When enabling interrupts fails, the interrupt enable bit remains set
in i8042_ctr. Later writes of i8042_ctr to the hardware could
accidentally retry enabling interrupts. Clear the bit on failure.
Signed-off-by: Markus Armbruster <[email protected]>
Acked-by: Steven Rostedt <[email protected]>
---
Some time ago Steven Rostedt and I went over this changeset:
commit de9ce703c6b807b1dfef5942df4f2fadd0fdb67a
Author: Dmitry Torokhov <[email protected]>
Date: Sun Sep 10 21:57:21 2006 -0400
Input: i8042 - get rid of polling timer
Remove polling timer that was used to detect keybord/mice hotplug and
register both IRQs right away instead of waiting for a driver to
attach to a port.
Signed-off-by: Dmitry Torokhov <[email protected]>
Steven pointed out to me that it changes behavior when enabling IRQ
fails.
The old code enabled IRQs this way:
i8042_ctr |= port->irqen;
if (i8042_command(&i8042_ctr, I8042_CMD_CTL_WCTR)) {
i8042_ctr &= ~port->irqen;
return -1;
}
i8042_ctr shadows the 8042's CTR. So, when enabling fails, the bit is
cleared in the shadow.
The new code does not clear the bit on the error path:
static int i8042_enable_kbd_port(void)
{
i8042_ctr &= ~I8042_CTR_KBDDIS;
i8042_ctr |= I8042_CTR_KBDINT;
if (i8042_command(&i8042_ctr, I8042_CMD_CTL_WCTR)) {
printk(KERN_ERR "i8042.c: Failed to enable KBD port.\n");
return -EIO;
}
return 0;
}
Same for i8042_enable_aux_port().
This leads to the question whether there are later writes of i8042_ctr
(possibly with other bits altered) to the hardware, which could
accidentally retry enabling interrupts.
I believe this possible, but unlikely (perhaps not so unlikely on
virtual machines). Scenarios involve enable succeeding the first
time, failing the second time, and succeeding the third time. I can
provide details, but the point I'd like to make is not that this is
broken (although it is, strictly speaking), but that it is not
obviously correct where it easily could be: just clear the interrupt
enable bits when writing them to the hardware failed, like the old
code did.
diff --git a/drivers/input/serio/i8042.c b/drivers/input/serio/i8042.c
index db9cca3..71a7e39 100644
--- a/drivers/input/serio/i8042.c
+++ b/drivers/input/serio/i8042.c
@@ -385,6 +385,7 @@ static int i8042_enable_kbd_port(void)
i8042_ctr |= I8042_CTR_KBDINT;
if (i8042_command(&i8042_ctr, I8042_CMD_CTL_WCTR)) {
+ i8042_ctr &= ~I8042_CTR_KBDINT;
printk(KERN_ERR "i8042.c: Failed to enable KBD port.\n");
return -EIO;
}
@@ -402,6 +403,7 @@ static int i8042_enable_aux_port(void)
i8042_ctr |= I8042_CTR_AUXINT;
if (i8042_command(&i8042_ctr, I8042_CMD_CTL_WCTR)) {
+ i8042_ctr &= ~I8042_CTR_AUXINT;
printk(KERN_ERR "i8042.c: Failed to enable AUX port.\n");
return -EIO;
}
--
On Mon, 10 Sep 2007, Markus Armbruster wrote:
>
> I believe this possible, but unlikely (perhaps not so unlikely on
> virtual machines). Scenarios involve enable succeeding the first
> time, failing the second time, and succeeding the third time. I can
> provide details, but the point I'd like to make is not that this is
> broken (although it is, strictly speaking), but that it is not
> obviously correct where it easily could be: just clear the interrupt
> enable bits when writing them to the hardware failed, like the old
> code did.
>
I also want to stress that this is more of a clean up for "technically
correct" code than a bug fix. This bug probably would never happen on
baremetal unless it was running on broken hardware.
BUT!!!
With more and more systems going to a virtual environment, having a bug or
some other anomaly can trigger the error that this patch prevents. The
patch will also trigger a print in the case of running on a buggy virtual
machine, which would help out the developers of that virtual machine to
fix their code.
Please apply.
-- Steve
Hi Steven, Markus,
On 9/10/07, Steven Rostedt <[email protected]> wrote:
>
> --
> On Mon, 10 Sep 2007, Markus Armbruster wrote:
> >
> > I believe this possible, but unlikely (perhaps not so unlikely on
> > virtual machines). Scenarios involve enable succeeding the first
> > time, failing the second time, and succeeding the third time. I can
> > provide details, but the point I'd like to make is not that this is
> > broken (although it is, strictly speaking), but that it is not
> > obviously correct where it easily could be: just clear the interrupt
> > enable bits when writing them to the hardware failed, like the old
> > code did.
> >
>
> I also want to stress that this is more of a clean up for "technically
> correct" code than a bug fix. This bug probably would never happen on
> baremetal unless it was running on broken hardware.
>
> BUT!!!
>
> With more and more systems going to a virtual environment, having a bug or
> some other anomaly can trigger the error that this patch prevents. The
> patch will also trigger a print in the case of running on a buggy virtual
> machine, which would help out the developers of that virtual machine to
> fix their code.
>
The patch is in my tree and will be merged in the next window.
--
Dmitry