This part was inherited from haproxy-1.5. But since a while (at least 1.8),
the backend applet, once created, is no longer changed. Thus there is no
reason to still check if the target has changed. And in fact, if it was
still possible, there would be a memory leak because the old applet would be
lost and never released.
There is no reason to backport this fix because the leak only exists on a
dead code path.
si_register_applet() and si_applet_release() are renamed
cs_register_applet() and cs_applet_release() and now manipulate a
conn-stream instead of a stream-inteface.
si_shutr(), si_shutw(), si_chk_rcv() and si_chk_snd() are moved in the
conn-stream scope and renamed, respectively, cs_shutr(), cs_shutw(),
cs_chk_rcv(), cs_chk_snd() and manipulate a conn-stream instead of a
stream-interface.
To be able to move wait_event from the stream-interface to the conn-stream,
we must be prepare to handle errors when a mux is attached to a conn-stream.
Indeed, the wait_event's tasklet will be allocated when both a mux and a
stream will be both attached to a stream. So, we must be prepared to handle
allocation errors.
si_connect() is moved in backend.c and renamed as do_connect_server(). In
addition, the function now manipulate a stream instead of a
stream-interface.
The stream-interface state (SI_ST_*) is now in the conn-stream. It is a
mechanical replacement for now. Nothing special. SI_ST_* and SI_SB_* were
renamed accordingly. Utils functions to manipulate these infos were moved
under the conn-stream scope.
But it could be good to keep in mind that this part should be
reworked. Indeed, at the CS level, we only need to know if it is ready to
receive or to send. The state of conn-stream from INI to EST is only used on
the server side. The client CS is immediately set to EST. Thus current
SI_ST_* states should probably be moved to the stream to reflect the server
connection state during the establishment stage.
Only the server side is concerned by the stream-interface error type. It is
useless to have an err_type field on the client side. So, it is now move to
the stream. SI_ET_* are renames STRM_ET_* and moved in stream-t.h header
file.
Flag to get the source ip/port with getsockname is now handled at the stream
level. Thus SI_FL_SRC_ADDR stream-int flag is replaced by SF_SRC_ADDR stream
flag.
Flags to disable lingering and half-close are now handled at the conn-stream
level. Thus SI_FL_NOLINGER and SI_FL_NOHALF stream-int flags are replaced by
CS_FL_NOLINGER and CS_FL_NOHALF conn-stream flags.
Instead of relying on the conn-stream error, via CS_FL_ERR flags, we now
directly use the error at the endpoint level with the flag CS_EP_ERROR. It
should be safe to do so. But we must be careful because it is still possible
that an error is processed too early. Anyway, a conn-stream has always a
valid endpoint, maybe detached from any endpoint, but valid.
SI_FL_ERR is removed and replaced by CS_FL_ERROR. It is a transient patch
because the idea is to rely on the endpoint to handle errors at this
level. But if for any reason it is not possible, the stream-interface flags
will still be replaced.
The expiration date in the stream-interface was only used on the server side
to set the connect, queue or turn-around timeout. It was checked on the
frontend stream-interface, but never used concretely. So it was removed and
replaced by a connect expiration date in the stream itself. Thus, SI_FL_EXP
flag in stream-interfaces is replaced by a stream flag, SF_CONN_EXP.
The source and destination addresses at the applicative layer are moved from
the stream-interface to the conn-stream. This simplifies a bit the code and
it is a logicial step to remove the stream-interface.
The conn_retries counter was set to the max value and decremented at each
connection retry. Thus the counter reflected the number of retries left and
not the real number of retries. All calculations of redispatch or reporting
of number of retries experienced were made using subtracts from the
configured retries, which was complicated and didn't bring any benefit.
Now, this counter is set to 0 and incremented at each retry. We know we've
reached the maximum allowed connection retries by comparing it to the
configured value. In all other cases, we directly use the counter.
This patch should address the feature request #1608.
The conn_retries counter may be moved into the stream structure. It only
concerns the connection establishment. The frontend stream-interface does not
use it. So it is a logical change.
At many places, we now use the new CS functions to get a stream or a channel
from a conn-stream instead of using the stream-interface API. It is the
first step to reduce the scope of the stream-interfaces. The main change
here is about the applet I/O callback functions. Before the refactoring, the
stream-interface was the appctx owner. Thus, it was heavily used. Now, as
far as possible,the conn-stream is used. Of course, it remains many calls to
the stream-interface API.
All old flags CS_FL_* are now moved in the endpoint scope and renamed
CS_EP_* accordingly. It is a systematic replacement. There is no true change
except for the health-check and the endpoint reset. Here it is a bit special
because the same conn-stream is reused. Thus, we must handle endpoint
allocation errors. To do so, cs_reset_endp() has been adapted.
Thanks to this last change, it will now be possible to simplify the
multiplexer and probably the applets too. A review must also be performed to
remove some flags in the channel or the stream-interface. The HTX will
probably be simplified too. Finally, there is now some place in the
conn-stream to move info from the stream-interface.
The conn-stream endpoint is now shared between the conn-stream and the
applet or the multiplexer. If the mux or the applet is created first, it is
responsible to also create the endpoint and share it with the conn-stream.
If the conn-stream is created first, it is the opposite.
When the endpoint is only owned by an applet or a mux, it is called an
orphan endpoint (there is no conn-stream). When it is only owned by a
conn-stream, it is called a detached endpoint (there is no mux/applet).
The last entity that owns an endpoint is responsible to release it. When a
mux or an applet is detached from a conn-stream, the conn-stream
relinquishes the endpoint to recreate a new one. This way, the endpoint
state is never lost for the mux or the applet.
It is a transient commit to prepare next changes. Now, when a conn-stream is
created from an applet or a multiplexer, an endpoint is always provided. In
addition, the API to create a conn-stream was specialized to have one
function per type.
The next step will be to share the endpoint structure.
Some CS flags, only related to the endpoint, are moved into the endpoint
struct. More will probably moved later. Those ones are not critical. So it
is pretty safe to move them now and this will ease next changes.
This change is only significant for the multiplexer part. For the applets,
the context and the endpoint are the same. Thus, there is no much change. For
the multiplexer part, the connection was used to set the conn-stream
endpoint and the mux's stream was the context. But it is a bit strange
because once a mux is installed, it takes over the connection. In a
wonderful world, the connection should be totally hidden behind the mux. The
stream-interface and, in a lesser extent, the stream, still access the
connection because that was inherited from the pre-multiplexer era.
Now, the conn-stream endpoint is the mux's stream (an opaque entity for the
conn-stream) and the connection is the context. Dedicated functions have
been added to attached an applet or a mux to a conn-stream.
Thanks to all previous changes, it is now possible to move the
stream-interface into the conn-stream. To do so, some SI functions are
removed and their conn-stream counterparts are added. In addition, the
conn-stream is now responsible to create and release the
stream-interface. While the stream-interfaces were inlined in the stream
structure, there is now a pointer in the conn-stream. stream-interfaces are
now dynamically allocated. Thus a dedicated pool is added. It is a temporary
change because, at the end, the stream-interface structure will most
probably disappear.
To be able to move the stream-interface from the stream to the conn-stream,
all access to the SI is done via the conn-stream. This patch is limited to
the backend part.
frontend and backend conn-streams are now directly accesible from the
stream. This way, and with some other changes, it will be possible to remove
the stream-interfaces from the stream structure.
Thanks to previous changes, it is now possible to set an appctx as endpoint
for a conn-stream. This means the appctx is no longer linked to the
stream-interface but to the conn-stream. Thus, a pointer to the conn-stream
is explicitly stored in the stream-interface. The endpoint (connection or
appctx) can be retrieved via the conn-stream.
The backend conn-stream is no longer released on connection retry. This
means the conn-stream is detached from the underlying connection but not
released. Thus, during connection retries, the stream has always an
allocated conn-stream with no connection. All previous changes were made to
make this possible.
Note that .attach() mux callback function was changed to get the conn-stream
as argument. The muxes are no longer responsible to create the conn-stream
when a server connection is attached to a stream.
The conn-stream will progressively replace the stream-interface. Thus, a
stream will have to allocate the backend conn-stream during its
creation. This means it will be possible to have a conn-stream with no
connection. To prepare this change, we test the conn-stream's connection
when we retrieve it.
At many places we use construct such as:
if (objt_server(blah))
do_something(objt_server(blah));
At -O2 the compiler manages to simplify the operation and see that the
second one returns the same result as the first one. But at -O1 that's
not always the case, and the compiler is able to emit a second
expression and sees the potential null that results from it, and may
warn about a potential null deref (e.g. with gcc-6.5). There are two
solutions to this:
- either the result of the first test has to be passed to a local
variable
- or the second reference ought to be unchecked using the __objt_*
variant.
This patch fixes all occurrences at once by taking the second approach
(the least intrusive). For constructs like:
objt_server(blah) ? objt_server(blah)->name : "no name"
a macro could be useful. It would for example take the object type
(server), the field name (name) and the default value. But there
are probably not enough occurrences across the whole code for this
to really matter.
This should be backported wherever it applies.
In alloc_dst_address(), the client destination address must only be
retrieved when we are sure to use it. Most of time, this save a syscall to
getsockname(). It is not a bugfix in itself. But it revealed a bug in the
QUIC part. The CO_FL_ADDR_TO_SET flag is not set when the destination
address is create for anew quic client connection.
Handle properly websocket streams if the server uses an ALPN with both
h1 and h2. Add a new field h2_ws in the server structure. If set to off,
reuse is automatically disable on backend and ALPN is forced to http1.x
if possible. Nothing is done if on.
Implement a mechanism to be able to use a different http version for
websocket streams. A new server member <ws> represents the algorithm to
select the protocol. This can overrides the server <proto>
configuration. If the connection uses ALPN for proto selection, it is
updated for websocket streams to select the right protocol.
Three mode of selection are implemented :
- auto : use the same protocol between non-ws and ws streams. If ALPN is
use, try to update it to "http/1.1"; this is only done if the server
ALPN contains "http/1.1".
- h1 : use http/1.1
- h2 : use http/2.0; this requires the server to support RFC8441 or an
error will be returned by haproxy.
Add a new parameter force_mux_ops. This will be useful to specify an
alternative to the srv->mux_proto field. If non-NULL, it will be use to
force the mux protocol wether srv->mux_proto is set or not.
This argument will become useful to install a mux for non-standard
streams, most notably websocket streams.
At a few places we were still using protocol_by_family() instead of
the richer protocol_lookup(). The former is limited as it enforces
SOCK_STREAM and a stream protocol at the control layer. At least with
protocol_lookup() we don't have this limitationn. The values were still
set for now but later we can imagine making them configurable on the
fly.
Client source and destination addresses at stream level are used to initiate
the connections to a server. For now, stream-interface addresses are never
set. So, thanks to the fallback mechanism, no changes are expected with this
patch. But its purpose is to rely on addresses at the appropriate level when
set instead of those at the connection level.
Skip the hash connection calcul when reuse must not be used in
connect_server() : this is the case for TCP proxies. This should result
in slightly better performance when using this use-case.
In connect_server(), if http-reuse always is set, the backend connection
is inserted into the available tree as soon as created. However, the
hash connection field is only set later at the end of the function.
This seems to have no impact as the hash connection field is always
position before a lookup. However, this is not a proper usage of ebmb
API. Fix this by setting the hash connection field before the insertion
into the avail tree.
This must be backported up to 2.4.
Add traces in connect_server() to debug idle connection reuse. These
are attached to stream trace module, as it's already in use in
backend.c with the macro TRACE_SOURCE.
It is not yet used but thanks to this patch, it will be possible to resolve
arguments found in defaults sections. However, there is some restrictions:
* For FE (frontend) or BE (backend) arguments, if the proxy is explicity
defined, there is no change. But for implicit proxy (not specified), the
argument points on the default proxy. when a sample fetch using this
kind of argument is evaluated, the default proxy replaced by the current
one.
* For SRV (server) and TAB (stick-table)arguments, the proxy must always
be specified. Otherwise an error is reported.
This patch is mandatory to support TCP/HTTP rules in defaults sections.
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).
backend.c, all muxes, backend.c started manipulating ebmb_nodes with
the introduction of idle conns but the types were inherited through
other includes. Let's add ebmbtree.h there.
One was in backend.c and the other one in hlua.c. No other candidate
was found with "git grep '^#if\s*USE'". It's worth noting that 3
other such tests exist for SSL_OP_NO_{SSLv3,TLSv1_1,TLSv1_2} but
that these ones are properly set to 0 in openssl-compat.h when not
defined.
This option had always been broken in HTX, which means that the first
breakage appeared in 1.9, that it was broken by default in 2.0 and that
no workaround existed starting with 2.1. The way this option works is
praticularly unfit to the rest of the configuration and to the internal
architecture. It had some uses when it was introduced 14 years ago but
nowadays it's possible to do much better and more reliable using a
set of "http-request set-dst" and "http-request set-uri" rules, which
additionally are compatible with DNS resolution (via do-resolve) and
are not exclusive to normal load balancing. The "option-http_proxy"
example config file was updated to reflect this.
The option is still parsed so that an error message gives hints about
what to look for.
Replace http_get_path by the http_uri_parser API. The new functions is
renamed http_parse_path. Replace duplicated code for scheme and
authority parsing by invocations to http_parse_scheme/authority.
If no scheme is found for an URI detected as an absolute-uri/authority,
consider it to be an authority format : no path will be found. For an
absolute-uri or absolute-path, use the remaining of the string as the
path. A new http_uri_parser state is declared to mark the path parsing
as done.
This basically undoes the API changes that were performed by commit
0274286dd ("BUG/MAJOR: server: fix deadlock when changing maxconn via
agent-check") to address the deadlock issue: since process_srv_queue()
doesn't use the server lock anymore, it doesn't need the "server_locked"
argument, so let's get rid of it before it gets used again.
This reverts commit c83e45e9b0.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
In 1.4, consistent hashing was brought by commit 6b2e11be1 ("[MEDIUM]
backend: implement consistent hashing variation") which took care of
replacing all direct calls to map_get_server_rr() with an alternate
call to chash_get_next_server() if consistent hash was being used.
One of them, however, cannot happen because a preliminary test for
static round-robin is being done prior to the call, so we're certain
that if it matches it cannot use a consistent hash tree.
Let's remove it.
This basically undoes the API changes that were performed by commit
0274286dd ("BUG/MAJOR: server: fix deadlock when changing maxconn via
agent-check") to address the deadlock issue: since process_srv_queue()
doesn't use the server lock anymore, it doesn't need the "server_locked"
argument, so let's get rid of it before it gets used again.
This essentially reverts commit 2b4370078 ("MINOR: lb/api: let callers
of take_conn/drop_conn tell if they have the lock") that was merged
during 2.4 before the various locks could be eliminated at the lower
layers. Passing that information complicates the cleanup of the queuing
code and it's become useless.
The server_parse_maxconn_change_request locks the server lock. However,
this function can be called via agent-checks or lua code which already
lock it. This bug has been introduced by the following commit :
commit 79a88ba3d0
BUG/MAJOR: server: prevent deadlock when using 'set maxconn server'
This commit tried to fix another deadlock with can occur because
previoulsy server_parse_maxconn_change_request requires the server lock
to be held. However, it may call internally process_srv_queue which also
locks the server lock. The locking policy has thus been updated. The fix
is functional for the CLI 'set maxconn' but fails to address the
agent-check / lua counterparts.
This new issue is fixed in two steps :
- changes from the above commit have been reverted. This means that
server_parse_maxconn_change_request must again be called with the
server lock.
- to counter the deadlock fixed by the above commit, process_srv_queue
now takes an argument to render the server locking optional if the
caller already held it. This is only used by
server_parse_maxconn_change_request.
The above commit was subject to backport up to 1.8. Thus this commit
must be backported in every release where it is already present.
When reusing a backend connection, do not reapply the SNI on the
connection. It should already be defined when the connection was
instantiated on a previous connect_server invocation. As the SNI is a
parameter used to select a connection, only connection with same value
can be reused.
The impact of this bug is unknown and may be null. No memory leak has
been reported by valgrind. So this is more a cleaning fix.
This commit relies on the SF_SRV_REUSED flag and thus depends on the
following fix :
BUG/MINOR: backend: restore the SF_SRV_REUSED flag original purpose
This should be backported up to 2.4.
The SF_SRV_REUSED flag was set if a stream reused a backend connection.
One of its purpose is to count the total reuse on the backend in
opposition to newly instantiated connection.
However, the flag was diverted from its original purpose since the
following commit :
e8f5f5d8b2
BUG/MEDIUM: servers: Only set SF_SRV_REUSED if the connection if fully ready.
With this change, the flag is not set anymore if the mux is not ready
when a connection is picked for reuse. This can happen for multiplexed
connections which are inserted in the available list as soon as created
in http-reuse always mode. The goal of this change is to not retry
immediately this request in case on an error on the same server if the
reused connection is not fully ready.
This change is justified for the retry timeout handling but it breaks
other places which still uses the flag for its original purpose. Mainly,
in this case the wrong 'connect' backend counter is incremented instead
of the 'reuse' one. The flag is also used in http_return_srv_error and
may have an impact if a http server error is replied for this stream.
To fix this problem, the original purpose of the flag is restored by
setting it unconditionaly when a connection is reused. Additionally, a
new flag SF_SRV_REUSED_ANTICIPATED is created. This flag is set when the
connection is reused but the mux is not ready yet. For the timeout
handling on error, the request is retried immediately only if the stream
reused a connection without this newly anticipated flag.
This must be backported up to 2.1.
In 2.3, a significant improvement was brought against situations where
the queue was heavily used, because some LB algos were still checked
for no reason before deciding to put the request into the queue. This
was commit 82cd5c13a ("OPTIM: backend: skip LB when we know the backend
is full").
As seen in previous commit ("BUG/MAJOR: queue: set SF_ASSIGNED when
setting strm->target on dequeue") the dequeuing code is extremely
tricky, and the optimization above tends to emphasize transient issues
by making them permanent until the next reload, which is not acceptable
as the code must always be robust against any bad situation.
This commit brings a protection against such a situation by slightly
relaxing the test. Instead of checking that there are pending connections
in the backend queue, it also verifies that the backend's connections are
not solely composed of queued connections, which would then indicate we
are in this situation. This is not rocket science, but at least if the
situation happens, we know that it will unlock by itself once the streams
have left, as new requests will be allowed to reach the servers and to
flush the queue again.
This needs to be backported to 2.4 and 2.3.
Thanks to the previous patch (822decfd "BUG/MAJOR: stream-int: Release SI
endpoint on server side ASAP on retry"), it is now useless to release any
existing connection in connect_server() because it was already done in
back_handle_st_cer() if necessary.
This patch is not a CLEANUP because it may introduce some bugs in edge
cases. There is no reason to backport it for now except if it is required to
fix a bug.
When a connection attempt failed, if a retry is possible, the SI endpoint on
the server side is immediately released, instead of waiting to establish a
new connection to a server. Thus, when the backend SI is switched from
SI_ST_CER state to SI_ST_REQ, SI_ST_ASS or SI_ST_TAR, its endpoint is
released. It is expected because the SI is moved to a state prior to the
connection stage ( < SI_ST_CONN). So it seems logical to not have any server
connection.
It is especially important if the retry is delayed (SI_ST_TAR or
SI_ST_QUE). Because, if the server connection is preserved, any error at the
connection level is unexpectedly relayed to the stream, via the
stream-interface, leading to an infinite loop in process_stream(). if
SI_FL_ERR flag is set on the backend SI in another state than SI_ST_CLO, an
internal goto is performed to resync the stream-interfaces. In addtition,
some ressources are not released ASAP.
This bug is quite old and was reported 1 or 2 times per years since the 2.2
(at least) with not enough information to catch it. It must be backported as
far as 2.2 with a special care because this part has moved several times and
after some observation period and feedback from users to be sure. For info,
in 2.0 and prior, the connection is released when an error is encountered in
SI_ST_CON or SI_ST_RDY states.
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.
Currently our atomic ops return a value but it's never known whether
the fetch is done before or after the operation, which causes some
confusion each time the value is desired. Let's create an explicit
variant of these operations suffixed with _FETCH to explicitly mention
that the fetch occurs after the operation, and make use of it at the
few call places.
We now use the stream instead of the proxy to know if we are processing HTTP
data or not. If the stream is an HTX stream, it means we are dealing with
HTTP data. It is more accurate than the proxy mode because when an HTTP
upgrade is performed, the proxy is not changed and only the stream may be
used.
Note that it was not a problem to rely on the proxy because HTTP upgrades
may only happen when an HTTP backend was set. But, we will add the support
of HTTP upgrades on the frontend side, after te tcp-request rules
evaluation. In this context, we cannot rely on the proxy mode.
Commit b1adf03df ("MEDIUM: backend: use a trylock when trying to grab an
idle connection") solved a contention issue on the backend under normal
condition, but there is another one further, which only happens when the
number of FDs in use is considered too high, and which obviously causes
random crashes with just 16 threads once the number of FDs is about to be
exhausted.
Like the aforementioned patch, this one should be backported to 2.3.
Release the lock before calling mux destroy in connect_server when
trying to kill an idle connection because the pool high count has been
reached.
The lock must be released because the mux destroy will call
srv_release_conn which also takes the lock to remove the connection from
the tree. As the connection was already deleted from the tree at this
stage, it is safe to release the lock, and the removal in
srv_release_conn will be a noop.
It does not need to be backported because it is only present in the
current release. It has been introduced by
5c7086f6b0
MEDIUM: connection: protect idle conn lists with locks
Introduce a new XPRT method, start(). The init() method will now only
initialize whatever is needed for the XPRT to run, but any action the XPRT
has to do before being ready, such as handshakes, will be done in the new
start() method. That way, we will be sure the full stack of xprt will be
initialized before attempting to do anything.
The init() call is also moved to conn_prepare(). There's no longer any reason
to wait for the ctrl to be ready, any action will be deferred until start(),
anyway. This means conn_xprt_init() is no longer needed.
This commit is a fix/complement to the following one :
08d87b3f49
BUG/MEDIUM: backend: never reuse a connection for tcp mode
It fixes the check for the early insertion of backend connections in
the reuse lists if the backend mode is HTTP.
The impact of this bug seems limited because :
- in tcp mode, no insertion is done in the avail list as mux_pt does not
support multiple streams.
- in http mode, muxes are also responsible to insert backend connections
in lists in their detach functions. Prior to this fix the reuse rate
could be slightly inferior.
It can be backported to 2.3.
Currently, there seems to be no way to have the transport layer ready
but not the mux in the function connect_server. Add a BUG_ON to report
if this implicit condition is not true anymore.
This should fix coverity report from github issue #1120.
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).
Sometime, a server connection may be performed synchronously. Most of time on
TCP socket, it does not happen. It is easier to have sync connect with unix
socket. When it happens, if we are not waiting for any hanshake completion, we
must be sure to have a mux installed before leaving the connect_server()
function because an attempt to send may be done before the I/O connection
handler have a chance to be executed to install the mux, if not already done.
For now, It is not expected to perform a send with no mux installed, leading to
a crash if it happens.
This patch should fix the issue #762 and probably #779 too. It must be
backported as far as 1.9.
Commit 08016ab82 ("MEDIUM: connection: Add private connections
synchronously in session server list") introduced a build warning about
a potential null dereference which is actually true: in case a reuse
fails an we fail to allocate a new connection, we could crash. The
issue was already present earlier but the compiler couldn't detect
it since it was guarded by an independent condition.
This should be carefully backported to older versions (at least 2.2
and maybe 2.1), the change consists in only adding a test on srv_conn.
The whole sequence of "if" blocks is ugly there and would deserve being
cleaned up so that the !srv_conn condition is matched ASAP and the
assignment is done later. This would remove complicated conditions.
The following sample fetches have been added :
* srv_iweight : returns the initial server's weight
* srv_uweight : returns the user-visible server's weight
* srv_weight : returns the current (or effetctive) server's weight
The requested server must be passed as argument, evnetually preceded by the
backend name. For instance :
srv_weight(back-http/www1)
The srv_use_idle_conn() function is now responsible to update the server
counters and the connection flags when an idle connection is reused. The same
function is called when a new connection is created. This simplifies a bit the
connect_server() function.
When a new connection is created, its target is always set just after. So the
connection target may set when it is created instead, during its initialisation
to be precise. It is the purpose of this patch. Now, conn_new() function is
called with the connection target as parameter. The target is then passed to
conn_init(). It means the target must be passed when cs_new() is called. In this
case, the target is only used when the conn-stream is created with no
connection. This only happens for tcpchecks for now.
The session_get_conn() must now be used to look for an available connection
matching a specific target for a given session. This simplifies a bit the
connect_server() function.
When a connection is marked as private, it is now added in the session server
list. We don't wait a stream is detached from the mux to do so. When the
connection is created, this happens after the mux creation. Otherwise, it is
performed when the connection is marked as private.
To allow that, when a connection is created, the session is systematically set
as the connectin owner. Thus, a backend connection has always a owner during its
creation. And a private connection has always a owner until its death.
Note that outside the detach() callback, if the call to session_add_conn()
failed, the error is ignored. In this situation, we retry to add the connection
into the session server list in the detach() callback. If this fails at this
step, the multiplexer is destroyed and the connection is closed.
To set a connection as private, the conn_set_private() function must now be
called. It sets the CO_FL_PRIVATE flags, but it also remove the connection from
the available connection list, if necessary. For now, it never happens because
only HTTP/1 connections may be set as private after their creation. And these
connections are never inserted in the available connection list.
When a new connection is created, it may immediatly be set as private if
http-reuse never is configured for the backend. There is no reason to wait the
call to mux->detach() to do so.
If an expression is configured to set the SNI on a server connection, the
connection is marked as private. To not needlessly add it in the available
connection list when the mux is installed, the SNI is now set on the connection
before installing the mux, just after the call to si_connect().
When a connection is added to an idle list, it's already detached and
cannot be seen by two threads at once, so there's no point using
TRY_ADDQ, there will never be any conflict. Let's just use the cheaper
ADDQ.
Initially when mt_lists were added, their purpose was to be used with
the scheduler, where anyone may concurrently add the same tasklet, so
it sounded natural to implement a check in MT_LIST_ADD{,Q}. Later their
usage was extended and MT_LIST_ADD{,Q} started to be used on situations
where the element to be added was exclusively owned by the one performing
the operation so a conflict was impossible. This became more obvious with
the idle connections and the new macro was called MT_LIST_ADDQ_NOCHECK.
But this remains confusing and at many places it's not expected that
an MT_LIST_ADD could possibly fail, and worse, at some places we start
by initializing it before adding (and the test is superflous) so let's
rename them to something more conventional to denote the presence of the
check or not:
MT_LIST_ADD{,Q} : inconditional operation, the caller owns the
element, and doesn't care about the element's
current state (exactly like LIST_ADD)
MT_LIST_TRY_ADD{,Q}: only perform the operation if the element is not
already added or in the process of being added.
This means that the previously "safe" MT_LIST_ADD{,Q} are not "safe"
anymore. This also means that in case of backport mistakes in the
future causing this to be overlooked, the slower and safer functions
will still be used by default.
Note that the missing unchecked MT_LIST_ADD macro was added.
The rest of the code will have to be reviewed so that a number of
callers of MT_LIST_TRY_ADDQ are changed to MT_LIST_ADDQ to remove
the unneeded test.
When a connection is created and the multiplexer is installed, if the connection
is marked as private, don't consider it as available, regardless the number of
available streams. This test is performed when the mux is installed when the
connection is created, in connect_server(), and when the mux is installed after
the handshakes stage.
No backport needed, this is 2.2-dev.
When a connection is picked from the session server list because the proxy or
the session are marked to use the last requested server, if it is idle, we must
marked it as used removing the CO_FL_SESS_IDLE flag and decrementing the session
idle_conns counter.
This patch must be backported as far as 1.9.
In the various takeover() methods, make sure we schedule the old tasklet
on the old thread, as we don't want it to run on our own thread! This
was causing a very rare crash when building with DEBUG_STRICT, seeing
that either an FD's thread mask didn't match the thread ID in h1_io_cb(),
or that stream_int_notify() would try to queue a task with the wrong
tid_bit.
In order to reproduce this, it is necessary to maintain many connections
(typically 30k) at a high request rate flowing over H1+SSL between two
proxies, the second of which would randomly reject ~1% of the incoming
connection and randomly killing some idle ones using a very short client
timeout. The request rate must be adjusted so that the CPUs are nearly
saturated, but never reach 100%. It's easier to reproduce this by skipping
local connections and always picking from other threads. The issue
should happen in less than 20s otherwise it's necessary to restart to
reset the idle connections lists.
No backport is needed, takeover() is 2.2 only.
Enables ('on') or disables ('off') sharing of idle connection pools between
threads for a same server. The default is to share them between threads in
order to minimize the number of persistent connections to a server, and to
optimize the connection reuse rate. But to help with debugging or when
suspecting a bug in HAProxy around connection reuse, it can be convenient to
forcefully disable this idle pool sharing between multiple threads, and force
this option to "off". The default is on.
This could have been nice to have during the idle connections debugging,
but it's not too late to add it!
The next thread walking algorithm in commit 566df309c ("MEDIUM:
connections: Attempt to get idle connections from other threads.")
proved to be sufficient for most cases, but it still has some rough
edges when threads are unevenly loaded. If one thread wakes up with
10 streams to process in a burst, it will mainly take over connections
from the next one until it doesn't have anymore.
This patch implements a rotating index that is stored into the server
list and that any thread taking over a connection is responsible for
updating. This way it starts mostly random and avoids always picking
from the same place. This results in a smoother distribution overall
and a slightly lower takeover rate.
There's a tricky behavior that was lost when the idle connections were
made sharable between thread in commit 566df309c ("MEDIUM: connections:
Attempt to get idle connections from other threads."), it is the ability
to retry from the safe list when looking for any type of idle connection
and not finding one in the idle list.
It is already important when dealing with long-lived connections since
they ultimately all become safe, but that case is already covered by
the fact that safe conns not being used end up closing and are not
looked up anymore since connect_server() sees there are none.
But it's even more important when using server-side connections which
periodically close, because the new connections may spend half of their
time in safe state and the other half in the idle state, and failing
to grab one such connection from the right list results in establishing
a new connection.
This patch makes sure that a failure to find an idle connection results
in a new attempt at finding one from the safe list if available. In order
to avoid locking twice, connections are attempted alternatively from the
idle then safe list when picking from siblings. Tests have shown a ~2%
performance increase by avoiding to lock twice.
A typical test with 10000 connections over 16 threads with 210 servers
having a 1 millisecond response time and closing every 5 requests shows
a degrading performance starting at 120k req/s down to 60-90k and an
average reuse rate of 44%. After the fix, the reuse rate raises to 79%
and the performance becomes stable at 254k req/s. Similarly the previous
test with full keep-alive has now increased from 96% reuse rate to 99%
and from 352k to 375k req/s.
No backport is needed as this is 2.2-only.
The problem with the way idle connections currently work is that it's
easy for a thread to steal all of its siblings' connections, then release
them, then it's done by another one, etc. This happens even more easily
due to scheduling latencies, or merged events inside the same pool loop,
which, when dealing with a fast server responding in sub-millisecond
delays, can really result in one thread being fully at work at a time.
In such a case, we perform a huge amount of takeover() which consumes
CPU and requires quite some locking, sometimes resulting in lower
performance than expected.
In order to fight against this problem, this patch introduces a new server
setting "pool-low-conn", whose purpose is to dictate when it is allowed to
steal connections from a sibling. As long as the number of idle connections
remains at least as high as this value, it is permitted to take over another
connection. When the idle connection count becomes lower, a thread may only
use its own connections or create a new one. By proceeding like this even
with a low number (typically 2*nbthreads), we quickly end up in a situation
where all active threads have a few connections. It then becomes possible
to connect to a server without bothering other threads the vast majority
of the time, while still being able to use these connections when the
number of available FDs becomes low.
We also use this threshold instead of global.nbthread in the connection
release logic, allowing to keep more extra connections if needed.
A test performed with 10000 concurrent HTTP/1 connections, 16 threads
and 210 servers with 1 millisecond of server response time showed the
following numbers:
haproxy 2.1.7: 185000 requests per second
haproxy 2.2: 314000 requests per second
haproxy 2.2 lowconn 32: 352000 requests per second
The takeover rate goes down from 300k/s to 13k/s. The difference is
further amplified as the response time shrinks.
In conn_backend_get() we can avoid locking other servers when trying
to steal their connections when we know for sure they will not have
one, so let's do it to lower the contention on the lock.
Starting with commit 079cb9a ("MEDIUM: connections: Revamp the way idle
connections are killed") we started to improve the way to compute the
need for idle connections. But the condition to keep a connection idle
or drop it when releasing it was not updated. This often results in
storms of close when certain thresholds are met, and long series of
takeover() when there aren't enough connections left for a thread on
a server.
This patch tries to improve the situation this way:
- it keeps an estimate of the number of connections needed for a server.
This estimate is a copy of the max over previous purge period, or is a
max of what is seen over current period; it differs from max_used_conns
in that this one is a counter that's reset on each purge period ;
- when releasing, if the number of current idle+used connections is
lower than this last estimate, then we'll keep the connection;
- when releasing, if the current thread's idle conns head is empty,
and we don't exceed the estimate by the number of threads, then
we'll keep the connection.
- when cleaning up connections, we consider the max of the last two
periods to avoid killing too many idle conns when facing bursty
traffic.
Thanks to this we can better converge towards a situation where, provided
there are enough FDs, each active server keeps at least one idle connection
per thread all the time, with a total number close to what was needed over
the previous measurement period (as defined by pool-purge-delay).
On tests with large numbers of concurrent connections (30k) and many
servers (200), this has quite smoothed the CPU usage pattern, increased
the reuse rate and roughly halved the takeover rate.
There's a minor glitch with the way idle connections start to be evicted.
The lookup always goes from thread 0 to thread N-1. This causes depletion
of connections on the first threads and abundance on the last ones. This
is visible with the takeover() stats below:
$ socat - /tmp/sock1 <<< "show activity"|grep ^fd ; \
sleep 10 ; \
socat -/tmp/sock1 <<< "show activity"|grep ^fd
fd_takeover: 300144 [ 91887 84029 66254 57974 ]
fd_takeover: 359631 [ 111369 99699 79145 69418 ]
There are respectively 19k, 15k, 13k and 11k takeovers for only 4 threads,
indicating that the first thread needs a foreign FD twice more often than
the 4th one.
This patch changes this si that all threads are scanned in round robin
starting with the current one. The takeovers now happen in a much more
distributed way (about 4 times 9k) :
fd_takeover: 1420081 [ 359562 359453 346586 354480 ]
fd_takeover: 1457044 [ 368779 368429 355990 363846 ]
There is no need to backport this, as this happened along a few patches
that were merged during 2.2 development.
The FD takeover operation might have certain impacts explaining
unexpected activities, so it's important to report such a counter
there. We thus count the number of times a thread has stolen an
FD from another thread.
In connect_server(), we want to increase curr_used_conns only if the
connection is new, or if it comes from an idle_pool, otherwise it means
the connection is already used by at least one another stream, and it is
already accounted for.
We used to have 3 thread-based arrays for toremove_lock, idle_cleanup,
and toremove_connections. The problem is that these items are small,
and that this creates false sharing between threads since it's possible
to pack up to 8-16 of these values into a single cache line. This can
cause real damage where there is contention on the lock.
This patch creates a new array of struct "idle_conns" that is aligned
on a cache line and which contains all three members above. This way
each thread has access to its variables without hindering the other
ones. Just doing this increased the HTTP/1 request rate by 5% on a
16-thread machine.
The definition was moved to connection.{c,h} since it appeared a more
natural evolution of the ongoing changes given that there was already
one of them declared in connection.h previously.
This patch fixes all the leftovers from the include cleanup campaign. There
were not that many (~400 entries in ~150 files) but it was definitely worth
doing it as it revealed a few duplicates.
This one was not easy because it was embarking many includes with it,
which other files would automatically find. At least global.h, arg.h
and tools.h were identified. 93 total locations were identified, 8
additional includes had to be added.
In the rare files where it was possible to finalize the sorting of
includes by adjusting only one or two extra lines, it was done. But
all files would need to be rechecked and cleaned up now.
It was the last set of files in types/ and proto/ and these directories
must not be reused anymore.
extern struct dict server_name_dict was moved from the type file to the
main file. A handful of inlined functions were moved at the bottom of
the file. Call places were updated to use server-t.h when relevant, or
to simply drop the entry when not needed.
