On Wed, Mar 08, 2023 at 04:54:45PM +0000, Pavel Begunkov wrote: > On 3/8/23 16:17, Ming Lei wrote: > > On Wed, Mar 08, 2023 at 02:46:48PM +0000, Pavel Begunkov wrote: > > > On 3/8/23 02:10, Ming Lei wrote: > > > > On Tue, Mar 07, 2023 at 05:17:04PM +0000, Pavel Begunkov wrote: > > > > > On 3/7/23 15:37, Pavel Begunkov wrote: > > > > > > On 3/7/23 14:15, Ming Lei wrote: > > > > > > > Hello, > > > > > > > > > > > > > > Add IORING_OP_FUSED_CMD, it is one special URING_CMD, which has to > > > > > > > be SQE128. The 1st SQE(master) is one 64byte URING_CMD, and the 2nd > > > > > > > 64byte SQE(slave) is another normal 64byte OP. For any OP which needs > > > > > > > to support slave OP, io_issue_defs[op].fused_slave needs to be set as 1, > > > > > > > and its ->issue() can retrieve/import buffer from master request's > > > > > > > fused_cmd_kbuf. The slave OP is actually submitted from kernel, part of > > > > > > > this idea is from Xiaoguang's ublk ebpf patchset, but this patchset > > > > > > > submits slave OP just like normal OP issued from userspace, that said, > > > > > > > SQE order is kept, and batching handling is done too. > > > > > > > > > > > > From a quick look through patches it all looks a bit complicated > > > > > > and intrusive, all over generic hot paths. I think instead we > > > > > > should be able to use registered buffer table as intermediary and > > > > > > reuse splicing. Let me try it out > > > > > > > > > > Here we go, isolated in a new opcode, and in the end should work > > > > > with any file supporting splice. It's a quick prototype, it's lacking > > > > > and there are many obvious fatal bugs. It also needs some optimisations, > > > > > improvements on how executed by io_uring and extra stuff like > > > > > memcpy ops and fixed buf recv/send. I'll clean it up. > > > > > > > > > > I used a test below, it essentially does zc recv. > > > > > > > > > > https://github.com/isilence/liburing/commit/81fe705739af7d9b77266f9aa901c1ada870739d > > > > > > > > [...] > > > > > +int io_splice_from(struct io_kiocb *req, unsigned int issue_flags) > > > > > +{ > > > > > + struct io_splice_from *sp = io_kiocb_to_cmd(req, struct io_splice_from); > > > > > + loff_t *ppos = (sp->off == -1) ? NULL : &sp->off; > > > > > + struct io_mapped_ubuf *imu; > > > > > + struct pipe_inode_info *pi; > > > > > + struct io_ring_ctx *ctx; > > > > > + unsigned int pipe_tail; > > > > > + int ret, i, nr_pages; > > > > > + u16 index; > > > > > + > > > > > + if (!sp->file->f_op->splice_read) > > > > > + return -ENOTSUPP; > > > > > + > > > > > + pi = alloc_pipe_info(); > > > > > > > > The above should be replaced with direct pipe, otherwise every time > > > > allocating one pipe inode really hurts performance. > > > > > > We don't even need to alloc it dynanically, could be just > > > on stack. There is a long list of TODOs I can add, e.g. > > > polling support, retries, nowait, caching imu and so on. > > > > > > [...] > > > > Your patch looks like transferring pages ownership to io_uring fixed > > > > buffer, but unfortunately it can't be done in this way. splice is > > > > supposed for moving data, not transfer buffer ownership. > > > > > > Borrowing rather than transferring. It's not obvious since it's > > > not implemented in the patch, but the buffer should be eventually > > > returned using the splice's ->release callback. > > > > What is the splice's ->release() callback? Is pipe buffer's > > release()? If yes, there is at least the following two problems: > > Right > > > 1) it requires the buffer to be saved(for calling its callback and use its private > > data to return back the whole buffer) in the pipe until it is consumed, which becomes > > one sync interface like splice syscall, and can't cross multiple io_uring OPs or > > per-buffer pipe inode is needed > > We don't mix data from different sources, it's reasonable to expect > that all buffers will have the same callback, then it'll be saved > in struct io_mapped_ubuf. That's sth should definitely be checked and > rejected if happens. It seems reasonable to expect ->release() &->confirm() to be same, but not sure if pipe_buffer->flags & pipe_buffer->private are same, and I think pipe_buffer->private has to be saved too, since ->release() could use it. ->read_splice() is one generic interface, if we support it in the new io_uring OP, all existed ->read_splice() have to be supported. Trust me, I did think about this approach, and the main difference is to add the buffer into io_uring ctx fixed buffer. > > > 2) pipe buffer's get()/release() works on per-buffer/page level, but > > we need to borrow the whole buffer, and the whole buffer could be used > > Surely that can be improved. > > > by arbitrary number of OPs, such as one IO buffer needs to be used for > > handling mirror or stripped targets, so when we know the buffer can be released? > > There is a separate efficient lifetime semantic for io_uring's registered > buffers, which don't involve any get/put. It'll be freed according to it, > i.e. when the userspace asks it to be removed and there are no more > inflight requests. Then one new OP need to be added for removing the buffer explicitly. Then at least three OPs(SQEs)(one for adding the buffer, one or more for consuming it, one for removing it) are involved for handling single IO, which can't be efficient. > > > And basically it can't be known by kernel, and only application knows > > when to release it. > > > > Anyway, please post the whole patch, otherwise it is hard to see > > the whole picture, and devil is always in details, especially Linus > > mentioned splice can't be used in this way. > > Sure > > > > > > > > > > https://lore.kernel.org/linux-block/CAJfpeguQ3xn2-6svkkVXJ88tiVfcDd-eKi1evzzfvu305fMoyw@xxxxxxxxxxxxxx/ > > > > > > > > 1) pages are actually owned by device side(ublk, here: sp->file), but we want to > > > > loan them to io_uring normal OPs. > > > > > > > > 2) after these pages are used by io_uring normal OPs, these pages have > > > > been returned back to sp->file, and the notification has to be done > > > > explicitly, because page is owned by sp->file of splice_read(). > > > > > > Right, see above, they're going to be returned back via ->release. > > > > How? > > I admit, I shouldn't have skipped it even for a quick POC. It'll save > ->release() in struct io_mapped_ubuf and call it when the buffer is > freed from io_uring perspective, that is there are no more requests > using it and the user requested it to be removed. It isn't enough to save ->release() only(->confirm, ->flags, ->private), and all existed ->read_splice() has to be considered. Basically the io_mapped_ubuf becomes one per-buffer (thin)pipe, and the extra allocation could be big. > > > > > 3) pages RW direction has to limited strictly, and in case of ublk/fuse, > > > > device pages can only be read or write which depends on user io request > > > > direction. > > > > > > Yes, I know, and directions will be needed anyway for DMA mappings and > > > different p2p cases in the future, but again a bunch of things is > > > omitted here. > > > > Please don't omitted it and it is one fundamental security problem. > > That's not interesting for a design concept with a huge warning. > io_import_fixed() already takes a dir argument, we just need to check > it against the buffer's one. The dir in io_import_fixed() can't be validated, and only the buffer owner knows if the buffer can be read or write. The main issue is that for one ublk/fuse read io request, the buffer can only be filled with data, and can't be read to somewhere, otherwise kernel data could be leaked to userspace. And the check has to be done in ->read_splcie() or new pipe_buffer->flags has to be added, that is the exact topic we discussed before, and NAKed by Linus. https://lore.kernel.org/linux-block/CAJfpeguQ3xn2-6svkkVXJ88tiVfcDd-eKi1evzzfvu305fMoyw@xxxxxxxxxxxxxx/ > > > > > > Also IMO it isn't good to add one buffer to ctx->user_bufs[] oneshot and > > > > retrieve it oneshot, and it can be set via req->imu simply in one fused > > > > command. > > > > > > That's one of the points though. It's nice if not necessary (for a generic > > > feature) to be able to do multiple ops on the data. For instance, if we > > > have a memcpy request, we can link it to this splice / zc recv, memcpy > > > necessary headers to the userspace and let it decide how to proceed with > > > data. > > > > I feel it could be one big problem for buffer borrowing to cross more than one > > OPs, and when can the buffer be returned back? > > Described above > > > memory copy can be done simply by device's read/write interface, please see > > patch 15. > > I don't think I understand how it looks in the userspace, maybe it's > only applicable to ublk? but it seems that the concept of having one op > producing a buffer and another consuming it don't go well with multi > use in general case, especially stretched in time. So far, the interface is only for ublk/fuse, and the use case for ublk is generic, such as: we have one logic volume manager ublk driver, and now we need to mirror the io request to multiple underlying devices, then the buffer needs to be consumed for each device. Same with stripped, or distribute network storage(one same request need to be sent to more than one remote machine). > > E.g. you recv data, some of which is an application protocol header > that should be looked at by the user and the rest is data that might > be sent out somewhere else. Yeah, see the example[1] in ublk/nbd, application protocol is recv/send by its own SQE, and io data is handled in standalone SQE. Here we only deal with ublk block io data, which isn't changed most of times. [1] https://github.com/ming1/ubdsrv/blob/master/nbd/tgt_nbd.cpp Thanks, Ming
