Davide,
Like reported by Michael Kerrisk, POLLRDHUP handling was not consistent[...]
between epoll and poll/select, since in epoll it was unmaskeable. This
patch brings uniformity in POLLRDHUP handling.
diff -Nru linux-2.6.16/fs/eventpoll.c linux-2.6.16.mod/fs/eventpoll.c
--- linux-2.6.16/fs/eventpoll.c 2006-04-03 20:08:23.000000000 -0700
+++ linux-2.6.16.mod/fs/eventpoll.c 2006-04-03 20:09:51.000000000 -0700
@@ -599,7 +599,7 @@
switch (op) {
case EPOLL_CTL_ADD:
if (!epi) {
- epds.events |= POLLERR | POLLHUP | POLLRDHUP;
+ epds.events |= POLLERR | POLLHUP;
error = ep_insert(ep, &epds, tfile, fd);
} else
@@ -613,7 +613,7 @@
break;
case EPOLL_CTL_MOD:
if (epi) {
- epds.events |= POLLERR | POLLHUP | POLLRDHUP;
+ epds.events |= POLLERR | POLLHUP;
error = ep_modify(ep, epi, &epds);
} else
error = -ENOENT;
This makes things consistent -- but in the opposite way
from what I thought they might be. (The alternative would of
course have been to make POLLRDHUP un-maskable in both epoll
and poll().)
So I'm curious: what is the rationale for making POLLRDHUP
maskable when POLLHUP is not? Is it an issue of ABI
compatibility; or something else?