Re: [PATCH] fuse: Send FORGET over io_uring when ring is ready

[Bernd Schubert <bernd@bsbernd.com>]

On 4/1/26 15:41, Horst Birthelmer wrote:
> On Wed, Apr 01, 2026 at 11:52:28AM +0000, Bernd Schubert wrote:
>>
>>
>> On 4/1/26 10:40, Li Wang wrote:
>>> Once the FUSE io_uring is registered and marked ready, most request
>>> types are delivered through io_uring, while FORGET notifications were still
>>> queued with fuse_dev_queue_forget() and only consumed through the legacy
>>> path on /dev/fuse.
>>>
>>> Deliver single FORGET operations through fuse_uring_queue_fuse_req() when
>>> the ring is ready. Otherwise, fall back to fuse_dev_queue_forget()
>>> so behavior matches the previous implementation.
>>>
>>> Benefits:
>>> - While io-uring is active, the daemon can handle forgets in the same
>>>   commit/fetch loop as other opcodes instead of also draining a separate
>>>   /dev/fuse read path for forget traffic.
>>> - Reduces split-brain transport for high-volume forgets (eviction,
>>>   unmount) when the ring is already the primary channel, which simplifies
>>>   userspace and keeps teardown forgets on the same completion path as
>>>   other uring-backed work.
>>> - Reuses the same per-queue io-uring machinery and noreply/force request
>>>   setup (creds, FR_WAITING/FR_FORCE, etc.) already used for similar
>>>   kernel-initiated traffic.
>>>
>>> Signed-off-by: Li Wang <liwang@kylinos.cn>
>>> ---
>>>  fs/fuse/dev.c        | 84 ++++++++++++++++++++++++++++++++++++++++++++
>>>  fs/fuse/dev_uring.c  |  2 +-
>>>  fs/fuse/fuse_dev_i.h |  4 +++
>>>  3 files changed, 89 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/fs/fuse/dev.c b/fs/fuse/dev.c
>>> index b212565a78cf..f58abc80fd7b 100644
>>> --- a/fs/fuse/dev.c
>>> +++ b/fs/fuse/dev.c
>>> @@ -665,6 +665,90 @@ static void fuse_args_to_req(struct fuse_req *req, struct fuse_args *args)
>>>  		__set_bit(FR_ASYNC, &req->flags);
>>>  }
>>>  
>>> +#ifdef CONFIG_FUSE_IO_URING
>>> +struct fuse_forget_uring_data {
>>> +	struct fuse_args args;
>>> +	struct fuse_forget_in inarg;
>>> +};
>>> +
>>> +static void fuse_forget_uring_free(struct fuse_mount *fm, struct fuse_args *args,
>>> +				   int error)
>>> +{
>>> +	struct fuse_forget_uring_data *d =
>>> +		container_of(args, struct fuse_forget_uring_data, args);
>>> +
>>> +	kfree(d);
>>> +}
>>> +
>>> +/*
>>> + * Send FUSE_FORGET through the io-uring ring when active; same payload as
>>> + * fuse_read_single_forget(), with userspace committing like any other request.
>>> + */
>>> +void fuse_io_uring_send_forget(struct fuse_iqueue *fiq,
>>> +				struct fuse_forget_link *forget)
>>> +{
>>> +	struct fuse_conn *fc = container_of(fiq, struct fuse_conn, iq);
>>> +	struct fuse_mount *fm;
>>> +	struct fuse_req *req;
>>> +	struct fuse_forget_uring_data *d;
>>> +
>>> +	if (!fuse_uring_ready(fc)) {
>>> +		fuse_dev_queue_forget(fiq, forget);
>>> +		return;
>>> +	}
>>> +
>>> +	down_read(&fc->killsb);
>>> +	if (list_empty(&fc->mounts)) {
>>> +		up_read(&fc->killsb);
>>> +		fuse_dev_queue_forget(fiq, forget);
>>> +		return;
>>> +	}
>>> +	fm = list_first_entry(&fc->mounts, struct fuse_mount, fc_entry);
>>> +	up_read(&fc->killsb);
>>> +
>>> +	d = kmalloc(sizeof(*d), GFP_KERNEL);
>>> +	if (!d)
>>> +		goto fallback;
>>> +
>>> +	atomic_inc(&fc->num_waiting);
>>> +	req = fuse_request_alloc(fm, GFP_KERNEL);
>>> +	if (!req) {
>>> +		kfree(d);
>>> +		fuse_drop_waiting(fc);
>>> +		goto fallback;
>>> +	}
>>> +
>>> +	memset(&d->args, 0, sizeof(d->args));
>>> +	d->inarg.nlookup = forget->forget_one.nlookup;
>>> +	d->args.opcode = FUSE_FORGET;
>>> +	d->args.nodeid = forget->forget_one.nodeid;
>>> +	d->args.in_numargs = 1;
>>> +	d->args.in_args[0].size = sizeof(d->inarg);
>>> +	d->args.in_args[0].value = &d->inarg;
>>> +	d->args.force = true;
>>> +	d->args.noreply = true;
>>> +	d->args.end = fuse_forget_uring_free;
>>> +
>>> +	kfree(forget);
>>> +
>>> +	fuse_force_creds(req);
>>> +	__set_bit(FR_WAITING, &req->flags);
>>> +	if (!d->args.abort_on_kill)
>>> +		__set_bit(FR_FORCE, &req->flags);
>>> +	fuse_adjust_compat(fc, &d->args);
>>> +	fuse_args_to_req(req, &d->args);
>>> +	req->in.h.len = sizeof(struct fuse_in_header) +
>>> +		fuse_len_args(req->args->in_numargs,
>>> +			      (struct fuse_arg *)req->args->in_args);
>>> +
>>> +	fuse_uring_queue_fuse_req(fiq, req);
>>> +	return;
>>> +
>>> +fallback:
>>> +	fuse_dev_queue_forget(fiq, forget);
>>> +}
>>> +#endif
>>> +
>>>  ssize_t __fuse_simple_request(struct mnt_idmap *idmap,
>>>  			      struct fuse_mount *fm,
>>>  			      struct fuse_args *args)
>>> diff --git a/fs/fuse/dev_uring.c b/fs/fuse/dev_uring.c
>>> index 7b9822e8837b..a96539ea400a 100644
>>> --- a/fs/fuse/dev_uring.c
>>> +++ b/fs/fuse/dev_uring.c
>>> @@ -1360,7 +1360,7 @@ bool fuse_uring_remove_pending_req(struct fuse_req *req)
>>>  
>>>  static const struct fuse_iqueue_ops fuse_io_uring_ops = {
>>>  	/* should be send over io-uring as enhancement */
>>> -	.send_forget = fuse_dev_queue_forget,
>>> +	.send_forget = fuse_io_uring_send_forget,
>>
>> I will check the other parts more thoroughly in the evening, but please
>> take a look into fuse_uring_register(), it also also overrides other
>> pointers at startup - I would like leave it here as it is, move the
>> function above into dev_uring.c and then update this part in dev_uring.c
>>
>> static const struct fuse_iqueue_ops fuse_io_uring_ops = {
>>         /* should be send over io-uring as enhancement */
>>         .send_forget = fuse_dev_queue_forget,
> 
> Hi Bernd,
> 
> I have never asked the question before, but now I'm a bit intrigued ...
> Why wasn't this not done before? Was it a performance thing?

Hi Horst,

never had a priority for me - I didn't consider it performance relevant.


Cheers,
Bernd