1 .. SPDX-License-Identifier: BSD-3-Clause
2 Copyright(c) 2017 Intel Corporation.
4 Cryptodev Scheduler Poll Mode Driver Library
5 ============================================
7 Scheduler PMD is a software crypto PMD, which has the capabilities of
8 attaching hardware and/or software cryptodevs, and distributes ingress
9 crypto ops among them in a certain manner.
11 .. figure:: img/scheduler-overview.*
13 Cryptodev Scheduler Overview
16 The Cryptodev Scheduler PMD library (**librte_crypto_scheduler**) acts as
17 a software crypto PMD and shares the same API provided by librte_cryptodev.
18 The PMD supports attaching multiple crypto PMDs, software or hardware, as
19 workers, and distributes the crypto workload to them with certain behavior.
20 The behaviors are categorizes as different "modes". Basically, a scheduling
21 mode defines certain actions for scheduling crypto ops to its workers.
23 The librte_crypto_scheduler library exports a C API which provides an API
24 for attaching/detaching workers, set/get scheduling modes, and enable/disable
25 crypto ops reordering.
30 * Sessionless crypto operation is not supported
31 * OOP crypto operation is not supported when the crypto op reordering feature
38 To use the PMD in an application, user must:
40 * Call rte_vdev_init("crypto_scheduler") within the application.
42 * Use --vdev="crypto_scheduler" in the EAL options, which will call
43 rte_vdev_init() internally.
46 The following parameters (all optional) can be provided in the previous
49 * socket_id: Specify the socket where the memory for the device is going
50 to be allocated (by default, socket_id will be the socket where the core
51 that is creating the PMD is running on).
53 * max_nb_sessions: Specify the maximum number of sessions that can be
54 created. This value may be overwritten internally if there are too
55 many devices are attached.
57 * worker: If a cryptodev has been initialized with specific name, it can be
58 attached to the scheduler using this parameter, simply filling the name
59 here. Multiple cryptodevs can be attached initially by presenting this
60 parameter multiple times.
62 * mode: Specify the scheduling mode of the PMD. The supported scheduling
63 mode parameter values are specified in the "Cryptodev Scheduler Modes
66 * mode_param: Specify the mode-specific parameter. Some scheduling modes
67 may be initialized with specific parameters other than the default ones,
68 such as the **threshold** packet size of **packet-size-distr** mode. This
69 parameter fulfills the purpose.
71 * ordering: Specify the status of the crypto operations ordering feature.
72 The value of this parameter can be "enable" or "disable". This feature
73 is disabled by default.
77 .. code-block:: console
79 ... --vdev "crypto_aesni_mb0,name=aesni_mb_1" --vdev "crypto_aesni_mb1,name=aesni_mb_2" --vdev "crypto_scheduler,worker=aesni_mb_1,worker=aesni_mb_2" ...
83 * The scheduler cryptodev cannot be started unless the scheduling mode
84 is set and at least one worker is attached. Also, to configure the
85 scheduler in the run-time, like attach/detach worker(s), change
86 scheduling mode, or enable/disable crypto op ordering, one should stop
87 the scheduler first, otherwise an error will be returned.
89 * The crypto op reordering feature requires using the userdata field of
90 every mbuf to be processed to store temporary data. By the end of
91 processing, the field is set to pointing to NULL, any previously
92 stored value of this field will be lost.
95 Cryptodev Scheduler Modes Overview
96 ----------------------------------
98 Currently the Crypto Scheduler PMD library supports following modes of
101 * **CDEV_SCHED_MODE_ROUNDROBIN:**
103 *Initialization mode parameter*: **round-robin**
105 Round-robin mode, which distributes the enqueued burst of crypto ops
106 among its workers in a round-robin manner. This mode may help to fill
107 the throughput gap between the physical core and the existing cryptodevs
108 to increase the overall performance.
110 * **CDEV_SCHED_MODE_PKT_SIZE_DISTR:**
112 *Initialization mode parameter*: **packet-size-distr**
114 Packet-size based distribution mode, which works with 2 workers, the primary
115 worker and the secondary worker, and distributes the enqueued crypto
116 operations to them based on their data lengths. A crypto operation will be
117 distributed to the primary worker if its data length is equal to or bigger
118 than the designated threshold, otherwise it will be handled by the secondary
121 A typical usecase in this mode is with the QAT cryptodev as the primary and
122 a software cryptodev as the secondary worker. This may help applications to
123 process additional crypto workload than what the QAT cryptodev can handle on
124 its own, by making use of the available CPU cycles to deal with smaller
127 The threshold is set to 128 bytes by default. It can be updated by calling
128 function **rte_cryptodev_scheduler_option_set**. The parameter of
129 **option_type** must be **CDEV_SCHED_OPTION_THRESHOLD** and **option** should
130 point to a rte_cryptodev_scheduler_threshold_option structure filled with
131 appropriate threshold value. Please NOTE this threshold has be a power-of-2
132 unsigned integer. It is possible to use **mode_param** initialization
133 parameter to achieve the same purpose. For example:
135 ... --vdev "crypto_scheduler,mode=packet-size-distr,mode_param=threshold:512" ...
137 The above parameter will overwrite the threshold value to 512.
139 * **CDEV_SCHED_MODE_FAILOVER:**
141 *Initialization mode parameter*: **fail-over**
143 Fail-over mode, which works with 2 workers, the primary worker and the
144 secondary worker. In this mode, the scheduler will enqueue the incoming
145 crypto operation burst to the primary worker. When one or more crypto
146 operations fail to be enqueued, then they will be enqueued to the secondary
149 * **CDEV_SCHED_MODE_MULTICORE:**
151 *Initialization mode parameter*: **multi-core**
153 Multi-core mode, which distributes the workload with several (up to eight)
154 worker cores. The enqueued bursts are distributed among the worker cores in a
155 round-robin manner. If scheduler cannot enqueue entire burst to the same worker,
156 it will enqueue the remaining operations to the next available worker.
157 For pure small packet size (64 bytes) traffic however the multi-core mode is not
158 an optimal solution, as it doesn't give significant per-core performance improvement.
159 For mixed traffic (IMIX) the optimal number of worker cores is around 2-3.
160 For large packets (1.5 kbytes) scheduler shows linear scaling in performance
162 Each worker uses its own cryptodev. Only software cryptodevs
163 are supported. Only the same type of cryptodevs should be used concurrently.
165 The multi-core mode uses one extra parameter:
167 * corelist: Semicolon-separated list of logical cores to be used as workers.
168 The number of worker cores should be equal to the number of worker cryptodevs.
169 These cores should be present in EAL core list parameter and
170 should not be used by the application or any other process.
173 ... --vdev "crypto_aesni_mb1,name=aesni_mb_1" --vdev "crypto_aesni_mb_pmd2,name=aesni_mb_2" \
174 --vdev "crypto_scheduler,worker=aesni_mb_1,worker=aesni_mb_2,mode=multi-core,corelist=23;24" ...