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.
When testing strong queue contention on a 48-thread machine, some crashes
would frequently happen due to process_srv_queue() never leaving and
processing pending requests forever. A dump showed more than 500000
loops at once. The problem is that other threads find it working so
they don't do anything and are free to process their pending requests.
Because of this, the dequeuing thread can be kept busy forever and does
not process its own requests anymore (fortunately the watchdog stops it).
This patch adds a limit to the number of rounds, it limits it to
maxpollevents, which is reasonable because it's also an indicator of
latency and batches size. However there's a catch. If all requests
are finished when the thread ends the loop, there might not be enough
anymore to restart processing the queue. Thus we tolerate to re-enter
the loop to process one request at a time when it doesn't have any
anymore. This way we're leaving more room for another thread to take
on this task, and we're sure to eventually end this loop.
Doing this has also improved the overall dequeuing performance by ~20%
in highly contended situations with 48 threads.
It should be backported at least to 2.4, maybe even 2.2 since issues
were faced in the past on machines having many cores.
Now that the inter-task wakeups are cheap, there's no point in using
task_instant_wakeup() anymore when dequeueing tasks. The use of the
regular task_wakeup() is sufficient there and will preserve a better
fairness: the test that went from 40k to 570k RPS now gives 580k RPS
(down from 585k RPS with previous commit). This essentially reverts
commit 27fab1dcb ("MEDIUM: queue: use tasklet_instant_wakeup() to
wake tasks").
This flag is no longer needed now that it must always match the presence
of a destination address on the backend conn_stream. Worse, before previous
patch, if it were to be accidently removed while the address is present, it
could result in a leak of that address since alloc_dst_address() would first
be called to flush it.
Its usage has a long history where addresses were stored in an area shared
with the connection, but as this is no longer the case, there's no reason
for putting this burden onto application-level code that should not focus
on setting obscure flags.
The only place where that made a small difference is in the dequeuing code
in case of queue redistribution, because previously the code would first
clear the flag, and only later when trying to deal with the queue, would
release the address. It's not even certain whether there would exist a
code path going to connect_server() without calling pendconn_dequeue()
first (e.g. retries on queue timeout maybe?).
Now the pendconn_dequeue() code will rely on SF_ASSIGNED to decide to
clear and release the address, since that flag is always set while in
a server's queue, and its clearance implies that we don't want to keep
the address. At least it remains consistent and there's no more risk of
leaking it.
It's long been known that queues didn't scale with threads for various
reasons ranging from the cost of the queue lock to the cost of the
massive amount of inter-thread wakeups.
But some recent reports showing deplorable perfs with threads used at
100% CPU helped us notice that the two elements above add on top of
each other:
- with plenty of inter-thread wakeups, the scheduler takes a lot of
time to dequeue pending tasks from the shared queue ;
- the lock held by the scheduler to do this slows down subsequent
task_wakeup() calls from the the queue that are made under the
queue's lock
- the queue's lock slows down addition of new requests to the queue
and adds up to the number of needed queue entries for a steady
traffic.
But the cost of the share queue has no reason for being paid because
it had already been paid when process_stream() added the request to
the queue. As such an instant wakeup is perfectly fit for this.
This is exactly what this patch does, it uses tasklet_instant_wakeup()
to dequeue pending requests, which has the effect of not bloating the
shared queue, hence not requiring the global queue lock, which in turn
results in the wakeup to be much faster, and the queue lock to be much
shorter. In the end, a test with 4k concurrent connections that was
being limited to 40-80k requests/s before with 16 threads, some of
which were stuck at 100% CPU now reaches 570k req/s with 4% idle.
Given that it's been found that it was possible to trigger the watchdog
on the queue lock under extreme conditions, and that such conditions
could happen when users want to protect their servers during a DoS, it
would definitely make sense to backport it to the most recent releases
(2.5 and 2.4 seem like good candidates especially because their scheduler
is modern enough to receive the change above). If a backport is performed,
the following patch is needed:
MINOR: task: add a new task_instant_wakeup() function
The source and destination addresses at the applicative layer are moved from
the stream-interface to the conn-stream. This simplifies a bit the code and
it is a logicial step to remove the stream-interface.
To be able to move the stream-interface from the stream to the conn-stream,
all access to the SI is done via the conn-stream. This patch is limited to
the backend part.
A subtle change of target address allocation was introduced with commit
68cf3959b ("MINOR: backend: rewrite alloc of stream target address") in
2.4. Prior to this patch, a target address was allocated by function
assign_server_address() only if none was previously allocated. After
the change, the allocation became unconditional. Most of the time it
makes no difference, except when we pass multiple times through
connect_server() with SF_ADDR_SET cleared.
The most obvious fix would be to avoid allocating that address there
when already set, but the root cause is that since introduction of
dynamically allocated addresses, the SF_ADDR_SET flag lies. It can
be cleared during redispatch or during a queue redistribution without
the address being released.
This patch instead gives back all its correct meaning to SF_ADDR_SET
and guarantees that when not set no address is allocated, by freeing
that address at the few places the flag is cleared. The flag could
even be removed so that only the address is checked but that would
require to touch many areas for no benefit.
The easiest way to test it is to send requests to a proxy with l7
retries enabled, which forwards to a server returning 500:
defaults
mode http
timeout client 1s
timeout server 1s
timeout connect 1s
retry-on all-retryable-errors
retries 1
option redispatch
listen proxy
bind *:5000
server app 0.0.0.0:5001
frontend dummy-app
bind :5001
http-request return status 500
Issuing "show pools" on the CLI will show that pool "sockaddr" grows
as requests are redispatched, and remains stable with the fix. Even
"ps" will show that the process' RSS grows by ~160B per request.
This fix will need to be backported to 2.4. Note that before 2.5,
there's no strm->si[1].dst, strm->target_addr must be used instead.
This addresses github issue #1499. Special thanks to Daniil Leontiev
for providing a well-documented reproducer.
The locking in the dequeuing process was significantly improved by commit
49667c14b ("MEDIUM: queue: take the proxy lock only during the px queue
accesses") in that it tries hard to limit the time during which the
proxy's queue lock is held to the strict minimum. Unfortunately it's not
enough anymore, because we take up the task and manipulate a few pendconn
elements after releasing the proxy's lock (while we're under the server's
lock) but the task will not necessarily hold the server lock since it may
not have successfully found one (e.g. timeout in the backend queue). As
such, stream_free() calling pendconn_free() may release the pendconn
immediately after the proxy's lock is released while the other thread
currently proceeding with the dequeuing tries to wake up the owner's
task and dies in task_wakeup().
One solution consists in releasing le proxy's lock later. But tests have
shown that we'd have to sacrifice a significant share of the performance
gained with the patch above (roughly a 20% loss).
This patch takes another approach. It adds a "del_lock" to each pendconn
struct, that allows to keep it referenced while the proxy's lock is being
released. It's mostly a serialization lock like a refcount, just to maintain
the pendconn alive till the task_wakeup() call is complete. This way we can
continue to release the proxy's lock early while keeping this one. It had
to be added to the few points where we're about to free a pendconn, namely
in pendconn_dequeue() and pendconn_unlink(). This way we continue to
release the proxy's lock very early and there is no performance degradation.
This lock may only be held under the queue's lock to prevent lock
inversion.
No backport is needed since the patch above was merged in 2.5-dev only.
A dedicated queue lock was added by commit 16fbdda3c ("MEDIUM: queue:
use a dedicated lock for the queues (v2)") but during its rebase, some
labels were lost and left to SERVER_LOCK / PROXY_LOCK instead of
QUEUE_LOCK. It's harmless but can confuse the lock debugger, so better
fix it.
No backport is needed.
Commit ae0b12ee0 ("MEDIUM: queue: use a trylock on the server's queue")
introduced a hard to trigger bug that's more visible with a single thread:
if a server dequeues a connection and finds another free slot with no
connection to place there, process_srv_queue() will never break out of
the loop. In multi-thread it almost does not happen because other threads
bring new connections.
No backport is needed as it's only in -dev.
Since the code paths became exactly the same except for what log field
to update, let's simplify the code and move further code out of the
lock. The queue position update and the test for server vs proxy do not
need to be inside the lock.
Now we directly use p->queue to get to the queue, which is much more
straightforward. The performance on 100 servers and 16 threads
increased from 560k to 574k RPS, or 2.5%.
A lot more simplifications are possible, but the minimum was done at
this point.
Doing so makes sure that threads attempting to wake up new connections
for a server will give up early if another thread is already in charge
of this. The goal is to avoid unneeded contention on low server counts.
Now with a single server with 16 threads in roundrobin we get the same
performance as with multiple servers, i.e. ~575kreq/s instead of ~496k
before. Leastconn is seeing a similar jump, from ~460 to ~560k (the
difference being the calls to fwlc_srv_reposition).
The overhead of process_srv_queue() is now around 2% instead of ~20%
previously.
There's no point keeping the proxy lock held for a long time, it's
only needed when checking the proxy's queue, and keeping it prevents
multiple servers from dequeuing in parallel. Let's move it into
pendconn_process_next_strm() and release it ASAP. The pendconn
remains under the server queue lock's protection, guaranteeing that
no stream will release it while it's being touched.
For roundrobin, the performance increases by 76% (327k to 575k) on
16 threads. Even with a single server and maxconn=100, the performance
increases from 398 to 496 kreq/s. For leastconn, almost no change is
visible (less than one percent) but this is expected since most of the
time there is spent in fwlc_reposition() and fwlc_get_next_server().
Doing so allows to retrieve and update the pendconn's queue index outside
of the queue's lock and to save one more percent CPU on a highly-contented
backend.
The code only differed by the nbpend_max counter. Let's have a pointer
to it and merge the two variants to always use a generic queue. It was
initially considered to put the max inside the queue structure itself,
but the stats support clearing values and maxes and this would have been
the only counter having to be handled separately there. Given that we
don't need this max anywhere outside stats, let's keep it where it is
and have a pointer to it instead.
The CAS loop to update the max remains. It was naively thought that it
would have been faster without atomic ops inside the lock, but this is
not the case for the simple reason that it is a max, it converges very
quickly and never has to perform the check anymore. Thus this code is
better out of the lock.
The queue_idx is still updated inside the lock since that's where the
idx is updated, though it could be performed using atomic ops given
that it's only used to roughly count places for logging.
This basically undoes the API changes that were performed by commit
0274286dd ("BUG/MAJOR: server: fix deadlock when changing maxconn via
agent-check") to address the deadlock issue: since process_srv_queue()
doesn't use the server lock anymore, it doesn't need the "server_locked"
argument, so let's get rid of it before it gets used again.
Till now whenever a server or proxy's queue was touched, this server
or proxy's lock was taken. Not only this requires distinct code paths,
but it also causes unnecessary contention with other uses of these locks.
This patch adds a lock inside the "queue" structure that will be used
the same way by the server and the proxy queuing code. The server used
to use a spinlock and the proxy an rwlock, though the queue only used
it for locked writes. This new version uses a spinlock since we don't
need the read lock part here. Tests have not shown any benefit nor cost
in using this one versus the rwlock so we could change later if needed.
The lower contention on the locks increases the performance from 362k
to 374k req/s on 16 threads with 20 servers and leastconn. The gain
with roundrobin even increases by 9%.
This is tagged medium because the lock is changed, but no other part of
the code touches the queues, with nor without locking, so this should
remain invisible.
There's no point doing atomic incs over px->served/px->totpend under the
locks from the inner loop, as this value is used by the LB algorithms but
not during the dequeuing step. In addition, the LB algo's take_conn()
doesn't need to be refreshed for each and every connection taken
under the lock, it can be performed once at the end and out of the
lock.
While the gain on roundrobin is not noticeable (only the atomic inc),
on leastconn which uses take_conn(), the performance increases from
355k to 362k req/s on 16 threads.
This reverts commit 5304669e1b.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
This reverts commit 3e92a31783.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
This reverts commit 1b648c857b.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
This reverts commit fcb8bf8650.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
This reverts commit c83e45e9b0.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
This reverts commit 3eecdb65c5.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
This reverts commit 1335eb9867.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
This reverts commit de814dd422.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
This reverts commit 9a6d0ddbd6.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
This reverts commit 5b39275311.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
This reverts commit 772e968b06.
The recent changes since 5304669e1 MEDIUM: queue: make
pendconn_process_next_strm() only return the pendconn opened a tiny race
condition between stream_free() and process_srv_queue(), as the pendconn
is accessed outside of the lock, possibly while it's being freed. A
different approach is required.
Dealing with the queue lock in the caller remains complicated. Let's
change pendconn_first() to take the queue instead of the tree head,
and handle the lock itself. It now returns an element with a locked
queue or no element with an unlocked queue. It can avoid locking if
the queue is already empty.
There's no point keeping the server's queue lock after seeing that the
server's queue is empty, just like there's no need to keep the proxy's
lock when its queue is empty. This patch checks for emptiness and
releases these locks as soon as possible.
With this the performance increased from 524k to 530k on 16 threads
with round-robin.
By placing the lock there, it becomes possible to lock the proxy
later and to unlock it earlier. The server unlocking also happens slightly
earlier.
The performance on roundrobin increases from 481k to 524k req/s on 16
threads. Leastconn shows about 513k req/s (the difference being the
take_conn() call).
The performance profile changes from this:
9.32% hap-pxok [.] process_srv_queue
7.56% hap-pxok [.] pendconn_dequeue
6.90% hap-pxok [.] pendconn_add
to this:
7.42% haproxy [.] process_srv_queue
5.61% haproxy [.] pendconn_dequeue
4.95% haproxy [.] pendconn_add
By doing so we can move some evaluations outside of the lock and the
loop. In the round robin case, the performance increases from 497k to
505k rps on 16 threads with 100 servers.
Doing so allows to retrieve and update the pendconn's queue index outside
of the queue's lock and to save one more percent CPU on a highly-contented
backend.
The code only differed by the nbpend_max counter. Let's have a pointer
to it and merge the two variants to always use a generic queue. It was
initially considered to put the max inside the queue structure itself,
but the stats support clearing values and maxes and this would have been
the only counter having to be handled separately there. Given that we
don't need this max anywhere outside stats, let's keep it where it is
and have a pointer to it instead.
The CAS loop to update the max remains. It was naively thought that it
would have been faster without atomic ops inside the lock, but this is
not the case for the simple reason that it is a max, it converges very
quickly and never has to perform the check anymore. Thus this code is
better out of the lock.
The queue_idx is still updated inside the lock since that's where the
idx is updated, though it could be performed using atomic ops given
that it's only used to roughly count places for logging.
This basically undoes the API changes that were performed by commit
0274286dd ("BUG/MAJOR: server: fix deadlock when changing maxconn via
agent-check") to address the deadlock issue: since process_srv_queue()
doesn't use the server lock anymore, it doesn't need the "server_locked"
argument, so let's get rid of it before it gets used again.
Till now whenever a server or proxy's queue was touched, this server
or proxy's lock was taken. Not only this requires distinct code paths,
but it also causes unnecessary contention with other uses of these locks.
This patch adds a lock inside the "queue" structure that will be used
the same way by the server and the proxy queuing code. The server used
to use a spinlock and the proxy an rwlock, though the queue only used
it for locked writes. This new version uses a spinlock since we don't
need the read lock part here. Tests have not shown any benefit nor cost
in using this one versus the rwlock so we could change later if needed.
The lower contention on the locks increases the performance from 491k
to 507k req/s on 16 threads with 20 servers and leastconn. The gain
with roundrobin even increases by 6%.
The performance profile changes from this:
13.03% haproxy [.] fwlc_srv_reposition
8.08% haproxy [.] fwlc_get_next_server
3.62% haproxy [.] process_srv_queue
1.78% haproxy [.] pendconn_dequeue
1.74% haproxy [.] pendconn_add
to this:
11.95% haproxy [.] fwlc_srv_reposition
7.57% haproxy [.] fwlc_get_next_server
3.51% haproxy [.] process_srv_queue
1.74% haproxy [.] pendconn_dequeue
1.70% haproxy [.] pendconn_add
At this point the differences are mostly measurement noise.
This is tagged medium because the lock is changed, but no other part of
the code touches the queues, with nor without locking, so this should
remain invisible.
This essentially reverts commit 2b4370078 ("MINOR: lb/api: let callers
of take_conn/drop_conn tell if they have the lock") that was merged
during 2.4 before the various locks could be eliminated at the lower
layers. Passing that information complicates the cleanup of the queuing
code and it's become useless.
The lock in process_srv_queue() was placed around the whole loop to
avoid the cost of taking/releasing it multiple times. But in practice
almost all calls to this function only dequeue a single connection, so
that argument doesn't really stand. However by placing the lock inside
the loop, we'd make it possible to release it before manipulating the
pendconn and waking the task up. That's what this patch does.
This increases the performance from 431k to 491k req/s on 16 threads
with 20 servers under leastconn.
The performance profile changes from this:
14.09% haproxy [.] process_srv_queue
10.22% haproxy [.] fwlc_srv_reposition
6.39% haproxy [.] fwlc_get_next_server
3.97% haproxy [.] pendconn_dequeue
3.84% haproxy [.] pendconn_add
to this:
13.03% haproxy [.] fwlc_srv_reposition
8.08% haproxy [.] fwlc_get_next_server
3.62% haproxy [.] process_srv_queue
1.78% haproxy [.] pendconn_dequeue
1.74% haproxy [.] pendconn_add
The difference is even slightly more visible in roundrobin which
does not have take_conn() call.
