Skip to contentSkip to conversation

RDMA/rxe: Correct race conditions

PATCH Superseded Bob Pearson 2 participants 13 messages Dec 16, 2021 – Jan 7, 2022

Subsystems (2)

v8 of 9 Superseded Latest is v9 →

lore
Share & tools
lore
#0 lore raw

v8 cover RDMA/rxe: Correct race conditions

There are several race conditions discovered in the current rdma_rxe driver. They mostly relate to races between normal operations and destroying objects. This patch series
- Makes several minor cleanups in rxe_pool.[ch]
- Replaces the red-black trees currently used by xarrays for indices
- Simplifies the API for keyed objects
- Corrects several reference counting errors
- Adds wait for completions to the paths in verbs APIs which destroy objects.

This patch series applies cleanly to current for-next. commit c8f476da84ad ("Merge branch 'mlx5-next' of
git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux")

---
v8
Fixed an additional race in 3/8 which was not handled correctly.
v7
Corrected issues reported by Jason Gunthorpe

v6
Fixed a kzalloc flags bug.
Fixed comment bug reported by 'Kernel Test Robot'.
Changed type of rxe_pool.c in __rxe_fini().
v5
Removed patches already accepted into for-next and addressed comments from Jason Gunthorpe.
v4
Restructured patch series to change to xarray earlier which greatly simplified the changes.
Rebased to current for-next
v3
Changed rxe_alloc to use GFP_KERNEL
Addressed other comments by Jason Gunthorp
Merged the previous 06/10 and 07/10 patches into one since they overlapped Added some minor cleanups as 10/10
v2
Rebased to current for-next.
Added 4 additional patches

Bob Pearson (8):
RDMA/rxe: Replace RB tree by xarray for indexes RDMA/rxe: Reverse the sense of RXE_POOL_NO_ALLOC RDMA/rxe: Cleanup pool APIs for keyed objects RDMA/rxe: Fix ref error in rxe_av.c
RDMA/rxe: Replace mr by rkey in responder resources RDMA/rxe: Minor cleanups in rxe_pool.c/rxe_pool.h RDMA/rxe: Replace rxe_alloc by kzalloc for rxe_mc_elem RDMA/rxe: Add wait for completion to obj destruct

 drivers/infiniband/sw/rxe/rxe.c       | 101 +----
 drivers/infiniband/sw/rxe/rxe_av.c    |  19 +-
 drivers/infiniband/sw/rxe/rxe_loc.h   |  10 +-
 drivers/infiniband/sw/rxe/rxe_mcast.c |  76 ++--
 drivers/infiniband/sw/rxe/rxe_mr.c    |   3 +-
 drivers/infiniband/sw/rxe/rxe_mw.c    |   7 +-
 drivers/infiniband/sw/rxe/rxe_net.c   |  17 +-
 drivers/infiniband/sw/rxe/rxe_pool.c  | 529 ++++++++++++--------------
 drivers/infiniband/sw/rxe/rxe_pool.h  | 110 ++----
 drivers/infiniband/sw/rxe/rxe_qp.c    |  10 +-
 drivers/infiniband/sw/rxe/rxe_req.c   |  55 +--
 drivers/infiniband/sw/rxe/rxe_resp.c  | 125 ++++--
 drivers/infiniband/sw/rxe/rxe_verbs.c |  72 ++--
 drivers/infiniband/sw/rxe/rxe_verbs.h |   3 -
 14 files changed, 515 insertions(+), 622 deletions(-)
Signature · 1 line
2.32.0
#1 lore raw

PATCH v8 1/8 RDMA/rxe: Replace RB tree by xarray for indexes

Currently the rxe driver uses red-black trees to add indices and keys to the rxe object pool. Linux xarrays provide a better way to implement the same functionality for indices but not keys. This patch replaces red-black trees by xarrays for indexed objects. Since caller managed locks for indexed objects are not used these APIs are deleted as well.

To avoid double locking since xarray already includes a spinlock replace the rxe_pool rwlock by the spinlock included in xarray.

The RDMA objects are created and destroyed by verbs calls from rdma_core but are looked up from indices or keys from soft IRQs so _bh style locks are the correct type to use.

---
 drivers/infiniband/sw/rxe/rxe.c       | 100 ++----------
 drivers/infiniband/sw/rxe/rxe_mcast.c |   6 +-
 drivers/infiniband/sw/rxe/rxe_mr.c    |   1 -
 drivers/infiniband/sw/rxe/rxe_mw.c    |   4 -
 drivers/infiniband/sw/rxe/rxe_pool.c  | 221 ++++++++------------------
 drivers/infiniband/sw/rxe/rxe_pool.h  |  77 ++++-----
 drivers/infiniband/sw/rxe/rxe_verbs.c |  12 --
 7 files changed, 112 insertions(+), 309 deletions(-)
drivers/infiniband/sw/rxe/rxe.c +14 -86
index 8e0f9c489cab..09c73a0d8513 100644
@@ -116,97 +116,31 @@ static void rxe_init_ports(struct rxe_dev *rxe)
 }
 
 /* init pools of managed objects */
-static int rxe_init_pools(struct rxe_dev *rxe)
+static void rxe_init_pools(struct rxe_dev *rxe)
 {
-	int err;
-
-	err = rxe_pool_init(rxe, &rxe->uc_pool, RXE_TYPE_UC,
-			    rxe->max_ucontext);
-	if (err)
-		goto err1;
-
-	err = rxe_pool_init(rxe, &rxe->pd_pool, RXE_TYPE_PD,
-			    rxe->attr.max_pd);
-	if (err)
-		goto err2;
-
-	err = rxe_pool_init(rxe, &rxe->ah_pool, RXE_TYPE_AH,
-			    rxe->attr.max_ah);
-	if (err)
-		goto err3;
-
-	err = rxe_pool_init(rxe, &rxe->srq_pool, RXE_TYPE_SRQ,
-			    rxe->attr.max_srq);
-	if (err)
-		goto err4;
-
-	err = rxe_pool_init(rxe, &rxe->qp_pool, RXE_TYPE_QP,
-			    rxe->attr.max_qp);
-	if (err)
-		goto err5;
-
-	err = rxe_pool_init(rxe, &rxe->cq_pool, RXE_TYPE_CQ,
-			    rxe->attr.max_cq);
-	if (err)
-		goto err6;
-
-	err = rxe_pool_init(rxe, &rxe->mr_pool, RXE_TYPE_MR,
-			    rxe->attr.max_mr);
-	if (err)
-		goto err7;
-
-	err = rxe_pool_init(rxe, &rxe->mw_pool, RXE_TYPE_MW,
-			    rxe->attr.max_mw);
-	if (err)
-		goto err8;
-
-	err = rxe_pool_init(rxe, &rxe->mc_grp_pool, RXE_TYPE_MC_GRP,
+	rxe_pool_init(rxe, &rxe->uc_pool, RXE_TYPE_UC, rxe->max_ucontext);
+	rxe_pool_init(rxe, &rxe->pd_pool, RXE_TYPE_PD, rxe->attr.max_pd);
+	rxe_pool_init(rxe, &rxe->ah_pool, RXE_TYPE_AH, rxe->attr.max_ah);
+	rxe_pool_init(rxe, &rxe->srq_pool, RXE_TYPE_SRQ, rxe->attr.max_srq);
+	rxe_pool_init(rxe, &rxe->qp_pool, RXE_TYPE_QP, rxe->attr.max_qp);
+	rxe_pool_init(rxe, &rxe->cq_pool, RXE_TYPE_CQ, rxe->attr.max_cq);
+	rxe_pool_init(rxe, &rxe->mr_pool, RXE_TYPE_MR, rxe->attr.max_mr);
+	rxe_pool_init(rxe, &rxe->mw_pool, RXE_TYPE_MW, rxe->attr.max_mw);
+	rxe_pool_init(rxe, &rxe->mc_grp_pool, RXE_TYPE_MC_GRP,
 			    rxe->attr.max_mcast_grp);
-	if (err)
-		goto err9;
-
-	err = rxe_pool_init(rxe, &rxe->mc_elem_pool, RXE_TYPE_MC_ELEM,
+	rxe_pool_init(rxe, &rxe->mc_elem_pool, RXE_TYPE_MC_ELEM,
 			    rxe->attr.max_total_mcast_qp_attach);
-	if (err)
-		goto err10;
-
-	return 0;
-
-err10:
-	rxe_pool_cleanup(&rxe->mc_grp_pool);
-err9:
-	rxe_pool_cleanup(&rxe->mw_pool);
-err8:
-	rxe_pool_cleanup(&rxe->mr_pool);
-err7:
-	rxe_pool_cleanup(&rxe->cq_pool);
-err6:
-	rxe_pool_cleanup(&rxe->qp_pool);
-err5:
-	rxe_pool_cleanup(&rxe->srq_pool);
-err4:
-	rxe_pool_cleanup(&rxe->ah_pool);
-err3:
-	rxe_pool_cleanup(&rxe->pd_pool);
-err2:
-	rxe_pool_cleanup(&rxe->uc_pool);
-err1:
-	return err;
 }
 
 /* initialize rxe device state */
-static int rxe_init(struct rxe_dev *rxe)
+static void rxe_init(struct rxe_dev *rxe)
 {
-	int err;
-
 	/* init default device parameters */
 	rxe_init_device_param(rxe);
 
 	rxe_init_ports(rxe);
 
-	err = rxe_init_pools(rxe);
-	if (err)
-		return err;
+	rxe_init_pools(rxe);
 
 	/* init pending mmap list */
 	spin_lock_init(&rxe->mmap_offset_lock);
@@ -214,8 +148,6 @@ static int rxe_init(struct rxe_dev *rxe)
 	INIT_LIST_HEAD(&rxe->pending_mmaps);
 
 	mutex_init(&rxe->usdev_lock);
-
-	return 0;
 }
 
 void rxe_set_mtu(struct rxe_dev *rxe, unsigned int ndev_mtu)
@@ -237,11 +169,7 @@ void rxe_set_mtu(struct rxe_dev *rxe, unsigned int ndev_mtu)
  */
 int rxe_add(struct rxe_dev *rxe, unsigned int mtu, const char *ibdev_name)
 {
-	int err;
-
-	err = rxe_init(rxe);
-	if (err)
-		return err;
+	rxe_init(rxe);
 
 	rxe_set_mtu(rxe, mtu);
 
drivers/infiniband/sw/rxe/rxe_mcast.c +3 -3
index bd1ac88b8700..1692526c5b57 100644
@@ -44,7 +44,7 @@ int rxe_mcast_get_grp(struct rxe_dev *rxe, union ib_gid *mgid,
 	if (rxe->attr.max_mcast_qp_attach == 0)
 		return -EINVAL;
 
-	write_lock_bh(&pool->pool_lock);
+	rxe_pool_lock_bh(pool);
 
 	grp = rxe_pool_get_key_locked(pool, mgid);
 	if (grp)
@@ -52,13 +52,13 @@ int rxe_mcast_get_grp(struct rxe_dev *rxe, union ib_gid *mgid,
 
 	grp = create_grp(rxe, pool, mgid);
 	if (IS_ERR(grp)) {
-		write_unlock_bh(&pool->pool_lock);
+		rxe_pool_unlock_bh(pool);
 		err = PTR_ERR(grp);
 		return err;
 	}
 
 done:
-	write_unlock_bh(&pool->pool_lock);
+	rxe_pool_unlock_bh(pool);
 	*grp_p = grp;
 	return 0;
 }
drivers/infiniband/sw/rxe/rxe_mr.c +0 -1
index 25c78aade822..3c4390adfb80 100644
@@ -693,7 +693,6 @@ int rxe_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
 
 	mr->state = RXE_MR_STATE_INVALID;
 	rxe_drop_ref(mr_pd(mr));
-	rxe_drop_index(mr);
 	rxe_drop_ref(mr);
 
 	return 0;
drivers/infiniband/sw/rxe/rxe_mw.c +0 -4
index 32dd8c0b8b9e..3ae981d77c25 100644
@@ -20,7 +20,6 @@ int rxe_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
 		return ret;
 	}
 
-	rxe_add_index(mw);
 	mw->rkey = ibmw->rkey = (mw->elem.index << 8) | rxe_get_next_key(-1);
 	mw->state = (mw->ibmw.type == IB_MW_TYPE_2) ?
 			RXE_MW_STATE_FREE : RXE_MW_STATE_VALID;
@@ -332,7 +331,4 @@ struct rxe_mw *rxe_lookup_mw(struct rxe_qp *qp, int access, u32 rkey)
 
 void rxe_mw_cleanup(struct rxe_pool_elem *elem)
 {
-	struct rxe_mw *mw = container_of(elem, typeof(*mw), elem);
-
-	rxe_drop_index(mw);
 }
drivers/infiniband/sw/rxe/rxe_pool.c +62 -159

334 lines not on this page · raw message

drivers/infiniband/sw/rxe/rxe_pool.h +33 -44

119 lines not on this page · raw message

drivers/infiniband/sw/rxe/rxe_verbs.c +0 -12
index 07ca169110bf..e3f64eae088c 100644
@@ -181,7 +181,6 @@ static int rxe_create_ah(struct ib_ah *ibah,
 		return err;
 
 	/* create index > 0 */
-	rxe_add_index(ah);
 	ah->ah_num = ah->elem.index;
 
 	if (uresp) {
@@ -189,7 +188,6 @@ static int rxe_create_ah(struct ib_ah *ibah,
 		err = copy_to_user(&uresp->ah_num, &ah->ah_num,
 					 sizeof(uresp->ah_num));
 		if (err) {
-			rxe_drop_index(ah);
 			rxe_drop_ref(ah);
 			return -EFAULT;
 		}
@@ -230,7 +228,6 @@ static int rxe_destroy_ah(struct ib_ah *ibah, u32 flags)
 {
 	struct rxe_ah *ah = to_rah(ibah);
 
-	rxe_drop_index(ah);
 	rxe_drop_ref(ah);
 	return 0;
 }
@@ -437,7 +434,6 @@ static int rxe_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init,
 	if (err)
 		return err;
 
-	rxe_add_index(qp);
 	err = rxe_qp_from_init(rxe, qp, pd, init, uresp, ibqp->pd, udata);
 	if (err)
 		goto qp_init;
@@ -445,7 +441,6 @@ static int rxe_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init,
 	return 0;
 
 qp_init:
-	rxe_drop_index(qp);
 	rxe_drop_ref(qp);
 	return err;
 }
@@ -490,7 +485,6 @@ static int rxe_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 	struct rxe_qp *qp = to_rqp(ibqp);
 
 	rxe_qp_destroy(qp);
-	rxe_drop_index(qp);
 	rxe_drop_ref(qp);
 	return 0;
 }
@@ -893,7 +887,6 @@ static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access)
 	if (!mr)
 		return ERR_PTR(-ENOMEM);
 
-	rxe_add_index(mr);
 	rxe_add_ref(pd);
 	rxe_mr_init_dma(pd, access, mr);
 
@@ -917,7 +910,6 @@ static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd,
 		goto err2;
 	}
 
-	rxe_add_index(mr);
 
 	rxe_add_ref(pd);
 
@@ -929,7 +921,6 @@ static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd,
 
 err3:
 	rxe_drop_ref(pd);
-	rxe_drop_index(mr);
 	rxe_drop_ref(mr);
 err2:
 	return ERR_PTR(err);
@@ -952,8 +943,6 @@ static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
 		goto err1;
 	}
 
-	rxe_add_index(mr);
-
 	rxe_add_ref(pd);
 
 	err = rxe_mr_init_fast(pd, max_num_sg, mr);
@@ -964,7 +953,6 @@ static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
 
 err2:
 	rxe_drop_ref(pd);
-	rxe_drop_index(mr);
 	rxe_drop_ref(mr);
 err1:
 	return ERR_PTR(err);
Signature · 1 line
2.32.0
maintainer
#2 lore raw

Re: [PATCH for-next v8 1/8] RDMA/rxe: Replace RB tree by xarray for indexes

Quoted 12 lines · Bob Pearson
drivers/infiniband/sw/rxe/rxe_pool.c +10 -0
+/**
+ * rxe_pool_get_index - lookup object from index
+ * @pool: the object pool
+ * @index: the index of the object
+ *
+ * Returns: the object if the index exists in the pool
+ * and the reference count on the object is positive
+ * else NULL
+ */
+void *rxe_pool_get_index(struct rxe_pool *pool, u32 index)
 {
 	struct rxe_pool_elem *elem;

drivers/infiniband/sw/rxe/rxe_pool.c void rxe_elem_release(struct kref *kref) · line 353

Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_pool.c +3 -0
 	void *obj;
 
+	elem = xa_load(&pool->xarray.xa, index);
+	if (elem && kref_get_unless_zero(&elem->ref_cnt))
 		obj = elem->obj;
+	else
 		obj = NULL;
 
 	return obj;
 }

This has to hold some kind of lock to protect elem from free while calling kref_get_unless_zero().

You can hold the xa_lock() around this, or use rcu_read_lock(), but that requires using kfree_rcu to free the memory.

Jason

#3 lore raw

PATCH v8 2/8 RDMA/rxe: Reverse the sense of RXE_POOL_NO_ALLOC

In reply to Bob Pearson

Since most rxe objects are now allocated in rdma-core change the sense of RXE_POOL_NO_ALLOC to RXE_POOL_ALLOC. This makes the code easier to understand.

---
 drivers/infiniband/sw/rxe/rxe_pool.c | 18 ++++++++----------
 drivers/infiniband/sw/rxe/rxe_pool.h |  2 +-
 2 files changed, 9 insertions(+), 11 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +8 -10
index ff7ca2dbcb0a..d1981309aa23 100644
@@ -23,19 +23,17 @@ static const struct rxe_type_info {
 		.name		= "rxe-uc",
 		.size		= sizeof(struct rxe_ucontext),
 		.elem_offset	= offsetof(struct rxe_ucontext, elem),
-		.flags          = RXE_POOL_NO_ALLOC,
 	},
 	[RXE_TYPE_PD] = {
 		.name		= "rxe-pd",
 		.size		= sizeof(struct rxe_pd),
 		.elem_offset	= offsetof(struct rxe_pd, elem),
-		.flags		= RXE_POOL_NO_ALLOC,
 	},
 	[RXE_TYPE_AH] = {
 		.name		= "rxe-ah",
 		.size		= sizeof(struct rxe_ah),
 		.elem_offset	= offsetof(struct rxe_ah, elem),
-		.flags		= RXE_POOL_INDEX | RXE_POOL_NO_ALLOC,
+		.flags		= RXE_POOL_INDEX,
 		.min_index	= RXE_MIN_AH_INDEX,
 		.max_index	= RXE_MAX_AH_INDEX,
 	},
@@ -43,7 +41,7 @@ static const struct rxe_type_info {
 		.name		= "rxe-srq",
 		.size		= sizeof(struct rxe_srq),
 		.elem_offset	= offsetof(struct rxe_srq, elem),
-		.flags		= RXE_POOL_INDEX | RXE_POOL_NO_ALLOC,
+		.flags		= RXE_POOL_INDEX,
 		.min_index	= RXE_MIN_SRQ_INDEX,
 		.max_index	= RXE_MAX_SRQ_INDEX,
 	},
@@ -52,7 +50,7 @@ static const struct rxe_type_info {
 		.size		= sizeof(struct rxe_qp),
 		.elem_offset	= offsetof(struct rxe_qp, elem),
 		.cleanup	= rxe_qp_cleanup,
-		.flags		= RXE_POOL_INDEX | RXE_POOL_NO_ALLOC,
+		.flags		= RXE_POOL_INDEX,
 		.min_index	= RXE_MIN_QP_INDEX,
 		.max_index	= RXE_MAX_QP_INDEX,
 	},
@@ -60,7 +58,6 @@ static const struct rxe_type_info {
 		.name		= "rxe-cq",
 		.size		= sizeof(struct rxe_cq),
 		.elem_offset	= offsetof(struct rxe_cq, elem),
-		.flags          = RXE_POOL_NO_ALLOC,
 		.cleanup	= rxe_cq_cleanup,
 	},
 	[RXE_TYPE_MR] = {
@@ -68,7 +65,7 @@ static const struct rxe_type_info {
 		.size		= sizeof(struct rxe_mr),
 		.elem_offset	= offsetof(struct rxe_mr, elem),
 		.cleanup	= rxe_mr_cleanup,
-		.flags		= RXE_POOL_INDEX,
+		.flags		= RXE_POOL_INDEX | RXE_POOL_ALLOC,
 		.min_index	= RXE_MIN_MR_INDEX,
 		.max_index	= RXE_MAX_MR_INDEX,
 	},
@@ -77,7 +74,7 @@ static const struct rxe_type_info {
 		.size		= sizeof(struct rxe_mw),
 		.elem_offset	= offsetof(struct rxe_mw, elem),
 		.cleanup	= rxe_mw_cleanup,
-		.flags		= RXE_POOL_INDEX | RXE_POOL_NO_ALLOC,
+		.flags		= RXE_POOL_INDEX,
 		.min_index	= RXE_MIN_MW_INDEX,
 		.max_index	= RXE_MAX_MW_INDEX,
 	},
@@ -86,7 +83,7 @@ static const struct rxe_type_info {
 		.size		= sizeof(struct rxe_mc_grp),
 		.elem_offset	= offsetof(struct rxe_mc_grp, elem),
 		.cleanup	= rxe_mc_cleanup,
-		.flags		= RXE_POOL_KEY,
+		.flags		= RXE_POOL_KEY | RXE_POOL_ALLOC,
 		.key_offset	= offsetof(struct rxe_mc_grp, mgid),
 		.key_size	= sizeof(union ib_gid),
 	},
@@ -94,6 +91,7 @@ static const struct rxe_type_info {
 		.name		= "rxe-mc_elem",
 		.size		= sizeof(struct rxe_mc_elem),
 		.elem_offset	= offsetof(struct rxe_mc_elem, elem),
+		.flags		= RXE_POOL_ALLOC,
 	},
 };
 
@@ -321,7 +319,7 @@ void rxe_elem_release(struct kref *kref)
 	if (pool->cleanup)
 		pool->cleanup(elem);
 
-	if (!(pool->flags & RXE_POOL_NO_ALLOC)) {
+	if (pool->flags & RXE_POOL_ALLOC) {
 		obj = elem->obj;
 		kfree(obj);
 	}
drivers/infiniband/sw/rxe/rxe_pool.h +1 -1
index fbef351783dc..514be1582bce 100644
@@ -10,7 +10,7 @@
 enum rxe_pool_flags {
 	RXE_POOL_INDEX		= BIT(1),
 	RXE_POOL_KEY		= BIT(2),
-	RXE_POOL_NO_ALLOC	= BIT(4),
+	RXE_POOL_ALLOC		= BIT(4),
 };
 
 enum rxe_elem_type {
Signature · 1 line
2.32.0
#4 lore raw

PATCH v8 3/8 RDMA/rxe: Cleanup pool APIs for keyed objects

In reply to Bob Pearson

Simplify the rxe pool APIs for keyed objects. Eliminate xxx_locked() APIs. Merge rxe_drop_key into rxe_drop_ref. Replace separate rxe_get_key, and add_key by one call rxe_add_key which looks up and if necessary allocates a new object. This routine is written to avoid using GFP_ATOMIC by maintaining a free list of objects allocated without holding a spinlock.

---
v8
Corrected an additional race condition where another tread uses the remaining objects on the free list after this thread creates them before they are used.

 drivers/infiniband/sw/rxe/rxe_loc.h   |   5 +-
 drivers/infiniband/sw/rxe/rxe_mcast.c |  46 ++--
 drivers/infiniband/sw/rxe/rxe_pool.c  | 297 ++++++++++++++------------
 drivers/infiniband/sw/rxe/rxe_pool.h  |  59 +----
 4 files changed, 187 insertions(+), 220 deletions(-)
drivers/infiniband/sw/rxe/rxe_loc.h +1 -4
index b1e174afb1d4..6558602be751 100644
@@ -40,17 +40,14 @@ void rxe_cq_disable(struct rxe_cq *cq);
 void rxe_cq_cleanup(struct rxe_pool_elem *arg);
 
 /* rxe_mcast.c */
+int rxe_init_grp(struct rxe_pool_elem *elem);
 int rxe_mcast_get_grp(struct rxe_dev *rxe, union ib_gid *mgid,
 		      struct rxe_mc_grp **grp_p);
-
 int rxe_mcast_add_grp_elem(struct rxe_dev *rxe, struct rxe_qp *qp,
 			   struct rxe_mc_grp *grp);
-
 int rxe_mcast_drop_grp_elem(struct rxe_dev *rxe, struct rxe_qp *qp,
 			    union ib_gid *mgid);
-
 void rxe_drop_all_mcast_groups(struct rxe_qp *qp);
-
 void rxe_mc_cleanup(struct rxe_pool_elem *arg);
 
 /* rxe_mmap.c */
drivers/infiniband/sw/rxe/rxe_mcast.c +12 -34

84 lines not on this page · raw message

drivers/infiniband/sw/rxe/rxe_pool.c +166 -131

389 lines not on this page · raw message

drivers/infiniband/sw/rxe/rxe_pool.h +8 -51

95 lines not on this page · raw message

Signature · 1 line
2.32.0
maintainer
#5 lore raw

Re: [PATCH for-next v8 3/8] RDMA/rxe: Cleanup pool APIs for keyed objects

Quoted 6 lines · Bob Pearson
drivers/infiniband/sw/rxe/rxe_pool.c +6 -0
+/**
+ * rxe_elem_release() - cleanup pool element when last reference dropped
+ * @kref: address of the kref contained in pool element
+ *
+ * Caller should hold pool lock
+ */

drivers/infiniband/sw/rxe/rxe_pool.c void *rxe_pool_get_key_locked(struct rxe_pool *pool, void *key) · line 421

Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_pool.c +10 -0
+void rxe_elem_release(struct kref *kref)
+{
+	struct rxe_pool_elem *elem =
+		container_of(kref, struct rxe_pool_elem, ref_cnt);
+	struct rxe_pool *pool = elem->pool;
+
+	if (pool->flags & RXE_POOL_INDEX)
+		xa_erase(&pool->xarray.xa, elem->index);
+	else if (pool->flags & RXE_POOL_KEY)
+		rb_erase(&elem->key_node, &pool->key.tree);

This doesn't hold a lock anymore before calling rb_erase()

Jason

#6 lore raw

PATCH v8 4/8 RDMA/rxe: Fix ref error in rxe_av.c

In reply to Bob Pearson

The commit referenced below can take a reference to the AH which is never dropped. This only happens in the UD request path. This patch optionally passes that AH back to the caller so that it can hold the reference while the AV is being accessed and then drop it. Code to do this is added to rxe_req.c. The AV is also passed to rxe_prepare in rxe_net.c as an optimization.

Fixes e2fe06c90806 ("RDMA/rxe: Lookup kernel AH from ah index in UD WQEs")
---
 drivers/infiniband/sw/rxe/rxe_av.c   | 19 +++++++++-
 drivers/infiniband/sw/rxe/rxe_loc.h  |  5 ++-
 drivers/infiniband/sw/rxe/rxe_net.c  | 17 +++++----
 drivers/infiniband/sw/rxe/rxe_req.c  | 55 +++++++++++++++++-----------
 drivers/infiniband/sw/rxe/rxe_resp.c |  2 +-
 5 files changed, 63 insertions(+), 35 deletions(-)
drivers/infiniband/sw/rxe/rxe_av.c +17 -2
index 38c7b6fb39d7..360a567159fe 100644
@@ -99,11 +99,14 @@ void rxe_av_fill_ip_info(struct rxe_av *av, struct rdma_ah_attr *attr)
 	av->network_type = type;
 }
 
-struct rxe_av *rxe_get_av(struct rxe_pkt_info *pkt)
+struct rxe_av *rxe_get_av(struct rxe_pkt_info *pkt, struct rxe_ah **ahp)
 {
 	struct rxe_ah *ah;
 	u32 ah_num;
 
+	if (ahp)
+		*ahp = NULL;
+
 	if (!pkt || !pkt->qp)
 		return NULL;
 
@@ -117,10 +120,22 @@ struct rxe_av *rxe_get_av(struct rxe_pkt_info *pkt)
 	if (ah_num) {
 		/* only new user provider or kernel client */
 		ah = rxe_pool_get_index(&pkt->rxe->ah_pool, ah_num);
-		if (!ah || ah->ah_num != ah_num || rxe_ah_pd(ah) != pkt->qp->pd) {
+		if (!ah) {
 			pr_warn("Unable to find AH matching ah_num\n");
 			return NULL;
 		}
+
+		if (rxe_ah_pd(ah) != pkt->qp->pd) {
+			pr_warn("PDs don't match for AH and QP\n");
+			rxe_drop_ref(ah);
+			return NULL;
+		}
+
+		if (ahp)
+			*ahp = ah;
+		else
+			rxe_drop_ref(ah);
+
 		return &ah->av;
 	}
 
drivers/infiniband/sw/rxe/rxe_loc.h +3 -2
index 6558602be751..02d57c894e34 100644
@@ -19,7 +19,7 @@ void rxe_av_to_attr(struct rxe_av *av, struct rdma_ah_attr *attr);
 
 void rxe_av_fill_ip_info(struct rxe_av *av, struct rdma_ah_attr *attr);
 
-struct rxe_av *rxe_get_av(struct rxe_pkt_info *pkt);
+struct rxe_av *rxe_get_av(struct rxe_pkt_info *pkt, struct rxe_ah **ahp);
 
 /* rxe_cq.c */
 int rxe_cq_chk_attr(struct rxe_dev *rxe, struct rxe_cq *cq,
@@ -99,7 +99,8 @@ void rxe_mw_cleanup(struct rxe_pool_elem *arg);
 /* rxe_net.c */
 struct sk_buff *rxe_init_packet(struct rxe_dev *rxe, struct rxe_av *av,
 				int paylen, struct rxe_pkt_info *pkt);
-int rxe_prepare(struct rxe_pkt_info *pkt, struct sk_buff *skb);
+int rxe_prepare(struct rxe_av *av, struct rxe_pkt_info *pkt,
+		struct sk_buff *skb);
 int rxe_xmit_packet(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
 		    struct sk_buff *skb);
 const char *rxe_parent_name(struct rxe_dev *rxe, unsigned int port_num);
drivers/infiniband/sw/rxe/rxe_net.c +9 -8
index 2cb810cb890a..456e960cacd7 100644
@@ -293,13 +293,13 @@ static void prepare_ipv6_hdr(struct dst_entry *dst, struct sk_buff *skb,
 	ip6h->payload_len = htons(skb->len - sizeof(*ip6h));
 }
 
-static int prepare4(struct rxe_pkt_info *pkt, struct sk_buff *skb)
+static int prepare4(struct rxe_av *av, struct rxe_pkt_info *pkt,
+		    struct sk_buff *skb)
 {
 	struct rxe_qp *qp = pkt->qp;
 	struct dst_entry *dst;
 	bool xnet = false;
 	__be16 df = htons(IP_DF);
-	struct rxe_av *av = rxe_get_av(pkt);
 	struct in_addr *saddr = &av->sgid_addr._sockaddr_in.sin_addr;
 	struct in_addr *daddr = &av->dgid_addr._sockaddr_in.sin_addr;
 
@@ -319,11 +319,11 @@ static int prepare4(struct rxe_pkt_info *pkt, struct sk_buff *skb)
 	return 0;
 }
 
-static int prepare6(struct rxe_pkt_info *pkt, struct sk_buff *skb)
+static int prepare6(struct rxe_av *av, struct rxe_pkt_info *pkt,
+		    struct sk_buff *skb)
 {
 	struct rxe_qp *qp = pkt->qp;
 	struct dst_entry *dst;
-	struct rxe_av *av = rxe_get_av(pkt);
 	struct in6_addr *saddr = &av->sgid_addr._sockaddr_in6.sin6_addr;
 	struct in6_addr *daddr = &av->dgid_addr._sockaddr_in6.sin6_addr;
 
@@ -344,16 +344,17 @@ static int prepare6(struct rxe_pkt_info *pkt, struct sk_buff *skb)
 	return 0;
 }
 
-int rxe_prepare(struct rxe_pkt_info *pkt, struct sk_buff *skb)
+int rxe_prepare(struct rxe_av *av, struct rxe_pkt_info *pkt,
+		struct sk_buff *skb)
 {
 	int err = 0;
 
 	if (skb->protocol == htons(ETH_P_IP))
-		err = prepare4(pkt, skb);
+		err = prepare4(av, pkt, skb);
 	else if (skb->protocol == htons(ETH_P_IPV6))
-		err = prepare6(pkt, skb);
+		err = prepare6(av, pkt, skb);
 
-	if (ether_addr_equal(skb->dev->dev_addr, rxe_get_av(pkt)->dmac))
+	if (ether_addr_equal(skb->dev->dev_addr, av->dmac))
 		pkt->mask |= RXE_LOOPBACK_MASK;
 
 	return err;
drivers/infiniband/sw/rxe/rxe_req.c +33 -22

121 lines not on this page · raw message

drivers/infiniband/sw/rxe/rxe_resp.c +1 -1
index e8f435fa6e4d..f589f4dde35c 100644
@@ -632,7 +632,7 @@ static struct sk_buff *prepare_ack_packet(struct rxe_qp *qp,
 	if (ack->mask & RXE_ATMACK_MASK)
 		atmack_set_orig(ack, qp->resp.atomic_orig);
 
-	err = rxe_prepare(ack, skb);
+	err = rxe_prepare(&qp->pri_av, ack, skb);
 	if (err) {
 		kfree_skb(skb);
 		return NULL;
Signature · 1 line
2.32.0
#7 lore raw

PATCH v8 5/8 RDMA/rxe: Replace mr by rkey in responder resources

In reply to Bob Pearson

Currently rxe saves a copy of MR in responder resources for RDMA reads. Since the responder resources are never freed just over written if more are needed this MR may not have a reference freed until the QP is destroyed. This patch uses the rkey instead of the MR and on subsequent packets of a multipacket read reply message it looks up the MR from the rkey for each packet. This makes it possible for a user to deregister an MR or unbind a MW on the fly and get correct behaviour.

---
 drivers/infiniband/sw/rxe/rxe_qp.c    |  10 +--
 drivers/infiniband/sw/rxe/rxe_resp.c  | 123 ++++++++++++++++++--------
 drivers/infiniband/sw/rxe/rxe_verbs.h |   1 -
 3 files changed, 87 insertions(+), 47 deletions(-)
drivers/infiniband/sw/rxe/rxe_qp.c +2 -8
index afe11f475b8c..0558253fe734 100644
@@ -135,12 +135,8 @@ static void free_rd_atomic_resources(struct rxe_qp *qp)
 
 void free_rd_atomic_resource(struct rxe_qp *qp, struct resp_res *res)
 {
-	if (res->type == RXE_ATOMIC_MASK) {
+	if (res->type == RXE_ATOMIC_MASK)
 		kfree_skb(res->atomic.skb);
-	} else if (res->type == RXE_READ_MASK) {
-		if (res->read.mr)
-			rxe_drop_ref(res->read.mr);
-	}
 	res->type = 0;
 }
 
@@ -817,10 +813,8 @@ static void rxe_qp_do_cleanup(struct work_struct *work)
 	if (qp->pd)
 		rxe_drop_ref(qp->pd);
 
-	if (qp->resp.mr) {
+	if (qp->resp.mr)
 		rxe_drop_ref(qp->resp.mr);
-		qp->resp.mr = NULL;
-	}
 
 	if (qp_type(qp) == IB_QPT_RC)
 		sk_dst_reset(qp->sk->sk);
drivers/infiniband/sw/rxe/rxe_resp.c +85 -38

154 lines not on this page · raw message

drivers/infiniband/sw/rxe/rxe_verbs.h +0 -1
index caf1ce118765..022abba4fb6b 100644
@@ -157,7 +157,6 @@ struct resp_res {
 			struct sk_buff	*skb;
 		} atomic;
 		struct {
-			struct rxe_mr	*mr;
 			u64		va_org;
 			u32		rkey;
 			u32		length;
Signature · 1 line
2.32.0
#8 lore raw

PATCH v8 6/8 RDMA/rxe: Minor cleanups in rxe_pool.c/rxe_pool.h

In reply to Bob Pearson

This patch includes a couple of minor cleanups in rxe_pool.c and rxe_pool.h

---
 drivers/infiniband/sw/rxe/rxe_pool.c | 10 +++-------
 drivers/infiniband/sw/rxe/rxe_pool.h |  1 -
 2 files changed, 3 insertions(+), 8 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +3 -7
index a5ff2fd692c9..dcfbc2b932af 100644
@@ -97,11 +97,8 @@ static const struct rxe_type_info {
 	},
 };
 
-void rxe_pool_init(
-	struct rxe_dev		*rxe,
-	struct rxe_pool		*pool,
-	enum rxe_elem_type	type,
-	unsigned int		max_elem)
+void rxe_pool_init(struct rxe_dev *rxe, struct rxe_pool *pool,
+		   enum rxe_elem_type type, unsigned int max_elem)
 {
 	const struct rxe_type_info *info = &rxe_type_info[type];
 
@@ -109,7 +106,6 @@ void rxe_pool_init(
 
 	pool->rxe		= rxe;
 	pool->name		= info->name;
-	pool->type		= type;
 	pool->max_elem		= max_elem;
 	pool->elem_size		= ALIGN(info->size, RXE_POOL_ALIGN);
 	pool->elem_offset	= info->elem_offset;
@@ -227,7 +223,7 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
 }
 
 /**
- * rxe_pool_get_index - lookup object from index
+ * rxe_pool_get_index() - lookup object from index
  * @pool: the object pool
  * @index: the index of the object
  *
drivers/infiniband/sw/rxe/rxe_pool.h +0 -1
index 894ffef4d6bd..be3b962d1c78 100644
@@ -46,7 +46,6 @@ struct rxe_pool {
 	int			(*init)(struct rxe_pool_elem *elem);
 	void			(*cleanup)(struct rxe_pool_elem *elem);
 	enum rxe_pool_flags	flags;
-	enum rxe_elem_type	type;
 	struct list_head	free_list;
 
 	unsigned int		max_elem;
Signature · 1 line
2.32.0
#9 lore raw

PATCH v8 7/8 RDMA/rxe: Replace rxe_alloc by kzalloc for rxe_mc_elem

In reply to Bob Pearson

Currently rxe_mc_elem structs are treated as rdma objects which is unneeded. This patch replaces rxe_alloc and rxe_drop_ref by kzalloc and kfree for these structs which hold associatons between multicast groups and QPs.

---
 drivers/infiniband/sw/rxe/rxe.c       |  3 ---
 drivers/infiniband/sw/rxe/rxe_mcast.c | 22 ++++++++++++++--------
 drivers/infiniband/sw/rxe/rxe_pool.c  |  6 ------
 drivers/infiniband/sw/rxe/rxe_pool.h  |  1 -
 drivers/infiniband/sw/rxe/rxe_verbs.h |  2 --
 5 files changed, 14 insertions(+), 20 deletions(-)
drivers/infiniband/sw/rxe/rxe.c +0 -3
index 09c73a0d8513..20a925aed29c 100644
@@ -31,7 +31,6 @@ void rxe_dealloc(struct ib_device *ib_dev)
 	rxe_pool_cleanup(&rxe->mr_pool);
 	rxe_pool_cleanup(&rxe->mw_pool);
 	rxe_pool_cleanup(&rxe->mc_grp_pool);
-	rxe_pool_cleanup(&rxe->mc_elem_pool);
 
 	if (rxe->tfm)
 		crypto_free_shash(rxe->tfm);
@@ -128,8 +127,6 @@ static void rxe_init_pools(struct rxe_dev *rxe)
 	rxe_pool_init(rxe, &rxe->mw_pool, RXE_TYPE_MW, rxe->attr.max_mw);
 	rxe_pool_init(rxe, &rxe->mc_grp_pool, RXE_TYPE_MC_GRP,
 			    rxe->attr.max_mcast_grp);
-	rxe_pool_init(rxe, &rxe->mc_elem_pool, RXE_TYPE_MC_ELEM,
-			    rxe->attr.max_total_mcast_qp_attach);
 }
 
 /* initialize rxe device state */
drivers/infiniband/sw/rxe/rxe_mcast.c +14 -8
index e110c4d3fbf4..b935634f86cd 100644
@@ -63,14 +63,15 @@ int rxe_mcast_add_grp_elem(struct rxe_dev *rxe, struct rxe_qp *qp,
 		goto out;
 	}
 
-	elem = rxe_alloc(&rxe->mc_elem_pool);
+	elem = kzalloc(sizeof(*elem), GFP_KERNEL);
 	if (!elem) {
 		err = -ENOMEM;
 		goto out;
 	}
 
-	/* each qp holds a ref on the grp */
+	/* each elem holds a ref on the grp and the qp */
 	rxe_add_ref(grp);
+	rxe_add_ref(qp);
 
 	grp->num_qp++;
 	elem->qp = qp;
@@ -91,6 +92,7 @@ int rxe_mcast_drop_grp_elem(struct rxe_dev *rxe, struct rxe_qp *qp,
 {
 	struct rxe_mc_grp *grp;
 	struct rxe_mc_elem *elem, *tmp;
+	int ret = -EINVAL;
 
 	grp = rxe_pool_get_key(&rxe->mc_grp_pool, mgid);
 	if (!grp)
@@ -107,18 +109,21 @@ int rxe_mcast_drop_grp_elem(struct rxe_dev *rxe, struct rxe_qp *qp,
 
 			spin_unlock_bh(&grp->mcg_lock);
 			spin_unlock_bh(&qp->grp_lock);
-			rxe_drop_ref(elem);
-			rxe_drop_ref(grp);	/* ref held by QP */
-			rxe_drop_ref(grp);	/* ref from get_key */
-			return 0;
+			kfree(elem);
+			rxe_drop_ref(qp);	/* ref held by elem */
+			rxe_drop_ref(grp);	/* ref held by elem */
+			ret = 0;
+			goto out_drop_ref;
 		}
 	}
 
 	spin_unlock_bh(&grp->mcg_lock);
 	spin_unlock_bh(&qp->grp_lock);
+
+out_drop_ref:
 	rxe_drop_ref(grp);			/* ref from get_key */
 err1:
-	return -EINVAL;
+	return ret;
 }
 
 void rxe_drop_all_mcast_groups(struct rxe_qp *qp)
@@ -142,8 +147,9 @@ void rxe_drop_all_mcast_groups(struct rxe_qp *qp)
 		list_del(&elem->qp_list);
 		grp->num_qp--;
 		spin_unlock_bh(&grp->mcg_lock);
+		rxe_drop_ref(qp);
 		rxe_drop_ref(grp);
-		rxe_drop_ref(elem);
+		kfree(elem);
 	}
 }
 
drivers/infiniband/sw/rxe/rxe_pool.c +0 -6
index dcfbc2b932af..0f785d14f646 100644
@@ -89,12 +89,6 @@ static const struct rxe_type_info {
 		.key_offset	= offsetof(struct rxe_mc_grp, mgid),
 		.key_size	= sizeof(union ib_gid),
 	},
-	[RXE_TYPE_MC_ELEM] = {
-		.name		= "rxe-mc_elem",
-		.size		= sizeof(struct rxe_mc_elem),
-		.elem_offset	= offsetof(struct rxe_mc_elem, elem),
-		.flags		= RXE_POOL_ALLOC,
-	},
 };
 
 void rxe_pool_init(struct rxe_dev *rxe, struct rxe_pool *pool,
drivers/infiniband/sw/rxe/rxe_pool.h +0 -1
index be3b962d1c78..ccb923b10276 100644
@@ -23,7 +23,6 @@ enum rxe_elem_type {
 	RXE_TYPE_MR,
 	RXE_TYPE_MW,
 	RXE_TYPE_MC_GRP,
-	RXE_TYPE_MC_ELEM,
 	RXE_NUM_TYPES,		/* keep me last */
 };
 
drivers/infiniband/sw/rxe/rxe_verbs.h +0 -2
index 022abba4fb6b..9f39b097a976 100644
@@ -364,7 +364,6 @@ struct rxe_mc_grp {
 };
 
 struct rxe_mc_elem {
-	struct rxe_pool_elem	elem;
 	struct list_head	qp_list;
 	struct list_head	grp_list;
 	struct rxe_qp		*qp;
@@ -402,7 +401,6 @@ struct rxe_dev {
 	struct rxe_pool		mr_pool;
 	struct rxe_pool		mw_pool;
 	struct rxe_pool		mc_grp_pool;
-	struct rxe_pool		mc_elem_pool;
 
 	spinlock_t		pending_lock; /* guard pending_mmaps */
 	struct list_head	pending_mmaps;
Signature · 1 line
2.32.0
#10 lore raw

PATCH v8 8/8 RDMA/rxe: Add wait for completion to obj destruct

In reply to Bob Pearson

This patch adds code to wait until pending activity on RDMA objects has completed before freeing or returning to rdma-core where the object may be freed.

---
 drivers/infiniband/sw/rxe/rxe_mcast.c | 10 +++++
 drivers/infiniband/sw/rxe/rxe_mr.c    |  2 +
 drivers/infiniband/sw/rxe/rxe_mw.c    |  3 +-
 drivers/infiniband/sw/rxe/rxe_pool.c  | 31 +++++++++++++-
 drivers/infiniband/sw/rxe/rxe_pool.h  |  4 ++
 drivers/infiniband/sw/rxe/rxe_verbs.c | 60 ++++++++++++++++++---------
 6 files changed, 89 insertions(+), 21 deletions(-)
drivers/infiniband/sw/rxe/rxe_mcast.c +10 -0
index b935634f86cd..d91c2e30665a 100644
@@ -122,6 +122,11 @@ int rxe_mcast_drop_grp_elem(struct rxe_dev *rxe, struct rxe_qp *qp,
 
 out_drop_ref:
 	rxe_drop_ref(grp);			/* ref from get_key */
+	/* when grp ref count drops to zero
+	 * go ahead and free it
+	 */
+	if (grp->elem.complete.done)
+		rxe_fini(grp);
 err1:
 	return ret;
 }
@@ -149,6 +154,11 @@ void rxe_drop_all_mcast_groups(struct rxe_qp *qp)
 		spin_unlock_bh(&grp->mcg_lock);
 		rxe_drop_ref(qp);
 		rxe_drop_ref(grp);
+		/* when grp ref count drops to zero
+		 * go ahead and free it
+		 */
+		if (grp->elem.complete.done)
+			rxe_fini(grp);
 		kfree(elem);
 	}
 }
drivers/infiniband/sw/rxe/rxe_mr.c +2 -0
index 3c4390adfb80..5f8c08da352d 100644
@@ -695,6 +695,8 @@ int rxe_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
 	rxe_drop_ref(mr_pd(mr));
 	rxe_drop_ref(mr);
 
+	rxe_fini(mr);
+
 	return 0;
 }
 
drivers/infiniband/sw/rxe/rxe_mw.c +2 -1
index 3ae981d77c25..666a641f9934 100644
@@ -60,8 +60,9 @@ int rxe_dealloc_mw(struct ib_mw *ibmw)
 	rxe_do_dealloc_mw(mw);
 	spin_unlock_bh(&mw->lock);
 
-	rxe_drop_ref(mw);
 	rxe_drop_ref(pd);
+	rxe_drop_ref(mw);
+	rxe_fini(mw);
 
 	return 0;
 }
drivers/infiniband/sw/rxe/rxe_pool.c +30 -1
index 0f785d14f646..b23c58281eb6 100644
@@ -6,6 +6,8 @@
 
 #include "rxe.h"
 
+/* timeout in jiffies for pool element to complete */
+#define RXE_POOL_TIMEOUT	(100)
 #define RXE_POOL_ALIGN		(16)
 
 static const struct rxe_type_info {
@@ -160,6 +162,7 @@ void *rxe_alloc(struct rxe_pool *pool)
 	elem->pool = pool;
 	elem->obj = obj;
 	kref_init(&elem->ref_cnt);
+	init_completion(&elem->complete);
 
 	if (pool->init) {
 		err = pool->init(elem);
@@ -194,6 +197,7 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
 	elem->pool = pool;
 	elem->obj = (u8 *)elem - pool->elem_offset;
 	kref_init(&elem->ref_cnt);
+	init_completion(&elem->complete);
 
 	if (pool->init) {
 		err = pool->init(elem);
@@ -413,8 +417,33 @@ void rxe_elem_release(struct kref *kref)
 	if (pool->cleanup)
 		pool->cleanup(elem);
 
+	atomic_dec(&pool->num_elem);
+
+	complete(&elem->complete);
+}
+
+/**
+ * __rxe_fini() - wait for completion of pool element
+ * @elem: the pool elem
+ *
+ * Wait until the reference count of an object drops to zero when
+ * rxe_elem_release() will complete the object and then, if locally
+ * allocated, free the memory containing the object and return
+ *
+ * Returns: non-zero if the object completed successfully else zero
+ */
+int __rxe_fini(struct rxe_pool_elem *elem)
+{
+	struct rxe_pool *pool = elem->pool;
+	int ret;
+
+	ret = wait_for_completion_timeout(&elem->complete, RXE_POOL_TIMEOUT);
+	if (!ret)
+		pr_warn("Timed out waiting for %s#%d to complete\n",
+				pool->name, elem->index);
+
 	if (pool->flags & RXE_POOL_ALLOC)
 		kfree(elem->obj);
 
-	atomic_dec(&pool->num_elem);
+	return ret;
 }
drivers/infiniband/sw/rxe/rxe_pool.h +4 -0
index ccb923b10276..252f25bf5a1a 100644
@@ -30,6 +30,7 @@ struct rxe_pool_elem {
 	struct rxe_pool		*pool;
 	void			*obj;
 	struct kref		ref_cnt;
+	struct completion	complete;
 	struct list_head	list;
 
 	/* only used if keyed */
@@ -100,4 +101,7 @@ static inline bool __rxe_drop_ref(struct rxe_pool_elem *elem)
 }
 #define rxe_drop_ref(obj) __rxe_drop_ref(&(obj)->elem)
 
+int __rxe_fini(struct rxe_pool_elem *elem);
+#define rxe_fini(obj) __rxe_fini(&(obj)->elem)
+
 #endif /* RXE_POOL_H */
drivers/infiniband/sw/rxe/rxe_verbs.c +41 -19

200 lines not on this page · raw message

Signature · 1 line
2.32.0
maintainer
#11 lore raw

Re: [PATCH for-next v8 8/8] RDMA/rxe: Add wait for completion to obj destruct

Quoted 18 lines · Bob Pearson

This patch adds code to wait until pending activity on RDMA objects has completed before freeing or returning to rdma-core where the object may be freed.

 drivers/infiniband/sw/rxe/rxe_mcast.c | 10 +++++
 drivers/infiniband/sw/rxe/rxe_mr.c    |  2 +
 drivers/infiniband/sw/rxe/rxe_mw.c    |  3 +-
 drivers/infiniband/sw/rxe/rxe_pool.c  | 31 +++++++++++++-
 drivers/infiniband/sw/rxe/rxe_pool.h  |  4 ++
 drivers/infiniband/sw/rxe/rxe_verbs.c | 60 ++++++++++++++++++---------
 6 files changed, 89 insertions(+), 21 deletions(-)
drivers/infiniband/sw/rxe/rxe_mcast.c +1 -0
index b935634f86cd..d91c2e30665a 100644
+++ b/drivers/infiniband/sw/rxe/rxe_mcast.c

drivers/infiniband/sw/rxe/rxe_mcast.c int rxe_mcast_drop_grp_elem(struct rxe_dev *rxe, struct rxe_qp *qp, · line 129

Quoted 8 lines
drivers/infiniband/sw/rxe/rxe_mcast.c +5 -0
@@ -122,6 +122,11 @@ int rxe_mcast_drop_grp_elem(struct rxe_dev *rxe, struct rxe_qp *qp,
 
 out_drop_ref:
 	rxe_drop_ref(grp);			/* ref from get_key */
+	/* when grp ref count drops to zero
+	 * go ahead and free it
+	 */
+	if (grp->elem.complete.done)
+		rxe_fini(grp);

When using the completion pattern with refcounts there has to be a designated freer that *always* frees the memory. This is how all the other users of rxe_fini() are working.

You can't mix the completion pattern with multiple users that can free, it just becomes racy, eg two threads run the above concurrently, one will free the grp and the other will use-after-free checking done.

In this case I think you want to just run the kfree from the kref release.

Jason

maintainer
#12 lore raw
In reply to Bob Pearson
Quoted 9 lines · Bob Pearson

There are several race conditions discovered in the current rdma_rxe driver. They mostly relate to races between normal operations and destroying objects. This patch series
- Makes several minor cleanups in rxe_pool.[ch]
- Replaces the red-black trees currently used by xarrays for indices
- Simplifies the API for keyed objects
- Corrects several reference counting errors
- Adds wait for completions to the paths in verbs APIs which destroy objects.

I think this will work better for you you strip out the RXE_POOL_KEY stuff that is only use for multicast and move it to the mcast file. They are so different now they shouldn't be sharing much of anything.

Jason