Re: [PATCH v20 02/12] Add infrastructure for copy offload in block and request layer.

[Bart Van Assche]

On 5/29/24 12:48 AM, Damien Le Moal wrote:
> On 5/29/24 15:17, Nitesh Shetty wrote:
> > On 24/05/24 01:33PM, Bart Van Assche wrote:
> > > On 5/20/24 03:20, Nitesh Shetty wrote:
> > > > We add two new opcode REQ_OP_COPY_DST, REQ_OP_COPY_SRC.
> > > > Since copy is a composite operation involving src and dst sectors/lba,
> > > > each needs to be represented by a separate bio to make it compatible
> > > > with device mapper.
> > > > We expect caller to take a plug and send bio with destination information,
> > > > followed by bio with source information.
> > > > Once the dst bio arrives we form a request and wait for source
> > > > bio. Upon arrival of source bio we merge these two bio's and send
> > > > corresponding request down to device driver.
> > > > Merging non copy offload bio is avoided by checking for copy specific
> > > > opcodes in merge function.
> > >
> > > In this patch I don't see any changes for blk_attempt_bio_merge(). Does
> > > this mean that combining REQ_OP_COPY_DST and REQ_OP_COPY_SRC will never
> > > happen if the QUEUE_FLAG_NOMERGES request queue flag has been set?
> > Yes, in this case copy won't work, as both src and dst bio reach driver
> > as part of separate requests.
> > We will add this as part of documentation.
>
> So that means that 2 major SAS HBAs which set this flag (megaraid and mpt3sas)
> will not get support for copy offload ? Not ideal, by far.

QUEUE_FLAG_NOMERGES can also be set through sysfs (see also
queue_nomerges_store()). This is one of the reasons why using the merge
infrastructure for combining REQ_OP_COPY_DST and REQ_OP_COPY_SRC is
unacceptable.

Thanks,

Bart.