On Tue, Jun 14 2016 at  9:39pm -0400,
Martin K. Petersen [off-list ref] wrote:

quoted

quoted

quoted

quoted

quoted

"Hannes" == Hannes Reinecke [off-list ref] writes:

Hannes> Well, the primary issue is that 'blk_cloned_rq_check_limits()'
Hannes> doesn't check for BLOCK_PC,

Yes it does. It calls blk_rq_get_max_sectors() which has an explicit
check for this:

static inline unsigned int blk_rq_get_max_sectors(struct request *rq)
{
        struct request_queue *q = rq->q;

        if (unlikely(rq->cmd_type != REQ_TYPE_FS))
                        return q->limits.max_hw_sectors;
[...]

Hannes> The max_segments count, OTOH, _might_ change during failover
Hannes> (different hardware has different max_segments setting, and this
Hannes> is being changed during sg mapping), so there is some value to
Hannes> be had from testing it here.

Oh, this happens during failover? Are you sure it's not because DM is
temporarily resetting the queue limits? max_sectors is going to be a
single page in that case. I just discussed a backport regression in this
department with Mike at LSF/MM. But that was for an older kernel.

Not aware of any limits reset issue now...

Accidentally resetting the limits during table swaps has happened a
couple of times over the years. We trip it instantly with the database
in failover testing.

But feel free to throw your DB in failover tests (w/ dm-mpath) at a
recent kernel ;)