lkml.org 
[lkml]   [2015]   [Jan]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH v2 04/10] locks: move flock locks to file_lock_context
On Fri, 9 Jan 2015 06:31:55 -0800
Christoph Hellwig <hch@infradead.org> wrote:

> > void ceph_count_locks(struct inode *inode, int *fcntl_count, int *flock_count)
> > {
> > struct file_lock *lock;
> > + struct file_lock_context *ctx;
> >
> > *fcntl_count = 0;
> > *flock_count = 0;
> >
> > + spin_lock(&inode->i_lock);
>
> Seems like moving the locking around is unrelated to this patch.
>

Yeah that could be split out into a separate cleanup patch first. I'll
do that on the next iteration.

> > + list_for_each_entry(fl, &flctx->flc_flock, fl_list) {
> > + if (nfs_file_open_context(fl->fl_file)->state != state)
> > + continue;
> > + spin_unlock(&inode->i_lock);
> > + status = ops->recover_lock(state, fl);
> > + switch (status) {
> > + case 0:
> > + break;
> > + case -ESTALE:
> > + case -NFS4ERR_ADMIN_REVOKED:
> > + case -NFS4ERR_STALE_STATEID:
> > + case -NFS4ERR_BAD_STATEID:
> > + case -NFS4ERR_EXPIRED:
> > + case -NFS4ERR_NO_GRACE:
> > + case -NFS4ERR_STALE_CLIENTID:
> > + case -NFS4ERR_BADSESSION:
> > + case -NFS4ERR_BADSLOT:
> > + case -NFS4ERR_BAD_HIGH_SLOT:
> > + case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
> > + goto out;
> > + default:
> > + printk(KERN_ERR "NFS: %s: unhandled error %d\n",
> > + __func__, status);
> > + case -ENOMEM:
> > + case -NFS4ERR_DENIED:
> > + case -NFS4ERR_RECLAIM_BAD:
> > + case -NFS4ERR_RECLAIM_CONFLICT:
> > + /* kill_proc(fl->fl_pid, SIGLOST, 1); */
> > + status = 0;
> > + }
>
> Instead of duplicating this huge body of code it seems like a good idea
> to add a preparatory patch to factor it out into a helper function.
>

Sigh, I tried to do that first but the result was just too ugly. The
above logic is too deeply entwined into this function for that to work
well. I'm not usually a fan of cut and paste, but in this case I think
it's the best way to do this. The good news is that the duplication
goes away with the next patch in the series.

> > +static bool
> > +is_whole_file_wrlock(struct file_lock *fl)
> > +{
> > + return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX && fl->fl_type == F_WRLCK;
> > +}
>
> Please break this into multiple lines to stay under 80 characters.

Will do. I've probably violated that rule several times in this series
-- mea culpa. I'll clean that up for the next iteration.

--
Jeff Layton <jlayton@primarydata.com>


\
 
 \ /
  Last update: 2015-01-09 16:01    [W:1.107 / U:0.024 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site