Date: Mon, 1 Feb 2010 22:53:41 -0800
From: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
To: Al Viro <viro@ZenIV.linux.org.uk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>,
       linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org
Subject: Re: [PATCH][RFC] %pd - for printing dentry name
Message-ID: <20100202065341.GF6292@linux.vnet.ibm.com>
Reply-To: paulmck@linux.vnet.ibm.com
References: <20100201222511.GA12882@ZenIV.linux.org.uk> <alpine.LFD.2.00.1002011431490.3664@localhost.localdomain> <20100201231847.GC12882@ZenIV.linux.org.uk>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20100201231847.GC12882@ZenIV.linux.org.uk>
User-Agent: Mutt/1.5.15+20070412 (2007-04-11)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3162
Lines: 66

On Mon, Feb 01, 2010 at 11:18:47PM +0000, Al Viro wrote:
> On Mon, Feb 01, 2010 at 02:37:32PM -0800, Linus Torvalds wrote:
> 
> > > * don't use %pd under dentry->d_lock, use dentry->d_name.name instead; in
> > > that case it *is* safe.  Incidentally, ->d_lock isn't held a lot.
> > 
> > I realize we can just call it a rule, and yes, d_lock is held much less 
> > than something like console_lock etc that we've had ABBA issues with, but 
> > still..
> 
> > Quite frankly, I'd _much_ rather see something like just always freeing 
> > the dentry names (when they aren't inlined) using RCU. The VFS layer quite 
> > possibly would want to do that anyway at some point (eg Nick's VFS 
> > scalability patches), and then we could make it just a RCU read-lock or 
> > whatever (interrupt disable, what-not) instead.
> > 
> > And I'm much happier with printk doing that kind of thing, and wouldn't 
> > have issues with that kind of much weaker locking.
> 
> Ehh...  RCU will save you from stepping on freed memory, but it still will
> leave the joy of half-updated string with length out of sync with it, etc.
> We probably can get away with that, but we'll have to be a lot more careful
> with the order of updating these suckers in d_move_locked et.al.
> 
> I don't know...  Note that if we end up adding something extra to struct
> dentry, we might as well just add *another* spinlock, taken only under
> ->d_lock and only in two places in dcache.c that change d_name.  That kind
> of thing is trivial to enforce (just grep over the tree once in a while)
> and if it shares the cacheline with d_lock, we shouldn't get any real overhead
> in d_move()/d_materialise_unique().  I'm not particulary fond of that variant,
> but it's at least guaranteed to be devoid of subtleties.
> 
> If RCU folks can come up with a sane suggestions that would be robust and
> wouldn't bloat dentry - sure, I'm all for it.  If not...

Here is an approximation that might inspire someone to come up with a
real solution.

One approach would be to store the name length with the name, so that
struct qstr loses the "len" field, and so that its "name" field points
to a struct that has a "len" field followed by an array of const
unsigned char.  That way, the name and length are closely associated.
When you pick up a struct qstr's "name" pointer, you are guaranteed to
get a length that matches the name.

Unfortunately:

o	In theory, this leaves the length of the dentry unchanged, but
	alignment is a problem on 64-bit systems.  Also, the long names
	gain an extra four bytes.

o	If you get a pointer to the d_iname small-name field, rename
	might still change the name out from under you.  This could in
	theory be fixed by refusing to re-use the d_iname field until
	an RCU grace period had elapsed (using an external structure
	instead).  In practice, not sure if this is really a reasonable
	approach.

Thoughts?

							Thanx, Paul
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/