On Thu, 2024-10-17 at 15:17 +0800, David Gow wrote: > It turns out that this breaks the KUnit user alloc helpers on x86_64, > at least on my machine.
Yay, second bug from this ;-) > This can be reproduced with: > ./tools/testing/kunit/kunit.py run usercopy > > Though the 32-bit version works: > ./tools/testing/kunit/kunit.py run usercopy --kconfig_add CONFIG_64BIT=n > > The error we're getting is: > start_userspace : expected SIGSTOP, got status = 139 > Could not create userspace mm > > This basically is the result of the stub_exe segfaulting very early on > in its execution. > > It seems that this is due to the stack being misaligned, and so the > generated SSE instructions are faulting. The workarounds I've tested > here include: > a) Build the stub with -mno-sse > b) Decorate real_init() with __attribute__((force_align_arg_pointer)) > c) Decorate __start() with __attribute__((naked)) > > The last one seems to validate my theory as to why this is occurring: > __start's prologue is misaligning the stack, as __start is not > actually _called_ from anything, so there's no 8-byte misalignment to > hold the return address. > > If this makes sense, I'll send a patch out with whichever the > preferred fix(es) are. My guess is that (c) is the "proper" fix, > though I'd not _miss_ SSE if we chose to disable it for the handful of > instructions here anyway. Interesting. Actually somewhere here while reviewing this, though I don't remember precisely if it was _start, I thought we might need __attribute__((naked)) to ensure we don't get extra things, but I let that thought go since it seemed to work and I didn't wrap my head around it too much... I wonder now if the SSE instructions generated are memset() and that goes away with the patches that Nathan just sent to not have the memset (which was due to -ftrivial-auto-var-init) in the first place? But anyway all of these pretty much sound reasonable. We don't need much to happen in the stub here, simpler is better. johannes
