bbdev
play

(bbdev) Amr Mokhtar DPDK Summit Userspace - Dublin- 2017 why - PowerPoint PPT Presentation

Jan 31, 2024 •983 likes •1.15k views

Wireless Base Band Device (bbdev) Amr Mokhtar DPDK Summit Userspace - Dublin- 2017 why baseband..? MAC Tx Data Downlink * Reference: 3GPP TS 36.211 & 36.212 architecture Common programing framework for Application wireless

  1. Wireless Base Band Device (bbdev) Amr Mokhtar DPDK Summit Userspace - Dublin- 2017

  2. why baseband..? MAC Tx Data Downlink * Reference: 3GPP TS 36.211 & 36.212

  3. architecture  Common programing framework for Application wireless workloads Application facing API DPDK librte_bbdev  Seamless HW/SW abstraction interface for Driver facing API underlying operations bbdev HW driver bbdev SW driver bbdev SW driver bbdev SW driver  Pluggable driver support for various stages SW Modulation SW Turbo lib SW FFT/iFFT lib Mapper lib of wireless packet processing (new driver Software Hardware registers itself and reports its capabilities) FPGA or Fixed Function Accel.

  4. workflow Device stopped rte_eal_init() rte_bbdev_count() rte_bbdev_info_get() Device Configuration rte_bbdev_close() rte_bbdev_configure() rte_bbdev_configure() Device identified Device stopped Device configured rte_bbdev_stop() Queues configured rte_bbdev_start() rte_bbdev_queue_configure() rte_bbdev_enqueue_ops() Device running rte_bbdev_start() rte_bbdev_dequeue_ops()

  5. lookaside model - hardware Application calls the API to submit an offload request to the user-space 1. Enqueue Dequeue Thread device driver Thread Driver forms the descriptor in ring in memory, including pointers to data 2. 1. enqueue_ops() 8. dequeue_ops() buffers 4. Ret 10. Ret The driver enqueues the descriptor by writing to the relevant MMIO 3. Register 2. Form hw request 9. Consume response The driver returns from the API call back to the application thread descriptor(s) 4. descriptor(s), if available HW DMA reads the descriptor(s) created in step 2, and input data 5. 3. Enqueue request buffers descriptor(s) (MMIO write) N SW Descriptor … 1 Rings HW performs the operation(s) 6. 0 Once complete the HW will DMA write the output buffers and overwrite 7. DPDK the descriptor(s) indicating to SW that this request is complete. Software Application calls to API to check for completed requests (dequeue) 8. Hardware 5. DMA read descriptor, 7. DMA write results back and input data and update descriptor Driver checks if response descriptors have been written back 9. Driver returns results to application if descriptors have been written 10. back, or empty response if not. Turbo Dec/Enc 6. Perform Operation(s) * Enqueue thread and dequeue thread may be the same

  6. lookaside model - software Application calls the API to submit an offload request to the user-space 1. Enqueue Dequeue Thread Thread device driver Driver forms its internal structures and perform operation(s) 2. 1. enqueue_ops() 5. dequeue_ops() sequentially. 7. Ret 4. Ret The driver enqueues the outcomes to internal software rings. 3. 2. Perform Operation(s) Turbo Dec/Enc The driver returns from the API call back to the application thread. 4. 6. Consume from sw ring, if available 3. Produce to sw ring Application calls to API to check for completed requests (dequeue). 5. N Driver checks if some results were produced on the tip of the ring, then 6. Queue SW Rings … 1 pull it out. 0 Driver returns the pulled out results to application if there were any 7. DPDK available, or empty response if not. Software Hardware * Enqueue thread and dequeue thread may be the same

  7. Note on mbuf* usage in bbdev mbuf seg#2 mbuf seg#3 mbuf seg#1 mbuf seg#n ... CB #1 CB #2 CB #3 CB #n op_data->offset op_data->length Transport Block (TB) /** Data input and output buffer for Turbo operations */ struct rte_bbdev_op_data { struct rte_mbuf *data; /**< First mbuf segment with input/output data. Each segment represents * one Code Block. */ uint32_t offset; /**< The starting point for the Turbo input/output, in bytes, from the * start of the first segment's data buffer. It must be smaller than the * first segment's data_len! */ uint32_t length; /**< Length of Transport Block - number of bytes for Turbo Encode/Decode * operation for input; length of the output for output operation. */ }; * This mbuf formality is experimental and subject to change

  8. bbdevAPIs  Device Management APIs  Queue Management APIs  Operation Management APIs  Interrupts Support APIs  Statistics APIs

  9. bbdevAPIs >>  Device creation is based on the same principles as DPDK cryptodev and ethdev.  Register driver configuration structure with DPDK EAL using the existing RTE_PMD_REGISTER_PCI macro.  Physical devices are identified by PCI ID during the EAL PCI scan and allocated a unique device identifier.  Device initiation is also along the same principles as DPDK cryptodev and ethdev.  Devices are first configured  int rte_bbdev_configure (uint8_t dev_id, uint16_t num_queues, const struct rte_bbdev_conf *conf);  Devices queues are then configured before the device is started and used.  int rte_bbdev_queue_configure (uint8_t dev_id, uint16_t queue_id, const struct rte_bbdev_queue_conf *conf)

  10. bbdevAPIs – Device Management uint8_t rte_bbdev_count (void); bool rte_bbdev_is_valid (uint8_t dev_id); uint8_t rte_bbdev_next (uint8_t dev_id); int rte_bbdev_configure (uint8_t dev_id, uint16_t num_queues, const struct rte_bbdev_conf *conf); int rte_bbdev_info_get (uint8_t dev_id, struct rte_bbdev_info *dev_info); int rte_bbdev_start (uint8_t dev_id); int rte_bbdev_stop (uint8_t dev_id); int rte_bbdev_close (uint8_t dev_id);

  11. bbdevAPIs – Queue Management int rte_bbdev_queue_configure (uint8_t dev_id, uint16_t queue_id, const struct rte_bbdev_queue_conf *conf); /** Different operation types supported by the device */ enum rte_bbdev_op_type { int rte_bbdev_queue_start (uint8_t dev_id, uint16_t queue_id); RTE_BBDEV_OP_NONE , /**< Dummy operation that does nothing */ RTE_BBDEV_OP_TURBO_DEC , /**< Turbo decode */ RTE_BBDEV_OP_TURBO_ENC , /**< Turbo encode */ RTE_BBDEV_OP_TYPE_COUNT , /**< Count of different op types */ }; int rte_bbdev_queue_stop (uint8_t dev_id, uint16_t queue_id); int rte_bbdev_queue_info_get (uint8_t dev_id, uint16_t queue_id, struct rte_bbdev_queue_info *dev_info);

  12. DPDK BBDEV APIs – Operation Management static inline uint16_t rte_bbdev_enqueue_ops (uint8_t dev_id, uint16_t queue_id, struct rte_bbdev_op **ops, uint16_t num_ops) static inline uint16_t rte_bbdev_dequeue_ops (uint8_t dev_id, uint16_t queue_id, struct rte_bbdev_op **ops, uint16_t num_ops) /** Structure specifying a single operation */ struct rte_bbdev_op { enum rte_bbdev_op_type type; /**< Type of this operation */ int status; /**< Status of operation that was performed */ struct rte_mempool *mempool; /**< Mempool which op instance is in */ void *opaque_data; /**< Opaque pointer for user data */ union { struct rte_bbdev_op_turbo_dec *turbo_dec; struct rte_bbdev_op_turbo_enc *turbo_enc; }; };

  13. bbdevAPIs – Interrupt Support int rte_bbdev_callback_register (uint8_t dev_id, enum rte_bbdev_event_type event, rte_bbdev_cb_fn cb_fn, void *cb_arg); int rte_bbdev_callback_unregister (uint8_t dev_id, enum rte_bbdev_event_type event, rte_bbdev_cb_fn cb_fn, void *cb_arg); int rte_bbdev_queue_intr_enable (uint8_t dev_id, uint16_t queue_id); int rte_bbdev_queue_intr_disable (uint8_t dev_id, uint16_t queue_id); int rte_bbdev_queue_intr_ctl (uint8_t dev_id, uint16_t queue_id, int epfd, int op, void *data);

  14. bbdevAPIs – Statistics int rte_bbdev_stats_get (uint8_t dev_id, struct rte_bbdev_stats *stats); int rte_bbdev_stats_reset (uint8_t dev_id); int rte_bbdev_info_get (uint8_t dev_id, struct rte_bbdev_info *dev_info);

  15. Amr Mokhtar Questions? amr.mokhtar@intel.com

Recommend

More recommend