lkml.org 
[lkml]   [2022]   [Feb]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    Patch in this message
    /
    From
    Subject[RFC 9/9] ext4: fast_commit missing tracking updates to a file
    Date
    <DO NOT MERGE THIS YET>

    Testcase
    ==========
    1. i=0; while [ $i -lt 1000 ]; do xfs_io -f -c "pwrite -S 0xaa -b 32k 0 32k" -c "fsync" /mnt/$i; i=$(($i+1)); done && sudo ./src/godown -v /mnt && sudo umount /mnt && sudo mount /dev/loop2 /mnt'
    2. ls -alih /mnt/ -> In this you will observe one such file with 0 bytes (which ideally should not happen)

    ^^^ say if you don't see the issue because your underlying storage
    device is very fast, then maybe try with commit=1 mount option.

    Analysis
    ==========
    It seems a file's updates can be a part of two transaction tid.
    Below are the sequence of events which could cause this issue.

    jbd2_handle_start -> (t_tid = 38)
    __ext4_new_inode
    ext4_fc_track_template -> __track_inode -> (i_sync_tid = 38, t_tid = 38)
    <track more updates>
    jbd2_start_commit -> (t_tid = 38)

    jbd2_handle_start (tid = 39)
    ext4_fc_track_template -> __track_inode -> (i_sync_tid = 38, t_tid 39)
    -> ext4_fc_reset_inode & ei->i_sync_tid = t_tid

    ext4_fc_commit_start -> (will wait since jbd2 full commit is in progress)
    jbd2_end_commit (t_tid = 38)
    -> jbd2_fc_cleanup() -> this will cleanup entries in sbi->s_fc_q[FC_Q_MAIN]
    -> And the above could result inode size as 0 as after effect.
    ext4_fc_commit_stop

    You could find the logs for the above behavior for inode 979 at [1].

    -> So what is happening here is since the ei->i_fc_list is not empty
    (because it is already part of sb's MAIN queue), we don't add this inode
    again into neither sb's MAIN or STAGING queue.
    And after jbd2_fc_cleanup() is called from jbd2 full commit, we
    just remove this inode from the main queue.

    So as a simple fix, what I did below was to check if it is a jbd2 full commit
    in ext4_fc_cleanup(), and if the ei->i_sync_tid > tid, that means we
    need not remove that from MAIN queue. This is since neither jbd2 nor FC
    has committed updates of those inodes for this new txn tid yet.

    But below are some quick queries on this
    =========================================

    1. why do we call ext4_fc_reset_inode() when inode tid and
    running txn tid does not match?

    2. Also is this an expected behavior from the design perspective of
    fast_commit. i.e.
    a. the inode can be part of two tids?
    b. And that while a full commit is in progress, the inode can still
    receive updates but using a new transaction tid.

    Frankly speaking, since I was also working on other things, so I haven't
    yet got the chance to completely analyze the situation yet.
    Once I have those things sorted, I will spend more time on this, to
    understand it more. Meanwhile if you already have some answers to above
    queries/observations, please do share those here.

    Links
    =========
    [1] https://raw.githubusercontent.com/riteshharjani/LinuxStudy/master/ext4/fast_commit/fc_inode_missing_updates_ino_979.txt

    Signed-off-by: Ritesh Harjani <riteshh@linux.ibm.com>
    ---
    fs/ext4/fast_commit.c | 2 ++
    1 file changed, 2 insertions(+)

    diff --git a/fs/ext4/fast_commit.c b/fs/ext4/fast_commit.c
    index 8803ba087b07..769b584c2552 100644
    --- a/fs/ext4/fast_commit.c
    +++ b/fs/ext4/fast_commit.c
    @@ -1252,6 +1252,8 @@ static void ext4_fc_cleanup(journal_t *journal, int full, tid_t tid)
    spin_lock(&sbi->s_fc_lock);
    list_for_each_entry_safe(iter, iter_n, &sbi->s_fc_q[FC_Q_MAIN],
    i_fc_list) {
    + if (full && iter->i_sync_tid > tid)
    + continue;
    list_del_init(&iter->i_fc_list);
    ext4_clear_inode_state(&iter->vfs_inode,
    EXT4_STATE_FC_COMMITTING);
    --
    2.31.1
    \
     
     \ /
      Last update: 2022-02-22 21:36    [W:2.269 / U:0.044 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site