For both servers and proxies, use one connection queue per thread-group,
instead of only one. Having only one can lead to severe performance
issues on NUMA machines, it is actually trivial to get the watchdog to
trigger on an AMD machine, having a server with a maxconn of 96, and an
injector that uses 160 concurrent connections.
We now have one queue per thread-group, however when dequeueing, we're
dequeuing MAX_SELF_USE_QUEUE (currently 9) pendconns from our own queue,
before dequeueing one from another thread group, if available, to make
sure everybody is still running.
For both proxies and servers, properly calculates queueslength, which is
the total number of element in each queues (as they currently are only
using one queue, it is equivalent to the number of element of that
queue), and use it instead of the queue's length.
There is a small race condition, where a server would check if there is
something left in the proxy queue, and adding something to the proxy
queue. If the server checks just before the stream is added to the queue,
and it no longer has any stream to deal with, then nothing will take
care of the stream, that may stay in the queue forever.
This was worked around with commit 5541d4995d, by checking for that exact
condition after adding the stream to the queue, and trying again to get
a server assigned if it is detected.
That fix lead to multiple infinite loops, that got fixed, but it is not
unlikely that it could happen again. So let's fix the initial problem
differently : a single server may mark itself as ready, and it removes
itself once used. The principle is that when we discover that the just
queued stream is alone with no active request anywhere ot dequeue it,
instead of rebalancing it, it will be assigned to that current "ready"
server that is available to handle it. The extra cost of the atomic ops
is negligible since the situation is super rare.
The "served" field of struct server is used to know how many connections
are currently in use for a server. But served used to be incremented way
after the server was picked, so there were race conditions that could
lead more than maxconn connections to be allocated for one server. To
fix this, increment served way earlier, and make sure at the time that
it never goes past maxconn.
We now should never have more outgoing connections than set by maxconn.
Since commit 53f52e67a0 ("BUG/MEDIUM: queue: always dequeue the backend when
redistributing the last server"), we've got two reports again still showing
the theoretically impossible condition in pendconn_add(), including a single
threaded one.
Thanks to the traces, the issue could be tracked down to the redispatch part.
In fact, in non-determinist LB algorithms (RR, LC, FAS), we don't perform the
LB if there are pending connections in the backend, since it indicates that
previous attempts already failed, so we directly return SRV_STATUS_FULL. And
contrary to a previous belief, it is possible to meet this condition with
be->served==0 when redispatching (and likely with maxconn not greater than
the number of threads).
The problem is that in this case, the entry is queued and then the
pendconn_must_try_again() function checks if any connections are currently
being served to detect if we missed a race, and tries again, but that
situation is not caused by a concurrent thread and will never fix itself,
resulting in the loop.
All that part around pendconn_must_try_again() is still quite brittle, and
a safer approach would involve a sequence counter to detect new arrivals
and dequeues during the pendconn_add() call. But it's more sensitive work,
probably for a later fix.
This fix must be backported wherever the fix above was backported. Thanks
to Patrick Hemmer, as well as Damien Claisse and Basha Mougamadou from
Criteo for their help on tracking this one!
Instead of directly relying on the backend parameter to limit the number of
connection retries, we now use a per-stream value. This value is by default
inherited from the backend value when it is set. So for now, there is no
change except the stream value is used instead of the backend value. But
thanks to this change, it will be possible to dynamically change this value.
Multipath TCP (MPTCP), standardized in RFC8684 [1], is a TCP extension
that enables a TCP connection to use different paths.
Multipath TCP has been used for several use cases. On smartphones, MPTCP
enables seamless handovers between cellular and Wi-Fi networks while
preserving established connections. This use-case is what pushed Apple
to use MPTCP since 2013 in multiple applications [2]. On dual-stack
hosts, Multipath TCP enables the TCP connection to automatically use the
best performing path, either IPv4 or IPv6. If one path fails, MPTCP
automatically uses the other path.
To benefit from MPTCP, both the client and the server have to support
it. Multipath TCP is a backward-compatible TCP extension that is enabled
by default on recent Linux distributions (Debian, Ubuntu, Redhat, ...).
Multipath TCP is included in the Linux kernel since version 5.6 [3]. To
use it on Linux, an application must explicitly enable it when creating
the socket. No need to change anything else in the application.
This attached patch adds MPTCP per address support, to be used with:
mptcp{,4,6}@<address>[:port1[-port2]]
MPTCP v4 and v6 protocols have been added: they are mainly a copy of the
TCP ones, with small differences: names, proto, and receivers lists.
These protocols are stored in __protocol_by_family, as an alternative to
TCP, similar to what has been done with QUIC. By doing that, the size of
__protocol_by_family has not been increased, and it behaves like TCP.
MPTCP is both supported for the frontend and backend sides.
Also added an example of configuration using mptcp along with a backend
allowing to experiment with it.
Note that this is a re-implementation of Björn's work from 3 years ago
[4], when haproxy's internals were probably less ready to deal with
this, causing his work to be left pending for a while.
Currently, the TCP_MAXSEG socket option doesn't seem to be supported
with MPTCP [5]. This results in a warning when trying to set the MSS of
sockets in proto_tcp:tcp_bind_listener.
This can be resolved by adding two new variables:
sock_inet(6)_mptcp_maxseg_default that will hold the default
value of the TCP_MAXSEG option. Note that for the moment, this
will always be -1 as the option isn't supported. However, in the
future, when the support for this option will be added, it should
contain the correct value for the MSS, allowing to correctly
set the TCP_MAXSEG option.
Link: https://www.rfc-editor.org/rfc/rfc8684.html [1]
Link: https://www.tessares.net/apples-mptcp-story-so-far/ [2]
Link: https://www.mptcp.dev [3]
Link: https://github.com/haproxy/haproxy/issues/1028 [4]
Link: https://github.com/multipath-tcp/mptcp_net-next/issues/515 [5]
Co-authored-by: Dorian Craps <dorian.craps@student.vinci.be>
Co-authored-by: Matthieu Baerts (NGI0) <matttbe@kernel.org>
After checking that a server or backend is full, it remains possible
to call pendconn_add() just after the last pending requests finishes, so
that there's no more connection on the server for very low maxconn (typ 1),
leaving new ones in queue till the timeout.
The approach depends on where the request was queued, though:
- when queued on a server, we can simply detect that we may dequeue
pending requests and wake them up, it will wake our request and
that's fine. This needs to be done in srv_redispatch_connect() when
the server is set.
- when queued on a backend, it means that all servers are done with
their requests. It means that all servers were full before the
check and all were empty after. In practice this will only concern
configs with less servers than threads. It's where the issue was
first spotted, and it's very hard to reproduce with more than one
server. In this case we need to load-balance again in order to find
a spare server (or even to fail). For this, we call the newly added
dedicated function pendconn_must_try_again() that tells whether or
not a blocked pending request was dequeued and needs to be retried.
This should be backported along with pendconn_must_try_again() to all
stable versions, but with extreme care because over time the queue's
locking evolved.
In connect_server() function, there is a test to be able to reuse idle
connections for HTX streams only. Till now, only HTTP connections can be
idle. And this tests was added to be sure to now reuse idle connections for
legacy HTTP streams. But the legacy HTTP was removed in HAProxy-2.1. So we
can safely remove this test.
The related issue is #2502.
Implement pool-conn-name support for idle connection reuse. It replaces
SNI as arbitrary identifier for connections in the idle pool. Thus,
every SNI reference in this context have been replaced.
Main change occurs in connect_server() where pool-conn-name sample fetch
is now prehash to generate idle connection identifier. SNI is now solely
used in the context of SSL for ssl_sock_set_servername().
Update API for PROXY protocol header encoding. Previously, it requires
stream parameter to be set. Change make_proxy_line() and associated
functions to add an extra session parameter. This is useful in context
where no stream is instantiated. For example, this is the case for rhttp
preconnect.
This change allows to extend PROXY v2 TLV encoding. Replace
build_logline() which requires a stream instance and call directly
sess_build_logline().
Note that stream parameter is kept as it is necessary for unique ID
encoding.
This change has no functional change for standard connections. However,
it is necessary to support TLV encoding on rhttp preconnect.
Convert FN_AGE in stat_cols_px[] as generic columns. These values will
be automatically used for dump/preload of a stats-file.
Remove srv_lastsession() / be_lastsession() function which are now
useless as last_sess is calculated via me_generate_field().
last_change was a member present in both proxy and server struct. It is
used as an age statistics to report the last update of the object.
Move last_change into fe_counters/be_counters. This is necessary to be
able to manipulate it through generic stat column and report it into
stats-file.
Note that there is a change for proxy structure with now 2 different
last_change values, on frontend and backend side. Special care was taken
to ensure that the value is initialized only on the proxy side. The
other value is set to 0 unless a listen proxy is instantiated. For the
moment, only backend counter is reported in stats. However, with now two
distinct values, stats could be extended to report it on both side.
Move freq-ctr defined in proxy or server structures into their dedicated
fe_counters/be_counters struct.
Functionnaly no change here. This commit will allow to convert rate
stats column to generic one, which is mandatory to manipulate them in
the stats-file.
This commit is the first one of a series which aims to align counters
usage between frontends/listeners on one side and backends/servers on
the other.
Remove <down_trans> field from proxy structure. Use instead the same
name field from be_counters structure, which is already used for
servers.
As previously mentioned in cd352c0db ("MINOR: log/balance: rename
"log-sticky" to "sticky""), let's define a sticky algorithm that may be
used from any protocol. Sticky algorithm sticks on the same server as
long as it remains available.
The documentation was updated accordingly.
Extend takeover API both for MUX and XPRT with a new boolean argument
<release>. Its purpose is to signal if the connection will be freed
immediately after the takeover, rendering new resources allocation
unnecessary.
For the moment, release argument is always false. However, it will be
set to true on delete server CLI handler to proactively close server
idle connections.
By default, backend connections are attached to a server instance. This
allows to implement connection reuse. However, in some particular cases,
connection cannot be shared accross several clients. These connections
are considered and private and are attached to the session instance
instead.
These private connections are also indexed by the target server to not
mix them. All of this is implemented via a dedicated structure
previously named struct sess_srv_list.
Rename it to better reflect its usage to struct sess_priv_conns. Also
rename its internal members and all of the associated functions.
This commit is only a renaming, thus no functional impact is expected.
set-bc-{mark,tos} actions are pretty similar to set-fc-{mark,tos} to set
mark/tos on packets sent from haproxy to server: set-bc-{mark,tos} actions
act on the whole backend/srv connection: from connect() to connection
teardown, thus they may only be used before the connection to the server
is instantiated, meaning that they are only relevant for request-oriented
rules such as tcp-request or http-request rules. For now their use is
limited to content request rules, because tos and mark informations are
stored directly within the stream, thus it is required that the stream
already exists.
stream flags are used in combination with dedicated stream struct members
variables to pass 'tos' and 'mark' informations so that they are correctly
considered during stream connection assignment logic (prior to connecting
to actually connecting to the server)
'tos' and 'mark' fd sockopts are taken into account in conn hash
parameters for connection reuse mechanism.
The documentation was updated accordingly.
It's useful to be able to recognize certain functions that are often
present in backtraces as they call lower level functions, and for this
they must not be static. Let's remove "static" in front of these
functions:
sc_notify, sc_conn_recv, sc_conn_send, sc_conn_process,
sc_applet_process, back_establish, stream_update_both_sc,
httpclient_applet_io_handler, httpclient_applet_init,
httpclient_applet_release
Instead of the generic MUX_, we now use MUX_CTL_ prefix for all mux_ctl_type
value. This will avoid any ambiguities with other enums, especially with a
new one that will be added to get information on mux streams.
After giving it some thought, it could pretty well happen that other
protocols benefit from the sticky algorithm that some used to emulate
using a "stick-on int(0)" or things like this previously. So better
rename it to "sticky" right now instead of having to keep that "log-"
prefix forever. It's still limited to logs, of course, only the algo
is renamed in the config.
Previous commit renames 'proto_reverse_connect' module to 'proto_rhttp'.
This commits follows this by replacing various custom prefix by 'rhttp_'
to make the code uniform.
Note that 'reverse_' prefix was kept in connection module. This is
because if a new reversable protocol not based on HTTP is implemented,
it may be necessary to reused the same connection function which are
protocol agnostic.
"log-balance" directive was recently introduced to configure the
balancing algorithm to use when in a log backend. However, it is
confusing and it causes issues when used in default section.
In this patch, we take another approach: first we remove the
"log-balance" directive, and instead we rely on existing "balance"
directive to configure log load balancing in log backend.
Some algorithms such as roundrobin can be used as-is in a log backend,
and for log-only algorithms, they are implemented as "log-$name" inside
the "backend" directive.
The documentation was updated accordingly.
If less connections than threads are established on a reverse-http gateway
and these servers have a non-nul pool-min-conn, then conn_backend_get()
will refrain from picking available connections from other threads. But
this makes no sense for protocols for which there is no ->connect(),
since there's no way the current thread will manage to establish its own
connection. For such situations we should always accept to use another
thread's connection. That's precisely what this patch does.
Define a new function srv_add_to_avail_list(). This function is used to
centralize connection insertion in available tree. It reuses a BUG_ON()
statement to ensure the connection is not present in the idle list.
Since the following commit, idle conns are stored in a list as secondary
storage to retrieve them in usage order :
5afcb686b9
MAJOR: connection: purge idle conn by last usage
The list usage has been extended wherever connections lookup are done
both on idle and safe trees. This reduced the code size by replacing a
two tree loops by a single list loop.
LIST_ELEM() is used in this context to retrieve the first idle list
element from the server list head. However, macro usage was wrong due to
an extra '&' operator which returns an invalid connection reference.
This will most of the time caused a crash on conn_delete_from_tree() or
affiliated functions.
This bug only occurs if the FD pool is exhausted and some idle
connections are selected to be killed.
It can be reproduced using the following config and h2load command :
$ h2load -t 8 -c 800 -m 10 -n 800 "http://127.0.0.1:21080/?s=10k"
global
maxconn 100
defaults
mode http
timeout connect 20s
timeout client 20s
timeout server 20s
listen li
bind :21080 proto h2
server nginx 127.99.0.1:30080 proto h1
This bug has been introduced by the above commit. Thus no need to
backport this fix.
Note that LIST_ELEM() macro usage was slightly adjusted also in
srv_migrate_conns_to_remove(). The function used toremove_list instead
of idle_list connection list element. This is not a bug as they are
stored in the same union. However, the new code is clearer as it intends
to move connection from the idle_list only into the toremove_list
mt-list.
Idle connections are both stored in an idle/safe tree and in an idle
list. The list is used as a secondary storage to be able to retrieve
them by usage order.
If a connection is moved into the available tree, it must not be present
in the idle list. A BUG_ON() was written to check this but was placed at
the wrong code section. Fix this by removing the misplaced one and write
new ones for avail_conns tree insertion and lookup.
The impact of this bug is minor as the misplaced BUG_ON() did not seem
to be triggered.
No need to backport.
Because channel_is_empty() function does now only check the channel's
buffer, we can remove it and rely on co_data() instead. Of course, all tests
must be inverted.
channel_is_empty() is thus removed.
hash lb algorithm can be configured with the "log-balance hash <cnv_list>"
directive. With this algorithm, the user specifies a converter list with
<cnv_list>.
The produced log message will be passed as-is to the provided converter
list, and the resulting hash will be used to select the log server that
will receive the log message.
Instead of systematically computing the avalanche hash right after the
gen_hash() call, do it inside the gen_hash() function directly to ensure
avalanche setting is always considered.
Allow the use of the "none" hash-type function so that the key resulting
from the sample expression is directly used as the hash.
This can be useful to do the hashing manually using available hashing
converters, or even custom ones, and then inform haproxy that it can
directly rely on the sample expression result which is explictly handled
as an integer in this case.
In this patch we add basic support for the random algorithm:
random algorithm picks a random server using the result of the
statistical_prng() function as if it was a hash key to then compute the
related server ID.
There is no support for the <draw> parameter (which is implemented for
tcp/http load-balancing), because we don't have the required metrics to
evaluate server's load in log backends for the moment. Plus it would add
more complexity to the __do_send_log_backend() function so we'll keep it
this way for now but this might be needed in the future.
sticky algorithm always tries to send log messages to the first server in
the farm. The server will stay in front during queue and dequeue
operations (no other server can steal its place), unless it becomes
unavailable, in which case it will be replaced by another server from
the tree.
Using "mode log" in a backend section turns the proxy in a log backend
which can be used to log-balance logs between multiple log targets
(udp or tcp servers)
log backends can be used as regular log targets using the log directive
with "backend@be_name" prefix, like so:
| log backend@mybackend local0
A log backend will distribute log messages to servers according to the
log load-balancing algorithm that can be set using the "log-balance"
option from the log backend section. For now, only the roundrobin
algorithm is supported and set by default.
Refactor alloc_bind_address() function which is used to allocate a
sockaddr if a connection to a target server relies on a specific source
address setting.
The main objective of this change is to be able to use this function
outside of backend module, namely for preconnections using a reverse
server. As such, this function is now exported globally.
For reverse connect, there is no stream instance. As such, the function
parts which relied on it were reduced to the minimal. Now, stream is
only used if a non-static address is configured which is useful for
usesrc client|clientip|hdr_ip. These options have no sense for reverse
connect so it should be safe to use the same function.
Backend idle connections are purged on a recurring occurence during the
process lifetime. An estimated number of needed connections is
calculated and the excess is removed periodically.
Before this patch, purge was done directly using the idle then the safe
connection tree of a server instance. This has a major drawback to take
no account of a specific ordre and it may removed functional connections
while leaving ones which will fail on the next reuse.
The problem can be worse when using criteria to differentiate idle
connections such as the SSL SNI. In this case, purge may remove
connections with a high rate of reusing while leaving connections with
criteria never matched once, thus reducing drastically the reuse rate.
To improve this, introduce an alternative storage for idle connection
used in parallel of the idle/safe trees. Now, each connection inserted
in one of this tree is also inserted in the new list at
`srv_per_thread.idle_conn_list`. This guarantees that recently used
connection is present at the end of the list.
During the purge, use this list instead of idle/safe trees. Remove first
connection in front of the list which were not reused recently. This
will ensure that connection that are frequently reused are not purged
and should increase the reuse rate, particularily if distinct idle
connection criterias are in used.
Small change of API for conn_delete_from_tree(). Now the connection
instance is taken as argument instead of its inner node.
No functional change introduced with this commit. This simplifies
slightly invocation of conn_delete_from_tree(). The most useful changes
is that this function will be extended in the next patch to be able to
remove the connection from its new idle list at the same time as in its
idle tree.
A reverse server relies solely on its pool of idle connection to
transfer requests which will be populated through a new tcp-request rule
'attach-srv'.
Several changes are required on connect_server() to implement this.
First, reuse mode is forced to always for this type of server. Then, if
no idle connection is found, the request will be aborted. This results
with a 503 HTTP error code, similarly to when no server is available.
A connection contains extra elements which are only used for the backend
side. Regroup their allocation and deallocation in two new functions
named conn_backend_init() and conn_backend_deinit().
No functional change is introduced with this commit. The new functions
are reused in place of manual alloc/dealloc in conn_new() / conn_free().
This patch will be useful for reverse connect support with connection
conversion from backend to frontend side and vice-versa.
This puts an end to the occasional confusion between the "now" date
that is internal, monotonic and not synchronized with the system's
date, and "date" which is the system's date and not necessarily
monotonic. Variable "now" was removed and replaced with a 64-bit
integer "now_ns" which is a counter of nanoseconds. It wraps every
585 years, so if all goes well (i.e. if humanity does not need
haproxy anymore in 500 years), it will just never wrap. This implies
that now_ns is never nul and that the zero value can reliably be used
as "not set yet" for a timestamp if needed. This will also simplify
date checks where it becomes possible again to do "date1<date2".
All occurrences of "tv_to_ns(&now)" were simply replaced by "now_ns".
Due to the intricacies between now, global_now and now_offset, all 3
had to be turned to nanoseconds at once. It's not a problem since all
of them were solely used in 3 functions in clock.c, but they make the
patch look bigger than it really is.
The clock_update_local_date() and clock_update_global_date() functions
are now much simpler as there's no need anymore to perform conversions
nor to round the timeval up or down.
The wrapping continues to happen by presetting the internal offset in
the short future so that the 32-bit now_ms continues to wrap 20 seconds
after boot.
The start_time used to calculate uptime can still be turned to
nanoseconds now. One interrogation concerns global_now_ms which is used
only for the freq counters. It's unclear whether there's more value in
using two variables that need to be synchronized sequentially like today
or to just use global_now_ns divided by 1 million. Both approaches will
work equally well on modern systems, the difference might come from
smaller ones. Better not change anyhting for now.
One benefit of the new approach is that we now have an internal date
with a resolution of the nanosecond and the precision of the microsecond,
which can be useful to extend some measurements given that timestamps
also have this resolution.
Instead we're using ns_to_sec(tv_to_ns(&now)) which allows the tv_sec
part to disappear. At this point, "now" is only used as a timeval in
clock.c where it is updated.
Let's get rid of timeval in storage of internal timestamps so that they
are no longer mistaken for wall clock time. These were exclusively used
subtracted from each other or to/from "now" after being converted to ns,
so this patch removes the tv_to_ns() conversion to use them natively. Two
occurrences of tv_isge() were turned to a regular wrapping subtract.
Instead of operating on {sec, usec} now we convert both operands to
ns then subtract them and convert to ms. This is a first step towards
dropping timeval from these timestamps.
Interestingly, tv_ms_elapsed() and tv_ms_remain() are no longer used at
all and could be removed.
When compiled in debug mode, HAProxy prints a debug message at the beginning
of assign_server(). It is pretty annoying and useless because, in debug
mode, we can active stream traces. Thus, just remove it.
SE_FL_ERROR flag is no longer set when an error is detected durign the
connection establishment. SC_FL_ERROR flag is set instead. So it is safe to
remove test on SE_FL_ERROR to detect connection establishment error.
We can now fully rely on SC_FL_ERROR flag from the stream. The first step is
to stop to set the SE_FL_ERROR flag. Only endpoints are responsible to set
this flag. It was a design limitation. It is now fixed.
From the stream, when SE_FL_ERROR flag is tested, we now also test the
SC_FL_ERROR flag. Idea is to stop to rely on the SE descriptor to detect
errors.
