Persistent Memory Development Kit

The libpmem2 library

libpmem2 provides low-level persistent memory support. In particular, support for the persistent memory instructions for flushing changes to pmem is provided.

This library is provided for software that tracks every store to pmem and needs to flush those changes to durability. Most developers will find higher level libraries like libpmemobj to be much more convenient.

Man pages that contain a list of the Linux interfaces provided:

Man pages that contain a list of the Windows interfaces provided:

libpmem2 Example

The Basics

If you’ve decided to handle persistent memory allocation and consistency across program interruption yourself, you will find the functions in libpmem2 useful. It is important to understand that programming to raw pmem means you must create your own transactions or convince yourself you don’t care if a system or program crash leaves your pmem files in an inconsistent state. Interfaces in the libpmem2 are non-transactional, but can be used to build transactional interfaces like libpmemobj.

To illustrate the basics, let’s walk through the man page example first:

The example starts, as shown above, by including the necessary headers. Line 18 (the highlighted line) shows the header file you need to include to use libpmem2: libpmem2.h.

For this simple example, we’re going to write a string to a memory-mapped file. First, we need to prepare a config and mapping based on the source file.

The lines above create the file, prepare config and source structures, set required granularity and map the file into memory.

This illustrates basic functions in the libpmem2: First of them, pmem2_config_new(), creates config struct that will be used for mapping. Config is an object with which we define the parameters of the target mapping to be created.

In this example, we only set the granularity in the config, other values remain default. The granularity is the only required argument in the config. In addition to the granularity setting, libpmem2 provides multiple optional functions to configure target mapping, e.g. pmem2_config_set_length() to set length which will be used for mapping, or pmem2_config_set_offset which will be used to map the contents from the specified location of the source.

The second highlighted line contains a call to pmem2_source_from_fd(), which takes a file descriptor and creates a new instance of the source structure that describes the data source used for mapping. In this particular example, mapping source comes from a file descriptor, however, libpmem2 also provides functions to use a source from the file handle or create an anonymous map.

The next key step in this example is to set the granularity using pmem2_config_set_required_store_granularity. Granularity must be one of the three values: PMEM2_GRANULARITY_BYTE, PMEM2_GRANULARITY_CACHE_LINE, PMEM2_GRANULARITY_PAGE. In this case, we set a maximum permitted granularity to PMEM2_GRANULARITY_PAGE. Logically, by setting the granularity to page, the application indicates that it’s going to continue functioning even if the underlying device is block-based.

More detailed information about granularity concept and each option can be found here

The last step is to create a mapping using described above config and source. Underlying function of pmem2_map() calls mmap(2) on POSIX or CreateFileMapping() on Windows to memory map the entire file concerning the permissible granularity.

When the mapping is created we are ready to write into it.

Using the getters like pmem2_map_get_size or pmem2_map_get_address we can easily read information about created mapping.

The novel thing about pmem is you can copy to it directly, like any memory. The strcpy() call shown on line 63 above is just the usual libc function that stores a string to memory. If this example program were to be interrupted either during or just after the strcpy() call, you can’t be sure which parts of the string made it all the way to the media. It might be none of the string, all of the string, or somewhere in-between. Also, there’s no guarantee the string will make it to the media in the order it was stored! For longer ranges, it is just as likely that portions copied later make it to the media before earlier portions. (So don’t write code like the example above and then expect to check for zeros to see how much of the string was written.)

How can a string get stored in a seemingly random order? The reason is that until a flush function like msync() has returned successfully, the normal cache pressure that happens on an active system can push changes out to the media at any time in any order. Most processors have barrier instructions (like SFENCE on the Intel platform) but those instructions deal with ordering in the visibility of stores to other threads, not with the order that changes reach persistence. The only barrier for flushing to persistence is function returned by pmem2_get_persist_fn() as shown below.

The libpmem2 function pmem2_get_persist_fn automatically decides what’s the most appropriate mechanism for flushing data onto the underlying storage. This means that if the mapping does not support user-space flushing, the persist function will fall back to using the OS primitives for synchronizing data.

The persist() function above will, if possible, perform the flush directly from user space, without calling into the OS. This is made possible on the Intel platform using instructions like CLWB and CLFLUSHOPT which are described in Intel’s manuals. Of course, you are free to use these instructions directly in your program, but the program will crash with an undefined opcode if you try to use the instructions on a platform that doesn’t support them. This is where libpmem2 helps you out, by checking the platform capabilities on start-up and choosing the best instructions for each operation it supports.

To avoid leaks in our example the last thing to do is unmap the mapping, free source and config structures and close file descriptor.

Buildable source for the libpmem2 basic.c example above is available in the PMDK repository.

Copying to Persistent Memory

Another feature of libpmem2 is a set of routines for optimally copying to persistent memory. These functions perform the same functions as the libc functions memcpy(), memset(), and memmove(), but they are optimized for copying to pmem. On the Intel platform, this is done using the non-temporal store instructions which bypass the processor caches (eliminating the need to flush that portion of the data path).

Below example illustrates that the following code:

        memset(dest, c, len);
        pmem2_persist_fn persist_fn = pmem2_get_persist_fn(map);
        persist_fn(dest, len);

is functionally equivalent to:

        pmem2_memset_fn memset_fn = pmem2_get_memset_fn(map);
        memset_fn(dest, c, len);

The second part of the code above shows how memset_fn() is used just like memset(3) except that libpmem2 handles flushing the data to persistence as part of the set.

Separating the Flush Steps

There are two steps in flushing to persistence. The first step is to flush the processor caches, or bypass them entirely as explained in the previous example. The second step is to wait for any hardware buffers to drain, to ensure writes have reached the media. These steps are performed together when function returned by pmem2_get_persist_fn() is called:

	pmem2_persist_fn persist_fn = pmem2_get_persist_fn(map);
	persist_fn(addr, len);

or they can be called individually by calling function from pmem2_get_flush_fn() for the first step and function from pmem2_get_drain_fn() for the second:

	pmem2_flush_fn flush_fn = pmem2_get_flush_fn(map);
	pmem2_drain_fn drain_fn = pmem2_get_drain_fn(map);

	flush_fn(addr, len);

Note that either of these steps may be unnecessary on a given platform, and the library knows how to check for that and do the right thing. For example, on Intel platforms with eADR, flusn_fn() is an empty function.

When does it make sense to break flushing into steps? For example, if a program is using multiple calls to memcpy(), it can copy data and only performs the flush, postponing the final drain step to the end. This works because unlike the flush step, the drain step does not take an address range - it is a system-wide drain operation so can happen at the end of the loop that copies individual blocks of data.