trace,x86: Do not call local_irq_save() in load_current_idt()
authorSteven Rostedt (Red Hat) <rostedt@goodmis.org>
Sat, 22 Jun 2013 17:16:19 +0000 (13:16 -0400)
committerSteven Rostedt <rostedt@goodmis.org>
Sat, 22 Jun 2013 17:16:19 +0000 (13:16 -0400)
As load_current_idt() is now what is used to update the IDT for the
switches needed for NMI, lockdep debug, and for tracing, it must not
call local_irq_save(). This is because one of the users of this is
lockdep, which does tracing of local_irq_save() and when the debug
trap is hit, we need to update the IDT before tracing interrupts
being disabled. As load_current_idt() is used to do this, calling
local_irq_save() which lockdep traces, defeats the point of calling
load_current_idt().

As interrupts are already disabled when used by lockdep and NMI, the
only other user is tracing that can disable interrupts itself. Simply
have the tracing update disable interrupts before calling load_current_idt()
instead of breaking the other users.

Here's the dump that happened:

------------[ cut here ]------------
WARNING: at /work/autotest/nobackup/linux-test.git/kernel/fork.c:1196 copy_process+0x2c3/0x1398()
DEBUG_LOCKS_WARN_ON(!p->hardirqs_enabled)
Modules linked in:
CPU: 1 PID: 4570 Comm: gdm-simple-gree Not tainted 3.10.0-rc3-test+ #5
Hardware name:                  /DG965MQ, BIOS MQ96510J.86A.0372.2006.0605.1717 06/05/2006
 ffffffff81d2a7a5 ffff88006ed13d50 ffffffff8192822b ffff88006ed13d90
 ffffffff81035f25 ffff8800721c6000 ffff88006ed13da0 0000000001200011
 0000000000000000 ffff88006ed5e000 ffff8800721c6000 ffff88006ed13df0
Call Trace:
 [<ffffffff8192822b>] dump_stack+0x19/0x1b
 [<ffffffff81035f25>] warn_slowpath_common+0x67/0x80
 [<ffffffff81035fe1>] warn_slowpath_fmt+0x46/0x48
 [<ffffffff812bfc5d>] ? __raw_spin_lock_init+0x31/0x52
 [<ffffffff810341f7>] copy_process+0x2c3/0x1398
 [<ffffffff8103539d>] do_fork+0xa8/0x260
 [<ffffffff810ca7b1>] ? trace_preempt_on+0x2a/0x2f
 [<ffffffff812afb3e>] ? trace_hardirqs_on_thunk+0x3a/0x3f
 [<ffffffff81937fe7>] ? sysret_check+0x1b/0x56
 [<ffffffff81937fe7>] ? sysret_check+0x1b/0x56
 [<ffffffff810355cf>] SyS_clone+0x16/0x18
 [<ffffffff81938369>] stub_clone+0x69/0x90
 [<ffffffff81937fc2>] ? system_call_fastpath+0x16/0x1b
---[ end trace 8b157a9d20ca1aa2 ]---

in fork.c:

 #ifdef CONFIG_PROVE_LOCKING
DEBUG_LOCKS_WARN_ON(!p->hardirqs_enabled); <-- bug here
DEBUG_LOCKS_WARN_ON(!p->softirqs_enabled);
 #endif

Cc: Seiji Aguchi <seiji.aguchi@hds.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
arch/x86/include/asm/desc.h
arch/x86/kernel/tracepoint.c

index 1377ecb29d8d2769108e178d8dff2c5bcaeaa649..b90e5dfeee462b55bd942b4bdc5fab2eb4817403 100644 (file)
@@ -497,21 +497,19 @@ static inline void load_trace_idt(void)
 #endif
 
 /*
- * the load_current_idt() is called with interrupt disabled by local_irq_save()
+ * The load_current_idt() must be called with interrupts disabled
  * to avoid races. That way the IDT will always be set back to the expected
- * descriptor.
+ * descriptor. It's also called when a CPU is being initialized, and
+ * that doesn't need to disable interrupts, as nothing should be
+ * bothering the CPU then.
  */
 static inline void load_current_idt(void)
 {
-       unsigned long flags;
-
-       local_irq_save(flags);
        if (is_debug_idt_enabled())
                load_debug_idt();
        else if (is_trace_idt_enabled())
                load_trace_idt();
        else
                load_idt((const struct desc_ptr *)&idt_descr);
-       local_irq_restore(flags);
 }
 #endif /* _ASM_X86_DESC_H */
index 1423efe98fbcf1bfbbe505d42dfe55bcc5745d78..4e584a8d6eddf1fa2dbd8c540ce0f6542850bb79 100644 (file)
@@ -29,7 +29,11 @@ static void set_trace_idt_ctr(int val)
 
 static void switch_idt(void *arg)
 {
+       unsigned long flags;
+
+       local_irq_save(flags);
        load_current_idt();
+       local_irq_restore(flags);
 }
 
 void trace_irq_vector_regfunc(void)