#asahi-dev · 2022-02-16 17:19
axboedid some checking here, and the only fast flushes I see are the ones where there has been no writes since the last flush#asahi-dev · 2022-02-16 17:19
axboesven: curious on how much you looked at the slow FLUSH#asahi-dev · 2022-02-16 16:20
axboesven: seems that grabbing the cq lock really is required, still see timeouts/hangs with just serializing the issue itself#asahi-dev · 2022-02-16 15:45
axboewas poking at the slower flush#asahi-dev · 2022-02-16 15:45
axboesven: still triggers hang with rising edge, just a bit harder to trigger#asahi-dev · 2022-02-16 14:48
axboe17-18ms#asahi-dev · 2022-02-16 14:47
axboelooking at blktrace, some FWFW writes do indeed take forever#asahi-dev · 2022-02-16 14:45
axboeGlanzmann: lol#asahi-dev · 2022-02-16 14:44
axboethat is miserable#asahi-dev · 2022-02-16 14:44
axboeGlanzmann: 56 iops indeed :)#asahi-dev · 2022-02-16 13:39
axboehaha I totally know that feeling#asahi-dev · 2022-02-16 13:36
axboesven: huh that does seem to work, at least the usual "start all the things" combined with the full find and a randread iops test didn't cause any issues#asahi-dev · 2022-02-16 13:31
axboelet's give it a shot...#asahi-dev · 2022-02-16 13:26
axboesven: or do you want both#asahi-dev · 2022-02-16 13:23
axboesven: s/IRQ_TYPE_LEVEL_HIGH/IRQ_TYPE_EDGE_RISING for nvme@393...?#asahi-dev · 2022-02-16 13:23
axboesven: I can test the dtsi change#asahi-dev · 2022-02-16 13:22
axboeas long as the writel is ordered with the memcpy, it doesn't matter if one issue ends up writing the tag before another#asahi-dev · 2022-02-16 13:22
axboebut that should be fine, we don't need ordering here#asahi-dev · 2022-02-16 13:13
axboej`ey: gotcha#asahi-dev · 2022-02-16 13:13
axboej`ey: I don't think this is an irq vs issue problem#asahi-dev · 2022-02-16 13:11
axboerunning with that now, seems fine too (as expected)#asahi-dev · 2022-02-16 13:11
axboehttps://pastebin.com/p89hyGUT#asahi-dev · 2022-02-16 13:11
axboesven: was thinking last night, what prevents two writel() from running at the same time? I just cannot convince myself that the current situation is safe#asahi-dev · 2022-02-15 21:48
axboeyeah I know, plan was to refactor around it but then I forgot the last bit#asahi-dev · 2022-02-15 21:45
axboewill fix#asahi-dev · 2022-02-15 21:44
axboethen forgot#asahi-dev · 2022-02-15 21:44
axboej`ey: heh, it's funny I even thought of that while doing it#asahi-dev · 2022-02-15 21:37
axboe(and what a mess of patches...)#asahi-dev · 2022-02-15 21:36
axboetop 3 commits there#asahi-dev · 2022-02-15 21:36
axboesven: https://git.kernel.dk/cgit/linux-block/log/?h=m1-test#asahi-dev · 2022-02-15 21:32
axboeyou can just fold in or however you prefer, I don't need attribution since it's not in tree yet#asahi-dev · 2022-02-15 21:29
axboeI'll toss you that tcb patch too in a bit#asahi-dev · 2022-02-15 21:29
axboeyeah I don't think you need it right now, but for suspend you probably will. so I left it alone#asahi-dev · 2022-02-15 21:05
axboesven: assuming you're keeping the ->enabled state for adding suspend, because I don't think it's needed right now#asahi-dev · 2022-02-15 21:02
axboeno new ideas on the other front, yet#asahi-dev · 2022-02-15 21:01
axboeunrelated#asahi-dev · 2022-02-15 21:01
axboecleanup of irq handling#asahi-dev · 2022-02-15 21:01
axboesven: https://pastebin.com/djdMXek9#asahi-dev · 2022-02-15 20:35
axboebut only explanation I can find is that we trigger an interrupt for completion and don't find it, then it gets polled later#asahi-dev · 2022-02-15 20:34
axboeyep#asahi-dev · 2022-02-15 20:34
axboenot sure how the nvme hw side works if first tag isn't written yet#asahi-dev · 2022-02-15 20:34
axboeyou have two issues going on, one is half way through memcpy of command, other happens faster and submits the tag#asahi-dev · 2022-02-15 20:33
axboesven: one idea#asahi-dev · 2022-02-15 20:33
axboeand now booted#asahi-dev · 2022-02-15 20:32
axboeI've got a patch for that too :)#asahi-dev · 2022-02-15 20:32
axboeseems we can kill both of those, and just set flags to zero#asahi-dev · 2022-02-15 20:32
axboesince we memset at alloc, do we really need to clear in both inval and submit?#asahi-dev · 2022-02-15 20:31
axboeanother thing I just noticed, there seems to be excessive tcb clearing in there#asahi-dev · 2022-02-15 20:28
axboeiirc you grab the lock around the whole irq handling#asahi-dev · 2022-02-15 20:28
axboethe spin unlock is a barrier, and as irq grabs the lock too first, I think it'd be safe