Now that we can wake one thread sleeping in the poller, we don't have to
use THREAD_WANT_SYNC any more.
This gives a significant performance boost on highly contended accesses
(servers with maxconn 1), showing a jump from 21k to 31k conn/s on a
test involving 8 threads.
This lock was necessary to manipulate the pendconn element between
concurrent places, but was causing great difficulties in the list walk
by having to iterate over multiple entries instead of being able to
safely pick the first one (in fact the first element was always the
right one but the locking model was hard to prove).
Here since we know we can always rely on the queue's locks, we take
the queue's lock every time we need to modify the element. In practice
it was already the case everywhere except in pendconn_dequeue() which
only works on an element that was already detached. This function had
to be protected against the risk of meeting an incompletely detached
element (which could be unlinked but not yet assigned). By taking the
queue lock around the LIST_ISEMPTY test, it's enough to ensure that a
concurrent thread either didn't begin or had completed the operation.
The true benefit really is in pendconn_process_next_strm() where we
can again safely work with the first element of each queue. This will
significantly simplify next updates to this code.
The new pendconn_queue_lock() and pendconn_queue_unlock() functions are
made to make it more convenient to lock or unlock the pendconn queue
either at the proxy or the server depending on pendconn->srv. This way
it is possible to remove the open-coding of these locks at various places.
These ones have been used in pendconn_unlink() and pendconn_add(), thus
significantly simplifying the logic there.
The pendconn struct uses ->px and ->srv to designate where the element is
queued. There is something confusing regarding threads though, because we
have to lock the appropriate queue before inserting/removing elements, and
this queue may only be determined by looking at ->srv (if it's not NULL
it's the server, otherwise use the proxy). But pendconn_grab_from_px() and
pendconn_process_next_strm() both assign this ->srv field, making it
complicated to know what queue to lock before manipulating the element,
which is exactly why we have the pendconn_lock in the first place.
This commit introduces pendconn->target which is the target server that
the two aforementioned functions will set when assigning the server.
Thanks to this, the server pointer may always be relied on to determine
what queue to use.
pendconn_add() used to assign strm->pend_pos very late, after unlocking
the queue, so that a watching thread could see a random value in
pendconn->strm->pend_pos even while holding the lock on the element and
the queue itself. While there's currently nothing wrong with this, it
costs nothing to arrange it and will simplify code analysis later.
Now pendconn_free() takes a stream, checks that pend_pos is set, clears
it, and uses pendconn_unlink() to complete the job. It's cleaner and
centralizes all the bookkeeping work in pendconn_unlink() only and
ensures that there's a single place where the stream's position in the
queue is manipulated.
For now the pendconns may be dequeued at two places :
- pendconn_unlink(), which operates on a locked queue
- pendconn_free(), which operates on an unlocked queue and frees
everything.
Some changes are coming to the queue and we'll need to be able to be a
bit stricter regarding the places where we dequeue to keep the accounting
accurate. This first step renames the locked function __pendconn_unlink()
as it's for use by those aware of it, and introduces a new general purpose
pendconn_unlink() function which automatically grabs the necessary locks
before calling the former, and pendconn_cond_unlink() which additionally
checks the pointer and the presence in the queue.
src/queue.o: In function `pendconn_redistribute':
/home/ilia/haproxy/src/queue.c:272: undefined reference to `thread_want_sync'
src/queue.o: In function `pendconn_grab_from_px':
/home/ilia/haproxy/src/queue.c:311: undefined reference to `thread_want_sync'
src/queue.o: In function `process_srv_queue':
/home/ilia/haproxy/src/queue.c:184: undefined reference to `thread_want_sync'
collect2: error: ld returned 1 exit status
make: *** [Makefile:900: haproxy] Error 1
To be backported to 1.8.
The previous patch about queues (5cd4bbd7a "BUG/MAJOR: threads/queue: Fix
thread-safety issues on the queues management") revealed a performance drop when
multithreading is enabled (nbthread > 1). This happens when pending connections
handled by other theads are dequeued. If these other threads are blocked in the
poller, we have to wait the poller's timeout (or any I/O event) to process the
dequeued connections.
To fix the problem, at least temporarly, we "wake up" the threads by requesting
a synchronization. This may seem a bit overkill to use the sync point to do a
wakeup on threads, but it fixes this performance issue. So we can now think
calmly on the good way to address this kind of issues.
This patch should be backported in 1.8 with the commit 5cd4bbd7a ("BUG/MAJOR:
threads/queue: Fix thread-safety issues on the queues management").
The management of the servers and the proxies queues was not thread-safe at
all. First, the accesses to <strm>->pend_pos were not protected. So it was
possible to release it on a thread (for instance because the stream is released)
and to use it in same time on another one (because we redispatch pending
connections for a server). Then, the accesses to stream's information (flags and
target) from anywhere is forbidden. To be safe, The stream's state must always
be updated in the context of process_stream.
So to fix these issues, the queue module has been refactored. A lock has been
added in the pendconn structure. And now, when we try to dequeue a pending
connection, we start by unlinking it from the server/proxy queue and we wake up
the stream. Then, it is the stream reponsibility to really dequeue it (or
release it). This way, we are sure that only the stream can create and release
its <pend_pos> field.
However, be careful. This new implementation should be thread-safe
(hopefully...). But it is not optimal and in some situations, it could be really
slower in multi-threaded mode than in single-threaded one. The problem is that,
when we try to dequeue pending connections, we process it from the older one to
the newer one independently to the thread's affinity. So we need to wait the
other threads' wakeup to really process them. If threads are blocked in the
poller, this will add a significant latency. This problem happens when maxconn
values are very low.
This patch must be backported in 1.8.
pendconn_get_next_strm() is called from process_srv_queue() under the
server lock, and calls stream_add_srv_conn() with this lock held, while
the latter tries to take it again. This results in a deadlock when
a server's maxconn is reached and haproxy is built with thread support.
During the migration to the second version of the pools, the new
functions and pool pointers were all called "pool_something2()" and
"pool2_something". Now there's no more pool v1 code and it's a real
pain to still have to deal with this. Let's clean this up now by
removing the "2" everywhere, and by renaming the pool heads
"pool_head_something".
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.
The server state and weight was reworked to handle
"pending" values updated by checks/CLI/LUA/agent.
These values are commited to be propagated to the
LB stack.
In further dev related to multi-thread, the commit
will be handled into a sync point.
Pending values are named using the prefix 'next_'
Current values used by the LB stack are named 'cur_'
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.
srv_is_usable() is broader than srv_is_usable() as it not only considers
the weight but the server's state as well. Future changes will allow a
server to be in drain mode with a non-zero weight, so we should migrate
to use that function instead.
Checks.c has become a total mess. A number of proxy or server maintenance
and queue management functions were put there probably because they were
used there, but that makes the code untouchable. And that's without saying
that their names does not always relate to what they really do!
So let's do a first pass by moving these ones :
- set_backend_down() => backend.c
- redistribute_pending() => queue.c:pendconn_redistribute()
- check_for_pending() => queue.c:pendconn_grab_from_px()
- shutdown_sessions => server.c:srv_shutdown_sessions()
- shutdown_backup_sessions => server.c:srv_shutdown_backup_sessions()
All of them were moved at once.
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.
We used to call srv_is_usable() with either the current state and weights
or the previous ones. This causes trouble for future changes, so let's first
split it in two variants :
- srv_is_usable(srv) considers the current status
- srv_was_usable(srv) considers the previous status
Instead of storing a couple of (int, ptr) in the struct connection
and the struct session, we use a different method : we only store a
pointer to an integer which is stored inside the target object and
which contains a unique type identifier. That way, the pointer allows
us to retrieve the object type (by dereferencing it) and the object's
address (by computing the displacement in the target structure). The
NULL pointer always corresponds to OBJ_TYPE_NONE.
This reduces the size of the connection and session structs. It also
simplifies target assignment and compare.
In order to improve the generated code, we try to put the obj_type
element at the beginning of all the structs (listener, server, proxy,
si_applet), so that the original and target pointers are always equal.
A lot of code was touched by massive replaces, but the changes are not
that important.
It was reported that a server configured with a zero weight would
sometimes still take connections from the backend queue. This issue is
real, it happens this way :
1) the disabled server accepts a request with a cookie
2) many cookie-less requests accumulate in the backend queue
3) when the disabled server completes its request, it checks its own
queue and the backend's queue
4) the server takes a pending request from the backend queue and
processes it. In response, the server's cookie is assigned to
the client, which ensures that some requests will continue to
be served by this server, leading back to point 1 above.
The fix consists in preventing a zero-weight server from dequeuing pending
requests from the backend. Making use of srv_is_usable() in such tests makes
the tests more robust against future changes.
This fix must be backported to 1.4 and 1.3.
When reading the code, the "tracked" member of a server makes one
think the server is tracked while it's the opposite, it's a pointer
to the server being tracked. This is particularly true in constructs
such as :
if (srv->tracked) {
Since it's the second time I get caught misunderstanding it, let's
rename it to "track" to avoid the confusion.
The motivation for this is to allow iteration of all the connections
of a server without the expense of iterating over the global list
of connections.
The first use of this will be to implement an option to close connections
associated with a server when is is marked as being down or in maintenance
mode.
It's very annoying that frontend and backend stats are merged because we
don't know what we're observing. For instance, if a "listen" instance
makes use of a distinct backend, it's impossible to know what the bytes_out
means.
Some points take care of not updating counters twice if the backend points
to the frontend, indicating a "listen" instance. The thing becomes more
complex when we try to add support for server side keep-alive, because we
have to maintain a pointer to the backend used for last request, and to
update its stats. But we can't perform such comparisons anymore because
the counters will not match anymore.
So in order to get rid of this situation, let's have both frontend AND
backend stats in the "struct proxy". We simply update the relevant ones
during activity. Some of them are only accounted for in the backend,
while others are just for frontend. Maybe we can improve a bit on that
later, but the essential part is that those counters now reflect what
they really mean.
This one has been removed and is now totally superseded by ->target.
To get the server, one must use target_srv(&s->target) instead of
s->srv now.
The function ensures that non-server targets still return NULL.
When dealing with HTTP keep-alive, we'll have to know if we can reuse
an existing connection. For that, we'll have to check if the current
connection was made on the exact same target (referenced in the stream
interface).
Thus, we need to first assign the next target to the session, then
copy it to the stream interface upon connect(). Later we'll check for
equivalence between those two operations.
If a server is disabled or tracking a disabled server, it must not
dequeue requests pending in the proxy queue, it must only dequeue
its own ones.
The problem that was caused is that if a backend always had requests
in its queue, a disabled server would continue to take traffic forever.
(was commit 09d02aaf02d1f21c0c02672888f3a36a14bdd299 in 1.4)
There are a few remaining max values that need to move to counters.
Also, the counters are more often used than some config information,
so get them closer to the other useful struct members for better cache
efficiency.
Kai Krueger found that previous patch was incomplete, because there is
an unconditionnal call to process_srv_queue() in session_free() which
still causes a dead server to consume pending connections from the
backend.
This call was made unconditionnal so that we don't leave unserved
connections in the server queue, for instance connections coming
in with "option persist" which can bypass the server status check.
However, the server must not touch the backend's queue if it is down.
Another fear was that some connections might remain unserved when
the server is using a dynamic maxconn if the number of connections
to the backend is too low. Right now, srv_dynamic_maxconn() ensures
this cannot happen, so the call can remain conditionnal.
The fix consists in allowing a server to process it own queue whatever
its state, but not to touch the backend's queue if it is down. Its
queue should normally be empty when the server is down because it is
redistributed when the server goes down. The only remaining cases are
precisely the persistent connections with "option persist" set, coming
in after the queue has been redispatched. Those ones must still be
processed when a connection terminates.
(cherry picked from commit cd485c4480)
srv_dynamic_maxconn() is clearly documented as returning at least 1
possible connection under throttling. But the computation was wrong,
the minimum 1 was divided and got lost in case of very low maxconns.
Apply the MAX(1, max) before returning the result in order to ensure
that a newly appeared server will get some traffic.
(cherry picked from commit 819970098f)
It's very frequent to require some information about the
reason why a task is running. Some flags have been added
so that a task now knows if it got woken up due to I/O
completion, timeout, etc...
It should be stated as a rule that a C file should never
include types/xxx.h when proto/xxx.h exists, as it gives
less exposure to declaration conflicts (one of which was
caught and fixed here) and it complicates the file headers
for nothing.
Only types/global.h, types/capture.h and types/polling.h
have been found to be valid includes from C files.
The dequeuing logic was completely wrong. First, a task was assigned
to all servers to process the queue, but this task was never scheduled
and was only woken up on session free. Second, there was no reservation
of server entries when a task was assigned a server. This means that
as long as the task was not connected to the server, its presence was
not accounted for. This was causing trouble when detecting whether or
not a server had reached maxconn. Third, during a redispatch, a session
could lose its place at the server's and get blocked because another
session at the same moment would have stolen the entry. Fourth, the
redispatch option did not work when maxqueue was reached for a server,
and it was not possible to do so without indefinitely hanging a session.
The root cause of all those problems was the lack of pre-reservation of
connections at the server's, and the lack of tracking of servers during
a redispatch. Everything relied on combinations of flags which could
appear similarly in quite distinct situations.
This patch is a major rework but there was no other solution, as the
internal logic was deeply flawed. The resulting code is cleaner, more
understandable, uses less magics and is overall more robust.
As an added bonus, "option redispatch" now works when maxqueue has
been reached on a server.
The reported queue position in the logs was 0 for the first pending request
in the queue, which is wrong because it means that one request will have to
be completed before the queued one may execute. It caused the undesired side
effect that 0/0 was reported when either 0 or 1 request was pending in the
queue. Thus, we have to increment the queue size before reporting the value.
When a server terminates a connection, the next session in its
own queue was immediately processed. Because of this, if all
server queues are always filled, then no new anonymous request
will be processed. Consider oldest request between global and
server queues to choose from which to pick the request.
An improvement over this will consist in adding a configurable
offset when comparing expiration dates, so that cookie-less
requests can get either less or more priority.
The new 'slowstart' parameter for a server accepts a value in
milliseconds which indicates after how long a server which has
just come back up will run at full speed. The speed grows
linearly from 0 to 100% during this time. The limitation applies
to two parameters :
- maxconn: the number of connections accepted by the server
will grow from 1 to 100% of the usual dynamic limit defined
by (minconn,maxconn,fullconn).
- weight: when the backend uses a dynamic weighted algorithm,
the weight grows linearly from 1 to 100%. In this case, the
weight is updated at every health-check. For this reason, it
is important that the 'inter' parameter is smaller than the
'slowstart', in order to maximize the number of steps.
The slowstart never applies when haproxy starts, otherwise it
would cause trouble to running servers. It only applies when
a server has been previously seen as failed.
The timeout functions were difficult to manipulate because they were
rounding results to the millisecond. Thus, it was difficult to compare
and to check what expired and what did not. Also, the comparison
functions were heavy with multiplies and divides by 1000. Now, all
timeouts are stored in timevals, reducing the number of operations
for updates and leading to cleaner and more efficient code.
The rbtree-based wait queue consumes a lot of CPU. Use the ul2tree
instead. Lots of cleanups and code reorganizations made it possible
to reduce the task struct and simplify the code a bit.
The fiprm and beprm were added to ease the transition between
a single listener mode to frontends+backends. They are no longer
needed and make the code a bit more complicated. Remove them.
The maxconn argument is used only for the listeners, and the
fullconn is used only for the backends. If unset, it inherits
maxconn's value which itself can inherit the default or the
global value (we might need to change this).
The nbconn attribute in the proxies was not relevant anymore because
a frontend A may use backend B and both of them must account for their
respective connections. For this reason, there now are two separate
counters for frontend and backend connections.
The stats page has been updated to reflect the backend, but a separate
line entry for the frontend with error counts would be good.
Note that as of now, beconn may be higher than maxconn, because maxconn
applies to the frontend, while beconn may be increased due to sessions
passed from another frontend.
There was a confusion about the way to find filters and backend
parameters from sessions. The chaining has been changed between
the session and the proxy.
Now, a session knows only two proxies : one frontend (->fe) and
one backend (->be). Each proxy has a link to the proxy providing
filters and to the proxy providing backend parameters (both self
by default).
The captures (cookies and headers) have been attached to the
frontend's filters for now.
The uri_auth and the statistics are attached to the backend's
filters so that the uri can depend on a hostname for instance.
The references to the proxy from the session have been turned into
Frontend (fe), Filters (fi) and Backend (be). This should ease the
migration to the L7 switching features. Next step will be to kill
the struct proxy and have 3 independant structs instead, each
referenced from entities called listener, frontend, filters and
backend.
As suggested by Markus Elfring, a few "const char *" have replaced
some "char *" declarations where a function is not expected to
modify a value. It does not change the code but it helps detecting
coding errors.
The files are now stored under :
- include/haproxy for the generic includes
- include/types.h for the structures needed within prototypes
- include/proto.h for function prototypes and inline functions
- src/*.c for the C files
Most include files are now covered by LGPL. A last move still needs
to be done to put inline functions under GPL and not LGPL.
Version has been set to 1.3.0 in the code but some control still
needs to be done before releasing.
