RDMA/rxe: Fix a dead lock problem
Subsystems (2)
PATCH v5 1/2 RDMA/rxe: Fix a dead lock problem
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.
[ 296.806097] CPU0
[ 296.808550] ----
[ 296.811003] lock(&xa->xa_lock#15); <----- __rxe_add_to_pool
[ 296.814583] <Interrupt>
[ 296.817209] lock(&xa->xa_lock#15); <---- rxe_pool_get_index
[ 296.820961]
*** DEADLOCK ***---V4->V5: Commit logs are changed to avoid confusion. 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 | 20 ++++++++++++++------ 1 file changed, 14 insertions(+), 6 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +14 -6
@@ -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; }
@@ -138,8 +138,10 @@ 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 ); +xa_lock_irq (&pool->xa); + err =__xa_alloc_cyclic (&pool->xa, &elem->index, elem, pool->limit, + &pool->next,GFP_KERNEL ); +xa_unlock_irq (&pool->xa);if (err) goto err_free;
@@ -155,6 +157,7 @@ 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 +169,10 @@ 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;
@@ -200,8 +205,11 @@ 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; +unsigned long flags; -xa_erase (&pool->xa, elem->index); +xa_lock_irqsave (&pool->xa, flags); +__xa_erase (&pool->xa, elem->index); +xa_unlock_irqrestore (&pool->xa, flags);if (pool->cleanup) pool->cleanup (elem);
Signature · 1 line
2.27.0
PATCH v2 2/2 RDMA/rxe: Use different xa locks on different path
From: Zhu Yanjun <[email protected]>
The function __rxe_add_to_pool is called on different path, 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.
---V1->V2: Use pool type instead of gfp_t;
--- drivers/infiniband/sw/rxe/rxe_pool.c | 18 +++++++++++++----- 1 file changed, 13 insertions(+), 5 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +13 -5
@@ -157,7 +157,6 @@ 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 ;
@@ -169,10 +168,19 @@ 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 { +xa_lock_irq (&pool->xa); + err =__xa_alloc_cyclic (&pool->xa, &elem->index, elem, pool->limit, + &pool->next,GFP_KERNEL ); +xa_unlock_irq (&pool->xa); + }if (err) goto err_cnt;
Signature · 1 line
2.27.0
Re: [PATCHv2 2/2] RDMA/rxe: Use different xa locks on different path
Quoted 27 lines · On Sat, Apr 16, 2022 at 10:43:43PM -0400, [email protected] wrote:
From: Zhu Yanjun <[email protected]>
The function
__rxe_add_to_poolis called on different path, and the requirement of the locks is different. The functionrxe_create_ahrequires xa_lock_irqsave/irqrestore while others only requirexa_lock_irq.Signed-off-by Zhu Yanjun <[email protected]>
V1->V2: Use pool type instead ofgfp_t;drivers/infiniband/sw/rxe/rxe_pool.c | 18 +++++++++++++----- 1 file changed, 13 insertions(+), 5 deletions(-)drivers/infiniband/sw/rxe/rxe_pool.c +1
-2+++ b/drivers/infiniband/sw/rxe/rxe_pool.c@@ -157,7 +157,6 @@ 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 ;@@ -169,10 +168,19 @@ 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);
drivers/infiniband/sw/rxe/rxe_pool.c int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem) · line 177
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_pool.c +7
-3- 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);
This is still not right - you have to test RDMA_CREATE_AH_SLEEPABLE for the AH path as well.
Jason
Quoted 68 lines · On Sat, Apr 16, 2022 at 10:43:42PM -0400, [email protected] wrote:
From: Zhu Yanjun <[email protected]>
This is a dead lock problem.
Theah_poolxa_lockfirst 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_poolxa_lockis 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_lockis acquired. Then the function__rxe_add_to_poolis interrupted by softirq. The functionrxe_pool_get_indexwill also acquirexa_lock.Finally, the dead lock appears.
[ 296.806097] CPU0 [ 296.808550] ---- [ 296.811003] lock(&xa->xa_lock#15); <----- __rxe_add_to_pool [ 296.814583] <Interrupt> [ 296.817209] lock(&xa->xa_lock#15); <---- rxe_pool_get_index [ 296.820961] *** DEADLOCK ***Fixes 3225717f6dfa ("RDMA/rxe: Replace red-black trees by carrays")Reported-and-tested-by Yi Zhang <[email protected]>Signed-off-by Zhu Yanjun <[email protected]>---
V4->V5: Commit logs are changed to avoid confusion.V3->V4:xa_lock_irqlocks are used.V2->V3:__rxe_add_to_poolis betweenspin_lockandspin_unlock, so
GFP_ATOMICis used in__rxe_add_to_pool.V1->V2: ReplaceGFP_KERNELwithGFP_ATOMIC--- drivers/infiniband/sw/rxe/rxe_pool.c | 20 ++++++++++++++------ 1 file changed, 14 insertions(+), 6 deletions(-)drivers/infiniband/sw/rxe/rxe_pool.c +1
-1@@ -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; }
drivers/infiniband/sw/rxe/rxe_pool.c void *rxe_alloc(struct rxe_pool *pool) · line 144
Quoted 9 lines
drivers/infiniband/sw/rxe/rxe_pool.c +4
-2@@ -138,8 +138,10 @@ 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 ); +xa_lock_irq (&pool->xa); + err =__xa_alloc_cyclic (&pool->xa, &elem->index, elem, pool->limit, + &pool->next,GFP_KERNEL ); +xa_unlock_irq (&pool->xa);
I may admit that I didn't follow your previous discussions, so maybe you already explained it. But why do you need xa_lock_irq() here?
Thanks
Quoted 40 lines
drivers/infiniband/sw/rxe/rxe_pool.c +9
-3if (err) goto err_free;@@ -155,6 +157,7 @@ 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 +169,10 @@ 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;@@ -200,8 +205,11 @@ 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; +unsigned long flags; -xa_erase (&pool->xa, elem->index); +xa_lock_irqsave (&pool->xa, flags); +__xa_erase (&pool->xa, elem->index); +xa_unlock_irqrestore (&pool->xa, flags);if (pool->cleanup) pool->cleanup (elem);Signature · 1 line
2.27.0
Quoted 69 lines · Leon Romanovsky
Quoted 68 lines · On Sat, Apr 16, 2022 at 10:43:42PM -0400, [email protected] wrote:
From: Zhu Yanjun <[email protected]>
This is a dead lock problem.
Theah_poolxa_lockfirst 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_poolxa_lockis 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_lockis acquired. Then the function__rxe_add_to_poolis interrupted by softirq. The functionrxe_pool_get_indexwill also acquirexa_lock.Finally, the dead lock appears.
[ 296.806097] CPU0 [ 296.808550] ---- [ 296.811003] lock(&xa->xa_lock#15); <----- __rxe_add_to_pool [ 296.814583] <Interrupt> [ 296.817209] lock(&xa->xa_lock#15); <---- rxe_pool_get_index [ 296.820961] *** DEADLOCK ***Fixes 3225717f6dfa ("RDMA/rxe: Replace red-black trees by carrays")Reported-and-tested-by Yi Zhang <[email protected]>Signed-off-by Zhu Yanjun <[email protected]>
V4->V5: Commit logs are changed to avoid confusion.V3->V4:xa_lock_irqlocks are used.V2->V3:__rxe_add_to_poolis betweenspin_lockandspin_unlock, so
GFP_ATOMICis used in__rxe_add_to_pool.V1->V2: ReplaceGFP_KERNELwithGFP_ATOMICdrivers/infiniband/sw/rxe/rxe_pool.c | 20 ++++++++++++++------ 1 file changed, 14 insertions(+), 6 deletions(-)drivers/infiniband/sw/rxe/rxe_pool.c +2
-1+++ b/drivers/infiniband/sw/rxe/rxe_pool.c@@ -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; }@@ -138,8 +138,10 @@ void *rxe_alloc(struct rxe_pool *pool) elem->obj = obj;kref_init (&elem->ref_cnt);
Quoted 6 lines
drivers/infiniband/sw/rxe/rxe_pool.c void *rxe_alloc(struct rxe_pool *pool) · line 144
Quoted 6 lines
drivers/infiniband/sw/rxe/rxe_pool.c +4
-2- err =xa_alloc_cyclic (&pool->xa, &elem->index, elem, pool->limit, - &pool->next,GFP_KERNEL ); +xa_lock_irq (&pool->xa); + err =__xa_alloc_cyclic (&pool->xa, &elem->index, elem, pool->limit, + &pool->next,GFP_KERNEL ); +xa_unlock_irq (&pool->xa);
It should just use xa_alloc_cyclic_irq() and xa_erase_irq(). Don't open code the lock.
Quoted 2 lines
I may admit that I didn't follow your previous discussions, so maybe you already explained it. But why do you need
xa_lock_irq()here?
The spinlock type needs to be consistent in all users.
You can only use the naked version if the spinlock is always obtained from a process context.
You can only use bh version if the spinlock is always obtained from a process context or bh/softirq
You can always use the irq version
What I don't understand is why IRQ and not BH? AFAIK there is no case where rxe is called from a real IRQ, right? Or is it because you can't nest BH under the IRQ spinlock from CM?
Jason
Quoted 70 lines · Jason Gunthorpe
Quoted 69 lines · Leon Romanovsky
Quoted 68 lines · On Sat, Apr 16, 2022 at 10:43:42PM -0400, [email protected] wrote:
From: Zhu Yanjun <[email protected]>
This is a dead lock problem.
Theah_poolxa_lockfirst 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_poolxa_lockis 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_lockis acquired. Then the function__rxe_add_to_poolis interrupted by softirq. The functionrxe_pool_get_indexwill also acquirexa_lock.Finally, the dead lock appears.
[ 296.806097] CPU0 [ 296.808550] ---- [ 296.811003] lock(&xa->xa_lock#15); <----- __rxe_add_to_pool [ 296.814583] <Interrupt> [ 296.817209] lock(&xa->xa_lock#15); <---- rxe_pool_get_index [ 296.820961] *** DEADLOCK ***Fixes 3225717f6dfa ("RDMA/rxe: Replace red-black trees by carrays")Reported-and-tested-by Yi Zhang <[email protected]>Signed-off-by Zhu Yanjun <[email protected]>
V4->V5: Commit logs are changed to avoid confusion.V3->V4:xa_lock_irqlocks are used.V2->V3:__rxe_add_to_poolis betweenspin_lockandspin_unlock, so
GFP_ATOMICis used in__rxe_add_to_pool.V1->V2: ReplaceGFP_KERNELwithGFP_ATOMICdrivers/infiniband/sw/rxe/rxe_pool.c | 20 ++++++++++++++------ 1 file changed, 14 insertions(+), 6 deletions(-)drivers/infiniband/sw/rxe/rxe_pool.c +2
-1+++ b/drivers/infiniband/sw/rxe/rxe_pool.c@@ -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; }@@ -138,8 +138,10 @@ void *rxe_alloc(struct rxe_pool *pool) elem->obj = obj;kref_init (&elem->ref_cnt);
Quoted 9 lines
Quoted 6 lines
drivers/infiniband/sw/rxe/rxe_pool.c void *rxe_alloc(struct rxe_pool *pool) · line 144
Quoted 6 lines
drivers/infiniband/sw/rxe/rxe_pool.c +4
-2- err =xa_alloc_cyclic (&pool->xa, &elem->index, elem, pool->limit, - &pool->next,GFP_KERNEL ); +xa_lock_irq (&pool->xa); + err =__xa_alloc_cyclic (&pool->xa, &elem->index, elem, pool->limit, + &pool->next,GFP_KERNEL ); +xa_unlock_irq (&pool->xa);It should just use
xa_alloc_cyclic_irq()andxa_erase_irq(). Don't open code the lock.
Got it. I will use the above functions.
Quoted 9 lines
Quoted 3 lines
I may admit that I didn't follow your previous discussions, so maybe you already explained it. But why do you need
xa_lock_irq()here?The spinlock type needs to be consistent in all users.
You can only use the naked version if the spinlock is always obtained from a process context.
Quoted 8 lines
You can only use bh version if the spinlock is always obtained from a process context or bh/softirq
You can always use the irq version
What I don't understand is why IRQ and not BH? AFAIK there is no case where rxe is called from a real IRQ, right? Or is it because you can't nest BH under the IRQ spinlock from CM?
Sure. I will use IRQ spinlock. The reason is as below:
1. https://patchwork.kernel.org/project/linux-rdma/patch/[email protected]/
2. https://patchwork.kernel.org/project/linux-rdma/patch/[email protected]/
The above 2 links are why I used IRQ.
Zhu Yanjun
Quoted 2 lines
Jason