Date: Mon, 17 Oct 2016 15:00:05 +1100
From: Nicholas Piggin <npiggin@gmail.com>
To: Joel Fernandes <joelaf@google.com>
Cc: linux-kernel@vger.kernel.org, linux-rt-users@vger.kernel.org,
        Chris Wilson <chris@chris-wilson.co.uk>,
        Jisheng Zhang <jszhang@marvell.com>, John Dias <joaodias@google.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        linux-mm@kvack.org (open list:MEMORY MANAGEMENT)
Subject: Re: [PATCH v2] mm: vmalloc: Replace purge_lock spinlock with atomic
 refcount
Message-ID: <20161017150005.4c8f890d@roar.ozlabs.ibm.com>
In-Reply-To: <1476528162-21981-1-git-send-email-joelaf@google.com>
References: <1476528162-21981-1-git-send-email-joelaf@google.com>
Organization: IBM
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1035
Lines: 24

On Sat, 15 Oct 2016 03:42:42 -0700
Joel Fernandes <joelaf@google.com> wrote:

> The purge_lock spinlock causes high latencies with non RT kernel. This has been
> reported multiple times on lkml [1] [2] and affects applications like audio.
> 
> In this patch, I replace the spinlock with an atomic refcount so that
> preemption is kept turned on during purge. This Ok to do since [3] builds the
> lazy free list in advance and atomically retrieves the list so any instance of
> purge will have its own list it is purging. Since the individual vmap area
> frees are themselves protected by a lock, this is Ok.

This is a good idea, and good results, but that's not what the spinlock was
for -- it was for enforcing the sync semantics.

Going this route, you'll have to audit callers to expect changed behavior
and change documentation of sync parameter.

I suspect a better approach would be to instead use a mutex for this, and
require that all sync=1 callers be able to sleep. I would say that most
probably already can.

Thanks,
Nick