Skip to contentSkip to conversation

PCI: pciehp: clear cmd_busy bit when Command Completed in polling mode

PATCH In review Liguang Zhang 3 participants 11 messages Nov 11, 2021 – Feb 3, 2022 PCI

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 PCI: pciehp: clear cmd_busy bit when Command Completed in polling mode

This patch fixes this problem that on driver probe from system startup, pciehp checks the Presence Detect State bit in the Slot Status register to bring up an occupied slot or bring down an unoccupied slot. If empty slot's power status is on, turn power off. The Hot-Plug interrupt isn't requested yet, so avoid triggering a notification by calling pcie_disable_notification().

Both the CCIE and HPIE bits are masked in pcie_disable_notification(), when we issue a hotplug command, pcie_wait_cmd() will polling the Command Completed bit instead of waiting for an interrupt. But cmd_busy bit was not cleared when Command Completed which results in timeouts like this in pciehp_power_off_slot() and pcie_init_notification():

pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x01c0
(issued 2264 msec ago)
pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x05c0
(issued 2288 msec ago)

---
 drivers/pci/hotplug/pciehp_hpc.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/pci/hotplug/pciehp_hpc.c +2 -0
index 83a0fa119cae..8698aefc6041 100644
@@ -98,6 +98,8 @@ static int pcie_poll_cmd(struct controller *ctrl, int timeout)
 		if (slot_status & PCI_EXP_SLTSTA_CC) {
 			pcie_capability_write_word(pdev, PCI_EXP_SLTSTA,
 						   PCI_EXP_SLTSTA_CC);
+			ctrl->cmd_busy = 0;
+			smp_mb();
 			return 1;
 		}
 		msleep(10);
Signature · 1 line
2.19.1.6.gb485710b
author
#1 lore raw

Hi Bjorn & Lukas & Kuppuswamy & Amey,

Gentle ping! Any comments on this patch?

Quoted 36 lines · Liguang Zhang

This patch fixes this problem that on driver probe from system startup, pciehp checks the Presence Detect State bit in the Slot Status register to bring up an occupied slot or bring down an unoccupied slot. If empty slot's power status is on, turn power off. The Hot-Plug interrupt isn't requested yet, so avoid triggering a notification by calling pcie_disable_notification().

Both the CCIE and HPIE bits are masked in pcie_disable_notification(), when we issue a hotplug command, pcie_wait_cmd() will polling the Command Completed bit instead of waiting for an interrupt. But cmd_busy bit was not cleared when Command Completed which results in timeouts like this in pciehp_power_off_slot() and pcie_init_notification():

pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x01c0
(issued 2264 msec ago)
pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x05c0
(issued 2288 msec ago)

---
  drivers/pci/hotplug/pciehp_hpc.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/pci/hotplug/pciehp_hpc.c +2 -0
index 83a0fa119cae..8698aefc6041 100644
@@ -98,6 +98,8 @@ static int pcie_poll_cmd(struct controller *ctrl, int timeout)
  		if (slot_status & PCI_EXP_SLTSTA_CC) {
  			pcie_capability_write_word(pdev, PCI_EXP_SLTSTA,
  						   PCI_EXP_SLTSTA_CC);
+			ctrl->cmd_busy = 0;
+			smp_mb();
  			return 1;
  		}
  		msleep(10);
#2 lore raw
In reply to Liguang Zhang
Quoted 10 lines · Liguang Zhang

Both the CCIE and HPIE bits are masked in pcie_disable_notification(), when we issue a hotplug command, pcie_wait_cmd() will polling the Command Completed bit instead of waiting for an interrupt. But cmd_busy bit was not cleared when Command Completed which results in timeouts like this in pciehp_power_off_slot() and pcie_init_notification():

pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x01c0
(issued 2264 msec ago)
pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x05c0
(issued 2288 msec ago)

The first timeout occurs with the following bits set in ctrl->slot_ctrl:
PCI_EXP_SLTCTL_PWR_IND_ON | PCI_EXP_SLTCTL_ATTN_IND_OFF

Those bits are set by:
board_added()
pciehp_set_indicators()

The second timeout occurs with:
PCI_EXP_SLTCTL_PWR_IND_ON | PCI_EXP_SLTCTL_ATTN_IND_OFF | PCI_EXP_SLTCTL_PWR_OFF

This might be triggered by:
remove_board()
pciehp_power_off_slot()

So it seems Command Completed is not signaled when changing the Power Indicator, Attention Indicator and Power Controller Control bits in the Slot Control register. But apparently it works for the other bits.

We know there are hotplug controllers out there which suffer from broken Command Completed support. They support it for the bits mentioned above but not the others. So the inverse behavior from your hotplug controller. See this code comment in pcie_do_write_cmd():

/*
* Controllers with the Intel CF118 and similar errata advertise
* Command Completed support, but they only set Command Completed
* if we change the "Control" bits for power, power indicator,
* attention indicator, or interlock. If we only change the
* "Enable" bits, they never set the Command Completed bit. */

drivers/pci/hotplug/pciehp_hpc.c static int pcie_poll_cmd(struct controller *ctrl, int timeout) · line 105

Quoted 11 lines
drivers/pci/hotplug/pciehp_hpc.c +2 -0
@@ -98,6 +98,8 @@ static int pcie_poll_cmd(struct controller *ctrl, int timeout)
 		if (slot_status & PCI_EXP_SLTSTA_CC) {
 			pcie_capability_write_word(pdev, PCI_EXP_SLTSTA,
 						   PCI_EXP_SLTSTA_CC);
+			ctrl->cmd_busy = 0;
+			smp_mb();
 			return 1;
 		}
 		msleep(10);

I suspect that this patch merely papers over the problem and that the real solution would be to either apply quirk_cmd_compl or a similar quirk to your hotplug controller.

Please open a bug on bugzilla.kernel.org and attach full output of lspci -vv and dmesg. Be sure to add the following to the command line:
pciehp.pciehp_debug=1 dyndbg="file pciehp* +p"

Once you've done that, please report the bugzilla link here so that we can analyze the issue properly.

Thanks,

Lukas

author
#3 lore raw

Hi,thanks for your comments.

Quoted 7 lines · Lukas Wunner
Quoted 6 lines · Liguang Zhang

Both the CCIE and HPIE bits are masked in pcie_disable_notification(), when we issue a hotplug command, pcie_wait_cmd() will polling the Command Completed bit instead of waiting for an interrupt. But cmd_busy bit was not cleared when Command Completed which results in timeouts like this in pciehp_power_off_slot() and pcie_init_notification():

Quoted 6 lines
Quoted 4 lines

pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x01c0
(issued 2264 msec ago)
pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x05c0
(issued 2288 msec ago)

The first timeout occurs with the following bits set in ctrl->slot_ctrl:
PCI_EXP_SLTCTL_PWR_IND_ON | PCI_EXP_SLTCTL_ATTN_IND_OFF

After some debug, the first timeout occurs:

    pciehp_power_off_slot

        pcie_write_cmd(ctrl, PCI_EXP_SLTCTL_PWR_OFF, PCI_EXP_SLTCTL_PCC)

            pcie_do_write_cmd

                    /*
                     * Always wait for any previous command that might still be in progress */
                    pcie_wait_cmd(ctrl);  // at the beginning

                        if (!ctrl->cmd_busy)  // cmd_busy was not zero return;

Why cmd_busy was not flase, because in function pcie_disable_notification cmd_busy was not setting correctly:

    pcie_disable_notification  //  Both the CCIE and HPIE bits are masked

        pcie_write_cmd

            pcie_do_write_cmd

                ctrl->cmd_busy = 1  // cmd_busy was setting to 1

                pcie_wait_cmd

                    pcie_poll_cmd // pcie_wait_cmd() will polling the Command Completed bit instead of waiting for an interrupt

                         After debug we found Command Completed can be signaled without cmd_busy was setting to 0.

If we use interrupt mode pciehp_isr:

        pciehp_isr

            if (events & PCI_EXP_SLTSTA_CC) {
                ctrl->cmd_busy = 0;  //  cmd_busy was setting to zero, this was left in pcie_poll_cmd.

Thanks,

Liguang

Quoted 60 lines

Those bits are set by:
board_added()
pciehp_set_indicators()

The second timeout occurs with:
PCI_EXP_SLTCTL_PWR_IND_ON | PCI_EXP_SLTCTL_ATTN_IND_OFF | PCI_EXP_SLTCTL_PWR_OFF

This might be triggered by:
remove_board()
pciehp_power_off_slot()

So it seems Command Completed is not signaled when changing the Power Indicator, Attention Indicator and Power Controller Control bits in the Slot Control register. But apparently it works for the other bits.

We know there are hotplug controllers out there which suffer from broken Command Completed support. They support it for the bits mentioned above but not the others. So the inverse behavior from your hotplug controller. See this code comment in pcie_do_write_cmd():

/*
* Controllers with the Intel CF118 and similar errata advertise
* Command Completed support, but they only set Command Completed
* if we change the "Control" bits for power, power indicator,
* attention indicator, or interlock. If we only change the
* "Enable" bits, they never set the Command Completed bit. */

drivers/pci/hotplug/pciehp_hpc.c static int pcie_poll_cmd(struct controller *ctrl, int timeout) · line 105

Quoted 11 lines
drivers/pci/hotplug/pciehp_hpc.c +2 -0
@@ -98,6 +98,8 @@ static int pcie_poll_cmd(struct controller *ctrl, int timeout)
  		if (slot_status & PCI_EXP_SLTSTA_CC) {
  			pcie_capability_write_word(pdev, PCI_EXP_SLTSTA,
  						   PCI_EXP_SLTSTA_CC);
+			ctrl->cmd_busy = 0;
+			smp_mb();
  			return 1;
  		}
  		msleep(10);

I suspect that this patch merely papers over the problem and that the real solution would be to either apply quirk_cmd_compl or a similar quirk to your hotplug controller.

Please open a bug on bugzilla.kernel.org and attach full output of lspci -vv and dmesg. Be sure to add the following to the command line:
pciehp.pciehp_debug=1 dyndbg="file pciehp* +p"

Once you've done that, please report the bugzilla link here so that we can analyze the issue properly.

Thanks,

Lukas

#4 lore raw
Quoted 8 lines · luanshi

2021/11/19 20:00, Lukas Wunner:

Quoted 7 lines

Please open a bug on bugzilla.kernel.org and attach full output of lspci -vv and dmesg. Be sure to add the following to the command line:
pciehp.pciehp_debug=1 dyndbg="file pciehp* +p"

Once you've done that, please report the bugzilla link here so that we can analyze the issue properly.

I really need you to perform the above steps in order to analyze what's going on here.

Again, if you get such timeout messages, it's usually not caused by a bug in the driver but by an erratum in the hardware, i.e. the hardware neglected to signal Command Completed in response to a Slot Control register write.

Thanks,

Lukas

author
#5 lore raw

Sorry for the late reply, the machine is occupied by someone else.

For detailed information:

https://bugzilla.kernel.org/show_bug.cgi?id=215143

Thanks,

Liguang

Quoted 20 lines · Lukas Wunner
Quoted 8 lines · luanshi

2021/11/19 20:00, Lukas Wunner:

Quoted 7 lines

Please open a bug on bugzilla.kernel.org and attach full output of lspci -vv and dmesg. Be sure to add the following to the command line:
pciehp.pciehp_debug=1 dyndbg="file pciehp* +p"

Once you've done that, please report the bugzilla link here so that we can analyze the issue properly.

I really need you to perform the above steps in order to analyze what's going on here.

Again, if you get such timeout messages, it's usually not caused by a bug in the driver but by an erratum in the hardware, i.e. the hardware neglected to signal Command Completed in response to a Slot Control register write.

Thanks,

Lukas

Reviewed-by on v1 1/1
#6 lore raw
In reply to Liguang Zhang
Quoted 13 lines · Liguang Zhang

This patch fixes this problem that on driver probe from system startup, pciehp checks the Presence Detect State bit in the Slot Status register to bring up an occupied slot or bring down an unoccupied slot. If empty slot's power status is on, turn power off. The Hot-Plug interrupt isn't requested yet, so avoid triggering a notification by calling pcie_disable_notification().

Both the CCIE and HPIE bits are masked in pcie_disable_notification(), when we issue a hotplug command, pcie_wait_cmd() will polling the Command Completed bit instead of waiting for an interrupt. But cmd_busy bit was not cleared when Command Completed which results in timeouts like this in pciehp_power_off_slot() and pcie_init_notification():

Quoted 6 lines

pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x01c0
(issued 2264 msec ago)
pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x05c0
(issued 2288 msec ago)

Fixes a5dd4b4b0570 ("PCI: pciehp: Wait for hotplug command completion where necessary")
Cc [email protected] # v4.19+

Thanks a lot, that's a really good catch.

It's a somewhat intricate bug, so I'll try to explain in my own words:

If notification is disabled (HPIE or CCIE not set in the Slot Status register), we rely on pcie_poll_cmd() to poll for Command Completed. But once it's signaled, we neglect to clear ctrl->cmd_busy. (Normally it is cleared by the hardirq handler pciehp_isr() if notification is enabled.)

The result is that starting with the second Slot Control write, pciehp will gratuitously wait for a command to finish which has already finished and it will incorrectly report a timeout.

The bug was originally introduced in 2015 by commit a5dd4b4b0570 ("PCI: pciehp: Wait for hotplug command completion where necessary"), but didn't manifest itself because the first Slot Control Write already enabled notification and from that point on the hardirq handler would clear ctrl->cmd_busy. However I think the bug may have manifested itself with pciehp_poll_mode=1.

It wasn't until commit 4e6a13356f1c ("PCI: pciehp: Deduplicate presence check on probe & resume") that multiple consecutive Slot Control writes were performed on ->probe with notification disabled, so that's the commit which first exposed the bug with pciehp_poll_mode=0.

Thanks,

Lukas

Quoted 19 lines
---
 drivers/pci/hotplug/pciehp_hpc.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/pci/hotplug/pciehp_hpc.c +2 -0
index 83a0fa119cae..8698aefc6041 100644
@@ -98,6 +98,8 @@ static int pcie_poll_cmd(struct controller *ctrl, int timeout)
 		if (slot_status & PCI_EXP_SLTSTA_CC) {
 			pcie_capability_write_word(pdev, PCI_EXP_SLTSTA,
 						   PCI_EXP_SLTSTA_CC);
+			ctrl->cmd_busy = 0;
+			smp_mb();
 			return 1;
 		}
 		msleep(10);
Signature · 1 line
2.19.1.6.gb485710b
author
#8 lore raw
In reply to Lukas Wunner

Hi Bjorn & Lukas & Kuppuswamy & Amey,

Gentle ping! Any comments on this patch,should be merged?

Thanks,

Liguang

Quoted 74 lines · Lukas Wunner
Quoted 13 lines · Liguang Zhang

This patch fixes this problem that on driver probe from system startup, pciehp checks the Presence Detect State bit in the Slot Status register to bring up an occupied slot or bring down an unoccupied slot. If empty slot's power status is on, turn power off. The Hot-Plug interrupt isn't requested yet, so avoid triggering a notification by calling pcie_disable_notification().

Both the CCIE and HPIE bits are masked in pcie_disable_notification(), when we issue a hotplug command, pcie_wait_cmd() will polling the Command Completed bit instead of waiting for an interrupt. But cmd_busy bit was not cleared when Command Completed which results in timeouts like this in pciehp_power_off_slot() and pcie_init_notification():

Quoted 6 lines

pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x01c0
(issued 2264 msec ago)
pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x05c0
(issued 2288 msec ago)

Fixes a5dd4b4b0570 ("PCI: pciehp: Wait for hotplug command completion where necessary")
Cc [email protected] # v4.19+

Thanks a lot, that's a really good catch.

It's a somewhat intricate bug, so I'll try to explain in my own words:

If notification is disabled (HPIE or CCIE not set in the Slot Status register), we rely on pcie_poll_cmd() to poll for Command Completed. But once it's signaled, we neglect to clear ctrl->cmd_busy. (Normally it is cleared by the hardirq handler pciehp_isr() if notification is enabled.)

The result is that starting with the second Slot Control write, pciehp will gratuitously wait for a command to finish which has already finished and it will incorrectly report a timeout.

The bug was originally introduced in 2015 by commit a5dd4b4b0570 ("PCI: pciehp: Wait for hotplug command completion where necessary"), but didn't manifest itself because the first Slot Control Write already enabled notification and from that point on the hardirq handler would clear ctrl->cmd_busy. However I think the bug may have manifested itself with pciehp_poll_mode=1.

It wasn't until commit 4e6a13356f1c ("PCI: pciehp: Deduplicate presence check on probe & resume") that multiple consecutive Slot Control writes were performed on ->probe with notification disabled, so that's the commit which first exposed the bug with pciehp_poll_mode=0.

Thanks,

Lukas

Quoted 19 lines
---
  drivers/pci/hotplug/pciehp_hpc.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/pci/hotplug/pciehp_hpc.c +2 -0
index 83a0fa119cae..8698aefc6041 100644
@@ -98,6 +98,8 @@ static int pcie_poll_cmd(struct controller *ctrl, int timeout)
  		if (slot_status & PCI_EXP_SLTSTA_CC) {
  			pcie_capability_write_word(pdev, PCI_EXP_SLTSTA,
  						   PCI_EXP_SLTSTA_CC);
+			ctrl->cmd_busy = 0;
+			smp_mb();
  			return 1;
  		}
  		msleep(10);
Signature · 1 line
2.19.1.6.gb485710b
#9 lore raw
In reply to Lukas Wunner

Hi Bjorn,

the below patch is marked "Changes Requested" in patchwork:

https://patchwork.kernel.org/project/linux-pci/patch/[email protected]/

I think that might be erroneous because the patch is correct, I've provided a Reviewed-by and no change requests are recorded in patchwork or the mailing list archive.

If you've got a few minutes to spare, could you double-check the state in patchwork and provide Liguang Zhang with the changes you'd want (if any)?

Thanks!

Lukas

Quoted 75 lines · Lukas Wunner
Quoted 13 lines · Liguang Zhang

This patch fixes this problem that on driver probe from system startup, pciehp checks the Presence Detect State bit in the Slot Status register to bring up an occupied slot or bring down an unoccupied slot. If empty slot's power status is on, turn power off. The Hot-Plug interrupt isn't requested yet, so avoid triggering a notification by calling pcie_disable_notification().

Both the CCIE and HPIE bits are masked in pcie_disable_notification(), when we issue a hotplug command, pcie_wait_cmd() will polling the Command Completed bit instead of waiting for an interrupt. But cmd_busy bit was not cleared when Command Completed which results in timeouts like this in pciehp_power_off_slot() and pcie_init_notification():

Quoted 6 lines

pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x01c0
(issued 2264 msec ago)
pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x05c0
(issued 2288 msec ago)

Fixes a5dd4b4b0570 ("PCI: pciehp: Wait for hotplug command completion where necessary")

Thanks a lot, that's a really good catch.

It's a somewhat intricate bug, so I'll try to explain in my own words:

If notification is disabled (HPIE or CCIE not set in the Slot Status register), we rely on pcie_poll_cmd() to poll for Command Completed. But once it's signaled, we neglect to clear ctrl->cmd_busy. (Normally it is cleared by the hardirq handler pciehp_isr() if notification is enabled.)

The result is that starting with the second Slot Control write, pciehp will gratuitously wait for a command to finish which has already finished and it will incorrectly report a timeout.

The bug was originally introduced in 2015 by commit a5dd4b4b0570 ("PCI: pciehp: Wait for hotplug command completion where necessary"), but didn't manifest itself because the first Slot Control Write already enabled notification and from that point on the hardirq handler would clear ctrl->cmd_busy. However I think the bug may have manifested itself with pciehp_poll_mode=1.

It wasn't until commit 4e6a13356f1c ("PCI: pciehp: Deduplicate presence check on probe & resume") that multiple consecutive Slot Control writes were performed on ->probe with notification disabled, so that's the commit which first exposed the bug with pciehp_poll_mode=0.

Thanks,

Lukas

Quoted 19 lines
---
 drivers/pci/hotplug/pciehp_hpc.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/pci/hotplug/pciehp_hpc.c +2 -0
index 83a0fa119cae..8698aefc6041 100644
@@ -98,6 +98,8 @@ static int pcie_poll_cmd(struct controller *ctrl, int timeout)
 		if (slot_status & PCI_EXP_SLTSTA_CC) {
 			pcie_capability_write_word(pdev, PCI_EXP_SLTSTA,
 						   PCI_EXP_SLTSTA_CC);
+			ctrl->cmd_busy = 0;
+			smp_mb();
 			return 1;
 		}
 		msleep(10);
Signature · 1 line
2.19.1.6.gb485710b
Reviewed-by Lukas Wunner
#10 lore raw
Quoted 6 lines · Lukas Wunner

Hi Bjorn,

the below patch is marked "Changes Requested" in patchwork:

https://patchwork.kernel.org/project/linux-pci/patch/[email protected]/

Quoted 7 lines

I think that might be erroneous because the patch is correct, I've provided a Reviewed-by and no change requests are recorded in patchwork or the mailing list archive.

If you've got a few minutes to spare, could you double-check the state in patchwork and provide Liguang Zhang with the changes you'd want (if any)?

Thanks for reminding me about this. I don't remember why I marked it "changes requested" but I often do that if there's been significant discussion. In this case you also provided additional information (Fixes tag, bugzilla link, commit log elaboration) that should be included. Given that, I would typically wait for the author to repost it and incorporate the additional information.

But I applied it to pci/hotplug with the commit log below.

commit 92912b175178 ("PCI: pciehp: Clear cmd_busy bit in polling mode")

Author: Liguang Zhang <[email protected]>
Date:   Thu Nov 11 13:42:58 2021 +0800

PCI: pciehp: Clear cmd_busy bit in polling mode

Writes to a Downstream Port's Slot Control register are PCIe hotplug "commands." If the Port supports Command Completed events, software must wait for a command to complete before writing to Slot Control again.

pcie_do_write_cmd() sets ctrl->cmd_busy when it writes to Slot Control. If software notification is enabled, i.e., PCI_EXP_SLTCTL_HPIE and PCI_EXP_SLTCTL_CCIE are set, ctrl->cmd_busy is cleared by pciehp_isr().

But when software notification is disabled, as it is when pcie_init() powers off an empty slot, pcie_wait_cmd() uses pcie_poll_cmd() to poll for command completion, and it neglects to clear ctrl->cmd_busy, which leads to spurious timeouts:

pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x01c0 (issued 2264 msec ago) pcieport 0000:00:03.0: pciehp: Timeout on hotplug command 0x05c0 (issued 2288 msec ago)

Clear ctrl->cmd_busy in pcie_poll_cmd() when it detects a Command Completed event (PCI_EXP_SLTSTA_CC).