Ahoy,
I'm using Debian Trixie on a machine with an Intel Haswell integrated GPU with GNOME on Wayland. GNOME Maps nowadays has support for getting public transit directions in my area and displaying the possible routes on a map with vector tiles, potentially including all of the different bus stops as well. This is to say that it's probably fairly GPU-intensive when it draws the routes.
I was fumbling with a user interface quirk that made it tricky to set the departure time to my liking and so I quickly switched between the routes, changed departure times, and panned around on the map. Although this seems like a race condition I'm able to reproduce it pretty easily in a minute or two. I've seen it fail two different ways but the more typical (and insightful) way corresponds to the backtrace I'm attaching for invoking 'gjs /usr/bin/gnome-maps'.
To me, it looks like gjs uses a foreign function interface to call into GLib/GDK right within the same process, which is why gjs itself crashes. Words that stand out to me are "GskGLRenderer", "GskGLImage", and after what looks like a giant daisy chain of GSK GPU drawing operations, we find "_mesa_TexImage2D", "teximage_err", "st_AllocTextureImageBuffer", and then we get into Crocus functions before memset() tramples on the address space. (Sometimes the FORTIFY_SOURCE protections kick in and glibc intervenes before the kernel does.)
This sure does look an awful like the upstream Mesa issue 13426 I've referenced which is fixed only in Mesa 25.3.0 and newer—although because of the butterfly effect being at play, the backtrace does look kind of different every time on my end.
I am suspicious if I'm really hitting that same issue so I've marked this bug 'moreinfo' at least until I gather follow-up info. My backtrace shows 'compressed=0' but I think the upstream issue is about 'compressed=1' not being handled properly. I also set breakpoints on suspect functions mentioned in the upstream fix's commit message but I actually didn't hit any of them.
This is mainly a note for myself but my plan of attack is as follows:
• the Mesa version in trixie-backports should still be affected and I ought to confirm that
• then cherry pick the upstream fix into a local rebuild of the Debian package and see if I can reproduce the issue anymore: if not, I can confirm this and maybe we can talk about whether it deserves to get into trixie-(proposed-)updates?
• if that doesn't help, then I can fast-forward to whatever is in upstream Git—and maybe use some LD_PRELOAD or LD_LIBRARY_PATH voodoo so I don't have to deal with Debian packaging changes—and see if I can reproduce then. If I can't, then I can get to bisecting with Git. If the search space is big, I can probably script the keyboard shortcuts to provoke GNOME Maps using xdotool with Xwayland (if the issue even happens then?) or wtype.
Please note: I haven't done a search of the (potentially open) issues filed upstream yet.