lustre/kernel_patches/patches/vfs-pdirops-2.6.7.patch

   1  fs/inode.c         |    1
   2  fs/namei.c         |   66 ++++++++++++++++++++++++++++++++++++++---------------
   3  include/linux/fs.h |   11 ++++----
   4  3 files changed, 54 insertions(+), 24 deletions(-)
   5
   6 Index: linux-2.6.7/fs/namei.c
   7 ===================================================================
   8 --- linux-2.6.7.orig/fs/namei.c 2004-09-06 22:53:19.000000000 +0400
   9 +++ linux-2.6.7/fs/namei.c      2004-09-07 14:39:44.000000000 +0400
  10 @@ -103,6 +103,38 @@
  11   * any extra contention...
  12   */
  13
  14 +void *lock_dir(struct inode *dir, struct qstr *name)
  15 +{
  16 +       unsigned long hash;
  17 +
  18 +       if (!IS_PDIROPS(dir)) {
  19 +               down(&dir->i_sem);
  20 +               return 0;
  21 +       }
  22 +
  23 +       /* OK. fs understands parallel directory operations.
  24 +        * so, we try to acquire lock for hash of requested
  25 +        * filename in order to prevent any operations with
  26 +        * same name in same time -bzzz */
  27 +
  28 +       /* calculate name hash */
  29 +       hash = full_name_hash(name->name, name->len);
  30 +
  31 +       /* lock this hash */
  32 +       return dynlock_lock(&dir->i_dcache_lock, hash, 1, GFP_ATOMIC);
  33 +}
  34 +EXPORT_SYMBOL(lock_dir);
  35 +
  36 +void unlock_dir(struct inode *dir, void *lock)
  37 +{
  38 +       if (!IS_PDIROPS(dir)) {
  39 +               up(&dir->i_sem);
  40 +               return;
  41 +       }
  42 +       dynlock_unlock(&dir->i_dcache_lock, lock);
  43 +}
  44 +EXPORT_SYMBOL(unlock_dir);
  45 +
  46  /* In order to reduce some races, while at the same time doing additional
  47   * checking and hopefully speeding things up, we copy filenames to the
  48   * kernel data space before using them..
  49 @@ -362,10 +394,11 @@
  50         struct dentry * result;
  51         struct inode *dir = parent->d_inode;
  52         int counter = 0;
  53 +       void *lock;
  54
  55  again:
  56         counter++;
  57 -       down(&dir->i_sem);
  58 +       lock = lock_dir(dir, name);
  59         /*
  60          * First re-do the cached lookup just in case it was created
  61          * while we waited for the directory semaphore..
  62 @@ -391,7 +424,7 @@
  63                         else
  64                                 result = dentry;
  65                 }
  66 -               up(&dir->i_sem);
  67 +               unlock_dir(dir, lock);
  68                 return result;
  69         }
  70
  71 @@ -399,7 +432,7 @@
  72          * Uhhuh! Nasty case: the cache was re-populated while
  73          * we waited on the semaphore. Need to revalidate.
  74          */
  75 -       up(&dir->i_sem);
  76 +       unlock_dir(dir, lock);
  77         if (result->d_op && result->d_op->d_revalidate) {
  78                 if (!result->d_op->d_revalidate(result, nd) && !d_invalidate(result)) {
  79                         dput(result);
  80 @@ -1367,7 +1400,7 @@
  81
  82         dir = nd->dentry;
  83         nd->flags &= ~LOOKUP_PARENT;
  84 -       down(&dir->d_inode->i_sem);
  85 +       nd->lock = lock_dir(dir->d_inode, &nd->last);
  86         nd->flags |= LOOKUP_LAST;
  87         dentry = __lookup_hash(&nd->last, nd->dentry, nd);
  88         nd->flags &= ~LOOKUP_LAST;
  89 @@ -1375,7 +1408,7 @@
  90  do_last:
  91         error = PTR_ERR(dentry);
  92         if (IS_ERR(dentry)) {
  93 -               up(&dir->d_inode->i_sem);
  94 +               unlock_dir(dir->d_inode, nd->lock);
  95                 goto exit;
  96         }
  97
  98 @@ -1384,7 +1417,7 @@
  99                 if (!IS_POSIXACL(dir->d_inode))
 100                         mode &= ~current->fs->umask;
 101                 error = vfs_create(dir->d_inode, dentry, mode, nd);
 102 -               up(&dir->d_inode->i_sem);
 103 +               unlock_dir(dir->d_inode, nd->lock);
 104                 dput(nd->dentry);
 105                 nd->dentry = dentry;
 106                 if (error)
 107 @@ -1398,7 +1431,7 @@
 108         /*
 109          * It already exists.
 110          */
 111 -       up(&dir->d_inode->i_sem);
 112 +       unlock_dir(dir->d_inode, nd->lock);
 113
 114         error = -EEXIST;
 115         if (flag & O_EXCL)
 116 @@ -1474,7 +1507,7 @@
 117                 goto exit;
 118         }
 119         dir = nd->dentry;
 120 -       down(&dir->d_inode->i_sem);
 121 +       nd->lock = lock_dir(dir->d_inode, &nd->last);
 122         nd->flags |= LOOKUP_LAST;
 123         dentry = __lookup_hash(&nd->last, nd->dentry, nd);
 124         nd->flags &= ~LOOKUP_LAST;
 125 @@ -1494,7 +1527,7 @@
 126  {
 127         struct dentry *dentry;
 128
 129 -       down(&nd->dentry->d_inode->i_sem);
 130 +       nd->lock = lock_dir(nd->dentry->d_inode, &nd->last);
 131         dentry = ERR_PTR(-EEXIST);
 132         if (nd->last_type != LAST_NORM)
 133                 goto fail;
 134 @@ -1589,7 +1622,7 @@
 135                 }
 136                 dput(dentry);
 137         }
 138 -       up(&nd.dentry->d_inode->i_sem);
 139 +       unlock_dir(nd.dentry->d_inode, nd.lock);
 140  out2:
 141         path_release(&nd);
 142  out:
 143 @@ -1652,7 +1685,7 @@
 144                         error = vfs_mkdir(nd.dentry->d_inode, dentry, mode);
 145                         dput(dentry);
 146                 }
 147 -               up(&nd.dentry->d_inode->i_sem);
 148 +               unlock_dir(nd.dentry->d_inode, nd.lock);
 149  out2:
 150                 path_release(&nd);
 151  out:
 152 @@ -1765,14 +1798,14 @@
 153                         goto exit1;
 154         }
 155
 156 -       down(&nd.dentry->d_inode->i_sem);
 157 +       nd.lock = lock_dir(nd.dentry->d_inode, &nd.last);
 158         dentry = lookup_hash(&nd.last, nd.dentry);
 159         error = PTR_ERR(dentry);
 160         if (!IS_ERR(dentry)) {
 161                 error = vfs_rmdir(nd.dentry->d_inode, dentry);
 162                 dput(dentry);
 163         }
 164 -       up(&nd.dentry->d_inode->i_sem);
 165 +       unlock_dir(nd.dentry->d_inode, nd.lock);
 166  exit1:
 167         path_release(&nd);
 168  exit:
 169 @@ -1842,7 +1875,7 @@
 170                 if (error != -EOPNOTSUPP)
 171                         goto exit1;
 172         }
 173 -       down(&nd.dentry->d_inode->i_sem);
 174 +       nd.lock = lock_dir(nd.dentry->d_inode, &nd.last);
 175         dentry = lookup_hash(&nd.last, nd.dentry);
 176         error = PTR_ERR(dentry);
 177         if (!IS_ERR(dentry)) {
 178 @@ -1856,7 +1889,7 @@
 179         exit2:
 180                 dput(dentry);
 181         }
 182 -       up(&nd.dentry->d_inode->i_sem);
 183 +       unlock_dir(nd.dentry->d_inode, nd.lock);
 184  exit1:
 185         path_release(&nd);
 186  exit:
 187 @@ -1927,7 +1960,7 @@
 188                         error = vfs_symlink(nd.dentry->d_inode, dentry, from, S_IALLUGO);
 189                         dput(dentry);
 190                 }
 191 -               up(&nd.dentry->d_inode->i_sem);
 192 +               unlock_dir(nd.dentry->d_inode, nd.lock);
 193  out2:
 194                 path_release(&nd);
 195  out:
 196 @@ -2021,7 +2054,7 @@
 197                 error = vfs_link(old_nd.dentry, nd.dentry->d_inode, new_dentry);
 198                 dput(new_dentry);
 199         }
 200 -       up(&nd.dentry->d_inode->i_sem);
 201 +       unlock_dir(nd.dentry->d_inode, nd.lock);
 202  out_release:
 203         path_release(&nd);
 204  out:
 205 Index: linux-2.6.7/fs/inode.c
 206 ===================================================================
 207 --- linux-2.6.7.orig/fs/inode.c 2004-09-06 22:53:19.000000000 +0400
 208 +++ linux-2.6.7/fs/inode.c      2004-09-07 14:39:44.000000000 +0400
 209 @@ -165,6 +165,7 @@
 210                 }
 211                 memset(&inode->u, 0, sizeof(inode->u));
 212                 inode->i_mapping = mapping;
 213 +               dynlock_init(&inode->i_dcache_lock);
 214         }
 215         return inode;
 216  }
 217 Index: linux-2.6.7/include/linux/fs.h
 218 ===================================================================
 219 --- linux-2.6.7.orig/include/linux/fs.h 2004-09-07 14:12:12.000000000 +0400
 220 +++ linux-2.6.7/include/linux/fs.h      2004-09-07 14:39:44.000000000 +0400
 221 @@ -19,6 +19,7 @@
 222  #include <linux/prio_tree.h>
 223  #include <linux/kobject.h>
 224  #include <asm/atomic.h>
 225 +#include <linux/dynlocks.h>
 226
 227  struct iovec;
 228  struct nameidata;
 229 @@ -141,6 +142,7 @@
 230  #define S_NOQUOTA      64      /* Inode is not counted to quota */
 231  #define S_DIRSYNC      128     /* Directory modifications are synchronous */
 232  #define S_NOCMTIME     256     /* Do not update file c/mtime */
 233 +#define S_PDIROPS      256     /* Parallel directory operations */
 234
 235  /*
 236   * Note that nosuid etc flags are inode-specific: setting some file-system
 237 @@ -172,6 +174,7 @@
 238  #define IS_NODIRATIME(inode)   __IS_FLG(inode, MS_NODIRATIME)
 239  #define IS_POSIXACL(inode)     __IS_FLG(inode, MS_POSIXACL)
 240  #define IS_ONE_SECOND(inode)   __IS_FLG(inode, MS_ONE_SECOND)
 241 +#define IS_PDIROPS(inode)      __IS_FLG(inode, S_PDIROPS)
 242
 243  #define IS_DEADDIR(inode)      ((inode)->i_flags & S_DEAD)
 244  #define IS_NOCMTIME(inode)     ((inode)->i_flags & S_NOCMTIME)
 245 @@ -463,6 +466,7 @@
 246         atomic_t                i_writecount;
 247         void                    *i_security;
 248         __u32                   i_generation;
 249 +       struct dynlock          i_dcache_lock;  /* for parallel directory ops */
 250         union {
 251                 void            *generic_ip;
 252         } u;
 253 Index: linux-2.6.7/include/linux/namei.h
 254 ===================================================================
 255 --- linux-2.6.7.orig/include/linux/namei.h      2004-09-06 22:53:19.000000000 +0400
 256 +++ linux-2.6.7/include/linux/namei.h   2004-09-07 14:40:06.000000000 +0400
 257 @@ -52,6 +52,7 @@
 258         unsigned int    flags;
 259         int             last_type;
 260         struct lookup_intent intent;
 261 +       void *lock;
 262  };
 263
 264  /*