Heterogeneous Memory Management (HMM)

Provide infrastructure and helpers to integrate non-conventional memory (device memory like GPU on board memory) into regular kernel path, with the cornerstone of this being specialized struct page for such memory (see sections 5 to 7 of this document).

HMM also provides optional helpers for SVM (Share Virtual Memory), i.e., allowing a device to transparently access program address coherently with the CPU meaning that any valid pointer on the CPU is also a valid pointer for the device. This is becoming mandatory to simplify the use of advanced heterogeneous computing where GPU, DSP, or FPGA are used to perform various computations on behalf of a process.

This document is divided as follows: in the first section I expose the problems related to using device specific memory allocators. In the second section, I expose the hardware limitations that are inherent to many platforms. The third section gives an overview of the HMM design. The fourth section explains how CPU page-table mirroring works and the purpose of HMM in this context. The fifth section deals with how device memory is represented inside the kernel. Finally, the last section presents a new migration helper that allows lever- aging the device DMA engine.

Problems of using a device specific memory allocator

Devices with a large amount of on board memory (several gigabytes) like GPUs have historically managed their memory through dedicated driver specific APIs. This creates a disconnect between memory allocated and managed by a device driver and regular application memory (private anonymous, shared memory, or regular file backed memory). From here on I will refer to this aspect as split address space. I use shared address space to refer to the opposite situation: i.e., one in which any application memory region can be used by a device transparently.

Split address space happens because device can only access memory allocated through device specific API. This implies that all memory objects in a program are not equal from the device point of view which complicates large programs that rely on a wide set of libraries.

Concretely this means that code that wants to leverage devices like GPUs needs to copy object between generically allocated memory (malloc, mmap private, mmap share) and memory allocated through the device driver API (this still ends up with an mmap but of the device file).

For flat data sets (array, grid, image, ...) this isn’t too hard to achieve but complex data sets (list, tree, ...) are hard to get right. Duplicating a complex data set needs to re-map all the pointer relations between each of its elements. This is error prone and program gets harder to debug because of the duplicate data set and addresses.

Split address space also means that libraries cannot transparently use data they are getting from the core program or another library and thus each library might have to duplicate its input data set using the device specific memory allocator. Large projects suffer from this and waste resources because of the various memory copies.

Duplicating each library API to accept as input or output memory allocated by each device specific allocator is not a viable option. It would lead to a combinatorial explosion in the library entry points.

Finally, with the advance of high level language constructs (in C++ but in other languages too) it is now possible for the compiler to leverage GPUs and other devices without programmer knowledge. Some compiler identified patterns are only do-able with a shared address space. It is also more reasonable to use a shared address space for all other patterns.

I/O bus, device memory characteristics

I/O buses cripple shared address spaces due to a few limitations. Most I/O buses only allow basic memory access from device to main memory; even cache coherency is often optional. Access to device memory from CPU is even more limited. More often than not, it is not cache coherent.

If we only consider the PCIE bus, then a device can access main memory (often through an IOMMU) and be cache coherent with the CPUs. However, it only allows a limited set of atomic operations from device on main memory. This is worse in the other direction: the CPU can only access a limited range of the device memory and cannot perform atomic operations on it. Thus device memory cannot be considered the same as regular memory from the kernel point of view.

Another crippling factor is the limited bandwidth (~32GBytes/s with PCIE 4.0 and 16 lanes). This is 33 times less than the fastest GPU memory (1 TBytes/s). The final limitation is latency. Access to main memory from the device has an order of magnitude higher latency than when the device accesses its own memory.

Some platforms are developing new I/O buses or additions/modifications to PCIE to address some of these limitations (OpenCAPI, CCIX). They mainly allow two- way cache coherency between CPU and device and allow all atomic operations the architecture supports. Sadly, not all platforms are following this trend and some major architectures are left without hardware solutions to these problems.

So for shared address space to make sense, not only must we allow devices to access any memory but we must also permit any memory to be migrated to device memory while device is using it (blocking CPU access while it happens).

Shared address space and migration

HMM intends to provide two main features. First one is to share the address space by duplicating the CPU page table in the device page table so the same address points to the same physical memory for any valid main memory address in the process address space.

To achieve this, HMM offers a set of helpers to populate the device page table while keeping track of CPU page table updates. Device page table updates are not as easy as CPU page table updates. To update the device page table, you must allocate a buffer (or use a pool of pre-allocated buffers) and write GPU specific commands in it to perform the update (unmap, cache invalidations, and flush, ...). This cannot be done through common code for all devices. Hence why HMM provides helpers to factor out everything that can be while leaving the hardware specific details to the device driver.

The second mechanism HMM provides is a new kind of ZONE_DEVICE memory that allows allocating a struct page for each page of the device memory. Those pages are special because the CPU cannot map them. However, they allow migrating main memory to device memory using existing migration mechanisms and everything looks like a page is swapped out to disk from the CPU point of view. Using a struct page gives the easiest and cleanest integration with existing mm mech- anisms. Here again, HMM only provides helpers, first to hotplug new ZONE_DEVICE memory for the device memory and second to perform migration. Policy decisions of what and when to migrate things is left to the device driver.

Note that any CPU access to a device page triggers a page fault and a migration back to main memory. For example, when a page backing a given CPU address A is migrated from a main memory page to a device page, then any CPU access to address A triggers a page fault and initiates a migration back to main memory.

With these two features, HMM not only allows a device to mirror process address space and keeping both CPU and device page table synchronized, but also lever- ages device memory by migrating the part of the data set that is actively being used by the device.

Address space mirroring implementation and API

Address space mirroring’s main objective is to allow duplication of a range of CPU page table into a device page table; HMM helps keep both synchronized. A device driver that wants to mirror a process address space must start with the registration of an hmm_mirror struct:

int hmm_mirror_register(struct hmm_mirror *mirror,
                        struct mm_struct *mm);
int hmm_mirror_register_locked(struct hmm_mirror *mirror,
                               struct mm_struct *mm);

The locked variant is to be used when the driver is already holding mmap_sem of the mm in write mode. The mirror struct has a set of callbacks that are used to propagate CPU page tables:

struct hmm_mirror_ops {
    /* sync_cpu_device_pagetables() - synchronize page tables
     * @mirror: pointer to struct hmm_mirror
     * @update_type: type of update that occurred to the CPU page table
     * @start: virtual start address of the range to update
     * @end: virtual end address of the range to update
     * This callback ultimately originates from mmu_notifiers when the CPU
     * page table is updated. The device driver must update its page table
     * in response to this callback. The update argument tells what action
     * to perform.
     * The device driver must not return from this callback until the device
     * page tables are completely updated (TLBs flushed, etc); this is a
     * synchronous call.
     void (*update)(struct hmm_mirror *mirror,
                    enum hmm_update action,
                    unsigned long start,
                    unsigned long end);

The device driver must perform the update action to the range (mark range read only, or fully unmap, ...). The device must be done with the update before the driver callback returns.

When the device driver wants to populate a range of virtual addresses, it can use either:

int hmm_vma_get_pfns(struct vm_area_struct *vma,
                    struct hmm_range *range,
                    unsigned long start,
                    unsigned long end,
                    hmm_pfn_t *pfns);
int hmm_vma_fault(struct vm_area_struct *vma,
                  struct hmm_range *range,
                  unsigned long start,
                  unsigned long end,
                  hmm_pfn_t *pfns,
                  bool write,
                  bool block);

The first one (hmm_vma_get_pfns()) will only fetch present CPU page table entries and will not trigger a page fault on missing or non-present entries. The second one does trigger a page fault on missing or read-only entry if the write parameter is true. Page faults use the generic mm page fault code path just like a CPU page fault.

Both functions copy CPU page table entries into their pfns array argument. Each entry in that array corresponds to an address in the virtual range. HMM provides a set of flags to help the driver identify special CPU page table entries.

Locking with the update() callback is the most important aspect the driver must respect in order to keep things properly synchronized. The usage pattern is:

int driver_populate_range(...)
     struct hmm_range range;
     ret = hmm_vma_get_pfns(vma, &range, start, end, pfns);
     if (ret)
         return ret;
     if (!hmm_vma_range_done(vma, &range)) {
         goto again;

     // Use pfns array content to update device page table

     return 0;

The driver->update lock is the same lock that the driver takes inside its update() callback. That lock must be held before hmm_vma_range_done() to avoid any race with a concurrent CPU page table update.

HMM implements all this on top of the mmu_notifier API because we wanted a simpler API and also to be able to perform optimizations latter on like doing concurrent device updates in multi-devices scenario.

HMM also serves as an impedance mismatch between how CPU page table updates are done (by CPU write to the page table and TLB flushes) and how devices update their own page table. Device updates are a multi-step process. First, appropriate commands are written to a buffer, then this buffer is scheduled for execution on the device. It is only once the device has executed commands in the buffer that the update is done. Creating and scheduling the update command buffer can happen concurrently for multiple devices. Waiting for each device to report commands as executed is serialized (there is no point in doing this concurrently).

Represent and manage device memory from core kernel point of view

Several different designs were tried to support device memory. First one used a device specific data structure to keep information about migrated memory and HMM hooked itself in various places of mm code to handle any access to addresses that were backed by device memory. It turns out that this ended up replicating most of the fields of struct page and also needed many kernel code paths to be updated to understand this new kind of memory.

Most kernel code paths never try to access the memory behind a page but only care about struct page contents. Because of this, HMM switched to directly using struct page for device memory which left most kernel code paths unaware of the difference. We only need to make sure that no one ever tries to map those pages from the CPU side.

HMM provides a set of helpers to register and hotplug device memory as a new region needing a struct page. This is offered through a very simple API:

struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
                                  struct device *device,
                                  unsigned long size);
void hmm_devmem_remove(struct hmm_devmem *devmem);

The hmm_devmem_ops is where most of the important things are:

struct hmm_devmem_ops {
    void (*free)(struct hmm_devmem *devmem, struct page *page);
    int (*fault)(struct hmm_devmem *devmem,
                 struct vm_area_struct *vma,
                 unsigned long addr,
                 struct page *page,
                 unsigned flags,
                 pmd_t *pmdp);

The first callback (free()) happens when the last reference on a device page is dropped. This means the device page is now free and no longer used by anyone. The second callback happens whenever the CPU tries to access a device page which it cannot do. This second callback must trigger a migration back to system memory.

Migration to and from device memory

Because the CPU cannot access device memory, migration must use the device DMA engine to perform copy from and to device memory. For this we need a new migration helper:

int migrate_vma(const struct migrate_vma_ops *ops,
                struct vm_area_struct *vma,
                unsigned long mentries,
                unsigned long start,
                unsigned long end,
                unsigned long *src,
                unsigned long *dst,
                void *private);

Unlike other migration functions it works on a range of virtual address, there are two reasons for that. First, device DMA copy has a high setup overhead cost and thus batching multiple pages is needed as otherwise the migration overhead makes the whole exercise pointless. The second reason is because the migration might be for a range of addresses the device is actively accessing.

The migrate_vma_ops struct defines two callbacks. First one (alloc_and_copy()) controls destination memory allocation and copy operation. Second one is there to allow the device driver to perform cleanup operations after migration:

struct migrate_vma_ops {
    void (*alloc_and_copy)(struct vm_area_struct *vma,
                           const unsigned long *src,
                           unsigned long *dst,
                           unsigned long start,
                           unsigned long end,
                           void *private);
    void (*finalize_and_map)(struct vm_area_struct *vma,
                             const unsigned long *src,
                             const unsigned long *dst,
                             unsigned long start,
                             unsigned long end,
                             void *private);

It is important to stress that these migration helpers allow for holes in the virtual address range. Some pages in the range might not be migrated for all the usual reasons (page is pinned, page is locked, ...). This helper does not fail but just skips over those pages.

The alloc_and_copy() might decide to not migrate all pages in the range (for reasons under the callback control). For those, the callback just has to leave the corresponding dst entry empty.

Finally, the migration of the struct page might fail (for file backed page) for various reasons (failure to freeze reference, or update page cache, ...). If that happens, then the finalize_and_map() can catch any pages that were not migrated. Note those pages were still copied to a new page and thus we wasted bandwidth but this is considered as a rare event and a price that we are willing to pay to keep all the code simpler.

Memory cgroup (memcg) and rss accounting

For now device memory is accounted as any regular page in rss counters (either anonymous if device page is used for anonymous, file if device page is used for file backed page or shmem if device page is used for shared memory). This is a deliberate choice to keep existing applications, that might start using device memory without knowing about it, running unimpacted.

A drawback is that the OOM killer might kill an application using a lot of device memory and not a lot of regular system memory and thus not freeing much system memory. We want to gather more real world experience on how applications and system react under memory pressure in the presence of device memory before deciding to account device memory differently.

Same decision was made for memory cgroup. Device memory pages are accounted against same memory cgroup a regular page would be accounted to. This does simplify migration to and from device memory. This also means that migration back from device memory to regular memory cannot fail because it would go above memory cgroup limit. We might revisit this choice latter on once we get more experience in how device memory is used and its impact on memory resource control.

Note that device memory can never be pinned by device driver nor through GUP and thus such memory is always free upon process exit. Or when last reference is dropped in case of shared memory or file backed memory.