[v2,5/5] lib/stack: remove pop cas release ordering
diff mbox series

Message ID 20200925174340.10014-6-steven.lariau@arm.com
State Accepted
Delegated to: David Marchand
Headers show
Series
  • lib/stack: improve lockfree C11 implementation
Related show

Checks

Context Check Description
ci/iol-mellanox-Performance success Performance Testing PASS
ci/iol-intel-Performance success Performance Testing PASS
ci/travis-robot success Travis build: passed
ci/iol-testing success Testing PASS
ci/iol-intel-Functional success Functional Testing PASS
ci/iol-broadcom-Functional success Functional Testing PASS
ci/iol-broadcom-Performance success Performance Testing PASS
ci/Intel-compilation success Compilation OK
ci/checkpatch warning coding style issues

Commit Message

Steven Lariau Sept. 25, 2020, 5:43 p.m. UTC
Replace the store-release by relaxed for the CAS success at the end of
pop. Release isn't needed, because there is not write to data that need
to be synchronized.
The only preceding write is when the length is decreased, but the length
CAS loop already ensures the right synchronization.
The situation to avoid is when a thread sees the old length but the new
list, that doesn't have enough items for pop to success.
But the CAS success on length before the pop loop ensures any core reads
and updates the latest length, preventing this situation.

The store-release is also used to make sure that the items are read
before the head is updated, in order to prevent a core in pop to read an
incorrect value because another core rewrites it with push.
But this isn't needed, because items are read only when removed from the
used list. Right after this, they are pushed to the free list, and the
store-release in push makes sure the items are read before they are
visible in the free list.

Signed-off-by: Steven Lariau <steven.lariau@arm.com>
Reviewed-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
---
 lib/librte_stack/rte_stack_lf_c11.h | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)

Comments

Eads, Gage Sept. 25, 2020, 5:57 p.m. UTC | #1
> -----Original Message-----
> From: Steven Lariau <steven.lariau@arm.com>
> Sent: Friday, September 25, 2020 12:44 PM
> To: Eads, Gage <gage.eads@intel.com>; Olivier Matz <olivier.matz@6wind.com>
> Cc: dev@dpdk.org; nd@arm.com; Steven Lariau <steven.lariau@arm.com>
> Subject: [PATCH v2 5/5] lib/stack: remove pop cas release ordering
> 
> Replace the store-release by relaxed for the CAS success at the end of
> pop. Release isn't needed, because there is not write to data that need
> to be synchronized.
> The only preceding write is when the length is decreased, but the length
> CAS loop already ensures the right synchronization.
> The situation to avoid is when a thread sees the old length but the new
> list, that doesn't have enough items for pop to success.
> But the CAS success on length before the pop loop ensures any core reads
> and updates the latest length, preventing this situation.
> 
> The store-release is also used to make sure that the items are read
> before the head is updated, in order to prevent a core in pop to read an
> incorrect value because another core rewrites it with push.
> But this isn't needed, because items are read only when removed from the
> used list. Right after this, they are pushed to the free list, and the
> store-release in push makes sure the items are read before they are
> visible in the free list.
> 
> Signed-off-by: Steven Lariau <steven.lariau@arm.com>
> Reviewed-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
> Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>

Acked-by: Gage Eads <gage.eads@intel.com>

Thanks,
Gage

Patch
diff mbox series

diff --git a/lib/librte_stack/rte_stack_lf_c11.h b/lib/librte_stack/rte_stack_lf_c11.h
index adb9f590d..8403196d5 100644
--- a/lib/librte_stack/rte_stack_lf_c11.h
+++ b/lib/librte_stack/rte_stack_lf_c11.h
@@ -141,11 +141,25 @@  __rte_stack_lf_pop_elems(struct rte_stack_lf_list *list,
 		new_head.top = tmp;
 		new_head.cnt = old_head.cnt + 1;
 
+		/*
+		 * The CAS should have release semantics to ensure that
+		 * items are read before the head is updated.
+		 * But this function is internal, and items are read
+		 * only when __rte_stack_lf_pop calls this function to
+		 * pop items from used list.
+		 * Then, those items are pushed to the free list.
+		 * Push uses a CAS store-release on head, which makes
+		 * sure that items are read before they are pushed to
+		 * the free list, without need for a CAS release here.
+		 * This CAS could also be used to ensure that the new
+		 * length is visible before the head update, but
+		 * acquire semantics on the length update is enough.
+		 */
 		success = rte_atomic128_cmp_exchange(
 				(rte_int128_t *)&list->head,
 				(rte_int128_t *)&old_head,
 				(rte_int128_t *)&new_head,
-				0, __ATOMIC_RELEASE,
+				0, __ATOMIC_RELAXED,
 				__ATOMIC_RELAXED);
 	} while (success == 0);