hugetlb: make free_huge_page irq safe
authorMike Kravetz <mike.kravetz@oracle.com>
Wed, 5 May 2021 01:35:07 +0000 (18:35 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 5 May 2021 18:27:22 +0000 (11:27 -0700)
Commit c77c0a8ac4c5 ("mm/hugetlb: defer freeing of huge pages if in
non-task context") was added to address the issue of free_huge_page being
called from irq context.  That commit hands off free_huge_page processing
to a workqueue if !in_task.  However, this doesn't cover all the cases as
pointed out by 0day bot lockdep report [1].

:  Possible interrupt unsafe locking scenario:
:
:        CPU0                    CPU1
:        ----                    ----
:   lock(hugetlb_lock);
:                                local_irq_disable();
:                                lock(slock-AF_INET);
:                                lock(hugetlb_lock);
:   <Interrupt>
:     lock(slock-AF_INET);

Shakeel has later explained that this is very likely TCP TX zerocopy from
hugetlb pages scenario when the networking code drops a last reference to
hugetlb page while having IRQ disabled.  Hugetlb freeing path doesn't
disable IRQ while holding hugetlb_lock so a lock dependency chain can lead
to a deadlock.

This commit addresses the issue by doing the following:
 - Make hugetlb_lock irq safe. This is mostly a simple process of
   changing spin_*lock calls to spin_*lock_irq* calls.
 - Make subpool lock irq safe in a similar manner.
 - Revert the !in_task check and workqueue handoff.

[1] https://lore.kernel.org/linux-mm/000000000000f1c03b05bc43aadc@google.com/

Link: https://lkml.kernel.org/r/20210409205254.242291-8-mike.kravetz@oracle.com
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Muchun Song <songmuchun@bytedance.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Cc: "Aneesh Kumar K . V" <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <song.bao.hua@hisilicon.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: HORIGUCHI NAOYA <naoya.horiguchi@nec.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Waiman Long <longman@redhat.com>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/hugetlb.c
mm/hugetlb_cgroup.c

index d7db93639787f3a86a434db019a12d0dab79911d..47f56e2cc8041821458c658dea6c24103efe15a0 100644 (file)
@@ -94,9 +94,10 @@ static inline bool subpool_is_free(struct hugepage_subpool *spool)
        return true;
 }
 
-static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
+static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
+                                               unsigned long irq_flags)
 {
-       spin_unlock(&spool->lock);
+       spin_unlock_irqrestore(&spool->lock, irq_flags);
 
        /* If no pages are used, and no other handles to the subpool
         * remain, give up any reservations based on minimum size and
@@ -135,10 +136,12 @@ struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
 
 void hugepage_put_subpool(struct hugepage_subpool *spool)
 {
-       spin_lock(&spool->lock);
+       unsigned long flags;
+
+       spin_lock_irqsave(&spool->lock, flags);
        BUG_ON(!spool->count);
        spool->count--;
-       unlock_or_release_subpool(spool);
+       unlock_or_release_subpool(spool, flags);
 }
 
 /*
@@ -157,7 +160,7 @@ static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
        if (!spool)
                return ret;
 
-       spin_lock(&spool->lock);
+       spin_lock_irq(&spool->lock);
 
        if (spool->max_hpages != -1) {          /* maximum size accounting */
                if ((spool->used_hpages + delta) <= spool->max_hpages)
@@ -184,7 +187,7 @@ static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
        }
 
 unlock_ret:
-       spin_unlock(&spool->lock);
+       spin_unlock_irq(&spool->lock);
        return ret;
 }
 
@@ -198,11 +201,12 @@ static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
                                       long delta)
 {
        long ret = delta;
+       unsigned long flags;
 
        if (!spool)
                return delta;
 
-       spin_lock(&spool->lock);
+       spin_lock_irqsave(&spool->lock, flags);
 
        if (spool->max_hpages != -1)            /* maximum size accounting */
                spool->used_hpages -= delta;
@@ -223,7 +227,7 @@ static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
         * If hugetlbfs_put_super couldn't free spool due to an outstanding
         * quota reference, free it now.
         */
-       unlock_or_release_subpool(spool);
+       unlock_or_release_subpool(spool, flags);
 
        return ret;
 }
@@ -1412,7 +1416,7 @@ struct hstate *size_to_hstate(unsigned long size)
        return NULL;
 }
 
-static void __free_huge_page(struct page *page)
+void free_huge_page(struct page *page)
 {
        /*
         * Can't pass hstate in here because it is called from the
@@ -1422,6 +1426,7 @@ static void __free_huge_page(struct page *page)
        int nid = page_to_nid(page);
        struct hugepage_subpool *spool = hugetlb_page_subpool(page);
        bool restore_reserve;
+       unsigned long flags;
 
        VM_BUG_ON_PAGE(page_count(page), page);
        VM_BUG_ON_PAGE(page_mapcount(page), page);
@@ -1450,7 +1455,7 @@ static void __free_huge_page(struct page *page)
                        restore_reserve = true;
        }
 
-       spin_lock(&hugetlb_lock);
+       spin_lock_irqsave(&hugetlb_lock, flags);
        ClearHPageMigratable(page);
        hugetlb_cgroup_uncharge_page(hstate_index(h),
                                     pages_per_huge_page(h), page);
@@ -1461,66 +1466,18 @@ static void __free_huge_page(struct page *page)
 
        if (HPageTemporary(page)) {
                remove_hugetlb_page(h, page, false);
-               spin_unlock(&hugetlb_lock);
+               spin_unlock_irqrestore(&hugetlb_lock, flags);
                update_and_free_page(h, page);
        } else if (h->surplus_huge_pages_node[nid]) {
                /* remove the page from active list */
                remove_hugetlb_page(h, page, true);
-               spin_unlock(&hugetlb_lock);
+               spin_unlock_irqrestore(&hugetlb_lock, flags);
                update_and_free_page(h, page);
        } else {
                arch_clear_hugepage_flags(page);
                enqueue_huge_page(h, page);
-               spin_unlock(&hugetlb_lock);
-       }
-}
-
-/*
- * As free_huge_page() can be called from a non-task context, we have
- * to defer the actual freeing in a workqueue to prevent potential
- * hugetlb_lock deadlock.
- *
- * 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_huge_page() 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;
-       struct page *page;
-
-       node = llist_del_all(&hpage_freelist);
-
-       while (node) {
-               page = container_of((struct address_space **)node,
-                                    struct page, mapping);
-               node = node->next;
-               __free_huge_page(page);
-       }
-}
-static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
-
-void free_huge_page(struct page *page)
-{
-       /*
-        * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
-        */
-       if (!in_task()) {
-               /*
-                * 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);
-               return;
+               spin_unlock_irqrestore(&hugetlb_lock, flags);
        }
-
-       __free_huge_page(page);
 }
 
 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
@@ -1530,11 +1487,11 @@ static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
        hugetlb_set_page_subpool(page, NULL);
        set_hugetlb_cgroup(page, NULL);
        set_hugetlb_cgroup_rsvd(page, NULL);
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        h->nr_huge_pages++;
        h->nr_huge_pages_node[nid]++;
        ClearHPageFreed(page);
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
 }
 
 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
@@ -1780,7 +1737,7 @@ retry:
        if (!PageHuge(page))
                return 0;
 
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        if (!PageHuge(page)) {
                rc = 0;
                goto out;
@@ -1797,7 +1754,7 @@ retry:
                 * when it is dissolved.
                 */
                if (unlikely(!HPageFreed(head))) {
-                       spin_unlock(&hugetlb_lock);
+                       spin_unlock_irq(&hugetlb_lock);
                        cond_resched();
 
                        /*
@@ -1821,12 +1778,12 @@ retry:
                }
                remove_hugetlb_page(h, page, false);
                h->max_huge_pages--;
-               spin_unlock(&hugetlb_lock);
+               spin_unlock_irq(&hugetlb_lock);
                update_and_free_page(h, head);
                return 0;
        }
 out:
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
        return rc;
 }
 
@@ -1868,16 +1825,16 @@ static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
        if (hstate_is_gigantic(h))
                return NULL;
 
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
                goto out_unlock;
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
 
        page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
        if (!page)
                return NULL;
 
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        /*
         * We could have raced with the pool size change.
         * Double check that and simply deallocate the new page
@@ -1887,7 +1844,7 @@ static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
         */
        if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
                SetHPageTemporary(page);
-               spin_unlock(&hugetlb_lock);
+               spin_unlock_irq(&hugetlb_lock);
                put_page(page);
                return NULL;
        } else {
@@ -1896,7 +1853,7 @@ static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
        }
 
 out_unlock:
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
 
        return page;
 }
@@ -1946,17 +1903,17 @@ struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
                nodemask_t *nmask, gfp_t gfp_mask)
 {
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        if (h->free_huge_pages - h->resv_huge_pages > 0) {
                struct page *page;
 
                page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
                if (page) {
-                       spin_unlock(&hugetlb_lock);
+                       spin_unlock_irq(&hugetlb_lock);
                        return page;
                }
        }
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
 
        return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
 }
@@ -2004,7 +1961,7 @@ static int gather_surplus_pages(struct hstate *h, long delta)
 
        ret = -ENOMEM;
 retry:
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
        for (i = 0; i < needed; i++) {
                page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
                                NUMA_NO_NODE, NULL);
@@ -2021,7 +1978,7 @@ retry:
         * After retaking hugetlb_lock, we need to recalculate 'needed'
         * because either resv_huge_pages or free_huge_pages may have changed.
         */
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        needed = (h->resv_huge_pages + delta) -
                        (h->free_huge_pages + allocated);
        if (needed > 0) {
@@ -2061,12 +2018,12 @@ retry:
                enqueue_huge_page(h, page);
        }
 free:
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
 
        /* Free unnecessary surplus pages to the buddy allocator */
        list_for_each_entry_safe(page, tmp, &surplus_list, lru)
                put_page(page);
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
 
        return ret;
 }
@@ -2116,9 +2073,9 @@ static void return_unused_surplus_pages(struct hstate *h,
        }
 
 out:
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
        update_and_free_pages_bulk(h, &page_list);
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
 }
 
 
@@ -2352,7 +2309,7 @@ struct page *alloc_huge_page(struct vm_area_struct *vma,
        if (ret)
                goto out_uncharge_cgroup_reservation;
 
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        /*
         * glb_chg is passed to indicate whether or not a page must be taken
         * from the global free pool (global change).  gbl_chg == 0 indicates
@@ -2360,7 +2317,7 @@ struct page *alloc_huge_page(struct vm_area_struct *vma,
         */
        page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
        if (!page) {
-               spin_unlock(&hugetlb_lock);
+               spin_unlock_irq(&hugetlb_lock);
                page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
                if (!page)
                        goto out_uncharge_cgroup;
@@ -2368,7 +2325,7 @@ struct page *alloc_huge_page(struct vm_area_struct *vma,
                        SetHPageRestoreReserve(page);
                        h->resv_huge_pages--;
                }
-               spin_lock(&hugetlb_lock);
+               spin_lock_irq(&hugetlb_lock);
                list_add(&page->lru, &h->hugepage_activelist);
                /* Fall through */
        }
@@ -2381,7 +2338,7 @@ struct page *alloc_huge_page(struct vm_area_struct *vma,
                                                  h_cg, page);
        }
 
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
 
        hugetlb_set_page_subpool(page, spool);
 
@@ -2593,9 +2550,9 @@ static void try_to_free_low(struct hstate *h, unsigned long count,
        }
 
 out:
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
        update_and_free_pages_bulk(h, &page_list);
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
 }
 #else
 static inline void try_to_free_low(struct hstate *h, unsigned long count,
@@ -2660,7 +2617,7 @@ 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);
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
 
        /*
         * Check for a node specific request.
@@ -2691,7 +2648,7 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
         */
        if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
                if (count > persistent_huge_pages(h)) {
-                       spin_unlock(&hugetlb_lock);
+                       spin_unlock_irq(&hugetlb_lock);
                        mutex_unlock(&h->resize_lock);
                        NODEMASK_FREE(node_alloc_noretry);
                        return -EINVAL;
@@ -2721,14 +2678,14 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
                 * page, free_huge_page will handle it by freeing the page
                 * and reducing the surplus.
                 */
-               spin_unlock(&hugetlb_lock);
+               spin_unlock_irq(&hugetlb_lock);
 
                /* yield cpu to avoid soft lockup */
                cond_resched();
 
                ret = alloc_pool_huge_page(h, nodes_allowed,
                                                node_alloc_noretry);
-               spin_lock(&hugetlb_lock);
+               spin_lock_irq(&hugetlb_lock);
                if (!ret)
                        goto out;
 
@@ -2767,9 +2724,9 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
                list_add(&page->lru, &page_list);
        }
        /* free the pages after dropping lock */
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
        update_and_free_pages_bulk(h, &page_list);
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
 
        while (count < persistent_huge_pages(h)) {
                if (!adjust_pool_surplus(h, nodes_allowed, 1))
@@ -2777,7 +2734,7 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
        }
 out:
        h->max_huge_pages = persistent_huge_pages(h);
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
        mutex_unlock(&h->resize_lock);
 
        NODEMASK_FREE(node_alloc_noretry);
@@ -2933,9 +2890,9 @@ static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
        if (err)
                return err;
 
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        h->nr_overcommit_huge_pages = input;
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
 
        return count;
 }
@@ -3522,9 +3479,9 @@ int hugetlb_overcommit_handler(struct ctl_table *table, int write,
                goto out;
 
        if (write) {
-               spin_lock(&hugetlb_lock);
+               spin_lock_irq(&hugetlb_lock);
                h->nr_overcommit_huge_pages = tmp;
-               spin_unlock(&hugetlb_lock);
+               spin_unlock_irq(&hugetlb_lock);
        }
 out:
        return ret;
@@ -3620,7 +3577,7 @@ static int hugetlb_acct_memory(struct hstate *h, long delta)
        if (!delta)
                return 0;
 
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        /*
         * When cpuset is configured, it breaks the strict hugetlb page
         * reservation as the accounting is done on a global variable. Such
@@ -3659,7 +3616,7 @@ static int hugetlb_acct_memory(struct hstate *h, long delta)
                return_unused_surplus_pages(h, (unsigned long) -delta);
 
 out:
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
        return ret;
 }
 
@@ -5687,7 +5644,7 @@ bool isolate_huge_page(struct page *page, struct list_head *list)
 {
        bool ret = true;
 
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        if (!PageHeadHuge(page) ||
            !HPageMigratable(page) ||
            !get_page_unless_zero(page)) {
@@ -5697,16 +5654,16 @@ bool isolate_huge_page(struct page *page, struct list_head *list)
        ClearHPageMigratable(page);
        list_move_tail(&page->lru, list);
 unlock:
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
        return ret;
 }
 
 void putback_active_hugepage(struct page *page)
 {
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        SetHPageMigratable(page);
        list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
        put_page(page);
 }
 
@@ -5740,12 +5697,12 @@ void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
                 */
                if (new_nid == old_nid)
                        return;
-               spin_lock(&hugetlb_lock);
+               spin_lock_irq(&hugetlb_lock);
                if (h->surplus_huge_pages_node[old_nid]) {
                        h->surplus_huge_pages_node[old_nid]--;
                        h->surplus_huge_pages_node[new_nid]++;
                }
-               spin_unlock(&hugetlb_lock);
+               spin_unlock_irq(&hugetlb_lock);
        }
 }
 
index 726b85f4f30382b43e54c50813c50c27b5e483d1..5383023d0ccadde9424e4aa57f957ee5a817d0cb 100644 (file)
@@ -204,11 +204,11 @@ static void hugetlb_cgroup_css_offline(struct cgroup_subsys_state *css)
        do {
                idx = 0;
                for_each_hstate(h) {
-                       spin_lock(&hugetlb_lock);
+                       spin_lock_irq(&hugetlb_lock);
                        list_for_each_entry(page, &h->hugepage_activelist, lru)
                                hugetlb_cgroup_move_parent(idx, h_cg, page);
 
-                       spin_unlock(&hugetlb_lock);
+                       spin_unlock_irq(&hugetlb_lock);
                        idx++;
                }
                cond_resched();
@@ -784,7 +784,7 @@ void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage)
        if (hugetlb_cgroup_disabled())
                return;
 
-       spin_lock(&hugetlb_lock);
+       spin_lock_irq(&hugetlb_lock);
        h_cg = hugetlb_cgroup_from_page(oldhpage);
        h_cg_rsvd = hugetlb_cgroup_from_page_rsvd(oldhpage);
        set_hugetlb_cgroup(oldhpage, NULL);
@@ -794,7 +794,7 @@ void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage)
        set_hugetlb_cgroup(newhpage, h_cg);
        set_hugetlb_cgroup_rsvd(newhpage, h_cg_rsvd);
        list_move(&newhpage->lru, &h->hugepage_activelist);
-       spin_unlock(&hugetlb_lock);
+       spin_unlock_irq(&hugetlb_lock);
        return;
 }