Re: [PATCH 6/8] btrfs: rework wake_all_tickets

[Nikolay Borisov <nborisov@suse.com>]

On 19.08.19 г. 18:06 ч., Josef Bacik wrote:
> On Mon, Aug 19, 2019 at 05:49:45PM +0300, Nikolay Borisov wrote:
>>
>>
>> On 16.08.19 г. 17:19 ч., Josef Bacik wrote:
>>> Now that we no longer partially fill tickets we need to rework
>>> wake_all_tickets to call btrfs_try_to_wakeup_tickets() in order to see
>>> if any subsequent tickets are able to be satisfied.  If our tickets_id
>>> changes we know something happened and we can keep flushing.
>>>
>>> Also if we find a ticket that is smaller than the first ticket in our
>>> queue then we want to retry the flushing loop again in case
>>> may_commit_transaction() decides we could satisfy the ticket by
>>> committing the transaction.
>>>
>>> Signed-off-by: Josef Bacik <josef@toxicpanda.com>
>>> ---
>>>  fs/btrfs/space-info.c | 34 +++++++++++++++++++++++++++-------
>>>  1 file changed, 27 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/fs/btrfs/space-info.c b/fs/btrfs/space-info.c
>>> index 8a1c7ada67cb..bd485be783b8 100644
>>> --- a/fs/btrfs/space-info.c
>>> +++ b/fs/btrfs/space-info.c
>>> @@ -676,19 +676,39 @@ static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
>>>  		!test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
>>>  }
>>>  
>>> -static bool wake_all_tickets(struct list_head *head)
>>> +static bool wake_all_tickets(struct btrfs_fs_info *fs_info,
>>> +			     struct btrfs_space_info *space_info)
>>>  {
>>>  	struct reserve_ticket *ticket;
>>> +	u64 tickets_id = space_info->tickets_id;
>>> +	u64 first_ticket_bytes = 0;
>>> +
>>> +	while (!list_empty(&space_info->tickets) &&
>>> +	       tickets_id == space_info->tickets_id) {
>>> +		ticket = list_first_entry(&space_info->tickets,
>>> +					  struct reserve_ticket, list);
>>> +
>>> +		/*
>>> +		 * may_commit_transaction will avoid committing the transaction
>>> +		 * if it doesn't feel like the space reclaimed by the commit
>>> +		 * would result in the ticket succeeding.  However if we have a
>>> +		 * smaller ticket in the queue it may be small enough to be
>>> +		 * satisified by committing the transaction, so if any
>>> +		 * subsequent ticket is smaller than the first ticket go ahead
>>> +		 * and send us back for another loop through the enospc flushing
>>> +		 * code.
>>> +		 */
>>> +		if (first_ticket_bytes == 0)
>>> +			first_ticket_bytes = ticket->bytes;
>>> +		else if (first_ticket_bytes > ticket->bytes)
>>> +			return true;
>>>  
>>> -	while (!list_empty(head)) {
>>> -		ticket = list_first_entry(head, struct reserve_ticket, list);
>>>  		list_del_init(&ticket->list);
>>>  		ticket->error = -ENOSPC;
>>>  		wake_up(&ticket->wait);
>>> -		if (ticket->bytes != ticket->orig_bytes)
>>> -			return true;
>>> +		btrfs_try_to_wakeup_tickets(fs_info, space_info);
>>
>> So the change in this logic is directly related to the implementation of
>> btrfs_try_to_wakeup_tickets. Because when we fail and remove a ticket in
>> this function we give a chance that the next ticket *could* be
>> satisfied. But how well does that work in practice, given you fail
>> normal prio tickets here, whereas btrfs_try_to_wakeup_tickets first
>> checks the prio ticket. So even if you are failing normal ticket but
>> there is one unsatifiable prio ticket that won't really change anything.
> 
> In practice we don't get to this state with high priority tickets on the list.
> Anything that would be long-ish term on the priority list is evict, and we wait
> for iput()'s in the normal flushing code.  At the point we hit wake_all_tickets
> we generally should only have tickets on the normal list.

Be that as it may, I think this assumption needs to be codified via an
assert or WARN_ON.

> 
> I suppose we could possibly get into this situation, but again the high priority
> tickets are going to be evict, truncate block, and relocate, which all have
> significantly lower reservation amounts than things like create or unlink.  If
> those things are unable to get reservations then we are truly out of space.
> Thanks,
> 
> Josef
>