We were doing this because of (correct) reports that NetBSD gives an
EBADF when you try to add the write side of a pipe for which the
read side has been closed. But on most kqueue platforms, that
doesn't happen, and on *all* kqueue platforms, reporting a
nonexistent fd (which we usually have if we have seen EBADF) as
readable tends to give programs a case of the vapors.
Nicholas Marriott wrote the original patch here; I did the comment
fixes.
This is not a perfect fix, but it's much much better than the
current buggy behavior, which could lead to filtering SSL
connections that just stopped reading.
Based on ideas by Maseeb Abdul Qadir and Mark Ellzey.
While re-adding all the events, event_reinit() could add a signal
event, which could then cause evsig_add() to add the
base->sig.ev_signal event. Later on its merry path through
base->eventqueue, event_reinit() would find that same event and give
it to event_io_add a second time. This would make the ev_io_next
list for that fd become circular. Ouch!
As originally written, the test would only pass if the accept()
callbacks for the evconnlistener were all invoked before the last of
the CONNECTED/ERROR callbacks for the connecting/resolving bufferevent
had its call to event_base_loopexit() complete. But this was only
accidentally true in 2.0, and might not be true at all in 2.1 where
we schedule event_base_once() callbacks more aggressively.
Found by Sebastian Hahn.
Older Linuxes sometimes respond to some nmap probes by having accept()
return a success but with socklen 0. That can lead to confusing behavior
when you go to process the sockaddr.
I thought we'd fixed the cases where this could come up, but
apparently having an event_base_break() happen while processing
signal events could get us in trouble.
Found by Remi Gacogne. Sourceforge issue 3451433 .
Previously, if some sender were generating data to read on an
OpenSSL connection as fast as we could process it, we could easily
wind up looping on an openssl do_read operation without ever
considering other sockets.
The difference between this and the original method in
consider_reading() is that it only loops for a single completed
*frame* instead of looping until fd is drained or an error condition
was triggered.
{Patch split out by nickm}
OpenSSL bufferevents with deferred callbacks enabled under high load will
spinlock in the function consider_reading(). This loop continues until all
data has been read.
Because of this condition; openssl bufferevents will never return back into
event_base_loop() until SSL_read has determined data is no longer ready.
As of yet I have not found a reason why this while loop exists, so this patch
just swaps out while for if.
If needed I can write same code which would trigger this effect; optionally
libevhtp has a test.c program which can be run with the following flags:
./test -s <keyfile.pem>
curl -vvvv -k -d@<HUGE_ASS_FILE> https://127.0.0.1:8081/
The return data will include the number of times the readcb got data and the
length of that read.
Without this patch, you are likely to see a small amount of "bytes read....",
otherwise the "bytes read..." return data should show much more reasonable
numbers.
