Thank you Andreas and Ted for reviewing the patch. Ted, thank you for
pointing out about the subject line when sending a new patch - I'll be
careful next time. I have handled all the review comments. I will send
out a patch with handled comments. About testing this patch, I have
created a new xfstest that verifies that directory shrinks after
deleting files in it. I'll send that patch out too.
On Tue, Jan 29, 2019 at 5:58 PM Andreas Dilger <adilger@xxxxxxxxx> wrote:
>
> On Jan 23, 2019, at 11:32 AM, harshadshirwadkar@xxxxxxxxx wrote:
> >
> > From: Harshad Shirwadkar <harshadshirwadkar@xxxxxxxxx>
> >
> > This patch is the first step towards shrinking htree based directories
> > when files are deleted. We truncate directory inode when a directory
> > entry removal causes last directory block to be empty. This patch just
> > removes the last block. We may end up in a situation where many
> > intermediate dirent blocks in directory inode are empty. Removing
> > those blocks would be handled later.
> >
> > Signed-off-by: Harshad Shirwadkar <harshadshirwadkar@xxxxxxxxx>
> > ---
> > fs/ext4/namei.c | 130 ++++++++++++++++++++++++++++++++++++++++--------
> > 1 file changed, 109 insertions(+), 21 deletions(-)
> >
> > diff --git a/fs/ext4/namei.c b/fs/ext4/namei.c
> > index 2a4c25c4681d..79cb88ba898a 100644
> > --- a/fs/ext4/namei.c
> > +++ b/fs/ext4/namei.c
> > @@ -273,7 +273,8 @@ static int ext4_htree_next_block(struct inode *dir, __u32 hash,
> > __u32 *start_hash);
> > static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
> > struct ext4_filename *fname,
> > - struct ext4_dir_entry_2 **res_dir);
> > + struct ext4_dir_entry_2 **res_dir,
> > + struct dx_frame *dx_frame);
>
> (style) these should align after '(' on the first line
>
> > @@ -380,6 +381,7 @@ static void ext4_dirent_csum_set(struct inode *inode,
> > (void *)t - (void *)dirent);
> > }
> >
> > +
> > int ext4_handle_dirty_dirent_node(handle_t *handle,
> > struct inode *inode,
> > struct buffer_head *bh)
>
> (style) No need for this hunk.
>
> > @@ -866,6 +868,42 @@ dx_probe(struct ext4_filename *fname, struct inode *dir,
> > return ret_err;
> > }
> >
> > +static void
> > +ext4_dx_delete_entry(handle_t *handle, struct inode *dir,
> > + struct dx_frame *dx_frame, __le64 block)
> > +{
> > + struct dx_entry *entries;
> > + unsigned int count, limit;
> > + int err, i;
> > +
> > + entries = dx_frame->entries;
> > + count = dx_get_count(entries);
> > + limit = dx_get_limit(entries);
> > +
> > + for (i = 0; i < count; i++)
> > + if (entries[i].block == block)
> > + break;
> > +
> > + if (i >= count)
> > + return;
> > +
> > + err = ext4_journal_get_write_access(handle, dx_frame->bh);
> > + if (err) {
> > + ext4_std_error(dir->i_sb, err);
> > + return;
> > + }
> > +
> > + for (; i < count - 1; i++)
> > + entries[i] = entries[i + 1];
> > +
> > + dx_set_count(entries, count - 1);
> > + dx_set_limit(entries, limit);
>
> I don't think you need to update the limit just because of the count changing?
> This is just setting it back to the same value that dx_get_limit() returned earlier.
>
> > @@ -1309,6 +1347,14 @@
> > +static int is_empty_dirent_block(struct inode *dir, struct buffer_head *bh)
>
> (style) this function should be type bool. The older "is_*" functions were written
> before "bool" was available in the kernel, though a separate cleanup patch to convert
> them to bool would be welcome.
>
> > +{
> > + struct ext4_dir_entry_2 *de = (struct ext4_dir_entry_2 *)bh->b_data;
> > +
> > + return (ext4_rec_len_from_disk(de->rec_len, dir->i_sb->s_blocksize)
> > + == dir->i_sb->s_blocksize);
>
> (style) no need for () around return value.
>
> >
> > @@ -1339,7 +1385,8 @@ static int is_dx_internal_node(struct inode *dir, ext4_lblk_t block,
> > static struct buffer_head * ext4_find_entry (struct inode *dir,
> > const struct qstr *d_name,
> > struct ext4_dir_entry_2 **res_dir,
> > - int *inlined)
> > + int *inlined,
> > + struct dx_frame *dx_frame)
>
> (style) this can fit on the previous line, and they can all align after '(' on
> the first line. Please also remove the space after '*' in the function declaration.
>
> > @@ -1486,9 +1533,10 @@ static struct buffer_head * ext4_find_entry (struct inode *dir,
> > return ret;
> > }
> >
> > -static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
> > - struct ext4_filename *fname,
> > - struct ext4_dir_entry_2 **res_dir)
> > +static struct buffer_head *ext4_dx_find_entry(
> > + struct inode *dir, struct ext4_filename *fname,
>
> (style) it isn't clear why you moved "struct inode *dir" from the previous line?
> The continued lines look like they can properly align after '(' on the previous line.
>
> > @@ -2337,9 +2389,13 @@ int ext4_generic_delete_entry(handle_t *handle,
> > static int ext4_delete_entry(handle_t *handle,
> > struct inode *dir,
> > struct ext4_dir_entry_2 *de_del,
> > - struct buffer_head *bh)
> > + struct buffer_head *bh,
> > + struct dx_frame *dx_frame)
> > {
> > + struct ext4_map_blocks map;
> > int err, csum_size = 0;
> > + int should_truncate = 0;
>
> (style) bool should_truncate
>
> > @@ -2363,11 +2419,35 @@ static int ext4_delete_entry(handle_t *handle,
> > if (err)
> > goto out;
> >
> > + if (dx_frame && dx_frame->bh && is_empty_dirent_block(dir, bh)) {
> > +
>
> (style) no blank line here.
>
> > + map.m_lblk = (dir->i_size - 1) >>
> > + (dir->i_sb->s_blocksize_bits);
>
> (style) no need for parenthesis around dir->i_sb->s_blocksize_bits.
> (style) this can fit on the previous line and still be under 80 columns.
>
> > + map.m_len = 1;
> > + err = ext4_map_blocks(handle, dir, &map, 0);
> > +
> > + if ((err > 0) && (bh->b_blocknr == map.m_pblk)) {
>
> (style) no need for extra parenthesis around "err > 0" and "b_blocknr == m_pblk".
>
> > + ext4_dx_delete_entry(handle, dir, dx_frame,
> > + cpu_to_le64(map.m_lblk));
> > + should_truncate = 1;
> > + }
> > + }
> > +
> > BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
> > +
> > err = ext4_handle_dirty_dirent_node(handle, dir, bh);
> > if (unlikely(err))
> > goto out;
> >
> > + if (should_truncate) {
> > + dir->i_size -= dir->i_sb->s_blocksize;
> > + ext4_truncate(dir);
>
> Since we have already mapped the last block above, I wonder if we can do something
> lighter weight than calling ext4_truncate() here? Maybe moving the ext4_truncate()
> guts into a helper function like ext4_truncate_internal() that avoids all of the
> complexity (orphan list, resizing the transaction, inline data, etc).
>
> > + if (dir->i_size == dir->i_sb->s_blocksize) {
> > + ext4_clear_inode_flag(dir, EXT4_INODE_INDEX);
> > + ext4_mark_inode_dirty(handle, dir);
>
> Are there any places in the code that assume an indexed directory will remain as such?
> The tricky part here is that any valid htree directory will have 3 blocks (dx_root,
> plus two leaf blocks), so it might be problematic to go back to a single non-htree
> block, especially since this doesn't look like it has any locking. It is probably
> worthwhile to check the code for this.
After handling Ted's comment about not deleting the last block if it
makes intermediate node empty, this patch will never result in a
situation where dir->i_size == blocksize. So, this hunk is not needed
anymore.
>
> > @@ -2985,6 +3065,9 @@ static int ext4_unlink(struct inode *dir, struct dentry *dentry)
> > struct buffer_head *bh;
> > struct ext4_dir_entry_2 *de;
> > handle_t *handle = NULL;
> > + struct dx_frame dx_frame;
> > +
> > + memset(&dx_frame, 0, sizeof(dx_frame));
>
> Rather than an explicit memset(), this can use a struct initializer when it is declared:
>
> struct dx_frame dx_frame = { NULL };
>
> Are there existing xfstests that cover this case? Certainly deleting files from a
> directory, but it would be useful to have something that checks whether the directory
> size is shrinking and this code is working.
>
> Cheers, Andreas
>
>
>
>
>
