On 08/17/2018 10:44 AM, Javier Gonzalez wrote:
On 17 Aug 2018, at 10.21, Matias Bjørling <mb@xxxxxxxxxxx> wrote:
On 08/16/2018 05:53 PM, Javier Gonzalez wrote:
On 16 Aug 2018, at 13.34, Matias Bjørling <mb@xxxxxxxxxxx> 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...
No. OCSSD 1.2 is deprecated. Others that have a derivative of 1.2 have
their own storage stack and spec that they will continue development on,
which can not be expected to be compatible with the OCSSD 1.2 that is
implemented in the lightnvm subsystem.
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.
I think you may be assuming that all blocks are open. My assumption is
that this is very rare (given the NAND characteristics). At most a
couple of blocks may be open per die. That leads me to the time quoted.
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?
I'll pull this in when you're tested it with your 1.2 implementation.
