2015-04-08 12:24:28

by Peng Tao

Subject: [PATCH] nfs: fix DIO good bytes calculation

When adding completed bytes, we should just use hdr->good_bytes.
Otherwise if we send multiple requests for one direct IO (due to
rsize/wsize limitation), we'll count completed bytes incorrectly.

This fixes 5fadeb47("nfs: count DIO good bytes correctly with mirroring").

Reported-by: Jean Spector <[email protected]>
Cc: <[email protected]> # v3.19+
Signed-off-by: Peng Tao <[email protected]>
---
fs/nfs/direct.c | 6 ++----
1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
index e907c8c..645109e 100644
--- a/fs/nfs/direct.c
+++ b/fs/nfs/direct.c
@@ -132,10 +132,8 @@ nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
WARN_ON_ONCE(hdr->pgio_mirror_idx >= dreq->mirror_count);

count = dreq->mirrors[hdr->pgio_mirror_idx].count;
- if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
- count = hdr->io_start + hdr->good_bytes - dreq->io_start;
- dreq->mirrors[hdr->pgio_mirror_idx].count = count;
- }
+ if (count + dreq->io_start < hdr->io_start + hdr->good_bytes)
+ dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;

/* update the dreq->count by finding the minimum agreed count from all
* mirrors */
--
1.9.1

2015-04-09 09:42:22

by Peng Tao

Subject: Re: [PATCH] nfs: fix DIO good bytes calculation

On Wed, Apr 8, 2015 at 8:24 PM, Peng Tao <[email protected]> wrote:
> When adding completed bytes, we should just use hdr->good_bytes.
> Otherwise if we send multiple requests for one direct IO (due to
> rsize/wsize limitation), we'll count completed bytes incorrectly.
>
> This fixes 5fadeb47("nfs: count DIO good bytes correctly with mirroring").
>
> Reported-by: Jean Spector <[email protected]>
> Cc: <[email protected]> # v3.19+
> Signed-off-by: Peng Tao <[email protected]>
> ---
> fs/nfs/direct.c | 6 ++----
> 1 file changed, 2 insertions(+), 4 deletions(-)
>
> diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
> index e907c8c..645109e 100644
> --- a/fs/nfs/direct.c
> +++ b/fs/nfs/direct.c
> @@ -132,10 +132,8 @@ nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
> WARN_ON_ONCE(hdr->pgio_mirror_idx >= dreq->mirror_count);
>
> count = dreq->mirrors[hdr->pgio_mirror_idx].count;
> - if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
> - count = hdr->io_start + hdr->good_bytes - dreq->io_start;
> - dreq->mirrors[hdr->pgio_mirror_idx].count = count;
> - }
> + if (count + dreq->io_start < hdr->io_start + hdr->good_bytes)
> + dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
oh, speaking too soon. This would actually fail concurrent DIO writers
if split requests complete randomly. Please don't merge it! I'll send
a updated version. sorry for the noise!

Thanks,
Tao