Re: [PATCH rdma-next v3 3/4] RDMA/core: Introduce new GID table query API

[Jason Gunthorpe <jgg@nvidia.com>]

On Wed, Sep 23, 2020 at 07:50:14PM +0300, Leon Romanovsky wrote:
> From: Avihai Horon <avihaih@nvidia.com>
> 
> Introduce rdma_query_gid_table which enables querying all the GID tables
> of a given device and copying the attributes of all valid GID entries to
> a provided buffer.
> 
> This API provides a faster way to query a GID table using single call and
> will be used in libibverbs to improve current approach that requires
> multiple calls to open, close and read multiple sysfs files for a single
> GID table entry.
> 
> Signed-off-by: Avihai Horon <avihaih@nvidia.com>
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
>  drivers/infiniband/core/cache.c         | 73 ++++++++++++++++++++++++-
>  include/rdma/ib_cache.h                 |  3 +
>  include/uapi/rdma/ib_user_ioctl_verbs.h |  8 +++
>  3 files changed, 81 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/infiniband/core/cache.c b/drivers/infiniband/core/cache.c
> index cf49ac0b0aa6..211b88d17bc7 100644
> +++ b/drivers/infiniband/core/cache.c
> @@ -1247,6 +1247,74 @@ rdma_get_gid_attr(struct ib_device *device, u8 port_num, int index)
>  }
>  EXPORT_SYMBOL(rdma_get_gid_attr);
>  
> +/**
> + * rdma_query_gid_table - Reads GID table entries of all the ports of a device up to max_entries.
> + * @device: The device to query.
> + * @entries: Entries where GID entries are returned.
> + * @max_entries: Maximum number of entries that can be returned.
> + * Entries array must be allocated to hold max_entries number of entries.
> + * @num_entries: Updated to the number of entries that were successfully read.
> + *
> + * Returns number of entries on success or appropriate error code.
> + */
> +ssize_t rdma_query_gid_table(struct ib_device *device,
> +			     struct ib_uverbs_gid_entry *entries,
> +			     size_t max_entries)
> +{
> +	const struct ib_gid_attr *gid_attr;
> +	ssize_t num_entries = 0, ret;
> +	struct ib_gid_table *table;
> +	unsigned int port_num, i;
> +	struct net_device *ndev;
> +	unsigned long flags;
> +
> +	rdma_for_each_port(device, port_num) {
> +		if (!rdma_ib_or_roce(device, port_num))
> +			continue;
> +
> +		table = rdma_gid_table(device, port_num);
> +		read_lock_irqsave(&table->rwlock, flags);
> +		for (i = 0; i < table->sz; i++) {
> +			if (!is_gid_entry_valid(table->data_vec[i]))
> +				continue;
> +			if (num_entries >= max_entries) {
> +				ret = -EINVAL;
> +				goto err;
> +			}
> +
> +			gid_attr = &table->data_vec[i]->attr;
> +
> +			memcpy(&entries->gid, &gid_attr->gid,
> +			       sizeof(gid_attr->gid));
> +			entries->gid_index = gid_attr->index;
> +			entries->port_num = gid_attr->port_num;
> +			entries->gid_type = gid_attr->gid_type;

> +			rcu_read_lock();
> +			ndev = rdma_read_gid_attr_ndev_rcu(gid_attr);

This can't call rdma_read_gid_attr_ndev_rcu(), that also obtains the
rwlock. rwlock can't be nested.

Why didn't lockdep explode on this?

This whole thing can just be:

    ndev = rcu_dereference_protected(gid_attr->ndev, lockdep_is_held(&table->rwlock))
    if (ndev)
         entries->netdev_ifindex = ndev->ifindex;

Jason