Re: [f2fs-dev] [PATCH 6/6] f2fs: set I_LINKABLE early to avoid wrong access by vfs

From: Jaegeuk Kim
Date: Thu Dec 12 2019 - 11:55:17 EST


On 12/11, Chao Yu wrote:
> On 2019/12/11 9:31, Jaegeuk Kim wrote:
> > On 12/11, Chao Yu wrote:
> >> On 2019/12/11 9:21, Jaegeuk Kim wrote:
> >>> On 12/10, Chao Yu wrote:
> >>>> On 2019/12/10 6:23, Jaegeuk Kim wrote:
> >>>>> This patch moves setting I_LINKABLE early in rename2(whiteout) to avoid the
> >>>>> below warning.
> >>>>>
> >>>>> [ 3189.163385] WARNING: CPU: 3 PID: 59523 at fs/inode.c:358 inc_nlink+0x32/0x40
> >>>>> [ 3189.246979] Call Trace:
> >>>>> [ 3189.248707] f2fs_init_inode_metadata+0x2d6/0x440 [f2fs]
> >>>>> [ 3189.251399] f2fs_add_inline_entry+0x162/0x8c0 [f2fs]
> >>>>> [ 3189.254010] f2fs_add_dentry+0x69/0xe0 [f2fs]
> >>>>> [ 3189.256353] f2fs_do_add_link+0xc5/0x100 [f2fs]
> >>>>> [ 3189.258774] f2fs_rename2+0xabf/0x1010 [f2fs]
> >>>>> [ 3189.261079] vfs_rename+0x3f8/0xaa0
> >>>>> [ 3189.263056] ? tomoyo_path_rename+0x44/0x60
> >>>>> [ 3189.265283] ? do_renameat2+0x49b/0x550
> >>>>> [ 3189.267324] do_renameat2+0x49b/0x550
> >>>>> [ 3189.269316] __x64_sys_renameat2+0x20/0x30
> >>>>> [ 3189.271441] do_syscall_64+0x5a/0x230
> >>>>> [ 3189.273410] entry_SYSCALL_64_after_hwframe+0x49/0xbe
> >>>>> [ 3189.275848] RIP: 0033:0x7f270b4d9a49
> >>>>>
> >>>>> Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> >>>>> ---
> >>>>> fs/f2fs/namei.c | 27 +++++++++++++--------------
> >>>>> 1 file changed, 13 insertions(+), 14 deletions(-)
> >>>>>
> >>>>> diff --git a/fs/f2fs/namei.c b/fs/f2fs/namei.c
> >>>>> index a1c507b0b4ac..5d9584281935 100644
> >>>>> --- a/fs/f2fs/namei.c
> >>>>> +++ b/fs/f2fs/namei.c
> >>>>> @@ -797,6 +797,7 @@ static int __f2fs_tmpfile(struct inode *dir, struct dentry *dentry,
> >>>>>
> >>>>> if (whiteout) {
> >>>>> f2fs_i_links_write(inode, false);
> >>>>> + inode->i_state |= I_LINKABLE;
> >>>>> *whiteout = inode;
> >>>>> } else {
> >>>>> d_tmpfile(dentry, inode);
> >>>>> @@ -867,6 +868,12 @@ static int f2fs_rename(struct inode *old_dir, struct dentry *old_dentry,
> >>>>> F2FS_I(old_dentry->d_inode)->i_projid)))
> >>>>> return -EXDEV;
> >>>>>
> >>>>> + if (flags & RENAME_WHITEOUT) {
> >>>>> + err = f2fs_create_whiteout(old_dir, &whiteout);
> >>>>> + if (err)
> >>>>> + return err;
> >>>>> + }
> >>>>
> >>>> To record quota info correctly, we need to create whiteout inode after
> >>>> dquot_initialize(old_dir)?
> >>>
> >>> __f2fs_tmpfile() will do it.
> >>
> >> Okay.
> >>
> >> Any comments on below question?
> >>
> >>>
> >>>>
> >>>>> +
> >>>>> err = dquot_initialize(old_dir);
> >>>>> if (err)
> >>>>> goto out;
> >>>>> @@ -898,17 +905,11 @@ static int f2fs_rename(struct inode *old_dir, struct dentry *old_dentry,
> >>>>> }
> >>>>> }
> >>>>>
> >>>>> - if (flags & RENAME_WHITEOUT) {
> >>>>> - err = f2fs_create_whiteout(old_dir, &whiteout);
> >>>>> - if (err)
> >>>>> - goto out_dir;
> >>>>> - }
> >>>>> -
> >>>>> if (new_inode) {
> >>>>>
> >>>>> err = -ENOTEMPTY;
> >>>>> if (old_dir_entry && !f2fs_empty_dir(new_inode))
> >>>>> - goto out_whiteout;
> >>>>> + goto out_dir;
> >>>>>
> >>>>> err = -ENOENT;
> >>>>> new_entry = f2fs_find_entry(new_dir, &new_dentry->d_name,
> >>>>> @@ -916,7 +917,7 @@ static int f2fs_rename(struct inode *old_dir, struct dentry *old_dentry,
> >>>>> if (!new_entry) {
> >>>>> if (IS_ERR(new_page))
> >>>>> err = PTR_ERR(new_page);
> >>>>> - goto out_whiteout;
> >>>>> + goto out_dir;
> >>>>> }
> >>>>>
> >>>>> f2fs_balance_fs(sbi, true);
> >>>>> @@ -948,7 +949,7 @@ static int f2fs_rename(struct inode *old_dir, struct dentry *old_dentry,
> >>>>> err = f2fs_add_link(new_dentry, old_inode);
> >>>>> if (err) {
> >>>>> f2fs_unlock_op(sbi);
> >>>>> - goto out_whiteout;
> >>>>> + goto out_dir;
> >>>>> }
> >>>>>
> >>>>> if (old_dir_entry)
> >>>>> @@ -972,7 +973,7 @@ static int f2fs_rename(struct inode *old_dir, struct dentry *old_dentry,
> >>>>> if (IS_ERR(old_page))
> >>>>> err = PTR_ERR(old_page);
> >>>>> f2fs_unlock_op(sbi);
> >>>>> - goto out_whiteout;
> >>>>> + goto out_dir;
> >>>>> }
> >>>>> }
> >>>>> }
> >>>>> @@ -991,7 +992,6 @@ static int f2fs_rename(struct inode *old_dir, struct dentry *old_dentry,
> >>>>> f2fs_delete_entry(old_entry, old_page, old_dir, NULL);
> >>>>>
> >>>>> if (whiteout) {
> >>>>> - whiteout->i_state |= I_LINKABLE;
> >>>>> set_inode_flag(whiteout, FI_INC_LINK);
> >>>>> err = f2fs_add_link(old_dentry, whiteout);
> >>>>
> >>>> [ 3189.256353] f2fs_do_add_link+0xc5/0x100 [f2fs]
> >>>> [ 3189.258774] f2fs_rename2+0xabf/0x1010 [f2fs]
> >>>>
> >>>> Does the call stack point here? if so, we have set I_LINKABLE before
> >>>> f2fs_add_link(), why the warning still be triggered?
> >>
> >> Am I missing something?
> >
> > Not sure exactly tho, I suspect some races before/after unlock_new_inode().
>
> Alright, I doubt some races on whiteout->i_state updating, as we set I_LINKABLE
> w/o holding inode.i_lock.
>
> Could you have a try with holding i_lock?

I don't see the warning with this patch, and jumped to another issue. I'd like
to take a look at that later.

Thanks,

>
> Thanks,
>
> >
> >>
> >> Thanks,
> >>
> >>>>
> >>>> Thanks,
> >>>>
> >>>>> if (err)
> >>>>> @@ -1027,15 +1027,14 @@ static int f2fs_rename(struct inode *old_dir, struct dentry *old_dentry,
> >>>>> f2fs_unlock_op(sbi);
> >>>>> if (new_page)
> >>>>> f2fs_put_page(new_page, 0);
> >>>>> -out_whiteout:
> >>>>> - if (whiteout)
> >>>>> - iput(whiteout);
> >>>>> out_dir:
> >>>>> if (old_dir_entry)
> >>>>> f2fs_put_page(old_dir_page, 0);
> >>>>> out_old:
> >>>>> f2fs_put_page(old_page, 0);
> >>>>> out:
> >>>>> + if (whiteout)
> >>>>> + iput(whiteout);
> >>>>> return err;
> >>>>> }
> >>>>>
> >>>>>
> >>> .
> >>>
> > .
> >