[dpdk-dev] [PATCH v9 03/10] baseband/acc100: add info get function

[Chautru, Nicolas]

Hi Tom, 

> From: Tom Rix <trix at redhat.com>
> On 9/28/20 5:29 PM, Nicolas Chautru wrote:
> > Add in the "info_get" function to the driver, to allow us to query the
> > device.
> > No processing capability are available yet.
> > Linking bbdev-test to support the PMD with null capability.
> >
> > Signed-off-by: Nicolas Chautru <nicolas.chautru at intel.com>
> > Acked-by: Liu Tianjiao <Tianjiao.liu at intel.com>
> > ---
> >  app/test-bbdev/meson.build               |   3 +
> >  drivers/baseband/acc100/rte_acc100_cfg.h |  96 +++++++++++++
> > drivers/baseband/acc100/rte_acc100_pmd.c | 225
> +++++++++++++++++++++++++++++++
> >  drivers/baseband/acc100/rte_acc100_pmd.h |   3 +
> >  4 files changed, 327 insertions(+)
> >  create mode 100644 drivers/baseband/acc100/rte_acc100_cfg.h
> >
> > diff --git a/app/test-bbdev/meson.build b/app/test-bbdev/meson.build
> > index 18ab6a8..fbd8ae3 100644
> > --- a/app/test-bbdev/meson.build
> > +++ b/app/test-bbdev/meson.build
> > @@ -12,3 +12,6 @@ endif
> >  if dpdk_conf.has('RTE_LIBRTE_PMD_BBDEV_FPGA_5GNR_FEC')
> >  	deps += ['pmd_bbdev_fpga_5gnr_fec']
> >  endif
> > +if dpdk_conf.has('RTE_LIBRTE_PMD_BBDEV_ACC100')
> > +	deps += ['pmd_bbdev_acc100']
> > +endif
> > \ No newline at end of file
> > diff --git a/drivers/baseband/acc100/rte_acc100_cfg.h
> > b/drivers/baseband/acc100/rte_acc100_cfg.h
> > new file mode 100644
> > index 0000000..73bbe36
> > --- /dev/null
> > +++ b/drivers/baseband/acc100/rte_acc100_cfg.h
> > @@ -0,0 +1,96 @@
> > +/* SPDX-License-Identifier: BSD-3-Clause
> > + * Copyright(c) 2020 Intel Corporation  */
> > +
> > +#ifndef _RTE_ACC100_CFG_H_
> > +#define _RTE_ACC100_CFG_H_
> > +
> > +/**
> > + * @file rte_acc100_cfg.h
> > + *
> > + * Functions for configuring ACC100 HW, exposed directly to applications.
> > + * Configuration related to encoding/decoding is done through the
> > + * librte_bbdev library.
> > + *
> > + * @warning
> > + * @b EXPERIMENTAL: this API may change without prior notice
> When will this experimental tag be removed ?

I have pushed a patch to remove it. But the feedback from some of the community was to wait a bit more until next year.

> > + */
> > +
> > +#include <stdint.h>
> > +#include <stdbool.h>
> > +
> > +#ifdef __cplusplus
> > +extern "C" {
> > +#endif
> > +/**< Number of Virtual Functions ACC100 supports */ #define
> > +RTE_ACC100_NUM_VFS 16
> This is already defined with ACC100_NUM_VFS

Thanks. 

> > +
> > +/**
> > + * Definition of Queue Topology for ACC100 Configuration
> > + * Some level of details is abstracted out to expose a clean
> > +interface
> > + * given that comprehensive flexibility is not required  */ struct
> > +rte_q_topology_t {
> > +	/** Number of QGroups in incremental order of priority */
> > +	uint16_t num_qgroups;
> > +	/**
> > +	 * All QGroups have the same number of AQs here.
> > +	 * Note : Could be made a 16-array if more flexibility is really
> > +	 * required
> > +	 */
> > +	uint16_t num_aqs_per_groups;
> > +	/**
> > +	 * Depth of the AQs is the same of all QGroups here. Log2 Enum : 2^N
> > +	 * Note : Could be made a 16-array if more flexibility is really
> > +	 * required
> > +	 */
> > +	uint16_t aq_depth_log2;
> > +	/**
> > +	 * Index of the first Queue Group Index - assuming contiguity
> > +	 * Initialized as -1
> > +	 */
> > +	int8_t first_qgroup_index;
> > +};
> > +
> > +/**
> > + * Definition of Arbitration related parameters for ACC100
> > +Configuration  */ struct rte_arbitration_t {
> > +	/** Default Weight for VF Fairness Arbitration */
> > +	uint16_t round_robin_weight;
> > +	uint32_t gbr_threshold1; /**< Guaranteed Bitrate Threshold 1 */
> > +	uint32_t gbr_threshold2; /**< Guaranteed Bitrate Threshold 2 */ };
> > +
> > +/**
> > + * Structure to pass ACC100 configuration.
> > + * Note: all VF Bundles will have the same configuration.
> > + */
> > +struct acc100_conf {
> > +	bool pf_mode_en; /**< 1 if PF is used for dataplane, 0 for VFs */
> > +	/** 1 if input '1' bit is represented by a positive LLR value, 0 if '1'
> > +	 * bit is represented by a negative value.
> > +	 */
> > +	bool input_pos_llr_1_bit;
> > +	/** 1 if output '1' bit is represented by a positive value, 0 if '1'
> > +	 * bit is represented by a negative value.
> > +	 */
> > +	bool output_pos_llr_1_bit;
> > +	uint16_t num_vf_bundles; /**< Number of VF bundles to setup */
> > +	/** Queue topology for each operation type */
> > +	struct rte_q_topology_t q_ul_4g;
> > +	struct rte_q_topology_t q_dl_4g;
> > +	struct rte_q_topology_t q_ul_5g;
> > +	struct rte_q_topology_t q_dl_5g;
> > +	/** Arbitration configuration for each operation type */
> > +	struct rte_arbitration_t arb_ul_4g[RTE_ACC100_NUM_VFS];
> > +	struct rte_arbitration_t arb_dl_4g[RTE_ACC100_NUM_VFS];
> > +	struct rte_arbitration_t arb_ul_5g[RTE_ACC100_NUM_VFS];
> > +	struct rte_arbitration_t arb_dl_5g[RTE_ACC100_NUM_VFS]; };
> > +
> > +#ifdef __cplusplus
> > +}
> > +#endif
> > +
> > +#endif /* _RTE_ACC100_CFG_H_ */
> > diff --git a/drivers/baseband/acc100/rte_acc100_pmd.c
> > b/drivers/baseband/acc100/rte_acc100_pmd.c
> > index 1b4cd13..7807a30 100644
> > --- a/drivers/baseband/acc100/rte_acc100_pmd.c
> > +++ b/drivers/baseband/acc100/rte_acc100_pmd.c
> > @@ -26,6 +26,184 @@
> >  RTE_LOG_REGISTER(acc100_logtype, pmd.bb.acc100, NOTICE);  #endif
> >
> > +/* Read a register of a ACC100 device */ static inline uint32_t
> > +acc100_reg_read(struct acc100_device *d, uint32_t offset) {
> > +
> > +	void *reg_addr = RTE_PTR_ADD(d->mmio_base, offset);
> > +	uint32_t ret = *((volatile uint32_t *)(reg_addr));
> > +	return rte_le_to_cpu_32(ret);
> > +}
> > +
> > +/* Calculate the offset of the enqueue register */ static inline
> > +uint32_t queue_offset(bool pf_device, uint8_t vf_id, uint8_t qgrp_id,
> > +uint16_t aq_id) {
> > +	if (pf_device)
> > +		return ((vf_id << 12) + (qgrp_id << 7) + (aq_id << 3) +
> > +				HWPfQmgrIngressAq);
> > +	else
> > +		return ((qgrp_id << 7) + (aq_id << 3) +
> > +				HWVfQmgrIngressAq);
> Could you add *QmrIngressAq to the acc100_registry_addr and skip the if
> (pf_device) check ?

I am not convinced. That acc100_registry_addr is not kept with the driver, you
still need to check pf_device flag to now what registers are to be used.


> > +}
> > +
> > +enum {UL_4G = 0, UL_5G, DL_4G, DL_5G, NUM_ACC};
> > +
> > +/* Return the queue topology for a Queue Group Index */ static inline
> > +void qtopFromAcc(struct rte_q_topology_t **qtop, int acc_enum,
> > +		struct acc100_conf *acc100_conf)
> > +{
> > +	struct rte_q_topology_t *p_qtop;
> > +	p_qtop = NULL;
> > +	switch (acc_enum) {
> > +	case UL_4G:
> > +		p_qtop = &(acc100_conf->q_ul_4g);
> > +		break;
> > +	case UL_5G:
> > +		p_qtop = &(acc100_conf->q_ul_5g);
> > +		break;
> > +	case DL_4G:
> > +		p_qtop = &(acc100_conf->q_dl_4g);
> > +		break;
> > +	case DL_5G:
> > +		p_qtop = &(acc100_conf->q_dl_5g);
> > +		break;
> > +	default:
> > +		/* NOTREACHED */
> > +		rte_bbdev_log(ERR, "Unexpected error evaluating
> qtopFromAcc");
> Use in fetch_acc100_config does not check for NULL.

Yes because it can't be null. This function is called explictly for supported values. 

> > +		break;
> > +	}
> > +	*qtop = p_qtop;
> > +}
> > +
> > +static void
> > +initQTop(struct acc100_conf *acc100_conf) {
> > +	acc100_conf->q_ul_4g.num_aqs_per_groups = 0;
> > +	acc100_conf->q_ul_4g.num_qgroups = 0;
> > +	acc100_conf->q_ul_4g.first_qgroup_index = -1;
> > +	acc100_conf->q_ul_5g.num_aqs_per_groups = 0;
> > +	acc100_conf->q_ul_5g.num_qgroups = 0;
> > +	acc100_conf->q_ul_5g.first_qgroup_index = -1;
> > +	acc100_conf->q_dl_4g.num_aqs_per_groups = 0;
> > +	acc100_conf->q_dl_4g.num_qgroups = 0;
> > +	acc100_conf->q_dl_4g.first_qgroup_index = -1;
> > +	acc100_conf->q_dl_5g.num_aqs_per_groups = 0;
> > +	acc100_conf->q_dl_5g.num_qgroups = 0;
> > +	acc100_conf->q_dl_5g.first_qgroup_index = -1; }
> > +
> > +static inline void
> > +updateQtop(uint8_t acc, uint8_t qg, struct acc100_conf *acc100_conf,
> > +		struct acc100_device *d) {
> > +	uint32_t reg;
> > +	struct rte_q_topology_t *q_top = NULL;
> > +	qtopFromAcc(&q_top, acc, acc100_conf);
> > +	if (unlikely(q_top == NULL))
> > +		return;
> as above, this error is not handled by caller fetch_acc100_config

It cannot really fail for fetch_acc100_config. If you insist I can add. 

> > +	uint16_t aq;
> > +	q_top->num_qgroups++;
> > +	if (q_top->first_qgroup_index == -1) {
> > +		q_top->first_qgroup_index = qg;
> > +		/* Can be optimized to assume all are enabled by default */
> > +		reg = acc100_reg_read(d, queue_offset(d->pf_device,
> > +				0, qg, ACC100_NUM_AQS - 1));
> > +		if (reg & QUEUE_ENABLE) {
> > +			q_top->num_aqs_per_groups = ACC100_NUM_AQS;
> > +			return;
> > +		}
> > +		q_top->num_aqs_per_groups = 0;
> > +		for (aq = 0; aq < ACC100_NUM_AQS; aq++) {
> > +			reg = acc100_reg_read(d, queue_offset(d-
> >pf_device,
> > +					0, qg, aq));
> > +			if (reg & QUEUE_ENABLE)
> > +				q_top->num_aqs_per_groups++;
> > +		}
> > +	}
> > +}
> > +
> > +/* Fetch configuration enabled for the PF/VF using MMIO Read (slow)
> > +*/ static inline void fetch_acc100_config(struct rte_bbdev *dev) {
> > +	struct acc100_device *d = dev->data->dev_private;
> > +	struct acc100_conf *acc100_conf = &d->acc100_conf;
> > +	const struct acc100_registry_addr *reg_addr;
> > +	uint8_t acc, qg;
> > +	uint32_t reg, reg_aq, reg_len0, reg_len1;
> > +	uint32_t reg_mode;
> > +
> > +	/* No need to retrieve the configuration is already done */
> > +	if (d->configured)
> > +		return;
> Warn ?

No this can genuinely happen on a regular basis, just no need to fetch it all again. 

> > +
> > +	/* Choose correct registry addresses for the device type */
> > +	if (d->pf_device)
> > +		reg_addr = &pf_reg_addr;
> > +	else
> > +		reg_addr = &vf_reg_addr;
> > +
> > +	d->ddr_size = (1 + acc100_reg_read(d, reg_addr->ddr_range)) << 10;
> > +
> > +	/* Single VF Bundle by VF */
> > +	acc100_conf->num_vf_bundles = 1;
> > +	initQTop(acc100_conf);
> > +
> > +	struct rte_q_topology_t *q_top = NULL;
> > +	int qman_func_id[5] = {0, 2, 1, 3, 4};
> Do these magic numbers need #defines ?

ok. 

> > +	reg = acc100_reg_read(d, reg_addr->qman_group_func);
> > +	for (qg = 0; qg < ACC100_NUM_QGRPS_PER_WORD; qg++) {
> > +		reg_aq = acc100_reg_read(d,
> > +				queue_offset(d->pf_device, 0, qg, 0));
> > +		if (reg_aq & QUEUE_ENABLE) {
> > +			acc = qman_func_id[(reg >> (qg * 4)) & 0x7];
> 0x7 and [5], this could overflow.

ok thanks, I can add exception handling. Not clear to me right now why it did not trigger tool warning. 

> > +			updateQtop(acc, qg, acc100_conf, d);
> > +		}
> > +	}
> > +
> > +	/* Check the depth of the AQs*/
> > +	reg_len0 = acc100_reg_read(d, reg_addr->depth_log0_offset);
> > +	reg_len1 = acc100_reg_read(d, reg_addr->depth_log1_offset);
> > +	for (acc = 0; acc < NUM_ACC; acc++) {
> > +		qtopFromAcc(&q_top, acc, acc100_conf);
> > +		if (q_top->first_qgroup_index <
> ACC100_NUM_QGRPS_PER_WORD)
> > +			q_top->aq_depth_log2 = (reg_len0 >>
> > +					(q_top->first_qgroup_index * 4))
> > +					& 0xF;
> > +		else
> > +			q_top->aq_depth_log2 = (reg_len1 >>
> > +					((q_top->first_qgroup_index -
> > +					ACC100_NUM_QGRPS_PER_WORD) *
> 4))
> > +					& 0xF;
> > +	}
> > +
> > +	/* Read PF mode */
> > +	if (d->pf_device) {
> > +		reg_mode = acc100_reg_read(d, HWPfHiPfMode);
> > +		acc100_conf->pf_mode_en = (reg_mode == 2) ? 1 : 0;
> 
> 2 is a magic number, consider a #define
> 

ok

> Tom
> 

Thanks
Nic

> > +	}
> > +
> > +	rte_bbdev_log_debug(
> > +			"%s Config LLR SIGN IN/OUT %s %s QG %u %u %u %u
> AQ %u %u %u %u Len %u %u %u %u\n",
> > +			(d->pf_device) ? "PF" : "VF",
> > +			(acc100_conf->input_pos_llr_1_bit) ? "POS" : "NEG",
> > +			(acc100_conf->output_pos_llr_1_bit) ? "POS" :
> "NEG",
> > +			acc100_conf->q_ul_4g.num_qgroups,
> > +			acc100_conf->q_dl_4g.num_qgroups,
> > +			acc100_conf->q_ul_5g.num_qgroups,
> > +			acc100_conf->q_dl_5g.num_qgroups,
> > +			acc100_conf->q_ul_4g.num_aqs_per_groups,
> > +			acc100_conf->q_dl_4g.num_aqs_per_groups,
> > +			acc100_conf->q_ul_5g.num_aqs_per_groups,
> > +			acc100_conf->q_dl_5g.num_aqs_per_groups,
> > +			acc100_conf->q_ul_4g.aq_depth_log2,
> > +			acc100_conf->q_dl_4g.aq_depth_log2,
> > +			acc100_conf->q_ul_5g.aq_depth_log2,
> > +			acc100_conf->q_dl_5g.aq_depth_log2);
> > +}
> > +
> >  /* Free 64MB memory used for software rings */  static int
> > acc100_dev_close(struct rte_bbdev *dev  __rte_unused) @@ -33,8
> +211,55
> > @@
> >  	return 0;
> >  }
> >
> > +/* Get ACC100 device info */
> > +static void
> > +acc100_dev_info_get(struct rte_bbdev *dev,
> > +		struct rte_bbdev_driver_info *dev_info) {
> > +	struct acc100_device *d = dev->data->dev_private;
> > +
> > +	static const struct rte_bbdev_op_cap bbdev_capabilities[] = {
> > +		RTE_BBDEV_END_OF_CAPABILITIES_LIST()
> > +	};
> > +
> > +	static struct rte_bbdev_queue_conf default_queue_conf;
> > +	default_queue_conf.socket = dev->data->socket_id;
> > +	default_queue_conf.queue_size = MAX_QUEUE_DEPTH;
> > +
> > +	dev_info->driver_name = dev->device->driver->name;
> > +
> > +	/* Read and save the populated config from ACC100 registers */
> > +	fetch_acc100_config(dev);
> > +
> > +	/* This isn't ideal because it reports the maximum number of queues
> but
> > +	 * does not provide info on how many can be uplink/downlink or
> different
> > +	 * priorities
> > +	 */
> > +	dev_info->max_num_queues =
> > +			d->acc100_conf.q_dl_5g.num_aqs_per_groups *
> > +			d->acc100_conf.q_dl_5g.num_qgroups +
> > +			d->acc100_conf.q_ul_5g.num_aqs_per_groups *
> > +			d->acc100_conf.q_ul_5g.num_qgroups +
> > +			d->acc100_conf.q_dl_4g.num_aqs_per_groups *
> > +			d->acc100_conf.q_dl_4g.num_qgroups +
> > +			d->acc100_conf.q_ul_4g.num_aqs_per_groups *
> > +			d->acc100_conf.q_ul_4g.num_qgroups;
> > +	dev_info->queue_size_lim = MAX_QUEUE_DEPTH;
> > +	dev_info->hardware_accelerated = true;
> > +	dev_info->max_dl_queue_priority =
> > +			d->acc100_conf.q_dl_4g.num_qgroups - 1;
> > +	dev_info->max_ul_queue_priority =
> > +			d->acc100_conf.q_ul_4g.num_qgroups - 1;
> > +	dev_info->default_queue_conf = default_queue_conf;
> > +	dev_info->cpu_flag_reqs = NULL;
> > +	dev_info->min_alignment = 64;
> > +	dev_info->capabilities = bbdev_capabilities;
> > +	dev_info->harq_buffer_size = d->ddr_size; }
> > +
> >  static const struct rte_bbdev_ops acc100_bbdev_ops = {
> >  	.close = acc100_dev_close,
> > +	.info_get = acc100_dev_info_get,
> >  };
> >
> >  /* ACC100 PCI PF address map */
> > diff --git a/drivers/baseband/acc100/rte_acc100_pmd.h
> > b/drivers/baseband/acc100/rte_acc100_pmd.h
> > index cd77570..662e2c8 100644
> > --- a/drivers/baseband/acc100/rte_acc100_pmd.h
> > +++ b/drivers/baseband/acc100/rte_acc100_pmd.h
> > @@ -7,6 +7,7 @@
> >
> >  #include "acc100_pf_enum.h"
> >  #include "acc100_vf_enum.h"
> > +#include "rte_acc100_cfg.h"
> >
> >  /* Helper macro for logging */
> >  #define rte_bbdev_log(level, fmt, ...) \ @@ -520,6 +521,8 @@ struct
> > acc100_registry_addr {
> >  /* Private data structure for each ACC100 device */  struct
> > acc100_device {
> >  	void *mmio_base;  /**< Base address of MMIO registers (BAR0) */
> > +	uint32_t ddr_size; /* Size in kB */
> > +	struct acc100_conf acc100_conf; /* ACC100 Initial configuration */
> >  	bool pf_device; /**< True if this is a PF ACC100 device */
> >  	bool configured; /**< True if this ACC100 device is configured */
> > };