fsync bug

fsync bug

[Eric Blake]

fsync() is required to work on read-only fds (in theory, you can sync 
the atime metadata, which is a write operation triggered by a read-only 
fd).  But cygwin rejects this program, which works on Linux:

$ cat foo.c
#include <errno.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <fcntl.h>
int main (void)
{
   int fd = open("file", O_CREAT|O_EXCL|O_WRONLY, 0600);
   if (fd < 0)
     return 1;
   if (close(fd))
     return 2;
   fd = open("file", O_RDONLY);
   if (fd < 0)
     return 3;
   if (fsync(fd))
     return 4;
   if (unlink("file"))
     return 5;
   puts("success");
   return 0;
}
$ rm -f file && ./foo; echo $?
4

-- 
Eric Blake   eblake@redhat.com    +1-801-349-2682
Libvirt virtualization library http://libvirt.org

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple

Re: fsync bug

[Christopher Faylor]

On Wed, Sep 21, 2011 at 02:43:21PM -0600, Eric Blake wrote:
>fsync() is required to work on read-only fds (in theory, you can sync 
>the atime metadata, which is a write operation triggered by a read-only 
>fd).  But cygwin rejects this program, which works on Linux:
>
>$ cat foo.c
>#include <errno.h>
>#include <unistd.h>
>#include <stdio.h>
>#include <stdlib.h>
>#include <assert.h>
>#include <fcntl.h>
>int main (void)
>{
>   int fd = open("file", O_CREAT|O_EXCL|O_WRONLY, 0600);
>   if (fd < 0)
>     return 1;
>   if (close(fd))
>     return 2;
>   fd = open("file", O_RDONLY);
>   if (fd < 0)
>     return 3;
>   if (fsync(fd))
>     return 4;
>   if (unlink("file"))
>     return 5;
>   puts("success");
>   return 0;
>}
>$ rm -f file && ./foo; echo $?
>4

http://msdn.microsoft.com/en-us/library/aa364439%28VS.85%29.aspx

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple

Re: fsync bug

[Eric Blake]

On 09/21/2011 02:54 PM, Christopher Faylor wrote:
> On Wed, Sep 21, 2011 at 02:43:21PM -0600, Eric Blake wrote:
>> fsync() is required to work on read-only fds (in theory, you can sync
>> the atime metadata, which is a write operation triggered by a read-only
>> fd).  But cygwin rejects this program, which works on Linux:
>>
>> $ cat foo.c
>> #include<errno.h>
>> #include<unistd.h>
>> #include<stdio.h>
>> #include<stdlib.h>
>> #include<assert.h>
>> #include<fcntl.h>
>> int main (void)
>> {
>>    int fd = open("file", O_CREAT|O_EXCL|O_WRONLY, 0600);
>>    if (fd<  0)
>>      return 1;
>>    if (close(fd))
>>      return 2;
>>    fd = open("file", O_RDONLY);
>>    if (fd<  0)
>>      return 3;
>>    if (fsync(fd))
>>      return 4;
>>    if (unlink("file"))
>>      return 5;
>>    puts("success");
>>    return 0;
>> }
>> $ rm -f file&&  ./foo; echo $?
>> 4
>
> http://msdn.microsoft.com/en-us/library/aa364439%28VS.85%29.aspx

So what's wrong with making fsync() a no-op for read-only fds, on the 
grounds that windows doesn't give us a way to flush that data, so the 
best emulation we can do is by ignoring the request rather than 
propagating a spurious failure back to the user?

Also, FlushFileBuffers sounds more like fdatasync(); I'm wondering if 
fsync() should also be flushing the containing directory of any given 
file, in case file metadata (such as timestamps or changes in size via 
truncate) still need to be flushed to disk by flushing the directory.

-- 
Eric Blake   eblake@redhat.com    +1-801-349-2682
Libvirt virtualization library http://libvirt.org

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple

Re: fsync bug

[Christopher Faylor]

On Wed, Sep 21, 2011 at 03:07:21PM -0600, Eric Blake wrote:
>On 09/21/2011 02:54 PM, Christopher Faylor wrote:
>> On Wed, Sep 21, 2011 at 02:43:21PM -0600, Eric Blake wrote:
>>> fsync() is required to work on read-only fds (in theory, you can sync
>>> the atime metadata, which is a write operation triggered by a read-only
>>> fd).  But cygwin rejects this program, which works on Linux:
>>>
>>> $ cat foo.c
>>> #include<errno.h>
>>> #include<unistd.h>
>>> #include<stdio.h>
>>> #include<stdlib.h>
>>> #include<assert.h>
>>> #include<fcntl.h>
>>> int main (void)
>>> {
>>>    int fd = open("file", O_CREAT|O_EXCL|O_WRONLY, 0600);
>>>    if (fd<  0)
>>>      return 1;
>>>    if (close(fd))
>>>      return 2;
>>>    fd = open("file", O_RDONLY);
>>>    if (fd<  0)
>>>      return 3;
>>>    if (fsync(fd))
>>>      return 4;
>>>    if (unlink("file"))
>>>      return 5;
>>>    puts("success");
>>>    return 0;
>>> }
>>> $ rm -f file&&  ./foo; echo $?
>>> 4
>>
>> http://msdn.microsoft.com/en-us/library/aa364439%28VS.85%29.aspx
>
>So what's wrong with making fsync() a no-op for read-only fds, on the 
>grounds that windows doesn't give us a way to flush that data, so the 
>best emulation we can do is by ignoring the request rather than 
>propagating a spurious failure back to the user?

It's hard to see how lying that fsync worked would satisfy the
requirements to write metadata.  But, then, it isn't clear that this is
even an issue on Windows.

But, in any event, you know where the source code is.

cgf

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple