Dump at least partial stack trace for stack overflow

[janvorli]

When stack overflow happens, it is sometimes hard to determine the real cause, since at stack overflow, the stack trace is not printed.
Since we've recently switched to allocating the alternate stack space using mmap, we could actually reserve larger VM space and commit just the size needed by the regular sigsegv handling. On stack overflow, we could commit more of the space so that we have enough to dump at least partial stack trace.

[markusschaber]

I currently have a problem where I cannot even get Stack Trace with Visual Studio debugger... So anything which could help us to get a clue would be welcome... :-)

[janvorli]

@markusschaber - are you talking about debugging Windows process or Unix one?

[markusschaber]

@janvorli I had the problem on Windows 10, using rather recent VS 2017 (Version 15.8.x or 15.9.0), running dotnet.exe (2.1, x64) on the same machine. As we were using p/invoke (but not in the code path involved), I had native debugging enabled. I was not able to get a stack trace of the stack involved, neither when running the service in the debugger, nor when attaching to it later or via "just in time" debugging.

[danmoseley]

@janvorli @tommcdon do you think this is possible in the 5.0 timeframe? Arguably this falls under "cloud native" since it requires you set up and attach a debugger to get an idea of the problem, which is not always easy if it only reproes in production. A callstack could be the clue you need to avoid that.

Pretty much every SO I have seen repeats after a few frames (if not 1) so even a few frames would give you almost all the full stack.

[GSPP]

@danmosemsft that is a good point. And without any stack you are pretty much lost. Stack overflows are tough to debug.

[tommcdon]

@janvorli this seems like a reasonable request for 5.0. We have other issues on the debugger where SOS PE on the StackOverflowException object and SOS ClrStack commeands do not work either. dotnet/diagnostics#135 dotnet/diagnostics#66
@sdmaclea @mikem8361 @sergiy-k

[yyjdelete]

Besides the top N stack, the bottom N stack maybe also needed to make an diag(to see what is called before the stack loop) and reproduce.

[janvorli]

I have made a working prototype that dumps the whole stack trace some time ago already, but I was waiting for the fix for dotnet/coreclr#21061 to finalize it, as the prototype was working only for cases when the stack overflow happened in methods without stack probing.
I was thinking it might be nice to try to "compress" the stack trace in some way to make it shorter and easier to read. E.g. in the case when a single frame repeats n times we could print the frame just once with the repeat count or we might even identify a repeated sequence of the same frames. But these were just quick ideas and I haven't thought much about the feasibility or complexity of doing that.

[danmoseley]

If it's feasible to include the start of that stack as well (sounds like it, which is great) then I think compressing it as you describe is inevitable. The output would be huge and unwieldy otherwise.

Incidentally what would be the experience in VS? There is not always a console attached to dump it to - would it show up in the debugger output window or the stack window?

[janvorli]

@danmosemsft It would dump to whatever output the fail fast output goes. I am not sure how that behaves in VS.

[mikem8361]

Will what you are doing for this issue fix the DAC problems in dotnet/diagnostics#66? The stack unwinding code doesn't like that we are on the alternate stack. This is way SOS, etc. doesn't work that well for stack overflow.

[janvorli]

@mikem8361 it would not fix that problem. I think that a solution to that one is to disable the asserts. I had to do it in couple of places in my prototype too. We cannot extend the stack in the original continuous VA range. Although we could possibly add the alternate stack ranges to the Thread and check SP against both the alternate and the regular stack ranges, it seems like an overkill just to make an assert happy.