]> git.openfabrics.org - ~emulex/infiniband.git/commitdiff
f2fs: enhance alloc_nid and build_free_nids flows
authorJaegeuk Kim <jaegeuk.kim@samsung.com>
Thu, 25 Apr 2013 07:05:51 +0000 (16:05 +0900)
committerJaegeuk Kim <jaegeuk.kim@samsung.com>
Mon, 29 Apr 2013 02:19:21 +0000 (11:19 +0900)
In order to avoid build_free_nid lock contention, let's change the order of
function calls as follows.

At first, check whether there is enough free nids.
 - If available, just get a free nid with spin_lock without any overhead.
 - Otherwise, conduct build_free_nids.
  : scan nat pages, journal nat entries, and nat cache entries.

We should consider carefullly not to serve free nids intermediately made by
build_free_nids.
We can get stable free nids only after build_free_nids is done.

Reviewed-by: Namjae Jeon <namjae.jeon@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
fs/f2fs/f2fs.h
fs/f2fs/node.c

index 6283c8d77c2e3287a55ab49adaf9e493caa1044d..20aab02f2a427181a4240ad99ad2b8200a235509 100644 (file)
@@ -190,7 +190,6 @@ static inline void set_raw_extent(struct extent_info *ext,
 struct f2fs_nm_info {
        block_t nat_blkaddr;            /* base disk address of NAT */
        nid_t max_nid;                  /* maximum possible node ids */
-       nid_t init_scan_nid;            /* the first nid to be scanned */
        nid_t next_scan_nid;            /* the next nid to be scanned */
 
        /* NAT cache management */
@@ -360,6 +359,7 @@ struct f2fs_sb_info {
        struct mutex writepages;                /* mutex for writepages() */
        unsigned char next_lock_num;            /* round-robin global locks */
        int por_doing;                          /* recovery is doing or not */
+       int on_build_free_nids;                 /* build_free_nids is doing */
 
        /* for orphan inode management */
        struct list_head orphan_inode_list;     /* orphan inode list */
index c8f48d436487cbbafcde27f97e1467d44d1270ca..aede91071f715e4e1418cfd83752aeb0d2f81b66 100644 (file)
@@ -1309,14 +1309,14 @@ static void build_free_nids(struct f2fs_sb_info *sbi)
        struct f2fs_nm_info *nm_i = NM_I(sbi);
        struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_HOT_DATA);
        struct f2fs_summary_block *sum = curseg->sum_blk;
-       nid_t nid = 0;
-       bool is_cycled = false;
-       int fcnt = 0;
-       int i;
+       int fcnt = 0, i = 0;
+       nid_t nid = nm_i->next_scan_nid;
 
-       nid = nm_i->next_scan_nid;
-       nm_i->init_scan_nid = nid;
+       /* Enough entries */
+       if (nm_i->fcnt > NAT_ENTRY_PER_BLOCK)
+               return;
 
+       /* readahead nat pages to be scanned */
        ra_nat_pages(sbi, nid);
 
        while (1) {
@@ -1326,19 +1326,15 @@ static void build_free_nids(struct f2fs_sb_info *sbi)
                f2fs_put_page(page, 1);
 
                nid += (NAT_ENTRY_PER_BLOCK - (nid % NAT_ENTRY_PER_BLOCK));
-
-               if (nid >= nm_i->max_nid) {
+               if (nid >= nm_i->max_nid)
                        nid = 0;
-                       is_cycled = true;
-               }
-               if (fcnt > MAX_FREE_NIDS)
-                       break;
-               if (is_cycled && nm_i->init_scan_nid <= nid)
+
+               if (i++ == FREE_NID_PAGES)
                        break;
        }
 
-       /* go to the next nat page in order to reuse free nids first */
-       nm_i->next_scan_nid = nm_i->init_scan_nid + NAT_ENTRY_PER_BLOCK;
+       /* go to the next free nat pages to find free nids abundantly */
+       nm_i->next_scan_nid = nid;
 
        /* find free nids from current sum_pages */
        mutex_lock(&curseg->curseg_mutex);
@@ -1375,41 +1371,36 @@ bool alloc_nid(struct f2fs_sb_info *sbi, nid_t *nid)
        struct free_nid *i = NULL;
        struct list_head *this;
 retry:
-       mutex_lock(&nm_i->build_lock);
-       if (!nm_i->fcnt) {
-               /* scan NAT in order to build free nid list */
-               build_free_nids(sbi);
-               if (!nm_i->fcnt) {
-                       mutex_unlock(&nm_i->build_lock);
-                       return false;
-               }
-       }
-       mutex_unlock(&nm_i->build_lock);
+       if (sbi->total_valid_node_count + 1 >= nm_i->max_nid)
+               return false;
 
-       /*
-        * We check fcnt again since previous check is racy as
-        * we didn't hold free_nid_list_lock. So other thread
-        * could consume all of free nids.
-        */
        spin_lock(&nm_i->free_nid_list_lock);
-       if (!nm_i->fcnt) {
-               spin_unlock(&nm_i->free_nid_list_lock);
-               goto retry;
-       }
 
-       BUG_ON(list_empty(&nm_i->free_nid_list));
-       list_for_each(this, &nm_i->free_nid_list) {
-               i = list_entry(this, struct free_nid, list);
-               if (i->state == NID_NEW)
-                       break;
-       }
+       /* We should not use stale free nids created by build_free_nids */
+       if (nm_i->fcnt && !sbi->on_build_free_nids) {
+               BUG_ON(list_empty(&nm_i->free_nid_list));
+               list_for_each(this, &nm_i->free_nid_list) {
+                       i = list_entry(this, struct free_nid, list);
+                       if (i->state == NID_NEW)
+                               break;
+               }
 
-       BUG_ON(i->state != NID_NEW);
-       *nid = i->nid;
-       i->state = NID_ALLOC;
-       nm_i->fcnt--;
+               BUG_ON(i->state != NID_NEW);
+               *nid = i->nid;
+               i->state = NID_ALLOC;
+               nm_i->fcnt--;
+               spin_unlock(&nm_i->free_nid_list_lock);
+               return true;
+       }
        spin_unlock(&nm_i->free_nid_list_lock);
-       return true;
+
+       /* Let's scan nat pages and its caches to get free nids */
+       mutex_lock(&nm_i->build_lock);
+       sbi->on_build_free_nids = 1;
+       build_free_nids(sbi);
+       sbi->on_build_free_nids = 0;
+       mutex_unlock(&nm_i->build_lock);
+       goto retry;
 }
 
 /*
@@ -1696,7 +1687,6 @@ static int init_node_manager(struct f2fs_sb_info *sbi)
        spin_lock_init(&nm_i->free_nid_list_lock);
        rwlock_init(&nm_i->nat_tree_lock);
 
-       nm_i->init_scan_nid = le32_to_cpu(sbi->ckpt->next_free_nid);
        nm_i->next_scan_nid = le32_to_cpu(sbi->ckpt->next_free_nid);
        nm_i->bitmap_size = __bitmap_size(sbi, NAT_BITMAP);
        version_bitmap = __bitmap_ptr(sbi, NAT_BITMAP);