Ceph

I forgot to mention, the other problem is that the number
of pages required in ceph_writepages_start() is computed
as (essentially):

    max_size = 1 << inode->i_blkbits;
    if (fsc->mount_options->wsize && fsc->mount_options->wsize < max_size)
        max_size = fsc->mount_options->wsize;
    if (max_size < PAGE_CACHE_SIZE)
        max_size = PAGE_CACHE_SIZE;
    pages_required = max_size >> PAGE_CACHE_SHIFT;

So as mentioned above, we need at least one page. But in addition
if the "wsize" mount option has value 0 (which is the default), then
the number of pages required could be as much as

    inode->i_blkbits >> PAGE_CACHE_SHIFT

And that is initialized in fs/ceph/inode.c:fill_inode() as:

        inode->i_blkbits = fls(le32_to_cpu(info->layout.fl_stripe_unit)) - 1;

The layout comes from the mds, but it looks to me like the
stripe unit therein can be set without limit by an ioctl() via
CEPH_MDS_OP_SETLAYOUT or CEPH_MDS_OP_SETDIRLAYOUT.

So bottom line of all this is that inside ceph_writepages_start()
if we are going to fall back to using the memory pool to satisfy
our request for a page pointer array, the size of the array in
the pool needs to be sufficient to hold the maximum sized stripe
unit value (which is apparently 0xffffffff).

My fix addresses the more simple issues described initially. I
do not intend to address this broader problem, I'll leave it to
someone else (or at least some other time) to fix that.

It's possible the maximum number of pages we'll ever need
is PAGEVEC_SIZE. But I'm not going to spend any time right
now trying to verify. It looks a bit like that might be
an upper bound though, so thought I'd mention that here.

On the osd, it looks to me like CEPH_MDS_OP_SETLAYOUT uses
ceph_file_layout_is_valid() to verify the layout supplied
is valid. The constraints defined for the stripe unit
there are:
- it must be non-zero; and
- it must be a multiple of CEPH_MIN_STRIPE_UNIT=65536

Since the stripe unit is represented in 32 bits, that
means the maximum valid stripe unit is 65535 or 0xffff.

Whoops, mean to update http://tracker.ceph.com/issues/4646.