Re: [netcdfgroup] a possible problem in libsrc/posixio.c

• To: Cedric Roux <cedric.roux@xxxxxxxxxx>
• Subject: Re: [netcdfgroup] a possible problem in libsrc/posixio.c
• From: Russ Rew <russ@xxxxxxxxxxxxxxxx>
• Date: Wed, 02 Dec 2009 11:11:22 -0700

Hi Cédric,

> there is a little problem in libsrc/posixio.c
> (I think).
> 
> In px_pgout, you do:
> -----
>         if(write(nciop->fd, vp, extent) != (ssize_t) extent)
>         {
>                 return errno;
>         }
> -----
> Problem is 'write' may write less bytes than you request.
> In which case errno will be ENOERR, so the caller of px_pgout
> thinks everything went fine, 'extent' bytes were written
> to the file.
> On the next call to px_pgout, the following assert will fail.
> -----
>         assert(*posp == OFF_NONE || *posp == lseek(nciop->fd, 0,
> SEEK_CUR));
> -----
> because the previous call didn't update posp. (Or other functions
> where a similar assert is done will fail.)
> 
> You should loop the write, something like
> (adapt to match your taste):
> -----
>   count = 0;
>   while ((n = write(XXX)) != -1) {
>     count += n;
>     if (count == extent) break;
>   }
>   if (n == -1) return errno;
> -----

Thanks for diagnosing the problem and suggesting a fix!  You are
absolutely right about this, and we have already tested the fix, which
will be in the upcoming 4.1 release.  This is an important bug fix that
explains several previously reported assertion violations when writing
to a disk that was nearly full.

--Russ

• References:
  • [netcdfgroup] a possible problem in libsrc/posixio.c
    • From: Cedric Roux