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.
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.
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".
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.
This commit introduces HTTP compression using the zlib library.
http_response_forward_body has been modified to call the compression
functions.
This feature includes 3 algorithms: identity, gzip and deflate:
* identity: this is mostly for debugging, and it was useful for
developping the compression feature. With Content-Length in input, it
is making each chunk with the data available in the current buffer.
With chunks in input, it is rechunking, the output chunks will be
bigger or smaller depending of the size of the input chunk and the
size of the buffer. Identity does not apply any change on data.
* gzip: same as identity, but applying a gzip compression. The data
are deflated using the Z_NO_FLUSH flag in zlib. When there is no more
data in the input buffer, it flushes the data in the output buffer
(Z_SYNC_FLUSH). At the end of data, when it receives the last chunk in
input, or when there is no more data to read, it writes the end of
data with Z_FINISH and the ending chunk.
* deflate: same as gzip, but with deflate algorithm and zlib format.
Note that this algorithm has ambiguous support on many browsers and
no support at all from recent ones. It is strongly recommended not
to use it for anything else than experimentation.
You can't choose the compression ratio at the moment, it will be set to
Z_BEST_SPEED (1), as tests have shown very little benefit in terms of
compression ration when going above for HTML contents, at the cost of
a massive CPU impact.
Compression will be activated depending of the Accept-Encoding request
header. With identity, it does not take care of that header.
To build HAProxy with zlib support, use USE_ZLIB=1 in the make
parameters.
This work was initially started by David Du Colombier at Exceliance.
Most calls to channel_forward() are performed with short byte counts and
are already optimized in channel_forward() taking just a few instructions.
Thus it's a waste of CPU cycles to call a function for this, let's just
inline the short byte count case and fall back to the common one for
remaining situations.
Doing so has increased the chunked encoding parser's performance by 12% !
ACL and sample fetches use args list and it is really not convenient to
check for null args everywhere. Now for empty args we pass a constant
list of end of lists. It will allow us to remove many useless checks.
This field was used to trace precisely where a session was terminated
but it did not survive code rearchitecture and was not used at all
anymore. Let's get rid of it.
With this commit, we now separate the channel from the buffer. This will
allow us to replace buffers on the fly without touching the channel. Since
nobody is supposed to keep a reference to a buffer anymore, doing so is not
a problem and will also permit some copy-less data manipulation.
Interestingly, these changes have shown a 2% performance increase on some
workloads, probably due to a better cache placement of data.
These two new log-format tags report the SSL protocol version (%sslv) and the
SSL ciphers (%sslc) used for the connection with the client. For instance, to
append these information just after the client's IP/port address information
on an HTTP log line, use the following configuration :
log-format %Ci:%Cp\ %sslv:%sslc\ [%t]\ %ft\ %b/%s\ %Tq/%Tw/%Tc/%Tr/%Tt\ %st\ %B\ %cc\ \ %cs\ %tsc\ %ac/%fc/%bc/%sc/%rc\ %sq/%bq\ %hr\ %hs\ %{+Q}r
It will report a line such as the following one :
Oct 12 20:47:30 haproxy[9643]: 127.0.0.1:43602 TLSv1:AES-SHA [12/Oct/2012:20:47:30.303] stick2~ stick2/s1 7/0/12/0/19 200 145 - - ---- 0/0/0/0/0 0/0 "GET /?t=0 HTTP/1.0"
When we start logging SSL information, we need the SSL struct to be
present even past the conn_xprt_close() call. In order to achieve this,
we should use refcounting on the connection and the transport layer. At
the moment it's not worth using plain refcounting as only the logs require
this, so instead of real refcounting we just use a flag which will be set
by the log subsystem when SSL data need to be logged.
What happens then is that the xprt->close() call is ignored and the
transport layer is closed again during session_free(), after the log
line is emitted.
When calling conn_xprt_close(), we always clear the transport pointer
so that all transport layers leave the connection in the same state after
a close. This will also make it safer and cheaper to call conn_xprt_close()
multiple times if needed.
Just like with the "bind" lines, we'll switch the "server" line
parsing to keyword registration. The code is essentially the same
as for bind keywords, with minor changes such as support for the
default-server keywords and support for variable argument count.
This callback sends a PROXY protocol line on the outgoing connection,
with the local and remote endpoint information. This is used for local
connections (eg: health checks) where the other end needs to have a
valid address and no connection is relayed.
It was previously in frontend.c but there is no reason for this anymore
considering that all the information involved is in the connection itself
only. Theorically this should be in the socket layer but we don't have
this yet.
We absolutely want to disable FD polling after an error is detected,
otherwise the data layer has to do it and it's far from being obvious
at these layers.
The way we did it was a bit tricky in conn_update_*_polling and
conn_*_polling_changes. However it has almost no impact on performance
and code size both for the fast and slow path.
We'll now be able to remove some flag updates in the stream interface.
The connection flags have progressively been added one after the other
and were not very well organized. Some of them are often used together
and a number of operations are performed on the DATA/SOCK ENA/POL flags.
Thus, they have been reorganized so that flags that work together are
close to each other (allows immediate operands on ARM) and that polling
changes can be detected with fewer operations using a simple shift and
xor. The handshakes are now the last ones so that it will be easier to
add new ones after without risking a collision. All activity-related
flags are also grouped together.
The generic data-layer init callback is now used after the transport
layer is complete and before calling the data layer recv/send callbacks.
This allows the session to switch from the embryonic session data layer
to the complete stream interface data layer, by making conn_session_complete()
the data layer's init callback.
It sill looks awkwards that the init() callback must be used opon error,
but except by adding yet another one, it does not seem to be mergeable
into another function (eg: it should probably not be merged with ->wake
to avoid unneeded calls during the handshake, though semantically that
would make sense).
Instead of calling conn_notify_si() from the connection handler, we
now call data->wake(), which will allow us to use a different callback
with health checks.
Note that we still rely on a flag in order to decide whether or not
to call this function. The reason is that with embryonic sessions,
the callback is already initialized to si_conn_cb without the flag,
and we can't call the SI notify function in the leave path before
the stream interface is initialized.
This issue should be addressed by involving a different data_cb for
embryonic sessions and for stream interfaces, that would be changed
during session_complete() for the final data_cb.
Now conn->data will designate the data layer which is the client for
the transport layer. In practice it's the stream interface and will
soon also be the health checks.
While working on the changes required to make the health checks use the
new connections, it started to become obvious that some naming was not
logical at all in the connections. Specifically, it is not logical to
call the "data layer" the layer which is in charge for all the handshake
and which does not yet provide a data layer once established until a
session has allocated all the required buffers.
In fact, it's more a transport layer, which makes much more sense. The
transport layer offers a medium on which data can transit, and it offers
the functions to move these data when the upper layer requests this. And
it is the upper layer which iterates over the transport layer's functions
to move data which should be called the data layer.
The use case where it's obvious is with embryonic sessions : an incoming
SSL connection is accepted. Only the connection is allocated, not the
buffers nor stream interface, etc... The connection handles the SSL
handshake by itself. Once this handshake is complete, we can't use the
data functions because the buffers and stream interface are not there
yet. Hence we have to first call a specific function to complete the
session initialization, after which we'll be able to use the data
functions. This clearly proves that SSL here is only a transport layer
and that the stream interface constitutes the data layer.
A similar change will be performed to rename app_cb => data, but the
two could not be in the same commit for obvious reasons.
It removes dependencies with futex or mutex but ssl performances decrease
using nbproc > 1 because switching process force session renegotiation.
This can be useful on small systems which never intend to run in multi-process
mode.
We don't needa to lock the memory when there is a single process. This can
make a difference on small systems where locking is much more expensive than
just a test.
This will be needed to find the stream interface from the connection
once they're detached, but in the more immediate term, we'll need this
for health checks since they don't use a stream interface.
Unix permissions are per-bind configuration line and not per listener,
so let's concretize this in the way the config is stored. This avoids
some unneeded loops to set permissions on all listeners.
The access level is not part of the unix perms so it has been moved
away. Once we can use str2listener() to set all listener addresses,
we'll have a bind keyword parser for this one.
Navigating through listeners was very inconvenient and error-prone. Not to
mention that listeners were linked in reverse order and reverted afterwards.
In order to definitely get rid of these issues, we now do the following :
- frontends have a dual-linked list of bind_conf
- frontends have a dual-linked list of listeners
- bind_conf have a dual-linked list of listeners
- listeners have a pointer to their bind_conf
This way we can now navigate from anywhere to anywhere and always find the
proper bind_conf for a given listener, as well as find the list of listeners
for a current bind_conf.
When an unknown "bind" keyword is detected, dump the list of all
registered keywords. Unsupported default alternatives are also reported
as "not supported".
Registering new SSL bind keywords was not particularly handy as it required
many #ifdef in cfgparse.c. Now the code has moved to ssl_sock.c which calls
a register function for all the keywords.
Error reporting was also improved by this move, because the called functions
build an error message using memprintf(), which can span multiple lines if
needed, and each of these errors will be displayed indented in the context of
the bind line being processed. This is important when dealing with certificate
directories which can report multiple errors.
With the arrival of SSL, the "bind" keyword has received even more options,
all of which are processed in cfgparse in a cumbersome way. So it's time to
let modules register their own bind options. This is done very similarly to
the ACLs with a small difference in that we make the difference between an
unknown option and a known, unimplemented option.
Some settings need to be merged per-bind config line and are not necessarily
SSL-specific. It becomes quite inconvenient to have this ssl_conf SSL-specific,
so let's replace it with something more generic.
A side effect of this change is that the "ssl" keyword on "bind" lines is now
just a boolean and that "crt" is needed to designate certificate files or
directories.
Note that much refcounting was needed to have the free() work correctly due to
the number of cert aliases which can make a context be shared by multiple names.
SSL config holds many parameters which are per bind line and not per
listener. Let's use a per-bind line config instead of having it
replicated for each listener.
At the moment we only do this for the SSL part but this should probably
evolved to handle more of the configuration and maybe even the state per
bind line.
This SSL session cache was developped at Exceliance and is the same that
was proposed for stunnel and stud. It makes use of a shared memory area
between the processes so that sessions can be handled by any process. It
is only useful when haproxy runs with nbproc > 1, but it does not hurt
performance at all with nbproc = 1. The aim is to totally replace OpenSSL's
internal cache.
The cache is optimized for Linux >= 2.6 and specifically for x86 platforms.
On Linux/x86, it makes use of futexes for inter-process locking, with some
x86 assembly for the locked instructions. On other architectures, GCC
builtins are used instead, which are available starting from gcc 4.1.
On other operating systems, the locks fall back to pthread mutexes so
libpthread is automatically linked. It is not recommended since pthreads
are much slower than futexes. The lib is only linked if SSL is enabled.
CVE-2009-3555 suggests that client-initiated renegociation should be
prevented in the middle of data. The workaround here consists in having
the SSL layer notify our callback about a handshake occurring, which in
turn causes the connection to be marked in the error state if it was
already considered established (which means if a previous handshake was
completed). The result is that the connection with the client is immediately
aborted and any pending data are dropped.
This data layer supports socket-to-buffer and buffer-to-socket operations.
No sock-to-pipe nor pipe-to-sock functions are provided, since splicing does
not provide any benefit with data transformation. At best it could save a
memcpy() and avoid keeping a buffer allocated but that does not seem very
useful.
An init function and a close function are provided because the SSL context
needs to be allocated/freed.
A data-layer shutw() function is also provided because upon successful
shutdown, we want to store the SSL context in the cache in order to reuse
it for future connections and avoid a new key generation.
The handshake function is directly called from the connection handler.
At this point it is not certain whether this will remain this way or
if a new ->handshake callback will be added to the data layer so that
the connection handler doesn't care about SSL.
The sock-to-buf and buf-to-sock functions are all capable of enabling
the SSL handshake at any time. This also implies polling in the opposite
direction to what was expected. The upper layers must take that into
account (it is OK right now with the stream interface).
It appears that fd.h includes a number of unneeded files and was
included from standard.h, and as such served as an intermediary
to provide almost everything to everyone.
By removing its useless includes, a long dependency chain broke
but could easily be fixed.
Polling flags were set for data and sock layer, but while this does make
sense for the ENA flag, it does not for the POL flag which translates the
detection of an EAGAIN condition. So now we remove the {DATA,SOCK}_POL*
flags and instead introduce two new layer-independant flags (WANT_RD and
WANT_WR). These flags are only set when an EAGAIN is encountered so that
polling can be enabled.
In order for these flags to have any meaning they are not persistent and
have to be cleared by the connection handler before calling the I/O and
data callbacks. For this reason, changes detection has been slightly
improved. Instead of comparing the WANT_* flags with CURR_*_POL, we only
check if the ENA status changes, or if the polling appears, since we don't
want to detect the useless poll to ena transition. Tests show that this
has eliminated one useless call to __fd_clr().
Finally the conn_set_polling() function which was becoming complex and
required complex operations from the caller was split in two and replaced
its two only callers (conn_update_data_polling and conn_update_sock_polling).
The two functions are now much smaller due to the less complex conditions.
Note that it would be possible to re-merge them and only pass a mask but
this does not appear much interesting.
The PROXY protocol is now decoded in the connection before other
handshakes. This means that it may be extracted from a TCP stream
before SSL is decoded from this stream.
