A vmemmap diet for HugeTLB and Device DAX

HugeTLB

This section is to explain how HugeTLB Vmemmap Optimization (HVO) works.

The struct page structures are used to describe a physical page frame. By default, there is a one-to-one mapping from a page frame to its corresponding struct page.

HugeTLB pages consist of multiple base page size pages and is supported by many architectures. See HugeTLB Pages for more details. On the x86-64 architecture, HugeTLB pages of size 2MB and 1GB are currently supported. Since the base page size on x86 is 4KB, a 2MB HugeTLB page consists of 512 base pages and a 1GB HugeTLB page consists of 262144 base pages. For each base page, there is a corresponding struct page.

Within the HugeTLB subsystem, only the first 4 struct page are used to contain unique information about a HugeTLB page. __NR_USED_SUBPAGE provides this upper limit. The only ‘useful’ information in the remaining struct page is the compound_head field, and this field is the same for all tail pages.

By removing redundant struct page for HugeTLB pages, memory can be returned to the buddy allocator for other uses.

Different architectures support different HugeTLB pages. For example, the following table is the HugeTLB page size supported by x86 and arm64 architectures. Because arm64 supports 4k, 16k, and 64k base pages and supports contiguous entries, so it supports many kinds of sizes of HugeTLB page.

Architecture

Page Size

HugeTLB Page Size

x86-64

4KB

2MB

1GB

arm64

4KB

64KB

2MB

32MB

1GB

16KB

2MB

32MB

1GB

64KB

2MB

512MB

16GB

When the system boot up, every HugeTLB page has more than one struct page structs which size is (unit: pages):

struct_size = HugeTLB_Size / PAGE_SIZE * sizeof(struct page) / PAGE_SIZE

Where HugeTLB_Size is the size of the HugeTLB page. We know that the size of the HugeTLB page is always n times PAGE_SIZE. So we can get the following relationship:

HugeTLB_Size = n * PAGE_SIZE

Then:

struct_size = n * PAGE_SIZE / PAGE_SIZE * sizeof(struct page) / PAGE_SIZE
            = n * sizeof(struct page) / PAGE_SIZE

We can use huge mapping at the pud/pmd level for the HugeTLB page.

For the HugeTLB page of the pmd level mapping, then:

struct_size = n * sizeof(struct page) / PAGE_SIZE
            = PAGE_SIZE / sizeof(pte_t) * sizeof(struct page) / PAGE_SIZE
            = sizeof(struct page) / sizeof(pte_t)
            = 64 / 8
            = 8 (pages)

Where n is how many pte entries which one page can contains. So the value of n is (PAGE_SIZE / sizeof(pte_t)).

This optimization only supports 64-bit system, so the value of sizeof(pte_t) is 8. And this optimization also applicable only when the size of struct page is a power of two. In most cases, the size of struct page is 64 bytes (e.g. x86-64 and arm64). So if we use pmd level mapping for a HugeTLB page, the size of struct page structs of it is 8 page frames which size depends on the size of the base page.

For the HugeTLB page of the pud level mapping, then:

struct_size = PAGE_SIZE / sizeof(pmd_t) * struct_size(pmd)
            = PAGE_SIZE / 8 * 8 (pages)
            = PAGE_SIZE (pages)

Where the struct_size(pmd) is the size of the struct page structs of a HugeTLB page of the pmd level mapping.

E.g.: A 2MB HugeTLB page on x86_64 consists in 8 page frames while 1GB HugeTLB page consists in 4096.

Next, we take the pmd level mapping of the HugeTLB page as an example to show the internal implementation of this optimization. There are 8 pages struct page structs associated with a HugeTLB page which is pmd mapped.

Here is how things look before optimization:

   HugeTLB                  struct pages(8 pages)         page frame(8 pages)
+-----------+ ---virt_to_page---> +-----------+   mapping to   +-----------+
|           |                     |     0     | -------------> |     0     |
|           |                     +-----------+                +-----------+
|           |                     |     1     | -------------> |     1     |
|           |                     +-----------+                +-----------+
|           |                     |     2     | -------------> |     2     |
|           |                     +-----------+                +-----------+
|           |                     |     3     | -------------> |     3     |
|           |                     +-----------+                +-----------+
|           |                     |     4     | -------------> |     4     |
|    PMD    |                     +-----------+                +-----------+
|   level   |                     |     5     | -------------> |     5     |
|  mapping  |                     +-----------+                +-----------+
|           |                     |     6     | -------------> |     6     |
|           |                     +-----------+                +-----------+
|           |                     |     7     | -------------> |     7     |
|           |                     +-----------+                +-----------+
|           |
|           |
|           |
+-----------+

The value of page->compound_head is the same for all tail pages. The first page of struct page (page 0) associated with the HugeTLB page contains the 4 struct page necessary to describe the HugeTLB. The only use of the remaining pages of struct page (page 1 to page 7) is to point to page->compound_head. Therefore, we can remap pages 1 to 7 to page 0. Only 1 page of struct page will be used for each HugeTLB page. This will allow us to free the remaining 7 pages to the buddy allocator.

Here is how things look after remapping:

   HugeTLB                  struct pages(8 pages)         page frame(8 pages)
+-----------+ ---virt_to_page---> +-----------+   mapping to   +-----------+
|           |                     |     0     | -------------> |     0     |
|           |                     +-----------+                +-----------+
|           |                     |     1     | ---------------^ ^ ^ ^ ^ ^ ^
|           |                     +-----------+                  | | | | | |
|           |                     |     2     | -----------------+ | | | | |
|           |                     +-----------+                    | | | | |
|           |                     |     3     | -------------------+ | | | |
|           |                     +-----------+                      | | | |
|           |                     |     4     | ---------------------+ | | |
|    PMD    |                     +-----------+                        | | |
|   level   |                     |     5     | -----------------------+ | |
|  mapping  |                     +-----------+                          | |
|           |                     |     6     | -------------------------+ |
|           |                     +-----------+                            |
|           |                     |     7     | ---------------------------+
|           |                     +-----------+
|           |
|           |
|           |
+-----------+

When a HugeTLB is freed to the buddy system, we should allocate 7 pages for vmemmap pages and restore the previous mapping relationship.

For the HugeTLB page of the pud level mapping. It is similar to the former. We also can use this approach to free (PAGE_SIZE - 1) vmemmap pages.

Apart from the HugeTLB page of the pmd/pud level mapping, some architectures (e.g. aarch64) provides a contiguous bit in the translation table entries that hints to the MMU to indicate that it is one of a contiguous set of entries that can be cached in a single TLB entry.

The contiguous bit is used to increase the mapping size at the pmd and pte (last) level. So this type of HugeTLB page can be optimized only when its size of the struct page structs is greater than 1 page.

Notice: The head vmemmap page is not freed to the buddy allocator and all tail vmemmap pages are mapped to the head vmemmap page frame. So we can see more than one struct page struct with PG_head (e.g. 8 per 2 MB HugeTLB page) associated with each HugeTLB page. The compound_head() can handle this correctly. There is only one head struct page, the tail struct page with PG_head are fake head struct page. We need an approach to distinguish between those two different types of struct page so that compound_head() can return the real head struct page when the parameter is the tail struct page but with PG_head.

Device DAX

The device-dax interface uses the same tail deduplication technique explained in the previous chapter, except when used with the vmemmap in the device (altmap).

The following page sizes are supported in DAX: PAGE_SIZE (4K on x86_64), PMD_SIZE (2M on x86_64) and PUD_SIZE (1G on x86_64). For powerpc equivalent details see Device DAX

The differences with HugeTLB are relatively minor.

It only use 3 struct page for storing all information as opposed to 4 on HugeTLB pages.

There’s no remapping of vmemmap given that device-dax memory is not part of System RAM ranges initialized at boot. Thus the tail page deduplication happens at a later stage when we populate the sections. HugeTLB reuses the the head vmemmap page representing, whereas device-dax reuses the tail vmemmap page. This results in only half of the savings compared to HugeTLB.

Deduplicated tail pages are not mapped read-only.

Here’s how things look like on device-dax after the sections are populated:

+-----------+ ---virt_to_page---> +-----------+   mapping to   +-----------+
|           |                     |     0     | -------------> |     0     |
|           |                     +-----------+                +-----------+
|           |                     |     1     | -------------> |     1     |
|           |                     +-----------+                +-----------+
|           |                     |     2     | ----------------^ ^ ^ ^ ^ ^
|           |                     +-----------+                   | | | | |
|           |                     |     3     | ------------------+ | | | |
|           |                     +-----------+                     | | | |
|           |                     |     4     | --------------------+ | | |
|    PMD    |                     +-----------+                       | | |
|   level   |                     |     5     | ----------------------+ | |
|  mapping  |                     +-----------+                         | |
|           |                     |     6     | ------------------------+ |
|           |                     +-----------+                           |
|           |                     |     7     | --------------------------+
|           |                     +-----------+
|           |
|           |
|           |
+-----------+