We've reached a point where the global pools represent a significant
bottleneck with threads. On a 64-core machine, the performance was
divided by 8 between 32 and 64 H2 connections only because there were
not enough entries in the local caches to avoid picking from the global
pools, and the contention on the list there was very high. It becomes
obvious that we need to have an array of lists, but that will require
more changes.
In parallel, standard memory allocators have improved, with tcmalloc
and jemalloc finding their ways through mainstream systems, and glibc
having upgraded to a thread-aware ptmalloc variant, keeping this level
of contention here isn't justified anymore when we have both the local
per-thread pool caches and a fast process-wide allocator.
For these reasons, this patch introduces a new compile time setting
CONFIG_HAP_NO_GLOBAL_POOLS which is set by default when threads are
enabled with thread local pool caches, and we know we have a fast
thread-aware memory allocator (currently set for glibc>=2.26). In this
case we entirely bypass the global pool and directly use the standard
memory allocator when missing objects from the local pools. It is also
possible to force it at compile time when a good allocator is used with
another setup.
It is still possible to re-enable the global pools using
CONFIG_HAP_GLOBAL_POOLS, if a corner case is discovered regarding the
operating system's default allocator, or when building with a recent
libc but a different allocator which provides other benefits but does
not scale well with threads.
There finally is a way to support register pairs on aarch64 assembly
under gcc, it's just undocumented, like many of the options there :-(
As indicated below, it's possible to pass "%H" to mention the high
part of a register pair (e.g. "%H0" to go with "%0"):
https://patchwork.ozlabs.org/project/gcc/patch/59368A74.2060908@foss.arm.com/
By making local variables from pairs of registers via a struct (as
is used in IST for example), we can let gcc choose the correct register
pairs and avoid a few moves in certain situations. The code is now slightly
more efficient than the previous one on AWS' Graviton2 platform, and
noticeably smaller (by 4.5kB approx). A few tests on older releases show
that even Linaro's gcc-4.7 used to support such register pairs and %H,
and by then ATOMICS were not supported so this should not cause build
issues, and as such this patch replaces the earlier implementation.
This variant uses the CASP instruction available on armv8.1-a CPU cores,
which is detected when __ARM_FEATURE_ATOMICS is set (gcc-linaro >= 7,
mainline >= 9). This one was tested on cortex-A55 (S905D3) and on AWS'
Graviton2 CPUs.
The instruction performs way better on high thread counts since it
guarantees some forward progress when facing extreme contention while
the original LL/SC approach is light on low-thread counts but doesn't
guarantee progress.
The implementation is not the most optimal possible. In particular since
the instruction requires to work on register pairs and there doesn't seem
to be a way to force gcc to emit register pairs, we have to decide to force
to use the pair (x0,x1) to store the old value, and (x2,x3) to store the
new one, and this necessarily involves some extra moves. But at least it
does improve the situation with 16 threads and more. See issue #958 for
more context.
Note, a first implementation of this function was making use of an
input/output constraint passed using "+Q"(*(void**)target), which was
resulting in smaller overall code than passing "target" as an input
register only. It turned out that the cause was directly related to
whether the function was inlined or not, hence the "forceinline"
attribute. Any changes to this code should still pay attention to this
important factor.
On highly threaded machines it is possible to occasionally trigger the
watchdog on certain contended areas like the server's connection list,
because while the mechanism inherently cannot guarantee a constant
progress, it lacks CPU relax calls which are absolutely necessary in
this situation to let a thread finish its job.
The loop's "while (1)" was changed to use a "for" statement calling
__ha_cpu_relax() as its continuation expression. This way the "continue"
statements jump to the unique place containing the pause without
excessively inflating the code.
This was sufficient to definitely fix the problem on 64-core ARM Graviton2
machines. This patch should probably be backported once it's confirmed it
also helps on many-cores x86 machines since some people are facing
contention in these environments. This patch depends on previous commit
"REORG: atomic: reimplement pl_cpu_relax() from atomic-ops.h".
An attempt was made to first read the value before exchanging, and it
significantly degraded the performance. It's very likely that this caused
other cores to lose exclusive ownership on their line and slow down their
next xchg operation.
In addition it was found that MT_LIST_ADD is significantly faster than
MT_LIST_ADDQ under high contention, because it fails one step earlier
when conflicting with an adjacent MT_LIST_DEL(). It might be worth
switching some operations' order to favor MT_LIST_ADDQ() instead.
There is some confusion here as we need to place some cpu_relax statements
in some loops where it's not easily possible to condition them on the use
of threads. That's what atomic.h already does. So let's take the various
pl_cpu_relax() implementations from there and place them in atomic.h under
the name __ha_cpu_relax() and let them adapt to the presence or absence of
threads and to the architecture (currently only x86 and aarch64 use a barrier
instruction), though it's very likely that arm would work well with a cache
flushing ISB instruction as well).
This time they were implemented as expressions returning 1 rather than
statements, in order to ease their placement as the loop condition or the
continuation expression inside "for" loops. We should probably do the same
with barriers and a few such other ones.
If dispatch mode or transparent backend is used, the backend connection
target is a proxy instead of a server. In these cases, the reuse of
backend connections is not consistent.
With the default behavior, no reuse is done and every new request uses a
new connection. However, if http-reuse is set to never, the connection
are stored by the mux in the session and can be reused for future
requests in the same session.
As no server is used for these connections, no reuse can be made outside
of the session, similarly to http-reuse never mode. A different
http-reuse config value should not have an impact. To achieve this, mark
these connections as private to have a defined behavior.
For this feature to properly work, the connection hash has been slightly
adjusted. The server pointer as an input as been replaced by a generic
target pointer to refer to the server or proxy instance. The hash is
always calculated on connect_server even if the connection target is not
a server. This also requires to allocate the connection hash node for
every backend connections, not just the one with a server target.
Fix such compilation issues:
include/haproxy/quic_tls.h:157:10: error: implicit conversion from
'enum ssl_encryption_level_t' to 'enum quic_tls_enc_level'
[-Werror=enum-conversion]
157 | return ssl_encryption_application;
| ^~~~~~~~~~~~~~~~~~~~~~~~~~
src/xprt_quic.c: In function 'quic_conn_enc_level_init':
src/xprt_quic.c:2358:13: error: implicit conversion from
'enum quic_tls_enc_level' to 'enum ssl_encryption_level_t'
[-Werror=enum-conversion]
2358 | qel->level = quic_to_ssl_enc_level(level);
| ^
Not detected by all the compilators.
This makes the code more readable and less prone to copy-paste errors.
In addition, it allows to place some __builtin_constant_p() predicates
to trigger a link-time error in case the compiler knows that the freed
area is constant. It will also produce compile-time error if trying to
free something that is not a regular pointer (e.g. a function).
The DEBUG_MEM_STATS macro now also defines an instance for ha_free()
so that all these calls can be checked.
178 occurrences were converted. The vast majority of them were handled
by the following Coccinelle script, some slightly refined to better deal
with "&*x" or with long lines:
@ rule @
expression E;
@@
- free(E);
- E = NULL;
+ ha_free(&E);
It was verified that the resulting code is the same, more or less a
handful of cases where the compiler optimized slightly differently
the temporary variable that holds the copy of the pointer.
A non-negligible amount of {free(str);str=NULL;str_len=0;} are still
present in the config part (mostly header names in proxies). These
ones should also be cleaned for the same reasons, and probably be
turned into ist strings.
A network may be specified to avoid header addition for "forwardfor" and
"orignialto" option via the "except" parameter. However, only IPv4
networks/addresses are supported. This patch adds the support of IPv6.
To do so, the net_addr structure is used to store the parameter value in the
proxy structure. And ipcmp2net() function is used to perform the comparison.
This patch should fix the issue #1145. It depends on the following commit:
* c6ce0ab MINOR: tools: Add function to compare an address to a network address
* 5587287 MINOR: tools: Add net_addr structure describing a network addess
ipcmp2net() function may be used to compare an addres (struct
sockaddr_storage) to a network address (struct net_addr). Among other
things, this function will be used to add support of IPv6 for "except"
parameter of "forwardfor" and "originalto" options.
The net_addr structure describes a IPv4 or IPv6 address. Its ip and mask are
represented. Among other things, this structure will be used to add support
of IPv6 for "except" parameter of "forwardfor" and "originalto" options.
This class will be used exclusively for heavy processing tasklets. It
will be cleaner than mixing them with the bulk ones. For now it's
allocated ~1% of the CPU bandwidth.
The largest part of the patch consists in re-arranging the fields in the
task_per_thread structure to preserve a clean alignment with one more
list head. Since we're now forced to increase the struct past a second
cache line, it now uses 4 cache lines (for easy multiplying) with the
first two ones being exclusively used by local operations and the third
one mostly by atomic operations. Interestingly, this better arrangement
causes less stress and reduced the response time by 8 microseconds at
1 million requests per second.
These function names are unbearably long, they don't even fit into the
screen in "show profiling", let's trim the "_connections" to "_conns",
which happens to match the name of the lists there.
While the scheduler is priority-aware and class-aware, and consistently
tries to maintain fairness between all classes, it doesn't make use of a
fine execution budget to compensate for high-latency tasks such as TLS
handshakes. This can result in many subsequent calls adding multiple
milliseconds of latency between the various steps of other tasklets that
don't even depend on this.
An ideal solution would be to add a 4th queue, have all tasks announce
their estimated cost upfront and let the scheduler maintain an auto-
refilling budget to pick from the most suitable queue.
But it turns out that a very simplified version of this already provides
impressive gains with very tiny changes and could easily be backported.
The principle is to reserve a new task flag "TASK_HEAVY" that indicates
that a task is expected to take a lot of time without yielding (e.g. an
SSL handshake typically takes 700 microseconds of crypto computation).
When the scheduler sees this flag when queuing a tasklet, it will place
it into the bulk queue. And during dequeuing, we accept only one of
these in a full round. This means that the first one will be accepted,
will not prevent other lower priority tasks from running, but if a new
one arrives, then the queue stops here and goes back to the polling.
This will allow to collect more important updates for other tasks that
will be batched before the next call of a heavy task.
Preliminary tests consisting in placing this flag on the SSL handshake
tasklet show that response times under SSL stress fell from 14 ms
before the patch to 3.0 ms with the patch, and even 1.8 ms if
tune.sched.low-latency is set to "on".
Some static functions are now exported and renamed to follow the same
pattern of other exported functions. Here is the list :
* update_server_fqdn: Renamed to srv_update_fqdn and exported
* update_server_check_addr_port: renamed to srv_update_check_addr_port and exported
* update_server_agent_addr_port: renamed to srv_update_agent_addr_port and exported
* update_server_addr: renamed to srv_update_addr
* update_server_addr_potr: renamed to srv_update_addr_port
* srv_prepare_for_resolution: exported
This change is mandatory to move all functions dealing with the server-state
files in a separate file.
Parsed parameters are now stored in the tree of server-state lines. This
way, a line from the global server-state file is only parsed once. Before,
it was parsed a first time to store it in the tree and one more time to load
the server state. To do so, the server-state line object must be allocated
before parsing a line. This means its size must no longer depend on the
length of first parsed parameters (backend and server names). Thus the node
type was changed to use a hashed key instead of a string.
The structure used to store a server-state line in an eb-tree has a too
generic name. Instead of state_line, the structure is renamed as
server_state_line.
<state_line.name_name> field is a node in an eb-tree. Thus, instead of
"name_name", we now use "node" to name this field. If is a more explicit
name and not too strange.
It is extremely useful to be able to observe the wakeup latency of some
important I/O operations, so let's accept to inflate the tasklet struct
by 8 extra bytes when DEBUG_TASK is set. With just this we have enough
to get live reports like this:
$ socat - /tmp/sock1 <<< "show profiling"
Per-task CPU profiling : on # set profiling tasks {on|auto|off}
Tasks activity:
function calls cpu_tot cpu_avg lat_tot lat_avg
si_cs_io_cb 8099492 4.833s 596.0ns 8.974m 66.48us
h1_io_cb 7460365 11.55s 1.548us 2.477m 19.92us
process_stream 7383828 22.79s 3.086us 18.39m 149.5us
h1_timeout_task 4157 - - 348.4ms 83.81us
srv_cleanup_toremove_connections751 39.70ms 52.86us 10.54ms 14.04us
srv_cleanup_idle_connections 21 1.405ms 66.89us 30.82us 1.467us
task_run_applet 16 1.058ms 66.13us 446.2us 27.89us
accept_queue_process 7 34.53us 4.933us 333.1us 47.58us
Instead of decrementing grq_total once per task picked from the global
run queue, let's do it at once after the loop like we do for other
counters. This simplifies the code everywhere. It is not expected to
bring noticeable improvements however, since global tasks tend to be
less common nowadays.
The function htx_reserve_max_data() should be used to get an HTX DATA block
with the max possible size. A current block may be extended or a new one
created, depending on the HTX message state. But the idea is to let the
caller to copy a bunch of data without requesting many new blocks. It is its
responsibility to resize the block at the end, to set the final block size.
This function will be used to parse messages with small chunks. Indeed, we
can have more than 2700 1-byte chunks in a 16Kb of input data. So it is easy
to understand how this function may help to improve the parsing of chunk
messages.
When a SRV record was created, it used to register the regular server name
resolution callbacks. That said, SRV records and regular server name
resolution don't work the same way, furthermore on error management.
This patch introduces a new call back to manage DNS errors related to
the SRV queries.
this fixes github issue #50.
Backport status: 2.3, 2.2, 2.1, 2.0
In fd_set_running_excl() we don't reset the old mask in the CAS loop,
so if we fail on the first round, we'll forcefully take the FD on the
next one.
In practice it's used bu fd_insert() and fd_delete() only, none of which
is supposed to be passed an FD which is still in use since in practice,
given that for now only listeners may be enabled on multiple threads at
once.
This can be backported to 2.2 but shouldn't result in fixing any user
visible bug for now.
This function has become large with the multi-queue scheduler. We need
to keep the fast path and the debugging parts inlined, but the rest now
moves to task.c just like was done for task_wakeup(). This has reduced
the code size by 6kB due to less inlining of large parts that are always
context-dependent, and as a side effect, has increased the overall
performance by 1%.
The nb_tasks counter was still global and gets incremented and decremented
for each task_new()/task_free(), and was read in process_runnable_tasks().
But it's only used for stats reporting, so doing this this often is
pointless and expensive. Let's move it to the task_per_thread struct and
have the stats sum it when needed.
Historically we used to call __task_wakeup() with a known tree root but
this is not the case and the code has remained needlessly complicated
with the root calculation in task_wakeup() passed in argument to
__task_wakeup() which compares it again.
Let's get rid of this and just move the detection code there. This
eliminates some ifdefs and allows to simplify the test conditions quite
a bit.
This one is systematically misunderstood due to its unclear name. It
is in fact the number of tasks in the local tasklet list. Let's call
it "tasks_in_list" to remove some of the confusion.
This one is exclusively used as a boolean nowadays and is non-zero only
when the thread-local run queue is not empty. Better check the root tree's
pointer and avoid updating this counter all the time.
This counter is solely used for reporting in the stats and is the hottest
thread contention point to date. Moving it to the scheduler and having a
separate one for the global run queue dramatically improves the performance,
showing a 12% boost on the request rate on 16 threads!
In addition, the thread debugging output which used to rely on rqueue_size
was not totally accurate as it would only report task counts. Now we can
return the exact thread's run queue length.
It is also interesting to note that there are still a few other task/tasklet
counters in the scheduler that are not efficiently updated because some cover
a single area and others cover multiple areas. It looks like having a distinct
counter for each of the following entries would help and would keep the code
a bit cleaner:
- global run queue (tree)
- per-thread run queue (tree)
- per-thread shared tasklets list
- per-thread local lists
Maybe even splitting the shared tasklets lists between pure tasklets and
tasks instead of having the whole and tasks would simplify the code because
there remain a number of places where several counters have to be updated.
The lock was still used exclusively to deal with the concurrency between
the "show sess" release handler and a stream_new() or stream_free() on
another thread. All other accesses made by "show sess" are already done
under thread isolation. The release handler only requires to unlink its
node when stopping in the middle of a dump (error, timeout etc). Let's
just isolate the thread to deal with this case so that it's compatible
with the dump conditions, and remove all remaining locking on the streams.
This effectively kills the streams lock. The measured gain here is around
1.6% with 4 threads (374krps -> 380k).
The global streams list is exclusively used for "show sess", to look up
a stream to shut down, and for the hard-stop. Having all of them in a
single list is extremely expensive in terms of locking when using threads,
with performance losses as high as 7% having been observed just due to
this.
This patch makes the list per-thread, since there's no need to have a
global one in this situation. All call places just iterate over all
threads. The most "invasive" changes was in "show sess" where the end
of list needs to go back to the beginning of next thread's list until
the last thread is seen. For now the lock was maintained to keep the
code auditable but a next commit should get rid of it.
The observed performance gain here with only 4 threads is already 7%
(350krps -> 374krps).
The "show sess" CLI command currently lists all streams and needs to
stop at a given position to avoid dumping forever. Since 2.2 with
commit c6e7a1b8e ("MINOR: cli: make "show sess" stop at the last known
session"), a hack consists in unlinking the stream running the applet
and linking it again at the current end of the list, in order to serve
as a delimiter. But this forces the stream list to be global, which
affects scalability.
This patch introduces an epoch, which is a global 32-bit counter that
is incremented by the "show sess" command, and which is copied by newly
created streams. This way any stream can know whether any other one is
newer or older than itself.
For now it's only stored and not exploited.
RAND_keep_random_devices_open is OpenSSL specific function, not
implemented in LibreSSL and BoringSSL. Let us define guard
HAVE_SSL_RAND_KEEP_RANDOM_DEVICES_OPEN in include/haproxy/openssl-compat.h
That guard does not depend anymore on HA_OPENSSL_VERSION
The runqueue_ticks counts the number of task wakeups and is used to
position new tasks in the run queue, but since we've had per-thread
run queues, the values there are not very relevant anymore and the
nice value doesn't apply well if some threads are more loaded than
others. In addition, letting all threads compete over a shared counter
is not smart as this may cause some excessive contention.
Let's move this index close to the run queues themselves, i.e. one per
thread and a global one. In addition to improving fairness, this has
increased global performance by 2% on 16 threads thanks to the lower
contention on rqueue_ticks.
Fairness issues were not observed, but if any were to be, this patch
could be backported as far as 2.0 to address them.
Historically this function would try to wake the most accurate number of
process_stream() waiters. But since the introduction of filters which could
also require buffers (e.g. for compression), things started not to be as
accurate anymore. Nowadays muxes and transport layers also use buffers, so
the runqueue size has nothing to do anymore with the number of supposed
users to come.
In addition to this, the threshold was compared to the number of free buffer
calculated as allocated minus used, but this didn't work anymore with local
pools since these counts are not updated upon alloc/free!
Let's clean this up and pass the number of released buffers instead, and
consider that each waiter successfully called counts as one buffer. This
is not rocket science and will not suddenly fix everything, but at least
it cannot be as wrong as it is today.
This could have been marked as a bug given that the current situation is
totally broken regarding this, but this probably doesn't completely fix
it, it only goes in a better direction. It is possible however that it
makes sense in the future to backport this as part of a larger series if
the situation significantly improves.
The buffer wait queue used to be global historically but this doest not
make any sense anymore given that the most common use case is to have
thread-local pools. Thus there's no point waking up waiters of other
threads after releasing an entry, as they won't benefit from it.
Let's move the queue head to the thread_info structure and use
ti->buffer_wq from now on.
This revert the commit 63e6cba12 ("MEDIUM: server: add server-states version
2"), but keeping all recent features added to the server-sate file. Instead
of adding a 2nd version for the server-state file format to handle the 5 new
fields added during the 2.4 development, these fields are considered as
optionnal during the parsing. So it is possible to load a server-state file
from HAProxy 2.3. However, from 2.4, these new fields are always dumped in
the server-state file. But it should not be a problem to load it on the 2.3.
This patch seems a bit huge but the diff ignoring the space is much smaller.
The version 2 of the server-state file format is reserved for a real
refactoring to address all issues of the current format.
Remove ebmb_node entry from struct connection and create a dedicated
struct conn_hash_node. struct connection contains now only a pointer to
a conn_hash_node, allocated only for connections where target is of type
OBJ_TYPE_SERVER. This will reduce memory footprints for every
connections that does not need http-reuse such as frontend connections.
In MT_LIST_TRY_ADDQ(), deal with the "prev" field of the element before the
"next". If the element is the first in the list, then its next will
already have been locked when we locked list->prev->next, so locking it
again will fail, and we'll start over and over.
This should be backported to 2.3.
The maximum number of connections accepted at once by a thread for a single
listener used to default to 64 divided by the number of processes but the
tasklet-based model is much more scalable and benefits from smaller values.
Experimentation has shown that 4 gives the highest accept rate for all
thread values, and that 3 and 5 come very close, as shown below (HTTP/1
connections forwarded per second at multi-accept 4 and 64):
ac\thr| 1 2 4 8 16
------+------------------------------
4| 80k 106k 168k 270k 336k
64| 63k 89k 145k 230k 274k
Some tests were also conducted on SSL and absolutely no change was observed.
The value was placed into a define because it used to be spread all over the
code.
It might be useful at some point to backport this to 2.3 and 2.2 to help
those who observed some performance regressions from 1.6.
Since a lot of internal callbacks were turned to tasklets, the runqueue
depth had not been readjusted from the default 200 which was initially
used to favor batched processing. But nowadays it appears too large
already based on the following tests conducted on a 8c16t machine with
a simple config involving "balance leastconn" and one server. The setup
always involved the two threads of a same CPU core except for 1 thread,
and the client was running over 1000 concurrent H1 connections. The
number of requests per second is reported for each (runqueue-depth,
nbthread) couple:
rq\thr| 1 2 4 8 16
------+------------------------------
32| 120k 159k 276k 477k 698k
40| 122k 160k 276k 478k 722k
48| 121k 159k 274k 482k 720k
64| 121k 160k 274k 469k 710k
200| 114k 150k 247k 415k 613k <-- default
It's possible to save up to about 18% performance by lowering the
default value to 40. One possible explanation to this is that checking
I/Os more frequently allows to flush buffers faster and to smooth the
I/O wait time over multiple operations instead of alternating phases
of processing, waiting for locks and waiting for new I/Os.
The total round trip time also fell from 1.62ms to 1.40ms on average,
among which at least 0.5ms is attributed to the testing tools since
this is the minimum attainable on the loopback.
After some observation it would be nice to backport this to 2.3 and
2.2 which observe similar improvements, since some users have already
observed some perf regressions between 1.6 and 2.2.
SCTL (signed certificate timestamp list) specified in RFC6962
was implemented in c74ce24cd22e8c683ba0e5353c0762f8616e597d, let
us introduce macro HAVE_SSL_SCTL for the HAVE_SSL_SCTL sake,
which in turn is based on SN_ct_cert_scts, which comes in the same commit
smp_is_safe() function is used to be sure a sample may be safely
modified. For string samples, a test is performed to verify if there is a
null-terminated byte. If not, one is added, if possible. It means if the
sample is not const and if there is some free space in the buffer, after
data. However, we must not try to read the null-terminated byte if the
string sample is too long (data >= size) or if the size is equal to
zero. This last test was not performed. Thus it was possible to consider a
string sample as safe by testing a byte outside the buffer.
Now, a zero size string sample is always considered as unsafe and is
duplicated when smp_make_safe() is called.
This patch must be backported in all stable versions.
It's pretty easy to pre-initialize the index, change it on free() and check
it during the wakeup, so let's do this to ease detection of any accidental
task_wakeup() after a task_free() or tasklet_wakeup() after a tasklet_free().
If this would ever happen we'd then get a backtrace and a core now. The
index's parity is respected so that the call history remains exploitable.
The idea is to know who woke a task up, by recording the last two
callers in a rotating mode. For now it's trivial with task_wakeup()
but tasklet_wakeup_on() will require quite some more changes.
This typically gives this from the debugger:
(gdb) p t->debug
$2 = {
caller_file = {0x0, 0x8c0d80 "src/task.c"},
caller_line = {0, 260},
caller_idx = 1
}
or this:
(gdb) p t->debug
$6 = {
caller_file = {0x7fffe40329e0 "", 0x885feb "src/stream.c"},
caller_line = {284, 284},
caller_idx = 1
}
But it also provides a trivial macro allowing to simply place a call in
a task/tasklet handler that needs to be observed:
DEBUG_TASK_PRINT_CALLER(t);
Then starting haproxy this way would trivially yield such info:
$ ./haproxy -db -f test.cfg | sort | uniq -c | sort -nr
199992 h1_io_cb woken up from src/sock.c:797
51764 h1_io_cb woken up from src/mux_h1.c:3634
65 h1_io_cb woken up from src/connection.c:169
45 h1_io_cb woken up from src/sock.c:777
The two algos defining these functions (first and leastconn) do not need the
server's lock. However it's already present in pendconn_process_next_strm()
so the API must be updated so that the functions may take it if needed and
that the callers indicate whether they already own it.
As such, the call places (backend.c and stream.c) now do not take it
anymore, queue.c was unchanged since it's already held, and both "first"
and "leastconn" were updated to take it if not already held.
A quick test on the "first" algo showed a jump from 432 to 565k rps by
just dropping the lock in stream.c!
This reverts commit 8f1f177ed0.
Repeated tests have shown a small perforamnce degradation of ~1.8%
caused by this patch at high request rates on 16 threads. The exact
cause is not yet perfectly known but it probably stems in slower
accesses for non-64-bit aligned atomic accesses.
