librpmem man page

librpmem -- remote persistent memory support library (EXPERIMENTAL)


#include <librpmem.h>
cc ... -lrpmem

Library API versioning

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

Error handling

const char *rpmem_errormsg(void);

Other library functions

A description of other librpmem functions can be found on the following manual pages:


librpmem provides low-level support for remote access to persistent memory (pmem) utilizing RDMA-capable RNICs. The library can be used to remotely replicate a memory region over the RDMA protocol. It utilizes an appropriate persistency mechanism based on the remote node's platform capabilities. librpmem utilizes the ssh(1) client to authenticate a user on the remote node, and for encryption of the connection's out-of-band configuration data. See SSH, below, for details.

The maximum replicated memory region size can not be bigger than the maximum locked-in-memory address space limit. See memlock in limits.conf(5) for more details.

This library is for applications that use remote persistent memory directly, without the help of any library-supplied transactions or memory allocation. Higher-level libraries that build on libpmem(7) are available and are recommended for most applications, see:

Target Node Address Format


The target node address is described by the hostname which the client connects to, with an optional user name. The user must be authorized to authenticate to the remote machine without querying for password/passphrase. The optional port number is used to establish the SSH connection. The default port number is 22.

Remote Pool Attributes

The rpmem_pool_attr structure describes a remote pool and is stored in remote pool's metadata. This structure must be passed to the rpmem_create(3) function by caller when creating a pool on remote node. When opening the pool using rpmem_open(3) function the appropriate fields are read from pool's metadata and returned back to the caller.


struct rpmem_pool_attr {
    char signature[RPMEM_POOL_HDR_SIG_LEN];
    uint32_t major;
    uint32_t compat_features;
    uint32_t incompat_features;
    uint32_t ro_compat_features;
    unsigned char poolset_uuid[RPMEM_POOL_HDR_UUID_LEN];
    unsigned char uuid[RPMEM_POOL_HDR_UUID_LEN];
    unsigned char next_uuid[RPMEM_POOL_HDR_UUID_LEN];
    unsigned char prev_uuid[RPMEM_POOL_HDR_UUID_LEN];
    unsigned char user_flags[RPMEM_POOL_USER_FLAGS_LEN];

The signature field is an 8-byte field which describes the pool's on-media format.

The major field is a major version number of the pool's on-media format.

The compat_features field is a mask describing compatibility of pool's on-media format optional features.

The incompat_features field is a mask describing compatibility of pool's on-media format required features.

The ro_compat_features field is a mask describing compatibility of pool's on-media format features. If these features are not available, the pool shall be opened in read-only mode.

The poolset_uuid field is an UUID of the pool which the remote pool is associated with.

The uuid field is an UUID of a first part of the remote pool. This field can be used to connect the remote pool with other pools in a list.

The next_uuid and prev_uuid fields are UUIDs of next and previous replicas respectively. These fields can be used to connect the remote pool with other pools in a list.

The user_flags field is a 16-byte user-defined flags.


librpmem utilizes the ssh(1) client to login and execute the rpmemd(1) process on the remote node. By default, ssh(1) is executed with the -4 option, which forces using IPv4 addressing.

For debugging purposes, both the ssh client and the commands executed on the remote node may be overridden by setting the RPMEM_SSH and RPMEM_CMD environment variables, respectively. See Environment for details.


The ssh(1) client is executed by rpmem_open(3) and rpmem_create(3) after forking a child process using fork(2). The application must take this into account when using wait(2) and waitpid(2), which may return the PID of the ssh(1) process executed by librpmem.

If fork(2) support is not enabled in libibverbs, rpmem_open(3) and rpmem_create(3) will fail. By default, fabric(7) initializes libibverbs with fork(2) support by calling the ibv_fork_init(3) function. See fi_verbs(7) for more details.


librpmem relies on the library destructor being called from the main thread. For this reason, all functions that might trigger destruction (e.g. dlclose(3)) should be called in the main thread. Otherwise some of the resources associated with that thread might not be cleaned up properly.

librpmem registers a pool as a single memory region. A Chelsio T4 and T5 hardware can not handle a memory region greater than or equal to 8GB due to a hardware bug. So pool_size value for rpmem_create(3) and rpmem_open(3) using this hardware can not be greater than or equal to 8GB.

Library API Versioning

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

The rpmem_check_version() function is used to see if the installed librpmem 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 <librpmem.h>, like this:

reason = rpmem_check_version(RPMEM_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 rpmem_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 rpmem_check_version() must not be modified or freed.


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

Setting this environment variable overrides the default ssh(1) client command name.

Setting this environment variable overrides the default command executed on the remote node using either ssh(1) or the alternative remote shell command specified by RPMEM_SSH.

RPMEM_CMD can contain multiple commands separated by a vertical bar (|). Each consecutive command is executed on the remote node in order read from a pool set file. This environment variable is read when the library is initialized, so RPMEM_CMD must be set prior to application launch (or prior to dlopen(3) if librpmem is being dynamically loaded).

Setting this variable to 1 enables using fi_sockets(7) provider for in-band RDMA connection. The sockets provider does not support IPv6. It is required to disable IPv6 system wide if RPMEM_ENABLE_SOCKETS == 1 and target == localhost (or any other loopback interface address) and SSH_CONNECTION variable (see ssh(1) for more details) contains IPv6 address after ssh to loopback interface. By default the sockets provider is disabled.

Setting this variable to 0 disables using fi_verbs(7) provider for in-band RDMA connection. The verbs provider is enabled by default.

Limit the maximum number of lanes to num. See LANES, in rpmem_create(3), for details.

Debugging and Error Handling

If an error is detected during the call to a librpmem function, the application may retrieve an error message describing the reason for the failure from rpmem_errormsg(). This function returns a pointer to a static buffer containing the last error message logged for the current thread. If errno was set, the error message may include a description of the corresponding error code 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 a librpmem 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.

Two versions of librpmem are typically available on a development system. The normal version, accessed when a program is linked using the -lrpmem option, is optimized for performance. That version skips checks that impact performance and never logs any trace information or performs any run-time assertions.

A second version of librpmem, accessed when a program uses the libraries under /usr/lib/pmdk_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/pmdk_debug or /usr/lib64/pmdk_debug, as appropriate. Debugging output is controlled using the following environment variables. These variables have no effect on the non-debug version of the library.

The value of RPMEM_LOG_LEVEL enables trace points in the debug version of the library, as follows:

Unless RPMEM_LOG_FILE is set, debugging output is written to stderr.

Specifies the name of a file 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 RPMEM_LOG_FILE is not set, logging output is written to stderr.


The following example uses librpmem to create a remote pool on given target node identified by given pool set name. The associated local memory pool is zeroed and the data is made persistent on remote node. Upon success the remote pool is closed.

#include <stdio.h>
#include <string.h>

#include <librpmem.h>

#define POOL_SIZE    (32 * 1024 * 1024)
#define NLANES        4
unsigned char pool[POOL_SIZE];

main(int argc, char *argv[])
    int ret;
    unsigned nlanes = NLANES;

    /* fill pool_attributes */
    struct rpmem_pool_attr pool_attr;
    memset(&pool_attr, 0, sizeof(pool_attr));

    /* create a remote pool */
    RPMEMpool *rpp = rpmem_create("localhost", "pool.set",
            pool, POOL_SIZE, &nlanes, &pool_attr);
    if (!rpp) {
        fprintf(stderr, "rpmem_create: %s\n", rpmem_errormsg());
        return 1;

    /* store data on local pool */
    memset(pool, 0, POOL_SIZE);

    /* make local data persistent on remote node */
    ret = rpmem_persist(rpp, 0, POOL_SIZE, 0);
    if (ret) {
        fprintf(stderr, "rpmem_persist: %s\n", rpmem_errormsg());
        return 1;

    /* close the remote pool */
    ret = rpmem_close(rpp);
    if (ret) {
        fprintf(stderr, "rpmem_close: %s\n", rpmem_errormsg());
        return 1;

    return 0;


The librpmem API is experimental and may be subject to change in the future. However, using the remote replication in libpmemobj(7) is safe and backward compatibility will be preserved.


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

See Also

rpmemd(1), ssh(1), fork(2), dlclose(3), dlopen(3), ibv_fork_init(3), rpmem_create(3), rpmem_open(3), rpmem_persist(3), strerror(3), limits.conf(5), fabric(7), fi_sockets(7), fi_verbs(7), libpmem(7), libpmemblk(7), libpmemcto(7), libpmemlog(7), libpmemobj(7) and <>

Referenced By

pmempool_rm(3), pmempool-rm(1), rpmem_create(3), rpmemd(1).

The man pages rpmem_check_version(3) and rpmem_errormsg(3) are aliases of librpmem(7).

2018-03-13 PMDK - rpmem API version 1.2 PMDK Programmer's Manual