08:57 mlankhorst: Looks like drm-tip was silently broken in xe_svm.c with a double definition of migrate_try_count, added fixup
11:19 stsquad: digetx: so is the vkmark issue with the system virglrenderer just an issue of missaligned headers?
11:55 stsquad: https://paste.rs/go42r
11:55 stsquad: digetx indeed with venus enabled and ASAN we detect virglrenderer crashing as a result of the commands ^
15:24 digetx: not sure, I wasn't able to reproduce the crash so far
15:27 stsquad: with debug symbols it looks like the line: ret = proxy_renderer_init(&proxy_cbs, flags | VIRGL_RENDERER_NO_VIRGL); is the failure
15:27 stsquad: digetx: I have an x86_64 image that crashes if you want
15:31 digetx: virgl_renderer_init() is invoked when qemu boots or restarts, so I assume this last crash trace isn't related to vkmark?
15:32 digetx: or it only segfaults with asan?
16:50 zmike: mareko: why does GenFramebuffers sync glthread? isn't it possible to allocate a new id without a sync?
17:07 zmike: I guess maybe it doesn't matter since CheckFramebufferStatus would have to sync anyway
20:34 mareko: zmike: glthread syncs for all functions that return something except when there is an async fast path for specific functions
20:35 mareko: zmike: if we want to make glGen async, the ID allocation and hash table would have to be moved or duplicated in front of glthread
20:36 mareko: VAO tracking already does that
20:37 zmike: right, but that wouldn't be super useful unless we also add a fastpath for CheckFramebufferStatus
20:37 zmike: i.e., something you could set for an app you know behaves and so the entrypoint will always return success without syncing
20:38 mareko: it still syncs for glGen VAOs
20:38 zmike: yeah I'm just seeing a case where framebuffer ops in particular are triggering big syncs
20:38 mareko: you can set glthread_nop_check_framebuffer_status to make CheckFramebufferStatus no-op
20:38 zmike: I don't know that it has any real perf impact
20:39 zmike: oh yeah I forgot about that
20:39 zmike: I'll have to recheck and see if it makes a difference
20:40 zmike: kinda skeptical
20:40 mareko: async glGen would be tricker because it still has to return usable and predictable IDs
20:41 zmike: isn't that only an issue with multi-context?
20:43 mareko: I don't think multi-context has anything to do with it
20:44 mareko: predictability is for traces like apitrace and viewperf
20:44 zmike: ah I see
20:44 zmike: well I'll tinker a bit if I get time and see
20:51 zzyiwei: stsquad: could you file a mesa or virglrenderer issue so folks not looking at chats here can be made aware as well. thanks