[patch, validator] fix files_lock related deadlock

[Ingo Molnar]

> to solve this we must either change files_lock to be softirq-safe too 
> (bleh!), or we must forbid remove_proc_entry() use from softirq 
> contexts. Neither is a happy solution - remove_proc_entry() is used 
> within free_irq(), and who knows how many drivers do free_irq() in 
> softirq/tasklet context ...
> 
> Andrew, this needs to be resolved before v2.6.16, correct? Steve's 
> patch solves a real bug in the upstream kernel.

the patch below does the easier and safer change: it makes files_lock 
softirq-safe. (A quick test shows that the validator does not complain 
when this patch is applied too - so it seems the 'softirq effect' does 
not spread to other VFS locks.)

	Ingo

-----
the validator just found another problem with this lock, pointing out 
that files_lock nests inside of proc_subdir_lock, and that files_lock is 
a softirq-unsafe lock, creating another (unlikely but possible) deadlock 
scenario:

 =====================================
 [ BUG: lock inversion bug detected! ]
 -------------------------------------
 grep/2290 just changed the state of lock {proc_subdir_lock} at:
  [<c0196e53>] remove_proc_entry+0x33/0x1f0
 but this lock took lock {files_lock} in the past,
 acquired at: [<c0196ece>] remove_proc_entry+0xae/0x1f0
 and interrupts could create an inverse lock dependency between them,
 which could lead to deadlocks!
 other info that might help in debugging this:
 ------------------------------
 | showing all locks held by: |  (grep/2290 [c321c790, 125]):
 ------------------------------

  [<c010432d>] show_trace+0xd/0x10
  [<c0104347>] dump_stack+0x17/0x20
  [<c0137b11>] check_no_lock_2_mask+0x131/0x180
  [<c0137ffb>] mark_lock+0xfb/0x2a0
  [<c01387b3>] debug_lock_chain+0x613/0x10d0
  [<c01392ad>] debug_lock_chain_spin+0x3d/0x60
  [<c02656ed>] _raw_spin_lock+0x2d/0x90
  [<c04d88d2>] _spin_lock_bh+0x12/0x20
  [<c0196e53>] remove_proc_entry+0x33/0x1f0
  [<c01427c9>] unregister_handler_proc+0x19/0x20
  [<c0141f8b>] free_irq+0x7b/0xe0
  [<c02f2302>] floppy_release_irq_and_dma+0x1b2/0x210
  [<c02f07f7>] set_dor+0xc7/0x1b0
  [<c02f3871>] motor_off_callback+0x21/0x30
  [<c01273a5>] run_timer_softirq+0xf5/0x1f0
  [<c0122cf7>] __do_softirq+0x97/0x130
  [<c0105519>] do_softirq+0x69/0x100
  =======================
  [<c01229a9>] irq_exit+0x39/0x50
  [<c010f4cc>] smp_apic_timer_interrupt+0x4c/0x50
  [<c010393b>] apic_timer_interrupt+0x27/0x2c

to solve this we must either change files_lock to be softirq-safe too 
(bleh!), or we must forbid remove_proc_entry() use from softirq 
contexts. Neither is a happy solution - remove_proc_entry() is used 
within free_irq(), and who knows how many drivers do free_irq() in 
softirq/tasklet context ...

the patch below makes files_lock softirq-safe.

Signed-off-by: Ingo Molnar <mingo@xxxxxxx>

----

 include/linux/fs.h |    8 ++++++--
 1 files changed, 6 insertions(+), 2 deletions(-)

Index: linux/include/linux/fs.h
===================================================================
--- linux.orig/include/linux/fs.h
+++ linux/include/linux/fs.h
@@ -648,9 +648,13 @@ struct file {
 #endif /* #ifdef CONFIG_EPOLL */
 	struct address_space	*f_mapping;
 };
+/*
+ * files_lock can also be taken from softirq context:
+ */
 extern spinlock_t files_lock;
-#define file_list_lock() spin_lock(&files_lock);
-#define file_list_unlock() spin_unlock(&files_lock);
+
+#define file_list_lock()	spin_lock_bh(&files_lock);
+#define file_list_unlock()	spin_unlock_bh(&files_lock);
 
 #define get_file(x)	atomic_inc(&(x)->f_count)
 #define file_count(x)	atomic_read(&(x)->f_count)
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/