From: Xiaoyun Li <xiaoyun.li@intel.com>
To: beilei.xing@intel.com, qi.z.zhang@intel.com, ferruh.yigit@intel.com
Cc: dev@dpdk.org, zhiyong.yang@intel.com, thomas@monjalon.net,
bruce.richardson@intel.com, Xiaoyun Li <xiaoyun.li@intel.com>
Subject: [dpdk-dev] [PATCH v7] net/i40e: add interface to use latest vec path
Date: Tue, 18 Sep 2018 10:22:40 +0800 [thread overview]
Message-ID: <20180918022240.49376-1-xiaoyun.li@intel.com> (raw)
In-Reply-To: <1535595399-430873-1-git-send-email-xiaoyun.li@intel.com>
For IA, the AVX2 vector path is only recommended to be used on later
platforms (identified by AVX512 support, like SKL etc.) This is because
performance benchmark shows downgrade when running AVX2 vector path on
early platform (BDW/HSW) in some cases. But we still observe perf gain
with some real work loading.
So this patch introduced the new devarg use-latest-supported-vec to
force the driver always selecting the latest supported vec path. Then
apps are able to take AVX2 path on early platforms. And this logic can
be re-used if we will have AVX512 vec path in future.
This patch only affects IA platforms. The selected vec path would be
like the following:
Without devarg/devarg = 0:
Machine vPMD
AVX512F AVX2
AVX2 SSE4.2
SSE4.2 SSE4.2
<SSE4.2 Not Supported
With devarg = 1
Machine vPMD
AVX512F AVX2
AVX2 AVX2
SSE4.2 SSE4.2
<SSE4.2 Not Supported
Other platforms can also apply the same logic if necessary in future.
Signed-off-by: Xiaoyun Li <xiaoyun.li@intel.com>
---
v7:
* Use uint_8 instead of bool type for struct member.
v6:
* Polish the doc and commit log.
* Use rte_kvargs_process instead of directly kvlist internals.
v5:
* Simpify the rx set function.
v4:
* Polish the codes.
v3:
* Polish the doc and commit log.
v2:
* Correct the calling of the wrong function last time.
* Fix seg fault bug.
---
doc/guides/nics/i40e.rst | 8 ++
doc/guides/rel_notes/release_18_11.rst | 4 +
drivers/net/i40e/i40e_ethdev.c | 65 ++++++++++-
drivers/net/i40e/i40e_ethdev.h | 3 +
drivers/net/i40e/i40e_rxtx.c | 143 +++++++++++++------------
5 files changed, 155 insertions(+), 68 deletions(-)
diff --git a/doc/guides/nics/i40e.rst b/doc/guides/nics/i40e.rst
index 5d8500cef..6393baf0b 100644
--- a/doc/guides/nics/i40e.rst
+++ b/doc/guides/nics/i40e.rst
@@ -163,6 +163,14 @@ Runtime Config Options
Currently hot-plugging of representor ports is not supported so all required
representors must be specified on the creation of the PF.
+- ``Use latest supported vector`` (default ``disable``)
+
+ Latest supported vector path may not always get the best perf so vector path was
+ recommended to use only on later platform. But users may want the latest vector path
+ since it can get better perf in some real work loading cases. So ``devargs`` param
+ ``use-latest-supported-vec`` is introduced, for example::
+ -w 84:00.0,use-latest-supported-vec=1
+
Driver compilation and testing
------------------------------
diff --git a/doc/guides/rel_notes/release_18_11.rst b/doc/guides/rel_notes/release_18_11.rst
index 8c4bb5447..49fdcdfc8 100644
--- a/doc/guides/rel_notes/release_18_11.rst
+++ b/doc/guides/rel_notes/release_18_11.rst
@@ -67,6 +67,10 @@ New Features
SR-IOV option in Hyper-V and Azure. This is an alternative to the previous
vdev_netvsc, tap, and failsafe drivers combination.
+* **Added a devarg to use the latest supported vector path.**
+ A new devarg ``use-latest-supported-vec`` was introduced to allow users to
+ choose the latest vector path that the platform supported. For example, users
+ can use AVX2 vector path on BDW/HSW to get better performance.
API Changes
-----------
diff --git a/drivers/net/i40e/i40e_ethdev.c b/drivers/net/i40e/i40e_ethdev.c
index 24d73f2ff..51ad847e8 100644
--- a/drivers/net/i40e/i40e_ethdev.c
+++ b/drivers/net/i40e/i40e_ethdev.c
@@ -44,6 +44,7 @@
#define ETH_I40E_FLOATING_VEB_LIST_ARG "floating_veb_list"
#define ETH_I40E_SUPPORT_MULTI_DRIVER "support-multi-driver"
#define ETH_I40E_QUEUE_NUM_PER_VF_ARG "queue-num-per-vf"
+#define ETH_I40E_USE_LATEST_VEC "use-latest-supported-vec"
#define I40E_CLEAR_PXE_WAIT_MS 200
@@ -409,6 +410,7 @@ static const char *const valid_keys[] = {
ETH_I40E_FLOATING_VEB_LIST_ARG,
ETH_I40E_SUPPORT_MULTI_DRIVER,
ETH_I40E_QUEUE_NUM_PER_VF_ARG,
+ ETH_I40E_USE_LATEST_VEC,
NULL};
static const struct rte_pci_id pci_id_i40e_map[] = {
@@ -1202,6 +1204,64 @@ i40e_aq_debug_write_global_register(struct i40e_hw *hw,
return i40e_aq_debug_write_register(hw, reg_addr, reg_val, cmd_details);
}
+static int
+i40e_parse_latest_vec_handler(__rte_unused const char *key,
+ const char *value,
+ void *opaque)
+{
+ struct i40e_adapter *ad;
+ int use_latest_vec;
+
+ ad = (struct i40e_adapter *)opaque;
+
+ use_latest_vec = atoi(value);
+
+ if (use_latest_vec != 0 && use_latest_vec != 1)
+ PMD_DRV_LOG(WARNING, "Value should be 0 or 1, set it as 1!");
+
+ ad->use_latest_vec = (uint8_t)use_latest_vec;
+
+ return 0;
+}
+
+static int
+i40e_use_latest_vec(struct rte_eth_dev *dev)
+{
+ struct i40e_adapter *ad =
+ I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
+ struct rte_kvargs *kvlist;
+ int kvargs_count;
+
+ ad->use_latest_vec = false;
+
+ if (!dev->device->devargs)
+ return 0;
+
+ kvlist = rte_kvargs_parse(dev->device->devargs->args, valid_keys);
+ if (!kvlist)
+ return -EINVAL;
+
+ kvargs_count = rte_kvargs_count(kvlist, ETH_I40E_USE_LATEST_VEC);
+ if (!kvargs_count) {
+ rte_kvargs_free(kvlist);
+ return 0;
+ }
+
+ if (kvargs_count > 1)
+ PMD_DRV_LOG(WARNING, "More than one argument \"%s\" and only "
+ "the first invalid or last valid one is used !",
+ ETH_I40E_USE_LATEST_VEC);
+
+ if (rte_kvargs_process(kvlist, ETH_I40E_USE_LATEST_VEC,
+ i40e_parse_latest_vec_handler, ad) < 0) {
+ rte_kvargs_free(kvlist);
+ return -EINVAL;
+ }
+
+ rte_kvargs_free(kvlist);
+ return 0;
+}
+
#define I40E_ALARM_INTERVAL 50000 /* us */
static int
@@ -1266,6 +1326,8 @@ eth_i40e_dev_init(struct rte_eth_dev *dev, void *init_params __rte_unused)
/* Check if need to support multi-driver */
i40e_support_multi_driver(dev);
+ /* Check if users want the latest supported vec path */
+ i40e_use_latest_vec(dev);
/* Make sure all is clean before doing PF reset */
i40e_clear_hw(hw);
@@ -12599,4 +12661,5 @@ RTE_PMD_REGISTER_PARAM_STRING(net_i40e,
ETH_I40E_FLOATING_VEB_ARG "=1"
ETH_I40E_FLOATING_VEB_LIST_ARG "=<string>"
ETH_I40E_QUEUE_NUM_PER_VF_ARG "=1|2|4|8|16"
- ETH_I40E_SUPPORT_MULTI_DRIVER "=1");
+ ETH_I40E_SUPPORT_MULTI_DRIVER "=1"
+ ETH_I40E_USE_LATEST_VEC "=0|1");
diff --git a/drivers/net/i40e/i40e_ethdev.h b/drivers/net/i40e/i40e_ethdev.h
index 3fffe5a55..7d197f6aa 100644
--- a/drivers/net/i40e/i40e_ethdev.h
+++ b/drivers/net/i40e/i40e_ethdev.h
@@ -1078,6 +1078,9 @@ struct i40e_adapter {
uint64_t pctypes_tbl[I40E_FLOW_TYPE_MAX] __rte_cache_min_aligned;
uint64_t flow_types_mask;
uint64_t pctypes_mask;
+
+ /* For devargs */
+ uint8_t use_latest_vec;
};
/**
diff --git a/drivers/net/i40e/i40e_rxtx.c b/drivers/net/i40e/i40e_rxtx.c
index 5e78d489e..7c986d535 100644
--- a/drivers/net/i40e/i40e_rxtx.c
+++ b/drivers/net/i40e/i40e_rxtx.c
@@ -2909,6 +2909,35 @@ i40e_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
qinfo->conf.offloads = txq->offloads;
}
+static eth_rx_burst_t
+i40e_get_latest_rx_vec(bool scatter)
+{
+#ifdef RTE_ARCH_X86
+ if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
+ return scatter ? i40e_recv_scattered_pkts_vec_avx2 :
+ i40e_recv_pkts_vec_avx2;
+#endif
+ return scatter ? i40e_recv_scattered_pkts_vec :
+ i40e_recv_pkts_vec;
+}
+
+static eth_rx_burst_t
+i40e_get_recommend_rx_vec(bool scatter)
+{
+#ifdef RTE_ARCH_X86
+ /*
+ * since AVX frequency can be different to base frequency, limit
+ * use of AVX2 version to later plaforms, not all those that could
+ * theoretically run it.
+ */
+ if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
+ return scatter ? i40e_recv_scattered_pkts_vec_avx2 :
+ i40e_recv_pkts_vec_avx2;
+#endif
+ return scatter ? i40e_recv_scattered_pkts_vec :
+ i40e_recv_pkts_vec;
+}
+
void __attribute__((cold))
i40e_set_rx_function(struct rte_eth_dev *dev)
{
@@ -2940,57 +2969,17 @@ i40e_set_rx_function(struct rte_eth_dev *dev)
}
}
- if (dev->data->scattered_rx) {
- /* Set the non-LRO scattered callback: there are Vector and
- * single allocation versions.
- */
- if (ad->rx_vec_allowed) {
- PMD_INIT_LOG(DEBUG, "Using Vector Scattered Rx "
- "callback (port=%d).",
- dev->data->port_id);
-
- dev->rx_pkt_burst = i40e_recv_scattered_pkts_vec;
-#ifdef RTE_ARCH_X86
- /*
- * since AVX frequency can be different to base
- * frequency, limit use of AVX2 version to later
- * plaforms, not all those that could theoretically
- * run it.
- */
- if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
- dev->rx_pkt_burst =
- i40e_recv_scattered_pkts_vec_avx2;
-#endif
- } else {
- PMD_INIT_LOG(DEBUG, "Using a Scattered with bulk "
- "allocation callback (port=%d).",
- dev->data->port_id);
- dev->rx_pkt_burst = i40e_recv_scattered_pkts;
- }
- /* If parameters allow we are going to choose between the following
- * callbacks:
- * - Vector
- * - Bulk Allocation
- * - Single buffer allocation (the simplest one)
- */
- } else if (ad->rx_vec_allowed) {
- PMD_INIT_LOG(DEBUG, "Vector rx enabled, please make sure RX "
- "burst size no less than %d (port=%d).",
- RTE_I40E_DESCS_PER_LOOP,
- dev->data->port_id);
-
- dev->rx_pkt_burst = i40e_recv_pkts_vec;
-#ifdef RTE_ARCH_X86
- /*
- * since AVX frequency can be different to base
- * frequency, limit use of AVX2 version to later
- * plaforms, not all those that could theoretically
- * run it.
- */
- if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
- dev->rx_pkt_burst = i40e_recv_pkts_vec_avx2;
-#endif
- } else if (ad->rx_bulk_alloc_allowed) {
+ if (ad->rx_vec_allowed) {
+ /* Vec Rx path */
+ PMD_INIT_LOG(DEBUG, "Vector Rx path will be used on port=%d.",
+ dev->data->port_id);
+ if (ad->use_latest_vec)
+ dev->rx_pkt_burst =
+ i40e_get_latest_rx_vec(dev->data->scattered_rx);
+ else
+ dev->rx_pkt_burst =
+ i40e_get_recommend_rx_vec(dev->data->scattered_rx);
+ } else if (!dev->data->scattered_rx && ad->rx_bulk_alloc_allowed) {
PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
"satisfied. Rx Burst Bulk Alloc function "
"will be used on port=%d.",
@@ -2998,12 +2987,12 @@ i40e_set_rx_function(struct rte_eth_dev *dev)
dev->rx_pkt_burst = i40e_recv_pkts_bulk_alloc;
} else {
- PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are not "
- "satisfied, or Scattered Rx is requested "
- "(port=%d).",
+ /* Simple Rx Path. */
+ PMD_INIT_LOG(DEBUG, "Simple Rx path will be used on port=%d.",
dev->data->port_id);
-
- dev->rx_pkt_burst = i40e_recv_pkts;
+ dev->rx_pkt_burst = dev->data->scattered_rx ?
+ i40e_recv_scattered_pkts :
+ i40e_recv_pkts;
}
/* Propagate information about RX function choice through all queues. */
@@ -3049,6 +3038,31 @@ i40e_set_tx_function_flag(struct rte_eth_dev *dev, struct i40e_tx_queue *txq)
txq->queue_id);
}
+static eth_tx_burst_t
+i40e_get_latest_tx_vec(void)
+{
+#ifdef RTE_ARCH_X86
+ if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
+ return i40e_xmit_pkts_vec_avx2;
+#endif
+ return i40e_xmit_pkts_vec;
+}
+
+static eth_tx_burst_t
+i40e_get_recommend_tx_vec(void)
+{
+#ifdef RTE_ARCH_X86
+ /*
+ * since AVX frequency can be different to base frequency, limit
+ * use of AVX2 version to later plaforms, not all those that could
+ * theoretically run it.
+ */
+ if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
+ return i40e_xmit_pkts_vec_avx2;
+#endif
+ return i40e_xmit_pkts_vec;
+}
+
void __attribute__((cold))
i40e_set_tx_function(struct rte_eth_dev *dev)
{
@@ -3073,17 +3087,12 @@ i40e_set_tx_function(struct rte_eth_dev *dev)
if (ad->tx_simple_allowed) {
if (ad->tx_vec_allowed) {
PMD_INIT_LOG(DEBUG, "Vector tx finally be used.");
- dev->tx_pkt_burst = i40e_xmit_pkts_vec;
-#ifdef RTE_ARCH_X86
- /*
- * since AVX frequency can be different to base
- * frequency, limit use of AVX2 version to later
- * plaforms, not all those that could theoretically
- * run it.
- */
- if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
- dev->tx_pkt_burst = i40e_xmit_pkts_vec_avx2;
-#endif
+ if (ad->use_latest_vec)
+ dev->tx_pkt_burst =
+ i40e_get_latest_tx_vec();
+ else
+ dev->tx_pkt_burst =
+ i40e_get_recommend_tx_vec();
} else {
PMD_INIT_LOG(DEBUG, "Simple tx finally be used.");
dev->tx_pkt_burst = i40e_xmit_pkts_simple;
--
2.17.1
next prev parent reply other threads:[~2018-09-18 2:31 UTC|newest]
Thread overview: 19+ messages / expand[flat|nested] mbox.gz Atom feed top
2018-08-30 2:16 [dpdk-dev] [PATCH] net/i40e: add interface to choose latest vector path Xiaoyun Li
2018-08-31 11:24 ` [dpdk-dev] [PATCH v2] " Xiaoyun Li
2018-09-04 11:39 ` [dpdk-dev] [PATCH v3] " Xiaoyun Li
2018-09-05 12:21 ` Zhang, Qi Z
2018-09-07 9:23 ` Li, Xiaoyun
2018-09-10 10:17 ` [dpdk-dev] [PATCH v4] " Xiaoyun Li
2018-09-12 7:45 ` Zhang, Qi Z
2018-09-12 7:50 ` Zhang, Qi Z
2018-09-12 10:12 ` [dpdk-dev] [PATCH v5] " Xiaoyun Li
2018-09-13 1:38 ` Zhang, Qi Z
2018-09-13 13:27 ` Ferruh Yigit
2018-09-17 7:12 ` Li, Xiaoyun
2018-09-17 9:58 ` [dpdk-dev] [PATCH v6] net/i40e: add interface to use latest vec path Xiaoyun Li
2018-09-17 14:14 ` Ferruh Yigit
2018-09-17 14:37 ` Thomas Monjalon
2018-09-18 1:28 ` Li, Xiaoyun
2018-09-18 2:22 ` Xiaoyun Li [this message]
2018-09-18 13:01 ` [dpdk-dev] [PATCH v7] " Ferruh Yigit
2018-09-18 13:52 ` Zhang, Qi Z
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20180918022240.49376-1-xiaoyun.li@intel.com \
--to=xiaoyun.li@intel.com \
--cc=beilei.xing@intel.com \
--cc=bruce.richardson@intel.com \
--cc=dev@dpdk.org \
--cc=ferruh.yigit@intel.com \
--cc=qi.z.zhang@intel.com \
--cc=thomas@monjalon.net \
--cc=zhiyong.yang@intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).