On 4/15/21 3:35 AM, Oscar Salvador wrote: > alloc_contig_range will fail if it ever sees a HugeTLB page within the > range we are trying to allocate, even when that page is free and can be > easily reallocated. > This has proved to be problematic for some users of alloc_contic_range, > e.g: CMA and virtio-mem, where those would fail the call even when those > pages lay in ZONE_MOVABLE and are free. > > We can do better by trying to replace such page. > > Free hugepages are tricky to handle so as to no userspace application > notices disruption, we need to replace the current free hugepage with > a new one. > > In order to do that, a new function called alloc_and_dissolve_huge_page > is introduced. > This function will first try to get a new fresh hugepage, and if it > succeeds, it will replace the old one in the free hugepage pool. > > The free page replacement is done under hugetlb_lock, so no external > users of hugetlb will notice the change. > To allocate the new huge page, we use alloc_buddy_huge_page(), so we > do not have to deal with any counters, and prep_new_huge_page() is not > called. This is valulable because in case we need to free the new page, > we only need to call __free_pages(). > > Once we know that the page to be replaced is a genuine 0-refcounted > huge page, we remove the old page from the freelist by remove_hugetlb_page(). > Then, we can call __prep_new_huge_page() and __prep_account_new_huge_page() > for the new huge page to properly initialize it and increment the > hstate->nr_huge_pages counter (previously decremented by > remove_hugetlb_page()). > Once done, the page is enqueued by enqueue_huge_page() and it is ready > to be used. > > There is one tricky case when > page's refcount is 0 because it is in the process of being released. > A missing PageHugeFreed bit will tell us that freeing is in flight so > we retry after dropping the hugetlb_lock. The race window should be > small and the next retry should make a forward progress. > > E.g: > > CPU0 CPU1 > free_huge_page() isolate_or_dissolve_huge_page > PageHuge() == T > alloc_and_dissolve_huge_page > alloc_buddy_huge_page() > spin_lock_irq(hugetlb_lock) > // PageHuge() && !PageHugeFreed && > // !PageCount() > spin_unlock_irq(hugetlb_lock) > spin_lock_irq(hugetlb_lock) > 1) update_and_free_page > PageHuge() == F > __free_pages() > 2) enqueue_huge_page > SetPageHugeFreed() > spin_unlock_irq(&hugetlb_lock) > spin_lock_irq(hugetlb_lock) > 1) PageHuge() == F (freed by case#1 from > CPU0) > 2) PageHuge() == T > PageHugeFreed() == T > - proceed with replacing the page > > In the case above we retry as the window race is quite small and we have high > chances to succeed next time. > > With regard to the allocation, we restrict it to the node the page belongs > to with __GFP_THISNODE, meaning we do not fallback on other node's zones. > > Note that gigantic hugetlb pages are fenced off since there is a cyclic > dependency between them and alloc_contig_range. > > Signed-off-by: Oscar Salvador <osalva...@suse.de> > Acked-by: Michal Hocko <mho...@suse.com>
Reviewed-by: Mike Kravetz <mike.krav...@oracle.com> -- Mike Kravetz
