Skip to contentSkip to conversation

sunxi a20 mmc timeout fix

PATCH Superseded Michal Suchanek 5 participants 9 messages May 24, 2015 – May 28, 2015

Subsystems (2)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#5 lore raw

PATCH 2/2 mmc: sunxi: Also set SDXC_LOW_POWER_ON

In reply to Michal Suchanek

The function sunxi_mmc_oclk_onoff filters out the SDXC_LOW_POWER_ON flag but never sets it.

Set SDXC_LOW_POWER_ON when oclk is disabled.

---
 drivers/mmc/host/sunxi-mmc.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/mmc/host/sunxi-mmc.c +2 -0
index 7cdeecd..e957888 100644
@@ -603,6 +603,8 @@ static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
 
 	if (oclk_en)
 		rval |= SDXC_CARD_CLOCK_ON;
+	else
+		rval |= SDXC_LOW_POWER_ON;
 
 	start = jiffies;
 	end = start + msecs_to_jiffies(750);
Signature · 1 line
2.1.4
#1 lore raw

PATCH 1/2 mmc: sunxi: fix timeout in sunxi_mmc_oclk_onoff

In reply to Michal Suchanek

The 250ms timeout is too short.

On my system enabling the oclk takes under 50ms and disabling slightly over 100ms when idle. Under load disabling the clock can take over 350ms.

This does not make mmc clock gating look like good option to have on sunxi but the system should not crash with mmc clock gating enabled nonetheless.

This patch sets the timeout to 750ms and adds debug prints which show how long enabling/disabling the clock took so more data can be collected from other systems.

---
 drivers/mmc/host/sunxi-mmc.c | 22 +++++++++++++++++++---
 1 file changed, 19 insertions(+), 3 deletions(-)
drivers/mmc/host/sunxi-mmc.c +19 -3
index 4d3e1ff..7cdeecd 100644
@@ -595,7 +595,7 @@ static irqreturn_t sunxi_mmc_handle_manual_stop(int irq, void *dev_id)
 
 static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
 {
-	unsigned long expire = jiffies + msecs_to_jiffies(250);
+	unsigned long start, end;
 	u32 rval;
 
 	rval = mmc_readl(host, REG_CLKCR);
@@ -604,6 +604,8 @@ static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
 	if (oclk_en)
 		rval |= SDXC_CARD_CLOCK_ON;
 
+	start = jiffies;
+	end = start + msecs_to_jiffies(750);
 	mmc_writel(host, REG_CLKCR, rval);
 
 	rval = SDXC_START | SDXC_UPCLK_ONLY | SDXC_WAIT_PRE_OVER;
@@ -611,15 +613,29 @@ static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
 
 	do {
 		rval = mmc_readl(host, REG_CMDR);
-	} while (time_before(jiffies, expire) && (rval & SDXC_START));
+	} while (time_before(jiffies, end) && (rval & SDXC_START));
+	end = jiffies;
 
 	/* clear irq status bits set by the command */
 	mmc_writel(host, REG_RINTR,
 		   mmc_readl(host, REG_RINTR) & ~SDXC_SDIO_INTERRUPT);
 
 	if (rval & SDXC_START) {
-		dev_err(mmc_dev(host->mmc), "fatal err update clk timeout\n");
+		dev_err(mmc_dev(host->mmc),
+			"fatal err update oclk timeout. Could not %s in %ims.\n",
+			oclk_en ? "enable" : "disable",
+			jiffies_to_msecs(end-start));
 		return -EIO;
+	} else {
+		int msecs = jiffies_to_msecs(end-start);
+		const char *ing = oclk_en ? "enabling" : "disabling";
+
+		if ((msecs > 150) || (oclk_en && (msecs > 50)))
+			dev_warn(mmc_dev(host->mmc),
+				 "%s oclk took %ims", ing, msecs);
+		else
+			dev_dbg(mmc_dev(host->mmc),
+				 "%s oclk took %ims", ing, msecs);
 	}
 
 	return 0;
Signature · 1 line
2.1.4
#0 lore raw

cover sunxi a20 mmc timeout fix

Hello,

I had serious problems with root filesustem on SD card on Cubieboard2 when clock gating was on.

It turns out that sunxi_mmc_oclk_onoff has way too short timeout ofr setting the clock.

So here are some patches for

- bumping the timout
- spamming dmesg when the time to set the clock seems too long
- setting SDXC_LOW_POWER_ON which is unset but never set

Also feel free to just bump the timeout to reasonably high value so the board can boot reliably with mmc clock gating enabled.

Thanks

Michal

Michal Suchanek (2):
mmc: sunxi: fix timeout in sunxi_mmc_oclk_onoff
mmc: sunxi: Also set SDXC_LOW_POWER_ON

 drivers/mmc/host/sunxi-mmc.c | 24 +++++++++++++++++++++---
 1 file changed, 21 insertions(+), 3 deletions(-)
Signature · 1 line
2.1.4
#2 lore raw

Re: [linux-sunxi] [PATCH 1/2] mmc: sunxi: fix timeout in sunxi_mmc_oclk_onoff

In reply to Michal Suchanek

Hi Michal,

Quoted 33 lines · Michal Suchanek

The 250ms timeout is too short.

On my system enabling the oclk takes under 50ms and disabling slightly over 100ms when idle. Under load disabling the clock can take over 350ms.

This does not make mmc clock gating look like good option to have on sunxi but the system should not crash with mmc clock gating enabled nonetheless.

This patch sets the timeout to 750ms and adds debug prints which show how long enabling/disabling the clock took so more data can be collected from other systems.

---
 drivers/mmc/host/sunxi-mmc.c | 22 +++++++++++++++++++---
 1 file changed, 19 insertions(+), 3 deletions(-)
drivers/mmc/host/sunxi-mmc.c +1 -1
index 4d3e1ff..7cdeecd 100644
@@ -595,7 +595,7 @@ static irqreturn_t sunxi_mmc_handle_manual_stop(int irq, void *dev_id)
 
 static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
 {
-       unsigned long expire = jiffies + msecs_to_jiffies(250);
+       unsigned long start, end;
        u32 rval;
 
        rval = mmc_readl(host, REG_CLKCR);

drivers/mmc/host/sunxi-mmc.c static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en) · line 608

Quoted 5 lines
drivers/mmc/host/sunxi-mmc.c +2 -0
@@ -604,6 +604,8 @@ static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
        if (oclk_en)
                rval |= SDXC_CARD_CLOCK_ON;
 
+       start = jiffies;
+       end = start + msecs_to_jiffies(750);

You might want to make the timeout value a #define so it's not buried in the code like this.

Thanks,

Signature · 4 lines
Julian Calaby

Email: julian.calaby at gmail.com
Profile: http://www.google.com/profiles/julian.calaby/
#3 lore raw

Re: [PATCH 1/2] mmc: sunxi: fix timeout in sunxi_mmc_oclk_onoff

In reply to Michal Suchanek

Hi,

Quoted 10 lines · Michal Suchanek

The 250ms timeout is too short.

On my system enabling the oclk takes under 50ms and disabling slightly over 100ms when idle. Under load disabling the clock can take over 350ms.

This does not make mmc clock gating look like good option to have on sunxi but the system should not crash with mmc clock gating enabled nonetheless.

Quoted 5 lines

This patch sets the timeout to 750ms and adds debug prints which show how long enabling/disabling the clock took so more data can be collected from other systems.

Looks good: Reviewed-by: Hans de Goede <[email protected]>

Regards,

Hans

Quoted 59 lines
---
  drivers/mmc/host/sunxi-mmc.c | 22 +++++++++++++++++++---
  1 file changed, 19 insertions(+), 3 deletions(-)
drivers/mmc/host/sunxi-mmc.c +19 -3
index 4d3e1ff..7cdeecd 100644
@@ -595,7 +595,7 @@ static irqreturn_t sunxi_mmc_handle_manual_stop(int irq, void *dev_id)
 
  static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
  {
-	unsigned long expire = jiffies + msecs_to_jiffies(250);
+	unsigned long start, end;
  	u32 rval;
 
  	rval = mmc_readl(host, REG_CLKCR);
@@ -604,6 +604,8 @@ static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
  	if (oclk_en)
  		rval |= SDXC_CARD_CLOCK_ON;
 
+	start = jiffies;
+	end = start + msecs_to_jiffies(750);
  	mmc_writel(host, REG_CLKCR, rval);
 
  	rval = SDXC_START | SDXC_UPCLK_ONLY | SDXC_WAIT_PRE_OVER;
@@ -611,15 +613,29 @@ static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
 
  	do {
  		rval = mmc_readl(host, REG_CMDR);
-	} while (time_before(jiffies, expire) && (rval & SDXC_START));
+	} while (time_before(jiffies, end) && (rval & SDXC_START));
+	end = jiffies;
 
  	/* clear irq status bits set by the command */
  	mmc_writel(host, REG_RINTR,
  		   mmc_readl(host, REG_RINTR) & ~SDXC_SDIO_INTERRUPT);
 
  	if (rval & SDXC_START) {
-		dev_err(mmc_dev(host->mmc), "fatal err update clk timeout\n");
+		dev_err(mmc_dev(host->mmc),
+			"fatal err update oclk timeout. Could not %s in %ims.\n",
+			oclk_en ? "enable" : "disable",
+			jiffies_to_msecs(end-start));
  		return -EIO;
+	} else {
+		int msecs = jiffies_to_msecs(end-start);
+		const char *ing = oclk_en ? "enabling" : "disabling";
+
+		if ((msecs > 150) || (oclk_en && (msecs > 50)))
+			dev_warn(mmc_dev(host->mmc),
+				 "%s oclk took %ims", ing, msecs);
+		else
+			dev_dbg(mmc_dev(host->mmc),
+				 "%s oclk took %ims", ing, msecs);
  	}
 
  	return 0;
#6 lore raw

Re: [PATCH 2/2] mmc: sunxi: Also set SDXC_LOW_POWER_ON

In reply to Michal Suchanek

Hi,

Quoted 4 lines · Michal Suchanek

The function sunxi_mmc_oclk_onoff filters out the SDXC_LOW_POWER_ON flag but never sets it.

Set SDXC_LOW_POWER_ON when oclk is disabled.

Nack, looking at the datasheet I do not thing this patch actually does anything, according to the datasheet setting this bit to 1 results
in: "Turn off card clock when FSM in IDLE state", iow this does mmc clock gating on idle automatically in hardware, since we completely disable the clock on clock-off by clearing SDXC_CARD_CLOCK_ON setting this bit on clock-off is a nop.

We could consider actually setting this to safe power when setting the clock on, for doing that it would be good to look at the android code and see if it ever sets this bit and if so when.

WRT CONFIG_MMC_CLKGATE we are probably better off using SDXC_LOW_POWER_ON then that when it is enabled. So maybe we should set SDXC_LOW_POWER_ON based on #ifdef WRT CONFIG_MMC_CLKGATE, and have some way to tell the mmc core to not do clock gating on this host ?

Regards,

Hans

Quoted 20 lines
---
  drivers/mmc/host/sunxi-mmc.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/mmc/host/sunxi-mmc.c +2 -0
index 7cdeecd..e957888 100644
@@ -603,6 +603,8 @@ static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
 
  	if (oclk_en)
  		rval |= SDXC_CARD_CLOCK_ON;
+	else
+		rval |= SDXC_LOW_POWER_ON;
 
  	start = jiffies;
  	end = start + msecs_to_jiffies(750);
#4 lore raw

Re: [PATCH 1/2] mmc: sunxi: fix timeout in sunxi_mmc_oclk_onoff

In reply to Michal Suchanek

Hi Michal,

The patch itself looks good, but there's a few style issues.

Quoted 49 lines · Michal Suchanek

The 250ms timeout is too short.

On my system enabling the oclk takes under 50ms and disabling slightly over 100ms when idle. Under load disabling the clock can take over 350ms.

This does not make mmc clock gating look like good option to have on sunxi but the system should not crash with mmc clock gating enabled nonetheless.

This patch sets the timeout to 750ms and adds debug prints which show how long enabling/disabling the clock took so more data can be collected from other systems.

---
 drivers/mmc/host/sunxi-mmc.c | 22 +++++++++++++++++++---
 1 file changed, 19 insertions(+), 3 deletions(-)
drivers/mmc/host/sunxi-mmc.c +5 -2
index 4d3e1ff..7cdeecd 100644
@@ -595,7 +595,7 @@ static irqreturn_t sunxi_mmc_handle_manual_stop(int irq, void *dev_id)
 
 static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
 {
-	unsigned long expire = jiffies + msecs_to_jiffies(250);
+	unsigned long start, end;
 	u32 rval;
 
 	rval = mmc_readl(host, REG_CLKCR);
@@ -604,6 +604,8 @@ static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
 	if (oclk_en)
 		rval |= SDXC_CARD_CLOCK_ON;
 
+	start = jiffies;
+	end = start + msecs_to_jiffies(750);
 	mmc_writel(host, REG_CLKCR, rval);
 
 	rval = SDXC_START | SDXC_UPCLK_ONLY | SDXC_WAIT_PRE_OVER;
@@ -611,15 +613,29 @@ static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
 
 	do {
 		rval = mmc_readl(host, REG_CMDR);
-	} while (time_before(jiffies, expire) && (rval & SDXC_START));
+	} while (time_before(jiffies, end) && (rval & SDXC_START));
+	end = jiffies;
 

drivers/mmc/host/sunxi-mmc.c static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en) · line 627

Quoted 10 lines
drivers/mmc/host/sunxi-mmc.c +4 -1
 	/* clear irq status bits set by the command */
 	mmc_writel(host, REG_RINTR,
 		   mmc_readl(host, REG_RINTR) & ~SDXC_SDIO_INTERRUPT);
 
 	if (rval & SDXC_START) {
-		dev_err(mmc_dev(host->mmc), "fatal err update clk timeout\n");
+		dev_err(mmc_dev(host->mmc),
+			"fatal err update oclk timeout. Could not %s in %ims.\n",
+			oclk_en ? "enable" : "disable",
+			jiffies_to_msecs(end-start));

You should have spaces around that operator.

drivers/mmc/host/sunxi-mmc.c static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en) · line 630

Quoted 3 lines
drivers/mmc/host/sunxi-mmc.c +2 -0
 		return -EIO;
+	} else {
+		int msecs = jiffies_to_msecs(end-start);

Ditto.

drivers/mmc/host/sunxi-mmc.c static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en) · line 638

Quoted 8 lines
drivers/mmc/host/sunxi-mmc.c +8 -0
+		const char *ing = oclk_en ? "enabling" : "disabling";
+
+		if ((msecs > 150) || (oclk_en && (msecs > 50)))
+			dev_warn(mmc_dev(host->mmc),
+				 "%s oclk took %ims", ing, msecs);
+		else
+			dev_dbg(mmc_dev(host->mmc),
+				 "%s oclk took %ims", ing, msecs);

And this should be aligned with the opening parenthesis.

Thanks!
Maxime

Signature · 10 lines
Maxime Ripard, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: Digital signature
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20150525/989fadfb/attachment.sig>
#7 lore raw

Re: [PATCH 2/2] mmc: sunxi: Also set SDXC_LOW_POWER_ON

In reply to Hans de Goede
Quoted 17 lines · Hans de Goede

Hi,

Quoted 5 lines · Michal Suchanek

The function sunxi_mmc_oclk_onoff filters out the SDXC_LOW_POWER_ON flag but never sets it.

Set SDXC_LOW_POWER_ON when oclk is disabled.

Nack, looking at the datasheet I do not thing this patch actually does anything, according to the datasheet setting this bit to 1 results
in: "Turn off card clock when FSM in IDLE state", iow this does mmc clock gating on idle automatically in hardware, since we completely disable the clock on clock-off by clearing SDXC_CARD_CLOCK_ON setting this bit on clock-off is a nop.

Quoted 8 lines

We could consider actually setting this to safe power when setting the clock on, for doing that it would be good to look at the android code and see if it ever sets this bit and if so when.

WRT CONFIG_MMC_CLKGATE we are probably better off using SDXC_LOW_POWER_ON then that when it is enabled. So maybe we should set SDXC_LOW_POWER_ON based on #ifdef WRT CONFIG_MMC_CLKGATE, and have some way to tell the mmc core to not do clock gating on this host ?

I am planning to remove the entire thing for CONFIG_MMC_CLKGATE, maybe it's time to do that now!?

Most hosts use runtime PM to deal with clock gating at idle and browsing def configs tells me that CONFIG_MMC_CLKGATE isn't much used.

Kind regards
Uffe

#8 lore raw

Re: [PATCH 2/2] mmc: sunxi: Also set SDXC_LOW_POWER_ON

Hi,

Quoted 22 lines · Ulf Hansson
Quoted 21 lines · Hans de Goede

Hi,

Quoted 5 lines · Michal Suchanek

The function sunxi_mmc_oclk_onoff filters out the SDXC_LOW_POWER_ON flag but never sets it.

Set SDXC_LOW_POWER_ON when oclk is disabled.

Nack, looking at the datasheet I do not thing this patch actually does anything, according to the datasheet setting this bit to 1 results
in: "Turn off card clock when FSM in IDLE state", iow this does mmc clock gating on idle automatically in hardware, since we completely disable the clock on clock-off by clearing SDXC_CARD_CLOCK_ON setting this bit on clock-off is a nop.

We could consider actually setting this to safe power when setting the clock on, for doing that it would be good to look at the android code and see if it ever sets this bit and if so when.

Quoted 7 lines
Quoted 4 lines

WRT CONFIG_MMC_CLKGATE we are probably better off using SDXC_LOW_POWER_ON then that when it is enabled. So maybe we should set SDXC_LOW_POWER_ON based on #ifdef WRT CONFIG_MMC_CLKGATE, and have some way to tell the mmc core to not do clock gating on this host ?

I am planning to remove the entire thing for CONFIG_MMC_CLKGATE, maybe it's time to do that now!?

Removing it now sounds like a good plan to me, we would still need to look into maybe enabling SDXC_LOW_POWER_ON on sunxi in some cases to save power, but that is an orthogonal problem.

Regards,

Hans