linux-kernel - Re: [RFC][PATCH 6/7] freelist: Lock less freelist

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20200828144650.GF28468@redhat.com>
Date:   Fri, 28 Aug 2020 16:46:52 +0200
From:   Oleg Nesterov <oleg@...hat.com>
To:     Peter Zijlstra <peterz@...radead.org>
Cc:     linux-kernel@...r.kernel.org, mhiramat@...nel.org,
        Eddy_Wu@...ndmicro.com, x86@...nel.org, davem@...emloft.net,
        rostedt@...dmis.org, naveen.n.rao@...ux.ibm.com,
        anil.s.keshavamurthy@...el.com, linux-arch@...r.kernel.org,
        cameron@...dycamel.com, will@...nel.org, paulmck@...nel.org
Subject: Re: [RFC][PATCH 6/7] freelist: Lock less freelist

On 08/27, Peter Zijlstra wrote:
>
>  1 file changed, 129 insertions(+)

129 lines! And I spent more than 2 hours trying to understand these
129 lines ;) looks correct...

However, I still can't understand the usage of _acquire/release ops
in this code.

> +static inline void __freelist_add(struct freelist_node *node, struct freelist_head *list)
> +{
> +	/*
> +	 * Since the refcount is zero, and nobody can increase it once it's
> +	 * zero (except us, and we run only one copy of this method per node at
> +	 * a time, i.e. the single thread case), then we know we can safely
> +	 * change the next pointer of the node; however, once the refcount is
> +	 * back above zero, then other threads could increase it (happens under
> +	 * heavy contention, when the refcount goes to zero in between a load
> +	 * and a refcount increment of a node in try_get, then back up to
> +	 * something non-zero, then the refcount increment is done by the other
> +	 * thread) -- so if the CAS to add the node to the actual list fails,
> +	 * decrese the refcount and leave the add operation to the next thread
> +	 * who puts the refcount back to zero (which could be us, hence the
> +	 * loop).
> +	 */
> +	struct freelist_node *head = READ_ONCE(list->head);
> +
> +	for (;;) {
> +		WRITE_ONCE(node->next, head);
> +		atomic_set_release(&node->refs, 1);
> +
> +		if (!try_cmpxchg_release(&list->head, &head, node)) {

OK, these 2 _release above look understandable, they pair with
atomic_try_cmpxchg_acquire/try_cmpxchg_acquire in freelist_try_get().

> +			/*
> +			 * Hmm, the add failed, but we can only try again when
> +			 * the refcount goes back to zero.
> +			 */
> +			if (atomic_fetch_add_release(REFS_ON_FREELIST - 1, &node->refs) == 1)
> +				continue;

Do we really need _release ? Why can't atomic_fetch_add_relaxed() work?

> +static inline struct freelist_node *freelist_try_get(struct freelist_head *list)
> +{
> +	struct freelist_node *prev, *next, *head = smp_load_acquire(&list->head);
> +	unsigned int refs;
> +
> +	while (head) {
> +		prev = head;
> +		refs = atomic_read(&head->refs);
> +		if ((refs & REFS_MASK) == 0 ||
> +		    !atomic_try_cmpxchg_acquire(&head->refs, &refs, refs+1)) {
> +			head = smp_load_acquire(&list->head);
> +			continue;
> +		}
> +
> +		/*
> +		 * Good, reference count has been incremented (it wasn't at
> +		 * zero), which means we can read the next and not worry about
> +		 * it changing between now and the time we do the CAS.
> +		 */
> +		next = READ_ONCE(head->next);
> +		if (try_cmpxchg_acquire(&list->head, &head, next)) {
> +			/*
> +			 * Yay, got the node. This means it was on the list,
> +			 * which means should-be-on-freelist must be false no
> +			 * matter the refcount (because nobody else knows it's
> +			 * been taken off yet, it can't have been put back on).
> +			 */
> +			WARN_ON_ONCE(atomic_read(&head->refs) & REFS_ON_FREELIST);
> +
> +			/*
> +			 * Decrease refcount twice, once for our ref, and once
> +			 * for the list's ref.
> +			 */
> +			atomic_fetch_add(-2, &head->refs);

Do we the barriers implied by _fetch_? Why can't atomic_sub(2, refs) work?

> +		/*
> +		 * OK, the head must have changed on us, but we still need to decrement
> +		 * the refcount we increased.
> +		 */
> +		refs = atomic_fetch_add(-1, &prev->refs);

Cosmetic, but why not atomic_fetch_dec() ?

Oleg.