gro: cleanup
authorJiayu Hu <jiayu.hu@intel.com>
Wed, 10 Jan 2018 14:03:10 +0000 (22:03 +0800)
committerThomas Monjalon <thomas@monjalon.net>
Thu, 11 Jan 2018 22:18:24 +0000 (23:18 +0100)
This patch updates codes as follows:
- change appropriate names for internal structures, variants and functions
- update comments and the content of the gro programmer guide for better
  understanding
- remove needless check and redundant comments

Signed-off-by: Jiayu Hu <jiayu.hu@intel.com>
Reviewed-by: Junjie Chen <junjie.j.chen@intel.com>
doc/guides/prog_guide/generic_receive_offload_lib.rst
doc/guides/prog_guide/img/gro-key-algorithm.svg [new file with mode: 0644]
lib/librte_gro/gro_tcp4.c
lib/librte_gro/gro_tcp4.h
lib/librte_gro/rte_gro.c
lib/librte_gro/rte_gro.h

index 22e50ec..1652e64 100644 (file)
@@ -32,128 +32,154 @@ Generic Receive Offload Library
 ===============================
 
 Generic Receive Offload (GRO) is a widely used SW-based offloading
-technique to reduce per-packet processing overhead. It gains performance
-by reassembling small packets into large ones. To enable more flexibility
-to applications, DPDK implements GRO as a standalone library. Applications
-explicitly use the GRO library to merge small packets into large ones.
-
-The GRO library assumes all input packets have correct checksums. In
-addition, the GRO library doesn't re-calculate checksums for merged
-packets. If input packets are IP fragmented, the GRO library assumes
-they are complete packets (i.e. with L4 headers).
-
-Currently, the GRO library implements TCP/IPv4 packet reassembly.
-
-Reassembly Modes
-----------------
-
-The GRO library provides two reassembly modes: lightweight and
-heavyweight mode. If applications want to merge packets in a simple way,
-they can use the lightweight mode API. If applications want more
-fine-grained controls, they can choose the heavyweight mode API.
-
-Lightweight Mode
-~~~~~~~~~~~~~~~~
-
-The ``rte_gro_reassemble_burst()`` function is used for reassembly in
-lightweight mode. It tries to merge N input packets at a time, where
-N should be less than or equal to ``RTE_GRO_MAX_BURST_ITEM_NUM``.
-
-In each invocation, ``rte_gro_reassemble_burst()`` allocates temporary
-reassembly tables for the desired GRO types. Note that the reassembly
-table is a table structure used to reassemble packets and different GRO
-types (e.g. TCP/IPv4 GRO and TCP/IPv6 GRO) have different reassembly table
-structures. The ``rte_gro_reassemble_burst()`` function uses the reassembly
-tables to merge the N input packets.
-
-For applications, performing GRO in lightweight mode is simple. They
-just need to invoke ``rte_gro_reassemble_burst()``. Applications can get
-GROed packets as soon as ``rte_gro_reassemble_burst()`` returns.
-
-Heavyweight Mode
-~~~~~~~~~~~~~~~~
-
-The ``rte_gro_reassemble()`` function is used for reassembly in heavyweight
-mode. Compared with the lightweight mode, performing GRO in heavyweight mode
-is relatively complicated.
-
-Before performing GRO, applications need to create a GRO context object
-by calling ``rte_gro_ctx_create()``. A GRO context object holds the
-reassembly tables of desired GRO types. Note that all update/lookup
-operations on the context object are not thread safe. So if different
-processes or threads want to access the same context object simultaneously,
-some external syncing mechanisms must be used.
-
-Once the GRO context is created, applications can then use the
-``rte_gro_reassemble()`` function to merge packets. In each invocation,
-``rte_gro_reassemble()`` tries to merge input packets with the packets
-in the reassembly tables. If an input packet is an unsupported GRO type,
-or other errors happen (e.g. SYN bit is set), ``rte_gro_reassemble()``
-returns the packet to applications. Otherwise, the input packet is either
-merged or inserted into a reassembly table.
-
-When applications want to get GRO processed packets, they need to use
-``rte_gro_timeout_flush()`` to flush them from the tables manually.
+technique to reduce per-packet processing overheads. By reassembling
+small packets into larger ones, GRO enables applications to process
+fewer large packets directly, thus reducing the number of packets to
+be processed. To benefit DPDK-based applications, like Open vSwitch,
+DPDK also provides own GRO implementation. In DPDK, GRO is implemented
+as a standalone library. Applications explicitly use the GRO library to
+reassemble packets.
+
+Overview
+--------
+
+In the GRO library, there are many GRO types which are defined by packet
+types. One GRO type is in charge of process one kind of packets. For
+example, TCP/IPv4 GRO processes TCP/IPv4 packets.
+
+Each GRO type has a reassembly function, which defines own algorithm and
+table structure to reassemble packets. We assign input packets to the
+corresponding GRO functions by MBUF->packet_type.
+
+The GRO library doesn't check if input packets have correct checksums and
+doesn't re-calculate checksums for merged packets. The GRO library
+assumes the packets are complete (i.e., MF==0 && frag_off==0), when IP
+fragmentation is possible (i.e., DF==0). Additionally, it requires IPv4
+ID to be increased by one.
 
-TCP/IPv4 GRO
-------------
+Currently, the GRO library provides GRO supports for TCP/IPv4 packets.
+
+Two Sets of API
+---------------
+
+For different usage scenarios, the GRO library provides two sets of API.
+The one is called the lightweight mode API, which enables applications to
+merge a small number of packets rapidly; the other is called the
+heavyweight mode API, which provides fine-grained controls to
+applications and supports to merge a large number of packets.
+
+Lightweight Mode API
+~~~~~~~~~~~~~~~~~~~~
+
+The lightweight mode only has one function ``rte_gro_reassemble_burst()``,
+which process N packets at a time. Using the lightweight mode API to
+merge packets is very simple. Calling ``rte_gro_reassemble_burst()`` is
+enough. The GROed packets are returned to applications as soon as it
+finishes.
+
+In ``rte_gro_reassemble_burst()``, table structures of different GRO
+types are allocated in the stack. This design simplifies applications'
+operations. However, limited by the stack size, the maximum number of
+packets that ``rte_gro_reassemble_burst()`` can process in an invocation
+should be less than or equal to ``RTE_GRO_MAX_BURST_ITEM_NUM``.
+
+Heavyweight Mode API
+~~~~~~~~~~~~~~~~~~~~
+
+Compared with the lightweight mode, using the heavyweight mode API is
+relatively complex. Firstly, applications need to create a GRO context
+by ``rte_gro_ctx_create()``. ``rte_gro_ctx_create()`` allocates tables
+structures in the heap and stores their pointers in the GRO context.
+Secondly, applications use ``rte_gro_reassemble()`` to merge packets.
+If input packets have invalid parameters, ``rte_gro_reassemble()``
+returns them to applications. For example, packets of unsupported GRO
+types or TCP SYN packets are returned. Otherwise, the input packets are
+either merged with the existed packets in the tables or inserted into the
+tables. Finally, applications use ``rte_gro_timeout_flush()`` to flush
+packets from the tables, when they want to get the GROed packets.
+
+Note that all update/lookup operations on the GRO context are not thread
+safe. So if different processes or threads want to access the same
+context object simultaneously, some external syncing mechanisms must be
+used.
+
+Reassembly Algorithm
+--------------------
 
-TCP/IPv4 GRO supports merging small TCP/IPv4 packets into large ones,
-using a table structure called the TCP/IPv4 reassembly table.
+The reassembly algorithm is used for reassembling packets. In the GRO
+library, different GRO types can use different algorithms. In this
+section, we will introduce an algorithm, which is used by TCP/IPv4 GRO.
 
-TCP/IPv4 Reassembly Table
-~~~~~~~~~~~~~~~~~~~~~~~~~
+Challenges
+~~~~~~~~~~
 
-A TCP/IPv4 reassembly table includes a "key" array and an "item" array.
-The key array keeps the criteria to merge packets and the item array
-keeps the packet information.
+The reassembly algorithm determines the efficiency of GRO. There are two
+challenges in the algorithm design:
 
-Each key in the key array points to an item group, which consists of
-packets which have the same criteria values but can't be merged. A key
-in the key array includes two parts:
+- a high cost algorithm/implementation would cause packet dropping in a
+  high speed network.
 
-* ``criteria``: the criteria to merge packets. If two packets can be
-  merged, they must have the same criteria values.
+- packet reordering makes it hard to merge packets. For example, Linux
+  GRO fails to merge packets when encounters packet reordering.
 
-* ``start_index``: the item array index of the first packet in the item
-  group.
+The above two challenges require our algorithm is:
 
-Each element in the item array keeps the information of a packet. An item
-in the item array mainly includes three parts:
+- lightweight enough to scale fast networking speed
 
-* ``firstseg``: the mbuf address of the first segment of the packet.
+- capable of handling packet reordering
 
-* ``lastseg``: the mbuf address of the last segment of the packet.
+In DPDK GRO, we use a key-based algorithm to address the two challenges.
 
-* ``next_pkt_index``: the item array index of the next packet in the same
-  item group. TCP/IPv4 GRO uses ``next_pkt_index`` to chain the packets
-  that have the same criteria value but can't be merged together.
+Key-based Reassembly Algorithm
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+:numref:`figure_gro-key-algorithm` illustrates the procedure of the
+key-based algorithm. Packets are classified into "flows" by some header
+fields (we call them as "key"). To process an input packet, the algorithm
+searches for a matched "flow" (i.e., the same value of key) for the
+packet first, then checks all packets in the "flow" and tries to find a
+"neighbor" for it. If find a "neighbor", merge the two packets together.
+If can't find a "neighbor", store the packet into its "flow". If can't
+find a matched "flow", insert a new "flow" and store the packet into the
+"flow".
+
+.. note::
+        Packets in the same "flow" that can't merge are always caused
+        by packet reordering.
+
+The key-based algorithm has two characters:
+
+- classifying packets into "flows" to accelerate packet aggregation is
+  simple (address challenge 1).
+
+- storing out-of-order packets makes it possible to merge later (address
+  challenge 2).
+
+.. _figure_gro-key-algorithm:
+
+.. figure:: img/gro-key-algorithm.*
+   :align: center
+
+   Key-based Reassembly Algorithm
+
+TCP/IPv4 GRO
+------------
 
-Procedure to Reassemble a Packet
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+The table structure used by TCP/IPv4 GRO contains two arrays: flow array
+and item array. The flow array keeps flow information, and the item array
+keeps packet information.
 
-To reassemble an incoming packet needs three steps:
+Header fields used to define a TCP/IPv4 flow include:
 
-#. Check if the packet should be processed. Packets with one of the
-   following properties aren't processed and are returned immediately:
+- source and destination: Ethernet and IP address, TCP port
 
-   * FIN, SYN, RST, URG, PSH, ECE or CWR bit is set.
+- TCP acknowledge number
 
-   * L4 payload length is 0.
+TCP/IPv4 packets whose FIN, SYN, RST, URG, PSH, ECE or CWR bit is set
+won't be processed.
 
-#.  Traverse the key array to find a key which has the same criteria
-    value with the incoming packet. If found, go to the next step.
-    Otherwise, insert a new key and a new item for the packet.
+Header fields deciding if two packets are neighbors include:
 
-#. Locate the first packet in the item group via ``start_index``. Then
-   traverse all packets in the item group via ``next_pkt_index``. If a
-   packet is found which can be merged with the incoming one, merge them
-   together. If one isn't found, insert the packet into this item group.
-   Note that to merge two packets is to link them together via mbuf's
-   ``next`` field.
+- TCP sequence number
 
-When packets are flushed from the reassembly table, TCP/IPv4 GRO updates
-packet header fields for the merged packets. Note that before reassembling
-the packet, TCP/IPv4 GRO doesn't check if the checksums of packets are
-correct. Also, TCP/IPv4 GRO doesn't re-calculate checksums for merged
-packets.
+- IPv4 ID. The IPv4 ID fields of the packets should be increased by 1.
diff --git a/doc/guides/prog_guide/img/gro-key-algorithm.svg b/doc/guides/prog_guide/img/gro-key-algorithm.svg
new file mode 100644 (file)
index 0000000..94e42f5
--- /dev/null
@@ -0,0 +1,223 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.0//EN" "http://www.w3.org/TR/2001/REC-SVG-20010904/DTD/svg10.dtd">
+<!-- Generated by Microsoft Visio 11.0, SVG Export, v1.0 gro-key-algorithm.svg Page-1 -->
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:ev="http://www.w3.org/2001/xml-events"
+               xmlns:v="http://schemas.microsoft.com/visio/2003/SVGExtensions/" width="6.06163in" height="2.66319in"
+               viewBox="0 0 436.438 191.75" xml:space="preserve" color-interpolation-filters="sRGB" class="st10">
+       <v:documentProperties v:langID="1033" v:viewMarkup="false"/>
+
+       <style type="text/css">
+       <![CDATA[
+               .st1 {fill:url(#grad30-4);stroke:#404040;stroke-linecap:round;stroke-linejoin:round;stroke-width:0.25}
+               .st2 {fill:#000000;font-family:Calibri;font-size:1.00001em}
+               .st3 {font-size:1em;font-weight:bold}
+               .st4 {fill:#000000;font-family:Calibri;font-size:1.00001em;font-weight:bold}
+               .st5 {font-size:1em;font-weight:normal}
+               .st6 {marker-end:url(#mrkr5-38);stroke:#404040;stroke-linecap:round;stroke-linejoin:round;stroke-width:1}
+               .st7 {fill:#404040;fill-opacity:1;stroke:#404040;stroke-opacity:1;stroke-width:0.28409090909091}
+               .st8 {fill:none;stroke:none;stroke-linecap:round;stroke-linejoin:round;stroke-width:0.25}
+               .st9 {fill:#000000;font-family:Calibri;font-size:0.833336em}
+               .st10 {fill:none;fill-rule:evenodd;font-size:12px;overflow:visible;stroke-linecap:square;stroke-miterlimit:3}
+       ]]>
+       </style>
+
+       <defs id="Patterns_And_Gradients">
+               <linearGradient id="grad30-4" v:fillPattern="30" v:foreground="#c6d09f" v:background="#d1dab4" x1="0" y1="1" x2="0" y2="0">
+                       <stop offset="0" style="stop-color:#c6d09f;stop-opacity:1"/>
+                       <stop offset="1" style="stop-color:#d1dab4;stop-opacity:1"/>
+               </linearGradient>
+               <linearGradient id="grad30-35" v:fillPattern="30" v:foreground="#f0f0f0" v:background="#ffffff" x1="0" y1="1" x2="0" y2="0">
+                       <stop offset="0" style="stop-color:#f0f0f0;stop-opacity:1"/>
+                       <stop offset="1" style="stop-color:#ffffff;stop-opacity:1"/>
+               </linearGradient>
+       </defs>
+       <defs id="Markers">
+               <g id="lend5">
+                       <path d="M 2 1 L 0 0 L 1.98117 -0.993387 C 1.67173 -0.364515 1.67301 0.372641 1.98465 1.00043 " style="stroke:none"/>
+               </g>
+               <marker id="mrkr5-38" class="st7" v:arrowType="5" v:arrowSize="2" v:setback="6.16" refX="-6.16" orient="auto"
+                               markerUnits="strokeWidth" overflow="visible">
+                       <use xlink:href="#lend5" transform="scale(-3.52,-3.52) "/>
+               </marker>
+       </defs>
+       <g v:mID="0" v:index="1" v:groupContext="foregroundPage">
+               <title>Page-1</title>
+               <v:pageProperties v:drawingScale="1" v:pageScale="1" v:drawingUnits="0" v:shadowOffsetX="9" v:shadowOffsetY="-9"/>
+               <v:layer v:name="Connector" v:index="0"/>
+               <g id="shape1-1" v:mID="1" v:groupContext="shape" transform="translate(0.25,-117.25)">
+                       <title>Rounded rectangle</title>
+                       <desc>Categorize into an existed “flow”</desc>
+                       <v:userDefs>
+                               <v:ud v:nameU="visVersion" v:val="VT0(14):26"/>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <v:textBlock v:margins="rect(4,4,4,4)"/>
+                       <v:textRect cx="90" cy="173.75" width="180" height="36"/>
+                       <path d="M171 191.75 A9.00007 9.00007 -180 0 0 180 182.75 L180 164.75 A9.00007 9.00007 -180 0 0 171 155.75 L9 155.75
+                                                A9.00007 9.00007 -180 0 0 -0 164.75 L0 182.75 A9.00007 9.00007 -180 0 0 9 191.75 L171 191.75 Z"
+                                       class="st1"/>
+                       <text x="8.91" y="177.35" class="st2" v:langID="1033"><v:paragraph v:horizAlign="1"/><v:tabList/>Categorize into an <tspan
+                                               class="st3">existed</tspan><tspan class="st3" v:langID="2052"> </tspan>“<tspan class="st3">flow</tspan>”</text>             </g>
+               <g id="shape2-9" v:mID="2" v:groupContext="shape" transform="translate(0.25,-58.75)">
+                       <title>Rounded rectangle.2</title>
+                       <desc>Search for a “neighbor”</desc>
+                       <v:userDefs>
+                               <v:ud v:nameU="visVersion" v:val="VT0(14):26"/>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <v:textBlock v:margins="rect(4,4,4,4)"/>
+                       <v:textRect cx="90" cy="173.75" width="180" height="36"/>
+                       <path d="M171 191.75 A9.00007 9.00007 -180 0 0 180 182.75 L180 164.75 A9.00007 9.00007 -180 0 0 171 155.75 L9 155.75
+                                                A9.00007 9.00007 -180 0 0 -0 164.75 L0 182.75 A9.00007 9.00007 -180 0 0 9 191.75 L171 191.75 Z"
+                                       class="st1"/>
+                       <text x="32.19" y="177.35" class="st2" v:langID="1033"><v:paragraph v:horizAlign="1"/><v:tabList/>Search for a “<tspan
+                                               class="st3">neighbor</tspan>”</text>          </g>
+               <g id="shape3-14" v:mID="3" v:groupContext="shape" transform="translate(225.813,-117.25)">
+                       <title>Rounded rectangle.3</title>
+                       <desc>Insert a new “flow” and store the packet</desc>
+                       <v:userDefs>
+                               <v:ud v:nameU="visVersion" v:val="VT0(14):26"/>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <v:textBlock v:margins="rect(4,4,4,4)"/>
+                       <v:textRect cx="105.188" cy="173.75" width="210.38" height="36"/>
+                       <path d="M201.37 191.75 A9.00007 9.00007 -180 0 0 210.37 182.75 L210.37 164.75 A9.00007 9.00007 -180 0 0 201.37 155.75
+                                                L9 155.75 A9.00007 9.00007 -180 0 0 -0 164.75 L0 182.75 A9.00007 9.00007 -180 0 0 9 191.75 L201.37 191.75
+                                                Z" class="st1"/>
+                       <text x="5.45" y="177.35" class="st2" v:langID="1033"><v:paragraph v:horizAlign="1"/><v:tabList/>Insert a <tspan
+                                               class="st3">new </tspan>“<tspan class="st3">flow</tspan>” and <tspan class="st3">store </tspan>the packet</text>            </g>
+               <g id="shape4-21" v:mID="4" v:groupContext="shape" transform="translate(225.25,-58.75)">
+                       <title>Rounded rectangle.4</title>
+                       <desc>Store the packet</desc>
+                       <v:userDefs>
+                               <v:ud v:nameU="visVersion" v:val="VT0(14):26"/>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <v:textBlock v:margins="rect(4,4,4,4)"/>
+                       <v:textRect cx="83.25" cy="173.75" width="166.5" height="36"/>
+                       <path d="M157.5 191.75 A9.00007 9.00007 -180 0 0 166.5 182.75 L166.5 164.75 A9.00007 9.00007 -180 0 0 157.5 155.75 L9
+                                                155.75 A9.00007 9.00007 -180 0 0 -0 164.75 L0 182.75 A9.00007 9.00007 -180 0 0 9 191.75 L157.5 191.75 Z"
+                                       class="st1"/>
+                       <text x="42.81" y="177.35" class="st4" v:langID="1033"><v:paragraph v:horizAlign="1"/><v:tabList/>Store <tspan
+                                               class="st5">the packet</tspan></text>           </g>
+               <g id="shape5-26" v:mID="5" v:groupContext="shape" transform="translate(0.25,-0.25)">
+                       <title>Rounded rectangle.5</title>
+                       <desc>Merge the packet</desc>
+                       <v:userDefs>
+                               <v:ud v:nameU="visVersion" v:val="VT0(14):26"/>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <v:textBlock v:margins="rect(4,4,4,4)"/>
+                       <v:textRect cx="90" cy="173.75" width="180" height="36"/>
+                       <path d="M171 191.75 A9.00007 9.00007 -180 0 0 180 182.75 L180 164.75 A9.00007 9.00007 -180 0 0 171 155.75 L9 155.75
+                                                A9.00007 9.00007 -180 0 0 -0 164.75 L0 182.75 A9.00007 9.00007 -180 0 0 9 191.75 L171 191.75 Z"
+                                       class="st1"/>
+                       <text x="46.59" y="177.35" class="st4" v:langID="1033"><v:paragraph v:horizAlign="1"/><v:tabList/>Merge <tspan
+                                               class="st5">the packet</tspan></text>           </g>
+               <g id="shape6-31" v:mID="6" v:groupContext="shape" v:layerMember="0" transform="translate(81.25,-175.75)">
+                       <title>Dynamic connector</title>
+                       <v:userDefs>
+                               <v:ud v:nameU="visVersion" v:val="VT0(14):26"/>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <path d="M9 191.75 L9 208.09" class="st6"/>
+               </g>
+               <g id="shape7-39" v:mID="7" v:groupContext="shape" v:layerMember="0" transform="translate(81.25,-117.25)">
+                       <title>Dynamic connector.7</title>
+                       <v:userDefs>
+                               <v:ud v:nameU="visVersion" v:val="VT0(14):26"/>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <path d="M9 191.75 L9 208.09" class="st6"/>
+               </g>
+               <g id="shape8-45" v:mID="8" v:groupContext="shape" v:layerMember="0" transform="translate(81.25,-58.75)">
+                       <title>Dynamic connector.8</title>
+                       <v:userDefs>
+                               <v:ud v:nameU="visVersion" v:val="VT0(14):26"/>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <path d="M9 191.75 L9 208.09" class="st6"/>
+               </g>
+               <g id="shape9-51" v:mID="9" v:groupContext="shape" v:layerMember="0" transform="translate(180.25,-126.25)">
+                       <title>Dynamic connector.9</title>
+                       <v:userDefs>
+                               <v:ud v:nameU="visVersion" v:val="VT0(14):26"/>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <path d="M0 182.75 L39.4 182.75" class="st6"/>
+               </g>
+               <g id="shape10-57" v:mID="10" v:groupContext="shape" v:layerMember="0" transform="translate(180.25,-67.75)">
+                       <title>Dynamic connector.10</title>
+                       <v:userDefs>
+                               <v:ud v:nameU="visVersion" v:val="VT0(14):26"/>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <path d="M0 182.75 L38.84 182.75" class="st6"/>
+               </g>
+               <g id="shape11-63" v:mID="11" v:groupContext="shape" transform="translate(65.5,-173.5)">
+                       <title>Sheet.11</title>
+                       <desc>packet</desc>
+                       <v:userDefs>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <v:textBlock v:margins="rect(4,4,4,4)"/>
+                       <v:textRect cx="24.75" cy="182.75" width="49.5" height="18"/>
+                       <rect x="0" y="173.75" width="49.5" height="18" class="st8"/>
+                       <text x="8.46" y="186.35" class="st2" v:langID="1033"><v:paragraph v:horizAlign="1"/><v:tabList/>packet</text>          </g>
+               <g id="shape14-66" v:mID="14" v:groupContext="shape" transform="translate(98.125,-98.125)">
+                       <title>Sheet.14</title>
+                       <desc>find a “flow”</desc>
+                       <v:userDefs>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <v:textBlock v:margins="rect(4,4,4,4)"/>
+                       <v:textRect cx="32.0625" cy="183.875" width="64.13" height="15.75"/>
+                       <rect x="0" y="176" width="64.125" height="15.75" class="st8"/>
+                       <text x="6.41" y="186.88" class="st9" v:langID="1033"><v:paragraph v:horizAlign="1"/><v:tabList/>find a “flow”</text>               </g>
+               <g id="shape15-69" v:mID="15" v:groupContext="shape" transform="translate(99.25,-39.625)">
+                       <title>Sheet.15</title>
+                       <desc>find a “neighbor”</desc>
+                       <v:userDefs>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <v:textBlock v:margins="rect(4,4,4,4)"/>
+                       <v:textRect cx="40.5" cy="183.875" width="81" height="15.75"/>
+                       <rect x="0" y="176" width="81" height="15.75" class="st8"/>
+                       <text x="5.48" y="186.88" class="st9" v:langID="1033"><v:paragraph v:horizAlign="1"/><v:tabList/>find a “neighbor”</text>           </g>
+               <g id="shape13-72" v:mID="13" v:groupContext="shape" transform="translate(181.375,-79)">
+                       <title>Sheet.13</title>
+                       <desc>not find</desc>
+                       <v:userDefs>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <v:textBlock v:margins="rect(4,4,4,4)"/>
+                       <v:textRect cx="21.375" cy="183.875" width="42.75" height="15.75"/>
+                       <rect x="0" y="176" width="42.75" height="15.75" class="st8"/>
+                       <text x="5.38" y="186.88" class="st9" v:langID="1033"><v:paragraph v:horizAlign="1"/><v:tabList/>not find</text>                </g>
+               <g id="shape12-75" v:mID="12" v:groupContext="shape" transform="translate(181.375,-137.5)">
+                       <title>Sheet.12</title>
+                       <desc>not find</desc>
+                       <v:userDefs>
+                               <v:ud v:nameU="msvThemeColors" v:val="VT0(36):26"/>
+                               <v:ud v:nameU="msvThemeEffects" v:val="VT0(16):26"/>
+                       </v:userDefs>
+                       <v:textBlock v:margins="rect(4,4,4,4)"/>
+                       <v:textRect cx="21.375" cy="183.875" width="42.75" height="15.75"/>
+                       <rect x="0" y="176" width="42.75" height="15.75" class="st8"/>
+                       <text x="5.38" y="186.88" class="st9" v:langID="1033"><v:paragraph v:horizAlign="1"/><v:tabList/>not find</text>                </g>
+       </g>
+</svg>
index 03e5ccf..a38a06e 100644 (file)
@@ -44,20 +44,20 @@ gro_tcp4_tbl_create(uint16_t socket_id,
        }
        tbl->max_item_num = entries_num;
 
-       size = sizeof(struct gro_tcp4_key) * entries_num;
-       tbl->keys = rte_zmalloc_socket(__func__,
+       size = sizeof(struct gro_tcp4_flow) * entries_num;
+       tbl->flows = rte_zmalloc_socket(__func__,
                        size,
                        RTE_CACHE_LINE_SIZE,
                        socket_id);
-       if (tbl->keys == NULL) {
+       if (tbl->flows == NULL) {
                rte_free(tbl->items);
                rte_free(tbl);
                return NULL;
        }
-       /* INVALID_ARRAY_INDEX indicates empty key */
+       /* INVALID_ARRAY_INDEX indicates an empty flow */
        for (i = 0; i < entries_num; i++)
-               tbl->keys[i].start_index = INVALID_ARRAY_INDEX;
-       tbl->max_key_num = entries_num;
+               tbl->flows[i].start_index = INVALID_ARRAY_INDEX;
+       tbl->max_flow_num = entries_num;
 
        return tbl;
 }
@@ -69,7 +69,7 @@ gro_tcp4_tbl_destroy(void *tbl)
 
        if (tcp_tbl) {
                rte_free(tcp_tbl->items);
-               rte_free(tcp_tbl->keys);
+               rte_free(tcp_tbl->flows);
        }
        rte_free(tcp_tbl);
 }
@@ -81,50 +81,46 @@ gro_tcp4_tbl_destroy(void *tbl)
  * the original packet.
  */
 static inline int
-merge_two_tcp4_packets(struct gro_tcp4_item *item_src,
+merge_two_tcp4_packets(struct gro_tcp4_item *item,
                struct rte_mbuf *pkt,
-               uint16_t ip_id,
+               int cmp,
                uint32_t sent_seq,
-               int cmp)
+               uint16_t ip_id)
 {
        struct rte_mbuf *pkt_head, *pkt_tail, *lastseg;
-       uint16_t tcp_datalen;
+       uint16_t hdr_len;
 
        if (cmp > 0) {
-               pkt_head = item_src->firstseg;
+               pkt_head = item->firstseg;
                pkt_tail = pkt;
        } else {
                pkt_head = pkt;
-               pkt_tail = item_src->firstseg;
+               pkt_tail = item->firstseg;
        }
 
-       /* check if the packet length will be beyond the max value */
-       tcp_datalen = pkt_tail->pkt_len - pkt_tail->l2_len -
-               pkt_tail->l3_len - pkt_tail->l4_len;
-       if (pkt_head->pkt_len - pkt_head->l2_len + tcp_datalen >
-                       TCP4_MAX_L3_LENGTH)
+       /* check if the IPv4 packet length is greater than the max value */
+       hdr_len = pkt_head->l2_len + pkt_head->l3_len + pkt_head->l4_len;
+       if (unlikely(pkt_head->pkt_len - pkt_head->l2_len + pkt_tail->pkt_len -
+                               hdr_len > MAX_IPV4_PKT_LENGTH))
                return 0;
 
-       /* remove packet header for the tail packet */
-       rte_pktmbuf_adj(pkt_tail,
-                       pkt_tail->l2_len +
-                       pkt_tail->l3_len +
-                       pkt_tail->l4_len);
+       /* remove the packet header for the tail packet */
+       rte_pktmbuf_adj(pkt_tail, hdr_len);
 
        /* chain two packets together */
        if (cmp > 0) {
-               item_src->lastseg->next = pkt;
-               item_src->lastseg = rte_pktmbuf_lastseg(pkt);
+               item->lastseg->next = pkt;
+               item->lastseg = rte_pktmbuf_lastseg(pkt);
                /* update IP ID to the larger value */
-               item_src->ip_id = ip_id;
+               item->ip_id = ip_id;
        } else {
                lastseg = rte_pktmbuf_lastseg(pkt);
-               lastseg->next = item_src->firstseg;
-               item_src->firstseg = pkt;
+               lastseg->next = item->firstseg;
+               item->firstseg = pkt;
                /* update sent_seq to the smaller value */
-               item_src->sent_seq = sent_seq;
+               item->sent_seq = sent_seq;
        }
-       item_src->nb_merged++;
+       item->nb_merged++;
 
        /* update mbuf metadata for the merged packet */
        pkt_head->nb_segs += pkt_tail->nb_segs;
@@ -133,45 +129,46 @@ merge_two_tcp4_packets(struct gro_tcp4_item *item_src,
        return 1;
 }
 
+/*
+ * Check if two TCP/IPv4 packets are neighbors.
+ */
 static inline int
 check_seq_option(struct gro_tcp4_item *item,
-               struct tcp_hdr *tcp_hdr,
-               uint16_t tcp_hl,
-               uint16_t tcp_dl,
+               struct tcp_hdr *tcph,
+               uint32_t sent_seq,
                uint16_t ip_id,
-               uint32_t sent_seq)
+               uint16_t tcp_hl,
+               uint16_t tcp_dl)
 {
-       struct rte_mbuf *pkt0 = item->firstseg;
-       struct ipv4_hdr *ipv4_hdr0;
-       struct tcp_hdr *tcp_hdr0;
-       uint16_t tcp_hl0, tcp_dl0;
-       uint16_t len;
-
-       ipv4_hdr0 = (struct ipv4_hdr *)(rte_pktmbuf_mtod(pkt0, char *) +
-                       pkt0->l2_len);
-       tcp_hdr0 = (struct tcp_hdr *)((char *)ipv4_hdr0 + pkt0->l3_len);
-       tcp_hl0 = pkt0->l4_len;
-
-       /* check if TCP option fields equal. If not, return 0. */
-       len = RTE_MAX(tcp_hl, tcp_hl0) - sizeof(struct tcp_hdr);
-       if ((tcp_hl != tcp_hl0) ||
-                       ((len > 0) && (memcmp(tcp_hdr + 1,
-                                       tcp_hdr0 + 1,
+       struct rte_mbuf *pkt_orig = item->firstseg;
+       struct ipv4_hdr *iph_orig;
+       struct tcp_hdr *tcph_orig;
+       uint16_t len, tcp_hl_orig;
+
+       iph_orig = (struct ipv4_hdr *)(rte_pktmbuf_mtod(pkt_orig, char *) +
+                       pkt_orig->l2_len);
+       tcph_orig = (struct tcp_hdr *)((char *)iph_orig + pkt_orig->l3_len);
+       tcp_hl_orig = pkt_orig->l4_len;
+
+       /* Check if TCP option fields equal */
+       len = RTE_MAX(tcp_hl, tcp_hl_orig) - sizeof(struct tcp_hdr);
+       if ((tcp_hl != tcp_hl_orig) ||
+                       ((len > 0) && (memcmp(tcph + 1, tcph_orig + 1,
                                        len) != 0)))
                return 0;
 
        /* check if the two packets are neighbors */
-       tcp_dl0 = pkt0->pkt_len - pkt0->l2_len - pkt0->l3_len - tcp_hl0;
-       if ((sent_seq == (item->sent_seq + tcp_dl0)) &&
-                       (ip_id == (item->ip_id + 1)))
+       len = pkt_orig->pkt_len - pkt_orig->l2_len - pkt_orig->l3_len -
+               tcp_hl_orig;
+       if ((sent_seq == item->sent_seq + len) && (ip_id == item->ip_id + 1))
                /* append the new packet */
                return 1;
-       else if (((sent_seq + tcp_dl) == item->sent_seq) &&
-                       ((ip_id + item->nb_merged) == item->ip_id))
+       else if ((sent_seq + tcp_dl == item->sent_seq) &&
+                       (ip_id + item->nb_merged == item->ip_id))
                /* pre-pend the new packet */
                return -1;
-       else
-               return 0;
+
+       return 0;
 }
 
 static inline uint32_t
@@ -187,13 +184,13 @@ find_an_empty_item(struct gro_tcp4_tbl *tbl)
 }
 
 static inline uint32_t
-find_an_empty_key(struct gro_tcp4_tbl *tbl)
+find_an_empty_flow(struct gro_tcp4_tbl *tbl)
 {
        uint32_t i;
-       uint32_t max_key_num = tbl->max_key_num;
+       uint32_t max_flow_num = tbl->max_flow_num;
 
-       for (i = 0; i < max_key_num; i++)
-               if (tbl->keys[i].start_index == INVALID_ARRAY_INDEX)
+       for (i = 0; i < max_flow_num; i++)
+               if (tbl->flows[i].start_index == INVALID_ARRAY_INDEX)
                        return i;
        return INVALID_ARRAY_INDEX;
 }
@@ -201,10 +198,10 @@ find_an_empty_key(struct gro_tcp4_tbl *tbl)
 static inline uint32_t
 insert_new_item(struct gro_tcp4_tbl *tbl,
                struct rte_mbuf *pkt,
-               uint16_t ip_id,
-               uint32_t sent_seq,
+               uint64_t start_time,
                uint32_t prev_idx,
-               uint64_t start_time)
+               uint32_t sent_seq,
+               uint16_t ip_id)
 {
        uint32_t item_idx;
 
@@ -221,7 +218,7 @@ insert_new_item(struct gro_tcp4_tbl *tbl,
        tbl->items[item_idx].nb_merged = 1;
        tbl->item_num++;
 
-       /* if the previous packet exists, chain the new one with it */
+       /* if the previous packet exists, chain them together. */
        if (prev_idx != INVALID_ARRAY_INDEX) {
                tbl->items[item_idx].next_pkt_idx =
                        tbl->items[prev_idx].next_pkt_idx;
@@ -237,7 +234,7 @@ delete_item(struct gro_tcp4_tbl *tbl, uint32_t item_idx,
 {
        uint32_t next_idx = tbl->items[item_idx].next_pkt_idx;
 
-       /* set NULL to firstseg to indicate it's an empty item */
+       /* NULL indicates an empty item */
        tbl->items[item_idx].firstseg = NULL;
        tbl->item_num--;
        if (prev_item_idx != INVALID_ARRAY_INDEX)
@@ -247,44 +244,42 @@ delete_item(struct gro_tcp4_tbl *tbl, uint32_t item_idx,
 }
 
 static inline uint32_t
-insert_new_key(struct gro_tcp4_tbl *tbl,
-               struct tcp4_key *key_src,
+insert_new_flow(struct gro_tcp4_tbl *tbl,
+               struct tcp4_flow_key *src,
                uint32_t item_idx)
 {
-       struct tcp4_key *key_dst;
-       uint32_t key_idx;
+       struct tcp4_flow_key *dst;
+       uint32_t flow_idx;
 
-       key_idx = find_an_empty_key(tbl);
-       if (key_idx == INVALID_ARRAY_INDEX)
+       flow_idx = find_an_empty_flow(tbl);
+       if (unlikely(flow_idx == INVALID_ARRAY_INDEX))
                return INVALID_ARRAY_INDEX;
 
-       key_dst = &(tbl->keys[key_idx].key);
+       dst = &(tbl->flows[flow_idx].key);
 
-       ether_addr_copy(&(key_src->eth_saddr), &(key_dst->eth_saddr));
-       ether_addr_copy(&(key_src->eth_daddr), &(key_dst->eth_daddr));
-       key_dst->ip_src_addr = key_src->ip_src_addr;
-       key_dst->ip_dst_addr = key_src->ip_dst_addr;
-       key_dst->recv_ack = key_src->recv_ack;
-       key_dst->src_port = key_src->src_port;
-       key_dst->dst_port = key_src->dst_port;
+       ether_addr_copy(&(src->eth_saddr), &(dst->eth_saddr));
+       ether_addr_copy(&(src->eth_daddr), &(dst->eth_daddr));
+       dst->ip_src_addr = src->ip_src_addr;
+       dst->ip_dst_addr = src->ip_dst_addr;
+       dst->recv_ack = src->recv_ack;
+       dst->src_port = src->src_port;
+       dst->dst_port = src->dst_port;
 
-       /* non-INVALID_ARRAY_INDEX value indicates this key is valid */
-       tbl->keys[key_idx].start_index = item_idx;
-       tbl->key_num++;
+       tbl->flows[flow_idx].start_index = item_idx;
+       tbl->flow_num++;
 
-       return key_idx;
+       return flow_idx;
 }
 
+/*
+ * Check if two TCP/IPv4 packets belong to the same flow.
+ */
 static inline int
-is_same_key(struct tcp4_key k1, struct tcp4_key k2)
+is_same_tcp4_flow(struct tcp4_flow_key k1, struct tcp4_flow_key k2)
 {
-       if (is_same_ether_addr(&k1.eth_saddr, &k2.eth_saddr) == 0)
-               return 0;
-
-       if (is_same_ether_addr(&k1.eth_daddr, &k2.eth_daddr) == 0)
-               return 0;
-
-       return ((k1.ip_src_addr == k2.ip_src_addr) &&
+       return (is_same_ether_addr(&k1.eth_saddr, &k2.eth_saddr) &&
+                       is_same_ether_addr(&k1.eth_daddr, &k2.eth_daddr) &&
+                       (k1.ip_src_addr == k2.ip_src_addr) &&
                        (k1.ip_dst_addr == k2.ip_dst_addr) &&
                        (k1.recv_ack == k2.recv_ack) &&
                        (k1.src_port == k2.src_port) &&
@@ -292,7 +287,7 @@ is_same_key(struct tcp4_key k1, struct tcp4_key k2)
 }
 
 /*
- * update packet length for the flushed packet.
+ * update the packet length for the flushed packet.
  */
 static inline void
 update_header(struct gro_tcp4_item *item)
@@ -315,27 +310,31 @@ gro_tcp4_reassemble(struct rte_mbuf *pkt,
        struct ipv4_hdr *ipv4_hdr;
        struct tcp_hdr *tcp_hdr;
        uint32_t sent_seq;
-       uint16_t tcp_dl, ip_id;
+       uint16_t tcp_dl, ip_id, hdr_len;
 
-       struct tcp4_key key;
+       struct tcp4_flow_key key;
        uint32_t cur_idx, prev_idx, item_idx;
-       uint32_t i, max_key_num;
+       uint32_t i, max_flow_num, remaining_flow_num;
        int cmp;
+       uint8_t find;
 
        eth_hdr = rte_pktmbuf_mtod(pkt, struct ether_hdr *);
        ipv4_hdr = (struct ipv4_hdr *)((char *)eth_hdr + pkt->l2_len);
        tcp_hdr = (struct tcp_hdr *)((char *)ipv4_hdr + pkt->l3_len);
+       hdr_len = pkt->l2_len + pkt->l3_len + pkt->l4_len;
 
        /*
-        * if FIN, SYN, RST, PSH, URG, ECE or
-        * CWR is set, return immediately.
+        * Don't process the packet which has FIN, SYN, RST, PSH, URG, ECE
+        * or CWR set.
         */
        if (tcp_hdr->tcp_flags != TCP_ACK_FLAG)
                return -1;
-       /* if payload length is 0, return immediately */
-       tcp_dl = rte_be_to_cpu_16(ipv4_hdr->total_length) - pkt->l3_len -
-               pkt->l4_len;
-       if (tcp_dl == 0)
+       /*
+        * Don't process the packet whose payload length is less than or
+        * equal to 0.
+        */
+       tcp_dl = pkt->pkt_len - hdr_len;
+       if (tcp_dl <= 0)
                return -1;
 
        ip_id = rte_be_to_cpu_16(ipv4_hdr->packet_id);
@@ -349,25 +348,34 @@ gro_tcp4_reassemble(struct rte_mbuf *pkt,
        key.dst_port = tcp_hdr->dst_port;
        key.recv_ack = tcp_hdr->recv_ack;
 
-       /* search for a key */
-       max_key_num = tbl->max_key_num;
-       for (i = 0; i < max_key_num; i++) {
-               if ((tbl->keys[i].start_index != INVALID_ARRAY_INDEX) &&
-                               is_same_key(tbl->keys[i].key, key))
-                       break;
+       /* Search for a matched flow. */
+       max_flow_num = tbl->max_flow_num;
+       remaining_flow_num = tbl->flow_num;
+       find = 0;
+       for (i = 0; i < max_flow_num && remaining_flow_num; i++) {
+               if (tbl->flows[i].start_index != INVALID_ARRAY_INDEX) {
+                       if (is_same_tcp4_flow(tbl->flows[i].key, key)) {
+                               find = 1;
+                               break;
+                       }
+                       remaining_flow_num--;
+               }
        }
 
-       /* can't find a key, so insert a new key and a new item. */
-       if (i == tbl->max_key_num) {
-               item_idx = insert_new_item(tbl, pkt, ip_id, sent_seq,
-                               INVALID_ARRAY_INDEX, start_time);
+       /*
+        * Fail to find a matched flow. Insert a new flow and store the
+        * packet into the flow.
+        */
+       if (find == 0) {
+               item_idx = insert_new_item(tbl, pkt, start_time,
+                               INVALID_ARRAY_INDEX, sent_seq, ip_id);
                if (item_idx == INVALID_ARRAY_INDEX)
                        return -1;
-               if (insert_new_key(tbl, &key, item_idx) ==
+               if (insert_new_flow(tbl, &key, item_idx) ==
                                INVALID_ARRAY_INDEX) {
                        /*
-                        * fail to insert a new key, so
-                        * delete the inserted item
+                        * Fail to insert a new flow, so delete the
+                        * stored packet.
                         */
                        delete_item(tbl, item_idx, INVALID_ARRAY_INDEX);
                        return -1;
@@ -375,24 +383,26 @@ gro_tcp4_reassemble(struct rte_mbuf *pkt,
                return 0;
        }
 
-       /* traverse all packets in the item group to find one to merge */
-       cur_idx = tbl->keys[i].start_index;
+       /*
+        * Check all packets in the flow and try to find a neighbor for
+        * the input packet.
+        */
+       cur_idx = tbl->flows[i].start_index;
        prev_idx = cur_idx;
        do {
                cmp = check_seq_option(&(tbl->items[cur_idx]), tcp_hdr,
-                               pkt->l4_len, tcp_dl, ip_id, sent_seq);
+                               sent_seq, ip_id, pkt->l4_len, tcp_dl);
                if (cmp) {
                        if (merge_two_tcp4_packets(&(tbl->items[cur_idx]),
-                                               pkt, ip_id,
-                                               sent_seq, cmp))
+                                               pkt, cmp, sent_seq, ip_id))
                                return 1;
                        /*
-                        * fail to merge two packets since the packet
-                        * length will be greater than the max value.
-                        * So insert the packet into the item group.
+                        * Fail to merge the two packets, as the packet
+                        * length is greater than the max value. Store
+                        * the packet into the flow.
                         */
-                       if (insert_new_item(tbl, pkt, ip_id, sent_seq,
-                                               prev_idx, start_time) ==
+                       if (insert_new_item(tbl, pkt, start_time, prev_idx,
+                                               sent_seq, ip_id) ==
                                        INVALID_ARRAY_INDEX)
                                return -1;
                        return 0;
@@ -401,12 +411,9 @@ gro_tcp4_reassemble(struct rte_mbuf *pkt,
                cur_idx = tbl->items[cur_idx].next_pkt_idx;
        } while (cur_idx != INVALID_ARRAY_INDEX);
 
-       /*
-        * can't find a packet in the item group to merge,
-        * so insert the packet into the item group.
-        */
-       if (insert_new_item(tbl, pkt, ip_id, sent_seq, prev_idx,
-                               start_time) == INVALID_ARRAY_INDEX)
+       /* Fail to find a neighbor, so store the packet into the flow. */
+       if (insert_new_item(tbl, pkt, start_time, prev_idx, sent_seq,
+                               ip_id) == INVALID_ARRAY_INDEX)
                return -1;
 
        return 0;
@@ -420,44 +427,33 @@ gro_tcp4_tbl_timeout_flush(struct gro_tcp4_tbl *tbl,
 {
        uint16_t k = 0;
        uint32_t i, j;
-       uint32_t max_key_num = tbl->max_key_num;
+       uint32_t max_flow_num = tbl->max_flow_num;
 
-       for (i = 0; i < max_key_num; i++) {
-               /* all keys have been checked, return immediately */
-               if (tbl->key_num == 0)
+       for (i = 0; i < max_flow_num; i++) {
+               if (unlikely(tbl->flow_num == 0))
                        return k;
 
-               j = tbl->keys[i].start_index;
+               j = tbl->flows[i].start_index;
                while (j != INVALID_ARRAY_INDEX) {
                        if (tbl->items[j].start_time <= flush_timestamp) {
                                out[k++] = tbl->items[j].firstseg;
                                if (tbl->items[j].nb_merged > 1)
                                        update_header(&(tbl->items[j]));
                                /*
-                                * delete the item and get
-                                * the next packet index
+                                * Delete the packet and get the next
+                                * packet in the flow.
                                 */
-                               j = delete_item(tbl, j,
-                                               INVALID_ARRAY_INDEX);
+                               j = delete_item(tbl, j, INVALID_ARRAY_INDEX);
+                               tbl->flows[i].start_index = j;
+                               if (j == INVALID_ARRAY_INDEX)
+                                       tbl->flow_num--;
 
-                               /*
-                                * delete the key as all of
-                                * packets are flushed
-                                */
-                               if (j == INVALID_ARRAY_INDEX) {
-                                       tbl->keys[i].start_index =
-                                               INVALID_ARRAY_INDEX;
-                                       tbl->key_num--;
-                               } else
-                                       /* update start_index of the key */
-                                       tbl->keys[i].start_index = j;
-
-                               if (k == nb_out)
+                               if (unlikely(k == nb_out))
                                        return k;
                        } else
                                /*
-                                * left packets of this key won't be
-                                * timeout, so go to check other keys.
+                                * The left packets in this flow won't be
+                                * timeout. Go to check other flows.
                                 */
                                break;
                }
index d129523..49e03b4 100644 (file)
@@ -9,13 +9,13 @@
 #define GRO_TCP4_TBL_MAX_ITEM_NUM (1024UL * 1024UL)
 
 /*
- * the max L3 length of a TCP/IPv4 packet. The L3 length
- * is the sum of ipv4 header, tcp header and L4 payload.
+ * The max length of a IPv4 packet, which includes the length of the L3
+ * header, the L4 header and the data payload.
  */
-#define TCP4_MAX_L3_LENGTH UINT16_MAX
+#define MAX_IPV4_PKT_LENGTH UINT16_MAX
 
-/* criteria of mergeing packets */
-struct tcp4_key {
+/* Header fields representing a TCP/IPv4 flow */
+struct tcp4_flow_key {
        struct ether_addr eth_saddr;
        struct ether_addr eth_daddr;
        uint32_t ip_src_addr;
@@ -26,41 +26,38 @@ struct tcp4_key {
        uint16_t dst_port;
 };
 
-struct gro_tcp4_key {
-       struct tcp4_key key;
+struct gro_tcp4_flow {
+       struct tcp4_flow_key key;
        /*
-        * the index of the first packet in the item group.
-        * If the value is INVALID_ARRAY_INDEX, it means
-        * the key is empty.
+        * The index of the first packet in the flow.
+        * INVALID_ARRAY_INDEX indicates an empty flow.
         */
        uint32_t start_index;
 };
 
 struct gro_tcp4_item {
        /*
-        * first segment of the packet. If the value
+        * The first MBUF segment of the packet. If the value
         * is NULL, it means the item is empty.
         */
        struct rte_mbuf *firstseg;
-       /* last segment of the packet */
+       /* The last MBUF segment of the packet */
        struct rte_mbuf *lastseg;
        /*
-        * the time when the first packet is inserted
-        * into the table. If a packet in the table is
-        * merged with an incoming packet, this value
-        * won't be updated. We set this value only
-        * when the first packet is inserted into the
-        * table.
+        * The time when the first packet is inserted into the table.
+        * This value won't be updated, even if the packet is merged
+        * with other packets.
         */
        uint64_t start_time;
        /*
-        * we use next_pkt_idx to chain the packets that
-        * have same key value but can't be merged together.
+        * next_pkt_idx is used to chain the packets that
+        * are in the same flow but can't be merged together
+        * (e.g. caused by packet reordering).
         */
        uint32_t next_pkt_idx;
-       /* the sequence number of the packet */
+       /* TCP sequence number of the packet */
        uint32_t sent_seq;
-       /* the IP ID of the packet */
+       /* IPv4 ID of the packet */
        uint16_t ip_id;
        /* the number of merged packets */
        uint16_t nb_merged;
@@ -72,31 +69,31 @@ struct gro_tcp4_item {
 struct gro_tcp4_tbl {
        /* item array */
        struct gro_tcp4_item *items;
-       /* key array */
-       struct gro_tcp4_key *keys;
+       /* flow array */
+       struct gro_tcp4_flow *flows;
        /* current item number */
        uint32_t item_num;
-       /* current key num */
-       uint32_t key_num;
+       /* current flow num */
+       uint32_t flow_num;
        /* item array size */
        uint32_t max_item_num;
-       /* key array size */
-       uint32_t max_key_num;
+       /* flow array size */
+       uint32_t max_flow_num;
 };
 
 /**
  * This function creates a TCP/IPv4 reassembly table.
  *
  * @param socket_id
- *  socket index for allocating TCP/IPv4 reassemble table
+ *  Socket index for allocating the TCP/IPv4 reassemble table
  * @param max_flow_num
- *  the maximum number of flows in the TCP/IPv4 GRO table
+ *  The maximum number of flows in the TCP/IPv4 GRO table
  * @param max_item_per_flow
- *  the maximum packet number per flow.
+ *  The maximum number of packets per flow
  *
  * @return
- *  if create successfully, return a pointer which points to the
- *  created TCP/IPv4 GRO table. Otherwise, return NULL.
+ *  - Return the table pointer on success.
+ *  - Return NULL on failure.
  */
 void *gro_tcp4_tbl_create(uint16_t socket_id,
                uint16_t max_flow_num,
@@ -106,62 +103,56 @@ void *gro_tcp4_tbl_create(uint16_t socket_id,
  * This function destroys a TCP/IPv4 reassembly table.
  *
  * @param tbl
- *  a pointer points to the TCP/IPv4 reassembly table.
+ *  Pointer pointing to the TCP/IPv4 reassembly table.
  */
 void gro_tcp4_tbl_destroy(void *tbl);
 
 /**
- * This function searches for a packet in the TCP/IPv4 reassembly table
- * to merge with the inputted one. To merge two packets is to chain them
- * together and update packet headers. Packets, whose SYN, FIN, RST, PSH
- * CWR, ECE or URG bit is set, are returned immediately. Packets which
- * only have packet headers (i.e. without data) are also returned
- * immediately. Otherwise, the packet is either merged, or inserted into
- * the table. Besides, if there is no available space to insert the
- * packet, this function returns immediately too.
+ * This function merges a TCP/IPv4 packet. It doesn't process the packet,
+ * which has SYN, FIN, RST, PSH, CWR, ECE or URG set, or doesn't have
+ * payload.
  *
- * This function assumes the inputted packet is with correct IPv4 and
- * TCP checksums. And if two packets are merged, it won't re-calculate
- * IPv4 and TCP checksums. Besides, if the inputted packet is IP
- * fragmented, it assumes the packet is complete (with TCP header).
+ * This function doesn't check if the packet has correct checksums and
+ * doesn't re-calculate checksums for the merged packet. Additionally,
+ * it assumes the packets are complete (i.e., MF==0 && frag_off==0),
+ * when IP fragmentation is possible (i.e., DF==0). It returns the
+ * packet, if the packet has invalid parameters (e.g. SYN bit is set)
+ * or there is no available space in the table.
  *
  * @param pkt
- *  packet to reassemble.
+ *  Packet to reassemble
  * @param tbl
- *  a pointer that points to a TCP/IPv4 reassembly table.
+ *  Pointer pointing to the TCP/IPv4 reassembly table
  * @start_time
- *  the start time that the packet is inserted into the table
+ *  The time when the packet is inserted into the table
  *
  * @return
- *  if the packet doesn't have data, or SYN, FIN, RST, PSH, CWR, ECE
- *  or URG bit is set, or there is no available space in the table to
- *  insert a new item or a new key, return a negative value. If the
- *  packet is merged successfully, return an positive value. If the
- *  packet is inserted into the table, return 0.
+ *  - Return a positive value if the packet is merged.
+ *  - Return zero if the packet isn't merged but stored in the table.
+ *  - Return a negative value for invalid parameters or no available
+ *    space in the table.
  */
 int32_t gro_tcp4_reassemble(struct rte_mbuf *pkt,
                struct gro_tcp4_tbl *tbl,
                uint64_t start_time);
 
 /**
- * This function flushes timeout packets in a TCP/IPv4 reassembly table
- * to applications, and without updating checksums for merged packets.
- * The max number of flushed timeout packets is the element number of
- * the array which is used to keep flushed packets.
+ * This function flushes timeout packets in a TCP/IPv4 reassembly table,
+ * and without updating checksums.
  *
  * @param tbl
- *  a pointer that points to a TCP GRO table.
+ *  TCP/IPv4 reassembly table pointer
  * @param flush_timestamp
- *  this function flushes packets which are inserted into the table
- *  before or at the flush_timestamp.
+ *  Flush packets which are inserted into the table before or at the
+ *  flush_timestamp.
  * @param out
- *  pointer array which is used to keep flushed packets.
+ *  Pointer array used to keep flushed packets
  * @param nb_out
- *  the element number of out. It's also the max number of timeout
+ *  The element number in 'out'. It also determines the maximum number of
  *  packets that can be flushed finally.
  *
  * @return
- *  the number of packets that are returned.
+ *  The number of flushed packets
  */
 uint16_t gro_tcp4_tbl_timeout_flush(struct gro_tcp4_tbl *tbl,
                uint64_t flush_timestamp,
@@ -173,10 +164,10 @@ uint16_t gro_tcp4_tbl_timeout_flush(struct gro_tcp4_tbl *tbl,
  * reassembly table.
  *
  * @param tbl
- *  pointer points to a TCP/IPv4 reassembly table.
+ *  TCP/IPv4 reassembly table pointer
  *
  * @return
- *  the number of packets in the table
+ *  The number of packets in the table
  */
 uint32_t gro_tcp4_tbl_pkt_count(void *tbl);
 #endif
index d6b8cd1..0b64866 100644 (file)
@@ -23,11 +23,14 @@ static gro_tbl_destroy_fn tbl_destroy_fn[RTE_GRO_TYPE_MAX_NUM] = {
 static gro_tbl_pkt_count_fn tbl_pkt_count_fn[RTE_GRO_TYPE_MAX_NUM] = {
                        gro_tcp4_tbl_pkt_count, NULL};
 
+#define IS_IPV4_TCP_PKT(ptype) (RTE_ETH_IS_IPV4_HDR(ptype) && \
+               ((ptype & RTE_PTYPE_L4_TCP) == RTE_PTYPE_L4_TCP))
+
 /*
- * GRO context structure, which is used to merge packets. It keeps
- * many reassembly tables of desired GRO types. Applications need to
- * create GRO context objects before using rte_gro_reassemble to
- * perform GRO.
+ * GRO context structure. It keeps the table structures, which are
+ * used to merge packets, for different GRO types. Before using
+ * rte_gro_reassemble(), applications need to create the GRO context
+ * first.
  */
 struct gro_ctx {
        /* GRO types to perform */
@@ -85,8 +88,6 @@ rte_gro_ctx_destroy(void *ctx)
        uint64_t gro_type_flag;
        uint8_t i;
 
-       if (gro_ctx == NULL)
-               return;
        for (i = 0; i < RTE_GRO_TYPE_MAX_NUM; i++) {
                gro_type_flag = 1ULL << i;
                if ((gro_ctx->gro_types & gro_type_flag) == 0)
@@ -103,62 +104,54 @@ rte_gro_reassemble_burst(struct rte_mbuf **pkts,
                uint16_t nb_pkts,
                const struct rte_gro_param *param)
 {
-       uint16_t i;
-       uint16_t nb_after_gro = nb_pkts;
-       uint32_t item_num;
-
        /* allocate a reassembly table for TCP/IPv4 GRO */
        struct gro_tcp4_tbl tcp_tbl;
-       struct gro_tcp4_key tcp_keys[RTE_GRO_MAX_BURST_ITEM_NUM];
+       struct gro_tcp4_flow tcp_flows[RTE_GRO_MAX_BURST_ITEM_NUM];
        struct gro_tcp4_item tcp_items[RTE_GRO_MAX_BURST_ITEM_NUM] = {{0} };
 
        struct rte_mbuf *unprocess_pkts[nb_pkts];
-       uint16_t unprocess_num = 0;
+       uint32_t item_num;
        int32_t ret;
-       uint64_t current_time;
+       uint16_t i, unprocess_num = 0, nb_after_gro = nb_pkts;
 
-       if ((param->gro_types & RTE_GRO_TCP_IPV4) == 0)
+       if (unlikely((param->gro_types & RTE_GRO_TCP_IPV4) == 0))
                return nb_pkts;
 
-       /* get the actual number of packets */
+       /* Get the maximum number of packets */
        item_num = RTE_MIN(nb_pkts, (param->max_flow_num *
-                       param->max_item_per_flow));
+                               param->max_item_per_flow));
        item_num = RTE_MIN(item_num, RTE_GRO_MAX_BURST_ITEM_NUM);
 
        for (i = 0; i < item_num; i++)
-               tcp_keys[i].start_index = INVALID_ARRAY_INDEX;
+               tcp_flows[i].start_index = INVALID_ARRAY_INDEX;
 
-       tcp_tbl.keys = tcp_keys;
+       tcp_tbl.flows = tcp_flows;
        tcp_tbl.items = tcp_items;
-       tcp_tbl.key_num = 0;
+       tcp_tbl.flow_num = 0;
        tcp_tbl.item_num = 0;
-       tcp_tbl.max_key_num = item_num;
+       tcp_tbl.max_flow_num = item_num;
        tcp_tbl.max_item_num = item_num;
 
-       current_time = rte_rdtsc();
-
        for (i = 0; i < nb_pkts; i++) {
-               if ((pkts[i]->packet_type & (RTE_PTYPE_L3_IPV4 |
-                                       RTE_PTYPE_L4_TCP)) ==
-                               (RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_TCP)) {
-                       ret = gro_tcp4_reassemble(pkts[i],
-                                       &tcp_tbl,
-                                       current_time);
+               if (IS_IPV4_TCP_PKT(pkts[i]->packet_type)) {
+                       /*
+                        * The timestamp is ignored, since all packets
+                        * will be flushed from the tables.
+                        */
+                       ret = gro_tcp4_reassemble(pkts[i], &tcp_tbl, 0);
                        if (ret > 0)
                                /* merge successfully */
                                nb_after_gro--;
-                       else if (ret < 0) {
-                               unprocess_pkts[unprocess_num++] =
-                                       pkts[i];
-                       }
+                       else if (ret < 0)
+                               unprocess_pkts[unprocess_num++] = pkts[i];
                } else
                        unprocess_pkts[unprocess_num++] = pkts[i];
        }
 
-       /* re-arrange GROed packets */
        if (nb_after_gro < nb_pkts) {
-               i = gro_tcp4_tbl_timeout_flush(&tcp_tbl, current_time,
-                               pkts, nb_pkts);
+               /* Flush all packets from the tables */
+               i = gro_tcp4_tbl_timeout_flush(&tcp_tbl, 0, pkts, nb_pkts);
+               /* Copy unprocessed packets */
                if (unprocess_num > 0) {
                        memcpy(&pkts[i], unprocess_pkts,
                                        sizeof(struct rte_mbuf *) *
@@ -174,31 +167,28 @@ rte_gro_reassemble(struct rte_mbuf **pkts,
                uint16_t nb_pkts,
                void *ctx)
 {
-       uint16_t i, unprocess_num = 0;
        struct rte_mbuf *unprocess_pkts[nb_pkts];
        struct gro_ctx *gro_ctx = ctx;
+       void *tcp_tbl;
        uint64_t current_time;
+       uint16_t i, unprocess_num = 0;
 
-       if ((gro_ctx->gro_types & RTE_GRO_TCP_IPV4) == 0)
+       if (unlikely((gro_ctx->gro_types & RTE_GRO_TCP_IPV4) == 0))
                return nb_pkts;
 
+       tcp_tbl = gro_ctx->tbls[RTE_GRO_TCP_IPV4_INDEX];
        current_time = rte_rdtsc();
 
        for (i = 0; i < nb_pkts; i++) {
-               if ((pkts[i]->packet_type & (RTE_PTYPE_L3_IPV4 |
-                                       RTE_PTYPE_L4_TCP)) ==
-                               (RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_TCP)) {
-                       if (gro_tcp4_reassemble(pkts[i],
-                                               gro_ctx->tbls
-                                               [RTE_GRO_TCP_IPV4_INDEX],
+               if (IS_IPV4_TCP_PKT(pkts[i]->packet_type)) {
+                       if (gro_tcp4_reassemble(pkts[i], tcp_tbl,
                                                current_time) < 0)
                                unprocess_pkts[unprocess_num++] = pkts[i];
                } else
                        unprocess_pkts[unprocess_num++] = pkts[i];
        }
        if (unprocess_num > 0) {
-               memcpy(pkts, unprocess_pkts,
-                               sizeof(struct rte_mbuf *) *
+               memcpy(pkts, unprocess_pkts, sizeof(struct rte_mbuf *) *
                                unprocess_num);
        }
 
@@ -224,6 +214,7 @@ rte_gro_timeout_flush(void *ctx,
                                flush_timestamp,
                                out, max_nb_out);
        }
+
        return 0;
 }
 
@@ -232,19 +223,20 @@ rte_gro_get_pkt_count(void *ctx)
 {
        struct gro_ctx *gro_ctx = ctx;
        gro_tbl_pkt_count_fn pkt_count_fn;
+       uint64_t gro_types = gro_ctx->gro_types, flag;
        uint64_t item_num = 0;
-       uint64_t gro_type_flag;
        uint8_t i;
 
-       for (i = 0; i < RTE_GRO_TYPE_MAX_NUM; i++) {
-               gro_type_flag = 1ULL << i;
-               if ((gro_ctx->gro_types & gro_type_flag) == 0)
+       for (i = 0; i < RTE_GRO_TYPE_MAX_NUM && gro_types; i++) {
+               flag = 1ULL << i;
+               if ((gro_types & flag) == 0)
                        continue;
 
+               gro_types ^= flag;
                pkt_count_fn = tbl_pkt_count_fn[i];
-               if (pkt_count_fn == NULL)
-                       continue;
-               item_num += pkt_count_fn(gro_ctx->tbls[i]);
+               if (pkt_count_fn)
+                       item_num += pkt_count_fn(gro_ctx->tbls[i]);
        }
+
        return item_num;
 }
index 81a2eac..85d8143 100644 (file)
@@ -31,8 +31,8 @@ extern "C" {
 /**< TCP/IPv4 GRO flag */
 
 /**
- * A structure which is used to create GRO context objects or tell
- * rte_gro_reassemble_burst() what reassembly rules are demanded.
+ * Structure used to create GRO context objects or used to pass
+ * application-determined parameters to rte_gro_reassemble_burst().
  */
 struct rte_gro_param {
        uint64_t gro_types;
@@ -78,26 +78,23 @@ void rte_gro_ctx_destroy(void *ctx);
 
 /**
  * This is one of the main reassembly APIs, which merges numbers of
- * packets at a time. It assumes that all inputted packets are with
- * correct checksums. That is, applications should guarantee all
- * inputted packets are correct. Besides, it doesn't re-calculate
- * checksums for merged packets. If inputted packets are IP fragmented,
- * this function assumes them are complete (i.e. with L4 header). After
- * finishing processing, it returns all GROed packets to applications
- * immediately.
+ * packets at a time. It doesn't check if input packets have correct
+ * checksums and doesn't re-calculate checksums for merged packets.
+ * It assumes the packets are complete (i.e., MF==0 && frag_off==0),
+ * when IP fragmentation is possible (i.e., DF==0). The GROed packets
+ * are returned as soon as the function finishes.
  *
  * @param pkts
- *  a pointer array which points to the packets to reassemble. Besides,
- *  it keeps mbuf addresses for the GROed packets.
+ *  Pointer array pointing to the packets to reassemble. Besides, it
+ *  keeps MBUF addresses for the GROed packets.
  * @param nb_pkts
- *  the number of packets to reassemble.
+ *  The number of packets to reassemble
  * @param param
- *  applications use it to tell rte_gro_reassemble_burst() what rules
- *  are demanded.
+ *  Application-determined parameters for reassembling packets.
  *
  * @return
- *  the number of packets after been GROed. If no packets are merged,
- *  the returned value is nb_pkts.
+ *  The number of packets after been GROed. If no packets are merged,
+ *  the return value is equals to nb_pkts.
  */
 uint16_t rte_gro_reassemble_burst(struct rte_mbuf **pkts,
                uint16_t nb_pkts,
@@ -107,32 +104,28 @@ uint16_t rte_gro_reassemble_burst(struct rte_mbuf **pkts,
  * @warning
  * @b EXPERIMENTAL: this API may change without prior notice
  *
- * Reassembly function, which tries to merge inputted packets with
- * the packets in the reassembly tables of a given GRO context. This
- * function assumes all inputted packets are with correct checksums.
- * And it won't update checksums if two packets are merged. Besides,
- * if inputted packets are IP fragmented, this function assumes they
- * are complete packets (i.e. with L4 header).
+ * Reassembly function, which tries to merge input packets with the
+ * existed packets in the reassembly tables of a given GRO context.
+ * It doesn't check if input packets have correct checksums and doesn't
+ * re-calculate checksums for merged packets. Additionally, it assumes
+ * the packets are complete (i.e., MF==0 && frag_off==0), when IP
+ * fragmentation is possible (i.e., DF==0).
  *
- * If the inputted packets don't have data or are with unsupported GRO
- * types etc., they won't be processed and are returned to applications.
- * Otherwise, the inputted packets are either merged or inserted into
- * the table. If applications want get packets in the table, they need
- * to call flush API.
+ * If the input packets have invalid parameters (e.g. no data payload,
+ * unsupported GRO types), they are returned to applications. Otherwise,
+ * they are either merged or inserted into the table. Applications need
+ * to flush packets from the tables by flush API, if they want to get the
+ * GROed packets.
  *
  * @param pkts
- *  packet to reassemble. Besides, after this function finishes, it
- *  keeps the unprocessed packets (e.g. without data or unsupported
- *  GRO types).
+ *  Packets to reassemble. It's also used to store the unprocessed packets.
  * @param nb_pkts
- *  the number of packets to reassemble.
+ *  The number of packets to reassemble
  * @param ctx
- *  a pointer points to a GRO context object.
+ *  GRO context object pointer
  *
  * @return
- *  return the number of unprocessed packets (e.g. without data or
- *  unsupported GRO types). If all packets are processed (merged or
- *  inserted into the table), return 0.
+ *  The number of unprocessed packets.
  */
 uint16_t rte_gro_reassemble(struct rte_mbuf **pkts,
                uint16_t nb_pkts,
@@ -142,29 +135,28 @@ uint16_t rte_gro_reassemble(struct rte_mbuf **pkts,
  * @warning
  * @b EXPERIMENTAL: this API may change without prior notice
  *
- * This function flushes the timeout packets from reassembly tables of
- * desired GRO types. The max number of flushed timeout packets is the
- * element number of the array which is used to keep the flushed packets.
+ * This function flushes the timeout packets from the reassembly tables
+ * of desired GRO types. The max number of flushed packets is the
+ * element number of 'out'.
  *
- * Besides, this function won't re-calculate checksums for merged
- * packets in the tables. That is, the returned packets may be with
- * wrong checksums.
+ * Additionally, the flushed packets may have incorrect checksums, since
+ * this function doesn't re-calculate checksums for merged packets.
  *
  * @param ctx
- *  a pointer points to a GRO context object.
+ *  GRO context object pointer.
  * @param timeout_cycles
- *  max TTL for packets in reassembly tables, measured in nanosecond.
+ *  The max TTL for packets in reassembly tables, measured in nanosecond.
  * @param gro_types
- *  this function only flushes packets which belong to the GRO types
- *  specified by gro_types.
+ *  This function flushes packets whose GRO types are specified by
+ *  gro_types.
  * @param out
- *  a pointer array that is used to keep flushed timeout packets.
+ *  Pointer array used to keep flushed packets.
  * @param max_nb_out
- *  the element number of out. It's also the max number of timeout
+ *  The element number of 'out'. It's also the max number of timeout
  *  packets that can be flushed finally.
  *
  * @return
- *  the number of flushed packets. If no packets are flushed, return 0.
+ *  The number of flushed packets.
  */
 uint16_t rte_gro_timeout_flush(void *ctx,
                uint64_t timeout_cycles,
@@ -180,10 +172,10 @@ uint16_t rte_gro_timeout_flush(void *ctx,
  * of a given GRO context.
  *
  * @param ctx
- *  pointer points to a GRO context object.
+ *  GRO context object pointer.
  *
  * @return
- *  the number of packets in all reassembly tables.
+ *  The number of packets in the tables.
  */
 uint64_t rte_gro_get_pkt_count(void *ctx);