Re: [PATCH 1/5] btrfs: drop log root for dropped roots

[Josef Bacik <josef@toxicpanda.com>]

On 12/10/19 4:19 PM, Nikolay Borisov wrote:
> 
> 
> On 10.12.19 г. 22:05 ч., Josef Bacik wrote:
>> On 12/6/19 10:03 AM, Nikolay Borisov wrote:
>>>
>>>
>>> On 6.12.19 г. 16:37 ч., Josef Bacik wrote:
>>>> If we fsync on a subvolume and create a log root for that volume, and
>>>> then later delete that subvolume we'll never clean up its log root.  Fix
>>>> this by making switch_commit_roots free the log for any dropped roots we
>>>> encounter.
>>>>
>>>> Signed-off-by: Josef Bacik <josef@toxicpanda.com>
>>>> ---
>>>>    fs/btrfs/transaction.c | 22 ++++++++++++----------
>>>>    1 file changed, 12 insertions(+), 10 deletions(-)
>>>>
>>>> diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
>>>> index cfc08ef9b876..55d8fd68775a 100644
>>>> --- a/fs/btrfs/transaction.c
>>>> +++ b/fs/btrfs/transaction.c
>>>> @@ -147,13 +147,14 @@ void btrfs_put_transaction(struct
>>>> btrfs_transaction *transaction)
>>>>        }
>>>>    }
>>>>    -static noinline void switch_commit_roots(struct btrfs_transaction
>>>> *trans)
>>>> +static noinline void switch_commit_roots(struct btrfs_trans_handle
>>>> *trans)
>>>>    {
>>>> +    struct btrfs_transaction *cur_trans = trans->transaction;
>>>>        struct btrfs_fs_info *fs_info = trans->fs_info;
>>>>        struct btrfs_root *root, *tmp;
>>>>          down_write(&fs_info->commit_root_sem);
>>>> -    list_for_each_entry_safe(root, tmp, &trans->switch_commits,
>>>> +    list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits,
>>>>                     dirty_list) {
>>>>            list_del_init(&root->dirty_list);
>>>>            free_extent_buffer(root->commit_root);
>>>> @@ -165,16 +166,17 @@ static noinline void switch_commit_roots(struct
>>>> btrfs_transaction *trans)
>>>>        }
>>>>          /* We can free old roots now. */
>>>> -    spin_lock(&trans->dropped_roots_lock);
>>>> -    while (!list_empty(&trans->dropped_roots)) {
>>>> -        root = list_first_entry(&trans->dropped_roots,
>>>> +    spin_lock(&cur_trans->dropped_roots_lock);
>>>> +    while (!list_empty(&cur_trans->dropped_roots)) {
>>>> +        root = list_first_entry(&cur_trans->dropped_roots,
>>>>                        struct btrfs_root, root_list);
>>>>            list_del_init(&root->root_list);
>>>> -        spin_unlock(&trans->dropped_roots_lock);
>>>> +        spin_unlock(&cur_trans->dropped_roots_lock);
>>>> +        btrfs_free_log(trans, root);
>>>
>>> THis patch should really have been this line and converting
>>> switch_commit_roots to taking a trans handle another patch. Otherwise
>>> this is lost in the mechanical refactoring.
>>>
>>
>> We need the trans handle to even call btrfs_free_log, we're just fixing
>> it so the trans handle can be passed in, making its separate is just
>> superfluous.  Thanks,
> 
> Actually no because callees handle the case when trans is not passed
> (i.e. walk_log_tree and walk_(up|down)_log_tree. If passing valid
> trances changes the call logic then this needs to be explained in the
> changelog. And there is currently one caller calling that function
> without a trans - btrfs_drop_and_free_fs_root in BTRFS_FS_STATE_ERROR case.
> 

Yeah, it's clear that NULL is used only in the error case.  I'm not going to 
explain the entirety of how the log tree works in a basic fix for not freeing up 
a tree log when we should be doing it.  Thanks,

Josef