Re: [PATCH v8 09/12] NFSD: Handle both offset and memory alignment for direct I/O

[Mike Snitzer <snitzer@kernel.org>]

On Fri, Oct 31, 2025 at 09:21:27AM -0400, Chuck Lever wrote:
> On 10/31/25 9:19 AM, Christoph Hellwig wrote:
> > On Mon, Oct 27, 2025 at 11:46:27AM -0400, Chuck Lever wrote:
> >> From: Chuck Lever <chuck.lever@oracle.com>
> >>
> >> Currently, nfsd_is_write_dio_possible() only considers file offset
> >> alignment (nf_dio_offset_align) when splitting an NFS WRITE request
> >> into segments. This leaves accounting for memory buffer alignment
> >> (nf_dio_mem_align) until nfsd_setup_write_dio_iters(). If this
> >> second check fails, the code falls back to cached I/O entirely,
> >> wasting the opportunity to use direct I/O for the bulk of the
> >> request.
> >>
> >> Enhance the logic to find a beginning segment size that satisfies
> >> both alignment constraints simultaneously. The search algorithm
> >> starts with the file offset alignment requirement and steps through
> >> multiples of offset_align, checking memory alignment at each step.
> >> The search is bounded by lcm(offset_align, mem_align) to ensure that
> >> it always terminates.
> > 
> > How likely is that going to happen?  After the first bvec the
> > alignment constrains won't change, so how are we going to succeed
> > then?
> > 
> 
> I was hoping that this algorithm would improve the likelihood of
> finding a middle segment alignment for NFS WRITE on TCP. I'm not
> entirely sure it has been effective.
> 
> Given the complexity, I'm wondering if I want to keep this one.

I don't think its worth the complexity.  And I don't see how it'd help
make TCP's potentially unaligned buffer, for the NFS WRITE's payload,
able to work within the required ondisk logical offset granularity.

The mlperf_npz_tool.py I provided does a really good job of exposing
this worse case scenario, especially when the NFS client is forced to
use DIO (but even if just using buffered IO it has some issue just
that the NFS client's use of the page cache causes the IO that's sent
over the wire to NFSD to be much larger, so its head and tail of that
large IO are misaligned).  But when the NFS client does use O_DIRECT
each and every 1MB IO is offset in memory such that every single
logical_block_size isn't aligned relative to associated page -- each
spans multiple pages.