Hi, all.
Recently syzbot reported a bug as following:
kernel BUG at fs/f2fs/inode.c:896! CPU: 1 UID: 0 PID: 5217 Comm: syz-executor605 Not tainted 6.11.0-rc4-syzkaller-00033-g872cf28b8df9 #0 RIP: 0010:f2fs_evict_inode+0x1598/0x15c0 fs/f2fs/inode.c:896 Call Trace: <TASK> evict+0x532/0x950 fs/inode.c:704 dispose_list fs/inode.c:747 [inline] evict_inodes+0x5f9/0x690 fs/inode.c:797 generic_shutdown_super+0x9d/0x2d0 fs/super.c:627 kill_block_super+0x44/0x90 fs/super.c:1696 kill_f2fs_super+0x344/0x690 fs/f2fs/super.c:4898 deactivate_locked_super+0xc4/0x130 fs/super.c:473 cleanup_mnt+0x41f/0x4b0 fs/namespace.c:1373 task_work_run+0x24f/0x310 kernel/task_work.c:228 ptrace_notify+0x2d2/0x380 kernel/signal.c:2402 ptrace_report_syscall include/linux/ptrace.h:415 [inline] ptrace_report_syscall_exit include/linux/ptrace.h:477 [inline] syscall_exit_work+0xc6/0x190 kernel/entry/common.c:173 syscall_exit_to_user_mode_prepare kernel/entry/common.c:200 [inline] __syscall_exit_to_user_mode_work kernel/entry/common.c:205 [inline] syscall_exit_to_user_mode+0x279/0x370 kernel/entry/common.c:218 do_syscall_64+0x100/0x230 arch/x86/entry/common.c:89 entry_SYSCALL_64_after_hwframe+0x77/0x7f
The syzbot constructed the following scenario: concurrently creating directories and setting the file system to read-only. In this case, while f2fs was making dir, the filesystem switched to readonly, and when it tried to clear the dirty flag, it triggered this code path: f2fs_mkdir()-> f2fs_sync_fs()->f2fs_write_checkpoint() ->f2fs_readonly(). This resulted FI_DIRTY_INODE flag not being cleared, which eventually led to a bug being triggered during the FI_DIRTY_INODE check in f2fs_evict_inode().
In this case, we cannot do anything further, so if filesystem is readonly, do not trigger the BUG. Instead, clean up resources to the best of our ability to prevent triggering subsequent resource leak checks.
If there is anything important I'm missing, please let me know, thanks.
Reported-by: syzbot+ebea2790904673d7c618@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=ebea2790904673d7c618 Fixes: ca7d802a7d8e ("f2fs: detect dirty inode in evict_inode") CC: stable@vger.kernel.org Signed-off-by: Julian Sun sunjunchao2870@gmail.com --- fs/f2fs/inode.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/fs/f2fs/inode.c b/fs/f2fs/inode.c index aef57172014f..ebf825dba0a5 100644 --- a/fs/f2fs/inode.c +++ b/fs/f2fs/inode.c @@ -892,7 +892,8 @@ void f2fs_evict_inode(struct inode *inode) atomic_read(&fi->i_compr_blocks));
if (likely(!f2fs_cp_error(sbi) && - !is_sbi_flag_set(sbi, SBI_CP_DISABLED))) + !is_sbi_flag_set(sbi, SBI_CP_DISABLED)) && + !f2fs_readonly(sbi->sb)) f2fs_bug_on(sbi, is_inode_flag_set(inode, FI_DIRTY_INODE)); else f2fs_inode_synced(inode);
On 2024/8/29 0:54, Julian Sun wrote:
Hi, all.
Recently syzbot reported a bug as following:
kernel BUG at fs/f2fs/inode.c:896! CPU: 1 UID: 0 PID: 5217 Comm: syz-executor605 Not tainted 6.11.0-rc4-syzkaller-00033-g872cf28b8df9 #0 RIP: 0010:f2fs_evict_inode+0x1598/0x15c0 fs/f2fs/inode.c:896 Call Trace:
<TASK> evict+0x532/0x950 fs/inode.c:704 dispose_list fs/inode.c:747 [inline] evict_inodes+0x5f9/0x690 fs/inode.c:797 generic_shutdown_super+0x9d/0x2d0 fs/super.c:627 kill_block_super+0x44/0x90 fs/super.c:1696 kill_f2fs_super+0x344/0x690 fs/f2fs/super.c:4898 deactivate_locked_super+0xc4/0x130 fs/super.c:473 cleanup_mnt+0x41f/0x4b0 fs/namespace.c:1373 task_work_run+0x24f/0x310 kernel/task_work.c:228 ptrace_notify+0x2d2/0x380 kernel/signal.c:2402 ptrace_report_syscall include/linux/ptrace.h:415 [inline] ptrace_report_syscall_exit include/linux/ptrace.h:477 [inline] syscall_exit_work+0xc6/0x190 kernel/entry/common.c:173 syscall_exit_to_user_mode_prepare kernel/entry/common.c:200 [inline] __syscall_exit_to_user_mode_work kernel/entry/common.c:205 [inline] syscall_exit_to_user_mode+0x279/0x370 kernel/entry/common.c:218 do_syscall_64+0x100/0x230 arch/x86/entry/common.c:89 entry_SYSCALL_64_after_hwframe+0x77/0x7f
The syzbot constructed the following scenario: concurrently creating directories and setting the file system to read-only. In this case, while f2fs was making dir, the filesystem switched to readonly, and when it tried to clear the dirty flag, it triggered this code path: f2fs_mkdir()-> f2fs_sync_fs()->f2fs_write_checkpoint() ->f2fs_readonly(). This resulted FI_DIRTY_INODE flag not being cleared, which eventually led to a bug being triggered during the FI_DIRTY_INODE check in f2fs_evict_inode().
In this case, we cannot do anything further, so if filesystem is readonly, do not trigger the BUG. Instead, clean up resources to the best of our ability to prevent triggering subsequent resource leak checks.
If there is anything important I'm missing, please let me know, thanks.
Reported-by: syzbot+ebea2790904673d7c618@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=ebea2790904673d7c618 Fixes: ca7d802a7d8e ("f2fs: detect dirty inode in evict_inode") CC: stable@vger.kernel.org Signed-off-by: Julian Sun sunjunchao2870@gmail.com
fs/f2fs/inode.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/fs/f2fs/inode.c b/fs/f2fs/inode.c index aef57172014f..ebf825dba0a5 100644 --- a/fs/f2fs/inode.c +++ b/fs/f2fs/inode.c @@ -892,7 +892,8 @@ void f2fs_evict_inode(struct inode *inode) atomic_read(&fi->i_compr_blocks)); if (likely(!f2fs_cp_error(sbi) &&
!is_sbi_flag_set(sbi, SBI_CP_DISABLED)))
!is_sbi_flag_set(sbi, SBI_CP_DISABLED)) &&
!f2fs_readonly(sbi->sb))
Is it fine to drop this dirty inode? Since once it remounts f2fs as rw one, previous updates on such inode may be lost? Or am I missing something?
Thanks,
f2fs_bug_on(sbi, is_inode_flag_set(inode, FI_DIRTY_INODE));
else f2fs_inode_synced(inode);
On Mon, 2024-09-02 at 16:13 +0800, Chao Yu wrote:
On 2024/8/29 0:54, Julian Sun wrote:
Hi, all.
Recently syzbot reported a bug as following:
kernel BUG at fs/f2fs/inode.c:896! CPU: 1 UID: 0 PID: 5217 Comm: syz-executor605 Not tainted 6.11.0-rc4-syzkaller-00033-g872cf28b8df9 #0 RIP: 0010:f2fs_evict_inode+0x1598/0x15c0 fs/f2fs/inode.c:896 Call Trace: <TASK> evict+0x532/0x950 fs/inode.c:704 dispose_list fs/inode.c:747 [inline] evict_inodes+0x5f9/0x690 fs/inode.c:797 generic_shutdown_super+0x9d/0x2d0 fs/super.c:627 kill_block_super+0x44/0x90 fs/super.c:1696 kill_f2fs_super+0x344/0x690 fs/f2fs/super.c:4898 deactivate_locked_super+0xc4/0x130 fs/super.c:473 cleanup_mnt+0x41f/0x4b0 fs/namespace.c:1373 task_work_run+0x24f/0x310 kernel/task_work.c:228 ptrace_notify+0x2d2/0x380 kernel/signal.c:2402 ptrace_report_syscall include/linux/ptrace.h:415 [inline] ptrace_report_syscall_exit include/linux/ptrace.h:477 [inline] syscall_exit_work+0xc6/0x190 kernel/entry/common.c:173 syscall_exit_to_user_mode_prepare kernel/entry/common.c:200 [inline] __syscall_exit_to_user_mode_work kernel/entry/common.c:205 [inline] syscall_exit_to_user_mode+0x279/0x370 kernel/entry/common.c:218 do_syscall_64+0x100/0x230 arch/x86/entry/common.c:89 entry_SYSCALL_64_after_hwframe+0x77/0x7f
The syzbot constructed the following scenario: concurrently creating directories and setting the file system to read-only. In this case, while f2fs was making dir, the filesystem switched to readonly, and when it tried to clear the dirty flag, it triggered this code path: f2fs_mkdir()-> f2fs_sync_fs()-
f2fs_write_checkpoint()
->f2fs_readonly(). This resulted FI_DIRTY_INODE flag not being cleared, which eventually led to a bug being triggered during the FI_DIRTY_INODE check in f2fs_evict_inode().
In this case, we cannot do anything further, so if filesystem is readonly, do not trigger the BUG. Instead, clean up resources to the best of our ability to prevent triggering subsequent resource leak checks.
If there is anything important I'm missing, please let me know, thanks.
Reported-by: syzbot+ebea2790904673d7c618@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=ebea2790904673d7c618 Fixes: ca7d802a7d8e ("f2fs: detect dirty inode in evict_inode") CC: stable@vger.kernel.org Signed-off-by: Julian Sun sunjunchao2870@gmail.com
fs/f2fs/inode.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/fs/f2fs/inode.c b/fs/f2fs/inode.c index aef57172014f..ebf825dba0a5 100644 --- a/fs/f2fs/inode.c +++ b/fs/f2fs/inode.c @@ -892,7 +892,8 @@ void f2fs_evict_inode(struct inode *inode) atomic_read(&fi->i_compr_blocks)); if (likely(!f2fs_cp_error(sbi) && - !is_sbi_flag_set(sbi, SBI_CP_DISABLED))) + !is_sbi_flag_set(sbi, SBI_CP_DISABLED)) && + !f2fs_readonly(sbi->sb))
Is it fine to drop this dirty inode? Since once it remounts f2fs as rw one, previous updates on such inode may be lost? Or am I missing something?
The purpose of calling this here is mainly to avoid triggering the f2fs_bug_on(sbi, 1); statement in the subsequent f2fs_put_super() due to a reference count check failure. I would say it's possible, but there doesn't seem to be much more we can do in this scenario: the inode is about to be freed, and the file system is read-only. Or do we need a mechanism to save the inode that is about to be freed and then write it back to disk at the appropriate time after the file system becomes rw again? But such a mechanism sounds somewhat complex and a little bit of weird... Do you have any suggestions?
Thanks,
f2fs_bug_on(sbi, is_inode_flag_set(inode, FI_DIRTY_INODE)); else f2fs_inode_synced(inode);
Thanks,
On 2024/9/2 21:01, Julian Sun wrote:
On Mon, 2024-09-02 at 16:13 +0800, Chao Yu wrote:
On 2024/8/29 0:54, Julian Sun wrote:
Hi, all.
Recently syzbot reported a bug as following:
kernel BUG at fs/f2fs/inode.c:896! CPU: 1 UID: 0 PID: 5217 Comm: syz-executor605 Not tainted 6.11.0-rc4-syzkaller-00033-g872cf28b8df9 #0 RIP: 0010:f2fs_evict_inode+0x1598/0x15c0 fs/f2fs/inode.c:896 Call Trace: <TASK> evict+0x532/0x950 fs/inode.c:704 dispose_list fs/inode.c:747 [inline] evict_inodes+0x5f9/0x690 fs/inode.c:797 generic_shutdown_super+0x9d/0x2d0 fs/super.c:627 kill_block_super+0x44/0x90 fs/super.c:1696 kill_f2fs_super+0x344/0x690 fs/f2fs/super.c:4898 deactivate_locked_super+0xc4/0x130 fs/super.c:473 cleanup_mnt+0x41f/0x4b0 fs/namespace.c:1373 task_work_run+0x24f/0x310 kernel/task_work.c:228 ptrace_notify+0x2d2/0x380 kernel/signal.c:2402 ptrace_report_syscall include/linux/ptrace.h:415 [inline] ptrace_report_syscall_exit include/linux/ptrace.h:477 [inline] syscall_exit_work+0xc6/0x190 kernel/entry/common.c:173 syscall_exit_to_user_mode_prepare kernel/entry/common.c:200 [inline] __syscall_exit_to_user_mode_work kernel/entry/common.c:205 [inline] syscall_exit_to_user_mode+0x279/0x370 kernel/entry/common.c:218 do_syscall_64+0x100/0x230 arch/x86/entry/common.c:89 entry_SYSCALL_64_after_hwframe+0x77/0x7f
The syzbot constructed the following scenario: concurrently creating directories and setting the file system to read-only. In this case, while f2fs was making dir, the filesystem switched to readonly, and when it tried to clear the dirty flag, it triggered
Go back to the root cause, I have no idea why it can leave dirty inode while mkdir races w/ readonly remount, due to the two operations should be exclusive, IIUC.
- mkdir - do_mkdirat - filename_create - mnt_want_write - mnt_get_write_access - mount - do_remount - reconfigure_super - sb_prepare_remount_readonly - mnt_hold_writers - vfs_mkdir - f2fs_mkdir
But when I try to reproduce this bug w/ reproducer provided by syzbot, I have found a clue in the log:
"skip recovering inline_dots inode..."
So I doubt the root cause is __recover_dot_dentries() in f2fs_lookup() generates dirty data/meta, in this path, we will not grab related lock to exclude readonly remount.
Let me try to verify below patch:
https://git.kernel.org/pub/scm/linux/kernel/git/chao/linux.git/commit/?h=wip...
Thanks,
this code path: f2fs_mkdir()-> f2fs_sync_fs()-
f2fs_write_checkpoint()
->f2fs_readonly(). This resulted FI_DIRTY_INODE flag not being cleared, which eventually led to a bug being triggered during the FI_DIRTY_INODE check in f2fs_evict_inode().
In this case, we cannot do anything further, so if filesystem is readonly, do not trigger the BUG. Instead, clean up resources to the best of our ability to prevent triggering subsequent resource leak checks.
If there is anything important I'm missing, please let me know, thanks.
Reported-by: syzbot+ebea2790904673d7c618@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=ebea2790904673d7c618 Fixes: ca7d802a7d8e ("f2fs: detect dirty inode in evict_inode") CC: stable@vger.kernel.org Signed-off-by: Julian Sun sunjunchao2870@gmail.com
fs/f2fs/inode.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/fs/f2fs/inode.c b/fs/f2fs/inode.c index aef57172014f..ebf825dba0a5 100644 --- a/fs/f2fs/inode.c +++ b/fs/f2fs/inode.c @@ -892,7 +892,8 @@ void f2fs_evict_inode(struct inode *inode) atomic_read(&fi->i_compr_blocks)); if (likely(!f2fs_cp_error(sbi) && - !is_sbi_flag_set(sbi, SBI_CP_DISABLED))) + !is_sbi_flag_set(sbi, SBI_CP_DISABLED)) && + !f2fs_readonly(sbi->sb))
Is it fine to drop this dirty inode? Since once it remounts f2fs as rw one, previous updates on such inode may be lost? Or am I missing something?
The purpose of calling this here is mainly to avoid triggering the f2fs_bug_on(sbi, 1); statement in the subsequent f2fs_put_super() due to a reference count check failure. I would say it's possible, but there doesn't seem to be much more we can do in this scenario: the inode is about to be freed, and the file system is read-only. Or do we need a mechanism to save the inode that is about to be freed and then write it back to disk at the appropriate time after the file system becomes rw again? But such a mechanism sounds somewhat complex and a little bit of weird... Do you have any suggestions?
Thanks,
f2fs_bug_on(sbi, is_inode_flag_set(inode, FI_DIRTY_INODE)); else f2fs_inode_synced(inode);
Thanks,
On 09/03, Chao Yu wrote:
On 2024/9/2 21:01, Julian Sun wrote:
On Mon, 2024-09-02 at 16:13 +0800, Chao Yu wrote:
On 2024/8/29 0:54, Julian Sun wrote:
Hi, all.
Recently syzbot reported a bug as following:
kernel BUG at fs/f2fs/inode.c:896! CPU: 1 UID: 0 PID: 5217 Comm: syz-executor605 Not tainted 6.11.0-rc4-syzkaller-00033-g872cf28b8df9 #0 RIP: 0010:f2fs_evict_inode+0x1598/0x15c0 fs/f2fs/inode.c:896 Call Trace: <TASK> evict+0x532/0x950 fs/inode.c:704 dispose_list fs/inode.c:747 [inline] evict_inodes+0x5f9/0x690 fs/inode.c:797 generic_shutdown_super+0x9d/0x2d0 fs/super.c:627 kill_block_super+0x44/0x90 fs/super.c:1696 kill_f2fs_super+0x344/0x690 fs/f2fs/super.c:4898 deactivate_locked_super+0xc4/0x130 fs/super.c:473 cleanup_mnt+0x41f/0x4b0 fs/namespace.c:1373 task_work_run+0x24f/0x310 kernel/task_work.c:228 ptrace_notify+0x2d2/0x380 kernel/signal.c:2402 ptrace_report_syscall include/linux/ptrace.h:415 [inline] ptrace_report_syscall_exit include/linux/ptrace.h:477 [inline] syscall_exit_work+0xc6/0x190 kernel/entry/common.c:173 syscall_exit_to_user_mode_prepare kernel/entry/common.c:200 [inline] __syscall_exit_to_user_mode_work kernel/entry/common.c:205 [inline] syscall_exit_to_user_mode+0x279/0x370 kernel/entry/common.c:218 do_syscall_64+0x100/0x230 arch/x86/entry/common.c:89 entry_SYSCALL_64_after_hwframe+0x77/0x7f
The syzbot constructed the following scenario: concurrently creating directories and setting the file system to read-only. In this case, while f2fs was making dir, the filesystem switched to readonly, and when it tried to clear the dirty flag, it triggered
Go back to the root cause, I have no idea why it can leave dirty inode while mkdir races w/ readonly remount, due to the two operations should be exclusive, IIUC.
Wait, we can think of writable disk mounted as fs-readonly. In that case, IIRC, we allow to recover files/data by roll-forward and so on, which can make some dirty inodes. Can we check if there's any missing path which does not flush dirty inode?
- mkdir
- do_mkdirat
- filename_create
- mnt_want_write
- mnt_get_write_access
- mount - do_remount - reconfigure_super - sb_prepare_remount_readonly - mnt_hold_writers
- vfs_mkdir
- f2fs_mkdir
But when I try to reproduce this bug w/ reproducer provided by syzbot, I have found a clue in the log:
"skip recovering inline_dots inode..."
So I doubt the root cause is __recover_dot_dentries() in f2fs_lookup() generates dirty data/meta, in this path, we will not grab related lock to exclude readonly remount.
Let me try to verify below patch:
https://git.kernel.org/pub/scm/linux/kernel/git/chao/linux.git/commit/?h=wip...
Thanks,
this code path: f2fs_mkdir()-> f2fs_sync_fs()- > f2fs_write_checkpoint() ->f2fs_readonly(). This resulted FI_DIRTY_INODE flag not being cleared, which eventually led to a bug being triggered during the FI_DIRTY_INODE check in f2fs_evict_inode().
In this case, we cannot do anything further, so if filesystem is readonly, do not trigger the BUG. Instead, clean up resources to the best of our ability to prevent triggering subsequent resource leak checks.
If there is anything important I'm missing, please let me know, thanks.
Reported-by: syzbot+ebea2790904673d7c618@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=ebea2790904673d7c618 Fixes: ca7d802a7d8e ("f2fs: detect dirty inode in evict_inode") CC: stable@vger.kernel.org Signed-off-by: Julian Sun sunjunchao2870@gmail.com
fs/f2fs/inode.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/fs/f2fs/inode.c b/fs/f2fs/inode.c index aef57172014f..ebf825dba0a5 100644 --- a/fs/f2fs/inode.c +++ b/fs/f2fs/inode.c @@ -892,7 +892,8 @@ void f2fs_evict_inode(struct inode *inode) atomic_read(&fi->i_compr_blocks)); if (likely(!f2fs_cp_error(sbi) && - !is_sbi_flag_set(sbi, SBI_CP_DISABLED))) + !is_sbi_flag_set(sbi, SBI_CP_DISABLED)) && + !f2fs_readonly(sbi->sb))
Is it fine to drop this dirty inode? Since once it remounts f2fs as rw one, previous updates on such inode may be lost? Or am I missing something?
The purpose of calling this here is mainly to avoid triggering the f2fs_bug_on(sbi, 1); statement in the subsequent f2fs_put_super() due to a reference count check failure. I would say it's possible, but there doesn't seem to be much more we can do in this scenario: the inode is about to be freed, and the file system is read-only. Or do we need a mechanism to save the inode that is about to be freed and then write it back to disk at the appropriate time after the file system becomes rw again? But such a mechanism sounds somewhat complex and a little bit of weird... Do you have any suggestions?
Thanks,
f2fs_bug_on(sbi, is_inode_flag_set(inode, FI_DIRTY_INODE)); else f2fs_inode_synced(inode);
Thanks,
On 2024/9/4 5:20, Jaegeuk Kim wrote:
On 09/03, Chao Yu wrote:
On 2024/9/2 21:01, Julian Sun wrote:
On Mon, 2024-09-02 at 16:13 +0800, Chao Yu wrote:
On 2024/8/29 0:54, Julian Sun wrote:
> Hi, all. > > Recently syzbot reported a bug as following: > > kernel BUG at fs/f2fs/inode.c:896! > CPU: 1 UID: 0 PID: 5217 Comm: syz-executor605 Not tainted > 6.11.0-rc4-syzkaller-00033-g872cf28b8df9 #0 > RIP: 0010:f2fs_evict_inode+0x1598/0x15c0 fs/f2fs/inode.c:896 > Call Trace: > <TASK> > evict+0x532/0x950 fs/inode.c:704 > dispose_list fs/inode.c:747 [inline] > evict_inodes+0x5f9/0x690 fs/inode.c:797 > generic_shutdown_super+0x9d/0x2d0 fs/super.c:627 > kill_block_super+0x44/0x90 fs/super.c:1696 > kill_f2fs_super+0x344/0x690 fs/f2fs/super.c:4898 > deactivate_locked_super+0xc4/0x130 fs/super.c:473 > cleanup_mnt+0x41f/0x4b0 fs/namespace.c:1373 > task_work_run+0x24f/0x310 kernel/task_work.c:228 > ptrace_notify+0x2d2/0x380 kernel/signal.c:2402 > ptrace_report_syscall include/linux/ptrace.h:415 [inline] > ptrace_report_syscall_exit include/linux/ptrace.h:477 > [inline] > syscall_exit_work+0xc6/0x190 kernel/entry/common.c:173 > syscall_exit_to_user_mode_prepare kernel/entry/common.c:200 > [inline] > __syscall_exit_to_user_mode_work kernel/entry/common.c:205 > [inline] > syscall_exit_to_user_mode+0x279/0x370 > kernel/entry/common.c:218 > do_syscall_64+0x100/0x230 arch/x86/entry/common.c:89 > entry_SYSCALL_64_after_hwframe+0x77/0x7f > > The syzbot constructed the following scenario: concurrently > creating directories and setting the file system to read-only. > In this case, while f2fs was making dir, the filesystem > switched to > readonly, and when it tried to clear the dirty flag, it > triggered
Go back to the root cause, I have no idea why it can leave dirty inode while mkdir races w/ readonly remount, due to the two operations should be exclusive, IIUC.
Wait, we can think of writable disk mounted as fs-readonly. In that case, IIRC, we allow to recover files/data by roll-forward and so on, which can
We will remove SB_RDONLY flag from sb->s_flags intentionally before recovery, so that following write_checkpoint() or sync_filesystem() won't skip flushing due to sb is readonly.
static bool f2fs_recover_quota_begin(struct f2fs_sb_info *sbi) { ... if (readonly) { sbi->sb->s_flags &= ~SB_RDONLY; set_sbi_flag(sbi, SBI_IS_WRITABLE); } ... }
make some dirty inodes. Can we check if there's any missing path which does not flush dirty inode?
I guess the root cause of this issue is like this:
- f2fs_lookup - __recover_dot_dentries() - clear_inode_flag(dir, FI_INLINE_DOTS) - __mark_inode_dirty_flag() - remount rdonly - sb->s_flags |= SB_RDONLY
- umount - kill_f2fs_super - kill_block_super - generic_shutdown_super - sync_filesystem skips due to sb_rdonly is true - evict_inodes - dispose_list - f2fs_evict_inode panic
So how about this?
https://git.kernel.org/pub/scm/linux/kernel/git/chao/linux.git/commit/?h=wip...
Thanks,
- mkdir
- do_mkdirat
- filename_create
- mnt_want_write
- mnt_get_write_access
- mount - do_remount - reconfigure_super - sb_prepare_remount_readonly - mnt_hold_writers
- vfs_mkdir
- f2fs_mkdir
But when I try to reproduce this bug w/ reproducer provided by syzbot, I have found a clue in the log:
"skip recovering inline_dots inode..."
So I doubt the root cause is __recover_dot_dentries() in f2fs_lookup() generates dirty data/meta, in this path, we will not grab related lock to exclude readonly remount.
Let me try to verify below patch:
https://git.kernel.org/pub/scm/linux/kernel/git/chao/linux.git/commit/?h=wip...
Thanks,
> this > code path: f2fs_mkdir()-> f2fs_sync_fs()- >> f2fs_write_checkpoint() > ->f2fs_readonly(). This resulted FI_DIRTY_INODE flag not being > cleared, > which eventually led to a bug being triggered during the > FI_DIRTY_INODE > check in f2fs_evict_inode(). > > In this case, we cannot do anything further, so if filesystem > is > readonly, > do not trigger the BUG. Instead, clean up resources to the best > of > our > ability to prevent triggering subsequent resource leak checks. > > If there is anything important I'm missing, please let me know, > thanks. > > Reported-by: > syzbot+ebea2790904673d7c618@syzkaller.appspotmail.com > Closes: > https://syzkaller.appspot.com/bug?extid=ebea2790904673d7c618 > Fixes: ca7d802a7d8e ("f2fs: detect dirty inode in evict_inode") > CC: stable@vger.kernel.org > Signed-off-by: Julian Sun sunjunchao2870@gmail.com > --- > fs/f2fs/inode.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/fs/f2fs/inode.c b/fs/f2fs/inode.c > index aef57172014f..ebf825dba0a5 100644 > --- a/fs/f2fs/inode.c > +++ b/fs/f2fs/inode.c > @@ -892,7 +892,8 @@ void f2fs_evict_inode(struct inode *inode) > atomic_read(&fi->i_compr_blocks)); > if (likely(!f2fs_cp_error(sbi) && > - !is_sbi_flag_set(sbi, > SBI_CP_DISABLED))) > + !is_sbi_flag_set(sbi, > SBI_CP_DISABLED)) && > + !f2fs_readonly(sbi->sb))
Is it fine to drop this dirty inode? Since once it remounts f2fs as rw one, previous updates on such inode may be lost? Or am I missing something?
The purpose of calling this here is mainly to avoid triggering the f2fs_bug_on(sbi, 1); statement in the subsequent f2fs_put_super() due to a reference count check failure. I would say it's possible, but there doesn't seem to be much more we can do in this scenario: the inode is about to be freed, and the file system is read-only. Or do we need a mechanism to save the inode that is about to be freed and then write it back to disk at the appropriate time after the file system becomes rw again? But such a mechanism sounds somewhat complex and a little bit of weird... Do you have any suggestions?
Thanks,
> f2fs_bug_on(sbi, is_inode_flag_set(inode, > FI_DIRTY_INODE)); > else > f2fs_inode_synced(inode);
Thanks,
On Mon, 2024-09-02 at 16:13 +0800, Chao Yu wrote:
On 2024/8/29 0:54, Julian Sun wrote:
Hi, all.
Recently syzbot reported a bug as following:
kernel BUG at fs/f2fs/inode.c:896! CPU: 1 UID: 0 PID: 5217 Comm: syz-executor605 Not tainted 6.11.0- rc4-syzkaller-00033-g872cf28b8df9 #0 RIP: 0010:f2fs_evict_inode+0x1598/0x15c0 fs/f2fs/inode.c:896 Call Trace: <TASK> evict+0x532/0x950 fs/inode.c:704 dispose_list fs/inode.c:747 [inline] evict_inodes+0x5f9/0x690 fs/inode.c:797 generic_shutdown_super+0x9d/0x2d0 fs/super.c:627 kill_block_super+0x44/0x90 fs/super.c:1696 kill_f2fs_super+0x344/0x690 fs/f2fs/super.c:4898 deactivate_locked_super+0xc4/0x130 fs/super.c:473 cleanup_mnt+0x41f/0x4b0 fs/namespace.c:1373 task_work_run+0x24f/0x310 kernel/task_work.c:228 ptrace_notify+0x2d2/0x380 kernel/signal.c:2402 ptrace_report_syscall include/linux/ptrace.h:415 [inline] ptrace_report_syscall_exit include/linux/ptrace.h:477 [inline] syscall_exit_work+0xc6/0x190 kernel/entry/common.c:173 syscall_exit_to_user_mode_prepare kernel/entry/common.c:200 [inline] __syscall_exit_to_user_mode_work kernel/entry/common.c:205 [inline] syscall_exit_to_user_mode+0x279/0x370 kernel/entry/common.c:218 do_syscall_64+0x100/0x230 arch/x86/entry/common.c:89 entry_SYSCALL_64_after_hwframe+0x77/0x7f
The syzbot constructed the following scenario: concurrently creating directories and setting the file system to read-only. In this case, while f2fs was making dir, the filesystem switched to readonly, and when it tried to clear the dirty flag, it triggered this code path: f2fs_mkdir()-> f2fs_sync_fs()->f2fs_write_checkpoint() ->f2fs_readonly(). This resulted FI_DIRTY_INODE flag not being cleared, which eventually led to a bug being triggered during the FI_DIRTY_INODE check in f2fs_evict_inode().
In this case, we cannot do anything further, so if filesystem is readonly, do not trigger the BUG. Instead, clean up resources to the best of our ability to prevent triggering subsequent resource leak checks.
If there is anything important I'm missing, please let me know, thanks.
Reported-by: syzbot+ebea2790904673d7c618@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=ebea2790904673d7c618 Fixes: ca7d802a7d8e ("f2fs: detect dirty inode in evict_inode") CC: stable@vger.kernel.org Signed-off-by: Julian Sun sunjunchao2870@gmail.com
fs/f2fs/inode.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/fs/f2fs/inode.c b/fs/f2fs/inode.c index aef57172014f..ebf825dba0a5 100644 --- a/fs/f2fs/inode.c +++ b/fs/f2fs/inode.c @@ -892,7 +892,8 @@ void f2fs_evict_inode(struct inode *inode) atomic_read(&fi->i_compr_blocks)); if (likely(!f2fs_cp_error(sbi) && - !is_sbi_flag_set(sbi, SBI_CP_DISABLED))) + !is_sbi_flag_set(sbi, SBI_CP_DISABLED)) && + !f2fs_readonly(sbi->sb))
Is it fine to drop this dirty inode? Since once it remounts f2fs as rw one, previous updates on such inode may be lost? Or am I missing something?
Hi, Chao.
I believe the issue you pointed out goes beyond the scope of the problem reported by syzbot, as I have seen this issue in some existing code as well, such as in the handling of read-only file systems in f2fs_mark_inode_dirty_sync(), f2fs_write_checkpoint(), f2fs_do_sync_file(), etc.
If you do believe the problem needs to be fixed, we can discuss it further, but it should be addressed in a separate patch.
If there is anything important I'm missing, please let me know, thanks.
Thanks,
f2fs_bug_on(sbi, is_inode_flag_set(inode, FI_DIRTY_INODE)); else f2fs_inode_synced(inode);
Thanks,
linux-stable-mirror@lists.linaro.org