On 6/03/19 12:00 PM, Faiz Abbas wrote: > Adrian, > > On 25/02/19 1:47 PM, Adrian Hunter wrote: >> On 15/02/19 9:20 PM, Faiz Abbas wrote: >>> sdhci.c has two bottom halves implemented. A threaded_irq for handling >>> card insert/remove operations and a tasklet for finishing mmc requests. >>> With the addition of external dma support, dmaengine APIs need to >>> terminate in non-atomic context before unmapping the dma buffers. >>> >>> To facilitate this, remove the finish_tasklet and move the call of >>> sdhci_request_done() to the threaded_irq() callback. >> >> The irq thread has a higher latency than the tasklet. The performance drop >> is measurable on the system I tried: >> >> Before: >> >> # dd if=/dev/mmcblk1 of=/dev/null bs=1G count=1 & >> 1+0 records in >> 1+0 records out >> 1073741824 bytes (1.1 GB) copied, 4.44502 s, 242 MB/s >> >> After: >> >> # dd if=/dev/mmcblk1 of=/dev/null bs=1G count=1 & >> 1+0 records in >> 1+0 records out >> 1073741824 bytes (1.1 GB) copied, 4.50898 s, 238 MB/s >> >> So we only want to resort to the thread for the error case. >> > > Sorry for the late response here, but this is about 1.6% decrease. I > tried out the same commands on a dra7xx board here (with about 5 > consecutive dd of 1GB) and the average decrease was 0.3%. I believe you > will also find a lesser percentage change if you average over multiple > dd commands. > > Is this really so significant that we have to maintain two different > bottom halves and keep having difficulty with adding APIs that can sleep?
It is a performance drop that can be avoided, so it might as well be. Splitting the success path from the failure path is common for I/O drivers for similar reasons as here: the success path can be optimized whereas the failure path potentially needs to sleep. > > Also I am not sure how to implement only the error handling part in the > threaded_irq. We need to enter sdhci_request_done() and get the current > mrq before we can check for error conditions like I've done in patch 2: > > /* Terminate and synchronize dma in case of an error */ > if (data && (mrq->cmd->error || data->error) && > host->use_external_dma) { > struct dma_chan *chan = sdhci_external_dma_channel(host, data); > dmaengine_terminate_sync(chan); > } > > On a related note, do we really need to protect everything in > sdhci_request_done() with spinlocks? > In patch 2 I have only removed lock > for the terminate_sync() parts that I added but the whole > dma_unmap/dma_sync parts should be left unprotected IMO. As it is written, synchronization is needed to stop the same mrq being finished twice. I suggest doing the dmaengine_terminate_sync() before calling sdhci_request_done(). Perhaps you could record the channel that needs to be sync'd and then do: struct dma_chan *chan = READ_ONCE(host->sync_chan); if (chan) { dmaengine_terminate_sync(chan); host->sync_chan = NULL; }