Re: [PATCH/RFC] z3fold: use per-page read/write lock

From: Peter Zijlstra <peterz@infradead.org>
To: Vitaly Wool <vitalywool@gmail.com>
Cc: Linux-MM <linux-mm@kvack.org>,
	linux-kernel@vger.kernel.org, Dan Streetman <ddstreet@ieee.org>,
	Andrew Morton <akpm@linux-foundation.org>,
	Andi Kleen <andi@firstfloor.org>
Subject: Re: [PATCH/RFC] z3fold: use per-page read/write lock
Date: Thu, 10 Nov 2016 20:38:22 +0100	[thread overview]
Message-ID: <20161110193822.GK3175@twins.programming.kicks-ass.net> (raw)
In-Reply-To: <20161105144946.3b4be0ee799ae61a82e1d918@gmail.com>

On Sat, Nov 05, 2016 at 02:49:46PM +0100, Vitaly Wool wrote:
> +/* Read-lock a z3fold page */
> +static void z3fold_page_rlock(struct z3fold_header *zhdr)
> +{
> +	while (!atomic_add_unless(&zhdr->page_lock, 1, Z3FOLD_PAGE_WRITE_FLAG))
> +		cpu_relax();
> +	smp_mb();
> +}
> +
> +/* Read-unlock a z3fold page */
> +static void z3fold_page_runlock(struct z3fold_header *zhdr)
> +{
> +	atomic_dec(&zhdr->page_lock);
> +	smp_mb();
> +}
> +
> +/* Write-lock a z3fold page */
> +static void z3fold_page_wlock(struct z3fold_header *zhdr)
> +{
> +	while (atomic_cmpxchg(&zhdr->page_lock, 0, Z3FOLD_PAGE_WRITE_FLAG) != 0)
> +		cpu_relax();
> +	smp_mb();
> +}
> +
> +/* Write-unlock a z3fold page */
> +static void z3fold_page_wunlock(struct z3fold_header *zhdr)
> +{
> +	atomic_sub(Z3FOLD_PAGE_WRITE_FLAG, &zhdr->page_lock);
> +	smp_mb();
> +}

This is trivially broken. What Andi said, don't roll your own locks.

The unlocks want: smp_mb__before_atomic() _before_ the atomic for
'obvious' reasons.

Also, this lock has serious starvation issues and is reader biased.