From 48e15fc2de29276f0c93482f5175b95e50557fbf Mon Sep 17 00:00:00 2001 From: Nathan Froyd Date: Fri, 29 Oct 2010 07:48:57 -0700 Subject: linux-user: fix memory leaks with NPTL emulation Running programs that create large numbers of threads, such as this snippet from libstdc++'s pthread7-rope.cc: const int max_thread_count = 4; const int max_loop_count = 10000; ... for (int j = 0; j < max_loop_count; j++) { ... for (int i = 0; i < max_thread_count; i++) pthread_create (&tid[i], NULL, thread_main, 0); for (int i = 0; i < max_thread_count; i++) pthread_join (tid[i], NULL); } in user-mode emulation will quickly run out of memory. This is caused by a failure to free memory in do_syscall prior to thread exit: /* TODO: Free CPU state. */ pthread_exit(NULL); The first step in fixing this is to make all TaskStates used by QEMU dynamically allocated. The TaskState used by the initial thread was not, as it was allocated on main's stack. So fix that, free the cpu_env, free the TaskState, and we're home free, right? Not exactly. When we create a thread, we do: ts = qemu_mallocz(sizeof(TaskState) + NEW_STACK_SIZE); ... new_stack = ts->stack; ... ret = pthread_attr_setstack(&attr, new_stack, NEW_STACK_SIZE); If we blindly free the TaskState, then, we yank the current (host) thread's stack out from underneath it while it still has things to do, like calling pthread_exit. That causes problems, as you might expect. The solution adopted here is to let the C library allocate the thread's stack (so the C library can properly clean it up at pthread_exit) and provide a hint that we want NEW_STACK_SIZE bytes of stack. With those two changes, we're done, right? Well, almost. You see, we're creating all these host threads and their parent threads never bother to check that their children are finished. There's no good place for the parent threads to do so. Therefore, we need to create the threads in a detached state so the parent thread doesn't have to call pthread_join on the child to release the child's resources; the child does so automatically. With those three major changes, we can comfortably run programs like the above without exhausting memory. We do need to delete 'stack' from the TaskState structure. Signed-off-by: Nathan Froyd Signed-off-by: Riku Voipio --- linux-user/qemu.h | 2 -- 1 file changed, 2 deletions(-) (limited to 'linux-user/qemu.h') diff --git a/linux-user/qemu.h b/linux-user/qemu.h index 708021e006..00c6549d88 100644 --- a/linux-user/qemu.h +++ b/linux-user/qemu.h @@ -126,8 +126,6 @@ typedef struct TaskState { struct sigqueue sigqueue_table[MAX_SIGQUEUE_SIZE]; /* siginfo queue */ struct sigqueue *first_free; /* first free siginfo queue entry */ int signal_pending; /* non zero if a signal may be pending */ - - uint8_t stack[0]; } __attribute__((aligned(16))) TaskState; extern char *exec_path; -- cgit v1.2.3