There is a theorical problem in the wait queue, which is that with many
threads, one could spend a lot of time looping on the newly expired tasks,
causing a lot of contention on the global wq_lock and on the global
rq_lock. This initially sounds bening, but if another thread does just
a task_schedule() or task_queue(), it might end up waiting for a long
time on this lock, and this wait time will count on its execution budget,
degrading the end user's experience and possibly risking to trigger the
watchdog if that lasts too long.
The simplest (and backportable) solution here consists in bounding the
number of expired tasks that may be picked from the global wait queue at
once by a thread, given that all other ones will do it as well anyway.
We don't need to pick more than global.tune.runqueue_depth tasks at once
as we won't process more, so this counter is updated for both the local
and the global queues: threads with more local expired tasks will pick
less global tasks and conversely, keeping the load balanced between all
threads. This will guarantee a much lower latency if/when wakeup storms
happen (e.g. hundreds of thousands of synchronized health checks).
Note that some crashes have been witnessed with 1/4 of the threads in
wake_expired_tasks() and, while the issue might or might not be related,
not having reasonable bounds here definitely justifies why we can spend
so much time there.
This patch should be backported, probably as far as 2.0 (maybe with
some adaptations).
The proxy stopping mechanism was changed with commit 322b9b94e ("MEDIUM:
proxy: make stop_proxy() now use stop_listener()") so that it's now
entirely driven by the listeners. One thing was forgotten though, which
is that pure backends will not stop anymore since they don't have any
listener, and that it's necessary to stop them in order to stop the
health checks.
No backport is needed.
We don't need to specify the handler anymore since it's set in the
receiver. Let's remove this argument from the function and clean up
the remains of code that were still setting it.
Now we define a new sock_accept_iocb() for socket-based stream protocols
and use it as a wrapper for listener_accept() which now takes a listener
and not an FD anymore. This will allow the receiver's I/O cb to be
redefined during registration, and more specifically to get rid of the
hard-coded hacks in protocol_bind_all() made for syslog.
The previous ->accept() callback in the protocol was removed since it
doesn't have anything to do with accept() anymore but is more generic.
A few places where listener_accept() was compared against the FD's IO
callback for debugging purposes on the CLI were updated.
For now we're still using the protocol's default accept() function as
the I/O callback registered by the receiver into the poller. While
this is usable for most TCP connections where a listener is needed,
this is not suitable for UDP where a different handler is needed.
Let's make this configurable in the receiver just like the upper layer
is configurable for listeners. In order to ease stream protocols
handling, the protocols will now provide a default I/O callback
which will be preset into the receivers upon allocation so that
almost none of them has to deal with it.
The receiver FDs must not be manipulated by the listener_accept()
function anymore, it must exclusively rely on the job performed by
its listeners, as it is also the only way to keep the receivers
working for established connections regardless of the listener's
state (typically for multiplexed protocols like QUIC). This used
to be necessary when the FDs were adjusted at once only but now
that fd_done() is gone and the need for polling enabled by the
accept_conn() function which detects the EAGAIN, we have nothing
to do there to fixup any possible previous bad decision anymore.
Interestingly, as a side effect of making the code not depend on
the FD anymore, it also removes the need for a second lock, which
increase the accept rate by about 1% on 8 threads.
Now listener_accept() doesn't have to deal with the incoming FD anymore
(except for a little bit of side band stuff). It directly retrieves a
valid connection from the protocol layer, or receives a well-defined
error code that helps it decide how to proceed. This removes a lot of
hardly maintainable low-level code and opens the function to receive
new protocol stacks.
This is the same as previous commit, but this time for the sockpair-
specific stuff, relying on recv_fd_uxst() instead of accept(), so the
code is simpler. The various errno cases are handled like for regular
sockets, though some of them will probably never happen, but this does
not hurt.
The socket-specific accept() code in listener_accept() has nothing to
do there. Let's move it to sock.c where it can be significantly cleaned
up. It will now directly return an accepted connection and provide a
status code instead of letting listener_accept() deal with various errno
values. Note that this doesn't support the sockpair specific code.
The function is now responsible for dealing with its own receiver's
polling state and calling fd_cant_recv() when facing EAGAIN.
One tiny change from the previous implementation is that the connection's
sockaddr is now allocated before trying accept(), which saves a memcpy()
of the resulting address for each accept at the expense of a cheap
pool_alloc/pool_free on the final accept returning EAGAIN. This still
apparently slightly improves accept performance in microbencharks.
This call was introduced by commit 5ced3e887 ("MINOR: sock: add
sock_accept_conn() to test a listening socket") but is actually quite
confusing because it makes one think the socket will accept a connection
(which is what we want to have in a new function) while it only tells
whether it's configured to accept connections. Let's call it
sock_accepting_conn() instead.
The same change was applied to sockpair which had the same issue.
Now that this function is always called with an initialized connection
and that the control layer is always initialized, we don't need to play
games with fdtab[] to decide how to close, we can simply rely on the
regular close path using conn_ctrl_close(), which can be fused with
conn_xprt_close() into conn_full_close().
The code is cleaner because the FD is now used only for some
protocol-specific setup (that will eventually have to move) and to
try to send a hard-coded HTTP 500 error message on raw sockets.
Till now we would keep a per-thread queue of pending incoming connections
for which we would store:
- the listener
- the accepted FD
- the source address
- the source address' length
And these elements were first used in session_accept_fd() running on the
target thread to allocate a connection and duplicate them again. Doing
this induces various problems. The first one is that session_accept_fd()
may only run on file descriptors and cannot be reused for QUIC. The second
issue is that it induces lots of memory copies and that the listerner
queue thrashes a lot of cache, consuming 64 bytes per entry.
This patch changes this by allocating the connection before queueing it,
and by only placing the connection's pointer into the queue. Indeed, the
first two calls used to initialize the connection already store all the
information above, which can be retrieved from the connection pointer
alone. So we just have to pop one pointer from the target thread, and
pass it to session_accept_fd() which only needs the FD for the final
settings.
This starts to make the accept path a bit more transport-agnostic, and
saves memory and CPU cycles at the same time (1% connection rate increase
was noticed with 4 threads). Thanks to dividing the accept-queue entry
size from 64 to 8 bytes, its size could be increased from 256 to 1024
connections while still dividing the overall size by two. No single
queue full condition was met.
One minor drawback is that connection may be allocated from one thread's
pool to be used into another one. But this already happens a lot with
connection reuse so there is really nothing new here.
Roughly half of the calls to sockadr_alloc() are made to copy an already
known address. Let's optionally pass it in argument so that the function
can handle the copy at the same time, this slightly simplifies its usage.
fd_done_recv() used to be useful with the FD cache because it used to
allow to keep a file descriptor active in the poller without being
marked as ready in the cache, saving it from ringing immediately,
without incurring any system call. It was a way to make it yield
to wait for new events leaving a bit of time for others. The only
user left was the connection accepter (listen_accept()). We used
to suspect that with the FD cache removal it had become totally
useless since changing its readiness or not wouldn't change its
status regarding the poller itself, which would be the only one
deciding to report it again.
Careful tests showed that it indeed has exactly zero effect nowadays,
the syscall numbers are exactly the same with and without, including
when enabling edge-triggered polling.
Given that there's no more API available to manipulate it and that it
was directly called as an optimization from listener_accept(), it's
about time to remove it.
No protocol defines it anymore. The last user used to be the monitor-net
stuff that got partially broken already when the tcp_drain() function
moved to conn_sock_drain() with commit e215bba95 ("MINOR: connection:
make conn_sock_drain() work for all socket families") in 1.9-dev2.
A part of this will surely move back later when non-socket connections
arrive with QUIC but better keep the API clean and implement what's
needed in time instead.
As discussed here during 2.1-dev, "monitor-net" is totally obsolete:
https://www.mail-archive.com/haproxy@formilux.org/msg35204.html
It's fundamentally incompatible with usage of SSL, and imposes the
presence of file descriptors with hard-coded syscalls directly in the
generic accept path.
It's very unlikely that anyone has used it in the last 10 years for
anything beyond testing. In the worst case if anyone would depend
on it, replacing it with "http-request return status 200 if ..." and
"mode http" would certainly do the trick.
The keyword is still detected as special by the config parser to help
users update their configurations appropriately.
As discussed here during 2.1-dev, "mode health" is totally obsolete:
https://www.mail-archive.com/haproxy@formilux.org/msg35204.html
It's fundamentally incompatible with usage of SSL, doesn't support
source filtering, and imposes the presence of file descriptors with
hard-coded syscalls directly in the generic accept path.
It's very unlikely that anyone has used it in the last 10 years for
anything beyond testing. In the worst case if anyone would depend
on it, replacing it with "http-request return status 200" and "mode
http" would certainly do the trick.
The keyword is still detected as special by the config parser to help
users update their configurations appropriately.
FCGI mux is marked with HOL blocking. On safe reuse mode, the connection
using it are placed on the sessions instead of the available lists to
avoid sharing it with several clients. On detach, if they are no more
streams, remove the connection from the session before adding it to the
idle list. If there is still used streams, do not add it to available
list as it should be already on the session list.
H2 mux is marked with HOL blocking. On safe reuse mode, the connection
using it are placed on the sessions instead of the available lists to
avoid sharing it with several clients. On detach, if they are no more
streams, remove the connection from the session before adding it to the
idle list. If there is still used streams, do not add it to available
list as it should be already on the session list.
If a connection is using a mux protocol subject to HOL blocking, add it
to the session instead of the available list to avoid sharing it with
other clients on connection reuse.
When allocating a new session on connect_server, if the mux protocol is
marked as subject of HOL blocking, add it into session instead of
available list to avoid sharing it with other clients.
On server connection migration from one thread to another, the wrong
idle thread-specific counter is decremented. This bug was introduced
since commit 3d52f0f1f828acb2a74b3f13fcc3fa069106d09f due to the
factorization with srv_use_idle_conn. However, this statement is only
executed from conn_backend_get. Extract the decrement from
srv_use_idle_conn in conn_backend_get and use the correct
thread-specific counter.
Rename the function to srv_use_conn to better reflect its purpose as it
is also used with a newly initialized connection not in the idle list.
As a side change, the connection insertion to available list has also
been extracted to conn_backend_get. This will be useful to be able to
specify an alternative list for protocol subject to HOL risk that should
not be shared between several clients.
This bug is only present in this release and thus do not need a backport.
The loop always missed one iteration due to the incrementation done on
the for check. Move the incrementation on the loop last statement to fix
this behaviour.
This bug has a very limited impact, not at all visible to the user, but
could be backported to 2.2.
When running a pure config check (haproxy -c) we go through the deinit
phase without having allocated fdtab, so we can't blindly dereference
it. The issue was added by recent commit ae7bc4a23 ("MEDIUM: deinit:
close all receivers/listeners before scanning proxies"), no backport is
needed.
In issue #894, Coverity suspects uninitialized values for a socket's
address whose family is AF_UNSPEC but it doesn't know that the address
is not used in this case. It's not on a critical path and working around
it is trivial, let's fully declare the address. We're doing it for both
TCP and UDP, because the same principle appears at two places.
It may happen that during a temporary listener pause resulting from a
SIGTTOU, one process gets one of its sockets disabled by another process
and will not be able to recover from this situation by itself. For the
protocols supporting this (TCPv4 and TCPv6 at the moment) this situation
is detectable, so when this happens, let's put the listener into the
PAUSED state so that it remains consistent with the real socket state.
One nice effect is that just sending the SIGTTIN signal to the process
is enough to recover the socket in this case.
There is no need to backport this, this behavior has been there forever
and the fix requires to reimplement the getsockopt() call there.
For socket pairs we don't rely on a real listening socket but we need to
have a properly connected UNIX stream socket. This is what the new
sockpair_accept_conn() tries to report. Some corner cases like half
shutdown will still not be detected but that should be sufficient for
most cases we really care about.
Now we introdce a new .rx_listening() function to report if a receiver is
actually a listening socket. The reason for this is to help detect shared
sockets that might have been broken by sibling processes.
At several places we need to check if a socket is still valid and still
willing to accept connections. Instead of open-coding this, each time,
let's add a new function for this.
Currently the suspend/resume mechanism for listeners only works on Linux
and we resort to a number of tricks involving shutdown+listen+shutdown
to try to detect failures on other operating systems that do not support
it. But on Linux connect(AF_UNSPEC) also works pretty well and is much
cleaner. It still doesn't work on other operating systems but the error
is easier to detect and appears safer. So let's switch to this.
When starting with a huge maxconn (say 1 billion), the only error seen
is "No polling mechanism available". This doesn't help at all to resolve
the problem. Let's add specific alerts for the failed mallocs. Now we can
get this instead:
[ALERT] 286/154439 (23408) : Not enough memory to allocate 2000000033 entries for fdtab!
This may be backported as far as 2.0 as it helps debugging bad configurations.
There are reports of a few "SC" in logs during reloads when H2 is used
on the backend side. Christopher analysed this as being caused by the
proxy disabled test in h2_process(). As the comment says, this was done
for frontends only, and must absolutely not send a GOAWAY to the backend,
as all it will result in is to make newly queued streams fail.
The fix consists in simply testing the connection side before deciding
to send the GOAWAY.
This may be backported as far as 2.0, though for whatever reason it seems
to manifest itself only since 2.2 (probably due to changes in the outgoing
connection setup sequence).
On some operating systems, RLIM_INFINITY is set to -1 so that when the
hard limit on the number of FDs is set to unlimited, taking the MAX
of both values keeps rlim_fd_cur and everything works. But on other
systems this values is defined as the highest positive integer. This
is what was observed on a 32-bit AIX 5.1. The effect is that maxsock
becomes 2^31-1 and that fdtab allocation fails.
Note that a simple workaround consists in manually setting maxconn in
the global section.
Let's ignore unlimited as soon as we retrieve rlim_fd_max so that all
systems behave consistently.
This may be backported as far as 2.0, though it doesn't seem like it
has annoyed anyone.
This patch adds "coll" new counter and the heartbeat timer values to "show peers"
command. It also adds the elapsed time since the last handshake to new "last_hdshk"
new peer dump field.
When factoring out the pause/resume error messages in commit 775e00158
("MAJOR: signals: use protocol_pause_all() and protocol_resume_all()")
I forgot that ha_warning() and send_log() take a format string and not
just a const string. No backport is needed, this is 2.3-dev.
This one was scheduled for removal in 2.3 since 2.2-dev3 by commit
1b85785bc ("MINOR: config: mark global.debug as deprecated"). Let's
remove it now. It remains totally possible to use -d on the command
line though.
This was introduced 15 years ago or so to delay the stopping of some
services so that a monitoring device could detect its port being down
before services were stopped. Since then, clean reloads were implemented
and this doesn't cope well with reload at all, preventing the new process
from seamlessly binding, and forcing processes to coexist with half-baked
configurations.
Now it has become a real problem because there's a significant code
portion in the proxies that is solely dedicated to this obsolete feature,
and dealing with its special cases eases the introduction of bugs in
other places so it's about time that it goes.
We could tentatively schedule its removal for 2.4 with a hard deadline
for 2.5 in any case.
Now we have ->suspend() and ->resume() for listeners at the protocol
level. This means that it now becomes possible for a protocol to redefine
its own way to suspend and resume. The default functions are provided for
TCP, UDP and unix, and they are pass-through to the receiver equivalent
as it used to be till now. Nothing was defined for sockpair since it does
not need to suspend/resume during reloads, hence it will succeed.
The inner part now goes into the protocol and is used to decide how to
unbind a given protocol's listener. The existing code which is able to
also unbind the receiver was provided as a default function that we
currently use everywhere. Some complex listeners like QUIC will use this
to decide how to unbind without impacting existing connections, possibly
by setting up other incoming paths for the traffic.
This is used as a generic way to unbind a receiver at the end of
do_unbind_listener(). This allows to considerably simplify that function
since we can now let the protocol perform the cleanup. The generic code
was moved to sock.c, along with the conditional rx_disable() call. Now
the code also supports that the ->disable() function of the protocol
which acts on the listener performs the close itself and adjusts the
RX_F_BUOND flag accordingly.
This listener flag indicates whether the receiver part of the listener
is specific to the master or to the workers. In practice it's only used
by the master's CLI right now. It's used to know whether or not the FD
must be closed before forking the workers. For this reason it's way more
of a receiver's property than a listener's property, so let's move it
there under the name RX_F_MWORKER. The rest of the code remains
unchanged.
And also remove it from its callers. This subtle distinction was added as
sort of a hack for the seamless reload feature but is not needed anymore
since the do_close turned unused since commit previous commit ("MEDIUM:
listener: let do_unbind_listener() decide whether to close or not").
This also removes the unbind_listener_no_close() function.
The listener contains all the information needed to decide to close on
unbind or not. The rule is the following (when we're not stopping):
- worker process unbinding from a worker's FD with socket transfer enabled => keep
- master process unbinding from a master's inherited FD => keep
- master process unbinding from a master's FD => close
- master process unbinding from a worker's FD => close
- worker process unbinding from a master's FD => close
- worker process unbinding from a worker's FD => close
Let's translate that into the function and stop using the do_close
argument that is a bit obscure for callers. It was not yet removed
to ease code testing.
BROKEN: the failure rate on reg-tests/seamless-reload/abns_socket.vtc has
significantly increased for no obvious reason. It fails 99% of the time vs
10% before.
do_unbind_listener() is not logical and is not even idempotent. It must
not touch the fd if already -1, which also means not touch the receiver.
In addition, when performing a partial stop on a socket (not closing),
we know the socket remains in the listening state yet it's marked as
LI_ASSIGNED, which is confusing as it doesn't translate its real state.
With this change, we make sure that FDs marked for close end up in
ASSIGNED state and that those which are really bound and on which a
listen() was made (i.e. not pause) remain in LISTEN state. This is what
is closest to reality.
Ideally this function should become a default proto->unbind() one but
it may still keep a bit too much state logic to become generalized to
other protocols (e.g. QUIC).
Right now in enable_listener(), we used to start all enabled
listeners then kill from the workers those that were for the master.
But this is incomplete. We must also close from the master the
listeners that are solely for workers, and do it before we even
start them. Otherwise we end up with a master responding to the
worker CLI connections if the listener remains in listen mode to
translate the socket's real state.
It doesn't seem like it could have caused bugs in the past because we
used to aggressively mark disabled listeners as LI_ASSIGNED despite
the fact that they were still bound and listening. If this patch were
ever seen as a candidate solution for any obscure bug, be careful in
that it subtly relies on the fact that fd_delete() doesn't close
inherited FDs anymore, otherwise that could break the master's ability
to pass inherited FDs on reloads.
