Skip to contentSkip to conversation

RDMA/rxe: Fix dead lock caused by __rxe_add_to_pool interrupted by rxe_pool_get_index

PATCH In review [email protected] 5 participants 18 messages Apr 22, 2022 – Jul 23, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v6 1/4 RDMA/rxe: Fix dead lock caused by __rxe_add_to_pool interrupted by rxe_pool_get_index

From: Zhu Yanjun <[email protected]>

This is a dead lock problem.
The ah_pool xa_lock first is acquired in this:

{SOFTIRQ-ON-W} state was registered at:

  lock_acquire+0x1d2/0x5a0
  _raw_spin_lock+0x33/0x80
  __rxe_add_to_pool+0x183/0x230 [rdma_rxe]

Then ah_pool xa_lock is acquired in this:

{IN-SOFTIRQ-W}:

Call Trace:
 <TASK>
  dump_stack_lvl+0x44/0x57
  mark_lock.part.52.cold.79+0x3c/0x46
  __lock_acquire+0x1565/0x34a0
  lock_acquire+0x1d2/0x5a0
  _raw_spin_lock_irqsave+0x42/0x90
  rxe_pool_get_index+0x72/0x1d0 [rdma_rxe]
  rxe_get_av+0x168/0x2a0 [rdma_rxe]
</TASK>

From the above, in the function __rxe_add_to_pool, xa_lock is acquired. Then the function __rxe_add_to_pool is interrupted by softirq. The function rxe_pool_get_index will also acquire xa_lock.

Finally, the dead lock appears.

        CPU0
        ----
   lock(&xa->xa_lock#15);  <----- __rxe_add_to_pool
   <Interrupt>
     lock(&xa->xa_lock#15); <---- rxe_pool_get_index

                 *** DEADLOCK ***
Fixes 3225717f6dfa ("RDMA/rxe: Replace red-black trees by carrays")
Reported-and-tested-by Yi Zhang <[email protected]>

---
V5->V6: One dead lock fix in one commit
V4->V5: Commit logs are changed.
V3->V4: xa_lock_irq locks are used.
V2->V3: __rxe_add_to_pool is between spin_lock and spin_unlock, so
GFP_ATOMIC is used in __rxe_add_to_pool.
V1->V2: Replace GFP_KERNEL with GFP_ATOMIC

---
 drivers/infiniband/sw/rxe/rxe_pool.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +7 -4
index 87066d04ed18..67f1d4733682 100644
@@ -106,7 +106,7 @@ void rxe_pool_init(struct rxe_dev *rxe, struct rxe_pool *pool,
 
 	atomic_set(&pool->num_elem, 0);
 
-	xa_init_flags(&pool->xa, XA_FLAGS_ALLOC);
+	xa_init_flags(&pool->xa, XA_FLAGS_ALLOC | XA_FLAGS_LOCK_IRQ);
 	pool->limit.min = info->min_index;
 	pool->limit.max = info->max_index;
 }
@@ -155,6 +155,7 @@ void *rxe_alloc(struct rxe_pool *pool)
 int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
 {
 	int err;
+	unsigned long flags;
 
 	if (WARN_ON(pool->flags & RXE_POOL_ALLOC))
 		return -EINVAL;
@@ -166,8 +167,10 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
 	elem->obj = (u8 *)elem - pool->elem_offset;
 	kref_init(&elem->ref_cnt);
 
-	err = xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
-			      &pool->next, GFP_KERNEL);
+	xa_lock_irqsave(&pool->xa, flags);
+	err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
+				&pool->next, GFP_ATOMIC);
+	xa_unlock_irqrestore(&pool->xa, flags);
 	if (err)
 		goto err_cnt;
 
@@ -201,7 +204,7 @@ static void rxe_elem_release(struct kref *kref)
 	struct rxe_pool_elem *elem = container_of(kref, typeof(*elem), ref_cnt);
 	struct rxe_pool *pool = elem->pool;
 
-	xa_erase(&pool->xa, elem->index);
+	xa_erase_irq(&pool->xa, elem->index);
 
 	if (pool->cleanup)
 		pool->cleanup(elem);
Signature · 1 line
2.27.0
#1 lore raw

PATCH 2/4 RDMA/rxe: Fix dead lock caused by rxe_alloc interrupted by rxe_pool_get_index

From: Zhu Yanjun <[email protected]>

The ah_pool xa_lock first is acquired in this:

{SOFTIRQ-ON-W} state was registered at:

  lock_acquire+0x1d2/0x5a0
  _raw_spin_lock+0x33/0x80
  rxe_alloc+0x1be/0x290 [rdma_rxe]

Then ah_pool xa_lock is acquired in this:

{IN-SOFTIRQ-W}:

  <TASK>
  __lock_acquire+0x1565/0x34a0
  lock_acquire+0x1d2/0x5a0
  _raw_spin_lock_irqsave+0x42/0x90
  rxe_pool_get_index+0x72/0x1d0 [rdma_rxe]
  </TASK>

From the above, in the function rxe_alloc,
xa_lock is acquired. Then the function rxe_alloc
is interrupted by softirq. The function
rxe_pool_get_index will also acquire xa_lock.

Finally, the dead lock appears.

        CPU0
        ----
   lock(&xa->xa_lock#15);  <----- rxe_alloc
   <Interrupt>
     lock(&xa->xa_lock#15); <---- rxe_pool_get_index

    *** DEADLOCK ***
Fixes 3225717f6dfa ("RDMA/rxe: Replace red-black trees by carrays")
Reported-and-tested-by Yi Zhang <[email protected]>
---
 drivers/infiniband/sw/rxe/rxe_pool.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +2 -2
index 67f1d4733682..7b12a52fed35 100644
@@ -138,8 +138,8 @@ void *rxe_alloc(struct rxe_pool *pool)
 	elem->obj = obj;
 	kref_init(&elem->ref_cnt);
 
-	err = xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
-			      &pool->next, GFP_KERNEL);
+	err = xa_alloc_cyclic_irq(&pool->xa, &elem->index, elem, pool->limit,
+				  &pool->next, GFP_KERNEL);
 	if (err)
 		goto err_free;
 
Signature · 1 line
2.27.0
#2 lore raw

PATCH 3/4 RDMA/rxe: Use different xa locks on different path

In reply to [email protected]

From: Zhu Yanjun <[email protected]>

The function __rxe_add_to_pool is called on different paths, and the requirement of the locks is different. The function rxe_create_ah requires xa_lock_irqsave/irqrestore while others only require xa_lock_irq.

---
 drivers/infiniband/sw/rxe/rxe_pool.c | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +11 -5
index 7b12a52fed35..3f3fa2123f30 100644
@@ -155,7 +155,6 @@ void *rxe_alloc(struct rxe_pool *pool)
 int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
 {
 	int err;
-	unsigned long flags;
 
 	if (WARN_ON(pool->flags & RXE_POOL_ALLOC))
 		return -EINVAL;
@@ -167,10 +166,17 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
 	elem->obj = (u8 *)elem - pool->elem_offset;
 	kref_init(&elem->ref_cnt);
 
-	xa_lock_irqsave(&pool->xa, flags);
-	err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
-				&pool->next, GFP_ATOMIC);
-	xa_unlock_irqrestore(&pool->xa, flags);
+	if (pool->type == RXE_TYPE_AH) {
+		unsigned long flags;
+
+		xa_lock_irqsave(&pool->xa, flags);
+		err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
+					&pool->next, GFP_ATOMIC);
+		xa_unlock_irqrestore(&pool->xa, flags);
+	} else {
+		err = xa_alloc_cyclic_irq(&pool->xa, &elem->index, elem, pool->limit,
+					  &pool->next, GFP_KERNEL);
+	}
 	if (err)
 		goto err_cnt;
 
Signature · 1 line
2.27.0
#3 lore raw

PATCH 4/4 RDMA/rxe: Check RDMA_CREATE_AH_SLEEPABLE in creating AH

In reply to [email protected]

From: Zhu Yanjun <[email protected]>

During creating AH, the flag RDMA_CREATE_AH_SLEEPABLE should be tested.

---
 drivers/infiniband/sw/rxe/rxe_mw.c    |  2 +-
 drivers/infiniband/sw/rxe/rxe_pool.c  | 14 ++++++++------
 drivers/infiniband/sw/rxe/rxe_pool.h  |  4 ++--
 drivers/infiniband/sw/rxe/rxe_verbs.c | 18 ++++++++++++------
 4 files changed, 23 insertions(+), 15 deletions(-)
drivers/infiniband/sw/rxe/rxe_mw.c +1 -1
index c86b2efd58f2..9d72dcc9060d 100644
@@ -14,7 +14,7 @@ int rxe_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
 
 	rxe_get(pd);
 
-	ret = rxe_add_to_pool(&rxe->mw_pool, mw);
+	ret = rxe_add_to_pool(&rxe->mw_pool, mw, GFP_KERNEL);
 	if (ret) {
 		rxe_put(pd);
 		return ret;
drivers/infiniband/sw/rxe/rxe_pool.c +8 -6
index 3f3fa2123f30..5555060702fd 100644
@@ -152,7 +152,7 @@ void *rxe_alloc(struct rxe_pool *pool)
 	return NULL;
 }
 
-int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
+int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem, gfp_t gfp)
 {
 	int err;
 
@@ -166,16 +166,18 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
 	elem->obj = (u8 *)elem - pool->elem_offset;
 	kref_init(&elem->ref_cnt);
 
-	if (pool->type == RXE_TYPE_AH) {
+	if ((pool->type == RXE_TYPE_AH) && (gfp & GFP_ATOMIC)) {
 		unsigned long flags;
 
 		xa_lock_irqsave(&pool->xa, flags);
-		err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
-					&pool->next, GFP_ATOMIC);
+		err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem,
+					pool->limit, &pool->next,
+					GFP_ATOMIC);
 		xa_unlock_irqrestore(&pool->xa, flags);
 	} else {
-		err = xa_alloc_cyclic_irq(&pool->xa, &elem->index, elem, pool->limit,
-					  &pool->next, GFP_KERNEL);
+		err = xa_alloc_cyclic_irq(&pool->xa, &elem->index, elem,
+					  pool->limit, &pool->next,
+					  GFP_KERNEL);
 	}
 	if (err)
 		goto err_cnt;
drivers/infiniband/sw/rxe/rxe_pool.h +2 -2
index 24bcc786c1b3..12986622088b 100644
@@ -62,9 +62,9 @@ void rxe_pool_cleanup(struct rxe_pool *pool);
 void *rxe_alloc(struct rxe_pool *pool);
 
 /* connect already allocated object to pool */
-int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem);
+int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem, gfp_t gfp);
 
-#define rxe_add_to_pool(pool, obj) __rxe_add_to_pool(pool, &(obj)->elem)
+#define rxe_add_to_pool(pool, obj, gfp) __rxe_add_to_pool(pool, &(obj)->elem, gfp)
 
 /* lookup an indexed object from index. takes a reference on object */
 void *rxe_pool_get_index(struct rxe_pool *pool, u32 index);
drivers/infiniband/sw/rxe/rxe_verbs.c +12 -6
index 67184b0281a0..dce665e74fa7 100644
@@ -108,7 +108,7 @@ static int rxe_alloc_ucontext(struct ib_ucontext *ibuc, struct ib_udata *udata)
 	struct rxe_dev *rxe = to_rdev(ibuc->device);
 	struct rxe_ucontext *uc = to_ruc(ibuc);
 
-	return rxe_add_to_pool(&rxe->uc_pool, uc);
+	return rxe_add_to_pool(&rxe->uc_pool, uc, GFP_KERNEL);
 }
 
 static void rxe_dealloc_ucontext(struct ib_ucontext *ibuc)
@@ -142,7 +142,7 @@ static int rxe_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 	struct rxe_dev *rxe = to_rdev(ibpd->device);
 	struct rxe_pd *pd = to_rpd(ibpd);
 
-	return rxe_add_to_pool(&rxe->pd_pool, pd);
+	return rxe_add_to_pool(&rxe->pd_pool, pd, GFP_KERNEL);
 }
 
 static int rxe_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
@@ -162,6 +162,7 @@ static int rxe_create_ah(struct ib_ah *ibah,
 	struct rxe_ah *ah = to_rah(ibah);
 	struct rxe_create_ah_resp __user *uresp = NULL;
 	int err;
+	gfp_t gfp;
 
 	if (udata) {
 		/* test if new user provider */
@@ -176,7 +177,12 @@ static int rxe_create_ah(struct ib_ah *ibah,
 	if (err)
 		return err;
 
-	err = rxe_add_to_pool(&rxe->ah_pool, ah);
+	if (init_attr->flags & RDMA_CREATE_AH_SLEEPABLE)
+		gfp = GFP_KERNEL;
+	else
+		gfp = GFP_ATOMIC;
+
+	err = rxe_add_to_pool(&rxe->ah_pool, ah, gfp);
 	if (err)
 		return err;
 
@@ -299,7 +305,7 @@ static int rxe_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init,
 	if (err)
 		goto err1;
 
-	err = rxe_add_to_pool(&rxe->srq_pool, srq);
+	err = rxe_add_to_pool(&rxe->srq_pool, srq, GFP_KERNEL);
 	if (err)
 		goto err1;
 
@@ -431,7 +437,7 @@ static int rxe_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init,
 		qp->is_user = false;
 	}
 
-	err = rxe_add_to_pool(&rxe->qp_pool, qp);
+	err = rxe_add_to_pool(&rxe->qp_pool, qp, GFP_KERNEL);
 	if (err)
 		return err;
 
@@ -800,7 +806,7 @@ static int rxe_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 	if (err)
 		return err;
 
-	return rxe_add_to_pool(&rxe->cq_pool, cq);
+	return rxe_add_to_pool(&rxe->cq_pool, cq, GFP_KERNEL);
 }
 
 static int rxe_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
Signature · 1 line
2.27.0
#4 lore raw

Re: [PATCH 4/4] RDMA/rxe: Check RDMA_CREATE_AH_SLEEPABLE in creating AH

drivers/infiniband/sw/rxe/rxe_pool.c int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem) · line 170

Quoted 7 lines · On Fri, Apr 22, 2022 at 03:44:16PM -0400, [email protected] wrote:
drivers/infiniband/sw/rxe/rxe_pool.c +1 -1
@@ -166,16 +166,18 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
 	elem->obj = (u8 *)elem - pool->elem_offset;
 	kref_init(&elem->ref_cnt);
 
-	if (pool->type == RXE_TYPE_AH) {
+	if ((pool->type == RXE_TYPE_AH) && (gfp & GFP_ATOMIC)) {
 		unsigned long flags;

No test for AH should be here, just gfp.

Jason

author
#5 lore raw

Re: [PATCH 4/4] RDMA/rxe: Check RDMA_CREATE_AH_SLEEPABLE in creating AH

Quoted 9 lines · Jason Gunthorpe

drivers/infiniband/sw/rxe/rxe_pool.c int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem) · line 170

Quoted 7 lines · On Fri, Apr 22, 2022 at 03:44:16PM -0400, [email protected] wrote:
drivers/infiniband/sw/rxe/rxe_pool.c +1 -1
@@ -166,16 +166,18 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
  	elem->obj = (u8 *)elem - pool->elem_offset;
  	kref_init(&elem->ref_cnt);
 
-	if (pool->type == RXE_TYPE_AH) {
+	if ((pool->type == RXE_TYPE_AH) && (gfp & GFP_ATOMIC)) {
  		unsigned long flags;

No test for AH should be here, just gfp.

Agree. Got it.

Zhu Yanjun

Quoted 2 lines

Jason

#6 lore raw

PATCH v2 4/4 RDMA/rxe: Check RDMA_CREATE_AH_SLEEPABLE in creating AH

In reply to Jason Gunthorpe

From: Zhu Yanjun <[email protected]>

During creating AH, the flag RDMA_CREATE_AH_SLEEPABLE should be tested.

---
V1->V2: Remove the RXE_TYPE_AH test

---
 drivers/infiniband/sw/rxe/rxe_mw.c    |  2 +-
 drivers/infiniband/sw/rxe/rxe_pool.c  | 14 ++++++++------
 drivers/infiniband/sw/rxe/rxe_pool.h  |  4 ++--
 drivers/infiniband/sw/rxe/rxe_verbs.c | 18 ++++++++++++------
 4 files changed, 23 insertions(+), 15 deletions(-)
drivers/infiniband/sw/rxe/rxe_mw.c +1 -1
index c86b2efd58f2..9d72dcc9060d 100644
@@ -14,7 +14,7 @@ int rxe_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
 
 	rxe_get(pd);
 
-	ret = rxe_add_to_pool(&rxe->mw_pool, mw);
+	ret = rxe_add_to_pool(&rxe->mw_pool, mw, GFP_KERNEL);
 	if (ret) {
 		rxe_put(pd);
 		return ret;
drivers/infiniband/sw/rxe/rxe_pool.c +8 -6
index 3f3fa2123f30..793df1569ff1 100644
@@ -152,7 +152,7 @@ void *rxe_alloc(struct rxe_pool *pool)
 	return NULL;
 }
 
-int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
+int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem, gfp_t gfp)
 {
 	int err;
 
@@ -166,16 +166,18 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
 	elem->obj = (u8 *)elem - pool->elem_offset;
 	kref_init(&elem->ref_cnt);
 
-	if (pool->type == RXE_TYPE_AH) {
+	if (gfp & GFP_ATOMIC) {
 		unsigned long flags;
 
 		xa_lock_irqsave(&pool->xa, flags);
-		err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
-					&pool->next, GFP_ATOMIC);
+		err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem,
+					pool->limit, &pool->next,
+					GFP_ATOMIC);
 		xa_unlock_irqrestore(&pool->xa, flags);
 	} else {
-		err = xa_alloc_cyclic_irq(&pool->xa, &elem->index, elem, pool->limit,
-					  &pool->next, GFP_KERNEL);
+		err = xa_alloc_cyclic_irq(&pool->xa, &elem->index, elem,
+					  pool->limit, &pool->next,
+					  GFP_KERNEL);
 	}
 	if (err)
 		goto err_cnt;
drivers/infiniband/sw/rxe/rxe_pool.h +2 -2
index 24bcc786c1b3..12986622088b 100644
@@ -62,9 +62,9 @@ void rxe_pool_cleanup(struct rxe_pool *pool);
 void *rxe_alloc(struct rxe_pool *pool);
 
 /* connect already allocated object to pool */
-int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem);
+int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem, gfp_t gfp);
 
-#define rxe_add_to_pool(pool, obj) __rxe_add_to_pool(pool, &(obj)->elem)
+#define rxe_add_to_pool(pool, obj, gfp) __rxe_add_to_pool(pool, &(obj)->elem, gfp)
 
 /* lookup an indexed object from index. takes a reference on object */
 void *rxe_pool_get_index(struct rxe_pool *pool, u32 index);
drivers/infiniband/sw/rxe/rxe_verbs.c +12 -6
index 67184b0281a0..dce665e74fa7 100644
@@ -108,7 +108,7 @@ static int rxe_alloc_ucontext(struct ib_ucontext *ibuc, struct ib_udata *udata)
 	struct rxe_dev *rxe = to_rdev(ibuc->device);
 	struct rxe_ucontext *uc = to_ruc(ibuc);
 
-	return rxe_add_to_pool(&rxe->uc_pool, uc);
+	return rxe_add_to_pool(&rxe->uc_pool, uc, GFP_KERNEL);
 }
 
 static void rxe_dealloc_ucontext(struct ib_ucontext *ibuc)
@@ -142,7 +142,7 @@ static int rxe_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 	struct rxe_dev *rxe = to_rdev(ibpd->device);
 	struct rxe_pd *pd = to_rpd(ibpd);
 
-	return rxe_add_to_pool(&rxe->pd_pool, pd);
+	return rxe_add_to_pool(&rxe->pd_pool, pd, GFP_KERNEL);
 }
 
 static int rxe_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
@@ -162,6 +162,7 @@ static int rxe_create_ah(struct ib_ah *ibah,
 	struct rxe_ah *ah = to_rah(ibah);
 	struct rxe_create_ah_resp __user *uresp = NULL;
 	int err;
+	gfp_t gfp;
 
 	if (udata) {
 		/* test if new user provider */
@@ -176,7 +177,12 @@ static int rxe_create_ah(struct ib_ah *ibah,
 	if (err)
 		return err;
 
-	err = rxe_add_to_pool(&rxe->ah_pool, ah);
+	if (init_attr->flags & RDMA_CREATE_AH_SLEEPABLE)
+		gfp = GFP_KERNEL;
+	else
+		gfp = GFP_ATOMIC;
+
+	err = rxe_add_to_pool(&rxe->ah_pool, ah, gfp);
 	if (err)
 		return err;
 
@@ -299,7 +305,7 @@ static int rxe_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init,
 	if (err)
 		goto err1;
 
-	err = rxe_add_to_pool(&rxe->srq_pool, srq);
+	err = rxe_add_to_pool(&rxe->srq_pool, srq, GFP_KERNEL);
 	if (err)
 		goto err1;
 
@@ -431,7 +437,7 @@ static int rxe_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init,
 		qp->is_user = false;
 	}
 
-	err = rxe_add_to_pool(&rxe->qp_pool, qp);
+	err = rxe_add_to_pool(&rxe->qp_pool, qp, GFP_KERNEL);
 	if (err)
 		return err;
 
@@ -800,7 +806,7 @@ static int rxe_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 	if (err)
 		return err;
 
-	return rxe_add_to_pool(&rxe->cq_pool, cq);
+	return rxe_add_to_pool(&rxe->cq_pool, cq, GFP_KERNEL);
 }
 
 static int rxe_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
Signature · 1 line
2.27.0
#7 lore raw
In reply to [email protected]

Use of rcu_read_lock solves this problem. Rcu_read_lock and spinlock on same data can Co-exist at the same time. That is the whole point. All this is going away soon.

Bob

-----Original Message-----
From: [email protected] <[email protected]>
Sent: Friday, April 22, 2022 2:44 PM
To: [email protected]; [email protected]; [email protected]; [email protected]
Cc: Yi Zhang <[email protected]>
Subject: [PATCHv6 1/4] RDMA/rxe: Fix dead lock caused by __rxe_add_to_pool interrupted by rxe_pool_get_index

From: Zhu Yanjun <[email protected]>

This is a dead lock problem.
The ah_pool xa_lock first is acquired in this:

{SOFTIRQ-ON-W} state was registered at:

  lock_acquire+0x1d2/0x5a0
  _raw_spin_lock+0x33/0x80
  __rxe_add_to_pool+0x183/0x230 [rdma_rxe]

Then ah_pool xa_lock is acquired in this:

{IN-SOFTIRQ-W}:

Call Trace:
 <TASK>
  dump_stack_lvl+0x44/0x57
  mark_lock.part.52.cold.79+0x3c/0x46
  __lock_acquire+0x1565/0x34a0
  lock_acquire+0x1d2/0x5a0
  _raw_spin_lock_irqsave+0x42/0x90
  rxe_pool_get_index+0x72/0x1d0 [rdma_rxe]
  rxe_get_av+0x168/0x2a0 [rdma_rxe]
</TASK>

From the above, in the function __rxe_add_to_pool, xa_lock is acquired. Then the function __rxe_add_to_pool is interrupted by softirq. The function rxe_pool_get_index will also acquire xa_lock.

Finally, the dead lock appears.

        CPU0
        ----
   lock(&xa->xa_lock#15);  <----- __rxe_add_to_pool
   <Interrupt>
     lock(&xa->xa_lock#15); <---- rxe_pool_get_index

                 *** DEADLOCK ***
Fixes 3225717f6dfa ("RDMA/rxe: Replace red-black trees by carrays")
Reported-and-tested-by Yi Zhang <[email protected]>

---
V5->V6: One dead lock fix in one commit
V4->V5: Commit logs are changed.
V3->V4: xa_lock_irq locks are used.
V2->V3: __rxe_add_to_pool is between spin_lock and spin_unlock, so
GFP_ATOMIC is used in __rxe_add_to_pool.
V1->V2: Replace GFP_KERNEL with GFP_ATOMIC

---
 drivers/infiniband/sw/rxe/rxe_pool.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +7 -4
index 87066d04ed18..67f1d4733682 100644
@@ -106,7 +106,7 @@ void rxe_pool_init(struct rxe_dev *rxe, struct rxe_pool *pool,
 
 	atomic_set(&pool->num_elem, 0);
 
-	xa_init_flags(&pool->xa, XA_FLAGS_ALLOC);
+	xa_init_flags(&pool->xa, XA_FLAGS_ALLOC | XA_FLAGS_LOCK_IRQ);
 	pool->limit.min = info->min_index;
 	pool->limit.max = info->max_index;
 }
@@ -155,6 +155,7 @@ void *rxe_alloc(struct rxe_pool *pool)  int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)  {
 	int err;
+	unsigned long flags;
 
 	if (WARN_ON(pool->flags & RXE_POOL_ALLOC))
 		return -EINVAL;
@@ -166,8 +167,10 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
 	elem->obj = (u8 *)elem - pool->elem_offset;
 	kref_init(&elem->ref_cnt);
 
-	err = xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
-			      &pool->next, GFP_KERNEL);
+	xa_lock_irqsave(&pool->xa, flags);
+	err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
+				&pool->next, GFP_ATOMIC);
+	xa_unlock_irqrestore(&pool->xa, flags);
 	if (err)
 		goto err_cnt;
 
@@ -201,7 +204,7 @@ static void rxe_elem_release(struct kref *kref)
 	struct rxe_pool_elem *elem = container_of(kref, typeof(*elem), ref_cnt);
 	struct rxe_pool *pool = elem->pool;
 
-	xa_erase(&pool->xa, elem->index);
+	xa_erase_irq(&pool->xa, elem->index);
 
 	if (pool->cleanup)
 		pool->cleanup(elem);

--
2.27.0

author
#8 lore raw
Quoted 2 lines · Robert B

Use of rcu_read_lock solves this problem. Rcu_read_lock and spinlock on same data can Co-exist at the same time. That is the whole point. All this is going away soon.

This is based on your unproved assumption.

Zhu Yanjun

Quoted 106 lines

Bob

-----Original Message-----
From: [email protected] <[email protected]>
Sent: Friday, April 22, 2022 2:44 PM
To: [email protected]; [email protected]; [email protected]; [email protected]
Cc: Yi Zhang <[email protected]>
Subject: [PATCHv6 1/4] RDMA/rxe: Fix dead lock caused by __rxe_add_to_pool interrupted by rxe_pool_get_index

From: Zhu Yanjun <[email protected]>

This is a dead lock problem.
The ah_pool xa_lock first is acquired in this:

{SOFTIRQ-ON-W} state was registered at:

   lock_acquire+0x1d2/0x5a0
   _raw_spin_lock+0x33/0x80
   __rxe_add_to_pool+0x183/0x230 [rdma_rxe]

Then ah_pool xa_lock is acquired in this:

{IN-SOFTIRQ-W}:

Call Trace:
  <TASK>
   dump_stack_lvl+0x44/0x57
   mark_lock.part.52.cold.79+0x3c/0x46
   __lock_acquire+0x1565/0x34a0
   lock_acquire+0x1d2/0x5a0
   _raw_spin_lock_irqsave+0x42/0x90
   rxe_pool_get_index+0x72/0x1d0 [rdma_rxe]
   rxe_get_av+0x168/0x2a0 [rdma_rxe]
</TASK>

From the above, in the function __rxe_add_to_pool, xa_lock is acquired. Then the function __rxe_add_to_pool is interrupted by softirq. The function rxe_pool_get_index will also acquire xa_lock.

Finally, the dead lock appears.

         CPU0
         ----
    lock(&xa->xa_lock#15);  <----- __rxe_add_to_pool
    <Interrupt>
      lock(&xa->xa_lock#15); <---- rxe_pool_get_index

                  *** DEADLOCK ***
Fixes 3225717f6dfa ("RDMA/rxe: Replace red-black trees by carrays")
Reported-and-tested-by Yi Zhang <[email protected]>

---
V5->V6: One dead lock fix in one commit
V4->V5: Commit logs are changed.
V3->V4: xa_lock_irq locks are used.
V2->V3: __rxe_add_to_pool is between spin_lock and spin_unlock, so
GFP_ATOMIC is used in __rxe_add_to_pool.
V1->V2: Replace GFP_KERNEL with GFP_ATOMIC

---
  drivers/infiniband/sw/rxe/rxe_pool.c | 11 +++++++----
  1 file changed, 7 insertions(+), 4 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +7 -5
index 87066d04ed18..67f1d4733682 100644
@@ -106,7 +106,7 @@ void rxe_pool_init(struct rxe_dev *rxe, struct rxe_pool *pool,
 
  	atomic_set(&pool->num_elem, 0);
 
-	xa_init_flags(&pool->xa, XA_FLAGS_ALLOC);
+	xa_init_flags(&pool->xa, XA_FLAGS_ALLOC | XA_FLAGS_LOCK_IRQ);
  	pool->limit.min = info->min_index;
  	pool->limit.max = info->max_index;
  }
@@ -155,6 +155,7 @@ void *rxe_alloc(struct rxe_pool *pool)  int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)  {
  	int err;
+	unsigned long flags;
 
  	if (WARN_ON(pool->flags & RXE_POOL_ALLOC))
  		return -EINVAL;
@@ -166,8 +167,10 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
  	elem->obj = (u8 *)elem - pool->elem_offset;
  	kref_init(&elem->ref_cnt);
 
-	err = xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
-			      &pool->next, GFP_KERNEL);
+	xa_lock_irqsave(&pool->xa, flags);
+	err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
+				&pool->next, GFP_ATOMIC);
+	xa_unlock_irqrestore(&pool->xa, flags);
  	if (err)
  		goto err_cnt;
 
@@ -201,7 +204,7 @@ static void rxe_elem_release(struct kref *kref)
  	struct rxe_pool_elem *elem = container_of(kref, typeof(*elem), ref_cnt);
  	struct rxe_pool *pool = elem->pool;
 
-	xa_erase(&pool->xa, elem->index);
+	xa_erase_irq(&pool->xa, elem->index);
 
  	if (pool->cleanup)
  		pool->cleanup(elem);
--

2.27.0

#9 lore raw
Quoted 5 lines · Yanjun Zhu
Quoted 2 lines · Robert B

Use of rcu_read_lock solves this problem. Rcu_read_lock and spinlock on same data can Co-exist at the same time. That is the whole point. All this is going away soon.

This is based on your unproved assumption.

We are running the same tests (pyverbs, rping, blktests, perftest, etc.) with the same configurations set (lockdep) and I am using rcu_read_lock for rxe_pool_get_index. I do not see any deadlocks or lockdep warnings. The idea of RCU is to separate accesses to a shared data structure into write operations and read operations. For write operations a normal spinlock is used to allow only one writer at a time to access the data. For RCU, unlike rwlocks, the reader is *not* locked at all and the writers are written so that they can change the data structure underneath the readers and the reader will either see the old data or the new data and not some combination of the two. This requires some care but there is library support for this usually denoted by xxx_rcu(). In the case of xarrays the whole library is RCU enabled so one can safely use xa_load() without holding a spinlock. The rcu_read_lock() API marks the critical section so the writers can make sure to wait long enough that there are no readers in the critical section before freeing the data. The rcu_read_lock() is also recursive. It just expands the critical section. In the case of rxe_pool.c the only rcu reader is rxe_pool_get_index() which looks up an object from its index by calling xa_load(). This normally runs in a tasklet but sometimes could run in process context. The only writers (as we have discussed) run in process context in a create or destroy verbs call. The writers sometimes call while holding a _saveirq spinlock from ib_create_ah() so lockdep issues warnings unless the spinlock in rxe_add_to_pool() is also an _irq spinlock. However it does not issue this warning when RCU is used because the reader *doesn't take a lock and therefore can't deadlock*. Thus the default xa_alloc_xxx() which takes a spin_lock() it OK and the sequence xa_lock_irq(); __xa_alloc_xxx(); xa_unlock_irq() is not needed.

Quoted 103 lines

Zhu Yanjun

Quoted 99 lines

Bob

-----Original Message-----
From: [email protected] <[email protected]>
Sent: Friday, April 22, 2022 2:44 PM
To: [email protected]; [email protected]; [email protected]; [email protected]
Cc: Yi Zhang <[email protected]>
Subject: [PATCHv6 1/4] RDMA/rxe: Fix dead lock caused by __rxe_add_to_pool interrupted by rxe_pool_get_index

From: Zhu Yanjun <[email protected]>

This is a dead lock problem.
The ah_pool xa_lock first is acquired in this:

{SOFTIRQ-ON-W} state was registered at:

   lock_acquire+0x1d2/0x5a0
   _raw_spin_lock+0x33/0x80
   __rxe_add_to_pool+0x183/0x230 [rdma_rxe]

Then ah_pool xa_lock is acquired in this:

{IN-SOFTIRQ-W}:

Call Trace:
  <TASK>
   dump_stack_lvl+0x44/0x57
   mark_lock.part.52.cold.79+0x3c/0x46
   __lock_acquire+0x1565/0x34a0
   lock_acquire+0x1d2/0x5a0
   _raw_spin_lock_irqsave+0x42/0x90
   rxe_pool_get_index+0x72/0x1d0 [rdma_rxe]
   rxe_get_av+0x168/0x2a0 [rdma_rxe]
</TASK>

 From the above, in the function __rxe_add_to_pool, xa_lock is acquired. Then the function __rxe_add_to_pool is interrupted by softirq. The function rxe_pool_get_index will also acquire xa_lock.

Finally, the dead lock appears.

         CPU0
         ----
    lock(&xa->xa_lock#15);  <----- __rxe_add_to_pool <Interrupt>
      lock(&xa->xa_lock#15); <---- rxe_pool_get_index

                  *** DEADLOCK ***

Fixes 3225717f6dfa ("RDMA/rxe: Replace red-black trees by carrays")
Reported-and-tested-by Yi Zhang <[email protected]>

---
V5->V6: One dead lock fix in one commit
V4->V5: Commit logs are changed.
V3->V4: xa_lock_irq locks are used.
V2->V3: __rxe_add_to_pool is between spin_lock and spin_unlock, so GFP_ATOMIC is used in __rxe_add_to_pool. V1->V2: Replace GFP_KERNEL with GFP_ATOMIC
---
  drivers/infiniband/sw/rxe/rxe_pool.c | 11 +++++++----

  1 file changed, 7 insertions(+), 4 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +6 -1
index 87066d04ed18..67f1d4733682 100644
@@ -106,7 +106,7 @@ void rxe_pool_init(struct rxe_dev *rxe, struct rxe_pool *pool,
         atomic_set(&pool->num_elem, 0);
   -    xa_init_flags(&pool->xa, XA_FLAGS_ALLOC);
+    xa_init_flags(&pool->xa, XA_FLAGS_ALLOC | XA_FLAGS_LOCK_IRQ);
       pool->limit.min = info->min_index;
       pool->limit.max = info->max_index;
   }
@@ -155,6 +155,7 @@ void *rxe_alloc(struct rxe_pool *pool)  int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)  {
       int err;
+    unsigned long flags;
         if (WARN_ON(pool->flags & RXE_POOL_ALLOC))
           return -EINVAL;
@@ -166,8 +167,10 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
       elem->obj = (u8 *)elem - pool->elem_offset;
       kref_init(&elem->ref_cnt);
   -    err = xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
-                  &pool->next, GFP_KERNEL);
+    xa_lock_irqsave(&pool->xa, flags);
+    err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
+                &pool->next, GFP_ATOMIC);
+    xa_unlock_irqrestore(&pool->xa, flags);
       if (err)
           goto err_cnt;
   @@ -201,7 +204,7 @@ static void rxe_elem_release(struct kref *kref)

      struct rxe_pool_elem *elem = container_of(kref, typeof(*elem), ref_cnt); struct rxe_pool *pool = elem->pool;
  -    xa_erase(&pool->xa, elem->index);
+    xa_erase_irq(&pool->xa, elem->index);
        if (pool->cleanup)
          pool->cleanup(elem);

Signature · 1 line
2.27.0
#10 lore raw
In reply to Yanjun Zhu
Quoted 5 lines · Yanjun Zhu
Quoted 2 lines · Robert B

Use of rcu_read_lock solves this problem. Rcu_read_lock and spinlock on same data can Co-exist at the same time. That is the whole point. All this is going away soon.

This is based on your unproved assumption.

No, Bob is right, RCU avoids the need for a BH lock on this XA, the only remaining issue is the AH creation atomic call path.

Jason

author
#11 lore raw
Quoted 7 lines · Jason Gunthorpe
Quoted 4 lines · Yanjun Zhu
Quoted 2 lines · Robert B

Use of rcu_read_lock solves this problem. Rcu_read_lock and spinlock on same data can Co-exist at the same time. That is the whole point. All this is going away soon.

This is based on your unproved assumption.

No, Bob is right, RCU avoids the need for a BH lock on this XA, the only remaining issue is the AH creation atomic call path.

If RCU is used, the similar issues like AH creation atomic call path will become more.

Zhu Yanjun

Quoted 2 lines

Jason

#12 lore raw
Quoted 4 lines · Yanjun Zhu
Quoted 2 lines · Jason Gunthorpe
Quoted 1 line · Yanjun Zhu

在 2022/4/22 23:57, Pearson, Robert B 写道:

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines
Quoted 2 lines

Use of rcu_read_lock solves this problem. Rcu_read_lock and spinlock on same data can Co-exist at the same time. That is the whole point. All this is going away soon.

This is based on your unproved assumption.

No, Bob is right, RCU avoids the need for a BH lock on this XA, the only remaining issue is the AH creation atomic call path.

If RCU is used, the similar issues like AH creation atomic call path will become more.

AH creation is unique, there will not be more cases like it.

Jason

#13 lore raw
In reply to [email protected]

Hi Yanjun, Bob

Could you tell me if the dead lock issue has been fixed by the following issue:
[PATCH v16 2/2] RDMA/rxe: Convert read side locking to rcu

Best Regards,
Xiao Yang

Quoted 98 lines · On 2022/4/23 3:44, [email protected] 写道:

From: Zhu Yanjun <[email protected]>

This is a dead lock problem.
The ah_pool xa_lock first is acquired in this:

{SOFTIRQ-ON-W} state was registered at:

   lock_acquire+0x1d2/0x5a0
   _raw_spin_lock+0x33/0x80
   __rxe_add_to_pool+0x183/0x230 [rdma_rxe]

Then ah_pool xa_lock is acquired in this:

{IN-SOFTIRQ-W}:

Call Trace:
  <TASK>
   dump_stack_lvl+0x44/0x57
   mark_lock.part.52.cold.79+0x3c/0x46
   __lock_acquire+0x1565/0x34a0
   lock_acquire+0x1d2/0x5a0
   _raw_spin_lock_irqsave+0x42/0x90
   rxe_pool_get_index+0x72/0x1d0 [rdma_rxe]
   rxe_get_av+0x168/0x2a0 [rdma_rxe]
</TASK>

From the above, in the function __rxe_add_to_pool,
xa_lock is acquired. Then the function __rxe_add_to_pool is interrupted by softirq. The function rxe_pool_get_index will also acquire xa_lock.

Finally, the dead lock appears.

         CPU0
         ----
    lock(&xa->xa_lock#15);  <----- __rxe_add_to_pool
    <Interrupt>
      lock(&xa->xa_lock#15); <---- rxe_pool_get_index

                  *** DEADLOCK ***
Fixes 3225717f6dfa ("RDMA/rxe: Replace red-black trees by carrays")
Reported-and-tested-by Yi Zhang <[email protected]>

---
V5->V6: One dead lock fix in one commit
V4->V5: Commit logs are changed.
V3->V4: xa_lock_irq locks are used.
V2->V3: __rxe_add_to_pool is between spin_lock and spin_unlock, so
GFP_ATOMIC is used in __rxe_add_to_pool.
V1->V2: Replace GFP_KERNEL with GFP_ATOMIC

---
  drivers/infiniband/sw/rxe/rxe_pool.c | 11 +++++++----
  1 file changed, 7 insertions(+), 4 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +7 -4
index 87066d04ed18..67f1d4733682 100644
@@ -106,7 +106,7 @@ void rxe_pool_init(struct rxe_dev *rxe, struct rxe_pool *pool,
 
  	atomic_set(&pool->num_elem, 0);
 
-	xa_init_flags(&pool->xa, XA_FLAGS_ALLOC);
+	xa_init_flags(&pool->xa, XA_FLAGS_ALLOC | XA_FLAGS_LOCK_IRQ);
  	pool->limit.min = info->min_index;
  	pool->limit.max = info->max_index;
  }
@@ -155,6 +155,7 @@ void *rxe_alloc(struct rxe_pool *pool)
  int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
  {
  	int err;
+	unsigned long flags;
 
  	if (WARN_ON(pool->flags & RXE_POOL_ALLOC))
  		return -EINVAL;
@@ -166,8 +167,10 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
  	elem->obj = (u8 *)elem - pool->elem_offset;
  	kref_init(&elem->ref_cnt);
 
-	err = xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
-			      &pool->next, GFP_KERNEL);
+	xa_lock_irqsave(&pool->xa, flags);
+	err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
+				&pool->next, GFP_ATOMIC);
+	xa_unlock_irqrestore(&pool->xa, flags);
  	if (err)
  		goto err_cnt;
 
@@ -201,7 +204,7 @@ static void rxe_elem_release(struct kref *kref)
  	struct rxe_pool_elem *elem = container_of(kref, typeof(*elem), ref_cnt);
  	struct rxe_pool *pool = elem->pool;
 
-	xa_erase(&pool->xa, elem->index);
+	xa_erase_irq(&pool->xa, elem->index);
 
  	if (pool->cleanup)
  		pool->cleanup(elem);
author
#14 lore raw
Quoted 5 lines · 在 2022/7/22 14:51, [email protected] 写道:

Hi Yanjun, Bob

Could you tell me if the dead lock issue has been fixed by the following issue:
[PATCH v16 2/2] RDMA/rxe: Convert read side locking to rcu

Hi, Xiao

Normally I applied this "RDMA/rxe: Fix dead lock caused by __rxe_add_to_pool interrupted by rxe_pool_get_index" patch

series to fix this problem. And I am not sure if this problem is fixed by "[PATCH v16 2/2] RDMA/rxe: Convert read side locking to rcu".

Zhu Yanjun

Quoted 103 lines

Best Regards,
Xiao Yang

Quoted 98 lines · On 2022/4/23 3:44, [email protected] 写道:

From: Zhu Yanjun <[email protected]>

This is a dead lock problem.
The ah_pool xa_lock first is acquired in this:

{SOFTIRQ-ON-W} state was registered at:

    lock_acquire+0x1d2/0x5a0
    _raw_spin_lock+0x33/0x80
    __rxe_add_to_pool+0x183/0x230 [rdma_rxe]

Then ah_pool xa_lock is acquired in this:

{IN-SOFTIRQ-W}:

Call Trace:
   <TASK>
    dump_stack_lvl+0x44/0x57
    mark_lock.part.52.cold.79+0x3c/0x46
    __lock_acquire+0x1565/0x34a0
    lock_acquire+0x1d2/0x5a0
    _raw_spin_lock_irqsave+0x42/0x90
    rxe_pool_get_index+0x72/0x1d0 [rdma_rxe]
    rxe_get_av+0x168/0x2a0 [rdma_rxe]
</TASK>

From the above, in the function __rxe_add_to_pool,
xa_lock is acquired. Then the function __rxe_add_to_pool is interrupted by softirq. The function rxe_pool_get_index will also acquire xa_lock.

Finally, the dead lock appears.

          CPU0
          ----
     lock(&xa->xa_lock#15);  <----- __rxe_add_to_pool
     <Interrupt>
       lock(&xa->xa_lock#15); <---- rxe_pool_get_index

                   *** DEADLOCK ***
Fixes 3225717f6dfa ("RDMA/rxe: Replace red-black trees by carrays")
Reported-and-tested-by Yi Zhang <[email protected]>

---
V5->V6: One dead lock fix in one commit
V4->V5: Commit logs are changed.
V3->V4: xa_lock_irq locks are used.
V2->V3: __rxe_add_to_pool is between spin_lock and spin_unlock, so
GFP_ATOMIC is used in __rxe_add_to_pool.
V1->V2: Replace GFP_KERNEL with GFP_ATOMIC

---
   drivers/infiniband/sw/rxe/rxe_pool.c | 11 +++++++----
   1 file changed, 7 insertions(+), 4 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +7 -4
index 87066d04ed18..67f1d4733682 100644
@@ -106,7 +106,7 @@ void rxe_pool_init(struct rxe_dev *rxe, struct rxe_pool *pool,
 
   	atomic_set(&pool->num_elem, 0);
 
-	xa_init_flags(&pool->xa, XA_FLAGS_ALLOC);
+	xa_init_flags(&pool->xa, XA_FLAGS_ALLOC | XA_FLAGS_LOCK_IRQ);
   	pool->limit.min = info->min_index;
   	pool->limit.max = info->max_index;
   }
@@ -155,6 +155,7 @@ void *rxe_alloc(struct rxe_pool *pool)
   int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
   {
   	int err;
+	unsigned long flags;
 
   	if (WARN_ON(pool->flags & RXE_POOL_ALLOC))
   		return -EINVAL;
@@ -166,8 +167,10 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
   	elem->obj = (u8 *)elem - pool->elem_offset;
   	kref_init(&elem->ref_cnt);
 
-	err = xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
-			      &pool->next, GFP_KERNEL);
+	xa_lock_irqsave(&pool->xa, flags);
+	err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
+				&pool->next, GFP_ATOMIC);
+	xa_unlock_irqrestore(&pool->xa, flags);
   	if (err)
   		goto err_cnt;
 
@@ -201,7 +204,7 @@ static void rxe_elem_release(struct kref *kref)
   	struct rxe_pool_elem *elem = container_of(kref, typeof(*elem), ref_cnt);
   	struct rxe_pool *pool = elem->pool;
 
-	xa_erase(&pool->xa, elem->index);
+	xa_erase_irq(&pool->xa, elem->index);
 
   	if (pool->cleanup)
   		pool->cleanup(elem);
#15 lore raw
Quoted 9 lines · Yanjun Zhu

Hi, Xiao

Normally I applied this "RDMA/rxe: Fix dead lock caused by __rxe_add_to_pool interrupted by rxe_pool_get_index" patch

series to fix this problem. And I am not sure if this problem is fixed by "[PATCH v16 2/2] RDMA/rxe: Convert read side locking to rcu".

Zhu Yanjun

Hi Yanjun,

I have confirmed that the problem has been fixed by:
[PATCH v16 2/2] RDMA/rxe: Convert read side locking to rcu

Best Regards,
Xiao Yang

#16 lore raw
Quoted 6 lines · On Fri, Jul 22, 2022 at 03:14:25PM +0000, [email protected] wrote:
Quoted 5 lines · Yanjun Zhu

Hi, Xiao

Normally I applied this "RDMA/rxe: Fix dead lock caused by __rxe_add_to_pool interrupted by rxe_pool_get_index" patch

Quoted 8 lines
Quoted 4 lines

series to fix this problem. And I am not sure if this problem is fixed by "[PATCH v16 2/2] RDMA/rxe: Convert read side locking to rcu".

Zhu Yanjun

Hi Yanjun,

I have confirmed that the problem has been fixed by:
[PATCH v16 2/2] RDMA/rxe: Convert read side locking to rcu

Thanks!

Jason

author
#17 lore raw
In reply to [email protected]
Quoted 6 lines · 在 2022/7/22 23:14, [email protected] 写道:
Quoted 5 lines · Yanjun Zhu

Hi, Xiao

Normally I applied this "RDMA/rxe: Fix dead lock caused by __rxe_add_to_pool interrupted by rxe_pool_get_index" patch

Quoted 8 lines
Quoted 4 lines

series to fix this problem. And I am not sure if this problem is fixed by "[PATCH v16 2/2] RDMA/rxe: Convert read side locking to rcu".

Zhu Yanjun

Hi Yanjun,

I have confirmed that the problem has been fixed by:
[PATCH v16 2/2] RDMA/rxe: Convert read side locking to rcu

Thanks.

Zhu Yanjun

Quoted 3 lines

Best Regards,
Xiao Yang