linux_dsm_epyc7002/Documentation/vm/hmm.txt

.. hmm:

=====================================
Heterogeneous Memory Management (HMM)
=====================================

Transparently allow any component of a program to use any memory region of said
program with a device without using device specific memory allocator. This is
becoming a requirement to simplify the use of advance heterogeneous computing
where GPU, DSP or FPGA are use to perform various computations.

This document is divided as follow, in the first section i expose the problems
related to the use of a device specific allocator. The second section i expose
the hardware limitations that are inherent to many platforms. The third section
gives an overview of HMM designs. The fourth section explains how CPU page-
table mirroring works and what is HMM purpose in this context. Fifth section
deals with how device memory is represented inside the kernel. Finaly the last
section present the new migration helper that allow to leverage the device DMA
engine.

.. contents:: :local:

Problems of using device specific memory allocator
==================================================

Device with large amount of on board memory (several giga bytes) like GPU have
historically manage their memory through dedicated driver specific API. This
creates a disconnect between memory allocated and managed by device driver and
regular application memory (private anonymous, share memory or regular file
back memory). From here on i will refer to this aspect as split address space.
I use share address space to refer to the opposite situation ie one in which
any memory region can be use by device transparently.

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

Concretly this means that code that wants to leverage device like GPU need to
copy object between genericly allocated memory (malloc, mmap private/share/)
and memory allocated through the device driver API (this still end up with an
mmap but of the device file).

For flat dataset (array, grid, image, ...) this isn't too hard to achieve but
complex data-set (list, tree, ...) are hard to get right. Duplicating a complex
data-set need 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.

Split address space also means that library can not transparently use data they
are getting from core program or other library and thus each library might have
to duplicate its input data-set using specific memory allocator. Large project
suffer from this and waste resources because of the various memory copy.

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

Finaly with the advance of high level language constructs (in C++ but in other
language too) it is now possible for compiler to leverage GPU or other devices
without even the programmer knowledge. Some of compiler identified patterns are
only do-able with a share address. It is as well more reasonable to use a share
address space for all the other patterns.


System bus, device memory characteristics
=========================================

System bus cripple share address due to few limitations. Most system bus 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, most
often than not it is not cache coherent.

If we only consider the PCIE bus than 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 operation from device on main memory. This is worse
in the other direction the CPUs can only access a limited range of the device
memory and can not perform atomic operations on it. Thus device memory can not
be consider like regular memory from 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 that 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 access its own memory.

Some platform are developing new system bus or additions/modifications to PCIE
to address some of those limitations (OpenCAPI, CCIX). They mainly allow two
way cache coherency between CPU and device and allow all atomic operations the
architecture supports. Saddly not all platform are following this trends and
some major architecture are left without hardware solutions to those problems.

So for share address space to make sense not only we must allow device to
access any memory 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).


Share address space and migration
=================================

HMM intends to provide two main features. First one is to share the address
space by duplication the CPU page table into the device page table so same
address point to same memory and this for any valid main memory address in
the process address space.

To achieve this, HMM offer 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
allow a buffer (or use a pool of pre-allocated buffer) and write GPU specifics
commands in it to perform the update (unmap, cache invalidations and flush,
...). This can not be done through common code for all device. Hence why HMM
provides helpers to factor out everything that can be while leaving the gory
details to the device driver.

The second mechanism HMM provide is a new kind of ZONE_DEVICE memory that does
allow to allocate a struct page for each page of the device memory. Those page
are special because the CPU can not map them. They however allow to migrate
main memory to device memory using exhisting migration mechanism and everything
looks like if page was swap out to disk from CPU point of view. Using a struct
page gives the easiest and cleanest integration with existing mm mechanisms.
Again here HMM only provide helpers, first to hotplug new ZONE_DEVICE memory
for the device memory and second to perform migration. Policy decision of what
and when to migrate things is left to the device driver.

Note that any CPU access to a device page trigger a page fault and a migration
back to main memory ie when a page backing an given address A is migrated from
a main memory page to a device page then any CPU access to address A trigger a
page fault and initiate a migration back to main memory.


With this two features, HMM not only allow a device to mirror a process address
space and keeps both CPU and device page table synchronize, but also allow to
leverage device memory by migrating part of data-set that is actively use by a
device.


Address space mirroring implementation and API
==============================================

Address space mirroring main objective is to allow to duplicate range of CPU
page table into a device page table and HMM helps keeping both synchronize. A
device driver that want 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 use when the driver is already holding the mmap_sem
of the mm in write mode. The mirror struct has a set of callback that are use
to propagate CPU page table::

 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);
 };

Device driver must perform update to the range following action (turn range
read only, or fully unmap, ...). Once driver callback returns the device must
be done with the update.


When device driver wants to populate a range of virtual address 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);

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

Both function copy CPU page table into their pfns array argument. Each entry in
that array correspond to an address in the virtual range. HMM provide a set of
flags to help 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 synchronize. The usage pattern is::

 int driver_populate_range(...)
 {
      struct hmm_range range;
      ...
 again:
      ret = hmm_vma_get_pfns(vma, &range, start, end, pfns);
      if (ret)
          return ret;
      take_lock(driver->update);
      if (!hmm_vma_range_done(vma, &range)) {
          release_lock(driver->update);
          goto again;
      }

      // Use pfns array content to update device page table

      release_lock(driver->update);
      return 0;
 }

The driver->update lock is the same lock that driver takes inside its update()
callback. That lock must be call 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 to a
simpler API and also to be able to perform optimization latter own like doing
concurrent device update in multi-devices scenario.

HMM also serve as an impedence missmatch between how CPU page table update are
done (by CPU write to the page table and TLB flushes) from how device update
their own page table. Device update is a multi-step process, first appropriate
commands are write to a buffer, then this buffer is schedule 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 update command buffer can happen
concurrently for multiple devices. Waiting for each device to report commands
as executed is serialize (there is no point in doing this concurrently).


Represent and manage device memory from core kernel point of view
=================================================================

Several differents design were try to support device memory. First one use
device specific data structure to keep information about migrated memory and
HMM hooked itself in various place of mm code to handle any access to address
that were back by device memory. It turns out that this ended up replicating
most of the fields of struct page and also needed many kernel code path to be
updated to understand this new kind of memory.

Thing is most kernel code path never try to access the memory behind a page
but only care about struct page contents. Because of this HMM switchted to
directly using struct page for device memory which left most kernel code path
un-aware of the difference. We only need to make sure that no one ever try to
map those page from the CPU side.

HMM provide a set of helpers to register and hotplug device memory as a new
region needing struct page. This is offer 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
drop. This means the device page is now free and no longer use by anyone. The
second callback happens whenever CPU try to access a device page which it can
not do. This second callback must trigger a migration back to system memory.


Migrate to and from device memory
=================================

Because CPU can not access device memory, migration must use 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 function it works on a range of virtual address, there
is 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
make the whole excersie pointless. The second reason is because driver trigger
such migration base on range of address the device is actively accessing.

The migrate_vma_ops struct define two callbacks. First one (alloc_and_copy())
control destination memory allocation and copy operation. Second one is there
to allow device driver to perform cleanup operation 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 this migration helpers allow for hole in the
virtual address range. Some pages in the range might not be migrated for all
the usual reasons (page is pin, page is lock, ...). This helper does not fail
but just skip over those pages.

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

Finaly the migration of the struct page might fails (for file back 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 was not
migrated. Note those page were still copied to 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 use for anonymous, file if device page is use for
file back page or shmem if device page is use for share memory). This is a
deliberate choice to keep existing application that might start using device
memory without knowing about it to keep runing unimpacted.

Drawbacks is that 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 application 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 page 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 can not 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 use and its impact on memory
resource control.


Note that device memory can never be pin nor by device driver nor through GUP
and thus such memory is always free upon process exit. Or when last reference
is drop in case of share memory or file back memory.