It's pointless to requeue the task when we're closing, so swap the
order of the task_queue() and h2_release(). It also matches what
was written in the comment regarding re-arming the timer.
h2_detach() is called after a stream was closed, and it evaluates if it's
worth closing the connection. The issue there is that the connection is
closed too early in case there's demand for closing after the last stream,
even if some data remain in the mux. Let's change the condition to check
for this.
When the assignment of the connection state was moved into h2c_error(),
3 of them were missed because they were wrong, using H2_SS_ERROR instead.
This resulted in the connection's state being set to H2_CS_ERROR2 in fact,
so the error was not properly sent.
This one was created to maintain the knowledge that a stream was closed
after having sent an RST_STREAM frame but that's not needed anymore and
it confuses certain conditions on the error processing path. It's time
to get rid of it.
The call to xprt->snd_buf() was not conditionned on the presence of
data in the buffer, resulting in snd_buf() returning 0 and never
disabling the polling. It was revealed by the previous bug on error
processing but must properly be handled.
Some stream errors are detected on the MUX path (eg: H1 response
encoding). The ones forgot to emit an RST_STREAM frame, causing the
client to wait and/or to see the connection being immediately closed.
This is now fixed.
This flag was added after the GOAWAY flags were introduced and mistakenly
placed in the connection, but that doesn't make sense as it's specific to
the stream. The main impact is the risk of returning a DATA0+ES frame for
an error instead of an RST_STREAM.
From first-hand experience I realized that using encrypted passwords in
userlists can quickly become overwhelming for busy sites. In my case
just about 100 rq/s were enough to drive (user) CPU usage from 2-3% up
to >90%. While it is perfectly explicable why this is the case, having
it mentioned in the relevant documentation section might spare someone
some confusion in the future.
snr_check_ip_callback() may be called with the server lock, so don't attempt
to lock it again, instead, make sure the callers always have the lock before
calling it.
The scheduler is complex and uses local queues to amortize the cost of
locks. But all this comes with a cost that is quite observable with
single-thread workloads.
The purpose of this patch is to reimplement the much simpler scheduler
for the case where threads are not used. The code is very small and
simple. It doesn't impact the multi-threaded performance at all, and
provides a nice 10% performance increase in single-thread by reaching
606kreq/s on the tests that showed 550kreq/s before.
process_runnable_tasks() needs to requeue or wake up tasks after
processing them in batches. By only refilling the existing ones, we
avoid revisiting all the queue. The performance gain is measurable
starting with two threads, where the request rate climbs to 657k/s
compared to 644k.
The spin locks used to rely on W locks, which involve a loop waiting
for readers to leave, and this doesn't happen here. It's more efficient
to use S locks instead, which are also mutually exclusive and do not
have this loop. This saves one test per spinlock and a few tens of
bytes allowing certain functions to be inlined.
This patch slightly rearranges the loop to pack the locked code a little
bit, and to try to concentrate accesses to the tree together to benefit
more from the cache.
It also fixes how the loop handles the right margin : now that is guaranteed
that the retrieved nodes are filtered to only match the current thread, we
don't need to rewind every 16 entries. Instead we can rewind each time we
reach the right margin again.
With this change, we now achieve the following performance for 10 H2 conns
each containing 100 streams :
1 thread : 550kreq/s
2 thread : 644kreq/s
3 thread : 598kreq/s
This function is sensitive, let's make it shorter by factoring out the
unlock and leave code. This reduced the function's size by a few tens
of bytes and increased the overall performance by about 1%.
Currently the task scheduler suffers from an O(n) lookup when
skipping tasks that are not for the current thread. The reason
is that eb32_lookup_ge() has no information about the current
thread so it always revisits many tasks for other threads before
finding its own tasks.
This is particularly visible with HTTP/2 since the number of
concurrent streams created at once causes long series of tasks
for the same stream in the scheduler. With only 10 connections
and 100 streams each, by running on two threads, the performance
drops from 640kreq/s to 11.2kreq/s! Lookup metrics show that for
only 200000 task lookups, 430 million skips had to be performed,
which means that on average, each lookup leads to 2150 nodes to
be visited.
This commit backports the principle of scope lookups for ebtrees
from the ebtree_v7 development tree. The idea is that each node
contains a mask indicating the union of the scopes for the nodes
below it, which is fed during insertion, and used during lookups.
Then during lookups, branches that do not contain any leaf matching
the requested scope are simply ignored. This perfectly matches a
thread mask, allowing a thread to only extract the tasks it cares
about from the run queue, and to always find them in O(log(n))
instead of O(n). Thus the scheduler uses tid_bit and
task->thread_mask as the ebtree scope here.
Doing this has recovered most of the performance, as can be seen on
the test below with two threads, 10 connections, 100 streams each,
and 1 million requests total :
Before After Gain
test duration : 89.6s 4.73s x19
HTTP requests/s (DEBUG) : 11200 211300 x19
HTTP requests/s (PROD) : 15900 447000 x28
spin_lock time : 85.2s 0.46s /185
time per lookup : 13us 40ns /325
Even when going to 6 threads (on 3 hyperthreaded CPU cores), the
performance stays around 284000 req/s, showing that the contention
is much lower.
A test showed that there's no benefit in using this for the wait queue
though.
In the scheduler we always have to loop back to the beginning after
we don't find the last entry, so let's implement this in a new lookup
function instead. The resulting code is slightly faster, mostly due
to the fact that there's much less inlined code in the fast path.
Now when looking up a node via eb32sc_first(), eb32sc_next(), and
eb32sc_lookup_ge(), we only focus on the branches matching the requested
scope. The code must be careful to miss no branch. It changes a little
bit from the previous one because the scope stored on the intermediary
nodes is not exact (since we don't propagate upwards during deletion),
so in case a lookup fails, we have to walk up and pick the next matching
entry.
During a delete operation, if the deleted node is above its leaf's
parent, this parent will replace the node and then go up. In this
case it is important to update the new parent's scope to reflect
the presence of other branches.
It's worth noting that in theory we should precisely recompute the
exact node value, but it seems that it's not worth it for the rare
cases there is a mismatch.
A new kind of tree nodes is currently being developed in ebtree v7,
consisting in storing a scope in each node indicating a visibility
mask so that certain nodes are not reported on certain lookups. The
initial goal was to make this usable with a multi-thread scheduler.
Since the ebtree v7 code is completely different from v6, this patch
instead copies the minimally required functions from eb32 and ebtree
and calls them "eb32sc_*". At the moment the scope is not implemented,
it's only passed in arguments.
The first pid in the pidfile is now the parent, it's more convenient for
supervising the processus.
You can now reload haproxy in master-worker mode with convenient command
like: kill -USR2 $(head -1 /tmp/haproxy.pid)
The __appctx_wakeup() function already does it. It matters with threads
enabled because it simplifies the code in appctx_res_wakeup() to get rid
of this test.
Commit 0493149 ("MINOR: thread: report multi-thread support in haproxy -vv")
added information about thread support in haproxy -vv output but accidently
marked the message as "must_free" while it's a constant. This causes a segv
on the old process on clean exit if threads are enabled. It doesn't affect
the stability during operations however.
unbind_listener() takes the listener lock, which is already held by
enable_listener(). This situation happens when starting with nbproc > 1
with some bind lines limited to a certain process, because in this case
enable_listener() tries to stop unneeded listeners.
This commit introduces __do_unbind_listeners() which must be called with
the lock held, and makes enable_listener() use this one. Given that the
only return code has never been used and that it starts to make the code
more complicated to propagate it before throwing it to the trash, the
function's return type was changed to void.
The spin_unlock() was called just before setting the expiry to
TICK_ETERNITY, so if another thread has the time to perform its
update and set a timeout, this would would clear it.
Commit c3680ec ("MINOR: add severity information to cli feedback messages")
introduced a severity level to CLI messages, but one of them was missed
on "set server addr". No backport is needed.
Given that all spinning loops we've had since 1.8-rc1 were caused by
unbalanced lock/unlock, let's get rid of all return statements in the
locked check functions and only exit via a a single unlock place.
The "set server <srv> check-port" CLI handler forgot to return after
detecting an error on the port number, and still proceeds with the action.
This needs to be backported to 1.7.
Some unlocks were missing, resulting in deadlocks even with a single thread.
We really need to make these functions safer by getting rid of all those
remaining "return" calls and only leave using a goto!
Released version 1.8-rc2 with the following main changes :
- BUG/MINOR: send-proxy-v2: fix dest_len in make_tlv call
- BUG/MINOR: send-proxy-v2: string size must include ('\0')
- MINOR: mux: Only define pipe functions on linux.
- MINOR: cache: Remove useless test for nonzero.
- MINOR: cache: Don't confuse act_return and act_parse_ret.
- BUG/MEDIUM: h2: don't try to parse incomplete H1 responses
- BUG/MEDIUM: checks/mux: always enable send-polling after connecting
- BUG/MAJOR: fix deadlock on healthchecks.
- BUG/MINOR: thread: fix a typo in the debug code
- BUILD: shctx: allow to be built without openssl
- BUG/MEDIUM: cache: don't try to resolve wrong filters
- BUG/MAJOR: buffers: fix get_buffer_nc() for data at end of buffer
- BUG/MINOR: freq: fix infinite loop on freq_ctr_period.
- BUG/MINOR: stdarg.h inclusion
- BUG/MINOR: dns: fix missing lock protection on server.
- BUG/MINOR: lua: fix missing lock protection on server.
- BUILD: enable USE_THREAD for OpenBSD build.
- BUG/MAJOR: mux_pt: don't dereference a connstream after ->wake()
- MINOR: thread: report multi-thread support in haproxy -vv
