hulk inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I9K8D1 CVE: NA

--------------------------------

do_migrate_range() will firstly isolate pages and then migrate pages. If a pages is failed to isolate, it may not be successful in a short time even if tried many times. Since migrate_pages() has already tried 10 times to migrate the pages, there is also no need to migrate pages many times.

So just call do_migrate_range() once. This will reduce the log of migration failures.

Fixes: cdbeee51d044 ("mm/dynamic_hugetlb: add migration function") Signed-off-by: Liu Shixin liushixin2@huawei.com --- mm/dynamic_hugetlb.c | 33 +++++++++++++++++---------------- 1 file changed, 17 insertions(+), 16 deletions(-)

diff --git a/mm/dynamic_hugetlb.c b/mm/dynamic_hugetlb.c index 802a36081c1b..dc4cf48a332e 100644 --- a/mm/dynamic_hugetlb.c +++ b/mm/dynamic_hugetlb.c @@ -231,9 +231,6 @@ static void reclaim_pages_from_percpu_pool(struct dhugetlb_pool *hpool, } }

-/* We only try 5 times to reclaim pages */ -#define HPOOL_RECLAIM_RETRIES 5 - static int hpool_merge_page(struct dhugetlb_pool *hpool, int hpages_pool_idx, bool force_merge) { struct huge_pages_pool *hpages_pool, *src_hpages_pool; @@ -242,7 +239,8 @@ static int hpool_merge_page(struct dhugetlb_pool *hpool, int hpages_pool_idx, bo struct page *page, *next, *p; struct percpu_pages_pool *percpu_pool; bool need_migrate = false, need_initial = false; - int i, try; + bool tried_migrate, can_merge; + int i; LIST_HEAD(wait_page_list);

lockdep_assert_held(&hpool->lock); @@ -269,9 +267,11 @@ static int hpool_merge_page(struct dhugetlb_pool *hpool, int hpages_pool_idx, bo return -ENOMEM;

list_for_each_entry_safe(split_page, split_next, &hpages_pool->hugepage_splitlists, head_pages) { - try = 0; + tried_migrate = false;

merge: + can_merge = true; + /* * If we are merging 4K page to 2M page, we need to get * lock of percpu pool sequentially and clear percpu pool. @@ -290,8 +290,14 @@ static int hpool_merge_page(struct dhugetlb_pool *hpool, int hpages_pool_idx, bo for (i = 0; i < nr_pages; i+= block_size) { p = pfn_to_page(split_page->start_pfn + i); if (PagePool(p)) { - if (!need_migrate) - goto next; + /* + * Some pages still in use, can't merge. + * If don't need migration or have tried, + * then skip merging these pages. + */ + can_merge = false; + if (!need_migrate || tried_migrate) + break; else goto migrate; } @@ -303,6 +309,9 @@ static int hpool_merge_page(struct dhugetlb_pool *hpool, int hpages_pool_idx, bo */ dhugetlb_percpu_pool_unlock_all(hpool);

+ if (!can_merge) + continue; + list_del(&split_page->head_pages); hpages_pool->split_normal_pages--; for (i = 0; i < nr_pages; i+= block_size) { @@ -333,16 +342,8 @@ static int hpool_merge_page(struct dhugetlb_pool *hpool, int hpages_pool_idx, bo add_new_page_to_pool(hpool, page, hpages_pool_idx); trace_dynamic_hugetlb_split_merge(hpool, page, DHUGETLB_MERGE, page_size(page)); return 0; -next: - if (hpages_pool_idx == HUGE_PAGES_POOL_2M) - /* Unlock percpu pool before try next */ - dhugetlb_percpu_pool_unlock_all(hpool); - - continue; migrate: - /* page migration only used for HUGE_PAGES_POOL_2M */ - if (try++ >= HPOOL_RECLAIM_RETRIES) - goto next; + tried_migrate = true;

/* Isolate free page first. */ INIT_LIST_HEAD(&wait_page_list);

hulk inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I9K8D1 CVE: NA

--------------------------------

Since the pages in pagevec will failed to isolate, call lru_add_drain_all() before migration can increase the probability of migration success.

Signed-off-by: Liu Shixin liushixin2@huawei.com --- mm/dynamic_hugetlb.c | 2 ++ 1 file changed, 2 insertions(+)

diff --git a/mm/dynamic_hugetlb.c b/mm/dynamic_hugetlb.c index fd8a994ecf63..bf217db405fb 100644 --- a/mm/dynamic_hugetlb.c +++ b/mm/dynamic_hugetlb.c @@ -7,6 +7,7 @@ #include <linux/migrate.h> #include <linux/memblock.h> #include <linux/memory_hotplug.h> +#include <linux/swap.h> #include <linux/dynamic_hugetlb.h>

#include "internal.h" @@ -365,6 +366,7 @@ static int hpool_merge_page(struct dhugetlb_pool *hpool, int hpages_pool_idx, bo p = pfn_to_page(split_page->start_pfn + i); if (PagePool(p)) { cond_resched(); + lru_add_drain_all(); /* * TODO: fatal migration failures should bail * out

hulk inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I9K8D1 CVE: NA

--------------------------------

Like commit db71ef79b59b ("hugetlb: make free_huge_page irq safe") said, the spinlock may be interrupt by hugetlb freeing path, which will results deadlock. This patch fix such deadlock scenarios by changing spin_lock to spin_lock_irq in every path.

Fixes: a8a836a36072 ("mm/dynamic_hugetlb: establish the dynamic hugetlb feature framework") Signed-off-by: Liu Shixin liushixin2@huawei.com --- mm/dynamic_hugetlb.c | 33 ++++++++++++++++++--------------- mm/hugetlb.c | 4 ++-- 2 files changed, 20 insertions(+), 17 deletions(-)

diff --git a/mm/dynamic_hugetlb.c b/mm/dynamic_hugetlb.c index dbdd8546994e..c40d436b1d2b 100644 --- a/mm/dynamic_hugetlb.c +++ b/mm/dynamic_hugetlb.c @@ -30,7 +30,7 @@ static inline void dhugetlb_percpu_pool_lock_all(struct dhugetlb_pool *hpool) int i;

for (i = 0; i < NR_PERCPU_POOL; i++) - spin_lock(&hpool->percpu_pool[i].lock); + spin_lock_nested(&hpool->percpu_pool[i].lock, i); }

static inline void dhugetlb_percpu_pool_unlock_all(struct dhugetlb_pool *hpool) @@ -55,6 +55,7 @@ static inline void dhugetlb_percpu_pool_unlock_all(struct dhugetlb_pool *hpool) */ static inline void dhugetlb_lock_all(struct dhugetlb_pool *hpool) { + local_irq_disable(); dhugetlb_percpu_pool_lock_all(hpool); spin_lock(&hpool->lock); } @@ -65,6 +66,7 @@ static inline void dhugetlb_unlock_all(struct dhugetlb_pool *hpool)

spin_unlock(&hpool->lock); dhugetlb_percpu_pool_unlock_all(hpool); + local_irq_enable(); }

#define hugepage_index(pfn) ((pfn) >> (PUD_SHIFT - PAGE_SHIFT)) @@ -273,7 +275,7 @@ static int hpool_merge_page(struct dhugetlb_pool *hpool, int hpages_pool_idx, bo merge: can_merge = true;

- spin_unlock(&hpool->lock); + spin_unlock_irq(&hpool->lock); cond_resched(); /* * If we are merging 4K page to 2M page, we need to get @@ -287,7 +289,7 @@ static int hpool_merge_page(struct dhugetlb_pool *hpool, int hpages_pool_idx, bo percpu_pool->free_pages); } } else { - spin_lock(&hpool->lock); + spin_lock_irq(&hpool->lock); }

page = pfn_to_page(split_page->start_pfn); @@ -374,7 +376,7 @@ static int hpool_merge_page(struct dhugetlb_pool *hpool, int hpages_pool_idx, bo do_migrate_range(page_to_pfn(p), page_to_pfn(p) + block_size); } } - spin_lock(&hpool->lock); + spin_lock_irq(&hpool->lock);

/* * Move all isolate pages to src_hpages_pool and then try @@ -733,7 +735,7 @@ int dhugetlb_acct_memory(struct hstate *h, long delta, struct hugetlbfs_inode_in if (delta == 0) return 0;

- spin_lock(&hpool->lock); + spin_lock_irq(&hpool->lock); if (hstate_is_gigantic(h)) hpages_pool = &hpool->hpages_pool[HUGE_PAGES_POOL_1G]; else @@ -750,7 +752,7 @@ int dhugetlb_acct_memory(struct hstate *h, long delta, struct hugetlbfs_inode_in trace_dynamic_hugetlb_acct_memory(hpool, hpages_pool->resv_huge_pages, DHUGETLB_UNRESV, huge_page_size(h)); } - spin_unlock(&hpool->lock); + spin_unlock_irq(&hpool->lock);

return ret; } @@ -801,6 +803,7 @@ void free_huge_page_to_dhugetlb_pool(struct page *page, bool restore_reserve) struct hstate *h = page_hstate(page); struct huge_pages_pool *hpages_pool; struct dhugetlb_pool *hpool; + unsigned long flags;

hpool = find_hpool_by_dhugetlb_pagelist(page);

@@ -809,7 +812,7 @@ void free_huge_page_to_dhugetlb_pool(struct page *page, bool restore_reserve) return; }

- spin_lock(&hpool->lock); + spin_lock_irqsave(&hpool->lock, flags); /* * memory_failure will free the hwpoison hugepage, and then try to * dissolve it and free subpage to buddy system. Since the page in @@ -835,7 +838,7 @@ void free_huge_page_to_dhugetlb_pool(struct page *page, bool restore_reserve) trace_dynamic_hugetlb_alloc_free(hpool, page, hpages_pool->free_huge_pages, DHUGETLB_FREE, huge_page_size(h)); out: - spin_unlock(&hpool->lock); + spin_unlock_irqrestore(&hpool->lock, flags); put_hpool(hpool); }

@@ -851,7 +854,7 @@ static int alloc_hugepage_from_hugetlb(struct dhugetlb_pool *hpool, if (!h) return -ENOMEM;

- spin_lock(&hpool->lock); + spin_lock_irq(&hpool->lock); spin_lock(&hugetlb_lock); if (h->free_huge_pages_node[nid] - h->resv_huge_pages_node[nid] < nr_pages) { ret = -ENOMEM; @@ -877,7 +880,7 @@ static int alloc_hugepage_from_hugetlb(struct dhugetlb_pool *hpool,

out_unlock: spin_unlock(&hugetlb_lock); - spin_unlock(&hpool->lock); + spin_unlock_irq(&hpool->lock); return ret; }

@@ -983,15 +986,15 @@ int hugetlb_pool_destroy(struct cgroup *cgrp) */ mem_cgroup_force_empty(hpool->attach_memcg);

- spin_lock(&hpool->lock); + spin_lock_irq(&hpool->lock); ret = hugetlb_pool_merge_all_pages(hpool); if (ret) { - spin_unlock(&hpool->lock); + spin_unlock_irq(&hpool->lock); return -ENOMEM; } ret = free_hugepage_to_hugetlb(hpool); memcg->hpool = NULL; - spin_unlock(&hpool->lock); + spin_unlock_irq(&hpool->lock); put_hpool(hpool); return ret; } @@ -1058,7 +1061,7 @@ static ssize_t update_reserved_pages(struct mem_cgroup *memcg, char *buf, int hp return -EINVAL;

mutex_lock(&hpool->reserved_lock); - spin_lock(&hpool->lock); + spin_lock_irq(&hpool->lock); hpages_pool = &hpool->hpages_pool[hpages_pool_idx]; if (nr_pages > hpages_pool->nr_huge_pages) { delta = nr_pages - hpages_pool->nr_huge_pages; @@ -1092,7 +1095,7 @@ static ssize_t update_reserved_pages(struct mem_cgroup *memcg, char *buf, int hp hpages_pool->free_huge_pages -= delta; hpages_pool->free_normal_pages += delta; } - spin_unlock(&hpool->lock); + spin_unlock_irq(&hpool->lock); mutex_unlock(&hpool->reserved_lock); put_hpool(hpool); return 0; diff --git a/mm/hugetlb.c b/mm/hugetlb.c index a333caf66d15..e04c21f735ad 100644 --- a/mm/hugetlb.c +++ b/mm/hugetlb.c @@ -1597,14 +1597,14 @@ void free_huge_page(struct page *page) ClearHPageRestoreReserve(page);

if (dhugetlb_enabled && PagePool(page)) { - spin_lock(&hugetlb_lock); + spin_lock_irqsave(&hugetlb_lock, flags); ClearHPageMigratable(page); list_del(&page->lru); hugetlb_cgroup_uncharge_page(hstate_index(h), pages_per_huge_page(h), page); hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), pages_per_huge_page(h), page); - spin_unlock(&hugetlb_lock); + spin_unlock_irqrestore(&hugetlb_lock, flags); free_huge_page_to_dhugetlb_pool(page, restore_reserve); return; }