On 10 May 2017 at 10:24, Linus Walleij <linus.walleij@xxxxxxxxxx> wrote: > This is a series that starts to untangle the MMC "big host lock", > i.e. what is taken by issueing mmc_claim_host() which usually happens > through mmc_get_card(). > > The host lock is standing in the way of a bunch of modernizations, > because the block layer interface takes this lock when a new request > arrives, then will not release it until the queue is empty, which > is detected by getting NULL from blk_fetch_request(). > > This does not work with the new MQ block layer because it issues > requests asynchrously and there is no way of telling if the submit > queue is empty or not. Block requests must always be served. Trying > to introduce an interface for figuring out if the queue is empty > is probably generally a bad idea as it will require cross-talk > between all CPUs and then we hit Amdahl's Law again when scaling > upwards with CPUs. > > So Arnd Bergmann suggested that whatever parallel requests the > host lock is protecting, maybe these requests can be funneled > through the block layer itself? > > It turns out that most block drivers already do this, by using the > "special" block requests REQ_OP_DRV_IN and REQ_OP_DRV_OUT. And > it is what we should have done from the beginning. > > To do this, the per-request extra data (which I think is also > referred to as "tag") need to be handled the same way that all > other block drivers do it: allocate it through the block layer. > In the MMC stack, this extra per-request data (tag) is called > struct mmc_queue_req. > > So the second patch convert to using the generic block layer > mechanism to allocate tags. This makes the core simpler IMO and > that is why the patch series has negative line count. We move > stuff over to the block layer. > > The first patch cleans out bounce buffer configurability and > move that to be a property of the host rather than a Kconfig option > in order to make it cleaner to do the rest of the refactorings. > > The last two patches moves the ioctl() calls over to use the > new per-request tags and removes two instances of mmc_get_card(), > so we start to untangle the big host lock. > > This approach can be used to get rid of the debugfs mmc_get_card() > as well but I want to start with the ioctl()s. > > It already fixes a problem: userspace and the block layer could > essentially starve each other by bombing requests from either > block access or ioctl() congesting the host lock. With this, > ioctl() operations get scheduled by the block layer with > everything else. Not that I know how the block layer prioritizes > REQ_OP_DRV_IN and REQ_OP_DRV_OUT requests, but I am confident > it does a better job than the first-come-first-served host lock. > > This drives a truck through mine and Adrians patch sets for > multiqueue and command queueing respectively, but I think that > both of our patch series will get easier to implement if we > exploit these patches as a base for future work, so if they can > get positive reviews and does not make everything explode, I > would suggest we merge them as a starter for the v4.13 kernel > cycle. > > Linus Walleij (5): > mmc: core: Delete bounce buffer Kconfig option > mmc: core: Allocate per-request data using the block layer core > mmc: block: Tag is_rpmb as bool > mmc: block: move single ioctl() commands to block requests > mmc: block: move multi-ioctl() to use block layer > > drivers/mmc/core/Kconfig | 18 ---- > drivers/mmc/core/block.c | 126 +++++++++++++++---------- > drivers/mmc/core/queue.c | 235 ++++++++++++---------------------------------- > drivers/mmc/core/queue.h | 26 +++-- > drivers/mmc/host/cavium.c | 3 + > drivers/mmc/host/pxamci.c | 6 ++ > include/linux/mmc/card.h | 2 - > include/linux/mmc/host.h | 1 + > 8 files changed, 161 insertions(+), 256 deletions(-) > > -- > 2.9.3 > Really nice work you have done here! I have currently looked over the series and it seems promising. Allow me to get of couple of more days for review though. Also, hopefully we can get someone more to help to test it, but I am rather eager to apply this once rc1 is out. Kind regards Uffe
