Re: [block] allow blk_flush_policy to return REQ_FSEQ_DATA independent of *FLUSH

[Tejun Heo <tj@kernel.org>]

Hello,

On Tue, Jan 08, 2013 at 10:04:23AM -0800, ajithb.kumar@gmail.com wrote:
> Hi,
> Could you please provide clarity on the following.
> ">   Hmmm... yes, this can become a correctness issue if (and only if)
> >   blk_queue_flush() is called to change q->flush_flags while requests
> >   are in-flight;"
>
> Could you please clarify as to why is it a correctness issue only if
> blk_queue_flush() is used to change flush_flags when requests are in
> flight ?  As I understand, XFS does set WRITE_FLUSH_FUA flag in
> _xfs_buf_ioapply() function irrespective of whether the underlying
> device supports flush capabilities or not which will flow into
> blk_insert_flush().  Is my reading of the code correct and is there
> a general correctness issue here which potentially results in XFS
> file system corruption in case of an abrupt shutdown independent of
> q->flush_flags getting changed while request is in flight.

My memory is kinda fuzzy at this point but if a queue doesn't support
flush, its flush_flags should be zero and
generic_make_request_checks() will clear REQ_FLUSH|REQ_FUA from
bio->bi_rw so we never hit blk_insert_flush() and the request will be
processed as a normal IO one; however, if REQ_FLUSH goes off after a
request passed generic_make_request_checks() but before
blk_flush_policy(), it'll become null op and its data payload won't
get written out to the underlying device, which is data corruption.

Thanks.

-- 
tejun