Add fast trace collection & use it for instrumenting guest memory operations #103
Conversation
syntactically
force-pushed
the
lm/tracing
branch
from
e37690f to
cbe1370
Compare
syntactically
force-pushed
the
lm/tracing
branch
from
cbe1370 to
88a6083
Compare
In the future, the outb handler will need to take a Hypervisor instance in order to be able to access register and memory state of the VM, so it doesn't make sense for these interfaces to be more public than the `Hypervisor` trait. Nobody outside of Hyperlight seems to use these at the moment, so it's probably simplest to restrict these to `pub(crate)`. Signed-off-by: Lucy Menon <[email protected]>
syntactically
force-pushed
the
lm/tracing
branch
2 times, most recently
from
538b2d8 to
becb995
Compare
This adds (unused) support for creating trace files for sandboxes and passing them around to relevant sandbox event handler code. This will be used for collecting debug trace and profiling information. Signed-off-by: Lucy Menon <[email protected]>
This will be useful in the near future, when it will allow transforming the exe_info into unwind information without an extra copy. Signed-off-by: Lucy Menon <[email protected]>
This adds a new interface which tracing code can use to request the values of registers from the hypervisor supervising a sandbox. Signed-off-by: Lucy Menon <[email protected]>
Signed-off-by: Lucy Menon <[email protected]>
This allows for producing profiles of memory usage. Signed-off-by: Lucy Menon <[email protected]>
Signed-off-by: Lucy Menon <[email protected]>
syntactically
force-pushed
the
lm/tracing
branch
from
becb995 to
482ce23
Compare
There was a problem hiding this comment.
This looks great to me. Stack unwinding will definitely be very helpful. Btw, it's not totally clear to me how to use the two added features, nor the new trace_dump crate. Would it be possible to add a simple example or minimal docs somewhere? Lastly, should we try to compile some feature matrix of this in CI to make sure we don't break it in the future? Great job 👍
There was a problem hiding this comment.
This is really great @syntactically , I added a few review comments.
I agree with @ludfjig it would be nice to have some docs that explain how to use these features and how they interact (it isn't immediately obvious to me at least if the trace_guest feature is useful on its own).
I also think it would be useful to perhaps have a README for the trace_dump tool that specifies how to use it , last I think we should maybe include the trace_dump binary in the GH release?
This adds a framework for quickly tracing events of interest in a guest, including getting stack traces out of the guest based on dwarf unwinding information. This capability is used to allow collecting stack traces of memory allocations/frees, providing various statistics on what causes the maximum memory consumption for a sandbox.