Re: [dpdk-dev] [PATCH v5 05/16] dma/idxd: create dmadev instances on pci probe

[fengchengwen <fengchengwen@huawei.com>]

On 2021/9/17 23:24, Kevin Laatz wrote:
> When a suitable device is found during the PCI probe, create a dmadev
> instance for each HW queue. HW definitions required are also included.
> 
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> Signed-off-by: Kevin Laatz <kevin.laatz@intel.com>
> Reviewed-by: Conor Walsh <conor.walsh@intel.com>
> 
> ---
> v4: rebase changes
> ---
>  drivers/dma/idxd/idxd_hw_defs.h  |  71 ++++++++
>  drivers/dma/idxd/idxd_internal.h |  16 ++
>  drivers/dma/idxd/idxd_pci.c      | 278 ++++++++++++++++++++++++++++++-
>  3 files changed, 362 insertions(+), 3 deletions(-)
>  create mode 100644 drivers/dma/idxd/idxd_hw_defs.h
> 
> diff --git a/drivers/dma/idxd/idxd_hw_defs.h b/drivers/dma/idxd/idxd_hw_defs.h
> new file mode 100644
> index 0000000000..ea627cba6d
> --- /dev/null
> +++ b/drivers/dma/idxd/idxd_hw_defs.h
> @@ -0,0 +1,71 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright 2021 Intel Corporation
> + */
> +
> +#ifndef _IDXD_HW_DEFS_H_
> +#define _IDXD_HW_DEFS_H_
> +
> +/*** Definitions for Intel(R) Data Streaming Accelerator  ***/
> +
> +#define IDXD_CMD_SHIFT 20
> +enum rte_idxd_cmds {
> +	idxd_enable_dev = 1,
> +	idxd_disable_dev,
> +	idxd_drain_all,
> +	idxd_abort_all,
> +	idxd_reset_device,
> +	idxd_enable_wq,
> +	idxd_disable_wq,
> +	idxd_drain_wq,
> +	idxd_abort_wq,
> +	idxd_reset_wq,
> +};
> +
> +/* General bar0 registers */
> +struct rte_idxd_bar0 {
> +	uint32_t __rte_cache_aligned version;    /* offset 0x00 */
> +	uint64_t __rte_aligned(0x10) gencap;     /* offset 0x10 */
> +	uint64_t __rte_aligned(0x10) wqcap;      /* offset 0x20 */
> +	uint64_t __rte_aligned(0x10) grpcap;     /* offset 0x30 */
> +	uint64_t __rte_aligned(0x08) engcap;     /* offset 0x38 */
> +	uint64_t __rte_aligned(0x10) opcap;      /* offset 0x40 */
> +	uint64_t __rte_aligned(0x20) offsets[2]; /* offset 0x60 */
> +	uint32_t __rte_aligned(0x20) gencfg;     /* offset 0x80 */
> +	uint32_t __rte_aligned(0x08) genctrl;    /* offset 0x88 */
> +	uint32_t __rte_aligned(0x10) gensts;     /* offset 0x90 */
> +	uint32_t __rte_aligned(0x08) intcause;   /* offset 0x98 */
> +	uint32_t __rte_aligned(0x10) cmd;        /* offset 0xA0 */
> +	uint32_t __rte_aligned(0x08) cmdstatus;  /* offset 0xA8 */
> +	uint64_t __rte_aligned(0x20) swerror[4]; /* offset 0xC0 */
> +};
> +
> +/* workqueue config is provided by array of uint32_t. */
> +enum rte_idxd_wqcfg {
> +	wq_size_idx,       /* size is in first 32-bit value */
> +	wq_threshold_idx,  /* WQ threshold second 32-bits */
> +	wq_mode_idx,       /* WQ mode and other flags */
> +	wq_sizes_idx,      /* WQ transfer and batch sizes */
> +	wq_occ_int_idx,    /* WQ occupancy interrupt handle */
> +	wq_occ_limit_idx,  /* WQ occupancy limit */
> +	wq_state_idx,      /* WQ state and occupancy state */
> +};
> +
> +#define WQ_MODE_SHARED    0
> +#define WQ_MODE_DEDICATED 1
> +#define WQ_PRIORITY_SHIFT 4
> +#define WQ_BATCH_SZ_SHIFT 5
> +#define WQ_STATE_SHIFT 30
> +#define WQ_STATE_MASK 0x3
> +
> +struct rte_idxd_grpcfg {
> +	uint64_t grpwqcfg[4]  __rte_cache_aligned; /* 64-byte register set */
> +	uint64_t grpengcfg;  /* offset 32 */
> +	uint32_t grpflags;   /* offset 40 */
> +};
> +
> +#define GENSTS_DEV_STATE_MASK 0x03
> +#define CMDSTATUS_ACTIVE_SHIFT 31
> +#define CMDSTATUS_ACTIVE_MASK (1 << 31)
> +#define CMDSTATUS_ERR_MASK 0xFF
> +
> +#endif
> diff --git a/drivers/dma/idxd/idxd_internal.h b/drivers/dma/idxd/idxd_internal.h
> index fa6f053f72..cb3a68c69b 100644
> --- a/drivers/dma/idxd/idxd_internal.h
> +++ b/drivers/dma/idxd/idxd_internal.h
> @@ -5,6 +5,10 @@
>  #ifndef _IDXD_INTERNAL_H_
>  #define _IDXD_INTERNAL_H_
>  
> +#include <rte_spinlock.h>
> +
> +#include "idxd_hw_defs.h"
> +
>  /**
>   * @file idxd_internal.h
>   *
> @@ -24,6 +28,16 @@ extern int idxd_pmd_logtype;
>  #define IDXD_PMD_ERR(fmt, args...)    IDXD_PMD_LOG(ERR, fmt, ## args)
>  #define IDXD_PMD_WARN(fmt, args...)   IDXD_PMD_LOG(WARNING, fmt, ## args)
>  
> +struct idxd_pci_common {
> +	rte_spinlock_t lk;
> +
> +	uint8_t wq_cfg_sz;
> +	volatile struct rte_idxd_bar0 *regs;
> +	volatile uint32_t *wq_regs_base;
> +	volatile struct rte_idxd_grpcfg *grp_regs;
> +	volatile void *portals;
> +};
> +
>  struct idxd_dmadev {
>  	/* counters to track the batches */
>  	unsigned short max_batches;
> @@ -58,6 +72,8 @@ struct idxd_dmadev {
>  		struct {
>  			unsigned int dsa_id;
>  		} bus;
> +
> +		struct idxd_pci_common *pci;
>  	} u;
>  };
>  
> diff --git a/drivers/dma/idxd/idxd_pci.c b/drivers/dma/idxd/idxd_pci.c
> index 79e4aadcab..171e5ffc07 100644
> --- a/drivers/dma/idxd/idxd_pci.c
> +++ b/drivers/dma/idxd/idxd_pci.c
> @@ -3,6 +3,9 @@
>   */
>  
>  #include <rte_bus_pci.h>
> +#include <rte_devargs.h>
> +#include <rte_dmadev_pmd.h>
> +#include <rte_malloc.h>
>  
>  #include "idxd_internal.h"
>  
> @@ -16,17 +19,286 @@ const struct rte_pci_id pci_id_idxd_map[] = {
>  	{ .vendor_id = 0, /* sentinel */ },
>  };
>  
> +static inline int
> +idxd_pci_dev_command(struct idxd_dmadev *idxd, enum rte_idxd_cmds command)
> +{
> +	uint8_t err_code;
> +	uint16_t qid = idxd->qid;
> +	int i = 0;
> +
> +	if (command >= idxd_disable_wq && command <= idxd_reset_wq)
> +		qid = (1 << qid);
> +	rte_spinlock_lock(&idxd->u.pci->lk);
> +	idxd->u.pci->regs->cmd = (command << IDXD_CMD_SHIFT) | qid;
> +
> +	do {
> +		rte_pause();
> +		err_code = idxd->u.pci->regs->cmdstatus;
> +		if (++i >= 1000) {
> +			IDXD_PMD_ERR("Timeout waiting for command response from HW");
> +			rte_spinlock_unlock(&idxd->u.pci->lk);
> +			return err_code;
> +		}
> +	} while (idxd->u.pci->regs->cmdstatus & CMDSTATUS_ACTIVE_MASK);
> +	rte_spinlock_unlock(&idxd->u.pci->lk);
> +
> +	return err_code & CMDSTATUS_ERR_MASK;
> +}
> +
> +static uint32_t *
> +idxd_get_wq_cfg(struct idxd_pci_common *pci, uint8_t wq_idx)
> +{
> +	return RTE_PTR_ADD(pci->wq_regs_base,
> +			(uintptr_t)wq_idx << (5 + pci->wq_cfg_sz));
> +}
> +
> +static int
> +idxd_is_wq_enabled(struct idxd_dmadev *idxd)
> +{
> +	uint32_t state = idxd_get_wq_cfg(idxd->u.pci, idxd->qid)[wq_state_idx];
> +	return ((state >> WQ_STATE_SHIFT) & WQ_STATE_MASK) == 0x1;
> +}
> +
> +static const struct rte_dma_dev_ops idxd_pci_ops = {
> +
> +};
> +
> +/* each portal uses 4 x 4k pages */
> +#define IDXD_PORTAL_SIZE (4096 * 4)
> +
> +static int
> +init_pci_device(struct rte_pci_device *dev, struct idxd_dmadev *idxd,
> +		unsigned int max_queues)
> +{
> +	struct idxd_pci_common *pci;
> +	uint8_t nb_groups, nb_engines, nb_wqs;
> +	uint16_t grp_offset, wq_offset; /* how far into bar0 the regs are */
> +	uint16_t wq_size, total_wq_size;
> +	uint8_t lg2_max_batch, lg2_max_copy_size;
> +	unsigned int i, err_code;
> +
> +	pci = malloc(sizeof(*pci));
> +	if (pci == NULL) {
> +		IDXD_PMD_ERR("%s: Can't allocate memory", __func__);
> +		goto err;
> +	}
> +	rte_spinlock_init(&pci->lk);
> +
> +	/* assign the bar registers, and then configure device */
> +	pci->regs = dev->mem_resource[0].addr;
> +	grp_offset = (uint16_t)pci->regs->offsets[0];
> +	pci->grp_regs = RTE_PTR_ADD(pci->regs, grp_offset * 0x100);
> +	wq_offset = (uint16_t)(pci->regs->offsets[0] >> 16);
> +	pci->wq_regs_base = RTE_PTR_ADD(pci->regs, wq_offset * 0x100);
> +	pci->portals = dev->mem_resource[2].addr;
> +	pci->wq_cfg_sz = (pci->regs->wqcap >> 24) & 0x0F;
> +
> +	/* sanity check device status */
> +	if (pci->regs->gensts & GENSTS_DEV_STATE_MASK) {
> +		/* need function-level-reset (FLR) or is enabled */
> +		IDXD_PMD_ERR("Device status is not disabled, cannot init");
> +		goto err;
> +	}
> +	if (pci->regs->cmdstatus & CMDSTATUS_ACTIVE_MASK) {
> +		/* command in progress */
> +		IDXD_PMD_ERR("Device has a command in progress, cannot init");
> +		goto err;
> +	}
> +
> +	/* read basic info about the hardware for use when configuring */
> +	nb_groups = (uint8_t)pci->regs->grpcap;
> +	nb_engines = (uint8_t)pci->regs->engcap;
> +	nb_wqs = (uint8_t)(pci->regs->wqcap >> 16);
> +	total_wq_size = (uint16_t)pci->regs->wqcap;
> +	lg2_max_copy_size = (uint8_t)(pci->regs->gencap >> 16) & 0x1F;
> +	lg2_max_batch = (uint8_t)(pci->regs->gencap >> 21) & 0x0F;
> +
> +	IDXD_PMD_DEBUG("nb_groups = %u, nb_engines = %u, nb_wqs = %u",
> +			nb_groups, nb_engines, nb_wqs);
> +
> +	/* zero out any old config */
> +	for (i = 0; i < nb_groups; i++) {
> +		pci->grp_regs[i].grpengcfg = 0;
> +		pci->grp_regs[i].grpwqcfg[0] = 0;
> +	}
> +	for (i = 0; i < nb_wqs; i++)
> +		idxd_get_wq_cfg(pci, i)[0] = 0;
> +
> +	/* limit queues if necessary */
> +	if (max_queues != 0 && nb_wqs > max_queues) {
> +		nb_wqs = max_queues;
> +		if (nb_engines > max_queues)
> +			nb_engines = max_queues;
> +		if (nb_groups > max_queues)
> +			nb_engines = max_queues;
> +		IDXD_PMD_DEBUG("Limiting queues to %u", nb_wqs);
> +	}
> +
> +	/* put each engine into a separate group to avoid reordering */
> +	if (nb_groups > nb_engines)
> +		nb_groups = nb_engines;
> +	if (nb_groups < nb_engines)
> +		nb_engines = nb_groups;
> +
> +	/* assign engines to groups, round-robin style */
> +	for (i = 0; i < nb_engines; i++) {
> +		IDXD_PMD_DEBUG("Assigning engine %u to group %u",
> +				i, i % nb_groups);
> +		pci->grp_regs[i % nb_groups].grpengcfg |= (1ULL << i);
> +	}
> +
> +	/* now do the same for queues and give work slots to each queue */
> +	wq_size = total_wq_size / nb_wqs;
> +	IDXD_PMD_DEBUG("Work queue size = %u, max batch = 2^%u, max copy = 2^%u",
> +			wq_size, lg2_max_batch, lg2_max_copy_size);
> +	for (i = 0; i < nb_wqs; i++) {
> +		/* add engine "i" to a group */
> +		IDXD_PMD_DEBUG("Assigning work queue %u to group %u",
> +				i, i % nb_groups);
> +		pci->grp_regs[i % nb_groups].grpwqcfg[0] |= (1ULL << i);
> +		/* now configure it, in terms of size, max batch, mode */
> +		idxd_get_wq_cfg(pci, i)[wq_size_idx] = wq_size;
> +		idxd_get_wq_cfg(pci, i)[wq_mode_idx] = (1 << WQ_PRIORITY_SHIFT) |
> +				WQ_MODE_DEDICATED;
> +		idxd_get_wq_cfg(pci, i)[wq_sizes_idx] = lg2_max_copy_size |
> +				(lg2_max_batch << WQ_BATCH_SZ_SHIFT);
> +	}
> +
> +	/* dump the group configuration to output */
> +	for (i = 0; i < nb_groups; i++) {
> +		IDXD_PMD_DEBUG("## Group %d", i);
> +		IDXD_PMD_DEBUG("    GRPWQCFG: %"PRIx64, pci->grp_regs[i].grpwqcfg[0]);
> +		IDXD_PMD_DEBUG("    GRPENGCFG: %"PRIx64, pci->grp_regs[i].grpengcfg);
> +		IDXD_PMD_DEBUG("    GRPFLAGS: %"PRIx32, pci->grp_regs[i].grpflags);
> +	}
> +
> +	idxd->u.pci = pci;
> +	idxd->max_batches = wq_size;
> +
> +	/* enable the device itself */
> +	err_code = idxd_pci_dev_command(idxd, idxd_enable_dev);
> +	if (err_code) {
> +		IDXD_PMD_ERR("Error enabling device: code %#x", err_code);
> +		return err_code;
> +	}
> +	IDXD_PMD_DEBUG("IDXD Device enabled OK");
> +
> +	return nb_wqs;
> +
> +err:
> +	free(pci);
> +	return -1;
> +}
> +
>  static int
>  idxd_dmadev_probe_pci(struct rte_pci_driver *drv, struct rte_pci_device *dev)
>  {
> -	int ret = 0;
> +	struct idxd_dmadev idxd = {0};
> +	uint8_t nb_wqs;
> +	int qid, ret = 0;
>  	char name[PCI_PRI_STR_SIZE];
> +	unsigned int max_queues = 0;
>  
>  	rte_pci_device_name(&dev->addr, name, sizeof(name));
>  	IDXD_PMD_INFO("Init %s on NUMA node %d", name, dev->device.numa_node);
>  	dev->device.driver = &drv->driver;
>  
> -	return ret;
> +	if (dev->device.devargs && dev->device.devargs->args[0] != '\0') {
> +		/* if the number of devargs grows beyond just 1, use rte_kvargs */
> +		if (sscanf(dev->device.devargs->args,
> +				"max_queues=%u", &max_queues) != 1) {
> +			IDXD_PMD_ERR("Invalid device parameter: '%s'",
> +					dev->device.devargs->args);
> +			return -1;
> +		}
> +	}
> +
> +	ret = init_pci_device(dev, &idxd, max_queues);
> +	if (ret < 0) {
> +		IDXD_PMD_ERR("Error initializing PCI hardware");
> +		return ret;
> +	}
> +	if (idxd.u.pci->portals == NULL) {
> +		IDXD_PMD_ERR("Error, invalid portal assigned during initialization\n");
> +		return -EINVAL;
> +	}
> +	nb_wqs = (uint8_t)ret;
> +
> +	/* set up one device for each queue */
> +	for (qid = 0; qid < nb_wqs; qid++) {
> +		char qname[32];
> +
> +		/* add the queue number to each device name */
> +		snprintf(qname, sizeof(qname), "%s-q%d", name, qid);
> +		idxd.qid = qid;
> +		idxd.portal = RTE_PTR_ADD(idxd.u.pci->portals,
> +				qid * IDXD_PORTAL_SIZE);
> +		if (idxd_is_wq_enabled(&idxd))
> +			IDXD_PMD_ERR("Error, WQ %u seems enabled", qid);
> +		ret = idxd_dmadev_create(qname, &dev->device,
> +				&idxd, &idxd_pci_ops);
> +		if (ret != 0) {
> +			IDXD_PMD_ERR("Failed to create dmadev %s", name);
> +			if (qid == 0) /* if no devices using this, free pci */
> +				free(idxd.u.pci);
> +			return ret;
> +		}
> +	}
> +
> +	return 0;
> +}
> +
> +static int
> +idxd_dmadev_destroy(const char *name)
> +{
> +	int ret;
> +	uint8_t err_code;
> +	struct rte_dma_dev *dmadev;
> +	struct idxd_dmadev *idxd;
> +	int dev_id = rte_dma_get_dev_id(name);
> +
> +	if (!name) {
> +		IDXD_PMD_ERR("Invalid device name");
> +		return -EINVAL;
> +	}
> +
> +	if (dev_id < 0) {
> +		IDXD_PMD_ERR("Invalid device ID");
> +		return -EINVAL;
> +	}
> +
> +	dmadev = &rte_dma_devices[dev_id];
> +	if (!dmadev) {
> +		IDXD_PMD_ERR("Invalid device name (%s)", name);
> +		return -EINVAL;
> +	}
> +
> +	idxd = dmadev->dev_private;
> +	if (!idxd) {
> +		IDXD_PMD_ERR("Error getting dev_private");
> +		return -EINVAL;
> +	}
> +
> +	/* disable the device */
> +	err_code = idxd_pci_dev_command(idxd, idxd_disable_dev);
> +	if (err_code) {
> +		IDXD_PMD_ERR("Error disabling device: code %#x", err_code);
> +		return err_code;
> +	}
> +	IDXD_PMD_DEBUG("IDXD Device disabled OK");

Recommended: move the above disable idxd device to rte_dma_close() ops.
when create: mark state as READY
when destroy: direct call rte_dma_pmd_release(name), the lib will call rte_dma_close().

> +
> +	/* free device memory */
> +	IDXD_PMD_DEBUG("Freeing device driver memory");
> +	dmadev->dev_private = NULL;

The dmalib managed dev_private, so that driver could not do free again.

> +	rte_free(idxd->batch_idx_ring);
> +	rte_free(idxd->desc_ring);

Please move above free ops to rte_dma_close() ops.

> +
> +	/* rte_dma_close is called by pmd_release */
> +	ret = rte_dma_pmd_release(name);
> +	if (ret)
> +		IDXD_PMD_DEBUG("Device cleanup failed");
> +
> +	return 0;
>  }
>  
>  static int
> @@ -39,7 +311,7 @@ idxd_dmadev_remove_pci(struct rte_pci_device *dev)
>  	IDXD_PMD_INFO("Closing %s on NUMA node %d",
>  			name, dev->device.numa_node);
>  
> -	return 0;
> +	return idxd_dmadev_destroy(name);
>  }
>  
>  struct rte_pci_driver idxd_pmd_drv_pci = {
>