> Il giorno 22 nov 2019, alle ore 10:50, Arnd Bergmann <arnd@xxxxxxxx> ha scritto: > > (adding Paolo as well, maybe he has some more insights) > > On Mon, Nov 18, 2019 at 11:04 AM (Exiting) Baolin Wang > <baolin.wang@xxxxxxxxxx> wrote: >> On Tue, 12 Nov 2019 at 16:48, Baolin Wang <baolin.wang7@xxxxxxxxx> wrote: >>> On Tue, Nov 12, 2019 at 12:59 AM Arnd Bergmann <arnd@xxxxxxxx> wrote: >>>> On Mon, Nov 11, 2019 at 1:58 PM Baolin Wang <baolin.wang@xxxxxxxxxx> wrote: >>>>>> - With that change in place calling a blocking __mmc_claim_host() is >>>>>> still a problem, so there should still be a nonblocking mmc_try_claim_host() >>>>>> for the submission path, leading to a BLK_STS_DEV_RESOURCE (?) >>>>>> return code from mmc_mq_queue_rq(). Basically mmc_mq_queue_rq() >>>>>> should always return right away, either after having queued the next I/O >>>>>> or with an error, but not waiting for the device in any way. >>>>> >>>>> Actually not only the mmc_claim_host() will block the MMC request >>>>> processing, in this routine, the mmc_blk_part_switch() and >>>>> mmc_retune() can also block the request processing. Moreover the part >>>>> switching and tuning should be sync operations, and we can not move >>>>> them to a work or a thread. >>>> >>>> Ok, I see. >>>> >>>> Those would also cause requests to be sent to the device or the host >>>> controller, right? Maybe we can treat them as "a non-IO request >>> >>> Right. >>> >>>> has successfully been queued to the device" events, returning >>>> busy from the mmc_mq_queue_rq() function and then running >>>> the queue again when they complete? >>> >>> Yes, seems reasonable to me. >>> >>>> >>>>>> - For the packed requests, there is apparently a very simple way to implement >>>>>> that without a software queue: mmc_mq_queue_rq() is allowed to look at >>>>>> and dequeue all requests that are currently part of the request_queue, >>>>>> so it should take out as many as it wants to submit at once and send >>>>>> them all down to the driver together, avoiding the need for any further >>>>>> round-trips to blk_mq or maintaining a queue in mmc. >>>>> >>>>> You mean we can dispatch a request directly from >>>>> elevator->type->ops.dispatch_request()? but we still need some helper >>>>> functions to check if these requests can be packed (the package >>>>> condition), and need to invent new APIs to start a packed request (or >>>>> using cqe interfaces, which means we still need to implement some cqe >>>>> callbacks). >>>> >>>> I don't know how the dispatch_request() function fits in there, >>>> what Hannes told me is that in ->queue_rq() you can always >>>> look at the following requests that are already queued up >>>> and take the next ones off the list. Looking at bd->last >>>> tells you if there are additional requests. If there are, you can >>>> look at the next one from blk_mq_hw_ctx (not sure how, but >>>> should not be hard to find) >>>> >>>> I also see that there is a commit_rqs() callback that may >>>> go along with queue_rq(), implementing that one could make >>>> this easier as well. >>> >>> Yes, we can use queue_rq()/commit_rqs() and bd->last (now bd->last may >>> can not work well, see [1]), but like we talked before, for packed >>> request, we still need some new interfaces (for example, a interface >>> used to start a packed request, and a interface used to complete a >>> packed request), but at last we got a consensus that we should re-use >>> the CQE interfaces instead of new invention. >>> >>> [1] https://lore.kernel.org/patchwork/patch/1102897/ >>> >>>> >>>>>> - The DMA management (bounce buffer, map, unmap) that is currently >>>>>> done in mmc_blk_mq_issue_rq() should ideally be done in the >>>>>> init_request()/exit_request() (?) callbacks from mmc_mq_ops so this >>>>>> can be done asynchronously, out of the critical timing path for the >>>>>> submission. With this, there won't be any need for a software queue. >>>>> >>>>> This is not true, now the blk-mq will allocate some static request >>>>> objects (usually the static requests number should be the same with >>>>> the hardware queue depth) saved in struct blk_mq_tags. So the >>>>> init_request() is used to initialize the static requests when >>>>> allocating them, and call exit_request to free the static requests >>>>> when freeing the 'struct blk_mq_tags', such as the queue is dead. So >>>>> we can not move the DMA management into the init_request/exit_request. >>>> >>>> Ok, I must have misremembered which callback that is then, but I guess >>>> there is some other place to do it. >>> >>> I checked the 'struct blk_mq_ops', and I did not find a ops can be >>> used to do DMA management. And I also checked UFS driver, it also did >>> the DMA mapping in the queue_rq() (scsi_queue_rq() ---> >>> ufshcd_queuecommand() ---> ufshcd_map_sg()). Maybe I missed something? >>> >>> Moreover like I said above, for the packed request, we still need >>> implement something (like the software queue) based on the CQE >>> interfaces to help to handle packed requests. >> >> After some investigation and offline discussion with you, I still have >> some concerns about your suggestion. >> >> 1) Now blk-mq have not supplied some ops to prepare a request, which is >> used to do some DMA management asynchronously. But yes, we can >> introduce new ops for blk-mq. But there are still some remaining >> preparation in mmc_mq_queue_rq(), like mmc part switch. For software >> queue, we can prepare a request totally after issuing one. > > I suppose to make the submission non-blocking, all operations that > currently block in the submission path may have to be changed first. > > For the case of a partition switch (same for retune), I suppose > something like this can be done: > > - in queue_rq() check whether a partition switch is needed. If not, > submit the current rq > - if a partition switch is needed, submit the partition switch cmd > instead, and return busy status > - when the completion arrives for the partition switch, call back into > blk_mq to have it call queue_rq again. > > Or possibly even (this might not be possible without signifcant > restructuring): > > - when preparing a request that would require a partition switch, > insert another meta-request to switch the partition ahead of it. > > I do realize that this is a significant departure from how it was done > in the past, but it seems cleaner that way to me. > >> 2) I wonder if it is appropriate that using the irq threaded context >> to dispatch next request, actually we will still introduce a context >> switch here. Now we will complete a request in the hard irq handler >> and kick the softirq to do time-consuming operations, like DMA >> unmapping , and will start next request in the hard irq handler >> without context switch. Moreover if we remove the BLK_MQ_F_BLOCKING in >> future like you suggested, then we can remove all context switch. And >> I think we can dispatch next request in the softirq context (actually >> the CQE already did). > > I hope Hannes (or someone else) can comment here, as I don't > know exactly what his objection to kicking off the next cmd in the > hardirq was. > > I think generally, deferring all slow operations to an irqthread > rather than a softirq is a good idea, but I share your concern that > this can introduce an unnecessary latency between the the > the IRQ is signaled and the time the following cmd is sent to the > hardware. > > Doing everything in a single (irqthread) context is clearly simpler, > so this would need to be measured carefully to avoid unnecessary > complexity, but I think don't see anything stopping us from having > the fast-path where the low-level driver first checks for any possible > error conditions in hardirq context and the fires off a prepared cmd > right away whenever it can before triggering the irqthread that does > everything else. I think this has to be a per-driver optimization, so > the common case would just have an irqthread. > >> 3) For packed request support, I did not see an example that block >> driver can dispatch a request from the IO scheduler in queue_rq() and >> no APIs supported from blk-mq. And we do not know where can dispatch a >> request in queue_rq(), from IO scheduler? from ctx? or from >> hctx->dispatch list? and if this request can not be passed to host >> now, how to do it? Seems lots of complicated things. > > The only way I can see is the ->last flag, so if blk_mq submits multiple > requests in a row to queue_rq() with this flag cleared and calls > ->commit_rqs() after the last one. This seems to be what the scsi > disk driver and the nvme driver rely on, and we should be able to use > it the same way for packed cmds, by checking each time in queue_rq() > whether requests can/should be combined and reporting busy otherwise > (after preparing a combined mmc cmd). > blk_mq will then call commit_rqs, which should do the actual submission > to the hardware driver. > > Now as you point out, the *big* problem with this is that we never > get multiple requests together in practice, i.e. the last flag is almost > always set, and any optimization around it has no effect. > > This is where I'm a bit lost in the code as well, but it seems that > this is part of the current bfq design that only sends one request down > the driver stack at a time, and this would have to change first before > we can rely on this for packing requests. > > Paolo, can you comment on why this is currently done, or if it can > be changed? It seems to me that sending multiple requests at > once would also have a significant benefit on the per-request overhead > on NVMe devices with with bfq. > Hi, actually, "one request dispatched at a time" is not a peculiarity of bfq. Any scheduler can provide only one request at a time, with the current blk-mq API for I/O schedulers. Yet, when it is time to refill an hardware queue, blk-mq pulls as many requests as it deems appropriate from the scheduler, by invoking the latter multiple times. See blk_mq_do_dispatch_sched() in block/blk-mq-sched.c. I don't know where the glitch for MMC is with respect to this scheme. Thanks, Paolo >> Moreover, we still need some interfaces for the packed request >> handling, from previous discussion, we still need something like MMC >> software queue based on the CQE to help to handle the packed request. >> >> So I think I still need to introduce the MMC software queue, on the one >> hand is that it can really improve the performance from fio data and >> avoid a long latency, on the other hand we can expand it to support >> packed request easily in future. Thanks. >> >> (Anyway I will still post the V7 to address Adrian's comments and to >> see if we can get a consensus there). > > > Arnd