Re: [PATCH] fs: Disable preempt when acquire i_size_seqcount write lock

[Fan Du <fan.du@windriver.com>]

On 2013年01月11日 06:38, Andrew Morton wrote:
> On Wed, 9 Jan 2013 11:34:19 +0800
> Fan Du<fan.du@windriver.com>  wrote:
>
>> Two rt tasks bind to one CPU core.
>>
>> The higher priority rt task A preempts a lower priority rt task B which
>> has already taken the write seq lock, and then the higher priority
>> rt task A try to acquire read seq lock, it's doomed to lockup.
>>
>> rt task A with lower priority: call write
>> i_size_write                                        rt task B with higher priority: call sync, and preempt task A
>>    write_seqcount_begin(&inode->i_size_seqcount);    i_size_read
>>    inode->i_size = i_size;                             read_seqcount_begin<-- lockup here...
>>
>
> Ouch.
>
> And even if the preemping task is preemptible, it will spend an entire
> timeslice pointlessly spinning, which isn't very good.
>
>> So disable preempt when acquiring every i_size_seqcount *write* lock will
>> cure the problem.
>>
>> ...
>>
>> --- a/include/linux/fs.h
>> +++ b/include/linux/fs.h
>> @@ -758,9 +758,11 @@ static inline loff_t i_size_read(const struct inode *inode)
>>   static inline void i_size_write(struct inode *inode, loff_t i_size)
>>   {
>>   #if BITS_PER_LONG==32&&  defined(CONFIG_SMP)
>> +	preempt_disable();
>>   	write_seqcount_begin(&inode->i_size_seqcount);
>>   	inode->i_size = i_size;
>>   	write_seqcount_end(&inode->i_size_seqcount);
>> +	preempt_enable();
>>   #elif BITS_PER_LONG==32&&  defined(CONFIG_PREEMPT)
>>   	preempt_disable();
>>   	inode->i_size = i_size;
>
> afacit all write_seqcount_begin()/read_seqretry() sites are vulnerable
> to this problem.  Would it not be better to do the preempt_disable() in
> write_seqcount_begin()?

IMHO, write_seqcount_begin/write_seqcount_end are often wrapped by mutex,
this gives higher priority task a chance to sleep, and then lower priority task
get cpu to unlock, so avoid the problematic scenario this patch describing.

But in i_size_write case, I could only find disable preempt a good choice before
someone else has better idea :)

>
> Possible problems:
>
> - mm/filemap_xip.c does disk I/O under write_seqcount_begin().
>
> - dev_change_name() does GFP_KERNEL allocations under write_seqcount_begin()
>
> - I didn't review u64_stats_update_begin() callers.
>
> But I think calling schedule() under preempt_disable() is OK anyway?
>

-- 
浮沉随浪只记今朝笑

--fan