* [PATCH v2 3/7] net/ena: separate doorbell logic for Rx and Tx
2025-05-22 13:06 [PATCH v2 0/7] net/ena: release 2.13.0 Shai Brandes
2025-05-22 13:06 ` [PATCH v2 1/7] net/ena/base: avoid recalculating desc per entry Shai Brandes
2025-05-22 13:06 ` [PATCH v2 2/7] net/ena/base: coding style changes Shai Brandes
@ 2025-05-22 13:06 ` Shai Brandes
2025-05-22 13:06 ` [PATCH v2 4/7] net/ena: support fragment bypass mode Shai Brandes
3 siblings, 0 replies; 5+ messages in thread
From: Shai Brandes @ 2025-05-22 13:06 UTC (permalink / raw)
To: stephen; +Cc: dev, Shai Brandes
The function ena_com_write_sq_doorbell() currently
checks for LLQ mode using is_llq_max_tx_burst_exists()
which is relevant only for TX queues.
Since RX queues do not operate in LLQ mode, this check
is unnecessary for the RX path.
This patch separates the doorbell write logic into two
distinct handlers for RX and TX, eliminating the
irrelevant LLQ check in the RX path.
Signed-off-by: Shai Brandes <shaibran@amazon.com>
Reviewed-by: Amit Bernstein <amitbern@amazon.com>
Reviewed-by: Yosef Raisman <yraisman@amazon.com>
---
drivers/net/ena/base/ena_eth_com.h | 15 ++++++++++++++-
drivers/net/ena/ena_ethdev.c | 6 +++---
2 files changed, 17 insertions(+), 4 deletions(-)
diff --git a/drivers/net/ena/base/ena_eth_com.h b/drivers/net/ena/base/ena_eth_com.h
index 8a12ed5fba..9e0a7af325 100644
--- a/drivers/net/ena/base/ena_eth_com.h
+++ b/drivers/net/ena/base/ena_eth_com.h
@@ -159,7 +159,20 @@ static inline bool ena_com_is_doorbell_needed(struct ena_com_io_sq *io_sq,
return num_entries_needed > io_sq->entries_in_tx_burst_left;
}
-static inline int ena_com_write_sq_doorbell(struct ena_com_io_sq *io_sq)
+static inline int ena_com_write_rx_sq_doorbell(struct ena_com_io_sq *io_sq)
+{
+ u16 tail = io_sq->tail;
+
+ ena_trc_dbg(ena_com_io_sq_to_ena_dev(io_sq),
+ "Write submission queue doorbell for queue: %d tail: %d\n",
+ io_sq->qid, tail);
+
+ ENA_REG_WRITE32(io_sq->bus, tail, io_sq->db_addr);
+
+ return 0;
+}
+
+static inline int ena_com_write_tx_sq_doorbell(struct ena_com_io_sq *io_sq)
{
u16 max_entries_in_tx_burst = io_sq->llq_info.max_entries_in_tx_burst;
u16 tail = io_sq->tail;
diff --git a/drivers/net/ena/ena_ethdev.c b/drivers/net/ena/ena_ethdev.c
index a13506890f..d8ff6851d2 100644
--- a/drivers/net/ena/ena_ethdev.c
+++ b/drivers/net/ena/ena_ethdev.c
@@ -1835,7 +1835,7 @@ static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
/* When we submitted free resources to device... */
if (likely(i > 0)) {
/* ...let HW know that it can fill buffers with data. */
- ena_com_write_sq_doorbell(rxq->ena_com_io_sq);
+ ena_com_write_rx_sq_doorbell(rxq->ena_com_io_sq);
rxq->next_to_use = next_to_use;
}
@@ -3163,7 +3163,7 @@ static int ena_xmit_mbuf(struct ena_ring *tx_ring, struct rte_mbuf *mbuf)
PMD_TX_LOG_LINE(DEBUG,
"LLQ Tx max burst size of queue %d achieved, writing doorbell to send burst",
tx_ring->id);
- ena_com_write_sq_doorbell(tx_ring->ena_com_io_sq);
+ ena_com_write_tx_sq_doorbell(tx_ring->ena_com_io_sq);
tx_ring->tx_stats.doorbells++;
tx_ring->pkts_without_db = false;
}
@@ -3296,7 +3296,7 @@ static uint16_t eth_ena_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
/* If there are ready packets to be xmitted... */
if (likely(tx_ring->pkts_without_db)) {
/* ...let HW do its best :-) */
- ena_com_write_sq_doorbell(tx_ring->ena_com_io_sq);
+ ena_com_write_tx_sq_doorbell(tx_ring->ena_com_io_sq);
tx_ring->tx_stats.doorbells++;
tx_ring->pkts_without_db = false;
}
--
2.17.1
^ permalink raw reply [flat|nested] 5+ messages in thread* [PATCH v2 4/7] net/ena: support fragment bypass mode
2025-05-22 13:06 [PATCH v2 0/7] net/ena: release 2.13.0 Shai Brandes
` (2 preceding siblings ...)
2025-05-22 13:06 ` [PATCH v2 3/7] net/ena: separate doorbell logic for Rx and Tx Shai Brandes
@ 2025-05-22 13:06 ` Shai Brandes
3 siblings, 0 replies; 5+ messages in thread
From: Shai Brandes @ 2025-05-22 13:06 UTC (permalink / raw)
To: stephen; +Cc: dev, Shai Brandes, Yosef Raisman
Introduce devarg `enable_frag_bypass` to toggle the fragment bypass
mode for egress packets.
When enabled, this feature bypasses the PPS limit enforced by EC2 for
fragmented egress packets on every ENI. Note that enabling this might
negatively impact network performance.
By default, this feature is disabled. To enable it set
`enable_frag_bypass=1`. If it cannot be enabled, a warning will be
printed, but driver initialization will proceed as normal.
Signed-off-by: Yosef Raisman <yraisman@amazon.com>
Signed-off-by: Shai Brandes <shaibran@amazon.com>
Reviewed-by: Amit Bernstein <amitbern@amazon.com>
---
doc/guides/nics/ena.rst | 9 ++++
doc/guides/rel_notes/release_25_07.rst | 5 ++
drivers/net/ena/base/ena_com.c | 33 +++++++++++++
drivers/net/ena/base/ena_com.h | 8 ++++
.../net/ena/base/ena_defs/ena_admin_defs.h | 15 ++++++
drivers/net/ena/ena_ethdev.c | 46 +++++++++++++++++++
drivers/net/ena/ena_ethdev.h | 2 +
7 files changed, 118 insertions(+)
diff --git a/doc/guides/nics/ena.rst b/doc/guides/nics/ena.rst
index a34575dc9b..a42deccd81 100644
--- a/doc/guides/nics/ena.rst
+++ b/doc/guides/nics/ena.rst
@@ -141,6 +141,15 @@ Runtime Configuration
**A non-zero value for this devarg is mandatory for control path functionality
when binding ports to uio_pci_generic kernel module which lacks interrupt support.**
+ * **enable_frag_bypass** (default 0)
+
+ Enable fragment bypass mode for egress packets. This mode bypasses the PPS
+ limit enforced by EC2 for fragmented egress packets on every ENI. Note that
+ enabling it might negatively impact network performance.
+
+ 0 - Disabled (Default).
+
+ 1 - Enabled.
ENA Configuration Parameters
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
diff --git a/doc/guides/rel_notes/release_25_07.rst b/doc/guides/rel_notes/release_25_07.rst
index 467536fe05..07b5feabc4 100644
--- a/doc/guides/rel_notes/release_25_07.rst
+++ b/doc/guides/rel_notes/release_25_07.rst
@@ -55,6 +55,11 @@ New Features
Also, make sure to start the actual text at the margin.
=======================================================
+* **Updated Amazon ENA (Elastic Network Adapter) net driver.**
+
+ * Added support for enabling fragment bypass mode for egress packets.
+ This mode bypasses the PPS limit enforced by EC2 for fragmented egress packets on every ENI.
+
* **Added Mucse rnp net driver.**
Added a new network PMD which supports Mucse 10 Gigabit Ethernet NICs.
diff --git a/drivers/net/ena/base/ena_com.c b/drivers/net/ena/base/ena_com.c
index 588dc61387..9715a627c1 100644
--- a/drivers/net/ena/base/ena_com.c
+++ b/drivers/net/ena/base/ena_com.c
@@ -3459,3 +3459,36 @@ int ena_com_config_dev_mode(struct ena_com_dev *ena_dev,
return 0;
}
+int ena_com_set_frag_bypass(struct ena_com_dev *ena_dev, bool enable)
+{
+ struct ena_admin_set_feat_resp set_feat_resp;
+ struct ena_com_admin_queue *admin_queue;
+ struct ena_admin_set_feat_cmd cmd;
+ int ret;
+
+ if (!ena_com_check_supported_feature_id(ena_dev, ENA_ADMIN_FRAG_BYPASS)) {
+ ena_trc_dbg(ena_dev, "Feature %d isn't supported\n",
+ ENA_ADMIN_FRAG_BYPASS);
+ return ENA_COM_UNSUPPORTED;
+ }
+
+ memset(&cmd, 0x0, sizeof(cmd));
+ admin_queue = &ena_dev->admin_queue;
+
+ cmd.aq_common_descriptor.opcode = ENA_ADMIN_SET_FEATURE;
+ cmd.aq_common_descriptor.flags = 0;
+ cmd.feat_common.feature_id = ENA_ADMIN_FRAG_BYPASS;
+ cmd.feat_common.feature_version = ENA_ADMIN_FRAG_BYPASS_FEATURE_VERSION_0;
+ cmd.u.frag_bypass.enable = (u8)enable;
+
+ ret = ena_com_execute_admin_command(admin_queue,
+ (struct ena_admin_aq_entry *)&cmd,
+ sizeof(cmd),
+ (struct ena_admin_acq_entry *)&set_feat_resp,
+ sizeof(set_feat_resp));
+
+ if (unlikely(ret))
+ ena_trc_err(ena_dev, "Failed to enable frag bypass. error: %d\n", ret);
+
+ return ret;
+}
diff --git a/drivers/net/ena/base/ena_com.h b/drivers/net/ena/base/ena_com.h
index b2aede1be1..f064095fd1 100644
--- a/drivers/net/ena/base/ena_com.h
+++ b/drivers/net/ena/base/ena_com.h
@@ -1109,6 +1109,14 @@ static inline bool ena_com_get_missing_admin_interrupt(struct ena_com_dev *ena_d
return ena_dev->admin_queue.is_missing_admin_interrupt;
}
+/* ena_com_set_frag_bypass - set fragment bypass
+ * @ena_dev: ENA communication layer struct
+ * @enable: true if fragment bypass is enabled, false otherwise.
+ *
+ * @return - 0 on success, negative value on failure.
+ */
+int ena_com_set_frag_bypass(struct ena_com_dev *ena_dev, bool enable);
+
/* ena_com_io_sq_to_ena_dev - Extract ena_com_dev using contained field io_sq.
* @io_sq: IO submit queue struct
*
diff --git a/drivers/net/ena/base/ena_defs/ena_admin_defs.h b/drivers/net/ena/base/ena_defs/ena_admin_defs.h
index bdc6efadcf..d315014776 100644
--- a/drivers/net/ena/base/ena_defs/ena_admin_defs.h
+++ b/drivers/net/ena/base/ena_defs/ena_admin_defs.h
@@ -57,6 +57,7 @@ enum ena_admin_aq_feature_id {
ENA_ADMIN_EXTRA_PROPERTIES_STRINGS = 5,
ENA_ADMIN_EXTRA_PROPERTIES_FLAGS = 6,
ENA_ADMIN_MAX_QUEUES_EXT = 7,
+ ENA_ADMIN_FRAG_BYPASS = 8,
ENA_ADMIN_RSS_HASH_FUNCTION = 10,
ENA_ADMIN_STATELESS_OFFLOAD_CONFIG = 11,
ENA_ADMIN_RSS_INDIRECTION_TABLE_CONFIG = 12,
@@ -165,6 +166,11 @@ enum ena_admin_ena_srd_flags {
ENA_ADMIN_ENA_SRD_UDP_ORDERING_BYPASS_ENABLED = BIT(2),
};
+enum ena_admin_frag_bypass_feature_version {
+ /* Enable only */
+ ENA_ADMIN_FRAG_BYPASS_FEATURE_VERSION_0 = 0,
+};
+
struct ena_admin_aq_common_desc {
/* 11:0 : command_id
* 15:12 : reserved12
@@ -706,6 +712,13 @@ struct ena_admin_feature_llq_desc {
struct ena_admin_accel_mode_req accel_mode;
};
+struct ena_admin_feature_frag_bypass_desc {
+ /* Enable frag_bypass */
+ uint8_t enable;
+
+ uint8_t reserved[3];
+};
+
struct ena_admin_queue_ext_feature_fields {
uint32_t max_tx_sq_num;
@@ -1180,6 +1193,8 @@ struct ena_admin_set_feat_cmd {
/* PHC configuration */
struct ena_admin_feature_phc_desc phc;
+ /* Fragment bypass configuration */
+ struct ena_admin_feature_frag_bypass_desc frag_bypass;
} u;
};
diff --git a/drivers/net/ena/ena_ethdev.c b/drivers/net/ena/ena_ethdev.c
index d8ff6851d2..4caad9006e 100644
--- a/drivers/net/ena/ena_ethdev.c
+++ b/drivers/net/ena/ena_ethdev.c
@@ -106,6 +106,12 @@ struct ena_stats {
* driver.
*/
#define ENA_DEVARG_CONTROL_PATH_POLL_INTERVAL "control_path_poll_interval"
+/*
+ * Toggles fragment bypass mode. Fragmented egress packets are rate limited by
+ * EC2 per ENI; this devarg bypasses the PPS limit but may impact performance.
+ * Disabled by default.
+ */
+#define ENA_DEVARG_ENABLE_FRAG_BYPASS "enable_frag_bypass"
/*
* Each rte_memzone should have unique name.
@@ -306,6 +312,9 @@ static int ena_xstats_get_by_id(struct rte_eth_dev *dev,
static int ena_process_llq_policy_devarg(const char *key,
const char *value,
void *opaque);
+static int ena_process_bool_devarg(const char *key,
+ const char *value,
+ void *opaque);
static int ena_parse_devargs(struct ena_adapter *adapter,
struct rte_devargs *devargs);
static void ena_copy_customer_metrics(struct ena_adapter *adapter,
@@ -1336,6 +1345,12 @@ static int ena_start(struct rte_eth_dev *dev)
if (rc)
goto err_start_tx;
+ if (adapter->enable_frag_bypass) {
+ rc = ena_com_set_frag_bypass(&adapter->ena_dev, true);
+ if (rc)
+ PMD_DRV_LOG_LINE(WARNING, "Failed to enable frag bypass, rc: %d", rc);
+ }
+
if (adapter->edev_data->dev_conf.rxmode.mq_mode & RTE_ETH_MQ_RX_RSS_FLAG) {
rc = ena_rss_configure(adapter);
if (rc)
@@ -3764,6 +3779,31 @@ static int ena_process_llq_policy_devarg(const char *key, const char *value, voi
PMD_INIT_LOG_LINE(INFO,
"LLQ policy is %u [0 - disabled, 1 - device recommended, 2 - normal, 3 - large]",
adapter->llq_header_policy);
+
+ return 0;
+}
+
+static int ena_process_bool_devarg(const char *key, const char *value, void *opaque)
+{
+ struct ena_adapter *adapter = opaque;
+ bool bool_value;
+
+ /* Parse the value. */
+ if (strcmp(value, "1") == 0) {
+ bool_value = true;
+ } else if (strcmp(value, "0") == 0) {
+ bool_value = false;
+ } else {
+ PMD_INIT_LOG_LINE(ERR,
+ "Invalid value: '%s' for key '%s'. Accepted: '0' or '1'",
+ value, key);
+ return -EINVAL;
+ }
+
+ /* Now, assign it to the proper adapter field. */
+ if (strcmp(key, ENA_DEVARG_ENABLE_FRAG_BYPASS) == 0)
+ adapter->enable_frag_bypass = bool_value;
+
return 0;
}
@@ -3773,6 +3813,7 @@ static int ena_parse_devargs(struct ena_adapter *adapter, struct rte_devargs *de
ENA_DEVARG_LLQ_POLICY,
ENA_DEVARG_MISS_TXC_TO,
ENA_DEVARG_CONTROL_PATH_POLL_INTERVAL,
+ ENA_DEVARG_ENABLE_FRAG_BYPASS,
NULL,
};
struct rte_kvargs *kvlist;
@@ -3799,6 +3840,10 @@ static int ena_parse_devargs(struct ena_adapter *adapter, struct rte_devargs *de
ena_process_uint_devarg, adapter);
if (rc != 0)
goto exit;
+ rc = rte_kvargs_process(kvlist, ENA_DEVARG_ENABLE_FRAG_BYPASS,
+ ena_process_bool_devarg, adapter);
+ if (rc != 0)
+ goto exit;
exit:
rte_kvargs_free(kvlist);
@@ -4021,6 +4066,7 @@ RTE_PMD_REGISTER_PARAM_STRING(net_ena,
ENA_DEVARG_LLQ_POLICY "=<0|1|2|3> "
ENA_DEVARG_MISS_TXC_TO "=<uint>"
ENA_DEVARG_CONTROL_PATH_POLL_INTERVAL "=<0-1000>");
+ ENA_DEVARG_ENABLE_FRAG_BYPASS "=<0|1> ");
RTE_LOG_REGISTER_SUFFIX(ena_logtype_init, init, NOTICE);
RTE_LOG_REGISTER_SUFFIX(ena_logtype_driver, driver, NOTICE);
#ifdef RTE_ETHDEV_DEBUG_RX
diff --git a/drivers/net/ena/ena_ethdev.h b/drivers/net/ena/ena_ethdev.h
index 335028ad19..f4461733e9 100644
--- a/drivers/net/ena/ena_ethdev.h
+++ b/drivers/net/ena/ena_ethdev.h
@@ -348,6 +348,8 @@ struct ena_adapter {
/* Time (in microseconds) of the control path queues monitoring interval */
uint64_t control_path_poll_interval;
+ bool enable_frag_bypass;
+
/*
* Helper variables for holding the information about the supported
* metrics.
--
2.17.1
^ permalink raw reply [flat|nested] 5+ messages in thread