On Sun, 17 Jan 2021, Muchun Song wrote:
> In the subsequent patch, we should allocate the vmemmap pages when
> freeing HugeTLB pages. But update_and_free_page() is always called
> with holding hugetlb_lock, 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_hpage_vmemmap_workfn() is where the call to allocate
> vmemmmap pages will be inserted.
>
I think it's reasonable to assume that userspace can release free hugetlb
pages from the pool on oom conditions when reclaim has become too
expensive. This approach now requires that we can allocate vmemmap pages
in a potential oom condition as a prerequisite for freeing memory, which
seems less than ideal.
And, by doing this through a kworker, we can presumably get queued behind
another work item that requires memory to make forward progress in this
oom condition.
Two thoughts:
- We're going to be freeing the hugetlb page after we can allocate the
vmemmap pages, so why do we need to allocate with GFP_KERNEL? Can't we
simply dip into memory reserves using GFP_ATOMIC (and thus can be
holding hugetlb_lock) because we know we'll be freeing more memory than
we'll be allocating? I think requiring a GFP_KERNEL allocation to block
to free memory for vmemmap when we'll be freeing memory ourselves is
dubious. This simplifies all of this.
- If the answer is that we actually have to use GFP_KERNEL for other
reasons, what are your thoughts on pre-allocating the vmemmap as opposed
to deferring to a kworker? In other words, preallocate the necessary
memory with GFP_KERNEL and put it on a linked list in struct hstate
before acquiring hugetlb_lock.
> Signed-off-by: Muchun Song <songmuchun@xxxxxxxxxxxxx>
> Reviewed-by: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
> ---
> mm/hugetlb.c | 74 ++++++++++++++++++++++++++++++++++++++++++++++++++--
> mm/hugetlb_vmemmap.c | 12 ---------
> mm/hugetlb_vmemmap.h | 17 ++++++++++++
> 3 files changed, 89 insertions(+), 14 deletions(-)
>
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index 140135fc8113..c165186ec2cf 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -1292,15 +1292,85 @@ static inline void destroy_compound_gigantic_page(struct page *page,
> unsigned int order) { }
> #endif
>
> -static void update_and_free_page(struct hstate *h, struct page *page)
> +static void __free_hugepage(struct hstate *h, struct page *page);
> +
> +/*
> + * As update_and_free_page() is always called with holding hugetlb_lock, 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.
> + *
> + * The update_hpage_vmemmap_workfn() is where the call to allocate vmemmmap
> + * pages will be inserted.
> + *
> + * update_hpage_vmemmap_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 update_hpage_vmemmap_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_update_freelist);
> +
> +static void update_hpage_vmemmap_workfn(struct work_struct *work)
> {
> - int i;
> + struct llist_node *node;
> +
> + node = llist_del_all(&hpage_update_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);
> +
> + spin_lock(&hugetlb_lock);
> + __free_hugepage(h, page);
> + spin_unlock(&hugetlb_lock);
>
> + cond_resched();
Wouldn't it be better to hold hugetlb_lock for the iteration rather than
constantly dropping it and reacquiring it? Use
cond_resched_lock(&hugetlb_lock) instead?
