Re: [PATCH 14/15] libnvdimm: support read-only btt backing devices

[Dan Williams]

On Sun, Jun 21, 2015 at 6:54 AM, Christoph Hellwig <hch@xxxxxx> wrote:
> On Sun, Jun 21, 2015 at 06:21:50AM -0700, Dan Williams wrote:
>> This question has come up before.  Making btt an internal property of
>> a device makes some things cleaner and others more messy.  We lose the
>> ability to place a btt instance on top of a partition, rather than a
>> whole disk.
>
> I thought the addition of nfit labels avoids the need for a partition
> table now?
>

The labels only allow allocation of persistent media between pmem and
blk.  For a given dimm you may access in either mode and the label
records the decision.  We can have a btt on either the pmem or
blk-mode disk type, or partition thereof.

>> If we ever need to access the raw device we no longer
>> have a direct block device to reference.  Linux has been doing stacked
>> configurations to change the personality of block devices since
>> forever (md, dm, bcache...), why invent something new to handle the
>> btt-personality of ->rw_bytes() devices?
>
> Because the underlying abstraction really isn't a block device
> anymore, it's a byte addressable device.  This is more similar to
> for example how the mtd subsystem is structured.

Yes, it's this hybrid thing that mostly fits into the existing block
device model save for two new block_device_operations
->direct_access() and ->rw_bytes().  We then use property of a
block_device that allows it to be claimed for exclusive ownership by a
filesystem or another block_device to layer storage semantics on top
be it files+directories, raid, caching, or atomic sectors.  NVDIMM
devices don't present the same complexity as MTD devices.  The only
complexity they present is byte-address-ability, not erase-block-size,
wear-leveling, etc...

>> BTT precludes DAX, if you want both modes on one pmem disk placing BTT
>> on a partition of the disk for fs metadata and DAX-capable data on the
>> rest is our proposed solution.  We chose this architecture after a
>> conversation with Dave Chinner about XFS's need to have atomic sector
>> guarantees for its metadata and wanting to simultaneously enable
>> XFS-DAX.
>
> I can't see why a v5 XFS filesystem with CRCs on all metadata would need
> sector atomic updates any more.  But even in a case where it would it
> seem like whatever label you use for partioning should sit above the
> block layer.

Yes, we use standard partition labels to sub-divide a namespace.  A
namespace boundary is either set by a label internal to the dimm or
the NFIT directly (for dimms that do not support internal labeling).
Good to hear that we don't need BTT for XFS v5, can we make the
guarantee for all filesystems that may want to support DAX?  I still
think stacking is a natural fit for this problem.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
Please read the FAQ at  http://www.tux.org/lkml/