An SSL connection takes some memory when it exists and during handshakes.
We measured up to 16kB for an established endpoint, and up to 76 extra kB
during a handshake. The SSL layer stores these values into the global
struct during initialization. If other SSL libs are used, it's easy to
change these values. Anyway they'll only be used as gross estimates in
order to guess the max number of SSL conns that can be established when
memory is constrained and the limit is not set.
We'll need to know the number of SSL connections, their use and their
cost soon. In order to avoid getting tons of ifdefs everywhere, always
export SSL information in the global section. We add two flags to know
whether or not SSL is used in a frontend and in a backend.
This setting is used to limit memory usage without causing the alloc
failures caused by "-m". Unexpectedly, tests have shown a performance
boost of up to about 18% on HTTP traffic when limiting the number of
buffers to about 10% of the amount of concurrent connections.
tune.buffers.limit <number>
Sets a hard limit on the number of buffers which may be allocated per process.
The default value is zero which means unlimited. The minimum non-zero value
will always be greater than "tune.buffers.reserve" and should ideally always
be about twice as large. Forcing this value can be particularly useful to
limit the amount of memory a process may take, while retaining a sane
behaviour. When this limit is reached, sessions which need a buffer wait for
another one to be released by another session. Since buffers are dynamically
allocated and released, the waiting time is very short and not perceptible
provided that limits remain reasonable. In fact sometimes reducing the limit
may even increase performance by increasing the CPU cache's efficiency. Tests
have shown good results on average HTTP traffic with a limit to 1/10 of the
expected global maxconn setting, which also significantly reduces memory
usage. The memory savings come from the fact that a number of connections
will not allocate 2*tune.bufsize. It is best not to touch this value unless
advised to do so by an haproxy core developer.
We've already experimented with three wake up algorithms when releasing
buffers : the first naive one used to wake up far too many sessions,
causing many of them not to get any buffer. The second approach which
was still in use prior to this patch consisted in waking up either 1
or 2 sessions depending on the number of FDs we had released. And this
was still inaccurate. The third one tried to cover the accuracy issues
of the second and took into consideration the number of FDs the sessions
would be willing to use, but most of the time we ended up waking up too
many of them for nothing, or deadlocking by lack of buffers.
This patch completely removes the need to allocate two buffers at once.
Instead it splits allocations into critical and non-critical ones and
implements a reserve in the pool for this. The deadlock situation happens
when all buffers are be allocated for requests pending in a maxconn-limited
server queue, because then there's no more way to allocate buffers for
responses, and these responses are critical to release the servers's
connection in order to release the pending requests. In fact maxconn on
a server creates a dependence between sessions and particularly between
oldest session's responses and latest session's requests. Thus, it is
mandatory to get a free buffer for a response in order to release a
server connection which will permit to release a request buffer.
Since we definitely have non-symmetrical buffers, we need to implement
this logic in the buffer allocation mechanism. What this commit does is
implement a reserve of buffers which can only be allocated for responses
and that will never be allocated for requests. This is made possible by
the requester indicating how much margin it wants to leave after the
allocation succeeds. Thus it is a cooperative allocation mechanism : the
requester (process_session() in general) prefers not to get a buffer in
order to respect other's need for response buffers. The session management
code always knows if a buffer will be used for requests or responses, so
that is not difficult :
- either there's an applet on the initiator side and we really need
the request buffer (since currently the applet is called in the
context of the session)
- or we have a connection and we really need the response buffer (in
order to support building and sending an error message back)
This reserve ensures that we don't take all allocatable buffers for
requests waiting in a queue. The downside is that all the extra buffers
are really allocated to ensure they can be allocated. But with small
values it is not an issue.
With this change, we don't observe any more deadlocks even when running
with maxconn 1 on a server under severely constrained memory conditions.
The code becomes a bit tricky, it relies on the scheduler's run queue to
estimate how many sessions are already expected to run so that it doesn't
wake up everyone with too few resources. A better solution would probably
consist in having two queues, one for urgent requests and one for normal
requests. A failed allocation for a session dealing with an error, a
connection event, or the need for a response (or request when there's an
applet on the left) would go to the urgent request queue, while other
requests would go to the other queue. Urgent requests would be served
from 1 entry in the pool, while the regular ones would be served only
according to the reserve. Despite not yet having this, it works
remarkably well.
This mechanism is quite efficient, we don't perform too many wake up calls
anymore. For 1 million sessions elapsed during massive memory contention,
we observe about 4.5M calls to process_session() compared to 4.0M without
memory constraints. Previously we used to observe up to 16M calls, which
rougly means 12M failures.
During a test run under high memory constraints (limit enforced to 27 MB
instead of the 58 MB normally needed), performance used to drop by 53% prior
to this patch. Now with this patch instead it *increases* by about 1.5%.
The best effect of this change is that by limiting the memory usage to about
2/3 to 3/4 of what is needed by default, it's possible to increase performance
by up to about 18% mainly due to the fact that pools are reused more often
and remain hot in the CPU cache (observed on regular HTTP traffic with 20k
objects, buffers.limit = maxconn/10, buffers.reserve = limit/2).
Below is an example of scenario which used to cause a deadlock previously :
- connection is received
- two buffers are allocated in process_session() then released
- one is allocated when receiving an HTTP request
- the second buffer is allocated then released in process_session()
for request parsing then connection establishment.
- poll() says we can send, so the request buffer is sent and released
- process session gets notified that the connection is now established
and allocates two buffers then releases them
- all other sessions do the same till one cannot get the request buffer
without hitting the margin
- and now the server responds. stream_interface allocates the response
buffer and manages to get it since it's higher priority being for a
response.
- but process_session() cannot allocate the request buffer anymore
=> We could end up with all buffers used by responses so that none may
be allocated for a request in process_session().
When the applet processing leaves the session context, the test will have
to be changed so that we always allocate a response buffer regardless of
the left side (eg: H2->H1 gateway). A final improvement would consists in
being able to only retry the failed I/O operation without waking up a
task, but to date all experiments to achieve this have proven not to be
reliable enough.
Adds global statements 'ssl-default-server-options' and
'ssl-default-bind-options' to force on 'server' and 'bind' lines
some ssl options.
Currently available options are 'no-sslv3', 'no-tlsv10', 'no-tlsv11',
'no-tlsv12', 'force-sslv3', 'force-tlsv10', 'force-tlsv11',
'force-tlsv12', and 'no-tls-tickets'.
Example:
global
ssl-default-server-options no-sslv3
ssl-default-bind-options no-sslv3
With all the goodies supported by logformat, people find that the limit
of 1024 chars for log lines is too short. Some servers do not support
larger lines and can simply drop them, so changing the default value is
not always the best choice.
This patch takes a different approach. Log line length is specified per
log server on the "log" line, with a value between 80 and 65535. That
way it's possibly to satisfy all needs, even with some fat local servers
and small remote ones.
When no static DH parameters are specified, this patch makes haproxy
use standardized (rfc 2409 / rfc 3526) DH parameters with prime lenghts
of 1024, 2048, 4096 or 8192 bits for DHE key exchange. The size of the
temporary/ephemeral DH key is computed as the minimum of the RSA/DSA server
key size and the value of a new option named tune.ssl.default-dh-param.
One important aspect of SSL performance tuning is the cache size,
but there's no metric to know whether it's large enough or not. This
commit introduces two counters, one for the cache lookups and another
one for cache misses. These counters are reported on "show info" on
the stats socket. This way, it suffices to see the cache misses
counter constantly grow to know that a larger cache could possibly
help.
It's commonly needed to know how many SSL asymmetric keys are computed
per second on either side (frontend or backend), and to know the SSL
session reuse ratio. Now we compute these values and report them in
"show info".
Some consistency checks cannot be performed between frontends, backends
and peers at the moment because there is no way to check for intersection
between processes bound to some processes when the number of processes is
higher than the number of bits in a word.
So first, let's limit the number of processes to the machine's word size.
This means nbproc will be limited to 32 on 32-bit machines and 64 on 64-bit
machines. This is far more than enough considering that configs rarely go
above 16 processes due to scalability and management issues, so 32 or 64
should be fine.
This way we'll ensure we can always build a mask of all the processes a
section is bound to.
When compiled with USE_GETADDRINFO, make sure we use getaddrinfo(3) to
perform name lookups. On default dual-stack setups this will change the
behavior of using IPv6 first. Global configuration option
'nogetaddrinfo' can be used to revert to deprecated gethostbyname(3).
The new tune.idletimer value allows one to set a different value for
idle stream detection. The default value remains set to one second.
It is possible to disable it using zero, and to change the default
value at build time using DEFAULT_IDLE_TIMER.
If no CA file specified on a server line, the config parser will show an error.
Adds an cmdline option '-dV' to re-set verify 'none' as global default on
servers side (previous behavior).
Also adds 'ssl-server-verify' global statement to set global default to
'none' or 'required'.
WARNING: this changes the default verify mode from "none" to "required" on
the server side, and it *will* break insecure setups.
Just like the previous commit, we sometimes want to limit the rate of
incoming SSL connections. While it can be done for a frontend, it was
not possible for a whole process, which makes sense when multiple
processes are running on a system to server multiple customers.
The new global "maxsslrate" setting is usable to fix a limit on the
session rate going to the SSL frontends. The limits applies before
the SSL handshake and not after, so that it saves the SSL stack from
expensive key computations that would finally be aborted before being
accounted for.
The same setting may be changed at run time on the CLI using
"set rate-limit ssl-session global".
It's sometimes useful to be able to limit the connection rate on a machine
running many haproxy instances (eg: per customer) but it removes the ability
for that machine to defend itself against a DoS. Thus, better also provide a
limit on the session rate, which does not include the connections rejected by
"tcp-request connection" rules. This permits to have much higher limits on
the connection rate without having to raise the session rate limit to insane
values.
The limit can be changed on the CLI using "set rate-limit sessions global",
or in the global section using "maxsessrate".
Patrick Hemmer reported that using unique_id_format and logs did not
report the same unique ID counter since commit 9f09521 ("BUG/MEDIUM:
unique_id: HTTP request counter must be unique!"). This is because
the increment was done while producing the log message, so it was
performed twice.
A better solution consists in fetching a new value once per request
and saving it in the request or session context for all of this
request's life.
It happens that sessions already have a unique ID field which is used
for debugging and reporting errors, and which differs from the one
sent in logs and unique_id header.
So let's change this to reuse this field to have coherent IDs everywhere.
As of now, a session gets a new unique ID once it is instanciated. This
means that TCP sessions will also benefit from a unique ID that can be
logged. And this ID is renewed for each extra HTTP request received on
an existing session. Thus, all TCP sessions and HTTP requests will have
distinct IDs that will be stable along all their life, and coherent
between all places where they're used (logs, unique_id header,
"show sess", "show errors").
This feature is 1.5-specific, no backport to 1.4 is needed.
The HTTP request counter is incremented non atomically, which means that
many requests can log the same ID. Let's increment it when it is consumed
so that we avoid this case.
This bug was reported by Patrick Hemmer. It's 1.5-specific and does not
need to be backported.
Add new tunable "tune.ssl.maxrecord".
Over SSL/TLS, the client can decipher the data only once it has received
a full record. With large records, it means that clients might have to
download up to 16kB of data before starting to process them. Limiting the
record size can improve page load times on browsers located over high
latency or low bandwidth networks. It is suggested to find optimal values
which fit into 1 or 2 TCP segments (generally 1448 bytes over Ethernet
with TCP timestamps enabled, or 1460 when timestamps are disabled), keeping
in mind that SSL/TLS add some overhead. Typical values of 1419 and 2859
gave good results during tests. Use "strace -e trace=write" to find the
best value.
This trick was first suggested by Mike Belshe :
http://www.belshe.com/2010/12/17/performance-and-the-tls-record-size/
Then requested again by Ilya Grigorik who provides some hints here :
http://ofps.oreilly.com/titles/9781449344764/_transport_layer_security_tls.html#ch04_00000101
This patch adds a new option "-Ds" which is exactly like "-D", but instead of
forking n times to get n jobs running and then exiting, prefers to wait for all the
children it just created. With this done, haproxy becomes more systemd-compliant,
without changing anything for other systems.
Signed-off-by: Marc-Antoine Perennou <Marc-Antoine@Perennou.com>
Commit 290e63aa moved the unix parameters out of the global stats socket
to the bind_conf struct. As such the stats admin level was also moved
overthere, but it remained in the stats global section where it was not
used, except by a nasty memcpy() used to initialize the ux struct in the
bind_conf with too large data. Fortunately, the extra data copied were
the previous level over the new level so it did not have any impact, but
it could have been worse.
This bug is 1.5 specific, no backport is needed.
Reported-by: Dinko Korunic <dkorunic@reflected.net>
Some users need more than 64 characters to log large cookies. The limit
was set to 63 characters (and not 64 as previously documented). Now it
is possible to change this using the global "tune.http.cookielen" setting
if required.
The new "cpu-map" directive allows one to assign the CPU sets that
a process is allowed to bind to. This is useful in combination with
the "nbproc" and "bind-process" directives.
The support is implicit on Linux 2.6.28 and above.
Now that all pollers make use of speculative I/O, there is no point
having two epoll implementations, so replace epoll with the sepoll code
and remove sepoll which has just become the standard epoll method.
This patch adds input and output rate calcutation on the HTTP compresion
feature.
Compression can be limited with a maximum rate value in kilobytes per
second. The rate is set with the global 'maxcomprate' option. You can
change this value dynamicaly with 'set rate-limit http-compression
global' on the UNIX socket.
With the global maxzlibmem option, you are able ton control the maximum
amount of RAM usable for HTTP compression.
A test is done before each zlib allocation, if the there isn't available
memory, the test fail and so the zlib initialization, so data won't be
compressed.
The window size and the memlevel of the zlib are now configurable using
global options tune.zlib.memlevel and tune.zlib.windowsize.
It affects the memory consumption of the zlib.
The trash is used everywhere to store the results of temporary strings
built out of s(n)printf, or as a storage for a chunk when chunks are
needed.
Using global.tune.bufsize is not the most convenient thing either.
So let's replace trash with a chunk and directly use it as such. We can
then use trash.size as the natural way to get its size, and get rid of
many intermediary chunks that were previously used.
The patch is huge because it touches many areas but it makes the code
a lot more clear and even outlines places where trash was used without
being that obvious.
These ones are used to set the default ciphers suite on "bind" lines and
"server" lines respectively, instead of using OpenSSL's defaults. These
are probably mainly useful for distro packagers.
Now the stats socket is allocated when the 'stats socket' line is parsed,
and assigned using the standard str2listener(). This has two effects :
- more than one stats socket can now be declared
- stats socket now support protocols other than UNIX
The next step is to remove the duplicate bind config parsing.
SSL connections take a huge amount of memory, and unfortunately openssl
does not check malloc() returns and easily segfaults when too many
connections are used.
The only solution against this is to provide a global maxsslconn setting
to reject SSL connections above the limit in order to avoid reaching
unsafe limits.
Before it was possible to resize the buffers using global.tune.bufsize,
the trash has always been the size of a buffer by design. Unfortunately,
the recent buffer sizing at runtime forgot to adjust the trash, resulting
in it being too short for content rewriting if buffers were enlarged from
the default value.
The bug was encountered in 1.4 so the fix must be backported there.
%Fi: Frontend IP
%Fp: Frontend Port
%Si: Server IP
%Sp: Server Port
%Ts: Timestamp
%rt: HTTP request counter
%H: hostname
%pid: PID
+X: Hexadecimal represenation
The +X mode in logformat displays hexadecimal for the following flags
%Ci %Cp %Fi %Fp %Bi %Bp %Si %Sp %Ts %ct %pid
rename logformat_write_string() to lf_text()
Optimize size computation
This patch settles the 2 loggers limitation.
Loggers are now stored in linked lists.
Using "global log", the global loggers list content is added at the end
of the current proxy list. Each "log" entries are added at the end of
the proxy list.
"no log" flush a logger list.
