Skip to contentSkip to conversation

bpf: Mitigate latency spikes associated with freeing non-preallocated htab

PATCH Superseded Yafang Shao 2 participants 3 messages Mar 26, 2024 – Mar 27, 2024 BPF

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bpf: Mitigate latency spikes associated with freeing non-preallocated htab

Following the recent upgrade of one of our BPF programs, we encountered significant latency spikes affecting other applications running on the same host. After thorough investigation, we identified that these spikes were primarily caused by the prolonged duration required to free a non-preallocated htab with approximately 2 million keys.

Notably, our kernel configuration lacks the presence of CONFIG_PREEMPT. In scenarios where kernel execution extends excessively, other threads might be starved of CPU time, resulting in latency issues across the system. To mitigate this, we've adopted a proactive approach by incorporating cond_resched() calls within the kernel code. This ensures that during lengthy kernel operations, the scheduler is invoked periodically to provide opportunities for other threads to execute.

---
 kernel/bpf/hashtab.c | 1 +
 1 file changed, 1 insertion(+)
kernel/bpf/hashtab.c +1 -0
index 3a088a5349bc..d3d5aad045cc 100644
@@ -1489,6 +1489,7 @@ static void delete_all_elements(struct bpf_htab *htab)
 		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
 			hlist_nulls_del_rcu(&l->hash_node);
 			htab_elem_free(htab, l);
+			cond_resched();
 		}
 	}
 	migrate_enable();
Signature · 1 line
2.30.1 (Apple Git-130)
reviewer
#1 lore raw
Quoted 24 lines · Yafang Shao

Following the recent upgrade of one of our BPF programs, we encountered significant latency spikes affecting other applications running on the same host. After thorough investigation, we identified that these spikes were primarily caused by the prolonged duration required to free a non-preallocated htab with approximately 2 million keys.

Notably, our kernel configuration lacks the presence of CONFIG_PREEMPT. In scenarios where kernel execution extends excessively, other threads might be starved of CPU time, resulting in latency issues across the system. To mitigate this, we've adopted a proactive approach by incorporating cond_resched() calls within the kernel code. This ensures that during lengthy kernel operations, the scheduler is invoked periodically to provide opportunities for other threads to execute.

---
  kernel/bpf/hashtab.c | 1 +
  1 file changed, 1 insertion(+)
kernel/bpf/hashtab.c +0 -0
index 3a088a5349bc..d3d5aad045cc 100644

kernel/bpf/hashtab.c static void delete_all_elements(struct bpf_htab *htab) · line 1493

Quoted 5 lines
kernel/bpf/hashtab.c +1 -0
@@ -1489,6 +1489,7 @@ static void delete_all_elements(struct bpf_htab *htab)
  		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
  			hlist_nulls_del_rcu(&l->hash_node);
  			htab_elem_free(htab, l);
+			cond_resched();
  		}

should we put cond_resched() here inside the top 'for' loop, but outside the bucket loop? Do you really have a long link list for a particular bucket? Otherwise, the patch looks good to me. In hashtab.c, we have cond_resched() in some other places to mitigate similar issues.

Quoted 2 lines
kernel/bpf/hashtab.c +0 -0
  	}
  	migrate_enable();
#2 lore raw
Quoted 22 lines · Yonghong Song
Quoted 19 lines · Yafang Shao

Following the recent upgrade of one of our BPF programs, we encountered significant latency spikes affecting other applications running on the same host. After thorough investigation, we identified that these spikes were primarily caused by the prolonged duration required to free a non-preallocated htab with approximately 2 million keys.

Notably, our kernel configuration lacks the presence of CONFIG_PREEMPT. In scenarios where kernel execution extends excessively, other threads might be starved of CPU time, resulting in latency issues across the system. To mitigate this, we've adopted a proactive approach by incorporating cond_resched() calls within the kernel code. This ensures that during lengthy kernel operations, the scheduler is invoked periodically to provide opportunities for other threads to execute.

---
  kernel/bpf/hashtab.c | 1 +
  1 file changed, 1 insertion(+)
Quoted 13 lines

kernel/bpf/hashtab.c static void delete_all_elements(struct bpf_htab *htab) · line 1493

Quoted 10 lines
kernel/bpf/hashtab.c +1 -0
index 3a088a5349bc..d3d5aad045cc 100644
@@ -1489,6 +1489,7 @@ static void delete_all_elements(struct bpf_htab *htab)
              hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
                      hlist_nulls_del_rcu(&l->hash_node);
                      htab_elem_free(htab, l);
+                     cond_resched();
              }

should we put cond_resched() here inside the top 'for' loop, but outside the bucket loop? Do you really have a long link list for a particular bucket?

We haven't monitored a long list for a particular bucket. Will change it. Thanks for your suggestion.

Quoted 5 lines

Otherwise, the patch looks good to me. In hashtab.c, we have cond_resched() in some other places to mitigate similar issues.

Quoted 2 lines
kernel/bpf/hashtab.c +0 -0
      }
      migrate_enable();
Signature · 2 lines
Regards
Yafang