Re: [RFC][PATCH] f2fs: Avoid print false deadlock messages.

[majianpeng <majianpeng@gmail.com>]

On 05/16/2013 04:41 PM, Peter Zijlstra wrote:
> On Thu, May 16, 2013 at 09:16:45AM +0800, majianpeng wrote:
>
>>> There isn't. What you typically want to do is annotate the lock site.
>>> In particular it looks like mutex_lock_all() is the offensive piece of
>>> code (horrible function name though; the only redeeming thing being that
>>> f2fs.h isn't likely to be included elsewhere).
>>>
>>> One thing you can do here is modify it to look like:
>>>
>>> static inline void mutex_lock_all(struct f2fs_sb_info *sbi)
>>> {
>>> 	int i;
>>>
>>> 	for (i = 0; i < NR_GLOBAL_LOCKS; i++) {
>>> 		/*
>>> 		 * This is the only time we take multiple fs_lock[]
>>> 		 * instances; the order is immaterial since we
>>> 		 * always hold cp_mutex, which serializes multiple
>>> 		 * such operations.
>>> 		 */
>>> 		mutex_lock_nest_lock(&sbi->fs_lock[i], &sbi->cp_mutex);
>>> 	}
>>> }
>>>
>>> That tells the lock validator that it is ok to lock multiple instances
>>> of the fs_lock[i] class because the lock order is guarded by cp_mutex.
>>> While your patch also works, it has multiple down-sides; its easy to get
>>> out of sync when you modify NR_GLOBAL_LOCKS; also it consumes more
>>> static lockdep resources (lockdep has to allocate all its resources
>>> from static arrays since allocating memory also uses locks -- recursive
>>> problem).
>>>
>> Yes, but there is a problem if fs_block[] met deadlock. How to find which one?
>> Because the lock->name is the same.
> The most useful part of the lockdep report are the call traces that tell you
> where locks where acquired; the names are secondary. That is, while they are at
> times helpful in finding the right lock site, they're rarely _that_ important.
>
> Remember, your code will very likely not have the exact number hardcoded either.
> It'll be a variable. So having the number in the lockdep output will not help
> you find the offending code any sooner.
>
> Suppose there's another site that acquires two fs_block[] locks; currently this
> would generate another such warning as this thread started with -- lockdep
> doesn't look at lock instances but at classes; and it cannot differentiate
> between two locks of the same class and thus reports the possible deadlock.
>
> The way to find the offending code is to look at the "locks held" section of
> the lockdep report along with the call traces.
>
> Once you find the way in which the two locks nest the specific numbers are
> irrelevant. Your aim then is to prove your locking scheme is indeed deadlock
> free and then tell lockdep about it.
>
>
Thanks very much! I'll take times to understand.
Can you send a patch about this?

Thanks!
Jianpeng Ma