Patchwork [Bug,668799] Re: qemu-arm segfaults executing msgmerge (gettext)

login
register
mail settings
Submitter Peter Maydell
Date Nov. 19, 2010, 5:50 p.m.
Message ID <20101119175045.31464.28699.malone@wampee.canonical.com>
Download mbox | patch
Permalink /patch/72280/
State New
Headers show

Comments

Peter Maydell - Nov. 19, 2010, 5:50 p.m.
The following patch stops the segfault (which happens because
cpu_unlink_tb() is fiddling with the links between tbs without taking
the tb_lock, so another thread can come in via eg tb_add_jump() and
cause corruption of the linked lists). However, there are a number of
comments in the TB handling code about things being non-thread-safe or
not SMP safe, so I need to have a more careful think about the whole
thing.

Patch

diff --git a/exec.c b/exec.c
index db9ff55..5f4a50b 100644
--- a/exec.c
+++ b/exec.c
@@ -1606,9 +1606,8 @@  static void cpu_unlink_tb(CPUState *env)
        emulation this often isn't actually as bad as it sounds.  Often
        signals are used primarily to interrupt blocking syscalls.  */
     TranslationBlock *tb;
-    static spinlock_t interrupt_lock = SPIN_LOCK_UNLOCKED;
 
-    spin_lock(&interrupt_lock);
+    spin_lock(&tb_lock);
     tb = env->current_tb;
     /* if the cpu is currently executing code, we must unlink it and
        all the potentially executing TB */
@@ -1616,7 +1615,7 @@  static void cpu_unlink_tb(CPUState *env)
         env->current_tb = NULL;
         tb_reset_jump_recursive(tb);
     }
-    spin_unlock(&interrupt_lock);
+    spin_unlock(&tb_lock);
 }
 
 /* mask must never be zero, except for A20 change call */