jffs2: Improve post-mount CRC scan efficiency
authorDavid Woodhouse <David.Woodhouse@intel.com>
Mon, 1 Feb 2016 12:00:25 +0000 (12:00 +0000)
committerDavid Woodhouse <David.Woodhouse@intel.com>
Mon, 29 Feb 2016 22:29:10 +0000 (22:29 +0000)
We need to finish doing the CRC checks before we can allow writes to
happen, and we currently process the inodes in order. This means a call
to jffs2_get_ino_cache() for each possible inode# up to c->highest_ino.

There may be a lot of lookups which fail, if the inode# space is used
sparsely. And the inode# space is *often* used sparsely, if a file
system contains a lot of stuff that was put there in the original
image, followed by lots of creation and deletion of new files.

Instead of processing them numerically with a lookup each time, just
walk the hash buckets instead.

[fix locking typo reported by Dan Carpenter]
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
fs/jffs2/gc.c
fs/jffs2/jffs2_fs_sb.h
fs/jffs2/nodemgmt.c

index 5a2dec2..61c5547 100644 (file)
@@ -134,37 +134,59 @@ int jffs2_garbage_collect_pass(struct jffs2_sb_info *c)
        if (mutex_lock_interruptible(&c->alloc_sem))
                return -EINTR;
 
+
        for (;;) {
+               /* We can't start doing GC until we've finished checking
+                  the node CRCs etc. */
+               int bucket, want_ino;
+
                spin_lock(&c->erase_completion_lock);
                if (!c->unchecked_size)
                        break;
-
-               /* We can't start doing GC yet. We haven't finished checking
-                  the node CRCs etc. Do it now. */
-
-               /* checked_ino is protected by the alloc_sem */
-               if (c->checked_ino > c->highest_ino && xattr) {
-                       pr_crit("Checked all inodes but still 0x%x bytes of unchecked space?\n",
-                               c->unchecked_size);
-                       jffs2_dbg_dump_block_lists_nolock(c);
-                       spin_unlock(&c->erase_completion_lock);
-                       mutex_unlock(&c->alloc_sem);
-                       return -ENOSPC;
-               }
-
                spin_unlock(&c->erase_completion_lock);
 
                if (!xattr)
                        xattr = jffs2_verify_xattr(c);
 
                spin_lock(&c->inocache_lock);
+               /* Instead of doing the inodes in numeric order, doing a lookup
+                * in the hash for each possible number, just walk the hash
+                * buckets of *existing* inodes. This means that we process
+                * them out-of-order, but it can be a lot faster if there's
+                * a sparse inode# space. Which there often is. */
+               want_ino = c->check_ino;
+               for (bucket = c->check_ino % c->inocache_hashsize ; bucket < c->inocache_hashsize; bucket++) {
+                       for (ic = c->inocache_list[bucket]; ic; ic = ic->next) {
+                               if (ic->ino < want_ino)
+                                       continue;
+
+                               if (ic->state != INO_STATE_CHECKEDABSENT &&
+                                   ic->state != INO_STATE_PRESENT)
+                                       goto got_next; /* with inocache_lock held */
+
+                               jffs2_dbg(1, "Skipping ino #%u already checked\n",
+                                         ic->ino);
+                       }
+                       want_ino = 0;
+               }
 
-               ic = jffs2_get_ino_cache(c, c->checked_ino++);
+               /* Point c->check_ino past the end of the last bucket. */
+               c->check_ino = ((c->highest_ino + c->inocache_hashsize + 1) &
+                               ~c->inocache_hashsize) - 1;
 
-               if (!ic) {
-                       spin_unlock(&c->inocache_lock);
-                       continue;
-               }
+               spin_unlock(&c->inocache_lock);
+
+               pr_crit("Checked all inodes but still 0x%x bytes of unchecked space?\n",
+                       c->unchecked_size);
+               jffs2_dbg_dump_block_lists_nolock(c);
+               mutex_unlock(&c->alloc_sem);
+               return -ENOSPC;
+
+       got_next:
+               /* For next time round the loop, we want c->checked_ino to indicate
+                * the *next* one we want to check. And since we're walking the
+                * buckets rather than doing it sequentially, it's: */
+               c->check_ino = ic->ino + c->inocache_hashsize;
 
                if (!ic->pino_nlink) {
                        jffs2_dbg(1, "Skipping check of ino #%d with nlink/pino zero\n",
@@ -176,8 +198,6 @@ int jffs2_garbage_collect_pass(struct jffs2_sb_info *c)
                switch(ic->state) {
                case INO_STATE_CHECKEDABSENT:
                case INO_STATE_PRESENT:
-                       jffs2_dbg(1, "Skipping ino #%u already checked\n",
-                                 ic->ino);
                        spin_unlock(&c->inocache_lock);
                        continue;
 
@@ -196,7 +216,7 @@ int jffs2_garbage_collect_pass(struct jffs2_sb_info *c)
                                  ic->ino);
                        /* We need to come back again for the _same_ inode. We've
                         made no progress in this case, but that should be OK */
-                       c->checked_ino--;
+                       c->check_ino = ic->ino;
 
                        mutex_unlock(&c->alloc_sem);
                        sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
index 046fee8..778275f 100644 (file)
@@ -49,7 +49,7 @@ struct jffs2_sb_info {
        struct mtd_info *mtd;
 
        uint32_t highest_ino;
-       uint32_t checked_ino;
+       uint32_t check_ino;             /* *NEXT* inode to be checked */
 
        unsigned int flags;
 
index b6bd4af..cda0774 100644 (file)
@@ -846,8 +846,8 @@ int jffs2_thread_should_wake(struct jffs2_sb_info *c)
                return 1;
 
        if (c->unchecked_size) {
-               jffs2_dbg(1, "jffs2_thread_should_wake(): unchecked_size %d, checked_ino #%d\n",
-                         c->unchecked_size, c->checked_ino);
+               jffs2_dbg(1, "jffs2_thread_should_wake(): unchecked_size %d, check_ino #%d\n",
+                         c->unchecked_size, c->check_ino);
                return 1;
        }