numam-spdk/doc/memory.md

Memory Management for User Space Drivers

The following is an attempt to explain why all data buffers passed to SPDK must be allocated using spdk_dma_malloc() or its siblings, and why SPDK relies on DPDK's proven base functionality to implement memory management.

Computing platforms generally carve physical memory up into 4KiB segments called pages. They number the pages from 0 to N starting from the beginning of addressable memory. Operating systems then overlay 4KiB virtual memory pages on top of these physical pages using arbitrarily complex mappings. See Virtual Memory for an overview.

Physical memory is attached on channels, where each memory channel provides some fixed amount of bandwidth. To optimize total memory bandwidth, the physical addressing is often set up to automatically interleave between channels. For instance, page 0 may be located on channel 0, page 1 on channel 1, page 2 on channel 2, etc. This is so that writing to memory sequentially automatically utilizes all available channels. In practice, interleaving is done at a much more granular level than a full page.

Modern computing platforms support hardware acceleration for virtual to physical translation inside of their Memory Management Unit (MMU). The MMU often supports multiple different page sizes. On recent x86_64 systems, 4KiB, 2MiB, and 1GiB pages are supported. Typically, operating systems use 4KiB pages by default.

NVMe devices transfer data to and from system memory using Direct Memory Access (DMA). Specifically, they send messages across the PCI bus requesting data transfers. In the absence of an IOMMU, these messages contain physical memory addresses. These data transfers happen without involving the CPU, and the MMU is responsible for making access to memory coherent.

NVMe devices also may place additional requirements on the physical layout of memory for these transfers. The NVMe 1.0 specification requires all physical memory to be describable by what is called a PRP list. To be described by a PRP list, memory must have the following properties:

• The memory is broken into physical 4KiB pages, which we'll call device pages.
• The first device page can be a partial page starting at any 4-byte aligned address. It may extend up to the end of the current physical page, but not beyond.
• If there is more than one device page, the first device page must end on a physical 4KiB page boundary.
• The last device page begins on a physical 4KiB page boundary, but is not required to end on a physical 4KiB page boundary.

The specification allows for device pages to be other sizes than 4KiB, but all known devices as of this writing use 4KiB.

The NVMe 1.1 specification added support for fully flexible scatter gather lists, but the feature is optional and most devices available today do not support it.

User space drivers run in the context of a regular process and so have access to virtual memory. In order to correctly program the device with physical addresses, some method for address translation must be implemented.

The simplest way to do this on Linux is to inspect /proc/self/pagemap from within a process. This file contains the virtual address to physical address mappings. As of Linux 4.0, accessing these mappings requires root privileges. However, operating systems make absolutely no guarantee that the mapping of virtual to physical pages is static. The operating system has no visibility into whether a PCI device is directly transferring data to a set of physical addresses, so great care must be taken to coordinate DMA requests with page movement. When an operating system flags a page such that the virtual to physical address mapping cannot be modified, this is called pinning the page.

There are several reasons why the virtual to physical mappings may change, too. By far the most common reason is due to page swapping to disk. However, the operating system also moves pages during a process called compaction, which collapses identical virtual pages onto the same physical page to save memory. Some operating systems are also capable of doing transparent memory compression. It is also increasingly possible to hot-add additional memory, which may trigger a physical address rebalance to optimize interleaving.

POSIX provides the mlock call that forces a virtual page of memory to always be backed by a physical page. In effect, this is disabling swapping. This does not guarantee, however, that the virtual to physical address mapping is static. The mlock call should not be confused with a pin call, and it turns out that POSIX does not define an API for pinning memory. Therefore, the mechanism to allocate pinned memory is operating system specific.

SPDK relies on DPDK to allocate pinned memory. On Linux, DPDK does this by allocating hugepages (by default, 2MiB). The Linux kernel treats hugepages differently than regular 4KiB pages. Specifically, the operating system will never change their physical location. This is not by intent, and so things could change in future versions, but it is true today and has been for a number of years (see the later section on the IOMMU for a future-proof solution). DPDK goes through great pains to allocate hugepages such that it can string together the longest runs of physical pages possible, such that it can accomodate physically contiguous allocations larger than a single page.

With this explanation, hopefully it is now clear why all data buffers passed to SPDK must be allocated using spdk_dma_malloc() or its siblings. The buffers must be allocated specifically so that they are pinned and so that physical addresses are known.

IOMMU Support

Many platforms contain an extra piece of hardware called an I/O Memory Management Unit (IOMMU). An IOMMU is much like a regular MMU, except it provides virtualized address spaces to peripheral devices (i.e. on the PCI bus). The MMU knows about virtual to physical mappings per process on the system, so the IOMMU associates a particular device with one of these mappings and then allows the user to assign arbitrary bus addresses to virtual addresses in their process. All DMA operations between the PCI device and system memory are then translated through the IOMMU by converting the bus address to a virtual address and then the virtual address to the physical address. This allows the operating system to freely modify the virtual to physical address mapping without breaking ongoing DMA operations. Linux provides a device driver, vfio-pci, that allows a user to configure the IOMMU with their current process.

This is a future-proof, hardware-accelerated solution for performing DMA operations into and out of a user space process and forms the long-term foundation for SPDK and DPDK's memory management strategy. We highly recommend that applications are deployed using vfio and the IOMMU enabled, which is fully supported today.