On 26.07.2016 12:21, Kent Overstreet wrote:

On Tue, Jul 26, 2016 at 11:51:25AM +0200, Stefan Bader wrote:

quoted

On 21.07.2016 10:58, Stefan Bader wrote:

quoted

I was pointed at the thread which seems to address the same after
I wrote most of below text. Did not want to re-write this so please
bear with the odd layout.

https://www.redhat.com/archives/dm-devel/2016-June/msg00015.html

Zhengyuan tries to fix the problem by relocating the superblock on
disk. But I am not sure whether there is really any guarantee about
how __bread fills data into the buffer_head. What if there is the next
odd arch with 128K pages?

So below is an attempt to be more generic. Still I don't feel completely
happy with the way that a page moves (or is shared) between buffer_head
and biovec. What I tried to outline below is to let the register functions
allocate bio+biovec memory and use the in-memory sb_cache data to initialize
the biovec buffer.

Any opinions here? Also adding LKML as I don't seem to get through moderation on
dm-devel.

The correct solution is to rip out the __bread() and just read the superblock by
issuing a bio, the same way all the other IO in bcache is done.

This is the way it's done in the bcache-dev branch - unfortunately, the patch
that does that in bcache-dev is big and invasive and probably not worth the
hassle to backport:

https://evilpiepirate.org/git/linux-bcache.git/commit/?h=bcache-dev&id=303eb67bffad57b4d9e71523e7df04bf258e66d1

I agree that this looks better and also rather large.

Probably best to just do something small and localized.

So what did you think about the change I did? It seemed to be ok for 4K and 64K
at least and is rather small. And I believe that, compared to Zhengyuan's
approach this would have the benefit of not changing the superblock sector. So
it would be compatible with previously created superblocks.

-Stefan