pmem API

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.


libpmem - persistent memory support library


#include <libpmem.h>

cc ... -lpmem

Most commonly used functions:

int pmem_is_pmem(const void *addr, size_t len);
void pmem_persist(const void *addr, size_t len);
int pmem_msync(const void *addr, size_t len);
void *pmem_map_file(const char *path, size_t len, int flags,
 mode_t mode, size_t *mapped_lenp, int *is_pmemp);
int pmem_unmap(void *addr, size_t len);

Partial flushing operations:

void pmem_flush(const void *addr, size_t len);
void pmem_drain(void);
int pmem_has_hw_drain(void);

Copying to persistent memory:

void *pmem_memmove_persist(void *pmemdest, const void *src, size_t len);
void *pmem_memcpy_persist(void *pmemdest, const void *src, size_t len);
void *pmem_memset_persist(void *pmemdest, int c, size_t len);
void *pmem_memmove_nodrain(void *pmemdest, const void *src, size_t len);
void *pmem_memcpy_nodrain(void *pmemdest, const void *src, size_t len);
void *pmem_memset_nodrain(void *pmemdest, int c, size_t len);

Library API versioning:

const char *pmem_check_version(
 unsigned major_required,
 unsigned minor_required);

Error handling:



libpmem provides low-level persistent memory (pmem) support for applications using direct access storage (DAX), which is storage that supports load/store access without paging blocks from a block storage device. Some types of non-volatile memory DIMMs (NVDIMMs) provide this type of byte addressable access to storage. A persistent memory aware file system is typically used to expose the direct access to applications. Memory mapping a file from this type of file system results in the load/store, non-paged access to pmem.

This library is for applications that use persistent memory directly, without the help of any library-supplied transactions or memory allocation. Higher-level libraries that build on libpmem are available and are recommended for most applications, see:

libpmemobj(3), a general use persistent memory API, providing memory allocation and transactional operations on variable-sized objects.

libpmemblk(3), providing pmem-resident arrays of fixed-sized blocks with atomic updates.

libpmemlog(3), providing a pmem-resident log file.

Under normal usage, libpmem will never print messages or intentionally cause the process to exit. The only exception to this is the debugging information, when enabled, as described under DEBUGGING AND ERROR HANDLING below.


Most pmem-aware applications will take advantage of higher level libraries that alleviate the application from calling into libpmem directly. Application developers that wish to access raw memory mapped persistence directly (via mmap(2)) and that wish to take on the responsibility for flushing stores to persistence will find the functions described in this section to be the most commonly used.

int pmem_is_pmem(const void *addr, size_t len);

The pmem_is_pmem() function returns true only if the entire range [addr, addr+len) consists of persistent memory. A true return from pmem_is_pmem() means it is safe to use pmem_persist() and the related functions below to make changes durable for that memory range.

The implementation of pmem_is_pmem() requires a non-trivial amount of work to determine if the given range is entirely persistent memory. For this reason, it is better to call pmem_is_pmem() once when a range of memory is first encountered, save the result, and use the saved result to determine whether pmem_persist() or msync(2) is appropriate for flushing changes to persistence. Calling pmem_is_pmem() each time changes are flushed to persistence will not perform well.

WARNING: Using pmem_persist() on a range where pmem_is_pmem() returns false may not do anything useful -- use msync(2) instead.

void pmem_persist(const void *addr, size_t len);

Force any changes in the range [addr, addr+len) to be stored durably in persistent memory. This is equivalent to calling msync(2) but may be more optimal and will avoid calling into the kernel if possible. There are no alignment restrictions on the range described by addr and len, but pmem_persist() may expand the range as necessary to meet platform alignment requirements.

WARNING: Like msync(2), there is nothing atomic or transactional about this call. Any unwritten stores in the given range will be written, but some stores may have already been written by virtue of normal cache eviction/replacement policies. Correctly written code must not depend on stores waiting until pmem_persist() is called to become persistent -- they can become persistent at any time before pmem_persist() is called.

int pmem_msync(const void *addr, size_t len);

The function pmem_msync() is like pmem_persist() in that it forces any changes in the range [addr, addr+len) to be stored durably. Since it calls msync(), this function works on either persistent memory or a memory mapped file on traditional storage. pmem_msync() takes steps to ensure the alignment of addresses and lengths passed to msync() meet the requirements of that system call. It calls msync() with the MS_SYNC flag as described in msync(2). Typically the application only checks for the existence of persistent memory once, and then uses that result throughout the program, for example:

    /* do this call once, after the pmem is memory mapped */
    int is_pmem = pmem_is_pmem(rangeaddr, rangelen);

    /* ... make changes to a range of pmem ... */

    /* make the changes durable */
    if (is_pmem)
        pmem_persist(subrangeaddr, subrangelen);
        pmem_msync(subrangeaddr, subrangelen);

    /* ... */

The return value of pmem_msync() is the return value of msync(), which can return -1 and set errno to indicate an error.

void *pmem_map_file(const char *path, size_t len, int flags,
** mode_t mode, size_t *mapped_lenp, int *is_pmemp);**

Given a path, pmem_map_file() function creates a new read/write mapping for the named file. It will map the file using mmap(2), but it also takes extra steps to make large page mappings more likely.

On success, pmem_map_file() returns a pointer to mapped area. If mapped_lenp is not NULL, the length of the mapping is also stored at the address it points to. The is_pmemp argument, if non-NULL, points to a flag that pmem_is_pmem() sets to say if the mapped file is actual pmem, or if msync() must be used to flush writes for the mapped range. On error, NULL is returned, errno is set appropriately, and mapped_lenp and is_pmemp are left untouched.

The flags argument can be 0 or bitwise OR of one or more of the following file creation flags:

PMEM_FILE_CREATE - Create the named file if it does not exist. len must be non-zero and specifies the size of the file to be created. mode has the same meaning as for open(2) and specifies the mode to use in case a new file is created. If neither PMEM_FILE_CREATE nor PMEM_FILE_TMPFILE is specified, then mode is ignored.

PMEM_FILE_EXCL - Same meaning as O_EXCL on open(2) - Ensure that this call creates the file. If this flag is specified in conjunction with PMEM_FILE_CREATE, and pathname already exists, then pmem_map_file() will fail.

PMEM_FILE_TMPFILE - Same meaning as O_TMPFILE on open(2). Create a mapping for an unnamed temporary file. PMEM_FILE_CREATE and len must be specified and path must be an existing directory name.

PMEM_FILE_SPARSE - When creating a file, create a sparse (holey) file instead of calling posix_fallocate(2). Valid only if specified in conjunction with PMEM_FILE_CREATE or PMEM_FILE_TMPFILE, otherwise ignored.

If creation flags are not supplied, then pmem_map_file() creates a mapping for an existing file. In such case, len should be zero. The entire file is mapped to memory; its length is used as the length of the mapping and returned via mapped_lenp.

To delete mappings created with pmem_map_file(), use pmem_unmap().

int pmem_unmap(void *addr, size_t len);

The pmem_unmap() function deletes all the mappings for the specified address range, and causes further references to addresses within the range to generate invalid memory references. It will use the address specified by the parameter addr, where addr must be a previously mapped region. pmem_unmap() will delete the mappings using the munmap(2), On success, pmem_unmap() returns zero. On error, -1 is returned, and errno is set appropriately.


The functions in this section provide access to the stages of flushing to persistence, for the less common cases where an application needs more control of the flushing operations than the pmem_persist() function described above.

void pmem_flush(const void *addr, size_t len);
void pmem_drain(void);

These functions provide partial versions of the pmem_persist() function described above. pmem_persist() can be thought of as this:

pmem_persist(const void *addr, size_t len)
    /* flush the processor caches */
    pmem_flush(addr, len);

    /* wait for any pmem stores to drain from HW buffers */

These functions allow advanced programs to create their own variations of pmem_persist(). For example, a program that needs to flush several discontiguous ranges can call pmem_flush() for each range and then follow up by calling pmem_drain() once.

NOTE: Some software is designed for custom platforms that obviate the need for using PCOMMIT (perhaps the platform issues PCOMMIT on shutdown or something similar). Even in such cases, it is recommended that applications using libpmem do not skip the step of calling pmem_drain(), either directly or by using pmem_persist(). The recommended way to inhibit use of the PCOMMIT instruction is by setting the PMEM_NO_PCOMMIT environment variable as described in the ENVIRONMENT VARIABLES section.

int pmem_has_hw_drain(void);

The pmem_has_hw_drain() function returns true if the machine supports the hardware drain function for persistent memory, such as that provided by the PCOMMIT instruction on Intel processors. If support for hardware drain is not found, or cannot be detected by the library, pmem_has_hw_drain() will return false. Although it is typically an administrative task to provide the correct platform configuration for persistent memory, this function is provided for the less common cases where an application needs to ensure this feature is available. Note that the lack of this feature means that calling pmem_persist() may not fully ensure stores are durable, without additional platform features such as Asynchronous DRAM Refresh (ADR) or something similar.


The functions in this section provide optimized copying to persistent memory.

void *pmem_memmove_persist(void *pmemdest, const void *src,
** size_t len);**
void *pmem_memcpy_persist(void *pmemdest, const void *src, size_t len);
void *pmem_memset_persist(void *pmemdest, int c, size_t len);

The pmem_memmove_persist(), pmem_memcpy_persist(), and pmem_memset_persist(), functions provide the same memory copying as their namesakes memmove(3) memcpy(3), and memset(3), and ensure that the result has been flushed to persistence before returning. For example, the following code is functionally equivalent to pmem_memmove_persist():

void *
pmem_memmove_persist(void *pmemdest, const void *src, size_t len)
    void *retval = memmove(pmemdest, src, len);

    pmem_persist(pmemdest, len);

    return retval;

Calling pmem_memmove_persist() may out-perform the above code, however, since the libpmem implementation may take advantage of the fact that pmemdest is persistent memory and use instructions such as non-temporal stores to avoid the need to flush processor caches.

WARNING: Using these functions where pmem_is_pmem() returns false may not do anything useful. Use the normal libc functions in that case.

void *pmem_memmove_nodrain(void *pmemdest, const void *src,
** size_t len);**
void *pmem_memcpy_nodrain(void *pmemdest, const void *src, size_t len);
void *pmem_memset_nodrain(void *pmemdest, int c, size_t len);

The pmem_memmove_nodrain(), pmem_memcpy_nodrain() and pmem_memset_nodrain() functions are similar to pmem_memmove_persist(), pmem_memcpy_persist(), and pmem_memset_persist() described above, except they skip the final pmem_drain() step. This allows applications to optimize cases where several ranges are being copied to persistent memory, followed by a single call to pmem_drain(). The following example illustrates how these functions might be used to avoid multiple calls to pmem_drain() when copying several ranges of memory to pmem:

    /* ... write several ranges to pmem ... */
    pmem_memcpy_nodrain(pmemdest1, src1, len1);
    pmem_memcpy_nodrain(pmemdest2, src2, len2);

    /* ... */

    /* wait for any pmem stores to drain from HW buffers */

WARNING: Using pmem_memmove_nodrain(), pmem_memcpy_nodrain() or pmem_memset_nodrain() on a destination where pmem_is_pmem() returns false may not do anything useful.


This section describes how the library API is versioned, allowing applications to work with an evolving API.

const char *pmem_check_version(
** unsigned major_required,
** unsigned minor_required);

The pmem_check_version() function is used to see if the installed libpmem supports the version of the library API required by an application. The easiest way to do this is for the application to supply the compile-time version information, supplied by defines in <libpmem.h>, like this:

reason = pmem_check_version(PMEM_MAJOR_VERSION,
if (reason != NULL) {
    /*  version check failed, reason string tells you why */

Any mismatch in the major version number is considered a failure, but a library with a newer minor version number will pass this check since increasing minor versions imply backwards compatibility.

An application can also check specifically for the existence of an interface by checking for the version where that interface was introduced. These versions are documented in this man page as follows: unless otherwise specified, all interfaces described here are available in version 1.0 of the library. Interfaces added after version 1.0 will contain the text introduced in version x.y in the section of this manual describing the feature.

When the version check performed by pmem_check_version() is successful, the return value is NULL. Otherwise the return value is a static string describing the reason for failing the version check. The string returned by pmem_check_version() must not be modified or freed.


Two versions of libpmem are typically available on a development system. The normal version, accessed when a program is linked using the -lpmem option, is optimized for performance. That version skips checks that impact performance and never logs any trace information or performs any run-time assertions. If an error is detected during the call to libpmem function, an application may retrieve an error message describing the reason of failure using the following function:


The pmem_errormsg() function returns a pointer to a static buffer containing the last error message logged for current thread. The error message may include description of the corresponding error code (if errno was set), as returned by strerror(3). The error message buffer is thread-local; errors encountered in one thread do not affect its value in other threads. The buffer is never cleared by any library function; its content is significant only when the return value of the immediately preceding call to libpmem function indicated an error, or if errno was set. The application must not modify or free the error message string, but it may be modified by subsequent calls to other library functions.

A second version of libpmem, accessed when a program uses the libraries under /usr/lib/nvml_debug, contains run-time assertions and trace points. The typical way to access the debug version is to set the environment variable LD_LIBRARY_PATH to /usr/lib/nvml_debug or /usr/lib64/nvml_debug depending on where the debug libraries are installed on the system. The trace points in the debug version of the library are enabled using the environment variable PMEM_LOG_LEVEL, which can be set to the following values:

  1. This is the default level when PMEM_LOG_LEVEL is not set. No log messages are emitted at this level.

  2. Additional details on any errors detected are logged (in addition to returning the errno-based errors as usual). The same information may be retrieved using pmem_errormsg().

  3. A trace of basic operations is logged.

  4. This level enables a very verbose amount of function call tracing in the library.

  5. This level enables voluminous and fairly obscure tracing information that is likely only useful to the libpmem developers.

The environment variable PMEM_LOG_FILE specifies a file name where all logging information should be written. If the last character in the name is "-", the PID of the current process will be appended to the file name when the log file is created. If PMEM_LOG_FILE is not set, the logging output goes to stderr.

Setting the environment variable PMEM_LOG_LEVEL has no effect on the non-debug version of libpmem.


libpmem can change its default behavior based on the following environment variables. These are largely intended for testing and are not normally required.


If val is 0 (zero), then pmem_is_pmem() will always return false. Setting val to 1 causes pmem_is_pmem() to always return true. This variable is mostly used for testing but can be used to force pmem behavior on a system where a range of pmem is not detectable as pmem for some reason.


Setting this environment variable to 1 forces libpmem to never issue the Intel PCOMMIT instruction. This can be used on platforms where the hardware drain function is performed some other way, like automatic flushing during a power failure.

WARNING: Using this environment variable incorrectly may impact program correctness.


Setting this environment variable to 1 forces libpmem to never issue the CLWB instruction on Intel hardware, falling back to other cache flush instructions instead (CLFLUSHOPT or CLFLUSH on Intel hardware). Without this environment variable, libpmem will always use the CLWB instruction for flushing processor caches on platforms that support the instruction. This variable is intended for use during library testing but may be required for some rare cases where using CLWB has a negative impact on performance.


Setting this environment variable to 1 forces libpmem to never issue the CLFLUSHOPT instruction on Intel hardware, falling back to the CLFLUSH instructions instead. Without this environment variable, libpmem will always use the CLFLUSHOPT instruction for flushing processor caches on platforms that support the instruction, but where CLWB is not available. This variable is intended for use during library testing.


Setting this environment variable to 1 forces libpmem to never use the non-temporal move instructions on Intel hardware. Without this environment variable, libpmem will use the non-temporal instructions for copying larger ranges to persistent memory on platforms that support the instructions. This variable is intended for use during library testing.


This environment variable allows overriding the minimal length of pmem_memcpy_*(), pmem_memmove_*() or pmem_memset_*() operations, for which libpmem uses non-temporal move instructions. Setting this environment variable to 0 forces libpmem to always use the non-temporal move instructions if available. It has no effect if PMEM_NO_MOVNT variable is set to 1. This variable is intended for use during library testing.

**PMEM_MMAP_HINT=**val This environment variable allows overriding the hint address used by pmem_map_file(). If set, it also disables mapping address randomization. This variable is intended for use during library testing and debugging. Setting it to some fairly large value (i.e. 0x10000000000) will very likely result in mapping the file at the specified address (if not used) or at the first unused region above given address, without adding any random offset. When debugging, this makes it easier to calculate the actual address of the persistent memory block, based on its offset in the file. In case of libpmemobj it simplifies conversion of a persistent object identifier (OID) into a direct pointer to the object. NOTE: Setting this environment variable affects all the NVM libraries, disabling mapping address randomization and causing the specified address to be used as a hint about where to place the mapping.


The following example uses libpmem to flush changes made to raw, memory-mapped persistent memory.

WARNING: there is nothing transactional about the pmem_persist() or pmem_msync() calls in this example. Interrupting the program may result in a partial write to pmem. Use a transactional library such as libpmemobj(3) to avoid torn updates.

#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <stdio.h>
#include <errno.h>
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <libpmem.h>

/* using 4k of pmem for this example */
#define	PMEM_LEN 4096

#define	PATH "/pmem-fs/myfile"

main(int argc, char *argv[])
	char *pmemaddr;
	size_t mapped_len;
	int is_pmem;

	/* create a pmem file and memory map it */
	if ((pmemaddr = pmem_map_file(PATH, PMEM_LEN, PMEM_FILE_CREATE,
				0666, &mapped_len, &is_pmem)) == NULL) {

	/* store a string to the persistent memory */
	strcpy(pmemaddr, "hello, persistent memory");

	/* flush above strcpy to persistence */
	if (is_pmem)
		pmem_persist(pmemaddr, mapped_len);
		pmem_msync(pmemaddr, mapped_len);

	 * Delete the mappings. The region is also
	 * automatically unmapped when the process is
	 * terminated.
	pmem_unmap(pmemaddr, mapped_len);

See for more examples using the libpmem API.


libpmem builds on the persistent memory programming model recommended by the SNIA NVM Programming Technical Work Group:


open(2), mmap(2), munmap(2), msync(2), strerror(3), libpmemobj(3), libpmemblk(3), libpmemlog(3), libvmem(3) and

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