Re: [patch] cpufreq: mark cpufreq_tsc() as core_initcall_sync

From: Jens Axboe <jens.axboe@oracle.com>
To: Alan Stern <stern@rowland.harvard.edu>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>,
	Oleg Nesterov <oleg@tv-sign.ru>,
	Kernel development list <linux-kernel@vger.kernel.org>
Subject: Re: [patch] cpufreq: mark cpufreq_tsc() as core_initcall_sync
Date: Mon, 20 Nov 2006 21:13:35 +0100	[thread overview]
Message-ID: <20061120201334.GE8055@kernel.dk> (raw)
In-Reply-To: <Pine.LNX.4.44L0.0611201436230.7569-100000@iolanthe.rowland.org>

On Mon, Nov 20 2006, Alan Stern wrote:
> On Mon, 20 Nov 2006, Jens Axboe wrote:
> 
> > On Mon, Nov 20 2006, Alan Stern wrote:
> > > Paul:
> > > 
> > > Here's my version of your patch from yesterday.  It's basically the same, 
> > > but I cleaned up the code in a few places and fixed a bug (the sign of idx 
> > > in srcu_read_unlock).  Also I changed the init routine back to void, since 
> > > it's no longer an error if the per-cpu allocation fails.
> > > 
> > > More importantly, I added a static initializer and included the fast-path 
> > > in synchronize_srcu.  It's protected by the new symbol 
> > > SMP__STORE_MB_LOAD_WORKS, which should be defined in arch-specific headers 
> > > for those architectures where the store-mb-load pattern is safe.
> > 
> > Must we introduce memory allocations in srcu_read_lock()? It makes it
> > much harder and nastier for me to use. I'd much prefer a failing
> > init_srcu(), seems like a much better API.
> 
> Paul agrees with you that allocation failures in init_srcu() should be 
> passed back to the caller, and I certainly don't mind doing so.
> 
> However we can't remove the memory allocation in srcu_read_lock().  That
> was the point which started this whole thread: the per-cpu allocation
> cannot be done statically, and some users of a static SRCU structure can't
> easily call init_srcu() early enough.
> 
> Once the allocation succeeds, the overhead in srcu_read_lock() is minimal.

It's not about the overhead, it's about a potentially problematic
allocation.

-- 
Jens Axboe