perf(tracing): cache OTel tid syscall

tests/ext/pcntl/otel_process_context_fork.phpt

@@ -16,6 +16,8 @@ DD_TRACE_GENERATE_ROOT_SPAN=0
 --FILE--
 <?php
 
+require __DIR__ . '/../includes/otel_thread_context.inc';
+
 function processContextMappings(): array
 {
     $mappings = [];
@@ -81,10 +83,15 @@ echo "Parent mapping is MADV_DONTFORK: "; var_dump(
 );
 $parentRuntimeId = processContextRuntimeId($ffi, $parentMappings[0]);
 echo "Parent has runtime ID: "; var_dump($parentRuntimeId !== '');
+$threadContext = new OtelThreadContext();
 $parentRoot = DDTrace\start_span();
 echo "Parent runtime ID matches tracer: "; var_dump(
     $parentRoot->meta['runtime-id'] === $parentRuntimeId
 );
+$parentThreadId = $threadContext->attributes()[4] ?? null;
+echo "Parent thread ID matches process ID: "; var_dump(
+    $parentThreadId === (string) getmypid()
+);
 DDTrace\close_span();
 
 $pid = pcntl_fork();
@@ -106,6 +113,11 @@ if ($pid === 0) {
     echo "Child mapping was republished: "; var_dump(
         $childRuntimeId !== $parentRuntimeId
     );
+    $childThreadId = $threadContext->attributes()[4] ?? null;
+    echo "Child thread ID was refreshed: "; var_dump(
+        $childThreadId === (string) getmypid()
+        && $childThreadId !== $parentThreadId
+    );
     DDTrace\close_span();
     exit;
 }
@@ -121,8 +133,10 @@ Parent has one mapping: bool(true)
 Parent mapping is MADV_DONTFORK: bool(true)
 Parent has runtime ID: bool(true)
 Parent runtime ID matches tracer: bool(true)
+Parent thread ID matches process ID: bool(true)
 Child has one mapping: bool(true)
 Child mapping is MADV_DONTFORK: bool(true)
 Child runtime ID matches tracer: bool(true)
 Child mapping was republished: bool(true)
+Child thread ID was refreshed: bool(true)
 Child exited successfully: bool(true)

tracer/ddtrace.c

@@ -686,15 +686,15 @@ bool ddtrace_update_remote_config_flags(ddog_RemoteConfigFlags *flags) {
 #define JOIN_BGS_BEFORE_FORK 1
 #endif
 
-void ddtrace_internal_handle_prefork() {
+void ddtrace_internal_handle_prefork(void) {
 #if JOIN_BGS_BEFORE_FORK
     if (!get_global_DD_TRACE_SIDECAR_TRACE_SENDER()) {
         ddtrace_coms_flush_shutdown_writer_synchronous();
     }
 #endif
 }
 
-void ddtrace_internal_handle_postfork() {
+void ddtrace_internal_handle_postfork(void) {
 #if JOIN_BGS_BEFORE_FORK
     if (!get_global_DD_TRACE_SIDECAR_TRACE_SENDER()) {
         ddtrace_coms_restart_writer();
@@ -705,6 +705,7 @@ void ddtrace_internal_handle_postfork() {
 void ddtrace_internal_handle_fork() {
 #ifdef __linux__
     ddtrace_otel_detach();
+    ddtrace_otel_tid_fork_handler();
 #endif
     if (DATADOG_G(sidecar)) {
         // Unconditionally send, even if root span is NULL

tracer/otel_context.c

@@ -37,6 +37,15 @@ _Static_assert(offsetof(ddtrace_root_span_data, otel_context) % 8 == 0, "unexpec
 _Static_assert((offsetof(ddtrace_root_span_data, otel_context) + offsetof(datadog_otel_thr_ctx_rec, span_id)) % 8 == 0,
                "unexpected OTel thread context span_id placement");
 
+typedef struct {
+    // The length of the used bytes in .value, excluding null.
+    uint8_t len;
+    // On Linux tids are pid_t, which are `int`, so the most negative value is
+    // the longest when printed (because of sign bit).
+    char value[sizeof("-2147483648")];
+} ddtrace_otel_cached_tid;
+
+static __thread ddtrace_otel_cached_tid otel_tid = {0};
 __thread void *otel_thread_ctx_v1 __attribute__((visibility("default"), tls_model("global-dynamic")));
 
 static void ddtrace_otel_record_begin_update(datadog_otel_thr_ctx_rec *record);
@@ -203,10 +212,20 @@ static void ddtrace_otel_record_set_attrs(datadog_otel_thr_ctx_rec *record, ddtr
     offset = ddtrace_otel_record_write_attr_zstr(record, offset, DDTRACE_OTEL_ATTR_DEPLOYMENT_ENVIRONMENT_NAME, ddtrace_otel_attr_zstr(env));
     offset = ddtrace_otel_record_write_attr_zstr(record, offset, DDTRACE_OTEL_ATTR_SERVICE_VERSION, ddtrace_otel_attr_zstr(version));
 
-    char thread_id[32];
-    int thread_id_len = snprintf(thread_id, sizeof(thread_id), "%llu", (unsigned long long)syscall(SYS_gettid));
-    if (thread_id_len > 0) {
-        offset = ddtrace_otel_record_write_attr(record, offset, DDTRACE_OTEL_ATTR_THREAD_ID, thread_id, (size_t)thread_id_len);
+    if (UNEXPECTED(otel_tid.len == 0)) { // should only happen 1x per thread
+        long tid = syscall(SYS_gettid);
+        if (EXPECTED(tid > 0)) { // this syscall is fundamental and shouldn't fail
+            // tids are just pid_t but there's no safety in casting down to int
+            // here, if the buffer is big enough, then we're good.
+            int tid_len = snprintf(otel_tid.value, sizeof otel_tid.value, "%ld", tid);
+            // Should fit, tids are `int` on Linux.
+            if (EXPECTED(tid_len > 0 && (size_t)tid_len < sizeof otel_tid.value)) {
+                otel_tid.len = (uint8_t)tid_len;
+            }
+        }
+    }
+    if (EXPECTED(otel_tid.len != 0)) {
+        offset = ddtrace_otel_record_write_attr(record, offset, DDTRACE_OTEL_ATTR_THREAD_ID, otel_tid.value, (size_t)otel_tid.len);
     }
     record->attrs_data_size = (uint16_t)offset;
 
@@ -241,3 +260,7 @@ static size_t ddtrace_otel_record_write_attr(datadog_otel_thr_ctx_rec *record, s
 }
 
 static zend_string *ddtrace_otel_attr_zstr(zend_string *value) { return value ? value : ZSTR_EMPTY_ALLOC(); }
+
+void ddtrace_otel_tid_fork_handler(void) {
+    otel_tid.len = 0;
+}

tracer/otel_context.h

@@ -63,6 +63,9 @@ void ddtrace_otel_attach_stack(ddtrace_span_stack *stack);
 /** Detach the current thread's record. */
 void ddtrace_otel_detach(void);
 
+/** Reset the tid after a fork. */
+void ddtrace_otel_tid_fork_handler(void);
+
 END_EXTERN_C()
 
 #endif  // DDTRACE_OTEL_CONTEXT_H