Re: [f2fs-dev] [PATCH] f2fs: synchronize atomic write aborts

From: Chao Yu
Date: Tue Jan 31 2023 - 20:40:25 EST


Hi Daeho,

On 2023/2/1 6:34, Daeho Jeong wrote:
On Tue, Jan 31, 2023 at 1:57 PM Daeho Jeong <daeho43@xxxxxxxxx> wrote:

On Tue, Jan 31, 2023 at 1:38 PM Daeho Jeong <daeho43@xxxxxxxxx> wrote:

On Tue, Jan 31, 2023 at 11:13 AM Daeho Jeong <daeho43@xxxxxxxxx> wrote:

Hi Chao,

On Tue, Jan 31, 2023 at 3:37 AM Chao Yu <chao@xxxxxxxxxx> 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.

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@xxxxxxxxxx> wrote:

Hi Daeho, Jaegeuk,

Please take a look at patchset in below link:

https://lore.kernel.org/linux-f2fs-devel/20230109034453.490176-1-chao@xxxxxxxxxx/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@xxxxxxxxxx>

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@xxxxxxxxxxxxxxxxxxxxxxxxx
Fixes: 3db1de0e582c ("f2fs: change the current atomic write way")
Signed-off-by: Daeho Jeong <daehojeong@xxxxxxxxxx>
---
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);