RE: [PATCH] aio: fix kernel bug when page is temporally busy

["Ananiev, Leonid I" <leonid.i.ananiev@intel.com>]

> It returns -EIOCBRETRY without guaranteeing that kick_iocb() will be  
> called.  This can lead to operations hanging

If EIOCBRETRY then generic_file_aio_write() will be recalled for the
same iocb.

> It overwrites -EIOCBQUEUED, leading to an aio_complete() while a  
> retry is happening.

EIOCBQUEUED or EIOCBRETRY does not lead to aio_complete() call:
        if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
                aio_complete(iocb, ret, 0);

> This can lead to reference count confusion.
But just reference count confusion was deleted by patch. Isn't it?

Leonid

-----Original Message-----
From: Zach Brown [mailto:zach.brown@oracle.com] 
Sent: Thursday, February 15, 2007 9:25 PM
To: Ananiev, Leonid I
Cc: Ken Chen; suparna@in.ibm.com; Andrew Morton;
linux-kernel@vger.kernel.org; linux-aio; Chris Mason
Subject: Re: [PATCH] aio: fix kernel bug when page is temporally busy

> If invalidate_inode_pages2_range() will return EIOCBRETRY as the patch
> "aio: fix kernel bug when page is temporally busy"

Sorry Leonid, this patch is not safe.

It returns -EIOCBRETRY without guaranteeing that kick_iocb() will be  
called.  This can lead to operations hanging, both AIO and calls that  
come through do_sync_{read,write}.

It overwrites -EIOCBQUEUED, leading to an aio_complete() while a  
retry is happening.  This can lead to reference count confusion.   
Double-frees, referencing freed memory, that kind of thing.  This  
isn't a new problem.  The current code that overwrites with -EIO has  
this problem.  But moving to -EIOCBRETRY does introduce new behaviour  
of aio_complete() and the retry path racing.

I'll have a candidate patch to address the problem of EIO being  
raised on the way back up from a path which has returned -EIOCBQUEUED.

- z