rdma_rxe: call comp_handler without holding cq->cq_lock
Subsystems (2)
PATCH v2 1/1 rdma_rxe: call comp_handler without holding cq->cq_lock
Allow the comp_handler callback implementation to call ib_poll_cq(). A call to ib_poll_cq() calls rxe_poll_cq() with the rdma_rxe driver. And rxe_poll_cq() locks cq->cq_lock. That leads to a spinlock deadlock.
The Mellanox and Intel drivers allow a comp_handler callback implementation to call ib_poll_cq().
Avoid the deadlock by calling the comp_handler callback without holding cq->cq_lock.
Changelog:
v1: https://lore.kernel.org/all/[email protected]/
v1 -> v2:
- Only reset cq->notify to 0 when invoking the comp_handler
====================
--- drivers/infiniband/sw/rxe/rxe_cq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/infiniband/sw/rxe/rxe_cq.c +5 -1
@@ -88,6 +88,7 @@ int rxe_cq_post(struct rxe_cq *cq, struct rxe_cqe *cqe, int solicited)int full;void *addr;unsigned long flags; +bool invoke_handler = false;spin_lock_irqsave (&cq->cq_lock, flags);
@@ -113,11 +114,14 @@ int rxe_cq_post(struct rxe_cq *cq, struct rxe_cqe *cqe, int solicited)if ((cq->notify &IB_CQ_NEXT_COMP ) || (cq->notify &IB_CQ_SOLICITED && solicited)) { cq->notify =0 ; - cq->ibcq.comp_handler (&cq->ibcq, cq->ibcq.cq_context); + invoke_handler = true; }spin_unlock_irqrestore (&cq->cq_lock, flags); +if (invoke_handler) + cq->ibcq.comp_handler (&cq->ibcq, cq->ibcq.cq_context); +return 0 ; }
Signature · 1 line
2.50.1
Quoted 3 lines · Philipp Reisner
Allow the
comp_handlercallback implementation to callib_poll_cq(). A call toib_poll_cq()callsrxe_poll_cq()with therdma_rxedriver. Andrxe_poll_cq()lockscq->cq_lock. That leads to a spinlock deadlock.
Can you please be more specific about the deadlock? Please write call stack to describe it.
Quoted 49 lines
The Mellanox and Intel drivers allow a
comp_handlercallback implementation to callib_poll_cq().Avoid the deadlock by calling the
comp_handlercallback without holdingcq->cq_lock.Changelog:
v1: https://lore.kernel.org/all/[email protected]/
v1 -> v2:
- Only resetcq->notifyto 0 when invoking thecomp_handler
====================Signed-off-by Philipp Reisner <[email protected]>Reviewed-by Zhu Yanjun <[email protected]>--- drivers/infiniband/sw/rxe/rxe_cq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/infiniband/sw/rxe/rxe_cq.c +5
-1@@ -88,6 +88,7 @@ int rxe_cq_post(struct rxe_cq *cq, struct rxe_cqe *cqe, int solicited)int full;void *addr;unsigned long flags; +bool invoke_handler = false;spin_lock_irqsave (&cq->cq_lock, flags);@@ -113,11 +114,14 @@ int rxe_cq_post(struct rxe_cq *cq, struct rxe_cqe *cqe, int solicited)if ((cq->notify &IB_CQ_NEXT_COMP ) || (cq->notify &IB_CQ_SOLICITED && solicited)) { cq->notify =0 ; - cq->ibcq.comp_handler (&cq->ibcq, cq->ibcq.cq_context); + invoke_handler = true; }spin_unlock_irqrestore (&cq->cq_lock, flags); +if (invoke_handler) + cq->ibcq.comp_handler (&cq->ibcq, cq->ibcq.cq_context); +return 0 ; }Signature · 1 line
2.50.1
Quoted 9 lines · Leon Romanovsky
Quoted 3 lines · Philipp Reisner
Allow the
comp_handlercallback implementation to callib_poll_cq(). A call toib_poll_cq()callsrxe_poll_cq()with therdma_rxedriver. Andrxe_poll_cq()lockscq->cq_lock. That leads to a spinlock deadlock.Can you please be more specific about the deadlock? Please write call stack to describe it.
Instead of a call stack, I write it from top to bottom:
The line numbers in the .c files are valid for Linux-6.16:
1 rxe_cq_post() [rxe_cq.c:85] 2 spin_lock_irqsave() [rxe_cq.c:93] 3 cq->ibcq.comp_handler() [rxe_cq.c:116] 4 some_comp_handler() 5 ib_poll_cq() 6 cq->device->ops.poll_cq() [ib_verbs.h:4037] 7 rxe_poll_cq() [rxe_verbs.c:1165] 8 spin_lock_irqsave() [rxe_verbs.c:1172]
In line 8 of this call graph, it deadlocks because the spinlock was already acquired in line 2 of the call graph.
This patch changes that call graph to:
(Line numbers now valid with the patch in discussion applied)
1 rxe_cq_post() [rxe_cq.c:85] 2 spin_lock_irqsave() [rxe_cq.c:93] 3 spin_unlock_irqrestore() [rxe_cq.c:120] 4 cq->ibcq.comp_handler() [rxe_cq.c:123] 5 some_comp_handler() 6 ib_poll_cq() 7 cq->device->ops.poll_cq() [ib_verbs.h:4037] 8 rxe_poll_cq() [rxe_verbs.c:1165] 9 spin_lock_irqsave() [rxe_verbs.c:1172]
With the patch, there is no deadlock in line 9 of the call graph, as the spinlock was released in line 3.
I hope that helps.
[...]
Quoted 23 lines · Philipp Reisner
Quoted 9 lines · Leon Romanovsky
Quoted 3 lines · Philipp Reisner
Allow the
comp_handlercallback implementation to callib_poll_cq(). A call toib_poll_cq()callsrxe_poll_cq()with therdma_rxedriver. Andrxe_poll_cq()lockscq->cq_lock. That leads to a spinlock deadlock.Can you please be more specific about the deadlock? Please write call stack to describe it.
Instead of a call stack, I write it from top to bottom:
The line numbers in the .c files are valid for Linux-6.16:
1 rxe_cq_post() [rxe_cq.c:85] 2 spin_lock_irqsave() [rxe_cq.c:93] 3 cq->ibcq.comp_handler() [rxe_cq.c:116] 4 some_comp_handler() 5 ib_poll_cq() 6 cq->device->ops.poll_cq() [ib_verbs.h:4037] 7 rxe_poll_cq() [rxe_verbs.c:1165] 8 spin_lock_irqsave() [rxe_verbs.c:1172]
Quoted 2 lines
In line 8 of this call graph, it deadlocks because the spinlock was already acquired in line 2 of the call graph.
Is this even legal in verbs? I'm not sure you can do pull cq from a interrupt driven comp handler.. Is something already doing this intree?
Jason
Quoted 25 lines · Jason Gunthorpe
Quoted 23 lines · Philipp Reisner
Quoted 9 lines · Leon Romanovsky
Quoted 3 lines · Philipp Reisner
Allow the
comp_handlercallback implementation to callib_poll_cq(). A call toib_poll_cq()callsrxe_poll_cq()with therdma_rxedriver. Andrxe_poll_cq()lockscq->cq_lock. That leads to a spinlock deadlock.Can you please be more specific about the deadlock? Please write call stack to describe it.
Instead of a call stack, I write it from top to bottom:
The line numbers in the .c files are valid for Linux-6.16:
1 rxe_cq_post() [rxe_cq.c:85] 2 spin_lock_irqsave() [rxe_cq.c:93] 3 cq->ibcq.comp_handler() [rxe_cq.c:116] 4 some_comp_handler() 5 ib_poll_cq() 6 cq->device->ops.poll_cq() [ib_verbs.h:4037] 7 rxe_poll_cq() [rxe_verbs.c:1165] 8 spin_lock_irqsave() [rxe_verbs.c:1172]
Quoted 6 lines
Quoted 2 lines
In line 8 of this call graph, it deadlocks because the spinlock was already acquired in line 2 of the call graph.
Is this even legal in verbs? I'm not sure you can do pull cq from a interrupt driven comp handler.. Is something already doing this intree?
The file drivers/infiniband/sw/rdmavt/cq.c has this comment:
/*
* The completion handler will most likely rearm the notification
* and poll for all pending entries. If a new completion entry
* is added while we are in this routine, queue_work()
* won't call us again until we return so we check triggered to
* see if we need to call the handler again.
*/
Also, Intel and Mellanox cards and drivers allow calling ib_poll_cq() from the completion handler.
The problem exists only with the RXE driver.
Quoted 42 lines · Philipp Reisner
Quoted 25 lines · Jason Gunthorpe
Quoted 23 lines · Philipp Reisner
Quoted 9 lines · Leon Romanovsky
Quoted 3 lines · Philipp Reisner
Allow the
comp_handlercallback implementation to callib_poll_cq(). A call toib_poll_cq()callsrxe_poll_cq()with therdma_rxedriver. Andrxe_poll_cq()lockscq->cq_lock. That leads to a spinlock deadlock.Can you please be more specific about the deadlock? Please write call stack to describe it.
Instead of a call stack, I write it from top to bottom:
The line numbers in the .c files are valid for Linux-6.16:
1 rxe_cq_post() [rxe_cq.c:85] 2 spin_lock_irqsave() [rxe_cq.c:93] 3 cq->ibcq.comp_handler() [rxe_cq.c:116] 4 some_comp_handler() 5 ib_poll_cq() 6 cq->device->ops.poll_cq() [ib_verbs.h:4037] 7 rxe_poll_cq() [rxe_verbs.c:1165] 8 spin_lock_irqsave() [rxe_verbs.c:1172]Quoted 6 lines
Quoted 2 lines
In line 8 of this call graph, it deadlocks because the spinlock was already acquired in line 2 of the call graph.
Is this even legal in verbs? I'm not sure you can do pull cq from a interrupt driven comp handler.. Is something already doing this intree?
The file
drivers/infiniband/sw/rdmavt/cq.chas this comment:
/*
* The completion handler will most likely rearm the notification
* and poll for all pending entries. If a new completion entry
* is added while we are in this routine,queue_work()
* won't call us again until we return so we check triggered to
* see if we need to call the handler again.
*/
Quoted 2 lines
Also, Intel and Mellanox cards and drivers allow calling
ib_poll_cq()from the completion handler.
And do these drivers drop CQ lock like you are proposing here?
Quoted 2 lines
The problem exists only with the RXE driver.
Jason Gunthorpe
maintainerQuoted 9 lines · Philipp Reisner
Allow the
comp_handlercallback implementation to callib_poll_cq(). A call toib_poll_cq()callsrxe_poll_cq()with therdma_rxedriver. Andrxe_poll_cq()lockscq->cq_lock. That leads to a spinlock deadlock.The Mellanox and Intel drivers allow a
comp_handlercallback implementation to callib_poll_cq().Avoid the deadlock by calling the
comp_handlercallback without holdingcq->cq_lock.
I spent some time looking at this, and I think the basic statement above is right. The comp_handler should be able to call poll_cq/etc
rxe holding a lock it used to push a CQE is not correct.
However! The comp_handler is also supposed to be single threaded by the driver, I don't think ULPs are prepared to handle concurrent calls to comp_handler.
Other HW drivers run their comp_handlers from an EQ which is both single threaded and does not exclude poll_cq/etc.
So while removing the cq lock here is correct from the perspective of allowing poll_cq, I could not find any locking in rxe that made do_complete() be single threaded.
Please send a v2, either explain how the do_complete is single threaded in a comment above the comp_handler call, or make it be single threaded.
Thanks,
Jason
Hi Jason,
Quoted 23 lines · Jason Gunthorpe
Quoted 9 lines · Philipp Reisner
Allow the
comp_handlercallback implementation to callib_poll_cq(). A call toib_poll_cq()callsrxe_poll_cq()with therdma_rxedriver. Andrxe_poll_cq()lockscq->cq_lock. That leads to a spinlock deadlock.The Mellanox and Intel drivers allow a
comp_handlercallback implementation to callib_poll_cq().Avoid the deadlock by calling the
comp_handlercallback without holdingcq->cq_lock.I spent some time looking at this, and I think the basic statement above is right. The
comp_handlershould be able to call poll_cq/etcrxe holding a lock it used to push a CQE is not correct.
However! The
comp_handleris also supposed to be single threaded by the driver, I don't think ULPs are prepared to handle concurrent calls tocomp_handler.Other HW drivers run their
comp_handlersfrom an EQ which is both single threaded and does not exclude poll_cq/etc.
Quoted 8 lines
So while removing the cq lock here is correct from the perspective of allowing
poll_cq, I could not find any locking in rxe that madedo_complete()be single threaded.Please send a v2, either explain how the
do_completeis single threaded in a comment above thecomp_handlercall, or make it be single threaded.
Thanks for following up. Sure, I will send a new version of it, it will be v3, as this is already the discussion on v2.
Best regards,
Philipp