd_req *req = (void *)&sqe->cmd[0];
> > +
> > +    fuse_uring_sqe_prepare(sqe, q, FUSE_IO_URING_CMD_REGISTER);
> > +
> > +    sqe->addr = (uint64_t)(q->ent.iov);
> > +    sqe->len = 2;
> > +
> > +    fuse_uring_sqe_set_req_data(req, q->qid, 0);
> > +}
> > +
> > +static void fuse_uring_submit_register(void *opaque)
> > +{
> > +    FuseQueue *q = opaque;
> > +    FuseExport *exp = q->exp;
> > +
> > +
> > +    aio_add_sqe(fuse_uring_prep_sqe_register, q, 
> > &(q->ent.fuse_cqe_handler));
> 
> I think there might be a tricky issue with the io_uring integration in QEMU.
> Currently, when the number of IOThreads goes above ~6 or 7, there’s a pretty
> high chance of a hang. I added some debug logging in the kernel’s
> fuse_uring_cmd() registration part, and noticed that the number of register
> calls is less than the total number of entries in the queue. In theory, we
> should be registering each entry for each queue.
> 
> On the userspace side, everything seems normal, the number of aio_add_sqe()
> calls matches the number of IOThreads. But here’s the weird part: if I add a
> printf inside the while loop in fdmon-io_uring.c::fdmon_io_uring_wait(),
> suddenly everything works fine, and the kernel receives registration
> requests for all entries as expected.
> 
>     do {
>         ret = io_uring_submit_and_wait(&ctx->fdmon_io_uring, wait_nr);
>         fprintf(stderr, "io_uring_submit_and_wait ret: %d\n", ret);
>     } while (ret == -EINTR);
> 
> My guess is that printf is just slowing down the loop, or maybe there’s some
> implicit memory barrier happening. Obviously, the right fix isn’t to
> sprinkle fprintfs around. I suspect there might be a subtle
> synchronization/race issue here.

Strange, your fprintf(3) is after io_uring_submit_and_wait(3). I'm not
sure how that would influence timing because there should be num_cpus
IOThreads independently submitting 1 REGISTER uring_cmd.

Debugging ideas:

- When QEMU hangs, cat /proc/<pid>/fdinfo/<fd> for each IOThread's
  io_uring file descriptor. That shows you what the kernel sees,
  including the state of the SQ/CQ rings. If userspace has filled in the
  SQE then the output should reflect that.

- Replace the REGISTER uring_cmd SQE with a IORING_OP_NOP SQE. This way
  you eliminate FUSE and can focus purely on testing io_uring. If the
  CQE is still missing then there is probably a bug in QEMU's
  aio_add_sqe() API.

Stefan

> 
> Brian
> 

Attachment: signature.asc
Description: PGP signature

Reply via email to