On 9/14/21 10:24 PM, Jeff Layton wrote:
On Tue, 2021-09-14 at 13:40 +0800, Xiubo Li wrote:
On 9/14/21 3:34 AM, Jeff Layton wrote:
[...]
I'll have to think about whether that's still racy. Part of the problem
is that once the client doesn't have caps, it doesn't have a way to
ensure that fscrypt_file (whatever it holds) doesn't change while it's
doing that zeroing.
As my understanding, if clientA want to read a file, it will open it
with RD mode, then it will get the fscrypt_file meta and "Fr" caps, then
it can safely read the file contents and do the zeroing for that block.
Ok, so in this case, the client is just zeroing out the appropriate
region in the resulting read data, and not on the OSD. That should be
ok.
Then the opened file shouldn't worry whether the fscrypt_file will be
changed by others during it still holding the Fr caps, because if the
clientB wants to update the fscrypt_file it must acquire the Fw caps
first, before that the Fr caps must be revoked from clientA and at the
same time the read pagecache in clientA will be invalidated.
Are you certain that Fw caps is enough to ensure that no other client
holds Fr caps?
I spent hours and went through the mds Locker related code on the weekends.
From the mds/lock.cc code, for mds filelock for example in the LOCK_MIX
state and some interim transition states to LOCK_MIX it will allow
different clients could hold any of Fw or Fr caps. But the Fb/Fc will be
disabled. Checked the mds/Locker.cc code, found that the mds filelock
could to switch LOCK_MIX state in some cases when there has one client
wants Fw and another client wants any of Fr and Fw.
In this case I think the Linux advisory or mandatory locks are necessary
to keep the file contents concurrency. In multiple processes' concurrent
read/write or write/write cases without the Linux advisory/mandatory
locks the file contents' concurrency won't be guaranteed, so the logic
is the same here ?
If so, couldn't we just assume the Fw vs Fw and Fr vs Fw caps should be
exclusive in correct use case ? For example, just after the mds filelock
state switches to LOCK_MIX, if clientA gets the advisory file lock and
the Fw caps, and even another clientB could be successfully issued the
Fr caps, the clientB won't do any read because it should be still stuck
and be waiting for the advisory file lock.
I'm not sure I like that idea. Basically, that would change the meaning
of the what Frw caps represent, in a way that is not really consistent
with how they have been used before.
We could gate that new behavior on the new feature flags, but it sounds
pretty tough.
I think we have a couple of options:
1) we could just make the clients request and wait on Fx caps when they
do a truncate. They might stall for a bit if there is contention, but it
would ensure consistency and the client could be completely in charge of
the truncate. [a]
Yeah, for my defer RMW approach we need to held the Fx caps every time
when writing/truncating files, and the Fs caps every time when reading.
While currently almost all the read/write code have ignored them because
read/write do not need them in most cases.
Note that we already cache truncate operations when we have Fx.
Yeah, only when we have Fx and the attr.ia_size > isize will it cache
the truncate operation.
For this I am a little curious why couldn't we cache truncate operations
when attr.ia_size >= isize ?
See
__ceph_setattr. Do we even need to change the read path here, or is the
existing code just wrong?
This is info I've been trying to get a handle on since I started working
on cephfs. The rules for FILE caps are still extremely unclear.
I am still check this logic from MDS side. Still not very clear.
[...]
2) we could rev the protocol, and have the client send along the last
block to be written along with the SETATTR request. Maybe we even
consider just adding a new TRUNCATE call independent of SETATTR. The MDS
would remain in complete control of it at that point.
This approach seems much better, since the last block size will always
less than or equal to 4K(CEPH_FSCRYPT_BLOCK_SIZE) and the truncate
should be rare in normal use cases (?), with extra ~4K data in the
SETATTR should be okay when truncating the file.
So when truncating a file, in kclient it should read that block, which
needs to do the RMW, first, and then do the truncate locally and encrypt
it again, and then together with SETATTR request send it to MDS. And the
MDS will update that block just before truncating the file.
This approach could also keep the fscrypt logic being opaque for the MDS.
Note that you'll need to guard against races on the RMW. For instance,
another client could issue a write to that last block just after we do
the read for the rmw.
If you do this, then you'll probably need to send along the object
version that you got from the read and have the MDS validate that before
it does the truncate and writes out the updated crypto block.
If something changed in the interim, the MDS can just return -EAGAIN or
whatever to the client and it can re-do the whole thing. It's a mess,
but it should be consistent.
I think we probably want a new call for this too instead of overloading
SETATTR (which is already extremely messy).
Okay, I will check this more carefully.
The other ideas I've considered seem more complex and don't offer any
significant advantages that I can see.
[a]: Side question: why does buffering a truncate require Fx and not Fb?
How do Fx and Fb interact?
For my defer RMW approach we need the Fx caps every time when writing
the file, and the Fw caps is the 'need' caps for write, while the Fb is
the 'want' caps. If the Fb caps is not allowed or issued by the MDS, it
will write-through data to the osd, after that the Fxw could be safely
released. If the client gets the Fb caps, the client must also hold the
Fx caps until the buffer has been writen back.
The problem there is that will effectively serialize all writers of a
file -- even ones that are writing to completely different backend
objects.
That will almost certainly regress performance. I think we want to try
to avoid that. I'd rather that truncate be extremely slow and expensive
than slow down writes.
Agree.
Thanks.
