This function was designed for haproxy while testing other functions
in the past. Initially it was not planned to be used given the not
very interesting numbers it showed on real URL data : it is not as
smooth as the other ones. But later tests showed that the other ones
are extremely sensible to the server count and the type of input data,
especially DJB2 which must not be used on numeric input. So in fact
this function is still a generally average performer and it can make
sense to merge it in the end, as it can provide an alternative to
sdbm+avalanche or djb2+avalanche for consistent hashing or when hashing
on numeric data such as a source IP address or a visitor identifier in
a URL parameter.
Summary:
Avalanche is supported not as a native hashing choice, but a modifier
on the hashing function. Note that this means that possible configs
written after 1.5-dev4 using "hash-type avalanche" will get an informative
error instead. But as discussed on the mailing list it seems nobody ever
used it anyway, so let's fix it before the final 1.5 release.
The default values were selected for backward compatibility with previous
releases, as discussed on the mailing list, which means that the consistent
hashing will still apply the avalanche hash by default when no explicit
algorithm is specified.
Examples
(default) hash-type map-based
Map based hashing using sdbm without avalanche
(default) hash-type consistent
Consistent hashing using sdbm with avalanche
Additional Examples:
(a) hash-type map-based sdbm
Same as default for map-based above
(b) hash-type map-based sdbm avalanche
Map based hashing using sdbm with avalanche
(c) hash-type map-based djb2
Map based hashing using djb2 without avalanche
(d) hash-type map-based djb2 avalanche
Map based hashing using djb2 with avalanche
(e) hash-type consistent sdbm avalanche
Same as default for consistent above
(f) hash-type consistent sdbm
Consistent hashing using sdbm without avalanche
(g) hash-type consistent djb2
Consistent hashing using djb2 without avalanche
(h) hash-type consistent djb2 avalanche
Consistent hashing using djb2 with avalanche
Summary:
In testing at tumblr, we found that using djb2 hashing instead of the
default sdbm hashing resulted is better workload distribution to our backends.
This commit implements a change, that allows the user to specify the hash
function they want to use. It does not limit itself to consistent hashing
scenarios.
The supported hash functions are sdbm (default), and djb2.
For a discussion of the feature and analysis, see mailing list thread
"Consistent hashing alternative to sdbm" :
http://marc.info/?l=haproxy&m=138213693909219
Note: This change does NOT make changes to new features, for instance,
applying an avalance hashing always being performed before applying
consistent hashing.
A call to free_pattern_tree() upon exit() is made to free all ACL
patterns allocated in a tree (strings or IP addresses). Unfortunately
it happens that this function has been bogus from the beginning, it
walks over the whole tree, frees the nodes but forgets to remove them
from the tree prior to freeing them. So after visiting a leaf, the
next eb_next() call will require to revisit some of the upper nodes
that were just freed. This can remain unnoticed for a long time because
free() often just marks the area as free. But in cases of aggressive
memory freeing, the location will not be mapped anymore and the process
segfaults.
Note that the bug has no impact other than polluting kernel logs and
frightening sysadmins, since it happens just before exit().
Simply adding the debug code below makes it easier to reproduce the
same bug :
while (node) {
next = eb_next(node);
+ node->node_p = (void *)-1;
free(node);
node = next;
}
Many thanks to the StackExchange team for their very detailed bug report
that permitted to quickly understand this non-obvious bug!
This fix should be backported to 1.4 which introduced the bug.
If the dumped length of buffer is not multiple of 16, the last output line can
be seen as below:
Dumping contents from byte 0 to byte 125
0 1 2 3 4 5 6 7 8 9 a b c d e f
0000: 47 45 54 20 2f 69 6e 64 - 65 78 2e 68 74 6d 20 48 GET /index.htm H
0010: 54 54 50 2f 31 2e 30 0d - 0a 55 73 65 72 2d 41 67 TTP/1.0..User-Ag
...
0060: 30 0d 0a 43 6f 6e 6e 65 - 63 74 69 6f 6e 3a 20 4b 0..Connection: K
0070: 65 65 70 2d 41 6c 69 76 - 65 0d 0a 0d 0a eep-Alive....
Yes, the hex column will be overlapped by the text column. Both the hex and
text column should be aligned at their own area as below:
Dumping contents from byte 0 to byte 125
0 1 2 3 4 5 6 7 8 9 a b c d e f
0000: 47 45 54 20 2f 69 6e 64 - 65 78 2e 68 74 6d 20 48 GET /index.htm H
0010: 54 54 50 2f 31 2e 30 0d - 0a 55 73 65 72 2d 41 67 TTP/1.0..User-Ag
...
0060: 30 0d 0a 43 6f 6e 6e 65 - 63 74 69 6f 6e 3a 20 4b 0..Connection: K
0070: 65 65 70 2d 41 6c 69 76 - 65 0d 0a 0d 0a eep-Alive....
Signed-off-by: Godbach <nylzhaowei@gmail.com>
It's quite common to write directives like the following :
tcp-request reject if WAIT_END { sc0_inc_gpc0 }
This one will never reject, because sc0_inc_gpc0 is provided no value
to compare against. The proper form should have been something like this :
tcp-request reject if WAIT_END { sc0_inc_gpc0 gt 0 }
or :
tcp-request reject if WAIT_END { sc0_inc_gpc0 -m found }
Now we detect the absence of any argument on the command line and emit
a warning suggesting alternatives or the use of "--" to really avoid
matching anything (might be used when debugging).
When a condition does something like :
action if A B C || D E F
If B returns a miss (can't tell true or false), C must not
be evaluated. This is important when C has a side effect
(eg: sc*_inc_gpc0). However the second part after the ||
can still be evaluated.
The track-sc* tcp rules are bogus. The test to verify if the
tracked counter was already assigned is performed in the same
condition as the test for the action. The effect is that a
rule which tracks a counter that is already being tracked
is implicitly converted to an accept because the default
rule is an accept.
This bug only affects 1.5-dev releases.
In session_accept(), if we face a memory allocation error, we try to
emit an HTTP 500 error message in HTTP mode. The problem is that we
must not use http_error_message() for this since it dereferences the
session which can be NULL in this case.
We don't need the session to build the error message anyway since
this function only uses it to retrieve the backend and frontend to
get the most suited error message. Let's pick it ourselves, we're
at the beginning of the session, only the frontend is relevant.
This bug is 1.5-specific.
Currently url_decode returns 1 or 0 depending on whether it could decode
the string or not. For some future use cases, it will be needed to get the
decoded string length after a successful decoding, so let's make it return
that value, and fall back to a negative one in case of error.
If haproxy is compiled with the USE_PCRE_JIT option, the length of the
string is used. If it is compiled without this option the function doesn't
use the length and expects a null terminated string.
The prototype of the function is ambiguous, and depends on the
compilation option. The developer can think that the length is always
used, and many bugs can be created.
This patch makes sure that the length is used. The regex_exec function
adds the final '\0' if it is needed.
William Lallemand reported a bug which happens when an ACL keyword using an
implicit argument (eg: a proxy name) is used : the keyword is not properly
set in the arglist field, resulting in an error about the previous keyword
being returned, or "(null)" if the faulty ACL appears first.
The bug only affects error reporting and is 1.5-specific, so no backport is
nedeed.
Bertrand Jacquin reported a but when using tcp_request content rules
on large POST HTTP requests. The issue is that smp_prefetch_http()
first tries to validate an input buffer, but only if the buffer is
not full. This test is wrong since it must only be performed after
the parsing has failed, otherwise we don't accept POST requests which
fill the buffer as valid HTTP requests.
This bug is 1.5-specific, no backport needed.
The current file "regex.h" define an abstraction for the regex. It
provides the same struct name and the same "regexec" function for the
3 regex types supported: standard libc, basic pcre and jit pcre.
The regex compilation function is not provided by this file. If the
developper wants to use regex, he must write regex compilation code
containing "#define *JIT*".
This patch provides a unique regex compilation function according to
the compilation options.
In addition, the "regex.h" file checks the presence of the "#define
PCRE_CONFIG_JIT" when "USE_PCRE_JIT" is enabled. If this flag is not
present, the pcre lib doesn't support JIT and "#error" is emitted.
Though si_conn_send_loop() does not loop over ->snd_buf() after commit ed7f836,
there is still some codes left which use `while` but only execute once. This
commit does the cleanup job and rename si_conn_send_loop() to si_conn_send().
Signed-off-by: Godbach <nylzhaowei@gmail.com>
RFC6125 does not specify if wildcard matches empty strings but
classical browsers implementations does.
After the fix foo*bar.exemple.om matches foobar.exemple.com.
Both static-rr and hash with type map-based call init_server_map() to allocate
server map, so the server map should be freed while doing cleanup if one of
the above load balance algorithms is used.
Signed-off-by: Godbach <nylzhaowei@gmail.com>
[wt: removed the unneeded "if" before the free]
This minor bug was found using the coccinelle script "da.cocci". The
len was initialized twice instead of setting the size. It's harmless
since no operations are performed on this empty string but needs to
be fixed anyway.
At the moment, HTTP response time is computed after response headers are
processed. This can misleadingly assign to the server some heavy local
processing (eg: regex), and also prevents response headers from passing
information related to the response time (which can sometimes be useful
for stats).
Let's retrieve the reponse time before processing the headers instead.
Note that in order to remain compatible with what was previously done,
we disable the response time when we get a 502 or any bad response. This
should probably be changed in 1.6 since it does not make sense anymore
to lose this information.
health_adjust() should requeue the task after changing its expire timer.
I noticed it on devel servers without load. We have long inter (10 seconds)
and short fasinter (100ms). But according to webserver logs, after a failed
request next check request was called with same 10s interval.
This patch should probably be backported to 1.4 which has the same feature.
This fetch method returns the response buffer len, similarly
to req.len for the request. Previously it was only possible
to rely on "res.payload(0,size) -m found" to find if at least
that amount of data was available, which was a bit tricky.
This new action immediately closes the connection with the server
when the condition is met. The first such rule executed ends the
rules evaluation. The main purpose of this action is to force a
connection to be finished between a client and a server after an
exchange when the application protocol expects some long time outs
to elapse first. The goal is to eliminate idle connections which
take signifiant resources on servers with certain protocols.
When a process with large stick tables is replaced by a new one and remains
present until the last connection finishes, it keeps these data in memory
for nothing since they will never be used anymore by incoming connections,
except during syncing with the new process. This is especially problematic
when dealing with long session protocols such as WebSocket as it becomes
possible to stack many processes and eat a lot of memory.
So the idea here is to know if a table still needs to be synced or not,
and to purge all unused entries once the sync is complete. This means that
after a few hundred milliseconds when everything has been synchronized with
the new process, only a few entries will remain allocated (only the ones
held by sessions during the restart) and all the remaining memory will be
freed.
Note that we carefully do that only after the grace period is expired so as
not to impact a possible proxy that needs to accept a few more connections
before leaving.
Doing this required to add a sync counter to the stick tables, to know how
many peer sync sessions are still in progress in order not to flush the entries
until all synchronizations are completed.
David Berard reported that send-proxy was broken on FreeBSD and tracked the
issue to be an error returned by send(). We already had the same issue in
the past in another area which was addressed by the following commit :
0ea0cf6 BUG: raw_sock: also consider ENOTCONN in addition to EAGAIN
In fact, on Linux send() returns EAGAIN when the connection is not yet
established while other OSes return ENOTCONN. Let's consider ENOTCONN for
send-proxy there as the same as EAGAIN.
David confirmed that this change properly fixed the issue.
Another place was affected as well (health checks with send-proxy), and
was fixed.
This fix does not need any backport since it only affects 1.5.
verifyhost allows you to specify a hostname that the remote server's
SSL certificate must match. Connections that don't match will be
closed with an SSL error.
With a facily of 2 or 1 digit, the send size was wrong and bytes with
unknown value were sent.
The size was calculated using the start of the buffer and not the start
of the data which varies with the number of digits of the facility.
This bug was reported by Samuel Stoller and reported by Lukas Tribus.
When a request fail, the unique_id was allocated but not generated.
The string was not initialized and junk was printed in the log with %ID.
This patch changes the behavior of the unique_id. The unique_id is now
generated when a request failed.
This bug was reported by Patrick Hemmer.
The HTTP request counter is incremented non atomically, which means that
many requests can log the same ID. Let's increment it when it is consumed
so that we avoid this case.
This bug was reported by Patrick Hemmer. It's 1.5-specific and does not
need to be backported.
Mathew Levett reported an issue which is a bit nasty and hard to track
down. RDP cookies contain both the IP and the port, and haproxy matches
them exactly. So if a server has no port specified (or a remapped port),
it will never match a port specified in a cookie. Better warn the user
when this is detected.
Apollon Oikonomopoulos reported a build failure on Hurd where PATH_MAX
is not defined. The only place where it is referenced is ssl_sock.c,
all other places use MAXPATHLEN instead, with a fallback to 128 when
the OS does not define it. So let's switch to MAXPATHLEN as well.
Mark Brooks reported the following issue :
"My table looks like this -
0x24a8294: key=192.168.136.10 use=0 exp=1761492 server_id=3
0x24a8344: key=192.168.136.11 use=0 exp=1761506 server_id=2
0x24a83f4: key=192.168.136.12 use=0 exp=1761520 server_id=3
0x24a84a4: key=192.168.136.13 use=0 exp=1761534 server_id=2
0x24a8554: key=192.168.136.14 use=0 exp=1761548 server_id=3
0x24a8604: key=192.168.136.15 use=0 exp=1761563 server_id=2
0x24a86b4: key=192.168.136.16 use=0 exp=1761580 server_id=3
0x24a8764: key=192.168.136.17 use=0 exp=1761592 server_id=2
0x24a8814: key=192.168.136.18 use=0 exp=1761607 server_id=3
0x24a88c4: key=192.168.136.19 use=0 exp=1761622 server_id=2
0x24a8974: key=192.168.136.20 use=0 exp=1761636 server_id=3
0x24a8a24: key=192.168.136.21 use=0 exp=1761649 server_id=2
im running the command -
socat unix-connect:/var/run/haproxy.stat stdio <<< 'clear table VIP_Name-2 data.server_id eq 2'
Id assume that the entries with server_id = 2 would be removed but its
removing everything each time."
The cause of the issue is a missing test for skip_entry when deciding
whether to clear the key or not. The test was present when only the
last node is to be removed, so removing only the first node from a
list of two always did the right thing, explaining why it remained
unnoticed in basic unit tests.
The bug was introduced by commit 8fa52f4e which attempted to fix a
previous issue with this feature where only the last node was removed.
This bug is 1.5-specific and does not require any backport.
Such load balance algorithms as roundrobin, leastconn and first will check the
server after being selected with the following condition:
if (!s->maxconn || (!s->nbpend && s->served < srv_dynamic_maxconn(s)))
But static-rr uses the different one in map_get_server_rr() as below:
if (!srv->maxconn || srv->cur_sess < srv_dynamic_maxconn(srv))
After viewing this difference, it is a better choice for static-rr to use the
same check condition as other algorithms.
This change will only affect static-rr. Though all hash algorithms with type
map-based will use the same server map as static-rr, they call another function
map_get_server_hash() to get server.
Signed-off-by: Godbach <nylzhaowei@gmail.com>
When using req.payload and res.payload to look up for specific content at an
arbitrary location, we're often facing the problem of not knowing the input
buffer length. If the length argument is larger than the buffer length, the
function did not match, and if they're smaller, there is a risk of not getting
the expected content. This is especially true when looking for data in SOAP
requests.
So let's make some provisions for scanning the whole buffer by specifying a
length of 0 bytes. This greatly simplifies the processing of random-sized
input data.
The "set table" statement allows to create new entries with their respective
values. Till now it was limited to a single data type per line, requiring as
many "set table" statements as the desired data types to be set. Since this
is only a parser limitation, this patch gets rid of it. It also allows the
creation of a key with no data types (all reset to their default values).
Since commit 654694e1, it has been possible to feed some data into
stick tables from the CLI. That commit considered that frequency
counters would only have their previous value set, so that they
progressively fade out. But this does not match any real world
use case in fact. The only reason for feeding a freq counter is
to pass some data learned outside. We certainly don't want to see
such data start to vanish immediately, otherwise it will force the
external scripts to loop very frequently to limit the losses.
So let's set the current value instead in order to guarantee that
the data remains stable over the full period, then starts to fade
out between 1* and 2* the period.
sc_* sample fetches now take an optional parameter which allows to look
the key in an alternate table. This is convenient to pass multiple
information for the same key at once (eg: have multiple gpc0 for the
same key, or support being fed complementary information from the CLI).
Example :
listen front
bind :8000
tcp-request content track-sc0 src table local-ip
http-response set-header src-id %[sc0_get_gpc0]+%[sc0_get_gpc0(global-ip)]
server dummy 127.0.0.1:8001
backend local-ip
stick-table size 1k type ip store gpc0
backend global-ip
stick-table size 1k type ip store gpc0
One very annoying issue when trying to extend the sticky counters beyond
the current 3 counters is that it requires a massive copy-paste of fetch
functions (we don't have to copy-paste code anymore), just so that the
fetch names exist.
So let's have an alternate form like "sc_*(num)" to allow passing the
counter number as an argument without having to redefine new fetch names.
The MAX_SESS_STKCTR macro defines the number of usable sticky counters,
which defaults to 3.
In preparation of more flexibility in the stick counters, make their
number configurable. It still defaults to 3 which is the minimum
accepted value. Changing the value alone is not sufficient to get
more counters, some bitfields still need to be updated and the TCP
actions need to be updated as well, but this update tries to be
easier, which is nice for experimentation purposes.
smp_fetch_sc0_trackers, smp_fetch_sc1_trackers and smp_fetch_sc2_trackers
were merged into a single function which relies on the fetch name to decide
what to return.
This is also a bug fix for this feature which has never worked till its bogus
introduction by commit "2406db4 MEDIUM: counters: add sc1_trackers/sc2_trackers"
(1.5-dev10).
Instead of returning the value in the sample, it was returned as the fetch
result!
There is no need to backport this fix anyway since it's 1.5-specific and
nobody uses the feature.
smp_fetch_sc0_bytes_out_rate, smp_fetch_sc1_bytes_out_rate, smp_fetch_sc2_bytes_out_rate,
smp_fetch_src_bytes_out_rate and smp_fetch_bytes_out_rate were merged into a single
function which relies on the fetch name to decide what to return.
smp_fetch_sc0_kbytes_out, smp_fetch_sc1_kbytes_out, smp_fetch_sc2_kbytes_out,
smp_fetch_src_kbytes_out and smp_fetch_kbytes_out were merged into a single
function which relies on the fetch name to decide what to return.
smp_fetch_sc0_bytes_in_rate, smp_fetch_sc1_bytes_in_rate, smp_fetch_sc2_bytes_in_rate,
smp_fetch_src_bytes_in_rate and smp_fetch_bytes_in_rate were merged into a single
function which relies on the fetch name to decide what to return.
smp_fetch_sc0_kbytes_in, smp_fetch_sc1_kbytes_in, smp_fetch_sc2_kbytes_in,
smp_fetch_src_kbytes_in and smp_fetch_kbytes_in were merged into a single
function which relies on the fetch name to decide what to return.
