Revert "f2fs: use percpu_rw_semaphore"
authorJaegeuk Kim <jaegeuk@kernel.org>
Thu, 4 Aug 2016 18:38:25 +0000 (11:38 -0700)
committerJaegeuk Kim <jaegeuk@kernel.org>
Fri, 19 Aug 2016 02:15:08 +0000 (11:15 +0900)
LKP reported -36.3% regression of fsmark.files_per_sec due to this patch.
I've confirmed that fxmark [1] has also slight regression for DWAL.

[1] https://github.com/sslab-gatech/fxmark

This reverts commit ec795418c41850056feb956534edf059dc1155d4.

fs/f2fs/f2fs.h
fs/f2fs/node.c
fs/f2fs/super.c

index 675fa79d86f6500d779917b7ca7da3e62f14fda7..14f5fe2b841e2e7d4e699072cc2aeb6e1c14c645 100644 (file)
@@ -538,7 +538,7 @@ struct f2fs_nm_info {
        /* NAT cache management */
        struct radix_tree_root nat_root;/* root of the nat entry cache */
        struct radix_tree_root nat_set_root;/* root of the nat set cache */
-       struct percpu_rw_semaphore nat_tree_lock;       /* protect nat_tree_lock */
+       struct rw_semaphore nat_tree_lock;      /* protect nat_tree_lock */
        struct list_head nat_entries;   /* cached nat entry list (clean) */
        unsigned int nat_cnt;           /* the # of cached nat entries */
        unsigned int dirty_nat_cnt;     /* total num of nat entries in set */
@@ -787,7 +787,7 @@ struct f2fs_sb_info {
        struct f2fs_checkpoint *ckpt;           /* raw checkpoint pointer */
        struct inode *meta_inode;               /* cache meta blocks */
        struct mutex cp_mutex;                  /* checkpoint procedure lock */
-       struct percpu_rw_semaphore cp_rwsem;            /* blocking FS operations */
+       struct rw_semaphore cp_rwsem;           /* blocking FS operations */
        struct rw_semaphore node_write;         /* locking node writes */
        wait_queue_head_t cp_wait;
        unsigned long last_time[MAX_TIME];      /* to store time in jiffies */
@@ -1074,22 +1074,22 @@ static inline void clear_ckpt_flags(struct f2fs_checkpoint *cp, unsigned int f)
 
 static inline void f2fs_lock_op(struct f2fs_sb_info *sbi)
 {
-       percpu_down_read(&sbi->cp_rwsem);
+       down_read(&sbi->cp_rwsem);
 }
 
 static inline void f2fs_unlock_op(struct f2fs_sb_info *sbi)
 {
-       percpu_up_read(&sbi->cp_rwsem);
+       up_read(&sbi->cp_rwsem);
 }
 
 static inline void f2fs_lock_all(struct f2fs_sb_info *sbi)
 {
-       percpu_down_write(&sbi->cp_rwsem);
+       down_write(&sbi->cp_rwsem);
 }
 
 static inline void f2fs_unlock_all(struct f2fs_sb_info *sbi)
 {
-       percpu_up_write(&sbi->cp_rwsem);
+       up_write(&sbi->cp_rwsem);
 }
 
 static inline int __get_cp_reason(struct f2fs_sb_info *sbi)
index b2fa4b615925b894a92377b215780edb33354571..f75d197d5beb05366e876876e09521db7ac3930c 100644 (file)
@@ -206,14 +206,14 @@ int need_dentry_mark(struct f2fs_sb_info *sbi, nid_t nid)
        struct nat_entry *e;
        bool need = false;
 
-       percpu_down_read(&nm_i->nat_tree_lock);
+       down_read(&nm_i->nat_tree_lock);
        e = __lookup_nat_cache(nm_i, nid);
        if (e) {
                if (!get_nat_flag(e, IS_CHECKPOINTED) &&
                                !get_nat_flag(e, HAS_FSYNCED_INODE))
                        need = true;
        }
-       percpu_up_read(&nm_i->nat_tree_lock);
+       up_read(&nm_i->nat_tree_lock);
        return need;
 }
 
@@ -223,11 +223,11 @@ bool is_checkpointed_node(struct f2fs_sb_info *sbi, nid_t nid)
        struct nat_entry *e;
        bool is_cp = true;
 
-       percpu_down_read(&nm_i->nat_tree_lock);
+       down_read(&nm_i->nat_tree_lock);
        e = __lookup_nat_cache(nm_i, nid);
        if (e && !get_nat_flag(e, IS_CHECKPOINTED))
                is_cp = false;
-       percpu_up_read(&nm_i->nat_tree_lock);
+       up_read(&nm_i->nat_tree_lock);
        return is_cp;
 }
 
@@ -237,13 +237,13 @@ bool need_inode_block_update(struct f2fs_sb_info *sbi, nid_t ino)
        struct nat_entry *e;
        bool need_update = true;
 
-       percpu_down_read(&nm_i->nat_tree_lock);
+       down_read(&nm_i->nat_tree_lock);
        e = __lookup_nat_cache(nm_i, ino);
        if (e && get_nat_flag(e, HAS_LAST_FSYNC) &&
                        (get_nat_flag(e, IS_CHECKPOINTED) ||
                         get_nat_flag(e, HAS_FSYNCED_INODE)))
                need_update = false;
-       percpu_up_read(&nm_i->nat_tree_lock);
+       up_read(&nm_i->nat_tree_lock);
        return need_update;
 }
 
@@ -284,7 +284,7 @@ static void set_node_addr(struct f2fs_sb_info *sbi, struct node_info *ni,
        struct f2fs_nm_info *nm_i = NM_I(sbi);
        struct nat_entry *e;
 
-       percpu_down_write(&nm_i->nat_tree_lock);
+       down_write(&nm_i->nat_tree_lock);
        e = __lookup_nat_cache(nm_i, ni->nid);
        if (!e) {
                e = grab_nat_entry(nm_i, ni->nid);
@@ -334,7 +334,7 @@ static void set_node_addr(struct f2fs_sb_info *sbi, struct node_info *ni,
                        set_nat_flag(e, HAS_FSYNCED_INODE, true);
                set_nat_flag(e, HAS_LAST_FSYNC, fsync_done);
        }
-       percpu_up_write(&nm_i->nat_tree_lock);
+       up_write(&nm_i->nat_tree_lock);
 }
 
 int try_to_free_nats(struct f2fs_sb_info *sbi, int nr_shrink)
@@ -342,7 +342,8 @@ int try_to_free_nats(struct f2fs_sb_info *sbi, int nr_shrink)
        struct f2fs_nm_info *nm_i = NM_I(sbi);
        int nr = nr_shrink;
 
-       percpu_down_write(&nm_i->nat_tree_lock);
+       if (!down_write_trylock(&nm_i->nat_tree_lock))
+               return 0;
 
        while (nr_shrink && !list_empty(&nm_i->nat_entries)) {
                struct nat_entry *ne;
@@ -351,7 +352,7 @@ int try_to_free_nats(struct f2fs_sb_info *sbi, int nr_shrink)
                __del_from_nat_cache(nm_i, ne);
                nr_shrink--;
        }
-       percpu_up_write(&nm_i->nat_tree_lock);
+       up_write(&nm_i->nat_tree_lock);
        return nr - nr_shrink;
 }
 
@@ -373,13 +374,13 @@ void get_node_info(struct f2fs_sb_info *sbi, nid_t nid, struct node_info *ni)
        ni->nid = nid;
 
        /* Check nat cache */
-       percpu_down_read(&nm_i->nat_tree_lock);
+       down_read(&nm_i->nat_tree_lock);
        e = __lookup_nat_cache(nm_i, nid);
        if (e) {
                ni->ino = nat_get_ino(e);
                ni->blk_addr = nat_get_blkaddr(e);
                ni->version = nat_get_version(e);
-               percpu_up_read(&nm_i->nat_tree_lock);
+               up_read(&nm_i->nat_tree_lock);
                return;
        }
 
@@ -403,11 +404,11 @@ void get_node_info(struct f2fs_sb_info *sbi, nid_t nid, struct node_info *ni)
        node_info_from_raw_nat(ni, &ne);
        f2fs_put_page(page, 1);
 cache:
-       percpu_up_read(&nm_i->nat_tree_lock);
+       up_read(&nm_i->nat_tree_lock);
        /* cache nat entry */
-       percpu_down_write(&nm_i->nat_tree_lock);
+       down_write(&nm_i->nat_tree_lock);
        cache_nat_entry(sbi, nid, &ne);
-       percpu_up_write(&nm_i->nat_tree_lock);
+       up_write(&nm_i->nat_tree_lock);
 }
 
 /*
@@ -1788,7 +1789,7 @@ void build_free_nids(struct f2fs_sb_info *sbi)
        ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), FREE_NID_PAGES,
                                                        META_NAT, true);
 
-       percpu_down_read(&nm_i->nat_tree_lock);
+       down_read(&nm_i->nat_tree_lock);
 
        while (1) {
                struct page *page = get_current_nat_page(sbi, nid);
@@ -1820,7 +1821,7 @@ void build_free_nids(struct f2fs_sb_info *sbi)
                        remove_free_nid(nm_i, nid);
        }
        up_read(&curseg->journal_rwsem);
-       percpu_up_read(&nm_i->nat_tree_lock);
+       up_read(&nm_i->nat_tree_lock);
 
        ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nm_i->next_scan_nid),
                                        nm_i->ra_nid_pages, META_NAT, false);
@@ -2209,7 +2210,7 @@ void flush_nat_entries(struct f2fs_sb_info *sbi)
        if (!nm_i->dirty_nat_cnt)
                return;
 
-       percpu_down_write(&nm_i->nat_tree_lock);
+       down_write(&nm_i->nat_tree_lock);
 
        /*
         * if there are no enough space in journal to store dirty nat
@@ -2232,7 +2233,7 @@ void flush_nat_entries(struct f2fs_sb_info *sbi)
        list_for_each_entry_safe(set, tmp, &sets, set_list)
                __flush_nat_entry_set(sbi, set);
 
-       percpu_up_write(&nm_i->nat_tree_lock);
+       up_write(&nm_i->nat_tree_lock);
 
        f2fs_bug_on(sbi, nm_i->dirty_nat_cnt);
 }
@@ -2268,8 +2269,7 @@ static int init_node_manager(struct f2fs_sb_info *sbi)
 
        mutex_init(&nm_i->build_lock);
        spin_lock_init(&nm_i->free_nid_list_lock);
-       if (percpu_init_rwsem(&nm_i->nat_tree_lock))
-               return -ENOMEM;
+       init_rwsem(&nm_i->nat_tree_lock);
 
        nm_i->next_scan_nid = le32_to_cpu(sbi->ckpt->next_free_nid);
        nm_i->bitmap_size = __bitmap_size(sbi, NAT_BITMAP);
@@ -2326,7 +2326,7 @@ void destroy_node_manager(struct f2fs_sb_info *sbi)
        spin_unlock(&nm_i->free_nid_list_lock);
 
        /* destroy nat cache */
-       percpu_down_write(&nm_i->nat_tree_lock);
+       down_write(&nm_i->nat_tree_lock);
        while ((found = __gang_lookup_nat_cache(nm_i,
                                        nid, NATVEC_SIZE, natvec))) {
                unsigned idx;
@@ -2351,9 +2351,8 @@ void destroy_node_manager(struct f2fs_sb_info *sbi)
                        kmem_cache_free(nat_entry_set_slab, setvec[idx]);
                }
        }
-       percpu_up_write(&nm_i->nat_tree_lock);
+       up_write(&nm_i->nat_tree_lock);
 
-       percpu_free_rwsem(&nm_i->nat_tree_lock);
        kfree(nm_i->nat_bitmap);
        sbi->nm_info = NULL;
        kfree(nm_i);
index 1b86d3f638efcf87435da17af79811237fc584a9..7f863a645ab1d0c46052875f35f603bebde10a33 100644 (file)
@@ -706,8 +706,6 @@ static void destroy_percpu_info(struct f2fs_sb_info *sbi)
                percpu_counter_destroy(&sbi->nr_pages[i]);
        percpu_counter_destroy(&sbi->alloc_valid_block_count);
        percpu_counter_destroy(&sbi->total_valid_inode_count);
-
-       percpu_free_rwsem(&sbi->cp_rwsem);
 }
 
 static void f2fs_put_super(struct super_block *sb)
@@ -1483,9 +1481,6 @@ static int init_percpu_info(struct f2fs_sb_info *sbi)
 {
        int i, err;
 
-       if (percpu_init_rwsem(&sbi->cp_rwsem))
-               return -ENOMEM;
-
        for (i = 0; i < NR_COUNT_TYPE; i++) {
                err = percpu_counter_init(&sbi->nr_pages[i], 0, GFP_KERNEL);
                if (err)
@@ -1686,6 +1681,7 @@ try_onemore:
                sbi->write_io[i].bio = NULL;
        }
 
+       init_rwsem(&sbi->cp_rwsem);
        init_waitqueue_head(&sbi->cp_wait);
        init_sb_info(sbi);