src/patches/suse-2.6.27.31/patches.suse/mnt_clone_write.patch

   1 From: Nick Piggin <npiggin@suse.de>
   2 Subject: fs: introduce mnt_clone_write
   3 References: bnc#436953
   4 Patch-upstream: no (could be submitted)
   5
   6 This patch speeds up lmbench lat_mmap test by about another 2% after the
   7 first patch.
   8
   9 Before:
  10  avg = 462.286
  11  std = 5.46106
  12
  13 After:
  14  avg = 453.12
  15  std = 9.58257
  16
  17 (50 runs of each, stddev gives a reasonable confidence)
  18
  19 It does this by introducing mnt_clone_write, which avoids some heavyweight
  20 operations of mnt_want_write if called on a vfsmount which we know already
  21 has a write count; and mnt_want_write_file, which can call mnt_clone_write
  22 if the file is open for write.
  23
  24 After these two patches, mnt_want_write and mnt_drop_write go from 7% on
  25 the profile down to 1.3% (including mnt_clone_write).
  26
  27 ---
  28  fs/file_table.c       |    2 -
  29  fs/inode.c            |    2 -
  30  fs/namespace.c        |   63 ++++++++++++++++++++++++++++++++++++++++++++++++++
  31  fs/open.c             |    4 +--
  32  fs/xattr.c            |    4 +--
  33  include/linux/mount.h |    5 +++
  34  6 files changed, 74 insertions(+), 6 deletions(-)
  35
  36 Index: linux-2.6.27/fs/file_table.c
  37 ===================================================================
  38 --- linux-2.6.27.orig/fs/file_table.c
  39 +++ linux-2.6.27/fs/file_table.c
  40 @@ -210,7 +210,7 @@ int init_file(struct file *file, struct
  41          */
  42         if ((mode & FMODE_WRITE) && !special_file(dentry->d_inode->i_mode)) {
  43                 file_take_write(file);
  44 -               error = mnt_want_write(mnt);
  45 +               error = mnt_clone_write_2(mnt);
  46                 WARN_ON(error);
  47         }
  48         return error;
  49 Index: linux-2.6.27/fs/inode.c
  50 ===================================================================
  51 --- linux-2.6.27.orig/fs/inode.c
  52 +++ linux-2.6.27/fs/inode.c
  53 @@ -1256,7 +1256,7 @@ void file_update_time(struct file *file)
  54         if (IS_NOCMTIME(inode))
  55                 return;
  56
  57 -       err = mnt_want_write(file->f_path.mnt);
  58 +       err = mnt_want_write_file(file->f_path.mnt, file);
  59         if (err)
  60                 return;
  61
  62 Index: linux-2.6.27/fs/namespace.c
  63 ===================================================================
  64 --- linux-2.6.27.orig/fs/namespace.c
  65 +++ linux-2.6.27/fs/namespace.c
  66 @@ -264,6 +264,69 @@ out:
  67  EXPORT_SYMBOL_GPL(mnt_want_write);
  68
  69  /**
  70 + * mnt_clone_write - get write access to a mount
  71 + * @mnt: the mount on which to take a write
  72 + *
  73 + * This is effectively like mnt_want_write, except
  74 + * it must only be used to take an extra write reference
  75 + * on a mountpoint that we already know has a write reference
  76 + * on it. This allows some optimisation.
  77 + *
  78 + * The caller should really check __mnt_is_readonly before callint
  79 + * mnt_clone_write. See mnt_clone_write_2.
  80 + *
  81 + * After finished, mnt_drop_write must be called as usual to
  82 + * drop the reference.
  83 + */
  84 +void mnt_clone_write(struct vfsmount *mnt)
  85 +{
  86 +       preempt_disable();
  87 +       inc_mnt_writers(mnt);
  88 +       preempt_enable();
  89 +}
  90 +EXPORT_SYMBOL_GPL(mnt_clone_write);
  91 +
  92 +/**
  93 + * mnt_clone_write_2 - get write access to a mount
  94 + * @mnt: the mount on which to take a write
  95 + *
  96 + * Same as mnt_clone_write, but it performs the __mnt_is_readonly
  97 + * check itself, and returns -error on failure. This is the preferred
  98 + * function. This is here to preserve kABI compatibility.
  99 + *
 100 + * After finished, mnt_drop_write must be called as usual to
 101 + * drop the reference.
 102 + */
 103 +int mnt_clone_write_2(struct vfsmount *mnt)
 104 +{
 105 +       /* superblock may be r/o */
 106 +       if (__mnt_is_readonly(mnt))
 107 +               return -EROFS;
 108 +       preempt_disable();
 109 +       inc_mnt_writers(mnt);
 110 +       preempt_enable();
 111 +       return 0;
 112 +}
 113 +EXPORT_SYMBOL_GPL(mnt_clone_write_2);
 114 +
 115 +/**
 116 + * mnt_want_write_file - get write access to a file's mount
 117 + * @file: the file who's mount on which to take a write
 118 + *
 119 + * This is like mnt_want_write, but it takes a file and can
 120 + * do some optimisations if the file is open for write already
 121 + */
 122 +int mnt_want_write_file(struct vfsmount *mnt, struct file *file)
 123 +{
 124 +       struct inode *inode = file->f_dentry->d_inode;
 125 +       if (!(file->f_mode & FMODE_WRITE) || special_file(inode->i_mode))
 126 +               return mnt_want_write(mnt);
 127 +       else
 128 +               return mnt_clone_write_2(mnt);
 129 +}
 130 +EXPORT_SYMBOL_GPL(mnt_want_write_file);
 131 +
 132 +/**
 133   * mnt_drop_write - give up write access to a mount
 134   * @mnt: the mount on which to give up write access
 135   *
 136 Index: linux-2.6.27/fs/open.c
 137 ===================================================================
 138 --- linux-2.6.27.orig/fs/open.c
 139 +++ linux-2.6.27/fs/open.c
 140 @@ -616,7 +616,7 @@ SYSCALL_DEFINE2(fchmod, unsigned int, fd
 141
 142         audit_inode(NULL, dentry);
 143
 144 -       err = mnt_want_write(file->f_path.mnt);
 145 +       err = mnt_want_write_file(file->f_path.mnt, file);
 146         if (err)
 147                 goto out_putf;
 148         mutex_lock(&inode->i_mutex);
 149 @@ -765,7 +765,7 @@ SYSCALL_DEFINE3(fchown, unsigned int, fd
 150         if (!file)
 151                 goto out;
 152
 153 -       error = mnt_want_write(file->f_path.mnt);
 154 +       error = mnt_want_write_file(file->f_path.mnt, file);
 155         if (error)
 156                 goto out_fput;
 157         dentry = file->f_path.dentry;
 158 Index: linux-2.6.27/fs/xattr.c
 159 ===================================================================
 160 --- linux-2.6.27.orig/fs/xattr.c
 161 +++ linux-2.6.27/fs/xattr.c
 162 @@ -301,7 +301,7 @@ SYSCALL_DEFINE5(fsetxattr, int, fd, cons
 163                 return error;
 164         dentry = f->f_path.dentry;
 165         audit_inode(NULL, dentry);
 166 -       error = mnt_want_write(f->f_path.mnt);
 167 +       error = mnt_want_write_file(f->f_path.mnt, f);
 168         if (!error) {
 169                 error = setxattr(dentry, name, value, size, flags);
 170                 mnt_drop_write(f->f_path.mnt);
 171 @@ -528,7 +528,7 @@ SYSCALL_DEFINE2(fremovexattr, int, fd, c
 172                 return error;
 173         dentry = f->f_path.dentry;
 174         audit_inode(NULL, dentry);
 175 -       error = mnt_want_write(f->f_path.mnt);
 176 +       error = mnt_want_write_file(f->f_path.mnt, f);
 177         if (!error) {
 178                 error = removexattr(dentry, name);
 179                 mnt_drop_write(f->f_path.mnt);
 180 Index: linux-2.6.27/include/linux/mount.h
 181 ===================================================================
 182 --- linux-2.6.27.orig/include/linux/mount.h
 183 +++ linux-2.6.27/include/linux/mount.h
 184 @@ -98,7 +98,12 @@ static inline struct vfsmount *mntget(st
 185         return mnt;
 186  }
 187
 188 +struct file;
 189 +
 190  extern int mnt_want_write(struct vfsmount *mnt);
 191 +extern int mnt_want_write_file(struct vfsmount *mnt, struct file *file);
 192 +extern void mnt_clone_write(struct vfsmount *mnt);
 193 +extern int mnt_clone_write_2(struct vfsmount *mnt);
 194  extern void mnt_drop_write(struct vfsmount *mnt);
 195  extern void mntput_no_expire(struct vfsmount *mnt);
 196  extern void mnt_pin(struct vfsmount *mnt);