On 5/31/24 21:10, Jens Axboe wrote:
> On 5/31/24 11:36 AM, Bernd Schubert wrote:
>> On 5/31/24 18:24, Jens Axboe wrote:
>>> On 5/29/24 12:00 PM, Bernd Schubert wrote:
>>>> This is to avoid using async completion tasks
>>>> (i.e. context switches) when not needed.
>>>>
>>>> Cc: io-uring@xxxxxxxxxxxxxxx
>>>> Signed-off-by: Bernd Schubert <bschubert@xxxxxxx>
>>>
>>> This patch is very confusing, even after having pulled the other
>>> changes. In general, would be great if the io_uring list was CC'ed on
>>
>> Hmm, let me try to explain. And yes, I definitely need to add these details
>> to the commit message
>>
>> Without the patch:
>>
>> <sending a struct fuse_req>
>>
>> fuse_uring_queue_fuse_req
>> fuse_uring_send_to_ring
>> io_uring_cmd_complete_in_task
>>
>> <async task runs>
>> io_uring_cmd_done()
>
> And this is a worthwhile optimization, you always want to complete it
> line if at all possible. But none of this logic or code belongs in fuse,
> it really should be provided by io_uring helpers.
>
> I would just drop this patch for now and focus on the core
> functionality. Send out a version with that, and then we'll be happy to
> help this as performant as it can be. This is where the ask on "how to
> reproduce your numbers" comes from - with that, it's usually trivial to
> spot areas where things could be improved. And I strongly suspect that
> will involve providing you with the right API to use here, and perhaps
> refactoring a bit on the fuse side. Making up issue_flags is _really_
> not something a user should do.
Great that you agree, I don't like the issue_flag handling in fuse code either.
I will also follow your suggestion to drop this patch.
>
>> 1) (current == queue->server_task)
>> fuse_uring_cmd (IORING_OP_URING_CMD) received a completion for a
>> previous fuse_req, after completion it fetched the next fuse_req and
>> wants to send it - for 'current == queue->server_task' issue flags
>> got stored in struct fuse_ring_queue::uring_cmd_issue_flags
>
> And queue->server_task is the owner of the ring? Then yes that is safe
Yeah, it is the thread that submits SQEs - should be the owner of the ring,
unless daemon side does something wrong (given that there are several
userspace implementation and not a single libfuse only, we need to expect
and handle implementation errors, though).
>>
>> 2) 'else if (current->io_uring)'
>>
>> (actually documented in the code)
>>
>> 2.1 This might be through IORING_OP_URING_CMD as well, but then server
>> side uses multiple threads to access the same ring - not nice. We only
>> store issue_flags into the queue for 'current == queue->server_task', so
>> we do not know issue_flags - sending through task is needed.
>
> What's the path leading to you not having the issue_flags?
We get issue flags here, but I want to keep changes to libfuse small and want
to avoid changing non uring related function signatures. Which is the the
why we store issue_flags for the presumed ring owner thread in the queue data
structure, but we don't have it for possible other threads then
Example:
IORING_OP_URING_CMD
fuse_uring_cmd
fuse_uring_commit_and_release
fuse_uring_req_end_and_get_next --> until here issue_flags passed
fuse_request_end -> generic fuse function, issue_flags not passed
req->args->end() / fuse_writepage_end
fuse_simple_background
fuse_request_queue_background
fuse_request_queue_background_uring
fuse_uring_queue_fuse_req
fuse_uring_send_to_ring
io_uring_cmd_done
I.e. we had issue_flags up to fuse_uring_req_end_and_get_next(), but then
call into generic fuse functions and stop passing through issue_flags.
For the ring-owner we take issue flags stored by fuse_uring_cmd()
into struct fuse_ring_queue, but if daemon side uses multiple threads to
access the ring we won't have that. Well, we could allow it and store
it into an array or rb-tree, but I don't like that multiple threads access
something that is optimized to have a thread per core already.
>
>> 2.2 This might be an application request through the mount point, through
>> the io-uring interface. We do know issue flags either.
>> (That one was actually a surprise for me, when xfstests caught it.
>> Initially I had a condition to send without the extra task then lockdep
>> caught that.
>
> In general, if you don't know the context (eg you don't have issue_flags
> passed in), you should probably assume the only way is to sanely proceed
> is to have it processed by the task itself.
>
>>
>> In both cases it has to use a tasks.
>>
>>
>> My question here is if 'current->io_uring' is reliable.
>
> Yes that will be reliable in the sense that it tells you that the
> current task has (at least) one io_uring context setup. But it doesn't
> tell you anything beyond that, like if it's the owner of this request.
Yeah, you can see that it just checks for current->io_uring and then
uses a task.
>
>> 3) everything else
>>
>> 3.1) For async requests, interesting are cached reads and writes here. At a minimum
>> writes a holding a spin lock and that lock conflicts with the mutex io-uring is taking -
>> we need a task as well
>>
>> 3.2) sync - no lock being hold, it can send without the extra task.
>
> As mentioned, let's drop this patch 19 for now. Send out what you have
> with instructions on how to test it, and I'll give it a spin and see
> what we can do about this.
>
>>> Outside of that, would be super useful to include a blurb on how you set
>>> things up for testing, and how you run the testing. That would really
>>> help in terms of being able to run and test it, and also to propose
>>> changes that might make a big difference.
>>>
>>
>> Will do in the next version.
>> You basically need my libfuse uring branch
>> (right now commit history is not cleaned up) and follow
>> instructions in <libfuse>/xfstests/README.md how to run xfstests.
>> Missing is a slight patch for that dir to set extra daemon parameters,
>> like direct-io (fuse' FOPEN_DIRECT_IO) and io-uring. Will add that libfuse
>> during the next days.
>
> I'll leave the xfstests to you for now, but running some perf testing
> just to verify how it's being used would be useful and help improve it
> for sure.
>
Ah you meant performance tests. I used libfuse/example/passthrough_hp from
my uring branch and then fio on top of that for reads/writes and mdtest from
the ior repo for metadata. Maybe I should upload my scripts somewhere.
Thanks,
Beernd
