[PATCH 0/3] hisi_sas: some CQ processing fixes

[PATCH 0/3] hisi_sas: some CQ processing fixes

[John Garry]

This patchset fixes some issues related to servicing of the
completion queue interrupt.
The major fix is that sensitive hisi_hba structures need to be
locked when free'ing a slot.
Another modification is that the v2 hw completion queue irq is
now serviced with a tasklet, as too much work was being done in
the ISR.

John Garry (3):
  scsi: hisi_sas: service v2 hw CQ ISR with tasklet
  scsi: hisi_sas: lock sensitive regions when servicing CQ interrupt
  scsi: hisi_sas: lock sensitive region in hisi_sas_slot_abort()

 drivers/scsi/hisi_sas/hisi_sas.h       |  1 +
 drivers/scsi/hisi_sas/hisi_sas_main.c  |  3 +++
 drivers/scsi/hisi_sas/hisi_sas_v1_hw.c |  2 ++
 drivers/scsi/hisi_sas/hisi_sas_v2_hw.c | 26 ++++++++++++++++++++------
 4 files changed, 26 insertions(+), 6 deletions(-)

-- 
1.9.1

[PATCH 1/3] scsi: hisi_sas: service v2 hw CQ ISR with tasklet

[John Garry]

Currently the all the slot processing for the completion
queue is done in ISR context. It is judged that the slot
processing can take a long time, especially when a SATA
NCQ completes (upto 32 slots).

So, as a solution, defer the bulk of the ISR processing
to tasklet context. Each CQ will have its down tasklet.

Signed-off-by: John Garry <john.garry@huawei.com>
Reviewed-by: Xiang Chen <chenxiang66@hisilicon.com>
---
 drivers/scsi/hisi_sas/hisi_sas.h       |  1 +
 drivers/scsi/hisi_sas/hisi_sas_v2_hw.c | 24 ++++++++++++++++++------
 2 files changed, 19 insertions(+), 6 deletions(-)

diff --git a/drivers/scsi/hisi_sas/hisi_sas.h b/drivers/scsi/hisi_sas/hisi_sas.h
index c0cd505..9216dea 100644
--- a/drivers/scsi/hisi_sas/hisi_sas.h
+++ b/drivers/scsi/hisi_sas/hisi_sas.h
@@ -95,6 +95,7 @@ struct hisi_sas_port {
 
 struct hisi_sas_cq {
 	struct hisi_hba *hisi_hba;
+	struct tasklet_struct tasklet;
 	int	rd_point;
 	int	id;
 };
diff --git a/drivers/scsi/hisi_sas/hisi_sas_v2_hw.c b/drivers/scsi/hisi_sas/hisi_sas_v2_hw.c
index b934aec..e506260 100644
--- a/drivers/scsi/hisi_sas/hisi_sas_v2_hw.c
+++ b/drivers/scsi/hisi_sas/hisi_sas_v2_hw.c
@@ -2481,20 +2481,17 @@ static irqreturn_t fatal_axi_int_v2_hw(int irq_no, void *p)
 	return IRQ_HANDLED;
 }
 
-static irqreturn_t cq_interrupt_v2_hw(int irq_no, void *p)
+static void cq_tasklet_v2_hw(unsigned long val)
 {
-	struct hisi_sas_cq *cq = p;
+	struct hisi_sas_cq *cq = (struct hisi_sas_cq *)val;
 	struct hisi_hba *hisi_hba = cq->hisi_hba;
 	struct hisi_sas_slot *slot;
 	struct hisi_sas_itct *itct;
 	struct hisi_sas_complete_v2_hdr *complete_queue;
-	u32 irq_value, rd_point = cq->rd_point, wr_point, dev_id;
+	u32 rd_point = cq->rd_point, wr_point, dev_id;
 	int queue = cq->id;
 
 	complete_queue = hisi_hba->complete_hdr[queue];
-	irq_value = hisi_sas_read32(hisi_hba, OQ_INT_SRC);
-
-	hisi_sas_write32(hisi_hba, OQ_INT_SRC, 1 << queue);
 
 	wr_point = hisi_sas_read32(hisi_hba, COMPL_Q_0_WR_PTR +
 				   (0x14 * queue));
@@ -2545,6 +2542,18 @@ static irqreturn_t cq_interrupt_v2_hw(int irq_no, void *p)
 	/* update rd_point */
 	cq->rd_point = rd_point;
 	hisi_sas_write32(hisi_hba, COMPL_Q_0_RD_PTR + (0x14 * queue), rd_point);
+}
+
+static irqreturn_t cq_interrupt_v2_hw(int irq_no, void *p)
+{
+	struct hisi_sas_cq *cq = p;
+	struct hisi_hba *hisi_hba = cq->hisi_hba;
+	int queue = cq->id;
+
+	hisi_sas_write32(hisi_hba, OQ_INT_SRC, 1 << queue);
+
+	tasklet_schedule(&cq->tasklet);
+
 	return IRQ_HANDLED;
 }
 
@@ -2726,6 +2735,8 @@ static int interrupt_init_v2_hw(struct hisi_hba *hisi_hba)
 
 	for (i = 0; i < hisi_hba->queue_count; i++) {
 		int idx = i + 96; /* First cq interrupt is irq96 */
+		struct hisi_sas_cq *cq = &hisi_hba->cq[i];
+		struct tasklet_struct *t = &cq->tasklet;
 
 		irq = irq_map[idx];
 		if (!irq) {
@@ -2742,6 +2753,7 @@ static int interrupt_init_v2_hw(struct hisi_hba *hisi_hba)
 				irq, rc);
 			return -ENOENT;
 		}
+		tasklet_init(t, cq_tasklet_v2_hw, (unsigned long)cq);
 	}
 
 	return 0;
-- 
1.9.1

[PATCH 2/3] scsi: hisi_sas: lock sensitive regions when servicing CQ interrupt

[John Garry]

There is a bug in the current driver in that certain hisi_hba
and port structure elements which we access when servicing
the CQ interrupt do not use thread-safe accesses; these include
hisi_sas_port linked-list of active slots (hisi_sas_port.entry),
bitmap of currently allocated IPTT (in hisi_hba.slot_index_tags),
and completion queue read pointer.

As a solution, lock these elements with the hisi_hba.lock.

Signed-off-by: John Garry <john.garry@huawei.com>
Reviewed-by: Xiang Chen <chenxiang66@hisilicon.com>
---
 drivers/scsi/hisi_sas/hisi_sas_v1_hw.c | 2 ++
 drivers/scsi/hisi_sas/hisi_sas_v2_hw.c | 2 ++
 2 files changed, 4 insertions(+)

diff --git a/drivers/scsi/hisi_sas/hisi_sas_v1_hw.c b/drivers/scsi/hisi_sas/hisi_sas_v1_hw.c
index 8a1be0b..854fbea 100644
--- a/drivers/scsi/hisi_sas/hisi_sas_v1_hw.c
+++ b/drivers/scsi/hisi_sas/hisi_sas_v1_hw.c
@@ -1596,6 +1596,7 @@ static irqreturn_t cq_interrupt_v1_hw(int irq, void *p)
 			hisi_hba->complete_hdr[queue];
 	u32 irq_value, rd_point = cq->rd_point, wr_point;
 
+	spin_lock(&hisi_hba->lock);
 	irq_value = hisi_sas_read32(hisi_hba, OQ_INT_SRC);
 
 	hisi_sas_write32(hisi_hba, OQ_INT_SRC, 1 << queue);
@@ -1628,6 +1629,7 @@ static irqreturn_t cq_interrupt_v1_hw(int irq, void *p)
 	/* update rd_point */
 	cq->rd_point = rd_point;
 	hisi_sas_write32(hisi_hba, COMPL_Q_0_RD_PTR + (0x14 * queue), rd_point);
+	spin_unlock(&hisi_hba->lock);
 
 	return IRQ_HANDLED;
 }
diff --git a/drivers/scsi/hisi_sas/hisi_sas_v2_hw.c b/drivers/scsi/hisi_sas/hisi_sas_v2_hw.c
index e506260..69b0f06 100644
--- a/drivers/scsi/hisi_sas/hisi_sas_v2_hw.c
+++ b/drivers/scsi/hisi_sas/hisi_sas_v2_hw.c
@@ -2493,6 +2493,7 @@ static void cq_tasklet_v2_hw(unsigned long val)
 
 	complete_queue = hisi_hba->complete_hdr[queue];
 
+	spin_lock(&hisi_hba->lock);
 	wr_point = hisi_sas_read32(hisi_hba, COMPL_Q_0_WR_PTR +
 				   (0x14 * queue));
 
@@ -2542,6 +2543,7 @@ static void cq_tasklet_v2_hw(unsigned long val)
 	/* update rd_point */
 	cq->rd_point = rd_point;
 	hisi_sas_write32(hisi_hba, COMPL_Q_0_RD_PTR + (0x14 * queue), rd_point);
+	spin_unlock(&hisi_hba->lock);
 }
 
 static irqreturn_t cq_interrupt_v2_hw(int irq_no, void *p)
-- 
1.9.1

[PATCH 3/3] scsi: hisi_sas: lock sensitive region in hisi_sas_slot_abort()

[John Garry]

When we call hisi_sas_slot_task_free() we should grab the
hisi_hba.lock, as hisi_sas_slot_task_free() accesses common
hisi_hba elements.
Function hisi_sas_slot_abort() is missing this, so add it.

Signed-off-by: John Garry <john.garry@huawei.com>
---
 drivers/scsi/hisi_sas/hisi_sas_main.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/scsi/hisi_sas/hisi_sas_main.c b/drivers/scsi/hisi_sas/hisi_sas_main.c
index d50e9cf..22dba01 100644
--- a/drivers/scsi/hisi_sas/hisi_sas_main.c
+++ b/drivers/scsi/hisi_sas/hisi_sas_main.c
@@ -146,6 +146,7 @@ static void hisi_sas_slot_abort(struct work_struct *work)
 	struct scsi_lun lun;
 	struct device *dev = &hisi_hba->pdev->dev;
 	int tag = abort_slot->idx;
+	unsigned long flags;
 
 	if (!(task->task_proto & SAS_PROTOCOL_SSP)) {
 		dev_err(dev, "cannot abort slot for non-ssp task\n");
@@ -159,7 +160,9 @@ static void hisi_sas_slot_abort(struct work_struct *work)
 	hisi_sas_debug_issue_ssp_tmf(task->dev, lun.scsi_lun, &tmf_task);
 out:
 	/* Do cleanup for this task */
+	spin_lock_irqsave(&hisi_hba->lock, flags);
 	hisi_sas_slot_task_free(hisi_hba, task, abort_slot);
+	spin_unlock_irqrestore(&hisi_hba->lock, flags);
 	if (task->task_done)
 		task->task_done(task);
 	if (sas_dev)
-- 
1.9.1

Re: [PATCH 1/3] scsi: hisi_sas: service v2 hw CQ ISR with tasklet

[zhangfei]

On 2017年01月03日 20:24, John Garry wrote:
> Currently the all the slot processing for the completion
> queue is done in ISR context. It is judged that the slot
> processing can take a long time, especially when a SATA
> NCQ completes (upto 32 slots).
>
> So, as a solution, defer the bulk of the ISR processing
> to tasklet context. Each CQ will have its down tasklet.
>
> Signed-off-by: John Garry <john.garry@huawei.com>
> Reviewed-by: Xiang Chen <chenxiang66@hisilicon.com>
Reviewed-by: Zhangfei Gao <zhangfei.gao@linaro.org>

Re: [PATCH 2/3] scsi: hisi_sas: lock sensitive regions when servicing CQ interrupt

[zhangfei]

On 2017年01月03日 20:24, John Garry wrote:
> There is a bug in the current driver in that certain hisi_hba
> and port structure elements which we access when servicing
> the CQ interrupt do not use thread-safe accesses; these include
> hisi_sas_port linked-list of active slots (hisi_sas_port.entry),
> bitmap of currently allocated IPTT (in hisi_hba.slot_index_tags),
> and completion queue read pointer.
>
> As a solution, lock these elements with the hisi_hba.lock.
>
> Signed-off-by: John Garry <john.garry@huawei.com>
> Reviewed-by: Xiang Chen <chenxiang66@hisilicon.com>
Reviewed-by: Zhangfei Gao <zhangfei.gao@linaro.org>

Re: [PATCH 3/3] scsi: hisi_sas: lock sensitive region in hisi_sas_slot_abort()

[zhangfei]

On 2017年01月03日 20:24, John Garry wrote:
> When we call hisi_sas_slot_task_free() we should grab the
> hisi_hba.lock, as hisi_sas_slot_task_free() accesses common
> hisi_hba elements.
> Function hisi_sas_slot_abort() is missing this, so add it.
>
> Signed-off-by: John Garry <john.garry@huawei.com>
Reviewed-by: Zhangfei Gao <zhangfei.gao@linaro.org>

Re: [PATCH 0/3] hisi_sas: some CQ processing fixes

[Hanjun Guo]

On 2017/1/3 20:24, John Garry wrote:
> This patchset fixes some issues related to servicing of the
> completion queue interrupt.
> The major fix is that sensitive hisi_hba structures need to be
> locked when free'ing a slot.
> Another modification is that the v2 hw completion queue irq is
> now serviced with a tasklet, as too much work was being done in
> the ISR.

Tested on v2 based sas hardware and the crashes are gone,

Tested-by: Hanjun Guo <hanjun.guo@linaro.org>

Thanks
Hanjun

Re: [PATCH 0/3] hisi_sas: some CQ processing fixes

[Martin K. Petersen]

>>>>> "John" == John Garry <john.garry@huawei.com> writes:

John> This patchset fixes some issues related to servicing of the
John> completion queue interrupt.  The major fix is that sensitive
John> hisi_hba structures need to be locked when free'ing a slot.
John> Another modification is that the v2 hw completion queue irq is now
John> serviced with a tasklet, as too much work was being done in the
John> ISR.

Applied to 4.11/scsi-queue.

-- 
Martin K. Petersen	Oracle Linux Engineering