Re: [PATCH RFC v2 00/19] fuse: fuse-over-io-uring

[Bernd Schubert <bernd.schubert@fastmail.fm>]

On 6/12/24 17:55, Kent Overstreet wrote:
> On Wed, Jun 12, 2024 at 03:40:14PM GMT, Bernd Schubert wrote:
>> On 6/12/24 16:19, Kent Overstreet wrote:
>>> On Wed, Jun 12, 2024 at 03:53:42PM GMT, Bernd Schubert wrote:
>>>> I will definitely look at it this week. Although I don't like the idea
>>>> to have a new kthread. We already have an application thread and have
>>>> the fuse server thread, why do we need another one?
>>>
>>> Ok, I hadn't found the fuse server thread - that should be fine.
>>>
>>>>>
>>>>> The next thing I was going to look at is how you guys are using splice,
>>>>> we want to get away from that too.
>>>>
>>>> Well, Ming Lei is working on that for ublk_drv and I guess that new approach
>>>> could be adapted as well onto the current way of io-uring.
>>>> It _probably_ wouldn't work with IORING_OP_READV/IORING_OP_WRITEV.
>>>>
>>>> https://lore.gnuweeb.org/io-uring/20240511001214.173711-6-ming.lei@redhat.com/T/
>>>>
>>>>>
>>>>> Brian was also saying the fuse virtio_fs code may be worth
>>>>> investigating, maybe that could be adapted?
>>>>
>>>> I need to check, but really, the majority of the new additions
>>>> is just to set up things, shutdown and to have sanity checks.
>>>> Request sending/completing to/from the ring is not that much new lines.
>>>
>>> What I'm wondering is how read/write requests are handled. Are the data
>>> payloads going in the same ringbuffer as the commands? That could work,
>>> if the ringbuffer is appropriately sized, but alignment is a an issue.
>>
>> That is exactly the big discussion Miklos and I have. Basically in my
>> series another buffer is vmalloced, mmaped and then assigned to ring entries.
>> Fuse meta headers and application payload goes into that buffer.
>> In both kernel/userspace directions. io-uring only allows 80B, so only a
>> really small request would fit into it.
> 
> Well, the generic ringbuffer would lift that restriction.

Yeah, kind of. Instead allocating the buffer in fuse, it would be now allocated
in that code. At least all that setup code would be moved out of fuse. I will
eventually come to your patches today.
Now we only need to convince Miklos that your ring is better ;)

> 
>> Legacy /dev/fuse has an alignment issue as payload follows directly as the fuse
>> header - intrinsically fixed in the ring patches.
> 
> *nod*
> 
> That's the big question, put the data inline (with potential alignment
> hassles) or manage (and map) a separate data structure.
> 
> Maybe padding could be inserted to solve alignment?

Right now I have this struct:

struct fuse_ring_req {
	union {
		/* The first 4K are command data */
		char ring_header[FUSE_RING_HEADER_BUF_SIZE];

		struct {
			uint64_t flags;

			/* enum fuse_ring_buf_cmd */
			uint32_t in_out_arg_len;
			uint32_t padding;

			/* kernel fills in, reads out */
			union {
				struct fuse_in_header in;
				struct fuse_out_header out;
			};
		};
	};

	char in_out_arg[];
};


Data go into in_out_arg, i.e. headers are padded by the union.
I actually wonder if FUSE_RING_HEADER_BUF_SIZE should be page size
and not a fixed 4K.

(I just see the stale comment 'enum fuse_ring_buf_cmd',
will remove it in the next series)


> 
> A separate data structure would only really be useful if it enabled zero
> copy, but that should probably be a secondary enhancement.
> 
>> I will now try without mmap and just provide a user buffer as pointer in the 80B
>> section.
>>
>>
>>>
>>> We just looked up the device DMA requirements and with modern NVME only
>>> 4 byte alignment is required, but the block layer likely isn't set up to
>>> handle that.
>>
>> I think existing fuse headers have and their data have a 4 byte alignment.
>> Maybe even 8 byte, I don't remember without looking through all request types.
>> If you try a simple O_DIRECT read/write to libfuse/example_passthrough_hp
>> without the ring patches it will fail because of alignment. Needs to be fixed
>> in legacy fuse and would also avoid compat issues we had in libfuse when the
>> kernel header was updated.
>>
>>>
>>> So - prearranged buffer? Or are you using splice to get pages that
>>> userspace has read into into the kernel pagecache?
>>
>> I didn't even try to use splice yet, because for the DDN (my employer) use case
>> we cannot use  zero copy, at least not without violating the rule that one
>> cannot access the application buffer in userspace.
> 
> DDN - lustre related?

I have bit of ancient Lustre background, also with DDN, then went to Fraunhofer
for FhGFS/BeeGFS (kind of competing with Lustre).
Back at DDN initially on IME (burst buffer) and now Infinia. Lustre is mostly HPC
only, Infina is kind of everything.