To reduce memory pressure, we can add light per-loop (netmgr worker)
memory pools for isc_nmsocket_t structures. This will help in
situations where there's a lot of churn creating and destroying the
nmsockets.
Embedding isc_nmsocket_h2_t directly inside isc_nmsocket_t had increased
the size of isc_nmsocket_t to 1840 bytes. Making the isc_nmsocket_h2_t
to be a pointer to the structure and allocated on demand allows us to
reduce the size to 1208 bytes. While there are still some possible
reductions in the isc_nmsocket_t (embedded tlsstream, streamdns
structures), this was the far biggest drop in the memory usage.
The uv_req union member of struct isc__nm_uvreq contained libuv request
types that we don't use. Turns out that uv_getnameinfo_t is 1000 bytes
big and unnecessarily enlarged the whole structure. Remove all the
unused members from the uv_req union.
This commit improves TLS messages framing by avoiding an extra call to
SSL_write_ex(). Before that we would use an extra SSL_write_ex() call
to pass DNS message length to OpenSSL. That could create an extra TLS
frame, increasing number of bytes sent due to frame header and
padding.
This commit fixes that by making the code pass both DNS message length
and data at once, just like old TLS code did.
It should improve compatibility with some buggy clients that expect
both DNS message length and data to be in one TLS frame.
Older TLS DNS code worked like this, too.
The statistics channel does not expose the current number of TCP clients
connected, only the highwater. Therefore, users did not have an easy
means to collect statistics about TCP clients served over time. This
information could only be measured as a seperate mechanism via rndc by
looking at the TCP quota filled.
In order to expose the exact current count of connected TCP clients
(tracked by the "tcp-clients" quota) as a statistics counter, an
extra, dedicated Network Manager callback would need to be
implemented for that purpose (a counterpart of ns__client_tcpconn()
that would be run when a TCP connection is torn down), which is
inefficient. Instead, track the number of currently-connected TCP
clients separately for IPv4 and IPv6, as Network Manager statistics.
This commit removes wrong INSIST() condition as the assumption that if
'csock->recv_cb != NULL' iff 'csock->statichandle != NULL' is wrong.
There is no direct relation between 'csock->statichandle' and
'csock->recv_cb', as 'csock->statichandle' gets set when allocating a
handle regardless of 'csock->recv_cb' not being NULL, as it is
possible to attach to the handle without starting a read operation (at
the very least, it is correct to start writing before reading).
That condition made `cipher-suites` system test fail with crash on
some platforms in FIPS mode (namely, Oracle Linux 9) despite not being
related to FIPS at all.
This commit modifies TLS Stream and DNS-over-HTTPS transports so that
they do not use the "sock->iface" and "sock->peer" of the lower level
transport directly.
That did not cause any problems before, as things worked as expected,
but with the introduction of PROXYv2 support we use handles to store
the information in both PROXY Stream and UDP Proxy
transports. Therefore, in order to propagate the information (like
addresses), extracted from PROXYv2 headers, from the lower level
transports to the higher-level ones, we need to get that information
from the lower-level handles rather than sockets. That means that we
should get the peer and interface addresses using the intended
APIs ("isc_nmhandle_peeraddr()" and "isc_nmhandle_localaddr()").
Add the new isc__nm_dump_active_manager() function that can be used
for debugging purposes: it dumps all active sockets withing the
network manager instance.
This commit adds a new transport that supports PROXYv2 over UDP. It is
built on top of PROXYv2 handling code (just like PROXY Stream). It
works by processing and stripping the PROXYv2 headers at the beginning
of a datagram (when accepting a datagram) or by placing a PROXYv2
header to the beginning of an outgoing datagram.
The transport is built in such a way that incoming datagrams are being
handled with minimal memory allocations and copying.
In the previous versions of the NM, detecting the case when worker is
shutting down was not that important and actual status code did not
matter much. However, that might be not the case all the time.
This commit makes necessary modifications to the code.
This commit makes it possible to use PROXY Stream not only over TCP,
but also over TLS. That is, now PROXY Stream can work in two modes as
far as TLS is involved:
1. PROXY over (plain) TCP - PROXYv2 headers are sent unencrypted before
TLS handshake messages. That is the main mode as described in the
PROXY protocol specification (as it is clearly stated there), and most
of the software expects PROXYv2 support to be implemented that
way (e.g. HAProxy);
2. PROXY over (encrypted) TLS - PROXYv2 headers are sent after the TLS
handshake has happened. For example, this mode is being used (only ?)
by "dnsdist". As far as I can see, that is, in fact, a deviation from
the spec, but I can certainly see how PROXYv2 could end up being
implemented this way elsewhere.
This commit modifies TLS Stream to make it possible to use over PROXY
Stream. That is required to add PROVYv2 support into TLS-based
transports (DNS over HTTP, DNS over TLS).
This commit adds a new stream-based transport with an interface
compatible with TCP. The transport is built on top of TCP transport
and the new PROXYv2 handling code. Despite being built on top of TCP,
it can be easily extended to work on top of any TCP-like stream-based
transport. The intention of having this transport is to add PROXYv2
support into all existing stream-based DNS transport (DNS over TCP,
DNS over TLS, DNS over HTTP) by making the work on top of this new
transport.
The idea behind the transport is simple after accepting the connection
or connecting to a remote server it enters PROXYv2 handling mode: that
is, it either attempts to read (when accepting the connection) or send
(when establishing a connection) a PROXYv2 header. After that it works
like a mere wrapper on top of the underlying stream-based
transport (TCP).
The Unix Domain Sockets support in BIND 9 has been completely disabled
since BIND 9.18 and it has been a fatal error since then. Cleanup the
code and the documentation that suggest that Unix Domain Sockets are
supported.
Use the new isc_mem_c*() calloc-like API for allocations that are
zeroed.
In turn, this also fixes couple of incorrect usage of the ISC_MEM_ZERO
for structures that need to be zeroed explicitly.
There are few places where isc_mem_cput() is used on structures with a
flexible member (or similar).
Instead of growing and never shrinking the list of the inactive
handles (to be reused mostly on the UDP connections), limit the number
of maximum number of inactive handles kept to 64. Instead of caching
the inactive handles for all listening sockets, enable the caching on on
UDP listening sockets. For TCP, the handles were cached for each
accepted socket thus reusing the handles only for long-standing TCP
connections, but not reusing the handles across different TCP streams.
when reading on a streamdns socket failed due to timeout, but
the dispatch was still waiting for other responses, it would
resume reading by calling isc_nm_read() again. this caused
an assertion because the socket was already reading.
we now check that either the socket is reading, or that it was
already reading on the same handle.
The isc_async API was using lock-free stack (where enqueue operation was
not wait-free). Change the isc_async to use cds_wfcqueue internally -
enqueue and splice (move the queue members from one list to another) is
nonblocking and wait-free.
When shutting down TCP sockets, the read callback calling logic was
flawed, it would call either one less callback or one extra. Fix the
logic in the way:
1. When isc_nm_read() has been called but isc_nm_read_stop() hasn't on
the handle, the read callback will be called with ISC_R_CANCELED to
cancel active reading from the socket/handle.
2. When isc_nm_read() has been called and isc_nm_read_stop() has been
called on the on the handle, the read callback will be called with
ISC_R_SHUTTINGDOWN to signal that the dormant (not-reading) socket
is being shut down.
3. The .reading and .recv_read flags are little bit tricky. The
.reading flag indicates if the outer layer is reading the data (that
would be uv_tcp_t for TCP and isc_nmsocket_t (TCP) for TLSStream),
the .recv_read flag indicates whether somebody is interested in the
data read from the socket.
Usually, you would expect that the .reading should be false when
.recv_read is false, but it gets even more tricky with TLSStream as
the TLS protocol might need to read from the socket even when sending
data.
Fix the usage of the .recv_read and .reading flags in the TLSStream
to their true meaning - which mostly consist of using .recv_read
everywhere and then wrapping isc_nm_read() and isc_nm_read_stop()
with the .reading flag.
4. The TLS failed read helper has been modified to resemble the TCP code
as much as possible, clearing and re-setting the .recv_read flag in
the TCP timeout code has been fixed and .recv_read is now cleared
when isc_nm_read_stop() has been called on the streaming socket.
5. The use of Network Manager in the named_controlconf, isccc_ccmsg, and
isc_httpd units have been greatly simplified due to the improved design.
6. More unit tests for TCP and TLS testing the shutdown conditions have
been added.
Co-authored-by: Ondřej Surý <ondrej@isc.org>
Co-authored-by: Artem Boldariev <artem@isc.org>
By inspecting the code, it was discovered that .sendbuf member of the
isc__nm_networker_t was unused and just consuming ~64k per worker.
Remove the member and the association allocation/deallocation.
In e185412872, the TCP accept quota code
became broken in a subtle way - the quota would get initialized on the
first accept for the server socket and then deleted from the server
socket, so it would never get applied again.
Properly fixing this required a bigger refactoring of the isc_quota API
code to make it much simpler. The new code decouples the ownership of
the quota and acquiring/releasing the quota limit.
After (during) the refactoring it became more clear that we need to use
the callback from the child side of the accepted connection, and not the
server side.
Instead of using isc_async_run() when closing StreamDNS handle, add
isc_job_t member to the isc_nmhandle_t structure and use isc_job_run()
to avoid allocation/deallocation on the StreamDNS hot-path.
If the quota callback is called on a thread matching the socket, call
the TCP accept function directly instead of using isc_async_run() which
allocates-deallocates memory.
Cleanup the remnants of MS Compiler bits from <isc/refcount.h>, printing
the information in named/main.c, and cleanup some comments about Windows
that no longer apply.
The bits in picohttpparser.{h,c} were left out, because it's not our
code.
The `isc_trampoline` module had a lot of machinery to support stable
thread IDs for use by hazard pointers. But the hazard pointer code
is gone, and the `isc_loop` module now has its own per-loop thread
IDs.
The trampoline machinery seems over-complicated for its remaining
tasks, so move the per-thread initialization into `isc/thread.c`,
and delete the rest.
This is a simple replacement using the semantic patch from the previous
commit and as added bonus, one removal of previously undetected unused
variable in named/server.c.
Instead of marking the unused entities with UNUSED(x) macro in the
function body, use a `ISC_ATTR_UNUSED` attribute macro that expans to
C23 [[maybe_unused]] or __attribute__((__unused__)) as fallback.
With the changes to tls_try_handshake() made in
2846888c57 there are some incorrect
INSISTS() related to handshake handling which better to be removed.
When accepting a TCP connection in the higher layers (tlsstream,
streamdns, and http) attach to the socket the connection was accepted
on, and use this socket instead of the parent listening socket.
This has an advantage - accessing the sock->listener now doesn't break
the thread boundaries, so we can properly check whether the socket is
being closed without requiring .closing member to be atomic_bool.
The last atomic_bool variable sock->active was converted to non-atomic
bool by properly handling the listening socket case where we were
checking parent socket instead of children sockets.
This is no longer necessary as we properly set the .active to false on
the children sockets.
Additionally, cleanup the .rchildren - the atomic variable was used for
mutex+condition to block until all children were listening, but that's
now being handled by a barrier.
Finally, just remove dead .self and .active_child_connections members of
the netmgr socket.
Now that everything runs on their own loop and we don't cross the thread
boundaries (with few exceptions), most of the atomic_bool variables used
to track the socket state have been unatomicized because they are always
accessed from the matching thread.
The remaining few have been relaxed: a) the sock->active is now using
acquire/release memory ordering; b) the various global limits are now
using relaxed memory ordering - we don't really care about the
synchronization for those.
Change the isc_job_run() to not-make any allocations. The caller must
make sure that it allocates isc_job_t - usually as part of the argument
passed to the callback.
For simple jobs, using isc_async_run() is advised as it allocates its
own separate isc_job_t.
Mark Andrews
Ondřej Surý
Artem Boldariev
Aydın Mercan
Tony FinchandOndřej Surý
Evan Hunt
