On Mon, 21 Mar 2022, Sebastian Andrzej Siewior wrote:
> The access to mlock_pvec is protected by disabling preemption via
> get_cpu_var() or implicit by having preemption disabled by the caller
> (in mlock_page_drain() case).
> This breaks on PREEMPT_RT since folio_lruvec_lock_irq() acquires a
> sleeping lock in this section.
>
> Create struct mlock_pvec which consits of the local_lock_t and the
> pagevec. Acquire the local_lock() before accessing the per-CPU pagevec.
> Replace mlock_page_drain() with a _local() version which is invoked on
> the local CPU and acquires the local_lock_t and a _remote() version
> which uses the pagevec from a remote CPU which offline.
>
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
Acked-by: Hugh Dickins <hughd@xxxxxxxxxx>
Thanks a lot for catching this: your patch looks exactly right to me.
Bad mistake on my part to have missed all the local_lock'ing in mm/swap.c
when I rebased my old mm/munlock implementation to the current tree.
> ---
> mm/internal.h | 6 ++++--
> mm/migrate.c | 2 +-
> mm/mlock.c | 46 ++++++++++++++++++++++++++++++++++++----------
> mm/page_alloc.c | 1 +
> mm/rmap.c | 4 ++--
> mm/swap.c | 4 +++-
> 6 files changed, 47 insertions(+), 16 deletions(-)
>
> diff --git a/mm/internal.h b/mm/internal.h
> index 9be0227ccc942..50c3fd71d7ddd 100644
> --- a/mm/internal.h
> +++ b/mm/internal.h
> @@ -456,7 +456,8 @@ static inline void munlock_vma_page(struct page *page,
> }
> void mlock_new_page(struct page *page);
> bool need_mlock_page_drain(int cpu);
> -void mlock_page_drain(int cpu);
> +void mlock_page_drain_local(void);
> +void mlock_page_drain_remote(int cpu);
>
> extern pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma);
>
> @@ -547,7 +548,8 @@ static inline void munlock_vma_page(struct page *page,
> struct vm_area_struct *vma, bool compound) { }
> static inline void mlock_new_page(struct page *page) { }
> static inline bool need_mlock_page_drain(int cpu) { return false; }
> -static inline void mlock_page_drain(int cpu) { }
> +static inline void mlock_page_drain_local(void) { }
> +static inline void mlock_page_drain_remote(int cpu) { }
> static inline void vunmap_range_noflush(unsigned long start, unsigned long end)
> {
> }
> diff --git a/mm/migrate.c b/mm/migrate.c
> index 3d60823afd2d3..de175e2fdba5d 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -246,7 +246,7 @@ static bool remove_migration_pte(struct folio *folio,
> set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
> }
> if (vma->vm_flags & VM_LOCKED)
> - mlock_page_drain(smp_processor_id());
> + mlock_page_drain_local();
>
> trace_remove_migration_pte(pvmw.address, pte_val(pte),
> compound_order(new));
> diff --git a/mm/mlock.c b/mm/mlock.c
> index efd2dd2943dec..6e503b8690f52 100644
> --- a/mm/mlock.c
> +++ b/mm/mlock.c
> @@ -28,7 +28,14 @@
>
> #include "internal.h"
>
> -static DEFINE_PER_CPU(struct pagevec, mlock_pvec);
> +struct mlock_pvec {
> + local_lock_t lock;
> + struct pagevec vec;
> +};
> +
> +static DEFINE_PER_CPU(struct mlock_pvec, mlock_pvec) = {
> + .lock = INIT_LOCAL_LOCK(lock),
> +};
>
> bool can_do_mlock(void)
> {
> @@ -203,18 +210,30 @@ static void mlock_pagevec(struct pagevec *pvec)
> pagevec_reinit(pvec);
> }
>
> -void mlock_page_drain(int cpu)
> +void mlock_page_drain_local(void)
> {
> struct pagevec *pvec;
>
> - pvec = &per_cpu(mlock_pvec, cpu);
> + local_lock(&mlock_pvec.lock);
> + pvec = this_cpu_ptr(&mlock_pvec.vec);
> + if (pagevec_count(pvec))
> + mlock_pagevec(pvec);
> + local_unlock(&mlock_pvec.lock);
> +}
> +
> +void mlock_page_drain_remote(int cpu)
> +{
> + struct pagevec *pvec;
> +
> + WARN_ON_ONCE(cpu_online(cpu));
> + pvec = &per_cpu(mlock_pvec.vec, cpu);
> if (pagevec_count(pvec))
> mlock_pagevec(pvec);
> }
>
> bool need_mlock_page_drain(int cpu)
> {
> - return pagevec_count(&per_cpu(mlock_pvec, cpu));
> + return pagevec_count(&per_cpu(mlock_pvec.vec, cpu));
> }
>
> /**
> @@ -223,7 +242,10 @@ bool need_mlock_page_drain(int cpu)
> */
> void mlock_folio(struct folio *folio)
> {
> - struct pagevec *pvec = &get_cpu_var(mlock_pvec);
> + struct pagevec *pvec;
> +
> + local_lock(&mlock_pvec.lock);
> + pvec = this_cpu_ptr(&mlock_pvec.vec);
>
> if (!folio_test_set_mlocked(folio)) {
> int nr_pages = folio_nr_pages(folio);
> @@ -236,7 +258,7 @@ void mlock_folio(struct folio *folio)
> if (!pagevec_add(pvec, mlock_lru(&folio->page)) ||
> folio_test_large(folio) || lru_cache_disabled())
> mlock_pagevec(pvec);
> - put_cpu_var(mlock_pvec);
> + local_unlock(&mlock_pvec.lock);
> }
>
> /**
> @@ -245,9 +267,11 @@ void mlock_folio(struct folio *folio)
> */
> void mlock_new_page(struct page *page)
> {
> - struct pagevec *pvec = &get_cpu_var(mlock_pvec);
> + struct pagevec *pvec;
> int nr_pages = thp_nr_pages(page);
>
> + local_lock(&mlock_pvec.lock);
> + pvec = this_cpu_ptr(&mlock_pvec.vec);
> SetPageMlocked(page);
> mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
> __count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
> @@ -256,7 +280,7 @@ void mlock_new_page(struct page *page)
> if (!pagevec_add(pvec, mlock_new(page)) ||
> PageHead(page) || lru_cache_disabled())
> mlock_pagevec(pvec);
> - put_cpu_var(mlock_pvec);
> + local_unlock(&mlock_pvec.lock);
> }
>
> /**
> @@ -265,8 +289,10 @@ void mlock_new_page(struct page *page)
> */
> void munlock_page(struct page *page)
> {
> - struct pagevec *pvec = &get_cpu_var(mlock_pvec);
> + struct pagevec *pvec;
>
> + local_lock(&mlock_pvec.lock);
> + pvec = this_cpu_ptr(&mlock_pvec.vec);
> /*
> * TestClearPageMlocked(page) must be left to __munlock_page(),
> * which will check whether the page is multiply mlocked.
> @@ -276,7 +302,7 @@ void munlock_page(struct page *page)
> if (!pagevec_add(pvec, page) ||
> PageHead(page) || lru_cache_disabled())
> mlock_pagevec(pvec);
> - put_cpu_var(mlock_pvec);
> + local_unlock(&mlock_pvec.lock);
> }
>
> static int mlock_pte_range(pmd_t *pmd, unsigned long addr,
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index f648decfe39d1..94f515845d53a 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -8355,6 +8355,7 @@ static int page_alloc_cpu_dead(unsigned int cpu)
> struct zone *zone;
>
> lru_add_drain_cpu(cpu);
> + mlock_page_drain_remote(cpu);
> drain_pages(cpu);
>
> /*
> diff --git a/mm/rmap.c b/mm/rmap.c
> index 3eb95fcde7000..90b4956096ef3 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -1727,7 +1727,7 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
> */
> page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
> if (vma->vm_flags & VM_LOCKED)
> - mlock_page_drain(smp_processor_id());
> + mlock_page_drain_local();
> folio_put(folio);
> }
>
> @@ -2005,7 +2005,7 @@ static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
> */
> page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
> if (vma->vm_flags & VM_LOCKED)
> - mlock_page_drain(smp_processor_id());
> + mlock_page_drain_local();
> folio_put(folio);
> }
>
> diff --git a/mm/swap.c b/mm/swap.c
> index bceff0cb559c9..7e320ec08c6ae 100644
> --- a/mm/swap.c
> +++ b/mm/swap.c
> @@ -624,7 +624,6 @@ void lru_add_drain_cpu(int cpu)
> pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
>
> activate_page_drain(cpu);
> - mlock_page_drain(cpu);
> }
>
> /**
> @@ -706,6 +705,7 @@ void lru_add_drain(void)
> local_lock(&lru_pvecs.lock);
> lru_add_drain_cpu(smp_processor_id());
> local_unlock(&lru_pvecs.lock);
> + mlock_page_drain_local();
> }
>
> /*
> @@ -720,6 +720,7 @@ static void lru_add_and_bh_lrus_drain(void)
> lru_add_drain_cpu(smp_processor_id());
> local_unlock(&lru_pvecs.lock);
> invalidate_bh_lrus_cpu();
> + mlock_page_drain_local();
> }
>
> void lru_add_drain_cpu_zone(struct zone *zone)
> @@ -728,6 +729,7 @@ void lru_add_drain_cpu_zone(struct zone *zone)
> lru_add_drain_cpu(smp_processor_id());
> drain_local_pages(zone);
> local_unlock(&lru_pvecs.lock);
> + mlock_page_drain_local();
> }
>
> #ifdef CONFIG_SMP
> --
> 2.35.1
>
>
