Re: epoll_wait returning EFAULT on Linux 3.2.78

Andres Freund <andres(at)anarazel(dot)de> writes:
> On 2016-06-02 18:41:00 +0100, Greg Stark wrote:
>> Well there's not *nothing* we can do. I thought I we were going to
>> have to go back and do manual offset calculations to get that right.

> The kernel accesses the elements as an array. If the array stride (by
> virtue of sizeof) were wrong, we couldn't fix that.

Right. The general rule in C is that sizeof(anything) is always a
multiple of the something's alignment requirement, so that if you
have a correctly aligned initial element of an array then later
elements are also correctly aligned. The problem in our existing
code is that sizeof(WaitEventSet) might not be a multiple of the
alignment requirement of WaitEvent, and either of those two might
not be a multiple of the alignment requirement of struct epoll_event,
etc. So we should make the code look like

sz += MAXALIGN(sizeof(WaitEventSet));
sz += MAXALIGN(sizeof(WaitEvent) * nevents);

#if defined(WAIT_USE_EPOLL)
sz += MAXALIGN(sizeof(struct epoll_event) * nevents);

etc, so that each of the subsidiary arrays starts on a MAXALIGN boundary.
Where the later array elements fall is taken care of given that.

regards, tom lane