interface: add the transmit queue infrastructure document

Type: docs

Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>
Change-Id: I57f27f8ec4be7f3d8dc3d13ff4ea6b1b21c3cf6b

docs/developer/corefeatures/index.rst

@@ -17,6 +17,7 @@ Core Features
     span_doc
     mtu
     gso
+    tx_queue
     sylog_doc
     eventviewer
     stats

docs/developer/corefeatures/tx_queue.rst

@@ -0,0 +1 @@
+../../../src/vnet/interface/tx_queue.rst

docs/spelling_wordlist.txt

@@ -1147,6 +1147,8 @@ tuntap
 tw
 twovppinstances
 tx
+txq
+TXQ
 txt
 typedef
 ubuntu
@@ -1178,6 +1180,7 @@ unmounting
 Unmounting
 unmounts
 unreachables
+unregister
 unrouteable
 unserialization
 unserialize

src/vnet/interface/tx_queue.rst

@@ -0,0 +1,159 @@
+.. _TX_Queue_doc:
+
+Transmit Queues
+===============
+
+Overview
+________
+
+VPP implements Transmit queues infra to access and manage them. It provides
+common registration functions to register or unregister interfaces’ transmit
+queues. It also provides functions for queues placement on given thread(s).
+
+The TXQ Infrastructure
+_______________________
+
+Infra registers each queue using a unique key which is formed by concatenating
+the hardware interface index ``hw_if_index`` and unique queue identifier for
+given interface ``queue_id``. As a result of registration of queue, infra
+returns back a unique global ``queue_index`` which can be used by driver to
+access that queue later.
+
+Interface output node uses pre-computed ``output_node_thread_runtime`` data
+which provides essential information related to queue placements on given
+thread of given interface. Transmit queue infra implements an algorithm to
+pre-compute this information. It also pre-computes scalar arguments of frame
+``vnet_hw_if_tx_frame_t``. It also pre-calculates a ``lookup_table`` for
+thread if there are multiple transmit queues are placed on that thread.
+Interface drivers call ``vnet_hw_if_update_runtime_data()`` to execute that
+algorithm after registering the transmit queues to TXQ infra.
+
+The algorithm makes the copy of existing runtime data and iterate through them
+for each vpp main and worker thread. In each iteration, algorithm loop through
+all the tx queues of given interface to fill the information in the frame data
+structure ``vnet_hw_if_tx_frame_t``. Algorithm also updates the information
+related to number of transmit queues of given interface on given vpp thread in
+data structure ``output_node_thread_runtime``. As a consequence of any update
+to the copy, triggers the function to update the actual working copy by taking
+the worker barrier and free the old copy of ``output_node_thread_runtime``.
+
+Multi-TXQ infra
+^^^^^^^^^^^^^^^
+
+Interface output node uses packet flow hash using hash infra in case of multi-txq
+on given thread. Each hardware interface class contains type of the hash required
+for interfaces from that hardware interface class i.e. ethernet interface hardware
+class contains type ``VNET_HASH_FN_TYPE_ETHERNET``. Though, the hash function
+itself is contained by hardware interface data structure of given interface. Default
+hashing function is selected upon interface creation based on priority. User can
+configure a different hash to an interface for multi-txq use case.
+
+Interface output node uses packet flow hash as an index to the pre-calculated lookup
+table to get the queue identifier for given transmit queue. Interface output node
+enqueues the packets to respective frame and also copies the ``vnet_hw_if_tx_frame_t``
+to frame scalar arguments. Drivers use scalar arguments ``vnet_hw_if_tx_frame_t``
+of the given frame to extract the information about the transmit queue to be used to
+transmit the packets. Drivers may need to acquire a lock on given queue before
+transmitting the packets based on the ``shared_queue`` bit status.
+
+Data structures
+^^^^^^^^^^^^^^^
+
+Queue information is stored in data structure ``vnet_hw_if_tx_queue_t``:
+
+.. code:: c
+
+  typedef struct
+  {
+    /* either this queue is shared among multiple threads */
+    u8 shared_queue : 1;
+    /* hw interface index */
+    u32 hw_if_index;
+
+    /* hardware queue identifier */
+    u32 queue_id;
+
+    /* bitmap of threads which use this queue */
+    clib_bitmap_t *threads;
+  } vnet_hw_if_tx_queue_t;
+
+
+Frame information is stored in data structure: ``vnet_hw_if_tx_frame_t``:
+
+.. code:: c
+
+  typedef enum
+  {
+    VNET_HW_IF_TX_FRAME_HINT_NOT_CHAINED = (1 << 0),
+    VNET_HW_IF_TX_FRAME_HINT_NO_GSO = (1 << 1),
+    VNET_HW_IF_TX_FRAME_HINT_NO_CKSUM_OFFLOAD = (1 << 2),
+  } vnet_hw_if_tx_frame_hint_t;
+
+  typedef struct
+  {
+    u8 shared_queue : 1;
+    vnet_hw_if_tx_frame_hint_t hints : 16;
+    u32 queue_id;
+  } vnet_hw_if_tx_frame_t;
+
+Output node runtime information is stored in data structure: ``output_node_thread_runtime``:
+
+.. code:: c
+
+  typedef struct
+  {
+    CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
+    vnet_hw_if_tx_frame_t *frame;
+    u32 *lookup_table;
+    u32 n_queues;
+  } vnet_hw_if_output_node_runtime_t;
+
+
+MultiTXQ API
+^^^^^^^^^^^^
+
+This API message is used to place tx queue of an interface to vpp main or worker(s) thread(s).
+
+.. code:: c
+
+  autoendian autoreply define sw_interface_set_tx_placement
+  {
+      u32 client_index;
+      u32 context;
+      vl_api_interface_index_t sw_if_index;
+      u32 queue_id;
+      u32 array_size;
+      u32 threads[array_size];
+      option vat_help = "<interface | sw_if_index <index>> queue <n> [threads <list> | mask <hex>]";
+  };
+
+Multi-TXQ CLI
+^^^^^^^^^^^^^
+
+::
+
+  set interface tx-queue                   set interface tx-queue <interface> queue <n> [threads <list>]
+  set interface tx-hash                    set interface tx-hash <interface> hash-name <hash-name>
+
+::
+
+  show hardware-interfaces
+
+         Name                Idx   Link  Hardware
+  tap0                        1     up   tap0
+    Link speed: unknown
+    RX Queues:
+      queue thread         mode
+      0     main (0)       polling
+    TX Queues:
+      TX Hash: [name: crc32c-5tuple  priority: 50 description: IPv4/IPv6 header and TCP/UDP ports]
+      queue shared thread(s)
+      0     no     0
+    Ethernet address 02:fe:27:69:5a:b5
+    VIRTIO interface
+       instance 0
+         RX QUEUE : Total Packets
+                0 : 0
+         TX QUEUE : Total Packets
+                0 : 0
+