The filter was using "now" in visible output in debug mode, that's
not correct, we should rather use "date" since it's visible. No
backport is needed as it was mostly emphasized by commit 28360dc
("MEDIUM: clock: force internal time to wrap early after boot")
in 2.8..
Since commit 3a4bedccc ("MEDIUM: vars: replace the global name index with
a hash") the names of HAProxy variables are no longer saved, ie their
64-bit hashes are saved instead.
This is very convenient for the HAProxy itself, but for the OpenTracing
module it is a problem because the names of the variables are important
when transferring the OpenTracing context. Namely, this context consists
of an unknown amount of data stored in a key-value format. The number
of these data (and the name of the variable used for this purpose) is
determined with the configuration of the OpenTracing filter, as well as
with the tracer used. The two previous sentences seem to be in conflict,
but that is only so at first glance. The function in the OpenTracing
filter used to read the context does not really know this, neither their
number nor its name. The only thing that function actually knows is the
prefix of the variable names used for context transfer, and by that it
could find all the necessary data. Of course, until the application of
the above-mentioned commit.
The problem is solved in a very simple way: in a common variable that
the filter always knows its name, the names of all variables that are the
product of the OpenTracing context are saved. The names of these context
variables can only be added to that common variable. When that variable
is no longer needed (when we no longer need context), it is deleted.
The format for saving data to this common variable is as follows:
+-----+---------------+-- .. --+-----+---------------+
| len | variable name | | len | variable name |
+-----+---------------+-- .. --+-----+---------------+
The amount of memory space used to store the length of the name is 1 byte,
with a sign (the minus sign is provided for inactive records, but this is
not currently used). This means that the maximum length of the variable
name to be saved is 127 characters, which is quite enough for use in the
filter. The buffer size for such data storage is global.tune.bufsize.
This patch must be backported in 2.5.
A display of the contents of the err variable has been added to the
FLT_OT_ERR() macro, once it has been set.
This patch must be backported as far as 2.4.
It's always good to replace "hard-coded" values with something that looks
less "hard-coded" (even though that doesn't change anything in the code).
This is done here using FLT_OT_PARSE_INVALID_enum enum.
This patch must be backported as far as 2.4.
If the OpenTracing filter is compiled using the 'OT_DEBUG=1' option, then
log messages are printed to stderr when the filter is running. In the log
one can then find (among other things) the order in which the function is
called and the value that the function returns (if it is not a void type).
Prior to applying this patch, no value returned by a function was logged.
Log output example:
[ 1] 0.038807 [OT]: flt_ot_init_per_thread(0x56365bd45ec0, 0x56365bd48210) {
[ 1] 0.038807 [OT]: ot_start(0x56365bd58920, 0x56365bd4e3a0, 0x7f561acba168:(nil)) {
[ 1] 0.038807 [OT]: } = 0
[ 1] 0.038807 [OT]: } = 0
This patch must be backported as far as 2.4.
This reverts commit 560c7b874a.
The ot.uuid variable should have the 'sess' scope because it is created
when an OpenTracing filter is attached to a stream. After that, the
stream processing is started and on that occasion the contexts for the
variables that have the range 'txn' and 'req' are initialized. This
means that we cannot use variables with the specified scopes before that
point.
This patch must be backported in 2.5.
The flt_ot_var_get() function is not used anywhere and is unnecessary
in the existing implementation of the OpenTracing filter.
This patch must be backported as far as 2.4.
The flt_ot_var_unset() function is not used anywhere and is unnecessary
in the existing implementation of the OpenTracing filter.
This patch must be backported as far as 2.4.
There is currently a problem related to time keeping. We're mixing
the functions to perform calculations with the os-dependent code
needed to retrieve and adjust the local time.
This patch extracts from time.{c,h} the parts that are solely dedicated
to time keeping. These are the "now" or "before_poll" variables for
example, as well as the various now_*() functions that make use of
gettimeofday() and clock_gettime() to retrieve the current time.
The "tv_*" functions moved there were also more appropriately renamed
to "clock_*".
Other parts used to compute stolen time are in other files, they will
have to be picked next.
At the suggestion of Willy Tarreau, the scope of the 'ot.uuid' variable was
changed from 'sess' to 'txn', so it is now limited to the transaction only.
To avoid duplicate source code, the original haproxy function is used to
generate the OpenTracing runtime context UUID.
Also, the structure flt_ot_runtime_context is simplified because the
detailed definition of UUID is removed from it (struct flt_ot_uuid),
ie the UUID is left only in the form of a string.
Due to a recent change in the handling of haproxy variables, their use for
OpenTracing context transfer has been excluded from the compilation process.
The use of variables can be re-enabled if the newly defined variable
OT_USE_VARS is set to 1 when calling the 'make' utility. However,
this should not be used for now as the compilation will end in error.
This change prevents the use of haproxy variables to convey the OpenTracing
context. This means that the 'use-vars' parameter cannot be used in the
OpenTracing filter configuration for 'inject' and 'extract' operations.
An example configuration that uses this feature is in the test/ctx
directory, while the script to run that test is test/run-ctx.sh.
Then, the 'sess.ot.uuid' variable is no longer set when initializing the
OpenTracing session. This means that this variable can still be used in
the OpenTracing configuration, but its contents will be empty.
This patch solves the problem reported in github issue #1204, where the
OpenTracing filter cannot communicate with the selected tracer if HAProxy
is run in daemon mode.
This commit also solves github issue #1274, where the problem manifests
itself when using the 'chroot' keyword in the HAProxy configuration.
This is solved so that the initialization of the OpenTracing plugin is
split into two operations, first the plugin (dynamic library) is loaded
before switching the HAProxy to daemon mode (or chroot) and then the
tracer thread is started.
This means that nothing is retrieved from the file system in runtime.
After applying this commit, opentracing C wrapper version 1.1.0 should be
used because the earlier version does not have separated initialization
functions.
This resolves GitHub issues #1204 and #1274.
The inclusion of header files proxy.h and tools.h was added to the
addons/ot/include/include.h file. Without this HAProxy cannot be
compiled if the OpenTracing filter is to be used.
The current "ADD" vs "ADDQ" is confusing because when thinking in terms
of appending at the end of a list, "ADD" naturally comes to mind, but
here it does the opposite, it inserts. Several times already it's been
incorrectly used where ADDQ was expected, the latest of which was a
fortunate accident explained in 6fa922562 ("CLEANUP: stream: explain
why we queue the stream at the head of the server list").
Let's use more explicit (but slightly longer) names now:
LIST_ADD -> LIST_INSERT
LIST_ADDQ -> LIST_APPEND
LIST_ADDED -> LIST_INLIST
LIST_DEL -> LIST_DELETE
The same is true for MT_LISTs, including their "TRY" variant.
LIST_DEL_INIT keeps its short name to encourage to use it instead of the
lazier LIST_DELETE which is often less safe.
The change is large (~674 non-comment entries) but is mechanical enough
to remain safe. No permutation was performed, so any out-of-tree code
can easily map older names to new ones.
The list doc was updated.
In order to enable the assignment of a context name, and yet exclude the
use of that name (prefix in this case) when extracting the context from
the HTTP header, a special character '-' has been added, which can be
specified at the beginning of the prefix.
So let's say if we look at examples of the fe-be configuration, we can
transfer the context via an HTTP header without a prefix like this:
fe/ot.cfg:
..
span "HAProxy session"
inject "" use-headers
event on-backend-http-request
Such a context can be read in another process using a name that has a
special '-' sign at the beginning:
be/ot.cfg:
ot-scope frontend_http_request
extract "-ot-ctx" use-headers
span "HAProxy session" child-of "-ot-ctx" root
..
This means that the context name will be '-ot-ctx' but it will not be
used when extracting data from HTTP headers.
Of course, if the context does not have a prefix set, all HTTP headers
will be inserted into the OpenTracing library as context. All of the
above will only work correctly if that library can figure out what is
relevant to the context and what is not.
This one is the last optional module to build with haproxy, so let's move
it to addons/. It was renamed to "ot" as it was the only one whose USE_*
option did not match the directory name, now this is consistent.
Few changes were required, only the Makefile, and doc were adjusted, as
the directory was already self-contained and relocatable.
Willy Tarreau
Miroslav Zagorac