[dpdk-dev] [PATCH 03/14] vhost: introduce vDPA devices class

[Gaëtan Rivet]

On 11/06/20 23:37 +0200, Maxime Coquelin wrote:
> This patch introduces vDPA device class. It will enable
> application to iterate over the vDPA devices.
> 
> Signed-off-by: Maxime Coquelin <maxime.coquelin at redhat.com>
> ---
>  lib/librte_vhost/vdpa.c | 116 +++++++++++++++++++++++++++++++++-------
>  1 file changed, 98 insertions(+), 18 deletions(-)
> 
> diff --git a/lib/librte_vhost/vdpa.c b/lib/librte_vhost/vdpa.c
> index b2b2a105f1..61ab9aadb4 100644
> --- a/lib/librte_vhost/vdpa.c
> +++ b/lib/librte_vhost/vdpa.c
> @@ -10,11 +10,12 @@
>  
>  #include <stdbool.h>
>  
> +#include <rte_class.h>
>  #include <rte_malloc.h>
>  #include "rte_vdpa.h"
>  #include "vhost.h"
>  
> -static struct rte_vdpa_device *vdpa_devices[MAX_VHOST_DEVICE];
> +static struct rte_vdpa_device vdpa_devices[MAX_VHOST_DEVICE];
>  static uint32_t vdpa_device_num;
>  
>  static bool
> @@ -46,35 +47,28 @@ rte_vdpa_register_device(struct rte_vdpa_dev_addr *addr,
>  		struct rte_vdpa_dev_ops *ops)
>  {
>  	struct rte_vdpa_device *dev;
> -	char device_name[MAX_VDPA_NAME_LEN];
>  	int i;
>  
>  	if (vdpa_device_num >= MAX_VHOST_DEVICE || addr == NULL || ops == NULL)
>  		return -1;
>  
>  	for (i = 0; i < MAX_VHOST_DEVICE; i++) {
> -		dev = vdpa_devices[i];
> -		if (dev && is_same_vdpa_device(&dev->addr, addr))
> +		dev = &vdpa_devices[i];
> +		if (dev->ops && is_same_vdpa_device(&dev->addr, addr))
>  			return -1;
>  	}
>  
>  	for (i = 0; i < MAX_VHOST_DEVICE; i++) {
> -		if (vdpa_devices[i] == NULL)
> +		if (vdpa_devices[i].ops == NULL)
>  			break;
>  	}
>  
>  	if (i == MAX_VHOST_DEVICE)
>  		return -1;
>  
> -	snprintf(device_name, sizeof(device_name), "vdpa-dev-%d", i);
> -	dev = rte_zmalloc(device_name, sizeof(struct rte_vdpa_device),
> -			RTE_CACHE_LINE_SIZE);
> -	if (!dev)
> -		return -1;
> -
> +	dev = &vdpa_devices[i];
>  	memcpy(&dev->addr, addr, sizeof(struct rte_vdpa_dev_addr));
>  	dev->ops = ops;
> -	vdpa_devices[i] = dev;
>  	vdpa_device_num++;
>  
>  	return i;
> @@ -83,11 +77,10 @@ rte_vdpa_register_device(struct rte_vdpa_dev_addr *addr,
>  int
>  rte_vdpa_unregister_device(int did)
>  {
> -	if (did < 0 || did >= MAX_VHOST_DEVICE || vdpa_devices[did] == NULL)
> +	if (did < 0 || did >= MAX_VHOST_DEVICE || vdpa_devices[did].ops == NULL)
>  		return -1;
>  
> -	rte_free(vdpa_devices[did]);
> -	vdpa_devices[did] = NULL;
> +	memset(&vdpa_devices[did], 0, sizeof(struct rte_vdpa_device));
>  	vdpa_device_num--;
>  
>  	return did;
> @@ -103,8 +96,11 @@ rte_vdpa_find_device_id(struct rte_vdpa_dev_addr *addr)
>  		return -1;
>  
>  	for (i = 0; i < MAX_VHOST_DEVICE; ++i) {
> -		dev = vdpa_devices[i];
> -		if (dev && is_same_vdpa_device(&dev->addr, addr))
> +		dev = &vdpa_devices[i];
> +		if (dev->ops == NULL)
> +			continue;
> +
> +		if (is_same_vdpa_device(&dev->addr, addr))
>  			return i;
>  	}
>  
> @@ -117,7 +113,7 @@ rte_vdpa_get_device(int did)
>  	if (did < 0 || did >= MAX_VHOST_DEVICE)
>  		return NULL;
>  
> -	return vdpa_devices[did];
> +	return &vdpa_devices[did];
>  }
>  
>  int
> @@ -227,3 +223,87 @@ rte_vdpa_relay_vring_used(int vid, uint16_t qid, void *vring_m)
>  		free_ind_table(idesc);
>  	return -1;
>  }
> +
> +static uint16_t
> +vdpa_dev_to_id(const struct rte_vdpa_device *dev)
> +{
> +	if (dev == NULL)
> +		return MAX_VHOST_DEVICE;

I know it comes from ethdev, this helper was introduced by 214ed1acd12
and changed the comparison type from ptrdiff_t to uin16_t.

This is incorrect, ideally it should instead be a ptrdiff_t, then
if within acceptable range it should be casted to uin16_t to be used as
an index.

> +	return dev - vdpa_devices;
> +}
> +
> +static int
> +vdpa_dev_match(struct rte_vdpa_device *dev,
> +	      const struct rte_device *rte_dev)
> +{
> +	struct rte_vdpa_dev_addr addr;
> +
> +	/*  Only PCI bus supported for now */
> +	if (strcmp(rte_dev->bus->name, "pci") != 0)
> +		return -1;
> +
> +	addr.type = VDPA_ADDR_PCI;
> +
> +	if (rte_pci_addr_parse(rte_dev->name, &addr.pci_addr) != 0)
> +		return -1;
> +
> +	if (!is_same_vdpa_device(&dev->addr, &addr))
> +		return -1;
> +
> +	return 0;
> +}
> +
> +/* Generic rte_vdpa_dev comparison function. */
> +typedef int (*rte_vdpa_cmp_t)(struct rte_vdpa_device *,
> +		const struct rte_device *rte_dev);
> +
> +static struct rte_vdpa_device *
> +vdpa_find_device(const struct rte_vdpa_device *start, rte_vdpa_cmp_t cmp,
> +		struct rte_device *rte_dev)
> +{
> +	struct rte_vdpa_device *dev;
> +	ptrdiff_t idx;

The ptrdiff_t was used in ethdev due to the original

   start - &rte_eth_devices[0] + 1;

The diff was removed by commit 214ed1acd12 . 'idx' should be
the proper type to hold a device id, here probably uint16_t.

Ethdev class should be cleaned up as well.

> +
> +	/* Avoid Undefined Behaviour */
> +	if (start != NULL &&
> +	    (start < &vdpa_devices[0] ||
> +	     start >= &vdpa_devices[MAX_VHOST_DEVICE]))
> +		return NULL;

Same as before, this guard made sense mostly for the diff following.
It is coupled to the "random ptr to static array" comparison happening,
it would be better in vpda_dev_to_id() I think.

> +
> +	if (start != NULL)
> +		idx = vdpa_dev_to_id(start) + 1;
> +	else
> +		idx = 0;
> +	for (; idx < MAX_VHOST_DEVICE; idx++) {
> +		dev = &vdpa_devices[idx];
> +		/*
> +		 * ToDo: Certainly better to introduce a state field,
> +		 * but rely on ops being set for now.
> +		 */
> +		if (dev->ops == NULL)
> +			continue;
> +		if (cmp(dev, rte_dev) == 0)
> +			return dev;
> +	}
> +	return NULL;
> +}
> +
> +static void *
> +vdpa_dev_iterate(const void *start,
> +		const char *str,
> +		const struct rte_dev_iterator *it)
> +{
> +	struct rte_vdpa_device *vdpa_dev = NULL;
> +
> +	RTE_SET_USED(str);
> +
> +	vdpa_dev = vdpa_find_device(start, vdpa_dev_match, it->device);
> +
> +	return vdpa_dev;
> +}
> +
> +static struct rte_class rte_class_vdpa = {
> +	.dev_iterate = vdpa_dev_iterate,
> +};
> +
> +RTE_REGISTER_CLASS(vdpa, rte_class_vdpa);
> -- 
> 2.26.2
> 

-- 
Gaëtan