* [PATCH] mm: hugetlb_vmemmap: use bulk allocator in alloc_vmemmap_page_list()
@ 2023-09-05 7:10 Kefeng Wang
2023-09-05 7:32 ` Kefeng Wang
0 siblings, 1 reply; 5+ messages in thread
From: Kefeng Wang @ 2023-09-05 7:10 UTC (permalink / raw)
To: Andrew Morton, Mike Kravetz, Muchun Song, linux-mm; +Cc: Kefeng Wang, Yuan Can
It is needed 4095 pages(1G) or 7 pages(2M) to be allocated once in
alloc_vmemmap_page_list(), so let's add a bulk allocator varietas
alloc_pages_bulk_list_node() and switch alloc_vmemmap_page_list()
to use it to accelerate page allocation.
Simple test on arm64's qemu with 1G Hugetlb, 870,842ns vs 3,845,252ns
despite the fluctuations, it is still a nice improvement.
Tested-by: Yuan Can <yuancan@huawei.com>
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
---
include/linux/gfp.h | 9 +++++++++
mm/hugetlb_vmemmap.c | 7 ++++++-
2 files changed, 15 insertions(+), 1 deletion(-)
diff --git a/include/linux/gfp.h b/include/linux/gfp.h
index 665f06675c83..d6e82f15b61f 100644
--- a/include/linux/gfp.h
+++ b/include/linux/gfp.h
@@ -195,6 +195,15 @@ alloc_pages_bulk_list(gfp_t gfp, unsigned long nr_pages, struct list_head *list)
return __alloc_pages_bulk(gfp, numa_mem_id(), NULL, nr_pages, list, NULL);
}
+static inline unsigned long
+alloc_pages_bulk_list_node(gfp_t gfp, int nid, unsigned long nr_pages, struct list_head *list)
+{
+ if (nid == NUMA_NO_NODE)
+ nid = numa_mem_id();
+
+ return __alloc_pages_bulk(gfp, nid, NULL, nr_pages, list, NULL);
+}
+
static inline unsigned long
alloc_pages_bulk_array(gfp_t gfp, unsigned long nr_pages, struct page **page_array)
{
diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
index 4b9734777f69..699c4fea6b9f 100644
--- a/mm/hugetlb_vmemmap.c
+++ b/mm/hugetlb_vmemmap.c
@@ -384,8 +384,13 @@ static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
int nid = page_to_nid((struct page *)start);
struct page *page, *next;
+ unsigned long nr_alloced;
- while (nr_pages--) {
+ nr_alloced = alloc_pages_bulk_list_node(gfp_mask, nid, nr_pages, list);
+ if (!nr_alloced)
+ return -ENOMEM;
+
+ while (nr_alloced < nr_pages) {
page = alloc_pages_node(nid, gfp_mask, 0);
if (!page)
goto out;
--
2.27.0
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] mm: hugetlb_vmemmap: use bulk allocator in alloc_vmemmap_page_list()
2023-09-05 7:10 [PATCH] mm: hugetlb_vmemmap: use bulk allocator in alloc_vmemmap_page_list() Kefeng Wang
@ 2023-09-05 7:32 ` Kefeng Wang
2023-09-05 9:23 ` Muchun Song
0 siblings, 1 reply; 5+ messages in thread
From: Kefeng Wang @ 2023-09-05 7:32 UTC (permalink / raw)
To: Andrew Morton, Mike Kravetz, Muchun Song, linux-mm; +Cc: Yuan Can
On 2023/9/5 15:10, Kefeng Wang wrote:
> It is needed 4095 pages(1G) or 7 pages(2M) to be allocated once in
> alloc_vmemmap_page_list(), so let's add a bulk allocator varietas
> alloc_pages_bulk_list_node() and switch alloc_vmemmap_page_list()
> to use it to accelerate page allocation.
>
> Simple test on arm64's qemu with 1G Hugetlb, 870,842ns vs 3,845,252ns
> despite the fluctuations, it is still a nice improvement.
>
> Tested-by: Yuan Can <yuancan@huawei.com>
> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
> ---
> include/linux/gfp.h | 9 +++++++++
> mm/hugetlb_vmemmap.c | 7 ++++++-
> 2 files changed, 15 insertions(+), 1 deletion(-)
>
> diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> index 665f06675c83..d6e82f15b61f 100644
> --- a/include/linux/gfp.h
> +++ b/include/linux/gfp.h
> @@ -195,6 +195,15 @@ alloc_pages_bulk_list(gfp_t gfp, unsigned long nr_pages, struct list_head *list)
> return __alloc_pages_bulk(gfp, numa_mem_id(), NULL, nr_pages, list, NULL);
> }
>
> +static inline unsigned long
> +alloc_pages_bulk_list_node(gfp_t gfp, int nid, unsigned long nr_pages, struct list_head *list)
> +{
> + if (nid == NUMA_NO_NODE)
> + nid = numa_mem_id();
> +
> + return __alloc_pages_bulk(gfp, nid, NULL, nr_pages, list, NULL);
> +}
> +
> static inline unsigned long
> alloc_pages_bulk_array(gfp_t gfp, unsigned long nr_pages, struct page **page_array)
> {
> diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
> index 4b9734777f69..699c4fea6b9f 100644
> --- a/mm/hugetlb_vmemmap.c
> +++ b/mm/hugetlb_vmemmap.c
> @@ -384,8 +384,13 @@ static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
> unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
> int nid = page_to_nid((struct page *)start);
> struct page *page, *next;
> + unsigned long nr_alloced;
>
> - while (nr_pages--) {
> + nr_alloced = alloc_pages_bulk_list_node(gfp_mask, nid, nr_pages, list);
> + if (!nr_alloced)
> + return -ENOMEM;
> +
eh， forget to inc nr_allocated in the fallback patch， will resend
> + while (nr_alloced < nr_pages) {
> page = alloc_pages_node(nid, gfp_mask, 0);
> if (!page)
> goto out;
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] mm: hugetlb_vmemmap: use bulk allocator in alloc_vmemmap_page_list()
2023-09-05 7:32 ` Kefeng Wang
@ 2023-09-05 9:23 ` Muchun Song
2023-09-05 9:34 ` Kefeng Wang
2023-09-05 10:12 ` Kefeng Wang
0 siblings, 2 replies; 5+ messages in thread
From: Muchun Song @ 2023-09-05 9:23 UTC (permalink / raw)
To: Kefeng Wang; +Cc: Andrew Morton, Mike Kravetz, Linux-MM, Yuan Can
> On Sep 5, 2023, at 15:32, Kefeng Wang <wangkefeng.wang@huawei.com> wrote:
> On 2023/9/5 15:10, Kefeng Wang wrote:
>> It is needed 4095 pages(1G) or 7 pages(2M) to be allocated once in
>> alloc_vmemmap_page_list(), so let's add a bulk allocator varietas
>> alloc_pages_bulk_list_node() and switch alloc_vmemmap_page_list()
>> to use it to accelerate page allocation.
>> Simple test on arm64's qemu with 1G Hugetlb, 870,842ns vs 3,845,252ns
>> despite the fluctuations, it is still a nice improvement.
>> Tested-by: Yuan Can <yuancan@huawei.com>
>> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
>> ---
>> include/linux/gfp.h | 9 +++++++++
>> mm/hugetlb_vmemmap.c | 7 ++++++-
>> 2 files changed, 15 insertions(+), 1 deletion(-)
>> diff --git a/include/linux/gfp.h b/include/linux/gfp.h
>> index 665f06675c83..d6e82f15b61f 100644
>> --- a/include/linux/gfp.h
>> +++ b/include/linux/gfp.h
>> @@ -195,6 +195,15 @@ alloc_pages_bulk_list(gfp_t gfp, unsigned long nr_pages, struct list_head *list)
>> return __alloc_pages_bulk(gfp, numa_mem_id(), NULL, nr_pages, list, NULL);
>> }
>> +static inline unsigned long
>> +alloc_pages_bulk_list_node(gfp_t gfp, int nid, unsigned long nr_pages, struct list_head *list)
>> +{
>> + if (nid == NUMA_NO_NODE)
>> + nid = numa_mem_id();
>> +
>> + return __alloc_pages_bulk(gfp, nid, NULL, nr_pages, list, NULL);
>> +}
>> +
>> static inline unsigned long
>> alloc_pages_bulk_array(gfp_t gfp, unsigned long nr_pages, struct page **page_array)
>> {
>> diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
>> index 4b9734777f69..699c4fea6b9f 100644
>> --- a/mm/hugetlb_vmemmap.c
>> +++ b/mm/hugetlb_vmemmap.c
>> @@ -384,8 +384,13 @@ static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
>> unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
>> int nid = page_to_nid((struct page *)start);
>> struct page *page, *next;
>> + unsigned long nr_alloced;
>> - while (nr_pages--) {
>> + nr_alloced = alloc_pages_bulk_list_node(gfp_mask, nid, nr_pages, list);
>> + if (!nr_alloced)
>> + return -ENOMEM;
>> +
>
> eh， forget to inc nr_allocated in the fallback patch， will resend
Do not change the judgement, "nr_pages -= nr_alloced;" is enough
and simple.
>> + while (nr_alloced < nr_pages) {
>> page = alloc_pages_node(nid, gfp_mask, 0);
>> if (!page)
>> goto out;
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] mm: hugetlb_vmemmap: use bulk allocator in alloc_vmemmap_page_list()
2023-09-05 9:23 ` Muchun Song
@ 2023-09-05 9:34 ` Kefeng Wang
2023-09-05 10:12 ` Kefeng Wang
1 sibling, 0 replies; 5+ messages in thread
From: Kefeng Wang @ 2023-09-05 9:34 UTC (permalink / raw)
To: Muchun Song; +Cc: Andrew Morton, Mike Kravetz, Linux-MM, Yuan Can
On 2023/9/5 17:23, Muchun Song wrote:
>
>
>> On Sep 5, 2023, at 15:32, Kefeng Wang <wangkefeng.wang@huawei.com> wrote:
>> On 2023/9/5 15:10, Kefeng Wang wrote:
>>> It is needed 4095 pages(1G) or 7 pages(2M) to be allocated once in
>>> alloc_vmemmap_page_list(), so let's add a bulk allocator varietas
>>> alloc_pages_bulk_list_node() and switch alloc_vmemmap_page_list()
>>> to use it to accelerate page allocation.
>>> Simple test on arm64's qemu with 1G Hugetlb, 870,842ns vs 3,845,252ns
>>> despite the fluctuations, it is still a nice improvement.
>>> Tested-by: Yuan Can <yuancan@huawei.com>
>>> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
>>> ---
>>> include/linux/gfp.h | 9 +++++++++
>>> mm/hugetlb_vmemmap.c | 7 ++++++-
>>> 2 files changed, 15 insertions(+), 1 deletion(-)
>>> diff --git a/include/linux/gfp.h b/include/linux/gfp.h
>>> index 665f06675c83..d6e82f15b61f 100644
>>> --- a/include/linux/gfp.h
>>> +++ b/include/linux/gfp.h
>>> @@ -195,6 +195,15 @@ alloc_pages_bulk_list(gfp_t gfp, unsigned long nr_pages, struct list_head *list)
>>> return __alloc_pages_bulk(gfp, numa_mem_id(), NULL, nr_pages, list, NULL);
>>> }
>>> +static inline unsigned long
>>> +alloc_pages_bulk_list_node(gfp_t gfp, int nid, unsigned long nr_pages, struct list_head *list)
>>> +{
>>> + if (nid == NUMA_NO_NODE)
>>> + nid = numa_mem_id();
>>> +
>>> + return __alloc_pages_bulk(gfp, nid, NULL, nr_pages, list, NULL);
>>> +}
>>> +
>>> static inline unsigned long
>>> alloc_pages_bulk_array(gfp_t gfp, unsigned long nr_pages, struct page **page_array)
>>> {
>>> diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
>>> index 4b9734777f69..699c4fea6b9f 100644
>>> --- a/mm/hugetlb_vmemmap.c
>>> +++ b/mm/hugetlb_vmemmap.c
>>> @@ -384,8 +384,13 @@ static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
>>> unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
>>> int nid = page_to_nid((struct page *)start);
>>> struct page *page, *next;
>>> + unsigned long nr_alloced;
>>> - while (nr_pages--) {
>>> + nr_alloced = alloc_pages_bulk_list_node(gfp_mask, nid, nr_pages, list);
>>> + if (!nr_alloced)
>>> + return -ENOMEM;
>>> +
>>
>> eh， forget to inc nr_allocated in the fallback patch， will resend
>
> Do not change the judgement, "nr_pages -= nr_alloced;" is enough
> and simple.
sure, thanks.
>
>>> + while (nr_alloced < nr_pages) {
>>> page = alloc_pages_node(nid, gfp_mask, 0);
>>> if (!page)
>>> goto out;
>
>
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] mm: hugetlb_vmemmap: use bulk allocator in alloc_vmemmap_page_list()
2023-09-05 9:23 ` Muchun Song
2023-09-05 9:34 ` Kefeng Wang
@ 2023-09-05 10:12 ` Kefeng Wang
1 sibling, 0 replies; 5+ messages in thread
From: Kefeng Wang @ 2023-09-05 10:12 UTC (permalink / raw)
To: Muchun Song; +Cc: Andrew Morton, Mike Kravetz, Linux-MM, Yuan Can
On 2023/9/5 17:23, Muchun Song wrote:
>
>
>> On Sep 5, 2023, at 15:32, Kefeng Wang <wangkefeng.wang@huawei.com> wrote:
>> On 2023/9/5 15:10, Kefeng Wang wrote:
>>> It is needed 4095 pages(1G) or 7 pages(2M) to be allocated once in
>>> alloc_vmemmap_page_list(), so let's add a bulk allocator varietas
>>> alloc_pages_bulk_list_node() and switch alloc_vmemmap_page_list()
>>> to use it to accelerate page allocation.
>>> Simple test on arm64's qemu with 1G Hugetlb, 870,842ns vs 3,845,252ns
>>> despite the fluctuations, it is still a nice improvement.
>>> Tested-by: Yuan Can <yuancan@huawei.com>
>>> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
>>> ---
>>> include/linux/gfp.h | 9 +++++++++
>>> mm/hugetlb_vmemmap.c | 7 ++++++-
>>> 2 files changed, 15 insertions(+), 1 deletion(-)
>>> diff --git a/include/linux/gfp.h b/include/linux/gfp.h
>>> index 665f06675c83..d6e82f15b61f 100644
>>> --- a/include/linux/gfp.h
>>> +++ b/include/linux/gfp.h
>>> @@ -195,6 +195,15 @@ alloc_pages_bulk_list(gfp_t gfp, unsigned long nr_pages, struct list_head *list)
>>> return __alloc_pages_bulk(gfp, numa_mem_id(), NULL, nr_pages, list, NULL);
>>> }
>>> +static inline unsigned long
>>> +alloc_pages_bulk_list_node(gfp_t gfp, int nid, unsigned long nr_pages, struct list_head *list)
>>> +{
>>> + if (nid == NUMA_NO_NODE)
>>> + nid = numa_mem_id();
>>> +
>>> + return __alloc_pages_bulk(gfp, nid, NULL, nr_pages, list, NULL);
>>> +}
>>> +
>>> static inline unsigned long
>>> alloc_pages_bulk_array(gfp_t gfp, unsigned long nr_pages, struct page **page_array)
>>> {
>>> diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
>>> index 4b9734777f69..699c4fea6b9f 100644
>>> --- a/mm/hugetlb_vmemmap.c
>>> +++ b/mm/hugetlb_vmemmap.c
>>> @@ -384,8 +384,13 @@ static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
>>> unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
>>> int nid = page_to_nid((struct page *)start);
>>> struct page *page, *next;
>>> + unsigned long nr_alloced;
>>> - while (nr_pages--) {
>>> + nr_alloced = alloc_pages_bulk_list_node(gfp_mask, nid, nr_pages, list);
>>> + if (!nr_alloced)
>>> + return -ENOMEM;
>>> +
>>
>> eh， forget to inc nr_allocated in the fallback patch， will resend
>
> Do not change the judgement, "nr_pages -= nr_alloced;" is enough
> and simple.
nr_pages = 7, nr_alloced = 4, new nr_pages = 3, the fallback won't
execute if nr_alloced not cleared, will add nr_allocated only alloc page
successfully.
while (nr_allocated < nr_pages) {
page = alloc_pages_node(nid, gfp_mask, 0);
if (!page)
goto out;
list_add_tail(&page->lru, list);
+ nr_allocated++;
}
>
>>> + while (nr_alloced < nr_pages) {
>>> page = alloc_pages_node(nid, gfp_mask, 0);
>>> if (!page)
>>> goto out;
>
>
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2023-09-05 10:12 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-09-05 7:10 [PATCH] mm: hugetlb_vmemmap: use bulk allocator in alloc_vmemmap_page_list() Kefeng Wang
2023-09-05 7:32 ` Kefeng Wang
2023-09-05 9:23 ` Muchun Song
2023-09-05 9:34 ` Kefeng Wang
2023-09-05 10:12 ` Kefeng Wang
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox