lkml.org 
[lkml]   [2016]   [Oct]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
Date
SubjectRe: [PATCH 3/3] ovl: redirect on rename-dir
On Wed, Oct 26, 2016 at 2:26 PM, Miklos Szeredi <miklos@szeredi.hu> wrote:
> On Tue, Oct 25, 2016 at 2:49 PM, Amir Goldstein <amir73il@gmail.com> wrote:
>
>>> @@ -880,31 +913,34 @@ static int ovl_rename(struct inode *olddir, struct dentry *old,
>>> if (WARN_ON(olddentry->d_inode == newdentry->d_inode))
>>> goto out_dput;
>>>
>>> - if (is_dir && !old_opaque && ovl_lower_positive(new)) {
>>> - err = ovl_set_opaque(olddentry);
>>> - if (err)
>>> - goto out_dput;
>>> - ovl_dentry_set_opaque(old, true);
>>> + if (is_dir) {
>>> + if (ovl_type_merge_or_lower(old)) {
>>> + err = ovl_set_redirect(old);
>>
>> There is a fair chance of getting ENOSPC/EDQUOT here and confuse user space.
>> Would it be better to convert these non fatal errors with EXDEV, so
>> user space will
>> gracefully fallback to recursive rename/clone/copy?
>
> Recursive copy up will surely consume more space than an xattr?

Surely. But that is not what I meant.
In Ext4, for example, the total size of extended attributes for an
inode is limited to a single block,
so you can get ENOSPC with alot of free space in the file system.

You can also get ERANGE if the redirect path length > 256.
I'm just saying that instead of failing the move with ERANGE or
uncalled for ENOSPC,
it is better for user space to get EXDEV, from which there is a sane fallback.
BTW, mv (from linux-utils) falls back from -EXDEV to recursive move,
which is quite
efficient with clone up.

>
>>> @@ -162,6 +223,23 @@ struct dentry *ovl_lookup(struct inode *dir, struct dentry *dentry,
>>> stack[ctr].dentry = this;
>>> stack[ctr].mnt = lowerpath.mnt;
>>> ctr++;
>>> +
>>> + if (!stop && i != poe->numlower - 1 &&
>>> + d_is_dir(this) && ovl_redirect_dir(dentry->d_sb)) {
>>> + err = ovl_check_redirect(this, &redirect);
>>> + if (err)
>>> + goto out_put;
>>> +
>>> + if (redirect && poe != dentry->d_sb->s_root->d_fsdata) {
>>> + poe = dentry->d_sb->s_root->d_fsdata;
>>> +
>>
>> Now you are about to continue looping until new value of poe->numlower,
>> which is >= then olf value of poe->numlower, but 'stack' was allocated
>> according to old value of poe->numlower, so aren't you in danger of
>> overflowing it?
>>
>> Please add a comment to explain the purpose of this loop rewind.
>
> We are jumping to a stack possibly wider than the current one and need
> to find the layer where to continue the downward traversal. I'll add
> the comment.
>
> BTW I don't remember having tested this, so it might possibly be
> buggy. Automatic multi-layer testing would really be good. What we
> basically need is:
>
> - create normal (two layer) overlay (with interesting constructs,
> whiteout, opaque dir, redirect)
> - umount
> - create three layer overlay where the two lower layers come from the
> previous upper/lower layers
> - do more interesting things
>
> There's one such test in xfstests but it would be good to have more.
>
> Thanks,
> Miklos

\
 
 \ /
  Last update: 2016-10-26 14:12    [W:0.106 / U:0.040 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site