As specified in the function comment, the function h1_skip_chunk_crlf() must not
change anything and return zero if not enough data are available. This must
include the case where there is no data at all. On this point, it must do the
same that other h1 parsing functions. This bug is made visible since the commit
91f77d599 ("BUG/MINOR: mux-h1: Process input even if the input buffer is
empty").
This patch must be backported to 1.9.
As by default we add all keepalive connections to the idle pool, if we run
into a pathological case, where all client don't do keepalive, but the server
does, and haproxy is configured to only reuse "safe" connections, we will
soon find ourself having lots of idling, unusable for new sessions, connections,
while we won't have any file descriptors available to create new connections.
To fix this, add 2 new global settings, "pool_low_ratio" and "pool_high_ratio".
pool-low-fd-ratio is the % of fds we're allowed to use (against the maximum
number of fds available to haproxy) before we stop adding connections to the
idle pool, and destroy them instead. The default is 20. pool-high-fd-ratio is
the % of fds we're allowed to use (against the maximum number of fds available
to haproxy) before we start killing idling connection in the event we have to
create a new outgoing connection, and no reuse is possible. The default is 25.
For most of the xprt methods, provide a xprt_ctx. This will be useful later
when we'll want to be able to stack xprts.
The init() method now has to create and provide the said xprt_ctx if needed.
task_delete() was never used without calling task_free() just after, and
task_free() was only used on error pathes to destroy a just-created task,
so merge them into task_destroy(), that will remove the task from the
wait queue, and make sure the task is either destroyed immediately if it's
not in the run queue, or destroyed when it's supposed to run.
The wakeup condition in task_wakeup() is redundant as it is already
validated by the CAS. Better move the __task_wakeup() call there, it
also has the merit of being easier to audit this way. This also reduces
the code size by around 1.8 kB :
$ size haproxy-?
text data bss dec hex filename
2153806 100208 1307676 3561690 3658da haproxy-1
2152094 100208 1307676 3559978 36522a haproxy-2
Commit 0c7a4b6 ("MINOR: tasks: Don't set the TASK_RUNNING flag when
adding in the tasklet list.") revealed a hole in the way tasks may
be freed : they could be removed while in the run queue when the
TASK_QUEUED flag was present but not the TASK_RUNNING one. But it
seems the issue was emphasized by commit cde7902 ("MEDIUM: tasks:
improve fairness between the local and global queues") though the
code it replaces was already affected given how late the TASK_RUNNING
flag was set after removal from the global queue.
At the moment the task is picked from the global run queue, if it
is the last one, the global run queue lock is dropped, and then
the TASK_RUNNING flag was added. In the mean time another thread
might have performed a task_free(), and immediately after, the
TASK_RUNNING flag was re-added to the task, which was then added
to the tasklet list. The unprotected window was extremely faint
but does definitely exist and inconsistent task lists have been
observed a few times during very intensive tests over the last few
days. From this point various options are possible, the task might
have been re-allocated while running, and assigned state 0 and/or
state QUEUED while it was still running, resulting in the tast not
being put back into the tree.
This commit simply makes sure that tests on TASK_RUNNING before removing
the task also cover TASK_QUEUED.
It must be backported to 1.9 along with the previous ones touching
that area.
Now that we have the warranty that a task won't be added in the runqueue
while the TASK_QUEUED or the TASK_RUNNING flag is set, don't bother trying
to lock the task by setting leaf_p to 0x1 while inserting it in the runqueue
or having it in the tasklet_list, as nobody else will attempt to add it.
Make sure we set TASK_QUEUED in every case before adding the task to the
run queue. task_wakeup() now checks if either TASK_QUEUED or TASK_RUNNING
is set, and if neither is set, add TASK_QUEUED and effectively add the task
to the runqueue.
No longer use __task_wakeup() anywhere except in task_wakeup(), always use
task_wakeup() instead.
With the old code, process_runnable_task() may re-add a task in the runqueue
without setting the TASK_QUEUED flag, and there were race conditions that could
lead to a task having the TASK_QUEUED flag but not in the runqueue, thus
being unschedulable.
This should be backported to 1.9.
As for smp_prefetch_http(), there is now a way to successfully perform a
prefetch in HTX, even if the message forwarding already begun. It is used for
the sample fetches "req.proto_http" and "method".
This patch must be backported to 1.9.
All HTTP samples are buggy because the channel tested in the prefetch functions
(HTX and legacy HTTP) is chosen depending on the sample direction and not the
keyword really used. It means the request channel is used if the sample is
called during the request analysis and the response channel is used if it is
called during the response analysis, regardless the sample really called. For
instance, if you use the sample "req.ver" in an http-response rule, the response
channel will be prefeched because it is called during the response analysis,
while the request channel should have been used instead. So some assumptions on
the validity of the sample may be made on the wrong channel. It is the first
bug.
Then the same error is done in some samples themselves. So fetches are performed
on the wrong channel. For instance, the header extraction (req.fhdr, res.fhdr,
req.hdr, res.hdr...). If the sample "req.hdr" is used in an http-response rule,
then the matching is done on the response headers and not the request ones. It
is the second bug.
Finally, the last one but not the least, in some samples, the right channel is
used. But because the prefetch was done on the wrong one, this channel may be in
a undefined state. For instance, using the sample "req.ver" in an http-response
rule leads to a matching on a posibility released buffer.
To fix all these bugs, the right channel is now chosen in sample fetches, before
the prefetch. If the same function is used to fetch requests and responses
elements, then the keyword is used to choose the right one. This channel is then
used by the functions smp_prefetch_htx() and smp_prefetch_http(). Of course, it
is also used by the samples themselves to extract information.
This patch must be backported to all supported versions. For version 1.8 and
priors, it must be totally refactored. First because there is no HTX into these
versions. Then the buffers API has changed in HAProxy 1.9. The files
http_fetch.{ch} doesn't exist on old versions.
It's always a pain to get a core dump when enabling user/group setting
(which disables the dumpable flag on Linux), when using a chroot and/or
when haproxy is started by a service management tool which requires
complex operations to just raise the core dump limit.
This patch introduces a new "set-dumpable" global directive to work
around these troubles by doing the following :
- remove file size limits (equivalent of ulimit -f unlimited)
- remove core size limits (equivalent of ulimit -c unlimited)
- mark the process dumpable again (equivalent of suid_dumpable=1)
Some of these will depend on the operating system. This way it becomes
much easier to retrieve a core file. Temporarily moving the chroot to
a user-writable place generally enough.
Since the introduction of the options field, we can use it to store the
type of process.
type = 'm' is replaced by PROC_O_TYPE_MASTER
type = 'w' is replaced by PROC_O_TYPE_WORKER
type = 'e' is replaced by PROC_O_TYPE_PROG
The old values are still used in the HAPROXY_PROCESSES environment
variable to pass the information during a reload.
This option is already the default, but its opposite 'no option
start-on-reload' allows the master to keep a previous instance of a
program and don't start a new one upon a reload.
The old program will then appear as a current one in "show proc" and
could also trigger an exit-on-failure upon a segfault.
Previously we were assuming than a process was in a leaving state when
its number of reload was greater than 0. With mworker programs it's not
the case anymore so we need to store a leaving state.
Maksim Kupriianov reported very strange crashes in fwrr_update_position()
which didn't make sense because of an apparent divide overflow except that
the value was not null in the core.
It happens that while the locking is correct in all the functions' call
graph, the uppermost one (fwrr_get_next_server()) incorrectly expected
that its target server was already locked when called. This stupid
assumption causd the server lock not to be held when calling the other
ones, explaining how it was possible to change the server's eweight by
calling srv_lb_commit_status() under the server lock yet collide with
its unprotected usage.
This commit makes sure that fwrr_get_server_from_group() retrieves a
locked server and that fwrr_get_next_server() is responsible for
unlocking the server before returning it. There is one subtlety in
this function which is that it builds a list of avoided servers that
were full while scanning the tree, and all of them are queued in a
full state so they must be unlocked upon return.
Many thanks to Maksim for providing detailed info allowing to narrow
down this bug.
This fix must be backported to 1.9. In 1.8 the lock seems much wider
and changes to the server's state are performed under the rendez-vous
point so this it doesn't seem possible that it happens there.
Implements "show peers [peers section]" new CLI command to dump information
about the peers and their stick-tables to be synchronized and others internal.
May be backported as far as 1.5.
Older compilers (like gcc-3.4) warn about the use of "const" on functions
returning a struct, which makes sense since the return may only be copied :
include/common/htx.h:233: warning: type qualifiers ignored on function return type
Let's simply drop "const" here.
Older compilers don't like to see "inline" placed after the type in a
function declaration, it must be "static inline <type>" only. This
patch touches various areas. The warnings were seen with gcc-3.4.
Move the definition of the various _HA_ATOMIC_* macros that use
__atomic_* in the #if GCC_VERSION >= 4.7, not just after it, so that we
can build with older versions of gcc again.
Whenever HAProxy was reloaded with rotated keys, the resumption would be
broken for previous encryption key. The bug was introduced with the addition
of 80 byte keys in 9e7547 (MINOR: ssl: add support of aes256 bits ticket keys
on file and cli.).
This fix needs to be backported to 1.9.
This one hasn't been used anymore since the scheduler changes after 1.8
but it kept being exported and maintained up to date while it's always
reset when scanning the trees. Let's stop exporting it and updating it.
The upgrade is performed when an H2 preface is detected when the first request
on a connection is parsed. The CS is destroyed by setting EOS flag on it. A
special flag is added on the HTX message to warn the HTX analyzers the stream
will be closed because of an upgrade. This way, no error and no log are
emitted. When the mux h1 is released, we create a mux h2, without any CS and
passing the buffer with the unparsed H2 preface.
This function will handle mux upgrades, for frontend connections only. It will
retrieve the best mux in the same way than conn_install_mux_fe except that the
mode and optionnally the proto are forced.
The new multiplexer is initialized using a new context and a specific input
buffer. Then, the old one is destroyed. If an error occurred, everything is
rolled back.
It is mandatory to handle mux upgrades, because during a mux upgrade, the
connection will be reassigned to another multiplexer. So when the old one is
destroyed, it does not own the connection anymore. Or in other words, conn->ctx
does not point to the old mux's context when its destroy() callback is
called. So we now rely on the multiplexer context do destroy it instead of the
connection.
In addition, h1_release() and h2_release() have also been updated in the same
way.
The mux's callback init() now take a pointer to a buffer as extra argument. It
must be used by the multiplexer as its input buffer. This buffer is always NULL
when a multiplexer is initialized with a fresh connection. But if a mux upgrade
is performed, it may be filled with existing data. Note that, for now, mux
upgrades are not supported. But this commit is mandatory to do so.
The flag SF_HTX has been added to know when a stream uses the HTX or not. It is
set when an HTX stream is created. There are 2 conditions to set it. The first
one is when the HTTP frontend enables the HTX. The second one is when the attached
conn_stream uses an HTX multiplexer.
A multiplexer must now set the flag MX_FL_HTX when it uses the HTX to structured
the data exchanged with channels. the muxes h1 and h2 set this flag. Of course,
for the mux h2, it is set on h2_htx_ops only.
The flag H1_MF_CLEAN_CONN_HDR has been added to let the H1 parser sanitize
connection headers. It means it will remove all "close" and "keep-alive" values
during the parsing. One noticeable effect is that connection headers may be
unfolded. In practice, this is not a problem because it is not frequent to have
multiple values for the connection headers.
If this flag is set, during the parsing The function
h1_parse_next_connection_header() is called in a loop instead of
h1_parse_conection_header().
No need to backport this patch
Because the option http-tunnel is now ignored in HTX, there is no longer any
need to adjust the transaction mode in HTX analyzers. A channel can still be
switch to the tunnel mode for legitimate cases (HTTP CONNECT or switching
protocols). So the function htx_adjust_conn_mode() is now useless.
This patch must be backported to 1.9. It is not strictly speaking required but
it will ease futur backports.
The "show activity" command reports the number of incoming connections
dispatched per thread but doesn't report the number of connections
received by each thread. It is important to be able to monitor this
value as it can show that for whatever reason a smaller set of threads
is receiving the connections and dispatching them to all other ones.
When creating a new initcall, don't forget to define the symbols, as it may
not be done automatically and that would lead to undefined symbols.
This should be backported to 1.9.
This patch implements the external binary support in the master worker.
To configure an external process, you need to use the program section,
for example:
program dataplane-api
command ./dataplane_api
Those processes are launched at the same time as the workers.
During a reload of HAProxy, those processes are dealing with the same
sequence as a worker:
- the master is re-executed
- the master sends a USR1 signal to the program
- the master launches a new instance of the program
During a stop, or restart, a SIGTERM is sent to the program.
The children variable is still used in haproxy, it is not required
anymore since we have the information about the current workers in the
mworker_proc linked list.
The oldpids array is also replaced by this linked list when we
generated the arguments for the master reexec.
I found on an (old) AIX 5.1 machine that stdint.h didn't exist while
inttypes.h which is expected to include it does exist and provides the
desired functionalities.
As explained here, stdint being just a subset of inttypes for use in
freestanding environments, it's probably always OK to switch to inttypes
instead:
https://pubs.opengroup.org/onlinepubs/009696799/basedefs/stdint.h.html
Also it's even clearer here in the autoconf doc :
https://www.gnu.org/software/autoconf/manual/autoconf-2.61/html_node/Header-Portability.html
"The C99 standard says that inttypes.h includes stdint.h, so there's
no need to include stdint.h separately in a standard environment.
Some implementations have inttypes.h but not stdint.h (e.g., Solaris
7), but we don't know of any implementation that has stdint.h but not
inttypes.h"
The current initcall implementation relies on dedicated sections (one
section per init stage) to store the initcall descriptors. Then upon
startup, these sections are scanned from beginning to end and all items
found there are called in sequence.
On platforms like AIX or Cygwin it seems difficult to figure the
beginning and end of sections as the linker doesn't seem to provide
the corresponding symbols. In order to replace this, this patch
simply implements an array of single linked (one per init stage)
which are fed using constructors for each register call. These
constructors are declared static, with a name depending on their
line number in the file, in order to avoid name clashes. The final
effect is the same, except that the method is slightly more expensive
in that it explicitly produces code to register these initcalls :
$ size haproxy.sections haproxy.constructor
text data bss dec hex filename
4060312 249176 1457652 5767140 57ffe4 haproxy.sections
4062862 260408 1457652 5780922 5835ba haproxy.constructor
This mechanism is enabled as an alternative to the default one when
build option USE_OBSOLETE_LINKER is set. This option is currently
enabled by default only on AIX and Cygwin, and may be attempted for
any target which fails to build complaining about missing symbols
__start_init_* and/or __stop_init_*.
Once confirmed as a reliable fix, this will likely have to be backported
to 1.9 where AIX and Cygwin do not build anymore.
Older Solaris and AIX versions do not have unsetenv(). This adds a
fairly simple implementation which scans the environment, for use
with those systems. It will simply require to pass the define in
the "DEFINE" macro at build time like this :
DEFINE="-Dunsetenv=my_unsetenv"
http_known_methods, HTTP_100 and HTTP_103 were not declared extern and
as such were multiply defined since they were in http.h. There was
apparently no more side effect but it may depend on the platform and
the linker.
This needs to be backported to 1.9.
Previous commit 3ea351368 ("BUG/MEDIUM: h2: Remove the tasklet from the
task list if unsubscribing.") uncovered an issue which needs to be
addressed in the scheduler's API. The function task_remove_from_task_list()
was initially designed to remove a task from the running tasklet list from
within the scheduler, and had to be used in h2 to abort pending I/O events.
However this function was not designed to be idempotent, occasionally
causing a double removal from the tasklet list, with the second doing
nothing but affecting the apparent tasks count and making haproxy use
100% CPU on some tests consisting in stopping the client during some
transfers. The h2_unsubscribe() function can sometimes be called upon
stream exit after an error where the tasklet was possibly already
removed, so it.
This patch does 2 things :
- it renames task_remove_from_task_list() to
__task_remove_from_tasklet_list() to discourage users from calling
it. Also note the fix in the naming since it's a tasklet list and
not a task list. This function is still uesd from the scheduler.
- it adds a new, idempotent, task_remove_from_tasklet_list() function
which does nothing if the task is already not in the tasklet list.
This patch will need to be backported where the commit above is backported.
The flag CF_EOI is now set on the input channel when the flag CS_FL_EOI is set
on the corresponding conn_stream. In addition, if a read activity is reported
when this flag is set, the stream is woken up.
This patch should be backported to 1.9.
Since the begining, in the H2 multiplexer, when the end of a message is reached,
the flag CS_FL_(R)EOS is set on the conn_stream to notify the upper layer that
all data were received and consumed and there is no longer any expected. The
stream-interface converts it into a shutdown read. But it leads to some
ambiguities with the real shutr. Once it was reported at the end of the message,
there is no way to report it when the read0 is received. For this reason, aborts
after the message was fully received cannot be reported. And on the channel
side, it is hard to make the difference between a shutr because the end of the
message was reached and a shutr because of an abort.
For these reasons, there is now a flag to mark the end of the message. It is
called CS_FL_EOI (end-of-input) because it is only used on the receipt path.
This flag is only declared and not used yet.
This patch will be used by future bug fixes and will have to be backported
to 1.9.
In the stats applet (in HTX and legacy HTTP), after a response is fully sent to
a client, the request is consumed. It is done at the end, after all the response
was copied into the channel's buffer. But only outgoing data at time the applet
is called are consumed. Then the applet is closed. If a request with a huge body
is sent, an error is triggerred because a SHUTW is catched for an unfinisehd
request.
Now, we consume request data until the end. In fact, we don't try to shutdown
the request's channel for write anymore.
This patch must be backported to 1.9 after some observation period. It should
probably be backported in prior versions too. But honnestly, with refactoring
on the connection layer and the stream interface in 1.9, it is probably safer
to not do so.
It's never easy to guess what services are built in. We currently have
the prometheus exporter in contrib/ which is the only extension for now.
Let's enumerate all available ones just like we do for filterr and pollers.
For conveniance, in HTTP muxes (h1 and h2), the end of the stream and the end of
the message are reported the same way to the stream, by setting the flag
CS_FL_EOS. In the stream-interface, when CS_FL_EOS is detected, a shutdown for
read is reported on the channel side. This is historical. With the legacy HTTP
layer, because the parsing is done by the stream in HTTP analyzers, the EOS
really means a shutdown for read.
Most of time, for muxes h1 and h2, it works pretty well, especially because the
keep-alive is handled by the muxes. The stream is only used for one
transaction. So mixing EOS and EOM is good enough. But not everytime. For now,
client aborts are only reported if it happens before the end of the request. It
is an error and it is properly handled. But because the EOS was already
reported, client aborts after the end of the request are silently
ignored. Eventually an error can be reported when the response is sent to the
client, if the sending fails. Otherwise, if the server does not reply fast
enough, an error is reported when the server timeout is reached. It is the
expected behaviour, excpect when the option abortonclose is set. In this case,
we must report an error when the client aborts. But as said before, this event
can be ignored. So to be short, for now, the abortonclose is broken.
In fact, it is a design problem and we have to rethink all channel's flags and
probably the conn-stream ones too. It is important to split EOS and EOM to not
loose information anymore. But it is not a small job and the refactoring will be
far from straightforward.
So for now, temporary flags are introduced. When the last read is received, the
flag CS_FL_READ_NULL is set on the conn-stream. This way, we can set the flag
SI_FL_READ_NULL on the stream interface. Both flags are persistant. And to be
sure to wake the stream, the event CF_READ_NULL is reported. So the stream will
always have the chance to handle the last read.
This patch must be backported to 1.9 because it will be used by another patch to
fix the option abortonclose.
The status codes definition (STAT_STATUS_*) and their string representation
stat_status_codes) have been moved in stats files. There is no reason to keep
them in proto_http files.
In task_unlink_rq, to decide if we should logk the global runqueue lock,
use the TASK_GLOBAL flag instead of relying on t->thread_mask being tid_bit,
as it could be so while still being in the global runqueue if another thread
woke that task for us.
This should be backported to 1.9.
Injecting on a saturated listener started to exhibit some deadlocks
again between LIST_POP_LOCKED() and LIST_DEL_LOCKED(). Olivier found
it was due to a leftover from a previous debugging session. This patch
fixes it.
This will have to be backported if the other LIST_*_LOCKED() patches
are backported.
Some packages used to rely on DEFAULT_MAXCONN to set the default global
maxconn value to use regardless of the initial ulimit. The recent changes
made the lowest bound set to 100 so that it is compatible with almost any
environment. Now that DEFAULT_MAXCONN is not needed for anything else, we
can use it for the lowest bound set when maxconn is not configured. This
way it retains its original purpose of setting the default maxconn value
eventhough most of the time the effective value will be higher thanks to
the automatic computation based on "ulimit -n".
This entry was still set to 2000 but never used anymore. The only places
where it appeared was as an alias to SYSTEM_MAXCONN which forces it, so
let's turn these ones to SYSTEM_MAXCONN and remove the default value for
DEFAULT_MAXCONN. SYSTEM_MAXCONN still defines the upper bound however.
Add variants of the HA_ATOMIC* macros, prefixed with a _, that do the
atomic operation with no barrier generated by the compiler. It is expected
the developer adds barriers manually if needed.
When using the new __atomic* API, ask the compiler to generate barriers.
A variant of those functions that don't generate barriers will be added later.
Before that, using HA_ATOMIC* would not generate any barrier, and some parts
of the code should be reviewed and missing barriers should be added.
This should probably be backported to 1.8 and 1.9.
Implement __ha_barrier functions to be used when trying to protect data
modified by atomic operations (except when using HA_ATOMIC_STORE).
On intel, atomic operations either use the LOCK prefix and xchg, and both
atc as full barrier, so there's no need to add an extra barrier.
In this function we end up with successive locked operations then a
store barrier, and in addition the compiler has to emit less efficient
code due to a longer jump. There's no need for absolutely updating the
tasks_run_queue counter before clearing the task's leaf pointer, so
let's swap the two operations and benefit from a single barrier as much
as possible. This code is on the hot path and shows about half a percent
of improvement with 8 threads.
Tests show that it's slightly faster to have this field in the listener.
The cache walk patterns are under heavy stress and having only this field
written to in the bind_conf was wasting a cache line that was heavily
read. Let's move this close to the other entries already written to in
the listener. Warning, the position does have an impact on peak performance.
Now that the P2C algorithm for the accept queue is removed, we don't
need to map a number to a thread bit anymore, so let's remove all
these fields which are taking quite some space for no reason.
We already have my_ffsl() to find the lowest bit set in a word, and
this patch implements the search for the highest bit set in a word.
On x86 it uses the bsr instruction and on other architectures it
uses an efficient implementation.
By picking two randoms following the P2C algorithm, we seldom observe
asymmetric loads on bursts of small session counts. This is typically
what makes h2load take a bit of time to complete the last 100% because
if a thread gets two connections while the other ones only have one,
it takes twice the time to complete its work.
This patch proposes a modification of the p2c algorithm which seems
more suitable to this case : it mixes a rotating index with a random.
This way, we're certain that all threads are consulted in turn and at
the same time we're not forced to use the ones we're giving a chance.
This significantly increases the traffic rate. Now h2load shows faster
completion and the average request rates on H2 and the TLS resume rate
increases by a bit more than 5% compared to pure p2c.
The index was placed into the struct bind_conf because 1) it's faster
there and it's the best place to optimally distribute traffic among a
group of listeners. It's the only runtime-modified element there and
it will be quite cache-hot.
By using LIST_DEL_INIT() instead of LIST_DEL()+LIST_INIT() we manage
to bump the peak connection rate by no less than 3% on 8 threads.
The perf top profile shows much less contention in this area which
suffered from the second reload.
It turns out that we call LIST_DEL+LIST_INIT very frequently and that
the compiler doesn't know what pointers get modified in the e->n->p
and e->p->n dance, so when LIST_INIT() is called, it reloads these
pointers, which is quite a bit of a mess in terms of performance.
This patch adds LIST_DEL_INIT() to perform the two operations at once
using local temporary variables so that the compiler knows these
pointers are left unaffected.
We move the code responsible of parsing protocol buffers messages
inside gRPC messages from sample.c to include/proto/protocol_buffers.h
so that to reuse it to cascade "ungrpc" converter.
For now on, "ungrpc" may take a second optional argument to provide
the protocol buffers types used to encode the field value to be extracted.
When absent the field value is extracted as a binary sample which may then
followed by others converters like "hex" which takes binary as input sample.
When this second argument is a type which does not match the one found by "ungrpc",
this field is considered as not found even if present.
With this patch we also remove the useless "varint" and "svarint" converters.
Update the documentation about "ungrpc" converters.
Parsing protocol buffer fields always consists in skip the field
if the field is not found or store the field value if found.
So, with this patch we factorize a little bit the code for "ungrpc" converter.
Well, that's becoming embarrassing. Now this fixes commit 4ef6801c
("BUG/MEDIUM: list: correct fix for LIST_POP_LOCKED's removal of last
element") which itself tried to fix commit 285192564. This fix only
works under low contention and was tested with the listener's queue.
With the idle conns it's obvious that it's still wrong since adding
more than one element to the list leaves a LLIST_BUSY pointer into
the list's head. This was visible when accumulating idle connections
in a server's list.
This new version of the fix almost goes back to the original code,
except that since then we addressed issues with expectedly idempotent
operations that were not. Now the code has been verified on paper again
and has survived 300 million connections spread over 4 threads.
This will have to be backported if the commit above is backported.
Let's keep a copy of these initial values. They will be useful to
compute automatic maxconn, as well as to restore proper limits when
doing an execve() on external checks.
This patch implements peer heartbeat feature to prevent any haproxy peer
from reconnecting too often, consuming sockets for nothing.
To do so, we add PEER_MSG_CTRL_HEARTBEAT new message to PEER_MSG_CLASS_CONTROL peers
control class of messages. A ->heartbeat field is added to peer structs
to store the heatbeat timeout value which is handled by the same function as for ->reconnect
to control the session timeouts. A 2-bytes heartbeat message is sent every 3s when
no updates have to be sent. This way, the peer which receives such a message is sure
the remote peer is still alive. So, it resets the ->reconnect peer session
timeout to its initial value (5s). This prevents any reconnection to an
already connected alive peer.
Historically the default frontend's maxconn used to be quite low (2000),
which was sufficient two decades ago but often proved to be a problem
when users had purposely set the global maxconn value but forgot to set
the frontend's.
There is no point in keeping this arbitrary limit for frontends : when
the global maxconn is lower, it's already too high and when the global
maxconn is much higher, it becomes a limiting factor which causes trouble
in production.
This commit allows the value to be set to zero, which becomes the new
default value, to mean it's not directly limited, or in fact it's set
to the global maxconn. Since this operation used to be performed before
computing a possibly automatic global maxconn based on memory limits,
the calculation of the maxconn value and its propagation to the backends'
fullconn has now moved to a dedicated function, proxy_adjust_all_maxconn(),
which is called once the global maxconn is stabilized.
This comes with two benefits :
1) a configuration missing "maxconn" in the defaults section will not
limit itself to a magically hardcoded value but will scale up to the
global maxconn ;
2) when the global maxconn is not set and memory limits are used instead,
the frontends' maxconn automatically adapts, and the backends' fullconn
as well.
In an attempt to try to provide automatic maxconn settings, we need to
decorrelate a listner's backlog and maxconn so that these values can be
independent. This introduces a listener_backlog() function which retrieves
the backlog value from the listener's backlog, the frontend's, the
listener's maxconn, the frontend's or falls back to 1024. This
corresponds to what was done in cfgparse.c to force a value there except
the last fallback which was not set since the frontend's maxconn is always
known.
As seen with Olivier, in the end the fix in commit 285192564 ("BUG/MEDIUM:
list: fix LIST_POP_LOCKED's removal of the last pointer") is wrong,
the code there was right but the bug was triggered by another bug in
LIST_ADDQ_LOCKED() which doesn't properly update the list's head by
inserting in the wrong order.
This will have to be backported if the commit above is backported.
There is a very difficult to reproduce race in the listener's accept
code, which is much easier to reproduce once connection limits are
properly enforced. It's an ABBA lock issue :
- the following functions take l->lock then lq_lock :
disable_listener, pause_listener, listener_full, limit_listener,
do_unbind_listener
- the following ones take lq_lock then l->lock :
resume_listener, dequeue_all_listener
This is because __resume_listener() only takes the listener's lock
and expects to be called with lq_lock held. The problem can easily
happen when listener_full() and limit_listener() are called a lot
while in parallel another thread releases sessions for the same
listener using listener_release() which in turn calls resume_listener().
This scenario is more prevalent in 2.0-dev since the removal of the
accept lock in listener_accept(). However in 1.9 and before, a different
but extremely unlikely scenario can happen :
thread1 thread2
............................ enter listener_accept()
limit_listener()
............................ long pause before taking the lock
session_free()
dequeue_all_listeners()
lock(lq_lock) [1]
............................ try_lock(l->lock) [2]
__resume_listener()
spin_lock(l->lock) =>WAIT[2]
............................ accept()
l->accept()
nbconn==maxconn =>
listener_full()
state==LI_LIMITED =>
lock(lq_lock) =>DEADLOCK[1]!
In practice it is almost impossible to trigger it because it requires
to limit both on the listener's maxconn and the frontend's rate limit,
at the same time, and to release the listener when the connection rate
goes below the limit between poll() returns the FD and the lock is
taken (a few nanoseconds). But maybe with threads competing on the
same core it has more chances to appear.
This patch removes the lq_lock and replaces it with a lockless queue
for the listener's wait queue (well, technically speaking a self-locked
queue) brought by commit a8434ec14 ("MINOR: lists: Implement locked
variations.") and its few subsequent fixes. This relieves us from the
need of the lq_lock and removes the deadlock. It also gets rid of the
distinction between __resume_listener() and resume_listener() since the
only difference was the lq_lock. All listener removals from the list
are now unconditional to avoid races on the state. It's worth noting
that the list used to never be initialized and that it used to work
only thanks to the state tests, so the initialization has now been
added.
This patch must carefully be backported to 1.9 and very likely 1.8.
It is mandatory to be careful about replacing all manipulations of
l->wait_queue, global.listener_queue and p->listener_queue.
Since LIST_DEL_LOCKED() and LIST_POP_LOCKED() now automatically reinitialize
the removed element, there's no need for keeping this LIST_INIT() call in the
idle connection code.
