mm: Free some vmemmap pages of HugeTLB page

\

```
Since Mike's patches (make hugetlb put_page safe for all calling contexts[1])
applied into the next-20210412. We can move forward on this patch series now.

This patch series will free some vmemmap pages(struct page structures)
associated with each HugeTLB page when preallocated to save memory.

In order to reduce the difficulty of the first version of code review.
>From this version, we disable PMD/huge page mapping of vmemmap if this
feature was enabled. This acutely eliminates a bunch of the complex code
doing page table manipulation. When this patch series is solid, we cam add
the code of vmemmap page table manipulation in the future.

The struct page structures (page structs) are used to describe a physical
page frame. By default, there is an one-to-one mapping from a page frame to
it's corresponding page struct.

The HugeTLB pages consist of multiple base page size pages and is supported
by many architectures. See hugetlbpage.rst in the Documentation directory
for more details. On the x86 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
4096 base pages. For each base page, there is a corresponding page struct.

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

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

When the system boot up, every 2M HugeTLB has 512 struct page structs which
size is 8 pages(sizeof(struct page) * 512 / PAGE_SIZE).

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

The value of page->compound_head is the same for all tail pages. The first
page of page structs (page 0) associated with the HugeTLB page contains the 4
page structs necessary to describe the HugeTLB. The only use of the remaining
pages of page structs (page 1 to page 7) is to point to page->compound_head.
Therefore, we can remap pages 2 to 7 to page 1. Only 2 pages of page structs
will be used for each HugeTLB page. This will allow us to free the remaining
6 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     | -------------> |     1     |
 |           |                     +-----------+                +-----------+
 |           |                     |     2     | ----------------^ ^ ^ ^ ^ ^
 |           |                     +-----------+                   | | | | |
 |           |                     |     3     | ------------------+ | | | |
 |           |                     +-----------+                     | | | |
 |           |                     |     4     | --------------------+ | | |
 |    2MB    |                     +-----------+                       | | |
 |           |                     |     5     | ----------------------+ | |
 |           |                     +-----------+                         | |
 |           |                     |     6     | ------------------------+ |
 |           |                     +-----------+                           |
 |           |                     |     7     | --------------------------+
 |           |                     +-----------+
 |           |
 |           |
 |           |
 +-----------+

When a HugeTLB is freed to the buddy system, we should allocate 6 pages for
vmemmap pages and restore the previous mapping relationship.

Apart from 2MB HugeTLB page, we also have 1GB HugeTLB page. It is similar
to the 2MB HugeTLB page. We also can use this approach to free the vmemmap
pages.

In this case, for the 1GB HugeTLB page, we can save 4094 pages. This is a
very substantial gain. On our server, run some SPDK/QEMU applications which
will use 1024GB HugeTLB page. With this feature enabled, we can save ~16GB
(1G hugepage)/~12GB (2MB hugepage) memory.

Because there are vmemmap page tables reconstruction on the freeing/allocating
path, it increases some overhead. Here are some overhead analysis.

1) Allocating 10240 2MB HugeTLB pages.

a) With this patch series applied:
   # time echo 10240 > /proc/sys/vm/nr_hugepages

real     0m0.166s
   user     0m0.000s
   sys      0m0.166s

# bpftrace -e 'kprobe:alloc_fresh_huge_page { @start[tid] = nsecs; }
     kretprobe:alloc_fresh_huge_page /@start[tid]/ { @latency = hist(nsecs -
     @start[tid]); delete(@start[tid]); }'
   Attaching 2 probes...

@latency:
   [8K, 16K)           5476 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
   [16K, 32K)          4760 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@       |
   [32K, 64K)             4 |                                                    |

b) Without this patch series:
   # time echo 10240 > /proc/sys/vm/nr_hugepages

real     0m0.067s
   user     0m0.000s
   sys      0m0.067s

# bpftrace -e 'kprobe:alloc_fresh_huge_page { @start[tid] = nsecs; }
     kretprobe:alloc_fresh_huge_page /@start[tid]/ { @latency = hist(nsecs -
     @start[tid]); delete(@start[tid]); }'
   Attaching 2 probes...

@latency:
   [4K, 8K)           10147 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
   [8K, 16K)             93 |                                                    |

Summarize: this feature is about ~2x slower than before.

2) Freeing 10240 2MB HugeTLB pages.

a) With this patch series applied:
   # time echo 0 > /proc/sys/vm/nr_hugepages

real     0m0.213s
   user     0m0.000s
   sys      0m0.213s

# bpftrace -e 'kprobe:free_pool_huge_page { @start[tid] = nsecs; }
     kretprobe:free_pool_huge_page /@start[tid]/ { @latency = hist(nsecs -
     @start[tid]); delete(@start[tid]); }'
   Attaching 2 probes...

@latency:
   [8K, 16K)              6 |                                                    |
   [16K, 32K)         10227 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
   [32K, 64K)             7 |                                                    |

b) Without this patch series:
   # time echo 0 > /proc/sys/vm/nr_hugepages

real     0m0.081s
   user     0m0.000s
   sys      0m0.081s

# bpftrace -e 'kprobe:free_pool_huge_page { @start[tid] = nsecs; }
     kretprobe:free_pool_huge_page /@start[tid]/ { @latency = hist(nsecs -
     @start[tid]); delete(@start[tid]); }'
   Attaching 2 probes...

@latency:
   [4K, 8K)            6805 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
   [8K, 16K)           3427 |@@@@@@@@@@@@@@@@@@@@@@@@@@                          |
   [16K, 32K)             8 |                                                    |

Summarize: The overhead of __free_hugepage is about ~2-3x slower than before.

Although the overhead has increased, the overhead is not significant. Like Mike
said, "However, remember that the majority of use cases create HugeTLB pages at
or shortly after boot time and add them to the pool. So, additional overhead is
at pool creation time. There is no change to 'normal run time' operations of
getting a page from or returning a page to the pool (think page fault/unmap)".

Despite the overhead and in addition to the memory gains from this series. The
following data is obtained by Joao Martins. Very thanks to his effort.

There's an additional benefit which is page (un)pinners will see an improvement
and Joao presumes because there are fewer memmap pages and thus the tail/head
pages are staying in cache more often.

Out of the box Joao saw (when comparing linux-next against linux-next + this series)
with gup_test and pinning a 16G HugeTLB file (with 1G pages):

get_user_pages(): ~32k -> ~9k
	unpin_user_pages(): ~75k -> ~70k

Usually any tight loop fetching compound_head(), or reading tail pages data (e.g.
compound_head) benefit a lot. There's some unpinning inefficiencies Joao was
fixing[2], but with that in added it shows even more:

unpin_user_pages(): ~27k -> ~3.8k

[1] https://lore.kernel.org/linux-mm/20210409205254.242291-1-mike.kravetz@oracle.com/
[2] https://lore.kernel.org/linux-mm/20210204202500.26474-1-joao.m.martins@oracle.com/

Todo:
  - Free all of the tail vmemmap pages
    Now for the 2MB HugrTLB page, we only free 6 vmemmap pages. we really can
    free 7 vmemmap pages. In this case, we can see 8 of the 512 struct page
    structures has beed set PG_head flag. If we can adjust compound_head()
    slightly and make compound_head() return the real head struct page when
    the parameter is the tail struct page but with PG_head flag set.

In order to make the code evolution route clearer. This feature can can be
    a separate patch after this patchset is solid.

- Support for other architectures (e.g. aarch64).
  - Enable PMD/huge page mapping of vmemmap even if this feature was enabled.
```

GVP openEuler/kernel

内容风险标识

评论 (11)

GVPopenEuler/kernel

内容风险标识