On 9/8/21 4:58 AM, Jeff Layton wrote:
On Tue, 2021-09-07 at 21:19 +0800, Xiubo Li wrote:
On 9/7/21 8:35 PM, Jeff Layton wrote:
On Fri, 2021-09-03 at 16:15 +0800, xiubli@xxxxxxxxxx wrote:
From: Xiubo Li <xiubli@xxxxxxxxxx>
This patch series is based Jeff's ceph-fscrypt-size-experimental
branch in https://git.kernel.org/pub/scm/linux/kernel/git/jlayton/linux.git.
This is just a draft patch and need to rebase or recode after Jeff
finished his huge patch set.
Post the patch out for advices and ideas. Thanks.
I'll take a look. Going forward though, it'd probably be best for you to
just take over development of the entire ceph-fscrypt-size series
instead of trying to develop on top of my branch.
That branch is _very_ rough anyway. Just clone the branch into your tree
and then you can drop or change patches in it as you see fit.
Sure.
====
This approach will not do the rmw immediately after the file is
truncated. If the truncate size is aligned to the BLOCK SIZE, so
there no need to do the rmw and only in unaligned case will the
rmw is needed.
And the 'fscrypt_file' field will be cleared after the rmw is done.
If the 'fscrypt_file' is none zero that means after the kclient
reading that block to local buffer or pagecache it needs to do the
zeroing of that block in range of [fscrypt_file, round_up(fscrypt_file,
BLOCK SIZE)).
Once any kclient has dirty that block and write it back to ceph, the
'fscrypt_file' field will be cleared and set to 0. More detail please
see the commit comments in the second patch.
That sounds odd. How do you know where the file ends once you zero out
fscrypt_file?
/me goes to look at the patches...
The code in the ceph_fill_inode() is not handling well for multiple
ftruncates case, need to be fixed.
Ok. It'd probably be best to do that fix first in a separate patch and
do the fscrypt work on top.
FWIW, I'd really like to see the existing truncate code simplified (or
at least, better documented). I'm very leery of adding yet more fields
to the inode to handle truncate/size. So far, we have all of this:
struct mutex i_truncate_mutex;
u32 i_truncate_seq; /* last truncate to smaller size */
u64 i_truncate_size; /* and the size we last truncated down to */
int i_truncate_pending; /* still need to call vmtruncate */
u64 i_max_size; /* max file size authorized by mds */
u64 i_reported_size; /* (max_)size reported to or requested of mds */
u64 i_wanted_max_size; /* offset we'd like to write too */
u64 i_requested_max_size; /* max_size we've requested */
Your patchset adds yet another new field with its own logic. I think we
need to aim to simplify this code rather than just piling more logic on
top.
Yeah, makes sense.
Maybe we need to change the 'fscrypt_file' field's logic and make it
opaqueness for MDS, then the MDS will use it to do the truncate instead
as I mentioned in the previous reply in your patch set.
Then we can do the defer rmw in any kclient when necessary as this patch
does.
I think you can't defer the rmw unless you have Fb caps. In that case,
you'd probably want to just truncate it in the pagecache, dirty the last
page in the inode, and issue the truncate to the MDS.
In the case where you don't have Fb caps, then I think you don't want to
defer anything, as you can't guarantee another client won't get in there
to read the object. On a truncate, you'll want to issue the truncate to
the MDS and do the RMW on the last page. I'm not sure what order you'd
want to do that in though. Maybe you can issue them simultaneously?
I am not sure I correctly understand this. If my understanding is correct:
If one kclient will ftruncate a file, the fscrypt_file will be recorded
in the metadata. So after that this kclient could just release the Fwb
caps if it has. And later for any kclient it should first get the
fscrypt_file, so when:
A), reading, it should be granted the Fr caps, then we it always zero
that specified block, which the contents needs to be truncated, just
after the readed data dencrypted by using the fscrypt_file.
B), writing, if kclient wants to write data back to a file, it should
always do the read-modify-write, right ? It will read the data to the
local page buffers first by zeroing that specified block. Since it can
buffer the data it should already have been granted the Fb caps. If that
specified block will be updated, then it should update that whole block
contents, and that whole block has already been truncated and modified.
Then we can reset the fscrypt_file value in MDS, and at the same time we
need to hold the Fx too. That means if encryption is enabled, when
writing it should always get the Fx caps.
If one kclient have held the Fb caps, will MDS allow any other kclient
to hold the Fr caps ?
For the Fb cap did I miss something ?
There also need on small work in Jeff's MDS PR in cap flushing code
to clear the 'fscrypt_file'.
Xiubo Li (2):
Revert "ceph: make client zero partial trailing block on truncate"
ceph: truncate the file contents when needed when file scrypted
fs/ceph/addr.c | 19 ++++++++++++++-
fs/ceph/caps.c | 24 ++++++++++++++++++
fs/ceph/file.c | 65 ++++++++++++++++++++++++++++++++++++++++++++++---
fs/ceph/inode.c | 48 +++++++++++++++++++-----------------
fs/ceph/super.h | 13 +++++++---
5 files changed, 138 insertions(+), 31 deletions(-)
