Hello! It turned out the security-uploaded build of samba on i386 is broken. There were several reports about smbd segfaulting at startup or other weirdness. This is specific to i386 build, the x64 build is fine (and I know nothing about the other architectures).
An example bugreport is #1006935 - it has several items in the list of broken things, but this list is definitely not complete. Also #1009855 #1002059. I tried to rebuild the same source package in a local clean bullseye chroot, apparently with the same versions of everything, in the same environment, and I'm getting *significantly* different binaries. Updating just one package - switching from debian-built one into my locally-built one immediately fixes the problem for me, samba starts working as it should without weird errors or crashes. The issue at hand seems to be the usage of python hashes in samba build procedure for everything including lists of include or library paths, lists of object files for link and everything. By default, python hashes are randomly-ordered, - this means the order of all this is unpredictable and each time we're getting VERY different binaries. Since samba overrides many system-provided functions, and the order of objects to link is important, in this particular build we got an order which made it use wrong functions in the end, and the thing started to behave in a weird way. Upstream tried to fix this by using PYTHONHASHSEED=1 (which *still* gives an order which depends on the architecture, but this is a different issue). So we have to fix this one in bullseye. I already prepared a bullseye-pu version of samba - #1009726 - the bug#), which does not include this fix. I'll update it today (the fix is trivial) and resubmit. The bullseye-pu version has some more fixes than is usually okay to bring to security@ but most of them are worth the effort. Now since the prob is quite serious, maybe we can fix this some faster way? Thanks, /mjt
