On Tue, 30 Apr 2019 at 14:06, Ludovic BARRE <ludovic.barre@xxxxxx> wrote: > > > > On 4/30/19 1:13 PM, Ulf Hansson wrote: > > On Fri, 26 Apr 2019 at 09:46, Ludovic Barre <ludovic.Barre@xxxxxx> wrote: > >> > >> From: Ludovic Barre <ludovic.barre@xxxxxx> > >> > >> This patch series adds busy detect for stm32 sdmmc variant. > >> Some adaptations are required: > >> -Avoid to check and poll busy status when is not expected. > >> -Clear busy status bit if busy_detect_flag and busy_detect_mask are > >> different. > >> -Add hardware busy timeout with MMCIDATATIMER register. > >> > >> V2: > >> -mmci_cmd_irq cleanup in separate patch. > >> -simplify the busy_detect_flag exclude > >> -replace sdmmc specific comment in > >> "mmc: mmci: avoid fake busy polling in mmci_irq" > >> to focus on common behavior > >> > >> Ludovic Barre (5): > >> mmc: mmci: cleanup mmci_cmd_irq for busy detect feature > >> mmc: mmci: avoid fake busy polling in mmci_irq > >> mmc: mmci: fix clear of busy detect status > >> mmc: mmci: add hardware busy timeout feature > >> mmc: mmci: add busy detect for stm32 sdmmc variant > >> > >> drivers/mmc/host/mmci.c | 61 ++++++++++++++++++++++++++++++++++++++----------- > >> drivers/mmc/host/mmci.h | 3 +++ > >> 2 files changed, 51 insertions(+), 13 deletions(-) > >> > >> -- > >> 2.7.4 > >> > > > > Ludovic, just wanted to let you know that I am reviewing and testing > > this series. > > > > However, while running some tests on Ux500 for validating the busy > > detection code, even without your series applied, I encounter some odd > > behaviors. I am looking into the problem to understand better and will > > let you know as soon as I have some more data to share. > > Oops, don't hesitate to share your status, if I could help. Thanks! Good and bad news here, then. I now understand what is going on - and there is certainly room for improvements here, but more importantly the actual mmci busy detection works as expected. When it comes to improvements, the main issue I have found is how we treat DATA WRITES. In many cases we simply don't use the HW busy detection at all, but instead rely on the mmc core to send CMD13 in a loop to poll. Well, then if the polling would have consisted of a couple of CMD13s that wouldn't be an issue, but my observations is rather that the numbers of CMD13 sent to poll is in the range or hundreds/thousands - per each WRITE request! I am going to send a patch (or two) that improves the behavior. It might even involve changing parts in core layer, not sure how the end result will look like yet. In any case, I have applied patch 1 and patch2 for next, as the tests turned out well at my side. I also took the liberty of updating some of the comments/changelogs, please have look and tell if there is something you want to change. I will continue with the rest of series next week. Kind regards Uffe
