Skip to contentSkip to conversation

mmc: pwrseq: Use proper reboot notifier path

PATCH In review Andrew Davis 3 participants 5 messages Jan 26, 2024 – Feb 1, 2024 MMC, SD and SDIO

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 mmc: pwrseq: Use proper reboot notifier path

This driver registers itself as a reboot handler, which means it claims it can reboot the system. It does this so it is called during the system reboot sequence. The correct way to be notified during the reboot sequence is to register a notifier with register_reboot_notifier(). Do this here.

Note this will be called during normal reboots but not emergency reboots. This is the expected behavior, emergency reboot means emergency, not go do some cleanup with emmc pins.. The reboot notifiers are intentionally not called in the emergency path for a reason and working around that by pretending to be a reboot handler is a hack.

---
 drivers/mmc/core/pwrseq_emmc.c | 8 +-------
 1 file changed, 1 insertion(+), 7 deletions(-)
drivers/mmc/core/pwrseq_emmc.c +1 -7
index 3b6d69cefb4eb..d5045fd1a02c1 100644
@@ -70,14 +70,8 @@ static int mmc_pwrseq_emmc_probe(struct platform_device *pdev)
 		return PTR_ERR(pwrseq->reset_gpio);
 
 	if (!gpiod_cansleep(pwrseq->reset_gpio)) {
-		/*
-		 * register reset handler to ensure emmc reset also from
-		 * emergency_reboot(), priority 255 is the highest priority
-		 * so it will be executed before any system reboot handler.
-		 */
 		pwrseq->reset_nb.notifier_call = mmc_pwrseq_emmc_reset_nb;
-		pwrseq->reset_nb.priority = 255;
-		register_restart_handler(&pwrseq->reset_nb);
+		register_reboot_notifier(&pwrseq->reset_nb);
 	} else {
 		dev_notice(dev, "EMMC reset pin tied to a sleepy GPIO driver; reset on emergency-reboot disabled\n");
 	}
Signature · 1 line
2.39.2
#1 lore raw
Quoted 7 lines · Andrew Davis

This driver registers itself as a reboot handler, which means it claims it can reboot the system. It does this so it is called during the system reboot sequence. The correct way to be notified during the reboot sequence is to register a notifier with register_reboot_notifier(). Do this here.

Quoted 5 lines

Note this will be called during normal reboots but not emergency reboots. This is the expected behavior, emergency reboot means emergency, not go do some cleanup with emmc pins.. The reboot notifiers are intentionally not called in the emergency path for a reason and working around that by pretending to be a reboot handler is a hack.

I understand the reason for the $subject patch, but it will not work, unfortunately.

For eMMC we need to manage emergency reboots too. The fiddling with GPIOs isn't a "cleanup", but tries to move the eMMC into a clean reset state. This is needed on some platforms with broken bootloaders (ROM code), that is expecting the eMMC to always start in a clean reset state.

So, we need both parts, as was discussed here [1] too.

Kind regards
Uffe

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

Quoted 29 lines
---
 drivers/mmc/core/pwrseq_emmc.c | 8 +-------
 1 file changed, 1 insertion(+), 7 deletions(-)
drivers/mmc/core/pwrseq_emmc.c +1 -8
index 3b6d69cefb4eb..d5045fd1a02c1 100644
@@ -70,14 +70,8 @@ static int mmc_pwrseq_emmc_probe(struct platform_device *pdev)
                return PTR_ERR(pwrseq->reset_gpio);
 
        if (!gpiod_cansleep(pwrseq->reset_gpio)) {
-               /*
-                * register reset handler to ensure emmc reset also from
-                * emergency_reboot(), priority 255 is the highest priority
-                * so it will be executed before any system reboot handler.
-                */
                pwrseq->reset_nb.notifier_call = mmc_pwrseq_emmc_reset_nb;
-               pwrseq->reset_nb.priority = 255;
-               register_restart_handler(&pwrseq->reset_nb);
+               register_reboot_notifier(&pwrseq->reset_nb);
        } else {
                dev_notice(dev, "EMMC reset pin tied to a sleepy GPIO driver; reset on emergency-reboot disabled\n");
        }
--

2.39.2

#2 lore raw
Quoted 14 lines · Ulf Hansson
Quoted 7 lines · Andrew Davis

This driver registers itself as a reboot handler, which means it claims it can reboot the system. It does this so it is called during the system reboot sequence. The correct way to be notified during the reboot sequence is to register a notifier with register_reboot_notifier(). Do this here.

Quoted 5 lines

Note this will be called during normal reboots but not emergency reboots. This is the expected behavior, emergency reboot means emergency, not go do some cleanup with emmc pins.. The reboot notifiers are intentionally not called in the emergency path for a reason and working around that by pretending to be a reboot handler is a hack.

Quoted 6 lines

I understand the reason for the $subject patch, but it will not work, unfortunately.

For eMMC we need to manage emergency reboots too. The fiddling with GPIOs isn't a "cleanup", but tries to move the eMMC into a clean reset state.

That is by definition a "cleanup", even if the cleanup is really important.

Quoted 4 lines

This is needed on some platforms with broken bootloaders (ROM code), that is expecting the eMMC to always start in a clean reset state.

I understand the reason, I don't agree with the method used to get the result.

Quoted 2 lines

So, we need both parts, as was discussed here [1] too.

In this thread I see a lot of discussion about the priority of the handler. You want this to run before any real reboot handlers are run. Luckily for you, all reboot "notifiers" are run before any "handlers" are run. So if you register as a "notifier" as this patch does, you will be run first, no super high priority settings needed.

The real issue is you want to be called even in the emergency_restart() path, which is fine. But from the docs for that function this type of restart is done:

Quoted 1 line

Without shutting down any hardware

So we have two options:

1. Add a new notifier list that *does* get called in the emergency_restart() path. Then register this driver with with that.

2. Remove emergency_restart() from the kernel. It only has a couple of callers, and most of those callers look like they should instead be using hw_protection_reboot() or panic(). That way all reboot paths activate the reboot notifiers. Kinda wondering why you think you need to handle the emergency_restart() case at all, will even be a thing on your hardware, i.e. is this a real problem at all?

Having this driver claim to be a real reboot handler to sneak around doing one of the above is preventing some cleanup I am working on. So if either of the above two options work for you just let me know, I'll help out in implementing them for you.

Thanks,
Andrew

Quoted 35 lines

Kind regards
Uffe

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

Quoted 29 lines
---
  drivers/mmc/core/pwrseq_emmc.c | 8 +-------
  1 file changed, 1 insertion(+), 7 deletions(-)
drivers/mmc/core/pwrseq_emmc.c +1 -8
index 3b6d69cefb4eb..d5045fd1a02c1 100644
@@ -70,14 +70,8 @@ static int mmc_pwrseq_emmc_probe(struct platform_device *pdev)
                 return PTR_ERR(pwrseq->reset_gpio);
 
         if (!gpiod_cansleep(pwrseq->reset_gpio)) {
-               /*
-                * register reset handler to ensure emmc reset also from
-                * emergency_reboot(), priority 255 is the highest priority
-                * so it will be executed before any system reboot handler.
-                */
                 pwrseq->reset_nb.notifier_call = mmc_pwrseq_emmc_reset_nb;
-               pwrseq->reset_nb.priority = 255;
-               register_restart_handler(&pwrseq->reset_nb);
+               register_reboot_notifier(&pwrseq->reset_nb);
         } else {
                 dev_notice(dev, "EMMC reset pin tied to a sleepy GPIO driver; reset on emergency-reboot disabled\n");
         }
--

2.39.2

#3 lore raw

+ Oleksij

Quoted 36 lines · Andrew Davis
Quoted 14 lines · Ulf Hansson
Quoted 7 lines · Andrew Davis

This driver registers itself as a reboot handler, which means it claims it can reboot the system. It does this so it is called during the system reboot sequence. The correct way to be notified during the reboot sequence is to register a notifier with register_reboot_notifier(). Do this here.

Quoted 5 lines

Note this will be called during normal reboots but not emergency reboots. This is the expected behavior, emergency reboot means emergency, not go do some cleanup with emmc pins.. The reboot notifiers are intentionally not called in the emergency path for a reason and working around that by pretending to be a reboot handler is a hack.

Quoted 6 lines

I understand the reason for the $subject patch, but it will not work, unfortunately.

For eMMC we need to manage emergency reboots too. The fiddling with GPIOs isn't a "cleanup", but tries to move the eMMC into a clean reset state.

That is by definition a "cleanup", even if the cleanup is really important.

Quoted 4 lines

This is needed on some platforms with broken bootloaders (ROM code), that is expecting the eMMC to always start in a clean reset state.

I understand the reason, I don't agree with the method used to get the result.

Quoted 2 lines

So, we need both parts, as was discussed here [1] too.

Quoted 6 lines

In this thread I see a lot of discussion about the priority of the handler. You want this to run before any real reboot handlers are run. Luckily for you, all reboot "notifiers" are run before any "handlers" are run. So if you register as a "notifier" as this patch does, you will be run first, no super high priority settings needed.

Right, I didn't say the solution we use for mmc is perfect, but it was the easiest solution at hand at the introduction.

Quoted 13 lines

The real issue is you want to be called even in the emergency_restart() path, which is fine. But from the docs for that function this type of restart is done:

Quoted 1 line

Without shutting down any hardware

So we have two options:

1. Add a new notifier list that *does* get called in the emergency_restart() path. Then register this driver with with that.

Quoted 7 lines

2. Remove emergency_restart() from the kernel. It only has a couple of callers, and most of those callers look like they should instead be using hw_protection_reboot() or panic(). That way all reboot paths activate the reboot notifiers. Kinda wondering why you think you need to handle the emergency_restart() case at all, will even be a thing on your hardware, i.e. is this a real problem at all?

The "emergency reset" is needed, due to broken bootloaders, as I pointed out earlier.

That said, I don't have any strong opinions around this, whatever option you pick to rework the code is fine by me. The important point is that we can continue to support the use cases we need for MMC.

BTW, there was a recent related discussion [1] that you may want to catch up with too, before you start doing the restructuring of the restart/reboot code. See the link below.

Quoted 5 lines

Having this driver claim to be a real reboot handler to sneak around doing one of the above is preventing some cleanup I am working on. So if either of the above two options work for you just let me know, I'll help out in implementing them for you.

That would be great, thanks!

Quoted 3 lines

Thanks,
Andrew

Kind regards
Uffe

[1]
PATCH v1 0/3] introduce priority-based shutdown support]
https://lore.kernel.org/lkml/2023112520-paper-image-ef5d@gregkh/T/#mb45749c3bc9b89caecfeca6e66da8721d920191b

#4 lore raw
In reply to Andrew Davis
Quoted 11 lines · Andrew Davis

This driver registers itself as a reboot handler, which means it claims it can reboot the system. It does this so it is called during the system reboot sequence. The correct way to be notified during the reboot sequence is to register a notifier with register_reboot_notifier(). Do this here.

Note this will be called during normal reboots but not emergency reboots. This is the expected behavior, emergency reboot means emergency, not go do some cleanup with emmc pins.. The reboot notifiers are intentionally not called in the emergency path for a reason and working around that by pretending to be a reboot handler is a hack.

Well, I'm the author of this 'hack' and unfortunately there was no other way to make emergency reboot working on boards requiring the eMMC pwrseq. IIRC this has been already discussed and the conclusion was to accept the hack with the comments explaining the problem.

Quoted 15 lines
---
  drivers/mmc/core/pwrseq_emmc.c | 8 +-------
  1 file changed, 1 insertion(+), 7 deletions(-)
drivers/mmc/core/pwrseq_emmc.c +0 -2
index 3b6d69cefb4eb..d5045fd1a02c1 100644
@@ -70,14 +70,8 @@ static int mmc_pwrseq_emmc_probe(struct platform_device *pdev)
  		return PTR_ERR(pwrseq->reset_gpio);
 
  	if (!gpiod_cansleep(pwrseq->reset_gpio)) {
-		/*
-		 * register reset handler to ensure emmc reset also from

drivers/mmc/core/pwrseq_emmc.c static int mmc_pwrseq_emmc_probe(struct platform_device *pdev) · line 77

Quoted 10 lines
drivers/mmc/core/pwrseq_emmc.c +1 -5
-		 * emergency_reboot(), priority 255 is the highest priority
-		 * so it will be executed before any system reboot handler.
-		 */
  		pwrseq->reset_nb.notifier_call = mmc_pwrseq_emmc_reset_nb;
-		pwrseq->reset_nb.priority = 255;
-		register_restart_handler(&pwrseq->reset_nb);
+		register_reboot_notifier(&pwrseq->reset_nb);
  	} else {
  		dev_notice(dev, "EMMC reset pin tied to a sleepy GPIO driver; reset on emergency-reboot disabled\n");
  	}

Best regards

Signature · 2 lines
Marek Szyprowski, PhD
Samsung R&D Institute Poland