Date: Tue, 5 Dec 2023 13:57:17 +0800
From: Zhu Yanjun <yanjun.zhu@...ux.dev>
To: Bob Pearson <rpearsonhpe@...il.com>, jgg@...dia.com,
 linux-rdma@...r.kernel.org, dsahern@...nel.org, davem@...emloft.net,
 netdev@...r.kernel.org
Subject: Re: [PATCH for-next v5 5/7] RDMA/rxe: Split multicast lock

Add  David S. Miller and  David Ahern.

They are the maintainers in netdev and very familiar with mcast.

Zhu Yanjun

在 2023/12/5 8:26, Bob Pearson 写道:
> Split rxe->mcg_lock into two locks. One to protect mcg->qp_list
> and one to protect rxe->mcg_tree (red-black tree) write side
> operations and provide serialization between rxe_attach_mcast
> and rxe_detach_mcast.
>
> Make the qp_list lock a spin_lock_irqsave lock and move to the
> mcg struct. It protects the qp_list from simultaneous access
> from rxe_mcast.c and rxe_recv.c when processing incoming multi-
> cast packets. In theory some ethernet driver could bypass NAPI
> so an irq lock is better than a bh lock.
>
> Make the mcg_tree lock a mutex since the attach/detach APIs are
> not called in atomic context. This allows some significant cleanup
> since we can call kzalloc while holding the mutex so some recheck
> code can be eliminated.
>
> Signed-off-by: Bob Pearson <rpearsonhpe@...il.com>
> ---
>   drivers/infiniband/sw/rxe/rxe.c       |   2 +-
>   drivers/infiniband/sw/rxe/rxe_mcast.c | 254 ++++++++++----------------
>   drivers/infiniband/sw/rxe/rxe_recv.c  |   5 +-
>   drivers/infiniband/sw/rxe/rxe_verbs.h |   3 +-
>   4 files changed, 105 insertions(+), 159 deletions(-)
>
> diff --git a/drivers/infiniband/sw/rxe/rxe.c b/drivers/infiniband/sw/rxe/rxe.c
> index 54c723a6edda..147cb16e937d 100644
> --- a/drivers/infiniband/sw/rxe/rxe.c
> +++ b/drivers/infiniband/sw/rxe/rxe.c
> @@ -142,7 +142,7 @@ static void rxe_init(struct rxe_dev *rxe)
>   	INIT_LIST_HEAD(&rxe->pending_mmaps);
>   
>   	/* init multicast support */
> -	spin_lock_init(&rxe->mcg_lock);
> +	mutex_init(&rxe->mcg_mutex);
>   	rxe->mcg_tree = RB_ROOT;
>   
>   	mutex_init(&rxe->usdev_lock);
> diff --git a/drivers/infiniband/sw/rxe/rxe_mcast.c b/drivers/infiniband/sw/rxe/rxe_mcast.c
> index 44948f9cb02b..ac8da0bc8428 100644
> --- a/drivers/infiniband/sw/rxe/rxe_mcast.c
> +++ b/drivers/infiniband/sw/rxe/rxe_mcast.c
> @@ -135,7 +135,7 @@ static int rxe_mcast_del(struct rxe_mcg *mcg)
>    * __rxe_insert_mcg - insert an mcg into red-black tree (rxe->mcg_tree)
>    * @mcg: mcg object with an embedded red-black tree node
>    *
> - * Context: caller must hold a reference to mcg and rxe->mcg_lock and
> + * Context: caller must hold a reference to mcg and rxe->mcg_mutex and
>    * is responsible to avoid adding the same mcg twice to the tree.
>    */
>   static void __rxe_insert_mcg(struct rxe_mcg *mcg)
> @@ -170,7 +170,7 @@ static void __rxe_insert_mcg(struct rxe_mcg *mcg)
>    * __rxe_remove_mcg - remove an mcg from red-black tree holding lock
>    * @mcg: mcast group object with an embedded red-black tree node
>    *
> - * Context: caller must hold a reference to mcg and rxe->mcg_lock
> + * Context: caller must hold a reference to mcg and rxe->mcg_mutex
>    */
>   static void __rxe_remove_mcg(struct rxe_mcg *mcg)
>   {
> @@ -210,34 +210,6 @@ struct rxe_mcg *rxe_lookup_mcg(struct rxe_dev *rxe,
>   	return mcg;
>   }
>   
> -/**
> - * __rxe_init_mcg - initialize a new mcg
> - * @rxe: rxe device
> - * @mgid: multicast address as a gid
> - * @mcg: new mcg object
> - *
> - * Context: caller should hold rxe->mcg lock
> - */
> -static void __rxe_init_mcg(struct rxe_dev *rxe, union ib_gid *mgid,
> -			   struct rxe_mcg *mcg)
> -{
> -	kref_init(&mcg->ref_cnt);
> -	memcpy(&mcg->mgid, mgid, sizeof(mcg->mgid));
> -	mcg->is_ipv6 = !ipv6_addr_v4mapped((struct in6_addr *)mgid);
> -	INIT_LIST_HEAD(&mcg->qp_list);
> -	mcg->rxe = rxe;
> -
> -	/* caller holds a ref on mcg but that will be
> -	 * dropped when mcg goes out of scope. We need to take a ref
> -	 * on the pointer that will be saved in the red-black tree
> -	 * by __rxe_insert_mcg and used to lookup mcg from mgid later.
> -	 * Inserting mcg makes it visible to outside so this should
> -	 * be done last after the object is ready.
> -	 */
> -	kref_get(&mcg->ref_cnt);
> -	__rxe_insert_mcg(mcg);
> -}
> -
>   /**
>    * rxe_get_mcg - lookup or allocate a mcg
>    * @rxe: rxe device object
> @@ -247,51 +219,48 @@ static void __rxe_init_mcg(struct rxe_dev *rxe, union ib_gid *mgid,
>    */
>   static struct rxe_mcg *rxe_get_mcg(struct rxe_dev *rxe, union ib_gid *mgid)
>   {
> -	struct rxe_mcg *mcg, *tmp;
> +	struct rxe_mcg *mcg;
>   	int err;
>   
> -	if (rxe->attr.max_mcast_grp == 0)
> -		return ERR_PTR(-EINVAL);
> -
> -	/* check to see if mcg already exists */
> +	mutex_lock(&rxe->mcg_mutex);
>   	mcg = rxe_lookup_mcg(rxe, mgid);
>   	if (mcg)
> -		return mcg;
> +		goto out;	/* nothing to do */
>   
> -	/* check to see if we have reached limit */
>   	if (atomic_inc_return(&rxe->mcg_num) > rxe->attr.max_mcast_grp) {
>   		err = -ENOMEM;
>   		goto err_dec;
>   	}
>   
> -	/* speculative alloc of new mcg */
>   	mcg = kzalloc(sizeof(*mcg), GFP_KERNEL);
>   	if (!mcg) {
>   		err = -ENOMEM;
>   		goto err_dec;
>   	}
>   
> -	spin_lock_bh(&rxe->mcg_lock);
> -	/* re-check to see if someone else just added it */
> -	tmp = rxe_lookup_mcg(rxe, mgid);
> -	if (tmp) {
> -		spin_unlock_bh(&rxe->mcg_lock);
> -		atomic_dec(&rxe->mcg_num);
> -		kfree(mcg);
> -		return tmp;
> -	}
> -
> -	__rxe_init_mcg(rxe, mgid, mcg);
> -	spin_unlock_bh(&rxe->mcg_lock);
> +	memcpy(&mcg->mgid, mgid, sizeof(mcg->mgid));
> +	mcg->is_ipv6 = !ipv6_addr_v4mapped((struct in6_addr *)mgid);
> +	mcg->rxe = rxe;
> +	kref_init(&mcg->ref_cnt);
> +	INIT_LIST_HEAD(&mcg->qp_list);
> +	spin_lock_init(&mcg->lock);
> +	kref_get(&mcg->ref_cnt);
> +	__rxe_insert_mcg(mcg);
>   
> -	/* add mcast address outside of lock */
>   	err = rxe_mcast_add(mcg);
> -	if (!err)
> -		return mcg;
> +	if (err)
> +		goto err_free;
> +
> +out:
> +	mutex_unlock(&rxe->mcg_mutex);
> +	return mcg;
>   
> +err_free:
> +	__rxe_remove_mcg(mcg);
>   	kfree(mcg);
>   err_dec:
>   	atomic_dec(&rxe->mcg_num);
> +	mutex_unlock(&rxe->mcg_mutex);
>   	return ERR_PTR(err);
>   }
>   
> @@ -307,10 +276,10 @@ void rxe_cleanup_mcg(struct kref *kref)
>   }
>   
>   /**
> - * __rxe_destroy_mcg - destroy mcg object holding rxe->mcg_lock
> + * __rxe_destroy_mcg - destroy mcg object holding rxe->mcg_mutex
>    * @mcg: the mcg object
>    *
> - * Context: caller is holding rxe->mcg_lock
> + * Context: caller is holding rxe->mcg_mutex
>    * no qp's are attached to mcg
>    */
>   static void __rxe_destroy_mcg(struct rxe_mcg *mcg)
> @@ -335,151 +304,123 @@ static void rxe_destroy_mcg(struct rxe_mcg *mcg)
>   	/* delete mcast address outside of lock */
>   	rxe_mcast_del(mcg);
>   
> -	spin_lock_bh(&mcg->rxe->mcg_lock);
> +	mutex_lock(&mcg->rxe->mcg_mutex);
>   	__rxe_destroy_mcg(mcg);
> -	spin_unlock_bh(&mcg->rxe->mcg_lock);
> +	mutex_unlock(&mcg->rxe->mcg_mutex);
>   }
>   
>   /**
> - * __rxe_init_mca - initialize a new mca holding lock
> + * rxe_attach_mcg - attach qp to mcg if not already attached
>    * @qp: qp object
>    * @mcg: mcg object
> - * @mca: empty space for new mca
> - *
> - * Context: caller must hold references on qp and mcg, rxe->mcg_lock
> - * and pass memory for new mca
>    *
>    * Returns: 0 on success else an error
>    */
> -static int __rxe_init_mca(struct rxe_qp *qp, struct rxe_mcg *mcg,
> -			  struct rxe_mca *mca)
> +static int rxe_attach_mcg(struct rxe_mcg *mcg, struct rxe_qp *qp)
>   {
> -	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
> -	int n;
> +	struct rxe_dev *rxe = mcg->rxe;
> +	struct rxe_mca *mca;
> +	unsigned long flags;
> +	int err;
>   
> -	n = atomic_inc_return(&rxe->mcg_attach);
> -	if (n > rxe->attr.max_total_mcast_qp_attach) {
> -		atomic_dec(&rxe->mcg_attach);
> -		return -ENOMEM;
> +	mutex_lock(&rxe->mcg_mutex);
> +	spin_lock_irqsave(&mcg->lock, flags);
> +	list_for_each_entry(mca, &mcg->qp_list, qp_list) {
> +		if (mca->qp == qp) {
> +			spin_unlock_irqrestore(&mcg->lock, flags);
> +			goto out;	/* nothing to do */
> +		}
>   	}
> +	spin_unlock_irqrestore(&mcg->lock, flags);
>   
> -	n = atomic_inc_return(&mcg->qp_num);
> -	if (n > rxe->attr.max_mcast_qp_attach) {
> -		atomic_dec(&mcg->qp_num);
> -		atomic_dec(&rxe->mcg_attach);
> -		return -ENOMEM;
> +	if (atomic_inc_return(&rxe->mcg_attach) >
> +	    rxe->attr.max_total_mcast_qp_attach) {
> +		err = -EINVAL;
> +		goto err_dec_attach;
>   	}
>   
> -	atomic_inc(&qp->mcg_num);
> +	if (atomic_inc_return(&mcg->qp_num) >
> +	    rxe->attr.max_mcast_qp_attach) {
> +		err = -EINVAL;
> +		goto err_dec_qp_num;
> +	}
> +
> +	mca = kzalloc(sizeof(*mca), GFP_KERNEL);
> +	if (!mca) {
> +		err = -ENOMEM;
> +		goto err_dec_qp_num;
> +	}
>   
> +	atomic_inc(&qp->mcg_num);
>   	rxe_get(qp);
>   	mca->qp = qp;
>   
> +	spin_lock_irqsave(&mcg->lock, flags);
>   	list_add_tail(&mca->qp_list, &mcg->qp_list);
> -
> +	spin_unlock_irqrestore(&mcg->lock, flags);
> +out:
> +	mutex_unlock(&rxe->mcg_mutex);
>   	return 0;
> +
> +err_dec_qp_num:
> +	atomic_dec(&mcg->qp_num);
> +err_dec_attach:
> +	atomic_dec(&rxe->mcg_attach);
> +	mutex_unlock(&rxe->mcg_mutex);
> +	return err;
>   }
>   
>   /**
> - * rxe_attach_mcg - attach qp to mcg if not already attached
> - * @qp: qp object
> + * rxe_detach_mcg - detach qp from mcg
>    * @mcg: mcg object
> + * @qp: qp object
>    *
> - * Context: caller must hold reference on qp and mcg.
> - * Returns: 0 on success else an error
> + * Returns: 0 on success else an error if qp is not attached.
>    */
> -static int rxe_attach_mcg(struct rxe_mcg *mcg, struct rxe_qp *qp)
> +static int rxe_detach_mcg(struct rxe_mcg *mcg, struct rxe_qp *qp)
>   {
>   	struct rxe_dev *rxe = mcg->rxe;
> -	struct rxe_mca *mca, *tmp;
> -	int err;
> +	struct rxe_mca *mca;
> +	unsigned long flags;
> +	int err = 0;
>   
> -	/* check to see if the qp is already a member of the group */
> -	spin_lock_bh(&rxe->mcg_lock);
> +	mutex_lock(&rxe->mcg_mutex);
> +	spin_lock_irqsave(&mcg->lock, flags);
>   	list_for_each_entry(mca, &mcg->qp_list, qp_list) {
>   		if (mca->qp == qp) {
> -			spin_unlock_bh(&rxe->mcg_lock);
> -			return 0;
> +			spin_unlock_irqrestore(&mcg->lock, flags);
> +			goto found;
>   		}
>   	}
> -	spin_unlock_bh(&rxe->mcg_lock);
> +	spin_unlock_irqrestore(&mcg->lock, flags);
>   
> -	/* speculative alloc new mca without using GFP_ATOMIC */
> -	mca = kzalloc(sizeof(*mca), GFP_KERNEL);
> -	if (!mca)
> -		return -ENOMEM;
> -
> -	spin_lock_bh(&rxe->mcg_lock);
> -	/* re-check to see if someone else just attached qp */
> -	list_for_each_entry(tmp, &mcg->qp_list, qp_list) {
> -		if (tmp->qp == qp) {
> -			kfree(mca);
> -			err = 0;
> -			goto out;
> -		}
> -	}
> -
> -	err = __rxe_init_mca(qp, mcg, mca);
> -	if (err)
> -		kfree(mca);
> -out:
> -	spin_unlock_bh(&rxe->mcg_lock);
> -	return err;
> -}
> +	/* we didn't find the qp on the list */
> +	err = -EINVAL;
> +	goto err_out;
>   
> -/**
> - * __rxe_cleanup_mca - cleanup mca object holding lock
> - * @mca: mca object
> - * @mcg: mcg object
> - *
> - * Context: caller must hold a reference to mcg and rxe->mcg_lock
> - */
> -static void __rxe_cleanup_mca(struct rxe_mca *mca, struct rxe_mcg *mcg)
> -{
> +found:
> +	spin_lock_irqsave(&mcg->lock, flags);
>   	list_del(&mca->qp_list);
> +	spin_unlock_irqrestore(&mcg->lock, flags);
>   
>   	atomic_dec(&mcg->qp_num);
>   	atomic_dec(&mcg->rxe->mcg_attach);
>   	atomic_dec(&mca->qp->mcg_num);
>   	rxe_put(mca->qp);
> -
>   	kfree(mca);
> -}
> -
> -/**
> - * rxe_detach_mcg - detach qp from mcg
> - * @mcg: mcg object
> - * @qp: qp object
> - *
> - * Returns: 0 on success else an error if qp is not attached.
> - */
> -static int rxe_detach_mcg(struct rxe_mcg *mcg, struct rxe_qp *qp)
> -{
> -	struct rxe_dev *rxe = mcg->rxe;
> -	struct rxe_mca *mca, *tmp;
>   
> -	spin_lock_bh(&rxe->mcg_lock);
> -	list_for_each_entry_safe(mca, tmp, &mcg->qp_list, qp_list) {
> -		if (mca->qp == qp) {
> -			__rxe_cleanup_mca(mca, mcg);
> -
> -			/* if the number of qp's attached to the
> -			 * mcast group falls to zero go ahead and
> -			 * tear it down. This will not free the
> -			 * object since we are still holding a ref
> -			 * from the caller
> -			 */
> -			if (atomic_read(&mcg->qp_num) <= 0)
> -				__rxe_destroy_mcg(mcg);
> -
> -			spin_unlock_bh(&rxe->mcg_lock);
> -			return 0;
> -		}
> -	}
> +	/* if the number of qp's attached to the
> +	 * mcast group falls to zero go ahead and
> +	 * tear it down. This will not free the
> +	 * object since we are still holding a ref
> +	 * from the caller
> +	 */
> +	if (atomic_read(&mcg->qp_num) <= 0)
> +		__rxe_destroy_mcg(mcg);
>   
> -	/* we didn't find the qp on the list */
> -	spin_unlock_bh(&rxe->mcg_lock);
> -	return -EINVAL;
> +err_out:
> +	mutex_unlock(&rxe->mcg_mutex);
> +	return err;
>   }
>   
>   /**
> @@ -497,6 +438,9 @@ int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
>   	struct rxe_qp *qp = to_rqp(ibqp);
>   	struct rxe_mcg *mcg;
>   
> +	if (rxe->attr.max_mcast_grp == 0)
> +		return -EINVAL;
> +
>   	/* takes a ref on mcg if successful */
>   	mcg = rxe_get_mcg(rxe, mgid);
>   	if (IS_ERR(mcg))
> diff --git a/drivers/infiniband/sw/rxe/rxe_recv.c b/drivers/infiniband/sw/rxe/rxe_recv.c
> index 7153de0799fc..6cf0da958864 100644
> --- a/drivers/infiniband/sw/rxe/rxe_recv.c
> +++ b/drivers/infiniband/sw/rxe/rxe_recv.c
> @@ -194,6 +194,7 @@ static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb)
>   	struct rxe_mca *mca;
>   	struct rxe_qp *qp;
>   	union ib_gid dgid;
> +	unsigned long flags;
>   	int err;
>   
>   	if (skb->protocol == htons(ETH_P_IP))
> @@ -207,7 +208,7 @@ static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb)
>   	if (!mcg)
>   		goto drop;	/* mcast group not registered */
>   
> -	spin_lock_bh(&rxe->mcg_lock);
> +	spin_lock_irqsave(&mcg->lock, flags);
>   
>   	/* this is unreliable datagram service so we let
>   	 * failures to deliver a multicast packet to a
> @@ -259,7 +260,7 @@ static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb)
>   		}
>   	}
>   
> -	spin_unlock_bh(&rxe->mcg_lock);
> +	spin_unlock_irqrestore(&mcg->lock, flags);
>   
>   	kref_put(&mcg->ref_cnt, rxe_cleanup_mcg);
>   
> diff --git a/drivers/infiniband/sw/rxe/rxe_verbs.h b/drivers/infiniband/sw/rxe/rxe_verbs.h
> index 8058e5039322..f21963dcb2c8 100644
> --- a/drivers/infiniband/sw/rxe/rxe_verbs.h
> +++ b/drivers/infiniband/sw/rxe/rxe_verbs.h
> @@ -351,6 +351,7 @@ struct rxe_mcg {
>   	struct list_head	qp_list;
>   	union ib_gid		mgid;
>   	atomic_t		qp_num;
> +	spinlock_t		lock;	/* protect qp_list */
>   	u32			qkey;
>   	u16			pkey;
>   	bool			is_ipv6;
> @@ -390,7 +391,7 @@ struct rxe_dev {
>   	struct rxe_pool		mw_pool;
>   
>   	/* multicast support */
> -	spinlock_t		mcg_lock;
> +	struct mutex		mcg_mutex;
>   	struct rb_root		mcg_tree;
>   	atomic_t		mcg_num;
>   	atomic_t		mcg_attach;

Powered by blists - more mailing lists