Florian Forster [Sat, 25 May 2013 07:46:36 +0000 (09:46 +0200)]
src/plugin.c: Make plugin_load() idempotent, i.e. don't load plugins twice.
Previously, multiple 'LoadPlugin "foo"' lines would dlopen the same .so file
twice and call its module_register function twice. This would eventually result
in complaints about callbacks with the same name being registered. With this
patch each plugin is only loaded only once, allowing people to write "self
contained" configuration segments.
Florian Forster [Sat, 11 May 2013 07:55:09 +0000 (09:55 +0200)]
mic plugin: Use the "core number" as plugin instance.
Previously, per-core CPU statistics would be dispatched as (e.g.):
<host>/mic-0/cpu-1-user
This makes graping hard, because it required to parse the type instance to
seperate out different graphs. The core number has therefore been moved to the
plugin instance, so the "mic" plugin behaves more like the "cpu" plugin, e.g.:
Florian Forster [Sat, 11 May 2013 07:40:05 +0000 (09:40 +0200)]
mic plugin: Remove the NUM_THERMS define.
Also some coding style changes: Limit line length, move temporary variables
to an innter scope, print the temperature name in an error message (rather
than its ID).
* Avoid mixed declarations.
* Use parse_value() rather than atoll().
* Comment static arguments to walk_directory().
* Return an error if the read() function can't locate the cgroup
mount point.
volume plugin: Fix a couple of issues from the code review.
* Use df_complex as data type.
* Submit LV sizes and "free", but not "total" and "used".
* Fix error handling; don't use variable if it is NULL, remove calls to
exit().
Build system: Don't ship the generated *.pb-c.[ch] files.
They depend on specific versions of the protobuf-c.h header file, so
shipping this will cause problems for people having a different version
of protoc-c installed. This reverts some of the work done in #269, but
distcheck should still work. Meh.
write_riemann plugin: Use the "Attributes" of an Event rather than abuse the "Tags".
Riemann 0.2.0 provides "Attributes", which are key-value pairs. Since
the plugin previously basically abused the tag system for adding this
kind of data, switch to the new and shiny thing :)
Chad Malfait [Sun, 7 Apr 2013 01:32:27 +0000 (19:32 -0600)]
added the use of df_complex and changed fprintf to ERROR for error
messages. Also changed variable names to follow recommended standard
from the website.
configure.in: Fix dependency of the pinba and write_riemann plugins.
Since the generated *.pb-c.[ch] files are now included in the
distribution, they only depend on the <google/protobuf-c/protobuf-c.h>
header file, not the protobuf compiler.
Sebastian Harl [Tue, 20 Nov 2012 13:44:43 +0000 (14:44 +0100)]
build system (Perl bindings): Make Makefile.PL's PREFIX/INSTALL_BASE optional.
Both options may cause problems:
- PREFIX might clash with INSTALL_BASE being set in $PERL_MM_OPT.
- INSTALL_BASE without any further options set will install into non-standard
directories in many cases.
Rather, use PREFIX=${prefix} as default but overwrite that if
--with-perl-bindings=... has been used.
Thanks to faxm0dem for reporting the INSTALL_BASE problems on IRC and
bvarner and nirik for reporting this in GH #177 and #196.
network plugin: Make sure gcrypt is initialized before using any of its functions.
@marekbecka found that sockent_open() is called during the configuration
phase, but gcrypt is initialized later, during the initialization phase.
This patch will initialize gcrypt in sockent_open() if any sockets with
security are configured.
- Java detection used TMPDIR as var to parse command output
- TMPDIR is used in libltdl later on in ./configure
- alternative fix is local var or subshell
- fixes issue introduced via #33163ee
This patch changes the snmp GETNEXT request behaviour implemented in snmp.c.
The old implementation requested all OIDs using GETNEXT requests until all OIDs
left their own subtree. In cases were trees in a Data template are much longer
than other trees the shorter subtrees were re-requested over and over again.
The new implementation will only request OIDs that did not already leave their
subtrees (see the oid_todo_list implementation for details). This renders the
function csnmp_check_res_left_subtree useless as the oid_todo_list keeps track
if all OIDs have finished.
During tests against Cat6500 (CatOS/IOS) as well as Nexus5k (NX-OS) it looks as
though GETNEXT requests (when requesting multiple OIDs like all 14 dot3Stats
errors from Etherlike-MIB) can take about 5-10ms (CatOS 30ms) longer if they wrap
to the next OID.
This does not sound much but when collecting data for the Etherlike-MIB (that only
has entries for physical interfaces) with a collectd "Instance" variable in IF-MIB
(that has entries for all physical as well as pseudo [SVIs, VLANs, ...] interfaces)
this can make a notable difference (e.g. for core routers that have all SVIs and
VLANs but only some switches attached):
After refactoring csnmp_read_table "Instance" and "Value" OIDs are now handled
consistently (so no pointer-forward foo needed). It doesn't change any logic
and data structures, though - so there should not be any impact to other
functions.
The refactored code also fixes GitHub bugs #235 and #258. This bug is due to
reusing the status variable in following code section which might lead to errors
if the subtrees are of different length:
1436 /* Calculate the current suffix. This is later used to check that the
1437 * suffix is increasing. This also checks if we left the subtree */
1438 status = csnmp_oid_suffix (&suffix, &vb_name, data->values + i);
projects / thirdparty / collectd.git / log
