On Wed, Jun 15, 2016 at 12:17:47PM +0200, Jens Axboe wrote:

I guess you are right, the fs path jumps in via __blk_mq_alloc_request(), 
so not fs fast path. But it'd still be nice to properly fix this. Can it 
wait until the rq->pc mess is resolved?

I'll resend the series removing the memset in the blk_get_request path,
which I think is very useful for now as the requests currently returned
from blk_get_request / blk_mq_alloc_request are inherently dangerous.
req->cmd zeroing is really just a workaround broken devices in practice,
as spec complicant scsi device never should look at the cmd array
beyond the command length.

I'll try to expedite the pc_request split, but I'd really like to get
all the NVMe over Fabrics work in first, as we'd be in merge hell otherwise.