2015-09-16 21:21:28

by J. Bruce Fields

Subject: [PATCH] nfs: fix v4.2 SEEK on files over 2 gigs

From: "J. Bruce Fields" <[email protected]>

We're incorrectly assigning a loff_t return to an int. If SEEK_HOLE or
SEEK_DATA returns an offset over 2^31 then the application will see a
weird lseek() result (usually -EIO).

Cc: [email protected]
Fixes: bdcc2cd14e4e "NFSv4.2: handle NFS-specific llseek errors"
Signed-off-by: J. Bruce Fields <[email protected]>
---
fs/nfs/nfs42proc.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/fs/nfs/nfs42proc.c b/fs/nfs/nfs42proc.c
index d731bbf974aa..0f020e4d8421 100644
--- a/fs/nfs/nfs42proc.c
+++ b/fs/nfs/nfs42proc.c
@@ -175,10 +175,12 @@ loff_t nfs42_proc_llseek(struct file *filep, loff_t offset, int whence)
{
struct nfs_server *server = NFS_SERVER(file_inode(filep));
struct nfs4_exception exception = { };
- int err;
+ loff_t err;

do {
err = _nfs42_proc_llseek(filep, offset, whence);
+ if (err >= 0)
+ break;
if (err == -ENOTSUPP)
return -EOPNOTSUPP;
err = nfs4_handle_exception(server, err, &exception);
--
2.4.3

2015-09-17 13:04:09

by Anna Schumaker

Subject: Re: [PATCH] nfs: fix v4.2 SEEK on files over 2 gigs

Hey Bruce,

I actually worked on a version of this patch on my own yesterday, too. Looks like you beat me to submitting it! :)

On 09/16/2015 05:21 PM, J. Bruce Fields wrote:
> From: "J. Bruce Fields" <[email protected]>
>
> We're incorrectly assigning a loff_t return to an int. If SEEK_HOLE or
> SEEK_DATA returns an offset over 2^31 then the application will see a
> weird lseek() result (usually -EIO).

I saw roughly the same thing with xfstests generic/285.

>
> Cc: [email protected]
> Fixes: bdcc2cd14e4e "NFSv4.2: handle NFS-specific llseek errors"
> Signed-off-by: J. Bruce Fields <[email protected]>

Reviewed-by: Anna Schumaker <[email protected]>

> ---
> fs/nfs/nfs42proc.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/fs/nfs/nfs42proc.c b/fs/nfs/nfs42proc.c
> index d731bbf974aa..0f020e4d8421 100644
> --- a/fs/nfs/nfs42proc.c
> +++ b/fs/nfs/nfs42proc.c
> @@ -175,10 +175,12 @@ loff_t nfs42_proc_llseek(struct file *filep, loff_t offset, int whence)
> {
> struct nfs_server *server = NFS_SERVER(file_inode(filep));
> struct nfs4_exception exception = { };
> - int err;
> + loff_t err;
>
> do {
> err = _nfs42_proc_llseek(filep, offset, whence);
> + if (err >= 0)
> + break;
> if (err == -ENOTSUPP)
> return -EOPNOTSUPP;
> err = nfs4_handle_exception(server, err, &exception);
>

2015-09-17 13:25:53

by J. Bruce Fields

Subject: Re: [PATCH] nfs: fix v4.2 SEEK on files over 2 gigs

On Thu, Sep 17, 2015 at 09:04:00AM -0400, Anna Schumaker wrote:
> Hey Bruce,
>
> I actually worked on a version of this patch on my own yesterday, too. Looks like you beat me to submitting it! :)

Oh, OK, well feel free to credit it however you'd like.

> On 09/16/2015 05:21 PM, J. Bruce Fields wrote:
> > From: "J. Bruce Fields" <[email protected]>
> >
> > We're incorrectly assigning a loff_t return to an int. If SEEK_HOLE or
> > SEEK_DATA returns an offset over 2^31 then the application will see a
> > weird lseek() result (usually -EIO).
>
> I saw roughly the same thing with xfstests generic/285.

I maybe should have included my reproducer in the changelog, which was:

git clone git://git.infradead.org/users/dedekind/bmap-tools.git
cd bmap-tools

mount overs=4.2 localhost:/exports /mnt/
dd if=/dev/zero of=/mnt/test1.image seek=2097152 bs=1K count=1

./bmaptool create /mnt/test1.image

In the good case it outputs some xml, in the bad case it aborts with an
IO error.

--b.

> > Cc: [email protected]
> > Fixes: bdcc2cd14e4e "NFSv4.2: handle NFS-specific llseek errors"
> > Signed-off-by: J. Bruce Fields <[email protected]>
>
> Reviewed-by: Anna Schumaker <[email protected]>
>
> > ---
> > fs/nfs/nfs42proc.c | 4 +++-
> > 1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/fs/nfs/nfs42proc.c b/fs/nfs/nfs42proc.c
> > index d731bbf974aa..0f020e4d8421 100644
> > --- a/fs/nfs/nfs42proc.c
> > +++ b/fs/nfs/nfs42proc.c
> > @@ -175,10 +175,12 @@ loff_t nfs42_proc_llseek(struct file *filep, loff_t offset, int whence)
> > {
> > struct nfs_server *server = NFS_SERVER(file_inode(filep));
> > struct nfs4_exception exception = { };
> > - int err;
> > + loff_t err;
> >
> > do {
> > err = _nfs42_proc_llseek(filep, offset, whence);
> > + if (err >= 0)
> > + break;
> > if (err == -ENOTSUPP)
> > return -EOPNOTSUPP;
> > err = nfs4_handle_exception(server, err, &exception);
> >

2015-09-17 13:44:47

by Anna Schumaker

Subject: Re: [PATCH] nfs: fix v4.2 SEEK on files over 2 gigs

On 09/17/2015 09:25 AM, J. Bruce Fields wrote:
> On Thu, Sep 17, 2015 at 09:04:00AM -0400, Anna Schumaker wrote:
>> Hey Bruce,
>>
>> I actually worked on a version of this patch on my own yesterday, too. Looks like you beat me to submitting it! :)
>
> Oh, OK, well feel free to credit it however you'd like.
>
>> On 09/16/2015 05:21 PM, J. Bruce Fields wrote:
>>> From: "J. Bruce Fields" <[email protected]>
>>>
>>> We're incorrectly assigning a loff_t return to an int. If SEEK_HOLE or
>>> SEEK_DATA returns an offset over 2^31 then the application will see a
>>> weird lseek() result (usually -EIO).
>>
>> I saw roughly the same thing with xfstests generic/285.
>
> I maybe should have included my reproducer in the changelog, which was:
>
> git clone git://git.infradead.org/users/dedekind/bmap-tools.git
> cd bmap-tools
>
> mount overs=4.2 localhost:/exports /mnt/
> dd if=/dev/zero of=/mnt/test1.image seek=2097152 bs=1K count=1
>
> ./bmaptool create /mnt/test1.image
>
> In the good case it outputs some xml, in the bad case it aborts with an
> IO error.

That's roughly what I see, too:

$ cat results/generic/285.out.bad
...
10. Test a huge file for offset overflow
10.01 SEEK_HOLE expected 2097152 or 8589934592, got 2097152. succ
10.02 SEEK_HOLE expected 2097152 or 8589934592, got 2097152. succ
10.03 SEEK_DATA expected 0 or 0, got 0. succ
10.04 SEEK_DATA expected 1 or 1, got 1. succ
10.05 SEEK_HOLE expected 8589934592 or 8589934592, got 0. FAIL
10.06 SEEK_DATA expected 8587837440 or 8587837440, got -1. FAIL
10.07 SEEK_DATA expected 8587837441 or 8587837441, got -1. FAIL
10.08 SEEK_DATA expected 8587837440 or 8587837440, got -1. FAIL

>
> --b.
>
>>> Cc: [email protected]
>>> Fixes: bdcc2cd14e4e "NFSv4.2: handle NFS-specific llseek errors"
>>> Signed-off-by: J. Bruce Fields <[email protected]>
>>
>> Reviewed-by: Anna Schumaker <[email protected]>
>>
>>> ---
>>> fs/nfs/nfs42proc.c | 4 +++-
>>> 1 file changed, 3 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/fs/nfs/nfs42proc.c b/fs/nfs/nfs42proc.c
>>> index d731bbf974aa..0f020e4d8421 100644
>>> --- a/fs/nfs/nfs42proc.c
>>> +++ b/fs/nfs/nfs42proc.c
>>> @@ -175,10 +175,12 @@ loff_t nfs42_proc_llseek(struct file *filep, loff_t offset, int whence)
>>> {
>>> struct nfs_server *server = NFS_SERVER(file_inode(filep));
>>> struct nfs4_exception exception = { };
>>> - int err;
>>> + loff_t err;
>>>
>>> do {
>>> err = _nfs42_proc_llseek(filep, offset, whence);
>>> + if (err >= 0)
>>> + break;
>>> if (err == -ENOTSUPP)
>>> return -EOPNOTSUPP;
>>> err = nfs4_handle_exception(server, err, &exception);
>>>