Date:   Tue, 7 Dec 2021 10:34:30 +0100
From:   Peter Zijlstra <peterz@infradead.org>
To:     Linus Torvalds <torvalds@linux-foundation.org>
Cc:     Kees Cook <keescook@chromium.org>, Jens Axboe <axboe@kernel.dk>,
        Christoph Hellwig <hch@infradead.org>,
        "linux-block@vger.kernel.org" <linux-block@vger.kernel.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] block: switch to atomic_t for request references
Message-ID: <Ya8qptlJ4yLVUSBi@hirez.programming.kicks-ass.net>
References: <9f2ad6f1-c1bb-dfac-95c8-7d9eaa7110cc@kernel.dk>
 <Ya2zfVAwh4aQ7KVd@infradead.org>
 <Ya3KZiLg5lYjsGcQ@hirez.programming.kicks-ass.net>
 <CAHk-=wjXmGt9-JQp-wvup4y2tFNUCVjvx2W7MHzuAaxpryP4mg@mail.gmail.com>
 <282666e2-93d4-0302-b2d0-47d03395a6d4@kernel.dk>
 <202112061247.C5CD07E3C@keescook>
 <CAHk-=wh0RhnMfZG6xQJ=yHTgmPTaxjQOo1Q2=r+_ZR56yiRi4A@mail.gmail.com>
 <202112061455.F23512C3CB@keescook>
 <CAHk-=whLU+dk7EmPu5UC6DDSd76_dO4bVd4BkvxmR4W5-mmAgg@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAHk-=whLU+dk7EmPu5UC6DDSd76_dO4bVd4BkvxmR4W5-mmAgg@mail.gmail.com>
Precedence: bulk

On Mon, Dec 06, 2021 at 04:13:00PM -0800, Linus Torvalds wrote:

> Fix it to not unnecessarily use expensive compare-and-exchange loops,
> when you can safely just race a bit, safe in the knowledge that you're
> not going to race 2**31 times.
> 
> IOW, I think that "try_get_page()" function is basically the *much*
> superior version of what is currently a broken "refcount_inc()".
> 
> And yes, it does warn about that overflow case that you claim only
> refcount_t does. And does so without the broken semantics that
> refcount h as.

I think you should perhaps look at the current code again. Those cmpxchg
loops are gone. The only cmpxchg loop that's still there is for
add_not_zero(), and that is the exact same loop we have for
atomic_inc_not_zero(v) := atomic_add_unless(v,1,0) :=
atomic_fetch_add_unless(v,1,0) != 0.

The rest is exactly that LOCK XADD and assume you don't race 2^31 bits
worth.

Now, it could be GCC generates atrociously bad code simply because it
doesn't know it can use the flags from the XADD in which case we can
look at doing an arch asm implementation.