Re: [PATCH] fs: Disable preempt when acquire i_size_seqcount writelock

[Fan Du]

On 2013å01æ11æ 06:38, Andrew Morton wrote:
> On Wed, 9 Jan 2013 11:34:19 +0800
> Fan Du<fan.du@xxxxxxxxxxxxx>  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?

--
ææéæåèäæç

--fan
--
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/