1 .. SPDX-License-Identifier: BSD-3-Clause
2 Copyright(c) 2020 Intel Corporation.
4 Driver for the Intel® Dynamic Load Balancer (DLB)
5 =================================================
7 The DPDK dlb poll mode driver supports the Intel® Dynamic Load Balancer.
12 Follow the DPDK :ref:`Getting Started Guide for Linux <linux_gsg>` to setup
13 the basic DPDK environment.
18 The DLB PF PMD is a user-space PMD that uses VFIO to gain direct
19 device access. To use this operation mode, the PCIe PF device must be bound
20 to a DPDK-compatible VFIO driver, such as vfio-pci.
25 The DLB provides the functions of a DPDK event device; specifically, it
26 supports atomic, ordered, and parallel scheduling events from queues to ports.
27 However, the DLB hardware is not a perfect match to the eventdev API. Some DLB
28 features are abstracted by the PMD (e.g. directed ports), some are only
29 accessible as vdev command-line parameters, and certain eventdev features are
30 not supported (e.g. the event flow ID is not maintained during scheduling).
32 In general the dlb PMD is designed for ease-of-use and does not require a
33 detailed understanding of the hardware, but these details are important when
34 writing high-performance code. This section describes the places where the
35 eventdev API and DLB misalign.
37 Scheduling Domain Configuration
38 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
40 There are 32 scheduling domainis the DLB.
41 When one is configured, it allocates load-balanced and
42 directed queues, ports, credits, and other hardware resources. Some
43 resource allocations are user-controlled -- the number of queues, for example
44 -- and others, like credit pools (one directed and one load-balanced pool per
45 scheduling domain), are not.
47 The DLB is a closed system eventdev, and as such the ``nb_events_limit`` device
48 setup argument and the per-port ``new_event_threshold`` argument apply as
49 defined in the eventdev header file. The limit is applied to all enqueues,
50 regardless of whether it will consume a directed or load-balanced credit.
55 The Eventdev API allows one to reconfigure a device, its ports, and its queues
56 by first stopping the device, calling the configuration function(s), then
57 restarting the device. The DLB does not support configuring an individual queue
58 or port without first reconfiguring the entire device, however, so there are
59 certain reconfiguration sequences that are valid in the eventdev API but not
62 Specifically, the PMD supports the following configuration sequence:
63 1. Configure and start the device
65 3. (Optional) Reconfigure the device
66 4. (Optional) If step 3 is run:
68 a. Setup queue(s). The reconfigured queue(s) lose their previous port links.
69 b. The reconfigured port(s) lose their previous queue links.
71 5. (Optional, only if steps 4a and 4b are run) Link port(s) to queue(s)
72 6. Restart the device. If the device is reconfigured in step 3 but one or more
73 of its ports or queues are not, the PMD will apply their previous
74 configuration (including port->queue links) at this time.
76 The PMD does not support the following configuration sequences:
77 1. Configure and start the device
79 3. Setup queue or setup port
82 This sequence is not supported because the event device must be reconfigured
83 before its ports or queues can be.
88 A load-balanced queue can support atomic and ordered scheduling, or atomic and
89 unordered scheduling, but not atomic and unordered and ordered scheduling. A
90 queue's scheduling types are controlled by the event queue configuration.
92 If the user sets the ``RTE_EVENT_QUEUE_CFG_ALL_TYPES`` flag, the
93 ``nb_atomic_order_sequences`` determines the supported scheduling types.
94 With non-zero ``nb_atomic_order_sequences``, the queue is configured for atomic
95 and ordered scheduling. In this case, ``RTE_SCHED_TYPE_PARALLEL`` scheduling is
96 supported by scheduling those events as ordered events. Note that when the
97 event is dequeued, its sched_type will be ``RTE_SCHED_TYPE_ORDERED``. Else if
98 ``nb_atomic_order_sequences`` is zero, the queue is configured for atomic and
99 unordered scheduling. In this case, ``RTE_SCHED_TYPE_ORDERED`` is unsupported.
101 If the ``RTE_EVENT_QUEUE_CFG_ALL_TYPES`` flag is not set, schedule_type
102 dictates the queue's scheduling type.
104 The ``nb_atomic_order_sequences`` queue configuration field sets the ordered
105 queue's reorder buffer size. DLB has 4 groups of ordered queues, where each
106 group is configured to contain either 1 queue with 1024 reorder entries, 2
107 queues with 512 reorder entries, and so on down to 32 queues with 32 entries.
109 When a load-balanced queue is created, the PMD will configure a new sequence
110 number group on-demand if num_sequence_numbers does not match a pre-existing
111 group with available reorder buffer entries. If all sequence number groups are
112 in use, no new group will be created and queue configuration will fail. (Note
113 that when the PMD is used with a virtual DLB device, it cannot change the
114 sequence number configuration.)
116 The queue's ``nb_atomic_flows`` parameter is ignored by the DLB PMD, because
117 the DLB does not limit the number of flows a queue can track. In the DLB, all
118 load-balanced queues can use the full 16-bit flow ID range.
120 Load-balanced and Directed Ports
121 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
123 DLB ports come in two flavors: load-balanced and directed. The eventdev API
124 does not have the same concept, but it has a similar one: ports and queues that
125 are singly-linked (i.e. linked to a single queue or port, respectively).
127 The ``rte_event_dev_info_get()`` function reports the number of available
128 event ports and queues (among other things). For the DLB PMD, max_event_ports
129 and max_event_queues report the number of available load-balanced ports and
130 queues, and max_single_link_event_port_queue_pairs reports the number of
131 available directed ports and queues.
133 When a scheduling domain is created in ``rte_event_dev_configure()``, the user
134 specifies ``nb_event_ports`` and ``nb_single_link_event_port_queues``, which
135 control the total number of ports (load-balanced and directed) and the number
136 of directed ports. Hence, the number of requested load-balanced ports is
137 ``nb_event_ports - nb_single_link_event_ports``. The ``nb_event_queues`` field
138 specifies the total number of queues (load-balanced and directed). The number
139 of directed queues comes from ``nb_single_link_event_port_queues``, since
140 directed ports and queues come in pairs.
142 When a port is setup, the ``RTE_EVENT_PORT_CFG_SINGLE_LINK`` flag determines
143 whether it should be configured as a directed (the flag is set) or a
144 load-balanced (the flag is unset) port. Similarly, the
145 ``RTE_EVENT_QUEUE_CFG_SINGLE_LINK`` queue configuration flag controls
146 whether it is a directed or load-balanced queue.
148 Load-balanced ports can only be linked to load-balanced queues, and directed
149 ports can only be linked to directed queues. Furthermore, directed ports can
150 only be linked to a single directed queue (and vice versa), and that link
151 cannot change after the eventdev is started.
153 The eventdev API does not have a directed scheduling type. To support directed
154 traffic, the dlb PMD detects when an event is being sent to a directed queue
155 and overrides its scheduling type. Note that the originally selected scheduling
156 type (atomic, ordered, or parallel) is not preserved, and an event's sched_type
157 will be set to ``RTE_SCHED_TYPE_ATOMIC`` when it is dequeued from a directed
163 The flow ID field is not preserved in the event when it is scheduled in the
164 DLB, because the DLB hardware control word format does not have sufficient
165 space to preserve every event field. As a result, the flow ID specified with
166 the enqueued event will not be in the dequeued event. If this field is
167 required, the application should pass it through an out-of-band path (for
168 example in the mbuf's udata64 field, if the event points to an mbuf) or
169 reconstruct the flow ID after receiving the event.
171 Also, the DLB hardware control word supports a 16-bit flow ID. Since struct
172 rte_event's flow_id field is 20 bits, the DLB PMD drops the most significant
173 four bits from the event's flow ID.
178 DLB uses a hardware credit scheme to prevent software from overflowing hardware
179 event storage, with each unit of storage represented by a credit. A port spends
180 a credit to enqueue an event, and hardware refills the ports with credits as the
181 events are scheduled to ports. Refills come from credit pools, and each port is
182 a member of a load-balanced credit pool and a directed credit pool. The
183 load-balanced credits are used to enqueue to load-balanced queues, and directed
184 credits are used for directed queues.
186 A DLB eventdev contains one load-balanced and one directed credit pool. These
187 pools' sizes are controlled by the nb_events_limit field in struct
188 rte_event_dev_config. The load-balanced pool is sized to contain
189 nb_events_limit credits, and the directed pool is sized to contain
190 nb_events_limit/4 credits. The directed pool size can be overridden with the
191 num_dir_credits vdev argument, like so:
193 .. code-block:: console
195 --vdev=dlb1_event,num_dir_credits=<value>
197 This can be used if the default allocation is too low or too high for the
198 specific application needs. The PMD also supports a vdev arg that limits the
199 max_num_events reported by rte_event_dev_info_get():
201 .. code-block:: console
203 --vdev=dlb1_event,max_num_events=<value>
205 By default, max_num_events is reported as the total available load-balanced
206 credits. If multiple DLB-based applications are being used, it may be desirable
207 to control how many load-balanced credits each application uses, particularly
208 when application(s) are written to configure nb_events_limit equal to the
209 reported max_num_events.
211 Each port is a member of both credit pools. A port's credit allocation is
212 defined by its low watermark, high watermark, and refill quanta. These three
213 parameters are calculated by the dlb PMD like so:
215 - The load-balanced high watermark is set to the port's enqueue_depth.
216 The directed high watermark is set to the minimum of the enqueue_depth and
217 the directed pool size divided by the total number of ports.
218 - The refill quanta is set to half the high watermark.
219 - The low watermark is set to the minimum of 16 and the refill quanta.
221 When the eventdev is started, each port is pre-allocated a high watermark's
222 worth of credits. For example, if an eventdev contains four ports with enqueue
223 depths of 32 and a load-balanced credit pool size of 4096, each port will start
224 with 32 load-balanced credits, and there will be 3968 credits available to
225 replenish the ports. Thus, a single port is not capable of enqueueing up to the
226 nb_events_limit (without any events being dequeued), since the other ports are
227 retaining their initial credit allocation; in short, all ports must enqueue in
228 order to reach the limit.
230 If a port attempts to enqueue and has no credits available, the enqueue
231 operation will fail and the application must retry the enqueue. Credits are
232 replenished asynchronously by the DLB hardware.
237 The DLB is a "closed system" event dev, and the DLB PMD layers a software
238 credit scheme on top of the hardware credit scheme in order to comply with
239 the per-port backpressure described in the eventdev API.
241 The DLB's hardware scheme is local to a queue/pipeline stage: a port spends a
242 credit when it enqueues to a queue, and credits are later replenished after the
243 events are dequeued and released.
245 In the software credit scheme, a credit is consumed when a new (.op =
246 RTE_EVENT_OP_NEW) event is injected into the system, and the credit is
247 replenished when the event is released from the system (either explicitly with
248 RTE_EVENT_OP_RELEASE or implicitly in dequeue_burst()).
250 In this model, an event is "in the system" from its first enqueue into eventdev
251 until it is last dequeued. If the event goes through multiple event queues, it
252 is still considered "in the system" while a worker thread is processing it.
254 A port will fail to enqueue if the number of events in the system exceeds its
255 ``new_event_threshold`` (specified at port setup time). A port will also fail
256 to enqueue if it lacks enough hardware credits to enqueue; load-balanced
257 credits are used to enqueue to a load-balanced queue, and directed credits are
258 used to enqueue to a directed queue.
260 The out-of-credit situations are typically transient, and an eventdev
261 application using the DLB ought to retry its enqueues if they fail.
262 If enqueue fails, DLB PMD sets rte_errno as follows:
264 - -ENOSPC: Credit exhaustion (either hardware or software)
265 - -EINVAL: Invalid argument, such as port ID, queue ID, or sched_type.
267 Depending on the pipeline the application has constructed, it's possible to
268 enter a credit deadlock scenario wherein the worker thread lacks the credit
269 to enqueue an event, and it must dequeue an event before it can recover the
270 credit. If the worker thread retries its enqueue indefinitely, it will not
271 make forward progress. Such deadlock is possible if the application has event
272 "loops", in which an event in dequeued from queue A and later enqueued back to
275 Due to this, workers should stop retrying after a time, release the events it
276 is attempting to enqueue, and dequeue more events. It is important that the
277 worker release the events and don't simply set them aside to retry the enqueue
278 again later, because the port has limited history list size (by default, twice
279 the port's dequeue_depth).
284 The DLB supports event priority and per-port queue service priority, as
285 described in the eventdev header file. The DLB does not support 'global' event
286 queue priority established at queue creation time.
288 DLB supports 8 event and queue service priority levels. For both priority
289 types, the PMD uses the upper three bits of the priority field to determine the
290 DLB priority, discarding the 5 least significant bits. The 5 least significant
291 event priority bits are not preserved when an event is enqueued.
293 Atomic Inflights Allocation
294 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
296 In the last stage prior to scheduling an atomic event to a CQ, DLB holds the
297 inflight event in a temporary buffer that is divided among load-balanced
298 queues. If a queue's atomic buffer storage fills up, this can result in
299 head-of-line-blocking. For example:
301 - An LDB queue allocated N atomic buffer entries
302 - All N entries are filled with events from flow X, which is pinned to CQ 0.
304 Until CQ 0 releases 1+ events, no other atomic flows for that LDB queue can be
305 scheduled. The likelihood of this case depends on the eventdev configuration,
306 traffic behavior, event processing latency, potential for a worker to be
307 interrupted or otherwise delayed, etc.
309 By default, the PMD allocates 16 buffer entries for each load-balanced queue,
310 which provides an even division across all 128 queues but potentially wastes
311 buffer space (e.g. if not all queues are used, or aren't used for atomic
314 The PMD provides a dev arg to override the default per-queue allocation. To
315 increase a vdev's per-queue atomic-inflight allocation to (for example) 64:
317 .. code-block:: console
319 --vdev=dlb1_event,atm_inflights=64
324 The DLB PMD's default behavior for managing a CQ is to "pop" the CQ once per
325 dequeued event before returning from rte_event_dequeue_burst(). This frees the
326 corresponding entries in the CQ, which enables the DLB to schedule more events
329 To support applications seeking finer-grained scheduling control -- for example
330 deferring scheduling to get the best possible priority scheduling and
331 load-balancing -- the PMD supports a deferred scheduling mode. In this mode,
332 the CQ entry is not popped until the *subsequent* rte_event_dequeue_burst()
333 call. This mode only applies to load-balanced event ports with dequeue depth of
336 To enable deferred scheduling, use the defer_sched vdev argument like so:
338 .. code-block:: console
340 --vdev=dlb1_event,defer_sched=on
git.droids-corp.org / dpdk.git / blob