It's more convenient to return the element than to return just 0 or 1,
as the next thing we'll want to do is to act on this element! In addition
it was using variable arguments instead of consts, causing some reuse
constraints which were also addressed. This doesn't change its use as
a boolean, hence why call places were not modified.
The maxage and smaxage variables were inadvertently assigned the
Cache-Control s-maxage and max-age values respectively when it should
have been the other way around.
This can be backported on all branches after 1.8 (included).
If an HTTP request or response had a "Cache-Control" header that had
multiple comma-separated subparts in its value (like "max-age=1,
no-store" for instance), we did not process the values correctly and
only parsed the first one. That made us store some HTTP responses in the
cache when they were explicitely uncacheable.
This patch replaces the way the values are parsed by an http_find_header
loop that manages every sub part of the value independently.
This patch should be backported to 2.2 and 2.1. The bug also exists on
previous versions but since the sources changed, a new commit will have
to be created.
[wla: This patch requires bb4582c ("MINOR: ist: Add a case insensitive
istmatch function"). Backporting for < 2.1 is not a requirement since it
works well enough for most cases, it was a known limitation of the
implementation of non-htx version too]
When no Cache-Control max-age or s-maxage information is present in a
cached response, we need to parse the Expires header value (RFC 7234#5.3).
An invalid Expires date value or a date earlier than the reception date
will make the cache_entry stale upon creation.
For now, the Cache-Control and Expires headers are parsed after the
insertion of the response in the cache so even if the parsing of the
Expires results in an already stale entry, the entry will exist in the
cache.
Memset the sample before using it through hlua_lua2smp. This function is
ORing the smp.flags, so this field need to be cleared before its use.
This was reported by a coverity warning.
Fixes the github issue #929.
This bug can be backported up to 1.8.
Adjust condition used to report down_time for statistics. There was a
tiny probabilty to have a negative downtime if last_change was superior
to now. If this is the case, return only down_time.
This bug can backported up to 1.8.
When a server is up after a failure, its downtime was reset to 0 on the
statistics. This is due to a wrong condition that causes srv.down_time
to never be set. Fix this by updating down_time each time the server is in
STARTING state.
Fixes the github issue #920.
This bug can be backported up to 1.8.
Implement counters for h2 protocol error on connection or stream level.
Also count the total number of rst_stream and goaway frames sent by the
mux in response to a detected error.
Add pointer to counters as a member for h2c structure. This pointer is
initialized on h2_init function. This is useful to quickly access and
manipulate the counters inside every h2 functions.
Res.cache_hit sample fetch returns a boolean which is true when the HTTP
response was built out of a cache. The cache's name is returned by the
res.cache_name sample_fetch.
This resolves GitHub issue #900.
If a client sends a conditional request containing an If-Modified-Since
header (and no If-None-Match header), we try to compare the date with
the one stored in the cache entry (coming either from a Last-Modified
head, or a Date header, or corresponding to the first response's
reception time). If the request's date is earlier than the stored one,
we send a "304 Not Modified" response back. Otherwise, the stored is sent
(through a 200 OK response).
This resolves GitHub issue #821.
In order to manage "If-Modified-Since" requests, we need to keep a
reference time for our cache entries (to which the conditional request's
date will be compared).
This reference is either extracted from the "Last-Modified" header, or
the "Date" header, or the reception time of the response (in decreasing
order of priority).
The date values are converted into seconds since epoch in order to ease
comparisons and to limit storage space.
BorinSSL pretends to be 1.1.1 version of OpenSSL. It messes some
version based feature presense checks. For example, OpenSSL specific
early data support.
Let us change that feature detction to SSL_READ_EARLY_DATA_SUCCESS
macro check instead of version comparision.
Previous commit ae32ac74db ("BUG/MINOR: log: fix memory leak on logsrv
parse error") addressed one issue and introduced another one, the logsrv
pointer may also be null at the end of the function so we must test it
before deciding to dereference it.
This should be backported along with the patch above to 2.2.
In case of parsing error on logsrv, we can leave parse_logsrv() without
releasing logsrv->ring_name or smp_rgs. Let's free them on the error path.
This should fix issue #926 detected by Coverity.
The impact is only a tiny leak just before reporting a fatal error, so it
will essentially annoy valgrind.
This can be backported to 2.0 (just drop the ring part).
It's a regression from b3201a3e "BUG/MINOR: disable dynamic OCSP load
with BoringSSL". The origin bug is link to 76b4a12 "BUG/MEDIUM: ssl:
memory leak of ocsp data at SSL_CTX_free()": ssl_sock_free_ocsp()
shoud be in #ifndef OPENSSL_IS_BORINGSSL.
To avoid long #ifdef for small code, the BoringSSL part for ocsp load
is isolated in a simple #ifdef.
This must be backported in 2.2 and 2.1
`att_beg` is assigned to `next` at the end of the `for` loop, but is
assigned to `prev` at the beginning of the loop, which is itself
assigned to `next` after each loop. So it represents a double
assignation for the same value. Also `att_beg` is not used after the end
of the loop.
this is a partial fix for github issue #923, all the others could
probably be marked as intentional to protect future changes.
no backport needed.
Signed-off-by: William Dauchy <wdauchy@gmail.com>
Issue #910 reports that we fail to check a few extchk_setenv() in the
child process. These are mostly harmless, but instead of counting on
the external check script to fail the dirty way, better fail cleanly
when detecting the failure.
This could probably be backported to all stable branches.
As reported by Coverity in issue #917, commit 96bca33 ("OPTIM: queue:
decrement the nbpend and totpend counters outside of the lock")
introduced a bug when moving the increments outside of the loop,
because we can't always rely on the pendconn "p" here as it may
be null. We can retrieve the proxy pointer directly from s->proxy
instead. The same is true for pendconn_redistribute(), though the
last "p" pointer there was still valid. This patch fixes both.
No backport is needed, this was introduced just before 2.3-dev8.
The "weight" column on the stats page is somewhat confusing when using
slowstart becaue it reports the effective weight, without being really
explicit about it. In some situations the user-configured weight is more
relevant (especially with long slowstarts where it's important to know
if the configured weight is correct).
This adds a new uweight stat which reports a server's user-configured
weight, and in a backend it receives the sum of all servers' uweights.
In addition it adds the mention of "effective" in a few descriptions
for the "weight" column (help and doc).
As a result, the list of servers in a backend is now always scanned
when dumping the stats. But this is not a problem given that these
servers are already scanned anyway and for way heavier processing.
In order to be compatible with the "set ssl cert" command of the CLI,
this patch restrict the ssl-load-extra-del-ext to files with a ".crt"
extension in the configuration.
Related to issue #785.
Should be backported where 8e8581e ("MINOR: ssl: 'ssl-load-extra-del-ext'
removes the certificate extension") was backported.
When dumping the stats page (or the CSV output), when many states are
mixed, it's hard to figure the number of up servers. But when showing
only the "up" servers or hiding the "maint" servers, there's no way to
know how many servers are configured, which is problematic when trying
to update server-templates.
What this patch does, for dumps in "up" or "no-maint" modes, is to add
after the backend's "UP" or "DOWN" state "(%d/%d)" indicating the number
of servers seen as UP to the total number of servers in the backend. As
such, seeing "UP (33/39)" immediately tells that there are 6 servers that
are not listed when using "up", or will let the client figure how many
servers are left once deducted the number of non-maintenance ones. It's
not done on default dumps so as not to disturb existing tools, which
already have all the information they need in the dump.
"no-maint" is a bit similar to "up" except that it will only hide
servers that are in maintenance (or disabled in the configuration), and
not those that are enabled but failed a check. One benefit here is to
significantly reduce the output of the "show stat" command when using
large server-templates containing entries that are not yet provisioned.
Note that the prometheus exporter also has such an option which does
the exact same.
We already had it on the HTTP interface but it was not accessible on the
CLI. It can be very convenient to hide servers which are down, do not
resolve, or are in maintenance.
Leastconn has the nice propery of being able to sort servers by their
current usage. It's really a shame to force all requests into the backend
queue when the algo would be able to also consider their current queue.
In order not to change existing behavior but extend it, this patch allows
leastconn to elect servers which are already full if they have an explicitly
configured maxqueue setting above zero and their queue hasn't reached that
threshold. This will significantly reduce the pressure in the backend queue
when queuing a lot with lots of servers.
A test on 8 threads with 100 servers configured with maxconn 1 jumped
from 165krps to 330krps with maxqueue 15 with this patch.
This partially undoes commit 82cd5c13a ("OPTIM: backend: skip LB when we
know the backend is full") but allows to scale much better even by setting
a single-digit maxqueue value. Some better heuristics could be used to
maintain the behavior of the bypass in the patch above, consisting in
keeping it if it's known that there is no server with a configured
maxqueue in the farm (or in the backend).
When servers are queued into the leastconn tree, it's important to also
consider their queue length. There could be some servers with lots of
queued requests that we don't want to hammer with extra connections. In
order not to add extra stress to the LB algorithm, we don't update the
value when adding to the queue, only when updating the connection count
(i.e. picking from the queue or releasing a connection). This will be
sufficient to significantly improve the fairness in such situations.
We don't need to do that inside the lock. However since the operation
used to be done in deep functions, we have to make it resurface closer
to visible parts. It remains reasonably self-contained in queue.c so
that's not that big of a deal. Some places (redistribute) could benefit
from a single operation for all counts at once. Others like
pendconn_process_next_strm() are still called with both locks held but
now it will be possible to change this.
Instead of incrementing, decrementing them and updating their max under
the lock, make them atomic and keep them out of the lock as much as
possible. For __pendconn_unlink_* it would be wide to decide to move
these counters outside of the function, inside the callers so that a
single atomic op can be done per counter even for groups of operations.
Similarly to previous changes, we know if we're dealing with a server
or proxy lock so let's directly lock at the finest possible places
there. It's worth noting that a part of the operation consisting in
an increment and update of a max could be done outside of the lock
using atomic ops and a CAS.
The function is called with the lock held and does too many tests for
things that are already known from its callers. Let's split it in two
so that its callers call either the per-server or per-proxy function
depending on where the element is (since they had to determine it
prior to taking the lock).
No need to use an exclusive lock on the proxy anymore when reading its
setting, a read lock is enough. A few other places continue to use a
write-lock when modifying simple flags only in order to let this
function see a consistent value all along. This might be changed in
the future using barriers and local copies.
This is an anticipation of finer grained locking for the queues. For now
all lock places take a write lock so that there is no difference at all
with previous code.
In h2_send(), if we are in a state where we know it is no longer possible to
send data, we must exit the sending loop to avoid any possiblity to loop
forever. It may happen if the mbuf ring is released while the H2_CF_MUX_MFULL
flag is still set. Here is a possible scenario to trigger the bug :
1) The mbuf ring is full because we are unable to send data. The
H2_CF_MUX_MFULL flag is set on the H2 connection.
2) At this stage, the task timeout expires because the H2 connection is
blocked. We enter in h2_timeout_task() function. Because the mbuf ring is
full, we cannot send the GOAWAY frame. Thus the H2_CF_GOAWAY_FAILED flag is
set. The H2 connection is not released yet because there is still a stream
attached. Here we leave h2_timeout_task() function.
3) A bit later, the H2 connection is woken up. If h2_process(), nothing is
performed by the first attempt to send data, in h2_send(). Then, because
the H2_CF_GOAWAY_FAILED flag is set, the mbuf ring is released. But the
H2_CF_MUX_MFULL flag is still there. At this step a second attempt to send
data is performed.
4) In h2_send(), we try to send data in a loop. To exist this loop, done
variable must be set to 1. Because the H2_CF_MUX_MFULL flag is set, we
don't call h2_process_mux() and done is not updated. Because the mbuf ring
is now empty, nothing is sent and the H2_CF_MUX_MFULL flag is never
removed. Now, we loop forever... waiting for the watchdog.
To fix the bug, we now exit the loop if one of these conditions is true :
- The H2_CF_GOAWAY_FAILED flag is set on the H2 connection
- The CO_FL_SOCK_WR_SH flag is set on the underlying connection
- The H2 connection is in the H2_CS_ERROR2 state
This patch should fix the issue #912 and most probably #875. It must be
backported as far as the 1.8.
When an internal response is returned to a client, the message payload must be
skipped if it is a reply to a HEAD request. The payload is removed from the HTX
message just before the message forwarding.
This bugs has been around for a long time. It was already there in the pre-HTX
versions. In legacy HTTP mode, internal errors are not parsed. So this bug
cannot be easily fixed. Thus, this patch should only be backported in all HTX
versions, as far as 2.0. However, the code has significantly changed in the
2.2. Thus in the 2.1 and 2.0, the patch must be entirely reworked.
Partial support of conditional HTTP requests. This commit adds the
support of the 'If-None-Match' header (see RFC 7232#3.2).
When a client specifies a list of ETags through one or more
'If-None-Match' headers, they are all compared to the one that might have
been stored in the corresponding http cache entry until one of them
matches.
If a match happens, a specific "304 Not Modified" response is
sent instead of the cached data. This response has all the stored
headers but no other data (see RFC 7232#4.1). Otherwise, the whole cached data
is sent.
Although unlikely in a GET/HEAD request, the "If-None-Match: *" syntax is
valid and also receives a "304 Not Modified" response (RFC 7434#4.3.2).
This resolves a part of GitHub issue #821.
When sent by a server for a given resource, the ETag header is
stored in the coresponding cache entry (as any other header). So in
order to perform future ETag comparisons (for subsequent conditional
HTTP requests), we keep the length of the ETag and its offset
relative to the start of the cache_entry.
If no ETag header exists, the length and offset are zero.
Add a function that compares two etags that might be of different types.
If any of them is weak, the 'W/' prefix is discarded and a strict string
comparison is performed.
Co-authored-by: Tim Duesterhus <tim@bastelstu.be>
If the slowstart value in a state file implies the latest state change
is within the slowstart period, we end up calling srv_update_status()
to reschedule the server's state change but its task is not yet
allocated and remains null, causing a crash on startup.
Make sure srv_update_status() supports being called with partially
initialized servers which do not yet have a task. If the task has to
be scheduled, it will necessarily happen after initialization since
it will result from a state change.
This should be backported wherever server-state is present.
In commit 5cd4bbd7a ("BUG/MAJOR: threads/queue: Fix thread-safety issues
on the queues management") the counter of transferred connections was
accidently lost, so that when a server goes down with connections in its
queue, it will always be reported that 0 connection were transferred.
This should be backported as far as 1.8 since the patch above was
backported there.
In issue #785, users are reporting that it's not convenient to load a
".crt.key" when the configuration contains a ".crt".
This option allows to remove the extension of the certificate before
trying to load any extra SSL file (.key, .ocsp, .sctl, .issuer etc.)
The patch changes a little bit the way ssl_sock_load_files_into_ckch()
looks for the file.
safer to close handle before the object is put back in the global pool.
this was introduced by commit 9378bbe0be ("MEDIUM: listener:
use protocol->accept_conn() to accept a connection")
this should fix github issue #902
no backport needed.
Signed-off-by: William Dauchy <wdauchy@gmail.com>
As previously discussed, nbproc usage is bad, deprecated, and scheduled
for removal in 2.5.
If "nbproc" is found with more than one process while nbthread is not
set, a warning will be emitted encouraging to remove it or to migrate
to nbthread instead. This makes sure the user has an opportunity to
both see the message and silence it.
This counter is only updated and never used, and in addition it's done
without any atomicity so it's very unlikely to be correct on multi-CPU
systems! Let's just remove it since it's not used.
It's a bit overkill to register an initcall to call a function to set
a lock to zero when not debugging, let's just declare the lock as
pre-initialized to zero.
When using a low hash-balance-factor value, it's possible to loop
many times trying to find the best server. Figures in the order of
100-300 times were observed for 1000 servers with a factor of 101
(which seems a bit excessive for such a large farm). Given that
there's nothing in that function that prevents multiple threads
from working in parallel, let's switch to a read lock. Tests on
8 threads show roughly a 2% performance increase with this.
The "first" algorithm creates a lot of contention because all threads
focus on the same server by definition (the first available one). By
turning the exclusive lock to a read lock in fas_get_next_server(),
the request rate increases by 16% for 8 threads when many servers are
getting close to their maxconn.
This function doesn't change the tree, it only looks for the first
usable server, so let's do that under a read lock to limit the
situations like the ones described in issue #881 where finding a
usable server when dealing with lots of saturated ones can be
expensive. At least threads will now be able to look up in
parallel.
It's interesting to note that s->served is not incremented during the
server choice, nor is the server repositionned. So right now already,
nothing prevents multiple threads from picking the same server. This
will not cause a significant imbalance anyway given that the server
will automatically be repositionned at the right place, but this might
be something to improve in the future if it doesn't come with too high
a cost.
It also looks like the way a server's weight is updated could be
revisited so that the write lock gets tighter at the expense of a
short part of inconsistency between weights and servers still present
in the tree.
- map_get_server_hash() doesn't need a write lock since it only
reads the array, let's only use a read lock here.
- map_get_server_rr() only needs exclusivity to adjust the rr_idx
while looking for its entry. Since this one is not used by
map_get_server_hash(), let's turn this lock to a seek lock that
doesn't block reads.
With 8 threads, no significant performance difference was noticed
given that lookups are usually instant with this LB algo so the
lock contention is rare.
It was previously a spinlock, and it happens that a number of LB algos
only lock it for lookups, without performing any modification. Let's
first turn it to an rwlock and w-lock it everywhere. This is strictly
identical.
It was carefully checked that every HA_SPIN_LOCK() was turned to
HA_RWLOCK_WRLOCK() and that HA_SPIN_UNLOCK() was turned to
HA_RWLOCK_WRUNLOCK() on this lock. _INIT and _DESTROY were updated too.
The server lock must be held when server_take_conn() and server_drop_conn()
lbprm callback functions are called. It is a documented prerequisite but it is
not always performed. It only affects leastconn and fas lb algorithm. Others
don't use these callback functions.
A race condition on the next pending effecive weight (next_eweight) may be
encountered with the leastconn lb algorithm. An agent check may set it to 0
while fwlc_srv_reposition() is called. The server is locked during the
next_eweight update. But because the server lock is not acquired when
fwlc_srv_reposition() is called, we may use it to recompute the server key,
leading to a division by 0.
This patch must be backported as far as 1.8.
It is not guaranteed that the backend connection has an owner. It is set when
the connection is created. But when the connection is moved in a server idle
list, the connection owner is set to NULL and may never be set again. On the
other hand, when a mux is created or when a CS is attached, the session is
always defined. The H1 stream always keep a reference on it when it is
created. Thus, when a bad message is captured we should not rely on the
connection owner to retrieve the session. Instead we should get it from the H1
stream.
If an agent try to set a variable with the NULL data type, an unset is perform
instead to avoid undefined behaviors. Once decoded, such data are translated to
a sample with the type SMP_T_ANY. It is unexpected in HAProxy. When a variable
is set with such sample, no data are attached to the variable. Thus, when the
variable is retrieved later in the transaction, the sample data are
uninitialized, leading to undefined behaviors depending on how it is used. For
instance, it leads to a crash if the debug converter is used on such variable.
This patch should fix the issue #855. It must be backported as far as 1.8.
Detect if the sni used a constant value and if so, allow to reuse this
connection for later sessions. Use a combination of SMP_USE_INTRN +
!SMP_F_VOLATILE to consider a sample as a constant value.
This features has been requested on github issue #371.
During a peers session collision (two peer sessions opened on both side) we must
mark the peer the session of which will be shutdown as alive, if not ->reconnect
timer will be set with a wrong value if the synchro task expires after the peer
has been reconnected. This possibly leads to unexpected deconnections during handshakes.
Furthermore, this patch cancels any heartbeat tranmimission when a reconnection
is prepared.
Right now when running a configuration with many global timers (e.g. many
health checks), there is a lot of contention on the global wait queue
lock because all threads queue up in front of it to scan it.
With 2000 servers checked every 10 milliseconds (200k checks per second),
after 23 seconds running on 8 threads, the lock stats were this high:
Stats about Lock TASK_WQ:
write lock : 9872564
write unlock: 9872564 (0)
wait time for write : 9208.409 msec
wait time for write/lock: 932.727 nsec
read lock : 240367
read unlock : 240367 (0)
wait time for read : 149.025 msec
wait time for read/lock : 619.991 nsec
i.e. ~5% of the total runtime spent waiting on this specific lock.
With upgradable locks we don't need to work like this anymore. We
can just try to upgade the read lock to a seek lock before scanning
the queue, then upgrade the seek lock to a write lock for each element
we want to delete there and immediately downgrade it to a seek lock.
The benefit is double:
- all other threads which need to call next_expired_task() before
polling won't wait anymore since the seek lock is compatible with
the read lock ;
- all other threads competing on trying to grab this lock will fail
on the upgrade attempt from read to seek, and will let the current
lock owner finish collecting expired entries.
Doing only this has reduced the wake_expired_tasks() CPU usage in a
very large servers test from 2.15% to 1.04% as reported by perf top,
and increased by 3% the health check rate (all threads being saturated).
This is expected to help against (and possibly solve) the problem
described in issue #875.
There is a theorical problem in the wait queue, which is that with many
threads, one could spend a lot of time looping on the newly expired tasks,
causing a lot of contention on the global wq_lock and on the global
rq_lock. This initially sounds bening, but if another thread does just
a task_schedule() or task_queue(), it might end up waiting for a long
time on this lock, and this wait time will count on its execution budget,
degrading the end user's experience and possibly risking to trigger the
watchdog if that lasts too long.
The simplest (and backportable) solution here consists in bounding the
number of expired tasks that may be picked from the global wait queue at
once by a thread, given that all other ones will do it as well anyway.
We don't need to pick more than global.tune.runqueue_depth tasks at once
as we won't process more, so this counter is updated for both the local
and the global queues: threads with more local expired tasks will pick
less global tasks and conversely, keeping the load balanced between all
threads. This will guarantee a much lower latency if/when wakeup storms
happen (e.g. hundreds of thousands of synchronized health checks).
Note that some crashes have been witnessed with 1/4 of the threads in
wake_expired_tasks() and, while the issue might or might not be related,
not having reasonable bounds here definitely justifies why we can spend
so much time there.
This patch should be backported, probably as far as 2.0 (maybe with
some adaptations).
The proxy stopping mechanism was changed with commit 322b9b94e ("MEDIUM:
proxy: make stop_proxy() now use stop_listener()") so that it's now
entirely driven by the listeners. One thing was forgotten though, which
is that pure backends will not stop anymore since they don't have any
listener, and that it's necessary to stop them in order to stop the
health checks.
No backport is needed.
We don't need to specify the handler anymore since it's set in the
receiver. Let's remove this argument from the function and clean up
the remains of code that were still setting it.
Now we define a new sock_accept_iocb() for socket-based stream protocols
and use it as a wrapper for listener_accept() which now takes a listener
and not an FD anymore. This will allow the receiver's I/O cb to be
redefined during registration, and more specifically to get rid of the
hard-coded hacks in protocol_bind_all() made for syslog.
The previous ->accept() callback in the protocol was removed since it
doesn't have anything to do with accept() anymore but is more generic.
A few places where listener_accept() was compared against the FD's IO
callback for debugging purposes on the CLI were updated.
For now we're still using the protocol's default accept() function as
the I/O callback registered by the receiver into the poller. While
this is usable for most TCP connections where a listener is needed,
this is not suitable for UDP where a different handler is needed.
Let's make this configurable in the receiver just like the upper layer
is configurable for listeners. In order to ease stream protocols
handling, the protocols will now provide a default I/O callback
which will be preset into the receivers upon allocation so that
almost none of them has to deal with it.
The receiver FDs must not be manipulated by the listener_accept()
function anymore, it must exclusively rely on the job performed by
its listeners, as it is also the only way to keep the receivers
working for established connections regardless of the listener's
state (typically for multiplexed protocols like QUIC). This used
to be necessary when the FDs were adjusted at once only but now
that fd_done() is gone and the need for polling enabled by the
accept_conn() function which detects the EAGAIN, we have nothing
to do there to fixup any possible previous bad decision anymore.
Interestingly, as a side effect of making the code not depend on
the FD anymore, it also removes the need for a second lock, which
increase the accept rate by about 1% on 8 threads.
Now listener_accept() doesn't have to deal with the incoming FD anymore
(except for a little bit of side band stuff). It directly retrieves a
valid connection from the protocol layer, or receives a well-defined
error code that helps it decide how to proceed. This removes a lot of
hardly maintainable low-level code and opens the function to receive
new protocol stacks.
This is the same as previous commit, but this time for the sockpair-
specific stuff, relying on recv_fd_uxst() instead of accept(), so the
code is simpler. The various errno cases are handled like for regular
sockets, though some of them will probably never happen, but this does
not hurt.
The socket-specific accept() code in listener_accept() has nothing to
do there. Let's move it to sock.c where it can be significantly cleaned
up. It will now directly return an accepted connection and provide a
status code instead of letting listener_accept() deal with various errno
values. Note that this doesn't support the sockpair specific code.
The function is now responsible for dealing with its own receiver's
polling state and calling fd_cant_recv() when facing EAGAIN.
One tiny change from the previous implementation is that the connection's
sockaddr is now allocated before trying accept(), which saves a memcpy()
of the resulting address for each accept at the expense of a cheap
pool_alloc/pool_free on the final accept returning EAGAIN. This still
apparently slightly improves accept performance in microbencharks.
This call was introduced by commit 5ced3e887 ("MINOR: sock: add
sock_accept_conn() to test a listening socket") but is actually quite
confusing because it makes one think the socket will accept a connection
(which is what we want to have in a new function) while it only tells
whether it's configured to accept connections. Let's call it
sock_accepting_conn() instead.
The same change was applied to sockpair which had the same issue.
Now that this function is always called with an initialized connection
and that the control layer is always initialized, we don't need to play
games with fdtab[] to decide how to close, we can simply rely on the
regular close path using conn_ctrl_close(), which can be fused with
conn_xprt_close() into conn_full_close().
The code is cleaner because the FD is now used only for some
protocol-specific setup (that will eventually have to move) and to
try to send a hard-coded HTTP 500 error message on raw sockets.
Till now we would keep a per-thread queue of pending incoming connections
for which we would store:
- the listener
- the accepted FD
- the source address
- the source address' length
And these elements were first used in session_accept_fd() running on the
target thread to allocate a connection and duplicate them again. Doing
this induces various problems. The first one is that session_accept_fd()
may only run on file descriptors and cannot be reused for QUIC. The second
issue is that it induces lots of memory copies and that the listerner
queue thrashes a lot of cache, consuming 64 bytes per entry.
This patch changes this by allocating the connection before queueing it,
and by only placing the connection's pointer into the queue. Indeed, the
first two calls used to initialize the connection already store all the
information above, which can be retrieved from the connection pointer
alone. So we just have to pop one pointer from the target thread, and
pass it to session_accept_fd() which only needs the FD for the final
settings.
This starts to make the accept path a bit more transport-agnostic, and
saves memory and CPU cycles at the same time (1% connection rate increase
was noticed with 4 threads). Thanks to dividing the accept-queue entry
size from 64 to 8 bytes, its size could be increased from 256 to 1024
connections while still dividing the overall size by two. No single
queue full condition was met.
One minor drawback is that connection may be allocated from one thread's
pool to be used into another one. But this already happens a lot with
connection reuse so there is really nothing new here.
Roughly half of the calls to sockadr_alloc() are made to copy an already
known address. Let's optionally pass it in argument so that the function
can handle the copy at the same time, this slightly simplifies its usage.
fd_done_recv() used to be useful with the FD cache because it used to
allow to keep a file descriptor active in the poller without being
marked as ready in the cache, saving it from ringing immediately,
without incurring any system call. It was a way to make it yield
to wait for new events leaving a bit of time for others. The only
user left was the connection accepter (listen_accept()). We used
to suspect that with the FD cache removal it had become totally
useless since changing its readiness or not wouldn't change its
status regarding the poller itself, which would be the only one
deciding to report it again.
Careful tests showed that it indeed has exactly zero effect nowadays,
the syscall numbers are exactly the same with and without, including
when enabling edge-triggered polling.
Given that there's no more API available to manipulate it and that it
was directly called as an optimization from listener_accept(), it's
about time to remove it.
No protocol defines it anymore. The last user used to be the monitor-net
stuff that got partially broken already when the tcp_drain() function
moved to conn_sock_drain() with commit e215bba95 ("MINOR: connection:
make conn_sock_drain() work for all socket families") in 1.9-dev2.
A part of this will surely move back later when non-socket connections
arrive with QUIC but better keep the API clean and implement what's
needed in time instead.
As discussed here during 2.1-dev, "monitor-net" is totally obsolete:
https://www.mail-archive.com/haproxy@formilux.org/msg35204.html
It's fundamentally incompatible with usage of SSL, and imposes the
presence of file descriptors with hard-coded syscalls directly in the
generic accept path.
It's very unlikely that anyone has used it in the last 10 years for
anything beyond testing. In the worst case if anyone would depend
on it, replacing it with "http-request return status 200 if ..." and
"mode http" would certainly do the trick.
The keyword is still detected as special by the config parser to help
users update their configurations appropriately.
As discussed here during 2.1-dev, "mode health" is totally obsolete:
https://www.mail-archive.com/haproxy@formilux.org/msg35204.html
It's fundamentally incompatible with usage of SSL, doesn't support
source filtering, and imposes the presence of file descriptors with
hard-coded syscalls directly in the generic accept path.
It's very unlikely that anyone has used it in the last 10 years for
anything beyond testing. In the worst case if anyone would depend
on it, replacing it with "http-request return status 200" and "mode
http" would certainly do the trick.
The keyword is still detected as special by the config parser to help
users update their configurations appropriately.
FCGI mux is marked with HOL blocking. On safe reuse mode, the connection
using it are placed on the sessions instead of the available lists to
avoid sharing it with several clients. On detach, if they are no more
streams, remove the connection from the session before adding it to the
idle list. If there is still used streams, do not add it to available
list as it should be already on the session list.
H2 mux is marked with HOL blocking. On safe reuse mode, the connection
using it are placed on the sessions instead of the available lists to
avoid sharing it with several clients. On detach, if they are no more
streams, remove the connection from the session before adding it to the
idle list. If there is still used streams, do not add it to available
list as it should be already on the session list.
If a connection is using a mux protocol subject to HOL blocking, add it
to the session instead of the available list to avoid sharing it with
other clients on connection reuse.
When allocating a new session on connect_server, if the mux protocol is
marked as subject of HOL blocking, add it into session instead of
available list to avoid sharing it with other clients.
On server connection migration from one thread to another, the wrong
idle thread-specific counter is decremented. This bug was introduced
since commit 3d52f0f1f8 due to the
factorization with srv_use_idle_conn. However, this statement is only
executed from conn_backend_get. Extract the decrement from
srv_use_idle_conn in conn_backend_get and use the correct
thread-specific counter.
Rename the function to srv_use_conn to better reflect its purpose as it
is also used with a newly initialized connection not in the idle list.
As a side change, the connection insertion to available list has also
been extracted to conn_backend_get. This will be useful to be able to
specify an alternative list for protocol subject to HOL risk that should
not be shared between several clients.
This bug is only present in this release and thus do not need a backport.
The loop always missed one iteration due to the incrementation done on
the for check. Move the incrementation on the loop last statement to fix
this behaviour.
This bug has a very limited impact, not at all visible to the user, but
could be backported to 2.2.
When running a pure config check (haproxy -c) we go through the deinit
phase without having allocated fdtab, so we can't blindly dereference
it. The issue was added by recent commit ae7bc4a23 ("MEDIUM: deinit:
close all receivers/listeners before scanning proxies"), no backport is
needed.
In issue #894, Coverity suspects uninitialized values for a socket's
address whose family is AF_UNSPEC but it doesn't know that the address
is not used in this case. It's not on a critical path and working around
it is trivial, let's fully declare the address. We're doing it for both
TCP and UDP, because the same principle appears at two places.
It may happen that during a temporary listener pause resulting from a
SIGTTOU, one process gets one of its sockets disabled by another process
and will not be able to recover from this situation by itself. For the
protocols supporting this (TCPv4 and TCPv6 at the moment) this situation
is detectable, so when this happens, let's put the listener into the
PAUSED state so that it remains consistent with the real socket state.
One nice effect is that just sending the SIGTTIN signal to the process
is enough to recover the socket in this case.
There is no need to backport this, this behavior has been there forever
and the fix requires to reimplement the getsockopt() call there.
For socket pairs we don't rely on a real listening socket but we need to
have a properly connected UNIX stream socket. This is what the new
sockpair_accept_conn() tries to report. Some corner cases like half
shutdown will still not be detected but that should be sufficient for
most cases we really care about.
Now we introdce a new .rx_listening() function to report if a receiver is
actually a listening socket. The reason for this is to help detect shared
sockets that might have been broken by sibling processes.
At several places we need to check if a socket is still valid and still
willing to accept connections. Instead of open-coding this, each time,
let's add a new function for this.
Currently the suspend/resume mechanism for listeners only works on Linux
and we resort to a number of tricks involving shutdown+listen+shutdown
to try to detect failures on other operating systems that do not support
it. But on Linux connect(AF_UNSPEC) also works pretty well and is much
cleaner. It still doesn't work on other operating systems but the error
is easier to detect and appears safer. So let's switch to this.
When starting with a huge maxconn (say 1 billion), the only error seen
is "No polling mechanism available". This doesn't help at all to resolve
the problem. Let's add specific alerts for the failed mallocs. Now we can
get this instead:
[ALERT] 286/154439 (23408) : Not enough memory to allocate 2000000033 entries for fdtab!
This may be backported as far as 2.0 as it helps debugging bad configurations.
There are reports of a few "SC" in logs during reloads when H2 is used
on the backend side. Christopher analysed this as being caused by the
proxy disabled test in h2_process(). As the comment says, this was done
for frontends only, and must absolutely not send a GOAWAY to the backend,
as all it will result in is to make newly queued streams fail.
The fix consists in simply testing the connection side before deciding
to send the GOAWAY.
This may be backported as far as 2.0, though for whatever reason it seems
to manifest itself only since 2.2 (probably due to changes in the outgoing
connection setup sequence).
