Re: [PATCH] lightnvm: pblk: Introduce hot-cold data separation

[Igor Konopko]

On 01.05.2019 22:20, Heiner Litz wrote:
> Javier, Igor,
> you are correct. The problem exists if we have a power loss and we
> have an open gc and an open user line and both contain the same LBA.
> In that case, I think we need to care about the 4 scenarios:
>
> 1. user_seq_id > gc_seq_id and user_write after gc_write: No issue
> 2. user_seq_id > gc_seq_id and gc_write > user_write: Cannot happen,
> open user lines are not gc'ed

Maybe it would be just a theoretical scenario, but I'm not seeing any 
reason why this cannot happen in pblk implementation:
Let assume that user line X+1 is opened when GC line X is already open 
and the user line is closed when GC line X is still in use. Then GC 
quickly choose user line X+1 as a GC victim and we are hitting 2nd case.

> 3. gc_seq_id > user_seq_id and user_write after gc_write: RACE
> 4. gc_seq_id > user_seq_id and gc_write after user_write: No issue
>
> To address 3.) we can do the following:
> Whenever a gc line is opened, determine all open user lines and store
> them in a field of pblk_line. When choosing a victim for GC, ignore
> those lines.

Your solution sounds right, but I would extend this based on my previous 
comment to 2nd case by sth like: during opening new user data also add 
this line ID to this "blacklist" for the GC selection.

Igor

> Let me know if that sounds good and I will send a v2
> Heiner
>
> On Tue, Apr 30, 2019 at 11:19 PM Javier GonzÃlez <javier@xxxxxxxxxxx> wrote:
> > > On 26 Apr 2019, at 18.23, Heiner Litz <hlitz@xxxxxxxx> wrote:
> > >
> > > Nice catch Igor, I hadn't thought of that.
> > >
> > > Nevertheless, here is what I think: In the absence of a flush we don't
> > > need to enforce ordering so we don't care about recovering the older
> > > gc'ed write. If we completed a flush after the user write, we should
> > > have already invalidated the gc mapping and hence will not recover it.
> > > Let me know if I am missing something.
> >
> > I think that this problem is orthogonal to a flush on the user path. For example
> >
> >     - Write to LBA0 + completion to host
> >     - [â]
> >     - GC LBA0
> >     - Write to LBA0 + completion to host
> >     - fsync() + completion
> >     - Power Failure
> >
> > When we power up and do recovery in the current implementation, you
> > might get the old LBA0 mapped correctly in the L2P table.
> >
> > If we enforce ID ordering for GC lines this problem goes away as we can
> > continue ordering lines based on ID and then recovering sequentially.
> >
> > Thoughts?
> >
> > Thanks,
> > Javier
> >
> > > On Fri, Apr 26, 2019 at 6:46 AM Igor Konopko <igor.j.konopko@xxxxxxxxx> wrote:
> > > > On 26.04.2019 12:04, Javier GonzÃlez wrote:
> > > > > > On 26 Apr 2019, at 11.11, Igor Konopko <igor.j.konopko@xxxxxxxxx> wrote:
> > > > > >
> > > > > > On 25.04.2019 07:21, Heiner Litz wrote:
> > > > > > > Introduce the capability to manage multiple open lines. Maintain one line
> > > > > > > for user writes (hot) and a second line for gc writes (cold). As user and
> > > > > > > gc writes still utilize a shared ring buffer, in rare cases a multi-sector
> > > > > > > write will contain both gc and user data. This is acceptable, as on a
> > > > > > > tested SSD with minimum write size of 64KB, less than 1% of all writes
> > > > > > > contain both hot and cold sectors.
> > > > > >
> > > > > > Hi Heiner
> > > > > >
> > > > > > Generally I really like this changes, I was thinking about sth similar since a while, so it is very good to see that patch.
> > > > > >
> > > > > > I have a one question related to this patch, since it is not very clear for me - how you ensure the data integrity in following scenarios:
> > > > > > -we have open line X for user data and line Y for GC
> > > > > > -GC writes LBA=N to line Y
> > > > > > -user writes LBA=N to line X
> > > > > > -we have power failure when both line X and Y were not written completely
> > > > > > -during pblk creation we are executing OOB metadata recovery
> > > > > > And here is the question, how we distinguish whether LBA=N from line Y or LBA=N from line X is the valid one?
> > > > > > Line X and Y might have seq_id either descending or ascending - this would create two possible scenarios too.
> > > > > >
> > > > > > Thanks
> > > > > > Igor
> > > > >
> > > > > You are right, I think this is possible in the current implementation.
> > > > >
> > > > > We need an extra constrain so that we only GC lines above the GC line
> > > > > ID. This way, when we order lines on recovery, we can guarantee
> > > > > consistency. This means potentially that we would need several open
> > > > > lines for GC to avoid padding in case this constrain forces to choose a
> > > > > line with an ID higher than the GC line ID.
> > > > >
> > > > > What do you think?
> > > >
> > > > I'm not sure yet about your approach, I need to think and analyze this a
> > > > little more.
> > > >
> > > > I also believe that probably we need to ensure that current user data
> > > > line seq_id is always above the current GC line seq_id or sth like that.
> > > > We cannot also then GC any data from the lines which are still open, but
> > > > I believe that this is a case even right now.
> > > >
> > > > > Thanks,
> > > > > Javier