There are multiple per-thread lists in the listeners, which isn't the
most efficient in terms of cache, and doesn't easily allow to store all
the per-thread stuff.
Now we introduce an srv_per_thread structure which the servers will have an
array of, and place the idle/safe/avail conns tree heads into. Overall this
was a fairly mechanical change, and the array is now always initialized for
all servers since we'll put more stuff there. It's worth noting that the Lua
code still has to deal with its own deinit by itself despite being in a
global list, because its server is not dynamically allocated.
The PRNG used by the "random" LB algorithm was the central one which tries
hard to produce "correct" (i.e. hardly predictable) values suitable for use
in UUIDs or cookies. It's much too expensive for pure load balancing where
a cheaper thread-local PRNG is sufficient, and the current PRNG is part of
the hot places when running with many threads.
Let's switch to the stastistical PRNG instead, it's thread-local, very
fast, and with a period of (2^32)-1 which is more than enough to decide
on a server.
In conn_backend_get() we can cause some extreme contention due to the
idle_conns_lock. Indeed, even though it's per-thread, it still causes
high contention when running with many threads. The reason is that all
threads which do not have any idle connections are quickly skipped,
till the point where there are still some, so the first reaching that
point will grab the lock and the other ones wait behind. From this
point, all threads are synchronized waiting on the same lock, and
will follow the leader in small jumps, all hindering each other.
Here instead of doing this we're using a trylock. This way when a thread
is already checking a list, other ones will continue to next thread. In
the worst case, a high contention will lead to a few new connections to
be set up, but this may actually be what is required to avoid contention
in the first place. With this change, the contention has mostly
disappeared on this lock (it's still present in muxes and transport
layers due to the takeover).
Surprisingly, checking for emptiness of the tree root before taking
the lock didn't address any contention.
A few improvements are still possible and desirable here. The first
one would be to avoid seeing all threads jump to the next one. We
could have each thread use a different prime number as the increment
so as to spread them across the entire table instead of keeping them
synchronized. The second one is that the lock in the muck layers
shouldn't be needed to check for the tasklet's context availability.
If dispatch mode or transparent backend is used, the backend connection
target is a proxy instead of a server. In these cases, the reuse of
backend connections is not consistent.
With the default behavior, no reuse is done and every new request uses a
new connection. However, if http-reuse is set to never, the connection
are stored by the mux in the session and can be reused for future
requests in the same session.
As no server is used for these connections, no reuse can be made outside
of the session, similarly to http-reuse never mode. A different
http-reuse config value should not have an impact. To achieve this, mark
these connections as private to have a defined behavior.
For this feature to properly work, the connection hash has been slightly
adjusted. The server pointer as an input as been replaced by a generic
target pointer to refer to the server or proxy instance. The hash is
always calculated on connect_server even if the connection target is not
a server. This also requires to allocate the connection hash node for
every backend connections, not just the one with a server target.
Fix a leak in connect_server which happens when a connection is reused
and a bind_addr was allocated because transparent mode is active. The
connection has already an allocated bind_addr so free the newly
allocated one.
No backport needed.
When the selected server has no address, the destination address of the
client is used. However, for now, only the address is set, not the
family. Thus depending on how the server is configured and the client's
destination address, the server address family may be wrong.
For instance, with such server :
server srv 0.0.0.0:0
The server address family is AF_INET. The server connection will fail if a
client is asking for an IPv6 destination.
To fix the bug, we take care to set the rigth family, the family of the
client destination address.
This patch should fix the issue #202. It must be backported to all stable
versions.
Remove ebmb_node entry from struct connection and create a dedicated
struct conn_hash_node. struct connection contains now only a pointer to
a conn_hash_node, allocated only for connections where target is of type
OBJ_TYPE_SERVER. This will reduce memory footprints for every
connections that does not need http-reuse such as frontend connections.
The two algos defining these functions (first and leastconn) do not need the
server's lock. However it's already present in pendconn_process_next_strm()
so the API must be updated so that the functions may take it if needed and
that the callers indicate whether they already own it.
As such, the call places (backend.c and stream.c) now do not take it
anymore, queue.c was unchanged since it's already held, and both "first"
and "leastconn" were updated to take it if not already held.
A quick test on the "first" algo showed a jump from 432 to 565k rps by
just dropping the lock in stream.c!
conn_hash_prehash does not need a nul-terminated string, thus it is only
needed to test if the sni sample is not null before using it as
connection hash input.
Moreover, a bug could be introduced between smp_make_safe and
ssl_sock_set_servername call. Indeed, smp_make_safe may call smp_dup
which duplicates the sample in the trash buffer. If another function
manipulates the trash buffer before the call to ssl_sock_set_servername,
the sni sample might be erased. Currently, no function seems to do that
except make_proxy_line in case proxy protocol is used simultaneously
with the sni on the server.
This does not need to be backported.
Use the proxy protocol frame if proxy protocol is activated on the
server line. Do not add anymore these connections in the private list.
If some requests are made with the same proxy fields, they can reuse
the idle connection.
The reg-tests proxy_protocol_send_unique_id must be adapted has it
relied on the side effect behavior that every requests from a same
connection reused a private server connection. Now, a new connection is
created as expected if the proxy protocol fields differ.
The source address is used as an input to the the server connection hash. The
address and port are used as separate hash inputs. Do not add anymore these
connections in the private list.
This parameter is set only if used in the transparent-proxy mode.
The destination address is used as an input to the server connection hash. The
address and port are used as separated hash inputs. Note that they are not used
when statically specified on the server line. This is only useful for dynamic
destination address.
This is typically used when the server address is dynamically set via the
set-dst action. The address and port are separated hash parameters.
Most notably, it should fixed set-dst use case (cf github issue #947).
Change the API of the function used to allocate the stream target
address. This is done in order to be able to allocate the destination
address and use it to reuse a connection sharing with the same address.
In particular, the flag stream SF_ADDR_SET is now set outside of the
function.
The sni parameter is an input to the server connection hash. Do not add
anymore connections with dynamic sni in the private list. Thus, it is
now possible to reuse a server connection if they use the same sni.
Compare the connection hash when reusing a connection from the session.
This ensures that a private connection is reused only if it shares the
same set of parameters.
The pointer of the target server is used as a first parameter for the
server connection hash calcul. This prevents the hash to be null when no
specific parameters are present, and can serve as a simple defense
against an attacker trying to reuse a non-conform connection.
This is a preliminary work for the calcul of the backend connection
hash. A structure conn_hash_params is the input for the operation,
containing the various specific parameters of a connection.
The high bits of the hash will reflect the parameters present as input.
A set of macros is written to manipulate the connection hash and extract
the parameters/payload.
With http-reuse always, if no matching safe connection is found, check
in idle tree for a matching one. This is needed because now idle
connections can be differentiated from each other.
If only the safe tree was checked because not empty, but did not contain
a matching connection, we could miss matching entry in idle tree.
If no matching connection is found on available, check on idle/safe
trees for a matching one. This is needed because now idle connections
can be differentiated from each other.
If only the available list was checked because not empty, but did not
contain a matching connection, we could miss matching entries in idle or
safe trees.
The server idle/safe/available connection lists are replaced with ebmb-
trees. This is used to store backend connections, with the new field
connection hash as the key. The hash is a 8-bytes size field, used to
reflect specific connection parameters.
This is a preliminary work to be able to reuse connection with SNI,
explicit src/dst address or PROXY protocol.
This is a preparation work for connection reuse with sni/proxy
protocol/specific src-dst addresses.
Protect every access to idle conn lists with a lock. This is currently
strictly not needed because the access to the list are made with atomic
operations. However, to be able to reuse connection with specific
parameters, the list storage will be converted to eb-trees. As this
structure does not have atomic operation, it is mandatory to protect it
with a lock.
For this, the takeover lock is reused. Its role was to protect during
connection takeover. As it is now extended to general idle conns usage,
it is renamed to idle_conns_lock. A new lock section is also
instantiated named IDLE_CONNS_LOCK to isolate its impact on performance.
The wrong lock seems to be held when trying to remove another thread
connection if max fd limit has been reached (locking the current thread
instead of the target thread lock).
This could be backported up to 2.0.
Do not consider reuse connection if available list is not allocated for
the target server. This will prevent a crash when using a standalone
server for an external purpose like socket_tcp/socket_ssl on hlua code.
For the idle/safe lists, they are considered allocated if
srv.max_idle_conns is not null.
Note that the hlua code is currently safe thanks to the additional
checks on proxy http mode and stream reuse policy not never. However,
this might not be sufficient for future code.
This patch should be backported in every branches containing the
following patch :
7f68d815af (2.4 tree)
REORG: backend: simplify conn_backend_get
The reuse of idle connections should only happen for a proxy with the
http mode. In case of a backend with the tcp mode, the reuse selection
and insertion in session list are skipped.
This behavior is present since commit :
MEDIUM: connection: Add private connections synchronously in session server list
It could also be further exagerated by :
MEDIUM: backend: add reused conn to sess if mux marked as HOL blocking
It can be backported up to 2.3.
Reorganize the conditions for the reuse of idle/safe connections :
- reduce code by using variable to store reuse mode and idle/safe conns
counts
- consider that idle/safe/avail lists are properly allocated if
max_idle_conns not null. An allocation failure prevents haproxy
startup.
This is from the output of codespell. It's done at once over a bunch
of files and only affects comments, so there is nothing user-visible.
No backport needed.
At a few places we used to rely on conn->owner to retrieve the session
while the session is already known. This is not correct because at some
of these points the reason the connection's owner was still the session
(instead of NULL) is a mistake. At one place a comparison is even made
between the session and conn->owner assuming it's valid without checking
if it's NULL. Let's clean this up to use the session all the time.
Note that this will be needed for a forthcoming fix and will have to be
backported.
Leastconn has the nice propery of being able to sort servers by their
current usage. It's really a shame to force all requests into the backend
queue when the algo would be able to also consider their current queue.
In order not to change existing behavior but extend it, this patch allows
leastconn to elect servers which are already full if they have an explicitly
configured maxqueue setting above zero and their queue hasn't reached that
threshold. This will significantly reduce the pressure in the backend queue
when queuing a lot with lots of servers.
A test on 8 threads with 100 servers configured with maxconn 1 jumped
from 165krps to 330krps with maxqueue 15 with this patch.
This partially undoes commit 82cd5c13a ("OPTIM: backend: skip LB when we
know the backend is full") but allows to scale much better even by setting
a single-digit maxqueue value. Some better heuristics could be used to
maintain the behavior of the bypass in the patch above, consisting in
keeping it if it's known that there is no server with a configured
maxqueue in the farm (or in the backend).
The server lock must be held when server_take_conn() and server_drop_conn()
lbprm callback functions are called. It is a documented prerequisite but it is
not always performed. It only affects leastconn and fas lb algorithm. Others
don't use these callback functions.
A race condition on the next pending effecive weight (next_eweight) may be
encountered with the leastconn lb algorithm. An agent check may set it to 0
while fwlc_srv_reposition() is called. The server is locked during the
next_eweight update. But because the server lock is not acquired when
fwlc_srv_reposition() is called, we may use it to recompute the server key,
leading to a division by 0.
This patch must be backported as far as 1.8.
Detect if the sni used a constant value and if so, allow to reuse this
connection for later sessions. Use a combination of SMP_USE_INTRN +
!SMP_F_VOLATILE to consider a sample as a constant value.
This features has been requested on github issue #371.
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.
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 3d52f0f1f8 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.
For some algos (roundrobin, static-rr, leastconn, first) we know that
if there is any request queued in the backend, it's because a previous
attempt failed at finding a suitable server after trying all of them.
This alone is sufficient to decide that the next request will skip the
LB algo and directly reach the backend's queue. Doing this alone avoids
an O(N) lookup when load-balancing on a saturated farm of N servers,
which starts to be very expensive for hundreds of servers, especially
under the lbprm lock. This change alone has increased the request rate
from 110k to 148k RPS for 200 saturated servers on 8 threads, and
fwlc_reposition_srv() doesn't show up anymore in perf top. See github
issue #880 for more context.
It could have been the same for random, except that random is performed
using a consistent hash and it only considers a small set of servers (2
by default), so it may result in queueing at the backend despite having
some free slots on unknown servers. It's no big deal though since random()
only performs two attempts by default.
For hashing algorithms this is pointless since we don't queue at the
backend, except when there's no hash key found, which is the least of
our concerns here.
If random() returns a server whose maxconn is reached or the queue is
used, instead of adding the request to the server's queue, better add
it to the backend queue so that it can be served by any server (hence
the fastest one).
Since we've fixed the way URIs are handled in 2.1, some users have started
to experience inconsistencies in "balance uri" between requests received
over H1 and the same ones received over H2. This is caused by the fact
that H1 rarely uses absolute URIs while H2 always uses them. Similar
issues were reported already around replace-uri etc, leading to "pathq"
recently being introduced, so this isn't new.
Here what this patch does is add a new option to "balance uri" to indicate
that the hashing should only start at the path and not cover the authority.
This makes H1 relative URIs and H2 absolute URI hashes equally again.
Some extra options could be added to normalize URIs by always hashing the
authority (or host) in front of them, which would make sure that both
absolute and relative requests provide the same hash. This is left for
later if needed.
In connect_server(), we can enter in a stupid situation:
- conn_install_mux_be() is called to install the mux. This one
subscribes for receiving and quits ;
- then we discover that a handshake is required on the connection
(e.g. send-proxy), so xprt_add_hs() is called and subscribes as
well.
- we crash in conn_subscribe() which gets a different subscriber.
And if BUG_ON is disabled, we'd likely lose one event.
Note that it doesn't seem to happen by default, but definitely does
if connect() rightfully performs fd_cant_recv(), so it's a matter of
who does what and in what order.
A simple reproducer consists in adding fd_cant_recv() after fd_cant_send()
in tcp_connect_server() and running it on this config, as discussed in issue
listen foo
bind :8181
mode http
server srv1 127.0.0.1:8888 send-proxy-v2
The root cause is that xprt_add_hs() installs an xprt layer underneath
the mux without taking over its subscriptions. Ideally if we want to
support this, we'd need to steal the connection's wait_events and
replace them by new ones. But there doesn't seem to be any case where
we're interested in doing this so better simply always install the
transport layer before installing the mux, that's safer and simpler.
This needs to be backported to 2.1 which is constructed the same way
and thus suffers from the same issue, though the code is slightly
different there.
In the connect_server() function, there is an optim to install the mux as soon
as possible. It is possible if we can determine the mux to use from the
configuration only. For instance if the mux is explicitly specified or if no ALPN
is set. This patch adds a new condition to preinstall the mux for non-ssl
connection. In this case, by default, we always use the mux_pt for raw
connections and the mux-h1 for HTTP ones.
This patch is related to the issue #762. It may be backported to 2.2 (and
possibly as far as 1.9 if necessary).
