On 2020/03/24 16:51, Hannes Reinecke wrote:
> On 3/24/20 4:52 AM, Damien Le Moal wrote:
>> +Bob who had proposed a similar change a last month.
>> On 2020/03/24 0:04, Hannes Reinecke wrote:
>>> Implement 'cache' zones which reside on a different device.
>>> The device is logically split into zones, which then will be
>>> used as 'cache' zones, similar to the existing randow write
>>> zones.
>> It does look like the new "cache" zones are really used exactly as 
>> conventional
>> zones of the SMR drive. So I wonder: why even define this new zone type ? We
>> could have the "cache" device split into random (conventional) zones added 
>> to a
>> single pool of random zones. We can simply add device awareness to the zone
>> allocator to avoid as much as possible using a random zone from the same 
>> drive
>> as the sequential zone it buffers. That would avoid repeating most of the 
>> code
>> for cache & random.
> Yes, indeed that was the idea to keep 'cache' and 'random' zones 
> essentially similar. But then as there is a need to differentiate 
> between them I thought it easier to introduce a new zone type.
> However, it's a nice idea to use the device to differentiate between 
> both. And it would even lend to a simpler reclaim mechanism; set the low 
> watermark when all random zones on the cache device are full, and set 
> the high watermark when half of the random zones on the SMR device are full.
> I'll give it a go and see where I end up.
>> Furthermore, this work is really great to support SMR drives with no
>> conventional zones (a lot of ask for these). And considering that the new 
>> WITH PRESET command is coming soon, a user will be able to reformat an SMR 
>> drive
>> with sequential zones only to maximize capacity. For these, the cache device
>> would need to hold the random zones, at which point the difference between 
>> cache
>> and rando goes away.
> I know :-)
>>> Signed-off-by: Hannes Reinecke <h...@suse.de>
>>> ---
>>>   drivers/md/dm-zoned-metadata.c | 174 ++++++++++++++++++++++++++++-----
>>>   drivers/md/dm-zoned-reclaim.c  |  76 +++++++++++---
>>>   drivers/md/dm-zoned-target.c   | 109 ++++++++++++++++++---
>>>   drivers/md/dm-zoned.h          |  31 +++++-
>>>   4 files changed, 339 insertions(+), 51 deletions(-)
>>> diff --git a/drivers/md/dm-zoned-metadata.c b/drivers/md/dm-zoned-metadata.c
>>> index 369de15c4e80..41cc3a29db0b 100644
>>> --- a/drivers/md/dm-zoned-metadata.c
>>> +++ b/drivers/md/dm-zoned-metadata.c
>>> @@ -132,6 +132,8 @@ struct dmz_sb {
>>>   struct dmz_metadata {
>>>     struct dmz_dev          *dev;
>>> +   struct dmz_cdev         *cdev;
>> Given the point above, we could have this generalized as an array of devices,
>> with the first one meeting the constraints:
>> * It contains the metadata
>> * It has random/conventional zones, or is a regular device (with all its
>> capacity used through emulated random zones)
>> I do not think that complicates the changes you did a lot. The reclaim part 
>> will
>> need some more love I guess to be efficient, but it may be as simple as 
>> defining
>> one work struct for each drive beside the first one.
>> Thoughts ?
> Rather not. Stringing several devices together essentially emulates a 
> RAID0 setup without any of the benefits. And the reclaim mechanism gets 
> infinitely more complex.

OK. Fair point.

> Another thing: I would need to update the metadata to hold the device 
> and zoneset UUID; both devices need to carry a metadata so that we can 
> stitch them together upon restart.
> But some bright soul put a crc in the middle of the metadata :-(
> So we can't easily extend the metadata with new fields as then the 
> meaning of the crc is unclear; as it stands it would only cover the old 
> fields, and not the new ones.

Haha ! OK. You got me. Not my finest hour on this one :)

> So I would propose a 'v2' metadata, holding the crc as the last entry of 
> the metadata. And adding a device UUID and cacheset UUID.
> And ensuring that the first metadata set is stored on the cache device, 
> and the backup one on the SMR device.

That would work.

> Cheers,
> Hannes

Damien Le Moal
Western Digital Research

dm-devel mailing list

Reply via email to