pmemobj API version 2.3

The PMDK repository on GitHub is the ultimate source of information on PMDK from release 2.0! For all questions and to submit eventual issues please follow to that repository. The PMDK documentation collected here should be valid up to the 1.13.1 release but is maintained only on a best-effort basis and may not reflect the latest state of the art.




pmemobj_ctl_get(), pmemobj_ctl_set(), pmemobj_ctl_exec()

  • Query and modify libpmemobj internal behavior (EXPERIMENTAL)


#include <libpmemobj.h>

int pmemobj_ctl_get(PMEMobjpool *pop, const char *name, void *arg); (EXPERIMENTAL)
int pmemobj_ctl_set(PMEMobjpool *pop, const char *name, void *arg); (EXPERIMENTAL)
int pmemobj_ctl_exec(PMEMobjpool *pop, const char *name, void *arg); (EXPERIMENTAL)


The pmemobj_ctl_get(), pmemobj_ctl_set() and pmemobj_ctl_exec() functions provide a uniform interface for querying and modifying the internal behavior of libpmemobj through the control (CTL) namespace.

The CTL namespace is organized in a tree structure. Starting from the root, each node can be either internal, containing other elements, or a leaf. Internal nodes themselves can only contain other nodes and cannot be entry points. There are two types of those nodes: named and indexed. Named nodes have string identifiers. Indexed nodes represent an abstract array index and have an associated string identifier. The index itself is provided by the user. A collection of indexes present on the path of an entry point is provided to the handler functions as name and index pairs.

The name argument specifies an entry point as defined in the CTL namespace specification. The entry point description specifies whether the extra arg is required. Those two parameters together create a CTL query. The pop argument is optional if the entry point resides in a global namespace (i.e., is shared for all the pools). The functions and the entry points are thread-safe unless indicated otherwise below. If there are special conditions for calling an entry point, they are explicitly stated in its description. The functions propagate the return value of the entry point. If either name or arg is invalid, -1 is returned.

Entry points are the leaves of the CTL namespace structure. Each entry point can read from the internal state, write to the internal state, exec a function or a combination of these operations.

The entry points are listed in the following format:

name | r(ead)w(rite)x(ecute) | global/- | read argument type | write argument type | exec argument type | config argument type



prefault.at_create | rw | global | int | int | - | boolean

If set, every page of the pool will be touched and written to when the pool is created, in order to trigger page allocation and minimize the performance impact of pagefaults. Affects only the pmemobj_create() function.

Always returns 0.

prefault.at_open | rw | global | int | int | - | boolean

If set, every page of the pool will be touched and written to when the pool is opened, in order to trigger page allocation and minimize the performance impact of pagefaults. Affects only the pmemobj_open() function.

Always returns 0.

tx.debug.skip_expensive_checks | rw | - | int | int | - | boolean

Turns off some expensive checks performed by the transaction module in “debug” builds. Ignored in “release” builds.

tx.cache.size | rw | - | long long | long long | - | integer

Size in bytes of the transaction snapshot cache. In a larger cache the frequency of persistent allocations is lower, but with higher fixed cost.

This should be set to roughly the sum of sizes of the snapshotted regions in an average transaction in the pool.

This value must be a in a range between 0 and PMEMOBJ_MAX_ALLOC_SIZE. If the current threshold is larger than the new cache size, the threshold will be made equal to the new size.

This entry point is not thread safe and should not be modified if there are any transactions currently running.

Returns 0 if successful, -1 otherwise.

tx.cache.threshold | rw | - | long long | long long | - | integer

Threshold in bytes, below which snapshots will use the cache. All larger snapshots will trigger a persistent allocation.

This value must be a in a range between 0 and tx.cache.size.

This entry point is not thread safe and should not be modified if there are any transactions currently running.

Returns 0 if successful, -1 otherwise.

tx.post_commit.queue_depth | rw | - | int | int | - | integer

Controls the depth of the post-commit tasks queue. A post-commit task is the collection of work items that need to be performed on the persistent state after a successfully completed transaction. This includes freeing no longer needed objects and cleaning up various caches. By default, this queue does not exist and the post-commit task is executed synchronously in the same thread that ran the transaction. By changing this parameter, one can offload this task to a separate worker. If the queue is full, the algorithm, instead of waiting, performs the post-commit in the current thread.

The task is performed on a finite resource (lanes, of which there are 1024), and if the worker threads that process this queue are unable to keep up with the demand, regular threads might start to block waiting for that resource. This will happen if the queue depth value is too large.

As a general rule, this value should be set to approximately 1024 minus the average number of threads in the application (not counting the post-commit workers); however, this may vary from workload to workload.

The queue depth value must also be a power of two.

This entry point is not thread-safe and must be called when no transactions are currently being executed.

Returns 0 if successful, -1 otherwise.

tx.post_commit.worker | r- | - | void * | - | - | -

The worker function launched in a thread to perform asynchronous processing of post-commit tasks. This function returns only after a stop entry point is called. There may be many worker threads at a time. If there is no work to be done, this function sleeps instead of polling.

Always returns 0.

tx.post_commit.stop | r- | - | void * | - | - | -

This function forces all the post-commit worker functions to exit and return control back to the calling thread. This should be called before the application terminates and the post commit worker threads need to be shutdown.

After the invocation of this entry point, the post-commit task queue can no longer be used. If worker threads must be restarted after a stop, the tx.post_commit.queue_depth needs to be set again.

This entry point must be called when no transactions are currently being executed.

Always returns 0.

heap.alloc_class.[class_id].desc | rw | - | struct pobj_alloc_class_desc | struct pobj_alloc_class_desc | - | integer, integer, integer, string

Describes an allocation class. Allows one to create or view the internal data structures of the allocator.

Creating custom allocation classes can be beneficial for both raw allocation throughput, scalability and, most importantly, fragmentation. By carefully constructing allocation classes that match the application workload, one can entirely eliminate external and internal fragmentation. For example, it is possible to easily construct a slab-like allocation mechanism for any data structure.

The [class_id] is an index field. Only values between 0-254 are valid. If setting an allocation class, but the class_id is already taken, the function will return -1. The values between 0-127 are reserved for the default allocation classes of the library and can be used only for reading.

The recommended method for retrieving information about all allocation classes is to call this entry point for all class ids between 0 and 254 and discard those results for which the function returns an error.

This entry point takes a complex argument.

struct pobj_alloc_class_desc {
	size_t unit_size;
	size_t alignment;
	unsigned units_per_block;
	enum pobj_header_type header_type;
	unsigned class_id;

The first field, unit_size, is an 8-byte unsigned integer that defines the allocation class size. While theoretically limited only by PMEMOBJ_MAX_ALLOC_SIZE, for most workloads this value should be between 8 bytes and 2 megabytes.

The alignment field is currently unsupported and must be set to 0. All objects have default alignment of 64 bytes, but the user data alignment is affected by the size of the chosen header.

The units_per_block field defines how many units a single block of memory contains. This value will be rounded up to match the internal size of the block (256 kilobytes or a multiple thereof). For example, given a class with a unit_size of 512 bytes and a units_per_block of 1000, a single block of memory for that class will have 512 kilobytes. This is relevant because the bigger the block size, the less frequently blocks need to be fetched, resulting in lower contention on global heap state. Keep in mind that object allocation is tracked in a bitmap with a limited number of entries, making it inefficient to create allocation classes smaller than 128 bytes.

The header_type field defines the header of objects from the allocation class. There are three types:

  • POBJ_HEADER_LEGACY, string value: legacy. Used for allocation classes prior to version 1.3 of the library. Not recommended for use. Incurs a 64 byte metadata overhead for every object. Fully supports all features.
  • POBJ_HEADER_COMPACT, string value: compact. Used as default for all predefined allocation classes. Incurs a 16 byte metadata overhead for every object. Fully supports all features.
  • POBJ_HEADER_NONE, string value: none. Header type that incurs no metadata overhead beyond a single bitmap entry. Can be used for very small allocation classes or when objects must be adjacent to each other. This header type does not support type numbers (type number is always 0) or allocations that span more than one unit.

The class_id field is an optional, runtime-only variable that allows the user to retrieve the identifier of the class. This will be equivalent to the provided [class_id]. This field cannot be set from a config file.

The allocation classes are a runtime state of the library and must be created after every open. It is highly recommended to use the configuration file to store the classes.

This structure is declared in the libpmemobj/ctl.h header file. Please refer to this file for an in-depth explanation of the allocation classes and relevant algorithms.

Allocation classes constructed in this way can be leveraged by explicitly specifying the class using POBJ_CLASS_ID(id) flag in pmemobj_tx_xalloc()/pmemobj_xalloc() functions.

Example of a valid alloc class query string:


This query, if executed, will create an allocation class with an id of 128 that has a unit size of 500 bytes, has at least 1000 units per block and uses a compact header.

For reading, function returns 0 if successful, if the allocation class does not exist it sets the errno to ENOENT and returns -1;

For writing, function returns 0 if the allocation class has been successfully created, -1 otherwise. | -w | - | - | struct pobj_alloc_class_desc | - | integer, integer, integer, string

Same as heap.alloc_class.[class_id].desc, but instead of requiring the user to provide the class_id, it automatically creates the allocation class with the first available identifier.

This should be used when it’s impossible to guarantee unique allocation class naming in the application (e.g. when writing a library that uses libpmemobj).

The required class identifier will be stored in the class_id field of the struct pobj_alloc_class_desc.

This function returns 0 if the allocation class has been successfully created, -1 otherwise.

stats.enabled | rw | - | int | int | - | boolean

Enables or disables runtime collection of statistics. Statistics are not recalculated after enabling; any operations that occur between disabling and re-enabling will not be reflected in subsequent values.

Statistics are disabled by default. Enabling them may have non-trivial performance impact.

Always returns 0.

stats.heap.curr_allocated | r- | - | int | - | - | -

Returns the number of bytes currently allocated in the heap. If statistics were disabled at any time in the lifetime of the heap, this value may be inaccurate.

heap.size.granularity | rw- | - | uint64_t | uint64_t | - | long long

Reads or modifies the granularity with which the heap grows when OOM. Valid only if the poolset has been defined with directories.

A granularity of 0 specifies that the pool will not grow automatically.

This function returns 0 if the granularity value is 0, or is larger than PMEMOBJ_MIN_PART, -1 otherwise.

heap.size.extend | –x | - | - | - | uint64_t | -

Extends the heap by the given size. Must be larger than PMEMOBJ_MIN_PART.

This function returns 0 if successful, -1 otherwise.


In addition to direct function call, each write entry point can also be set using two alternative methods.

The first method is to load a configuration directly from the PMEMOBJ_CONF environment variable. A properly formatted ctl config string is a single-line sequence of queries separated by ‘;’:


A single query is constructed from the name of the ctl write entry point and the argument, separated by ‘=’:


The entry point argument type is defined by the entry point itself, but there are three predefined primitives:

*) integer: represented by a sequence of [0-9] characters that form
	a single number.
*) boolean: represented by a single character: y/n/Y/N/0/1, each
	corresponds to true or false. If the argument contains any
	trailing characters, they are ignored.
*) string: a simple sequence of characters.

There are also complex argument types that are formed from the primitives separated by a ‘,’:


In summary, a full configuration sequence looks like this:

(first_entry_point)=(arguments, ...);...;(last_entry_point)=(arguments, ...);

As an example, to set both prefault at_open and at_create variables:


The second method of loading an external configuration is to set the PMEMOBJ_CONF_FILE environment variable to point to a file that contains a sequence of ctl queries. The parsing rules are all the same, but the file can also contain white-spaces and comments.

To create a comment, simply use ‘#’ anywhere in a line and everything afterwards, until a new line ‘\n’, will be ignored.

An example configuration file:

# My pmemobj configuration
# Global settings:
prefault. # modify the behavior of pre-faulting
	at_open = 1; # prefault when the pool is opened

	at_create = 0; # but don't prefault when it's created

# Per-pool settings:
# ...


libpmemobj(7) and

The contents of this web site and the associated GitHub repositories are BSD-licensed open source.