On 6/23/22 7:31 AM, Hao Xu wrote: > On 6/20/22 21:41, Jens Axboe wrote: >> On 6/20/22 6:01 AM, Hao Xu wrote: >>> Hi, >>> I've some thought on the way of doing async IO. The current model is: >>> (given we are using SQPOLL mode) >>> >>> the sqthread does: >>> (a) Issue a request with nowait/nonblock flag. >>> (b) If it would block, reutrn -EAGAIN >>> (c) The io_uring layer captures this -EAGAIN and wake up/create >>> a io-worker to execute the request synchronously. >>> (d) Try to issue other requests in the above steps again. >>> >>> This implementation has two downsides: >>> (1) we have to find all the block point in the IO stack manually and >>> change them into "nowait/nonblock friendly". >>> (2) when we raise another io-worker to do the request, we submit the >>> request from the very beginning. This isn't a little bit inefficient. >>> >>> >>> While I think we can actually do it in a reverse way: >>> (given we are using SQPOLL mode) >>> >>> the sqthread1 does: >>> (a) Issue a request in the synchronous way >>> (b) If it is blocked/scheduled soon, raise another sqthread2 >>> (c) sqthread2 tries to issue other requests in the same way. >>> >>> This solves problem (1), and may solve (2). >>> For (1), we just do the sqthread waken-up at the beginning of schedule() >>> just like what the io-worker and system-worker do. No need to find all >>> the block point. >>> For (2), we continue the blocked request from where it is blocked when >>> resource is satisfied. >>> >>> What we need to take care is making sure there is only one task >>> submitting the requests. >>> >>> To achieve this, we can maintain a pool of sqthread just like the iowq. >>> >>> I've done a very simple/ugly POC to demonstrate this: >>> >>> https://github.com/HowHsu/linux/commit/183be142493b5a816b58bd95ae4f0926227b587b >>> >>> I also wrote a simple test to test it, which submits two sqes, one >>> read(pipe), one nop request. The first one will be block since no data >>> in the pipe. Then a new sqthread was created/waken up to submit the >>> second one and then some data is written to the pipe(by a unrelated >>> user thread), soon the first sqthread is waken up and continues the >>> request. >>> >>> If the idea sounds no fatal issue I'll change the POC to real patches. >>> Any comments are welcome! >> >> One thing I've always wanted to try out is kind of similar to this, but >> a superset of it. Basically io-wq isn't an explicit offload mechanism, >> it just happens automatically if the issue blocks. This applies to both >> SQPOLL and non-SQPOLL. >> >> This takes a page out of the old syslet/threadlet that Ingo Molnar did >> way back in the day [1], but it never really went anywhere. But the >> pass-on-block primitive would apply very nice to io_uring. > > I've read a part of the syslet/threadlet patchset, seems it has > something that I need, my first idea about the new iowq offload is > just like syslet----if blocked, trigger a new worker, deliver the > context to it, and then update the current context so that we return > to the place of sqe submission. But I just didn't know how to do it. Exactly, what you mentioned was very close to what I had considered in the past, and what the syslet/threadlet attempted to do. Except it flips it upside down a bit, which I do think is probably the saner way to do it rather than have the original block and fork a new one. > By the way, may I ask why the syslet/threadlet is not merged to the > mainline. The mail thread is very long, haven't gotten a chance to > read all of it. Not quite sure, it's been a long time. IMHO it's a good idea looking for the right interface, which we now have. So the time may be ripe to do something like this, finally. > > For the approach I posted, I found it is actually SQPOLL-nonrelated. > The original conext just wake up a worker in the pool to do the > submission, and if one blocks, another one wakes up to do the > submission. It is definitely easier to implement than something like > syslet(context delivery) since the new worker naturally goes to the > place of submission thus no context delivery needed. but a downside is > every time we call io_uring_enter to submit a batch of sqes, there is a > wakeup at the beginning. > > I'll try if I can implement a context delivery version. Sounds good, thanks. -- Jens Axboe
