Listeners might be disabled by other threads while running in
listener_accept() due to a stopping condition or possibly a rebinding
error after a failed stop/start. When this happens, the listener's FD
is -1 and accesses made by the lower layers to fdtab[-1] do not end up
well. This can occasionally be noticed if running at high connection
rates in master-worker mode when compiled with ASAN and hammered with
10 reloads per second. From time to time an out-of-bounds error will
be reported.
One approach could consist in keeping a copy of critical information
such as the FD before proceeding but that's not correct since in case of
close() the FD might be reassigned to another connection for example.
In fact what is needed is to read-lock the listener during this operation
so that it cannot change while we're touching it.
Tests have shown that using a spinlock only does generally work well but
it doesn't scale much with threads and we can see listener_accept() eat
10-15% CPU on a 24 thread machine at 300k conn/s. For this reason the
lock was turned to an rwlock by previous commit and this patch only takes
the read lock to make sure other operations do not change the listener's
state while threads are accepting connections. With this approach, no
performance loss was noticed at all and listener_accept() doesn't appear
in perf top.
This ought to be backported to about all branches that make use of the
unlocked listeners, but in practice it seems to mostly concern 2.3 and
above, since 2.2 and older will take the FD in the argument (and the
race exists there, this FD could end up being reassigned in parallel
but there's not much that can be done there to prevent that race; at
least a permanent error will be reported).
For backports, the current approach is preferred, with a preliminary
backport of previous commit "MINOR: listener: replace the listener's
spinlock with an rwlock". However if for any reason this commit cannot
be backported, the current patch can be modified to simply take a
spinlock (tested and works), it will just impact high performance
workloads (like DDoS protection).
We'll need to lock the listener a little bit more during accept() and
tests show that a spinlock is a massive performance killer, so let's
first switch to an rwlock for this lock.
This patch might have to be backported for the next patch to work, and
if so, the change is almost mechanical (look for LISTENER_LOCK), but do
not forget about the few HA_SPIN_INIT() in the file. There's no reference
to this lock outside of listener.c nor listener-t.h.
Create a new structure li_per_thread. This is uses as an array in the
listener structure, with an entry allocated per thread. The new function
li_init_per_thr is responsible of the allocation.
For now, li_per_thread contains fields only useful for QUIC listeners.
As such, it is only allocated for QUIC listeners.
This flag is named RX_F_LOCAL_ACCEPT. It will be activated for special
receivers where connection balancing to threads is already handle
outside of listener_accept, such as with QUIC listeners.
Coverity reported in issue #1416 that label oom3 is not reachable in
function close_listener() added by commit 59a877dfd ("MINOR: listeners:
add clone_listener() to duplicate listeners at boot time"). The code
leading to it was removed during the development of the function, but
not the label itself.
When the clone_listener() function was added in commit 59a877dfd
("MINOR: listeners: add clone_listener() to duplicate listeners at
boot time"), a stupid bug was introduced when splitting the error
path because while the first case where calloc fails will leave NULL
in the output value, the other cases will return the pointer to a
freed area. This was reported by Coverity in issue #1416.
In practice nobody will face it (out-of-memory while checking config),
but let's fix it.
No backport is needed.
In multi-threaded mode, on operating systems supporting multiple listeners on
the same IP:port, this will automatically create this number of multiple
identical listeners for the same line, all bound to a fair share of the number
of the threads attached to this listener. This can sometimes be useful when
using very large thread counts where the in-kernel locking on a single socket
starts to cause a significant overhead. In this case the incoming traffic is
distributed over multiple sockets and the contention is reduced. Note that
doing this can easily increase the CPU usage by making more threads work a
little bit.
If the number of shards is higher than the number of available threads, it
will automatically be trimmed to the number of threads. A special value
"by-thread" will automatically assign one shard per thread.
This function's purpose will be to duplicate a listener in INIT state.
This will be used to ease declaration of listeners spanning multiple
groups, which will thus require multiple FDs hence multiple receivers.
With groups at some point we'll have to have distinct masks/groups in the
receiver and the bind_conf, because a single bind_conf might require to
instantiate multiple receivers (one per group).
Let's split the thread mask and group to have one for the bind_conf and
another one for the receiver while it remains easy to do. This will later
allow to use different storage for the bind_conf if needed (e.g. support
multiple groups).
This extends the "thread" statement of bind lines to support an optional
thread group number. When unspecified (0) it's an absolute thread range,
and when specified it's one relative to the thread group. Masks are still
used so no more than 64 threads may be specified at once, and a single
group is possible. The directive is not used for now.
The TI_FL_STUCK flag is manipulated by the watchdog and scheduler
and describes the apparent life/death of a thread so it changes
all the time and it makes sense to move it to the thread's context
for an active thread.
These functions have no reason for being inlined, and they require some
includes with long dependencies. Let's move them to listener.c and trim
unused includes in listener.h.
A number of files currently access activity counters but rely on their
definitions to be inherited from other files (task.c, backend.c hlua.c,
sock.c, pool.c, stats.c, fd.c).
We'll need to improve the API to pass other arguments in the future, so
let's start to adapt better to the current use cases. task_new() is used:
- 18 times as task_new(tid_bit)
- 18 times as task_new(MAX_THREADS_MASK)
- 2 times with a single bit (in a loop)
- 1 in the debug code that uses a mask
This patch provides 3 new functions to achieve this:
- task_new_here() to create a task on the calling thread
- task_new_anywhere() to create a task to be run anywhere
- task_new_on() to create a task to run on a specific thread
The change is trivial and will allow us to later concentrate the
required adaptations to these 3 functions only. It's still possible
to call task_new() if needed but a comment was added to encourage the
use of the new ones instead. The debug code was not changed and still
uses it.
The "process" directive on "bind" lines becomes quite confusing considering
that the only allowed value is 1 for the process, and that threads are
optional and come after the mandatory "1/".
Let's introduce a new "thread" directive to directly configure thread
numbers, and mark "process" as deprecated. Now "process" will emit a
warning and will suggest how to be replaced with "thread" instead.
The doc was updated accordingly (mostly a copy-paste of the previous
description which was already up to date).
This is marked as MEDIUM as it will impact users having "zero-warning"
and "process" specified.
Lots of places iterating over nbproc or comparing with nbproc could be
simplified. Further, "bind-process" and "process" parsing that was
already limited to process 1 or "all" or "odd" resulted in a bind_proc
field that was either 0 or 1 during the init phase and later always 1.
All the checks for compatibilities were removed since it's not possible
anymore to run a frontend and a backend on different processes or to
have peers and stick-tables bound on different ones. This is the largest
part of this patch.
The bind_proc field was removed from both the proxy and the receiver
structs.
Since the "process" and "bind-process" directives are still parsed,
configs making use of correct values allowing process 1 will continue
to work.
The current "ADD" vs "ADDQ" is confusing because when thinking in terms
of appending at the end of a list, "ADD" naturally comes to mind, but
here it does the opposite, it inserts. Several times already it's been
incorrectly used where ADDQ was expected, the latest of which was a
fortunate accident explained in 6fa922562 ("CLEANUP: stream: explain
why we queue the stream at the head of the server list").
Let's use more explicit (but slightly longer) names now:
LIST_ADD -> LIST_INSERT
LIST_ADDQ -> LIST_APPEND
LIST_ADDED -> LIST_INLIST
LIST_DEL -> LIST_DELETE
The same is true for MT_LISTs, including their "TRY" variant.
LIST_DEL_INIT keeps its short name to encourage to use it instead of the
lazier LIST_DELETE which is often less safe.
The change is large (~674 non-comment entries) but is mechanical enough
to remain safe. No permutation was performed, so any out-of-tree code
can easily map older names to new ones.
The list doc was updated.
This patch replaces roughly all occurrences of an HA_ATOMIC_ADD(&foo, 1)
or HA_ATOMIC_SUB(&foo, 1) with the equivalent HA_ATOMIC_INC(&foo) and
HA_ATOMIC_DEC(&foo) respectively. These are 507 changes over 45 files.
Just like with the server keywords, now's the turn of "bind" keywords.
The difference is that 100% of the bind keywords are registered, thus
we do not need the list of extra keywords.
There are multiple bind line parsers today, all were updated:
- peers
- log
- dgram-bind
- cli
$ printf "listen f\nbind :8000 tcut\n" | ./haproxy -c -f /dev/stdin
[NOTICE] 070/101358 (25146) : haproxy version is 2.4-dev11-7b8787-26
[NOTICE] 070/101358 (25146) : path to executable is ./haproxy
[ALERT] 070/101358 (25146) : parsing [/dev/stdin:2] : 'bind :8000' unknown keyword 'tcut'; did you mean 'tcp-ut' maybe ?
[ALERT] 070/101358 (25146) : Error(s) found in configuration file : /dev/stdin
[ALERT] 070/101358 (25146) : Fatal errors found in configuration.
The default proxy was passed as a variable to all parsers instead of a
const, which is not without risk, especially when some timeout parsers used
to make some int pointers point to the default values for comparisons. We
want to be certain that none of these parsers will modify the defaults
sections by accident, so it's important to mark this proxy as const.
This patch touches all occurrences found (89).
It's been too short for quite a while now and is now full. It's still
time to extend it to 32-bits since we have room for this without
wasting any space, so we now gained 16 new bits for future flags.
The values were not reassigned just in case there would be a few
hidden u16 or short somewhere in which these flags are placed (as
it used to be the case with stream->pending_events).
The patch is tagged MEDIUM because this required to update the task's
process() prototype to use an int instead of a short, that's quite a
bunch of places.
Refactoring performed with the following Coccinelle patch:
@@
char *s;
@@
(
- ist2(s, strlen(s))
+ ist(s)
|
- ist2(strdup(s), strlen(s))
+ ist(strdup(s))
)
Note that this replacement is safe even in the strdup() case, because `ist()`
will not call `strlen()` on a `NULL` pointer. Instead is inserts a length of
`0`, effectively resulting in `IST_NULL`.
The maximum number of connections accepted at once by a thread for a single
listener used to default to 64 divided by the number of processes but the
tasklet-based model is much more scalable and benefits from smaller values.
Experimentation has shown that 4 gives the highest accept rate for all
thread values, and that 3 and 5 come very close, as shown below (HTTP/1
connections forwarded per second at multi-accept 4 and 64):
ac\thr| 1 2 4 8 16
------+------------------------------
4| 80k 106k 168k 270k 336k
64| 63k 89k 145k 230k 274k
Some tests were also conducted on SSL and absolutely no change was observed.
The value was placed into a define because it used to be spread all over the
code.
It might be useful at some point to backport this to 2.3 and 2.2 to help
those who observed some performance regressions from 1.6.
move listen status to a helper, defining both status enum and string
definition.
this will be helpful to be reused in prometheus code. It also removes
this hard-to-read nested ternary.
Signed-off-by: William Dauchy <wdauchy@gmail.com>
This reverts commit 62e8aaa1bd5ca96089eaa88487c700c4af4617f4.
While is works extremely well to address SSL handshake floods, it prevents
establishment of new connections during regular traffic above 50-60 Gbps,
because for an unknown reason the queue seems to have ~1.7 active tasks
per connection all the time, which makes no sense as these ought to be
waiting on subscribed events. It might uncover a deeper issue but at least
for now a different solution is needed. cf issue #822.
The test is trivial to run, just start a config with tune.runqueue-depth 10
and inject on 1GB objects with more than 10 connections. Try to connect to
the stats socket, it only works once, then the listeners are not dequeued.
In github issue #822, user @ngaugler reported some performance problems when
dealing with many concurrent SSL connections on restarts, after migrating
from 1.6 to 2.2, indicating a long time required to re-establish connections.
The Run_queue metric in the traces showed an abnormally high number of tasks
in the run queue, likely indicating we were accepting faster than we could
process. And this is indeed one of the differences between 1.6 and 2.2, the
accept I/O loop and the TLS handshakes are totally independent, so much that
they can even run on different threads. In 1.6 the SSL handshake was handled
almost immediately after the accept(), so this was limiting the input rate.
With large maxconn values, as long as there are incoming connections, new
I/Os are scheduled and many of them pass before the handshake, being tagged
for low latency processing.
The result is that handshakes get postponed, and are further postponed as
new connections are accepted. When they are finally able to be processed,
some of them fail as the client is gone, and the client had already queued
new ones. This causes an excess number of apparent connections and total
number of handshakes to be processed, just because we were accepting
connections on a temporarily saturated machine.
The solution is to temporarily pause new incoming connections when the
load already indicates that more tasks are already queued than will be
handled in a poll loop. The difficulty with this usually is to be able
to come back to re-enable the operation, but given that the metric is
the run queue, we just have to queue the global_listener_queue task so
that it gets picked by any thread once the run queues get flushed.
Before this patch, injecting with SSL reneg with 10000 concurrent
connections resulted in 350k tasks in the run queue, and a majority of
handshake timeouts noticed by the client. With the patch, the run queue
fluctuates between 1-3x runqueue-depth, the process is constantly busy, the
accept rate is maximized and clients observe no error anymore.
It would be desirable to backport this patch to 2.3 and 2.2 after some more
testing, provided the accept loop there is compatible.
With the removal of the family-specific port setting, all protocol had
exactly the same implementation of ->add(). A generic one was created
with the name "default_add_listener" so that all other ones can now be
removed. The API was slightly adjusted so that the protocol and the
listener are passed instead of the listener and the port.
Note that all protocols continue to provide this ->add() method instead
of routinely calling default_add_listener() from create_listeners(). This
makes sure that any non-standard protocol will still be able to intercept
the listener addition if needed.
This could be backported to 2.3 along with the few previous patches on
listners as a pure code cleanup.
In create_listeners() we iterate over a port range and call the
protocol's ->add() function to add a new listener on the specified
port. Only tcp4/tcp6/udp4/udp6 support a port, the other ones ignore
it. Now that we can rely on the address family to properly set the
port, better do it this way directly from create_listeners() and
remove the family-specific case from the protocol layer.
Define a per-thread counters allocated with the greatest size of any
stat module counters. This variable is named trash_counters.
When using a proxy without allocated counters, return the trash counters
from EXTRA_COUNTERS_GET instead of a dangling pointer to prevent
segfault.
This is useful for all the proxies used internally and not
belonging to the global proxy list. As these objects does not appears on
the stat report, it does not matter to use the dummy counters.
For this fix to be functional, the extra counters are explicitly
initialized to NULL on proxy/server/listener init functions.
Most notably, the crash has already been detected with the following
vtc:
- reg-tests/lua/txn_get_priv.vtc
- reg-tests/peers/tls_basic_sync.vtc
- reg-tests/peers/tls_basic_sync_wo_stkt_backend.vtc
There is probably other parts that may be impacted (SPOE for example).
This bug was introduced in the current release and do not need to be
backported. The faulty commits are
"MINOR: ssl: count client hello for stats" and
"MINOR: ssl: add counters for ssl sessions".
The test on listener->state == LI_LISTEN is not sufficient to decide
if we need to enable a listener. Indeed, there is a very special case
which is the inherited FD shared, which has to reflect the real socket
state even after the previous test, and as such needs to remain in
LI_LISTEN state. In this case we don't want a worker to start the
master's listener nor conversely. Let's add a specific test for this.
This is an anticipation of finer grained locking for the queues. For now
all lock places take a write lock so that there is no difference at all
with previous code.
safer to close handle before the object is put back in the global pool.
this was introduced by commit 9378bbe0bef4005155d ("MEDIUM: listener:
use protocol->accept_conn() to accept a connection")
this should fix github issue #902
no backport needed.
Signed-off-by: William Dauchy <wdauchy@gmail.com>
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.
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.
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.
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.
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.
