Well, it turns out that when I tested it on my own machine, I actually
wasn't using a tmpfs - I misread the output of df and used /tmp¹ instead
of /run. Doing the test in /run does in fact give a bus error. Mea culpa.

¹ Apparently on my system /tmp isn't a tmpfs. Go figure.


> Also, if the filesystem doesn't support sparse files, this writes a
> lot to the disk (and if it crashes, you end up with a huge file).

You may be right; I hadn't thought about that possibility. My concern was
that the test suite isn't run with -M very often, so these sorts of tests
could often be broken for a while before someone found out. In the past,
none of the buildbots ran bigmem tests, so there was a real danger that
we wouldn't notice breakages. However, with the addition of the debian
bigmem buildbot, that is no longer the case, so this isn't such an issue
any more.

I'm okay with leaving the tests as they are in 3.3. Any objections?
If not, I'll also backport the change to 3.2 and 2.7.