Message-ID: <956344E2-D7B3-480D-9C1D-7C385B2EFA14@cnexlabs.com>
Date:   Fri, 17 Aug 2018 08:44:30 +0000
From:   Javier Gonzalez <javier@...xlabs.com>
To:     Matias Bjørling <mb@...htnvm.io>
CC:     "Konopko, Igor J" <igor.j.konopko@...el.com>,
        "Dziegielewski, Marcin" <marcin.dziegielewski@...el.com>,
        Hans Holmberg <hans.holmberg@...xlabs.com>,
        Heiner Litz <hlitz@...c.edu>,
        Young Tack Tack Jin <youngtack.jin@...cuitblvd.com>,
        "linux-block@...r.kernel.org" <linux-block@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [RFC PATCH v2 0/1] lightnvm: move bad block and chunk state logic
 to core

> On 17 Aug 2018, at 10.21, Matias Bjørling <mb@...htnvm.io> wrote:
> 
> On 08/16/2018 05:53 PM, Javier Gonzalez wrote:
>>> On 16 Aug 2018, at 13.34, Matias Bjørling <mb@...htnvm.io> wrote:
>>> 
>>> This patch moves the 1.2 and 2.0 block/chunk metadata retrieval to
>>> core.
>>> 
>>> Hi Javier, I did not end up using your patch. I had misunderstood what
>>> was implemented. Instead I implemented the detection of the each chunk by
>>> first sensing the first page, then the last page, and if the chunk
>>> is sensed as open, a per page scan will be executed to update the write
>>> pointer appropriately.
>> I see why you want to do it this way for maintaining the chunk
>> abstraction, but this is potentially very inefficient as blocks not used
>> by any target will be recovered unnecessarily.
> 
> True. It will up to the target to not ask for more metadata than necessary (similarly for 2.0)
> 
> Note that in 1.2, it is
>> expected that targets will need to recover the write pointer themselves.
>> What is more, in the normal path, this will be part of the metadata
>> being stored so no wp recovery is needed. Still, this approach forces
>> recovery on each 1.2 instance creation (also on factory reset). In this
>> context, you are right, the patch I proposed only addresses the double
>> erase issue, which was the original motivator, and left the actual
>> pointer recovery to the normal pblk recovery process.
>> Besides this, in order to consider this as a real possibility, we need
>> to measure the impact on startup time. For this, could you implement
>> nvm_bb_scan_chunk() and nvm_bb_chunk_sense() more efficiently by
>> recovering (i) asynchronously and (ii) concurrently across luns so that
>> we can establish the recovery cost more fairly? We can look at a
>> specific penalty ranges afterwards.
> 
> Honestly, 1.2 is deprecated.

For some...

> I don't care about the performance, I
> care about being easy to maintain, so it doesn't borg me down in the
> future.

This should be stated clear in the commit message.

> 
> Back of the envelope calculation for a 64 die SSD with 1024 blocks per
> die, and 60us read time, will take 4 seconds to scan if all chunks are
> free, a worst case something like ~10 seconds. -> Not a problem for
> me.
> 

Worst case is _much_ worse than 10s if you need to scan the block to
find the write pointer. We are talking minutes.

At least make the recovery reads asynchronous. It is low hanging fruit
and will help the average case significantly.

>> Also, the recovery scheme in pblk will change significantly by doing
>> this, so I assume you will send a followup patchset reimplementing
>> recovery for the 1.2 path?
> 
> The 1.2 path shouldn't be necessary after this. That is the idea of
> this work. Obviously, the set bad block interface will have to
> preserved and called.

If we base this patch on top of my 2.0 recovery, we will still need to
make changes to support all 1.2 corner cases.

How do you want to do it? We get this patch in shape and I rebase on top
or the other way around?

Javier

Download attachment "signature.asc" of type "application/pgp-signature" (834 bytes)

Powered by blists - more mailing lists