Add a path for BPF-accelerated async signal emulation. #3731
Conversation
src/PerfCounters.cc
Outdated
| static struct user_regs_struct* bpf_regs; | ||
|
|
||
| if (!fd_async_signal_accelerator.is_open()) { | ||
| if (!initialized) { |
There was a problem hiding this comment.
How about moving this BPF initialization code into its own function?
There was a problem hiding this comment.
It feel a bit ugly to be mashing the BPF program's global state in this function. And it's ugly to be mmapping that buffer and then leaking it to the global variable.
How hard would it be to put the BPF program and its state into its own class with proper ownership, and have each ReplaySession hold a shared pointer to an object of that class?
There was a problem hiding this comment.
Alright I reorganized this along those lines. The bpf singleton stuff lives in a BpfAccelerator class that's shared between the different PerfCounters instances.
|
|
||
| class BpfAccelerator { | ||
| public: | ||
| static std::shared_ptr<BpfAccelerator> get_or_create(); |
There was a problem hiding this comment.
I was thinking we could just create one BpfAccelerator in ReplaySession and copy the reference when we clone ReplaySessions so we don't need a static variable here.
There was a problem hiding this comment.
I'm not convinced this is a great idea. It means moving BpfAccelerator into the header so ReplaySession can get at it. Is that really better than a static singleton?
Starting in kernel 6.10 BPF filters can choose whether or not to trigger the SIGIO behavior for a perf event that becomes readable. We combine that with a hardware breakpoint and a BPF filter that matches the GPRs to produce an accelerated internal breakpoint type that can fast forward through loop iterations to deliver async signals. On one trace this reduced rr's replay overhead by 94%. This adds a runtime dependency on libbpf and a compile time dependency on clang --target bpf. rr also needs CAP_BPF and CAP_PERFMON to use this feature. Because of all of that, this isn't really suitable for wide use at this point and is instead a CMake feature usebpf. Set -Dusebpf=ON to test it.
Starting in kernel 6.10 BPF filters can choose whether or not to trigger the SIGIO behavior for a perf event that becomes readable. We combine that with a hardware breakpoint and a BPF filter that matches the GPRs to produce an accelerated internal breakpoint type that can fast forward through loop iterations to deliver async signals. On one trace this reduced rr's replay overhead by 94%.
This adds a runtime dependency on libbpf and a compile time dependency on clang --target bpf. rr also needs CAP_BPF and CAP_PERFMON to use this feature. Because of all of that, this isn't really suitable for wide use at this point and is instead a CMake feature usebpf. Set -Dusebpf=ON to test it.
(I think we should wait until the kernel side hits Linus's tree to merge this.)