This is a note to let you know that I've just added the patch titled
RDMA/hns: Fix an AEQE overflow error caused by untimely update of eq_db_ci
to the 6.1-stable tree which can be found at:
http://www.kernel.org/git/?p=linux/kernel/git/stable/stable-queue.git;a=summary
The filename of the patch is:
rdma-hns-fix-an-aeqe-overflow-error-caused-by-untime.patch
and it can be found in the queue-6.1 subdirectory.
If you, or anyone else, feels it should not be added to the stable tree,
please let <stable@xxxxxxxxxxxxxxx> know about it.
commit 0ad1eb3b6ebe2e0707f9891a40a0e837f70a699d
Author: wenglianfa <wenglianfa@xxxxxxxxxx>
Date: Thu Oct 24 20:39:56 2024 +0800
RDMA/hns: Fix an AEQE overflow error caused by untimely update of eq_db_ci
[ Upstream commit 571e4ab8a45e530623ab129803f090a844dd3fe9 ]
eq_db_ci is updated only after all AEQEs are processed in the AEQ
interrupt handler, which is not timely enough and may result in
AEQ overflow. Two optimization methods are proposed:
1. Set an upper limit for AEQE processing.
2. Move time-consuming operations such as printings to the bottom
half of the interrupt.
cmd events and flush_cqe events are still fully processed in the top half
to ensure timely handling.
Fixes: a5073d6054f7 ("RDMA/hns: Add eq support of hip08")
Signed-off-by: wenglianfa <wenglianfa@xxxxxxxxxx>
Signed-off-by: Junxian Huang <huangjunxian6@xxxxxxxxxxxxx>
Link: https://patch.msgid.link/20241024124000.2931869-2-huangjunxian6@xxxxxxxxxxxxx
Signed-off-by: Leon Romanovsky <leon@xxxxxxxxxx>
Signed-off-by: Sasha Levin <sashal@xxxxxxxxxx>
diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h
index a2bdfa026c560..5b8b68b2d69cf 100644
--- a/drivers/infiniband/hw/hns/hns_roce_device.h
+++ b/drivers/infiniband/hw/hns/hns_roce_device.h
@@ -1228,6 +1228,7 @@ void hns_roce_cq_completion(struct hns_roce_dev *hr_dev, u32 cqn);
void hns_roce_cq_event(struct hns_roce_dev *hr_dev, u32 cqn, int event_type);
void flush_cqe(struct hns_roce_dev *dev, struct hns_roce_qp *qp);
void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type);
+void hns_roce_flush_cqe(struct hns_roce_dev *hr_dev, u32 qpn);
void hns_roce_srq_event(struct hns_roce_dev *hr_dev, u32 srqn, int event_type);
u8 hns_get_gid_index(struct hns_roce_dev *hr_dev, u32 port, int gid_index);
void hns_roce_handle_device_err(struct hns_roce_dev *hr_dev);
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
index 56c0e87c494ec..8b371d6ecbac9 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
@@ -5977,11 +5977,10 @@ static int hns_roce_v2_query_mpt(struct hns_roce_dev *hr_dev, u32 key,
return ret;
}
-static void hns_roce_irq_work_handle(struct work_struct *work)
+static void dump_aeqe_log(struct hns_roce_work *irq_work)
{
- struct hns_roce_work *irq_work =
- container_of(work, struct hns_roce_work, work);
- struct ib_device *ibdev = &irq_work->hr_dev->ib_dev;
+ struct hns_roce_dev *hr_dev = irq_work->hr_dev;
+ struct ib_device *ibdev = &hr_dev->ib_dev;
switch (irq_work->event_type) {
case HNS_ROCE_EVENT_TYPE_PATH_MIG:
@@ -6025,6 +6024,8 @@ static void hns_roce_irq_work_handle(struct work_struct *work)
case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
ibdev_warn(ibdev, "DB overflow.\n");
break;
+ case HNS_ROCE_EVENT_TYPE_MB:
+ break;
case HNS_ROCE_EVENT_TYPE_FLR:
ibdev_warn(ibdev, "function level reset.\n");
break;
@@ -6035,8 +6036,46 @@ static void hns_roce_irq_work_handle(struct work_struct *work)
ibdev_err(ibdev, "invalid xrceth error.\n");
break;
default:
+ ibdev_info(ibdev, "Undefined event %d.\n",
+ irq_work->event_type);
break;
}
+}
+
+static void hns_roce_irq_work_handle(struct work_struct *work)
+{
+ struct hns_roce_work *irq_work =
+ container_of(work, struct hns_roce_work, work);
+ struct hns_roce_dev *hr_dev = irq_work->hr_dev;
+ int event_type = irq_work->event_type;
+ u32 queue_num = irq_work->queue_num;
+
+ switch (event_type) {
+ case HNS_ROCE_EVENT_TYPE_PATH_MIG:
+ case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
+ case HNS_ROCE_EVENT_TYPE_COMM_EST:
+ case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
+ case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
+ case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
+ case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
+ case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
+ case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
+ case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
+ hns_roce_qp_event(hr_dev, queue_num, event_type);
+ break;
+ case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
+ case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
+ hns_roce_srq_event(hr_dev, queue_num, event_type);
+ break;
+ case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
+ case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
+ hns_roce_cq_event(hr_dev, queue_num, event_type);
+ break;
+ default:
+ break;
+ }
+
+ dump_aeqe_log(irq_work);
kfree(irq_work);
}
@@ -6097,14 +6136,14 @@ static struct hns_roce_aeqe *next_aeqe_sw_v2(struct hns_roce_eq *eq)
static irqreturn_t hns_roce_v2_aeq_int(struct hns_roce_dev *hr_dev,
struct hns_roce_eq *eq)
{
- struct device *dev = hr_dev->dev;
struct hns_roce_aeqe *aeqe = next_aeqe_sw_v2(eq);
irqreturn_t aeqe_found = IRQ_NONE;
+ int num_aeqes = 0;
int event_type;
u32 queue_num;
int sub_type;
- while (aeqe) {
+ while (aeqe && num_aeqes < HNS_AEQ_POLLING_BUDGET) {
/* Make sure we read AEQ entry after we have checked the
* ownership bit
*/
@@ -6115,25 +6154,12 @@ static irqreturn_t hns_roce_v2_aeq_int(struct hns_roce_dev *hr_dev,
queue_num = hr_reg_read(aeqe, AEQE_EVENT_QUEUE_NUM);
switch (event_type) {
- case HNS_ROCE_EVENT_TYPE_PATH_MIG:
- case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
- case HNS_ROCE_EVENT_TYPE_COMM_EST:
- case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
- case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
- hns_roce_qp_event(hr_dev, queue_num, event_type);
- break;
- case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
- case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
- hns_roce_srq_event(hr_dev, queue_num, event_type);
- break;
- case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
- case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
- hns_roce_cq_event(hr_dev, queue_num, event_type);
+ hns_roce_flush_cqe(hr_dev, queue_num);
break;
case HNS_ROCE_EVENT_TYPE_MB:
hns_roce_cmd_event(hr_dev,
@@ -6141,12 +6167,7 @@ static irqreturn_t hns_roce_v2_aeq_int(struct hns_roce_dev *hr_dev,
aeqe->event.cmd.status,
le64_to_cpu(aeqe->event.cmd.out_param));
break;
- case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
- case HNS_ROCE_EVENT_TYPE_FLR:
- break;
default:
- dev_err(dev, "unhandled event %d on EQ %d at idx %u.\n",
- event_type, eq->eqn, eq->cons_index);
break;
}
@@ -6158,6 +6179,7 @@ static irqreturn_t hns_roce_v2_aeq_int(struct hns_roce_dev *hr_dev,
hns_roce_v2_init_irq_work(hr_dev, eq, queue_num);
aeqe = next_aeqe_sw_v2(eq);
+ ++num_aeqes;
}
update_eq_db(eq);
@@ -6687,6 +6709,9 @@ static int hns_roce_v2_init_eq_table(struct hns_roce_dev *hr_dev)
int ret;
int i;
+ if (hr_dev->caps.aeqe_depth < HNS_AEQ_POLLING_BUDGET)
+ return -EINVAL;
+
other_num = hr_dev->caps.num_other_vectors;
comp_num = hr_dev->caps.num_comp_vectors;
aeq_num = hr_dev->caps.num_aeq_vectors;
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.h b/drivers/infiniband/hw/hns/hns_roce_hw_v2.h
index 2b4dbbb06eb56..e24ddbcafe0fc 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.h
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.h
@@ -114,6 +114,11 @@
#define HNS_ROCE_V2_TABLE_CHUNK_SIZE (1 << 18)
+/* budget must be smaller than aeqe_depth to guarantee that we update
+ * the ci before we polled all the entries in the EQ.
+ */
+#define HNS_AEQ_POLLING_BUDGET 64
+
enum {
HNS_ROCE_CMD_FLAG_IN = BIT(0),
HNS_ROCE_CMD_FLAG_OUT = BIT(1),
diff --git a/drivers/infiniband/hw/hns/hns_roce_qp.c b/drivers/infiniband/hw/hns/hns_roce_qp.c
index e158ed9e6c83c..7af6631761044 100644
--- a/drivers/infiniband/hw/hns/hns_roce_qp.c
+++ b/drivers/infiniband/hw/hns/hns_roce_qp.c
@@ -39,6 +39,25 @@
#include "hns_roce_device.h"
#include "hns_roce_hem.h"
+static struct hns_roce_qp *hns_roce_qp_lookup(struct hns_roce_dev *hr_dev,
+ u32 qpn)
+{
+ struct device *dev = hr_dev->dev;
+ struct hns_roce_qp *qp;
+ unsigned long flags;
+
+ xa_lock_irqsave(&hr_dev->qp_table_xa, flags);
+ qp = __hns_roce_qp_lookup(hr_dev, qpn);
+ if (qp)
+ refcount_inc(&qp->refcount);
+ xa_unlock_irqrestore(&hr_dev->qp_table_xa, flags);
+
+ if (!qp)
+ dev_warn(dev, "async event for bogus QP %08x\n", qpn);
+
+ return qp;
+}
+
static void flush_work_handle(struct work_struct *work)
{
struct hns_roce_work *flush_work = container_of(work,
@@ -95,31 +114,28 @@ void flush_cqe(struct hns_roce_dev *dev, struct hns_roce_qp *qp)
void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
{
- struct device *dev = hr_dev->dev;
struct hns_roce_qp *qp;
- xa_lock(&hr_dev->qp_table_xa);
- qp = __hns_roce_qp_lookup(hr_dev, qpn);
- if (qp)
- refcount_inc(&qp->refcount);
- xa_unlock(&hr_dev->qp_table_xa);
-
- if (!qp) {
- dev_warn(dev, "async event for bogus QP %08x\n", qpn);
+ qp = hns_roce_qp_lookup(hr_dev, qpn);
+ if (!qp)
return;
- }
- if (event_type == HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR ||
- event_type == HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR ||
- event_type == HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR ||
- event_type == HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION ||
- event_type == HNS_ROCE_EVENT_TYPE_INVALID_XRCETH) {
- qp->state = IB_QPS_ERR;
+ qp->event(qp, (enum hns_roce_event)event_type);
- flush_cqe(hr_dev, qp);
- }
+ if (refcount_dec_and_test(&qp->refcount))
+ complete(&qp->free);
+}
- qp->event(qp, (enum hns_roce_event)event_type);
+void hns_roce_flush_cqe(struct hns_roce_dev *hr_dev, u32 qpn)
+{
+ struct hns_roce_qp *qp;
+
+ qp = hns_roce_qp_lookup(hr_dev, qpn);
+ if (!qp)
+ return;
+
+ qp->state = IB_QPS_ERR;
+ flush_cqe(hr_dev, qp);
if (refcount_dec_and_test(&qp->refcount))
complete(&qp->free);
