The code to support parallel Netlink queries (commit 3c7193f) made use
of nlmsg_len member from struct nlmsghdr to allocate and copy the
responses. Since NLMSG_NEXT is later used to parse these responses, they
must be aligned, or the results are undefined.
Signed-off-by: Thomas Egerer <thomas.egerer@secunet.com>
When querying SAs the keys will end up in this buffer (the allocated
messages that are returned are already wiped). The kernel also returns
XFRM_MSG_NEWSA as response to XFRM_MSG_ALLOCSPI but we can't distinguish
this here as we only see the response.
References #2388.
During initialization of the plugins the thread pool is not yet
initialized so there is no watcher thread that could handle the queued
Netlink message and the main thread will wait indefinitely for a
response.
Fixes#2199.
When constructing the result, all responses from Netlink were concatenated
iteratively, i.e. for each response, the previously acquired result was
copied to newly allocated memory and the current response appended to it.
This results in O(n^2) copy operations. Instead, we now check for the
total final length of the result and copy the individual responses to it
in one pass, i.e. in O(n) copy operations. In particular, this issue caused
very high CPU usage in memcpy() function as the result is copied over and
over. Common way how to hit the issue is when having 1000+ routes and 5+
connecting clients a second. In that case, the memcpy() function can
take 50%+ of one CPU thread on a decent CPU and the whole charon daemon
is stuck just reading routes and concatenating them together (connecting
clients are blocked in that particular case as this is done under mutex).
Closes strongswan/strongswan#65.
References #2055.
After adding the read callback the state is WATCHER_QUEUED and it is
switched to WATCHER_RUNNING only later by an asynchronous job. This means
that a thread that sent a Netlink message shortly after registration
might see the state as WATCHER_QUEUED. If it then tries to read the
response and the watcher thread is quicker to actually read the message
from the socket, it could block on recv() while still holding the lock.
And the asynchronous job that actually read the message and tries to queue
it will block while trying to acquire the lock, so we'd end up in a deadlock.
This is probably mostly a problem in the unit tests.