[dpdk-dev] [PATCH v4 3/3] lib/lpm: use atomic store to avoid partial update

[Ruifeng Wang (Arm Technology China)]

Hi Honnappa,

> -----Original Message-----
> From: Honnappa Nagarahalli <Honnappa.Nagarahalli at arm.com>
> Sent: Monday, July 8, 2019 12:57
> To: Ruifeng Wang (Arm Technology China) <Ruifeng.Wang at arm.com>;
> vladimir.medvedkin at intel.com; bruce.richardson at intel.com
> Cc: dev at dpdk.org; Gavin Hu (Arm Technology China) <Gavin.Hu at arm.com>;
> nd <nd at arm.com>; Ruifeng Wang (Arm Technology China)
> <Ruifeng.Wang at arm.com>; Honnappa Nagarahalli
> <Honnappa.Nagarahalli at arm.com>; nd <nd at arm.com>
> Subject: RE: [PATCH v4 3/3] lib/lpm: use atomic store to avoid partial update
> 
> >
> > Compiler could generate non-atomic stores for whole table entry updating.
> > This may cause incorrect nexthop to be returned, if the byte with
> > valid flag is updated prior to the byte with next hot is updated.
>                                                            ^^^^^^^ Should be nexthop
> 
> >
> > Changed to use atomic store to update whole table entry.
> >
> > Suggested-by: Medvedkin Vladimir <vladimir.medvedkin at intel.com>
> > Signed-off-by: Ruifeng Wang <ruifeng.wang at arm.com>
> > Reviewed-by: Gavin Hu <gavin.hu at arm.com>
> > ---
> > v4: initial version
> >
> >  lib/librte_lpm/rte_lpm.c | 34 ++++++++++++++++++++++++----------
> >  1 file changed, 24 insertions(+), 10 deletions(-)
> >
> > diff --git a/lib/librte_lpm/rte_lpm.c b/lib/librte_lpm/rte_lpm.c index
> > baa6e7460..5d1dbd7e6 100644
> > --- a/lib/librte_lpm/rte_lpm.c
> > +++ b/lib/librte_lpm/rte_lpm.c
> > @@ -767,7 +767,9 @@ add_depth_small_v20(struct rte_lpm_v20 *lpm,
> > uint32_t ip, uint8_t depth,
> >  					 * Setting tbl8 entry in one go to avoid
> >  					 * race conditions
> >  					 */
> > -					lpm->tbl8[j] = new_tbl8_entry;
> > +					__atomic_store(&lpm->tbl8[j],
> > +						&new_tbl8_entry,
> > +						__ATOMIC_RELAXED);
> >
> >  					continue;
> >  				}
> > @@ -837,7 +839,9 @@ add_depth_small_v1604(struct rte_lpm *lpm,
> > uint32_t ip, uint8_t depth,
> >  					 * Setting tbl8 entry in one go to avoid
> >  					 * race conditions
> >  					 */
> > -					lpm->tbl8[j] = new_tbl8_entry;
> > +					__atomic_store(&lpm->tbl8[j],
> > +						&new_tbl8_entry,
> > +						__ATOMIC_RELAXED);
> >
> >  					continue;
> >  				}
> > @@ -965,7 +969,8 @@ add_depth_big_v20(struct rte_lpm_v20 *lpm,
> > uint32_t ip_masked, uint8_t depth,
> >  				 * Setting tbl8 entry in one go to avoid race
> >  				 * condition
> >  				 */
> > -				lpm->tbl8[i] = new_tbl8_entry;
> > +				__atomic_store(&lpm->tbl8[i],
> > &new_tbl8_entry,
> > +						__ATOMIC_RELAXED);
> >
> >  				continue;
> >  			}
> > @@ -1100,7 +1105,8 @@ add_depth_big_v1604(struct rte_lpm *lpm,
> > uint32_t ip_masked, uint8_t depth,
> >  				 * Setting tbl8 entry in one go to avoid race
> >  				 * condition
> >  				 */
> > -				lpm->tbl8[i] = new_tbl8_entry;
> > +				__atomic_store(&lpm->tbl8[i],
> > &new_tbl8_entry,
> > +						__ATOMIC_RELAXED);
> >
> >  				continue;
> >  			}
> > @@ -1393,7 +1399,9 @@ delete_depth_small_v20(struct rte_lpm_v20
> *lpm,
> > uint32_t ip_masked,
> >
> > 	RTE_LPM_TBL8_GROUP_NUM_ENTRIES); j++) {
> >
> >  					if (lpm->tbl8[j].depth <= depth)
> > -						lpm->tbl8[j] =
> > new_tbl8_entry;
> > +						__atomic_store(&lpm-
> >tbl8[j],
> > +							&new_tbl8_entry,
> > +							__ATOMIC_RELAXED);
> >  				}
> >  			}
> >  		}
> > @@ -1490,7 +1498,9 @@ delete_depth_small_v1604(struct rte_lpm *lpm,
> > uint32_t ip_masked,
> >
> > 	RTE_LPM_TBL8_GROUP_NUM_ENTRIES); j++) {
> >
> >  					if (lpm->tbl8[j].depth <= depth)
> > -						lpm->tbl8[j] =
> > new_tbl8_entry;
> > +						__atomic_store(&lpm-
> >tbl8[j],
> > +							&new_tbl8_entry,
> > +							__ATOMIC_RELAXED);
> >  				}
> >  			}
> >  		}
> > @@ -1646,7 +1656,8 @@ delete_depth_big_v20(struct rte_lpm_v20 *lpm,
> > uint32_t ip_masked,
> >  		 */
> >  		for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
> >  			if (lpm->tbl8[i].depth <= depth)
> > -				lpm->tbl8[i] = new_tbl8_entry;
> > +				__atomic_store(&lpm->tbl8[i],
> > &new_tbl8_entry,
> > +						__ATOMIC_RELAXED);
> >  		}
> >  	}
> >
> > @@ -1677,7 +1688,8 @@ delete_depth_big_v20(struct rte_lpm_v20 *lpm,
> > uint32_t ip_masked,
> >  		/* Set tbl24 before freeing tbl8 to avoid race condition.
> >  		 * Prevent the free of the tbl8 group from hoisting.
> >  		 */
> > -		lpm->tbl24[tbl24_index] = new_tbl24_entry;
> > +		__atomic_store(&lpm->tbl24[tbl24_index],
> > &new_tbl24_entry,
> > +				__ATOMIC_RELAXED);
> >  		__atomic_thread_fence(__ATOMIC_RELEASE);
> >  		tbl8_free_v20(lpm->tbl8, tbl8_group_start);
> tbl8_alloc_v20/tbl8_free_v20 need to be updated to use __atomic_store
> 
tbl8_alloc_v20/tbl8_free_v20 updates a single field of table entry. The process
is already atomic. Do we really need to use __atomic_store?

> >  	}
> > @@ -1730,7 +1742,8 @@ delete_depth_big_v1604(struct rte_lpm *lpm,
> > uint32_t ip_masked,
> >  		 */
> >  		for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
> >  			if (lpm->tbl8[i].depth <= depth)
> > -				lpm->tbl8[i] = new_tbl8_entry;
> > +				__atomic_store(&lpm->tbl8[i],
> > &new_tbl8_entry,
> > +						__ATOMIC_RELAXED);
> >  		}
> >  	}
> >
> > @@ -1761,7 +1774,8 @@ delete_depth_big_v1604(struct rte_lpm *lpm,
> > uint32_t ip_masked,
> >  		/* Set tbl24 before freeing tbl8 to avoid race condition.
> >  		 * Prevent the free of the tbl8 group from hoisting.
> >  		 */
> > -		lpm->tbl24[tbl24_index] = new_tbl24_entry;
> > +		__atomic_store(&lpm->tbl24[tbl24_index],
> > &new_tbl24_entry,
> > +				__ATOMIC_RELAXED);
> >  		__atomic_thread_fence(__ATOMIC_RELEASE);
> >  		tbl8_free_v1604(lpm->tbl8, tbl8_group_start);
> tbl8_alloc_v1604 /tbl8_free_v1604 need to be updated to use
> __atomic_store
Ditto.

> 
> >  	}
> > --
> > 2.17.1