Re: [PATCH] lightnvm: pblk: assume that chunks are closed on 1.2 devices

[Javier Gonzalez]

> On 26 Jun 2018, at 13.44, Matias BjÃrling <mb@xxxxxxxxxxx> wrote:
> 
>> On 06/26/2018 01:31 PM, Hans Holmberg wrote:
>>> On Tue, Jun 26, 2018 at 1:38 PM, Matias BjÃrling <mb@xxxxxxxxxxx> wrote:
>>>> On 06/26/2018 11:37 AM, Javier Gonzalez wrote:
>>>> 
>>>> 
>>>> 
>>>>> On 26 Jun 2018, at 10.41, Matias BjÃrling <mb@xxxxxxxxxxx> wrote:
>>>>> 
>>>>> On 06/19/2018 11:06 AM, Hans Holmberg wrote:
>>>>>> 
>>>>>> From: Hans Holmberg <hans.holmberg@xxxxxxxxxxxx>
>>>>>> We can't know if a block is closed or not on 1.2 devices, so assume
>>>>>> closed state to make sure that blocks are erased before writing.
>>>>>> Fixes: 32ef9412c114 ("lightnvm: pblk: implement get log report chunk")
>>>>>> Signed-off-by: Hans Holmberg <hans.holmberg@xxxxxxxxxxxx>
>>>>>> ---
>>>>>> This patch applies on:
>>>>>> ssh://github.com/OpenChannelSSD/linux branch for-4.19/core
>>>>>>   drivers/lightnvm/pblk-init.c | 5 +++--
>>>>>>   1 file changed, 3 insertions(+), 2 deletions(-)
>>>>>> diff --git a/drivers/lightnvm/pblk-init.c b/drivers/lightnvm/pblk-init.c
>>>>>> index aa24264..3b8aa4a 100644
>>>>>> --- a/drivers/lightnvm/pblk-init.c
>>>>>> +++ b/drivers/lightnvm/pblk-init.c
>>>>>> @@ -717,10 +717,11 @@ static int pblk_setup_line_meta_12(struct pblk
>>>>>> *pblk, struct pblk_line *line,
>>>>>>                 /*
>>>>>>                  * In 1.2 spec. chunk state is not persisted by the
>>>>>> device. Thus
>>>>>> -                * some of the values are reset each time pblk is
>>>>>> instantiated.
>>>>>> +                * some of the values are reset each time pblk is
>>>>>> instantiated,
>>>>>> +                * so we have to assume that the block is closed.
>>>>>>                  */
>>>>>>                 if (lun_bb_meta[line->id] == NVM_BLK_T_FREE)
>>>>>> -                       chunk->state =  NVM_CHK_ST_FREE;
>>>>>> +                       chunk->state =  NVM_CHK_ST_CLOSED;
>>>>>>                 else
>>>>>>                         chunk->state = NVM_CHK_ST_OFFLINE;
>>>>>> 
>>>>> 
>>>>> pblk should scan (or the lightnvm subsystem) the blocks for their
>>>>> state, such that it doesn't have to reinitialize a full drive if it is
>>>>> already in a closed state. If marking closed, it does a full erase
>>>>> cycle on initialization, which should be avoided since it is a limited
>>>>> resource.
>>>> 
>>>> 
>>>> In 1.2 there is no such state unfortunately. However, pblk will never
>>>> attempt to reinitialize the whole drive - metadata for closed blocks
>>>> will be recovered and only those going to GC will be erased before
>>>> usage. In fact, a full close drive is the state pblk expects.
>>>> 
>>>> This patch only affects "unknown blocks", thus the only case in which
>>>> pblk would attempt to double erase is when blocks have been pre-erased
>>>> (e.g., factory or through liblightnvm). After an erase round though,
>>>> pblk will only erase pre-usage. One thing we could do is attempting to
>>>> read the first page of these unknown blocks and mark them as free if
>>>> "empty page" is returned. Is this what you mean?
>>> 
>>> 
>>> Yes, that is what I mean.
>>> 
>>> Note that this can be
>>>> 
>>>> costly on large drives; this is the reason we returned to the pre-2.0
>>>> behaviour with this patch. We are implementing a log that, among other
>>>> things, keeps the state so that pblk can have an accurate state for the
>>>> cases this can be a problem.
>>> 
>>> 
>>> Yep, it will take some time. Good to hear with the log.
>> Until we have a log in place, this patch unbreaks 1.2 support and has
>> no negative impact on performance (as compared to pre 2.0 support), so
>> please consider it for the next window.
>> The current state is that if a pblk instance is created on a 1.2 disk
>> with written blocks, writes will fail.
>>  / Hans
> 
> The negative impact is that all blocks are erased, even if they are in free state. This is a showstopper. We cannot throw out 1/X of the lifetime of the drive on each initialization. The 1.2 spec is made such that a scan can recover the block state accurately.

This fixes patch returns to the original behavior, so itâs not introducing a worse behavior than before 2.0. But youâre right, it is not the way it should be. 

Can you consider taking this as a fix for 4.18 to avoid writes failing on 1.2 devices and I promise to send a patch this week to implement the state based on reads? This new patch would be for 4.19. 

Javier