On 01/16/2015 05:40 PM, Xavier Hernandez wrote:
On 01/16/2015 04:58 AM, Pranith Kumar Karampuri wrote:
On 01/15/2015 10:53 PM, Xavier Hernandez wrote:
Hi,
currently eager locking is implemented by checking the open-fd-count
special xattr for each write. If there's more than one open on the
same file, eager locking is disabled to avoid starvation.
This works quite well for file writes, but makes eager locking
unusable for other request types that do not involve an open fd (in
fact, this method is only for writes on regular files, not reads or
directories). This may cause a performance problem for other
operations, like metadata.
To be able to use eager locking for other purposes, what do you think
about this proposal:
Instead of implementing open-fd-count on posix xlator, do something
similar but in locks xlator. The difference will be that locks xlator
can use the pending locking information to determine if there are
other processes waiting for a resource. If so, set a flag in the cbk
xdata to let high level xlators know that they should not use eager
locking (this can be done only upon request by xdata).
I think this way provides a more precise way to avoid starvation and
maximize performance at the same time, and it can be used for any
request even if it doesn't depend on an fd.
Another advantage is that if one file has been opened multiple times
but all of them from the same glusterfs client, that client could use
a single inodelk to manage all the accesses, not needing to release
the lock. Current implementation in posix xlator cannot differentiate
from opens from the same client or different clients.
What do you think ?
I like the idea. So basically we can propagate list_empty information of
'blocking_locks' list. And for sending locks, we need to use lk-owner
based on gfid so that locks from same client i.e. lkowner+transport are
granted irrespective of conflicting locks. The respective xls need to
make sure to order the fops so that they don't step on each other in a
single process. This can be used for entry-locks also.
I don't understand what are the benefits of checking for
lkowner+transport to grant a lock bypassing conflicts. It seems
dangerous and I don't see exactly how this can help the upper xlator.
If this xlator already needs to take care of fop ordering for each
inode, it can share the same lock. It seems there's no need to do
additional locking calls. I may be missing some detail though.
Afr at the time of 3.2 or 3.3 used to take full file lock for doing
self-heal. But this scheme was useless for VM healing. So we had to
migrate the locking in a backward compatible way, so this strategy was
employed where healing will take 128k chunk lock at a time heal that
chunk and move to next chunk, but at no point we needed another
self-heal to start. So the locking scheme came to be: Take a full file
lock, find good/bad copies, take a lock on chunk-1, unlock full lock,
heal chunk-1 then take a lock on chunk-2 unlock the lock on chunk-1 etc.
To have this we needed the locks to be granted even when there were
conflicting locks, so we chose (lk-owner+ transport) being same as a way
to grant conflicting locks. We found that this can lead to another
problem where truncate fop etc can hang, so we are moving to a different
mechanism now. You can find the complete lock evolution document here
https://github.com/gluster/glusterfs/blob/master/doc/code/xlators/cluster/afr/afr-locks-evolution.md
Thinking a litle more about the way to detect multiple accesses to the
same inode using the list of pending locks, there's a case where some
more logic must be added to avoid unnecessary delays.
Suppose you receive a request for an inode from one client. If there
isn't anyone else waiting, a flag is set into the answer indicating
that there's no conflict. After that the caller begins an eager lock
timer because there isn't anyone else waiting. During that timeout,
another client tries to access the same inode. It will block until the
eager lock timer expires (at this time it will release the inode lock)
or another request from the first client arrives (in this case the
request is served and the result will indicate that it should release
the lock since there are other clients waiting). When the lock is
released, it will be granted to the other client. It's possible that
this client completes the request before the first one tries to
acquire the lock again (because it had more requests pending), causing
that the second client initiates another eager lock timer because
there were no other client waiting at the moment of executing the
request. This is an unnecessary delay.
To avoid this problem, we could add a flag in the inodelk/entrylk
calls to indicate that the lock is released to let other clients to
proceed, but we will want the lock again as soon as possible. It would
be as a combined unlock and lock on the same inode sent in a single
request. This way we avoid one round trip and locks xlator has more up
to date information to decide if there are concurrent accesses or not.
The problem I see with the combined unlock+lock is the order of arrival
of these unlock requests + disconnects + possibility of deadlocks. i.e.
In best case scenario, the new-lock requests that come as part of the
unlock request go to pending-locks. But in the worst case it can end up
with granted lock on some bricks and pending-lock on some bricks because
of disconnects on just those bricks.
To prevent such cases, we need to wind the unlocks in an order. Which
will add to the latency. One other way to get the desired behaviour of
granted-lock from mount-2 to find that it shouldn't start eager-lock
because of pending locks from mount-1, mount-1 should send the lock with
different lk-owner before unlock of the previous one so that we are
guaranteed to notice the presence of pending-locks at the time of
granting lock from mount-2.
Pranith
Xavi
_______________________________________________
Gluster-devel mailing list
Gluster-devel@gluster.org
http://www.gluster.org/mailman/listinfo/gluster-devel
