On "add server", if a GUID is defined, guid_insert() is used to add the
entry into the global GUID tree. If a similar entry already exists, GUID
insertion fails and the server creation is eventually aborted.
A crash could occur in this case because of an invalid memory access via
guid_remove(). The latter is caused via free_server() as the server
insertion is rejected. The invalid occurs on GUID key.
The issue occurs because of guid_insert(). The function properly
deallocates the GUID key on duplicate insertion, but it failed to reset
<guid.node.key> to NULL. This caused the invalid memory access on
guid_remove(). To fix this, ensure that key member is properly resetted
on guid_insert() error path.
This must be backported up to 3.0.
Implement stress mode on "add server help". This ensures that the
command is fully reentrant on full output buffer.
For testing, it requires compilation with USE_STRESS and global setting
"stress-level 1".
Implement "help" as a sub-command for "add server" CLI. The objective is
to list all the keywords that are supported for dynamic servers. CLI IO
handler and add_srv_ctx are used to support reentrancy on full output
buffer.
Now that this command is implemented, the outdated keyword list on "add
server" from management documentation can be removed.
Extend "add server" to support an IO handler function named
cli_io_handler_add_server(). A context object is also defined whose
usage will depend on IO handler capabilities.
IO handler is skipped when "add server" is run in default mode, i.e. on
a dynamic server creation. Thus, currently IO handler is unneeded.
However, it will become useful to support sub-commands for "add server".
Note that return value of "add server" parser has been changed on server
creation success. Previously, it was used incorrectly to report if
server was inserted or not. In fact, parser return value is used by CLI
generic code to detect if command processing has been completed, or
should continue to the IO handler. Now, "add server" always returns 1 to
signal that CLI processing is completed. This is necessary to preserve
CLI output emitted by parser, even now that IO handler is defined for
the command. Previously, output was emitted in every situations due to
IO handler not defined. See below code snippet from cli.c for a better
overview :
if (kw->parse && kw->parse(args, payload, appctx, kw->private) != 0) {
ret = 1;
goto fail;
}
/* kw->parse could set its own io_handler or io_release handler */
if (!appctx->cli_ctx.io_handler) {
ret = 1;
goto fail;
}
appctx->st0 = CLI_ST_CALLBACK;
ret = 1;
goto end;
Currently there is "no-tls-tickets" that is also supported in the
ssl-default-bind-options directive, but there's no way to re-enable
them on a specific "bind" line. This patch simply provides the option
to re-enable them. Note that the flag is inverted because tickets are
enabled by default and the no-tls-ticket option sets the flag to
disable them.
Several users already reported that it would be nice to support
strict-sni in ssl-default-bind-options. However, in order to support
it, we also need an option to disable it.
This patch moves the setting of the option from the strict_sni field
to a flag in the ssl_options field so that it can be inherited from
the default bind options, and adds a new "no-strict-sni" directive to
allow to disable it on a specific "bind" line.
The test file "del_ssl_crt-list.vtc" which already tests both options
was updated to make use of the default option and the no- variant to
confirm everything continues to work.
In the Prometheus exporter, the last health check status is already exposed,
with its code and duration in seconds. The server status is also exposed.
But the information about the agent check are not available. It is not
really handy because when a server status is changed because of the agent,
it is not obvious by looking to the Prometheus metrics. Indeed, the server
may reported as DOWN for instance, while the health check status still
reports a success. Being able to get the agent status in that case could be
valuable.
So now, the last agent check status is exposed, with its code and duration
in seconds. Following metrics can be grabbe now:
* haproxy_server_agent_status
* haproxy_server_agent_code
* haproxy_server_agent_duration_seconds
Note that unlike the other metrics, no per-backend aggregated metric is
exposed.
This patch is related to issue #2983.
It was mentioned during the development of glitches that it would be
nice to support not killing misbehaving connections below a certain
CPU usage so that poor implementations that routinely misbehave without
impact are not killed. This is now possible by setting a CPU usage
threshold under which we don't kill them via this parameter. It defaults
to zero so that we continue to kill them by default.
Insert some missing includes statement in QUIC source files. This was
detected after the next commit which adjust the include list used in
quic_conn-t.h file.
quic-conn layer has to handle itself STREAM frames after MUX release. If
the stream was already seen, it is probably only a retransmitted frame
which can be safely ignored. For other streams, an active closure may be
needed.
Thus it's necessary that quic-conn layer knows the highest stream ID
already handled by the MUX after its release. Previously, this was done
via <nb_streams> member array in quic-conn structure.
Refactor this by replacing <nb_streams> by two members called
<stream_max_uni>/<stream_max_bidi>. Indeed, it is unnecessary for
quic-conn layer to monitor locally opened uni streams, as the peer
cannot by definition emit a STREAM frame on it. Also, bidirectional
streams are always opened by the remote side.
Previously, <nb_streams> were set by quic-stream layer. Now,
<stream_max_uni>/<stream_max_bidi> members are only set one time, just
prior to QUIC MUX release. This is sufficient as quic-conn do not use
them if the MUX is available.
Note that previously, IDs were used relatively to their type, thus
incremented by 1, after shifting the original value. For simplification,
use the plain stream ID, which is incremented by 4.
Move general function to check if a stream is uni or bidirectional from
QUIC MUX to quic_utils module. This should prevent unnecessary include
of QUIC MUX header file in other sources.
Stop emitting \n in errmsg for intermediate error messages, this was
emitting multiline logs and was returning to a new line in the middle of
sentences.
We don't need to emit them in acme_start_task() since the errmsg is
ouput in a send_log which already contains a \n or on the CLI which
also emits it.
When starting the ACME task with a ckch_conf which does not contain the
domains, the ACME task would segfault because it will try to dereference
a NULL in this case.
The patch fix the issue by emitting a warning when no domains are
configured. It's not done at configuration parsing because it is not
easy to emit the warning because there are is no callback system which
give access to the whole ckch_conf once a line is parsed.
No backport needed.
RX buffer allocation has been reworked in current dev tree. The
objective is to support multiple buffers per QCS to improve upload
throughput.
RX buffer allocation failure is handled simply : the whole connection is
closed. This is done via qcc_set_error(), with INTERNAL_ERROR as error
code. This function contains a BUG_ON() to ensure it is called only one
time per connection instance.
On RX buffer alloc failure, the aformentioned BUG_ON() crashes due to a
double invokation of qcc_set_error(). First by qcs_get_rxbuf(), and
immediately after it by qcc_recv(), which is the caller of the previous
one. This regression was introduced by the following commit.
60f64449fb
MAJOR: mux-quic: support multiple QCS RX buffers
To fix this, simply remove qcc_set_error() invocation in
qcs_get_rxbuf(). On buffer alloc failture, qcc_recv() is responsible to
set the error.
This does not need to be backported.
The build failed on mips32 with a 64-bit time_t here:
https://github.com/haproxy/haproxy/actions/runs/15150389164/job/42595310111
Let's just turn the "remain" variable used to show the remaining time
into a more portable ullong and use %llu for all format specifiers,
since long remains limited to 32-bit on 32-bit archs.
No backport needed.
When building on MIPS-32 with gcc-9.5 and glibc-2.31, I got this:
src/ssl_trace.c: In function 'ssl_trace':
src/ssl_trace.c:118:42: warning: format '%ld' expects argument of type 'long int', but argument 3 has type 'ssize_t' {aka 'const int'} [-Wformat=]
118 | chunk_appendf(&trace_buf, " : size=%ld", *size);
| ~~^ ~~~~~
| | |
| | ssize_t {aka const int}
| long int
| %d
Let's just cast the type. No backport needed.
With commit a06c215f08 ("MEDIUM: wdt: always make the faulty thread
report its own warnings"), when the TH_FL_STUCK flag was flipped on,
we'd then go to the panic code instead of giving a second chance like
before the commit. This can trigger rare cases that only happen with
moderate loads like was addressed by commit 24ce001771 ("BUG/MEDIUM:
wdt: fix the stuck detection for warnings"). This is in fact due to
the loss of the common "goto update_and_leave" that used to serve
both the warning code and the flag setting for probation, and it's
apparently what hit Christian in issue #2980.
Let's make sure we exit naturally when turning the bit on for the
first time. Let's also update the confusing comment at the end of
the check that was left over by latest change.
Since the first commit was backported to 3.1, this commit should be
backported there as well.
For an unidentified reason, SSL_do_hanshake() succeeds at its first call when 0-RTT
is enabled for the connection. This behavior looks very similar by the one encountered
by AWS-LC stack. That said, it was documented by AWS-LC. This issue leads the
connection to stop sending handshake packets after having release the handshake
encryption level. In fact, no handshake packets could even been sent leading
the handshake to always fail.
To fix this, this patch simulates a "handshake in progress" state waiting
for the application level read secret to be established by the TLS stack.
This may happen only after the QUIC listener has completed/confirmed the handshake
upon handshake CRYPTO data receipt from the peer.
A QUIC must sent its transport parameter using a TLS custom extention. This
extension is reset by SSL_set_SSL_CTX(). It can be restored calling
quic_ssl_set_tls_cbs() (which calls SSL_set_quic_tls_cbs()).
The quic_conn struct is modified for two reasons. The first one is to store
the encoded version of the local tranport parameter as this is done for
USE_QUIC_OPENSSL_COMPAT. Indeed, the local transport parameter "should remain
valid until after the parameters have been sent" as mentionned by
SSL_set_quic_tls_cbs(3) manual. In our case, the buffer is a static buffer
attached to the quic_conn object. qc_ssl_set_quic_transport_params() function
whose role is to call SSL_set_tls_quic_transport_params() (aliased by
SSL_set_quic_transport_params() to set these local tranport parameter into
the TLS stack from the buffer attached to the quic_conn struct.
The second quic_conn struct modification is the addition of the new ->prot_level
(SSL protection level) member added to the quic_conn struct to store "the most
recent write encryption level set via the OSSL_FUNC_SSL_QUIC_TLS_yield_secret_fn
callback (if it has been called)" as mentionned by SSL_set_quic_tls_cbs(3) manual.
This patches finally implements the five remaining callacks to make the haproxy
QUIC implementation work.
OSSL_FUNC_SSL_QUIC_TLS_crypto_send_fn() (ha_quic_ossl_crypto_send) is easy to
implement. It calls ha_quic_add_handshake_data() after having converted
qc->prot_level TLS protection level value to the correct ssl_encryption_level_t
(boringSSL API/quictls) value.
OSSL_FUNC_SSL_QUIC_TLS_crypto_recv_rcd_fn() (ha_quic_ossl_crypto_recv_rcd())
provide the non-contiguous addresses to the TLS stack, without releasing
them.
OSSL_FUNC_SSL_QUIC_TLS_crypto_release_rcd_fn() (ha_quic_ossl_crypto_release_rcd())
release these non-contiguous buffer relying on the fact that the list of
encryption level (qc->qel_list) is correctly ordered by SSL protection level
secret establishements order (by the TLS stack).
OSSL_FUNC_SSL_QUIC_TLS_yield_secret_fn() (ha_quic_ossl_got_transport_params())
is a simple wrapping function over ha_quic_set_encryption_secrets() which is used
by boringSSL/quictls API.
OSSL_FUNC_SSL_QUIC_TLS_got_transport_params_fn() (ha_quic_ossl_got_transport_params())
role is to store the peer received transport parameters. It simply calls
quic_transport_params_store() and set them into the TLS stack calling
qc_ssl_set_quic_transport_params().
Also add some comments for all the OpenSSL 3.5 QUIC API callbacks.
This patch have no impact on the other use of QUIC API provided by the others TLS
stacks.
This patch allows the use of the new OpenSSL 3.5.0 QUIC TLS API when it is
available and detected at compilation time. The detection relies on the presence of the
OSSL_FUNC_SSL_QUIC_TLS_CRYPTO_SEND macro from openssl-compat.h. Indeed this
macro is defined by OpenSSL since 3.5.0 version. It is not defined by quictls.
This helps in distinguishing these two TLS stacks. When the detection succeeds,
HAVE_OPENSSL_QUIC is also defined by openssl-compat.h. Then, this is this new macro
which is used to detect the availability of the new OpenSSL 3.5.0 QUIC TLS API.
Note that this detection is done only if USE_QUIC_OPENSSL_COMPAT is not asked.
So, USE_QUIC_OPENSSL_COMPAT and HAVE_OPENSSL_QUIC are exclusive.
At the same location, from openssl-compat.h, ssl_encryption_level_t enum is
defined. This enum was defined by quictls and expansively used by the haproxy
QUIC implementation. SSL_set_quic_transport_params() is replaced by
SSL_set_quic_tls_transport_params. SSL_set_quic_early_data_enabled() (quictls) is also replaced
by SSL_set_quic_tls_early_data_enabled() (OpenSSL). SSL_quic_read_level() (quictls)
is not defined by OpenSSL. It is only used by the traces to log the current
TLS stack decryption level (read). A macro makes it return -1 which is an
usused values.
The most of the differences between quictls and OpenSSL QUI APIs are in quic_ssl.c
where some callbacks must be defined for these two APIs. This is why this
patch modifies quic_ssl.c to define an array of OSSL_DISPATCH structs: <ha_quic_dispatch>.
Each element of this arry defines a callback. So, this patch implements these
six callabcks:
- ha_quic_ossl_crypto_send()
- ha_quic_ossl_crypto_recv_rcd()
- ha_quic_ossl_crypto_release_rcd()
- ha_quic_ossl_yield_secret()
- ha_quic_ossl_got_transport_params() and
- ha_quic_ossl_alert().
But at this time, these implementations which must return an int return 0 interpreted
as a failure by the OpenSSL QUIC API, except for ha_quic_ossl_alert() which
is implemented the same was as for quictls. The five remaining functions above
will be implemented by the next patches to come.
ha_quic_set_encryption_secrets() and ha_quic_add_handshake_data() have been moved
to be defined for both quictls and OpenSSL QUIC API.
These callbacks are attached to the SSL objects (sessions) calling qc_ssl_set_cbs()
new function. This latter callback the correct function to attached the correct
callbacks to the SSL objects (defined by <ha_quic_method> for quictls, and
<ha_quic_dispatch> for OpenSSL).
The calls to SSL_provide_quic_data() and SSL_process_quic_post_handshake()
have been also disabled. These functions are not defined by OpenSSL QUIC API.
At this time, the functions which call them are still defined when HAVE_OPENSSL_QUIC
is defined.
There were no traces to diagnose qc_ssl_sess_init() failures from QUIC traces.
This patch add calls to TRACE_DEVEL() into qc_ssl_sess_init() and its caller
(qc_alloc_ssl_sock_ctx()). This was useful at least to diagnose SSL context
initialization failures when porting QUIC to the new OpenSSL 3.5 QUIC API.
Should be easily backported as far as 2.6.
This code is there from QUIC implementation start. It was supposed to
initialize <ha_quic_meth> as a BIO_METHOD static object. But this
BIO_METHOD is not used at all!
Should be backported as far as 2.6 to help integrate the next patches to come.
Output a sink message when the certificate was renewed by the ACME
client.
The message is emitted on the "dpapi" sink, and ends by \n\0.
Since the message contains this binary character, the right -0 parameter
must be used when consulting the sink over the CLI:
Example:
$ echo "show events dpapi -nw -0" | socat -t9999 /tmp/haproxy.sock -
<0>2025-05-19T15:56:23.059755+02:00 acme newcert foobar.pem.rsa\n\0
When used with the master CLI, @@1 should be used instead of @1 in order
to keep the connection to the worker.
Example:
$ echo "@@1 show events dpapi -nw -0" | socat -t9999 /tmp/master.sock -
<0>2025-05-19T15:56:23.059755+02:00 acme newcert foobar.pem.rsa\n\0
On ARM with 80 cores and a single server, it's sometimes possible to see
a segfault in fwlc_get_next_server() around 600-700k RPS. It seldom
happens as well on x86 with 128 threads with the same config around 1M
rps. It turns out that in fwlc_get_next_server(), before calling
fwlc_srv_reposition(), we have to drop the lock and that one takes it
back again.
The problem is that anything can happen to our node during this time,
and it can be freed. Then when continuing our work, we later iterate
over it and its next to find a node with an acceptable key, and by
doing so we can visit either uninitialized memory or simply nodes that
are no longer in the tree.
A first attempt at fixing this consisted in artificially incrementing
the elements count before dropping the lock, but that turned out to be
even worse because other threads could loop forever on such an element
looking for an entry that does not exist. Maintaining a separate
refcount didn't work well either, and it required to deal with the
memory release while dropping it, which is really not convenient.
Here we're taking a different approach consisting in simply not
trusting this node anymore and going back to the beginning of the
loop, as is done at a few other places as well. This way we can
safely ignore the possibly released node, and the test runs reliably
both on the arm and the x86 platforms mentioned above. No performance
regression was observed either, likely because this operation is quite
rare.
No backport is needed since this appeared with the leastconn rework
in 3.2.
If there is an alloc failure during qc_new_conn(), cleaning is done via
quic_conn_release(). However, since the below commit, an unchecked
dereferencing of <qc.path> is performed in the latter.
e841164a44
MINOR: quic: account for global congestion window
To fix this, simply check <qc.path> before dereferencing it in
quic_conn_release(). This is safe as it is properly initialized to NULL
on qc_new_conn() first stage.
This does not need to be backported.
The queue length was moved to its own variable in commit 583303c48
("MINOR: proxies/servers: Calculate queueslength and use it."), however a
few places were missed in pendconn_unlink() and assign_server_and_queue()
resulting in never decreasing counts on aborted streams. This was
reproduced when injecting more connections than the total backend
could stand in TCP mode and letting some of them time out in the
queue. No backport is needed, this is only 3.2.
Since commit 9fe72bba3 ("MAJOR: leastconn; Revamp the way servers are
ordered."), there's no way to escape the loop visiting the mt_list heads
in fwlc_get_next_server if all servers in the list are saturated,
resulting in a watchdog panic. It can be reproduced with this config
and injecting with more than 2 concurrent conns:
balance leastconn
server s1 127.0.0.1:8000 maxconn 1
server s2 127.0.0.1:8000 maxconn 1
Here we count the number of saturated servers that were encountered, and
escape the loop once the number of remaining servers exceeds the number
of saturated ones. No backport is needed since this arrived in 3.2.
Building with clang 16 on MIPS64 yields this warning:
src/slz.c:931:24: warning: unused function 'crc32_uint32' [-Wunused-function]
static inline uint32_t crc32_uint32(uint32_t data)
^
Let's guard it using UNALIGNED_LE_OK which is the only case where it's
used. This saves us from introducing a possibly non-portable attribute.
This is libslz upstream commit f5727531dba8906842cb91a75c1ffa85685a6421.
Calling slz_rfc1950_finish() without emitting any data would result in
incorrectly emitting a gzip header (rfc1952) instead of a zlib header
(rfc1950) due to a copy-paste between the two wrappers. The impact is
almost inexistent since the zlib format is almost never used in this
context, and compressing totally empty messages is quite rare as well.
Let's take this opportunity for fixing another mistake on an RFC number
in a comment.
This is slz upstream commit 7f3fce4f33e8c2f5e1051a32a6bca58e32d4f818.
The current hash involves 3 simple shifts and additions so that it can
be mapped to a multiply on architecures having a fast multiply. This is
indeed what the compiler does on x86_64. A large range of values was
scanned to try to find more optimal factors on machines supporting such
a fast multiply, and it turned out that new factor 0x1af42f resulted in
smoother hashes that provided on average 0.4% better compression on both
the Silesia corpus and an mbox file composed of very compressible emails
and uncompressible attachments. It's even slightly better than CRC32C
while being faster on Skylake. This patch enables this factor on archs
with a fast multiply.
This is slz upstream commit 82ad1e75c13245a835c1c09764c89f2f6e8e2a40.
If building for sse4 and USE_CRC32C_HASH is defined, then we can use
crc32c to calculate the lookup hash. By default we don't do it because
even on skylake it's slower than the current hash, which only involves
a short multiply (~5% slower). But the gains are marginal (0.3%).
This is slz upstream commit 44ae4f3f85eb275adba5844d067d281e727d8850.
Note: this is not used by default and only merged in order to avoid
divergence between the code bases.
On 64-bit platforms, disassembling the code shows that send_huff() performs
a left shift followed by a right one, which are the result of integer
truncation and zero-extension caused solely by using different types at
different levels in the call chain. By making encode24() take a 64-bit
int on input and send_huff() take one optionally, we can remove one shift
in the hot path and gain 1% performance without affecting other platforms.
This is slz upstream commit fd165b36c4621579c5305cf3bb3a7f5410d3720b.
In continuation with 9a05c1f574 ("BUG/MEDIUM: h2/h3: reject some
forbidden chars in :authority before reassembly") and the discussion
in issue #2941, @DemiMarie rightfully suggested that Host should also
be sanitized, because it is sometimes used in concatenation, such as
this:
http-request set-url https://%[req.hdr(host)]%[pathq]
which was proposed as a workaround for h2 upstream servers that require
:authority here:
https://www.mail-archive.com/haproxy@formilux.org/msg43261.html
The current patch then adds the same check for forbidden chars in the
Host header, using the same function as for the patch above, since in
both cases we validate the host:port part of the authority. This way
we won't reconstruct ambiguous URIs by concatenating Host and path.
Just like the patch above, this can be backported afer a period of
observation.
Let's make sure we drop extraneous Host headers after having compared
them. That also works when :authority was already present. This way,
like for h1 and h2, we only keep one copy of it, while still making
sure that Host matches :authority. This way, if a request has both
:authority and Host, only one Host header will be produced (from
:authority). Note that due to the different organization of the code
and wording along the evolving RFCs, here we also check that all
duplicates are identical, while h2 ignores them as per RFC7540, but
this will be re-unified later.
This should be backported to stable versions, at least 2.8, though
thanks to the existing checks the impact is probably nul.
It is especially a problem with Lua filters, but it is important to disable
the 0-copy forwarding if a filter alters the payload, or at least to be able
to disable it. While the filter is registered on the data filtering, it is
not an issue (and it is the common case) because, there is now way to
fast-forward data at all. But it may be an issue if a filter decides to
alter the payload and to unregister from data filtering. In that case, the
0-copy forwarding can be re-enabled in a hardly precdictable state.
To fix the issue, a SC flags was added to do so. The HTTP compression filter
set it and lua filters too if the body length is changed (via
HTTPMessage.set_body_len()).
Note that it is an issue because of a bad design about the HTX. Many info
about the message are stored in the HTX structure itself. It must be
refactored to move several info to the stream-endpoint descriptor. This
should ease modifications at the stream level, from filter or a TCP/HTTP
rules.
This should be backported as far as 3.0. If necessary, it may be backported
on lower versions, as far as 2.6. In that case, it must be reviewed and
adapted.
There was no function for a lua filter to change the body length of an HTTP
Message. But it is mandatory to be able to alter the message payload. It is
not possible update to directly update the message headers because the
internal state of the message must also be updated accordingly.
It is the purpose of HTTPMessage.set_body_len() function. The new body
length myst be passed as argument. If it is an integer, the right
"Content-Length" header is set. If the "chunked" string is used, it forces
the message to be chunked-encoded and in that case the "Transfer-Encoding"
header.
This patch should fix the issue #2837. It could be backported as far as 2.6.
There's a configurable limit to the number of messages sent to a
peer (tune.peers.max-updates-at-once), but this one is not applied to
the receive side. While it can usually be OK with default settings,
setups involving a large tune.bufsize (1MB and above) regularly
experience high latencies and even watchdogs during reloads because
the full learning process sends a lot of data that manages to fill
the entire buffer, and due to the compactness of the protocol, 1MB
of buffer can contain more than 100k updates, meaning taking locks
etc during this time, which is not workable.
Let's make sure the receiving side also respects the max-updates-at-once
setting. For this it counts incoming updates, and refrains from
continuing once the limit is reached. It's a bit tricky to do because
after receiving updates we still have to send ours (and possibly some
ACKs) so we cannot just leave the loop.
This issue was reported on 3.1 but it should progressively be backported
to all versions having the max-updates-at-once option available.
using "send-proxy" or "send-proxy-v2" option on a ring server is not
relevant nor supported. Worse, on 2.4 it causes haproxy process to
crash as reported in GH #2965.
Let's be more explicit about the fact that this keyword is not supported
under "ring" context by ignoring the option and emitting a warning message
to inform the user about that.
Ideally, we should do the same for peers and log servers. The proper way
would be to check servers options during postparsing but we currently lack
proper cross-type server postparsing hooks. This will come later and thus
will give us a chance to perform the compatibilty checks for server
options depending on proxy type. But for now let's simply fix the "ring"
case since it is the only one that's known to cause a crash.
It may be backported to all stable versions.
Some trace messages were not really accurrate, reporting a CLOSED connection
while only an error was reported on it. In addition, an TRACE_ERROR() was
used to report a short read on HELLO/DISCONNECT frames header. But it is not
an error. a TRACE_DEVEL() should be used instead.
This patch could be backported to 3.1 to ease future backports.
When an read error is detected, no error must be reported on the SPOP
connection is there are still some data to parse. It is important to be sure
to process all data before reporting the error and be sure to not truncate
received frames. However, we must also take care to handle short read case
to not wait data that will never be received.
This patch must be backported to 3.1.
There was no test in the demux part to detect truncated frames and to report
an error at the connection level. The SPOP streams were properly switch to
half-closed state. But waiting the associated SPOE applets were woken up and
released, the SPOP connection could be woken up several times for nothing. I
never triggered the watchdog in that case, but it is not excluded.
Now, at the end of the demux function, if a specific test was added to
detect truncated frames to report an error and close the connection.
This patch must be backported to 3.1.
When a frame was not fully received, a short read must be reported on the
SPOP connection to help the demux to handle truncated frames. This was
performed for frames truncated on the header part but not on the payload
part. It is now properly detected.
This patch must be backported to 3.1.
The CLOSING state was not handled at all by the SPOP multiplexer while it is
mandatory when a DISCONNECT frame was sent and the mux should wait for the
DISCONNECT frame in reply from the agent. Thanks to this patch, it should be
fixed.
In addition, if an error occurres during the AGENT HELLO frame parsing, the
SPOP connection is no longer switched to CLOSED state and remains in ERROR
state instead. It is important to be able to send the DISCONNECT frame to
the agent instead of closing the TCP connection immediately.
This patch depends on following commits:
* BUG/MEDIUM: mux-spop: Remove frame parsing states from the SPOP connection state
* MINOR: mux-spop: Don't set SPOP connection state to FRAME_H after ACK parsing
* BUG/MINOR: mux-spop: Don't open new streams for SPOP connection on error
* BUG/MINOR: mux-spop: Make the demux stream ID a signed integer
All the series must be backported to 3.1.
SPOP_CS_FRAME_H and SPOP_CS_FRAME_P states, that were used to handle frame
parsing, were removed. The demux process now relies on the demux stream ID
to know if it is waiting for the frame header or the frame
payload. Concretly, when the demux stream ID is not set (dsi == -1), the
demuxer is waiting for the next frame header. Otherwise (dsi >= 0), it is
waiting for the frame payload. It is especially important to be able to
properly handle DISCONNECT frames sent by the agents.
SPOP_CS_RUNNING state is introduced to know the hello handshake was finished
and the SPOP connection is able to open SPOP streams and exchange NOTIFY/ACK
frames with the agents.
It depends on the following fixes:
* MINOR: mux-spop: Don't set SPOP connection state to FRAME_H after ACK parsing
* BUG/MINOR: mux-spop: Make the demux stream ID a signed integer
This change will be mandatory for the next fix. It must be backported to 3.1
with the commits above.
After the ACK frame was parsed, it is useless to set the SPOP connection
state to SPOP_CS_FRAME_H state because this will be automatically handled by
the demux function. If it is not an issue, but this will simplify changes
for the next commit.
Till now, only SPOP connections fully closed or those with a TCP connection on
error were concerned. But available streams could be reported for SPOP
connections in error or closing state. But in these states, no NOTIFY frames
will be sent and no ACK frames will be parsed. So, no new SPOP streams should be
opened.
This patch should be backported to 3.1.
The demux stream ID of a SPOP connection, used when received frames are
parsed, must be a signed integer because it is set to -1 when the SPOP
connection is initialized. It will be important for the next fix.
This patch must be backported to 3.1.
When a SPOP connection was closed or was in error, an error was
systematically reported on all its SPOP streams. However, SPOP streams that
already received their ACK frame must be excluded. Otherwise if an agent
sends a ACK and close immediately, the ACK will be ignored because the SPOP
stream will handle the error first.
This patch must be backported to 3.1.
When the SPOE applet was released, if a SPOE filter context was still
attached to it, an error was reported to the filter. However, there is no
reason to report an error if the ACK message was already received. Because
of this bug, if the ACK message is received and the SPOE connection is
immediately closed, this prevents the ACK message to be processed.
This patch should be backported to 3.1.
When the channel API was revisted, the both functions above was added. An
offset can be passed as argument. However, this parameter could be reported
to be out of range if there was not enough input data was received yet. It
is an issue, especially with a tcp rule, because more data could be
received. If an error is reported too early, this prevent the rule to be
reevaluated later. In fact, an error should only be reported if the offset
is part of the output data.
Another issue is about the conditions to report 'nil' instead of an empty
string. 'nil' was reported when no data was found. But it is not aligned
with the documentation. 'nil' must only be returned if no more data cannot
be received and there is no input data at all.
This patch should fix the issue #2716. It should be backported as far as 2.6.
Most of the time, machines made of multiple CPU types use the same L3
for them, and grouping CPUs by frequencies to form groups doesn't bring
any value and on the opposite can impair the incoming connection balancing.
This choice of grouping by cluster was made in order to constitute a good
choice on homogenous machines as well, so better rely on the per-CCX
grouping than the per-cluster one in this case. This will create less
clusters on machines where it counts without affecting other ones.
It doesn't seem necessary to change anything for the "resource" policy
since it selects a single cluster.
This is similar to the previous change to the "performance" policy but
it applies to the "efficiency" one. Here we're changing the sorting
method to sort CPU clusters by average per-CPU capacity, and we evict
clusters whose per-CPU capacity is above 125% of the previous one.
Per-core capacity allows to detect discrepancies between CPU cores,
and to continue to focus on efficient ones as a priority.
Running the "performance" policy on highly heterogenous systems yields
bad choices when there are sufficiently more small than big cores,
and/or when there are multiple cluster types, because on such setups,
the higher the frequency, the lower the number of cores, despite small
differences in frequencies. In such cases, we quickly end up with
"performance" only choosing the small or the medium cores, which is
contrary to the original intent, which was to select performance cores.
This is what happens on boards like the Orion O6 for example where only
the 4 medium cores and 2 big cores are choosen, evicting the 2 biggest
cores and the 4 smallest ones.
Here we're changing the sorting method to sort CPU clusters by average
per-CPU capacity, and we evict clusters whose per-CPU capacity falls
below 80% of the previous one. Per-core capacity allows to detect
discrepancies between CPU cores, and to continue to focus on high
performance ones as a priority.
The current per-capacity sorting function acts on a whole cluster, but
in some setups having many small cores and few big ones, it becomes
easy to observe an inversion of metrics where the many small cores show
a globally higher total capacity than the few big ones. This does not
necessarily fit all use cases. Let's add new a function to sort clusters
by their per-cpu average capacity to cover more use cases.
This cpu-policy will only consider CCX and not clusters. This makes
a difference on machines with heterogenous CPUs that generally share
the same L3 cache, where it's not desirable to create multiple groups
based on the CPU types, but instead create one with the different CPU
types. The variants "group-by-2/3/4-ccx" have also been added.
Let's also add some text explaining the difference between cluster
and CCX.
Some (rare) boards have their clusters in an erratic order. This is
the case for the Radxa Orion O6 where one of the big cores appears as
CPU0 due to booting from it, then followed by the small cores, then the
medium cores, then the remaining big cores. This results in clusters
appearing this order: 0,2,1,0.
The core in cpu_policy_group_by_cluster() expected ordered clusters,
and performs ordered comparisons to decide whether a CPU's cluster has
already been taken care of. On the board above this doesn't work, only
clusters 0 and 2 appear and 1 is skipped.
Let's replace the cluster number comparison with a cpuset to record
which clusters have been taken care of. Now the groups properly appear
like this:
Tgrp/Thr Tid CPU set
1/1-2 1-2 2: 0,11
2/1-4 3-6 4: 1-4
3/1-6 7-12 6: 5-10
No backport is needed, this is purely 3.2.
Complete stream output for "show quic" by displaying information from
its upper QCS. Note that QCS may be NULL if already released, so a
default output is also provided.
Add a new format for "show quic" command labelled as "stream". This is
an equivalent of "show sess", dedicated to the QUIC stack. Each active
QUIC streams are listed on a line with their related infos.
The main objective of this command is to ensure there is no freeze
streams remaining after a transfer.
Add counters to measure Rx buffers usage per QCS. This reused the newly
defined bdata_ctr type already used for Tx accounting.
Note that for now, <tot> value of bdata_ctr is not used. This is because
it is not easy to account for data accross contiguous buffers.
These values are displayed both on log/traces and "show quic" output.
Add accounting at qc_stream_desc level to be able to report the number
of allocated Tx buffers and the sum of their data. This represents data
ready for emission or already emitted and waiting on ACK.
To simplify this accounting, a new counter type bdata_ctr is defined in
quic_utils.h. This regroups both buffers and data counter, plus a
maximum on the buffer value.
These values are now displayed on QCS info used both on logline and
traces, and also on "show quic" output.
As discussed here:
https://github.com/httpwg/http2-spec/pull/936https://github.com/haproxy/haproxy/issues/2941
It's important to take care of some special characters in the :authority
pseudo header before reassembling a complete URI, because after assembly
it's too late (e.g. the '/'). This patch does this, both for h2 and h3.
The impact on H2 was measured in the worst case at 0.3% of the request
rate, while the impact on H3 is around 1%, but H3 was about 1% faster
than H2 before and is now on par.
It may be backported after a period of observation, and in this case it
relies on this previous commit:
MINOR: http: add a function to validate characters of :authority
Thanks to @DemiMarie for reviving this topic in issue #2941 and bringing
new potential interesting cases.
Last commit 7361515 ("BUG/MINOR: server: dont depend on proxy for server
cleanup in srv_drop()") introduced a regression because the lbprm
server_deinit is not evaluated anymore with dynamic servers, possibly
resulting in a memory leak.
To fix the issue, in addition to free_proxy(), the server deinit check
should be manually performed in cli_parse_delete_server() as well.
No backport needed.
In commit b5ee8bebfc ("MINOR: server: always call ssl->destroy_srv when
available"), we made it so srv_drop() doesn't depend on proxy to perform
server cleanup.
It turns out this is now mandatory, because during deinit, free_proxy()
can occur before the final srv_drop(). This is the case when using Lua
scripts for instance.
In 2a9436f96 ("MINOR: lbprm: Add method to deinit server and proxy") we
added a freeing check under srv_drop() that depends on the proxy.
Because of that UAF may occur during deinit when using a Lua script that
manipulate server objects.
To fix the issue, let's perform the lbprm server deinit logic under
free_proxy() directly, where the DEINIT server hooks are evaluated.
Also, to prevent similar bugs in the future, let's explicitly document
in srv_drop() that server cleanups should assume that the proxy may
already be freed.
No backport needed unless 2a9436f96 is.
OSS Fuzz found that the previous fix ebb19fb367 ("BUG/MINOR: cfgparse:
consider the special case of empty arg caused by \x00") was incomplete,
as the output can sometimes be larger than the input (due to variables
expansion) in which case the work around to try to report a bad arg will
fail. While the parse_line() function has been made more robust now in
order to avoid this condition, let's fix the handling of this special
case anyway by just pointing to the beginning of the line if the supposed
error location is out of the line's buffer.
All details here:
https://oss-fuzz.com/testcase-detail/5202563081502720
No backport is needed unless the fix above is backported.
The fix in 10e6d0bd57 ("BUG/MINOR: tools: only fill first empty arg when
not out of range") was not that good. It focused on protecting against
<arg> becoming out of range to detect we haven't emitted anything, but
it's not the right way to detect this. We're always maintaining arg_start
as a copy of outpos, and that later one is incremented when emitting a
char, so instead of testing args[arg] against out+arg_start, we should
instead check outpos against arg_start, thereby eliminating the <out>
offset and the need to access args[]. This way we now always know if
we've emitted an empty arg without dereferencing args[].
There's no need to backport this unless the fix above is also backported.
When thread support is disabled ("USE_THREAD=" or "USE_THREAD=0" when
building), soft-stop doesn't work as haproxy never ends after stopping
the proxies.
This used to work fine in the past but suddenly stopped working with
ef422ced91 ("MEDIUM: thread: make stopping_threads per-group and add
stopping_tgroups") because the "break;" instruction under the stopping
condition is never executed when support for multithreading is disabled.
To fix the issue, let's add an "else" block to run the "break;"
instruction when USE_THREAD is not defined.
It should be backported up to 2.8
When using a crt-list or a crt-store, keywords mentionned twice on the
same line overwritte the previous value.
This patch emits a warning when the same keyword is found another time
on the same line.
The ckch_conf_parse() function is the generic function which parses
crt-store keywords from the crt-store section, and also from a
crt-list.
When having multiple time the same keyword, a leak of the previous
value happens. This patch ensure that the previous value is always
freed before overwriting it.
This is the same problem as the previous "BUG/MINOR: ssl/ckch: always
free() the previous entry during parsing" patch, however this one
applies on PARSE_TYPE_ARRAY_SUBSTR.
No backport needed.
The ckch_conf_parse() function is the generic function which parses
crt-store keywords from the crt-store section, and also from a crt-list.
When having multiple time the same keyword, a leak of the previous value
happens. This patch ensure that the previous value is always freed
before overwriting it.
This patch should be backported as far as 3.0.
The 'ssl-f-use' implementation doesn't prevent to have multiple time the
'crt' keyword, which overwrite the previous value. Letting users think
that is it possible to use multiple certificates on the same line, which
is not the case.
This patch emits an alert when setting the 'crt' keyword multiple times
on the same ssl-f-use line.
Should fix issue #2966.
No backport needed.
Commit c7f29afc ("MEDIUM: ssl: replace "crt" lines by "ssl-f-use"
lines") forgot to remove an the allocation of the crt field which was
done with the first argument.
Since ssl-f-use takes keywords, this would put the first keyword in
"crt" instead of the certificate name.
IPv6 connectivity might start off (e.g. network not fully up when
haproxy starts), so for features like resolvers, it would be nice to
periodically recheck.
With this change, instead of having the resolvers code rely on a variable
indicating connectivity, it will now call a function that will check for
how long a connectivity check hasn't been run, and will perform a new one
if needed. The age was set to 30s which seems reasonable considering that
the DNS will cache results anyway. There's no saving in spacing it more
since the syscall is very check (just a connect() without any packet being
emitted).
The variables remain exported so that we could present them in show info
or anywhere else.
This way, "dns-accept-family auto" will now stay up to date. Warning
though, it does perform some caching so even with a refreshed IPv6
connectivity, an older record may be returned anyway.
This way we can preserve the entire contents of the released area for
later inspection. This automatically enables comparison at reallocation
time as well (like "integrity" does). If used in combination with
integrity, the comparison is disabled but the check of non-corruption
of the area mangled by integrity is still operated.
The automatic scheduler is useful but sometimes you don't want to use,
or schedule manually.
This patch adds an 'acme.scheduler' option in the global section, which
can be set to either 'auto' or 'off'. (auto is the default value)
This also change the ouput of the 'acme status' command so it does not
shows scheduled values. The state will be 'Stopped' instead of
'Scheduled'.
The new MEM_F_UAF flag can be set just after a pool's creation to make
this pool UAF for debugging purposes. This allows to maintain a better
overall performance required to reproduce issues while still having a
chance to catch UAF. It will only be used by developers who will manually
add it to areas worth being inspected, though.
Emission of flow-control frames have been recently modified. Now, each
frame is sent one by one, via a single entry list. If a failure occurs,
emission is interrupted and frame is reinserted into the original
<qcc.lfctl.frms> list.
This code is incorrect as it only checks if qcc_send_frames() returns an
error code to perform the reinsert operation. However, an error here
does not always mean that the frame was not properly emitted by lower
quic-conn layer. As such, an extra test LIST_ISEMPTY() must be performed
prior to reinsert the frame.
This bug would cause a heap overflow. Indeed, the reinsert frame would
be a random value. A crash would occur as soon as it would be
dereferenced via <qcc.lfctl.frms> list.
This was reproduced by issuing a POST with a big file and interrupt it
after just a few seconds. This results in a crash in about a third of
the tests. Here is an example command using ngtcp2 :
$ ngtcp2-client -q --no-quic-dump --no-http-dump \
-m POST -d ~/infra/html/1g 127.0.0.1 20443 "http://127.0.0.1:20443/post"
Heap overflow was detected via a BUG_ON() statement from qc_frm_free()
via qcc_release() caller :
FATAL: bug condition "!((&((*frm)->reflist))->n == (&((*frm)->reflist)))" matched at src/quic_frame.c:1270
This does not need to be backported.
The reporting of the empty arg location added with commit 08d3caf30
("MINOR: cfgparse: visually show the input line on empty args") falls
victim of a special case detected by OSS Fuzz:
https://issues.oss-fuzz.com/issues/415850462
In short, making an argument start with "\x00" doesn't make it empty for
the parser, but still emits an empty string which is detected and
displayed. Unfortunately in this case the error pointer is not set so
the sanitization function crashes.
What we're doing in this case is that we fall back to the position of
the output argument as an estimate of where it was located in the input.
It's clearly inexact (quoting etc) but will still help the user locate
the problem.
No backport is needed unless the commit above is backported.
The following patch simplifies qc_stream_desc_ack(). The qc_stream_desc
instance is not freed anymore, even if all data were acknowledged. As
implies by the commit message, the caller is responsible to perform this
cleaning operation.
f4a83fbb14
MINOR: quic: do not remove qc_stream_desc automatically on ACK handling
However, despite the commit instruction, qc_stream_desc_free()
invokation was not moved in the caller. This commit fixes this by adding
it after stream ACK handling. This is performed only when a transfer is
completed : all data is acknowledged and qc_stream_desc has been
released by its MUX stream instance counterpart.
This bug may cause a significant increase in memory usage when dealing
with long running connection. However, there is no memory leak, as every
qc_stream_desc attached to a connection are finally freed when quic_conn
instance is released.
This must be backported up to 3.1.
Since commit 388539faa ("MEDIUM: stick-tables: defer adding updates to a
tasklet"), between the entry creation and its arrival in the updates tree,
there is time for scheduling, and it now becomes possible for an stksess
entry to be requeued into the list while it's still in the tree as a remote
one. Only local updates were removed prior to being inserted. In this case
we would re-insert the entry, causing it to appear as the parent of two
distinct nodes or leaves, and to be visited from the first leaf during a
delete() after having already been removed and freed, causing a crash,
as Christian reported in issue #2959.
There's no reason to backport this as this appeared with the commit above
in 3.2-dev13.
commit 29b76cae4 ("BUG/MEDIUM: server/log: "mode log" after server
keyword causes crash") introduced some postparsing checks/tasks for
server
Initially they were mainly meant for "mode log" servers postparsing, but
we already have a check dedicated to "tcp/http" servers (ie: only tcp
proto supported)
However when dynamic servers are added they bypass _srv_postparse() since
the REGISTER_POST_SERVER_CHECK() is only executed for servers defined in
the configuration.
To ensure consistency between dynamic and static servers, and ensure no
post-check init routine is missed, let's manually invoke _srv_postparse()
after creating a dynamic server added via the cli.
d3f928944 ("BUG/MINOR: cli: Issue an error when too many args are passed
for a command") added a new check to prevent the command to run when
too many arguments are provided. In this case an error is reported.
However it turns out this check (despite marked for backports) was
ineffective prior to 20ec1de21 ("MAJOR: cli: Refacor parsing and
execution of pipelined commands") as 'p' pointer was reset to the end of
the buffer before the check was executed.
Now since 20ec1de21, the check works, but we have another issue: we may
read past initialized bytes in the buffer because 'p' pointer is always
incremented in a while loop without checking if we increment it past 'end'
(This was detected using valgrind)
To fix the issue introduced by 20ec1de21, let's only increment 'p' pointer
if p < end.
For 3.2 this is it, now for older versions, since d3f928944 was marked for
backport, a sligthly different approach is needed:
- conditional p increment must be done in the loop (as in this patch)
- max arg check must moved above "fill unused slots" comment where p is
assigned to the end of the buffer
This patch should be backported with d3f928944.
It might be possible not to see the element in the tree, then not to
see it in the update list, thus not to take the lock before deleting.
But an element in the list could have moved to the tree during the
check, and be removed later without the updt_lock.
Let's delete prior to checking the presence in the tree to avoid
this situation. No backport is needed since this arrived in -dev13
with the update list.
In peer_treat_updatemsg(), we call stktable_touch_remote() after
releasing the write lock on the TS, asking it to decrement the
refcnt, then we update ts->seen. Unfortunately this is racy and
causes the issue that Christian reported in issue #2959.
The sequence of events is very hard to trigger manually, but what happens
is the following:
T1. stktable_touch_remote(table, ts, 1);
-> at this point the entry is in the mt_list, and the refcnt is zero.
T2. stktable_trash_oldest() or process_table_expire()
-> these can run, because the refcnt is now zero.
The entry is cleanly deleted and freed.
T1. HA_ATOMIC_STORE(&ts->seen, 1)
-> we dereference freed memory.
A first attempt at a fix was made by keeping the refcnt held during
all the time the entry is in the mt_list, but this is expensive as
such entries cannot be purged, causing lots of skips during
trash_oldest_data(). This managed to trigger watchdogs, and was only
hiding the real cause of the problem.
The correct approach clearly is to maintain the ref_cnt until we
touch ->seen. That's what this patch does. It does not decrement
the refcnt, while calling stktable_touch_remote(), and does it
manually after touching ->seen. With this the problem is gone.
Note that a reproducer involves the following:
- a config with 10 stick-ctr tracking the same table with a
random key between 10M and 100M depending on the machine.
- the expiration should be between 10 and 20s. http_req_cnt
is stored and shared with the peers.
- 4 total processes with such a config on the local machine,
each corresponding to a different peer. 3 of the peers are
bound to half of the cores (all threads) and share the same
threads; the last process is bound to the other half with
its own threads.
- injecting at full load, ~256 conn, on the shared listening
port. After ~2x expiration time to 1 minute the lone process
should segfault in pools code due to a corrupted by_lru list.
This problem already exists in earlier versions but the race looks
narrower. Given how difficult it is to trigger on a given machine
in its current form, it's likely that it only happens once in a
while on stable branches. The fix must be backported wherever the
code is similar, and there's no hope to reproduce it to validate
the backport.
Thanks again to Christian for his amazing help!
Add a checks on received max_udp_payload transport parameters. As
defined per RFC 9000, values below 1200 are invalid, and thus the
connection must be closed with TRANSPORT_PARAMETER_ERROR code.
Prior to this patch, an invalid value was silently ignored.
This should be backported up to 2.6. Note that is relies on previous
patch "MINOR: quic: extend return value on TP parsing".
Checks are implemented on some received transport parameter values,
to reject invalid ones defined per RFC 9000. This is the case for
max_ack_delay parameter.
The check was not properly implemented as it only reject values strictly
greater than the limit set to 2^14. Fix this by rejecting values of 2^14
and above. Also, the proper error code TRANSPORT_PARAMETER_ERROR is now
set.
This should be backported up to 2.6. Note that is relies on previous
patch "MINOR: quic: extend return value on TP parsing".
As per RFC 9000, checks must be implemented to reject invalid values for
received transport parameters. Such values are dependent on the
parameter type.
Checks were already implemented for ack_delay_exponent and
active_connection_id_limit, accordingly with the QUIC specification.
However, the connection was closed with an incorrect error code. Fix
this to ensure that TRANSPORT_PARAMETER_ERROR code is used as expected.
This should be backported up to 2.6. Note that is relies on previous
patch "MINOR: quic: extend return value on TP parsing".
Close the connection on error if retry_source_connection_id transport
parameter is received. This is specified by RFC 9000 as this parameter
must not be emitted by a client. Previously, it was silently ignored.
This should be backported up to 2.6. Note that is relies on previous
patch "MINOR: quic: extend return value on TP parsing".
This commit is similar to the previous one. It fixes the error code
reported when dealing with invalid received transport parameters. This
time, it handles reception of original_destination_connection_id,
preferred_address and stateless_reset_token which must not be emitted by
the client.
This should be backported up to 2.6. Note that is relies on previous
patch "MINOR: quic: extend return value on TP parsing".
Handle missing received transport parameter value
initial_source_connection_id / original_destination_connection_id.
Previously, such case would result in an error reported via
quic_transport_params_store(), which triggers a TLS alert converted as
expected as a CONNECTION_CLOSE. The issue is that the error code
reported in the frame was incorrect.
Fix this by returning QUIC_TP_DEC_ERR_INVAL for such conditions. This is
directly handled via quic_transport_params_store() which set the proper
TRANSPORT_PARAMETER_ERROR code for the CONNECTION_CLOSE. However, no
error is reported so the SSL handshake is properly terminated without a
TLS alert. This is enough to ensure that the CONNECTION_CLOSE frame will
be emitted as expected.
This should be backported up to 2.6. Note that is relies on previous
patch "MINOR: quic: extend return value on TP parsing".
Extend API used for QUIC transport parameter decoding. This is done via
the introduction of a dedicated enum to report the various error
condition detected. No functional change should occur with this patch,
as the only returned code is QUIC_TP_DEC_ERR_TRUNC, which results in the
connection closure via a TLS alert.
This patch will be necessary to properly reject transport parameters
with the proper CONNECTION_CLOSE error code. As such, it should be
backported up to 2.6 with the following series.
However the doc purposely says the opposite, to encourage migrating away
from "ip". The goal is that in the future we change "ip" to mean "ipv6",
which seems to be what most users naturally expect. But we cannot break
configurations in the LTS version so for now "ipv4" is the alias.
The reason for not changing it in the table is that the type name is
used at a few places (look for "].kw"):
- dumps
- promex
We'd rather not change that output for 3.2, but only do it in 3.3.
This way, 3.2 can be made future-proof by using "ipv4" in the config
without any other side effect.
Please see github issue #2962 for updates on this transition.
Now with the improved backtraces, the lock history and details in the
mux layers, some dumps appear truncated or with some chars alone at
the beginning of the line. The issue is in fact caused by the limited
dump buffer size (2kB for stderr, 4kB for warning), that cannot hold
a complete trace anymore.
Let's jump bump them to 8kB, this will be plenty for a long time.
In commit 3f2c8af313 ("MINOR: tools: make parse_line() provide hints
about empty args") we've added the ability to record the position of
the first empty arg in parse_line(), but that check requires to
access the args[] array for the current arg, which is not valid in
case we stopped on too large an argument count. Let's just check the
arg's validity before doing so.
This was reported by OSS Fuzz:
https://issues.oss-fuzz.com/issues/415850462
No backport is needed since this was in the latest dev branch.
When declaring a certificate via the crt-store section, this certificate
can then be used 2 ways in a crt-list:
- only by using its name, without any crt-store options
- or by using the exact set of crt-list option that was defined in the
crt-store
Since ssl-f-use is generating a crt-list, this is suppose to behave the
same. To achieve this, ckch_conf_parse() will parse the keywords related
to the ckch_conf on the ssl-f-use line and use ckch_conf_cmp() to
compare it to the previous declaration from the crt-store. This
comparaison is only done when any ckch_conf keyword are present.
However, ckch_conf_parse() was done for the crt-list, and the crt-list
does not use the "crt" parameter to declare the name of the certificate,
since it's the first element of the line. So when used with ssl-f-use,
ckch_conf_parse() will always see a "crt" keyword which is a ckch_conf
one, and consider that it will always need to have the exact same set of
paremeters when using the same crt in a crt-store and an ssl-f-use line.
So a simple configuration like this:
crt-store web
load crt "foo.com.crt" key "foo.com.key" alias "foo"
frontend mysite
bind :443 ssl
ssl-f-use crt "@web/foo" ssl-min-ver TLSv1.2
Would lead to an error like this:
config : '@web/foo' in crt-list '(null)' line 0, is already defined with incompatible parameters:
- different parameter 'key' : previously 'foo.com.key' vs '(null)'
In order to fix the issue, this patch parses the "crt" parameter itself
for ssl-f-use instead of using ckch_conf_parse(), so the keyword would
never be considered as a ckch_conf keyword to compare.
This patch also take care of setting the CKCH_CONF_SET_CRTLIST flag only
if a ckch_conf keyword was found. This flag is used by ckch_conf_cmp()
to know if it has to compare or not.
No backport needed.
Fill cfg_crt_node with a filename and linenum so the post_section
callback can use it to emit errors.
This way the errors are emitted with the right filename and linenum
where ssl-f-use is used instead of (null):0
The extra call to pendconn_process_next_strm() made in commit cda7275ef5
("MEDIUM: queue: Handle the race condition between queue and dequeue
differently") was performed after releasing the server queue's lock,
which is incompatible with the calling convention for this function.
The result is random corruption of the server's streams list likely
due to picking old or incorrect pendconns from the queue, and in the
end infinitely looping on apparently already locked mt_list objects.
Just adding the lock fixes the problem.
It's very difficult to reproduce, it requires low maxconn values on
servers, stickiness on the servers (cookie), a long enough slowstart
(e.g. 10s), and regularly flipping servers up/down to re-trigger the
slowstart.
No backport is needed as this was only in 3.2.
Remove the 'acme ps' command which does not seem useful anymore with the
'acme status' command.
The big difference with the 'acme status' command is that it was only
displaying the running tasks instead of the status of all certificate.
The "acme status" command, shows the status of every certificates
configured with ACME, not only the running task like "acme ps".
The IO handler loops on the ckch_store tree and outputs a line for each
ckch_store which has an acme section set. This is still done under the
ckch_store lock and doesn't support resuming when the buffer is full,
but we need to change that in the future.
This reverts commit 53c3046898.
This patch introduced a regression leading to a loop on the frames
demultiplexing because a frame may be ignore but not consumed.
But outside this regression that can be fixed, there is a design issue that
was not totally fixed by the patch above. The SPOP connection state is mixed
with the status of the frames demultiplexer and this needlessly complexify
the connection management. Instead of fixing the fix, a better solution is
to revert it to work a a proper solution.
For the record, the idea is to deal with the spop connection state onlu
using 'state' field and to introduce a new field to handle the frames
demultiplexer state. This should ease the closing state management.
Another issue that must be fixed. We must take care to not abort a SPOP
stream when an error is detected on a SPOP connection or when the connection
is closed, if the ACK frame was already received for this stream. It is not
a common case, but it can be solved by saving the last known stream ID that
recieved a ACK.
This patch must be backported if the commit above is backported.
proxy_inc_fe_cum_sess_ver_ctr() was implemented in 9969adbc
("MINOR: stats: add by HTTP version cumulated number of sessions and
requests")
As its name suggests, it is meant to be called for frontends, not backends
Also, in 9969adbc, when used under h1_init(), a precaution is taken to
ensure that the function is only called with frontends.
However, this precaution was not applied in h2_init() and qc_init().
Due to this, it remains possible to have proxy_inc_fe_cum_sess_ver_ctr()
being called with a backend proxy as parameter. While it did not cause
known issues so far, it is not expected and could result in bugs in the
future. Better fix this by ensuring the function is only called with
frontends.
It may be backported up to 2.8
Based on the lock history, we can spot some locks that are still held
by checking the last operation that happened on them: if it's not an
unlock, then we know the lock is held. In this case we append the list
after "locked:" with their label and state like below:
U:QUEUE S:IDLE_CONNS U:IDLE_CONNS R:TASK_WQ U:TASK_WQ S:QUEUE S:QUEUE S:QUEUE locked: QUEUE(S)
S:IDLE_CONNS U:IDLE_CONNS S:TASK_RQ U:TASK_RQ S:QUEUE U:QUEUE S:IDLE_CONNS locked: IDLE_CONNS(S)
R:TASK_WQ S:TASK_WQ R:TASK_WQ S:TASK_WQ R:TASK_WQ S:TASK_WQ R:TASK_WQ locked: TASK_WQ(R)
W:STK_TABLE W:STK_TABLE_UPDT U:STK_TABLE_UPDT W:STK_TABLE W:STK_TABLE_UPDT U:STK_TABLE_UPDT W:STK_TABLE W:STK_TABLE_UPDT locked: STK_TABLE(W) STK_TABLE_UPDT(W)
The format is slightly different (label(status)) so as to easily
differentiate them visually from the history.
The fix in commit 09a325a4de ("BUG/MINOR: tools: always terminate empty
lines") is insufficient. While it properly addresses the lack of trailing
zero, it doesn't account for it in the returned outlen that is used to
allocate a larger line. This happens at boot if the very first line of
the test file is exactly a sharp with nothing else. In this case it will
return a length 0 and the caller (parse_cfg()) will try to re-allocate an
entry of size zero and will fail, bailing out a lack of memory. This time
it should really be OK.
It doesn't need to be backported, unless the patch above would be.
Since latest commit 7e4a2f39ef ("BUG/MINOR: tools: do not create an empty
arg from trailing spaces"), an empty line will no longer produce an arg
and no longer append a trailing zero to them. This was not visible because
one is already present in the input string, however all the trailing args
are set to out+outpos-1, which now points one char before the buffer since
nothing was emitted, and was noticed by ASAN, and/or when parsing garbage.
Let's make sure to always emit the zero for empty lines as well to address
this issue. No backport is needed unless the patch above gets backported.
Now when an empty arg is found on a line, we emit the sanitized
input line and the position of the first empty arg so as to help
the user figure the cause (likely an empty environment variable).
Co-authored-by: Valentine Krasnobaeva <vkrasnobaeva@haproxy.com>
In order to help parse_line() callers report the position of empty
args to the user, let's decide that if no error is emitted, then
we'll stuff the errptr with the position of the first empty arg
without affecting the return value.
Co-authored-by: Valentine Krasnobaeva <vkrasnobaeva@haproxy.com>
For historical reasons, the config parser relies on the trailing '\0'
to detect the end of the line being parsed. When the lines started to be
tokenized into arguments, this principle has been preserved, and now all
the parsers rely on *args[arg]='\0' to detect the end of a line. But as
reported in issue #2944, while most of the time it breaks the parsing
like below:
http-request deny if { path_dir '' }
it can also cause some elements to be silently ignored like below:
acl bad_path path_sub '%2E' '' '%2F'
This may also subtly happen with environment variables that don't exist
or which are empty:
acl bad_path path_sub '%2E' "$BAD_PATTERN" '%2F'
Fortunately, parse_line() returns the number of arguments found, so it's
easy from the callers to verify if any was empty. The goal of this commit
is not to perform sensitive changes, it's only to mention when parsing a
line that an empty argument was found and alert about its consequences
using a warning. Most of the time when this happens, the config does not
parse. But for examples as the ACLs above, there could be consequences
that are better detected early.
This patch depends on this previous fix:
BUG/MINOR: tools: do not create an empty arg from trailing spaces
Co-authored-by: Valentine Krasnobaeva <vkrasnobaeva@haproxy.com>
Trailing spaces on the lines of the config file create an empty arg
which makes it complicated to detect really empty args. Let's first
address this. Note that it is not user-visible but prevents from
fixing user-visible issues. No backport is needed.
The initial issue was introduced with this fix that already tried to
address it:
8a6767d266 ("BUG/MINOR: config: don't count trailing spaces as empty arg (v2)")
The current patch properly addresses leading and trailing spaces by
only counting arguments if non-lws chars were found on the line. LWS
do not cause a transition to a new arg anymore but they complete the
current one. The whole new code relies on a state machine to detect
when to create an arg (!in_arg->in_arg), and when to close the current
arg. A special care was taken for word expansion in the form of
"${ARGS[*]}" which still continue to emit individual arguments past
the first LWS. This example works fine:
ARGS="100 check inter 1000"
server name 192.168.1."${ARGS[*]}"
It properly results in 6 args:
"server", "name", "192.168.1.100", "check", "inter", "1000"
This fix should not have any visible user impact and is a bit tricky,
so it's best not to backport it, at least for a while.
Co-authored-by: Valentine Krasnobaeva <vkrasnobaeva@haproxy.com>
Previous patch 7251c13c7 ("MINOR: acme: move the acme task init in a dedicated
function") mistakenly returned the wrong error code when "acme renew" parsing
was successful, and tried to emit an error message.
This patch fixes the issue by returning 0 when the acme task was correctly
scheduled to start.
No backport needed.
As reported in GH #2958, commit 6c9b315 caused a regression with sc_*
fetches and tracked counter id > 9.
As such, the below configuration would cause a BUG_ON() to be triggered:
global
log stdout format raw local0
tune.stick-counters 11
defaults
log global
mode http
frontend www
bind *:8080
acl track_me bool(true)
http-request set-var(txn.track_var) str("a")
http-request track-sc10 var(txn.track_var) table rate_table if track_me
http-request set-var(txn.track_var_rate) sc_gpc_rate(0,10,rate_table)
http-request return status 200
backend rate_table
stick-table type string size 1k expire 5m store gpc_rate(1,1m)
While in 6c9b315 the src_fetch logic was removed from
smp_fetch_sc_stkctr(), num > 9 is indeed not expected anymore as
original num value. But what we didn't consider is that num is effectively
re-assigned for generic sc_* variant.
Thus the BUG_ON() is misplaced as it should only be evaluated for
non-generic fetches. It explains why it triggers with valid configurations
Thanks to GH user @tkjaer for his detailed report and bug analysis
No backport needed, this bug is specific to 3.2.
This patch implements a very basic scheduler for the ACME tasks.
The scheduler is a task which is started from the postparser function
when at least one acme section was configured.
The scheduler will loop over the certificates in the ckchs_tree, and for
each certificate will start an ACME task if the notAfter date is past
curtime + (notAfter - notBefore) / 12, or 7 days if notBefore is not
available.
Once the lookup over all certificates is terminated, the task will sleep
and will wakeup after 12 hours.
acme_start_task() is a dedicated function which starts an acme task
for a specified <store> certificate.
The initialization code was move from the "acme renew" command parser to
this function, in order to be called from a scheduler.
When env variable name or value are not provided for setenv/presetenv it's not
clear from the old error message shown at stderr, what exactly is missed. User
needs to search in it's configuration.
Let's add more explicit error messages about these inconsistencies.
No need to be backported.
In process_table_expire(), limit the number of entries we remove in one
call, and just reschedule the task if there's more to do. Removing
entries require to use the heavily contended update write lock, and we
don't want to hold it for too long.
This helps getting stick tables perform better under heavy load.
Limit the number of old entries we remove in one call of
stktable_trash_oldest(), as we do so while holding the heavily contended
update write lock, so we'd rather not hold it for too long.
This helps getting stick tables perform better under heavy load.
There is a lot of contention trying to add updates to the tree. So
instead of trying to add the updates to the tree right away, just add
them to a mt-list (with one mt-list per thread group, so that the
mt-list does not become the new point of contention that much), and
create a tasklet dedicated to adding updates to the tree, in batchs, to
avoid keeping the update lock for too long.
This helps getting stick tables perform better under heavy load.
In peer_send_msgs(), give up in order to retry later if we failed at
getting the update read lock.
Similarly, in __process_running_peer_sync(), give up and just reschedule
the task if we failed to get the peer lock. There is an heavy contention
on both those locks, so we could spend a lot of time trying to get them.
This helps getting peers perform better under heavy load.
tune.lua.bool-sample-conversion must be set before any lua-load or
lua-load-per-thread is used for it to be considered. Indeed, lua-load
directives are parsed on the fly and will cause some parts of the scripts
to be executed during init already (script body/init contexts).
As such, we cannot afford to have "tune.lua.bool-sample-conversion" set
after some Lua code was loaded, because it would mean that the setting
would be handled differently for Lua's code executed during or after
config parsing.
To avoid ambiguities, the documentation now states that the setting must
be set before any lua-load(-per-thread) directive, and if the setting
is met after some Lua was already loaded, the directive is ignored and
a warning informs about that.
It should fix GH #2957
It may be backported with 29b6d8af16 ("MINOR: hlua: rename
"tune.lua.preserve-smp-bool" to "tune.lua.bool-sample-conversion"")
There's no point reading t->state to check for a tasklet after we've
atomically read the state into the local "state" variable. Not only it's
more expensive, it's also less clear whether that state is supposed to
be atomic or not. And in any case, tasks and tasklets have their type
forever and the one reflected in state is correct and stable.
After recent commit b81c9390f ("MEDIUM: tasks: Mutualize the TASK_KILLED
code between tasks and tasklets"), the task accounting was no longer
correct for killed tasks due to the decrement of tasks in list that was
no longer done, resulting in infinite loops in process_runnable_tasks().
This just illustrates that this code remains complex and should be further
cleaned up. No backport is needed, as this was in 3.2.
quic_conn_release() may, or may not, free the tasklet associated with
the connection. So make it return 1 if it was, and 0 otherwise, so that
if it was called from the tasklet handler itself, the said handler can
act accordingly and return NULL if the tasklet was destroyed.
This should be backported if 9240cd4a27
is backported.
The next request was always leaving the task befor initializing the
httpclient. This patch optimize it by jumping to the next step at the
end of the current one. This way, only the httpclient is doing a
task_wakeup() to handle the response. But transiting from response to
the next request does not leave the task.
Add an extra parametre to conn_create_mux(), "closed_connection".
If a pointer is provided, then let it know if the connection was closed.
Callers have no way to determine that otherwise, and we need to know
that, at least in ssl_sock_io_cb(), as if the connection was closed we
need to return NULL, as the tasklet was free'd, otherwise that can lead
to memory corruption and crashes.
This should be backported if 9240cd4a27
is backported too.
The code to handle a task/tasklet when it's been killed before it were
to run is mostly identical, so move it outside of task and tasklet
specific code, and inside the common code.
This commit is just cosmetic, and should have no impact.
TASK_QUEUED was used to mean "the task has been scheduled to run",
TASK_IN_LIST was used to mean "the tasklet has been scheduled to run",
remove TASK_IN_LIST and just use TASK_QUEUED for tasklets instead.
This commit is just cosmetic, and should not have any impact.
There is some code that should run no matter if the task was killed or
not, and was needlessly duplicated, so only use one instance.
This also fixes a small bug when a tasklet that got killed before it
could run would still count as a tasklet that ran, when it should not,
which just means that we'd run one less useful task before going back to
the poller.
This commit is mostly cosmetic, and should not have any impact.
The code that checks if we're currently running, and waits if so, was
identical between tasks and tasklets, so move it in code common to tasks
and tasklets.
This commit is just cosmetic, and should not have any impact.
Use acme_ctx_destroy() instead of a simple free() upon error in the
"acme renew" error handling.
It's better to use this function to be sure than everything has been
been freed.
Implement a way to display the running acme tasks over the CLI.
It currently only displays a "Running" status with the certificate name
and the acme section from the configuration.
The displayed running tasks are limited to the size of a buffer for now,
it will require a backref list later to be called multiple times to
resume the list.
{listener,proxy,server}_get_stats() methods are know to be expensive,
expecially if used under an iteration. Indeed, while automatic yield
is performed every X lua instructions (defaults to 10k), computing an
object's stats 10K times in a single cpu loop is not desirable and
could create contention.
In this patch we leverage hlua_yield_asap() at the end of *_get_stats()
methods in order to force the automatic yield to occur ASAP after the
method returns. Hopefully this should help in similar scenarios as the
one described in GH #2903
When called, this function will try to enforce a yield (if available) as
soon as possible. Indeed, automatic yield is already enforced every X
Lua instructions. However, there may be some cases where we know after
running heavy operation that we should yield already to avoid taking too
much CPU at once.
This is what this function offers, instead of asking the user to manually
yield using "core.yield()" from Lua itself after using an expensive
Lua method offered by haproxy, we can directly enforce the yield without
the need to do it in the Lua script.
The previous commit has implemented a new calcul method for
MAX_STREAM_DATA frame emission. Now, a frame may be emitted as soon as a
buffer was consumed by a QCS instance.
This will probably increase the number of MAX_STREAM_DATA frame
emission. It may even cause a series of frame emitted for the same
stream with increasing values under high load, which is completely
unnecessary.
To improve this, limit the number of MAX_STREAM_DATA frames built to one
per QCS instance. This is implemented by storing a reference to this
frame in QCS structure via a new member <tx.msd_frm>.
Note that to properly reset QCS msd_frm member, emission of flow-control
frames have been changed. Now, each frame is emitted individually. On
one side, it is better as it prevent to emit frames related to different
streams in a single datagram, which is not desirable in case of packet
loss. However, this can also increase sendto() syscall invocation.
Recently, QCS Rx allocation buffer method has been improved. It is now
possible to allocate multiple buffers per QCS instances, which was
necessary to improve HTTP/3 POST throughput.
However, a limitation remained related to the emission of
MAX_STREAM_DATA. These frames are only emitted once at least half of the
receive capacity has been consumed by its QCS instance. This may be too
restrictive when a client need to upload a large payload.
Improve this by adjusting MAX_STREAM_DATA allocation. If QCS capacity is
still limited to 1 or 2 buffers max, the old calcul is still used. This
is necessary when user has limited upload throughput via their
configuration. If QCS capacity is more than 2 buffers, a new frame is
emitted if at least a buffer was consumed.
This patch has reduced number of STREAM_DATA_BLOCKED frames received in
POST tests with some specific clients.
Becaues of a typo, '||' was used instead of '|' to test the SPOP conneciton
flags and decide if the mux is ready or not. The regression was introduced
in the commit fd7ebf117 ("BUG/MEDIUM: mux-spop: Wait end of handshake to
declare a spop connection ready").
This patch must be backported to 3.1 with the commit above.
The newly added SSL traces require an extra 'conn' parameter to
ssl_sock_chose_sni_ctx which was added in the "regular" code but not in
the wolfssl specific one.
Wolfssl also has a different prototype for some getter functions
(SSL_get_servername for instance), which do not expect a const SSL while
openssl version does.
h1_destroy() may be called to release a H1C after a multiplexer upgrade. In
that case, the connection is no longer attached to the H1C. It must not be
used in the h1 trace message because the connection context is no longer a H1C.
Because of this bug, when a H1>H2 upgrade is performed, a crash may be
experienced if the H1 traces are enabled.
This patch must be backport to all stable versions.
When an applicative upgrade of the H1 multiplexer is performed, we must not
pretend the connection was released. Indeed, in that case, a H1 stream is
still their with a stream connector attached on it. It must be detached
first before releasing the H1 connection and the underlying connection. So
it is important to not pretend the connection was already released.
Concretely, in that case h1_process() must return 0 instead of -1. It is
minor error because, AFAIK, it is harmless. But it is not correct. So let's
fix it to avoid futur bugs.
To be clear, this happens when a TCP connection is upgraded to H1 connection
and a H2 preface is detected, leading to a second upgrade from H1 to H2.
This patch may be backport to all stable versions.
In the SPOE specification, when an error occurred on the SPOP connection,
HAProxy must send a DISCONNECT frame and wait for the agent DISCONNECT frame
in return before trully closing the connection.
However, this part was not properly handled by the SPOP multiplexer. In this
case, the SPOP connection should be in the CLOSING state. But this state was
not used at all. Depending on when the error was encountered, the connection
could be closed immediately, without sending any DISCONNECT frame. It was
the case when an early error was detected during the AGENT-HELLO frame
parsing. Or it could be moved from ERROR to FRAME_H state, as if no error
were detected. This case was less dramatic than it seemed because some flags
were also set to prevent any problem. But it was not obvious.
So now, the SPOP connection is properly switch to CLOSING state when an
DISCONNECT is sent to the agent to be able to wait for its DISCONNECT in
reply. spop_process_demux() was updated to parse frames in that state and
some validity checks was added.
This patch must be backport to 3.1.
A SPOP connection must not be considered as ready while the hello handshake
is not finished with success. In addition, no error or shutdown must have
been reported for the underlying connection. Otherwise a freshly openned
spop connexion may be reused while it is in fact dead, leading to a
connection retry.
This patch must be backported to 3.1.
We had to parse the sigAlg extension by hand in order to properly select
the certificate used by the SSL frontends. These traces allow to dump
the allowed sigAlg list sent by the client in its clientHello.
This callback allows to pick the used certificate on an SSL frontend.
The certificate selection is made according to the information sent by
the client in the clientHello. The traces that were added will allow to
better understand what certificate was chosen and why. It will also warn
us if the chosen certificate was the default one.
The actual certificate parsing happens in ssl_sock_chose_sni_ctx. It's
in this function that we actually get the filename of the certificate
used.
If OCSP stapling fails because of a missing or invalid OCSP response we
used to silently disable stapling for the given session. We can now know
a bit more what happened regarding OCSP stapling.
Those traces dump information about the multiple SSL_do_handshake calls
(renegotiation and regular call). Some errors coud also be dumped in
case of rejected early data.
Depending on the chosen verbosity, some information about the current
handshake can be dumped as well (servername, tls version, chosen cipher
for instance).
In case of failed handshake, the error codes and messages will also be
dumped in the log to ease debugging.
Add a dedicated trace for some unlikely allocation failures and async
errors. Those traces will ostly be used to identify the start and end of
a given SSL connection.
This function can be used to convert a TLSv1.3 sigAlg entry (2bytes)
from the signature_agorithms client hello extension into a string.
In order to ease debugging, some TLSv1.2 combinations can also be
dumped. In TLSv1.2 those signature algorithms pairs were built out of a
one byte signature identifier combined to a one byte hash identifier.
In TLSv1.3 those identifiers are two bytes blocs that must be treated as
such.
In relation to issue #2954, it appears that turning some size_t length
calculations to the int that uses my_strndup() upsets coverity a bit.
Instead of dealing with such warnings each time, better address it at
the root. An inspection of all call places show that the size passed
there is always positive so we can safely use an unsigned type, and
size_t will always suit it like for strndup() where it's available.
Since 91e785edc ("MINOR: stream: Rely on a per-stream max connection
retries value"), px->conn_retries may be ignored in the following cases:
* proxy not part of a list which gets properly post-init (ie: main proxy
list, log-forward list, sink list)
* proxy lacking the CAP_FE capability
Documenting such cases where the px->conn_retries is set but effectively
ignored, so that we either remove ignored statements or fix them in
the future if they are really needed. In fact all cases affected here are
automomous applets that already handle the retries themselves so the fact
that 91e785edc made ->conn_retries ineffective should not be a big deal
anyway.
when dns session callback (dns_session_release()) is called upon error
(ie: when some pending queries were not sent), we try our best to
re-create the applet in order to preserve the pending queries and give
them a chance to be retried. This is done at the end of
dns_session_release().
However, doing so exposes to an issue: if the error preventing queries
from being sent is still encountered over and over the dns session could
stay there indefinitely. Meanwhile, other dns sessions may be created on
the same dns_stream_server periodically. If previous failing dns sessions
don't terminate but we also keep creating new ones, we end up accumulating
failing sessions on a given dns_stream_server, which can eventually cause
ressource shortage.
This issue was found when trying to address ("BUG/MINOR: dns: add tempo
between 2 connection attempts for dns servers")
To fix it, we track the number of failed consecutive sessions for a given
dns server. When we reach the threshold (set to 100), we consider that the
link to the dns server is broken (at least temporarily) and we force
dns_session_new() to fail, so that we stop creating new sessions until one
of the existing one eventually succeeds.
A workaround for this fix consists in setting the "maxconn" parameter on
nameserver directive (under resolvers section) to a reasonnable value so
that no more than "maxconn" sessions may co-exist on the same server at
a given time.
This may be backported to all stable versions.
("CLEANUP: dns: remove unused dns_stream_server struct member") may be
backported to ease the backport.
As reported by Lukas Tribus on the mailing list [1], trying to connect to
a nameserver with invalid network settings causes haproxy to retry a new
connection attempt immediately which eventually causes unexpected CPU usage
on the thread responsible for the applet (namely 100% on one CPU will be
observed).
This can be reproduced with the test config below:
resolvers default
nameserver ns1 tcp4@8.8.8.8:53 source 192.168.99.99
listen listen
mode http
bind :8080
server s1 www.google.com resolvers default init-addr none
To fix this the issue, we add a temporisation of one second between a new
connection attempt is retried. We do this in dns_session_create() when we
know that the applet was created in the release callback (when previous
query attempt was unsuccessful), which means initial connection is not
affected.
[1]: https://www.mail-archive.com/haproxy@formilux.org/msg45665.html
This should fix GH #2909 and may be backported to all stable versions.
This patch depends on ("MINOR: applet: add appctx_schedule() macro")
"virt@acme" was the default map used during development, now this must
be configured in the acme section or it won't try to use any map.
This patch removes the references to virt@acme in the comments and the
code.
The stateless mode which was documented previously in the ACME example
is not convenient for all use cases.
First, when HAProxy generates the account key itself, you wouldn't be
able to put the thumbprint in the configuration, so you will have to get
the thumbprint and then reload.
Second, in the case you are using multiple account key, there are
multiple thumbprint, and it's not easy to know which one you want to use
when responding to the challenger.
This patch allows to configure a map in the acme section, which will be
filled by the acme task with the token corresponding to the challenge,
as the key, and the thumbprint as the value. This way it's easy to reply
the right thumbprint.
Example:
http-request return status 200 content-type text/plain lf-string "%[path,field(-1,/)].%[path,field(-1,/),map(virt@acme)]\n" if { path_beg '/.well-known/acme-challenge/' }
Define a new settings tune.quic.frontend.max-tot-window. It contains a
size argument which can be used to set a limit on the sum of all QUIC
connections congestion window. This is applied both on
quic_cc_path_set() and quic_cc_path_inc().
Note that this limitation cannot reduce a congestion window more than
the minimal limit which is set to 2 datagrams.
Use the newly defined cshared type to account for the sum of congestion
window of every QUIC connection. This value is stored in global counter
quic_mem_global defined in proto_quic module.
Congestion window is limit by a minimal and maximum values which can
never be exceeded. Min value is hardcoded to 2 datagrams as recommended
by the specification. Max value is specified via haproxy configuration.
These values must be respected each time the congestion window size is
adjusted. However, in some rare occasions, limit were not always
enforced. Fix this by implementing wrappers to set or increment the
congestion window. These functions ensure limits are always applied
after the operation.
Additionnally, wrappers also ensure that if window reached a new maximum
value, it is saved in <cwnd_last_max> field.
This should be backported up to 2.6, after a brief period of
observation.
Write minor adjustments to QUIC BBR functions. The objective is to
centralize every modification of path cwnd field.
No functional change. This patch will be useful to simplify
implementation of global QUIC Tx memory usage limitation.
There was some possible confusion between fields related to congestion
window size min and max limit which cannot be exceeded, and the maximum
value previously reached by the window.
Fix this by adopting a new naming scheme. Enforced limit are now renamed
<limit_max>/<limit_min>, while the previously reached max value is
renamed <cwnd_last_max>.
This should be backported up to 3.1.
The account creation was mistakenly ending the task instead of being
wakeup for the NewOrder state, it was preventing the creation of the
certificate, however the account was correctly created.
To fix this, only the jump to the end label need to be remove, the
standard leaving codepath of the function will allow to be wakeup.
No backport needed.
TCP_NOTSENT_LOWAT is very convenient as it indicates when to report
EAGAIN on the sending side. It takes a margin on top of the estimated
window, meaning that it's no longer needed to store too many data in
socket buffers. Instead there's just enough to fill the send window
and a little bit of margin to cover the scheduling time to restart
sending. Experiments on a 100ms network have shown a 10-fold reduction
in the memory used by socket buffers by just setting this value to
tune.bufsize, without noticing any performance degradation. Theoretically
the responsiveness on multiplexed protocols such as H2 should also be
improved.
The ping frame's pattern must be written at offset 9 (frame header
length), not 8. This was added in 3.2 with commit 4dcfe098a6 ("MINOR:
mux-h2: prepare to support PING emission"), so no backport is needed.
While humans can find it convenient to enter the worker process in prompt
mode, for external tools it will not be convenient to have to systematically
disable it. A better approach is to replicate the master socket's mode
there, since it has already been configured to suit the user: interactive,
prompt and timed modes are automatically passed to the worker process.
This makes the using the worker commands more natural from the master
process, without having to systematically adapt it for each new connection.
Support the same syntax in master mode as in worker mode in order to
configure the prompt. The only thing is that for now the master doesn't
have a non-interactive mode and it doesn't seem necessary to implement
it, so we only support the interactive and prompt modes. However the code
was written in a way that makes it easy to change this later if desired.
Now the prompt mode can more finely be configured between non-interactive
(default), interactive without prompt, and interactive with prompt. This
will ease the usage from automated tools which are not necessarily
interested in having to consume '> ' after each command nor displaying
"+" on payload lines. This can also be convenient when coming from the
master CLI to keep the same output format.
The CLI's "prompt" command toggles two distinct things:
- displaying or hiding the prompt at the beginning of the line
- single-command vs interactive mode
These are two independent concepts and the prompt mode doesn't
always cope well with tools that would like to upload data without
having to read the prompt on return. Also, the master command line
works in interactive mode by default with no prompt, which is not
consistent (and not convenient for tools). So let's start by splitting
the bit in two, and have a new APPCTX_CLI_ST1_INTER flag dedicated
to the interactive mode. For now the "prompt" command alone continues
to toggle the two at once.
Generate the private key on the account file when the file does not
exists. This generate a private key of the type and parameters
configured in the acme section.
This will display the lock labels and modes for each non-empty step
at the end of "show threads" when these are defined. This allows to
emit up to the last 8 locking operation for each thread on 64 bit
machines.
We now default the value to zero and make sure all tests properly take
care of values above zero. This is in preparation for supporting several
degrees of debugging.
Machines lacking CAS8B/DWCAS and emit a warning in lb_fwlc.c without
threads due to declaration ordering. Let's just move the variable
declaration into the block that uses it as a last variable. No
backport is needed.
Not all systems have strndup(), that's why we have our "my_strndup()",
so let's make use of it here. This fixes the build on Solaris 10. No
backport is needed.
It has been requested to have the current_session_rate exposed at the
frontend level. For now only the per-process value was exposed
(ST_I_INF_SESS_RATE).
Thanks to the work done lately to merge promex and stat_cols_px[]
array, let's simply defined an .alt_name for the ST_I_PX_RATE metric in
order to have promex exposing it as current_session_rate for relevant
contexts.
Add a -t option to 'show ssl sni', allowing to add an offset to the
current date so it would allow to check which certificates are expired
after a certain period of time.
Since we made it possible for a bind_conf to listen to multiple thread
groups with shards in 2.8 with commit 9d360604bd ("MEDIUM: listener:
rework thread assignment to consider all groups"), the per-listener
connection count was not properly transferred to the target listener
with the connection when switching to another thread group. This results
in one listener possibly reaching high values and another one possibly
reaching negative values. Usually it's not visible, unless a maxconn is
set on the bind_conf, in which case comparisons will quickly put an end
to the willingness to accept new connections.
This problem only happens when thread groups are enabled, and it seems
very hard to trigger it normally, it only impacts sockets having a single
shard, hence currently the CLI (or any conf with "bind ... shards 1"),
where it can be reproduced with a config having a very low "maxconn" on
the stats socket directive (here, 4), and issuing a few tens of
socat <<< "show activity" in parallel, or sending HTTP connections to a
single-shared listener. Very quickly, haproxy stops accepting connections
and eats CPU in the poller which tries to get its connections accepted.
A BUG_ON(l->nbconn<0) after HA_ATOMIC_DEC() in listener_release() also
helps spotting them better.
Many thanks to Christian Ruppert who once again provided a very accurate
report in GH #2951 with the required data permitting this analysis.
This fix must be backported to 2.8.
tasklets were originally designed to alway run on only one thread, so it
was not possible to have it run on 2 threads concurrently.
The API has been extended so that another thread may wake the tasklet,
the idea was still that we wanted to have it run on one thread only.
However, the way it's been done meant that unless a tasklet was bound to
a specific tid with tasklet_set_tid(), or we explicitely used
tasklet_wakeup_on() to specify the thread for the target to run on, it
would be scheduled to run on the current thread.
This is in fact a desirable feature. There is however a race condition
in which the tasklet would be scheduled on a thread, while it is running
on another. This could lead to the same tasklet to run on multiple
threads, which we do not want.
To fix this, just do what we already do for regular tasks, set the
"TASK_RUNNING" flag, and when it's time to execute the tasklet, wait
until that flag is gone.
Only one case has been found in the current code, where the tasklet
could run on different threads depending on who wakes it up, in the
leastconn load balancer, since commit
627280e15f.
It should not be a problem in practice, as the function called can be
called concurrently.
If a bug is eventually found in relation to this problem, and this patch
should be backported, the following patches should be backported too :
MEDIUM: quic: Make sure we return the tasklet from quic_accept_run
MEDIUM: quic: Make sure we return NULL in quic_conn_app_io_cb if needed
MEDIUM: quic: Make sure we return the tasklet from qcc_io_cb
MEDIUM: mux_fcgi: Make sure we return the tasklet from fcgi_deferred_shut
MEDIUM: listener: Make sure w ereturn the tasklet from accept_queue_process
MEDIUM: checks: Make sure we return the tasklet from srv_chk_io_cb
In quic_conn_app_io_cb, make sure we return NULL if the tasklet has been
destroyed, so that the scheduler knows. It is not yet needed, but will
be soon.
On systems where the network is not reachable at boot time (certain HA
systems for example, or dynamically addressed test machines), we'll want
to be able to periodically revalidate the IPv6 reachability status. The
current code makes it complicated because it sets the config bits once
for all at boot time. This commit changes this so that the config bits
are not changed, but instead we rely on a static inline function that
relies on sock_inet6_seems_reachable for every test (really cheap). This
also removes the now unneeded resolvers late init code.
This variable for now is still set at boot time but this will ease the
transition later, as the resolvers code is now ready for this.
The new adhoc parser for the '@@' prefix forgot to require the presence
of the LF character marking the end of the line. This is the reason why
entering incomplete commands would display garbage, because the line was
expected to have its LF character replaced with a zero.
The problem is well illustrated by using socat in raw mode:
socat /tmp/master.sock STDIO,raw,echo=0
then entering "@@1 show info" one character at a time would error just
after the second "@". The command must take care to report an incomplete
line and wait for more data in such a case.
This reverts commit 0e94339eaf.
This patch was in fact fixing the symptom, not the cause. The root cause
of the problem is that the parser was processing an incomplete line when
looking for '@@'. When the LF is present, this problem does not exist
as it's properly replaced with a zero. This can be verified using socat
in raw mode:
socat /tmp/master.sock STDIO,raw,echo=0
Then entering "@@1 show info" one character at a time will immediately
fail on "@@" without going further. A subsequent patch will fix this.
No backport is needed.
When the CLI applet was refactord in the commit 20ec1de21 ("MAJOR: cli:
Refacor parsing and execution of pipelined commands"), a regression was
introduced. The applet shutdown was not longer handled when the applet was
waiting for the next command line. It is especially visible when a client
timeout occurred because the client connexion is no longer closed.
To fix the issue, the test on the SE_FL_SHW flag was reintroduced in
CLI_ST_PARSE_CMDLINE state, but only is there is no pending input data.
It is a 3.2-specific issue. No backport needed.
When the ACME task is checking for the status of the challenge, it would
only succeed or retry upon failure.
However that's not the best way to do it, ACME objects contain an
"status" field which could have a final status or a in progress status,
so we need to be able to retry.
This patch adds an acme_ret enum which contains OK, RETRY and FAIL.
In the case of the CHKCHALLENGE, the ACME could return a "pending" or a
"processing" status, which basically need to be rechecked later with the
RETRY. However a "invalid" or "valid" status is final and will return
either a FAIL or a OK.
So instead of retrying in any case, the "invalid" status will ends the
task with an error.
Parse the Retry-After header in response and store it in order to use
the value as the next delay for the next retry, fallback to 3s if the
value couldn't be parse or does not exist.
Instead of always having to force IPv4 or IPv6, let's now also offer
"auto" which will only enable IPv6 if the system has a default gateway
for it. This means that properly configured dual-stack systems will
default to "ipv4,ipv6" while those lacking a gateway will only use
"ipv4". Note that no real connectivity test is performed, so firewalled
systems may still get it wrong and might prefer to rely on a manual
"ipv4" assignment.
In order to ease dual-stack deployments, we could at least try to
check if ipv6 seems to be reachable. For this we're adding a test
based on a UDP connect (no traffic) on port 53 to the base of
public addresses (2001::) and see if the connect() is permitted,
indicating that the routing table knows how to reach it, or fails.
Based on this result we're setting a global variable that other
subsystems might use to preset their defaults.
In order to ease troubleshooting and testing, the new "-4" command line
argument enforces queries and processing of "A" DNS records only, i.e.
those representing IPv4 addresses. This can be useful when a host lack
end-to-end dual-stack connectivity. This overrides the global
"dns-accept-family" directive and is equivalent to value "ipv4".
By default, DNS resolvers accept both IPv4 and IPv6 addresses. This can be
influenced by the "resolve-prefer" keywords on server lines as well as the
family argument to the "do-resolve" action, but that is only a preference,
which does not block the other family from being used when it's alone. In
some environments where dual-stack is not usable, stumbling on an unreachable
IPv6-only DNS record can cause significant trouble as it will replace a
previous IPv4 one which would possibly have continued to work till next
request. The "dns-accept-family" global option permits to enforce usage of
only one (or both) address families. The argument is a comma-delimited list
of the following words:
- "ipv4": query and accept IPv4 addresses ("A" records)
- "ipv6": query and accept IPv6 addresses ("AAAA" records)
When a single family is used, no request will be sent to resolvers for the
other family, and any response for the othe family will be ignored. The
default value is "ipv4,ipv6", which effectively enables both families.
When '@@' alone is sent on the master CLI (no trailing LF), we get an
error that displays anything past these two characters in the buffer
since there's no room for a \0. Let's make sure to limit the length of
the process name in this case. No backport is needed since this was added
with 00c967fac4 ("MINOR: master/cli: support bidirectional communications
with workers").
When a service is initialized, the "use-service" rule that was executed is
now saved in the stream, using "current_rule" field, instead of saving it
into the applet context. It is safe to do so becaues this field is unused at
this stage. To avoid any issue, it is reset after the service
initialization. Doing so, it is no longer necessary to save it in the applet
context. It was the last usage of the rule pointer in the applet context.
The init functions for TCP and HTTP lua services were updated accordingly.
The lua function name was used in error messages of HTTP/TCP lua services
while the applet name can be used. Concretely, this will not change
anything, because when a lua service is regiestered, the lua function name
is used to name the applet. But it is easier, cleaner and more logicial
because it is really the applet name that should be displayed in these error
messages.
Thanks to this change, when a response is delivered from the cache, it is no
longer necessary to get the cache filter configuration from the http
"use-cache" rule saved in the appctx to get the currently used cache. It was
a bit complex to get an info that can be directly and naturally stored in
the cache applet context.
There are several fields in the appctx structure only used by the CLI. To
make things cleaner, all these fields are now placed in a dedicated context
inside the appctx structure. The final goal is to move it in the service
context and add an API for cli commands to get a command coontext inside the
cli context.
Thanks to the CLI refactoring ("MAJOR: cli: Refacor parsing and execution of
pipelined commands"), it is possible to fix "show event" I/O handle function
to no longer use the SC.
When the applet API was refactored to no longer manipulate the channels or
the stream-connectors, this part was missed. However, without the patch
above, it could not be fixed. It is now possible so let's do it.
This patch must not be backported becaues it depends on refactoring of the
CLI applet.
Thanks to the CLI refactoring ("MAJOR: cli: Refacor parsing and execution of
pipelined commands"), it is possible to fix the I/O handler function used by
lua CLI commands to no longer use the SC.
When the applet API was refactored to no longer manipulate the channels or
the stream-connectors, this part was missed. However, without the patch
above, it could not be fixed. It is now possible so let's do it.
This patch must not be backported becaues it depends on refactoring of the
CLI applet.
CLI_ST_GETREQ state was renamed into CLI_ST_PARSE_CMDLINE and CLI_ST_PARSEREQ
into CLI_ST_PROCESS_CMDLINE to reflect the real action performed in these
states.
Before this patch, when pipelined commands were received, each command was
parsed and then excuted before moving to the next command. Pending commands
were not copied in the input buffer of the applet. The major issue with this
way to handle commands is the impossibility to consume inputs from commands
with an I/O handler, like "show events" for instance. It was working thanks
to a "bug" if such commands were the last one on the command line. But it
was impossible to use them followed by another command. And this prevents us
to implement any streaming support for CLI commands.
So we decided to refactor the command line parsing to have something similar
to a basic shell. Now an entire line is parsed, including the payload,
before starting commands execution. The command line is copied in a
dedicated buffer. "appctx->chunk" buffer is used for this purpose. It was an
unsed field, so it is safe to use it here. Once the command line copied, the
commands found on this line are executed. Because the applet input buffer
was flushed, any input can be safely consumed by the CLI applet and is
available for the command I/O handler. Thanks to this change, "show event
-w" command can be followed by a command. And in theory, it should be
possible to implement commands supporting input data streaming. For
instance, the Tetris like lua applet can be used on the CLI now.
Note that the payload, if any, is part of the command line and must be fully
received before starting the commands processing. It means there is still
the limitation to a buffer, but not only for the payload but for the whole
command line. The payload is still necessarily at the end of the command
line and is passed as argument to the last command. Internally, the
"appctx->cli_payload" field was introduced to point on the payload in the
command line buffer.
This patch is quite huge but it cannot easily be splitted. It should not
introduced significant changes.
When a bidirection connection with no command is establisehd with a worker
(so "@@<pid>" alone), a "prompt" command is automatically added to display
the worker's prompt and enter in interactive mode in the worker context.
However, till now, an unfinished command line is sent, with a semicolon
instead of a newline at the end. It is not exactly a bug because this
works. But it is not really expected and could be a problem for future
changes.
So now, a full command line is sent: the "prompt" command finished by a
newline character.
When a command is parsed to split it in an array of arguments, by default,
at most 64 arguments are supported. But no warning was emitted when there
were too many arguments. Instead, the arguments above the limit were
silently ignored. It could be an issue for some commands, like "add server",
because there was no way to know some arguments were ignored.
Now an error is issued when too many arguments are passed and the command is
not executed.
This patch should be backported to all stable versions.
Add an early return to qcc_decode_qcs() if QCC instance is flagged on
error and connection is scheduled for immediate closure.
The main objective is to ensure to not trigger BUG_ON() from
qcc_set_error() : if a stream decoding has set the connection error, do
not try to process decoding on other streams as they may also encounter
an error. Thus, the connection is closed asap with the first encountered
error case.
This should be backported up to 2.6, after a period of observation.
With the support of multiple Rx buffers per QCS instance, stream
decoding in qcc_io_recv() has been reworked for the next haproxy
release. An issue appears in a double while loop : a break statement is
used in the inner loop, which is not sufficient as it should instead
exit from the outer one.
Fix this by replacing break with a goto statement.
No need to backport this.
Remove return statement in h3_rcv_buf() in case of stream/connection
error. Instead, reuse already existing label err. This simplifies the
code path. It also fixes the missing leave trace for these cases.
Use a customized proxy for the ACME client.
The proxy is initialized at the first acme section parsed.
The proxy uses the httpsclient log format as ACME CA use HTTPS.
Add an experimental "https" log-format for the httpclient, it is not
used by the httpclient by default, but could be define in a customized
proxy.
The string is basically a httpslog, with some of the fields replaced by
their backend equivalent or - when not available:
"%ci:%cp [%tr] %ft -/- %TR/%Tw/%Tc/%Tr/%Ta %ST %B %CC %CS %tsc %ac/%fc/%bc/%sc/%rc %sq/%bq %hr %hs %{+Q}r %[bc_err]/%[ssl_bc_err,hex]/-/-/%[ssl_bc_is_resumed] -/-/-"
The 'pause' HTTP action can now be used to suspend for a moment the message
analysis. A timeout, expressed in milliseconds using a time-format
parameter, or an expression can be used. If an expression is used, errors
and invalid values are ignored.
Internally, the action will set the analysis expiration date on the
corresponding channel to the configured value and it will yield while it is
not expired.
The 'pause' action is available for 'http-request' and 'http-response'
rules.
When a SPOP connection is opened, the maximum size for frames is negociated.
This negociated size is properly used when a frame is received and if a too
big frame is detected, an error is triggered. However, the same was not
performed on the sending path. No check was performed on frames sent to the
agent. So it was possible to send frames bigger than the maximum size
supported by the the SPOE agent.
Now, the size of NOTIFY and DISCONNECT frames is checked before sending them
to the agent.
Thanks to Miroslav to have reported the issue.
This patch must be backported to 3.1.
Since the commit "MINOR: hlua/h1: Use http_parse_cont_len_header() to parse
content-length value", this function is no longer used. So it can be safely
removed.
Till now, h1_parse_cont_len_header() was used during the H1 message parsing and
by the lua HTTP applets to parse the content-length header value. But a more
generic function was added some years ago doing exactly the same operations. So
let's use it instead.
Thanks to the commit "MINOR: mux-h1: Don't remove custom "Content-Length: 0"
header in 1xx and 204 messages", we are now sure that 1xx and 204 responses
were sanitized during the parsing. So, if one of these headers are found in
such responses when sent to the client, it means it was added by hand, via a
"set-header" action for instance. In this context, we are able to make an
exception for the "Content-Length: 0" header, and only this one with this
value, to not break leagacy applications.
So now, a user can force the "Content-Length: 0" header to appear in 1xx and
204 responses by adding the right action in hist configuration.
"Transfer-Encoding" headers are still dropped as "Content-Length" headers
with another value than 0. Note, that in practice, only 101 and 204 are
concerned because other 1xx message are not subject to HTTP analysis.
This patch should fix the issue #2888. There is no reason to backport
it. But if we do so, the patch above must be backported too.
According to the RFC9110 and RFC9112, a server must not add 'Content-Length'
or 'Transfer-Encoding' headers into 1xx and 204 responses. So till now,
these headers were dropped from the response when it is sent to the client.
However, it seems more logical to remove it during the message parsing. In
addition to sanitize messages as early as possible, this will allow us to
apply some exception in some cases (This will be the subject of another
patch).
In this patch, 'Content-Length' and 'Transfer-Encoding' headers are removed
from 1xx and 204 responses during the parsing but the same is still
performed during the formatting stage.
In RFC9110, it is stated that trailers could be merged with the
headers. While it should be performed with a speicial care, it may be a
problem for some applications. To avoid any trouble with such applications,
two new options were added to drop trailers during the message forwarding.
On the backend, "http-drop-request-trailers" option can be enabled to drop
trailers from the requests before sending them to the server. And on the
frontend, "http-drop-response-trailers" option can be enabled to drop
trailers from the responses before sending them to the client. The options
can be defined in defaults sections and disabled with "no" keyword.
This patch should fix the issue #2930.
This changes commit d2a9149f0 ("BUG/MINOR: proxy: always detach a proxy
from the names tree on free()") to be cleaner. Aurélien spotted that
the free(p->id) was indeed already done in proxy_free_common(), which is
called before we delete the node. That's still a bit ugly and it only
works because ebpt_delete() does not dereference the key during the
operation. Better play safe and delete the entry before freeing it,
that's more future-proof.
Stephen Farrell reported in issue #2942 that recent haproxy versions
crash if there's no resolv.conf. A quick bisect with his reproducer
showed that it started with commit 4194f75 ("MEDIUM: tree-wide: avoid
manually initializing proxies") which reorders the proxies initialization
sequence a bit. The crash shows a corrupted tree, typically indicating a
use-after-free. With the help of ASAN it was possible to find that a
resolver proxy had been destroyed and freed before the name insertion
that causes the crash, very likely caused by the absence of the needed
resolv.conf:
#0 0x7ffff72a82f7 in free (/usr/local/lib64/libasan.so.5+0x1062f7)
#1 0x94c1fd in free_proxy src/proxy.c:436
#2 0x9355d1 in resolvers_destroy src/resolvers.c:2604
#3 0x93e899 in resolvers_create_default src/resolvers.c:3892
#4 0xc6ed29 in httpclient_resolve_init src/http_client.c:1170
#5 0xc6fbcf in httpclient_create_proxy src/http_client.c:1310
#6 0x4ae9da in ssl_ocsp_update_precheck src/ssl_ocsp.c:1452
#7 0xa1b03f in step_init_2 src/haproxy.c:2050
But free_proxy() doesn't delete the ebpt_node that carries the name,
which perfectly explains the situation. This patch simply deletes the
name node and Stephen confirmed that it fixed the problem for him as
well. Let's also free it since the key points to p->id which is never
freed either in this function!
No backport is needed since the patch above was first merged into
3.2-dev10.
In check_config_validity(), we initialize the proxy in several stages.
We do so for the sink list for stage1, but not for stage2. It may not be
needed right now, but it may become needed in the future, so do it
anyway.
Now that all there is one tree per thread group, all thread groups will
start on the same server. To prevent that, just insert the servers in a
different order for each thread group.
When using the round-robin load balancer, the major source of contention
is the lbprm lock, that has to be held every time we pick a server.
To mitigate that, make it so there are one tree per thread-group, and
one lock per thread-group. That means we now have a lb_fwrr_per_tgrp
structure that will contain the two lb_fwrr_groups (active and backup) as well
as the lock to protect them in the per-thread lbprm struct, and all
fields in the struct server are now moved to the per-thread structure
too.
Those changes are mostly mechanical, and brings good performances
improvment, on a 64-cores AMD CPU, with 64 servers configured, we could
process about 620000 requests par second, and we now can process around
1400000 requests per second.
Move the "next_weight" outside of fwrr_group, and inside struct lb_fwrr
directly, one for the active servers, one for the backup servers.
We will soon have one fwrr_group per thread group, but next_weight will
be global to all of them.
Add a pointer to the server into the struct srv_per_tgroup, so that if
we only have access to that srv_per_tgroup, we can come back to the
corresponding server.
Move the call to initialize the proxy's per-thread structure earlier
than currently done, so that they are usable when we're initializing the
load balancers.
