"[MINOR] session: add a pointer to the new target into the session" (664beb8)
introduced a regression by changing the type of a peer's target from
TARG_TYPE_PROXY to TARG_TYPE_NONE. The effect of this is that during
a soft-restart the new process no longer tries to connect to the
old process to replicate its stick tables.
This patch sets the type of a peer's target as TARG_TYPE_PROXY and
replication on soft-restart works once again.
It is now possible to filter by termination code with -tcn <termcode>, to be
able to track one kind of errors, for example after counting it with -tc.
Use -TCN <termcode> gives you the opposite.
Adding health checks has become a real pain, with cross-references to all
checks everywhere because they're all a single bit. Since they're all
exclusive, let's change this to have a check number only. We reserve 4
bits allowing up to 16 checks (15+tcp), only 7 of which are currently
used. The code has shrunk by almost 1kB and we saved a few option bits.
The "dispatch" option has been moved to px->options, making a few tests
a bit cleaner.
This patch provides a new "option redis-check" statement to enable server health checks based on redis PING request (http://www.redis.io/commands/ping).
The new "set maxconn frontend XXX" statement on the stats socket allows
the admin to change a frontend's maxconn value. If some connections are
queued, they will immediately be accepted up to the new limit. If the
limit is lowered, new connections acceptation might be delayed. This can
be used to temporarily reduce or increase the impact of a specific frontend's
traffic on the whole process.
This global task is used to periodically check for end of resource shortage
and to try to enable queued listeners again. This is important in case some
temporary system-wide shortage is encountered, so that we don't have to wait
for an existing connection to be released before checking the queue again.
For situations where listeners are queued due to the global maxconn being
reached, the task is woken up at least every second. For situations where
a system resource shortage is detected (memory, sockets, ...) the task is
woken up at least every 100 ms. That way, recovery from severe events can
still be achieved under acceptable conditions.
This was revealed with one of the very latest patches which caused
the listener_queue not to be initialized on the stats socket frontend.
And in fact a number of other ones were missing too. This is getting so
boring that now we'll always make use of the same function to initialize
any proxy. Doing so has even saved about 500 bytes on the binary due to
the avoided code redundancy.
No backport is needed.
This function is finally not needed anymore, as it has been replaced with
a per-proxy task that is scheduled when some limits are encountered on
incoming connections or when the process is stopping. The savings should
be noticeable on configs with a large number of proxies. The most important
point is that the rate limiting is now enforced in a clean and solid way.
Peers were stopped on every call to maintain_proxies when stopping=1,
while they should only be stopped once upon call to soft_stop(). This
bug has little impact, mostly increased CPU usage. It's not needed to
backport it.
Instead of waking a listener up then making it sleep, we only wake them up
if we know their rate limit is fine. In the future we could improve on top
of that by deciding to wake a proxy-specific task in XX milliseconds to
take care of enabling the listeners again.
Patch d9bbe17b used to limit the rate-limit to off-by-one to avoid
a busy loop when the limit is reached. Now that the listeners are
automatically disabled and queued when a limit is reached, we don't
need this workaround anymore and can bring back the most accurate
computation.
Those states have been replaced with PR_STFULL and PR_STREADY respectively,
as it is what matches them the best now. Also, two occurrences of PR_STIDLE
in peers.c have been removed as this did not provide any form of error recovery
anyway.
Now maintain_proxies() only changes proxies states and does not affect their
listeners anymore since they are autonomous. A proxy will switch between the
PR_STIDLE and PR_STRUN states depending whether it's saturated or not. Next
step will consist in renaming PR_STIDLE to PR_STFULL. This state is now only
used to report the proxy state in the stats.
All listeners that are limited by a proxy-specific resource are now
queued at the proxy's and not globally. This allows finer-grained
wakeups when releasing resource.
When an accept() fails because of a connection limit or a memory shortage,
we now disable it and queue it so that it's dequeued only when a connection
is released. This has improved the behaviour of the process near the fd limit
as now a listener with a no connection (eg: stats) will not loop forever
trying to get its connection accepted.
The solution is still not 100% perfect, as we'd like to have this used when
proxy limits are reached (use a per-proxy list) and for safety, we'd need
to have dedicated tasks to periodically re-enable them (eg: to overcome
temporary system-wide resource limitations when no connection is released).
When a listeners encounters a resource shortage, it currently stops until
one re-enables it. This is far from being perfect as it does not yet handle
the case where the single connection from the listener is rejected (eg: the
stats page).
Now we'll have a special status for resource limited listeners and we'll
queue them into one or multiple lists. That way, each time we have to stop
a listener because of a resource shortage, we can enqueue it and change its
state, so that it is dequeued once more resources are available.
This patch currently does not change any existing behaviour, it only adds
the basic building blocks for doing that.
For listeners that are not bound to a frontend, the limit on the
number of accepted connections is tested at the end of the accept()
loop, but we don't break out of the loop, meaning that if more
connections than what the listener allows are available and if this
is less than the proxy's limits and within the size of a batch, then
they could be accepted. In practice, this problem currently cannot
appear since all listeners are bound to a frontend, and it's a very
minor issue anyway.
1.4 has the same issue (which cannot happen there either), but there
is some code after it, so it's the code cleanup which revealed it.
Managing listeners state is difficult because they have their own state
and can at the same time have theirs dictated by their proxy. The pause
is not done properly, as the proxy code is fiddling with sockets. By
introducing new functions such as pause_listener()/resume_listener(), we
make it a bit more obvious how/when they're supposed to be used. The
listen_proxies() function was also renamed to resume_proxies() since
it's only used for pause/resume.
This patch is the first in a series aiming at getting rid of the maintain_proxies
mess. In the end, proxies should not call enable_listener()/disable_listener()
anymore.
When an accept() returns -1 ENFILE du to system limits, it leaves the
connection pending in the backlog and epoll() comes back immediately
afterwards trying to make it accept it again. This causes haproxy to
remain at 100% CPU until something makes an accept() possible again.
Now upon such resource shortage, we mark the listener FULL so that we
only enable it again once at least one connection has been released.
In fact we only do that if there are some active connections on this
proxy, so that it has a chance to be marked not full again. This makes
haproxy remain idle when all resources are used, which helps a lot
releasing those resource as fast as possible.
Backport to 1.4 might be desirable but difficult and tricky.
By default on a single process, we accept 100 connections at once. This is too
much on recent CPUs where the cache is constantly thrashing, because we visit
all those connections several times. We should batch the processing slightly
less so that all the accepted session may remain in cache during their initial
processing.
Lowering the batch size from 100 to 32 has changed the connection rate for
concurrencies between 5-10k from 67 kcps to 94 kcps on a Core i5 660 (4M L3),
and forward rates from 30k to 39.5k.
Tests on this hardware show that values between 10 and 30 seem to do the job fine.
When we fail to create a session because of memory shortage, let's at
least try to send a 500 message directly on the socket. Even if we don't
have any buffers left, the kernel's orphans management will take care of
delivering the message as long as there are socket buffers left.
Patch af5149 introduced an issue which can be detected only on out of
memory conditions : a LIST_DEL() may be performed on an uninitialized
struct member instead of a LIST_INIT() during the accept() phase,
causing crashes and memory corruption to occur.
This issue was detected and diagnosed by the Exceliance R&D team.
This is 1.5-specific and very recent, so no existing deployment should
be impacted.
The motivation for this is that when soft-restart is merged
it will be come more important to free all relevant memory in deinit()
Discovered using valgrind.
The motivation for this is that when soft-restart is merged
it will be come more important to free all relevant memory in deinit()
Discovered using valgrind.
The motivation for this is that when soft-restart is merged
it will be come more important to free all relevant memory in deinit()
Discovered using valgrind.
The motivation for this is that when soft-restart is merged
it will be come more important to free all relevant memory in deinit()
Discovered using valgrind.
There were too many filters, we were losing time in all the "if" statements.
By moving all the filters to independant functions, we made the code cleaner
and slightly faster (3%).
One minor bug was found, the -tc and -st options did not report the number
of output lines, but always zero.
Almost all filters first check the line format, which takes a lot of code
and requires parsing back and forth. By centralizing this test, we can
save about 15-20 more percent of performance for all filters.
Also, the test was wrong, it was checking that the source IP address was
starting with a digit, which is not always true with local IPv6 addresses.
Instead, we now check that the next field (accept field) starts with an
opening bracket and is followed by a digit between 0 and 3 (day of the
month). Doing this has contributed a 2% speedup because all other field
calculations were relative to a closer field.
Since many fields are relative and some are used a lot, try to cache them
the first time they're used in order to avoid skipping them twice. The
status counts with HTTP pre-check enabled has sped up by 40%.
The SKIP_CHAR fix caused a measurable performance drop. Since we can
consider all chars below 0x20 as delimiters, we can avoid a cache lookup
which requires a char to pointer conversion.
This is used to perform cookie-based stickiness with table replication
between multiple masters and across restarts. This partially overrides
some of the appsession capabilities.
Never add connections allocated to this sever to a stick-table.
This may be used in conjunction with backup to ensure that
stick-table persistence is disabled for backup servers.
