Re: [PATCH] f2fs: enlarge block plug coverage
From: Chao Yu
Date: Fri Apr 13 2018 - 03:14:10 EST
On 2018/4/13 12:07, Jaegeuk Kim wrote:
> On 04/13, Chao Yu wrote:
>> On 2018/4/13 9:06, Jaegeuk Kim wrote:
>>> On 04/10, Chao Yu wrote:
>>>> On 2018/4/10 12:10, Jaegeuk Kim wrote:
>>>>> On 04/10, Chao Yu wrote:
>>>>>> On 2018/4/10 2:02, Jaegeuk Kim wrote:
>>>>>>> On 04/08, Chao Yu wrote:
>>>>>>>> On 2018/4/5 11:51, Jaegeuk Kim wrote:
>>>>>>>>> On 04/04, Chao Yu wrote:
>>>>>>>>>> This patch enlarges block plug coverage in __issue_discard_cmd, in
>>>>>>>>>> order to collect more pending bios before issuing them, to avoid
>>>>>>>>>> being disturbed by previous discard I/O in IO aware discard mode.
>>>>>>>>>
>>>>>>>>> Hmm, then we need to wait for huge discard IO for over 10 secs, which
>>>>>>>>
>>>>>>>> We found that total discard latency is rely on total discard number we issued
>>>>>>>> last time instead of range or length discard covered. IMO, if we don't change
>>>>>>>> .max_requests value, we will not suffer longer latency.
>>>>>>>>
>>>>>>>>> will affect following read/write IOs accordingly. In order to avoid that,
>>>>>>>>> we actually need to limit the discard size.
>>>>>>
>>>>>> Do you mean limit discard count or discard length?
>>>>>
>>>>> Both of them.
>>>>>
>>>>>>
>>>>>>>>
>>>>>>>> If you are worry about I/O interference in between discard and rw, I suggest to
>>>>>>>> decrease .max_requests value.
>>>>>>>
>>>>>>> What do you mean? This will produce more pending requests in the queue?
>>>>>>
>>>>>> I mean after applying this patch, we can queue more discard IOs in plug inside
>>>>>> task, otherwise, previous issued discard in block layer can make is_idle() be false,
>>>>>> then it can stop IO awared user to issue pending discard command.
>>>>>
>>>>> Then, unplug will issue lots of discard commands, which affects the following rw
>>>>> latencies. My preference would be issuing discard commands one by one as much as
>>>>> possible.
>>>>
>>>> Hmm.. for you concern, we can turn down IO priority of discard from background?
>>>
>>> That makes much more sense to me. :P
>>
>> Then, this patch which enlarge plug coverage will not still a problem, right? ;)
>
> This is different one.
Yup, if there will be no IO interference as you concerned before, can we accept
it now?
Thanks,
>
>>
>> Thanks,
>>
>>>
>>>>
>>>> Thanks,
>>>>
>>>>>
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>> Thanks,
>>>>>>>>
>>>>>>>>>
>>>>>>>>> Thanks,
>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Signed-off-by: Chao Yu <yuchao0@xxxxxxxxxx>
>>>>>>>>>> ---
>>>>>>>>>> fs/f2fs/segment.c | 7 +++++--
>>>>>>>>>> 1 file changed, 5 insertions(+), 2 deletions(-)
>>>>>>>>>>
>>>>>>>>>> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
>>>>>>>>>> index 8f0b5ba46315..4287e208c040 100644
>>>>>>>>>> --- a/fs/f2fs/segment.c
>>>>>>>>>> +++ b/fs/f2fs/segment.c
>>>>>>>>>> @@ -1208,10 +1208,12 @@ static int __issue_discard_cmd(struct f2fs_sb_info *sbi,
>>>>>>>>>> pend_list = &dcc->pend_list[i];
>>>>>>>>>>
>>>>>>>>>> mutex_lock(&dcc->cmd_lock);
>>>>>>>>>> +
>>>>>>>>>> + blk_start_plug(&plug);
>>>>>>>>>> +
>>>>>>>>>> if (list_empty(pend_list))
>>>>>>>>>> goto next;
>>>>>>>>>> f2fs_bug_on(sbi, !__check_rb_tree_consistence(sbi, &dcc->root));
>>>>>>>>>> - blk_start_plug(&plug);
>>>>>>>>>> list_for_each_entry_safe(dc, tmp, pend_list, list) {
>>>>>>>>>> f2fs_bug_on(sbi, dc->state != D_PREP);
>>>>>>>>>>
>>>>>>>>>> @@ -1227,8 +1229,9 @@ static int __issue_discard_cmd(struct f2fs_sb_info *sbi,
>>>>>>>>>> if (++iter >= dpolicy->max_requests)
>>>>>>>>>> break;
>>>>>>>>>> }
>>>>>>>>>> - blk_finish_plug(&plug);
>>>>>>>>>> next:
>>>>>>>>>> + blk_finish_plug(&plug);
>>>>>>>>>> +
>>>>>>>>>> mutex_unlock(&dcc->cmd_lock);
>>>>>>>>>>
>>>>>>>>>> if (iter >= dpolicy->max_requests)
>>>>>>>>>> --
>>>>>>>>>> 2.15.0.55.gc2ece9dc4de6
>>>>>>>>>
>>>>>>>>> .
>>>>>>>>>
>>>>>>>
>>>>>>> .
>>>>>>>
>>>>>
>>>>> .
>>>>>
>>>
>>> .
>>>
>
> .
>