On Wed, Jun 01, 2016 at 11:05:20AM -0700, Eric Wheeler wrote:

On Wed, 1 Jun 2016, Zhengyuan Liu wrote:

quoted

Hi, I have created a mapped block device (bcach0) using make-bcache on
ARM64 server which has kernel enable 64K page size. However, the
bcach0 disappeared after the server reboot and there is no or dirty
metadata on super block of both cache device and back device . The
output of command  bcache-super-show was as bellow showed:
   [root@master Linux-4.4-LTS-storage]# bcache-super-show  /dev/sdb
   sb.magic bad magic
   Invalid superblock (bad magic)
/dev/sdb was the backing device and cache device got bad magic too.

I tried to traced the written process of super block in bcache source
code and found that is the issue of PAGE_SIZE. It seems that the
bcache  was designed only considering for 4K PAGE_SIZE and it works
right only on 4K PAGE_SIZE exactly. To make bcache work correctly on
64K PAGE_SIZE, I committed a patch as bellow showd:
    diff --git a/drivers/md/bcache/super.c b/drivers/md/bcache/super.c
    index 330cd6e..ef567cd 100644
    --- a/drivers/md/bcache/super.c
    +++ b/drivers/md/bcache/super.c
    @@ -224,6 +224,12 @@ static void __write_super(struct cache_sb
*sb, struct bio     *bio
        bio->bi_iter.bi_size    = SB_SIZE;
        bch_bio_map(bio, NULL);

    +#ifdef CONFIG_ARM64_64K_PAGES
    +       out = (struct cache_sb *)((char *)out + (SB_SECTOR<<9));
    +       pr_debug("sb_page_adress %x, sb_address %x,page_size
          %d\n",page_address(bio
    +       bio->bi_io_vec[0].bv_offset = (SB_SECTOR<<9);
    +#endif

        out->offset             = cpu_to_le64(sb->offset);
        out->version            = cpu_to_le64(sb->version);

Does it not recommend to use bcache on 64K PAGE_SIZE? or it only
considers for 4K PAGE_SIZE for bcache currently?
Maybe it is more suitable  for me to redefine some macro such as
SB_SECTOR, BDEV_DATA_START_DEFAULT to make bcache work correctly on
both 64K PAGE_SIZE and 4K PAGE_SIZE.

I think a patch to support arbitrary page size would be great.  Can 
you write the macros in terms of PAGE_SIZE or PAGE_SHIFT?  

It shouldn't be referencing PAGE_SIZE at all - creating a new macro
(BCH_SB_SIZE, perhaps) is the correct approach.

The code that allocates the buffer for the superblock will have to be fixed
too - right now it's probably using __get_free_page(), it should probably just
be switched to kmalloc().

(Out of curiosity, what ARM64 hardware are you using?)

Kent, this may affect bcachefs too.  Can you think of any other places 
that might have PAGE_SIZE!=4k issues?

Oh, there's probably a couple. There's probably some stuff that'll break if
btree_node_size is smaller than PAGE_SIZE, too...

If it turns out to be too much for Zhengyuan, I can probably fix upstream too
(but I don't have any hardware to test with).