Re: [PATCH] fsync.2: ERRORS: add EIO and ENOSPC

From: Michael Kerrisk (man-pages)
Date: Thu Sep 17 2020 - 03:02:32 EST

On 9/17/20 1:25 AM, NeilBrown wrote:
> On Thu, Sep 10 2020, Jeff Layton wrote:
>>> Regarding your "NOTES" addition, I don't feel comfortable with the
>>> "clean" language. I would prefer something like:
>>> When fsync() reports a failure (EIO, ENOSPC, EDQUOT) it must be assumed
>>> that any write requests initiated since the previous successful fsync
>>> was initiated may have failed, and that any cached data may have been
>>> lost. A future fsync() will not attempt to write out the same data
>>> again. If recovery is possible and desired, the application must
>>> repeat all the writes that may have failed.
>>> If the regions of a file that were written to prior to a failed fsync()
>>> are read, the content reported may not reflect the stored content, and
>>> subsequent reads may revert to the stored content at any time.
>> Much nicer.
> I guess someone should turn it into a patch....

That woud be great.

>> Should we make a distinction between usage and functional classes of
>> errors in this? The "usage" errors will probably not result in the pages
>> being tossed out, but the functional ones almost certainly will...
> Maybe. I think it is a useful distinction, but to be consistent it
> would be best to make it in all (section 2) man pages. Maybe not all at
> once though. It is really up to Michael if that is a direction he is
> interesting in following.

I think it's useful, and I'd accept patches that make such
distinctions. Of course, if we said *everything* should get fixed
at the same time, nothing would get fixed :-). So, I think
I'd just take individual patches that made such changes on an
ad hoc basis.



Michael Kerrisk
Linux man-pages maintainer;
Linux/UNIX System Programming Training: