2019-09-30 20:41:44

by Trond Myklebust

[permalink] [raw]
Subject: Re: [PATCH] NFS: Fix O_DIRECT read problem when another write is going on

Hi Su,

On Mon, 2019-09-30 at 17:11 +0800, Su Yanjun wrote:
> In xfstests generic/465 tests failed. Because O_DIRECT r/w use
> async rpc calls, when r/w rpc calls are running concurrently we
> may read partial data which is wrong.
>
> For example as follows.
> user buffer
> /--------\
> > |XXXX|
> rpc0 rpc1
>
> When rpc0 runs it encounters eof so return 0, then another writes
> something. When rpc1 runs it returns some data. The total data
> buffer contains wrong data.
>
> In this patch we check eof mark for each direct request. If
> encounters
> eof then set eof mark in the request, when we meet it again report
> -EAGAIN error. In nfs_direct_complete we convert -EAGAIN as if read
> nothing. When the reader issue another read it will read ok.
>
> Signed-off-by: Su Yanjun <[email protected]>
> ---
> fs/nfs/direct.c | 14 +++++++++++++-
> 1 file changed, 13 insertions(+), 1 deletion(-)
>
> diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
> index 222d711..7f737a3 100644
> --- a/fs/nfs/direct.c
> +++ b/fs/nfs/direct.c
> @@ -93,6 +93,7 @@ struct nfs_direct_req {
> bytes_left, /* bytes left to be
> sent */
> error; /* any reported error
> */
> struct completion completion; /* wait for i/o completion */
> + int eof; /* eof mark in the
> req */
>
> /* commit state */
> struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo
> */
> @@ -380,6 +381,12 @@ static void nfs_direct_complete(struct
> nfs_direct_req *dreq)
> {
> struct inode *inode = dreq->inode;
>
> + /* read partial data just as read nothing */
> + if (dreq->error == -EAGAIN) {
> + dreq->count = 0;
> + dreq->error = 0;
> + }
> +
> inode_dio_end(inode);
>
> if (dreq->iocb) {
> @@ -413,8 +420,13 @@ static void nfs_direct_read_completion(struct
> nfs_pgio_header *hdr)
> if (hdr->good_bytes != 0)
> nfs_direct_good_bytes(dreq, hdr);
>
> - if (test_bit(NFS_IOHDR_EOF, &hdr->flags))
> + if (dreq->eof)
> + dreq->error = -EAGAIN;
> +
> + if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
> dreq->error = 0;
> + dreq->eof = 1;
> + }
>
> spin_unlock(&dreq->lock);
>

Thanks for looking into this issue. I agree with your analysis of what
is going wrong in generic/465.

However, I think the problem is greater than just EOF. I think we also
need to look at the generic error handling, and ensure that it handles
a truncated RPC call in the middle of a series of calls correctly.

Please see the two patches I sent you just now and check if they fix
the problem for you.

--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[email protected]



2019-10-07 02:18:14

by Su, Yanjun

[permalink] [raw]
Subject: Re: [PATCH] NFS: Fix O_DIRECT read problem when another write is going on


在 2019/10/1 2:06, Trond Myklebust 写道:
> Hi Su,
>
> On Mon, 2019-09-30 at 17:11 +0800, Su Yanjun wrote:
>> In xfstests generic/465 tests failed. Because O_DIRECT r/w use
>> async rpc calls, when r/w rpc calls are running concurrently we
>> may read partial data which is wrong.
>>
>> For example as follows.
>> user buffer
>> /--------\
>>> |XXXX|
>> rpc0 rpc1
>>
>> When rpc0 runs it encounters eof so return 0, then another writes
>> something. When rpc1 runs it returns some data. The total data
>> buffer contains wrong data.
>>
>> In this patch we check eof mark for each direct request. If
>> encounters
>> eof then set eof mark in the request, when we meet it again report
>> -EAGAIN error. In nfs_direct_complete we convert -EAGAIN as if read
>> nothing. When the reader issue another read it will read ok.
>>
>> Signed-off-by: Su Yanjun <[email protected]>
>> ---
>> fs/nfs/direct.c | 14 +++++++++++++-
>> 1 file changed, 13 insertions(+), 1 deletion(-)
>>
>> diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
>> index 222d711..7f737a3 100644
>> --- a/fs/nfs/direct.c
>> +++ b/fs/nfs/direct.c
>> @@ -93,6 +93,7 @@ struct nfs_direct_req {
>> bytes_left, /* bytes left to be
>> sent */
>> error; /* any reported error
>> */
>> struct completion completion; /* wait for i/o completion */
>> + int eof; /* eof mark in the
>> req */
>>
>> /* commit state */
>> struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo
>> */
>> @@ -380,6 +381,12 @@ static void nfs_direct_complete(struct
>> nfs_direct_req *dreq)
>> {
>> struct inode *inode = dreq->inode;
>>
>> + /* read partial data just as read nothing */
>> + if (dreq->error == -EAGAIN) {
>> + dreq->count = 0;
>> + dreq->error = 0;
>> + }
>> +
>> inode_dio_end(inode);
>>
>> if (dreq->iocb) {
>> @@ -413,8 +420,13 @@ static void nfs_direct_read_completion(struct
>> nfs_pgio_header *hdr)
>> if (hdr->good_bytes != 0)
>> nfs_direct_good_bytes(dreq, hdr);
>>
>> - if (test_bit(NFS_IOHDR_EOF, &hdr->flags))
>> + if (dreq->eof)
>> + dreq->error = -EAGAIN;
>> +
>> + if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
>> dreq->error = 0;
>> + dreq->eof = 1;
>> + }
>>
>> spin_unlock(&dreq->lock);
>>
> Thanks for looking into this issue. I agree with your analysis of what
> is going wrong in generic/465.
>
> However, I think the problem is greater than just EOF. I think we also
> need to look at the generic error handling, and ensure that it handles
> a truncated RPC call in the middle of a series of calls correctly.
>
> Please see the two patches I sent you just now and check if they fix
> the problem for you.

The patchset you sent works for generic/465.

Thanks a lot