On 2022/1/13 00:56, Lukas Straub wrote:
On Sun, 9 Jan 2022 20:13:36 +0800
Qu Wenruo <quwenruo.btrfs@xxxxxxx> wrote:
On 2022/1/9 18:04, David Woodhouse wrote:
On Sun, 2022-01-09 at 07:55 +0800, Qu Wenruo wrote:
On 2022/1/9 04:29, Lukas Straub wrote:
But there is a even simpler solution for btrfs: It could just not touch
stripes that already contain data.
That would waste a lot of space, if the fs is fragemented.
Or we have to write into data stripes when free space is low.
That's why I'm trying to implement a PPL-like journal for btrfs RAID56.
PPL writes the P/Q of the unmodified chunks from the stripe, doesn't
it?
Did I miss something or the PPL isn't what I thought?
I thought PPL either:
a) Just write a metadata entry into the journal to indicate a full
stripe (along with its location) is going to be written.
b) Write a metadata entry into the journal about a non-full stripe
write, then write the new data and new P/Q into the journal
And this is before we start any data/P/Q write.
And after related data/P/Q write is finished, remove corresponding
metadata and data entry from the journal.
Or PPL have even better solution?
Yes, PPL is a bit better than a journal as you described it (md
supports both). Because a journal would need to be replicated to
multiple devices (raid1) in the array while the PPL is only written to
the drive containing the parity for the particular stripe. And since the
parity is distributed across all drives, the PPL overhead is also
distributed across all drives. However, PPL only works for raid5 as
you'll see.
PPL works like this:
Before any data/parity write either:
a) Just write a metadata entry into the PPL on the parity drive to
indicate a full stripe (along with its location) is going to be
written.
b) Write a metadata entry into the PPL on the parity drive about a
non-full stripe write, including which data chunks are going to be
modified, then write the XOR of chunks not modified by this write in
to the PPL.
This is a little different than I thought, and I guess that's why RAID6
is not supported.
My original assumption would be something like this for one RMW
(X = modified data, | | = unmodified data)
Data 1: |XXXXXXXXX| | |
Data 2: | | |XXXXXXXX|
P(1+2): |XXXXXXXXX| |XXXXXXXX|
In that case, modified Data 1 and 2 will go logged into PPL for the
corresponding disks.
Then for P(1+2), only the modified two parts will be logged into the device.
I'm wondering if we go this solution, wouldn't it be able to handle
RAID6 too?
Even we lost two disks, the remaining part in the PPL should still be
enough to recover whatever is lost, as long as the unmodified sectors
are really unmodified on-disk.
Although this would greatly make the PPL management much harder, as
different devices will have different PPL data usage.
To recover a inconsistent array with a lost drive:
In case a), the stripe consists only of newly written data, so it will
be affected by the write-hole (this is the trade-off that PPL makes) so
just standard parity recovery.
In case b), XOR what we wrote to the PPL (the XOR of chunks not
modified) with the modified data chunks to get our new (consistent)
parity. Then do standard parity recovery. This just works if we lost a
unmodified data chunk.
If we lost a modified data chunk this is not possible and just do
standard parity recovery from the beginning. Again, the newly written
data is affected by the write-hole but existing data is not.
If we lost the parity drive (containing the PPL) there is no need to
recover since all the data chunks are present.
Of course, this was a simplified explanation, see drivers/md/raid5-ppl.c
for details (it has good comments with examples). This also covers the
case where a data chunk is only partially modified and the unmodified
part of the chunk also needs to be protected (by working on a per-block
basis instead of per-chunk).
Thanks for the detailed explanation.
Qu
The PPL is not possible for raid6 AFAIK, because there it could happen
that you loose both a modified data chunk and a unmodified data chunk.
Regards,
Lukas Straub
An alternative in a true file system which can do its own block
allocation is to just calculate the P/Q of the final stripe after it's
been modified, and write those (and) the updated data out to newly-
allocated blocks instead of overwriting the original.
This is what Johannes is considering, but for a different purpose.
Johannes' idea is to support zoned device. As the physical location a
zoned append write will only be known after it's written.
So his idea is to maintain another mapping tree for zoned write, so that
full stripe update will also happen in that tree.
But that idea is still in the future, on the other hand I still prefer
some tried-and-true method, as I'm 100% sure there will be new
difficulties waiting us for the new mapping tree method.
Thanks,
Qu
Then the final step is to free the original data blocks and P/Q.
This means that your RAID stripes no longer have a fixed topology; you
need metadata to be able to *find* the component data and P/Q chunks...
it ends up being non-trivial, but it has attractive properties if we
can work it out.
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://listman.redhat.com/mailman/listinfo/dm-devel
