On Thu, Sep 01, 2022 at 02:46:48PM -0300, Guilherme G. Piccoli wrote:
> On 01/09/2022 13:44, Greg KH wrote:
> > [...]
> >>> How are we supposed to know this here?
> >>>
> >>
> >> Reading the code?
> >> Or you mean, in the commit description this should be mentioned?
> >
> > Yes, and in the comment as this type of call is very rare and should
> > almost never be used.
>
> OK, I can add that, for sure.
>
>
> >> [...]
> >> I don't think it is so simple - we are in the panic path.
> >
> > Great, then the lock doesn't matter :)
> >
> >> So, imagine the lock was taken in CPU0, where GSMI is doing some
> >> operation. During that operation, CPU1 panics!
> >>
> >> When that happens, panic() executes in CPU1, disabling CPU0 through
> >> "strong" mechanisms (NMI). So, CPU0 had the lock, it is now off, and
> >> when CPU1 goes through the panic notifiers, it'll eventually wait
> >> forever for this lock in the GSMI handler, unless we have this patch
> >> that would prevent the handler to run in such case.
> >> Makes sense?
> >
> > I'm trying to say "if you are in panic, never grab the lock in the first
> > place". So change the place when you grab the lock, not here.
> >
>
> Evan, any comment here?
> I think the patch is still well suited for this case. Suggestions on how
> to improve it are welcome, of course.
>
> I honestly didn't understand exactly what you're suggesting Greg...
> Mind clarifying?
Something like this totally untested code:
diff --git a/drivers/firmware/google/gsmi.c b/drivers/firmware/google/gsmi.c
index adaa492c3d2d..6ad41b22671c 100644
--- a/drivers/firmware/google/gsmi.c
+++ b/drivers/firmware/google/gsmi.c
@@ -19,6 +19,7 @@
#include <linux/dma-mapping.h>
#include <linux/fs.h>
#include <linux/slab.h>
+#include <linux/panic.h>
#include <linux/panic_notifier.h>
#include <linux/ioctl.h>
#include <linux/acpi.h>
@@ -611,6 +612,11 @@ static const struct attribute *gsmi_attrs[] = {
NULL,
};
+static bool panic_in_progress(void)
+{
+ return unlikely(atomic_read(&panic_cpu) != PANIC_CPU_INVALID);
+}
+
static int gsmi_shutdown_reason(int reason)
{
struct gsmi_log_entry_type_1 entry = {
@@ -629,7 +635,8 @@ static int gsmi_shutdown_reason(int reason)
if (saved_reason & (1 << reason))
return 0;
- spin_lock_irqsave(&gsmi_dev.lock, flags);
+ if (!panic_in_progress())
+ spin_lock_irqsave(&gsmi_dev.lock, flags);
saved_reason |= (1 << reason);
@@ -644,7 +651,8 @@ static int gsmi_shutdown_reason(int reason)
rc = gsmi_exec(GSMI_CALLBACK, GSMI_CMD_SET_EVENT_LOG);
- spin_unlock_irqrestore(&gsmi_dev.lock, flags);
+ if (!panic_in_progress())
+ spin_unlock_irqrestore(&gsmi_dev.lock, flags);
if (rc < 0)
printk(KERN_ERR "gsmi: Log Shutdown Reason failed\n");
That being said, are you sure spinlocks are still held in the panic
notifier? What about the call to bust_spinlocks() that is called in
panic() already? Wouldn't that have already dropped whatever you were
worried about here?
thanks,
greg k-h
