lkml.org 
[lkml]   [2022]   [Dec]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH] drivers/cros_ec: Handle CrOS EC panics
On Wed, Dec 21, 2022 at 1:07 PM Guenter Roeck <groeck@google.com> wrote:
>
> On Wed, Dec 21, 2022 at 10:56 AM Rob Barnes <robbarnes@google.com> wrote:
> >
> > Add handler for CrOS EC panic events. When a panic is reported,
> > poll EC log then force an orderly shutdown.
> >
> > This will preserve the EC log leading up to the crash.
> >
> > Signed-off-by: Rob Barnes <robbarnes@google.com>
> > ---
> > drivers/platform/chrome/cros_ec_debugfs.c | 24 +++++++++++++++++++++
> > drivers/platform/chrome/cros_ec_lpc.c | 10 +++++++++
> > include/linux/platform_data/cros_ec_proto.h | 9 ++++++++
> > 3 files changed, 43 insertions(+)
> >
> > diff --git a/drivers/platform/chrome/cros_ec_debugfs.c b/drivers/platform/chrome/cros_ec_debugfs.c
> > index 21d973fc6be2..31637a4e4cf9 100644
> > --- a/drivers/platform/chrome/cros_ec_debugfs.c
> > +++ b/drivers/platform/chrome/cros_ec_debugfs.c
> > @@ -49,6 +49,7 @@ struct cros_ec_debugfs {
> > struct delayed_work log_poll_work;
> > /* EC panicinfo */
> > struct debugfs_blob_wrapper panicinfo_blob;
> > + struct notifier_block notifier_panic;
> > };
> >
> > /*
> > @@ -437,6 +438,23 @@ static int cros_ec_create_panicinfo(struct cros_ec_debugfs *debug_info)
> > return ret;
> > }
> >
> > +static int cros_ec_debugfs_panic_event(struct notifier_block *nb,
> > + unsigned long queued_during_suspend,
> > + void *_notify)
> > +{
> > + struct cros_ec_debugfs *debug_info =
> > + container_of(nb, struct cros_ec_debugfs, notifier_panic);
> > +
> > + if (debug_info->log_buffer.buf) {
> > + /* Force log poll work to run immediately */
> > + mod_delayed_work(debug_info->log_poll_work.wq, &debug_info->log_poll_work, 0);
> > + /* Block until log poll work finishes */
> > + flush_delayed_work(&debug_info->log_poll_work);
> > + }
> > +
> > + return NOTIFY_DONE;
> > +}
> > +
> > static int cros_ec_debugfs_probe(struct platform_device *pd)
> > {
> > struct cros_ec_dev *ec = dev_get_drvdata(pd->dev.parent);
> > @@ -473,6 +491,12 @@ static int cros_ec_debugfs_probe(struct platform_device *pd)
> > debugfs_create_u16("suspend_timeout_ms", 0664, debug_info->dir,
> > &ec->ec_dev->suspend_timeout_ms);
> >
> > + debug_info->notifier_panic.notifier_call = cros_ec_debugfs_panic_event;
> > + ret = blocking_notifier_chain_register(&ec->ec_dev->panic_notifier,
> > + &debug_info->notifier_panic);
> > + if (ret)
> > + goto remove_debugfs;
> > +
> > ec->debug_info = debug_info;
> >
> > dev_set_drvdata(&pd->dev, ec);
> > diff --git a/drivers/platform/chrome/cros_ec_lpc.c b/drivers/platform/chrome/cros_ec_lpc.c
> > index 7fc8f82280ac..21958c3b0c28 100644
> > --- a/drivers/platform/chrome/cros_ec_lpc.c
> > +++ b/drivers/platform/chrome/cros_ec_lpc.c
> > @@ -21,6 +21,7 @@
> > #include <linux/platform_data/cros_ec_proto.h>
> > #include <linux/platform_device.h>
> > #include <linux/printk.h>
> > +#include <linux/reboot.h>
> > #include <linux/suspend.h>
> >
> > #include "cros_ec.h"
> > @@ -332,6 +333,15 @@ static void cros_ec_lpc_acpi_notify(acpi_handle device, u32 value, void *data)
> >
> > if (value == ACPI_NOTIFY_DEVICE_WAKE)
> > pm_system_wakeup();
> > +
> > + if (value == ACPI_NOTIFY_CROS_EC_PANIC) {
> > + dev_err(ec_dev->dev,
> > + "CrOS EC Panic Reported. Shutdown is imminent!");
>
> dev_emerg() would seem to be more appropriate, given that we'll be
> crashing a second later.

Makes sense. Will do.

>
> > + blocking_notifier_call_chain(&ec_dev->panic_notifier, 0,
> > + ec_dev);
> > + /* Begin orderly shutdown. Force shutdown after 1 second. */
> > + hw_protection_shutdown("CrOS EC Panic", 1000);
>
> That seems to be the wrong API. From its description:
>
> * Initiate an emergency system shutdown in order to protect hardware from
> * further damage. Usage examples include a thermal protection or a voltage or
> * current regulator failures.
>
> This is an EC crash that apparently can not be handled gracefully.
> That has nothing to do with hardware protection. Why not just call
> panic() ?

I'm adding the ability for the EC to somewhat gracefully handle a
class of panics. The EC will enter "system safe mode" aka limp mode.
All tasks are disabled except for a few critical ones. The purpose of
this mode is to give the OS a couple of seconds to extract info about
the EC panic and shutdown as gracefully as possible. EC will still
force a reset after 2 seconds.

Since the EC is in a precarious, non deterministic state and the EC is
at least partially responsible for thermal and power regulation, it is
arguably a hardware protection event.

The primary motivation for allowing the OS to attempt an orderly
shutdown is to sync user data and EC panic data to storage. Will panic
sync storage? Or will that need to be a seperate call?

>
> > + }
>
> It troubles me that this comes last. That means the entire
> mkbp_event_supported handling executes first. What troubles me even
> more is that 'value' is, with the exception of
> ACPI_NOTIFY_DEVICE_WAKE, not checked at all. Presumably the EC sends
> some well defined value if it has an event to report. I don't think it
> is a good idea to depend on cros_ec_get_next_event() returning
> something useful after the EC crashed.

Could certainly move the ACPI_NOTIFY_CROS_EC_PANIC to the top. I think
'value' will be 0x80 when there's MKBP events. I will validate this
assumption.

> How is this handled today ? Does the EC just reset the AP ?

EC just immediately resets the AP and itself. System safe mode (see
go/ec-system-safe-mode) will allow the EC to run a few seconds after
some types of crashes.



>
> > }
> >
> > static int cros_ec_lpc_probe(struct platform_device *pdev)
> > diff --git a/include/linux/platform_data/cros_ec_proto.h b/include/linux/platform_data/cros_ec_proto.h
> > index e43107e0bee1..1c4487271836 100644
> > --- a/include/linux/platform_data/cros_ec_proto.h
> > +++ b/include/linux/platform_data/cros_ec_proto.h
> > @@ -41,6 +41,13 @@
> > #define EC_MAX_REQUEST_OVERHEAD 1
> > #define EC_MAX_RESPONSE_OVERHEAD 32
> >
> > +/*
> > + * EC panic is not covered by the standard (0-F) ACPI notify values.
> > + * Arbitrarily choosing B0 to notify ec panic, which is in the 84-BF
> > + * device specific ACPI notify range.
> > + */
> > +#define ACPI_NOTIFY_CROS_EC_PANIC 0xB0
> > +
> > /*
> > * Command interface between EC and AP, for LPC, I2C and SPI interfaces.
> > */
> > @@ -176,6 +183,8 @@ struct cros_ec_device {
> > /* The platform devices used by the mfd driver */
> > struct platform_device *ec;
> > struct platform_device *pd;
> > +
> > + struct blocking_notifier_head panic_notifier;
> > };
> >
> > /**
> > --
> > 2.39.0.314.g84b9a713c41-goog
> >

\
 
 \ /
  Last update: 2023-03-26 23:17    [W:0.056 / U:1.264 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site