Enable per-thread unwind caching.

This is much more performant as libunwind will not also add another
serialization point, it will also trigger calls to sigprocmask for
every mutex lock/unlock.

diff --git a/malloctrace.cpp b/malloctrace.cpp
index d672ef1..cfcf2a8 100644 (file)
--- a/malloctrace.cpp
+++ b/malloctrace.cpp
@@ -407,6 +407,10 @@ void init()
     real_valloc = findReal<valloc_t>("valloc");
     real_aligned_alloc = findReal<aligned_alloc_t>("aligned_alloc");
 
+    if (unw_set_caching_policy(unw_local_addr_space, UNW_CACHE_PER_THREAD)) {
+        fprintf(stderr, "Failed to enable per-thread libunwind caching.\n");
+    }
+
     data.reset(new Data);
 }