PMDACACHE(3) Library Functions Manual PMDACACHE(3)

pmdaCacheStore, pmdaCacheStoreKey, pmdaCacheLookup, pmdaCacheLookupName, pmdaCacheLookupKey, pmdaCacheOp, pmdaCachePurge, pmdaCachePurgeCallback, pmdaCacheResize - manage a cache of instance domain information for a PMDA

#include <pcp/pmapi.h>
#include <pcp/pmda.h>


int pmdaCacheStore(pmInDom indom, int flags, const char *name, void *private);
int pmdaCacheStoreKey(pmInDom indom, int flags, const char *name, int keylen, const void *key, void *private);
int pmdaCacheLookup(pmInDom indom, int inst, char **name, void **private);
int pmdaCacheLookupName(pmInDom indom, const char *name, int *inst, void **private);
int pmdaCacheLookupKey(pmInDom indom, const char *name, int keylen, const void *key, char **oname, int *inst, void **private);
int pmdaCacheOp(pmInDom indom, int op);
int pmdaCachePurge(pmInDom indom, time_t recent);
int pmdaCachePurgeCallback(pmInDom indom, time_t recent, void (*callback)(void *));
int pmdaCacheResize(pmInDom indom, int maximum);


cc ... -lpcp_pmda -lpcp

The pmdaCache family of routines provide services to support the maintenance of complex instance domains for Performance Co-Pilot PMDAs. There is potentially one cache of information for each instance domain, and for each instance the cache maintains:

  • external instance name (supplied by the PMDA)
  • internal instance identifier (assigned by pmdaCacheStore or calculated from a ``hint'' by pmdaCacheStoreKey)
  • state, where active instances are visible and part of the current instance domain, and inactive instances are hidden, but not forgotten; pmdaCacheStore or pmdaCacheStoreKey may be used to change the state of an instance
  • an optional opaque pointer to data that is associated with the instance, but maintained by the PMDA
  • an optional opaque key that is used as a ``hint'' to pmdaCacheStoreKey when guessing the initial internal instance identifier
  • the last time the cache was saved and the instance had been marked as active at some point since the previous cache load or save operation

The semantics of a PCP instance domain require a number of rules to be followed, namely:

1.
Each internal instance identifier must be unique and in the range 0 to 2^31 - 1. This rule is enforced by the pmdaCache family of routines.
2.
The external instance name must be unique. When the instance name contains a space, it is further constrained such that the name to the left of the first space (the short name) must also be unique. Refer to the INSTANCE NAME MATCHING section below. The PMDA must honor this rule, the pmdaCache family of routines will detect attempts to violate this rule.
3.
Where an external instance name corresponds to some object or entity, there is an expectation that the association between the name and the object is fixed, e.g. ``/dev/hda'' is always the name of the same disk on a particular system. This rule is perhaps the responsibility of the PMDA, but is often a characteristic of the environment in which the PMDA runs.
4.
It is preferable, although not mandatory, for the association between an external instance name and an internal instance identifier to be persistent. This rule is supported by the pmdaCache family of routines.
5.
When opaque keys are used, the values of the keys must be unique across all instances within an instance domain. This rule is enforced by the pmdaCache family of routines.

The visible interface to the cache is oriented towards the PMDA developer who is most concerned about the names of instances, while the details of how the rest of the PCP infrastructure expects the internal instance identifiers to be managed is not relevant.

Instances are updated in the cache for instance domain indom by calling pmdaCacheStore or pmdaCacheStoreKey with the external name of the instance passed via name. The opaque pointer private may be used to associate additional data with the entry in the cache; if no such data is required, private should be NULL. Any manipulation of the additional data (including allocation or freeing) is the responsibility of the PMDA caller, as the cache simply maintains the pointer to the data (passed via private).

The upper bound for identifiers allocated for any given indom cache can be optionally reduced from the default (2^31 - 1) to some lesser maximum, using pmdaCacheResize. This maximum will then be persisted and restored in the usual manner, and can thus be associated permanently with a cache once set. This has applications when using these interfaces as general purpose identifier caches, and is less applicable when using them for instance domain caching.

For cases where the PMDA developer wishes to influence the allocation of internal instance identifiers, e.g. for instance domains with more than one natural dimension, or where there is a desire to allocate the same instance identifier each time the PMDA is started, even on different hosts, pmdaCacheStoreKey may be used. In this case, an initial ``hint'' for the instance identifier is provided as an opaque key via the first keylen bytes in key (which could be any sort of data, including binary values) else if keylen is less than 1 or key is NULL then name is used as the ``hint''. The ``hint'' is hashed to produce an initial instance identifier in the range 0 to 2^31 - 1 (or lesser maximum, if set). If this instance identifier is already allocated, then the value is rehashed. This procedure is repeated until an unallocated instance identifier is found, or pmdaCacheStoreKey gives up and returns PM_ERR_GENERIC. For each instance domain, the ``hint'' must be unique across all instances, else pmdaCacheStoreKey returns PM_ERR_INST.

The flags argument controls how the instance should be processed in the cache as follows:

Insert the entry into the cache if it is not already there and mark it active. If the entry is already in the cache mark it active.
Mark the entry in the cache as inactive, but remember the details of the association between the external instance name and the internal instance identifier. Entries that are inactive will be hidden from cache traversal via PMDA_CACHE_WALK_NEXT operations, but remain visible to pmdaCacheLookup, pmdaCacheLookupName and pmdaCacheLookupKey requests.
Remove the entry from the cache.

On success pmdaCacheStore or pmdaCacheStoreKey will return the internal instance identifier of the associated cache entry. Valid instance identifiers are guaranteed to be unique and non-negative. Failure will be indicated by a negative value (suitable for decoding with pmErrStr(3)) and most likely PM_ERR_INST to indicate the requested instance is not in the cache, or -EINVAL to indicate a potential violation of the short name uniqueness property (see the INSTANCE NAME MATCHING section below).

pmdaCacheLookup is used to search the entries in the cache based on the internal instance identifier inst.

On success the return value will be PMDA_CACHE_ACTIVE or PMDA_CACHE_INACTIVE (depending on the active or inactive state of the cache entry), name (if not NULL) and private (if not NULL) will be set to the external instance name and the associate additional data area as provided when the instance was last activated via pmdaCacheStore or pmdaCacheStoreKey.

pmdaCacheLookup failure is indicated by a negative return value suitable for decoding with pmErrStr(3).

The pmdaCacheLookup interface is required by the PMDA's fetch callback that is registered via pmdaSetFetchCallBack(3). Here the internal instance identifier is passed to the fetch callback to identifier for which instance a value is required. Typical usage is shown in the code fragment below.

static int
foo_callback(pmdaMetric *mdesc, unsigned int inst, pmAtomValue *atom)
{
    mydata   *mdp;
    char     *name;
    int      sts;
    sts = pmdaCacheLookup(mdesc->m_desc.indom, inst, &name, (void **)&mdp);
    /*
     * expect sts == PMDA_CACHE_ACTIVE except for cataclysmic events
     * use mdp as required, name may be useful for diagnostics
     */
    ...


pmdaCacheLookupName is used to search the entries in the cache based on the external instance name name.

On success the return value will be PMDA_CACHE_ACTIVE or PMDA_CACHE_INACTIVE (depending on the active or inactive state of the cache entry), inst (if not NULL) and private (if not NULL) will be set to the internal instance identifier and the associate additional data area as provided when the instance was last activated via pmdaCacheStore or pmdaCacheStoreKey.

pmdaCacheLookupName failure is indicated by a negative return value suitable for decoding with pmErrStr(3).

The pmdaCacheLookupName interface is useful for PMDAs wishing to update an instance domain based on the external instance names.

pmdaCacheLookupKey is used to search the entries in the cache based on an opaque key (or ``hint'') previously used in a call to pmdaCacheStoreKey. The ``hint'' is provided via the first keylen bytes in key. For symmetry with pmdaCacheStoreKey, if keylen is less than 1 or key is NULL then name is used as the ``hint'' (although the results will be the same as calling pmdaCacheLookupName in this case).

On success the return value will be PMDA_CACHE_ACTIVE or PMDA_CACHE_INACTIVE (depending on the active or inactive state of the cache entry), oname (if not NULL), inst (if not NULL) and private (if not NULL) will be set to the external instance name, the internal instance identifier and the associate additional data area as provided when the instance was last activated via pmdaCacheStore or pmdaCacheStoreKey.

pmdaCacheLookupKey failure is indicated by a negative return value suitable for decoding with pmErrStr(3).

To avoid a persistent cache growing without bound, pmdaCachePurge can be used to cull all entries that have not been active in the last recent seconds. For performance reasons, the time accounting is imprecise and the entries are timestamped at the time of the next cache save operation after the entry has been added or marked active (refer to PMDA_CACHE_SAVE, PMDA_CACHE_WRITE and PMDA_CACHE_SYNC below). On success pmdaCachePurge returns the number of culled entries, else in the case of an error the return value is negative (and suitable for decoding with pmErrStr(3)).

The pmdaCachePurgeCallback function is similar to pmdaCachePurge except that a callback function will also be called with the private data pointer associated with the cache entry to be culled. The callback is not made if private is NULL. This would typically be used to free the private data when the associated entry is purged in PMDAs that do not separately maintain any references to the private data.

pmdaCacheOp may be used to perform additional operations on the cache as follows:

The cache can optionally be maintained as a persistent external file, so that the mapping of instance names to instance identifiers is persistent across executions of a PMDA. This operation loads the cache from the external file, and then all new cache entries are marked inactive, and the additional data pointer is set to NULL. Entries loaded from the external file are checked against the current cache contents and if the instance name and instance identifiers match then the state in the cache (active or inactive) is not changed. Should a mismatch be found (same instance name and different instance identifier, or same instance identifier and different instance name, or some but not all of the instance identifier, the instance name and the ``hint'' match) then the entry from the external file is ignored and a warning is issued on stderr. Typically a PMDA would only perform this operation once per execution.
If any instance has been added to, or deleted from, the instance domain since the last PMDA_CACHE_LOAD, PMDA_CACHE_SAVE, PMDA_CACHE_WRITE or PMDA_CACHE_SYNC operation, the entire cache is written to the external file as a bulk operation. This operation is provided for PMDAs that are not interested in using pmdaCachePurge and simply want the external file to reflect the set of known instances without accurate details of when they were last marked active.

Returns the number of instances saved to the external file, else 0 if the external file was already up to date.

A variant of PMDA_CACHE_SAVE where the entire cache is unconditionally written to the external file as a bulk operation, independent of any previous cache operations or the state of the cache.
Annotates this cache as being a special-purpose cache used for string de-duplication in PMDAs exporting large numbers of string valued metrics. This can be used to reduce the memory footprint of the PMDA (duplicate strings hash to the same bucket, and are stored in memory once only). Key comparisons are not terminated at the first space but rather the entire string is used for matching. These are specialised caches not useful for general purpose instance domain handling.
Within an instance domain, if any instance has been added to, or deleted from, or marked active since the last PMDA_CACHE_LOAD, PMDA_CACHE_SAVE, PMDA_CACHE_WRITE or PMDA_CACHE_SYNC operation, the entire cache is written to the external file as a bulk operation. This operation is similar to PMDA_CACHE_SAVE, but will save the instance domain more frequently so the timestamps more accurately match the semantics expected by pmdaCachePurge.

Returns the number of instances saved to the external file, else 0 if the external file was already synchronized.

Returns 1 if a cache exists for the specified instance domain, else 0.
When a new instance is added to the cache, the default strategy is to assign instance identifiers in a monotonic increasing manner. Once the maximum possible instance identifier value has been assigned, the strategy changes to one where starting from 0, the next available unused instance identifier will be used. Calling pmdaCacheOp with PMDA_CACHE_REUSE forces an irreversible change to a second (reuse) strategy where the next unallocated instance identifier will be used. This may be useful in cases where there is a desire to restrict the allocated instance identifiers to smaller values. The prevailing strategy will be saved and restored across PMDA_CACHE_SAVE and PMDA_CACHE_LOAD operations. If pmdaCacheStoreKey is ever used, the associated instance domain will be changed to PMDA_CACHE_REUSE mode.
Reorganize the cache to allow faster retrieval of active entries, at the cost of slower retrieval for inactive entries, and reclaim any culled entries. The cache may be internally re-organized as entries are added, so this operation is not required for most PMDAs.
Prepares for a traversal of the cache in ascending instance identifier sequence.
Fetch the next active instance identifier from the cache. Requires a prior call using PMDA_CACHE_WALK_REWIND and will return -1 when all instances have been processed.

Only one cache walk can be active at any given time, nesting calls to PMDA_CACHE_WALK_NEXT and PMDA_CACHE_WALK_REWIND will interfere with each other.

Changes every inactive entry in the cache to be marked active.
Changes every active entry in the cache to be marked inactive.
Remove every entry from the cache.
Return the number of entries in the cache (includes active, inactive and any culled entries that have not yet been reclaimed).
Return the number of active entries in the cache.
Return the number of inactive entries in the cache.
Dump the current state of the cache on stderr.
Like PMDA_CACHE_DUMP, but also dump the internal hashing structures used to support lookup by instance name, lookup by instance identifier and the collision statistics for ``hint'' hashing from pmdaCacheStoreKey.

pmdaCacheOp returns a non-negative value on success, and failure is indicated by a negative return value (suitable for decoding with pmErrStr(3)).

When the pmdaCache routines are used for particular instance domain, pmdaInstance(3) and the instance domain enumeration behind pmdaFetch(3) will attempt to extract instance domain information from the cache, thereby avoiding reference to the pmdaIndom data structures that have historically been used to define instance domains and service instance requests. A PMDA can adopt a hybrid approach and choose to implement some instance domains via the traditional pmdaIndom method, and others via the pmdaCache approach, however attempts to manage the same instance domain by both methods will result in the pmdaCache method silently prevailing.

If all instances in a PMDA are to be serviced from a pmdaCache then a pmdaIndom is not required, and the pmdaInit(3) call becomes

    pmdaInit(dp, NULL, 0, metrictab, nmetrics);


However, the PMDA will need to explicitly initialize the indom field of the pmDesc in the metrictab entries, as this cannot be done by pmdaInit(3) if indomtab is missing entries for the instance domains maintained in the cache.

Independent of how the instance domain is being maintained, to refresh an instance domain prior to a fetch or an instance domain operation, the standard methods of a ``wrapper'' to the pmdaInstance(3) and pmdaFetch(3) methods should be used.

Refer to the simple PMDA source code for an example use of the pmdaCache routines.

When using pmdaCacheStoreKey, if there is a desire to ensure the given ``hint'' generates the same initial instance identifier across all platforms, then the caller should ensure the endian and word-size issues are considered, e.g. if the natural data structure used for the key is an array of 32-bit integers, then htonl(3) should be used on each element of the array before calling pmdaCacheStoreKey or pmdaCacheLookupKey.

The following table summarizes the ``short name'' matching semantics for an instance domain (caches other than PMDA_CACHE_STRINGS style).

name in pmdaCacheLookup result
cache name
foodle foo no match (PM_ERR_INST)
foo foodle no match (PM_ERR_INST)
foo foo match
foo bar foo match on short name (instance identifier)
foo bar foo bar match on full name (instance identifier)
foo foo bar bad match (-EINVAL)
foo bar foo blah bad match (-EINVAL)

Cache persistence uses files with names constructed from the indom within the $PCP_VAR_DIR/config/pmda directory.

BYTEORDER(3), PMAPI(3), PMDA(3), pmdaInit(3), pmdaInstance(3), pmdaFetch(3), pmdaLabel(3), pmErrStr(3) and pmGetInDom(3).

PCP Performance Co-Pilot