Hi, Ilya, Xiubo, Greg,
I'm trying to finish my patches to make ceph work with netfslib and I'm
wondering if snap handling on inodes can be made easier to work with. Also, I
think there may be a bug in the interaction between ceph_queue_cap_snap() and
writable mmaps.
What I would like to do is to make page/folio->private point at the
ceph_cap_snap struct instead of pointing to ceph_snap_context. This makes it
easier to fish the metadata details out in ceph when netfslib asks it to
perform a write operation.
Netfslib has the capability to pass an netfs_group struct through the API, and
I currently have this subclassed by ceph_snap_context, but that doesn't
directly carry sufficient information as I presume that's a global thing and
not an inode-specific thing.
However, it looks like capsnaps don't always exist, even on dirty inodes...
So what I'm thinking is:
(1) Make struct ceph_cap_snap a subclass of netfs_group. This would allow
netfslib to manipulate them and attach them to dirty pages and do
selective writeback.
(2) Always keep a ceph_cap_snap on a dirty inode. It can be treated
specially when it's the only snap and at the head.
(3) Offload some of the fields from ceph_inode_info into ceph_cap_snap
(eg. truncate_size and truncate_seq) and update them directly there.
(4) On entry to any sort of write routine, see if we need a new capsnap for
that inode and, if so, create one. This would include ->write_iter(),
->page_mkwrite(), ->setattr(), possibly ->setxattr(),
(5) In queue_realm_cap_snaps(), mark the capsnap as being obsolete and call
unmap_mapping_pages() on each inode to force ->page_mkwrite() to be
called[!] on further modification.
queue_realm_cap_snaps() doesn't then need to create a new snapcap; this
can be left to the various write routines.
[!] This would fix the aforementioned potential bug whereby someone can
continue writing to the inode even though a new snap has happened.
(6) ceph_writepages() calls netfs_writepages_group() to flush out pages with
the matching group, stepping through the capsnap list on the inode.
Any thoughts on whether this would work? If I can do this, I can reduce
get_oldest_context() to almost nothing and don't need the ceph_writeback_ctl
struct anymore (I think).
Thanks,
David
