Improve pledge() and avoid CUDA exit() in server

Author: jart

llama.cpp/ggml-backend.c

@@ -2248,7 +2248,7 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
 }
 
 GGML_CALL static void system_exit(int rc) {
-    exit(rc);
+    pthread_exit(0);
 }
 
 GGML_CALL static void system_free(void *p) {

llamafile/server/main.cpp

@@ -105,22 +105,6 @@ main(int argc, char* argv[])
     for (int i = 0; i < FLAG_workers; ++i)
         npassert(!g_server->spawn());
 
-    // install security
-    if (!FLAG_unsecure) {
-        const char* promises;
-        if (FLAG_www_root) {
-            promises = "stdio anet rpath";
-        } else {
-            promises = "stdio anet";
-        }
-        if (pledge(0, 0)) {
-            SLOG("warning: this OS doesn't support pledge() security");
-        } else if (pledge("stdio anet", 0)) {
-            perror("pledge");
-            exit(1);
-        }
-    }
-
     // run server
     signals_init();
     llama_backend_init();

llamafile/server/worker.cpp

@@ -24,8 +24,9 @@
 #include "llamafile/server/tokenbucket.h"
 #include "llamafile/threadlocal.h"
 #include "llamafile/trust.h"
-#include <cosmo.h>
+#include <atomic>
 #include <cassert>
+#include <cosmo.h>
 #include <exception>
 #include <pthread.h>
 
@@ -135,6 +136,28 @@ Worker::handle()
 void
 Worker::run()
 {
+    if (!FLAG_unsecure) {
+        static std::atomic<bool> once;
+        if (llamafile_has_gpu()) {
+            if (!once.exchange(true))
+                SLOG("warning: gpu mode disables pledge security");
+        } else {
+            const char* promises;
+            if (FLAG_www_root && !startswith(FLAG_www_root, "/zip/")) {
+                promises = "stdio anet rpath";
+            } else {
+                promises = "stdio anet";
+            }
+            if (pledge(0, 0)) {
+                if (!once.exchange(true))
+                    SLOG("warning: this OS doesn't support pledge() security");
+            } else if (pledge(promises, 0)) {
+                perror("pledge");
+                exit(1);
+            }
+        }
+    }
+
     server_->lock();
     dll_make_first(&server_->idle_workers, &elem_);
     server_->worker_count.fetch_add(1, std::memory_order_acq_rel);