resolve race between llap_shrinker and umount.

Branch b1_8
b=18773
i=jay
i=tappro

diff --git a/lustre/ChangeLog b/lustre/ChangeLog
index f872eee..ff8aca4 100644 (file)
--- a/lustre/ChangeLog
+++ b/lustre/ChangeLog
@@ -185,6 +185,13 @@ Details    : Do not put cancelled locks into replay list, hold references on
          information, please refer to bugzilla 17630.
 
 Severity   : normal
+Frequency  : rare, at shutdown
+Bugzilla   : 18773
+Descriptoin: panic at umount
+Details    : llap_shrinker can be raced with killing super block from list and
+             this produce panic with access to already freeded pointer
+
+Severity   : normal
 Frequency  : rare
 Bugzilla   : 18154
 Descriptoin: don't lose wakeup for imp_recovery_waitq

diff --git a/lustre/llite/llite_lib.c b/lustre/llite/llite_lib.c
index db80b2e..eb4cb3b 100644 (file)
--- a/lustre/llite/llite_lib.c
+++ b/lustre/llite/llite_lib.c
@@ -66,7 +66,7 @@ extern struct address_space_operations ll_dir_aops;
 #define log2(n) ffz(~(n))
 #endif
 
-static inline void ll_pglist_fini(struct ll_sb_info *sbi)
+static void ll_pglist_fini(struct ll_sb_info *sbi)
 {
         struct page *page;
         int i;
@@ -86,7 +86,7 @@ static inline void ll_pglist_fini(struct ll_sb_info *sbi)
         sbi->ll_pglist = NULL;
 }
 
-static inline int ll_pglist_init(struct ll_sb_info *sbi)
+static int ll_pglist_init(struct ll_sb_info *sbi)
 {
         struct ll_pglist_data *pd;
         unsigned long budget;
@@ -226,10 +226,11 @@ void ll_free_sbi(struct super_block *sb)
         ENTRY;
 
         if (sbi != NULL) {
-                ll_pglist_fini(sbi);
                 spin_lock(&ll_sb_lock);
                 list_del(&sbi->ll_list);
                 spin_unlock(&ll_sb_lock);
+                /* dont allow find cache via sb list first */
+                ll_pglist_fini(sbi);
                 lcounter_destroy(&sbi->ll_async_page_count);
                 OBD_FREE(sbi->ll_async_page_sample,
                          sizeof(long) * num_possible_cpus());
@@ -1296,8 +1297,11 @@ ll_shrink_cache(int priority, unsigned int gfp_mask)
         struct ll_sb_info *sbi;
         int count = 0;
 
+        /* don't race with umount */
+        spin_lock(&ll_sb_lock);
         list_for_each_entry(sbi, &ll_super_blocks, ll_list)
                 count += llap_shrink_cache(sbi, priority);
+        spin_unlock(&ll_sb_lock);
 
 #if defined(HAVE_CACHE_RETURN_INT)
         return count;

diff --git a/lustre/llite/rw.c b/lustre/llite/rw.c
index c2e84de..b8498d2 100644 (file)
--- a/lustre/llite/rw.c
+++ b/lustre/llite/rw.c
@@ -555,9 +555,10 @@ static inline int llap_shrink_cache_internal(struct ll_sb_info *sbi,
                 int keep;
 
                 if (unlikely(need_resched())) {
+                        list_del(&dummy_llap.llap_pglist_item);                
                         ll_pglist_cpu_unlock(sbi, cpu);
-                        cond_resched();
-                        ll_pglist_cpu_lock(sbi, cpu);
+                        /* vmscan::shrink_slab() have own schedule() */
+                        return count;
                 }
 
                 llap = llite_pglist_next_llap(head,