Re: [RFC][PATCH 7/7] kref: Implement using refcount_t

[Will Deacon]

On Fri, Nov 18, 2016 at 12:37:18PM +0100, Peter Zijlstra wrote:
> On Fri, Nov 18, 2016 at 10:07:26AM +0000, Reshetova, Elena wrote:
> > 
> > Peter do you have the changes to the refcount_t interface compare to
> > the version in this patch? 
> 
> > We are now starting working on atomic_t --> refcount_t conversions and
> > it would save a bit of work to have latest version from you that we
> > can be based upon. 
> 
> The latestest version below, mostly just comment changes since last
> time.
> 
> ---
> Subject: refcount_t: A special purpose refcount type
> From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Date: Mon Nov 14 18:06:19 CET 2016
> 
> Provide refcount_t, an atomic_t like primitive built just for
> refcounting.
> 
> It provides saturation semantics such that overflow becomes impossible
> and thereby 'spurious' use-after-free is avoided.
> 
> Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> ---
>  include/linux/refcount.h |  241 +++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 241 insertions(+)
> 
> --- /dev/null
> +++ b/include/linux/refcount.h
> @@ -0,0 +1,241 @@
> +#ifndef _LINUX_REFCOUNT_H
> +#define _LINUX_REFCOUNT_H
> +
> +/*
> + * Variant of atomic_t specialized for reference counts.
> + *
> + * The interface matches the atomic_t interface (to aid in porting) but only
> + * provides the few functions one should use for reference counting.
> + *
> + * It differs in that the counter saturates at UINT_MAX and will not move once
> + * there. This avoids wrapping the counter and causing 'spurious'
> + * use-after-free issues.
> + *
> + * Memory ordering rules are slightly relaxed wrt regular atomic_t functions
> + * and provide only what is strictly required for refcounts.
> + *
> + * The increments are fully relaxed; these will not provide ordering. The
> + * rationale is that whatever is used to obtain the object we're increasing the
> + * reference count on will provide the ordering. For locked data structures,
> + * its the lock acquire, for RCU/lockless data structures its the dependent
> + * load.
> + *
> + * Do note that inc_not_zero() provides a control dependency which will order
> + * future stores against the inc, this ensures we'll never modify the object
> + * if we did not in fact acquire a reference.
> + *
> + * The decrements will provide release order, such that all the prior loads and
> + * stores will be issued before, it also provides a control dependency, which
> + * will order us against the subsequent free().
> + *
> + * The control dependency is against the load of the cmpxchg (ll/sc) that
> + * succeeded. This means the stores aren't fully ordered, but this is fine
> + * because the 1->0 transition indicates no concurrency.
> + *
> + * Note that the allocator is responsible for ordering things between free()
> + * and alloc().
> + *
> + *
> + * Note: the implementation hard relies on increments, bigger than 1 additions
> + *       need explicit overflow -> saturation logic.
> + *
> + */
> +
> +#include <linux/atomic.h>
> +#include <linux/bug.h>
> +#include <linux/mutex.h>
> +#include <linux/spinlock.h>
> +
> +typedef struct refcount_struct {
> +	atomic_t refs;
> +} refcount_t;
> +
> +#define REFCOUNT_INIT(n)	{ .refs = ATOMIC_INIT(n), }
> +
> +static inline void refcount_set(refcount_t *r, int n)
> +{
> +	atomic_set(&r->refs, n);
> +}
> +
> +static inline unsigned int refcount_read(const refcount_t *r)
> +{
> +	return atomic_read(&r->refs);
> +}

Minor nit, but it might be worth being consistent in our usage of int
(parameter to refcount_set) and unsigned int (return value of
refcount_read).

> +
> +/*
> + * Similar to atomic_inc(), will saturate at UINT_MAX and WARN.
> + *
> + * Provides no memory ordering, it is assumed the caller already has a
> + * reference on the object, will WARN when this is not so.
> + */
> +static inline void refcount_inc(refcount_t *r)
> +{
> +	unsigned int old, new, val = atomic_read(&r->refs);
> +
> +	for (;;) {
> +		WARN(!val, "refcount_t: increment on 0; use-after-free.\n");
> +
> +		if (unlikely(val == UINT_MAX))
> +			return;
> +
> +		new = val + 1;
> +		old = atomic_cmpxchg_relaxed(&r->refs, val, new);
> +		if (old == val)
> +			break;
> +
> +		val = old;
> +	}
> +
> +	WARN(new == UINT_MAX, "refcount_t: saturated; leaking memory.\n");
> +}
> +
> +/*
> + * Similar to atomic_inc_not_zero(), will saturate at UINT_MAX and WARN.
> + *
> + * Provides no memory ordering, it is assumed the caller has guaranteed the
> + * object memory to be stable (RCU, etc.). It does provide a control dependency
> + * and thereby orders future stores. See the comment on top.
> + */
> +static inline __must_check
> +bool refcount_inc_not_zero(refcount_t *r)
> +{
> +	unsigned int old, new, val = atomic_read(&r->refs);
> +
> +	for (;;) {
> +		if (!val)
> +			return false;
> +
> +		if (unlikely(val == UINT_MAX))
> +			return true;
> +
> +		new = val + 1;
> +		old = atomic_cmpxchg_relaxed(&r->refs, val, new);
> +		if (old == val)
> +			break;
> +
> +		val = old;

Hmm, it's a shame this code is duplicated from refcount_inc, but I suppose
you can actually be racing against the counter going to zero here and really
need to check it each time round the loop. Humph. That said, given that
refcount_inc WARNs if the thing is zero, maybe that could just call
refcount_inc_not_zero and warn if it returns false? Does it matter that
we don't actually do the increment?

> +	}
> +
> +	WARN(new == UINT_MAX, "refcount_t: saturated; leaking memory.\n");
> +
> +	return true;
> +}
> +
> +/*
> + * Similar to atomic_dec_and_test(), it will WARN on underflow and fail to
> + * decrement when saturated at UINT_MAX.

It also fails to decrement in the underflow case (which is fine, but not
obvious from the comment). Same thing below.

> + *
> + * Provides release memory ordering, such that prior loads and stores are done
> + * before, and provides a control dependency such that free() must come after.
> + * See the comment on top.
> + */
> +static inline __must_check
> +bool refcount_dec_and_test(refcount_t *r)
> +{
> +	unsigned int old, new, val = atomic_read(&r->refs);
> +
> +	for (;;) {
> +		if (val == UINT_MAX)
> +			return false;
> +
> +		new = val - 1;
> +		if (WARN(new > val, "refcount_t: underflow; use-after-free.\n"))
> +			return false;

Wouldn't it be clearer to compare val with 0 before doing the decrement?

Will