Skip to contentSkip to conversation

usb: xhci: avoid VBus glitch during controller reset operation

PATCH In review Peter Chen 3 participants 6 messages Aug 21, 2019 – Aug 23, 2019

Subsystems (2)
lore
Share & tools
lore

Peter Chen

author
#0 lore raw

PATCH 1/1 usb: xhci: avoid VBus glitch during controller reset operation

According to xHCI 1.1 CH4.19.4 Port Power:
While Chip Hardware Reset or HCRST is asserted, the value of PP is undefined. If the xHC supports power switches (PPC = ‘1’) then VBus may be deasserted during this time. PP (and VBus) shall be enabled immediately upon exiting the reset condition.

The VBus glitch may cause some USB devices work abnormal, we observe it at NXP LS1012AFWRY/LS1043ARDB/LX2160AQDS/LS1088ARDB platforms. To avoid this Vbus glitch, we could set PP as 0 before HCRST, and the PP will back to 1 after HCRST according to spec.

---
 drivers/usb/host/xhci.c | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)
drivers/usb/host/xhci.c +14 -1
index 6b34a573c3d9..f5a7b5d63031 100644
@@ -167,7 +167,8 @@ int xhci_reset(struct xhci_hcd *xhci)
 {
 	u32 command;
 	u32 state;
-	int ret;
+	int ret, i;
+	u32 portsc;
 
 	state = readl(&xhci->op_regs->status);
 
@@ -181,6 +182,18 @@ int xhci_reset(struct xhci_hcd *xhci)
 		return 0;
 	}
 
+	/*
+	 * Keep PORTSC.PP as 0 before HCRST to eliminate
+	 * Vbus glitch, see CH 4.19.4.
+	 */
+	for (i = 0; i < HCS_MAX_PORTS(xhci->hcs_params1); i++) {
+		__le32 __iomem *port_addr = &xhci->op_regs->port_status_base +
+				NUM_PORT_REGS * i;
+		portsc = readl(port_addr);
+		portsc &= ~PORT_POWER;
+		writel(portsc, port_addr);
+	}
+
 	xhci_dbg_trace(xhci, trace_xhci_dbg_init, "// Reset the HC");
 	command = readl(&xhci->op_regs->command);
 	command |= CMD_RESET;
Signature · 1 line
2.17.1
#1 lore raw
Quoted 11 lines · Peter Chen

According to xHCI 1.1 CH4.19.4 Port Power:
While Chip Hardware Reset or HCRST is asserted,
the value of PP is undefined. If the xHC supports power switches (PPC = ‘1’) then VBus may be deasserted during this time. PP (and VBus) shall be enabled immediately upon exiting the reset condition.

The VBus glitch may cause some USB devices work abnormal, we observe it at NXP LS1012AFWRY/LS1043ARDB/LX2160AQDS/LS1088ARDB platforms. To avoid this Vbus glitch, we could set PP as 0 before HCRST, and the PP will back to 1 after HCRST according to spec.

Is this glitch causing issues already at the first xHC reset when we are loading the xhci driver, or only later resets, like xHC reset at resume?

Quoted 25 lines
---
  drivers/usb/host/xhci.c | 15 ++++++++++++++-
  1 file changed, 14 insertions(+), 1 deletion(-)
drivers/usb/host/xhci.c +2 -1
index 6b34a573c3d9..f5a7b5d63031 100644
@@ -167,7 +167,8 @@ int xhci_reset(struct xhci_hcd *xhci)
  {
  	u32 command;
  	u32 state;
-	int ret;
+	int ret, i;
+	u32 portsc;
 
  	state = readl(&xhci->op_regs->status);
 
@@ -181,6 +182,18 @@ int xhci_reset(struct xhci_hcd *xhci)
  		return 0;
  	}
 

drivers/usb/host/xhci.c int xhci_reset(struct xhci_hcd *xhci) · line 194

Quoted 10 lines
drivers/usb/host/xhci.c +10 -0
+	/*
+	 * Keep PORTSC.PP as 0 before HCRST to eliminate
+	 * Vbus glitch, see CH 4.19.4.
+	 */
+	for (i = 0; i < HCS_MAX_PORTS(xhci->hcs_params1); i++) {
+		__le32 __iomem *port_addr = &xhci->op_regs->port_status_base +
+				NUM_PORT_REGS * i;
+		portsc = readl(port_addr);
+		portsc &= ~PORT_POWER;
+		writel(portsc, port_addr);

Not all bits read from PORTSC should be written back, you might need portsc = xhci_port_state_to_neutral(portsc) here.

Normally I'd recommend using the xhci_set_port_power() helper instead, but
if this is is needed at driver load time then port arrays may not be set up yet. xhci_set_port_power() would take care of possible ACPI method calls, and add some debugging.

Not sure if this will impact some other platforms, maybe it would be better to move this to a separate function, and call it before xhci_reset() if a quirk is set.

-Mathias

author
#2 lore raw
Quoted 8 lines · Mathias Nyman
Quoted 7 lines · Peter Chen

According to xHCI 1.1 CH4.19.4 Port Power:
While Chip Hardware Reset or HCRST is asserted,
the value of PP is undefined. If the xHC supports power switches (PPC = ‘1’) then VBus may be deasserted during this time. PP (and VBus) shall be enabled immediately upon exiting the reset condition.

Quoted 7 lines
Quoted 4 lines

The VBus glitch may cause some USB devices work abnormal, we observe it at NXP LS1012AFWRY/LS1043ARDB/LX2160AQDS/LS1088ARDB platforms. To avoid this Vbus glitch, we could set PP as 0 before HCRST, and the PP will back to 1 after HCRST according to spec.

Is this glitch causing issues already at the first xHC reset when we are loading the xhci driver, or only later resets, like xHC reset at resume?

The first time, Ran would you please confirm?

Quoted 28 lines
Quoted 28 lines
---
  drivers/usb/host/xhci.c | 15 ++++++++++++++-
  1 file changed, 14 insertions(+), 1 deletion(-)
drivers/usb/host/xhci.c +7 -1
index 6b34a573c3d9..f5a7b5d63031 100644
@@ -167,7 +167,8 @@ int xhci_reset(struct xhci_hcd *xhci)
  {
  	u32 command;
  	u32 state;
-	int ret;
+	int ret, i;
+	u32 portsc;
  	state = readl(&xhci->op_regs->status);
@@ -181,6 +182,18 @@ int xhci_reset(struct xhci_hcd *xhci)
  		return 0;
  	}
+	/*
+	 * Keep PORTSC.PP as 0 before HCRST to eliminate
+	 * Vbus glitch, see CH 4.19.4.
+	 */
+	for (i = 0; i < HCS_MAX_PORTS(xhci->hcs_params1); i++) {
Quoted 8 lines

drivers/usb/host/xhci.c int xhci_reset(struct xhci_hcd *xhci) · line 194

Quoted 5 lines
drivers/usb/host/xhci.c +5 -0
+		__le32 __iomem *port_addr = &xhci->op_regs->port_status_base +
+				NUM_PORT_REGS * i;
+		portsc = readl(port_addr);
+		portsc &= ~PORT_POWER;
+		writel(portsc, port_addr);

Not all bits read from PORTSC should be written back, you might need portsc = xhci_port_state_to_neutral(portsc) here.

Will change.

Quoted 5 lines

Normally I'd recommend using the xhci_set_port_power() helper instead, but
if this is is needed at driver load time then port arrays may not be set up yet. xhci_set_port_power() would take care of possible ACPI method calls, and add some debugging.

It is needed at load time, so I did not use port array.

Quoted 3 lines

Not sure if this will impact some other platforms, maybe it would be better to move this to a separate function, and call it before xhci_reset() if a quirk is set.

It follows spec, not at quirk. We talked with Synopsis engineer (case: 8001235479), they confirmed this behaviour follows spec. Besides, I tried at both dwc3 and cadence3 xHCI platforms, the PORT_POWER will be set again after controller set.

What's potential issue you consider?

Thanks,
Peter

#3 lore raw

HI Peter, Mathias,

Sorry for the late review.

Quoted 12 lines · Peter Chen
Quoted 10 lines · Mathias Nyman
Quoted 9 lines · Peter Chen

According to xHCI 1.1 CH4.19.4 Port Power:
While Chip Hardware Reset or HCRST is asserted,
the value of PP is undefined. If the xHC supports power switches (PPC = ‘1’) then VBus may be deasserted during this time. PP (and VBus) shall be enabled immediately upon exiting the reset condition.

The VBus glitch may cause some USB devices work abnormal, we observe it at NXP LS1012AFWRY/LS1043ARDB/LX2160AQDS/LS1088ARDB platforms.

Quoted 8 lines

To

Quoted 5 lines
Quoted 2 lines

avoid this Vbus glitch, we could set PP as 0 before HCRST, and the PP will back to 1 after HCRST according to spec.

Is this glitch causing issues already at the first xHC reset when we are loading the xhci driver, or only later resets, like xHC reset at resume?

The first time, Ran would you please confirm?

My understand is this will happened whenever PP is set to 0, such as xHCI reset. So I think it might also be observed during resume if xHC do reset.

However, according to my previous testing, it might be too late to
do this port power off in xhci_reset(), actually the VBUS will assert once DWC3 driver set IP to host mode (before doing xHC reset). So the glitch still can be observed on the scope; I have more issue description in another patch discussion about this, please see

lore.kernel.org/patchwork/patch/1032425/
Quoted from it:
Actually I have done experiment like what you suggested (in xhci-plat.c), but the timing seems too late--making VBUS waveform look like a square wave as below:

Here DWC3 enable host mode, VBUS on
|
+5V /---------\ 40ms /---------------------------....

0V  ________/   90ms   \______/
                       |      |           
                       |      Here do xhci reset, VBUS back to +5V again

Here set all PORTSC[PP] to 0 in xhci_gen_setup()

So I am afraid the solution might have to be added in DWC3 core driver where just after host mode enabling code if want fix this :(

Regards,
Ran

Quoted 64 lines
Quoted 30 lines
Quoted 23 lines
---
  drivers/usb/host/xhci.c | 15 ++++++++++++++-
  1 file changed, 14 insertions(+), 1 deletion(-)
drivers/usb/host/xhci.c index renamed from drivers/usb/host/xhci.c +0 -0

6b34a573c3d9..f5a7b5d63031 100644

drivers/usb/host/xhci.c +2 -1
@@ -167,7 +167,8 @@ int xhci_reset(struct xhci_hcd *xhci)
  {
  	u32 command;
  	u32 state;
-	int ret;
+	int ret, i;
+	u32 portsc;
  	state = readl(&xhci->op_regs->status); @@ -181,6 +182,18 @@ int
 xhci_reset(struct xhci_hcd *xhci)
  		return 0;
  	}

drivers/usb/host/xhci.c int xhci_reset(struct xhci_hcd *xhci)

Quoted 6 lines
drivers/usb/host/xhci.c +6 -0
+	/*
+	 * Keep PORTSC.PP as 0 before HCRST to eliminate
+	 * Vbus glitch, see CH 4.19.4.
+	 */
+	for (i = 0; i < HCS_MAX_PORTS(xhci->hcs_params1); i++) {
+		__le32 __iomem *port_addr = &xhci->op_regs-

port_status_base +

Quoted 7 lines

drivers/usb/host/xhci.c int xhci_reset(struct xhci_hcd *xhci) · line 194

Quoted 4 lines
drivers/usb/host/xhci.c +4 -0
+				NUM_PORT_REGS * i;
+		portsc = readl(port_addr);
+		portsc &= ~PORT_POWER;
+		writel(portsc, port_addr);

Not all bits read from PORTSC should be written back, you might need portsc = xhci_port_state_to_neutral(portsc) here.

Will change.

Quoted 3 lines

Normally I'd recommend using the xhci_set_port_power() helper instead, but if this is is needed at driver load time then port arrays may not be set up

yet.

Quoted 1 line

xhci_set_port_power() would take care of possible ACPI method calls, and add

some debugging.

Quoted 1 line

It is needed at load time, so I did not use port array.

Quoted 2 lines

Not sure if this will impact some other platforms, maybe it would be better to move this to a separate function, and call it before xhci_reset() if a

quirk is set.

Quoted 1 line

It follows spec, not at quirk. We talked with Synopsis engineer (case: 8001235479), they confirmed this behaviour follows spec. Besides, I tried at both dwc3 and cadence3 xHCI platforms, the PORT_POWER will be set again after controller set.

What's potential issue you consider?

Thanks,
Peter

author
#4 lore raw
Quoted 21 lines · Ran Wang

HI Peter, Mathias,

Sorry for the late review.

Quoted 16 lines · Peter Chen
Quoted 10 lines · Mathias Nyman
Quoted 9 lines · Peter Chen

According to xHCI 1.1 CH4.19.4 Port Power:
While Chip Hardware Reset or HCRST is asserted,
the value of PP is undefined. If the xHC supports power switches (PPC = ‘1’) then VBus may be deasserted during this time. PP (and VBus) shall be enabled immediately upon exiting the reset condition.

The VBus glitch may cause some USB devices work abnormal, we observe it at NXP LS1012AFWRY/LS1043ARDB/LX2160AQDS/LS1088ARDB platforms.

To

Quoted 3 lines
Quoted 2 lines

avoid this Vbus glitch, we could set PP as 0 before HCRST, and the PP will back to 1 after HCRST according to spec.

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

Is this glitch causing issues already at the first xHC reset when we are loading the xhci driver, or only later resets, like xHC reset at resume?

The first time, Ran would you please confirm?

My understand is this will happened whenever PP is set to 0, such as xHCI reset. So I think it might also be observed during resume if xHC do reset.

Yes, Vbus glitch should exists whenever we do controller reset, I thought you only meet this issue during boots up.

Quoted 14 lines

However, according to my previous testing, it might be too late to
do this port power off in xhci_reset(), actually the VBUS will assert once DWC3 driver set IP to host mode (before doing xHC reset). So the glitch still can be observed on the scope; I have more issue description in another patch discussion about this, please see

lore.kernel.org/patchwork/patch/1032425/
Quoted from it:
Actually I have done experiment like what you suggested (in xhci-plat.c), but the timing seems too late--making VBUS waveform look like a square wave as below:

Here DWC3 enable host mode, VBUS on
|
+5V /---------\ 40ms /---------------------------....

Quoted 7 lines
0V  ________/   90ms   \______/
                       |      |           
                       |      Here do xhci reset, VBUS back to +5V again

Here set all PORTSC[PP] to 0 in xhci_gen_setup()

So I am afraid the solution might have to be added in DWC3 core driver where just after host mode enabling code if want fix this :(

Per spec 4.19.4 Port Power:

Whether an xHC implementation supports port power switches or not, it shall automatically enable VBus on all Root Hub ports after a Chip Hardware Reset or HCRST.

Ran's observation confirmed it, PP is asserted once xHCI is enabled. From the code, the HCRST will be at boots up and system resume. So, it seems we can't keep PP always asserted. For xHCI, to avoid Vbus glitch totally, we may have to control Vbus without PP (eg, configure PP pin as GPIO).

Peter

Quoted 67 lines

Regards,
Ran

Quoted 64 lines
Quoted 30 lines
Quoted 23 lines
---
  drivers/usb/host/xhci.c | 15 ++++++++++++++-
  1 file changed, 14 insertions(+), 1 deletion(-)
drivers/usb/host/xhci.c index renamed from drivers/usb/host/xhci.c +0 -0

6b34a573c3d9..f5a7b5d63031 100644

drivers/usb/host/xhci.c +2 -1
@@ -167,7 +167,8 @@ int xhci_reset(struct xhci_hcd *xhci)
  {
  	u32 command;
  	u32 state;
-	int ret;
+	int ret, i;
+	u32 portsc;
  	state = readl(&xhci->op_regs->status); @@ -181,6 +182,18 @@ int
 xhci_reset(struct xhci_hcd *xhci)
  		return 0;
  	}

drivers/usb/host/xhci.c int xhci_reset(struct xhci_hcd *xhci)

Quoted 6 lines
drivers/usb/host/xhci.c +6 -0
+	/*
+	 * Keep PORTSC.PP as 0 before HCRST to eliminate
+	 * Vbus glitch, see CH 4.19.4.
+	 */
+	for (i = 0; i < HCS_MAX_PORTS(xhci->hcs_params1); i++) {
+		__le32 __iomem *port_addr = &xhci->op_regs-

port_status_base +

Quoted 7 lines

drivers/usb/host/xhci.c int xhci_reset(struct xhci_hcd *xhci) · line 194

Quoted 4 lines
drivers/usb/host/xhci.c +4 -0
+				NUM_PORT_REGS * i;
+		portsc = readl(port_addr);
+		portsc &= ~PORT_POWER;
+		writel(portsc, port_addr);

Not all bits read from PORTSC should be written back, you might need portsc = xhci_port_state_to_neutral(portsc) here.

Will change.

Quoted 3 lines

Normally I'd recommend using the xhci_set_port_power() helper instead, but if this is is needed at driver load time then port arrays may not be set up

yet.

Quoted 1 line

xhci_set_port_power() would take care of possible ACPI method calls, and add

some debugging.

Quoted 1 line

It is needed at load time, so I did not use port array.

Quoted 2 lines

Not sure if this will impact some other platforms, maybe it would be better to move this to a separate function, and call it before xhci_reset() if a

quirk is set.

Quoted 1 line

It follows spec, not at quirk. We talked with Synopsis engineer (case: 8001235479), they confirmed this behaviour follows spec. Besides, I tried at both dwc3 and cadence3 xHCI platforms, the PORT_POWER will be set again after controller set.

What's potential issue you consider?

Thanks,
Peter

#5 lore raw

Hi Peter,

Quoted 64 lines · Peter Chen
Quoted 17 lines · Ran Wang

HI Peter, Mathias,

Sorry for the late review.

Quoted 12 lines · Peter Chen
Quoted 10 lines · Mathias Nyman
Quoted 9 lines · Peter Chen

According to xHCI 1.1 CH4.19.4 Port Power:
While Chip Hardware Reset or HCRST is asserted,
the value of PP is undefined. If the xHC supports power switches (PPC = ‘1’) then VBus may be deasserted during this time. PP (and VBus) shall be enabled immediately upon exiting the reset condition.

The VBus glitch may cause some USB devices work abnormal, we observe it at NXP LS1012AFWRY/LS1043ARDB/LX2160AQDS/LS1088ARDB

platforms.

Quoted 6 lines
Quoted 6 lines

To

Quoted 5 lines
Quoted 2 lines

avoid this Vbus glitch, we could set PP as 0 before HCRST, and the PP will back to 1 after HCRST according to spec.

Is this glitch causing issues already at the first xHC reset when we are loading the xhci driver, or only later resets, like xHC reset at

resume?

Quoted 5 lines
Quoted 2 lines

The first time, Ran would you please confirm?

My understand is this will happened whenever PP is set to 0, such as xHCI reset. So I think it might also be observed during resume if xHC do reset.

Yes, Vbus glitch should exists whenever we do controller reset, I thought you only meet this issue during boots up.

Quoted 11 lines

However, according to my previous testing, it might be too late to do this port power off in xhci_reset(), actually the VBUS will assert once DWC3 driver set IP to host mode (before doing xHC reset). So the glitch still can be observed on the scope; I have more issue description in another patch discussion about this, please see

lore.kernel.org/patchwork/patch/1032425/
Quoted from it:
Actually I have done experiment like what you suggested (in xhci-plat.c), but the timing seems too late--making VBUS waveform look like a

square wave as below:

Quoted 4 lines

Here DWC3 enable host mode, VBUS on
|
+5V /---------\ 40ms /---------------------------....

Quoted 9 lines
0V  ________/   90ms   \______/
                       |      |
                       |      Here do xhci reset, VBUS back to +5V again

Here set all PORTSC[PP] to 0 in
xhci_gen_setup()

So I am afraid the solution might have to be added in DWC3 core driver where just after host mode enabling code if want fix this :(

Per spec 4.19.4 Port Power:

Quoted 8 lines

Whether an xHC implementation supports port power switches or not, it shall automatically enable VBus on all Root Hub ports after a Chip Hardware Reset or HCRST.

Ran's observation confirmed it, PP is asserted once xHCI is enabled. From the code, the HCRST will be at boots up and system resume. So, it seems we can't keep PP always asserted. For xHCI, to avoid Vbus glitch totally, we may have to control Vbus without PP (eg, configure PP pin as GPIO).

Yes, another option is to design a better VBUS driving circuit on HW side to filter this glitch out. I have found some earlier version of LS1043ARDB board design is perfect on doing this.

Anyway, for boot, my suggestion is to do it in DWC3 driver once after enabling host mode. But that solution is ugly, I have to admit.
And for resume, frankly I didn’t notice this before (thanks for remind), would do further survey if can found a similar solution.

Regards,
Ran

Quoted 68 lines
Quoted 68 lines
Quoted 68 lines
Quoted 31 lines
Quoted 24 lines
---
  drivers/usb/host/xhci.c | 15 ++++++++++++++-
  1 file changed, 14 insertions(+), 1 deletion(-)
drivers/usb/host/xhci.c +0 -0

index
6b34a573c3d9..f5a7b5d63031 100644

drivers/usb/host/xhci.c +2 -1
@@ -167,7 +167,8 @@ int xhci_reset(struct xhci_hcd *xhci)
  {
  	u32 command;
  	u32 state;
-	int ret;
+	int ret, i;
+	u32 portsc;
  	state = readl(&xhci->op_regs->status); @@ -181,6 +182,18 @@
 int xhci_reset(struct xhci_hcd *xhci)
  		return 0;
  	}

drivers/usb/host/xhci.c int xhci_reset(struct xhci_hcd *xhci)

Quoted 6 lines
drivers/usb/host/xhci.c +6 -0
+	/*
+	 * Keep PORTSC.PP as 0 before HCRST to eliminate
+	 * Vbus glitch, see CH 4.19.4.
+	 */
+	for (i = 0; i < HCS_MAX_PORTS(xhci->hcs_params1); i++) {
+		__le32 __iomem *port_addr = &xhci->op_regs-

port_status_base +

Quoted 7 lines

drivers/usb/host/xhci.c int xhci_reset(struct xhci_hcd *xhci) · line 194

Quoted 4 lines
drivers/usb/host/xhci.c +4 -0
+				NUM_PORT_REGS * i;
+		portsc = readl(port_addr);
+		portsc &= ~PORT_POWER;
+		writel(portsc, port_addr);

Not all bits read from PORTSC should be written back, you might need portsc = xhci_port_state_to_neutral(portsc) here.

Will change.

Quoted 4 lines

Normally I'd recommend using the xhci_set_port_power() helper instead, but if this is is needed at driver load time then port arrays may not be set up

yet.

Quoted 2 lines

xhci_set_port_power() would take care of possible ACPI method calls, and add

some debugging.

Quoted 1 line

It is needed at load time, so I did not use port array.

Quoted 3 lines

Not sure if this will impact some other platforms, maybe it would be better to move this to a separate function, and call it before xhci_reset() if a

quirk is set.

Quoted 1 line

It follows spec, not at quirk. We talked with Synopsis engineer (case: 8001235479), they confirmed this behaviour follows spec. Besides, I tried at both dwc3 and cadence3 xHCI platforms, the PORT_POWER will be set again after controller set.

What's potential issue you consider?

Thanks,
Peter