Re: [External] Re: [PATCH v20 5/9] mm: hugetlb: defer freeing of HugeTLB pages
On Sat, Apr 17, 2021 at 7:56 AM Mike Kravetz wrote:
>
> On 4/15/21 1:40 AM, Muchun Song wrote:
> > In the subsequent patch, we should allocate the vmemmap pages when
> > freeing a HugeTLB page. But update_and_free_page() can be called
> > under any context, so we cannot use GFP_KERNEL to allocate vmemmap
> > pages. However, we can defer the actual freeing in a kworker to
> > prevent from using GFP_ATOMIC to allocate the vmemmap pages.
>
> Thanks! I knew we would need to introduce a kworker for this when I
> removed the kworker previously used in free_huge_page.
Yeah, but another choice is using GFP_ATOMIC to allocate vmemmap
pages when we are in an atomic context. If not atomic context, just
use GFP_KERNEL. In this case, we can drop kworker.
>
> > The __update_and_free_page() is where the call to allocate vmemmmap
> > pages will be inserted.
>
> This patch adds the functionality required for __update_and_free_page
> to potentially sleep and fail. More questions will come up in the
> subsequent patch when code must deal with the failures.
Right. More questions are welcome.
>
> >
> > Signed-off-by: Muchun Song
> > ---
> > mm/hugetlb.c | 73
> >
> > mm/hugetlb_vmemmap.c | 12 -
> > mm/hugetlb_vmemmap.h | 17
> > 3 files changed, 85 insertions(+), 17 deletions(-)
> >
> > diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> > index 923d05e2806b..eeb8f5480170 100644
> > --- a/mm/hugetlb.c
> > +++ b/mm/hugetlb.c
> > @@ -1376,7 +1376,7 @@ static void remove_hugetlb_page(struct hstate *h,
> > struct page *page,
> > h->nr_huge_pages_node[nid]--;
> > }
> >
> > -static void update_and_free_page(struct hstate *h, struct page *page)
> > +static void __update_and_free_page(struct hstate *h, struct page *page)
> > {
> > int i;
> > struct page *subpage = page;
> > @@ -1399,12 +1399,73 @@ static void update_and_free_page(struct hstate *h,
> > struct page *page)
> > }
> > }
> >
> > +/*
> > + * As update_and_free_page() can be called under any context, so we cannot
> > + * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
> > + * actual freeing in a workqueue to prevent from using GFP_ATOMIC to
> > allocate
> > + * the vmemmap pages.
> > + *
> > + * free_hpage_workfn() locklessly retrieves the linked list of pages to be
> > + * freed and frees them one-by-one. As the page->mapping pointer is going
> > + * to be cleared in free_hpage_workfn() anyway, it is reused as the
> > llist_node
> > + * structure of a lockless linked list of huge pages to be freed.
> > + */
> > +static LLIST_HEAD(hpage_freelist);
> > +
> > +static void free_hpage_workfn(struct work_struct *work)
> > +{
> > + struct llist_node *node;
> > +
> > + node = llist_del_all(&hpage_freelist);
> > +
> > + while (node) {
> > + struct page *page;
> > + struct hstate *h;
> > +
> > + page = container_of((struct address_space **)node,
> > + struct page, mapping);
> > + node = node->next;
> > + page->mapping = NULL;
> > + h = page_hstate(page);
>
> The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate is going to
> trigger because a previous call to remove_hugetlb_page() will
> set_compound_page_dtor(page, NULL_COMPOUND_DTOR)
Sorry, I did not realise that. Thanks for your reminder.
>
> Note how h(hstate) is grabbed before calling update_and_free_page in
> existing code.
>
> We could potentially drop the !PageHuge(page) in page_hstate. Or,
> perhaps just use 'size_to_hstate(page_size(page))' in free_hpage_workfn.
I prefer not to change the behavior of page_hstate(). So I
should use 'size_to_hstate(page_size(page))' directly.
Thanks Mike.
> --
> Mike Kravetz
Re: [PATCH v20 5/9] mm: hugetlb: defer freeing of HugeTLB pages
On 4/15/21 1:40 AM, Muchun Song wrote:
> In the subsequent patch, we should allocate the vmemmap pages when
> freeing a HugeTLB page. But update_and_free_page() can be called
> under any context, so we cannot use GFP_KERNEL to allocate vmemmap
> pages. However, we can defer the actual freeing in a kworker to
> prevent from using GFP_ATOMIC to allocate the vmemmap pages.
Thanks! I knew we would need to introduce a kworker for this when I
removed the kworker previously used in free_huge_page.
> The __update_and_free_page() is where the call to allocate vmemmmap
> pages will be inserted.
This patch adds the functionality required for __update_and_free_page
to potentially sleep and fail. More questions will come up in the
subsequent patch when code must deal with the failures.
>
> Signed-off-by: Muchun Song
> ---
> mm/hugetlb.c | 73
>
> mm/hugetlb_vmemmap.c | 12 -
> mm/hugetlb_vmemmap.h | 17
> 3 files changed, 85 insertions(+), 17 deletions(-)
>
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index 923d05e2806b..eeb8f5480170 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -1376,7 +1376,7 @@ static void remove_hugetlb_page(struct hstate *h,
> struct page *page,
> h->nr_huge_pages_node[nid]--;
> }
>
> -static void update_and_free_page(struct hstate *h, struct page *page)
> +static void __update_and_free_page(struct hstate *h, struct page *page)
> {
> int i;
> struct page *subpage = page;
> @@ -1399,12 +1399,73 @@ static void update_and_free_page(struct hstate *h,
> struct page *page)
> }
> }
>
> +/*
> + * As update_and_free_page() can be called under any context, so we cannot
> + * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
> + * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
> + * the vmemmap pages.
> + *
> + * free_hpage_workfn() locklessly retrieves the linked list of pages to be
> + * freed and frees them one-by-one. As the page->mapping pointer is going
> + * to be cleared in free_hpage_workfn() anyway, it is reused as the
> llist_node
> + * structure of a lockless linked list of huge pages to be freed.
> + */
> +static LLIST_HEAD(hpage_freelist);
> +
> +static void free_hpage_workfn(struct work_struct *work)
> +{
> + struct llist_node *node;
> +
> + node = llist_del_all(&hpage_freelist);
> +
> + while (node) {
> + struct page *page;
> + struct hstate *h;
> +
> + page = container_of((struct address_space **)node,
> + struct page, mapping);
> + node = node->next;
> + page->mapping = NULL;
> + h = page_hstate(page);
The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate is going to
trigger because a previous call to remove_hugetlb_page() will
set_compound_page_dtor(page, NULL_COMPOUND_DTOR)
Note how h(hstate) is grabbed before calling update_and_free_page in
existing code.
We could potentially drop the !PageHuge(page) in page_hstate. Or,
perhaps just use 'size_to_hstate(page_size(page))' in free_hpage_workfn.
--
Mike Kravetz
> +
> + __update_and_free_page(h, page);
> +
> + cond_resched();
> + }
> +}
> +static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
> +
> +static inline void flush_free_hpage_work(struct hstate *h)
> +{
> + if (free_vmemmap_pages_per_hpage(h))
> + flush_work(&free_hpage_work);
> +}
> +
> +static void update_and_free_page(struct hstate *h, struct page *page,
> + bool atomic)
> +{
> + if (!free_vmemmap_pages_per_hpage(h) || !atomic) {
> + __update_and_free_page(h, page);
> + return;
> + }
> +
> + /*
> + * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
> + *
> + * Only call schedule_work() if hpage_freelist is previously
> + * empty. Otherwise, schedule_work() had been called but the workfn
> + * hasn't retrieved the list yet.
> + */
> + if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
> + schedule_work(&free_hpage_work);
> +}
> +
> static void update_and_free_pages_bulk(struct hstate *h, struct list_head
> *list)
> {
> struct page *page, *t_page;
>
> list_for_each_entry_safe(page, t_page, list, lru) {
> - update_and_free_page(h, page);
> + update_and_free_page(h, page, false);
> cond_resched();
> }
> }
> @@ -1471,12 +1532,12 @@ void free_huge_page(struct page *page)
> if (HPageTemporary(page)) {
> remove_hugetlb_page(h, page, false);
> spin_unlock_irqrestore(&hugetlb_lock, flags);
> - update_and_free_page(h, page);
> + update_and_free_page(h, page, true);
> } else if (h->surplus_huge_pages_node[nid]) {
> /* remove the page from active list */
>
[PATCH v20 5/9] mm: hugetlb: defer freeing of HugeTLB pages
In the subsequent patch, we should allocate the vmemmap pages when
freeing a HugeTLB page. But update_and_free_page() can be called
under any context, so we cannot use GFP_KERNEL to allocate vmemmap
pages. However, we can defer the actual freeing in a kworker to
prevent from using GFP_ATOMIC to allocate the vmemmap pages.
The __update_and_free_page() is where the call to allocate vmemmmap
pages will be inserted.
Signed-off-by: Muchun Song
---
mm/hugetlb.c | 73
mm/hugetlb_vmemmap.c | 12 -
mm/hugetlb_vmemmap.h | 17
3 files changed, 85 insertions(+), 17 deletions(-)
diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 923d05e2806b..eeb8f5480170 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -1376,7 +1376,7 @@ static void remove_hugetlb_page(struct hstate *h, struct
page *page,
h->nr_huge_pages_node[nid]--;
}
-static void update_and_free_page(struct hstate *h, struct page *page)
+static void __update_and_free_page(struct hstate *h, struct page *page)
{
int i;
struct page *subpage = page;
@@ -1399,12 +1399,73 @@ static void update_and_free_page(struct hstate *h,
struct page *page)
}
}
+/*
+ * As update_and_free_page() can be called under any context, so we cannot
+ * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
+ * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
+ * the vmemmap pages.
+ *
+ * free_hpage_workfn() locklessly retrieves the linked list of pages to be
+ * freed and frees them one-by-one. As the page->mapping pointer is going
+ * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
+ * structure of a lockless linked list of huge pages to be freed.
+ */
+static LLIST_HEAD(hpage_freelist);
+
+static void free_hpage_workfn(struct work_struct *work)
+{
+ struct llist_node *node;
+
+ node = llist_del_all(&hpage_freelist);
+
+ while (node) {
+ struct page *page;
+ struct hstate *h;
+
+ page = container_of((struct address_space **)node,
+struct page, mapping);
+ node = node->next;
+ page->mapping = NULL;
+ h = page_hstate(page);
+
+ __update_and_free_page(h, page);
+
+ cond_resched();
+ }
+}
+static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
+
+static inline void flush_free_hpage_work(struct hstate *h)
+{
+ if (free_vmemmap_pages_per_hpage(h))
+ flush_work(&free_hpage_work);
+}
+
+static void update_and_free_page(struct hstate *h, struct page *page,
+bool atomic)
+{
+ if (!free_vmemmap_pages_per_hpage(h) || !atomic) {
+ __update_and_free_page(h, page);
+ return;
+ }
+
+ /*
+* Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
+*
+* Only call schedule_work() if hpage_freelist is previously
+* empty. Otherwise, schedule_work() had been called but the workfn
+* hasn't retrieved the list yet.
+*/
+ if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
+ schedule_work(&free_hpage_work);
+}
+
static void update_and_free_pages_bulk(struct hstate *h, struct list_head
*list)
{
struct page *page, *t_page;
list_for_each_entry_safe(page, t_page, list, lru) {
- update_and_free_page(h, page);
+ update_and_free_page(h, page, false);
cond_resched();
}
}
@@ -1471,12 +1532,12 @@ void free_huge_page(struct page *page)
if (HPageTemporary(page)) {
remove_hugetlb_page(h, page, false);
spin_unlock_irqrestore(&hugetlb_lock, flags);
- update_and_free_page(h, page);
+ update_and_free_page(h, page, true);
} else if (h->surplus_huge_pages_node[nid]) {
/* remove the page from active list */
remove_hugetlb_page(h, page, true);
spin_unlock_irqrestore(&hugetlb_lock, flags);
- update_and_free_page(h, page);
+ update_and_free_page(h, page, true);
} else {
arch_clear_hugepage_flags(page);
enqueue_huge_page(h, page);
@@ -1785,7 +1846,7 @@ int dissolve_free_huge_page(struct page *page)
remove_hugetlb_page(h, page, false);
h->max_huge_pages--;
spin_unlock_irq(&hugetlb_lock);
- update_and_free_page(h, head);
+ update_and_free_page(h, head, false);
return 0;
}
out:
@@ -2627,6 +2688,7 @@ static int set_max_huge_pages(struct hstate *h, unsigned
long count, int nid,
* pages in hstate via the proc/sysfs interfaces.
*/
mutex_lock(&h->resize_lock);
+ flush_free_hpage_work(h);
spin_lock_ir
