Re: [PATCH v2] fs/dcache.c: fix spin lockup issue on nlru->lock
On Thu, Jun 22, 2017 at 10:01:39PM +0530, Sahitya Tummala wrote: > > > On 6/21/2017 10:01 PM, Vladimir Davydov wrote: > > > >>index cddf397..c8ca150 100644 > >>--- a/fs/dcache.c > >>+++ b/fs/dcache.c > >>@@ -1133,10 +1133,11 @@ void shrink_dcache_sb(struct super_block *sb) > >>LIST_HEAD(dispose); > >>freed = list_lru_walk(&sb->s_dentry_lru, > >>- dentry_lru_isolate_shrink, &dispose, UINT_MAX); > >>+ dentry_lru_isolate_shrink, &dispose, 1024); > >>this_cpu_sub(nr_dentry_unused, freed); > >>shrink_dentry_list(&dispose); > >>+ cond_resched(); > >>} while (freed > 0); > >In an extreme case, a single invocation of list_lru_walk() can skip all > >1024 dentries, in which case 'freed' will be 0 forcing us to break the > >loop prematurely. I think we should loop until there's at least one > >dentry left on the LRU, i.e. > > > > while (list_lru_count(&sb->s_dentry_lru) > 0) > > > >However, even that wouldn't be quite correct, because list_lru_count() > >iterates over all memory cgroups to sum list_lru_one->nr_items, which > >can race with memcg offlining code migrating dentries off a dead cgroup > >(see memcg_drain_all_list_lrus()). So it looks like to make this check > >race-free, we need to account the number of entries on the LRU not only > >per memcg, but also per node, i.e. add list_lru_node->nr_items. > >Fortunately, list_lru entries can't be migrated between NUMA nodes. > It looks like list_lru_count() is iterating per node before iterating over > all memory > cgroups as below - > > unsigned long list_lru_count_node(struct list_lru *lru, int nid) > { > long count = 0; > int memcg_idx; > > count += __list_lru_count_one(lru, nid, -1); > if (list_lru_memcg_aware(lru)) { > for_each_memcg_cache_index(memcg_idx) > count += __list_lru_count_one(lru, nid, memcg_idx); > } > return count; > } > > The first call to __list_lru_count_one() is iterating all the items per node > i.e, nlru->lru->nr_items. lru->node[nid].lru.nr_items returned by __list_lru_count_one(lru, nid, -1) only counts items accounted to the root cgroup, not the total number of entries on the node. > Is my understanding correct? If not, could you please clarify on how to get > the lru items per node? What I mean is iterating over list_lru_node->memcg_lrus to count the number of entries on the node is racy. For example, suppose you have three cgroups with the following values of list_lru_one->nr_items: 0 0 10 While list_lru_count_node() is at #1, cgroup #2 is offlined and its list_lru_one is drained, i.e. its entries are migrated to the parent cgroup, which happens to be #0, i.e. we see the following picture: 10 0 0 ^^^ memcg_ids points here in list_lru_count_node() Then the count returned by list_lru_count_node() will be 0, although there are still 10 entries on the list. To avoid this race, we could keep list_lru_node->lock locked while walking over list_lru_node->memcg_lrus, but that's too heavy. I'd prefer adding list_lru_node->nr_count which would be equal to the total number of list_lru entries on the node, i.e. sum of list_lru_node->lru.nr_lrus and list_lru_node->memcg_lrus->lru[]->nr_items.
Re: [PATCH v2] fs/dcache.c: fix spin lockup issue on nlru->lock
On 6/21/2017 10:01 PM, Vladimir Davydov wrote: index cddf397..c8ca150 100644 --- a/fs/dcache.c +++ b/fs/dcache.c @@ -1133,10 +1133,11 @@ void shrink_dcache_sb(struct super_block *sb) LIST_HEAD(dispose); freed = list_lru_walk(&sb->s_dentry_lru, - dentry_lru_isolate_shrink, &dispose, UINT_MAX); + dentry_lru_isolate_shrink, &dispose, 1024); this_cpu_sub(nr_dentry_unused, freed); shrink_dentry_list(&dispose); + cond_resched(); } while (freed > 0); In an extreme case, a single invocation of list_lru_walk() can skip all 1024 dentries, in which case 'freed' will be 0 forcing us to break the loop prematurely. I think we should loop until there's at least one dentry left on the LRU, i.e. while (list_lru_count(&sb->s_dentry_lru) > 0) However, even that wouldn't be quite correct, because list_lru_count() iterates over all memory cgroups to sum list_lru_one->nr_items, which can race with memcg offlining code migrating dentries off a dead cgroup (see memcg_drain_all_list_lrus()). So it looks like to make this check race-free, we need to account the number of entries on the LRU not only per memcg, but also per node, i.e. add list_lru_node->nr_items. Fortunately, list_lru entries can't be migrated between NUMA nodes. It looks like list_lru_count() is iterating per node before iterating over all memory cgroups as below - unsigned long list_lru_count_node(struct list_lru *lru, int nid) { long count = 0; int memcg_idx; count += __list_lru_count_one(lru, nid, -1); if (list_lru_memcg_aware(lru)) { for_each_memcg_cache_index(memcg_idx) count += __list_lru_count_one(lru, nid, memcg_idx); } return count; } The first call to __list_lru_count_one() is iterating all the items per node i.e, nlru->lru->nr_items. Is my understanding correct? If not, could you please clarify on how to get the lru items per node? -- Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center, Inc. Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.
Re: [PATCH v2] fs/dcache.c: fix spin lockup issue on nlru->lock
On Wed, Jun 21, 2017 at 12:09:15PM +0530, Sahitya Tummala wrote: > __list_lru_walk_one() acquires nlru spin lock (nlru->lock) for > longer duration if there are more number of items in the lru list. > As per the current code, it can hold the spin lock for upto maximum > UINT_MAX entries at a time. So if there are more number of items in > the lru list, then "BUG: spinlock lockup suspected" is observed in > the below path - > > [] spin_bug+0x90 > [] do_raw_spin_lock+0xfc > [] _raw_spin_lock+0x28 > [] list_lru_add+0x28 > [] dput+0x1c8 > [] path_put+0x20 > [] terminate_walk+0x3c > [] path_lookupat+0x100 > [] filename_lookup+0x6c > [] user_path_at_empty+0x54 > [] SyS_faccessat+0xd0 > [] el0_svc_naked+0x24 > > This nlru->lock is acquired by another CPU in this path - > > [] d_lru_shrink_move+0x34 > [] dentry_lru_isolate_shrink+0x48 > [] __list_lru_walk_one.isra.10+0x94 > [] list_lru_walk_node+0x40 > [] shrink_dcache_sb+0x60 > [] do_remount_sb+0xbc > [] do_emergency_remount+0xb0 > [] process_one_work+0x228 > [] worker_thread+0x2e0 > [] kthread+0xf4 > [] ret_from_fork+0x10 > > Fix this lockup by reducing the number of entries to be shrinked > from the lru list to 1024 at once. Also, add cond_resched() before > processing the lru list again. > > Link: http://marc.info/?t=14972286491&r=1&w=2 > Fix-suggested-by: Jan kara > Fix-suggested-by: Vladimir Davydov > Signed-off-by: Sahitya Tummala > --- > v2: patch shrink_dcache_sb() instead of list_lru_walk() > --- > fs/dcache.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/fs/dcache.c b/fs/dcache.c > index cddf397..c8ca150 100644 > --- a/fs/dcache.c > +++ b/fs/dcache.c > @@ -1133,10 +1133,11 @@ void shrink_dcache_sb(struct super_block *sb) > LIST_HEAD(dispose); > > freed = list_lru_walk(&sb->s_dentry_lru, > - dentry_lru_isolate_shrink, &dispose, UINT_MAX); > + dentry_lru_isolate_shrink, &dispose, 1024); > > this_cpu_sub(nr_dentry_unused, freed); > shrink_dentry_list(&dispose); > + cond_resched(); > } while (freed > 0); In an extreme case, a single invocation of list_lru_walk() can skip all 1024 dentries, in which case 'freed' will be 0 forcing us to break the loop prematurely. I think we should loop until there's at least one dentry left on the LRU, i.e. while (list_lru_count(&sb->s_dentry_lru) > 0) However, even that wouldn't be quite correct, because list_lru_count() iterates over all memory cgroups to sum list_lru_one->nr_items, which can race with memcg offlining code migrating dentries off a dead cgroup (see memcg_drain_all_list_lrus()). So it looks like to make this check race-free, we need to account the number of entries on the LRU not only per memcg, but also per node, i.e. add list_lru_node->nr_items. Fortunately, list_lru entries can't be migrated between NUMA nodes. > } > EXPORT_SYMBOL(shrink_dcache_sb);