Skip to contentSkip to conversation

spi: spi-qcom-qspi: Avoid some per-transfer overhead

PATCH Superseded Douglas Anderson 4 participants 7 messages Jul 7, 2020 – Jul 9, 2020

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover spi: spi-qcom-qspi: Avoid some per-transfer overhead

Not to be confused with the similar series I posed for the _other_ Qualcomm SPI controller (spi-geni-qcom) [1], this one avoids the overhead on the Quad SPI controller.

It's based atop the current Qualcomm tree including Rajendra's ("spi:
spi-qcom-qspi: Use OPP API to set clk/perf state"). As discussed in individual patches, these could ideally land through the Qualcomm tree with Mark's Ack.

Measuring:
* Before OPP / Interconnect patches reading all flash takes: ~3.4 seconds
* After OPP / Interconnect patches reading all flash takes: ~4.7 seconds
* After this patch reading all flash takes: ~3.3 seconds

[1] https://lore.kernel.org/r/[email protected]
[2] https://lore.kernel.org/r/[email protected]

Douglas Anderson (2):
spi: spi-qcom-qspi: Avoid clock setting if not needed
spi: spi-qcom-qspi: Set an autosuspend delay of 250 ms

 drivers/spi/spi-qcom-qspi.c | 45 ++++++++++++++++++++++++++++---------
 1 file changed, 35 insertions(+), 10 deletions(-)
Signature · 1 line
2.27.0.383.g050319c2ae-goog
#1 lore raw

PATCH 1/2 spi: spi-qcom-qspi: Avoid clock setting if not needed

As per recent changes to the spi-qcom-qspi, now when we set the clock we'll call into the interconnect framework and also call the OPP API. Those are expensive operations. Let's avoid calling them if possible. This has a big impact on getting transfer rates back up to where they were (or maybe slightly better) before those patches landed.

Fixes cff80645d6d3 ("spi: spi-qcom-qspi: Add interconnect support")

---
This applies atop the Qualcomm tree after Rajendra's ("spi:
spi-qcom-qspi: Use OPP API to set clk/perf state") patch and I'd hope it could land there with Mark's Ack just like the patch it Fixes did.

 drivers/spi/spi-qcom-qspi.c | 43 ++++++++++++++++++++++++++++---------
 1 file changed, 33 insertions(+), 10 deletions(-)
drivers/spi/spi-qcom-qspi.c +33 -10
index 18a59aa23ef8..322b88c22a86 100644
@@ -144,6 +144,7 @@ struct qcom_qspi {
 	struct icc_path *icc_path_cpu_to_qspi;
 	struct opp_table *opp_table;
 	bool has_opp_table;
+	unsigned long last_speed;
 	/* Lock to protect data accessed by IRQs */
 	spinlock_t lock;
 };
@@ -226,19 +227,13 @@ static void qcom_qspi_handle_err(struct spi_master *master,
 	spin_unlock_irqrestore(&ctrl->lock, flags);
 }
 
-static int qcom_qspi_transfer_one(struct spi_master *master,
-				  struct spi_device *slv,
-				  struct spi_transfer *xfer)
+static int qcom_qspi_set_speed(struct qcom_qspi *ctrl, unsigned long speed_hz)
 {
-	struct qcom_qspi *ctrl = spi_master_get_devdata(master);
 	int ret;
-	unsigned long speed_hz;
-	unsigned long flags;
 	unsigned int avg_bw_cpu;
 
-	speed_hz = slv->max_speed_hz;
-	if (xfer->speed_hz)
-		speed_hz = xfer->speed_hz;
+	if (speed_hz == ctrl->last_speed)
+		return 0;
 
 	/* In regular operation (SBL_EN=1) core must be 4x transfer clock */
 	ret = dev_pm_opp_set_rate(ctrl->dev, speed_hz * 4);
@@ -259,6 +254,28 @@ static int qcom_qspi_transfer_one(struct spi_master *master,
 		return ret;
 	}
 
+	ctrl->last_speed = speed_hz;
+
+	return 0;
+}
+
+static int qcom_qspi_transfer_one(struct spi_master *master,
+				  struct spi_device *slv,
+				  struct spi_transfer *xfer)
+{
+	struct qcom_qspi *ctrl = spi_master_get_devdata(master);
+	int ret;
+	unsigned long speed_hz;
+	unsigned long flags;
+
+	speed_hz = slv->max_speed_hz;
+	if (xfer->speed_hz)
+		speed_hz = xfer->speed_hz;
+
+	ret = qcom_qspi_set_speed(ctrl, speed_hz);
+	if (ret)
+		return ret;
+
 	spin_lock_irqsave(&ctrl->lock, flags);
 
 	/* We are half duplex, so either rx or tx will be set */
@@ -602,7 +619,13 @@ static int __maybe_unused qcom_qspi_runtime_resume(struct device *dev)
 		return ret;
 	}
 
-	return clk_bulk_prepare_enable(QSPI_NUM_CLKS, ctrl->clks);
+	ret = clk_bulk_prepare_enable(QSPI_NUM_CLKS, ctrl->clks);
+	if (ret)
+		return ret;
+
+	dev_pm_opp_set_rate(dev, ctrl->last_speed * 4);
+
+	return 0;
 }
 
 static int __maybe_unused qcom_qspi_suspend(struct device *dev)
Signature · 1 line
2.27.0.383.g050319c2ae-goog
maintainer Acked-by on v1 1/2
#2 lore raw

Re: [PATCH 1/2] spi: spi-qcom-qspi: Avoid clock setting if not needed

Quoted 2 lines · Douglas Anderson

As per recent changes to the spi-qcom-qspi, now when we set the clock we'll call into the interconnect framework and also call the OPP API.

#3 lore raw

PATCH 2/2 spi: spi-qcom-qspi: Set an autosuspend delay of 250 ms

In reply to Douglas Anderson

In commit cff80645d6d3 ("spi: spi-qcom-qspi: Add interconnect support") the spi_geni_runtime_suspend() and spi_geni_runtime_resume() became a bit slower. Measuring on my hardware I see numbers in the hundreds of microseconds now.

Let's use autosuspend to help avoid some of the overhead. Now if we're doing a bunch of transfers we won't need to be constantly chruning.

The number 250 ms for the autosuspend delay was picked a bit arbitrarily, so if someone has measurements showing a better value we could easily change this.

Fixes cff80645d6d3 ("spi: spi-qcom-qspi: Add interconnect support")

---
This patch could go through the SPI tree or land in the Qualcomm tree. The patch it Fixes is currently in the Qualcomm tree so if it lands in the main SPI tree there'd be a bit of a perf regression in the Qualcomm tree until things merge together in mainline.

 drivers/spi/spi-qcom-qspi.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/spi/spi-qcom-qspi.c +2 -0
index 322b88c22a86..6c39b23222b8 100644
@@ -553,6 +553,8 @@ static int qcom_qspi_probe(struct platform_device *pdev)
 		goto exit_probe_master_put;
 	}
 
+	pm_runtime_use_autosuspend(dev);
+	pm_runtime_set_autosuspend_delay(dev, 250);
 	pm_runtime_enable(dev);
 
 	ret = spi_register_master(master);
Signature · 1 line
2.27.0.383.g050319c2ae-goog
maintainer Acked-by on v1 2/2
#4 lore raw

Re: [PATCH 2/2] spi: spi-qcom-qspi: Set an autosuspend delay of 250 ms

Quoted 4 lines · Douglas Anderson

In commit cff80645d6d3 ("spi: spi-qcom-qspi: Add interconnect support") the spi_geni_runtime_suspend() and spi_geni_runtime_resume() became a bit slower. Measuring on my hardware I see numbers in the hundreds of microseconds now.

Reviewed-by Tested-by on the v1 series
#5 lore raw
In reply to Douglas Anderson
Quoted 10 lines · Douglas Anderson

Not to be confused with the similar series I posed for the _other_ Qualcomm SPI controller (spi-geni-qcom) [1], this one avoids the overhead on the Quad SPI controller.

It's based atop the current Qualcomm tree including Rajendra's ("spi:
spi-qcom-qspi: Use OPP API to set clk/perf state"). As discussed in individual patches, these could ideally land through the Qualcomm tree with Mark's Ack.

Quoted 4 lines

Measuring:
* Before OPP / Interconnect patches reading all flash takes: ~3.4 seconds
* After OPP / Interconnect patches reading all flash takes: ~4.7 seconds
* After this patch reading all flash takes: ~3.3 seconds

Thanks Doug, I saw similar benefit on my setup with these patches. They do help reduce the (unnecessary) additional overhead so it makes sense to merge these along with the OPP/Interconnect patches in-order to avoid the regression.

Quoted 12 lines

[1] https://lore.kernel.org/r/[email protected]
[2] https://lore.kernel.org/r/[email protected]

Douglas Anderson (2):
spi: spi-qcom-qspi: Avoid clock setting if not needed
spi: spi-qcom-qspi: Set an autosuspend delay of 250 ms

  drivers/spi/spi-qcom-qspi.c | 45 ++++++++++++++++++++++++++++---------
  1 file changed, 35 insertions(+), 10 deletions(-)
Signature · 2 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation
Reviewed-by on the v1 series
#6 lore raw
In reply to Douglas Anderson
Quoted 18 lines · Douglas Anderson

Not to be confused with the similar series I posed for the _other_ Qualcomm SPI controller (spi-geni-qcom) [1], this one avoids the overhead on the Quad SPI controller.

It's based atop the current Qualcomm tree including Rajendra's ("spi:
spi-qcom-qspi: Use OPP API to set clk/perf state"). As discussed in individual patches, these could ideally land through the Qualcomm tree with Mark's Ack.

Measuring:
* Before OPP / Interconnect patches reading all flash takes: ~3.4 seconds
* After OPP / Interconnect patches reading all flash takes: ~4.7 seconds
* After this patch reading all flash takes: ~3.3 seconds

[1] https://lore.kernel.org/r/[email protected]
[2] https://lore.kernel.org/r/[email protected]

Quoted 6 lines

Douglas Anderson (2):
spi: spi-qcom-qspi: Avoid clock setting if not needed
spi: spi-qcom-qspi: Set an autosuspend delay of 250 ms

  drivers/spi/spi-qcom-qspi.c | 45 ++++++++++++++++++++++++++++---------
  1 file changed, 35 insertions(+), 10 deletions(-)