The problem:
Accelerator driver developers are forced to reinvent external MM subsystems
case by case, because Linux core MM only considers host memory resources.
These reinvented MM subsystems have similar orders of magnitude of LoC as
Linux MM (80K), e.g. Nvidia-UVM has 70K, AMD GPU has 14K and Huawei NPU has
30K. Meanwhile, more and more vendors are implementing their own
accelerators, e.g. Microsoft's Maia 100. At the same time,
application-level developers suffer from poor programmability -- they must
consider parallel address spaces and be careful about the limited device
DRAM capacity. This can be alleviated if a malloc()-ed virtual address can
be shared by the accelerator, or the abundant host DRAM can further
transparently backup the device local memory.
These external MM systems share similar mechanisms except for the
hardware-dependent part, so reinventing them is effectively introducing
redundant code (14K~70K for each case). Such developing/maintaining is not
cheap. Furthermore, to share a malloc()-ed virtual address, device drivers
need to deeply interact with Linux MM via low-level MM APIs, e.g. MMU
notifiers/HMM. This raises the bar for driver development, since developers
must understand how Linux MM works. Further, it creates code maintenance
problems -- any changes to Linux MM potentially require coordinated changes
to accelerator drivers using low-level MM APIs.
Putting a cache-coherent bus between host and device will not make these
external MM subsystems disappear. For example, a throughput-oriented
accelerator will not tolerate executing heavy memory access workload with
a host MMU/IOMMU via a remote bus. Therefore, devices will still have
their own MMU and pick a simpler page table format for lower address
translation overhead, requiring external MM subsystems.
--------------------
What GMEM (Generalized Memory Management [1]) does:
GMEM extends Linux MM to share its machine-independent MM code. Only
high-level interface is provided for device drivers. This prevents
accelerator drivers from reinventing the wheel, but relies on drivers to
implement their hardware-dependent functions declared by GMEM. GMEM's key
interface include gm_dev_create(), gm_as_create(), gm_as_attach() and
gm_dev_register_physmem(). Here briefly describe how a device driver
utilizes them:
1. At boot time, call gm_dev_create() and registers the implementation of
hardware-dependent functions as declared in struct gm_mmu.
- If the device has local DRAM, call gm_dev_register_physmem() to
register available physical addresses.
2. When a device context is initialized (e.g. triggered by ioctl), check if
the current CPU process has been attached to a gmem address space
(struct gm_as). If not, call gm_as_create() and point current->mm->gm_as
to it.
3. Call gm_as_attach() to attach the device context to a gmem address space.
4. Invoke gm_dev_fault() to resolve a page fault or prepare data before
device computation happens.
GMEM has changed the following assumptions in Linux MM:
1. An mm_struct not only handle a single CPU context, but may also handle
external memory contexts encapsulated as gm_context listed in
mm->gm_as. An external memory context can include a few or all of the
following parts: an external MMU (that requires TLB invalidation), an
external page table (that requires PTE manipulation) and external DRAM
(that requires physical memory management).
2. Faulting a MAP_PRIVATE VMA with no CPU PTE found does not necessarily
mean that a zero-filled physical page should be mapped. The virtual
page may have been mapped to an external memory device.
3. Unmapping a page may include sending device TLB invalidation (even if
its MMU shares CPU page table) and manipulating device PTEs.
--------------------
Semantics of new syscalls:
1. mmap(..., MAP_PRIVATE | MAP_PEER_SHARED)
Allocate virtual address that is shared between the CPU and all
attached devices. Data is guaranteed to be coherent whenever the
address is accessed by either CPU or any attached device. If the device
does not support page fault, then device driver is responsible for
faulting memory before data gets accessed. By default, the CPU DRAM is
can be used as a swap backup for the device local memory.
2. hmadvise(NUMA_id, va_start, size, memory_hint)
Issuing memory hint for a given VMA. This extends traditional madvise()
syscall with an extra argument so that programmers have better control
with heterogeneous devices registered as NUMA nodes. One useful memory
hint could be MADV_PREFETCH, which guarantees that the physical data of
the given VMA [VA, VA+size) is migrated to NUMA node #id. Another
useful memory hint is MADV_DONTNEED. This is helpful to increase device
memory utilization. It is worth considering extending the existing
madvise() syscall with one additional argument.
--------------------
Implementation details
1. New VMA flag: MAP_PEER_SHARED
This new flag helps isolate GMEM feature, so that common processes with
no device attached does not need to maintain any logical page table. It
can be deleted if the extra overhead from GMEM is acceptable.
2. MMU functions
The device driver must implement the MMU functions declared in struct
gm_mmu.
VA functions: peer_va_alloc_fixed(), peer_va_free()
They are used to negotiate a common available VMA between a host
process and a device process at the mmap() time. This is because some
accelerators like Intel Xeon Phi or Huawei's Ascend NPU have their
acceleration tasks executed within a device CPU process context. Some
accelerators may also choose a different format of virtual address
space.
PA functions: alloc_page(), free_page(), prepare_page()
Alloc_page() and free_page() are used to allocate and free device physical
pages. Prepare_page() is used to zero-fill or DMA the data of a physical
page. These functions were removed from the submitted patch, since GMEM
does not need to invoke them when testing Huawei's NPU accelerator. The NPU
accelerator has an OS running in the device that manages the device
physical memory. However, even for such a device it is better for the host
to directly manage device physical memory, which saves device HBM and
avoids synchronizing management status between the host and device.
Page-table functions: pmap_create()/destroy()/enter()/release()/protect()
They are used to create and destroy device page tables, install and
uninstall page table entries and to change the protection of page table
entries.
TLB-invalidation functions: tlb_invl(), tlb_invl_coalesced()
They are used to invalidate the TLB entries of a given range of VA or
invalidate a given list of VMAs.
Wrapper functions: peer_map() and peer_unmap()
These two functions are used to create or destroy a device mapping which
could include allocating physical memory and copying data. They effectively
wraps the PA functions, Page-table functions and TLB-invalidation
functions. Implementing these steps together allows devices to optimize the
communication cost between host and device. However, it requires the device
driver to correctly order these steps.
3. Tracking logical mappings:
Each process starts maintaining an xarray in mm->vm_obj->logical_page_table
at the first time a host process calls mmap(MAP_PRIVATE | MAP_PEER_SHARED).
When a virtual page gets touched, its mapping status is created and stored
in struct gm_mapping. The logical page table is utilized to query the
struct gm_mapping given a virtual address. GMEM extends Linux MM to update
and lookup these logical mappings. For example, in the patch set we modify
the page fault path of to additionally check the logical mapping of
MAP_PEER_SHARED VMAs and identify if a device page should be migrated.
Similarly, if the device driver wants to resolve a device page fault or
prefetch data, the driver should call gm_dev_fault(). This function
examines the mapping status and determines whether the device driver should
migrate a CPU page to device or install a zero-filled device page.
The logical mapping abstraction enhances the extensibility of Linux core MM
(a virtual page may be mapped to a device physical page without any CPU PTE
installed). The current implementation is not complete, since it only
focused on anonymous VMAs with MAP_PEER_SHARED flag. The future plan of
logical page table is to provide a generic abstraction layer that support
common anonymous memory (I am looking at you, transparent huge pages) and
file-backed memory.
--------------------
Use cases
GMEM has been tested over Huawei's NPU (neural process unit) device driver.
The original NPU device driver has approximately 30,000 lines of code for
memory management. On the contrary, the GMEM-based one has less than 30
lines of code calling GMEM API, with approximately 3,700 lines of code
implementing the MMU functions. This effectively saves over 26,200 lines
of MM code for one driver. Therefore, developers from accelerator vendors,
including Nvidia, AMD, Intel and other companies are welcome to discuss if
GMEM could be helpful.
Using GMEM-based driver, it is possible to write a C-style accelerator code
with malloc(), whose underlying mmap() syscall should include
MAP_PEER_SHARED according to current GMEM implementation. Importantly, GMEM
guarantees a coherent view of memory between the host and all attached
devices. This means that any data written by the CPU or any attached
accelerator can be seen by the next memory load instruction issued by any
attached accelerator or the CPU. Furthermore, the NPU device was able to
oversubscribe memory by swapping memory to host DDR. Note that this memory
oversubscription mechanism can be universal if the physical memory
management is provided by GMEM. Other potential use cases of GMEM could
include the IOMMU driver, KVM and RDMA drivers, as long as the device needs
to manage external memory resources like VMAs, MMUs or local DRAMs.
--------------------
Discussion
Physical memory management
Most accelerators require the host OS to manage device DRAM. Even
accelerators capable of running an OS inside the driver can benefit from
it, since it helps avoid synchronizing management status between the host
and device. In Linux OSS EU summit 2023, Hannes Reinecke from SUSE Labs
suggested that people are concerned with the memory consumption of struct
page (which considers all generic scenarios for the kernel). This leads to
a possible solution that, instead of reusing Linux struct page and
ZONE_DEVICE mechanism, GMEM can implement an isolated buddy allocator for
the device to instantiate and register. The isolation is useful because
device DRAM physical address space is independent. Furthermore, the
isolated buddy allocator can utilize a customized struct page that consumes
less memory. It is worth discussing if accelerator vendors desire this
solution.
MMU functions
The MMU functions peer_map() and peer_unmap() overlap other functions,
leaving a question if the MMU functions should be decoupled as more basic
operations. Decoupling them could potentially prevent device drivers
coalescing these basic steps within a single host-device communication
operation, while coupling them makes it more difficult for device drivers
to utilize GMEM interface.
The idea of GMEM was originated from Weixi's PhD study with
Prof. Scott Rixner and Prof. Alan L. Cox at Rice University.
[1] https://arxiv.org/abs/2310.12554.
Weixi Zhu (6):
mm/gmem: add heterogeneous NUMA node
mm/gmem: add arch-independent abstraction to track address mapping
status
mm/gmem: add GMEM (Generalized Memory Management) interface for
external accelerators
mm/gmem: add new syscall hmadvise() to issue memory hints for
heterogeneous NUMA nodes
mm/gmem: resolve VMA conflicts for attached peer devices
mm/gmem: extending Linux core MM to support unified virtual address
space
arch/arm64/include/asm/unistd.h | 2 +-
arch/arm64/include/asm/unistd32.h | 2 +
drivers/base/node.c | 6 +
fs/proc/task_mmu.c | 3 +
include/linux/gmem.h | 368 ++++++++++++
include/linux/mm.h | 8 +
include/linux/mm_types.h | 5 +
include/linux/nodemask.h | 10 +
include/uapi/asm-generic/mman-common.h | 4 +
include/uapi/asm-generic/unistd.h | 5 +-
init/main.c | 2 +
kernel/fork.c | 5 +
kernel/sys_ni.c | 2 +
mm/Kconfig | 14 +
mm/Makefile | 1 +
mm/gmem.c | 746 ++++++++++++++++++++++++
mm/huge_memory.c | 85 ++-
mm/memory.c | 42 +-
mm/mempolicy.c | 4 +
mm/mmap.c | 40 +-
mm/oom_kill.c | 2 +
mm/page_alloc.c | 3 +
mm/vm_object.c | 309 ++++++++++
tools/include/uapi/asm-generic/unistd.h | 5 +-
24 files changed, 1654 insertions(+), 19 deletions(-)
create mode 100644 include/linux/gmem.h
create mode 100644 mm/gmem.c
create mode 100644 mm/vm_object.c
--
2.25.1
