aboutsummaryrefslogtreecommitdiffstats
path: root/doc/guides/prog_guide/poll_mode_drv.rst
diff options
context:
space:
mode:
Diffstat (limited to 'doc/guides/prog_guide/poll_mode_drv.rst')
-rw-r--r--doc/guides/prog_guide/poll_mode_drv.rst15
1 files changed, 13 insertions, 2 deletions
diff --git a/doc/guides/prog_guide/poll_mode_drv.rst b/doc/guides/prog_guide/poll_mode_drv.rst
index 4987f70a..1ac8f7eb 100644
--- a/doc/guides/prog_guide/poll_mode_drv.rst
+++ b/doc/guides/prog_guide/poll_mode_drv.rst
@@ -84,7 +84,7 @@ Whenever needed and appropriate, asynchronous communication should be introduced
Avoiding lock contention is a key issue in a multi-core environment.
To address this issue, PMDs are designed to work with per-core private resources as much as possible.
-For example, a PMD maintains a separate transmit queue per-core, per-port.
+For example, a PMD maintains a separate transmit queue per-core, per-port, if the PMD is not ``DEV_TX_OFFLOAD_MT_LOCKFREE`` capable.
In the same way, every receive queue of a port is assigned to and polled by a single logical core (lcore).
To comply with Non-Uniform Memory Access (NUMA), memory management is designed to assign to each logical core
@@ -146,6 +146,16 @@ This is also true for the pipe-line model provided all logical cores used are lo
Multiple logical cores should never share receive or transmit queues for interfaces since this would require global locks and hinder performance.
+If the PMD is ``DEV_TX_OFFLOAD_MT_LOCKFREE`` capable, multiple threads can invoke ``rte_eth_tx_burst()``
+concurrently on the same tx queue without SW lock. This PMD feature found in some NICs and useful in the following use cases:
+
+* Remove explicit spinlock in some applications where lcores are not mapped to Tx queues with 1:1 relation.
+
+* In the eventdev use case, avoid dedicating a separate TX core for transmitting and thus
+ enables more scaling as all workers can send the packets.
+
+See `Hardware Offload`_ for ``DEV_TX_OFFLOAD_MT_LOCKFREE`` capability probing details.
+
Device Identification and Configuration
---------------------------------------
@@ -290,7 +300,8 @@ Hardware Offload
Depending on driver capabilities advertised by
``rte_eth_dev_info_get()``, the PMD may support hardware offloading
-feature like checksumming, TCP segmentation or VLAN insertion.
+feature like checksumming, TCP segmentation, VLAN insertion or
+lockfree multithreaded TX burst on the same TX queue.
The support of these offload features implies the addition of dedicated
status bit(s) and value field(s) into the rte_mbuf data structure, along