mirror of
https://github.com/torvalds/linux
synced 2024-09-06 18:01:05 +00:00
mm: hugetlb: free the 2nd vmemmap page associated with each HugeTLB page
Patch series "Free the 2nd vmemmap page associated with each HugeTLB page", v7. This series can minimize the overhead of struct page for 2MB HugeTLB pages significantly. It further reduces the overhead of struct page by 12.5% for a 2MB HugeTLB compared to the previous approach, which means 2GB per 1TB HugeTLB. It is a nice gain. Comments and reviews are welcome. Thanks. The main implementation and details can refer to the commit log of patch 1. In this series, I have changed the following four helpers, the following table shows the impact of the overhead of those helpers. +------------------+-----------------------+ | APIs | head page | tail page | +------------------+-----------+-----------+ | PageHead() | Y | N | +------------------+-----------+-----------+ | PageTail() | Y | N | +------------------+-----------+-----------+ | PageCompound() | N | N | +------------------+-----------+-----------+ | compound_head() | Y | N | +------------------+-----------+-----------+ Y: Overhead is increased. N: Overhead is _NOT_ increased. It shows that the overhead of those helpers on a tail page don't change between "hugetlb_free_vmemmap=on" and "hugetlb_free_vmemmap=off". But the overhead on a head page will be increased when "hugetlb_free_vmemmap=on" (except PageCompound()). So I believe that Matthew Wilcox's folio series will help with this. The users of PageHead() and PageTail() are much less than compound_head() and most users of PageTail() are VM_BUG_ON(), so I have done some tests about the overhead of compound_head() on head pages. I have tested the overhead of calling compound_head() on a head page, which is 2.11ns (Measure the call time of 10 million times compound_head(), and then average). For a head page whose address is not aligned with PAGE_SIZE or a non-compound page, the overhead of compound_head() is 2.54ns which is increased by 20%. For a head page whose address is aligned with PAGE_SIZE, the overhead of compound_head() is 2.97ns which is increased by 40%. Most pages are the former. I do not think the overhead is significant since the overhead of compound_head() itself is low. This patch (of 5): This patch minimizes the overhead of struct page for 2MB HugeTLB pages significantly. It further reduces the overhead of struct page by 12.5% for a 2MB HugeTLB compared to the previous approach, which means 2GB per 1TB HugeTLB (2MB type). After the feature of "Free sonme vmemmap pages of HugeTLB page" is enabled, the mapping of the vmemmap addresses associated with a 2MB HugeTLB page becomes the figure below. HugeTLB struct pages(8 pages) page frame(8 pages) +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+---> PG_head | | | 0 | -------------> | 0 | | | +-----------+ +-----------+ | | | 1 | -------------> | 1 | | | +-----------+ +-----------+ | | | 2 | ----------------^ ^ ^ ^ ^ ^ | | +-----------+ | | | | | | | | 3 | ------------------+ | | | | | | +-----------+ | | | | | | | 4 | --------------------+ | | | | 2MB | +-----------+ | | | | | | 5 | ----------------------+ | | | | +-----------+ | | | | | 6 | ------------------------+ | | | +-----------+ | | | | 7 | --------------------------+ | | +-----------+ | | | | | | +-----------+ As we can see, the 2nd vmemmap page frame (indexed by 1) is reused and remaped. However, the 2nd vmemmap page frame is also can be freed to the buddy allocator, then we can change the mapping from the figure above to the figure below. HugeTLB struct pages(8 pages) page frame(8 pages) +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+---> PG_head | | | 0 | -------------> | 0 | | | +-----------+ +-----------+ | | | 1 | ---------------^ ^ ^ ^ ^ ^ ^ | | +-----------+ | | | | | | | | | 2 | -----------------+ | | | | | | | +-----------+ | | | | | | | | 3 | -------------------+ | | | | | | +-----------+ | | | | | | | 4 | ---------------------+ | | | | 2MB | +-----------+ | | | | | | 5 | -----------------------+ | | | | +-----------+ | | | | | 6 | -------------------------+ | | | +-----------+ | | | | 7 | ---------------------------+ | | +-----------+ | | | | | | +-----------+ After we do this, all tail vmemmap pages (1-7) are mapped to the head vmemmap page frame (0). In other words, there are more than one page struct with PG_head associated with each HugeTLB page. We __know__ that there is only one head page struct, the tail page structs with PG_head are fake head page structs. We need an approach to distinguish between those two different types of page structs so that compound_head(), PageHead() and PageTail() can work properly if the parameter is the tail page struct but with PG_head. The following code snippet describes how to distinguish between real and fake head page struct. if (test_bit(PG_head, &page->flags)) { unsigned long head = READ_ONCE(page[1].compound_head); if (head & 1) { if (head == (unsigned long)page + 1) ==> head page struct else ==> tail page struct } else ==> head page struct } We can safely access the field of the @page[1] with PG_head because the @page is a compound page composed with at least two contiguous pages. [songmuchun@bytedance.com: restore lost comment changes] Link: https://lkml.kernel.org/r/20211101031651.75851-1-songmuchun@bytedance.com Link: https://lkml.kernel.org/r/20211101031651.75851-2-songmuchun@bytedance.com Signed-off-by: Muchun Song <songmuchun@bytedance.com> Reviewed-by: Barry Song <song.bao.hua@hisilicon.com> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Oscar Salvador <osalvador@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: David Hildenbrand <david@redhat.com> Cc: Chen Huang <chenhuang5@huawei.com> Cc: Bodeddula Balasubramaniam <bodeddub@amazon.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <willy@infradead.org> Cc: Xiongchun Duan <duanxiongchun@bytedance.com> Cc: Fam Zheng <fam.zheng@bytedance.com> Cc: Qi Zheng <zhengqi.arch@bytedance.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This commit is contained in:
parent
5c2a956c3e
commit
e7d324850b
|
@ -1625,7 +1625,7 @@
|
|||
[KNL] Reguires CONFIG_HUGETLB_PAGE_FREE_VMEMMAP
|
||||
enabled.
|
||||
Allows heavy hugetlb users to free up some more
|
||||
memory (6 * PAGE_SIZE for each 2MB hugetlb page).
|
||||
memory (7 * PAGE_SIZE for each 2MB hugetlb page).
|
||||
Format: { on | off (default) }
|
||||
|
||||
on: enable the feature
|
||||
|
|
|
@ -190,13 +190,69 @@ enum pageflags {
|
|||
|
||||
#ifndef __GENERATING_BOUNDS_H
|
||||
|
||||
#ifdef CONFIG_HUGETLB_PAGE_FREE_VMEMMAP
|
||||
extern bool hugetlb_free_vmemmap_enabled;
|
||||
|
||||
/*
|
||||
* If the feature of freeing some vmemmap pages associated with each HugeTLB
|
||||
* page is enabled, the head vmemmap page frame is reused and all of the tail
|
||||
* vmemmap addresses map to the head vmemmap page frame (furture details can
|
||||
* refer to the figure at the head of the mm/hugetlb_vmemmap.c). In other
|
||||
* words, there are more than one page struct with PG_head associated with each
|
||||
* HugeTLB page. We __know__ that there is only one head page struct, the tail
|
||||
* page structs with PG_head are fake head page structs. We need an approach
|
||||
* to distinguish between those two different types of page structs so that
|
||||
* compound_head() can return the real head page struct when the parameter is
|
||||
* the tail page struct but with PG_head.
|
||||
*
|
||||
* The page_fixed_fake_head() returns the real head page struct if the @page is
|
||||
* fake page head, otherwise, returns @page which can either be a true page
|
||||
* head or tail.
|
||||
*/
|
||||
static __always_inline const struct page *page_fixed_fake_head(const struct page *page)
|
||||
{
|
||||
if (!hugetlb_free_vmemmap_enabled)
|
||||
return page;
|
||||
|
||||
/*
|
||||
* Only addresses aligned with PAGE_SIZE of struct page may be fake head
|
||||
* struct page. The alignment check aims to avoid access the fields (
|
||||
* e.g. compound_head) of the @page[1]. It can avoid touch a (possibly)
|
||||
* cold cacheline in some cases.
|
||||
*/
|
||||
if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
|
||||
test_bit(PG_head, &page->flags)) {
|
||||
/*
|
||||
* We can safely access the field of the @page[1] with PG_head
|
||||
* because the @page is a compound page composed with at least
|
||||
* two contiguous pages.
|
||||
*/
|
||||
unsigned long head = READ_ONCE(page[1].compound_head);
|
||||
|
||||
if (likely(head & 1))
|
||||
return (const struct page *)(head - 1);
|
||||
}
|
||||
return page;
|
||||
}
|
||||
#else
|
||||
static inline const struct page *page_fixed_fake_head(const struct page *page)
|
||||
{
|
||||
return page;
|
||||
}
|
||||
#endif
|
||||
|
||||
static __always_inline int page_is_fake_head(struct page *page)
|
||||
{
|
||||
return page_fixed_fake_head(page) != page;
|
||||
}
|
||||
|
||||
static inline unsigned long _compound_head(const struct page *page)
|
||||
{
|
||||
unsigned long head = READ_ONCE(page->compound_head);
|
||||
|
||||
if (unlikely(head & 1))
|
||||
return head - 1;
|
||||
return (unsigned long)page;
|
||||
return (unsigned long)page_fixed_fake_head(page);
|
||||
}
|
||||
|
||||
#define compound_head(page) ((typeof(page))_compound_head(page))
|
||||
|
@ -231,12 +287,13 @@ static inline unsigned long _compound_head(const struct page *page)
|
|||
|
||||
static __always_inline int PageTail(struct page *page)
|
||||
{
|
||||
return READ_ONCE(page->compound_head) & 1;
|
||||
return READ_ONCE(page->compound_head) & 1 || page_is_fake_head(page);
|
||||
}
|
||||
|
||||
static __always_inline int PageCompound(struct page *page)
|
||||
{
|
||||
return test_bit(PG_head, &page->flags) || PageTail(page);
|
||||
return test_bit(PG_head, &page->flags) ||
|
||||
READ_ONCE(page->compound_head) & 1;
|
||||
}
|
||||
|
||||
#define PAGE_POISON_PATTERN -1l
|
||||
|
@ -695,7 +752,20 @@ static inline bool test_set_page_writeback(struct page *page)
|
|||
return set_page_writeback(page);
|
||||
}
|
||||
|
||||
__PAGEFLAG(Head, head, PF_ANY) CLEARPAGEFLAG(Head, head, PF_ANY)
|
||||
static __always_inline bool folio_test_head(struct folio *folio)
|
||||
{
|
||||
return test_bit(PG_head, folio_flags(folio, FOLIO_PF_ANY));
|
||||
}
|
||||
|
||||
static __always_inline int PageHead(struct page *page)
|
||||
{
|
||||
PF_POISONED_CHECK(page);
|
||||
return test_bit(PG_head, &page->flags) && !page_is_fake_head(page);
|
||||
}
|
||||
|
||||
__SETPAGEFLAG(Head, head, PF_ANY)
|
||||
__CLEARPAGEFLAG(Head, head, PF_ANY)
|
||||
CLEARPAGEFLAG(Head, head, PF_ANY)
|
||||
|
||||
/**
|
||||
* folio_test_large() - Does this folio contain more than one page?
|
||||
|
|
|
@ -124,9 +124,9 @@
|
|||
* 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
|
||||
* Therefore, we can remap pages 1 to 7 to page 0. Only 1 page of page structs
|
||||
* will be used for each HugeTLB page. This will allow us to free the remaining
|
||||
* 6 pages to the buddy allocator.
|
||||
* 7 pages to the buddy allocator.
|
||||
*
|
||||
* Here is how things look after remapping.
|
||||
*
|
||||
|
@ -134,30 +134,30 @@
|
|||
* +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+
|
||||
* | | | 0 | -------------> | 0 |
|
||||
* | | +-----------+ +-----------+
|
||||
* | | | 1 | -------------> | 1 |
|
||||
* | | +-----------+ +-----------+
|
||||
* | | | 2 | ----------------^ ^ ^ ^ ^ ^
|
||||
* | | +-----------+ | | | | |
|
||||
* | | | 3 | ------------------+ | | | |
|
||||
* | | +-----------+ | | | |
|
||||
* | | | 4 | --------------------+ | | |
|
||||
* | PMD | +-----------+ | | |
|
||||
* | level | | 5 | ----------------------+ | |
|
||||
* | mapping | +-----------+ | |
|
||||
* | | | 6 | ------------------------+ |
|
||||
* | | +-----------+ |
|
||||
* | | | 7 | --------------------------+
|
||||
* | | | 1 | ---------------^ ^ ^ ^ ^ ^ ^
|
||||
* | | +-----------+ | | | | | |
|
||||
* | | | 2 | -----------------+ | | | | |
|
||||
* | | +-----------+ | | | | |
|
||||
* | | | 3 | -------------------+ | | | |
|
||||
* | | +-----------+ | | | |
|
||||
* | | | 4 | ---------------------+ | | |
|
||||
* | PMD | +-----------+ | | |
|
||||
* | level | | 5 | -----------------------+ | |
|
||||
* | mapping | +-----------+ | |
|
||||
* | | | 6 | -------------------------+ |
|
||||
* | | +-----------+ |
|
||||
* | | | 7 | ---------------------------+
|
||||
* | | +-----------+
|
||||
* | |
|
||||
* | |
|
||||
* | |
|
||||
* +-----------+
|
||||
*
|
||||
* When a HugeTLB is freed to the buddy system, we should allocate 6 pages for
|
||||
* 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 - 2) vmemmap pages.
|
||||
* 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
|
||||
|
@ -166,7 +166,13 @@
|
|||
*
|
||||
* 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 2 pages.
|
||||
* 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 (more details refer to the comment above compound_head()).
|
||||
*/
|
||||
#define pr_fmt(fmt) "HugeTLB: " fmt
|
||||
|
||||
|
@ -175,19 +181,21 @@
|
|||
/*
|
||||
* There are a lot of struct page structures associated with each HugeTLB page.
|
||||
* For tail pages, the value of compound_head is the same. So we can reuse first
|
||||
* page of tail page structures. We map the virtual addresses of the remaining
|
||||
* pages of tail page structures to the first tail page struct, and then free
|
||||
* these page frames. Therefore, we need to reserve two pages as vmemmap areas.
|
||||
* page of head page structures. We map the virtual addresses of all the pages
|
||||
* of tail page structures to the head page struct, and then free these page
|
||||
* frames. Therefore, we need to reserve one pages as vmemmap areas.
|
||||
*/
|
||||
#define RESERVE_VMEMMAP_NR 2U
|
||||
#define RESERVE_VMEMMAP_NR 1U
|
||||
#define RESERVE_VMEMMAP_SIZE (RESERVE_VMEMMAP_NR << PAGE_SHIFT)
|
||||
|
||||
bool hugetlb_free_vmemmap_enabled = IS_ENABLED(CONFIG_HUGETLB_PAGE_FREE_VMEMMAP_DEFAULT_ON);
|
||||
bool hugetlb_free_vmemmap_enabled __read_mostly =
|
||||
IS_ENABLED(CONFIG_HUGETLB_PAGE_FREE_VMEMMAP_DEFAULT_ON);
|
||||
EXPORT_SYMBOL(hugetlb_free_vmemmap_enabled);
|
||||
|
||||
static int __init early_hugetlb_free_vmemmap_param(char *buf)
|
||||
{
|
||||
/* We cannot optimize if a "struct page" crosses page boundaries. */
|
||||
if ((!is_power_of_2(sizeof(struct page)))) {
|
||||
if (!is_power_of_2(sizeof(struct page))) {
|
||||
pr_warn("cannot free vmemmap pages because \"struct page\" crosses page boundaries\n");
|
||||
return 0;
|
||||
}
|
||||
|
@ -236,7 +244,6 @@ int alloc_huge_page_vmemmap(struct hstate *h, struct page *head)
|
|||
*/
|
||||
ret = vmemmap_remap_alloc(vmemmap_addr, vmemmap_end, vmemmap_reuse,
|
||||
GFP_KERNEL | __GFP_NORETRY | __GFP_THISNODE);
|
||||
|
||||
if (!ret)
|
||||
ClearHPageVmemmapOptimized(head);
|
||||
|
||||
|
@ -282,9 +289,8 @@ void __init hugetlb_vmemmap_init(struct hstate *h)
|
|||
|
||||
vmemmap_pages = (nr_pages * sizeof(struct page)) >> PAGE_SHIFT;
|
||||
/*
|
||||
* The head page and the first tail page are not to be freed to buddy
|
||||
* allocator, the other pages will map to the first tail page, so they
|
||||
* can be freed.
|
||||
* The head page is not to be freed to buddy allocator, the other tail
|
||||
* pages will map to the head page, so they can be freed.
|
||||
*
|
||||
* Could RESERVE_VMEMMAP_NR be greater than @vmemmap_pages? It is true
|
||||
* on some architectures (e.g. aarch64). See Documentation/arm64/
|
||||
|
|
|
@ -245,6 +245,26 @@ static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
|
|||
set_pte_at(&init_mm, addr, pte, entry);
|
||||
}
|
||||
|
||||
/*
|
||||
* How many struct page structs need to be reset. When we reuse the head
|
||||
* struct page, the special metadata (e.g. page->flags or page->mapping)
|
||||
* cannot copy to the tail struct page structs. The invalid value will be
|
||||
* checked in the free_tail_pages_check(). In order to avoid the message
|
||||
* of "corrupted mapping in tail page". We need to reset at least 3 (one
|
||||
* head struct page struct and two tail struct page structs) struct page
|
||||
* structs.
|
||||
*/
|
||||
#define NR_RESET_STRUCT_PAGE 3
|
||||
|
||||
static inline void reset_struct_pages(struct page *start)
|
||||
{
|
||||
int i;
|
||||
struct page *from = start + NR_RESET_STRUCT_PAGE;
|
||||
|
||||
for (i = 0; i < NR_RESET_STRUCT_PAGE; i++)
|
||||
memcpy(start + i, from, sizeof(*from));
|
||||
}
|
||||
|
||||
static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
|
||||
struct vmemmap_remap_walk *walk)
|
||||
{
|
||||
|
@ -258,6 +278,7 @@ static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
|
|||
list_del(&page->lru);
|
||||
to = page_to_virt(page);
|
||||
copy_page(to, (void *)walk->reuse_addr);
|
||||
reset_struct_pages(to);
|
||||
|
||||
set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
|
||||
}
|
||||
|
|
Loading…
Reference in a new issue