lustre/kernel_patches/patches/vfs-pdirops-2.4.21-suse2.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-2.4.21-suse2/fs/namei.c
   7 ===================================================================
   8 --- linux-2.4.21-suse2.orig/fs/namei.c  2004-08-19 13:45:04.000000000 +0400
   9 +++ linux-2.4.21-suse2/fs/namei.c       2004-08-19 15:57:51.000000000 +0400
  10 @@ -103,6 +103,38 @@
  11  }
  12  EXPORT_SYMBOL(intent_release);
  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 @@ -305,10 +337,11 @@
  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 @@ -332,7 +365,7 @@
  63                         else
  64                                 result = dentry;
  65                 }
  66 -               up(&dir->i_sem);
  67 +               unlock_dir(dir, lock);
  68                 return result;
  69         }
  70
  71 @@ -340,7 +373,7 @@
  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, flags) && !d_invalidate(result)) {
  79                         dput(result);
  80 @@ -1186,13 +1219,13 @@
  81                 goto exit;
  82
  83         dir = nd->dentry;
  84 -       down(&dir->d_inode->i_sem);
  85 +       nd->lock = lock_dir(dir->d_inode, &nd->last);
  86         dentry = lookup_hash_it(&nd->last, nd->dentry, it);
  87
  88  do_last:
  89         error = PTR_ERR(dentry);
  90         if (IS_ERR(dentry)) {
  91 -               up(&dir->d_inode->i_sem);
  92 +               unlock_dir(dir->d_inode, nd->lock);
  93                 goto exit;
  94         }
  95
  96 @@ -1202,7 +1235,7 @@
  97                 if (!IS_POSIXACL(dir->d_inode))
  98                         mode &= ~current->fs->umask;
  99                 error = vfs_create_it(dir->d_inode, dentry, mode, it);
 100 -               up(&dir->d_inode->i_sem);
 101 +               unlock_dir(dir->d_inode, nd->lock);
 102  #ifndef DENTRY_WASTE_RAM
 103                 if (error)
 104                         d_drop(dentry);
 105 @@ -1220,7 +1253,7 @@
 106         /*
 107          * It already exists.
 108          */
 109 -       up(&dir->d_inode->i_sem);
 110 +       unlock_dir(dir->d_inode, nd->lock);
 111
 112         error = -EEXIST;
 113         if (flag & O_EXCL)
 114 @@ -1367,7 +1400,7 @@
 115                 goto exit;
 116         }
 117         dir = nd->dentry;
 118 -       down(&dir->d_inode->i_sem);
 119 +       nd->lock = lock_dir(dir->d_inode, &nd->last);
 120         dentry = lookup_hash_it(&nd->last, nd->dentry, it);
 121         putname(nd->last.name);
 122         goto do_last;
 123 @@ -1385,7 +1418,7 @@
 124  {
 125         struct dentry *dentry;
 126
 127 -       down(&nd->dentry->d_inode->i_sem);
 128 +       nd->lock = lock_dir(nd->dentry->d_inode, &nd->last);
 129         dentry = ERR_PTR(-EEXIST);
 130         if (nd->last_type != LAST_NORM)
 131                 goto fail;
 132 @@ -1479,7 +1512,7 @@
 133                 }
 134                 dput(dentry);
 135         }
 136 -       up(&nd.dentry->d_inode->i_sem);
 137 +       unlock_dir(nd.dentry->d_inode, nd.lock);
 138  out2:
 139         path_release(&nd);
 140  out:
 141 @@ -1547,7 +1580,7 @@
 142                         error = vfs_mkdir(nd.dentry->d_inode, dentry, mode);
 143                         dput(dentry);
 144                 }
 145 -               up(&nd.dentry->d_inode->i_sem);
 146 +               unlock_dir(nd.dentry->d_inode, nd.lock);
 147  out2:
 148                 path_release(&nd);
 149  out:
 150 @@ -1657,14 +1690,14 @@
 151                 if (error != -EOPNOTSUPP)
 152                         goto exit1;
 153         }
 154 -       down(&nd.dentry->d_inode->i_sem);
 155 +       nd.lock = lock_dir(nd.dentry->d_inode, &nd.last);
 156         dentry = lookup_hash_it(&nd.last, nd.dentry, NULL);
 157         error = PTR_ERR(dentry);
 158         if (!IS_ERR(dentry)) {
 159                 error = vfs_rmdir(nd.dentry->d_inode, dentry);
 160                 dput(dentry);
 161         }
 162 -       up(&nd.dentry->d_inode->i_sem);
 163 +       unlock_dir(nd.dentry->d_inode, nd.lock);
 164  exit1:
 165         path_release(&nd);
 166  exit:
 167 @@ -1723,7 +1756,7 @@
 168                 if (error != -EOPNOTSUPP)
 169                         goto exit1;
 170         }
 171 -       down(&nd.dentry->d_inode->i_sem);
 172 +       nd.lock = lock_dir(nd.dentry->d_inode, &nd.last);
 173         dentry = lookup_hash_it(&nd.last, nd.dentry, NULL);
 174         error = PTR_ERR(dentry);
 175         if (!IS_ERR(dentry)) {
 176 @@ -1734,7 +1767,7 @@
 177         exit2:
 178                 dput(dentry);
 179         }
 180 -       up(&nd.dentry->d_inode->i_sem);
 181 +       unlock_dir(nd.dentry->d_inode, nd.lock);
 182  exit1:
 183         path_release(&nd);
 184  exit:
 185 @@ -1808,7 +1841,7 @@
 186                         error = vfs_symlink(nd.dentry->d_inode, dentry, from);
 187                         dput(dentry);
 188                 }
 189 -               up(&nd.dentry->d_inode->i_sem);
 190 +               unlock_dir(nd.dentry->d_inode, nd.lock);
 191         out2:
 192                 path_release(&nd);
 193         out:
 194 @@ -1904,7 +1937,7 @@
 195                         error = vfs_link(old_nd.dentry, nd.dentry->d_inode, new_dentry);
 196                         dput(new_dentry);
 197                 }
 198 -               up(&nd.dentry->d_inode->i_sem);
 199 +               unlock_dir(nd.dentry->d_inode, nd.lock);
 200  out_release:
 201                 path_release(&nd);
 202  out:
 203 Index: linux-2.4.21-suse2/fs/inode.c
 204 ===================================================================
 205 --- linux-2.4.21-suse2.orig/fs/inode.c  2004-08-19 13:45:04.000000000 +0400
 206 +++ linux-2.4.21-suse2/fs/inode.c       2004-08-19 13:51:49.000000000 +0400
 207 @@ -121,6 +121,7 @@
 208                 mapping->host = inode;
 209                 mapping->gfp_mask = GFP_HIGHUSER;
 210                 inode->i_mapping = mapping;
 211 +               dynlock_init(&inode->i_dcache_lock);
 212         }
 213         return inode;
 214  }
 215 Index: linux-2.4.21-suse2/include/linux/fs.h
 216 ===================================================================
 217 --- linux-2.4.21-suse2.orig/include/linux/fs.h  2004-08-19 13:45:04.000000000 +0400
 218 +++ linux-2.4.21-suse2/include/linux/fs.h       2004-08-19 13:51:49.000000000 +0400
 219 @@ -22,6 +22,7 @@
 220  #include <linux/stddef.h>
 221  #include <linux/string.h>
 222  #include <linux/frlock.h>
 223 +#include <linux/dynlocks.h>
 224
 225  #include <asm/atomic.h>
 226  #include <asm/bitops.h>
 227 @@ -141,6 +142,7 @@
 228  #define S_IMMUTABLE    16      /* Immutable file */
 229  #define S_DEAD         32      /* removed, but still open directory */
 230  #define S_NOQUOTA      64      /* Inode is not counted to quota */
 231 +#define S_PDIROPS      256     /* Parallel directory operations */
 232
 233  /*
 234   * Note that nosuid etc flags are inode-specific: setting some file-system
 235 @@ -168,6 +170,7 @@
 236  #define IS_NOATIME(inode)      (__IS_FLG(inode, MS_NOATIME) || ((inode)->i_flags & S_NOATIME))
 237  #define IS_NODIRATIME(inode)   __IS_FLG(inode, MS_NODIRATIME)
 238  #define IS_POSIXACL(inode)     __IS_FLG(inode, MS_POSIXACL)
 239 +#define IS_PDIROPS(inode)      __IS_FLG(inode, S_PDIROPS)
 240
 241  #define IS_DEADDIR(inode)      ((inode)->i_flags & S_DEAD)
 242
 243 @@ -525,6 +528,7 @@
 244         atomic_t                i_writecount;
 245         unsigned int            i_attr_flags;
 246         __u32                   i_generation;
 247 +       struct dynlock          i_dcache_lock;  /* for parallel directory ops */
 248         union {
 249                 struct minix_inode_info         minix_i;
 250                 struct ext2_inode_info          ext2_i;
 251 @@ -806,6 +810,7 @@
 252         unsigned int flags;
 253         int last_type;
 254         struct lookup_intent *intent;
 255 +       void *lock;
 256  };
 257
 258  /*
 259 @@ -1793,12 +1798,6 @@
 260         return dget(dentry->d_parent);
 261  }
 262
 263 -static inline void unlock_dir(struct dentry *dir)
 264 -{
 265 -       up(&dir->d_inode->i_sem);
 266 -       dput(dir);
 267 -}
 268 -
 269  /*
 270   * Whee.. Deadlock country. Happily there are only two VFS
 271   * operations that does this..