[squid-dev] ftruncate() failures on OS X (Darwin)

[Alex Rousskov]

On 01/21/2016 08:03 PM, Markus Mayer wrote:

> I left out the "flags" variable from the shm_open() call that you
> showed, since I didn't see it being used or declared anywhere.

Sorry about that leftover from a previous version of my sketch. You did
the right thing.


> I did test it, and it works "as advertised".

Your patch is solving two problems at once. You have described only one,
but the other problem[1] is arguably more severe because (a) it leads to
subtle bugs and crashes rather than startup failures and (b) most people
run Squid on OSes other than OS X.

[1] Item #1 at
http://lists.squid-cache.org/pipermail/squid-dev/2015-December/004112.html


IMO, we should document both problems in the code to minimize our
chances of repeating them during future refactoring. I suggest replacing
this comment:

> +    // OS X does not allow using O_TRUNC with shm_open. Also, OS X permits
> +    // ftruncate() to be called only once on a shared memory area. The call
> +    // will fail if the shared memory area was previously truncated. To
> +    // prevent this error, we delete and re-create the area if it existed
> +    // previously (i.e. from an unclean shutdown) and wasn't newly created.
> +    if (!createExclusive() && errno == EEXIST) {
> +        unlink();
> +        createExclusive();
> +    }


with something like this (shown here without formatting):

```
Why a brand new segment? Our placement-new code requires an all-0
segment. We could truncate and resize an old segment, but OS X does not
allow using O_TRUNC with shm_open() and does not support resizing after
ftruncate(0).
```

and perhaps renaming createExclusive() to createFresh() for clarity. The
above can be done during commit -- no need to repost IMO.


Your commit message will explain the details. I (or another committer)
will add text to state that problem [1] is also fixed by this change.


If there are no objections, I will commit your patch (with the above
polishing touches).


Thank you,

Alex.