cpu_profiler: avoid taking samples during context switches

If the profiler's signal lands while a `seastar::thread` is switching
stacks, backtracing walks inconsistent register state and can crash in
libgcc's unwinder. This was observed as a `SIGSEGV` in `uw_update_context`
on aarch64 ASan builds while a debug bundle's 30s cpu_profile request
sampled admin handlers running under ss::async: glibc's aarch64
swapcontext has a nop-only CFI program but clobbers the link register,
so the unwinder computed register-save addresses from a garbage CFA.

Check the context-switch flag introduced for the stall detector and
drop the sample instead, counting it under a new dedicated stat, the
same treatment as samples dropped during exception unwinding.

Author: WillemKauf

include/seastar/core/internal/cpu_profiler.hh

@@ -82,19 +82,22 @@ struct cpu_profiler_stats {
     unsigned dropped_samples_from_exceptions{0};
     unsigned dropped_samples_from_buffer_full{0};
     unsigned dropped_samples_from_mutex_contention{0};
+    unsigned dropped_samples_from_context_switches{0};
 
     void clear_dropped() {
         dropped_samples_from_manual_disablement = 0;
         dropped_samples_from_exceptions = 0;
         dropped_samples_from_buffer_full = 0;
         dropped_samples_from_mutex_contention = 0;
+        dropped_samples_from_context_switches = 0;
     }
 
     unsigned sum_dropped() const {
         return dropped_samples_from_manual_disablement
             + dropped_samples_from_buffer_full
             + dropped_samples_from_exceptions
-            + dropped_samples_from_mutex_contention;
+            + dropped_samples_from_mutex_contention
+            + dropped_samples_from_context_switches;
     }
 };
 

src/core/cpu_profiler.cc

@@ -25,6 +25,7 @@
 
 #include <seastar/core/internal/cpu_profiler.hh>
 #include <seastar/core/scheduling.hh>
+#include <seastar/core/thread_impl.hh>
 #include <seastar/util/log.hh>
 
 namespace seastar {
@@ -117,13 +118,20 @@ void cpu_profiler::on_signal() {
     // is currently being unwound and avoid taking a profiling sample if so.
     //
     // Note: this only protects against C++ exceptions, therefore foreign
-    // exceptions or long jumps could still cause segfaults within the profiler.
+    // exceptions could still cause segfaults within the profiler.
     const bool no_uncaught_exceptions = std::uncaught_exceptions() == 0;
 
     if (force_drop_stacktraces) {
         _stats.dropped_samples_from_manual_disablement++;
     } else if (!no_uncaught_exceptions) {
         _stats.dropped_samples_from_exceptions++;
+    } else if (thread_impl::is_context_switch_in_progress()) {
+        // While a seastar::thread switches stacks the register state is
+        // inconsistent with the unwind info (glibc's swapcontext/longjmp
+        // clobber the link register on aarch64 while their CFI still claims
+        // the return address lives there), so backtracing from here would
+        // walk garbage frames and can crash in libgcc's unwinder.
+        _stats.dropped_samples_from_context_switches++;
     } else if (auto guard_opt = _traces_mutex.try_lock(); guard_opt.has_value()) {
         // Skip the sample if the main thread is currently reading
         // _traces. This case shouldn't happen often though.