If an error a detected at the MUX layer, all remaining stream endpoints
should be closed asap with error set. This is now done by checking for
QC_CF_ERRL flag on qc_wake_some_streams() and qc_send_buf(). To complete
this, qc_wake_some_streams() is called by qc_process() if needed.
This should help to quickly release streams as soon as a new error is
detected locally by the MUX or APP layer. This allows to in turn free
the MUX instance itself. Previously, error would not have been
automatically reported until the transport layer closure would occur
on CONNECTION_CLOSE emission.
This should be backported up to 2.7.
When a fatal error is detected by the QUIC MUX or H3 layer, the
connection should be closed with a CONNECTION_CLOSE with an error code
as the reason.
Previously, a direct call was used to the quic_conn layer to try to
close the connection. This API was adjusted to be more flexible. Now,
when an error is detected, the function qcc_set_error() is called. This
set the flag QC_CF_ERRL with the error code stored by the MUX. The
connection will be closed soon so most of the operations are not
conducted anymore. Connection is then finally closed during qc_send()
via quic_conn layer if QC_CF_ERRL is set. This will set the flag
QC_CF_ERRL_DONE which indicates that the MUX instance can be freed.
This model is cleaner and brings the following improvments :
- interaction with quic_conn layer for closure is centralized on a
single function
- CO_FL_ERROR is not set anymore. This was incorrect as this should be
reserved to errors reported by the transport layer to be similar with
other haproxy components. As a consequence, qcc_is_dead() has been
adjusted to check for QC_CF_ERRL_DONE to release the MUX instance.
This should be backported up to 2.7.
When HTX content is transferred from qcs instance to upper stream
endpoint, a wakeup is conducted for MUX tasklet. However, this is only
necessary if demux was interrupted due to a full QCS HTX buffer.
This should be backported up to 2.7.
Add a dedicated trace event QMUX_EV_QCC_ERR. This is used for locally
detected error when a CONNECTION_CLOSE should be emitted.
This should be backported up to 2.7.
When MUX performs a graceful shutdown, quic_conn error code is set to a
"no error" code which depends on the application layer used. However,
this may overwrite a previous error code if quic_conn layer has detected
an error on its side.
In practice, this behavior has not been seen on production. In fact, it
may have undesirable effect only if this error code modification happens
between the quic_conn error detection and the emission of the
CONNECTION_CLOSE, so it should be pretty rare. However, there is still a
tiny possibility it may happen.
To prevent this, first check that quic_conn error code is not set before
setting it. Ideally, transport layer API should be adjusted to be able
to set this without fiddling with the quic_conn directly.
This should be backported up to 2.6.
The commit 2722c04b ("MEDIUM: mux-h2: Don't expect data from server as long
as request is unfinished") introduced a regression in the H2 multiplexer.
The end of the request is not systematically handled to state a H2 stream on
client side now expexts data from the server.
Indeed, while the client is uploading its request, the H2 stream warns it
does not expect data from the server. This way, no server timeout is applied
at this stage. When end of the request is detected, the H2 stream must state
it now expects the server response. This enables the server timeout.
However, it was only performed at one place while the end of the request can
be handled at different places. First, during a zero-copy in
h2_rcv_buf(). Then, when the SC is created with the full request. Because of
this bug, it is possible to totally disable the server timeout for H2
streams.
In h2_rcv_buf(), we now rely on h2s flags to detect the end of the request,
but only when the rxbuf was emptied.
It is a 2.8-specific bug. No backport needed.
Sometimes it's convenient to test the effect of tasks running under
isolation, e.g. to validate the contents of the crash dumps. Let's
add an optional "isolated" keyword to "debug dev loop" for this.
Thread dumps include a field "prof" for each thread that reports whether
task profiling is currently active or not. It turns out that in 2.7-dev1,
commit 680ed5f28 ("MINOR: task: move profiling bit to per-thread")
mistakenly replaced it with a check for the current thread's bit in the
thread dumps, which basically is the only place where another thread is
being watched. The same mistake was done a few lines later by confusing
threads_want_rdv_mask with the profiling mask. This mask disappeared
in 2.7-dev2 with commit 598cf3f22 ("MAJOR: threads: change thread_isolate
to support inter-group synchronization"), though instead we know the ID
of the isolated thread. This commit fixes this and now reports "isolated"
instead of "wantrdv".
This can be backported to 2.7.
16kB buffers are not enough to dump 4096 threads with up to 10 bytes value
on each line. By storing the column number in the applet's context, we can
now restart from the last attempted column. This requires to dump all values
as they are produced, but it doesn't cost that much: a 4096-thread output
from a fesh process produces 300kB of output in ~8ms, or ~400us per call
(19*16kB), most of which are spent in vfprintf(). Given that we don't print
more than needed, it doesn't really change anything.
The main caveat is that when interrupted on such large lines, there's a
great possibility that the total or average on the first column doesn't
match anymore the sum or average of all dumped values. In order to avoid
this whenever possible (typically less than ~1500 threads), we first try
to dump entire lines and only proceed one column at a time when we have
to retry a failed dump. This is already the same for other stats that are
dumped in an interruptible way anyway and there's little that can be done
about it at this point (and not much immediately perceived benefit in
doing this with extreme accuracy for >1500 threads).
When using many threads, it's difficult to see the end of "show activity"
due to the numerous columns which fill the buffer. For example a dump of
a 256-thread, freshly booted process yields around 15kB.
Here by arranging the dump in a loop around a switch/case block where
each case checks the code line number against the current dump position,
we have a restartable counter for free with a granularity of the line of
code, without having to maintain a matching between states and specific
lines. It just requires to reset the trash buffer for each line and to
try to dump it after each line.
Now dumping 256 threads after a few seconds of traffic happily emits 20kB.
Now each line of "show activity" will iterate over n+2 fields, one for
the line header, one for the total, and one per thread. This will soon
allow us to save the current state in a restartable way.
Commit 986798718 ("DEBUG: cli: add "debug dev task" to show/wake/expire/kill
tasks and tasklets") broke the build on windows due to this:
src/debug.c:940:95: error: array subscript has type char [-Werror=char-subscripts]
940 | caller && may_access(caller) && may_access(caller->func) && isalnum(*caller->func) ? caller->func : "0",
| ^~~~~~~~~~~~~
It's classical on platforms which implement ctype.h as macros instead of
functions, let's cast it as uchar. No backport is needed.
The x509_v_err_str converter now outputs the numerical value as a string
when the corresponding constant name was not found.
Must be backported as far as 2.7.
When analyzing certain types of bugs in field, sometimes it would be
nice to be able to wake up a task or tasklet to see how events progress
(e.g. to detect a missing wakeup condition), or expire or kill such a
task. This restricted command shows hte current state of a task or tasklet
and allows to manipulate it like this. However it must be used with extreme
care because while it does verify that the pointers are mapped, it cannot
know if they point to a real task, and performing such actions on something
not a task will easily lead to a crash. In addition, performing a "kill"
on a task has great chances of provoking a deferred crash due to a double
free and/or another kill that is not idempotent. Use with extreme care!
The "show sess" command displays the stream's age in synthetic form,
and also makes it appear in the long version (show sess all). But that
last one uses the wrong origin, it uses accept_date.tv_sec instead of
accept_ts (formerly known as tv_accept). This was introduced in 1.4.2
with the long format, with commit 66dc20a17 ("[MINOR] stats socket: add
show sess <id> to dump details about a session"), while the code that
split the two variables was introduced in 1.3.16 with commit b7f694f20
("[MEDIUM] implement a monotonic internal clock"). This problem was
revealed by recent change ad5a5f677 ("MEDIUM: tree-wide: replace timeval
with nanoseconds in tv_accept and tv_request") that made this value report
random garbage, and generally emphasized by the fact that in 2.8 the two
clocks have sufficiently large an offset for such mistakes to be noticeable
early.
Arguably a difference between date and accept_date could also make sense,
to indicate if the stream had been there for more than 49 days, but this
would introduce instabilities for most sockets (including negative times)
for extremely rare cases while the goal is essentially to see how much
longer than a configured timeout a stream has been there. And that's what
other locations (including the short form) provide.
This patch could be backported but most users will never notice. In case
of backport, tv_accept.tv_sec should be used instead of accept_date.tv_sec.
WolfSSL is enabling by default the CRL checks even if a CRL file wasn't
provided. This patch resets the default X509_STORE flags so this is
not checked by default.
Before this patch, global sending rate was measured on the QUIC lower
layer just after sendto(). This meant that all QUIC frames were
accounted for, including non STREAM frames and also retransmission.
To have a better reflection of the application data transferred, move
the incrementation into the MUX layer. This allows to account only for
STREAM frames payload on their first emission.
This should be backported up to 2.6.
Now that "now" is no more a timeval, there's no point keeping a copy
of it as a timeval, let's also switch start_time to nanoseconds, it
simplifies operations.
This puts an end to the occasional confusion between the "now" date
that is internal, monotonic and not synchronized with the system's
date, and "date" which is the system's date and not necessarily
monotonic. Variable "now" was removed and replaced with a 64-bit
integer "now_ns" which is a counter of nanoseconds. It wraps every
585 years, so if all goes well (i.e. if humanity does not need
haproxy anymore in 500 years), it will just never wrap. This implies
that now_ns is never nul and that the zero value can reliably be used
as "not set yet" for a timestamp if needed. This will also simplify
date checks where it becomes possible again to do "date1<date2".
All occurrences of "tv_to_ns(&now)" were simply replaced by "now_ns".
Due to the intricacies between now, global_now and now_offset, all 3
had to be turned to nanoseconds at once. It's not a problem since all
of them were solely used in 3 functions in clock.c, but they make the
patch look bigger than it really is.
The clock_update_local_date() and clock_update_global_date() functions
are now much simpler as there's no need anymore to perform conversions
nor to round the timeval up or down.
The wrapping continues to happen by presetting the internal offset in
the short future so that the 32-bit now_ms continues to wrap 20 seconds
after boot.
The start_time used to calculate uptime can still be turned to
nanoseconds now. One interrogation concerns global_now_ms which is used
only for the freq counters. It's unclear whether there's more value in
using two variables that need to be synchronized sequentially like today
or to just use global_now_ns divided by 1 million. Both approaches will
work equally well on modern systems, the difference might come from
smaller ones. Better not change anyhting for now.
One benefit of the new approach is that we now have an internal date
with a resolution of the nanosecond and the precision of the microsecond,
which can be useful to extend some measurements given that timestamps
also have this resolution.
Instead we're using ns_to_sec(tv_to_ns(&now)) which allows the tv_sec
part to disappear. At this point, "now" is only used as a timeval in
clock.c where it is updated.
Let's get rid of timeval in storage of internal timestamps so that they
are no longer mistaken for wall clock time. These were exclusively used
subtracted from each other or to/from "now" after being converted to ns,
so this patch removes the tv_to_ns() conversion to use them natively. Two
occurrences of tv_isge() were turned to a regular wrapping subtract.
Instead of operating on {sec, usec} now we convert both operands to
ns then subtract them and convert to ms. This is a first step towards
dropping timeval from these timestamps.
Interestingly, tv_ms_elapsed() and tv_ms_remain() are no longer used at
all and could be removed.
The "show info" help for "Start_time_sec" says "Start time in seconds"
so it's definitely the start date in human format, not the internal one
that is solely used to compute uptime. Since commit 28360dc ("MEDIUM:
clock: force internal time to wrap early after boot"), both are split
apart since the start time takes into account the offset needed to cause
the early wraparound, so we must only use start_date here.
No backport is needed.
The commit a664aa6a6 ("BUG/MINOR: tcpcheck: Be able to expect an empty
response") instroduced a regression for expect rules relying on a custom
function. Indeed, there is no check on the buffer to be sure it is not empty
before calling the custom function. But some of these functions expect to
have data and don't perform any test on the buffer emptiness.
So instead of fixing all custom functions, we just don't eval them if the
buffer is empty.
This patch must be backported but only if the commit above was backported
first.
It was a cut/paste typo during stream-interface to conn-stream
refactoring. sc_have_room() was used instead of sc_need_room().
This patch must be backported as far as 2.6.
It is possible to start too many applets on sporadic burst of events after
an inactivity period. It is due to the way we estimate if a new applet must
be created or not. It is based on a frequency counter. We compare the events
processing rate against the number of events currently processed (in
progress or waiting to be processed). But we should also take care of the
number of idle applets.
We already track the number of idle applets, but it is global and not
per-thread. Thus we now also track the number of idle applets per-thread. It
is not a big deal because this fills a hole in the spoe_agent structure.
Thanks to this counter, we can refrain applets creation if there is enough
idle applets to handle currently processed events.
This patch should be backported to every stable versions.
That's hopefully the last one affected by this. It was a bit trickier
because there's the promise in the doc that the date is monotonous, so
we continue to use now-start_time as the uptime value and add it to
start_date to get the current date. It was also emphasized by commit
28360dc ("MEDIUM: clock: force internal time to wrap early after boot"),
causing core.now() to return a date of Mar 20 on Apr 27. No backport is
needed.
Yet another case where "now" was used instead of "date" for a publicly
visible date that was already incorrect and became worse after commit
28360dc ("MEDIUM: clock: force internal time to wrap early after boot").
No backport is needed.
Since commit 28360dc ("MEDIUM: clock: force internal time to wrap early
after boot") we have a much clearer distinction between 'now' (the internal,
drifting clock) and 'date' (the wall clock time). The calltrace code was
using "now" instead of "date" since the value is displayed to humans.
No backport is needed.
This reverts commit aadcfc9ea6.
The parts affecting the DeviceAtlas addon were wrong actually, the
"now" variable was a local time_t in a file that's not compiled with
the haproxy binary (dadwsch). Only the fix to the calltrace is correct,
so better revert and fix the only one in a separate commit. No backport
is needed.
Another case where "now" was used instead of "date" for a publicly visible
date that was already incorrect and became worse after commit 28360dc
("MEDIUM: clock: force internal time to wrap early after boot"). No
backport is needed.
The debug messages were still emitted with a date taken from "now" instead
of "date", which was not correct a long time ago but which became worse in
2.8 since commit 28360dc ("MEDIUM: clock: force internal time to wrap early
after boot"). Let's fix it. No backport is needed.
Since commit 28360dc ("MEDIUM: clock: force internal time to wrap early
after boot") we have a much clearer distinction between 'now' (the internal,
drifting clock) and 'date' (the wall clock time). There were still a few
places where 'now' was being used for human consumption.
No backport is needed.
Each quic_conn are attached in a global thread-local quic_conns list
used for "show quic" command. During thread rebinding, a connection is
detached from its local list instance and moved to its new thread list.
However this operation is not thread-safe and may cause a race
condition.
To fix this, only remove the connection from its list inside
qc_set_tid_affinity(). The connection is inserted only after in
qc_finalize_affinity_rebind() on the new thread instance thus prevented
a race condition. One impact of this is that a connection will be
invisible during rebinding for "show quic".
A connection must not transition to closing state in between this two
steps or else cleanup via quic_handle_stopping() may not miss it. To
ensure this, this patch relies on the previous commit :
commit d6646dddcc
MINOR: quic: finalize affinity change as soon as possible
This should be backported up to 2.7.
During accept, a quic-conn is rebind to a new thread. This process is
done in two times :
* first on the original thread via qc_set_tid_affinity()
* then on the newly assigned thread via qc_finalize_affinity_rebind()
Most quic_conn operations (I/O tasklet, task and quic_conn FD socket
read) are reactivated ony after the second step. However, there is a
possibility that datagrams are handled before it via quic_dgram_parse()
when using listener sockets. This does not seem to cause any issue but
this may cause unexpected behavior in the future.
To simplify this, qc_finalize_affinity_rebind() will be called both by
qc_xprt_start() and quic_dgram_parse(). Only one invocation will be
performed thanks to the new flag QUIC_FL_CONN_AFFINITY_CHANGED.
This should be backported up to 2.7.
Sometimes it may be necessary to send an empty STREAM frame to signal
clean stream closure with FIN bit set. Prior to this change, a Tx buffer
was allocated unconditionnally even if no data is transferred.
Most of the times, allocation was not performed due to an older buffer
reused. But if data were already acknowledge, a new buffer is allocated.
No memory leak occurs as the buffer is properly released when the empty
frame acknowledge is received. But this allocation is unnecessary and it
consumes a connexion Tx buffer for nothing.
Improve this by skipping buffer allocation if no data to transfer.
qcs_build_stream_frm() is now able to deal with a NULL out argument.
This should be backported up to 2.6.
Previous patch fixes an issue occurring with empty STREAM frames without
payload. The crash was hidden in part because buf/data fields of
qf_stream were set even if no payload is referenced. This was not the
true cause of the crash but to ease future debugging, a STREAM frame
built with no payload now has its buf and data fields set to NULL.
This should be backported up to 2.6.
Sometimes it may be necessary to send empty STREAM frames with only the
FIN bit set. For these frames, memcpy is thus unnecessary as their
payload is empty. However, we did not prevent its invocation inside
quic_build_stream_frame().
Normally, memcpy invocation with length==0 is safe. However, there is an
extra condition in our function to handle data wrapping. For an empty
STREAM frame in the context of MUX emission, this is safe as the frame
points to a valid buffer which causes the wrapping condition to be false
and resulting in a memcpy with 0 length.
However, in the context of retransmission, this may lead to a crash.
Consider the following scenario : two STREAM frames A and B are
produced, one with payload and one empty with FIN set, pointing to the
same stream_desc buffer. If A is acknowledged by the peer, its buffer is
released as no more data is left in it. If B needs to be resent, the
wrapping condition will be messed up to a reuse of a freed buffer. Most
of the times, <wrap> will be a negative number, which results in a
memcpy invocation causing a buffer overflow.
To fix this, simply add an extra condition to skip memcpy and wrapping
check if STREAM frame length is null inside quic_build_stream_frame().
This crash is pretty rare as it relies on a lot of conditions difficult
to reproduce. It seems to be the cause for the latest crashes reported
under github issue #2120. In all the inspected dumps, the segfault
occurred during retransmission with an empty STREAM frame being used as
input. Thanks again to Tristan from Mangadex for his help and
investigation on it.
This should be backported up to 2.6.
Since the following mentioned patch, a send-list mechanism was
implemented to improve streams priorization on sending.
commit 20f2a425ff
MAJOR: mux-quic: rework stream sending priorization
This is done to prevent the same streams to always be used as first ones
on emission. However there is still a flaw on the algorithm. Once put in
the send-list, a streams is not removed until it has sent all of its
content. When a stream transfers a large object, it will remain in the
send-list during all the transfer and will soon monopolize the first
place. the stream does never leave its position until the transfer is
finished and will monopolize the first place. Other streams behind won't
have the opportunity to advance on their own transfers due to a Tx
buffer exhaustion.
This situation is especially problematic if a small timeout client is
used. As some streams won't advance on their transfer for a long period
of time, they will be aborted due to a stream layer timeout client
causing a RESET_STREAM emission.
To fix this, during sending each stream with at least some bytes
transferred from its tx.buf to qc_stream_desc out buffer is put at the
end of the send-list. This ensures that on the next iteration streams
that cannot transfer anything will be used in priority.
This patch improves significantly h2load benchmarks for large objects
with several streams opened in parallel on a single connection. Without
it, errors may be reported by h2load for aborted streams. For example,
this improved the following scenario on a 10mbit/s link with a 10s
timeout client :
$ ./build/bin/h2load --npn-list h3 -t 1 -c 1 -m 30 -n 30 https://198.18.10.11:20443/?s=500k
This fix may help with the github issue #2004 where chrome browser stop
to use QUIC after receiving RESET_STREAM frames.
This should be backported up to 2.7.
Some HTX responses may not always contain a EOM block. For example this
is the case if content-length header is missing from the HTTP server
response. Stream termination is thus signaled to QUIC mux via shutw
callback. However, this is interpreted inconditionnally as an early
close by the mux with a RESET_STREAM emission. Most of the times, QUIC
clients report this as an error.
To fix this, check if htx.extra is set to HTX_UNKOWN_PAYLOAD_LENGTH for
a qcs instance. If true, shutw will never be used to emit a
RESET_STREAM. Instead, the stream will be closed properly with a FIN
STREAM frame. If all data were already transfered, an empty STREAM frame
is sent.
This fix may help with the github issue #2004 where chrome browser stop
to use QUIC after receiving RESET_STREAM frames.
This issue was reported by Vladimir Zakharychev. Thanks to him for his
help and testing. It was also reproduced locally using httpterm with the
query string "/?s=1k&b=0&C=1".
This should be backported up to 2.7.
Make quic_stateless_reset_token_cpy(), quic_derive_cid() and quic_get_cid_tid()
be more readable: there is no struct buffer variable manipulated by these
functions.
Should be backported to 2.7.
There is no <buf> variable passed to this function.
Also rename <buf_end> to <end> to mimic others functions.
Rename <beg> to <first_byte> and <end> to <last_byte>.
Should be backported to 2.7.
Make quic_build_packet_long_header(), quic_build_packet_short_header() and
quic_apply_header_protection() be more readable: there is no struct buffer
variables used by these functions.
Should be backported to 2.7.
Remove the pointer to the connection passed as parameters to qc_purge_tx_buf()
and other similar function which came with qc_purge_tx_buf() implementation.
They were there do track the connection during tests.
Must be backported to 2.7.
Rename all frame variables with the suffix _frm. This helps to
differentiate frame instances from other internal objects.
This should be backported up to 2.7.
Each frame type used in quic_frame union has been renamed with the
following prefix "qf_". This helps to differentiate frame instances from
other internal objects.
This should be backported up to 2.7.
From the idle_timer_task(), the I/O handler must be woken up to send ack. But
there is no reason to do that in draining state or killing state. In draining
state this is even forbidden.
Must be backported to 2.7.
The timer task responsible of triggering probing retransmission did not inspect
the state of the connection before doing its job. But there is no need to
probe the peer when the connection is in draining or killing state. About the
draining state, this is even forbidden.
Must be backported to 2.7 and 2.6.
qc_dgrams_retransmit() prepares two list of frames to be retransmitted into
two datagrams. If the first datagram could not be sent, the TX buffer will
be purged with the prepared packet and its frames, but this was not the case for
the second list of frames.
Must be backported in 2.7.
This bug arrived with this commit which was not sufficient:
BUG/MEDIUM: quic: Missing TX buffer draining from qc_send_ppkts()
Indeed, there were also remaining allocated TX packets to be released and
their TX frames.
Implement qc_purge_tx_buf() to do so which depends on qc_free_tx_coalesced_pkts()
and qc_free_frm_list().
Must be backported to 2.7.
Sharding by-group is exactly identical to by-process for a single
group, and will use the same number of file descriptors for more than
one group, while significantly lowering the kernel's locking overhead.
Now that all special listeners (cli, peers) are properly handled, and
that support for SO_REUSEPORT is detected at runtime per protocol, there
should be no more reason for now switching to by-group by default.
That's what this patch does. It does only this and nothing else so that
it's easy to revert, should any issue be raised.
Testing on an AMD EPYC 74F3 featuring 24 cores and 48 threads distributed
into 8 core complexes of 3 cores each, shows that configuring 8 groups
(one per CCX) is sufficient to simply double the forwarded connection
rate from 112k to 214k/s, reducing kernel locking from 71 to 55%.
This new setting accepts "by-process", "by-group" and "by-thread" and
will dictate how listeners will be sharded by default when nothing is
specified. While the default remains "by-process", "by-group" should be
much more efficient with many threads, while not changing anything for
single-group setups.
Now that we're able to run listeners on any set of groups, we don't need
to maintain a special case about the stats socket anymore. It used to be
forced to group 1 only so as to avoid startup failures in case several
groups were configured, but if it's done now, it will automatically bind
the needed FDs to have one per group so this is no more an issue.
When testing if a protocol supports SO_REUSEPORT, we're now able to
verify if the OS does really support it. While it may be supported at
build time, it may possibly have been blocked in a container for
example so we'd rather know what it's like.
The new function _sock_supports_reuseport() will be used to check if a
protocol type supports SO_REUSEPORT or not. This will be useful to verify
that shards can really work.
The new function protocol_supports_flag() checks the protocol flags
to verify if some features are supported, but will support being
extended to refine the tests. Let's use it to check for REUSEPORT.
Now if multiple shards are explicitly requested, and the listener's
protocol doesn't support SO_REUSEPORT, sharding is disabled, which will
result in the socket being automatically duped if needed. A warning is
emitted when this happens. If "shards by-group" or "shards by-thread"
are used, these will automatically be turned down to 1 since we want
this to be possible easily using -dR on the command line without having
to djust the config. For "by-thread", a diag warning will be emitted to
help troubleshoot possible performance issues.
Some protocol support SO_REUSEPORT and others not. Some have such a
limitation in the kernel, and others in haproxy itself (e.g. sock_unix
cannot support multiple bindings since each one will unbind the previous
one). Also it's really protocol-dependent and not just family-dependent
because on Linux for some time it was supported for TCP and not UDP.
Let's move the definition to the protocols instead. Now it's preset in
tcp/udp/quic when SO_REUSEPORT is defined, and is otherwise left unset.
The enabled() config condition test validates IPv4 (generally sufficient),
and -dR / noreuseport all protocols at once.
We'll use these flags to know if some protocols are supported, and if
so, with what options/extensions. Reuseport will move there for example.
Two functions were added to globally set/clear a flag.
The listeners in peers sections were still not handing the thread
groups fine. Shards were silently ignored and if a listener was bound
to more than one group, it would simply fail. Now we can call the
dedicated function to resolve all this and possibly create the missing
extra listeners.
bind_complete_thread_setup() was adjusted to use the proxy_type_str()
instead of writing "proxy" at the only place where this word was still
hard-coded so that we continue to speak about peers sections when
relevant.
What used to be only two lines to apply a mask in a loop in
check_config_validity() grew into a 130-line block that performs deeply
listener-specific operations that do not have their place there anymore.
In addition it's worth noting that the peers code still doesn't support
shards nor being bound to more than one group, which is a second reason
for moving that code to its own function. Nothing was changed except
recreating the missing variables from the bind_conf itself (the fe only).
In 2.6-dev1, NUMA topology detection was enabled on FreeBSD with commit
f5d48f8b3 ("MEDIUM: cfgparse: numa detect topology on FreeBSD."). But
it suffers from a minor bug which is that it forgets to check for the
number of domains and always emits a confusing warning indicating that
multiple sockets were found while it's not the case.
This can be backported to 2.6.
The lib compatibility checks introduced in 2.8-dev6 with commit c3b297d5a
("MEDIUM: tools: further relax dlopen() checks too consider grouped
symbols") were partially incorrect in that they check at the same time
libcrypto and libssl. But if loading a library that only depends on
libcrypto, the ssl-only symbols will be missing and this might present
an inconsistency. This is what is observed on FreeBSD 13.1 when
libcrypto is being loaded, where it sees two symbols having disappeared.
The fix consists in splitting the checks for libcrypto and libssl.
No backport is needed, unless the patch above finally gets backported.
On FreeBSD 13.1 I noticed that thread balancing using shards was not
always working. Sometimes several threads would work, but most of the
time a single one was taking all the traffic. This is related to how
SO_REUSEPORT works on FreeBSD since version 12, as it seems there is
no guarantee that multiple sockets will receive the traffic. However
there is SO_REUSEPORT_LB that is designed exactly for this, so we'd
rather use it when available.
This patch may possibly be backported, but nobody complained and it's
not sure that many users rely on shards. So better wait for some feedback
before backporting this.
In 2.7-dev2, "stats bind-process" was removed by commit 94f763b5e
("MEDIUM: config: remove deprecated "bind-process" directives from
frontends") and an error message indicates that it's no more supported.
However it says "stats" is not supported instead of "stats bind-process",
making it a bit confusing.
This should be backported to 2.7.
By comparing the local thread's load with the least loaded thread's
load, we can further improve the fairness and at the same time also
improve locality since it allows a small ratio of connections not to
be migrated. This is visible on CPU usage with long connections on
very large thread counts (224) and high bandwidth (200G). The cost
of checking the local thread's load remains fairly low so there's no
reason not to do this. We continue to update the index if we select
the local thread, because it means that the two other threads were
both more loaded so we'd rather find better ones.
One limitation of the current thread index mechanism is that if the
values are assigned multiple times to the same thread and the index
loops, it can match again the old value, which will not prevent a
competing thread from finishing its CAS and assigning traffic to a
thread that's not the optimal one. The probability is low but the
solution is simple enough and consists in implementing an update
counter in the high bits of the index to force a mismatch in this
case (assuming we don't try to cover for extremely unlikely cases
where the update counter loops while the index remains equal). So
let's do that. In order to improve the situation a little bit, we
now set the index to a ulong so that in 32 bits we have 8 bits of
counter and in 64 bits we have 40 bits.
During heavy accept competition, the CAS will occasionally fail and
we'll have to go through all the calculation again. While the first
two loops look heavy, they're almost never taken so they're quite
cheap. However the rest of the operation is heavy because we have to
consult connection counts and queue indexes for other threads, so
better double-check if the index is still valid before continuing.
Tests show that it's more efficient do retry half-way like this.
Instead of seeing each listener use its own thr_idx, let's use the same
for all those from a shard. It should provide more accurate and smoother
thread allocation.
Till now threads were assigned in listener_accept() to other threads of
the same group only, using a single group mask. Now that we have all the
relevant info (array of listeners of the same shard), we can spread the
thr_idx to cover all assigned groups. The thread indexes now contain the
group number in their upper bits, and the indexes run over te whole list
of threads, all groups included.
One particular subtlety here is that switching to a thread from another
group also means switching the group, hence the listener. As such, when
changing the group we need to update the connection's owner to point to
the listener of the same shard that is bound to the target group.
There has always been a race when checking the length of an accept queue
to determine which one is more loaded that another, because the head and
tail are read at two different moments. This is not required, we can merge
them as two 16 bit numbers inside a single 32-bit index that is always
accessed atomically. This way we read both values at once and always have
a consistent measurement.
Now it's possible for a bind line to span multiple thread groups. When
this happens, the first one will become the reference and will be entirely
set up, and the subsequent ones will be duplicated from this reference,
so that they can be registered in distinct groups. The reference is
always setup and started first so it is always available when the other
ones are started.
The doc was updated to reflect this new possibility with its limitations
and impacts, and the differences with the "shards" option.
It's not strictly necessary, but it's still better to avoid setting
up the same socket multiple times when it's being duplicated to a few
FDs. We don't change that for inherited ones however since they may
really need to be set up, so we only skip duplicated ones.
The different protocol's ->bind() function will now check the receiver's
RX_F_MUST_DUP flag to decide whether to bind a fresh new listener from
scratch or reuse an existing one and just duplicate it. It turns out
that the existing code already supports reusing FDs since that was done
as part of the FD passing and inheriting mechanism. Here it's not much
different, we pass the FD of the reference receiver, it gets duplicated
and becomes the new receiver's FD.
These FDs are also marked RX_F_INHERITED so that they are not exported
and avoid being touched directly (only the reference should be touched).
In order to create multiple receivers for one multi-group shard, we'll
need some more info about the shard. Here we store:
- the number of groups (= number of receivers)
- the number of threads (will be used for accept LB)
- pointer to the reference rx (to get the FD and to find all threads)
- pointers to the other members (to iterate over all threads)
For now since there's only one group per shard it remains simple. The
listener deletion code already takes care of removing the current
member from its shards list and moving others' reference to the last
one if it was their reference (so as to avoid o(n^2) updates during
ordered deletes).
Since the vast majority of setups will not use multi-group shards, we
try to save memory usage by only allocating the shard_info when it is
needed, so the principle here is that a receiver shard_info==NULL is
alone and doesn't share its socket with another group.
Various approaches were considered and tests show that the management
of the listeners during boot makes it easier to just attach to or
detach from a shard_info and automatically allocate it if it does not
exist, which is what is being done here.
For now the attach code is not called, but detach is already called
on delete.
This new algorithm for rebalancing incoming connections to multiple
threads is simpler and instead of considering the threads load, it will
only cycle through all of them, offering a fair share of the traffic to
each thread. It may be well suited for short-lived connections but is
also convenient for very large thread counts where it's not always certain
that the least loaded thread will always be found.
There's a li_per_thread array in each listener for use with QUIC
listeners. Since thread groups were introduced, this array can be
allocated too large because global.nbthread is allocated for each
listener, while only no more than MIN(nbthread,MAX_THREADS_PER_GROUP)
may be used by a single listener. This was because the global thread
ID is used as the index instead of the local ID (since a listener may
only be used by a single group). Let's just switch to local ID and
reduce the allocated size.
