[Top] [All Lists]

Re: [PATCH 0/5] xfs: more patches for 3.13

To: Dave Chinner <david@xxxxxxxxxxxxx>, Ben Myers <bpm@xxxxxxx>
Subject: Re: [PATCH 0/5] xfs: more patches for 3.13
From: Eric Sandeen <sandeen@xxxxxxxxxxx>
Date: Tue, 12 Nov 2013 19:16:03 -0600
Cc: xfs@xxxxxxxxxxx
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <20131107015706.GM6188@dastard>
References: <1383280040-21979-1-git-send-email-david@xxxxxxxxxxxxx> <20131106230133.GX1935@xxxxxxx> <20131107015706.GM6188@dastard>
User-agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.9; rv:24.0) Gecko/20100101 Thunderbird/24.1.0
On 11/6/13, 7:57 PM, Dave Chinner wrote:
> On Wed, Nov 06, 2013 at 05:01:33PM -0600, Ben Myers wrote:
>> On Fri, Nov 01, 2013 at 03:27:15PM +1100, Dave Chinner wrote:
>>> Hi folks,
>>> The following series follows up the recently committed series of
>>> patches for 3.13. The first two patches are the remaining
>>> uncommitted patches from the previous series.
>>> The next two patches are tracing patches, one for AIL manipulations
>>> and the other for AGF and AGI read operations. Both of these were
>>> written during recent debugging sessions, and both proved useful so
>>> should be added to the menagerie of tracepoints we already have
>>> avaialble.
>>> The final patch is the increasing of the inode cluster size for v5
>>> filesystems. I'd like to get this into v5 filesystems for 3.13 so we
>>> get wider exposure of it ASAP so we have more data available to be
>>> able to make informed decisions about how to bring this back to v4
>>> filesystems in a safe and controlled manner.
>> Applied 3 and 4.  I still don't understand why the locking on patch 2 is
>> correct.  Seems like the readers of i_version hold different locks than we do
>> when we log the inode.  Maybe Christoph can help me with that.
> Readers don't need to hold a spinlock, and many don't. The spinlock
> is only there to prevent concurrent updates from "losing" an update
> due to races.  All modifications to XFS inodes occur via
> transactions, inodes are locked exclusively in transactions and
> hence we will never lose i_version updates due to races. Hence we
> don't need the spinlock during the update, either.

I'm not completely convinced that readers don't need to.  What happens when
we read in the middle of an update?  Especially when a 32-bit box reads the
64-bit value in the middle of an update?

NFS is the only reader we care about (right?)

I see a several paths to i_version reads in nfs; so far I'm finding locked 

<2 callers of nfs_refresh_inode_locked>
                                (... && inode->i_version == 
                if (inode->i_version != fattr->change_attr) {
                        (... && inode->i_version != fattr->change_attr)


        if (!cinfo->atomic || cinfo->before != dir->i_version)


         fattr->pre_change_attr = inode->i_version;


I haven't audited everything but do you have an example of an unlocked
reader (which is relevant to xfs)?


<Prev in Thread] Current Thread [Next in Thread>