Since multiple streams can share one session attached to one listener,
the listener_release() call must be done in session_free() and not in
stream_free(), otherwise we end up with a negative count in H2.
This callback will be used to release upper layers when a mux is in
use. Given that the mux can be asynchronously deleted, we need a way
to release the extra information such as the session.
This callback will be called directly by the mux upon releasing
everything and before the connection itself is released, so that
the callee can find its information inside the connection if needed.
The way it currently works is not perfect, and most likely this should
instead become a mux release callback, but for now we have no easy way
to add mux-specific stuff, and since there's one mux per connection,
it works fine this way.
When an incoming connection is made on an HTTP mode frontend, the
session now looks up the mux to use based on the ALPN token and the
proxy mode. This will allow easier mux registration, and we don't
need to hard-code the mux_pt_ops anymore.
For HTTP/2 and QUIC, we'll need to deal with multiplexed streams inside
a connection. After quite a long brainstorming, it appears that the
connection interface to the existing streams is appropriate just like
the connection interface to the lower layers. In fact we need to have
the mux layer in the middle of the connection, between the transport
and the data layer.
A mux can exist on two directions/sides. On the inbound direction, it
instanciates new streams from incoming connections, while on the outbound
direction it muxes streams into outgoing connections. The difference is
visible on the mux->init() call : in one case, an upper context is already
known (outgoing connection), and in the other case, the upper context is
not yet known (incoming connection) and will have to be allocated by the
mux. The session doesn't have to create the new streams anymore, as this
is performed by the mux itself.
This patch introduces this and creates a pass-through mux called
"mux_pt" which is used for all new connections and which only
calls the data layer's recv,send,wake() calls. One incoming stream
is immediately created when init() is called on the inbound direction.
There should not be any visible impact.
Note that the connection's mux is purposely not set until the session
is completed so that we don't accidently run with the wrong mux. This
must not cause any issue as the xprt_done_cb function is always called
prior to using mux's recv/send functions.
Now, each proxy contains a lock that must be used when necessary to protect
it. Moreover, all proxy's counters are now updated using atomic operations.
First, we use atomic operations to update jobs/totalconn/actconn variables,
listener's nbconn variable and listener's counters. Then we add a lock on
listeners to protect access to their information. And finally, listener queues
(global and per proxy) are also protected by a lock. Here, because access to
these queues are unusal, we use the same lock for all queues instead of a global
one for the global queue and a lock per proxy for others.
2 global locks have been added to protect, respectively, the run queue and the
wait queue. And a process mask has been added on each task. Like for FDs, this
mask is used to know which threads are allowed to process a task.
For many tasks, all threads are granted. And this must be your first intension
when you create a new task, else you have a good reason to make a task sticky on
some threads. This is then the responsibility to the process callback to lock
what have to be locked in the task context.
Nevertheless, all tasks linked to a session must be sticky on the thread
creating the session. It is important that I/O handlers processing session FDs
and these tasks run on the same thread to avoid conflicts.
When compiled with Openssl >= 1.1.1, before attempting to do the handshake,
try to read any early data. If any early data is present, then we'll create
the session, read the data, and handle the request before we're doing the
handshake.
For this, we add a new connection flag, CO_FL_EARLY_SSL_HS, which is not
part of the CO_FL_HANDSHAKE set, allowing to proceed with a session even
before an SSL handshake is completed.
As early data do have security implication, we let the origin server know
the request comes from early data by adding the "Early-Data" header, as
specified in this draft from the HTTP working group :
https://datatracker.ietf.org/doc/html/draft-ietf-httpbis-replay
These flags are not exactly for the data layer, they instead indicate
what is expected from the transport layer. Since we're going to split
the connection between the transport and the data layers to insert a
mux layer, it's important to have a clear idea of what each layer does.
All function conn_data_* used to manipulate these flags were renamed to
conn_xprt_*.
Commit bcb86ab ("MINOR: session: add a streams field to the session
struct") added this list of streams that is not needed anymore. Let's
get rid of it now.
There are several places where we see feconn++, feconn--, totalconn++ and
an increment on the frontend's number of connections and connection rate.
This is done exactly once per session in each direction, so better take
care of this counter in the session and simplify the callers. At least it
ensures a better symmetry. It also ensures consistency as till now the
lua/spoe/peers frontend didn't have these counters properly set, which can
be useful at least for troubleshooting.
session_accept_fd() may either successfully complete a session creation,
or defer it to conn_complete_session() depending of whether a handshake
remains to be performed or not. The problem is that all the code after
the handshake was duplicated between the two functions.
This patch make session_accept_fd() synchronously call
conn_complete_session() to finish the session creation. It is only needed
to check if the session's task has to be released or not at the end, which
is fairly minimal. This way there is now a single place where the sessions
are created.
Commit 8e3c6ce ("MEDIUM: connection: get rid of data->init() which was
not for data") simplified conn_complete_session() but introduced a
confusing check which cannot happen on CO_FL_HANDSHAKE. Make it clear
that this call is final and will either succeed and complete the
session or fail.
Instead of duplicating some sensitive listener-specific code in the
session and in the stream code, let's call listener_release() when
releasing a connection attached to a listener.
Each user of a session increments/decrements the jobs variable at its
own place, resulting in a real mess and inconsistencies between them.
Let's have session_new() increment jobs and session_free() decrement
it.
The session may need to enforce a timeout when waiting for a handshake.
Till now we used a trick to avoid allocating a pointer, we used to set
the connection's owner to the task and set the task's context to the
session, so that it was possible to circle between all of them. The
problem is that we'll really need to pass the pointer to the session
to the upper layers during initialization and that the only place to
store it is conn->owner, which is squatted for this trick.
So this patch moves the struct task* into the session where it should
always have been and ensures conn->owner points to the session until
the data layer is properly initialized.
Currently a task is allocated in session_new() and serves two purposes :
- either the handshake is complete and it is offered to the stream via
the second arg of stream_new()
- or the handshake is not complete and it's diverted to be used as a
timeout handler for the embryonic session and repurposed once we land
into conn_complete_session()
Furthermore, the task's process() function was taken from the listener's
handler in conn_complete_session() prior to being replaced by a call to
stream_new(). This will become a serious mess with the mux.
Since it's impossible to have a stream without a task, this patch removes
the second arg from stream_new() and make this function allocate its own
task. In session_accept_fd(), we now only allocate the task if needed for
the embryonic session and delete it later.
The ->init() callback of the connection's data layer was only used to
complete the session's initialisation since sessions and streams were
split apart in 1.6. The problem is that it creates a big confusion in
the layers' roles as the session has to register a dummy data layer
when waiting for a handshake to complete, then hand it off to the
stream which will replace it.
The real need is to notify that the transport has finished initializing.
This should enable a better splitting between these layers.
This patch thus introduces a connection-specific callback called
xprt_done_cb() which informs about handshake successes or failures. With
this, data->init() can disappear, CO_FL_INIT_DATA as well, and we don't
need to register a dummy data->wake() callback to be notified of errors.
Till now connections used to rely exclusively on file descriptors. It
was planned in the past that alternative solutions would be implemented,
leading to member "union t" presenting sock.fd only for now.
With QUIC, the connection will need to continue to exist but will not
rely on a file descriptor but a connection ID.
So this patch introduces a "connection handle" which is either a file
descriptor or a connection ID, to replace the existing "union t". We've
now removed the intermediate "struct sock" which was never used. There
is no functional change at all, though the struct connection was inflated
by 32 bits on 64-bit platforms due to alignment.
We now refrain from clearing a session's variables, counters, and from
releasing it as long as at least one stream references it. For now it
never happens but with H2 this will be mandatory to avoid double frees.
It doesn't make sense that stream_new() doesn't sets the target nor
analysers and that the caller has to do it even if it doesn't know
about streams (eg: in session_accept_fd()). This causes trouble for
H2 where the applet handling the protocol cannot properly change
these information during its init phase.
Let's ensure it's always set and that the callers don't set it anymore.
Note: peers and lua don't use analysers and that's properly handled.
The task_wakeup was called on stream_new, but the task/stream
wasn't fully initialized yet. The task_wakeup must be called
explicitly by the caller once the task/stream is initialized.
Now we exclusively use xprt_get(XPRT_RAW) instead of &raw_sock or
xprt_get(XPRT_SSL) for &ssl_sock. This removes a bunch of #ifdef and
include spread over a number of location including backend, cfgparse,
checks, cli, hlua, log, server and session.
Historically, all listeners have a pointer to the frontend. But since
the introduction of SSL, we now have an intermediary layer called
bind_conf corresponding to a "bind" line. It makes no sense to have
the frontend on each listener given that it's the same for all
listeners belonging to a same bind_conf. Also certain parts like
SSL can only operate on bind_conf and need the frontend.
This patch fixes this by moving the frontend pointer from the listener
to the bind_conf. The extra indirection is quite cheap given and the
places were this is used are very scarce.
A mistake was made when the socket layer was cut into proto and
transport, the transport was attached to the listener while all
listeners in a single "bind" line always have exactly the same
transport. It doesn't seem obvious but this is the reason why there
are so many #ifdefs USE_OPENSSL in cfgparse : a lot of operations
have to be open-coded because cfgparse only manipulates bind_conf
and we don't have the information of the transport layer here.
Very little code makes use of the transport layer, mainly session
setup and log. These places can afford an extra pointer indirection
(the listener points to the bind_conf). This change is thus very small,
it saves a little bit of memory (8B per listener) and makes the code
more flexible.
In 1.6-dev2, commit 32990b5 ("MEDIUM: session: remove the task pointer
from the session") introduced a bug which can sometimes crash the process
on resource shortage. When stream_complete() returns -1, it has already
reattached the connection to the stream, then kill_mini_session() is
called and still expects to find the task in conn->owner. Note that
since this commit, the code has moved a bit and is now in stream_new()
but the problem remains the same.
Given that we already know the task around these places, let's simply
pass the task to kill_mini_session().
The conditions currently at risk are :
- failure to initialize filters for the new stream (lack of memory or
any filter returning < 0 on attach())
- failure to attach filters (any filter returning < 0 on stream_start())
- frontend's accept() returning < 0 (allocation failure)
This fix is needed in 1.7 and 1.6.
There's no more reason to keep tcp rules processing inside proto_tcp.c
given that there is nothing in common there except these 3 letters : tcp.
The tcp rules are in fact connection, session and content processing rules.
Let's move them to "tcp-rules" and let them live their life there.
This commit introduces "tcp-request session" rules. These are very
much like "tcp-request connection" rules except that they're processed
after the handshake, so it is possible to consider SSL information and
addresses rewritten by the proxy protocol header in actions. This is
particularly useful to track proxied sources as this was not possible
before, given that tcp-request content rules are processed after each
HTTP request. Similarly it is possible to assign the proxied source
address or the client's cert to a variable.
This is in order to make integration of tcp-request-session cleaner :
- tcp_exec_req_rules() was renamed tcp_exec_l4_rules()
- LI_O_TCP_RULES was renamed LI_O_TCP_L4_RULES
(LI_O_*'s horrible indent was also fixed and a provision was left
for L5 rules).
When NetScaler application switch is used as L3+ switch, informations
regarding the original IP and TCP headers are lost as a new TCP
connection is created between the NetScaler and the backend server.
NetScaler provides a feature to insert in the TCP data the original data
that can then be consumed by the backend server.
Specifications and documentations from NetScaler:
https://support.citrix.com/article/CTX205670https://www.citrix.com/blogs/2016/04/25/how-to-enable-client-ip-in-tcpip-option-of-netscaler/
When CIP is enabled on the NetScaler, then a TCP packet is inserted just after
the TCP handshake. This is composed as:
- CIP magic number : 4 bytes
Both sender and receiver have to agree on a magic number so that
they both handle the incoming data as a NetScaler Client IP insertion
packet.
- Header length : 4 bytes
Defines the length on the remaining data.
- IP header : >= 20 bytes if IPv4, 40 bytes if IPv6
Contains the header of the last IP packet sent by the client during TCP
handshake.
- TCP header : >= 20 bytes
Contains the header of the last TCP packet sent by the client during TCP
handshake.
This patch adds the support of filters in HAProxy. The main idea is to have a
way to "easely" extend HAProxy by adding some "modules", called filters, that
will be able to change HAProxy behavior in a programmatic way.
To do so, many entry points has been added in code to let filters to hook up to
different steps of the processing. A filter must define a flt_ops sutrctures
(see include/types/filters.h for details). This structure contains all available
callbacks that a filter can define:
struct flt_ops {
/*
* Callbacks to manage the filter lifecycle
*/
int (*init) (struct proxy *p);
void (*deinit)(struct proxy *p);
int (*check) (struct proxy *p);
/*
* Stream callbacks
*/
void (*stream_start) (struct stream *s);
void (*stream_accept) (struct stream *s);
void (*session_establish)(struct stream *s);
void (*stream_stop) (struct stream *s);
/*
* HTTP callbacks
*/
int (*http_start) (struct stream *s, struct http_msg *msg);
int (*http_start_body) (struct stream *s, struct http_msg *msg);
int (*http_start_chunk) (struct stream *s, struct http_msg *msg);
int (*http_data) (struct stream *s, struct http_msg *msg);
int (*http_last_chunk) (struct stream *s, struct http_msg *msg);
int (*http_end_chunk) (struct stream *s, struct http_msg *msg);
int (*http_chunk_trailers)(struct stream *s, struct http_msg *msg);
int (*http_end_body) (struct stream *s, struct http_msg *msg);
void (*http_end) (struct stream *s, struct http_msg *msg);
void (*http_reset) (struct stream *s, struct http_msg *msg);
int (*http_pre_process) (struct stream *s, struct http_msg *msg);
int (*http_post_process) (struct stream *s, struct http_msg *msg);
void (*http_reply) (struct stream *s, short status,
const struct chunk *msg);
};
To declare and use a filter, in the configuration, the "filter" keyword must be
used in a listener/frontend section:
frontend test
...
filter <FILTER-NAME> [OPTIONS...]
The filter referenced by the <FILTER-NAME> must declare a configuration parser
on its own name to fill flt_ops and filter_conf field in the proxy's
structure. An exemple will be provided later to make it perfectly clear.
For now, filters cannot be used in backend section. But this is only a matter of
time. Documentation will also be added later. This is the first commit of a long
list about filters.
It is possible to have several filters on the same listener/frontend. These
filters are stored in an array of at most MAX_FILTERS elements (define in
include/types/filters.h). Again, this will be replaced later by a list of
filters.
The filter API has been highly refactored. Main changes are:
* Now, HA supports an infinite number of filters per proxy. To do so, filters
are stored in list.
* Because filters are stored in list, filters state has been moved from the
channel structure to the filter structure. This is cleaner because there is no
more info about filters in channel structure.
* It is possible to defined filters on backends only. For such filters,
stream_start/stream_stop callbacks are not called. Of course, it is possible
to mix frontend and backend filters.
* Now, TCP streams are also filtered. All callbacks without the 'http_' prefix
are called for all kind of streams. In addition, 2 new callbacks were added to
filter data exchanged through a TCP stream:
- tcp_data: it is called when new data are available or when old unprocessed
data are still waiting.
- tcp_forward_data: it is called when some data can be consumed.
* New callbacks attached to channel were added:
- channel_start_analyze: it is called when a filter is ready to process data
exchanged through a channel. 2 new analyzers (a frontend and a backend)
are attached to channels to call this callback. For a frontend filter, it
is called before any other analyzer. For a backend filter, it is called
when a backend is attached to a stream. So some processing cannot be
filtered in that case.
- channel_analyze: it is called before each analyzer attached to a channel,
expects analyzers responsible for data sending.
- channel_end_analyze: it is called when all other analyzers have finished
their processing. A new analyzers is attached to channels to call this
callback. For a TCP stream, this is always the last one called. For a HTTP
one, the callback is called when a request/response ends, so it is called
one time for each request/response.
* 'session_established' callback has been removed. Everything that is done in
this callback can be handled by 'channel_start_analyze' on the response
channel.
* 'http_pre_process' and 'http_post_process' callbacks have been replaced by
'channel_analyze'.
* 'http_start' callback has been replaced by 'http_headers'. This new one is
called just before headers sending and parsing of the body.
* 'http_end' callback has been replaced by 'channel_end_analyze'.
* It is possible to set a forwarder for TCP channels. It was already possible to
do it for HTTP ones.
* Forwarders can partially consumed forwardable data. For this reason a new
HTTP message state was added before HTTP_MSG_DONE : HTTP_MSG_ENDING.
Now all filters can define corresponding callbacks (http_forward_data
and tcp_forward_data). Each filter owns 2 offsets relative to buf->p, next and
forward, to track, respectively, input data already parsed but not forwarded yet
by the filter and parsed data considered as forwarded by the filter. A any time,
we have the warranty that a filter cannot parse or forward more input than
previous ones. And, of course, it cannot forward more input than it has
parsed. 2 macros has been added to retrieve these offets: FLT_NXT and FLT_FWD.
In addition, 2 functions has been added to change the 'next size' and the
'forward size' of a filter. When a filter parses input data, it can alter these
data, so the size of these data can vary. This action has an effet on all
previous filters that must be handled. To do so, the function
'filter_change_next_size' must be called, passing the size variation. In the
same spirit, if a filter alter forwarded data, it must call the function
'filter_change_forward_size'. 'filter_change_next_size' can be called in
'http_data' and 'tcp_data' callbacks and only these ones. And
'filter_change_forward_size' can be called in 'http_forward_data' and
'tcp_forward_data' callbacks and only these ones. The data changes are the
filter responsability, but with some limitation. It must not change already
parsed/forwarded data or data that previous filters have not parsed/forwarded
yet.
Because filters can be used on backends, when we the backend is set for a
stream, we add filters defined for this backend in the filter list of the
stream. But we must only do that when the backend and the frontend of the stream
are not the same. Else same filters are added a second time leading to undefined
behavior.
The HTTP compression code had to be moved.
So it simplifies http_response_forward_body function. To do so, the way the data
are forwarded has changed. Now, a filter (and only one) can forward data. In a
commit to come, this limitation will be removed to let all filters take part to
data forwarding. There are 2 new functions that filters should use to deal with
this feature:
* flt_set_http_data_forwarder: This function sets the filter (using its id)
that will forward data for the specified HTTP message. It is possible if it
was not already set by another filter _AND_ if no data was yet forwarded
(msg->msg_state <= HTTP_MSG_BODY). It returns -1 if an error occurs.
* flt_http_data_forwarder: This function returns the filter id that will
forward data for the specified HTTP message. If there is no forwarder set, it
returns -1.
When an HTTP data forwarder is set for the response, the HTTP compression is
disabled. Of course, this is not definitive.
We don't pass sess->origin anymore but the pointer to the previous step. Now
it should be much easier to chain elements together once applets are moved out
of streams. Indeed, the session is only used for configuration and not for the
dynamic chaining anymore.
It's not the stream's job to manipulate the connection's flags, it's
more related to the session that accepted the new connection. And the
only case where we have to do it conditionally is based on the frontend
which is known from the session, thus it makes sense to do it there.
When the stream is instanciated from an applet, it doesn't necessarily
have a listener. The listener was sparsely used there, just to retrieve
the task function, update the listeners' stats, and set the analysers
and default target, both of which are often zero from applets. Thus
these elements are now initialized with default values that the caller
is free to change if desired.
The frontend is generic and does not depend on a file descriptor,
so applying some socket options to the incoming fd is not its role.
Let's move the setsockopt() calls earlier in session_accept_fd()
where others are done as well.
The function was called stream_accept_session(), let's rename it
stream_new() and make it return the newly allocated pointer. It's
more convenient for some callers who need it.
This concerns everythins related to accepting a new session and
expiring the embryonic session. There's still a hard-coded call
to stream_accept_session() which could be set somewhere in the
frontend, but for now it's not a problem.
There is now a pointer to the session in the stream, which is NULL
for now. The session pool is created as well. Some parts will move
from the stream to the session now.
With HTTP/2, we'll have to support multiplexed streams. A stream is in
fact the largest part of what we currently call a session, it has buffers,
logs, etc.
In order to catch any error, this commit removes any reference to the
struct session and tries to rename most "session" occurrences in function
names to "stream" and "sess" to "strm" when that's related to a session.
The files stream.{c,h} were added and session.{c,h} removed.
The session will be reintroduced later and a few parts of the stream
will progressively be moved overthere. It will more or less contain
only what we need in an embryonic session.
Sample fetch functions and converters will have to change a bit so
that they'll use an L5 (session) instead of what's currently called
"L4" which is in fact L6 for now.
Once all changes are completed, we should see approximately this :
L7 - http_txn
L6 - stream
L5 - session
L4 - connection | applet
There will be at most one http_txn per stream, and a same session will
possibly be referenced by multiple streams. A connection will point to
a session and to a stream. The session will hold all the information
we need to keep even when we don't yet have a stream.
Some more cleanup is needed because some code was already far from
being clean. The server queue management still refers to sessions at
many places while comments talk about connections. This will have to
be cleaned up once we have a server-side connection pool manager.
Stream flags "SN_*" still need to be renamed, it doesn't seem like
any of them will need to move to the session.
This will be useful later to state that some listeners have to use
certain decoders (typically an HTTP/2 decoder) regardless of the
regular processing applied to other listeners. For now it simply
defaults to the frontend's default target, and it is used by the
session.
Some services such as peers and CLI pre-set the target applet immediately
during accept(), and for this reason they're forced to have a dedicated
accept() function which does not even properly follow everything the regular
one does (eg: sndbuf/rcvbuf/linger/nodelay are not set, etc).
Let's store the default target when known into the frontend's config so that
it's session_accept() which automatically sets it.
Now that we can get the session from the channel, let's simplify the
prototype of session_alloc_recv_buffer() to only require the channel.
Both the caller and the function are now simplified.
All functions dealing with connection establishment currently use a
pointer to the stream interface. Now we know it cannot change and is
always s->si[1].
In process_session, we had around 300 accesses to channels and stream-ints
from the session. Not only this inflates the code due to the large offsets
from the original pointer, but readability can be improved. Let's have 4
local variables for the channels and stream-ints.
These 4 combinations are needlessly complicated since the session already
has direct access to the associated stream interfaces without having to
check an indirect pointer.
The purpose of these two macros will be to pass via the session to
find the relevant stream interfaces so that we don't need to store
the ->cons nor ->prod pointers anymore. Currently they're only defined
so that all references could be removed.
Note that many places need a second pass of clean up so that we don't
have any chn_prod(&s->req) anymore and only &s->si[0] instead, and
conversely for the 3 other cases.
We go back to the session to get the owner. Here again it's very easy
and is just a matter of relative offsets. Since the owner always exists
and always points to the session's task, we can remove some unneeded
tests.
This new flag "SI_FL_ISBACK" is set only on the back SI and is cleared
on the front SI. That way it's possible only by looking at the SI to
know what side it is.
We'll soon remove direct references to the channels from the stream
interface since everything belongs to the same session, so let's
first not dereference si->ib / si->ob anymore and use macros instead.
The channels were pointers to outside structs and this is not needed
anymore since the buffers have moved, but this complicates operations.
Move them back into the session so that both channels and stream interfaces
are always allocated for a session. Some places (some early sample fetch
functions) used to validate that a channel was NULL prior to dereferencing
it. Now instead we check if chn->buf is NULL and we force it to remain NULL
until the channel is initialized.
Actually, the Lua context is always initilized in each
session, even if the session doesn't use Lua. This
behavior cause 5% performances loss.
This patch initilize the Lua only if it is use by the
session. The initialization is now on demand.
This class of functions permit to access to all the functions
associated with the transaction like http header, HAProxy internal
fetches, etc ...
This patch puts the skeleton of this class. The class will be
enhanced later.
Later, the processing of some actions needs to be interrupted and resumed
later. This patch permit to resume the actions. The actions that needs
to run with the resume mode are not yet avalaible. It will be soon with
Lua patches. So the code added by this patch is untestable for the moment.
The list of "tcp_exec_req_rules" cannot resme because is called by the
unresumable function "accept_session".
This behavior is already existing for the "WAIT_HTTP" analyzer,
this patch just extends the system to any analyzer that would
be waked up on response activity.
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.
A session doesn't need buffers all the time, especially when they're
empty. With this patch, we don't allocate buffers anymore when the
session is initialized, we only allocate them in two cases :
- during process_session()
- during I/O operations
During process_session(), we try hard to allocate both buffers at once
so that we know for sure that a started operation can complete. Indeed,
a previous version of this patch used to allocate one buffer at a time,
but it can result in a deadlock when all buffers are allocated for
requests for example, and there's no buffer left to emit error responses.
Here, if any of the buffers cannot be allocated, the whole operation is
cancelled and the session is added at the tail of the buffer wait queue.
At the end of process_session(), a call to session_release_buffers() is
done so that we can offer unused buffers to other sessions waiting for
them.
For I/O operations, we only need to allocate a buffer on the Rx path.
For this, we only allocate a single buffer but ensure that at least two
are available to avoid the deadlock situation. In case buffers are not
available, SI_FL_WAIT_ROOM is set on the stream interface and the session
is queued. Unused buffers resulting either from a successful send() or
from an unused read buffer are offered to pending sessions during the
->wake() callback.
When a session_alloc_buffers() fails to allocate one or two buffers,
it subscribes the session to buffer_wq, and waits for another session
to release buffers. It's then removed from the queue and woken up with
TASK_WAKE_RES, and can attempt its allocation again.
We decide to try to wake as many waiters as we release buffers so
that if we release 2 and two waiters need only once, they both have
their chance. We must never come to the situation where we don't wake
enough tasks up.
It's common to release buffers after the completion of an I/O callback,
which can happen even if the I/O could not be performed due to half a
failure on memory allocation. In this situation, we don't want to move
out of the wait queue the session that was just added, otherwise it
will never get any buffer. Thus, we only force ourselves out of the
queue when freeing the session.
Note: at the moment, since session_alloc_buffers() is not used, no task
is subscribed to the wait queue.
This patch introduces session_alloc_recv_buffer(), session_alloc_buffers()
and session_release_buffers() whose purpose will be to allocate missing
buffers and release unneeded ones around the process_session() and during
I/O operations.
I/O callbacks only need a single buffer for recv operations and none
for send. However we still want to ensure that we don't pick the last
buffer. That's what session_alloc_recv_buffer() is for.
This allocator is atomic in that it always ensures we can get 2 buffers
or fails. Here, if any of the buffers is not ready and cannot be
allocated, the operation is cancelled. The purpose is to guarantee that
we don't enter into the deadlock where all buffers are allocated by the
same size of all sessions.
A queue will have to be implemented for failed allocations. For now
they're just reported as failures.
We'll soon want to release buffers together upon failure so we need to
allocate them after the channels. Let's change this now. There's no
impact on the behaviour, only the error path is unrolled slightly
differently. The same was done in peers.
We don't call pool_free2(pool2_buffers) anymore, we only call b_free()
to do the job. This ensures that we can start to centralize the releasing
of buffers.
It's not clean to initialize the buffer before the channel since it
dereferences one pointer in the channel. Also we'll want to let the
channel pre-initialize the buffer, so let's ensure that the channel
is always initialized prior to the buffers.
b_alloc() now allocates a buffer and initializes it to the size specified
in the pool minus the size of the struct buffer itself. This ensures that
callers do not need to care about buffer details anymore. Also this never
applies memory poisonning, which is slow and useless on buffers.
We'll soon need to be able to switch buffers without touching the
channel, so let's move buffer initialization out of channel_init().
We had the same in compressoin.c.
Since embryonic sessions were introduced in 1.5-dev12 with commit
2542b53 ("MAJOR: session: introduce embryonic sessions"), a major
bug remained present. If haproxy cannot allocate memory during
session_complete() (for example, no more buffers), it will not
unlink the new session from the sessions list. This will cause
memory corruptions if the memory area from the session is reused
for anything else, and may also cause bogus output on "show sess"
on the CLI.
This fix must be backported to 1.5.
This patch makes it possible to create binds and servers in separate
namespaces. This can be used to proxy between multiple completely independent
virtual networks (with possibly overlapping IP addresses) and a
non-namespace-aware proxy implementation that supports the proxy protocol (v2).
The setup is something like this:
net1 on VLAN 1 (namespace 1) -\
net2 on VLAN 2 (namespace 2) -- haproxy ==== proxy (namespace 0)
net3 on VLAN 3 (namespace 3) -/
The proxy is configured to make server connections through haproxy and sending
the expected source/target addresses to haproxy using the proxy protocol.
The network namespace setup on the haproxy node is something like this:
= 8< =
$ cat setup.sh
ip netns add 1
ip link add link eth1 type vlan id 1
ip link set eth1.1 netns 1
ip netns exec 1 ip addr add 192.168.91.2/24 dev eth1.1
ip netns exec 1 ip link set eth1.$id up
...
= 8< =
= 8< =
$ cat haproxy.cfg
frontend clients
bind 127.0.0.1:50022 namespace 1 transparent
default_backend scb
backend server
mode tcp
server server1 192.168.122.4:2222 namespace 2 send-proxy-v2
= 8< =
A bind line creates the listener in the specified namespace, and connections
originating from that listener also have their network namespace set to
that of the listener.
A server line either forces the connection to be made in a specified
namespace or may use the namespace from the client-side connection if that
was set.
For more documentation please read the documentation included in the patch
itself.
Signed-off-by: KOVACS Tamas <ktamas@balabit.com>
Signed-off-by: Sarkozi Laszlo <laszlo.sarkozi@balabit.com>
Signed-off-by: KOVACS Krisztian <hidden@balabit.com>
We currently release all pools when a proxy is stopped, except the
connection, pendconn, and pipe pools. Doing so can improve further
reduce memory usage of old processes, eventhough the connection struct
is quite small, but there are a lot and they can participate to memory
fragmentation. The pipe pool is very small and limited, and not exported
so it's not done here.
stktable_fetch_key() does not indicate whether it returns NULL because
the input sample was not found or because it's unstable. It causes trouble
with track-sc* rules. Just like with sample_fetch_string(), we want it to
be able to give more information to the caller about what it found. Thus,
now we use the pointer to a sample passed by the caller, and fill it with
the information we have about the sample. That way, even if we return NULL,
the caller has the ability to check whether a sample was found and if it is
still changing or not.
This is the 3rd regression caused by the changes below. The latest to
date was reported by Finn Arne Gangstad. If a server responds with no
content-length and the client's FIN is never received, either we leak
the client-side FD or we spin at 100% CPU if timeout client-fin is set.
Enough is enough. The amount of tricks needed to cover these side-effects
starts to look like used toilet paper stacked over a chocolate cake. I
don't want to eat that cake anymore!
All this to avoid reporting a server-side timeout when a client stops
uploading data and haproxy expires faster than the server... A lot of
"ifs" resulting in a technically valid log that doesn't always please
users, and whose alternative causes that many issues for all others
users.
So let's revert this crap merged since 1.5-dev25 :
Revert "CLEANUP: http: don't clear CF_READ_NOEXP twice"
This reverts commit 1592d1e72a.
Revert "BUG/MEDIUM: http: clear CF_READ_NOEXP when preparing a new transaction"
This reverts commit 77d29029af.
Revert "BUG/MEDIUM: session: don't clear CF_READ_NOEXP if analysers are not called"
This reverts commit 0943757a21.
Revert "BUG/MEDIUM: http: disable server-side expiration until client has sent the body"
This reverts commit 3bed5e9337.
Revert "BUG/MEDIUM: http: correctly report request body timeouts"
This reverts commit b9edf8fbec.
Revert "BUG/MEDIUM: http/session: disable client-side expiration only after body"
This reverts commit b1982e27aa.
If a cleaner AND SAFER way to do something equivalent in 1.6-dev, we *might*
consider backporting it to 1.5, but given the vicious bugs that have surfaced
since, I doubt it will happen any time soon.
Fortunately, that crap never made it into 1.4 so no backport is needed.
Using the last rate counters, we now compute the queue, connect, response
and total times per server and per backend with a 95% accuracy over the last
1024 samples. The operation is cheap so we don't need to condition it.
As discussed with Dmitry Sivachenko, is a server farm has more than one
active server, uses a guaranteed non-determinist algorithm (round robin),
and a connection was initiated from a non-persistent connection, there's
no point insisting to reconnect to the same server after a connect failure,
better redispatch upon the very first retry instead of insisting on the same
server multiple times.
The retry delay is only useful when sticking to a same server. During
a redispatch, it's useless and counter-productive if we're sure to
switch to another server, which is almost guaranteed when there's
more than one server and the balancing algorithm is round robin, so
better not pass via the turn-around state in this case. It could be
done as well for leastconn, but there's a risk of always killing the
delay after the recovery of a server in a farm where it's almost
guaranteed to take most incoming traffic. So better only kill the
delay when using round robin.
As discussed with Dmitry Sivachenko, the default 1-second connect retry
delay can be large for situations where the connect timeout is much smaller,
because it means that an active connection reject will take more time to be
retried than a silent drop, and that does not make sense.
This patch changes this so that the retry delay is the minimum of 1 second
and the connect timeout. That way people running with sub-second connect
timeout will benefit from the shorter reconnect.
Servers used to have 3 flags to store a state, now they have 4 states
instead. This avoids lots of confusion for the 4 remaining undefined
states.
The encoding from the previous to the new states can be represented
this way :
SRV_STF_RUNNING
| SRV_STF_GOINGDOWN
| | SRV_STF_WARMINGUP
| | |
0 x x SRV_ST_STOPPED
1 0 0 SRV_ST_RUNNING
1 0 1 SRV_ST_STARTING
1 1 x SRV_ST_STOPPING
Note that the case where all bits were set used to exist and was randomly
dealt with. For example, the task was not stopped, the throttle value was
still updated and reported in the stats and in the http_server_state header.
It was the same if the server was stopped by the agent or for maintenance.
It's worth noting that the internal function names are still quite confusing.
Till now, the server's state and flags were all saved as a single bit
field. It causes some difficulties because we'd like to have an enum
for the state and separate flags.
This commit starts by splitting them in two distinct fields. The first
one is srv->state (with its counter-part srv->prev_state) which are now
enums, but which still contain bits (SRV_STF_*).
The flags now lie in their own field (srv->flags).
The function srv_is_usable() was updated to use the enum as input, since
it already used to deal only with the state.
Note that currently, the maintenance mode is still in the state for
simplicity, but it must move as well.
As more or less suspected, commit b1982e2 ("BUG/MEDIUM: http/session:
disable client-side expiration only after body") was hazardous. It
introduced a regression causing client side timeout to expire during
connection retries if it's lower than the time needed to cover the
amount of retries, so clients get a 408 when the connection to the
server fails to establish fast enough.
The reason is that the CF_READ_NOEXP flag is set after the MSG_DONE state
is reached, which protects the timeout from being re-armed, then during
the retries, process_session() clears the flag without calling the analyser
(since there's no activity for it), so the timeouts are rearmed.
Ideally, these one-shot flags should be per-analyser, and the analyser
which sets them would be responsible for clearing them, or they would
automatically be cleared when switching to another analyser. Unfortunately
this is not really possible currently.
What can be done however is to only clear them in the following situations :
- we're going to call analysers
- analysers have all been unsubscribed
This method seems reliable enough and approaches the ideal case well enough.
No backport is needed, this bug was introduced in 1.5-dev25.
Long-lived sessions are often subject to half-closed sessions resulting in
a lot of sessions appearing in FIN_WAIT state in the system tables, and no
way for haproxy to get rid of them. This typically happens because clients
suddenly disconnect without sending any packet (eg: FIN or RST was lost in
the path), and while the server detects this using an applicative heart
beat, haproxy does not close the connection.
This patch adds two new timeouts : "timeout client-fin" and
"timeout server-fin". The former allows one to override the client-facing
timeout when a FIN has been received or sent. The latter does the same for
server-facing connections, which is less useful.
John-Paul Bader reported a nasty segv which happens after a few hours
when SSL is enabled under a high load. Fortunately he could catch a
stack trace, systematically looking like this one :
(gdb) bt full
level = 6
conn = (struct connection *) 0x0
err_msg = <value optimized out>
s = (struct session *) 0x80337f800
conn = <value optimized out>
flags = 41997063
new_updt = <value optimized out>
old_updt = 1
e = <value optimized out>
status = 0
fd = 53999616
nbfd = 279
wait_time = <value optimized out>
updt_idx = <value optimized out>
en = <value optimized out>
eo = <value optimized out>
count = 78
sr = <value optimized out>
sw = <value optimized out>
rn = <value optimized out>
wn = <value optimized out>
The variable "flags" in conn_fd_handler() holds a copy of connection->flags
when entering the function. These flags indicate 41997063 = 0x0280d307 :
- {SOCK,DATA,CURR}_RD_ENA=1 => it's a handshake, waiting for reading
- {SOCK,DATA,CURR}_WR_ENA=0 => no need for writing
- CTRL_READY=1 => FD is still allocated
- XPRT_READY=1 => transport layer is initialized
- ADDR_FROM_SET=1, ADDR_TO_SET=0 => clearly it's a frontend connection
- INIT_DATA=1, WAKE_DATA=1 => processing a handshake (ssl I guess)
- {DATA,SOCK}_{RD,WR}_SH=0 => no shutdown
- ERROR=0, CONNECTED=0 => handshake not completed yet
- WAIT_L4_CONN=0 => normal
- WAIT_L6_CONN=1 => waiting for an L6 handshake to complete
- SSL_WAIT_HS=1 => the pending handshake is an SSL handshake
So this is a handshake is in progress. And the only way to reach line 88
is for the handshake to complete without error. So we know for sure that
ssl_sock_handshake() was called and completed the handshake then removed
the CO_FL_SSL_WAIT_HS flag from the connection. With these flags,
ssl_sock_handshake() does only call SSL_do_handshake() and retruns. So
that means that the problem is necessarily in data->init().
The fd is wrong as reported but is simply mis-decoded as it's the lower
half of the last function pointer.
What happens in practice is that there's an issue with the way we deal
with embryonic sessions during their conversion to regular sessions.
Since they have no stream interface at the beginning, the pointer to
the connection is temporarily stored into s->target. Then during their
conversion, the first stream interface is properly initialized and the
connection is attached to it, then s->target is set to NULL.
The problem is that if anything fails in session_complete(), the
session is left in this intermediate state where s->target is NULL,
and kill_mini_session() is called afterwards to perform the cleanup.
It needs the connection, that it finds in s->target which is NULL,
dereferences it and dies. The only reasons for dying here are a problem
on the TCP connection when doing the setsockopt(TCP_NODELAY) or a
memory allocation issue.
This patch implements a solution consisting in restoring s->target in
session_complete() on the error path. That way embryonic sessions that
were valid before calling it are still valid after.
The bug was introduced in 1.5-dev20 by commit f8a49ea ("MEDIUM: session:
attach incoming connection to target on embryonic sessions"). No backport
is needed.
Special thanks to John for his numerous tests and traces.
For a very long time, back in the v1.3 days, we used to rely on a trick
to avoid expiring the client side while transferring a payload to the
server. The problem was that if a client was able to quickly fill the
buffers, and these buffers took some time to reach the server, the
client should not expire while not sending anything.
In order to cover this situation, the client-side timeout was disabled
once the connection to the server was OK, since it implied that we would
at least expire on the server if required.
But there is a drawback to this : if a client stops uploading data before
the end, its timeout is not enforced and we only expire on the server's
timeout, so the logs report a 504.
Since 1.4, we have message body analysers which ensure that we know whether
all the expected data was received or not (HTTP_MSG_DATA or HTTP_MSG_DONE).
So we can fix this problem by disabling the client-side or server-side
timeout at the end of the transfer for the respective side instead of
having it unconditionally in session.c during all the transfer.
With this, the logs now report the correct side for the timeout. Note that
this patch is not enough, because another issue remains : the HTTP body
forwarders do not abort upon timeout, they simply rely on the generic
handling from session.c. So for now, the session is still aborted when
reaching the server timeout, but the culprit is properly reported. A
subsequent patch will address this specific point.
This bug was tagged MEDIUM because of the changes performed. The issue
it fixes is minor however. After some cooling down, it may be backported
to 1.4.
It was reported by and discussed with Rachel Chavez and Patrick Hemmer
on the mailing list.
