#asahi-dev · 2022-02-25 15:12
axboeusing a pro right now, got an m1 max on the way...#asahi-dev · 2022-02-25 15:12
axboemaz: indeed, just checked#asahi-dev · 2022-02-25 14:53
axboej`ey: thanks!#asahi-dev · 2022-02-25 14:52
axboeshould benefit scheduling intensive workloads too, like compiling#asahi-dev · 2022-02-25 14:51
axboecurious on this fast-ipi, is there a patch posted somewhere?#asahi-dev · 2022-02-22 13:13
axboemarcan: yeah we can wire that up, I can take a look#asahi-dev · 2022-02-18 14:52
axboemps: promising#asahi-dev · 2022-02-18 14:48
axboeapart from that, with those settings, I like how the touchpad behaves#asahi-dev · 2022-02-18 14:47
axboeit seems the "disable touchpad on typing" is probably what bit me, doesn't work so well#asahi-dev · 2022-02-18 14:47
axboe Glanzmann: I'm already using that, copied from you :)#asahi-dev · 2022-02-18 13:19
axboesorry, jumpy touchpad...#asahi-dev · 2022-02-18 13:19
axboet log#asahi-dev · 2022-02-17 20:54
axboeyeah#asahi-dev · 2022-02-17 20:47
axboethat part works a bit better on "normal" nvme, as we just have to do that one doorbell write with the sq tail#asahi-dev · 2022-02-17 20:47
axboeammortize the cost of grabbing the submit side lock#asahi-dev · 2022-02-17 20:47
axboeso it allows you to submit N ios at the time#asahi-dev · 2022-02-17 20:46
axboeidea there is based on how linux manages IO in an on-stack plug for submitting more than 1#asahi-dev · 2022-02-17 20:46
axboeand queue_rqs is actually in 5.16, so could be done#asahi-dev · 2022-02-17 20:46
axboethis is the perf-wip patches for persistent maps cherry-picked, then the apple-nvme bits needed on top#asahi-dev · 2022-02-17 20:41
axboesven: pushed the nvme test stuff there, just in case you are curious#asahi-dev · 2022-02-17 20:40
axboesven: https://git.kernel.dk/cgit/linux-block/log/?h=m1-test-nvme#asahi-dev · 2022-02-17 20:26
axboekettenis: yeah looks like the bsds took the "make it a system tunable" instead, which does make a lot more sense than a separate command#asahi-dev · 2022-02-17 20:14
axboelooks like freebsd/netbsd at least do it saner and provide a sysctl setting for it#asahi-dev · 2022-02-17 20:14
axboeI guess I can google#asahi-dev · 2022-02-17 20:13
axboeanyone know if F_FULLFSYNC is osx only, or do some of the bsd's use it too?#asahi-dev · 2022-02-17 20:13
axboeonly about 2x slower than a rotating hard drive, nice#asahi-dev · 2022-02-17 20:12
axboejannau: great thanks, seems it's doing its job then#asahi-dev · 2022-02-17 19:56
axboedefinitely not typing this on the m1#asahi-dev · 2022-02-17 19:55
axboesomeone needs to run the compiler on osx :)#asahi-dev · 2022-02-17 19:55
axboethanks for spotting that, pushed out a fix#asahi-dev · 2022-02-17 19:53
axboejannau: doh#asahi-dev · 2022-02-17 19:46
axboesven: thanks!#asahi-dev · 2022-02-17 19:40
axboeor I'll have to borrow the wife's laptop later#asahi-dev · 2022-02-17 19:40
axboebtw, if someone can test the above on osx, would be great#asahi-dev · 2022-02-17 19:33
axboebut don't judge, fio is probably around 20 years old at this point#asahi-dev · 2022-02-17 19:33
axboealyssa: it's a handrolled one ;)#asahi-dev · 2022-02-17 19:10
axboemarcan: https://git.kernel.dk/cgit/fio/commit/?id=a04e0665cb5d3a545ab1dbe2d2b7c150b404735d#asahi-dev · 2022-02-17 18:30
axboebeats real work, right?#asahi-dev · 2022-02-17 18:30
axboebut yeah, worth an experiment#asahi-dev · 2022-02-17 18:30
axboedoing 32 issues batched, no real gain :/#asahi-dev · 2022-02-17 18:29
axboeso not worth pursuing imho#asahi-dev · 2022-02-17 18:29
axboesven: got queue_rqs setup, and it's just cpu reduction at this point, iops are capped / controller limited#asahi-dev · 2022-02-17 17:33
axboeso guessing we're actually controller limited on iops anyway around that point#asahi-dev · 2022-02-17 17:32
axboebut with the changes, we get around the same perf, but at about half the CPU usage#asahi-dev · 2022-02-17 17:32
axboeoh well, guess we can't easily do that#asahi-dev · 2022-02-17 17:32
axboesven: ok queue_rqs is hard to support, since the sq door bell is writing the specific tag...#asahi-dev · 2022-02-17 17:14
axboeand then apple-nvme needs to support ->queue_rqs too#asahi-dev · 2022-02-17 17:14
axboeI should align my buffers better and not assume 4k page sizes :)#asahi-dev · 2022-02-17 17:14
axboeoh, it's the 16k page size#asahi-dev · 2022-02-17 17:10
axboelet's debug...