Well, that's becoming embarrassing. Now this fixes commit 4ef6801c
("BUG/MEDIUM: list: correct fix for LIST_POP_LOCKED's removal of last
element") which itself tried to fix commit 285192564. This fix only
works under low contention and was tested with the listener's queue.
With the idle conns it's obvious that it's still wrong since adding
more than one element to the list leaves a LLIST_BUSY pointer into
the list's head. This was visible when accumulating idle connections
in a server's list.
This new version of the fix almost goes back to the original code,
except that since then we addressed issues with expectedly idempotent
operations that were not. Now the code has been verified on paper again
and has survived 300 million connections spread over 4 threads.
This will have to be backported if the commit above is backported.
Let's keep a copy of these initial values. They will be useful to
compute automatic maxconn, as well as to restore proper limits when
doing an execve() on external checks.
This patch implements peer heartbeat feature to prevent any haproxy peer
from reconnecting too often, consuming sockets for nothing.
To do so, we add PEER_MSG_CTRL_HEARTBEAT new message to PEER_MSG_CLASS_CONTROL peers
control class of messages. A ->heartbeat field is added to peer structs
to store the heatbeat timeout value which is handled by the same function as for ->reconnect
to control the session timeouts. A 2-bytes heartbeat message is sent every 3s when
no updates have to be sent. This way, the peer which receives such a message is sure
the remote peer is still alive. So, it resets the ->reconnect peer session
timeout to its initial value (5s). This prevents any reconnection to an
already connected alive peer.
Historically the default frontend's maxconn used to be quite low (2000),
which was sufficient two decades ago but often proved to be a problem
when users had purposely set the global maxconn value but forgot to set
the frontend's.
There is no point in keeping this arbitrary limit for frontends : when
the global maxconn is lower, it's already too high and when the global
maxconn is much higher, it becomes a limiting factor which causes trouble
in production.
This commit allows the value to be set to zero, which becomes the new
default value, to mean it's not directly limited, or in fact it's set
to the global maxconn. Since this operation used to be performed before
computing a possibly automatic global maxconn based on memory limits,
the calculation of the maxconn value and its propagation to the backends'
fullconn has now moved to a dedicated function, proxy_adjust_all_maxconn(),
which is called once the global maxconn is stabilized.
This comes with two benefits :
1) a configuration missing "maxconn" in the defaults section will not
limit itself to a magically hardcoded value but will scale up to the
global maxconn ;
2) when the global maxconn is not set and memory limits are used instead,
the frontends' maxconn automatically adapts, and the backends' fullconn
as well.
In an attempt to try to provide automatic maxconn settings, we need to
decorrelate a listner's backlog and maxconn so that these values can be
independent. This introduces a listener_backlog() function which retrieves
the backlog value from the listener's backlog, the frontend's, the
listener's maxconn, the frontend's or falls back to 1024. This
corresponds to what was done in cfgparse.c to force a value there except
the last fallback which was not set since the frontend's maxconn is always
known.
As seen with Olivier, in the end the fix in commit 285192564 ("BUG/MEDIUM:
list: fix LIST_POP_LOCKED's removal of the last pointer") is wrong,
the code there was right but the bug was triggered by another bug in
LIST_ADDQ_LOCKED() which doesn't properly update the list's head by
inserting in the wrong order.
This will have to be backported if the commit above is backported.
There is a very difficult to reproduce race in the listener's accept
code, which is much easier to reproduce once connection limits are
properly enforced. It's an ABBA lock issue :
- the following functions take l->lock then lq_lock :
disable_listener, pause_listener, listener_full, limit_listener,
do_unbind_listener
- the following ones take lq_lock then l->lock :
resume_listener, dequeue_all_listener
This is because __resume_listener() only takes the listener's lock
and expects to be called with lq_lock held. The problem can easily
happen when listener_full() and limit_listener() are called a lot
while in parallel another thread releases sessions for the same
listener using listener_release() which in turn calls resume_listener().
This scenario is more prevalent in 2.0-dev since the removal of the
accept lock in listener_accept(). However in 1.9 and before, a different
but extremely unlikely scenario can happen :
thread1 thread2
............................ enter listener_accept()
limit_listener()
............................ long pause before taking the lock
session_free()
dequeue_all_listeners()
lock(lq_lock) [1]
............................ try_lock(l->lock) [2]
__resume_listener()
spin_lock(l->lock) =>WAIT[2]
............................ accept()
l->accept()
nbconn==maxconn =>
listener_full()
state==LI_LIMITED =>
lock(lq_lock) =>DEADLOCK[1]!
In practice it is almost impossible to trigger it because it requires
to limit both on the listener's maxconn and the frontend's rate limit,
at the same time, and to release the listener when the connection rate
goes below the limit between poll() returns the FD and the lock is
taken (a few nanoseconds). But maybe with threads competing on the
same core it has more chances to appear.
This patch removes the lq_lock and replaces it with a lockless queue
for the listener's wait queue (well, technically speaking a self-locked
queue) brought by commit a8434ec14 ("MINOR: lists: Implement locked
variations.") and its few subsequent fixes. This relieves us from the
need of the lq_lock and removes the deadlock. It also gets rid of the
distinction between __resume_listener() and resume_listener() since the
only difference was the lq_lock. All listener removals from the list
are now unconditional to avoid races on the state. It's worth noting
that the list used to never be initialized and that it used to work
only thanks to the state tests, so the initialization has now been
added.
This patch must carefully be backported to 1.9 and very likely 1.8.
It is mandatory to be careful about replacing all manipulations of
l->wait_queue, global.listener_queue and p->listener_queue.
Since LIST_DEL_LOCKED() and LIST_POP_LOCKED() now automatically reinitialize
the removed element, there's no need for keeping this LIST_INIT() call in the
idle connection code.
These operations previously used to return a "locked" element, which is
a constraint when multiple threads try to delete the same element, because
the second one will block indefinitely. Instead, let's make sure that both
LIST_DEL_LOCKED() and LIST_POP_LOCKED() always reinitialize the element
after deleting it. This ensures that the second thread will immediately
unblock and succeed with the removal. It also secures the pop vs delete
competition that may happen when trying to remove an element that's about
to be dequeued.
Commit a8434ec14 ("MINOR: lists: Implement locked variations.")
introduced locked lists which use the elements pointers as locks
for concurrent operations. Under heavy stress the lists occasionally
fail. The cause is a missing barrier at some points when updating
the list element and the head : nothing prevents the compiler (or
CPU) from updating the list head first before updating the element,
making another thread jump to a wrong location. This patch simply
adds the missing barriers before these two opeations.
This will have to be backported if the commit above is backported.
There was a typo making the last updated pointer be the pre-last element's
prev instead of the last's prev element. It didn't show up during early
tests because the contention is very rare on this one and it's implicitly
recovered when updating the pointers to go to the next element, but it was
clearly visible in the listener_accept() tests by having all threads block
on LIST_POP_LOCKED() with n==p==LLIST_BUSY.
This will have to be backported if commit a8434ec14 ("MINOR: lists:
Implement locked variations.") is backported.
Commit a8434ec14 ("MINOR: lists: Implement locked variations.")
introduced locked lists which use the elements pointers as locks
for concurrent operations. A copy-paste typo in LIST_ADDQ_LOCKED()
causes corruption in the list in case the next pointer is already
held, as it restores the previous pointer into the next one. It
may impact the server pools.
This will have to be backported if the commit above is backported.
Threads have long matured by now, still for most users their usage is
not trivial. It's about time to enable them by default on platforms
where we know the number of CPUs bound. This patch does this, it counts
the number of CPUs the process is bound to upon startup, and enables as
many threads by default. Of course, "nbthread" still overrides this, but
if it's not set the default behaviour is to start one thread per CPU.
The default number of threads is reported in "haproxy -vv". Simply using
"taskset -c" is now enough to adjust this number of threads so that there
is no more need for playing with cpu-map. And thanks to the previous
patches on the listener, the vast majority of configurations will not
need to duplicate "bind" lines with the "process x/y" statement anymore
either, so a simple config will automatically adapt to the number of
processors available.
tune.listener.multi-queue { on | off }
Enables ('on') or disables ('off') the listener's multi-queue accept which
spreads the incoming traffic to all threads a "bind" line is allowed to run
on instead of taking them for itself. This provides a smoother traffic
distribution and scales much better, especially in environments where threads
may be unevenly loaded due to external activity (network interrupts colliding
with one thread for example). This option is enabled by default, but it may
be forcefully disabled for troubleshooting or for situations where it is
estimated that the operating system already provides a good enough
distribution and connections are extremely short-lived.
It's important to monitor the accept queues to know if some incoming
connections had to be handled by their originating thread due to an
overflow. It's also important to be able to confirm thread fairness.
This patch adds "accq_pushed" to activity reporting, which reports
the number of connections that were successfully pushed into each
thread's queue, and "accq_full", which indicates the number of
connections that couldn't be pushed because the thread's queue was
full.
There is one point where we can migrate a connection to another thread
without taking risk, it's when we accept it : the new FD is not yet in
the fd cache and no task was created yet. It's still possible to assign
it a different thread than the one which accepted the connection. The
only requirement for this is to have one accept queue per thread and
their respective processing tasks that have to be woken up each time
an entry is added to the queue.
This is a multiple-producer, single-consumer model. Entries are added
at the queue's tail and the processing task is woken up. The consumer
picks entries at the head and processes them in order. The accept queue
contains the fd, the source address, and the listener. Each entry of
the accept queue was rounded up to 64 bytes (one cache line) to avoid
cache aliasing because tests have shown that otherwise performance
suffers a lot (5%). A test has shown that it's important to have at
least 256 entries for the rings, as at 128 it's still possible to fill
them often at high loads on small thread counts.
The processing task does almost nothing except calling the listener's
accept() function and updating the global session and SSL rate counters
just like listener_accept() does on synchronous calls.
At this point the accept queue is implemented but not used.
In order to quickly pick a thread ID when accepting a connection, we'll
need to know certain pre-computed values derived from the thread mask,
which are counts of bits per position multiples of 1, 2, 4, 8, 16 and
32. In practice it is sufficient to compute only the 4 first ones and
store them in the bind_conf. We update the count every time the
bind_thread value is adjusted.
The fields in the bind_conf struct have been moved around a little bit
to make it easier to group all thread bit values into the same cache
line.
The function used to return a thread number is bind_map_thread_id(),
and it maps a number between 0 and 31/63 to a thread ID between 0 and
31/63, starting from the left.
Function mask_find_rank_bit() returns the bit position in mask <m> of
the nth bit set of rank <r>, between 0 and LONGBITS-1 included, starting
from the left. For example ranks 0,1,2,3 for mask 0x55 will be 6, 4, 2
and 0 respectively. This algorithm is based on a popcount variant and
is described here : https://graphics.stanford.edu/~seander/bithacks.html.
Now that nbproc and nbthread are exclusive, we can still provide more
detailed explanations about what we've found in the config when a bind
line appears on multiple threads and processes at the same time, then
ignore the setting.
This patch reduces the listener's thread mask to a single mask instead
of an array of masks per process. Now we have only one thread mask and
one process mask per bind-conf. This removes ~504 bytes of RAM per
bind-conf and will simplify handling of thread masks.
If a "bind" line only refers to process numbers not found by its parent
frontend or not covered by the global nbproc directive, or to a thread
not covered by the global nbthread directive, a warning is emitted saying
what will be used instead.
In LIST_DEL_LOCKED(), initialize p2 to NULL, and only attempt to set it back
to its previous value if we had a previous element, and thus p2 is non-NULL.
Instead of having one task per thread and per server that does clean the
idling connections, have only one global task for every servers.
That tasks parses all the servers that currently have idling connections,
and remove half of them, to put them in a per-thread list of connections
to kill. For each thread that does have connections to kill, wake a task
to do so, so that the cleaning will be done in the context of said thread.
Use the locked macros when manipulating idle_orphan_conns, so that other
threads can remove elements from it.
It will be useful later to avoid having a task per server and per thread to
cleanup the orphan list.
Implement LIST_ADD_LOCKED(), LIST_ADDQ_LOCKED(), LIST_DEL_LOCKED() and
LIST_POP_LOCKED().
LIST_ADD_LOCKED, LIST_ADDQ_LOCKED and LIST_DEL_LOCKED work the same as
LIST_ADD, LIST_ADDQ and LIST_DEL, except before any manipulation it locks
the relevant elements of the list, so it's safe to manipulate the list
with multiple threads.
LIST_POP_LOCKED() removes the first element from the list, and returns its
data.
This patch implements "req.ungrpc" sample fetch method to decode and
parse a gRPC request. It takes only one argument: a protocol buffers
field number to identify the protocol buffers message number to be looked up.
This argument is a sort of path in dotted notation to the terminal field number
to be retrieved.
ex:
req.ungrpc(1.2.3.4)
This sample fetch catch the data in raw mode, without interpreting them.
Some protocol buffers specific converters may be used to convert the data
to the correct type.
This function is useful to parse strings made of unsigned integers
and to allocate a C array of unsigned integers from there.
For instance this function allocates this array { 1, 2, 3, 4, } from
this string: "1.2.3.4".
It is the HTX version of co_skip(). Internally, It uses the function htx_drain().
It will be used by other commits to fix bugs, so it must be backported to 1.9.
The function htx_drain() can now be used to drain data from an HTX message.
It will be used by other commits to fix bugs, so it must be backported to 1.9.
in co_skip(), the flag CF_WRITE_PARTIAL is set on the channel. The flag
CF_WROTE_DATA must also be set to notify the channel some data were sent.
This patch must be backported to 1.9.
Calculate if the fd or task should be locked once, before locking, and
reuse the calculation when determing when to unlock.
Fixes a race condition added in 87d54a9a for fds, and b20aa9ee for tasks,
released in 1.9-dev4. When one thread modifies thread_mask to be a single
thread for a task or fd while a second thread has locked or is waiting on a
lock for that task or fd, the second thread will not unlock it. For FDs,
this is observable when a listener is polled by multiple threads, and is
closed while those threads have events pending. For tasks, this seems
possible, where task_set_affinity is called, but I did not observe it.
This must be backported to 1.9.
This is a naive implementation of closefrom() which closes all FDs
starting from the one passed in argument. closefrom() is not provided
on all operating systems, and other versions will follow.
Add a per-thread counter of idling connections, and use it to determine
how many connections we should kill after the timeout, instead of using
the global counter, or we're likely to just kill most of the connections.
This should be backported to 1.9.
Use atomic operations when dealing with srv->curr_idle_conns, as it's shared
between threads, otherwise we could get inconsistencies.
This should be backported to 1.9.
1xx responses does not work in HTTP2 when the HTX is enabled. First of all, when
a response is parsed, only one HEADERS frame is expected. So when an interim
response is received, the flag H2_SF_HEADERS_RCVD is set and the next HEADERS
frame (for another interim repsonse or the final one) is parsed as a trailers
one. Then when the response is sent, because an EOM block is found at the end of
the interim HTX response, the ES flag is added on the frame, closing too early
the stream. Here, it is a design problem of the HTX. Iterim responses are
considered as full messages, leading to some ambiguities when HTX messages are
processed. This will not be fixed now, but we need to keep it in mind for future
improvements.
To fix the parsing bug, the flag H2_MSGF_RSP_1XX is added when the response
headers are decoded. When this flag is set, an EOM block is added into the HTX
message, despite the fact that there is no ES flag on the frame. And we don't
set the flag H2_SF_HEADERS_RCVD on the corresponding H2S. So the next HEADERS
frame will not be parsed as a trailers one.
To fix the sending bug, the ES flag is not set on the frame when an interim
response is processed and the flag H2_SF_HEADERS_SENT is not set on the
corresponding H2S.
This patch must be backported to 1.9.
Initialize ->srv peer field for all the peers, the local peer included.
Indeed, a haproxy process needs to connect to the local peer of a remote
process. Furthermore, when a "peer" or "server" line is parsed by parse_server()
the address must be copied to ->addr field of the peer object only if this address
has been also parsed by parse_server(). This is not the case if this address belongs
to the local peer and is provided on a "server" line.
After having parsed the "peer" or "server" lines of a peer
sections, the ->srv part of all the peer must be initialized for SSL, if
enabled. Same thing for the binding part.
Revert 1417f0b commit which is no more required.
No backport is needed, this is purely 2.0.
The existing threading flag in the 51Degrees API
(FIFTYONEDEGREES_NO_THREADING) has now been mapped to the HAProxy
threading flag (USE_THREAD), and the 51Degrees module code has been made
thread safe.
In Pattern, the cache is now locked with a spin lock from hathreads.h
using a new lable 'OTHER_LOCK'. The workset pool is now created with the
same size as the number of threads to avoid any time waiting on a
worket.
In Hash Trie, the global device offsets structure is only used in single
threaded operation. Multi threaded operation creates a new offsets
structure in each thread.
Commit 1055e687a ("MINOR: peers: Make outgoing connection to SSL/TLS
peers work.") introduced an "srv" field in the peers, which points to
the equivalent server to hold SSL settings. This one is not set when
the peer is local so we must always test it before testing p->srv->use_ssl
otherwise haproxy dies during reloads.
No backport is needed, this is purely 2.0.
For some embedded systems, it's pointless to have 32- or even 64- large
arrays of processes when it's known that much fewer processes will be
used in the worst case. Let's introduce this MAX_PROCS define which
contains the highest number of processes allowed to run at once. It
still defaults to LONGBITS but may be lowered.
This also depends on the nbthread count, so it must only be performed after
parsing the whole config file. As a side effect, this removes some code
duplication between servers and server-templates.
This must be backported to 1.9.
These two functions return either all_{proc,threads}_mask, or the argument.
This is used to default to all_proc_mask or all_threads_mask when not set
on bind_conf or proxies.
We'll call popcount() more often so better use a parallel method
than an iterative one. One optimal design is proposed at the site
below. It requires a fast multiplication though, but even without
it will still be faster than the iterative one, and all relevant
64 bit platforms do have a multiply unit.
https://graphics.stanford.edu/~seander/bithacks.html
Some unused fields were placed early and some important ones were on
the second cache line. Let's move the proto_list and name closer to
the end of the structure to bring accept() and default_target() into
the first cache line.
When compiling with DEBUG_FAIL_ALLOC, add a new option, tune.fail-alloc,
that gives the percentage of chances an allocation fails.
This is useful to check that allocation failures are always handled
gracefully.
With variable connection limits, it's not possible to accurately determine
whether the mux is still in use by comparing usage and max to be equal due
to the fact that one determines the capacity and the other one takes care
of the context. This can cause some connections to be dropped before they
reach their stream ID limit.
It seems it could also cause some connections to be terminated with
streams still alive if the limit was reduced to match the newly computed
avail_streams() value, though this cannot yet happen with existing muxes.
Instead let's switch to usage reports and simply check whether connections
are both unused and available before adding them to the idle list.
This should be backported to 1.9.
