fix(profiling): avoid stack size limits on musl for stack v2 threads (#12174)

Co-authored-by: Yun Kim <[email protected]>
Co-authored-by: Nicole Cybul <[email protected]>
Co-authored-by: Nick Ripley <[email protected]>
Co-authored-by: William Conti <[email protected]>
Co-authored-by: Christophe Papazian <[email protected]>
Co-authored-by: Munir Abdinur <[email protected]>
Co-authored-by: Taegyun Kim <[email protected]>

Author: 8 people

ddtrace/internal/datadog/profiling/stack_v2/include/sampler.hpp

@@ -25,9 +25,6 @@ class Sampler
     // Parameters
     uint64_t echion_frame_cache_size = g_default_echion_frame_cache_size;
 
-    // Helper function; implementation of the echion sampling thread
-    void sampling_thread(const uint64_t seq_num);
-
     // This is a singleton, so no public constructor
     Sampler();
 
@@ -37,7 +34,7 @@ class Sampler
   public:
     // Singleton instance
     static Sampler& get();
-    void start();
+    bool start();
     void stop();
     void register_thread(uint64_t id, uint64_t native_id, const char* name);
     void unregister_thread(uint64_t id);
@@ -46,6 +43,7 @@ class Sampler
                       PyObject* _asyncio_scheduled_tasks,
                       PyObject* _asyncio_eager_tasks);
     void link_tasks(PyObject* parent, PyObject* child);
+    void sampling_thread(const uint64_t seq_num);
 
     // The Python side dynamically adjusts the sampling rate based on overhead, so we need to be able to update our
     // own intervals accordingly.  Rather than a preemptive measure, we assume the rate is ~fairly stable and just

ddtrace/internal/datadog/profiling/stack_v2/src/sampler.cpp

@@ -10,6 +10,52 @@
 
 using namespace Datadog;
 
+// Helper class for spawning a std::thread with control over its default stack size
+#ifdef __linux__
+#include <sys/resource.h>
+#include <unistd.h>
+
+struct ThreadArgs
+{
+    Sampler* sampler;
+    uint64_t seq_num;
+};
+
+void*
+call_sampling_thread(void* args)
+{
+    ThreadArgs thread_args = *static_cast<ThreadArgs*>(args);
+    delete static_cast<ThreadArgs*>(args); // no longer needed, dynamic alloc
+    Sampler* sampler = thread_args.sampler;
+    sampler->sampling_thread(thread_args.seq_num);
+    return nullptr;
+}
+
+pthread_t
+create_thread_with_stack(size_t stack_size, Sampler* sampler, uint64_t seq_num)
+{
+    pthread_attr_t attr;
+    if (pthread_attr_init(&attr) != 0) {
+        return 0;
+    }
+    if (stack_size > 0) {
+        pthread_attr_setstacksize(&attr, stack_size);
+    }
+
+    pthread_t thread_id;
+    ThreadArgs* thread_args = new ThreadArgs{ sampler, seq_num };
+    int ret = pthread_create(&thread_id, &attr, call_sampling_thread, thread_args);
+
+    pthread_attr_destroy(&attr);
+
+    if (ret != 0) {
+        delete thread_args; // usually deleted in the thread, but need to clean it up here
+        return 0;
+    }
+    return thread_id;
+}
+#endif
+
 void
 Sampler::sampling_thread(const uint64_t seq_num)
 {
@@ -134,7 +180,7 @@ Sampler::unregister_thread(uint64_t id)
     thread_info_map.erase(id);
 }
 
-void
+bool
 Sampler::start()
 {
     static std::once_flag once;
@@ -143,8 +189,22 @@ Sampler::start()
     // Launch the sampling thread.
     // Thread lifetime is bounded by the value of the sequence number.  When it is changed from the value the thread was
     // launched with, the thread will exit.
-    std::thread t(&Sampler::sampling_thread, this, ++thread_seq_num);
-    t.detach();
+#ifdef __linux__
+    // We might as well get the default stack size and use that
+    rlimit stack_sz = {};
+    getrlimit(RLIMIT_STACK, &stack_sz);
+    if (create_thread_with_stack(stack_sz.rlim_cur, this, ++thread_seq_num) == 0) {
+        return false;
+    }
+#else
+    try {
+        std::thread t(&Sampler::sampling_thread, this, ++thread_seq_num);
+        t.detach();
+    } catch (const std::exception& e) {
+        return false;
+    }
+#endif
+    return true;
 }
 
 void

ddtrace/internal/datadog/profiling/stack_v2/src/stack_v2.cpp

@@ -21,8 +21,10 @@ _stack_v2_start(PyObject* self, PyObject* args, PyObject* kwargs)
     }
 
     Sampler::get().set_interval(min_interval_s);
-    Sampler::get().start();
-    Py_RETURN_NONE;
+    if (Sampler::get().start()) {
+        Py_RETURN_TRUE;
+    }
+    Py_RETURN_FALSE;
 }
 
 // Bypasses the old-style cast warning with an unchecked helper function

releasenotes/notes/fix-stackv2-musl-stack-size-7c265de9939ce2ce.yaml

@@ -0,0 +1,6 @@
+---
+fixes:
+  - |
+    Fixes an issue where Profiling native threads would respect the musl libc
+    default stack size, which could cause stack overflows in certain
+    configurations.

tests/smoke_test.py

@@ -82,9 +82,8 @@ def emit(self, record):
         platform.system() == "Windows"
         # libdatadog x86_64-apple-darwin has not yet been integrated to dd-trace-py
         or (platform.system() == "Darwin" and platform.machine() == "x86_64")
-        # echion crashes on musl linux with Python 3.12 for both x86_64 and
-        # aarch64
-        or (platform.system() == "Linux" and sys.version_info[:2] == (3, 12) and platform.libc_ver()[0] != "glibc")
+        # echion only works with 3.8+
+        or sys.version_info < (3, 8, 0)
     ):
         orig_env = os.environ.copy()
         copied_env = copy.deepcopy(orig_env)