On Mon, 2025-08-18 at 15:05 -0400, Mike Snitzer wrote: > On Mon, Aug 18, 2025 at 10:45:47AM -0400, Jeff Layton wrote: > > On Fri, 2025-08-15 at 10:46 -0400, Mike Snitzer wrote: > > > If NFSD_IO_DIRECT is used, expand any misaligned READ to the next > > > DIO-aligned block (on either end of the READ). The expanded READ is > > > verified to have proper offset/len (logical_block_size) and > > > dma_alignment checking. > > > > > > Must allocate and use a bounce-buffer page (called 'start_extra_page') > > > if/when expanding the misaligned READ requires reading extra partial > > > page at the start of the READ so that its DIO-aligned. Otherwise that > > > extra page at the start will make its way back to the NFS client and > > > corruption will occur. As found, and then this fix of using an extra > > > page verified, using the 'dt' utility: > > > dt of=/mnt/share1/dt_a.test passes=1 bs=47008 count=2 \ > > > iotype=sequential pattern=iot onerr=abort oncerr=abort > > > see: https://github.com/RobinTMiller/dt.git > > > > > > Any misaligned READ that is less than 32K won't be expanded to be > > > DIO-aligned (this heuristic just avoids excess work, like allocating > > > start_extra_page, for smaller IO that can generally already perform > > > well using buffered IO). > > > > > > Suggested-by: Jeff Layton <jlayton@xxxxxxxxxx> > > > Suggested-by: Chuck Lever <chuck.lever@xxxxxxxxxx> > > > Signed-off-by: Mike Snitzer <snitzer@xxxxxxxxxx> > > > --- > > > fs/nfsd/vfs.c | 200 +++++++++++++++++++++++++++++++++++-- > > > include/linux/sunrpc/svc.h | 5 +- > > > 2 files changed, 194 insertions(+), 11 deletions(-) > > > > > > diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c > > > index c340708fbab4d..64732dc8985d6 100644 > > > --- a/fs/nfsd/vfs.c > > > +++ b/fs/nfsd/vfs.c > > > @@ -19,6 +19,7 @@ > > > #include <linux/splice.h> > > > #include <linux/falloc.h> > > > #include <linux/fcntl.h> > > > +#include <linux/math.h> > > > #include <linux/namei.h> > > > #include <linux/delay.h> > > > #include <linux/fsnotify.h> > > > @@ -1073,6 +1074,153 @@ __be32 nfsd_splice_read(struct svc_rqst *rqstp, struct svc_fh *fhp, > > > return nfsd_finish_read(rqstp, fhp, file, offset, count, eof, host_err); > > > } > > > > > > +struct nfsd_read_dio { > > > + loff_t start; > > > + loff_t end; > > > + unsigned long start_extra; > > > + unsigned long end_extra; > > > + struct page *start_extra_page; > > > +}; > > > + > > > +static void init_nfsd_read_dio(struct nfsd_read_dio *read_dio) > > > +{ > > > + memset(read_dio, 0, sizeof(*read_dio)); > > > + read_dio->start_extra_page = NULL; > > > +} > > > + > > > +#define NFSD_READ_DIO_MIN_KB (32 << 10) > > > + > > > +static bool nfsd_analyze_read_dio(struct svc_rqst *rqstp, struct svc_fh *fhp, > > > + struct nfsd_file *nf, loff_t offset, > > > + unsigned long len, unsigned int base, > > > + struct nfsd_read_dio *read_dio) > > > +{ > > > + const u32 dio_blocksize = nf->nf_dio_read_offset_align; > > > + loff_t middle_end, orig_end = offset + len; > > > + > > > + if (WARN_ONCE(!nf->nf_dio_mem_align || !nf->nf_dio_read_offset_align, > > > + "%s: underlying filesystem has not provided DIO alignment info\n", > > > + __func__)) > > > + return false; > > > + if (WARN_ONCE(dio_blocksize > PAGE_SIZE, > > > + "%s: underlying storage's dio_blocksize=%u > PAGE_SIZE=%lu\n", > > > + __func__, dio_blocksize, PAGE_SIZE)) > > > + return false; > > > + > > > + /* Return early if IO is irreparably misaligned (len < PAGE_SIZE, > > > + * or base not aligned). > > > + * Ondisk alignment is implied by the following code that expands > > > + * misaligned IO to have a DIO-aligned offset and len. > > > + */ > > > + if (unlikely(len < dio_blocksize) || ((base & (nf->nf_dio_mem_align-1)) != 0)) > > > + return false; > > > > The small len check makes sense, but "base" at this point is the offset > > into the first page. Here you're bailing out early if that's not > > aligned. Isn't that contrary to what this patch is supposed to do > > (which is expand the range so that the I/O is aligned)? > > No matter whether we're expanding the read or not (that's a means to > make the area read from disk DIO-aligned): the memory alignment is > what it is -- so it isn't something that we can change (not without an > extra copy). But thankfully with RDMA the memory for the READ payload > is generally always aligned. > > Chcuk did say in reply to an earlier version of this patchset > (paraphrasing, rather not go splunking in the linux-nfs archive to > find it): a future improvement would be to make sure the READ > payload's memory is always aligned. > Oh right, I got confused between the mem and block alignment here. In light of that, you can add Reviewed-by: Jeff Layton <jlayton@xxxxxxxxxx>
