On 10/10/24 9:07 PM, Ming Lei wrote: > On Thu, Oct 10, 2024 at 08:39:12PM -0600, Jens Axboe wrote: >> On 10/10/24 8:30 PM, Ming Lei wrote: >>> Hi Jens, >>> >>> On Thu, Oct 10, 2024 at 01:31:21PM -0600, Jens Axboe wrote: >>>> Hi, >>>> >>>> Discussed this with Pavel, and on his suggestion, I tried prototyping a >>>> "buffer update" opcode. Basically it works like >>>> IORING_REGISTER_BUFFERS_UPDATE in that it can update an existing buffer >>>> registration. But it works as an sqe rather than being a sync opcode. >>>> >>>> The idea here is that you could do that upfront, or as part of a chain, >>>> and have it be generically available, just like any other buffer that >>>> was registered upfront. You do need an empty table registered first, >>>> which can just be sparse. And since you can pick the slot it goes into, >>>> you can rely on that slot afterwards (either as a link, or just the >>>> following sqe). >>>> >>>> Quick'n dirty obviously, but I did write a quick test case too to verify >>>> that: >>>> >>>> 1) It actually works (it seems to) >>> >>> It doesn't work for ublk zc since ublk needs to provide one kernel buffer >>> for fs rw & net send/recv to consume, and the kernel buffer is invisible >>> to userspace. But __io_register_rsrc_update() only can register userspace >>> buffer. >> >> I'd be surprised if this simple one was enough! In terms of user vs >> kernel buffer, you could certainly use the same mechanism, and just >> ensure that buffers are tagged appropriately. I need to think about that >> a little bit. > > It is actually same with IORING_OP_PROVIDE_BUFFERS, so the following > consumer OPs have to wait until this OP_BUF_UPDATE is completed. See below for the registered vs provided buffer confusion that seems to be a confusion issue here. > Suppose we have N consumers OPs which depends on OP_BUF_UPDATE. > > 1) all N OPs are linked with OP_BUF_UPDATE > > Or > > 2) submit OP_BUF_UPDATE first, and wait its completion, then submit N > OPs concurrently. Correct > But 1) and 2) may slow the IO handing. In 1) all N OPs are serialized, > and 1 extra syscall is introduced in 2). Yes you don't want do do #1. But the OP_BUF_UPDATE is cheap enough that you can just do it upfront. It's not ideal in terms of usage, and I get where the grouping comes from. But is it possible to do the grouping in a less intrusive fashion with OP_BUF_UPDATE? Because it won't change any of the other ops in terms of buffer consumption, they'd just need fixed buffer support and you'd flag the buffer index in sqe->buf_index. And the nice thing about that is that while fixed/registered buffers aren't really used on the networking side yet (as they don't bring any benefit yet), adding support for them could potentially be useful down the line anyway. > The same thing exists in the next OP_BUF_UPDATE which has to wait until > all the previous buffer consumers are done. So the same slow thing are > doubled. Not mention the application will become more complicated. It does not, you can do an update on a buffer that's already inflight. > Here the provided buffer is only visible among the N OPs wide, and making > it global isn't necessary, and slow things down. And has kbuf lifetime > issue. I was worried about it being too slow too, but the basic testing seems like it's fine. Yes with updates inflight it'll make it a tad bit slower, but really should not be a concern. I'd argue that even doing the very basic of things, which would be: 1) Submit OP_BUF_UPDATE, get completion 2) Do the rest of the ops would be totally fine in terms of performance. OP_BUF_UPDATE will _always_ completely immediately and inline, which means that it'll _always_ be immediately available post submission. The only think you'd ever have to worry about in terms of failure is a badly formed request, which is a programming issue, or running out of memory on the host. > Also it makes error handling more complicated, io_uring has to remove > the kernel buffer when the current task is exit, dependency or order with > buffer provider is introduced. Why would that be? They belong to the ring, so should be torn down as part of the ring anyway? Why would they be task-private, but not ring-private? >> There are certainly many different ways that can get propagated which >> would not entail a complicated mechanism. I really like the aspect of >> having the identifier being the same thing that we already use, and >> hence not needing to be something new on the side. >> >>> Also multiple OPs may consume the buffer concurrently, which can't be >>> supported by buffer select. >> >> Why not? You can certainly have multiple ops using the same registered >> buffer concurrently right now. > > Please see the above problem. > > Also I remember that the selected buffer is removed from buffer list, > see io_provided_buffer_select(), but maybe I am wrong. You're mixing up provided and registered buffers. Provided buffers are ones that the applications gives to the kernel, and the kernel grabs and consumes them. Then the application replenishes, repeat. Registered buffers are entirely different, those are registered with the kernel and we can do things like pre-gup the pages so we don't have to do them for every IO. They are entirely persistent, any multiple ops can keep using them, concurrently. They don't get consumed by an IO like provided buffers, they remain in place until they get unregistered (or updated, like my patch) at some point. -- Jens Axboe
