03:01 rhyskidd: might be of interest: http://www.cs.ucr.edu/~zhiyunq/pub/ccs18_gpu_side_channel.pdf
12:06 karolherbst: rhyskidd: yeah... but there is a easy workaround afaik: disable global perf counters
12:07 karolherbst: rhyskidd: thing is, they basically do pattern matching
12:07 karolherbst: which is kind of ... well
12:07 karolherbst: it only really works for targeted attacks
12:07 karolherbst: a browser update might break the attack
12:07 karolherbst: or the website changes something
12:08 karolherbst: anyway, those attacks are no surprise at all
12:30 mupuf: karolherbst: not sure performance counters would count as side channel
12:30 karolherbst: mupuf: well, they can be used
12:30 karolherbst: and why wouldn't it count?
12:30 mupuf: they are *meant* for this
12:30 mupuf: ;p
12:30 karolherbst: well, right
12:30 karolherbst: :D
12:30 mupuf: cache timings == true side channel
12:31 karolherbst: yeah...
12:31 mupuf: but anyway, that's still a lot of work they spent to make somewhat practical
12:31 mupuf: but update the driver and boom...
12:31 karolherbst: yeah
12:31 karolherbst: that's my "targeted attack" about
12:31 mupuf: yep
12:31 karolherbst: it is still bad though
12:31 karolherbst: but the suggestions are good enough
12:31 mupuf: not gonna be a problem with nouveau... still we don't use the global counters
12:32 karolherbst: or maybe you have to be root to be able to use them?
12:32 karolherbst: allthough that sucks as well
12:32 karolherbst: but is the case for CPUs, right?
12:32 Sarayan: istr that your process has to be authorized
12:33 Sarayan: which does not necessarily mean root
12:33 karolherbst: okay, sure
12:34 karolherbst: mupuf: ohh btw, somebody filed a CVE against nouveau...
12:34 Sarayan: ah no, it's worse, they're ring 0
12:34 karolherbst: mupuf: because you can freeze the desktop through webgl
12:34 karolherbst: remote DOS
12:34 mupuf: karolherbst: true
12:34 karolherbst: oh well
12:34 karolherbst: I mean, we know that already
12:34 mupuf: but... we have bigger problems
12:34 mupuf: exactly
12:35 karolherbst: well, that one is one of our bigger problems
12:35 karolherbst: it's the channel killed, but we don't recover
12:35 karolherbst: and X waits
12:35 karolherbst: -> desktop freeze
12:35 mupuf: yes, I mean the webgl part is not the problem
12:35 karolherbst: right
12:36 karolherbst: anyway, made progress with my multi context fixes
12:36 mupuf: Sarayan: I would not mind having the accurate version of the counters exposed through root only, and having general business exposed through sysfs for applications' consumption
12:36 karolherbst: I think it works good enough now. Just need to change some small detail
12:36 mupuf: cool! What is going on there?
12:36 karolherbst: but I was able to fix my shader eviciton bug
12:36 karolherbst: mupuf: per context pushbuf/fence list
12:37 karolherbst: works pretty well
12:37 karolherbst: was able to fix crashes in several applications with no regression
12:37 mupuf: well, maybe I will be able to start using nouveau at some point on my desktop
12:37 mupuf: although it is a pascal, so the performance will be disastrous with my 4k screen
12:37 mupuf: but the crashes were the worst
12:37 mupuf: (kde user here)
12:37 karolherbst: and further issues always happen in multi context applications which would have crashed anyway
12:37 karolherbst: mupuf: you could try my branch if you want to
12:38 Sarayan: mupuf: http://www.brendangregg.com/perf.html seems to be the state of the art on linux
12:38 karolherbst: best thing is, the changes are super small
12:38 karolherbst: mupuf: 34 files changed, 352 insertions(+), 236 deletions(-)
12:39 karolherbst: still need some love for nv50 and nv30, but nvc0 should just work
12:39 mupuf: oh! That is much smaller than what you were working on at XDC, right?
12:40 karolherbst: yeah
12:40 karolherbst: here is the branch: https://github.com/karolherbst/mesa/commits/mt_fixes_take2
12:50 mupuf:is filing bugs, sorry :)
12:51 mupuf: shit hit the fan in intel CI :) Good new features exposing more bugs