iam: 0. do not unlock leaf while advancing index part. 1. avoid binary search when...

diff --git a/lustre/kernel_patches/patches/ext3-iam-separate.patch b/lustre/kernel_patches/patches/ext3-iam-separate.patch
index 35b6db1..6db2c28 100644 (file)
--- a/lustre/kernel_patches/patches/ext3-iam-separate.patch
+++ b/lustre/kernel_patches/patches/ext3-iam-separate.patch
@@ -15,7 +15,7 @@ Index: iam/fs/ext3/iam.c
 ===================================================================
 --- iam.orig/fs/ext3/iam.c
 +++ iam/fs/ext3/iam.c
-@@ -0,0 +1,1433 @@
+@@ -0,0 +1,1432 @@
 +/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
 + * vim:expandtab:shiftwidth=8:tabstop=8:
 + *
@@ -897,7 +897,6 @@ Index: iam/fs/ext3/iam.c
 +                 */
 +                while (result == 0 && iam_leaf_at_end(leaf)) {
 +                        do_corr(schedule());
-+                        iam_leaf_unlock(leaf);
 +                        /* advance index portion of the path */
 +                        result = iam_index_next(iam_it_container(it), path);
 +                        if (result == 1) {

diff --git a/lustre/kernel_patches/patches/ext3-pdirops-2.6.9.patch b/lustre/kernel_patches/patches/ext3-pdirops-2.6.9.patch
index 9877b1c..0c3649c 100644 (file)
--- a/lustre/kernel_patches/patches/ext3-pdirops-2.6.9.patch
+++ b/lustre/kernel_patches/patches/ext3-pdirops-2.6.9.patch
@@ -101,7 +101,7 @@ Index: iam/fs/ext3/namei.c
                        BREAKPOINT();
        return 0;
                }
-@@ -241,12 +280,197 @@ struct stats dx_show_entries(struct dx_h
+@@ -241,12 +280,238 @@ struct stats dx_show_entries(struct dx_h
  }
  #endif /* DX_DEBUG */
  
@@ -225,6 +225,43 @@ Index: iam/fs/ext3/namei.c
 +}
 +
 +/*
++ * Fast check for frame consistency.
++ */
++static int dx_check_fast(struct iam_path *path, struct iam_frame *frame)
++{
++      struct iam_container *bag;
++      struct iam_entry *next;
++      struct iam_entry *last;
++      struct iam_entry *entries;
++      struct iam_entry *at;
++
++      assert_corr(!dx_index_is_compat(path));
++
++      bag     = path->ip_container;
++      at      = frame->at;
++      entries = frame->entries;
++      last    = iam_entry_shift(path, entries, dx_get_count(entries) - 1);
++
++      if (unlikely(at > last))
++              return -EAGAIN;
++
++      if (unlikely(dx_get_block(path, at) != frame->leaf))
++              return -EAGAIN;
++
++      if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, at),
++                               path->ip_ikey_target) > 0))
++              return -EAGAIN;
++
++      next = iam_entry_shift(path, at, +1);
++      if (next <= last) {
++              if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, next),
++                                       path->ip_ikey_target) <= 0))
++                      return -EAGAIN;
++      }
++      return 0;
++}
++
++/*
 + * returns 0 if path was unchanged, -EAGAIN otherwise.
 + */
 +static int dx_check_path(struct iam_path *path, struct iam_frame *frame)
@@ -232,7 +269,8 @@ Index: iam/fs/ext3/namei.c
 +      int equal;
 +
 +      dx_lock_bh(frame->bh);
-+      equal = frame->leaf == dx_find_ptr(path, frame);
++      equal = dx_check_fast(path, frame) == 0 ||
++              frame->leaf == dx_find_ptr(path, frame);
 +      DX_DEVAL(dx_lock_stats.dls_bh_again += !equal);
 +      dx_unlock_bh(frame->bh);
 +      
@@ -269,6 +307,9 @@ Index: iam/fs/ext3/namei.c
 +              if (search) {
 +                      struct iam_entry *pos;
 +
++                      if (dx_check_fast(path, scan) == 0)
++                              continue;
++
 +                      pos = dx_find_position(path, scan);
 +                      if (scan->leaf != dx_get_block(path, pos)) {
 +                              result = -EAGAIN;
@@ -301,7 +342,7 @@ Index: iam/fs/ext3/namei.c
  
        struct iam_descr *param;
        struct iam_frame *frame;
-@@ -255,20 +479,19 @@ int dx_lookup(struct iam_path *path)
+@@ -255,20 +520,19 @@ int dx_lookup(struct iam_path *path)
        param = iam_path_descr(path);
        c = path->ip_container;
        
@@ -332,7 +373,7 @@ Index: iam/fs/ext3/namei.c
                if (err != 0)
                        break;
  
-@@ -283,53 +506,83 @@ int dx_lookup(struct iam_path *path)
+@@ -283,53 +547,83 @@ int dx_lookup(struct iam_path *path)
                        break;
  
                assert_inv(dx_node_check(path, frame));
@@ -351,13 +392,27 @@ Index: iam/fs/ext3/namei.c
 -                              q = iam_entry_shift(path, m, -1);
 -                      else
 -                              p = iam_entry_shift(path, m, +1);
--              }
--
++              assert(dx_bug11027_check(path, frame));
++              /*
++               * splitting may change root index block and move hash we're
++               * looking for into another index block so, we have to check
++               * this situation and repeat from begining if path got changed
++               * -bzzz
++               */
++              if (i > 0) {
++                      err = dx_check_path(path, frame - 1);
++                      if (err != 0)
++                                      break;
+               }
+ 
 -              frame->at = iam_entry_shift(path, p, -1);
 -              if (EXT3_INVARIANT_ON) { // linear search cross check
 -                      unsigned n = count - 1;
 -                      struct iam_entry *at;
--
++              frame->at = dx_find_position(path, frame);
++              frame->curidx = ptr;
++              frame->leaf = ptr = dx_get_block(path, frame->at);
+ 
 -                      at = entries;
 -                      while (n--) {
 -                              dxtrace(printk(","));
@@ -371,26 +426,11 @@ Index: iam/fs/ext3/namei.c
 -                                                            path->ip_ikey_target));
 -                                      }
 -                                      at = iam_entry_shift(path, at, -1);
-+              assert(dx_bug11027_check(path, frame));
-+              /*
-+               * splitting may change root index block and move hash we're
-+               * looking for into another index block so, we have to check
-+               * this situation and repeat from begining if path got changed
-+               * -bzzz
-+               */
-+              if (i > 0) {
-+                      err = dx_check_path(path, frame - 1);
-+                      if (err != 0)
-                                       break;
-                               }
+-                                      break;
+-                              }
 -                      }
 -                      assert_corr(at == frame->at);
 -              }
-+
-+              frame->at = dx_find_position(path, frame);
-+              frame->curidx = ptr;
-+              frame->leaf = ptr = dx_get_block(path, frame->at);
-+
 +              dx_unlock_bh(frame->bh);
 +              do_corr(schedule());
        }
@@ -455,7 +495,7 @@ Index: iam/fs/ext3/namei.c
  /*
   * Probe for a directory leaf block to search.
   *
-@@ -339,7 +592,7 @@ int dx_lookup(struct iam_path *path)
+@@ -339,7 +633,7 @@ int dx_lookup(struct iam_path *path)
   * check for this error code, and make sure it never gets reflected
   * back to userspace.
   */
@@ -464,7 +504,7 @@ Index: iam/fs/ext3/namei.c
                    struct dx_hash_info *hinfo, struct iam_path *path)
  {
        int err;
-@@ -347,7 +600,7 @@ static int dx_probe(struct dentry *dentr
+@@ -347,7 +641,7 @@ static int dx_probe(struct dentry *dentr
        
        assert_corr(path->ip_data != NULL);
        ipc = container_of(path->ip_data, struct iam_path_compat, ipc_descr);
@@ -473,7 +513,7 @@ Index: iam/fs/ext3/namei.c
        ipc->ipc_hinfo = hinfo;
  
        assert_corr(dx_index_is_compat(path));
-@@ -356,6 +609,7 @@ static int dx_probe(struct dentry *dentr
+@@ -356,6 +650,7 @@ static int dx_probe(struct dentry *dentr
        return err;
  }
  
@@ -481,7 +521,7 @@ Index: iam/fs/ext3/namei.c
  /*
   * This function increments the frame pointer to search the next leaf
   * block, and reads in the necessary intervening nodes if the search
-@@ -391,10 +645,13 @@ static int ext3_htree_advance(struct ino
+@@ -391,10 +686,13 @@ static int ext3_htree_advance(struct ino
         * nodes need to be read.
         */
        while (1) {
@@ -496,7 +536,7 @@ Index: iam/fs/ext3/namei.c
                if (p == path->ip_frames)
                        return 0;
                num_frames++;
-@@ -409,7 +666,7 @@ static int ext3_htree_advance(struct ino
+@@ -409,7 +707,7 @@ static int ext3_htree_advance(struct ino
         * If the hash is 1, then continue only if the next page has a
         * continuation hash of any value.  This is used for readdir
         * handling.  Otherwise, check to see if the hash matches the
@@ -505,7 +545,7 @@ Index: iam/fs/ext3/namei.c
         * there's no point to read in the successive index pages.
         */
                iam_get_ikey(path, p->at, (struct iam_ikey *)&bhash);
-@@ -425,25 +682,91 @@ static int ext3_htree_advance(struct ino
+@@ -425,25 +723,91 @@ static int ext3_htree_advance(struct ino
         * block so no check is necessary
         */
        while (num_frames--) {
@@ -602,7 +642,7 @@ Index: iam/fs/ext3/namei.c
  }
  
  int ext3_htree_next_block(struct inode *dir, __u32 hash,
-@@ -649,14 +972,26 @@ void iam_insert_key(struct iam_path *pat
+@@ -649,14 +1013,26 @@ void iam_insert_key(struct iam_path *pat
        struct iam_entry *new = iam_entry_shift(path, frame->at, +1);
        int count = dx_get_count(entries);
  
@@ -629,7 +669,7 @@ Index: iam/fs/ext3/namei.c
  }
  
  void dx_insert_block(struct iam_path *path, struct iam_frame *frame,
-@@ -882,7 +1217,7 @@ static struct buffer_head * ext3_dx_find
+@@ -882,7 +1258,7 @@ static struct buffer_head * ext3_dx_find
        sb = dir->i_sb;
        /* NFS may look up ".." - look at dx_root directory block */
        if (namelen > 2 || name[0] != '.'||(name[1] != '.' && name[1] != '\0')){
@@ -638,7 +678,7 @@ Index: iam/fs/ext3/namei.c
                if (*err != 0)
                        return NULL;
        } else {
-@@ -1114,7 +1449,7 @@ struct ext3_dir_entry_2 *move_entries(st
+@@ -1114,7 +1490,7 @@ struct ext3_dir_entry_2 *move_entries(st
        hash2 = map[split].hash;
        continued = hash2 == map[split - 1].hash;
        dxtrace(printk("Split block %i at %x, %i/%i\n",
@@ -647,7 +687,7 @@ Index: iam/fs/ext3/namei.c
  
        /* Fancy dance to stay within two buffers */
        de2 = dx_move_dirents(data1, data2, map + split, count - split);
-@@ -1484,16 +1819,38 @@ static int shift_entries(struct iam_path
+@@ -1484,16 +1860,38 @@ static int shift_entries(struct iam_path
               (char *) iam_entry_shift(path, entries, count1),
               count2 * iam_entry_size(path));
  
@@ -689,7 +729,7 @@ Index: iam/fs/ext3/namei.c
  {
  
        struct iam_entry *entries;   /* old block contents */
-@@ -1501,6 +1858,8 @@ int split_index_node(handle_t *handle, s
+@@ -1501,6 +1899,8 @@ int split_index_node(handle_t *handle, s
        struct iam_frame *frame, *safe;
        struct buffer_head *bh_new[DX_MAX_TREE_HEIGHT] = {0};
        u32 newblock[DX_MAX_TREE_HEIGHT] = {0};
@@ -698,7 +738,7 @@ Index: iam/fs/ext3/namei.c
        struct inode *dir = iam_path_obj(path);
        struct iam_descr *descr;
        int nr_splet;
-@@ -1523,12 +1882,14 @@ int split_index_node(handle_t *handle, s
+@@ -1523,12 +1923,14 @@ int split_index_node(handle_t *handle, s
         *   - first allocate all necessary blocks
         *
         *   - insert pointers into them atomically.
@@ -717,7 +757,7 @@ Index: iam/fs/ext3/namei.c
        dxtrace(printk("using %u of %u node entries\n",
                       dx_get_count(entries), dx_get_limit(entries)));
  
-@@ -1536,6 +1897,7 @@ int split_index_node(handle_t *handle, s
+@@ -1536,6 +1938,7 @@ int split_index_node(handle_t *handle, s
        for (nr_splet = 0; frame >= path->ip_frames &&
             dx_get_count(frame->entries) == dx_get_limit(frame->entries);
             --frame, ++nr_splet) {
@@ -725,7 +765,7 @@ Index: iam/fs/ext3/namei.c
                if (nr_splet == DX_MAX_TREE_HEIGHT) {
                        ext3_warning(dir->i_sb, __FUNCTION__,
                                     "Directory index full!\n");
-@@ -1545,14 +1907,53 @@ int split_index_node(handle_t *handle, s
+@@ -1545,14 +1948,53 @@ int split_index_node(handle_t *handle, s
        }
  
        safe = frame;
@@ -780,7 +820,7 @@ Index: iam/fs/ext3/namei.c
                BUFFER_TRACE(frame->bh, "get_write_access");
                err = ext3_journal_get_write_access(handle, frame->bh);
                if (err)
-@@ -1560,6 +1961,7 @@ int split_index_node(handle_t *handle, s
+@@ -1560,6 +2002,7 @@ int split_index_node(handle_t *handle, s
        }
        /* Add "safe" node to transaction too */
        if (safe + 1 != path->ip_frames) {
@@ -788,7 +828,7 @@ Index: iam/fs/ext3/namei.c
                err = ext3_journal_get_write_access(handle, safe->bh);
                if (err)
                        goto journal_error;
-@@ -1596,16 +1998,21 @@ int split_index_node(handle_t *handle, s
+@@ -1596,16 +2039,21 @@ int split_index_node(handle_t *handle, s
  
                        assert_corr(i == 0);
  
@@ -810,7 +850,7 @@ Index: iam/fs/ext3/namei.c
                        /* Shift frames in the path */
                        memmove(frames + 2, frames + 1,
                                (sizeof path->ip_frames) - 2 * sizeof frames[0]);
-@@ -1613,18 +2020,22 @@ int split_index_node(handle_t *handle, s
+@@ -1613,18 +2061,22 @@ int split_index_node(handle_t *handle, s
                        frames[1].at = iam_entry_shift(path, entries2, idx);
                        frames[1].entries = entries = entries2;
                        frames[1].bh = bh2;
@@ -833,7 +873,7 @@ Index: iam/fs/ext3/namei.c
                        count = shift_entries(path, frame, count,
                                              entries, entries2, newblock[i]);
                        /* Which index block gets the new entry? */
-@@ -1634,33 +2045,44 @@ int split_index_node(handle_t *handle, s
+@@ -1634,33 +2086,44 @@ int split_index_node(handle_t *handle, s
                                frame->at = iam_entry_shift(path, entries2,
                                                            idx - count + d);
                                frame->entries = entries = entries2;
@@ -879,7 +919,7 @@ Index: iam/fs/ext3/namei.c
        if (nr_splet > 0) {
                /*
                 * Log ->i_size modification.
-@@ -1674,6 +2096,12 @@ journal_error:
+@@ -1674,6 +2137,12 @@ journal_error:
        ext3_std_error(dir->i_sb, err);
  
  cleanup:
@@ -892,7 +932,7 @@ Index: iam/fs/ext3/namei.c
        for (i = 0; i < ARRAY_SIZE(bh_new); ++i) {
                if (bh_new[i] != NULL)
                        brelse(bh_new[i]);
-@@ -1695,18 +2123,18 @@ static int ext3_dx_add_entry(handle_t *h
+@@ -1695,18 +2164,18 @@ static int ext3_dx_add_entry(handle_t *h
        struct buffer_head * bh = NULL;
        struct inode *dir = dentry->d_parent->d_inode;
        struct ext3_dir_entry_2 *de;
@@ -913,7 +953,7 @@ Index: iam/fs/ext3/namei.c
        isize = dir->i_size;
  
        err = param->id_ops->id_node_read(path->ip_container,
-@@ -1726,7 +2154,7 @@ static int ext3_dx_add_entry(handle_t *h
+@@ -1726,7 +2195,7 @@ static int ext3_dx_add_entry(handle_t *h
                goto cleanup;
        }
        
@@ -922,7 +962,7 @@ Index: iam/fs/ext3/namei.c
        if (err)
                goto cleanup;   
  
-@@ -1736,12 +2164,14 @@ static int ext3_dx_add_entry(handle_t *h
+@@ -1736,12 +2205,14 @@ static int ext3_dx_add_entry(handle_t *h
                goto cleanup;
  
        assert_inv(dx_node_check(path, frame));