This optimisation causes haproxy to time out requests that result
in two TCP packets, one packet containing the header, and one
packet containing the actual data. This is a very typical type
of response from a lot of servers.
[Willy: I suspect the fix might have an impact on the compression code
which I'm not sure completely handles calls with 0 bytes to forward]
CF_READ_DONTWAIT was designed to avoid getting an EAGAIN upon recv() when
very few data are expected. It prevents the reader from looping over
recv(). Unfortunately with speculative I/O, it is very common that the
same event has the time to be called twice before the task handles the
data and disables the recv(). This is because not all tasks are always
processed at once.
Instead of leaving the buffer free-wheeling and doing an EAGAIN, we
disable reading as soon as the first recv() succeeds. This way we're
sure that only the next wakeup of the task will re-enable it if needed.
Doing so has totally removed the EAGAIN we were seeing till now (30% of
recv).
At the moment sepoll is not 100% event-driven, because a call to fd_set()
on an event which is already being polled will not change its state.
This causes issues with OpenSSL because if some I/O processing is interrupted
after clearing the I/O event (eg: read all data from a socket, can't put it
all into the buffer), then there is no way to call the SSL_read() again once
the buffer releases some space.
The only real solution is to go 100% event-driven. The principle is to use
the spec list as an event cache and that each time an I/O event is reported
by epoll_wait(), this event is automatically scheduled for addition to the
spec list for future calls until the consumer explicitly asks for polling
or stopping.
Doing this is a bit tricky because sepoll used to provide a substantial
number of optimizations such as event merging. These optimizations have
been maintained : a dedicated update list is affected when events change,
but not the event list, so that updates may cancel themselves without any
side effect such as displacing events. A specific case was considered for
handling newly created FDs as soon as they are detected from within the
poll loop. This ensures that their read or write operation will always be
attempted as soon as possible, thus reducing the number of poll loops and
process_session wakeups. This is especially true for newly accepted fds
which immediately perform their first recv() call.
Two new flags were added to the fdtab[] struct to tag the fact that a file
descriptor already exists in the update list. One flag indicates that a
file descriptor is new and has just been created (fdtab[].new) and the other
one indicates that a file descriptor is already referenced by the update list
(fdtab[].updated). Even if the FD state changes during operations or if the
fd is closed and replaced, it's not an issue because the update flag remains
and is easily spotted during list walks. The flag must absolutely reflect the
presence of the fd in the update list in order to avoid overflowing the update
list with more events than there are distinct fds.
Note that this change also recovers the small performance loss introduced
by its connection counter-part and goes even beyond.
The CO_FL_WAIT_* flags were not cleared after updating polling flags.
This means that any caller of these functions that did not clear it
would enable polling instead of speculative I/O. This happens during
the stream interface update call which is performed from the session
handler for example.
As of now it's not a problem yet because speculative I/O and polling
are handled the same way. However with upcoming changes it does cause
some deadlocks because enabling read processing on a file descriptor
where everything was already read will do nothing until something new
happens on this FD.
The correct fix consists in clearing the flags while leaving the update
functions.
This fix does not need any backport as it was introduced with recent
connection changes (dev12) and not triggered until last commit.
This is the first step of a series of changes aiming at making the
polling totally event-driven. This first change consists in only
remembering at the connection level whether an FD was enabled or not,
regardless of the fact it was being polled or cached. From now on, an
EAGAIN will always be considered as a change so that the pollers are
able to manage a cache and to flush it based on such events. One of
the noticeable effect is that conn_fd_handler() is called once more
per session (6 instead of 5 min) but other update functions are less
called.
Note that the performance loss caused by this change at the moment is
quite significant, around 2.5%, but the change is needed to have SSL
working correctly in all situations, even when data were read from the
socket and stored in the invisible cache, waiting for some room in the
channel's buffer.
There is a small waste of CPU cycles when no handshake is required on an
accepted connection, because we had to perform one call to conn_fd_handler()
to mark the connection CONNECTED and to call process_session() again to say
that nothing happened.
By marking the connection CONNECTED when there is no pending handshake, we
avoid this extra call to process_session().
Having a global expiration timer for a task means that the tasks are regularly
woken up (at least after each expiration timer). It's totally useless and counter
productive to process the whole session upon each such wakeup, and it's fairly
easy to detect such wakeups, so let's just update the task's timer and return
to sleep when this happens.
For 100k concurrent connections with 10s of timeouts, this can save 10k wakeups
per second, which is not bad.
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.
Don't use the zlib allocator anymore, 5 pools are used for the zlib
compression. Their sizes depends of the window size and the memLevel in
deflateInit2.
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 build was dependent of the zlib.h header, regardless of the USE_ZLIB
option. The fix consists of several #ifdef in the source code.
It removes the overhead of the zstream structure in the session when you
don't use the option.
With extra-large buffers, it is possible that a lot of data are sent upon
connection establishment before the session is notified. The issue is how
to handle a send() error after some data were actually sent.
At the moment, only a connection error is reported, causing a new connection
attempt and send() to restart after the last data. We absolutely don't want
to retry the connect() if at least one byte was sent, because those data are
lost.
The solution consists in reporting exactly what happens, which is :
- a successful connection attempt
- a read/write error on the channel
That way we go on with sess_establish(), the response analysers are called
and report the appropriate connection state for the error (typically a server
abort while waiting for a response). This mechanism also guarantees that we
won't retry since it's a success. The logs also report the correct connect
time.
Note that 1.4 is not directly affected because it only attempts one send(),
so it cannot detect a send() failure here and distinguish it form a failed
connection attempt. So no backport is needed. Also, this is just a safe belt
we're taking, since this issue should not happen anymore since previous commit.
It is stupid to loop over ->snd_buf() because the snd_buf() itself already
loops and stops when system buffers are full. But looping again onto it,
we lose the information of the full buffers and perform one useless syscall.
Furthermore, this causes issues when dealing with large uploads while waiting
for a connection to establish, as it can report a server reject of some data
as a connection abort, which is wrong.
1.4 does not have this issue as it loops maximum twice (once for each buffer
half) and exists as soon as system buffers are full. So no backport is needed.
Some old browsers that have a user-agent starting with "Mozilla/4" do
not support compressison correctly, so disable compression for those.
Internet explorer 6 after Windows XP service pack 2, IE 7, and IE 8,
do however support compression and still have a user agent starting
with Mozilla/4, so we try to enable compression for those.
MSIE has a user-agent on this form:
Mozilla/4.0 (compatible; MSIE <version>; ...)
98% of MSIE 6 SP2 user agents start with
Mozilla/4.0 (compatible; MSIE 6.0; Windows NT 5.1; SV1
The remaining 2% have additional flags before "SV1".
This simplified matching looking for MSIE at exactly position 25
and SV1 at exacly position 51 gives a few false negatives, so sometimes
a compression opportunity is lost.
A test against 3 hours of traffic to around 3000 news sites worldwide
gives less than 0.007% (70ppm) missed compression opportunities.
Keys are copied from samples to stick_table_key. If a key is larger
than the stick_table_key, we have an overflow. In pratice it does not
happen because it requires :
1) a configuration with tune.bufsize larger than BUFSIZE (common)
2) a stick-table configured with keys strictly larger than buffers
3) extraction of data larger than BUFSIZE (eg: using payload())
Points 2 and 3 don't make any sense for a real world configuration. That
said the issue needs be fixed. The solution consists in allocating it the
same size as the global buffer size, just like the samples. This fixes the
issue.
Sample conversions rely on two alternative buffers which were previously
allocated as static bufs of size BUFSIZE. Now they're initialized to the
global buffer size. It was the same for HTTP authentication. Note that it
seems that none of them was prone to any mistake when dealing with the
buffer size, but better stay on the safe side by maintaining the old
assumption that a trash buffer is always "large enough".
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.
This function's naming was misleading as it is used to append data
at the end of a string, causing some surprizes when used for the
first time!
Add a chunk_printf() function which does what its name suggests.
This is a first step in avoiding to constantly reinitialize chunks.
It replaces the old chunk_reset() which was not properly named as it
used to drop everything and was only used by chunk_destroy(). It has
been renamed chunk_drop().
We don't want the lower layer to forward a close while we're compressing,
and we want the system to fuse outgoing TCP segments using MSG_MORE as
much as possible to save round trips that can emerge from sending short
packets with a PUSH flag.
A test on a remote busy DSL line consisting in compressing a 100MB file
on the fly full of zeroes only showed a transfer rate of a few kB/s due
to these round trips.
We will need to be able to switch server connections on a session and
to keep idle connections. In order to achieve this, the preliminary
requirement is that the connections can survive the session and be
detached from them.
Right now they're still allocated at exactly the same place, so when
there is a session, there are always 2 connections. We could soon
improve on this by allocating the outgoing connection only during a
connect().
This current patch touches a lot of code and intentionally does not
change any functionnality. Performance tests show no regression (even
a very minor improvement). The doc has not yet been updated.
Thomas Heil reported that health checks did not work anymore when a backend
or server has "usesrc clientip". This is because the source address is not
set and tcp_bind_socket() tries to bind to that address anyway.
The solution consists in explicitly clearing the source address in the checks
and to make tcp_bind_socket() avoid binding when the address is not set. This
also has an indirect benefit that a useless bind() syscall will be avoided
when using "source 0.0.0.0 usesrc clientip" in health checks.
From the beginning it has been said that -D must always be used on the
command line from startup scripts so that haproxy does not accidentally
stay in foreground when loaded from init script... Except that this has
not been true for a long time now.
The fix is easy and must be backported to 1.4 too which is affected.
ssl_c_notbefore: start date of client cert (string, eg: "121022182230Z" for YYMMDDhhmmss[Z])
ssl_c_notafter: end date of client cert (string, eg: "121022182230Z" for YYMMDDhhmmss[Z])
ssl_f_notbefore: start date of frontend cert (string, eg: "121022182230Z" for YYMMDDhhmmss[Z])
ssl_f_notafter: end date of frontend cert (string, eg: "121022182230Z" for YYMMDDhhmmss[Z])
ssl_c_key_alg: algo used to encrypt the client's cert key (ex: rsaEncryption)
ssl_f_key_alg: algo used to encrypt the frontend's cert key (ex: rsaEncryption)
ssl_c_s_dn : client cert subject DN (string)
ssl_c_i_dn : client cert issuer DN (string)
ssl_f_s_dn : frontend cert subject DN (string)
ssl_f_i_dn : frontend cert issuer DN (string)
Return either the full DN without params, or just the DN entry (first param) or
its specific occurrence (second param).
The crappy zlib and openssl libs both define a free_func as a different typedef.
That's a very clever idea to use such a generic name in general purpose libraries,
really... The zlib one is easier to redefine than openssl's, so let's only fix this
one.
Decreasing the deflateInit2's memLevel parameter from 9 to 8 does not
affect the compression ratio and increases the compression speed by 12%.
Lower values do not increase transfer speed but decrease the compression
ratio so it looks like 8 is optimal.
A number of older browsers have many issues with compressed contents. It
happens that all these older browsers announce themselves as "Mozilla/4"
and that despite not being all broken, the amount of working browsers
announcing themselves this way compared to all other ones is so tiny
that it's not worth wasting cycles trying to adapt to every specific
one.
So let's simply disable compression for these older browsers.
More information on this very detailed article :
http://zoompf.com/2012/02/lose-the-wait-http-compression
