Re: [PATCH 1/2] btrfs: try to satisfy metadata requests when every flush_space() returns

[Wang Xiaoguang <wangxg.fnst@cn.fujitsu.com>]

hi,

On 10/07/2016 09:16 PM, Josef Bacik wrote:
> On 09/21/2016 02:59 AM, Wang Xiaoguang wrote:
>> In flush_space()->shrink_delalloc(), if can_overcommit() returns true,
>> though no bytes added to space_info, we still may satisfy some requests.
>>
>> Signed-off-by: Wang Xiaoguang <wangxg.fnst@cn.fujitsu.com>
>> ---
>>  fs/btrfs/extent-tree.c | 39 +++++++++++++++++++++++++++++++++++++++
>>  1 file changed, 39 insertions(+)
>>
>> diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
>> index 38c2df8..fdfc97f 100644
>> --- a/fs/btrfs/extent-tree.c
>> +++ b/fs/btrfs/extent-tree.c
>> @@ -4960,6 +4960,43 @@ static void wake_all_tickets(struct list_head 
>> *head)
>>  }
>>
>>  /*
>> + * This function must be protected by btrfs_space_info's lock.
>> + */
>> +static void try_to_wake_tickets(struct btrfs_root *root,
>> +                struct btrfs_space_info *space_info)
>> +{
>> +    struct reserve_ticket *ticket;
>> +    struct list_head *head = &space_info->priority_tickets;
>> +    enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
>> +    u64 used;
>> +
>> +again:
>> +    while (!list_empty(head)) {
>> +        ticket = list_first_entry(head, struct reserve_ticket,
>> +                      list);
>> +        used = space_info->bytes_used +
>> +            space_info->bytes_reserved + space_info->bytes_pinned +
>> +            space_info->bytes_readonly + space_info->bytes_may_use;
>> +
>> +        if (used + ticket->bytes <= space_info->total_bytes ||
>> +            can_overcommit(root, space_info, ticket->bytes, flush)) {
>> +            space_info->bytes_may_use += ticket->bytes;
>> +            list_del_init(&ticket->list);
>> +            ticket->bytes = 0;
>> +            space_info->tickets_id++;
>> +            wake_up(&ticket->wait);
>> +        } else
>> +            return;
>> +    }
>> +
>> +    if (head == &space_info->priority_tickets) {
>> +        head = &space_info->tickets;
>> +        flush = BTRFS_RESERVE_FLUSH_ALL;
>> +        goto again;
>> +    }
>> +}
>> +
>> +/*
>>   * This is for normal flushers, we can wait all goddamned day if we 
>> want to.  We
>>   * will loop and continuously try to flush as long as we are making 
>> progress.
>>   * We count progress as clearing off tickets each time we have to loop.
>> @@ -4995,6 +5032,8 @@ static void 
>> btrfs_async_reclaim_metadata_space(struct work_struct *work)
>>          ret = flush_space(fs_info->fs_root, space_info, to_reclaim,
>>                  to_reclaim, flush_state);
>>          spin_lock(&space_info->lock);
>> +        if (!ret)
>> +            try_to_wake_tickets(fs_info->fs_root, space_info);
>>          if (list_empty(&space_info->tickets)) {
>>              space_info->flush = 0;
>>              spin_unlock(&space_info->lock);
>>
>
> So first off I have no problems with this patch, it seems reasonable 
> to me.
>
> However I'm curious to see where it helped.  The only time 
> can_overcommit() would suddenly start returning true where it wouldn't 
> have before without actually adding space to the space_info would be 
> when we've dropped an empty block group (or added a new drive) and 
> suddenly fs_info->free_chunk_space is larger than it was.  Is that 
> what you were observing?  Thanks,
I think you're right. I don't add new drive when doing my big files 
create and delete test, so
it maybe "remove useless chunks" causing can_overcommit() returns true, 
but I don't know
how to observe this in codes, sorry. And this patch can really help to 
fix my enospc issue :)

Meanwhile, in shrink_delalloc(), if can_overcommit() returns true, 
shrink_delalloc()
will not shrink delalloc bytes any more, in this case we should check 
whether
some tickcts' requests can overcommit, if some can, we can satisfy them 
timely and directly.

I notice original btrfs codes before your patch "Btrfs: introduce 
ticketed enospc infrastructure"
will have over_commit try when every flush_space() returns, so here we 
may also need to do it, thanks.

Regards,
Xiaoguang Wang

>
> Josef
>
>