silent truncation for large file offsets

silent truncation for large file offsets

[Christoph Hellwig]

Re: silent truncation for large file offsets

[J. Bruce Fields]

On Mon, Feb 08, 2016 at 10:33:28AM -0800, Christoph Hellwig wrote:
> >From a Linux client to a Linux server (in fact the same system in this
> example), NFSv4.1, XFS file system:
> 
> root@vm:~/xfstests# truncate --size 9223372036854775807 /mnt/nfs1/test
> root@vm:~/xfstests# ls -l /mnt/nfs1/test
> -rw-r--r-- 1 root root 9223372036854775806 Feb  8 18:30 /mnt/nfs1/test
> root@vm:~/xfstests# ls -l /mnt/test/test
> -rw-r--r-- 1 root root 9223372036854775807 Feb  8 18:30 /mnt/test/test
> 
> so the file gets created with the correct size on the server, but
> the clients shows the size truncated.
> 
> This is extraced from xfstests generic/911 which tests clone
> functionality and fails because of this issue.

Took a quick look at wireshark, and GETATTR is returning the correct
(larger) size.

Also FSINFO (this is v3) returns 9223372036854775807 as the maximum file
size.

--b.

Re: silent truncation for large file offsets

[Trond Myklebust]

On Mon, Feb 8, 2016 at 2:30 PM, J. Bruce Fields <bfields@fieldses.org> wrote:
> On Mon, Feb 08, 2016 at 10:33:28AM -0800, Christoph Hellwig wrote:
>> >From a Linux client to a Linux server (in fact the same system in this
>> example), NFSv4.1, XFS file system:
>>
>> root@vm:~/xfstests# truncate --size 9223372036854775807 /mnt/nfs1/test
>> root@vm:~/xfstests# ls -l /mnt/nfs1/test
>> -rw-r--r-- 1 root root 9223372036854775806 Feb  8 18:30 /mnt/nfs1/test
>> root@vm:~/xfstests# ls -l /mnt/test/test
>> -rw-r--r-- 1 root root 9223372036854775807 Feb  8 18:30 /mnt/test/test
>>
>> so the file gets created with the correct size on the server, but
>> the clients shows the size truncated.
>>
>> This is extraced from xfstests generic/911 which tests clone
>> functionality and fails because of this issue.
>
> Took a quick look at wireshark, and GETATTR is returning the correct
> (larger) size.
>
> Also FSINFO (this is v3) returns 9223372036854775807 as the maximum file
> size.
>

I'll bet it's this:

static inline loff_t nfs_size_to_loff_t(__u64 size)
{
        if (size > (__u64) OFFSET_MAX - 1)
                return OFFSET_MAX - 1;
        return (loff_t) size;
}

Should be "return OFFSET_MAX", no?

Re: silent truncation for large file offsets

[J. Bruce Fields]

On Mon, Feb 08, 2016 at 02:45:54PM -0500, Trond Myklebust wrote:
> On Mon, Feb 8, 2016 at 2:30 PM, J. Bruce Fields <bfields@fieldses.org> wrote:
> > On Mon, Feb 08, 2016 at 10:33:28AM -0800, Christoph Hellwig wrote:
> >> >From a Linux client to a Linux server (in fact the same system in this
> >> example), NFSv4.1, XFS file system:
> >>
> >> root@vm:~/xfstests# truncate --size 9223372036854775807 /mnt/nfs1/test
> >> root@vm:~/xfstests# ls -l /mnt/nfs1/test
> >> -rw-r--r-- 1 root root 9223372036854775806 Feb  8 18:30 /mnt/nfs1/test
> >> root@vm:~/xfstests# ls -l /mnt/test/test
> >> -rw-r--r-- 1 root root 9223372036854775807 Feb  8 18:30 /mnt/test/test
> >>
> >> so the file gets created with the correct size on the server, but
> >> the clients shows the size truncated.
> >>
> >> This is extraced from xfstests generic/911 which tests clone
> >> functionality and fails because of this issue.
> >
> > Took a quick look at wireshark, and GETATTR is returning the correct
> > (larger) size.
> >
> > Also FSINFO (this is v3) returns 9223372036854775807 as the maximum file
> > size.
> >
> 
> I'll bet it's this:
> 
> static inline loff_t nfs_size_to_loff_t(__u64 size)
> {
>         if (size > (__u64) OFFSET_MAX - 1)
>                 return OFFSET_MAX - 1;
>         return (loff_t) size;
> }
> 
> Should be "return OFFSET_MAX", no?

I guess so

(That's confusing, though--in general, shouldn't the maximum file size
be one *more* than the maximum offset?  But looks like st_size is
off_t, so, I guess this is just how it is.)

--b.

Re: silent truncation for large file offsets

[Christoph Hellwig]

On Mon, Feb 08, 2016 at 02:45:54PM -0500, Trond Myklebust wrote:
> I'll bet it's this:
> 
> static inline loff_t nfs_size_to_loff_t(__u64 size)
> {
>         if (size > (__u64) OFFSET_MAX - 1)
>                 return OFFSET_MAX - 1;
>         return (loff_t) size;
> }
> 
> Should be "return OFFSET_MAX", no?

Yes.  That (or rather a slighty nicer version using min_t) fixes the
test case for me.

Re: silent truncation for large file offsets

[Trond Myklebust]

On Mon, Feb 8, 2016 at 2:54 PM, J. Bruce Fields <bfields@fieldses.org> wrote:
> On Mon, Feb 08, 2016 at 02:45:54PM -0500, Trond Myklebust wrote:
>> On Mon, Feb 8, 2016 at 2:30 PM, J. Bruce Fields <bfields@fieldses.org> wrote:
>> > On Mon, Feb 08, 2016 at 10:33:28AM -0800, Christoph Hellwig wrote:
>> >> >From a Linux client to a Linux server (in fact the same system in this
>> >> example), NFSv4.1, XFS file system:
>> >>
>> >> root@vm:~/xfstests# truncate --size 9223372036854775807 /mnt/nfs1/test
>> >> root@vm:~/xfstests# ls -l /mnt/nfs1/test
>> >> -rw-r--r-- 1 root root 9223372036854775806 Feb  8 18:30 /mnt/nfs1/test
>> >> root@vm:~/xfstests# ls -l /mnt/test/test
>> >> -rw-r--r-- 1 root root 9223372036854775807 Feb  8 18:30 /mnt/test/test
>> >>
>> >> so the file gets created with the correct size on the server, but
>> >> the clients shows the size truncated.
>> >>
>> >> This is extraced from xfstests generic/911 which tests clone
>> >> functionality and fails because of this issue.
>> >
>> > Took a quick look at wireshark, and GETATTR is returning the correct
>> > (larger) size.
>> >
>> > Also FSINFO (this is v3) returns 9223372036854775807 as the maximum file
>> > size.
>> >
>>
>> I'll bet it's this:
>>
>> static inline loff_t nfs_size_to_loff_t(__u64 size)
>> {
>>         if (size > (__u64) OFFSET_MAX - 1)
>>                 return OFFSET_MAX - 1;
>>         return (loff_t) size;
>> }
>>
>> Should be "return OFFSET_MAX", no?
>
> I guess so
>
> (That's confusing, though--in general, shouldn't the maximum file size
> be one *more* than the maximum offset?  But looks like st_size is
> off_t, so, I guess this is just how it is.)

If you ever manage to write to offset OFFSET_MAX, you will never know...