lustre/kernel_patches/patches/vfs-pdirops-2.6.10-fc3.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-2.6.10/fs/inode.c
   7 ===================================================================
   8 --- linux-2.6.10.orig/fs/inode.c        2004-12-25 05:35:40.000000000 +0800
   9 +++ linux-2.6.10/fs/inode.c     2005-03-31 18:03:53.551688872 +0800
  10 @@ -166,6 +166,7 @@
  11                 }
  12                 memset(&inode->u, 0, sizeof(inode->u));
  13                 inode->i_mapping = mapping;
  14 +               dynlock_init(&inode->i_dcache_lock);
  15         }
  16         return inode;
  17  }
  18 Index: linux-2.6.10/fs/namei.c
  19 ===================================================================
  20 --- linux-2.6.10.orig/fs/namei.c        2005-03-31 17:57:10.767921312 +0800
  21 +++ linux-2.6.10/fs/namei.c     2005-03-31 18:05:52.839554360 +0800
  22 @@ -104,6 +104,38 @@
  23   * any extra contention...
  24   */
  25
  26 +void *lock_dir(struct inode *dir, struct qstr *name)
  27 +{
  28 +       unsigned long hash;
  29 +
  30 +       if (!IS_PDIROPS(dir)) {
  31 +               down(&dir->i_sem);
  32 +               return 0;
  33 +       }
  34 +
  35 +       /* OK. fs understands parallel directory operations.
  36 +        * so, we try to acquire lock for hash of requested
  37 +        * filename in order to prevent any operations with
  38 +        * same name in same time -bzzz */
  39 +
  40 +       /* calculate name hash */
  41 +       hash = full_name_hash(name->name, name->len);
  42 +
  43 +       /* lock this hash */
  44 +       return dynlock_lock(&dir->i_dcache_lock, hash, 1, GFP_ATOMIC);
  45 +}
  46 +EXPORT_SYMBOL(lock_dir);
  47 +
  48 +void unlock_dir(struct inode *dir, void *lock)
  49 +{
  50 +       if (!IS_PDIROPS(dir)) {
  51 +               up(&dir->i_sem);
  52 +               return;
  53 +       }
  54 +       dynlock_unlock(&dir->i_dcache_lock, lock);
  55 +}
  56 +EXPORT_SYMBOL(unlock_dir);
  57 +
  58  /* In order to reduce some races, while at the same time doing additional
  59   * checking and hopefully speeding things up, we copy filenames to the
  60   * kernel data space before using them..
  61 @@ -390,8 +422,9 @@
  62  {
  63         struct dentry * result;
  64         struct inode *dir = parent->d_inode;
  65 +       void *lock;
  66
  67 -       down(&dir->i_sem);
  68 +       lock = lock_dir(dir, name);
  69         /*
  70          * First re-do the cached lookup just in case it was created
  71          * while we waited for the directory semaphore..
  72 @@ -417,7 +450,7 @@
  73                         else
  74                                 result = dentry;
  75                 }
  76 -               up(&dir->i_sem);
  77 +               unlock_dir(dir, lock);
  78                 return result;
  79         }
  80
  81 @@ -425,7 +458,7 @@
  82          * Uhhuh! Nasty case: the cache was re-populated while
  83          * we waited on the semaphore. Need to revalidate.
  84          */
  85 -       up(&dir->i_sem);
  86 +       unlock_dir(dir, lock);
  87         if (result->d_op && result->d_op->d_revalidate) {
  88                 if (!result->d_op->d_revalidate(result, nd) && !d_invalidate(result)) {
  89                         dput(result);
  90 @@ -1461,7 +1494,7 @@
  91
  92         dir = nd->dentry;
  93         nd->flags &= ~LOOKUP_PARENT;
  94 -       down(&dir->d_inode->i_sem);
  95 +       nd->lock = lock_dir(dir->d_inode, &nd->last);
  96         nd->flags |= LOOKUP_LAST;
  97         dentry = __lookup_hash(&nd->last, nd->dentry, nd);
  98         nd->flags &= ~LOOKUP_LAST;
  99 @@ -1469,7 +1502,7 @@
 100  do_last:
 101         error = PTR_ERR(dentry);
 102         if (IS_ERR(dentry)) {
 103 -               up(&dir->d_inode->i_sem);
 104 +               unlock_dir(dir->d_inode, nd->lock);
 105                 goto exit;
 106         }
 107
 108 @@ -1478,7 +1511,7 @@
 109                 if (!IS_POSIXACL(dir->d_inode))
 110                         mode &= ~current->fs->umask;
 111                 error = vfs_create(dir->d_inode, dentry, mode, nd);
 112 -               up(&dir->d_inode->i_sem);
 113 +               unlock_dir(dir->d_inode, nd->lock);
 114                 dput(nd->dentry);
 115                 nd->dentry = dentry;
 116                 if (error)
 117 @@ -1492,7 +1525,7 @@
 118         /*
 119          * It already exists.
 120          */
 121 -       up(&dir->d_inode->i_sem);
 122 +       unlock_dir(dir->d_inode, nd->lock);
 123
 124         error = -EEXIST;
 125         if (flag & O_EXCL)
 126 @@ -1576,7 +1609,7 @@
 127                 goto exit;
 128         }
 129         dir = nd->dentry;
 130 -       down(&dir->d_inode->i_sem);
 131 +       nd->lock = lock_dir(dir->d_inode, &nd->last);
 132         nd->flags |= LOOKUP_LAST;
 133         dentry = __lookup_hash(&nd->last, nd->dentry, nd);
 134         nd->flags &= ~LOOKUP_LAST;
 135 @@ -1596,7 +1629,7 @@
 136  {
 137         struct dentry *dentry;
 138
 139 -       down(&nd->dentry->d_inode->i_sem);
 140 +       nd->lock = lock_dir(nd->dentry->d_inode, &nd->last);
 141         dentry = ERR_PTR(-EEXIST);
 142         if (nd->last_type != LAST_NORM)
 143                 goto fail;
 144 @@ -1688,7 +1721,7 @@
 145                 }
 146                 dput(dentry);
 147         }
 148 -       up(&nd.dentry->d_inode->i_sem);
 149 +       unlock_dir(nd.dentry->d_inode, nd.lock);
 150  out2:
 151         path_release(&nd);
 152  out:
 153 @@ -1747,7 +1780,7 @@
 154                         error = vfs_mkdir(nd.dentry->d_inode, dentry, mode);
 155                         dput(dentry);
 156                 }
 157 -               up(&nd.dentry->d_inode->i_sem);
 158 +               unlock_dir(nd.dentry->d_inode, nd.lock);
 159  out2:
 160                 path_release(&nd);
 161  out:
 162 @@ -1852,14 +1885,14 @@
 163                         error = -EBUSY;
 164                         goto exit1;
 165         }
 166 -       down(&nd.dentry->d_inode->i_sem);
 167 +       nd.lock = lock_dir(nd.dentry->d_inode, &nd.last);
 168         dentry = lookup_hash(&nd.last, nd.dentry);
 169         error = PTR_ERR(dentry);
 170         if (!IS_ERR(dentry)) {
 171                 error = vfs_rmdir(nd.dentry->d_inode, dentry);
 172                 dput(dentry);
 173         }
 174 -       up(&nd.dentry->d_inode->i_sem);
 175 +       unlock_dir(nd.dentry->d_inode, nd.lock);
 176  exit1:
 177         path_release(&nd);
 178  exit:
 179 @@ -1925,7 +1958,7 @@
 180         error = -EISDIR;
 181         if (nd.last_type != LAST_NORM)
 182                 goto exit1;
 183 -       down(&nd.dentry->d_inode->i_sem);
 184 +       nd.lock = lock_dir(nd.dentry->d_inode, &nd.last);
 185         dentry = lookup_hash(&nd.last, nd.dentry);
 186         error = PTR_ERR(dentry);
 187         if (!IS_ERR(dentry)) {
 188 @@ -1939,7 +1972,7 @@
 189         exit2:
 190                 dput(dentry);
 191         }
 192 -       up(&nd.dentry->d_inode->i_sem);
 193 +       unlock_dir(nd.dentry->d_inode, nd.lock);
 194         if (inode)
 195                 iput(inode);    /* truncate the inode here */
 196  exit1:
 197 @@ -2005,7 +2038,7 @@
 198                         error = vfs_symlink(nd.dentry->d_inode, dentry, from, S_IALLUGO);
 199                         dput(dentry);
 200                 }
 201 -               up(&nd.dentry->d_inode->i_sem);
 202 +               unlock_dir(nd.dentry->d_inode, nd.lock);
 203  out2:
 204                 path_release(&nd);
 205  out:
 206 @@ -2094,7 +2127,7 @@
 207                 error = vfs_link(old_nd.dentry, nd.dentry->d_inode, new_dentry);
 208                 dput(new_dentry);
 209         }
 210 -       up(&nd.dentry->d_inode->i_sem);
 211 +       unlock_dir(nd.dentry->d_inode, nd.lock);
 212  out_release:
 213         path_release(&nd);
 214  out:
 215 Index: linux-2.6.10/include/linux/fs.h
 216 ===================================================================
 217 --- linux-2.6.10.orig/include/linux/fs.h        2005-03-31 17:57:13.330531736 +0800
 218 +++ linux-2.6.10/include/linux/fs.h     2005-03-31 18:08:59.645155592 +0800
 219 @@ -19,6 +19,7 @@
 220  #include <linux/prio_tree.h>
 221  #include <linux/kobject.h>
 222  #include <asm/atomic.h>
 223 +#include <linux/dynlocks.h>
 224
 225  struct iovec;
 226  struct nameidata;
 227 @@ -151,7 +152,7 @@
 228  #define S_DIRSYNC      64      /* Directory modifications are synchronous */
 229  #define S_NOCMTIME     128     /* Do not update file c/mtime */
 230  #define S_SWAPFILE     256     /* Do not truncate: swapon got its bmaps */
 231 -
 232 +#define S_PDIROPS      512     /* Parallel directory operations */
 233  /*
 234   * Note that nosuid etc flags are inode-specific: setting some file-system
 235   * flags just means all the inodes inherit those flags by default. It might be
 236 @@ -181,6 +182,7 @@
 237  #define IS_NODIRATIME(inode)   __IS_FLG(inode, MS_NODIRATIME)
 238  #define IS_POSIXACL(inode)     __IS_FLG(inode, MS_POSIXACL)
 239  #define IS_ONE_SECOND(inode)   __IS_FLG(inode, MS_ONE_SECOND)
 240 +#define IS_PDIROPS(inode)      __IS_FLG(inode, S_PDIROPS)
 241
 242  #define IS_DEADDIR(inode)      ((inode)->i_flags & S_DEAD)
 243  #define IS_NOCMTIME(inode)     ((inode)->i_flags & S_NOCMTIME)
 244 @@ -482,6 +484,7 @@
 245
 246         atomic_t                i_writecount;
 247         void                    *i_security;
 248 +       struct dynlock          i_dcache_lock;  /* for parallel directory ops */
 249         union {
 250                 void            *generic_ip;
 251         } u;
 252 Index: linux-2.6.10/include/linux/namei.h
 253 ===================================================================
 254 --- linux-2.6.10.orig/include/linux/namei.h     2005-03-31 17:50:12.533502608 +0800
 255 +++ linux-2.6.10/include/linux/namei.h  2005-03-31 18:10:30.237383480 +0800
 256 @@ -63,7 +63,8 @@
 257         int             last_type;
 258         unsigned        depth;
 259         char *saved_names[MAX_NESTED_LINKS + 1];
 260 -
 261 +
 262 +       void *lock;
 263         /* Intent data */
 264         union {
 265                 struct open_intent open;
 266 @@ -91,7 +92,7 @@
 267  #define LOOKUP_ATOMIC          64
 268  #define LOOKUP_LAST            128
 269  #define LOOKUP_LINK_NOTLAST    256
 270 -+
 271 +
 272  /*
 273   * Intent data
 274   */