#asahi-gpu · 2023-08-01 11:08
marcanwell that and the canberra thing I guess#asahi-gpu · 2023-08-01 11:07
marcannobody's ever seen kernel lockups since the latest update right?#asahi-gpu · 2023-08-01 11:07
marcanso then we just have that rare Mesa UAF thing then#asahi-gpu · 2023-08-01 10:51
_jannau__yes, I assumed it's well known that it is broken. I thought it wouldn't be a problem for arm64 since the default in make oldconfig was PER_VMA_LOCK=n#asahi-gpu · 2023-08-01 10:47
marcanhttps://lore.kernel.org/linux-mm/CAJuCfpH4HyAVO0fVYKGR9xhZJu19Ek_dTMEBiCCSbOjkDxXD=w@mail.gmail.com/T/#asahi-gpu · 2023-08-01 10:33
_jannau__that explains the difference frequency. it is 1 in x days for me#asahi-gpu · 2023-08-01 10:33
_jannau__the Firefox crashes mps saw are probably caused by CONFIG_PER_VMA_LOCK=y and thus unrelated to the one I saw#asahi-gpu · 2023-07-31 22:28
alyssai wonder if there's some sort of instruction grouping mechanism we've somehow missed.#asahi-gpu · 2023-07-31 22:20
alyssawith .discards#asahi-gpu · 2023-07-31 22:20
alyssait will do ldcf-read with alu-write with alu#asahi-gpu · 2023-07-31 22:20
alyssaok, coincidence yeah#asahi-gpu · 2023-07-31 22:18
alyssacoincidence, I hope#asahi-gpu · 2023-07-31 22:18
alyssa..I don't see any ALU writes after ldcf writes to the third register...#asahi-gpu · 2023-07-31 22:16
alyssabut what about ldcf about with ALU?#asahi-gpu · 2023-07-31 22:16
alyssaldcf definitely handles WaW hazards with itself#asahi-gpu · 2023-07-31 22:15
alyssaI have no idea how to coax the blob into doing more interesting things with ldcf to figure out how it deals with the hazard#asahi-gpu · 2023-07-31 22:13
alyssa(bit 52 is set only on the first -- that's the helper invocation discard bit, much less interesting)#asahi-gpu · 2023-07-31 22:12
alyssaI notice that bit 46 is set unpredictably, no idea what it means#asahi-gpu · 2023-07-31 22:11
alyssasome blob output with lots of ldcf instructions https://paste.centos.org/view/84334941#asahi-gpu · 2023-07-31 21:56
alyssa(if the ldcf registers are non-overlapping and the moves are removed, it works..)#asahi-gpu · 2023-07-31 21:55
alyssathis is flaking#asahi-gpu · 2023-07-31 21:55
alyssahttps://rosenzweig.io/broken.txt#asahi-gpu · 2023-07-31 21:52
alyssadougall: oof. the write-after-write hazard woes are not behind us like I thought :(#asahi-gpu · 2023-07-30 18:22
alyssamew.#asahi-gpu · 2023-07-30 17:19
jannauno, bug still there#asahi-gpu · 2023-07-30 16:53
jannauI wonder whether this might explain explain the heisenbug, I don't think it explains the firefox crash#asahi-gpu · 2023-07-30 16:52
jannaualyssa: feel free to close when implement a helper in gallium#asahi-gpu · 2023-07-30 16:51
jannauasahi specific merge request for the constant buffer upload created#asahi-gpu · 2023-07-30 06:13
linajannau: Pushed a fixup directly to main#asahi-gpu · 2023-07-30 05:49
linaI guess I can just change the condition to check for agx_linear_allowed, and if it's not, shrug and let it be compressed/tiled#asahi-gpu · 2023-07-30 05:48
lina... why are we trying to allocate mipmapped or cubemaps as shared? That looks wrong...#asahi-gpu · 2023-07-30 04:44
jannaufailing with "../src/gallium/drivers/asahi/agx_pipe.c:503: agx_select_best_modifier: Assertion `agx_linear_allowed(pres)' failed."#asahi-gpu · 2023-07-30 04:44
jannaulina: "asahi: Force linear for SHARED buffers with no/implicit modifier" broke a bunch of (all?) dEQP-EGL.functional.image.*gles2_cubemap* tests#asahi-gpu · 2023-07-30 04:42
jannaulina: alyssa found the solution and I'm very unsure how to implement it in a general way#asahi-gpu · 2023-07-30 03:52
linahttps://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21685 was my attempt but...#asahi-gpu · 2023-07-30 03:51
linaLooks like you found a better solution ^^#asahi-gpu · 2023-07-30 03:51
linajannau: I found that UAF a long time ago, tried to fix it in core mesa code, and it broke a piglit, and I back burnered it because it wasn't breaking anything IRL...#asahi-gpu · 2023-07-30 02:39
alyssaan d#asahi-gpu · 2023-07-29 22:15
alyssathat's... good? maybe?#asahi-gpu · 2023-07-29 21:20
jannauthat was the only remaining ASAN error in deqp-{egl,gles2,gles3}#asahi-gpu · 2023-07-29 20:21
alyssaThanks >:#asahi-gpu · 2023-07-29 20:21
alyssaYup.#asahi-gpu · 2023-07-29 20:17
jannaucomment in llvmpipe's set_constant_buffer suggest asahi's implementation is not sound#asahi-gpu · 2023-07-29 20:09
alyssaif you don't get to it I'll probably do that on monday#asahi-gpu · 2023-07-29 20:09
alyssaso we can fix those across drivers#asahi-gpu · 2023-07-29 20:09
alyssaI wonder if what we really want is a gallium-level helper to upload any user buffers#asahi-gpu · 2023-07-29 20:09
alyssahmm#asahi-gpu · 2023-07-29 20:08
alyssathanks in advance for the bug fix :3#asahi-gpu · 2023-07-29 20:07
alyssaprobably fster too#asahi-gpu · 2023-07-29 20:07
alyssajannau: but for asahi, I would say use const_uploader in set_const_buffer to avoid the dangling ptr