1The QorIQ DPAA Ethernet Driver
2==============================
3
4Authors:
5Madalin Bucur <madalin.bucur@nxp.com>
6Camelia Groza <camelia.groza@nxp.com>
7
8Contents
9========
10
11	- DPAA Ethernet Overview
12	- DPAA Ethernet Supported SoCs
13	- Configuring DPAA Ethernet in your kernel
14	- DPAA Ethernet Frame Processing
15	- DPAA Ethernet Features
16	- DPAA IRQ Affinity and Receive Side Scaling
17	- Debugging
18
19DPAA Ethernet Overview
20======================
21
22DPAA stands for Data Path Acceleration Architecture and it is a
23set of networking acceleration IPs that are available on several
24generations of SoCs, both on PowerPC and ARM64.
25
26The Freescale DPAA architecture consists of a series of hardware blocks
27that support Ethernet connectivity. The Ethernet driver depends upon the
28following drivers in the Linux kernel:
29
30 - Peripheral Access Memory Unit (PAMU) (* needed only for PPC platforms)
31    drivers/iommu/fsl_*
32 - Frame Manager (FMan)
33    drivers/net/ethernet/freescale/fman
34 - Queue Manager (QMan), Buffer Manager (BMan)
35    drivers/soc/fsl/qbman
36
37A simplified view of the dpaa_eth interfaces mapped to FMan MACs:
38
39  dpaa_eth       /eth0\     ...       /ethN\
40  driver        |      |             |      |
41  -------------   ----   -----------   ----   -------------
42       -Ports  / Tx  Rx \    ...    / Tx  Rx \
43  FMan        |          |         |          |
44       -MACs  |   MAC0   |         |   MACN   |
45             /   dtsec0   \  ...  /   dtsecN   \ (or tgec)
46            /              \     /              \(or memac)
47  ---------  --------------  ---  --------------  ---------
48      FMan, FMan Port, FMan SP, FMan MURAM drivers
49  ---------------------------------------------------------
50      FMan HW blocks: MURAM, MACs, Ports, SP
51  ---------------------------------------------------------
52
53The dpaa_eth relation to the QMan, BMan and FMan:
54              ________________________________
55  dpaa_eth   /            eth0                \
56  driver    /                                  \
57  ---------   -^-   -^-   -^-   ---    ---------
58  QMan driver / \   / \   / \  \   /  | BMan    |
59             |Rx | |Rx | |Tx | |Tx |  | driver  |
60  ---------  |Dfl| |Err| |Cnf| |FQs|  |         |
61  QMan HW    |FQ | |FQ | |FQs| |   |  |         |
62             /   \ /   \ /   \  \ /   |         |
63  ---------   ---   ---   ---   -v-    ---------
64            |        FMan QMI         |         |
65            | FMan HW       FMan BMI  | BMan HW |
66              -----------------------   --------
67
68where the acronyms used above (and in the code) are:
69DPAA = Data Path Acceleration Architecture
70FMan = DPAA Frame Manager
71QMan = DPAA Queue Manager
72BMan = DPAA Buffers Manager
73QMI = QMan interface in FMan
74BMI = BMan interface in FMan
75FMan SP = FMan Storage Profiles
76MURAM = Multi-user RAM in FMan
77FQ = QMan Frame Queue
78Rx Dfl FQ = default reception FQ
79Rx Err FQ = Rx error frames FQ
80Tx Cnf FQ = Tx confirmation FQs
81Tx FQs = transmission frame queues
82dtsec = datapath three speed Ethernet controller (10/100/1000 Mbps)
83tgec = ten gigabit Ethernet controller (10 Gbps)
84memac = multirate Ethernet MAC (10/100/1000/10000)
85
86DPAA Ethernet Supported SoCs
87============================
88
89The DPAA drivers enable the Ethernet controllers present on the following SoCs:
90
91# PPC
92P1023
93P2041
94P3041
95P4080
96P5020
97P5040
98T1023
99T1024
100T1040
101T1042
102T2080
103T4240
104B4860
105
106# ARM
107LS1043A
108LS1046A
109
110Configuring DPAA Ethernet in your kernel
111========================================
112
113To enable the DPAA Ethernet driver, the following Kconfig options are required:
114
115# common for arch/arm64 and arch/powerpc platforms
116CONFIG_FSL_DPAA=y
117CONFIG_FSL_FMAN=y
118CONFIG_FSL_DPAA_ETH=y
119CONFIG_FSL_XGMAC_MDIO=y
120
121# for arch/powerpc only
122CONFIG_FSL_PAMU=y
123
124# common options needed for the PHYs used on the RDBs
125CONFIG_VITESSE_PHY=y
126CONFIG_REALTEK_PHY=y
127CONFIG_AQUANTIA_PHY=y
128
129DPAA Ethernet Frame Processing
130==============================
131
132On Rx, buffers for the incoming frames are retrieved from one of the three
133existing buffers pools. The driver initializes and seeds these, each with
134buffers of different sizes: 1KB, 2KB and 4KB.
135
136On Tx, all transmitted frames are returned to the driver through Tx
137confirmation frame queues. The driver is then responsible for freeing the
138buffers. In order to do this properly, a backpointer is added to the buffer
139before transmission that points to the skb. When the buffer returns to the
140driver on a confirmation FQ, the skb can be correctly consumed.
141
142DPAA Ethernet Features
143======================
144
145Currently the DPAA Ethernet driver enables the basic features required for
146a Linux Ethernet driver. The support for advanced features will be added
147gradually.
148
149The driver has Rx and Tx checksum offloading for UDP and TCP. Currently the Rx
150checksum offload feature is enabled by default and cannot be controlled through
151ethtool. Also, rx-flow-hash and rx-hashing was added. The addition of RSS
152provides a big performance boost for the forwarding scenarios, allowing
153different traffic flows received by one interface to be processed by different
154CPUs in parallel.
155
156The driver has support for multiple prioritized Tx traffic classes. Priorities
157range from 0 (lowest) to 3 (highest). These are mapped to HW workqueues with
158strict priority levels. Each traffic class contains NR_CPU TX queues. By
159default, only one traffic class is enabled and the lowest priority Tx queues
160are used. Higher priority traffic classes can be enabled with the mqprio
161qdisc. For example, all four traffic classes are enabled on an interface with
162the following command. Furthermore, skb priority levels are mapped to traffic
163classes as follows:
164
165	* priorities 0 to 3 - traffic class 0 (low priority)
166	* priorities 4 to 7 - traffic class 1 (medium-low priority)
167	* priorities 8 to 11 - traffic class 2 (medium-high priority)
168	* priorities 12 to 15 - traffic class 3 (high priority)
169
170tc qdisc add dev <int> root handle 1: \
171	 mqprio num_tc 4 map 0 0 0 0 1 1 1 1 2 2 2 2 3 3 3 3 hw 1
172
173DPAA IRQ Affinity and Receive Side Scaling
174==========================================
175
176Traffic coming on the DPAA Rx queues or on the DPAA Tx confirmation
177queues is seen by the CPU as ingress traffic on a certain portal.
178The DPAA QMan portal interrupts are affined each to a certain CPU.
179The same portal interrupt services all the QMan portal consumers.
180
181By default the DPAA Ethernet driver enables RSS, making use of the
182DPAA FMan Parser and Keygen blocks to distribute traffic on 128
183hardware frame queues using a hash on IP v4/v6 source and destination
184and L4 source and destination ports, in present in the received frame.
185When RSS is disabled, all traffic received by a certain interface is
186received on the default Rx frame queue. The default DPAA Rx frame
187queues are configured to put the received traffic into a pool channel
188that allows any available CPU portal to dequeue the ingress traffic.
189The default frame queues have the HOLDACTIVE option set, ensuring that
190traffic bursts from a certain queue are serviced by the same CPU.
191This ensures a very low rate of frame reordering. A drawback of this
192is that only one CPU at a time can service the traffic received by a
193certain interface when RSS is not enabled.
194
195To implement RSS, the DPAA Ethernet driver allocates an extra set of
196128 Rx frame queues that are configured to dedicated channels, in a
197round-robin manner. The mapping of the frame queues to CPUs is now
198hardcoded, there is no indirection table to move traffic for a certain
199FQ (hash result) to another CPU. The ingress traffic arriving on one
200of these frame queues will arrive at the same portal and will always
201be processed by the same CPU. This ensures intra-flow order preservation
202and workload distribution for multiple traffic flows.
203
204RSS can be turned off for a certain interface using ethtool, i.e.
205
206	# ethtool -N fm1-mac9 rx-flow-hash tcp4 ""
207
208To turn it back on, one needs to set rx-flow-hash for tcp4/6 or udp4/6:
209
210	# ethtool -N fm1-mac9 rx-flow-hash udp4 sfdn
211
212There is no independent control for individual protocols, any command
213run for one of tcp4|udp4|ah4|esp4|sctp4|tcp6|udp6|ah6|esp6|sctp6 is
214going to control the rx-flow-hashing for all protocols on that interface.
215
216Besides using the FMan Keygen computed hash for spreading traffic on the
217128 Rx FQs, the DPAA Ethernet driver also sets the skb hash value when
218the NETIF_F_RXHASH feature is on (active by default). This can be turned
219on or off through ethtool, i.e.:
220
221	# ethtool -K fm1-mac9 rx-hashing off
222	# ethtool -k fm1-mac9 | grep hash
223	receive-hashing: off
224	# ethtool -K fm1-mac9 rx-hashing on
225	Actual changes:
226	receive-hashing: on
227	# ethtool -k fm1-mac9 | grep hash
228	receive-hashing: on
229
230Please note that Rx hashing depends upon the rx-flow-hashing being on
231for that interface - turning off rx-flow-hashing will also disable the
232rx-hashing (without ethtool reporting it as off as that depends on the
233NETIF_F_RXHASH feature flag).
234
235Debugging
236=========
237
238The following statistics are exported for each interface through ethtool:
239
240	- interrupt count per CPU
241	- Rx packets count per CPU
242	- Tx packets count per CPU
243	- Tx confirmed packets count per CPU
244	- Tx S/G frames count per CPU
245	- Tx error count per CPU
246	- Rx error count per CPU
247	- Rx error count per type
248	- congestion related statistics:
249		- congestion status
250		- time spent in congestion
251		- number of time the device entered congestion
252		- dropped packets count per cause
253
254The driver also exports the following information in sysfs:
255
256	- the FQ IDs for each FQ type
257	/sys/devices/platform/dpaa-ethernet.0/net/<int>/fqids
258
259	- the IDs of the buffer pools in use
260	/sys/devices/platform/dpaa-ethernet.0/net/<int>/bpids
261