|
NAME | C SYNOPSIS | DESCRIPTION | EXAMPLE | METRIC LOOKUP | METRIC CHANGES | DIAGNOSTICS | CAVEAT | SEE ALSO | COLOPHON |
|
PMDAINIT(3) Library Functions Manual PMDAINIT(3)
pmdaInit, pmdaRehash, pmdaSetFlags, pmdaExtSetFlags - initialize a
PMDA
#include <pcp/pmapi.h>
#include <pcp/pmda.h>
void pmdaInit(pmdaInterface *dispatch, pmdaIndom *indoms,
int nindoms, pmdaMetric *metrics, int nmetrics);
void pmdaRehash(pmdaExt *pmda, pmdaMetric *metrics, int nmetrics);
void pmdaSetFlags(pmdaInterface *dispatch, int flags);
void pmdaExtSetFlags(pmdaExt *pmda, int flags);
cc ... -lpcp_pmda -lpcp
pmdaInit initializes a PMDA so that it is ready to receive PDUs from
pmcd(1). The function expects as arguments the instance domain table
(indoms) and the metric description table (metrics) that are
initialized by the PMDA. The arguments nindoms and nmetrics should be
set to the number of instances and metrics in the tables,
respectively.
Much of the pmdaInterface structure can be automatically initialized
with pmdaDaemon(3), pmdaGetOpt(3) and pmdaDSO(3). pmdaInit completes
the PMDA initialization phase with three operations. The first
operation adds the domain and instance numbers to the instance and
metric tables. Singular metrics (metrics without an instance domain)
should have the instance domain PM_INDOM_NULL set in the indom field
of the pmDesc structure (see pmLookupDesc(3)). Metrics with an
instance domain should set this field to be the serial number of the
instance domain in the indoms table.
The instance domain table may be made empty by setting indoms to NULL
and nindoms to 0. This allows the caller to provide custom Fetch and
Instance callback functions. The metric table may be made empty by
setting metrics to NULL and nmetrics to 0. This allows the caller to
provide custom Fetch and Descriptor callback functions.
For example, a PMDA has three metrics: A, B and C, and two instance
domains X and Y, with two instances in each instance domain. The
instance domain and metrics description tables could be defined as:
static pmdaInstid _X[] = {
{ 0, "X1" }, { 1, "X2" }
};
static pmdaInstid _Y[] = {
{ 0, "Y1" }, { 1, "Y2" }
};
static pmdaIndom indomtab[] = {
#define X_INDOM 0
{ X_INDOM, 2, _X },
#define Y_INDOM 3
{ Y_INDOM, 2, _Y }
};
static pmdaMetric metrictab[] = {
/* A */
{ (void *)0,
{ PMDA_PMID(0,0), PM_TYPE_U32, PM_INDOM_NULL, PM_SEM_INSTANT,
{ 0,0,0,0,0,0} }, },
/* B */
{ (void *)0,
{ PMDA_PMID(0,1), PM_TYPE_U32, X_INDOM, PM_SEM_INSTANT,
{ 0,0,0,0,0,0} }, },
/* C */
{ (void *)0,
{ PMDA_PMID(0,2), PM_TYPE_DOUBLE, Y_INDOM, PM_SEM_INSTANT,
{ 0,1,0,0,PM_TIME_SEC,0} }, }
};
The metric description table defines metric A with no instance
domain, metric B with instance domain X and metric C with instance
domain Y. Metric C has units of seconds, while the other metrics
have no units (simple counters). pmdaInit will take these structures
and assign the PMDA(3) domain number to the it_indom field of each
instance domain. This identifier also replaces the indom field of
all metrics which have that instance domain, so that they are
correctly associated.
The second stage opens the help text file, if one was specified with
the -h command line option (see pmdaGetOpt(3)) or as a helptext
argument to pmdaDSO(3) or pmdaDaemon(3).
The final stage involves preparing the metric table lookup strategy.
When fetch and descriptor requests are made of the PMDA, each
requested PMID must be mapped to a metric table entry. There are
currently three strategies for performing this mapping - direct,
linear and hashed. Each has its own set of tradeoffs and an
appropriate strategy should be selected for each PMDA.
If all of the metric PMID item numbers correspond to the position in
the metrics table, then direct mapping is used. This is the most
efficient of the lookup functions as it involves a direct array index
(no additional memory is required nor any additional processing
overhead). If the PMID numbering requirement is met by the PMDA, it
is ideal. This strategy can be explicitly requested by calling
pmdaSetFlags(pmda, PMDA_EXT_FLAG_DIRECT) before calling pmdaInit. In
this case, if the direct mapping is not possible (e.g. due to an
oversight on the part of the PMDA developer), a warning is logged and
the linear strategy is used instead.
The second strategy (linear search) is the default, when a direct
mapping cannot be established. This provides greater flexibility in
the PMID numbering scheme, as the PMDA item numbers do not have to be
unique (hence, the PMID cluster numbers can be used more freely,
which is often extremely convenient for the PMDA developer).
However, lookup involves a linear walk from the start of the metric
table until a matching PMID is found, for each requested PMID in a
request.
The third strategy (hash lookup) can be requested by calling
pmdaSetFlags(pmda, PMDA_EXT_FLAG_HASHED) before calling pmdaInit.
This strategy is most useful for PMDAs with large numbers of metrics
(many hundreds, or thousands). Such PMDAs will almost always use the
cluster numbering scheme, so the direct lookup scheme becomes
inappropriate. They may also be prepared to sacrifice a small amount
of additional memory for a hash table, mapping PMID to metric table
offsets, to speed up lookups in their vast metric tables.
This final strategy can also be used by PMDAs serving up dynamically
numbered metrics. For this case, the pmdaRehash function should be
used to replace the metric table when new metrics become available,
or existing metrics are removed. The PMID hash mapping will be
recomputed at the same time that the new metric table is installed.
It should be well understood by PMDA authors that metric metadata for
individual metrics is fixed, and ideally would not ever change. In
the situation where metadata is incorrect and is updated, such a
change requires correction to logged metrics using pmlogrewrite(1),
and as a result should be avoided whenever possible.
However, a PMDA may become aware of new domain metrics at runtime,
and in this case it is ideal to export them immediately (without any
collector system restart). In this situation, the PMDA can inform
all running PMAPI clients that may have already explored the metric
namespace (for example, using pmTraversePMNS(3)) of the change to the
metric namespace.
This is achieved using pmdaSetFlags(pmda, PMDA_EXT_NAMES_CHANGE)
which will result in the PMCD_NAMES_CHANGE state change notification
being sent to each PMAPI client on next fetch. If the newly
discovered metrics have label metadata associated, then the
PMDA_EXT_LABEL_CHANGE flag may also be set, which will result in the
PMCD_LABEL_CHANGE notification being sent as well.
pmdaExtSetFlags is equivalent to pmdaSetFlags, and is provided as a
convenience interface in situations where the pmdaExt is more readily
available than the pmdaInterface structure.
pmdaInit will set dispatch->status to a value less than zero if there
is an error that would prevent the PMDA(3) from successfully running.
pmcd(1) will terminate the connection to the PMDA(3) if this occurs.
pmdaInit may issue any of these messages:
PMDA interface version interface not supported
The interface version is not supported by pmdaInit.
Using pmdaFetch() but fetch call back not set
The fetch callback, pmdaFetch(3), requires an
additional callback to be provided using
pmdaSetFetchCallBack(3).
Illegal instance domain inst for metric pmid
The instance domain inst that was specified for metric
pmid is not within the range of the instance domain
table.
No help text path specified
The help text callback, pmdaText(3), requires a help
text file for the metrics to have been opened, however
no path to the help text was specified as a command
line option, or as an argument to pmdaDSO(3) or
pmdaDaemon(3). This message is only a warning.
Direct mapping for metrics disabled @ num
The unit numbers of the metrics did not correspond to
the index in the metric description table. The direct
mapping failed for metric number num in the metrics
table. This is less efficient but is not fatal and
the message is only a warning.
Hashed mapping for metrics disabled @ num
A memory allocation failure occurred while building
the hash table to index the metric description table.
This is a non-fatal warning message - a fallback to
linear searching will be automatically performed
should this situation arise.
The PMDA must be using PMDA_INTERFACE_2 or later, as specified in the
call to pmdaDSO(3) or pmdaDaemon(3) to use pmdaInit.
The PMDA must use PMDA_INTERFACE_7 or later to issue state change
notifications using pmdaSetFlags or pmdaExtSetFlags.
newhelp(1), pmcd(1), pmlogrewrite(1), PMAPI(3), PMDA(3),
pmdaDaemon(3), pmdaDSO(3), pmdaFetch(3), pmdaGetOpt(3), pmdaText(3),
pmLookupDesc(3) and pmTraversePMNS(3).
This page is part of the PCP (Performance Co-Pilot) project.
Information about the project can be found at ⟨http://www.pcp.io/⟩.
If you have a bug report for this manual page, send it to
pcp@groups.io. This page was obtained from the project's upstream
Git repository ⟨https://github.com/performancecopilot/pcp.git⟩ on
2018-02-02. (At that time, the date of the most recent commit that
was found in the repository was 2018-02-02.) If you discover any
rendering problems in this HTML version of the page, or you believe
there is a better or more up-to-date source for the page, or you have
corrections or improvements to the information in this COLOPHON
(which is not part of the original manual page), send a mail to
man-pages@man7.org
Performance Co-Pilot PCP PMDAINIT(3)
Pages that refer to this page: pmda(3), pmdacache(3), pmdaconnect(3), pmdadso(3), pmdafetch(3), pmdahelp(3), pmdalabel(3), pmdatext(3)