From: Muchun Song <hidden> Date: 2021-02-08 08:54:59
Hi all,
This patch series will free some vmemmap pages(struct page structures)
associated with each hugetlbpage 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 accutualy eliminate 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 a 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 hugetlbpage. 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[0], but with that in added it shows even more:
unpin_user_pages(): ~27k -> ~3.8k
[0] 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.
Changelog in v14 -> v15:
- Fix some issues suggested by Oscar. Thanks to Oscar.
- Add numbers which Joao Martins tested to cover letter. Thanks to his effort.
Changelog in v13 -> v14:
- Refuse to free the HugeTLB page when the system is under memory pressure.
- Use GFP_ATOMIC to allocate vmemmap pages instead of GFP_KERNEL.
- Rebase to linux-next 20210202.
- Fix and add some comments for vmemmap_remap_free().
Thanks to Oscar, Mike, David H and David R's suggestions and review.
Changelog in v12 -> v13:
- Remove VM_WARN_ON_PAGE macro.
- Add more comments in vmemmap_pte_range() and vmemmap_remap_free().
Thanks to Oscar and Mike's suggestions and review.
Changelog in v11 -> v12:
- Move VM_WARN_ON_PAGE to a separate patch.
- Call __free_hugepage() with hugetlb_lock (See patch #5.) to serialize
with dissolve_free_huge_page(). It is to prepare for patch #9.
- Introduce PageHugeInflight. See patch #9.
Changelog in v10 -> v11:
- Fix compiler error when !CONFIG_HUGETLB_PAGE_FREE_VMEMMAP.
- Rework some comments and commit changes.
- Rework vmemmap_remap_free() to 3 parameters.
Thanks to Oscar and Mike's suggestions and review.
Changelog in v9 -> v10:
- Fix a bug in patch #11. Thanks to Oscar for pointing that out.
- Rework some commit log or comments. Thanks Mike and Oscar for the suggestions.
- Drop VMEMMAP_TAIL_PAGE_REUSE in the patch #3.
Thank you very much Mike and Oscar for reviewing the code.
Changelog in v8 -> v9:
- Rework some code. Very thanks to Oscar.
- Put all the non-hugetlb vmemmap functions under sparsemem-vmemmap.c.
Changelog in v7 -> v8:
- Adjust the order of patches.
Very thanks to David and Oscar. Your suggestions are very valuable.
Changelog in v6 -> v7:
- Rebase to linux-next 20201130
- Do not use basepage mapping for vmemmap when this feature is disabled.
- Rework some patchs.
[PATCH v6 08/16] mm/hugetlb: Free the vmemmap pages associated with each hugetlb page
[PATCH v6 10/16] mm/hugetlb: Allocate the vmemmap pages associated with each hugetlb page
Thanks to Oscar and Barry.
Changelog in v5 -> v6:
- Disable PMD/huge page mapping of vmemmap if this feature was enabled.
- Simplify the first version code.
Changelog in v4 -> v5:
- Rework somme comments and code in the [PATCH v4 04/21] and [PATCH v4 05/21].
Thanks to Mike and Oscar's suggestions.
Changelog in v3 -> v4:
- Move all the vmemmap functions to hugetlb_vmemmap.c.
- Make the CONFIG_HUGETLB_PAGE_FREE_VMEMMAP default to y, if we want to
disable this feature, we should disable it by a boot/kernel command line.
- Remove vmemmap_pgtable_{init, deposit, withdraw}() helper functions.
- Initialize page table lock for vmemmap through core_initcall mechanism.
Thanks for Mike and Oscar's suggestions.
Changelog in v2 -> v3:
- Rename some helps function name. Thanks Mike.
- Rework some code. Thanks Mike and Oscar.
- Remap the tail vmemmap page with PAGE_KERNEL_RO instead of PAGE_KERNEL.
Thanks Matthew.
- Add some overhead analysis in the cover letter.
- Use vmemap pmd table lock instead of a hugetlb specific global lock.
Changelog in v1 -> v2:
- Fix do not call dissolve_compound_page in alloc_huge_page_vmemmap().
- Fix some typo and code style problems.
- Remove unused handle_vmemmap_fault().
- Merge some commits to one commit suggested by Mike.
Muchun Song (8):
mm: memory_hotplug: factor out bootmem core functions to
bootmem_info.c
mm: hugetlb: introduce a new config HUGETLB_PAGE_FREE_VMEMMAP
mm: hugetlb: free the vmemmap pages associated with each HugeTLB page
mm: hugetlb: alloc the vmemmap pages associated with each HugeTLB page
mm: hugetlb: add a kernel parameter hugetlb_free_vmemmap
mm: hugetlb: introduce nr_free_vmemmap_pages in the struct hstate
mm: hugetlb: gather discrete indexes of tail page
mm: hugetlb: optimize the code with the help of the compiler
Documentation/admin-guide/kernel-parameters.txt | 14 ++
Documentation/admin-guide/mm/hugetlbpage.rst | 3 +
arch/x86/mm/init_64.c | 13 +-
fs/Kconfig | 6 +
include/linux/bootmem_info.h | 65 +++++
include/linux/hugetlb.h | 43 +++-
include/linux/hugetlb_cgroup.h | 19 +-
include/linux/memory_hotplug.h | 27 --
include/linux/mm.h | 5 +
mm/Makefile | 2 +
mm/bootmem_info.c | 124 ++++++++++
mm/hugetlb.c | 23 +-
mm/hugetlb_vmemmap.c | 314 ++++++++++++++++++++++++
mm/hugetlb_vmemmap.h | 31 +++
mm/memory_hotplug.c | 116 ---------
mm/sparse-vmemmap.c | 280 +++++++++++++++++++++
mm/sparse.c | 1 +
17 files changed, 928 insertions(+), 158 deletions(-)
create mode 100644 include/linux/bootmem_info.h
create mode 100644 mm/bootmem_info.c
create mode 100644 mm/hugetlb_vmemmap.c
create mode 100644 mm/hugetlb_vmemmap.h
--
2.11.0
From: Muchun Song <hidden> Date: 2021-02-08 08:56:03
Move bootmem info registration common API to individual bootmem_info.c.
And we will use {get,put}_page_bootmem() to initialize the page for the
vmemmap pages or free the vmemmap pages to buddy in the later patch.
So move them out of CONFIG_MEMORY_HOTPLUG_SPARSE. This is just code
movement without any functional change.
Signed-off-by: Muchun Song <redacted>
Acked-by: Mike Kravetz <redacted>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <redacted>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
---
arch/x86/mm/init_64.c | 3 +-
include/linux/bootmem_info.h | 40 +++++++++++++
include/linux/memory_hotplug.h | 27 ---------
mm/Makefile | 1 +
mm/bootmem_info.c | 124 +++++++++++++++++++++++++++++++++++++++++
mm/memory_hotplug.c | 116 --------------------------------------
mm/sparse.c | 1 +
7 files changed, 168 insertions(+), 144 deletions(-)
create mode 100644 include/linux/bootmem_info.h
create mode 100644 mm/bootmem_info.c
@@ -1571,7 +1572,7 @@ int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,returnerr;}-#if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)+#ifdef CONFIG_HAVE_BOOTMEM_INFO_NODEvoidregister_page_bootmem_memmap(unsignedlongsection_nr,structpage*start_page,unsignedlongnr_pages){
@@ -18,18 +18,6 @@ struct vmem_altmap;#ifdef CONFIG_MEMORY_HOTPLUGstructpage*pfn_to_online_page(unsignedlongpfn);-/*-*Typesforfreebootmemstoredinpage->lru.next.Thesehavetobein-*somerandomrangeinunsignedlongspacefordebuggingpurposes.-*/-enum{-MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE=12,-SECTION_INFO=MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE,-MIX_SECTION_INFO,-NODE_INFO,-MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE=NODE_INFO,-};-/* Types for control the zone type of onlined and offlined memory */enum{/* Offline the memory. */
From: Muchun Song <hidden> Date: 2021-02-08 08:56:44
The option HUGETLB_PAGE_FREE_VMEMMAP allows for the freeing of
some vmemmap pages associated with pre-allocated HugeTLB pages.
For example, on X86_64 6 vmemmap pages of size 4KB each can be
saved for each 2MB HugeTLB page. 4094 vmemmap pages of size 4KB
each can be saved for each 1GB HugeTLB page.
When a HugeTLB page is allocated or freed, the vmemmap array
representing the range associated with the page will need to be
remapped. When a page is allocated, vmemmap pages are freed
after remapping. When a page is freed, previously discarded
vmemmap pages must be allocated before remapping.
The config option is introduced early so that supporting code
can be written to depend on the option. The initial version of
the code only provides support for x86-64.
Like other code which frees vmemmap, this config option depends on
HAVE_BOOTMEM_INFO_NODE. The routine register_page_bootmem_info() is
used to register bootmem info. Therefore, make sure
register_page_bootmem_info is enabled if HUGETLB_PAGE_FREE_VMEMMAP
is defined.
Signed-off-by: Muchun Song <redacted>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Acked-by: Mike Kravetz <redacted>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
---
arch/x86/mm/init_64.c | 2 +-
fs/Kconfig | 6 ++++++
2 files changed, 7 insertions(+), 1 deletion(-)
From: Muchun Song <hidden> Date: 2021-02-08 08:57:00
Every HugeTLB has more than one struct page structure. We __know__ that
we only use the first 4(HUGETLB_CGROUP_MIN_ORDER) struct page structures
to store metadata associated with each HugeTLB.
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.
When we allocate a HugeTLB page from the buddy, we can free some vmemmap
pages associated with each HugeTLB page. It is more appropriate to do it
in the prep_new_huge_page().
The free_vmemmap_pages_per_hpage(), which indicates how many vmemmap
pages associated with a HugeTLB page can be freed, returns zero for
now, which means the feature is disabled. We will enable it once all
the infrastructure is there.
Signed-off-by: Muchun Song <redacted>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
---
include/linux/bootmem_info.h | 27 +++++-
include/linux/mm.h | 3 +
mm/Makefile | 1 +
mm/hugetlb.c | 3 +
mm/hugetlb_vmemmap.c | 219 +++++++++++++++++++++++++++++++++++++++++++
mm/hugetlb_vmemmap.h | 20 ++++
mm/sparse-vmemmap.c | 207 ++++++++++++++++++++++++++++++++++++++++
7 files changed, 479 insertions(+), 1 deletion(-)
create mode 100644 mm/hugetlb_vmemmap.c
create mode 100644 mm/hugetlb_vmemmap.h
@@ -27,8 +27,215 @@#include<linux/spinlock.h>#include<linux/vmalloc.h>#include<linux/sched.h>+#include<linux/pgtable.h>+#include<linux/bootmem_info.h>+#include<asm/dma.h>#include<asm/pgalloc.h>+#include<asm/tlbflush.h>++/**+*vmemmap_remap_walk-walkvmemmappagetable+*+*@remap_pte:calledforeachlowest-levelentry(PTE).+*@reuse_page:thepagewhichisreusedforthetailvmemmappages.+*@reuse_addr:thevirtualaddressofthe@reuse_pagepage.+*@vmemmap_pages:thelistheadofthevmemmappagesthatcanbefreed.+*/+structvmemmap_remap_walk{+void(*remap_pte)(pte_t*pte,unsignedlongaddr,+structvmemmap_remap_walk*walk);+structpage*reuse_page;+unsignedlongreuse_addr;+structlist_head*vmemmap_pages;+};++staticvoidvmemmap_pte_range(pmd_t*pmd,unsignedlongaddr,+unsignedlongend,+structvmemmap_remap_walk*walk)+{+pte_t*pte;++pte=pte_offset_kernel(pmd,addr);++/*+*Thereuse_pageisfound'first'intablewalkbeforewestart+*remapping(whichiscalling@walk->remap_pte).+*/+if(!walk->reuse_page){+BUG_ON(pte_none(*pte));+BUG_ON(walk->reuse_addr!=addr);++walk->reuse_page=pte_page(*pte++);+/*+*Becausethereuseaddressispartoftherangethatweare+*walking,skipthereuseaddressrange.+*/+addr+=PAGE_SIZE;+}++for(;addr!=end;addr+=PAGE_SIZE,pte++){+BUG_ON(pte_none(*pte));++walk->remap_pte(pte,addr,walk);+}+}++staticvoidvmemmap_pmd_range(pud_t*pud,unsignedlongaddr,+unsignedlongend,+structvmemmap_remap_walk*walk)+{+pmd_t*pmd;+unsignedlongnext;++pmd=pmd_offset(pud,addr);+do{+BUG_ON(pmd_none(*pmd)||pmd_leaf(*pmd));++next=pmd_addr_end(addr,end);+vmemmap_pte_range(pmd,addr,next,walk);+}while(pmd++,addr=next,addr!=end);+}++staticvoidvmemmap_pud_range(p4d_t*p4d,unsignedlongaddr,+unsignedlongend,+structvmemmap_remap_walk*walk)+{+pud_t*pud;+unsignedlongnext;++pud=pud_offset(p4d,addr);+do{+BUG_ON(pud_none(*pud));++next=pud_addr_end(addr,end);+vmemmap_pmd_range(pud,addr,next,walk);+}while(pud++,addr=next,addr!=end);+}++staticvoidvmemmap_p4d_range(pgd_t*pgd,unsignedlongaddr,+unsignedlongend,+structvmemmap_remap_walk*walk)+{+p4d_t*p4d;+unsignedlongnext;++p4d=p4d_offset(pgd,addr);+do{+BUG_ON(p4d_none(*p4d));++next=p4d_addr_end(addr,end);+vmemmap_pud_range(p4d,addr,next,walk);+}while(p4d++,addr=next,addr!=end);+}++staticvoidvmemmap_remap_range(unsignedlongstart,unsignedlongend,+structvmemmap_remap_walk*walk)+{+unsignedlongaddr=start;+unsignedlongnext;+pgd_t*pgd;++VM_BUG_ON(!IS_ALIGNED(start,PAGE_SIZE));+VM_BUG_ON(!IS_ALIGNED(end,PAGE_SIZE));++pgd=pgd_offset_k(addr);+do{+BUG_ON(pgd_none(*pgd));++next=pgd_addr_end(addr,end);+vmemmap_p4d_range(pgd,addr,next,walk);+}while(pgd++,addr=next,addr!=end);++/*+*Weonlychangethemappingofthevmemmapvirtualaddressrange+*[@start+PAGE_SIZE,end),soweonlyneedtoflushtheTLBwhich+*belongstotherange.+*/+flush_tlb_kernel_range(start+PAGE_SIZE,end);+}++/*+*Freeavmemmappage.Avmemmappagecanbeallocatedfromthememblock+*allocatororbuddyallocator.IfthePG_reservedflagisset,itmeans+*thatitallocatedfromthememblockallocator,justfreeitviathe+*free_bootmem_page().Otherwise,use__free_page().+*/+staticinlinevoidfree_vmemmap_page(structpage*page)+{+if(PageReserved(page))+free_bootmem_page(page);+else+__free_page(page);+}++/* Free a list of the vmemmap pages */+staticvoidfree_vmemmap_page_list(structlist_head*list)+{+structpage*page,*next;++list_for_each_entry_safe(page,next,list,lru){+list_del(&page->lru);+free_vmemmap_page(page);+}+}++staticvoidvmemmap_remap_pte(pte_t*pte,unsignedlongaddr,+structvmemmap_remap_walk*walk)+{+/*+*Remapthetailpagesasread-onlytocatchillegalwriteoperation+*tothetailpages.+*/+pgprot_tpgprot=PAGE_KERNEL_RO;+pte_tentry=mk_pte(walk->reuse_page,pgprot);+structpage*page=pte_page(*pte);++list_add(&page->lru,walk->vmemmap_pages);+set_pte_at(&init_mm,addr,pte,entry);+}++/**+*vmemmap_remap_free-remapthevmemmapvirtualaddressrange[@start,@end)+*tothepagewhich@reuseismappedto,thenfreevmemmap+*whichtherangearemappedto.+*@start:startaddressofthevmemmapvirtualaddressrangethatwewant+*toremap.+*@end:endaddressofthevmemmapvirtualaddressrangethatwewantto+*remap.+*@reuse:reuseaddress.+*+*Note:Thisfunctiondependsonvmemmapbeingbasepagemapped.Pleasemake+*surethatwedisablePMDmappingofvmemmappageswhencallingthisfunction.+*/+voidvmemmap_remap_free(unsignedlongstart,unsignedlongend,+unsignedlongreuse)+{+LIST_HEAD(vmemmap_pages);+structvmemmap_remap_walkwalk={+.remap_pte=vmemmap_remap_pte,+.reuse_addr=reuse,+.vmemmap_pages=&vmemmap_pages,+};++/*+*Inordertomakeremappingroutinemostefficientforthehugepages,+*theroutineofvmemmappagetablewalkinghasthefollowingrules+*(seemoredetailsfromthevmemmap_pte_range()):+*+*-Therange[@start,@end)andtherange[@reuse,@reuse+PAGE_SIZE)+*shouldbecontinuous.+*-The@reuseaddressispartoftherange[@reuse,@end)thatweare+*walkingwhichispassedtovmemmap_remap_range().+*-The@reuseaddressisthefirstinthecompleterange.+*+*Soweneedtomakesurethat@startand@reusemeettheaboverules.+*/+BUG_ON(start-reuse!=PAGE_SIZE);++vmemmap_remap_range(reuse,end,&walk);+free_vmemmap_page_list(&vmemmap_pages);+}/**Allocateablockofmemorytobeusedtobackthevirtualmemorymap
From: Muchun Song <hidden> Date: 2021-02-08 08:58:36
All the infrastructure is ready, so we introduce nr_free_vmemmap_pages
field in the hstate to indicate how many vmemmap pages associated with
a HugeTLB page that can be freed to buddy allocator. And initialize it
in the hugetlb_vmemmap_init(). This patch is actual enablement of the
feature.
Signed-off-by: Muchun Song <redacted>
Acked-by: Mike Kravetz <redacted>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
---
include/linux/hugetlb.h | 3 +++
mm/hugetlb.c | 1 +
mm/hugetlb_vmemmap.c | 30 ++++++++++++++++++++++++++----
mm/hugetlb_vmemmap.h | 5 +++++
4 files changed, 35 insertions(+), 4 deletions(-)
From: Muchun Song <hidden> Date: 2021-02-08 08:58:37
When the "struct page size" crosses page boundaries we cannot
make use of this feature. Let free_vmemmap_pages_per_hpage()
return zero if that is the case, most of the functions can be
optimized away.
Signed-off-by: Muchun Song <redacted>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
---
include/linux/hugetlb.h | 3 ++-
mm/hugetlb_vmemmap.c | 13 +++++++++++++
2 files changed, 15 insertions(+), 1 deletion(-)
From: Muchun Song <hidden> Date: 2021-02-08 08:58:37
For HugeTLB page, there are more metadata to save in the struct page.
But the head struct page cannot meet our needs, so we have to abuse
other tail struct page to store the metadata. In order to avoid
conflicts caused by subsequent use of more tail struct pages, we can
gather these discrete indexes of tail struct page. In this case, it
will be easier to add a new tail page index later.
There are only (RESERVE_VMEMMAP_SIZE / sizeof(struct page)) struct
page structs that can be used when CONFIG_HUGETLB_PAGE_FREE_VMEMMAP,
so add a BUILD_BUG_ON to catch invalid usage of the tail struct page.
Signed-off-by: Muchun Song <redacted>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
---
include/linux/hugetlb.h | 20 ++++++++++++++++++--
include/linux/hugetlb_cgroup.h | 19 +++++++++++--------
mm/hugetlb_vmemmap.c | 8 ++++++++
3 files changed, 37 insertions(+), 10 deletions(-)
From: Muchun Song <hidden> Date: 2021-02-08 08:59:08
Add a kernel parameter hugetlb_free_vmemmap to enable the feature of
freeing unused vmemmap pages associated with each hugetlb page on boot.
We disables PMD mapping of vmemmap pages for x86-64 arch when this
feature is enabled. Because vmemmap_remap_free() depends on vmemmap
being base page mapped.
Signed-off-by: Muchun Song <redacted>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Barry Song <redacted>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
---
Documentation/admin-guide/kernel-parameters.txt | 14 ++++++++++++++
Documentation/admin-guide/mm/hugetlbpage.rst | 3 +++
arch/x86/mm/init_64.c | 8 ++++++--
include/linux/hugetlb.h | 19 +++++++++++++++++++
mm/hugetlb_vmemmap.c | 22 ++++++++++++++++++++++
5 files changed, 64 insertions(+), 2 deletions(-)
@@ -1577,6 +1577,20 @@ Documentation/admin-guide/mm/hugetlbpage.rst. Format: size[KMG]+ hugetlb_free_vmemmap=+ [KNL] When CONFIG_HUGETLB_PAGE_FREE_VMEMMAP is set,+ this controls freeing unused vmemmap pages associated+ with each HugeTLB page. When this option is enabled,+ we disable PMD/huge page mapping of vmemmap pages which+ increase page table pages. So if a user/sysadmin only+ uses a small number of HugeTLB pages (as a percentage+ of system memory), they could end up using more memory+ with hugetlb_free_vmemmap on as opposed to off.+ Format: { on | off (default) }++ on: enable the feature+ off: disable the feature+ hung_task_panic= [KNL] Should the hung task detector generate panics. Format: 0 | 1
@@ -145,6 +145,9 @@ default_hugepagesz will all result in 256 2M huge pages being allocated. Valid default huge page size is architecture dependent.+hugetlb_free_vmemmap+ When CONFIG_HUGETLB_PAGE_FREE_VMEMMAP is set, this enables freeing+ unused vmemmap pages associated with each HugeTLB page. When multiple huge page sizes are supported, ``/proc/sys/vm/nr_hugepages`` indicates the current number of pre-allocated huge pages of the default size.
@@ -1557,7 +1558,8 @@ int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,{interr;-if(end-start<PAGES_PER_SECTION*sizeof(structpage))+if((is_hugetlb_free_vmemmap_enabled()&&!altmap)||+end-start<PAGES_PER_SECTION*sizeof(structpage))err=vmemmap_populate_basepages(start,end,node,NULL);elseif(boot_cpu_has(X86_FEATURE_PSE))err=vmemmap_populate_hugepages(start,end,node,altmap);
@@ -1585,6 +1587,8 @@ void register_page_bootmem_memmap(unsigned long section_nr,pmd_t*pmd;unsignedintnr_pmd_pages;structpage*page;+boolbase_mapping=!boot_cpu_has(X86_FEATURE_PSE)||+is_hugetlb_free_vmemmap_enabled();for(;addr<end;addr=next){pte_t*pte=NULL;
@@ -1610,7 +1614,7 @@ void register_page_bootmem_memmap(unsigned long section_nr,}get_page_bootmem(section_nr,pud_page(*pud),MIX_SECTION_INFO);-if(!boot_cpu_has(X86_FEATURE_PSE)){+if(base_mapping){next=(addr+PAGE_SIZE)&PAGE_MASK;pmd=pmd_offset(pud,addr);if(pmd_none(*pmd))
From: Muchun Song <hidden> Date: 2021-02-08 08:59:15
When we free a HugeTLB page to the buddy allocator, we should allocate the
vmemmap pages associated with it. But we may cannot allocate vmemmap pages
when the system is under memory pressure, in this case, we just refuse to
free the HugeTLB page instead of looping forever trying to allocate the
pages.
Signed-off-by: Muchun Song <redacted>
---
include/linux/mm.h | 2 ++
mm/hugetlb.c | 19 ++++++++++++-
mm/hugetlb_vmemmap.c | 30 +++++++++++++++++++++
mm/hugetlb_vmemmap.h | 6 +++++
mm/sparse-vmemmap.c | 75 +++++++++++++++++++++++++++++++++++++++++++++++++++-
5 files changed, 130 insertions(+), 2 deletions(-)
@@ -1397,16 +1397,26 @@ static void __free_huge_page(struct page *page)h->resv_huge_pages++;if(HPageTemporary(page)){-list_del(&page->lru);ClearHPageTemporary(page);++if(alloc_huge_page_vmemmap(h,page)){+h->surplus_huge_pages++;+h->surplus_huge_pages_node[nid]++;+gotoenqueue;+}+list_del(&page->lru);update_and_free_page(h,page);}elseif(h->surplus_huge_pages_node[nid]){+if(alloc_huge_page_vmemmap(h,page))+gotoenqueue;+/* remove the page from active list */list_del(&page->lru);update_and_free_page(h,page);h->surplus_huge_pages--;h->surplus_huge_pages_node[nid]--;}else{+enqueue:arch_clear_hugepage_flags(page);enqueue_huge_page(h,page);}
@@ -237,6 +238,78 @@ void vmemmap_remap_free(unsigned long start, unsigned long end,free_vmemmap_page_list(&vmemmap_pages);}+staticvoidvmemmap_restore_pte(pte_t*pte,unsignedlongaddr,+structvmemmap_remap_walk*walk)+{+pgprot_tpgprot=PAGE_KERNEL;+structpage*page;+void*to;++BUG_ON(pte_page(*pte)!=walk->reuse_page);++page=list_first_entry(walk->vmemmap_pages,structpage,lru);+list_del(&page->lru);+to=page_to_virt(page);+copy_page(to,(void*)walk->reuse_addr);++set_pte_at(&init_mm,addr,pte,mk_pte(page,pgprot));+}++staticintalloc_vmemmap_page_list(unsignedlongstart,unsignedlongend,+gfp_tgfp_mask,structlist_head*list)+{+unsignedlongnr_pages=(end-start)>>PAGE_SHIFT;+intnid=page_to_nid((structpage*)start);+structpage*page,*next;++while(nr_pages--){+page=alloc_pages_node(nid,gfp_mask,0);+if(!page)+gotoout;+list_add_tail(&page->lru,list);+}++return0;+out:+list_for_each_entry_safe(page,next,list,lru)+__free_pages(page,0);+return-ENOMEM;+}++/**+*vmemmap_remap_alloc-remapthevmemmapvirtualaddressrange[@start,end)+*tothepagewhichisfromthe@vmemmap_pages+*respectively.+*@start:startaddressofthevmemmapvirtualaddressrangethatwewant+*toremap.+*@end:endaddressofthevmemmapvirtualaddressrangethatwewantto+*remap.+*@reuse:reuseaddress.+*@gpf_mask:GFPflagforallocatingvmemmappages.+*/+intvmemmap_remap_alloc(unsignedlongstart,unsignedlongend,+unsignedlongreuse,gfp_tgfp_mask)+{+LIST_HEAD(vmemmap_pages);+structvmemmap_remap_walkwalk={+.remap_pte=vmemmap_restore_pte,+.reuse_addr=reuse,+.vmemmap_pages=&vmemmap_pages,+};++/* See the comment in the vmemmap_remap_free(). */+BUG_ON(start-reuse!=PAGE_SIZE);++might_sleep_if(gfpflags_allow_blocking(gfp_mask));++if(alloc_vmemmap_page_list(start,end,gfp_mask,&vmemmap_pages))+return-ENOMEM;++vmemmap_remap_range(reuse,end,&walk);++return0;+}+/**Allocateablockofmemorytobeusedtobackthevirtualmemorymap*ortobackthepagetablesthatareusedtocreatethemapping.
From: Mike Kravetz <hidden> Date: 2021-02-11 18:21:12
On 2/8/21 12:50 AM, Muchun Song wrote:
When we free a HugeTLB page to the buddy allocator, we should allocate the
vmemmap pages associated with it. But we may cannot allocate vmemmap pages
when the system is under memory pressure, in this case, we just refuse to
free the HugeTLB page instead of looping forever trying to allocate the
pages.
Signed-off-by: Muchun Song <redacted>
---
include/linux/mm.h | 2 ++
mm/hugetlb.c | 19 ++++++++++++-
mm/hugetlb_vmemmap.c | 30 +++++++++++++++++++++
mm/hugetlb_vmemmap.h | 6 +++++
mm/sparse-vmemmap.c | 75 +++++++++++++++++++++++++++++++++++++++++++++++++++-
5 files changed, 130 insertions(+), 2 deletions(-)
Muchun has done a great job simplifying this patch series and addressing
issues as they are brought up. This patch addresses the issue which seems
to be the biggest stumbling block to this series. The need to allocate
vmemmap pages to dissolve a hugetlb page to the buddy allocator. The way
it is addressed in this patch is to simply fail to dissolve the hugetlb
page if the vmmemmap pages can not be allocated. IMO, this is an 'acceptable'
strategy. If we find ourselves in this situation then we are likely to be
hitting other corner cases in the system. I wish there was a perfect way
to address this issue, but we have been unable to come up with one.
There was a decent discussion about this is a previous version of the
series starting here:
https://lore.kernel.org/linux-mm/20210126092942.GA10602@linux/
In this thread various other options were suggested and discussed.
I would like to come to some agreement on an acceptable way to handle this
specific issue. IMO, it makes little sense to continue refining other
parts of this series if we can not figure out how to move forward on this
issue.
It would be great if David H, David R and Michal could share their opinions
on this. No need to review details the code yet (unless you want), but
let's start a discussion on how to move past this issue if we can.
--
Mike Kravetz
From: David Hildenbrand <hidden> Date: 2021-02-12 14:17:06
On 11.02.21 19:05, Mike Kravetz wrote:
On 2/8/21 12:50 AM, Muchun Song wrote:
quoted
When we free a HugeTLB page to the buddy allocator, we should allocate the
vmemmap pages associated with it. But we may cannot allocate vmemmap pages
when the system is under memory pressure, in this case, we just refuse to
free the HugeTLB page instead of looping forever trying to allocate the
pages.
Signed-off-by: Muchun Song <redacted>
---
include/linux/mm.h | 2 ++
mm/hugetlb.c | 19 ++++++++++++-
mm/hugetlb_vmemmap.c | 30 +++++++++++++++++++++
mm/hugetlb_vmemmap.h | 6 +++++
mm/sparse-vmemmap.c | 75 +++++++++++++++++++++++++++++++++++++++++++++++++++-
5 files changed, 130 insertions(+), 2 deletions(-)
Muchun has done a great job simplifying this patch series and addressing
issues as they are brought up. This patch addresses the issue which seems
to be the biggest stumbling block to this series. The need to allocate
vmemmap pages to dissolve a hugetlb page to the buddy allocator. The way
it is addressed in this patch is to simply fail to dissolve the hugetlb
page if the vmmemmap pages can not be allocated. IMO, this is an 'acceptable'
strategy. If we find ourselves in this situation then we are likely to be
hitting other corner cases in the system. I wish there was a perfect way
to address this issue, but we have been unable to come up with one.
There was a decent discussion about this is a previous version of the
series starting here:
https://lore.kernel.org/linux-mm/20210126092942.GA10602@linux/
In this thread various other options were suggested and discussed.
I would like to come to some agreement on an acceptable way to handle this
specific issue. IMO, it makes little sense to continue refining other
parts of this series if we can not figure out how to move forward on this
issue.
It would be great if David H, David R and Michal could share their opinions
on this. No need to review details the code yet (unless you want), but
let's start a discussion on how to move past this issue if we can.
So a summary from my side:
We might fail freeing a huge page at any point in time iff we are low on
kernel (!CMA, !ZONE_MOVABLE) memory. While we could play games with
allocating the vmemmap from a huge page itself in some cases (e.g.,
!CMA, !ZONE_MOVABLE), simply retrying is way easier and we don't turn
the huge page forever unusable.
Corner cases might be having many huge pages in ZONE_MOVABLE, freeing
them all at once and eating up a lot of kernel memory. But then, the
same setup would already be problematic nowadays where we simply always
consume that kernel memory for the vmemmap.
I think this problem only really becomes visible in corner cases. And
someone actively has to enable new behavior.
1. Failing to free a huge page triggered by the user (decrease nr_pages):
Bad luck. Try again later.
2. Failing to free a surplus huge page when freed by the application:
Bad luck. But who will try again later?
3. Failing to dissolve a free huge page on ZONE_MOVABLE via offline_pages()
This is a bit unfortunate if we have plenty of ZONE_MOVABLE memory but
are low on kernel memory. For example, migration of huge pages would
still work, however, dissolving the free page does not work. I'd say
this is a corner cases. When the system is that much under memory
pressure, offlining/unplug can be expected to fail.
4. Failing to dissolve a huge page on CMA/ZONE_MOVABLE via
alloc_contig_range() - once we have that handling in place. Mainly
affects CMA and virtio-mem.
Similar to 3. However, we didn't even take care of huge pages *at all*
for now (neither migrate nor dissolve). So actually don't make the
current state any worse. virito-mem will handle migration errors
gracefully. CMA might be able to fallback on other free areas within the
CMA region.
I'd say, document the changed behavior properly so people are aware that
there might be issues in corner cases with huge pages on CMA / ZONE_MOVABLE.
--
Thanks,
David / dhildenb
From: Michal Hocko <mhocko@suse.com> Date: 2021-02-12 15:33:45
On Mon 08-02-21 16:50:09, Muchun Song wrote:
When we free a HugeTLB page to the buddy allocator, we should allocate the
vmemmap pages associated with it. But we may cannot allocate vmemmap pages
when the system is under memory pressure, in this case, we just refuse to
free the HugeTLB page instead of looping forever trying to allocate the
pages.
Thanks for simplifying the implementation from your early proposal!
This will not be looping for ever. The allocation will usually trigger
the OOM killer and sooner or later there will be a memory to allocate
from or the system panics when there are no eligible tasks to kill. This
is just a side note.
I think the changelog could benefit from a more explicit documentation
of those error failures. There are different cases when the hugetlb page
is freed. It can be due to an admin intervention (decrease the pool),
overcommit, migration, dissolving and likely some others. Most of them
should be fine to stay in the pool which would just increase the surplus
pages in the pool. I am not so sure about dissolving path.
[...]
I do not think that this is a good allocation mode. GFP_ATOMIC is a non
sleeping allocation and a medium memory pressure might cause it to
fail prematurely. I do not think this is really an atomic context which
couldn't afford memory reclaim. I also do not think we want to grant
access to memory reserve is reasonable. Just think of a huge number of
hugetlb pages being freed which can deplete the memory reserve for
atomic allocations. I think that you want
GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN | __GFP_THISNODE
for an initial implementation. The justification being that the
allocation should at least try to reclaim but it shouldn't cause any
major disruption because the failure is not fatal. If the failure rate
would be impractically high then just drop NORETRY part. You can replace
it by __GFP_RETRY_MAYFAIL but that shouldn't be strictly necessary
because __GFP_THISNODE on its own implies on OOM killer, but that is
kinda ugly to rely on.
--
Michal Hocko
SUSE Labs
From: Muchun Song <hidden> Date: 2021-02-15 10:06:27
On Fri, Feb 12, 2021 at 11:32 PM Michal Hocko [off-list ref] wrote:
On Mon 08-02-21 16:50:09, Muchun Song wrote:
quoted
When we free a HugeTLB page to the buddy allocator, we should allocate the
vmemmap pages associated with it. But we may cannot allocate vmemmap pages
when the system is under memory pressure, in this case, we just refuse to
free the HugeTLB page instead of looping forever trying to allocate the
pages.
Thanks for simplifying the implementation from your early proposal!
This will not be looping for ever. The allocation will usually trigger
the OOM killer and sooner or later there will be a memory to allocate
from or the system panics when there are no eligible tasks to kill. This
is just a side note.
I think the changelog could benefit from a more explicit documentation
of those error failures. There are different cases when the hugetlb page
is freed. It can be due to an admin intervention (decrease the pool),
overcommit, migration, dissolving and likely some others. Most of them
should be fine to stay in the pool which would just increase the surplus
pages in the pool. I am not so sure about dissolving path.
I do not think that this is a good allocation mode. GFP_ATOMIC is a non
sleeping allocation and a medium memory pressure might cause it to
fail prematurely. I do not think this is really an atomic context which
couldn't afford memory reclaim. I also do not think we want to grant
Because alloc_huge_page_vmemmap is called under hugetlb_lock
now. So using GFP_ATOMIC indeed makes the code more simpler.
From the document of the kernel, I learned that __GFP_NOMEMALLOC
can be used to explicitly forbid access to emergency reserves. So if
we do not want to use the reserve memory. How about replacing it to
GFP_ATOMIC | __GFP_NOMEMALLOC | __GFP_NOWARN | __GFP_THISNODE
Thanks.
access to memory reserve is reasonable. Just think of a huge number of
hugetlb pages being freed which can deplete the memory reserve for
atomic allocations. I think that you want
GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN | __GFP_THISNODE
for an initial implementation. The justification being that the
allocation should at least try to reclaim but it shouldn't cause any
major disruption because the failure is not fatal. If the failure rate
would be impractically high then just drop NORETRY part. You can replace
it by __GFP_RETRY_MAYFAIL but that shouldn't be strictly necessary
because __GFP_THISNODE on its own implies on OOM killer, but that is
kinda ugly to rely on.
--
Michal Hocko
SUSE Labs
From: Michal Hocko <mhocko@suse.com> Date: 2021-02-15 10:33:58
On Mon 15-02-21 18:05:06, Muchun Song wrote:
On Fri, Feb 12, 2021 at 11:32 PM Michal Hocko [off-list ref] wrote:
[...]
quoted
quoted
+int alloc_huge_page_vmemmap(struct hstate *h, struct page *head)
+{
+ int ret;
+ unsigned long vmemmap_addr = (unsigned long)head;
+ unsigned long vmemmap_end, vmemmap_reuse;
+
+ if (!free_vmemmap_pages_per_hpage(h))
+ return 0;
+
+ vmemmap_addr += RESERVE_VMEMMAP_SIZE;
+ vmemmap_end = vmemmap_addr + free_vmemmap_pages_size_per_hpage(h);
+ vmemmap_reuse = vmemmap_addr - PAGE_SIZE;
+
+ /*
+ * The pages which the vmemmap virtual address range [@vmemmap_addr,
+ * @vmemmap_end) are mapped to are freed to the buddy allocator, and
+ * the range is mapped to the page which @vmemmap_reuse is mapped to.
+ * When a HugeTLB page is freed to the buddy allocator, previously
+ * discarded vmemmap pages must be allocated and remapping.
+ */
+ ret = vmemmap_remap_alloc(vmemmap_addr, vmemmap_end, vmemmap_reuse,
+ GFP_ATOMIC | __GFP_NOWARN | __GFP_THISNODE);
I do not think that this is a good allocation mode. GFP_ATOMIC is a non
sleeping allocation and a medium memory pressure might cause it to
fail prematurely. I do not think this is really an atomic context which
couldn't afford memory reclaim. I also do not think we want to grant
Because alloc_huge_page_vmemmap is called under hugetlb_lock
now. So using GFP_ATOMIC indeed makes the code more simpler.
You can have a preallocated list of pages prior taking the lock.
Moreover do we want to manipulate vmemmaps from under spinlock in
general. I have to say I have missed that detail when reviewing. Need to
think more.
From the document of the kernel, I learned that __GFP_NOMEMALLOC
can be used to explicitly forbid access to emergency reserves. So if
we do not want to use the reserve memory. How about replacing it to
GFP_ATOMIC | __GFP_NOMEMALLOC | __GFP_NOWARN | __GFP_THISNODE
The whole point of GFP_ATOMIC is to grant access to memory reserves so
the above is quite dubious. If you do not want access to memory reserves
then use GFP_NOWAIT instead. But failures are much more easier to happen
then.
NOMEMALLOC is meant to be used from paths which are allowed to consume
memory reserves - e.g. when invoked from the memory reclaim path.
--
Michal Hocko
SUSE Labs
From: Muchun Song <hidden> Date: 2021-02-15 11:53:03
On Mon, Feb 15, 2021 at 6:33 PM Michal Hocko [off-list ref] wrote:
On Mon 15-02-21 18:05:06, Muchun Song wrote:
quoted
On Fri, Feb 12, 2021 at 11:32 PM Michal Hocko [off-list ref] wrote:
[...]
quoted
quoted
quoted
+int alloc_huge_page_vmemmap(struct hstate *h, struct page *head)
+{
+ int ret;
+ unsigned long vmemmap_addr = (unsigned long)head;
+ unsigned long vmemmap_end, vmemmap_reuse;
+
+ if (!free_vmemmap_pages_per_hpage(h))
+ return 0;
+
+ vmemmap_addr += RESERVE_VMEMMAP_SIZE;
+ vmemmap_end = vmemmap_addr + free_vmemmap_pages_size_per_hpage(h);
+ vmemmap_reuse = vmemmap_addr - PAGE_SIZE;
+
+ /*
+ * The pages which the vmemmap virtual address range [@vmemmap_addr,
+ * @vmemmap_end) are mapped to are freed to the buddy allocator, and
+ * the range is mapped to the page which @vmemmap_reuse is mapped to.
+ * When a HugeTLB page is freed to the buddy allocator, previously
+ * discarded vmemmap pages must be allocated and remapping.
+ */
+ ret = vmemmap_remap_alloc(vmemmap_addr, vmemmap_end, vmemmap_reuse,
+ GFP_ATOMIC | __GFP_NOWARN | __GFP_THISNODE);
I do not think that this is a good allocation mode. GFP_ATOMIC is a non
sleeping allocation and a medium memory pressure might cause it to
fail prematurely. I do not think this is really an atomic context which
couldn't afford memory reclaim. I also do not think we want to grant
Because alloc_huge_page_vmemmap is called under hugetlb_lock
now. So using GFP_ATOMIC indeed makes the code more simpler.
You can have a preallocated list of pages prior taking the lock.
Moreover do we want to manipulate vmemmaps from under spinlock in
general. I have to say I have missed that detail when reviewing. Need to
think more.
quoted
From the document of the kernel, I learned that __GFP_NOMEMALLOC
can be used to explicitly forbid access to emergency reserves. So if
we do not want to use the reserve memory. How about replacing it to
GFP_ATOMIC | __GFP_NOMEMALLOC | __GFP_NOWARN | __GFP_THISNODE
The whole point of GFP_ATOMIC is to grant access to memory reserves so
the above is quite dubious. If you do not want access to memory reserves
Look at the code of gfp_to_alloc_flags().
static inline unsigned int gfp_to_alloc_flags(gfp_t gfp_mask)
{
[...]
if (gfp_mask & __GFP_ATOMIC) {
/*
* Not worth trying to allocate harder for __GFP_NOMEMALLOC even
* if it can't schedule.
*/
if (!(gfp_mask & __GFP_NOMEMALLOC))
alloc_flags |= ALLOC_HARDER;
[...]
}
Seems to allow this operation (GFP_ATOMIC | __GFP_NOMEMALLOC).
then use GFP_NOWAIT instead. But failures are much more easier to happen
then.
NOMEMALLOC is meant to be used from paths which are allowed to consume
memory reserves - e.g. when invoked from the memory reclaim path.
--
Michal Hocko
SUSE Labs
From: Muchun Song <hidden> Date: 2021-02-15 12:02:17
On Mon, Feb 15, 2021 at 7:51 PM Muchun Song [off-list ref] wrote:
On Mon, Feb 15, 2021 at 6:33 PM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 18:05:06, Muchun Song wrote:
quoted
On Fri, Feb 12, 2021 at 11:32 PM Michal Hocko [off-list ref] wrote:
[...]
quoted
quoted
quoted
+int alloc_huge_page_vmemmap(struct hstate *h, struct page *head)
+{
+ int ret;
+ unsigned long vmemmap_addr = (unsigned long)head;
+ unsigned long vmemmap_end, vmemmap_reuse;
+
+ if (!free_vmemmap_pages_per_hpage(h))
+ return 0;
+
+ vmemmap_addr += RESERVE_VMEMMAP_SIZE;
+ vmemmap_end = vmemmap_addr + free_vmemmap_pages_size_per_hpage(h);
+ vmemmap_reuse = vmemmap_addr - PAGE_SIZE;
+
+ /*
+ * The pages which the vmemmap virtual address range [@vmemmap_addr,
+ * @vmemmap_end) are mapped to are freed to the buddy allocator, and
+ * the range is mapped to the page which @vmemmap_reuse is mapped to.
+ * When a HugeTLB page is freed to the buddy allocator, previously
+ * discarded vmemmap pages must be allocated and remapping.
+ */
+ ret = vmemmap_remap_alloc(vmemmap_addr, vmemmap_end, vmemmap_reuse,
+ GFP_ATOMIC | __GFP_NOWARN | __GFP_THISNODE);
I do not think that this is a good allocation mode. GFP_ATOMIC is a non
sleeping allocation and a medium memory pressure might cause it to
fail prematurely. I do not think this is really an atomic context which
couldn't afford memory reclaim. I also do not think we want to grant
Because alloc_huge_page_vmemmap is called under hugetlb_lock
now. So using GFP_ATOMIC indeed makes the code more simpler.
You can have a preallocated list of pages prior taking the lock.
Moreover do we want to manipulate vmemmaps from under spinlock in
general. I have to say I have missed that detail when reviewing. Need to
think more.
quoted
From the document of the kernel, I learned that __GFP_NOMEMALLOC
can be used to explicitly forbid access to emergency reserves. So if
we do not want to use the reserve memory. How about replacing it to
GFP_ATOMIC | __GFP_NOMEMALLOC | __GFP_NOWARN | __GFP_THISNODE
The whole point of GFP_ATOMIC is to grant access to memory reserves so
the above is quite dubious. If you do not want access to memory reserves
Look at the code of gfp_to_alloc_flags().
static inline unsigned int gfp_to_alloc_flags(gfp_t gfp_mask)
{
[...]
if (gfp_mask & __GFP_ATOMIC) {
/*
* Not worth trying to allocate harder for __GFP_NOMEMALLOC even
* if it can't schedule.
*/
if (!(gfp_mask & __GFP_NOMEMALLOC))
alloc_flags |= ALLOC_HARDER;
[...]
}
Seems to allow this operation (GFP_ATOMIC | __GFP_NOMEMALLOC).
I also found similar users.
netdev_alloc_frag()
page_frag_alloc(GFP_ATOMIC)
__page_frag_cache_refill(GFP_ATOMIC)
alloc_pages_node(GFP_ATOMIC | __GFP_COMP | __GFP_NOWARN |
__GFP_NORETRY |
__GFP_NOMEMALLOC)
quoted
then use GFP_NOWAIT instead. But failures are much more easier to happen
then.
NOMEMALLOC is meant to be used from paths which are allowed to consume
memory reserves - e.g. when invoked from the memory reclaim path.
--
Michal Hocko
SUSE Labs
From: Michal Hocko <mhocko@suse.com> Date: 2021-02-15 12:19:50
On Mon 15-02-21 20:00:07, Muchun Song wrote:
On Mon, Feb 15, 2021 at 7:51 PM Muchun Song [off-list ref] wrote:
quoted
On Mon, Feb 15, 2021 at 6:33 PM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 18:05:06, Muchun Song wrote:
quoted
On Fri, Feb 12, 2021 at 11:32 PM Michal Hocko [off-list ref] wrote:
[...]
quoted
quoted
quoted
+int alloc_huge_page_vmemmap(struct hstate *h, struct page *head)
+{
+ int ret;
+ unsigned long vmemmap_addr = (unsigned long)head;
+ unsigned long vmemmap_end, vmemmap_reuse;
+
+ if (!free_vmemmap_pages_per_hpage(h))
+ return 0;
+
+ vmemmap_addr += RESERVE_VMEMMAP_SIZE;
+ vmemmap_end = vmemmap_addr + free_vmemmap_pages_size_per_hpage(h);
+ vmemmap_reuse = vmemmap_addr - PAGE_SIZE;
+
+ /*
+ * The pages which the vmemmap virtual address range [@vmemmap_addr,
+ * @vmemmap_end) are mapped to are freed to the buddy allocator, and
+ * the range is mapped to the page which @vmemmap_reuse is mapped to.
+ * When a HugeTLB page is freed to the buddy allocator, previously
+ * discarded vmemmap pages must be allocated and remapping.
+ */
+ ret = vmemmap_remap_alloc(vmemmap_addr, vmemmap_end, vmemmap_reuse,
+ GFP_ATOMIC | __GFP_NOWARN | __GFP_THISNODE);
I do not think that this is a good allocation mode. GFP_ATOMIC is a non
sleeping allocation and a medium memory pressure might cause it to
fail prematurely. I do not think this is really an atomic context which
couldn't afford memory reclaim. I also do not think we want to grant
Because alloc_huge_page_vmemmap is called under hugetlb_lock
now. So using GFP_ATOMIC indeed makes the code more simpler.
You can have a preallocated list of pages prior taking the lock.
Moreover do we want to manipulate vmemmaps from under spinlock in
general. I have to say I have missed that detail when reviewing. Need to
think more.
quoted
From the document of the kernel, I learned that __GFP_NOMEMALLOC
can be used to explicitly forbid access to emergency reserves. So if
we do not want to use the reserve memory. How about replacing it to
GFP_ATOMIC | __GFP_NOMEMALLOC | __GFP_NOWARN | __GFP_THISNODE
The whole point of GFP_ATOMIC is to grant access to memory reserves so
the above is quite dubious. If you do not want access to memory reserves
Look at the code of gfp_to_alloc_flags().
static inline unsigned int gfp_to_alloc_flags(gfp_t gfp_mask)
{
[...]
if (gfp_mask & __GFP_ATOMIC) {
/*
* Not worth trying to allocate harder for __GFP_NOMEMALLOC even
* if it can't schedule.
*/
if (!(gfp_mask & __GFP_NOMEMALLOC))
alloc_flags |= ALLOC_HARDER;
[...]
}
Seems to allow this operation (GFP_ATOMIC | __GFP_NOMEMALLOC).
Please read my response again more carefully. I am not claiming that
combination is not allowed. I have said it doesn't make any sense in
this context.
--
Michal Hocko
SUSE Labs
From: Muchun Song <hidden> Date: 2021-02-15 12:46:45
On Mon, Feb 15, 2021 at 8:18 PM Michal Hocko [off-list ref] wrote:
On Mon 15-02-21 20:00:07, Muchun Song wrote:
quoted
On Mon, Feb 15, 2021 at 7:51 PM Muchun Song [off-list ref] wrote:
quoted
On Mon, Feb 15, 2021 at 6:33 PM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 18:05:06, Muchun Song wrote:
quoted
On Fri, Feb 12, 2021 at 11:32 PM Michal Hocko [off-list ref] wrote:
[...]
quoted
quoted
quoted
+int alloc_huge_page_vmemmap(struct hstate *h, struct page *head)
+{
+ int ret;
+ unsigned long vmemmap_addr = (unsigned long)head;
+ unsigned long vmemmap_end, vmemmap_reuse;
+
+ if (!free_vmemmap_pages_per_hpage(h))
+ return 0;
+
+ vmemmap_addr += RESERVE_VMEMMAP_SIZE;
+ vmemmap_end = vmemmap_addr + free_vmemmap_pages_size_per_hpage(h);
+ vmemmap_reuse = vmemmap_addr - PAGE_SIZE;
+
+ /*
+ * The pages which the vmemmap virtual address range [@vmemmap_addr,
+ * @vmemmap_end) are mapped to are freed to the buddy allocator, and
+ * the range is mapped to the page which @vmemmap_reuse is mapped to.
+ * When a HugeTLB page is freed to the buddy allocator, previously
+ * discarded vmemmap pages must be allocated and remapping.
+ */
+ ret = vmemmap_remap_alloc(vmemmap_addr, vmemmap_end, vmemmap_reuse,
+ GFP_ATOMIC | __GFP_NOWARN | __GFP_THISNODE);
I do not think that this is a good allocation mode. GFP_ATOMIC is a non
sleeping allocation and a medium memory pressure might cause it to
fail prematurely. I do not think this is really an atomic context which
couldn't afford memory reclaim. I also do not think we want to grant
Because alloc_huge_page_vmemmap is called under hugetlb_lock
now. So using GFP_ATOMIC indeed makes the code more simpler.
You can have a preallocated list of pages prior taking the lock.
Moreover do we want to manipulate vmemmaps from under spinlock in
general. I have to say I have missed that detail when reviewing. Need to
think more.
quoted
From the document of the kernel, I learned that __GFP_NOMEMALLOC
can be used to explicitly forbid access to emergency reserves. So if
we do not want to use the reserve memory. How about replacing it to
GFP_ATOMIC | __GFP_NOMEMALLOC | __GFP_NOWARN | __GFP_THISNODE
The whole point of GFP_ATOMIC is to grant access to memory reserves so
the above is quite dubious. If you do not want access to memory reserves
Look at the code of gfp_to_alloc_flags().
static inline unsigned int gfp_to_alloc_flags(gfp_t gfp_mask)
{
[...]
if (gfp_mask & __GFP_ATOMIC) {
/*
* Not worth trying to allocate harder for __GFP_NOMEMALLOC even
* if it can't schedule.
*/
if (!(gfp_mask & __GFP_NOMEMALLOC))
alloc_flags |= ALLOC_HARDER;
[...]
}
Seems to allow this operation (GFP_ATOMIC | __GFP_NOMEMALLOC).
Please read my response again more carefully. I am not claiming that
combination is not allowed. I have said it doesn't make any sense in
this context.
I see you are worried that using GFP_ATOMIC will use reverse memory
unlimited. So I think that __GFP_NOMEMALLOC may be suitable for us.
Sorry, I may not understand the point you said. What I missed?
From: Michal Hocko <mhocko@suse.com> Date: 2021-02-15 13:20:22
On Mon 15-02-21 20:44:57, Muchun Song wrote:
On Mon, Feb 15, 2021 at 8:18 PM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 20:00:07, Muchun Song wrote:
quoted
On Mon, Feb 15, 2021 at 7:51 PM Muchun Song [off-list ref] wrote:
quoted
On Mon, Feb 15, 2021 at 6:33 PM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 18:05:06, Muchun Song wrote:
quoted
On Fri, Feb 12, 2021 at 11:32 PM Michal Hocko [off-list ref] wrote:
[...]
quoted
quoted
quoted
+int alloc_huge_page_vmemmap(struct hstate *h, struct page *head)
+{
+ int ret;
+ unsigned long vmemmap_addr = (unsigned long)head;
+ unsigned long vmemmap_end, vmemmap_reuse;
+
+ if (!free_vmemmap_pages_per_hpage(h))
+ return 0;
+
+ vmemmap_addr += RESERVE_VMEMMAP_SIZE;
+ vmemmap_end = vmemmap_addr + free_vmemmap_pages_size_per_hpage(h);
+ vmemmap_reuse = vmemmap_addr - PAGE_SIZE;
+
+ /*
+ * The pages which the vmemmap virtual address range [@vmemmap_addr,
+ * @vmemmap_end) are mapped to are freed to the buddy allocator, and
+ * the range is mapped to the page which @vmemmap_reuse is mapped to.
+ * When a HugeTLB page is freed to the buddy allocator, previously
+ * discarded vmemmap pages must be allocated and remapping.
+ */
+ ret = vmemmap_remap_alloc(vmemmap_addr, vmemmap_end, vmemmap_reuse,
+ GFP_ATOMIC | __GFP_NOWARN | __GFP_THISNODE);
I do not think that this is a good allocation mode. GFP_ATOMIC is a non
sleeping allocation and a medium memory pressure might cause it to
fail prematurely. I do not think this is really an atomic context which
couldn't afford memory reclaim. I also do not think we want to grant
Because alloc_huge_page_vmemmap is called under hugetlb_lock
now. So using GFP_ATOMIC indeed makes the code more simpler.
You can have a preallocated list of pages prior taking the lock.
Moreover do we want to manipulate vmemmaps from under spinlock in
general. I have to say I have missed that detail when reviewing. Need to
think more.
quoted
From the document of the kernel, I learned that __GFP_NOMEMALLOC
can be used to explicitly forbid access to emergency reserves. So if
we do not want to use the reserve memory. How about replacing it to
GFP_ATOMIC | __GFP_NOMEMALLOC | __GFP_NOWARN | __GFP_THISNODE
The whole point of GFP_ATOMIC is to grant access to memory reserves so
the above is quite dubious. If you do not want access to memory reserves
Look at the code of gfp_to_alloc_flags().
static inline unsigned int gfp_to_alloc_flags(gfp_t gfp_mask)
{
[...]
if (gfp_mask & __GFP_ATOMIC) {
/*
* Not worth trying to allocate harder for __GFP_NOMEMALLOC even
* if it can't schedule.
*/
if (!(gfp_mask & __GFP_NOMEMALLOC))
alloc_flags |= ALLOC_HARDER;
[...]
}
Seems to allow this operation (GFP_ATOMIC | __GFP_NOMEMALLOC).
Please read my response again more carefully. I am not claiming that
combination is not allowed. I have said it doesn't make any sense in
this context.
I see you are worried that using GFP_ATOMIC will use reverse memory
unlimited. So I think that __GFP_NOMEMALLOC may be suitable for us.
Sorry, I may not understand the point you said. What I missed?
OK, let me try to explain again. GFP_ATOMIC is not only a non-sleeping
allocation request. It also grants access to memory reserves. The later
is a bit more involved because there are more layers of memory reserves
to access but that is not really important. Non-sleeping semantic can be
achieved by GFP_NOWAIT which will not grant access to reserves unless
explicitly stated - e.g. by __GFP_HIGH or __GFP_ATOMIC.
Is that more clear?
Now again why I do not think access to memory reserves is suitable.
Hugetlb pages can be released in a large batches and that might cause a
peak depletion of memory reserves which are normally used by other
consumers as well. Other GFP_ATOMIC users might see allocation failures.
Those shouldn't be really fatal as nobody should be relying on those and
a failure usually mean a hand over to a different, less constrained,
context. So this concern is more about a more well behaved behavior from
the hugetlb side than a correctness.
Is that more clear?
There shouldn't be any real reason why the memory allocation for
vmemmaps, or handling vmemmap in general, has to be done from within the
hugetlb lock and therefore requiring a non-sleeping semantic. All that
can be deferred to a more relaxed context. If you want to make a
GFP_NOWAIT optimistic attempt in the direct free path then no problem
but you have to expect failures under memory pressure. If you want to
have a more robust allocation request then you have to go outside of the
spin lock and use GFP_KERNEL | __GFP_NORETRY or GFP_KERNEL |
__GFP_RETRY_MAYFAIL depending on how hard you want to try.
__GFP_THISNODE makes a slight difference here but something that I would
recommend not depending on.
Is that more clear?
--
Michal Hocko
SUSE Labs
From: Muchun Song <hidden> Date: 2021-02-15 15:39:55
On Mon, Feb 15, 2021 at 9:19 PM Michal Hocko [off-list ref] wrote:
On Mon 15-02-21 20:44:57, Muchun Song wrote:
quoted
On Mon, Feb 15, 2021 at 8:18 PM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 20:00:07, Muchun Song wrote:
quoted
On Mon, Feb 15, 2021 at 7:51 PM Muchun Song [off-list ref] wrote:
quoted
On Mon, Feb 15, 2021 at 6:33 PM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 18:05:06, Muchun Song wrote:
quoted
On Fri, Feb 12, 2021 at 11:32 PM Michal Hocko [off-list ref] wrote:
[...]
quoted
quoted
quoted
+int alloc_huge_page_vmemmap(struct hstate *h, struct page *head)
+{
+ int ret;
+ unsigned long vmemmap_addr = (unsigned long)head;
+ unsigned long vmemmap_end, vmemmap_reuse;
+
+ if (!free_vmemmap_pages_per_hpage(h))
+ return 0;
+
+ vmemmap_addr += RESERVE_VMEMMAP_SIZE;
+ vmemmap_end = vmemmap_addr + free_vmemmap_pages_size_per_hpage(h);
+ vmemmap_reuse = vmemmap_addr - PAGE_SIZE;
+
+ /*
+ * The pages which the vmemmap virtual address range [@vmemmap_addr,
+ * @vmemmap_end) are mapped to are freed to the buddy allocator, and
+ * the range is mapped to the page which @vmemmap_reuse is mapped to.
+ * When a HugeTLB page is freed to the buddy allocator, previously
+ * discarded vmemmap pages must be allocated and remapping.
+ */
+ ret = vmemmap_remap_alloc(vmemmap_addr, vmemmap_end, vmemmap_reuse,
+ GFP_ATOMIC | __GFP_NOWARN | __GFP_THISNODE);
I do not think that this is a good allocation mode. GFP_ATOMIC is a non
sleeping allocation and a medium memory pressure might cause it to
fail prematurely. I do not think this is really an atomic context which
couldn't afford memory reclaim. I also do not think we want to grant
Because alloc_huge_page_vmemmap is called under hugetlb_lock
now. So using GFP_ATOMIC indeed makes the code more simpler.
You can have a preallocated list of pages prior taking the lock.
Moreover do we want to manipulate vmemmaps from under spinlock in
general. I have to say I have missed that detail when reviewing. Need to
think more.
quoted
From the document of the kernel, I learned that __GFP_NOMEMALLOC
can be used to explicitly forbid access to emergency reserves. So if
we do not want to use the reserve memory. How about replacing it to
GFP_ATOMIC | __GFP_NOMEMALLOC | __GFP_NOWARN | __GFP_THISNODE
The whole point of GFP_ATOMIC is to grant access to memory reserves so
the above is quite dubious. If you do not want access to memory reserves
Look at the code of gfp_to_alloc_flags().
static inline unsigned int gfp_to_alloc_flags(gfp_t gfp_mask)
{
[...]
if (gfp_mask & __GFP_ATOMIC) {
/*
* Not worth trying to allocate harder for __GFP_NOMEMALLOC even
* if it can't schedule.
*/
if (!(gfp_mask & __GFP_NOMEMALLOC))
alloc_flags |= ALLOC_HARDER;
[...]
}
Seems to allow this operation (GFP_ATOMIC | __GFP_NOMEMALLOC).
Please read my response again more carefully. I am not claiming that
combination is not allowed. I have said it doesn't make any sense in
this context.
I see you are worried that using GFP_ATOMIC will use reverse memory
unlimited. So I think that __GFP_NOMEMALLOC may be suitable for us.
Sorry, I may not understand the point you said. What I missed?
OK, let me try to explain again. GFP_ATOMIC is not only a non-sleeping
allocation request. It also grants access to memory reserves. The later
is a bit more involved because there are more layers of memory reserves
to access but that is not really important. Non-sleeping semantic can be
achieved by GFP_NOWAIT which will not grant access to reserves unless
explicitly stated - e.g. by __GFP_HIGH or __GFP_ATOMIC.
Is that more clear?
Now again why I do not think access to memory reserves is suitable.
Hugetlb pages can be released in a large batches and that might cause a
peak depletion of memory reserves which are normally used by other
consumers as well. Other GFP_ATOMIC users might see allocation failures.
Those shouldn't be really fatal as nobody should be relying on those and
a failure usually mean a hand over to a different, less constrained,
context. So this concern is more about a more well behaved behavior from
the hugetlb side than a correctness.
Is that more clear?
Ok. It is very clear. Very thanks for your patient explanations.
There shouldn't be any real reason why the memory allocation for
vmemmaps, or handling vmemmap in general, has to be done from within the
hugetlb lock and therefore requiring a non-sleeping semantic. All that
can be deferred to a more relaxed context. If you want to make a
Yeah, you are right. We can put the freeing hugetlb routine to a
workqueue. Just like I do in the previous version (before v13) patch.
I will pick up these patches.
GFP_NOWAIT optimistic attempt in the direct free path then no problem
but you have to expect failures under memory pressure. If you want to
have a more robust allocation request then you have to go outside of the
spin lock and use GFP_KERNEL | __GFP_NORETRY or GFP_KERNEL |
__GFP_RETRY_MAYFAIL depending on how hard you want to try.
__GFP_THISNODE makes a slight difference here but something that I would
recommend not depending on.
Is that more clear?
OK. I will use GFP_KERNEL instead of GFP_ATOMIC. Thanks for your
suggestions.
From: Michal Hocko <mhocko@suse.com> Date: 2021-02-15 16:31:26
On Mon 15-02-21 23:36:49, Muchun Song wrote:
[...]
quoted
There shouldn't be any real reason why the memory allocation for
vmemmaps, or handling vmemmap in general, has to be done from within the
hugetlb lock and therefore requiring a non-sleeping semantic. All that
can be deferred to a more relaxed context. If you want to make a
Yeah, you are right. We can put the freeing hugetlb routine to a
workqueue. Just like I do in the previous version (before v13) patch.
I will pick up these patches.
I haven't seen your v13 and I will unlikely have time to revisit that
version. I just wanted to point out that the actual allocation doesn't
have to happen from under the spinlock. There are multiple ways to go
around that. Dropping the lock would be one of them. Preallocation
before the spin lock is taken is another. WQ is certainly an option but
I would take it as the last resort when other paths are not feasible.
--
Michal Hocko
SUSE Labs
From: Muchun Song <hidden> Date: 2021-02-15 17:50:32
On Tue, Feb 16, 2021 at 12:28 AM Michal Hocko [off-list ref] wrote:
On Mon 15-02-21 23:36:49, Muchun Song wrote:
[...]
quoted
quoted
There shouldn't be any real reason why the memory allocation for
vmemmaps, or handling vmemmap in general, has to be done from within the
hugetlb lock and therefore requiring a non-sleeping semantic. All that
can be deferred to a more relaxed context. If you want to make a
Yeah, you are right. We can put the freeing hugetlb routine to a
workqueue. Just like I do in the previous version (before v13) patch.
I will pick up these patches.
I haven't seen your v13 and I will unlikely have time to revisit that
version. I just wanted to point out that the actual allocation doesn't
have to happen from under the spinlock. There are multiple ways to go
around that. Dropping the lock would be one of them. Preallocation
before the spin lock is taken is another. WQ is certainly an option but
I would take it as the last resort when other paths are not feasible.
"Dropping the lock" and "Preallocation before the spin lock" can limit
the context of put_page to non-atomic context. I am not sure if there
is a page puted somewhere under an atomic context. e.g. compaction.
I am not an expert on this.
From: Muchun Song <hidden> Date: 2021-02-15 18:21:00
On Tue, Feb 16, 2021 at 1:48 AM Muchun Song [off-list ref] wrote:
On Tue, Feb 16, 2021 at 12:28 AM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 23:36:49, Muchun Song wrote:
[...]
quoted
quoted
There shouldn't be any real reason why the memory allocation for
vmemmaps, or handling vmemmap in general, has to be done from within the
hugetlb lock and therefore requiring a non-sleeping semantic. All that
can be deferred to a more relaxed context. If you want to make a
Yeah, you are right. We can put the freeing hugetlb routine to a
workqueue. Just like I do in the previous version (before v13) patch.
I will pick up these patches.
I haven't seen your v13 and I will unlikely have time to revisit that
version. I just wanted to point out that the actual allocation doesn't
have to happen from under the spinlock. There are multiple ways to go
around that. Dropping the lock would be one of them. Preallocation
before the spin lock is taken is another. WQ is certainly an option but
I would take it as the last resort when other paths are not feasible.
"Dropping the lock" and "Preallocation before the spin lock" can limit
the context of put_page to non-atomic context. I am not sure if there
is a page puted somewhere under an atomic context. e.g. compaction.
I am not an expert on this.
Using GFP_KERNEL will also use the current task cpuset to allocate
memory. Do we have an interface to ignore current task cpuset？If not,
WQ may be the only option and it also will not limit the context of
put_page. Right?
From: Michal Hocko <mhocko@suse.com> Date: 2021-02-15 19:40:22
On Tue 16-02-21 02:19:20, Muchun Song wrote:
On Tue, Feb 16, 2021 at 1:48 AM Muchun Song [off-list ref] wrote:
quoted
On Tue, Feb 16, 2021 at 12:28 AM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 23:36:49, Muchun Song wrote:
[...]
quoted
quoted
There shouldn't be any real reason why the memory allocation for
vmemmaps, or handling vmemmap in general, has to be done from within the
hugetlb lock and therefore requiring a non-sleeping semantic. All that
can be deferred to a more relaxed context. If you want to make a
Yeah, you are right. We can put the freeing hugetlb routine to a
workqueue. Just like I do in the previous version (before v13) patch.
I will pick up these patches.
I haven't seen your v13 and I will unlikely have time to revisit that
version. I just wanted to point out that the actual allocation doesn't
have to happen from under the spinlock. There are multiple ways to go
around that. Dropping the lock would be one of them. Preallocation
before the spin lock is taken is another. WQ is certainly an option but
I would take it as the last resort when other paths are not feasible.
"Dropping the lock" and "Preallocation before the spin lock" can limit
the context of put_page to non-atomic context. I am not sure if there
is a page puted somewhere under an atomic context. e.g. compaction.
I am not an expert on this.
Using GFP_KERNEL will also use the current task cpuset to allocate
memory. Do we have an interface to ignore current task cpuset？If not,
WQ may be the only option and it also will not limit the context of
put_page. Right?
Well, GFP_KERNEL is constrained to the task cpuset only if the said
cpuset is hardwalled IIRC. But I do not see why this is a problem.
--
Michal Hocko
SUSE Labs
From: Muchun Song <hidden> Date: 2021-02-16 04:36:21
On Tue, Feb 16, 2021 at 3:39 AM Michal Hocko [off-list ref] wrote:
On Tue 16-02-21 02:19:20, Muchun Song wrote:
quoted
On Tue, Feb 16, 2021 at 1:48 AM Muchun Song [off-list ref] wrote:
quoted
On Tue, Feb 16, 2021 at 12:28 AM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 23:36:49, Muchun Song wrote:
[...]
quoted
quoted
There shouldn't be any real reason why the memory allocation for
vmemmaps, or handling vmemmap in general, has to be done from within the
hugetlb lock and therefore requiring a non-sleeping semantic. All that
can be deferred to a more relaxed context. If you want to make a
Yeah, you are right. We can put the freeing hugetlb routine to a
workqueue. Just like I do in the previous version (before v13) patch.
I will pick up these patches.
I haven't seen your v13 and I will unlikely have time to revisit that
version. I just wanted to point out that the actual allocation doesn't
have to happen from under the spinlock. There are multiple ways to go
around that. Dropping the lock would be one of them. Preallocation
before the spin lock is taken is another. WQ is certainly an option but
I would take it as the last resort when other paths are not feasible.
"Dropping the lock" and "Preallocation before the spin lock" can limit
the context of put_page to non-atomic context. I am not sure if there
is a page puted somewhere under an atomic context. e.g. compaction.
I am not an expert on this.
Using GFP_KERNEL will also use the current task cpuset to allocate
memory. Do we have an interface to ignore current task cpuset？If not,
WQ may be the only option and it also will not limit the context of
put_page. Right?
Well, GFP_KERNEL is constrained to the task cpuset only if the said
cpuset is hardwalled IIRC. But I do not see why this is a problem.
I mean that if there are more than one node in the system,
but the current task cpuset only allows one node. If current
node has no memory and other nodes have enough memory.
We can fail to allocate vmemmap pages. But actually it is
suitable to allocate vmemmap pages from other nodes.
Right?
Thanks.
From: Michal Hocko <mhocko@suse.com> Date: 2021-02-16 08:17:20
On Tue 16-02-21 12:34:41, Muchun Song wrote:
On Tue, Feb 16, 2021 at 3:39 AM Michal Hocko [off-list ref] wrote:
[...]
quoted
quoted
Using GFP_KERNEL will also use the current task cpuset to allocate
memory. Do we have an interface to ignore current task cpuset？If not,
WQ may be the only option and it also will not limit the context of
put_page. Right?
Well, GFP_KERNEL is constrained to the task cpuset only if the said
cpuset is hardwalled IIRC. But I do not see why this is a problem.
I mean that if there are more than one node in the system,
but the current task cpuset only allows one node.
How would that cpuset get a huge pages from a node which is not part of
the cpuset? Well, that would be possible if the cpuset was dynamic but I
am not sure that such a configuration would be very sensible along with
hardwall setup.
If current
node has no memory and other nodes have enough memory.
We can fail to allocate vmemmap pages. But actually it is
suitable to allocate vmemmap pages from other nodes.
Right?
Falling back to a different node would be very suboptimal because then
you would have vmemmap back by remote pages. We do not want that.
--
Michal Hocko
SUSE Labs
From: David Hildenbrand <hidden> Date: 2021-02-16 08:22:54
quoted
If current
node has no memory and other nodes have enough memory.
We can fail to allocate vmemmap pages. But actually it is
suitable to allocate vmemmap pages from other nodes.
Right?
Falling back to a different node would be very suboptimal because then
you would have vmemmap back by remote pages. We do not want that.
... and we even warn when this happens right now:
mm/sparse-vmemmap.c:vmemmap_verify()
--
Thanks,
David / dhildenb
From: Muchun Song <hidden> Date: 2021-02-16 09:06:49
On Tue, Feb 16, 2021 at 4:15 PM Michal Hocko [off-list ref] wrote:
On Tue 16-02-21 12:34:41, Muchun Song wrote:
quoted
On Tue, Feb 16, 2021 at 3:39 AM Michal Hocko [off-list ref] wrote:
[...]
quoted
quoted
quoted
Using GFP_KERNEL will also use the current task cpuset to allocate
memory. Do we have an interface to ignore current task cpuset？If not,
WQ may be the only option and it also will not limit the context of
put_page. Right?
Well, GFP_KERNEL is constrained to the task cpuset only if the said
cpuset is hardwalled IIRC. But I do not see why this is a problem.
I mean that if there are more than one node in the system,
but the current task cpuset only allows one node.
How would that cpuset get a huge pages from a node which is not part of
the cpuset? Well, that would be possible if the cpuset was dynamic but I
am not sure that such a configuration would be very sensible along with
hardwall setup.
Got it. I didn't realize this before. Thanks.
quoted
If current
node has no memory and other nodes have enough memory.
We can fail to allocate vmemmap pages. But actually it is
suitable to allocate vmemmap pages from other nodes.
Right?
Falling back to a different node would be very suboptimal because then
you would have vmemmap back by remote pages. We do not want that.
--
Michal Hocko
SUSE Labs
From: Michal Hocko <mhocko@suse.com> Date: 2021-02-15 19:04:02
On Tue 16-02-21 01:48:29, Muchun Song wrote:
On Tue, Feb 16, 2021 at 12:28 AM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 23:36:49, Muchun Song wrote:
[...]
quoted
quoted
There shouldn't be any real reason why the memory allocation for
vmemmaps, or handling vmemmap in general, has to be done from within the
hugetlb lock and therefore requiring a non-sleeping semantic. All that
can be deferred to a more relaxed context. If you want to make a
Yeah, you are right. We can put the freeing hugetlb routine to a
workqueue. Just like I do in the previous version (before v13) patch.
I will pick up these patches.
I haven't seen your v13 and I will unlikely have time to revisit that
version. I just wanted to point out that the actual allocation doesn't
have to happen from under the spinlock. There are multiple ways to go
around that. Dropping the lock would be one of them. Preallocation
before the spin lock is taken is another. WQ is certainly an option but
I would take it as the last resort when other paths are not feasible.
"Dropping the lock" and "Preallocation before the spin lock" can limit
the context of put_page to non-atomic context. I am not sure if there
is a page puted somewhere under an atomic context. e.g. compaction.
I am not an expert on this.
Then do a due research or ask for a help from the MM community. Do
not just try to go around harder problems and somehow duct tape a
solution. I am sorry for sounding harsh here but this is a repetitive
pattern.
Now to the merit. put_page can indeed be called from all sorts of
contexts. And it might be indeed impossible to guarantee that hugetlb
pages are never freed up from an atomic context. Requiring that would be
even hard to maintain longterm. There are ways around that, I believe,
though.
The most simple one that I can think of right now would be using
in_atomic() rather than in_task() check free_huge_page. IIRC recent
changes would allow in_atomic to be reliable also on !PREEMPT kernels
(via RCU tree, not sure where this stands right now). That would make
__free_huge_page always run in a non-atomic context which sounds like an
easy enough solution.
Another way would be to keep a pool of ready pages to use in case of
GFP_NOWAIT allocation fails and have means to keep that pool replenished
when needed. Would it be feasible to reused parts of the freed page in
the worst case?
--
Michal Hocko
SUSE Labs
From: David Hildenbrand <hidden> Date: 2021-02-16 08:15:13
On 15.02.21 20:02, Michal Hocko wrote:
On Tue 16-02-21 01:48:29, Muchun Song wrote:
quoted
On Tue, Feb 16, 2021 at 12:28 AM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 23:36:49, Muchun Song wrote:
[...]
quoted
quoted
There shouldn't be any real reason why the memory allocation for
vmemmaps, or handling vmemmap in general, has to be done from within the
hugetlb lock and therefore requiring a non-sleeping semantic. All that
can be deferred to a more relaxed context. If you want to make a
Yeah, you are right. We can put the freeing hugetlb routine to a
workqueue. Just like I do in the previous version (before v13) patch.
I will pick up these patches.
I haven't seen your v13 and I will unlikely have time to revisit that
version. I just wanted to point out that the actual allocation doesn't
have to happen from under the spinlock. There are multiple ways to go
around that. Dropping the lock would be one of them. Preallocation
before the spin lock is taken is another. WQ is certainly an option but
I would take it as the last resort when other paths are not feasible.
"Dropping the lock" and "Preallocation before the spin lock" can limit
the context of put_page to non-atomic context. I am not sure if there
is a page puted somewhere under an atomic context. e.g. compaction.
I am not an expert on this.
Then do a due research or ask for a help from the MM community. Do
not just try to go around harder problems and somehow duct tape a
solution. I am sorry for sounding harsh here but this is a repetitive
pattern.
Now to the merit. put_page can indeed be called from all sorts of
contexts. And it might be indeed impossible to guarantee that hugetlb
pages are never freed up from an atomic context. Requiring that would be
even hard to maintain longterm. There are ways around that, I believe,
though.
The most simple one that I can think of right now would be using
in_atomic() rather than in_task() check free_huge_page. IIRC recent
changes would allow in_atomic to be reliable also on !PREEMPT kernels
(via RCU tree, not sure where this stands right now). That would make
__free_huge_page always run in a non-atomic context which sounds like an
easy enough solution.
Another way would be to keep a pool of ready pages to use in case of
GFP_NOWAIT allocation fails and have means to keep that pool replenished
when needed. Would it be feasible to reused parts of the freed page in
the worst case?
As already discussed, this is only possible when the huge page does not
reside on ZONE_MOVABLE/CMA.
In addition, we can no longer form a huge page at that memory location ever.
--
Thanks,
David / dhildenb
From: Michal Hocko <mhocko@suse.com> Date: 2021-02-16 08:22:36
On Tue 16-02-21 09:13:09, David Hildenbrand wrote:
On 15.02.21 20:02, Michal Hocko wrote:
quoted
Would it be feasible to reused parts of the freed page in
the worst case?
As already discussed, this is only possible when the huge page does not
reside on ZONE_MOVABLE/CMA.
Right. But usually this is not the case so it would be at least
something.
In addition, we can no longer form a huge page at that memory location ever.
Yes, that's why I am saying in the worst case. E.g. when dissolving is
really necessary like in hwpoison case.
Maybe we are really far from needing something like that. I just wanted
to mention this option and I was not aware this has been discussed
previously.
--
Michal Hocko
SUSE Labs
From: Mike Kravetz <hidden> Date: 2021-02-16 19:46:34
On 2/15/21 8:27 AM, Michal Hocko wrote:
On Mon 15-02-21 23:36:49, Muchun Song wrote:
[...]
quoted
quoted
There shouldn't be any real reason why the memory allocation for
vmemmaps, or handling vmemmap in general, has to be done from within the
hugetlb lock and therefore requiring a non-sleeping semantic. All that
can be deferred to a more relaxed context. If you want to make a
Yeah, you are right. We can put the freeing hugetlb routine to a
workqueue. Just like I do in the previous version (before v13) patch.
I will pick up these patches.
I haven't seen your v13 and I will unlikely have time to revisit that
version. I just wanted to point out that the actual allocation doesn't
have to happen from under the spinlock. There are multiple ways to go
around that. Dropping the lock would be one of them. Preallocation
before the spin lock is taken is another. WQ is certainly an option but
I would take it as the last resort when other paths are not feasible.
Sorry for jumping in late, Monday was a US holiday ...
IIRC, the point of moving the vmemmap allocations under the hugetlb_lock
was just for simplicity. The idea was to modify the allocations to be
non-blocking so that allocating pages and restoring vmemmap could be done
as part of normal huge page freeing where we are holding the lock. Perhaps
that is too simplistic of an approach.
IMO, using the workque approach as done in previous patches introduces
too much complexity.
Michal did bring up the question "Do we really want to do all the vmemmap
allocation (even non-blocking) and manipulation under the hugetlb lock?
I'm thinking the answer may be no. For 1G pages, this will require 4094
calls to alloc_pages. Even with non-blocking calls this seems like a long
time.
If we are not going to do the allocations under the lock, then we will need
to either preallocate or take the workqueue approach. One complication with
preallocation is that we do not for sure we will be freeing the huge page
to buddy until we take the hugetlb_lock. This is because the decision to
free or not is based on counters protected by the lock. We could of course
check counters without the lock to guess if we will be freeing the page,
and then check again after acquiring the lock. This may not be too bad in
the case of freeing a single page, but would become more complex when doing
bulk freeing. After a little thought, the workqueue approach may even end
up simpler. However, I would suggest a very simple workqueue implementation
with non-blocking allocations. If we can not quickly get vmemmap pages,
put the page back on the hugetlb free list and treat as a surplus page.
--
Mike Kravetz
From: Mike Kravetz <hidden> Date: 2021-02-18 01:02:29
On 2/17/21 12:13 AM, Michal Hocko wrote:
On Tue 16-02-21 11:44:34, Mike Kravetz wrote:
[...]
quoted
If we are not going to do the allocations under the lock, then we will need
to either preallocate or take the workqueue approach.
We can still drop the lock temporarily right? As we already do before
calling destroy_compound_gigantic_page...
Yes we can. I forgot about that.
Actually, very little of what update_and_free_page does needs to be done
under the lock. Perhaps, just decrementing the global count and clearing
the destructor so PageHuge() is no longer true.
--
Mike Kravetz
From: Muchun Song <hidden> Date: 2021-02-18 03:22:11
On Thu, Feb 18, 2021 at 9:00 AM Mike Kravetz [off-list ref] wrote:
On 2/17/21 12:13 AM, Michal Hocko wrote:
quoted
On Tue 16-02-21 11:44:34, Mike Kravetz wrote:
[...]
quoted
If we are not going to do the allocations under the lock, then we will need
to either preallocate or take the workqueue approach.
We can still drop the lock temporarily right? As we already do before
calling destroy_compound_gigantic_page...
Yes we can. I forgot about that.
Actually, very little of what update_and_free_page does needs to be done
under the lock. Perhaps, just decrementing the global count and clearing
the destructor so PageHuge() is no longer true.
Right. I have another question about using GFP flags. Michal
suggested using GFP_KERNEL instead of GFP_ATOMIC to
save reserve memory. From your last email, you suggested
using non-blocking allocation GFP flags (perhaps GFP_NOWAIT).
Hi Mike and Michal,
What is the consensus we finally reached? Thanks.
From: Michal Hocko <mhocko@suse.com> Date: 2021-02-18 08:26:10
On Thu 18-02-21 11:20:51, Muchun Song wrote:
On Thu, Feb 18, 2021 at 9:00 AM Mike Kravetz [off-list ref] wrote:
quoted
On 2/17/21 12:13 AM, Michal Hocko wrote:
quoted
On Tue 16-02-21 11:44:34, Mike Kravetz wrote:
[...]
quoted
If we are not going to do the allocations under the lock, then we will need
to either preallocate or take the workqueue approach.
We can still drop the lock temporarily right? As we already do before
calling destroy_compound_gigantic_page...
Yes we can. I forgot about that.
Actually, very little of what update_and_free_page does needs to be done
under the lock. Perhaps, just decrementing the global count and clearing
the destructor so PageHuge() is no longer true.
Right. I have another question about using GFP flags. Michal
suggested using GFP_KERNEL instead of GFP_ATOMIC to
save reserve memory. From your last email, you suggested
using non-blocking allocation GFP flags (perhaps GFP_NOWAIT).
Hi Mike and Michal,
What is the consensus we finally reached? Thanks.
If the lock can be dropped and you make sure the final put on page is
not called from an atomic context then use (for starter)
GFP_KERNEL | __GFP_NORETRY | __GFP_THISNODE. I have intentionaly dropped
__GFP_NOWARN because likely want to hear about the failure so that we
can evaluate how often this happens.
This would be my recommendation.
--
Michal Hocko
SUSE Labs
From: Michal Hocko <mhocko@suse.com> Date: 2021-02-15 12:26:13
On Mon 15-02-21 19:51:26, Muchun Song wrote:
On Mon, Feb 15, 2021 at 6:33 PM Michal Hocko [off-list ref] wrote:
quoted
On Mon 15-02-21 18:05:06, Muchun Song wrote:
quoted
On Fri, Feb 12, 2021 at 11:32 PM Michal Hocko [off-list ref] wrote:
[...]
quoted
quoted
quoted
+int alloc_huge_page_vmemmap(struct hstate *h, struct page *head)
+{
+ int ret;
+ unsigned long vmemmap_addr = (unsigned long)head;
+ unsigned long vmemmap_end, vmemmap_reuse;
+
+ if (!free_vmemmap_pages_per_hpage(h))
+ return 0;
+
+ vmemmap_addr += RESERVE_VMEMMAP_SIZE;
+ vmemmap_end = vmemmap_addr + free_vmemmap_pages_size_per_hpage(h);
+ vmemmap_reuse = vmemmap_addr - PAGE_SIZE;
+
+ /*
+ * The pages which the vmemmap virtual address range [@vmemmap_addr,
+ * @vmemmap_end) are mapped to are freed to the buddy allocator, and
+ * the range is mapped to the page which @vmemmap_reuse is mapped to.
+ * When a HugeTLB page is freed to the buddy allocator, previously
+ * discarded vmemmap pages must be allocated and remapping.
+ */
+ ret = vmemmap_remap_alloc(vmemmap_addr, vmemmap_end, vmemmap_reuse,
+ GFP_ATOMIC | __GFP_NOWARN | __GFP_THISNODE);
I do not think that this is a good allocation mode. GFP_ATOMIC is a non
sleeping allocation and a medium memory pressure might cause it to
fail prematurely. I do not think this is really an atomic context which
couldn't afford memory reclaim. I also do not think we want to grant
Because alloc_huge_page_vmemmap is called under hugetlb_lock
now. So using GFP_ATOMIC indeed makes the code more simpler.
You can have a preallocated list of pages prior taking the lock.
I do not see any real response to the pre-allocation argument except
that put_page can be called from an atomic context. Which might be true
in general but it is not the case for hugetlb pages. hugetlb_lock would
have to be irq safe otherwise. Also the whole operation can be scheduled
to a kworker context for a stronger allocation context.
--
Michal Hocko
SUSE Labs