On Wed, Sep 02, 2020 at 07:50:27PM +0100, Mark Rutland wrote: > On Wed, Sep 02, 2020 at 11:32:13AM +0200, Miroslav Benes wrote: > > > - start_backtrace(&frame, > > > - (unsigned long)__builtin_frame_address(0), > > > - (unsigned long)__save_stack_trace); > Oh whoops; I'm annoyed I didn't spot that. > With that gone this cannot work for (task == current && regs == NULL), as > we'll erroneously use stale values from the task struct. I remember somehow convincing myself at the time I originally did this that doing the above was redundant with the new interface but that was quite some time ago and I can't reconstruct my reasoning any more, I'm pretty sure I was just mistaken. I've added it back in, thanks for spotting this. > It looks like the LKDTM tests only trigger cases with non-NULL regs, but > IIUC this should show up with show_stack(NULL, NULL, KERN_INFO), as > drivers/tty/sysrq.c does for other cpus. show_stack() doesn't go through this bit of the stacktrace code, it goes through dump_backtrace() in traps.c which used the underlying arch specific unwinder directly so is unaffected by arch_stack_walk(). Actually now I look at LKDTM it's ending up using show_stack() mostly if not entirely so my testing with it was not exercising this change as much as might be expected anyway (the modified code was getting hit by other things like /proc/N/stack).
Attachment:
signature.asc
Description: PGP signature
