Hi, Attempting to summarize the findings relating to this bug as well as the comments in the mail thread, my understanding is this: In order to conform to the psABI, gcc was changed (back in 2004?) to assume 16-byte stack alignment on Linux/alpha. This seems to be the case for user-mode processes stacks, however not for the stack in kernel-mode. Some simple printouts of stack pointers in kernel mode suggest that the kernel stack is in fact only 8-byte aligned on Linux/alpha. In comparison, I've made similar checks on hppa, sparc and x86_64 and on these platforms the kernel stack seems to be 16-byte aligned (at least). If gcc assumes 16-byte alignment, and the code uses __attribute__((__aligned__(x))) gcc will generate assembly code that may cause stack corruption, if the stack at run-time is in fact only 8-byte aligned. A quick-fix/workaround for this might be to avoid using the __attribute__((__aligned__(x))) directive on variables/structs declared on the stack in the kernel code (at least for alpha), but to really get to the bottom of this, the kernel needs to be fixed so that the kernel stack is in fact always 16-byte aligned on alpha. This means that this bug is not really related to rcu or smp but rather a mismatch between gcc and linux-alpha regarding psABI compliance. /Magnus On Mon, Jan 13, 2025 at 5:52 PM Magnus Lindholm <linmag7@xxxxxxxxx> wrote: > > Hi, I've sprinkled some stack pointer printouts and its seems like (as > suspected) the kernel stack pointer is not 16-byte aligned, at least > not in kernel-mode. > > Example printouts: > SP: fffffc00059dfc08 > SP: fffffc00059dfe48 > SP: fffffc00059dfc08 > > I found some ML threads that seemed relevant one on stack alignment in > NetBSD-alpha after changes in GCC and one on x86_64 with linux/gcc > > https://mail-index.netbsd.org/port-alpha/2021/07/05/msg001145.html > > https://patchwork.kernel.org/project/linux-crypto/patch/20170110143340.GA3787@xxxxxxxxxxxxxxxxxxx/ > > > > /Magnus > > On Mon, Jan 13, 2025 at 6:59 AM Magnus Lindholm <linmag7@xxxxxxxxx> wrote: > > > > > Can you please retrieve the value of SP in `smp_call_function_single'? > > > Just something such as: > > > > > > printk("SP: %016lx\n", __builtin_frame_address(0)); > > > > > > > > > I will check! > > > > Also wondering if this may be relevant? > > > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=16660
