Date:   Tue, 31 Jan 2023 20:33:01 -0800
From:   Daeho Jeong <daeho43@...il.com>
To:     Chao Yu <chao@...nel.org>
Cc:     linux-kernel@...r.kernel.org,
        linux-f2fs-devel@...ts.sourceforge.net, kernel-team@...roid.com,
        Daeho Jeong <daehojeong@...gle.com>,
        syzbot+823000d23b3400619f7c@...kaller.appspotmail.com
Subject: Re: [f2fs-dev] [PATCH] f2fs: synchronize atomic write aborts

On Tue, Jan 31, 2023 at 5:40 PM Chao Yu <chao@...nel.org> wrote:
>
> Hi Daeho,
>
> On 2023/2/1 6:34, Daeho Jeong wrote:
> > On Tue, Jan 31, 2023 at 1:57 PM Daeho Jeong <daeho43@...il.com> wrote:
> >>
> >> On Tue, Jan 31, 2023 at 1:38 PM Daeho Jeong <daeho43@...il.com> wrote:
> >>>
> >>> On Tue, Jan 31, 2023 at 11:13 AM Daeho Jeong <daeho43@...il.com> wrote:
> >>>>
> >>>> Hi Chao,
> >>>>
> >>>> On Tue, Jan 31, 2023 at 3:37 AM Chao Yu <chao@...nel.org> wrote:
> >>>>>
> >>>>> Hi Daeho,
> >>>>>
> >>>>> On 2023/1/31 0:34, Daeho Jeong wrote:
> >>>>>> Hi Chao,
> >>>>>>
> >>>>>> I read your patch series now and I like it.
> >>>>>
> >>>>> Thank you for checking the patches. :)
> >>>>>
> >>>>>> However, how about a race condition between start_atomic_write and
> >>>>>> abort_atomic_write?
> >>>>>
> >>>>> Yup, I noticed that issue, I guess we can avoid this race condition by
> >>>>> covering these two flows w/ i_atomic_sem.
> >>>>>
> >>>>>> abort_atomic_write is called without inode_lock in closing filp scenarios.
> >>>>>> What do you think about this?
> >>>>>
> >>>>> I'm fine w/ your change as it's more clean, but it's better to drop cow_inode's
> >>>>> page cache if atomic_write is committed or aborted to avoid caching obsolete page?
> >>>>
> >>>> It's better to put that part in f2fs_abort_atomic_write().
> >>>> On top of that, maybe, we should move
> >>>> f2fs_do_truncate_blocks(fi->cow_inode, 0, true) part from
> >>>> f2fs_ioc_start_atomic_write() to f2fs_abort_atomic_write(), too.
> >>>
> >>> Oh, we shouldn't touch the f2fs_do_truncate_blocks() part, since there
> >>> might be some left writeback after aborting atomic write.
> >>> Plz. review it related to the timing of calling truncate_inode_pages_final().
> >>
> >> Looks like the scenario becomes too complicated if I think about more
> >> than one writer's scenario.
> >> How about we check writecount in commit_atomic_write ioctl and return
> >> EBUSY when it's not only one writer?
> >> In that case, we can make the scenario simple and effective, and we
> >> can release all the resources in abort_atomic_write().
> >
> > Oh, I totally forgot this. We don't use pages of COW inode. So, we
>
> Yes, after atomic_write is committed or aborted, we don't use pages of
> cow_inode, so they are obsolete, IMO, we'd better to reclaim them
> immediately to avoid unnecessary memory use.

I mean we don't use page cache of the COW inode at all during the
whole atomic write procedure.
We only use the page cache of the original file.

>
> The implementation may be:
>
> - f2fs_ioc_abort_atomic_write
>   - f2fs_abort_atomic_write
>    - truncate_inode_pages_final(cow_inode)
>
> - f2fs_ioc_commit_atomic_write
>   - f2fs_abort_atomic_write
>    - truncate_inode_pages_final(cow_inode)
>
> Thanks,
>
> > don't need to clean them up.
> >
> >>
> >>>
> >>>>
> >>>> Thanks,
> >>>>
> >>>>>
> >>>>> Thanks,
> >>>>>
> >>>>>>
> >>>>>> Thanks,
> >>>>>>
> >>>>>>
> >>>>>> On Fri, Jan 27, 2023 at 6:07 PM Chao Yu <chao@...nel.org> wrote:
> >>>>>>>
> >>>>>>> Hi Daeho, Jaegeuk,
> >>>>>>>
> >>>>>>> Please take a look at patchset in below link:
> >>>>>>>
> >>>>>>> https://lore.kernel.org/linux-f2fs-devel/20230109034453.490176-1-chao@kernel.org/T/#t
> >>>>>>>
> >>>>>>> In PATCH 4/5, I'm trying to fix the same issue w/ alternative way, let me
> >>>>>>> know your preference. :)
> >>>>>>>
> >>>>>>> One comment as below.
> >>>>>>>
> >>>>>>> On 2023/1/13 8:49, Daeho Jeong wrote:
> >>>>>>>> From: Daeho Jeong <daehojeong@...gle.com>
> >>>>>>>>
> >>>>>>>> To fix a race condition between atomic write aborts, I use the inode
> >>>>>>>> lock and make COW inode to be re-usable thoroughout the whole
> >>>>>>>> atomic file inode lifetime.
> >>>>>>>>
> >>>>>>>> Reported-by: syzbot+823000d23b3400619f7c@...kaller.appspotmail.com
> >>>>>>>> Fixes: 3db1de0e582c ("f2fs: change the current atomic write way")
> >>>>>>>> Signed-off-by: Daeho Jeong <daehojeong@...gle.com>
> >>>>>>>> ---
> >>>>>>>>     fs/f2fs/file.c    | 43 ++++++++++++++++++++++++++++---------------
> >>>>>>>>     fs/f2fs/inode.c   | 11 +++++++++--
> >>>>>>>>     fs/f2fs/segment.c |  3 ---
> >>>>>>>>     fs/f2fs/super.c   |  2 --
> >>>>>>>>     4 files changed, 37 insertions(+), 22 deletions(-)
> >>>>>>>>
> >>>>>>>> diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
> >>>>>>>> index ecbc8c135b49..ff072a9ed258 100644
> >>>>>>>> --- a/fs/f2fs/file.c
> >>>>>>>> +++ b/fs/f2fs/file.c
> >>>>>>>> @@ -1866,7 +1866,10 @@ static int f2fs_release_file(struct inode *inode, struct file *filp)
> >>>>>>>>                         atomic_read(&inode->i_writecount) != 1)
> >>>>>>>>                 return 0;
> >>>>>>>>
> >>>>>>>> +     inode_lock(inode);
> >>>>>>>>         f2fs_abort_atomic_write(inode, true);
> >>>>>>>> +     inode_unlock(inode);
> >>>>>>>> +
> >>>>>>>>         return 0;
> >>>>>>>>     }
> >>>>>>>>
> >>>>>>>> @@ -1880,8 +1883,11 @@ static int f2fs_file_flush(struct file *file, fl_owner_t id)
> >>>>>>>>          * until all the writers close its file. Since this should be done
> >>>>>>>>          * before dropping file lock, it needs to do in ->flush.
> >>>>>>>>          */
> >>>>>>>> -     if (F2FS_I(inode)->atomic_write_task == current)
> >>>>>>>> +     if (F2FS_I(inode)->atomic_write_task == current) {
> >>>>>>>> +             inode_lock(inode);
> >>>>>>>>                 f2fs_abort_atomic_write(inode, true);
> >>>>>>>> +             inode_unlock(inode);
> >>>>>>>> +     }
> >>>>>>>>         return 0;
> >>>>>>>>     }
> >>>>>>>>
> >>>>>>>> @@ -2087,19 +2093,28 @@ static int f2fs_ioc_start_atomic_write(struct file *filp, bool truncate)
> >>>>>>>>                 goto out;
> >>>>>>>>         }
> >>>>>>>>
> >>>>>>>> -     /* Create a COW inode for atomic write */
> >>>>>>>> -     pinode = f2fs_iget(inode->i_sb, fi->i_pino);
> >>>>>>>> -     if (IS_ERR(pinode)) {
> >>>>>>>> -             f2fs_up_write(&fi->i_gc_rwsem[WRITE]);
> >>>>>>>> -             ret = PTR_ERR(pinode);
> >>>>>>>> -             goto out;
> >>>>>>>> -     }
> >>>>>>>> +     /* Check if the inode already has a COW inode */
> >>>>>>>> +     if (fi->cow_inode == NULL) {
> >>>>>>>> +             /* Create a COW inode for atomic write */
> >>>>>>>> +             pinode = f2fs_iget(inode->i_sb, fi->i_pino);
> >>>>>>>> +             if (IS_ERR(pinode)) {
> >>>>>>>> +                     f2fs_up_write(&fi->i_gc_rwsem[WRITE]);
> >>>>>>>> +                     ret = PTR_ERR(pinode);
> >>>>>>>> +                     goto out;
> >>>>>>>> +             }
> >>>>>>>>
> >>>>>>>> -     ret = f2fs_get_tmpfile(mnt_userns, pinode, &fi->cow_inode);
> >>>>>>>> -     iput(pinode);
> >>>>>>>> -     if (ret) {
> >>>>>>>> -             f2fs_up_write(&fi->i_gc_rwsem[WRITE]);
> >>>>>>>> -             goto out;
> >>>>>>>> +             ret = f2fs_get_tmpfile(mnt_userns, pinode, &fi->cow_inode);
> >>>>>>>> +             iput(pinode);
> >>>>>>>> +             if (ret) {
> >>>>>>>> +                     f2fs_up_write(&fi->i_gc_rwsem[WRITE]);
> >>>>>>>> +                     goto out;
> >>>>>>>> +             }
> >>>>>>>> +
> >>>>>>>> +             set_inode_flag(fi->cow_inode, FI_COW_FILE);
> >>>>>>>> +             clear_inode_flag(fi->cow_inode, FI_INLINE_DATA);
> >>>>>>>> +     } else {
> >>>>>>>> +             /* Reuse the already created COW inode */
> >>>>>>>> +             f2fs_do_truncate_blocks(fi->cow_inode, 0, true);
> >>>>>>>>         }
> >>>>>>>>
> >>>>>>>>         f2fs_write_inode(inode, NULL);
> >>>>>>>> @@ -2107,8 +2122,6 @@ static int f2fs_ioc_start_atomic_write(struct file *filp, bool truncate)
> >>>>>>>>         stat_inc_atomic_inode(inode);
> >>>>>>>>
> >>>>>>>>         set_inode_flag(inode, FI_ATOMIC_FILE);
> >>>>>>>> -     set_inode_flag(fi->cow_inode, FI_COW_FILE);
> >>>>>>>> -     clear_inode_flag(fi->cow_inode, FI_INLINE_DATA);
> >>>>>>>>
> >>>>>>>>         isize = i_size_read(inode);
> >>>>>>>>         fi->original_i_size = isize;
> >>>>>>>> diff --git a/fs/f2fs/inode.c b/fs/f2fs/inode.c
> >>>>>>>> index ff6cf66ed46b..4921f7209e28 100644
> >>>>>>>> --- a/fs/f2fs/inode.c
> >>>>>>>> +++ b/fs/f2fs/inode.c
> >>>>>>>> @@ -766,11 +766,18 @@ int f2fs_write_inode(struct inode *inode, struct writeback_control *wbc)
> >>>>>>>>     void f2fs_evict_inode(struct inode *inode)
> >>>>>>>>     {
> >>>>>>>>         struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
> >>>>>>>> -     nid_t xnid = F2FS_I(inode)->i_xattr_nid;
> >>>>>>>> +     struct f2fs_inode_info *fi = F2FS_I(inode);
> >>>>>>>> +     nid_t xnid = fi->i_xattr_nid;
> >>>>>>>>         int err = 0;
> >>>>>>>>
> >>>>>>>>         f2fs_abort_atomic_write(inode, true);
> >>>>>>>>
> >>>>>>>> +     if (fi->cow_inode) {
> >>>>>>>> +             clear_inode_flag(fi->cow_inode, FI_COW_FILE);
> >>>>>>>> +             iput(fi->cow_inode);
> >>>>>>>> +             fi->cow_inode = NULL;
> >>>>>>>> +     }
> >>>>>>>> +
> >>>>>>>>         trace_f2fs_evict_inode(inode);
> >>>>>>>>         truncate_inode_pages_final(&inode->i_data);
> >>>>>>>>
> >>>>>>>> @@ -857,7 +864,7 @@ void f2fs_evict_inode(struct inode *inode)
> >>>>>>>>         stat_dec_inline_inode(inode);
> >>>>>>>>         stat_dec_compr_inode(inode);
> >>>>>>>>         stat_sub_compr_blocks(inode,
> >>>>>>>> -                     atomic_read(&F2FS_I(inode)->i_compr_blocks));
> >>>>>>>> +                     atomic_read(&fi->i_compr_blocks));
> >>>>>>>>
> >>>>>>>>         if (likely(!f2fs_cp_error(sbi) &&
> >>>>>>>>                                 !is_sbi_flag_set(sbi, SBI_CP_DISABLED)))
> >>>>>>>> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> >>>>>>>> index ae3c4e5474ef..536d7c674b04 100644
> >>>>>>>> --- a/fs/f2fs/segment.c
> >>>>>>>> +++ b/fs/f2fs/segment.c
> >>>>>>>> @@ -192,9 +192,6 @@ void f2fs_abort_atomic_write(struct inode *inode, bool clean)
> >>>>>>>>         if (!f2fs_is_atomic_file(inode))
> >>>>>>>>                 return;
> >>>>>>>>
> >>>>>>>> -     clear_inode_flag(fi->cow_inode, FI_COW_FILE);
> >>>>>>>> -     iput(fi->cow_inode);
> >>>>>>>> -     fi->cow_inode = NULL;
> >>>>>>>>         release_atomic_write_cnt(inode);
> >>>>>>>>         clear_inode_flag(inode, FI_ATOMIC_COMMITTED);
> >>>>>>>>         clear_inode_flag(inode, FI_ATOMIC_REPLACE);
> >>>>>>>> diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> >>>>>>>> index 1f812b9ce985..10463f084d30 100644
> >>>>>>>> --- a/fs/f2fs/super.c
> >>>>>>>> +++ b/fs/f2fs/super.c
> >>>>>>>> @@ -1430,8 +1430,6 @@ static int f2fs_drop_inode(struct inode *inode)
> >>>>>>>>                         atomic_inc(&inode->i_count);
> >>>>>>>>                         spin_unlock(&inode->i_lock);
> >>>>>>>>
> >>>>>>>> -                     f2fs_abort_atomic_write(inode, true);
> >>>>>>>
> >>>>>>> In order to avoid caching obsolete page of cow_inode, how about truncating
> >>>>>>> them here?
> >>>>>>>
> >>>>>>> if (f2fs_is_atomic_file() && cow_inode)
> >>>>>>>           truncate_inode_pages_final(&cow_inode->i_data);
> >>>>>>>
> >>>>>>> Thanks,
> >>>>>>>
> >>>>>>>> -
> >>>>>>>>                         /* should remain fi->extent_tree for writepage */
> >>>>>>>>                         f2fs_destroy_extent_node(inode);
> >>>>>>>>

Powered by blists - more mailing lists