On 4/8/21 12:17 PM, Mark Brown wrote: > On Mon, Apr 05, 2021 at 03:43:10PM -0500, madvenka@xxxxxxxxxxxxxxxxxxx wrote: > >> These checks will involve checking the return PC to see if it falls inside >> any special functions where the stack trace is considered unreliable. >> Implement the infrastructure needed for this. > > Following up again based on an off-list discussion with Mark Rutland: > while I think this is a reasonable implementation for specifically > listing functions that cause problems we could make life easier for > ourselves by instead using annotations at the call sites to put things > into sections which indicate that they're unsafe for unwinding, we can > then check for any address in one of those sections (or possibly do the > reverse and check for any address in a section we specifically know is > safe) rather than having to enumerate problematic functions in the > unwinder. This also has the advantage of not having a list that's > separate to the functions themselves so it's less likely that the > unwinder will get out of sync with the rest of the code as things evolve. > > We already have SYM_CODE_START() annotations in the code for assembly > functions that aren't using the standard calling convention which should > help a lot here, we could add a variant of that for things that we know > are safe on stacks (like those we expect to find at the bottom of > stacks). > As I already mentioned before, I like the idea of sections. The only reason that I did not try it was that I have to address FTRACE trampolines and the kretprobe_trampoline (and optprobes in the future). I have the following options: 1. Create a common section (I will have to come up with an appropriate name) and put all such functions in that one section. 2. Create one section for each logical type (exception section, ftrace section and kprobe section) or some such. 3. Use the section idea only for the el1 exceptions. For the others use the current special_functions[] approach. Which one do you and Mark Rutland prefer? Or, is there another choice? Madhavan
