Query cancellation on hot standby because of buffer pins

I'm seeing query cancellation on hot standby (Postgres 9.3.4, configured
with hot_standby_feedback=on) because of buffer pins being held for too
long.

Database message looks like this:

ERROR: canceling statement due to conflict with recovery
DETAIL: User was holding shared buffer pin for too long.
STATEMENT: ...

I don't understand what kind of activity (either on primary or on standby)
can cause this. The documentation about cancellation on standby lists
access exclusive locks, vacuum and drops of tablespace or database on the
primary as possible causes for cancellation on the standby.

I'm sure that neither of those things was happening on the primary when the
query was cancelled on standby. I also don't know what "too long" means?
That is, what's the timeout? I've looked at the source a bit and it seems
that the timeout is the value of deadlock_timeout parametar, but I'm not
very sure about that.

So can somebody explain why the query needs to be cancelled if it pins the
buffer for too long and what exactly "too long" is?

The only thing I managed to find on the net is:
http://t61224.db-postgresql-bugs.dbtalk.info/bug-7546-backups-on-hot-standby-cancelled-despite-hot-standby-on-t61224.html

But that doesn't explain why this kind of cancellation happens.