convert-DAX-faults-to-iomap-infrastructure

   1 ext4: convert DAX faults to iomap infrastructure
   2
   3 From: Jan Kara <jack@suse.cz>
   4
   5 Convert DAX faults to use iomap infrastructure. We would not have to start
   6 transaction in ext4_dax_fault() anymore since ext4_iomap_begin takes
   7 care of that but so far we do that to avoid lock inversion of
   8 transaction start with DAX entry lock which gets acquired in
   9 dax_iomap_fault() before calling ->iomap_begin handler.
  10
  11 Signed-off-by: Jan Kara <jack@suse.cz>
  12 Signed-off-by: Theodore Ts'o <tytso@mit.edu>
  13 ---
  14  fs/ext4/file.c  |  9 +++++----
  15  fs/ext4/inode.c | 14 +++++++++-----
  16  2 files changed, 14 insertions(+), 9 deletions(-)
  17
  18 diff --git a/fs/ext4/file.c b/fs/ext4/file.c
  19 index 1953fe34f9fe..b5f184493c57 100644
  20 --- a/fs/ext4/file.c
  21 +++ b/fs/ext4/file.c
  22 @@ -275,7 +275,7 @@ static int ext4_dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
  23         if (IS_ERR(handle))
  24                 result = VM_FAULT_SIGBUS;
  25         else
  26 -               result = dax_fault(vma, vmf, ext4_dax_get_block);
  27 +               result = dax_iomap_fault(vma, vmf, &ext4_iomap_ops);
  28
  29         if (write) {
  30                 if (!IS_ERR(handle))
  31 @@ -309,9 +309,10 @@ static int ext4_dax_pmd_fault(struct vm_area_struct *vma, unsigned long addr,
  32
  33         if (IS_ERR(handle))
  34                 result = VM_FAULT_SIGBUS;
  35 -       else
  36 -               result = dax_pmd_fault(vma, addr, pmd, flags,
  37 -                                        ext4_dax_get_block);
  38 +       else {
  39 +               result = dax_iomap_pmd_fault(vma, addr, pmd, flags,
  40 +                                            &ext4_iomap_ops);
  41 +       }
  42
  43         if (write) {
  44                 if (!IS_ERR(handle))
  45 diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
  46 index 3192ec0768d4..4d71c7bc3524 100644
  47 --- a/fs/ext4/inode.c
  48 +++ b/fs/ext4/inode.c
  49 @@ -3361,12 +3361,16 @@ static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
  50                 }
  51
  52                 /*
  53 -                * If we added blocks beyond i_size we need to make sure they
  54 +                * If we added blocks beyond i_size, we need to make sure they
  55                  * will get truncated if we crash before updating i_size in
  56 -                * ext4_iomap_end().
  57 +                * ext4_iomap_end(). For faults we don't need to do that (and
  58 +                * even cannot because for orphan list operations inode_lock is
  59 +                * required) - if we happen to instantiate block beyond i_size,
  60 +                * it is because we race with truncate which has already added
  61 +                * the inode to the orphan list.
  62                  */
  63 -               if (first_block + map.m_len >
  64 -                   (inode->i_size + (1 << blkbits) - 1) >> blkbits) {
  65 +               if (!(flags & IOMAP_FAULT) && first_block + map.m_len >
  66 +                   (i_size_read(inode) + (1 << blkbits) - 1) >> blkbits) {
  67                         int err;
  68
  69                         err = ext4_orphan_add(handle, inode);
  70 @@ -3412,7 +3416,7 @@ static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
  71         int blkbits = inode->i_blkbits;
  72         bool truncate = false;
  73
  74 -       if (!(flags & IOMAP_WRITE))
  75 +       if (!(flags & IOMAP_WRITE) || (flags & IOMAP_FAULT))
  76                 return 0;
  77
  78         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
  79 --
  80 2.6.6