There's still a lot of contention when accessing the backend's
totpend and queueslength for every request in may_dequeue_tasks(),
even when queues are not used. This only happens because it's stored
in the same cache line as >beconn which is being written by other
threads:
0.01 | call sess_change_server
0.02 | mov 0x188(%r15),%esi ## s->queueslength
| if (may_dequeue_tasks(srv, s->be))
0.00 | mov 0xa8(%r12),%rax
0.00 | mov -0x50(%rbp),%r11d
0.00 | mov -0x60(%rbp),%r10
0.00 | test %esi,%esi
| jne 3349
0.01 | mov 0xa00(%rax),%ecx ## p->queueslength
8.26 | test %ecx,%ecx
4.08 | je 288d
This patch moves queueslength and totpend to their own cache line,
thus adding 64 bytes to the struct proxy, but gaining 3.6% of RPS
on a 64-core EPYC thanks to the elimination of this false sharing.
process_stream() goes down from 3.88% to 3.26% in perf top, with
the next top users being inc/dec (s->served) and be->beconn.
This field is shared by all threads and must be in the shared area
instead, because where it's placed, it slows down access to other
fields of the struct by false sharing. Just moving this field gives
a steady 2% gain on the request rate (1.93 to 1.96 Mrps) on a 64-core
EPYC.
This option allows to disable the certificate compression (RFC 8879)
using OpenSSL >= 3.2.0.
This feature is known to permit some denial of services by causing extra
memory allocations of approximately 22MiB and extra CPU work per
connection with OpenSSL versions affected by CVE-2025-66199.
( https://openssl-library.org/news/vulnerabilities/index.html#CVE-2025-66199 )
Setting this to "off" permits to mitigate the problem.
Must be backported to every stable branches.
The SSL passphrase callback function was only called when loading
private keys from a dedicated file (separate from the corresponding
certificate) but not when both the certificate and the key were in the
same file.
We can now load them properly, regardless of how they are provided.
A flas had to be added in the 'passphrase_cb_data' structure because in
the 'ssl_sock_load_pem_into_ckch' function, when calling
'PEM_read_bio_PrivateKey' there might be no private key in the PEM file
which would mean that the callback never gets called (and cannot set the
'passphrase_idx' to -1).
This patch can be backported to 3.3.
Certain object sizes cannot be controlled at declaration time because
the resulting object size may be slightly extended (tag, caller),
aligned and rounded up, or even doubled depending on pool settings
(e.g. if backup is used).
This patch addresses this by enlarging the type in the pool registration
to 64-bit so that no info is lost from the declaration, and extra checks
for overflows can be performed during registration after various rounding
steps. This allows to catch issues such as these ones and to report a
suitable error:
global
tune.http.logurilen 2147483647
frontend
capture request header name len 2147483647
http-request capture src len 2147483647
tcp-request content capture src len 2147483647
This patch changes the handling of named defaults sections. Prior to
this patch, every unreferenced defaults proxies were removed on post
parsing. Now by default, these sections are kept after postparsing and
only purged on deinit. The objective is to allow reusing them as base
configuration for dynamic backends.
To implement this, refcount of every still addressable named sections is
incremented by one after parsing. This ensures that they won't be
removed even if referencing proxies are removed at runtime. This is done
via the new function proxy_ref_all_defaults().
To ensure defaults instances are still properly removed on deinit, the
inverse operation is performed : refcount is decremented by one on every
defaults sections via proxy_unref_all_defaults().
The original behavior can still be used by using the new global keyword
tune.defaults.purge. This is useful for users using configuration with
large number of defaults and not interested in dynamic backends
creation.
Defaults section are indexed by their name in defproxy_by_name tree. For
named sections, there is no duplicate : if two instances have the same
name, the older one is removed from the tree. However, this was not the
case for unnamed defaults which are all stored inconditionnally in
defproxy_by_name.
This commit introduces a new approach for unnamed defaults. Now, these
instances are never inserted in the defproxy_by_name tree. Indeed, this
is not needed as no tree lookup is performed with empty names. This may
optimize slightly config parsing with a huge number of named and unnamed
defaults sections, as the first ones won't fill up the tree needlessly.
However, defproxy_by_name tree is also used to purge unreferenced
defaults instances, both on postparsing and deinit. Thus, a new approach
is needed for unnamed sections cleanup. Now, each time a new defaults is
parsed, if the previous instance is unnamed, it is freed unless if
referenced by a proxy. When config parsing is ended, a similar operation
is performed to ensure the last unnamed defaults section won't stay in
memory. To implement this, last_defproxy static variable is now set to
global. Unnamed sections which cannot be removed due to proxies
referencing proxies will still be removed when such proxies are freed
themselves, at runtime or on deinit.
Defaults proxies instance are stored in a global name tree. When there
is a name conflict and the older entry cannot be simply discarded as it
is already referenced, the older entry is instead removed from the name
tree and inserted into the orphaned list.
The purpose of the orphaned list was to guarantee that any remaining
unreferenced defaults are purged either on postparsing or deinit.
However, this is in fact completely useless. Indeed on postparsing,
orphaned entries are always referenced. On deinit instead, defaults are
already freed along the cleanup of all frontend/backend instances clean
up, thanks to their refcounting.
This patch streamlines this by removing orphaned list. Instead, a
defaults section is inserted into a new global defaults_list during
their whole lifetime. This is not strictly necessary but it ensures that
defaults instances can still be accessed easily in the future if needed
even if not present in the name tree. On deinit, a BUG_ON() is added to
ensure that defaults_list is indeed emptied.
Another benefit from this patch is to simplify the defaults deletion
procedure. Orphaned simple list is replaced by a proper double linked
list implementation, so a single LIST_DELETE() is now performed. This
will be notably useful as defaults may be removed at runtime in the
future if backends deletion at runtime is implemented.
This patch renames functions which deal with defaults section. A common
"defaults_px_" prefix is defined. This serves as a marker to identify
functions which can only be used with proxies defaults capability. New
BUG_ON() are enforced to ensure this is valid.
Also, older proxy_unref_or_destroy_defaults() is renamed
defaults_px_detach().
Function proxy_preset_defaults() purpose has evolved over time.
Originally, it was only used to initialize defaults proxies instances.
Until today, it was extended so that all proxies use it. Its objective
is to initialize settings to common default values.
To remove the confusion, this function is now removed. Its content is
integrated directly into init_new_proxy().
Currently, variable names are only used during parsing and are not
stored at runtime. This makes it impossible to iterate through
variables and retrieve their names.
This patch adds infrastructure to store variable names:
- Add 'name' and 'name_len' fields to var_desc structure
- Add 'name' field to var structure
- Add VDF_NAME_ALLOCATED flag to track memory ownership
- Store names in vars_fill_desc(), var_set(), vars_check_arg(),
and parse_store()
- Free names in var_clear() and release_store_rule()
- Add ARGT_VAR handling in release_sample_arg() to free the
allocated name when the flag is set
This prepares the ground for implementing dump_all_vars() in the
next commit.
Tested with:
- ASAN-enabled build on Linux (TARGET=linux-glibc USE_OPENSSL=1
ARCH_FLAGS="-g -fsanitize=address")
- Regression tests: reg-tests/sample_fetches/vars.vtc
- Regression tests: reg-tests/startup/default_rules.vtc
This function takes a string appends it to a buffer in a format
compatible with most languages (double-quoted, with special characters
escaped). It handles standard escape sequences like \n, \r, \", \\.
This generic utility is desined to be used for logging or debugging
purposes where arbitrary string data needs to be safely emitted without
breaking the output format. It will be primarily used by the upcoming
dump_all_vars() sample fetch to dump variable contents safely.
This converter checks the validity and decrypts the content of a JWE
token that has a symetric "alg" algorithm. In such a case, we only
require a secret as parameter in order to decrypt the token.
A recent patch has introduced a new state for proxies : unpublished
backends. Such backends won't be eligilible for traffic, thus
use_backend/default_backend rules which target them won't match and
content switching rules processing will continue.
This patch defines a new frontend keywords 'force-be-switch'. This
keyword allows to ignore unpublished or disabled state. Thus,
use_backend/default_backend will match even if the target backend is
unpublished or disabled. This is useful to be able to test a backend
instance before exposing it outside.
This new keyword is converted into a persist rule of new type
PERSIST_TYPE_BE_SWITCH, stored in persist_rules list proxy member. This
is the only persist rule applicable to frontend side. Prior to this
commit, pure frontend proxies persist_rules list were always empty.
This new features requires adjustment in process_switching_rules(). Now,
when a use_backend/default_backend rule matches with an non eligible
backend, frontend persist_rules are inspected to detect if a
force-be-switch is present so that the backend may be selected.
Utility function warnif_cond_conflicts() is used when parsing an ACL.
Previously, the function directly calls ha_warning() to report an error.
Change the function so that it now takes the error message as argument.
Caller can then output it as wanted.
This change is necessary to use the function when parsing a keyword
registered as cfg_kw_list. The next patch will reuse it.
Define a new set of CLI commands publish/unpublish backend <be>. The
objective is to be able to change the status of a backend to
unpublished. Such a backend is considered ineligible to traffic : this
allows to skip use_backend rules which target it.
Note that contrary to disabled/stopped proxies, an unpublished backend
still has server checks running on it.
Internally, a new proxy flags PR_FL_BE_UNPUBLISHED is defined. CLI
commands handler "publish backend" and "unpublish backend" are executed
under thread isolation. This guarantees that the flag can safely be set
or remove in the CLI handlers, and read during content-switching
processing.
A proxy can be marked as disabled using the keyword with the same name.
The doc mentions that it won't process any traffic. However, this is not
really the case for backends as they may still be selected via switching
rules during stream processing.
In fact, currently access to disabled backends will be conducted up to
assign_server(). However, no eligible server is found at this stage,
resulting in a connection closure or an HTTP 503, which is expected. So
in the end, servers in disabled backends won't receive any traffic. But
this is only because post-parsing steps are not performed on such
backends. Thus, this can be considered as functional but only via
side-effects.
This patch clarifies the handling of disable backends, so that they are
never selected via switching rules. Now, process_switching_rules() will
ignore disable backends and continue rules evaluation.
As this is a behavior change, this patch is labelled as medium. The
documentation manuel for use_backend is updated accordingly.
If we want to be able to have more than 64 thread groups, we can no
longer use thread group masks as long.
One remaining place where it is done is in struct thread_set. However,
it is not really used as a mask anywhere, all we want is a thread group
counter, so convert that mask to a counter.
Now that it is unused, eliminate all_tgroups_mask, as we can't 64bits
masks to represent thread groups, if we want to be able to have more
than 64 thread groups.
Contrarily to what was previously believed, there are corner cases where
the counters may not be allocated, and we may want to make them optional
at a later date, so we have to check if those counters are there.
However, just checking that shared.tg is non-NULL is enough, we can then
assume that shared.tg[tgid - 1] has properly been allocated too.
Also modify the various COUNTER_SHARED_* macros to make sure they check
for that too.
Before updating counters, a few tests are made to check if the counters
exits. but those counters should always exist at this point, so just
remmove them.
This commit should have no impact, but can easily be reverted with no
functional impact if various crashes appear.
Instead of statically allocating the per-thread group counters,
based on the max number of thread groups available, allocate
them dynamically, based on the number of thread groups actually
used. That way we can increase the maximum number of thread
groups without using an unreasonable amount of memory.
Increase the size of the stored tgid in the stat file from 8bits to
32bits, so that we can have more than 256 thread group. 65536 should be
enough for some time.
This bumps thet stat file minor version, as the structure changes.
Instead of always allocating MAX_TGROUPS members, allocate them
dynamically, using the number of thread groups we'll use, so that
increasing MAX_TGROUPS will not have a huge impact on the structure
size.
This flag is used as of commit dcce936912
("MINOR: connections: Add a new CO_FL_SSL_NO_CACHED_INFO flag"). This patch
should be backported to 3.3. Apparently dcce936912 has been backported
to 3.2 and 3.1 already, with that change already applied, so no need for a
backport there.
The only purpose from tgroup_mask seems to be to calculate how many
tgroups share the same shard, but this is an information we can
calculate differently, we just have to increment the number when a new
receiver is added to the shard, and decrement it when one is detached
from the shard. Removing thread group masks will allow us to increase
the maximum number of thread groups past 64.
It's regularly needed to call getsockopt() on a connection, but each
time the calling code has to do all the job by itself. This commit adds
a "get_opt()" callback on the protocol struct, that directly calls
getsockopt() on the connection's FD. A generic implementation for
standard sockets is provided, though QUIC would likely require a
different approach, or maybe a mapping. Due to the overlap between
IP/TCP/socket option values, it is necessary for the caller to indicate
both the level and the option. An abstraction of the level could be
done, but the caller would nonetheless have to know the optname, which
is generally defined in the same include files. So for now we'll
consider that this callback is only for very specific use.
The levels and optnames are purposely passed as signed ints so that it
is possible to further extend the API by using negative levels for
internal namespaces.
This option enables TCP_SAVE_SYN on the listening socket, which will
cause the kernel to try to save a copy of the SYN packet header (L2,
IP and TCP are supported). This can permit to check the source MAC
address of a client, or find certain TCP options such as a source
address encapsulated using RFC7974. It could also be used as an
alternate approach to retrieving the source and destination addresses
and ports. For now setting the option is enabled, but sample fetch
functions and converters will be needed to extract info.
This makes a significant difference when loading large files and during
commit and clear operations, thanks to improved cache locality. In the
measurements below, master refers to the code before any of the changes
to the patterns code, not the code before this one commit.
Timing the replacement of 10M entries from the CLI with this command
which also reports timestamps at start, end of upload and end of clear:
$ (echo "prompt i"; echo "show activity"; echo "prepare acl #0";
awk '{print "add acl @1 #0",$0}' < bad-ip.map; echo "show activity";
echo "commit acl @1 #0"; echo "clear acl @0 #0";echo "show activity") |
socat -t 10 - /tmp/sock1 | grep ^uptim
master, on a 3.7 GHz EPYC, 3 samples:
uptime_now: 6.087030
uptime_now: 25.981777 => 21.9 sec insertion time
uptime_now: 29.286368 => 3.3 sec commit+clear
uptime_now: 5.748087
uptime_now: 25.740675 => 20.0s insertion time
uptime_now: 29.039023 => 3.3 s commit+clear
uptime_now: 7.065362
uptime_now: 26.769596 => 19.7s insertion time
uptime_now: 30.065044 => 3.3s commit+clear
And after this commit:
uptime_now: 6.119215
uptime_now: 25.023019 => 18.9 sec insertion time
uptime_now: 27.155503 => 2.1 sec commit+clear
uptime_now: 5.675931
uptime_now: 24.551035 => 18.9s insertion
uptime_now: 26.652352 => 2.1s commit+clear
uptime_now: 6.722256
uptime_now: 25.593952 => 18.9s insertion
uptime_now: 27.724153 => 2.1s commit+clear
Now timing the startup time with a 10M entries file (on another machine)
on master, 20 samples:
Standard Deviation, s: 0.061652677408033
Mean: 4.217
And after this commit:
Standard Deviation, s: 0.081821371548669
Mean: 3.78
Instead of a global list (and tree) of pattern reference elements, we
now have an intermediate pat_ref_gen structure and store the elements in
those. This simplifies the logic of some operations such as commit and
clear, and improves performance in some cases - numbers to be provided
in a subsequent commit after one important optimization is added.
A lot of the changes are due to adding an extra level of indirection,
changing many cases where we iterate over all elements to an outer loop
iterating over the generation and an inner one iterating over the
elements of the current generation. It is therefore easier to read this
patch using 'git diff -w'.
Safe and non-functional changes that only add currently unused
structures, field, functions and macros, in preparation of larger
changes that alter the way pattern reference elements are stored.
This includes code to create and lookup generation objects, and
macros to iterate over the generations of a pattern reference.
This guarantees that the compiler will not optimize away the memset()
call if it detects a dead store.
Use this to clear SSL passphrases.
No backport needed.
Add a new global keyword, max-threads-per-group. It sets the maximum number of
threads a thread group can contain. Unless the number of thread groups
is fixed with "thread-groups", haproxy will just create more thread
groups as needed.
The default and maximum value is 64.
The RX_F_INHERITED flag was ambiguous, as it was used to mark both
listeners inherited from the parent process and listeners duplicated
from another local receiver. This could lead to incorrect behavior
concerning socket unbinding and suspension.
This commit refactors the handling of inherited listeners by splitting
the RX_F_INHERITED flag into two more specific flags:
- RX_F_INHERITED_FD: Indicates a listener inherited from the parent
process via its file descriptor. These listeners should not be unbound
by the master.
- RX_F_INHERITED_SOCK: Indicates a listener that shares a socket with
another one, either by being inherited from the parent or by being
duplicated from another local listener. These listeners should not be
suspended or resumed individually.
Previously, the sharding code was unconditionally using RX_F_INHERITED
when duplicating a file descriptor. In HAProxy versions prior to 3.1,
this led to a file descriptor leak for duplicated unix stats sockets in
the master process. This would eventually cause the master to crash with
a BUG_ON in fd_insert() once the file descriptor limit was reached.
This must be backported as far as 3.0. Branches earlier than 3.0 are
affected but would need a different patch as the logic is different.
- Convert additional cases to use the automatic alignment feature for
the THREAD_ALIGN(ED) macros. This includes some cases that are less
obviously correct where it seems we wanted to align only in the
USE_THREAD case but were not using the thread specific macros.
- Also move some alignment requirements to the structure definition
instead of having it on variable declaration.
- Use the automatic alignment feature instead of hardcoding 64 all over
the code.
- This also converts a few bare __attribute__((aligned(X))) to using the
ALIGNED macro.
- It is now possible to use the THREAD_ALIGN and THREAD_ALIGNED macros
without a parameter. In this case, we automatically align on the cache
line size.
- The cache line size is set to 64 by default to match the current code,
but it can be overridden on the command line.
- This required moving the DEFVAL/DEFNULL/DEFZERO macros to compiler.h
instead of tools-t.h, to avoid namespace pollution if we included
tools-t.h from compiler.h.
Add a new flag to connections, CO_FL_SSL_NO_CACHED_INFO, and set it for
checks.
It lets the ssl layer know that he should not use cached informations,
such as the ALPN as stored in the server, or cached sessions.
This wlil be used for checks, as checks may target different servers, or
used a different SSL configuration, so we can't assume the stored
informations are correct.
This should be backported to 3.3, and may be backported up to 2.8 if the
attempts to do session resume by checks is proven to be a problem.
Thanks to the previous patch, "BUG/MEDIUM: ssl: Don't reuse TLS session
if the connection's SNI differs", it is no useless to store the SNI of
cached TLS sessions. This SNI is no longer tested and new connections
reusing a session must have the same SNI.
The main change here is for the ssl_sock_set_servername() function. It is no
longer possible to compare the SNI of the reused session with the one of the
new connection. So, the SNI is always set, with no other processing. Mainly,
the session is not destroyed when SNIs don't match. It means the commit
119a4084bf ("BUG/MEDIUM: ssl: for a handshake when server-side SNI changes")
is implicitly reverted.
It is good to note that it is unclear for me when and why the reused session
should be destroyed. Because I'm unable to reproduce any issue fixed by the
commit above.
This patch could be backported as far as 3.0 with the commit above.
When a SNI is set on a new connection, its hash is now saved in the
connection itself. To do so, a dedicated field was added into the connection
strucutre, called sni_hash. For now, this value is only used when the TLS
session is cached.
For cached TLS sessions, in addition to the SNI itself, its hash is now also
saved. No changes are expected here because this hash is not used for now.
This commit relies on:
* MINOR: ssl: Add a function to hash SNIs
This patch only adds the function ssl_sock_sni_hash() that can be used to
get the hash value corresponding to an SNI. A global seed, sni_hash_seed, is
used.
This patch reverts the commit efe60745b ("MINOR: quic: remove connection arg
from qc_new_conn()"). The connection will be mandatory when the QUIC
connection is created on backend side to fix an issue when we try to reuse a
TLS session.
So, the connection is again an argument of qc_new_conn(), the 4th
argument. It is NULL for frontend QUIC connections but there is no special
check on it.
This patch impacts both the QUIC frontends and listeners.
Note that "ssl-default-bind-ciphersuites", "ssl-default-bind-curves",
are not ignored by QUIC by the frontend. This is also the case for the
backends with "ssl-default-server-ciphersuites" and "ssl-default-server-curves".
These settings are set by ssl_sock_prepare_ctx() for the frontends and
by ssl_sock_prepare_srv_ssl_ctx() for the backends. But ssl_quic_initial_ctx()
first sets the default QUIC frontends (see <quic_ciphers> and <quic_groups>)
before these ssl_sock.c function are called, leading some TLS stack to
refuse them if they do not support them. This is the case for some OpenSSL 3.5
stack with FIPS support. They do not support X25519.
To fix this, set the default QUIC ciphersuites and curves only if not already
set by the settings mentioned above.
Rename <quic_ciphers> global variable to <default_quic_ciphersuites>
and <quic_groups> to <default_quic_curves> to reflect the OpenSSL API naming.
These options are taken into an account by ssl_quic_initial_ctx()
which inspects these four variable before calling SSL_CTX_set_ciphersuites()
with <default_quic_ciphersuites> as parameter and SSL_CTX_set_curves() with
<default_quic_curves> as parameter if needed, that is to say, if no ciphersuites
and curves were set by "ssl-default-bind-ciphersuites", "ssl-default-bind-curves"
as global options or "ciphersuites", "curves" as "bind" line options.
Note that the bind_conf struct is not modified when no "ciphersuites" or
"curves" option are used on "bind" lines.
On backend side, rely on ssl_sock_init_srv() to set the server ciphersuites
and curves. This function is modified to use respectively <default_quic_ciphersuites>
and <default_quic_curves> if no ciphersuites and curves were set by
"ssl-default-server-ciphersuites", "ssl-default-server-curves" as global options
or "ciphersuites", "curves" as "server" line options.
Thank to @rwagoner for having reported this issue in GH #3194 when using
an OpenSSL 3.5.4 stack with FIPS support.
Must be backported as far as 2.6
This bug was revealed on backend side by reg-tests/ssl/del_ssl_crt-list.vtc when
run wich QUIC connections. As expected by the test, a TLS alert is generated on
servsr side. This latter sands a CONNECTION_CLOSE frame with a CRYPTO error
(>= 0x100). In this case the client closes its QUIC connection. But
the stream connection was not informed. This leads the connection to
be closed after the server timeout expiration. It shouls be closed asap.
This is the reason why reg-tests/ssl/del_ssl_crt-list.vtc could succeeds
or failed, but only after a 5 seconds delay.
To fix this, mimic the ssl_sock_io_cb() for TCP/SSL connections. Call
the same code this patch implements with ssl_sock_handle_hs_error()
to correctly handle the handshake errors. Note that some SSL counters
were not incremented for both the backends and frontends. After such
errors, ssl_sock_io_cb() start the mux after the connection has been
flagged in error. This has as side effect to close the stream
in conn_create_mux().
Must be backported to 3.3 only for backends. This is not sure at this time
if this bug may impact the frontends.
Extend QUIC server configuration so that congestion algorithm and
maximum window size can be set on the server line. This can be achieved
using quic-cc-algo keyword with a syntax similar to a bind line.
This should be backported up to 3.3 as this feature is considered as
necessary for full QUIC backend support. Note that this relies on the
serie of previous commits which should be picked first.
Each QUIC congestion algorithm is defined as a structure with callbacks
in it. Every quic_conn has a member pointing to the configured
algorithm, inherited from the bind-conf keyword or to the default CUBIC
value.
Convert all these definitions to const. This ensures that there never
will be an accidental modification of a globally shared structure. This
also requires to mark quic_cc_algo field in bind_conf and quic_cc as
const.
