[v2.0.x] prov/efa: Skip rx pkt refill under certain threshold

Libfabric currently refill the rx pkt pool in every cq read when there are >0 pkts to post, which makes it have chance to post ibv_recv 1-by-1 if there is only 1 pkt to post per cq read. Such 1-by-1 post is less performant than having a batch post once. This patch improves this strategy by introducing a threshold for the refilling. When When the number of internal rx pkts to post is lower than this threshold, the refill will be skipped. Also introduced FI_EFA_INTERNAL_RX_REFILL_THRESHOLD that allows tuning this parameter. Signed-off-by: Shi Jin <[email protected]> (cherry picked from commit a149f51)
HewlettPackard · Dec 11, 2024 · 62dbef6 · 62dbef6
1 parent 6deda39
commit 62dbef6
Show file tree

Hide file tree

Showing 8 changed files with 106 additions and 1 deletion.
diff --git a/man/fi_efa.7.md b/man/fi_efa.7.md
@@ -338,6 +338,11 @@ for details.
 : Use device's unsolicited write recv functionality when it's available. (Default: 1).
 Setting this environment variable to 0 can disable this feature.
 
+*FI_EFA_INTERNAL_RX_REFILL_THRESHOLD*
+: The threshold that EFA provider will refill the internal rx pkt pool. (Default: 8).
+When the number of internal rx pkts to post is lower than this threshold,
+the refill will be skipped.
+
 # SEE ALSO
 
 [`fabric`(7)](fabric.7.html),

diff --git a/prov/efa/src/efa_env.c b/prov/efa/src/efa_env.c
@@ -39,6 +39,7 @@ struct efa_env efa_env = {
 	.use_sm2 = false,
 	.huge_page_setting = EFA_ENV_HUGE_PAGE_UNSPEC,
 	.use_unsolicited_write_recv = 1,
+	.internal_rx_refill_threshold = 8,
 };
 
 /**
@@ -132,6 +133,7 @@ void efa_env_param_get(void)
 			    &efa_mr_max_cached_size);
 	fi_param_get_size_t(&efa_prov, "tx_size", &efa_env.tx_size);
 	fi_param_get_size_t(&efa_prov, "rx_size", &efa_env.rx_size);
+	fi_param_get_size_t(&efa_prov, "internal_rx_refill_threshold", &efa_env.internal_rx_refill_threshold);
 	fi_param_get_bool(&efa_prov, "rx_copy_unexp",
 			  &efa_env.rx_copy_unexp);
 	fi_param_get_bool(&efa_prov, "rx_copy_ooo",
@@ -232,6 +234,8 @@ void efa_env_define()
 			"will use huge page unless FI_EFA_FORK_SAFE is set to 1/on/true.");
 	fi_param_define(&efa_prov, "use_unsolicited_write_recv", FI_PARAM_BOOL,
 			"Use device's unsolicited write recv functionality when it's available. (Default: true)");
+	fi_param_define(&efa_prov, "internal_rx_refill_threshold", FI_PARAM_SIZE_T,
+			"The threshold that EFA provider will refill the internal rx pkt pool. (Default: %zu)", efa_env.internal_rx_refill_threshold);
 }
 
 

diff --git a/prov/efa/src/efa_env.h b/prov/efa/src/efa_env.h
@@ -79,6 +79,12 @@ struct efa_env {
 	int use_sm2;
 	enum efa_env_huge_page_setting huge_page_setting;
 	int use_unsolicited_write_recv;
+	/**
+	 * The threshold that EFA provider will refill the internal rx pkt pool.
+	 * When the number of internal rx pkts to post is lower than this threshold,
+	 * the refill will be skipped.
+	 */
+	size_t internal_rx_refill_threshold;
 };
 
 extern struct efa_env efa_env;

diff --git a/prov/efa/src/rdm/efa_rdm_ep.h b/prov/efa/src/rdm/efa_rdm_ep.h
@@ -263,6 +263,8 @@ struct efa_domain *efa_rdm_ep_domain(struct efa_rdm_ep *ep)
 
 void efa_rdm_ep_post_internal_rx_pkts(struct efa_rdm_ep *ep);
 
+int efa_rdm_ep_bulk_post_internal_rx_pkts(struct efa_rdm_ep *ep);
+
 /**
  * @brief return whether this endpoint should write error cq entry for RNR.
  *

diff --git a/prov/efa/src/rdm/efa_rdm_ep_utils.c b/prov/efa/src/rdm/efa_rdm_ep_utils.c
@@ -741,7 +741,11 @@ int efa_rdm_ep_bulk_post_internal_rx_pkts(struct efa_rdm_ep *ep)
 {
 	int i, err;
 
-	if (ep->efa_rx_pkts_to_post == 0)
+	/**
+	 * When efa_env.internal_rx_refill_threshold > efa_rdm_ep_get_rx_pool_size(ep),
+	 * we should always refill when the pool is empty.
+	 */
+	if (ep->efa_rx_pkts_to_post < MIN(efa_env.internal_rx_refill_threshold, efa_rdm_ep_get_rx_pool_size(ep)))
 		return 0;
 
 	assert(ep->efa_rx_pkts_to_post + ep->efa_rx_pkts_posted <= ep->efa_max_outstanding_rx_ops);

diff --git a/prov/efa/test/efa_unit_test_ep.c b/prov/efa/test/efa_unit_test_ep.c
@@ -1219,3 +1219,83 @@ void test_efa_rdm_ep_post_handshake_error_handling_pke_exhaustion(struct efa_res
 
 	free(pkt_entry_vec);
 }
+
+static
+void test_efa_rdm_ep_rx_refill_impl(struct efa_resource **state, int threshold, int rx_size)
+{
+	struct efa_resource *resource = *state;
+	struct efa_rdm_ep *efa_rdm_ep;
+	struct efa_rdm_pke *pkt_entry;
+	int i;
+	size_t threshold_orig;
+
+	if (threshold < 4 || rx_size < 4) {
+		fprintf(stderr, "Too small threshold or rx_size for this test\n");
+		fail();
+	}
+
+	threshold_orig = efa_env.internal_rx_refill_threshold;
+
+	efa_env.internal_rx_refill_threshold = threshold;
+
+	resource->hints = efa_unit_test_alloc_hints(FI_EP_RDM);
+	assert_non_null(resource->hints);
+	resource->hints->rx_attr->size = rx_size;
+	efa_unit_test_resource_construct_with_hints(resource, FI_EP_RDM, FI_VERSION(1, 14),
+	                                            resource->hints, true, true);
+
+	efa_rdm_ep = container_of(resource->ep, struct efa_rdm_ep, base_ep.util_ep.ep_fid);
+	assert_int_equal(efa_rdm_ep_get_rx_pool_size(efa_rdm_ep), rx_size);
+
+	/* Grow the rx pool and post rx pkts */
+	efa_rdm_ep_post_internal_rx_pkts(efa_rdm_ep);
+	assert_int_equal(efa_rdm_ep->efa_rx_pkts_posted, efa_rdm_ep_get_rx_pool_size(efa_rdm_ep));
+
+	assert_int_equal(efa_rdm_ep->efa_rx_pkts_to_post, 0);
+	for (i = 0; i < 4; i++) {
+		pkt_entry = ofi_bufpool_get_ibuf(efa_rdm_ep->efa_rx_pkt_pool, i);
+		assert_non_null(pkt_entry);
+		efa_rdm_pke_release_rx(pkt_entry);
+	}
+	assert_int_equal(efa_rdm_ep->efa_rx_pkts_to_post, 4);
+
+	efa_rdm_ep_bulk_post_internal_rx_pkts(efa_rdm_ep);
+
+	/**
+	 * efa_rx_pkts_to_post < FI_EFA_RX_REFILL_THRESHOLD
+	 * pkts should NOT be refilled
+	 */
+	assert_int_equal(efa_rdm_ep->efa_rx_pkts_to_post, 4);
+	assert_int_equal(efa_rdm_ep->efa_rx_pkts_posted, rx_size);
+
+	/* releasing more pkts to reach the threshold or rx_size*/
+	for (i = 4; i < MIN(rx_size, threshold); i++) {
+		pkt_entry = ofi_bufpool_get_ibuf(efa_rdm_ep->efa_rx_pkt_pool, i);
+		assert_non_null(pkt_entry);
+		efa_rdm_pke_release_rx(pkt_entry);
+	}
+
+	assert_int_equal(efa_rdm_ep->efa_rx_pkts_to_post, i);
+
+	efa_rdm_ep_bulk_post_internal_rx_pkts(efa_rdm_ep);
+
+	/**
+	 * efa_rx_pkts_to_post == min(FI_EFA_RX_REFILL_THRESHOLD, FI_EFA_RX_SIZE)
+	 * pkts should be refilled
+	 */
+	assert_int_equal(efa_rdm_ep->efa_rx_pkts_to_post, 0);
+	assert_int_equal(efa_rdm_ep->efa_rx_pkts_posted, rx_size + i);
+
+	/* recover the original value */
+	efa_env.internal_rx_refill_threshold = threshold_orig;
+}
+
+void test_efa_rdm_ep_rx_refill_threshold_smaller_than_rx_size(struct efa_resource **state)
+{
+	test_efa_rdm_ep_rx_refill_impl(state, 8, 64);
+}
+
+void test_efa_rdm_ep_rx_refill_threshold_larger_than_rx_size(struct efa_resource **state)
+{
+	test_efa_rdm_ep_rx_refill_impl(state, 128, 64);
+}
diff --git a/prov/efa/test/efa_unit_tests.c b/prov/efa/test/efa_unit_tests.c
@@ -114,6 +114,8 @@ int main(void)
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_zcpy_recv_cancel, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_zcpy_recv_eagain, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_post_handshake_error_handling_pke_exhaustion, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
+		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_rx_refill_threshold_smaller_than_rx_size, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
+		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_rx_refill_threshold_larger_than_rx_size, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_dgram_cq_read_empty_cq, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_ibv_cq_ex_read_empty_cq, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_ibv_cq_ex_read_failed_poll, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),

diff --git a/prov/efa/test/efa_unit_tests.h b/prov/efa/test/efa_unit_tests.h
@@ -128,6 +128,8 @@ void test_efa_rdm_ep_close_discard_posted_recv();
 void test_efa_rdm_ep_zcpy_recv_cancel();
 void test_efa_rdm_ep_zcpy_recv_eagain();
 void test_efa_rdm_ep_post_handshake_error_handling_pke_exhaustion();
+void test_efa_rdm_ep_rx_refill_threshold_smaller_than_rx_size();
+void test_efa_rdm_ep_rx_refill_threshold_larger_than_rx_size();
 void test_dgram_cq_read_empty_cq();
 void test_ibv_cq_ex_read_empty_cq();
 void test_ibv_cq_ex_read_failed_poll();
-Original file line number
+Diff line change
@@ Expand Up / @@ -263,6 +263,8 @@ struct efa_domain *efa_rdm_ep_domain(struct efa_rdm_ep *ep) @@
     void efa_rdm_ep_post_internal_rx_pkts(struct efa_rdm_ep *ep);
+    int efa_rdm_ep_bulk_post_internal_rx_pkts(struct efa_rdm_ep *ep);
     /**
      * @brief return whether this endpoint should write error cq entry for RNR.
      *
@@ Expand Down @@