Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel
On Thu, Oct 20, 2011 at 12:30:08 -0500, Steve Wise wrote: On 10/20/2011 12:25 PM, parav.pan...@emulex.com wrote: You are right, cq_lock will result into dead lock. Should there be a additional compl_handler spin_lock? I was measuring performance impact for adding it, and, irq_save() and irq_restore() variant showed additional 200 cycles, which I believe should be o.k.? Parav Yea, I think adding a comp_handler spin lock would do the trick. I'll let Kumar come up with this change. I will cook up a patch and post soon. Thanks, Kumar. -Original Message- From: Steve Wise [mailto:sw...@opengridcomputing.com] Sent: Thursday, October 20, 2011 9:52 PM To: Roland Dreier Cc: Pandit, Parav; kuma...@chelsio.com; linux-rdma@vger.kernel.org; d...@chelsio.com Subject: Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel On 10/20/2011 11:06 AM, Roland Dreier wrote: On Thu, Oct 20, 2011 at 2:11 AM,parav.pan...@emulex.com wrote: http://lxr.linux.no/#linux+v3.0.4/Documentation/infiniband/core_locking.txt Line no 66 to 97 states that - at a given point of time, there should be only one callback per CQ should be active. Is this ensured? compl_handler() is invoked from multiple places flush_qp() and post_qp_event() to my mind. Yes, does look like an issue with this patch :( And I think this bug exists in the T3 driver too. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel
On 10/20/2011 04:11 AM, parav.pan...@emulex.com wrote:
http://lxr.linux.no/#linux+v3.0.4/Documentation/infiniband/core_locking.txt
Line no 66 to 97 states that - at a given point of time, there should be only
one callback per CQ should be active.
Is this ensured?
compl_handler() is invoked from multiple places flush_qp() and post_qp_event()
to my mind.
I think you're correct. post_qp_event() can be called indirectly on the ingress side of things concurrently with the
application moving the QP to a state that causes a flush. My first thought is to serialize the call to the
compl_handler with the CQ spinlock, but I'm pretty certain that we can't hold that around an upcall because the app most
likely turns around and calls ib_poll_cq() which would attempt to acquire this same lock.
Parav
-Original Message-
From: linux-rdma-ow...@vger.kernel.org
[mailto:linux-rdma-ow...@vger.kernel.org] On Behalf Of Steve Wise
Sent: Thursday, October 13, 2011 9:43 PM
To: Roland Dreier
Cc: Kumar Sanghvi; linux-rdma@vger.kernel.org; d...@chelsio.com
Subject: Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel
On 10/13/2011 11:01 AM, Roland Dreier wrote:
Would this generate a completion event even if no completion entries are queued?
I guess it can if the QP has no WRs posted at all.
Maybe I'm misunderstanding, but this sounds like a bandaid for broken
applications,
and a bandaid that other hardware drivers won't implement.
I'm not sure other drivers have this issue. For example, when a mlx or mthca
QP moves out of RTS, does the HW flush the
pending recv WRs? For Chelsio devices, the provider driver/library must handle
this.
This logic is needed to adhere to the iwarp verbs spec which states that when
the QP moves out of RTS, all WRs that are
pending get completed with FLUSH status. For T3/T4 devices, this is all done
in software. For user mode, the provider
library has to flush the QP (IE the kernel doesn't own the queue state).
The idea is that if an application is expecting a CQ event notification when
the QP moves out of RTS, and there are only
recv wrs posted, then the T4 (and T3 does this too) driver must post this CQ
notification, in addition to marking the CQ
as in error which means some QP bound to this CQ needs flushing. Then when
the app wakes up and polls the CQ, the
libcxgb4 code will flush the QPs in error and thus CQEs will be inserted into
the CQ. We have seen certain applications
that rely on this event to discover that a QP has moved out of RTS. IE they
don't look at async QP events nor RDMACM
events.
Steve.
On Thu, Oct 13, 2011 at 1:21 AM, Kumar Sanghvikuma...@chelsio.com wrote:
At the time when peer closes connection, iw_cxgb4 will not
send a cq event if ibqp.uobject exists. In that case, its possible
for user application to get blocked in ibv_get_cq_event.
To resolve this, call the cq's comp_handler to unblock any read
from ibv_get_cq_event.
Signed-off-by: Kumar Sanghvikuma...@chelsio.com
---
drivers/infiniband/hw/cxgb4/qp.c |6 +-
1 files changed, 5 insertions(+), 1 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index ec3ce67..b59b56c 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -970,8 +970,12 @@ static void flush_qp(struct c4iw_qp *qhp)
if (qhp-ibqp.uobject) {
t4_set_wq_in_error(qhp-wq);
t4_set_cq_in_error(rchp-cq);
- if (schp != rchp)
+ (*rchp-ibcq.comp_handler)(rchp-ibcq, rchp-ibcq.cq_context);
+ if (schp != rchp) {
t4_set_cq_in_error(schp-cq);
+ (*schp-ibcq.comp_handler)(schp-ibcq,
+ schp-ibcq.cq_context);
+ }
return;
}
__flush_qp(qhp, rchp, schp);
--
1.7.1
--
To unsubscribe from this list: send the line unsubscribe linux-rdma in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line unsubscribe linux-rdma in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line unsubscribe linux-rdma in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel
On Thu, Oct 20, 2011 at 2:11 AM, parav.pan...@emulex.com wrote: http://lxr.linux.no/#linux+v3.0.4/Documentation/infiniband/core_locking.txt Line no 66 to 97 states that - at a given point of time, there should be only one callback per CQ should be active. Is this ensured? compl_handler() is invoked from multiple places flush_qp() and post_qp_event() to my mind. Yes, does look like an issue with this patch :( - R. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel
On 10/20/2011 11:06 AM, Roland Dreier wrote: On Thu, Oct 20, 2011 at 2:11 AM,parav.pan...@emulex.com wrote: http://lxr.linux.no/#linux+v3.0.4/Documentation/infiniband/core_locking.txt Line no 66 to 97 states that - at a given point of time, there should be only one callback per CQ should be active. Is this ensured? compl_handler() is invoked from multiple places flush_qp() and post_qp_event() to my mind. Yes, does look like an issue with this patch :( And I think this bug exists in the T3 driver too. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
RE: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel
You are right, cq_lock will result into dead lock. Should there be a additional compl_handler spin_lock? I was measuring performance impact for adding it, and, irq_save() and irq_restore() variant showed additional 200 cycles, which I believe should be o.k.? Parav -Original Message- From: Steve Wise [mailto:sw...@opengridcomputing.com] Sent: Thursday, October 20, 2011 9:52 PM To: Roland Dreier Cc: Pandit, Parav; kuma...@chelsio.com; linux-rdma@vger.kernel.org; d...@chelsio.com Subject: Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel On 10/20/2011 11:06 AM, Roland Dreier wrote: On Thu, Oct 20, 2011 at 2:11 AM,parav.pan...@emulex.com wrote: http://lxr.linux.no/#linux+v3.0.4/Documentation/infiniband/core_locking.txt Line no 66 to 97 states that - at a given point of time, there should be only one callback per CQ should be active. Is this ensured? compl_handler() is invoked from multiple places flush_qp() and post_qp_event() to my mind. Yes, does look like an issue with this patch :( And I think this bug exists in the T3 driver too. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel
On 10/20/2011 12:25 PM, parav.pan...@emulex.com wrote: You are right, cq_lock will result into dead lock. Should there be a additional compl_handler spin_lock? I was measuring performance impact for adding it, and, irq_save() and irq_restore() variant showed additional 200 cycles, which I believe should be o.k.? Parav Yea, I think adding a comp_handler spin lock would do the trick. I'll let Kumar come up with this change. -Original Message- From: Steve Wise [mailto:sw...@opengridcomputing.com] Sent: Thursday, October 20, 2011 9:52 PM To: Roland Dreier Cc: Pandit, Parav; kuma...@chelsio.com; linux-rdma@vger.kernel.org; d...@chelsio.com Subject: Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel On 10/20/2011 11:06 AM, Roland Dreier wrote: On Thu, Oct 20, 2011 at 2:11 AM,parav.pan...@emulex.com wrote: http://lxr.linux.no/#linux+v3.0.4/Documentation/infiniband/core_locking.txt Line no 66 to 97 states that - at a given point of time, there should be only one callback per CQ should be active. Is this ensured? compl_handler() is invoked from multiple places flush_qp() and post_qp_event() to my mind. Yes, does look like an issue with this patch :( And I think this bug exists in the T3 driver too. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel
On Thu, Oct 13, 2011 at 9:12 AM, Steve Wise sw...@opengridcomputing.com wrote: I'm not sure other drivers have this issue. For example, when a mlx or mthca QP moves out of RTS, does the HW flush the pending recv WRs? For Chelsio devices, the provider driver/library must handle this. Yeah, mthca will generate the flush events in hardware IIRC. So OK, this change makes sense to me... only risk is a spurious event for empty queues. - R. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel
Acked-by: Steve Wise sw...@opengridcomputing.com -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel
Would this generate a completion event even if no completion entries are queued?
Maybe I'm misunderstanding, but this sounds like a bandaid for broken
applications,
and a bandaid that other hardware drivers won't implement.
On Thu, Oct 13, 2011 at 1:21 AM, Kumar Sanghvi kuma...@chelsio.com wrote:
At the time when peer closes connection, iw_cxgb4 will not
send a cq event if ibqp.uobject exists. In that case, its possible
for user application to get blocked in ibv_get_cq_event.
To resolve this, call the cq's comp_handler to unblock any read
from ibv_get_cq_event.
Signed-off-by: Kumar Sanghvi kuma...@chelsio.com
---
drivers/infiniband/hw/cxgb4/qp.c | 6 +-
1 files changed, 5 insertions(+), 1 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/qp.c
b/drivers/infiniband/hw/cxgb4/qp.c
index ec3ce67..b59b56c 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -970,8 +970,12 @@ static void flush_qp(struct c4iw_qp *qhp)
if (qhp-ibqp.uobject) {
t4_set_wq_in_error(qhp-wq);
t4_set_cq_in_error(rchp-cq);
- if (schp != rchp)
+ (*rchp-ibcq.comp_handler)(rchp-ibcq,
rchp-ibcq.cq_context);
+ if (schp != rchp) {
t4_set_cq_in_error(schp-cq);
+ (*schp-ibcq.comp_handler)(schp-ibcq,
+ schp-ibcq.cq_context);
+ }
return;
}
__flush_qp(qhp, rchp, schp);
--
1.7.1
--
To unsubscribe from this list: send the line unsubscribe linux-rdma in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCH] RDMA/cxgb4: Unblock reads on comp_channel
On 10/13/2011 11:01 AM, Roland Dreier wrote:
Would this generate a completion event even if no completion entries are queued?
I guess it can if the QP has no WRs posted at all.
Maybe I'm misunderstanding, but this sounds like a bandaid for broken
applications,
and a bandaid that other hardware drivers won't implement.
I'm not sure other drivers have this issue. For example, when a mlx or mthca QP moves out of RTS, does the HW flush the
pending recv WRs? For Chelsio devices, the provider driver/library must handle this.
This logic is needed to adhere to the iwarp verbs spec which states that when the QP moves out of RTS, all WRs that are
pending get completed with FLUSH status. For T3/T4 devices, this is all done in software. For user mode, the provider
library has to flush the QP (IE the kernel doesn't own the queue state).
The idea is that if an application is expecting a CQ event notification when the QP moves out of RTS, and there are only
recv wrs posted, then the T4 (and T3 does this too) driver must post this CQ notification, in addition to marking the CQ
as in error which means some QP bound to this CQ needs flushing. Then when the app wakes up and polls the CQ, the
libcxgb4 code will flush the QPs in error and thus CQEs will be inserted into the CQ. We have seen certain applications
that rely on this event to discover that a QP has moved out of RTS. IE they don't look at async QP events nor RDMACM
events.
Steve.
On Thu, Oct 13, 2011 at 1:21 AM, Kumar Sanghvikuma...@chelsio.com wrote:
At the time when peer closes connection, iw_cxgb4 will not
send a cq event if ibqp.uobject exists. In that case, its possible
for user application to get blocked in ibv_get_cq_event.
To resolve this, call the cq's comp_handler to unblock any read
from ibv_get_cq_event.
Signed-off-by: Kumar Sanghvikuma...@chelsio.com
---
drivers/infiniband/hw/cxgb4/qp.c |6 +-
1 files changed, 5 insertions(+), 1 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index ec3ce67..b59b56c 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -970,8 +970,12 @@ static void flush_qp(struct c4iw_qp *qhp)
if (qhp-ibqp.uobject) {
t4_set_wq_in_error(qhp-wq);
t4_set_cq_in_error(rchp-cq);
- if (schp != rchp)
+ (*rchp-ibcq.comp_handler)(rchp-ibcq, rchp-ibcq.cq_context);
+ if (schp != rchp) {
t4_set_cq_in_error(schp-cq);
+ (*schp-ibcq.comp_handler)(schp-ibcq,
+ schp-ibcq.cq_context);
+ }
return;
}
__flush_qp(qhp, rchp, schp);
--
1.7.1
--
To unsubscribe from this list: send the line unsubscribe linux-rdma in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
