On Tuesday 24 March 2009, Jesse Barnes wrote:
> On Tue, 24 Mar 2009 11:57:19 +1100
> Benjamin Herrenschmidt <benh@xxxxxxxxxxxxxxxxxxx> wrote:
>
> > On Mon, 2009-03-23 at 15:23 -0700, Jesse Barnes wrote:
> > > The thing I didn't like was that it made the radeon driver use an
> > > internal interface; I'd really prefer a proper return value from
> > > pci_set_power_state, which in turn means auditing all its current
> > > callers. But that doesn't seem worth it unless we see other drivers
> > > needing something similar...
> > >
> > > And if we did go with something like your first patch, I'd still
> > > rather see the timeout done in the driver, rather than having the
> > > attempts & delay included in the function...
> >
> > So what ? The driver would call pci_set_power_state() until it stops
> > failing ?
>
> Yeah, that's what I had in mind.
>
> > I'm not too fan of that, because it will change the access pattern
> > to the chip:
> >
> > - write PM to 2
> > - short delay
> > - read PM, see 0, return error
> > - driver does big delay
> > - write PM to 2
> > - short delay
> > - read PM ....
> >
> > vs. the current sequence which is
> >
> > - write PM to 2
> > - long delay
> > - read PM, be happy
> >
> > Which -seems- to be pretty much what happens in practice, though on
> > that chip, I don't know for sure about others.
> >
> > I'm worried of the possible side effects of the first sequence that
> > you propose since it would do 2 things potentially confusing to the
> > HW:
> >
> > - read PM after a short delay... it -should- be harmless but you know
> > HW as well as I do ...
> >
> > - write PM to 2 a second time after the long delay. Again, it
> > -should- be harmless since the chip at this stage should already be
> > in D2 state but god knows how the HW will react.
> >
> > I'm especially worried about the later in fact. Maybe we can minimize
> > it by having pci_set_power_state() dbl check the content of the PM
> > reg before writing to it...
>
> Honestly I'm not too happy about any of the approaches, but yeah I see
> your point. The main thing is to prevent any config space access for
> a specified time after the first D-state transition, which I think we
> do correctly in the core. Beyond that, we just have to make sure the
> core state is updated correctly; Rafael's first patch does that
> correctly I think.
In fact I have yet another idea. If we use the "retransmission with exponential
backoff" algorithm in pci_raw_set_power_state(), we won't have to add any
extra parameters to pci_set_power_state() and the radeon case will be covered
automatically. That should also cover any other devices having similar
problems IMO.
A patch to implement that is appended, please tell me what you think.
Thanks,
Rafael
---
drivers/pci/pci.c | 50 ++++++++++++++++++++++++++++++++++++++++----------
include/linux/pci.h | 1 +
2 files changed, 41 insertions(+), 10 deletions(-)
Index: linux-2.6/drivers/pci/pci.c
===================================================================
--- linux-2.6.orig/drivers/pci/pci.c
+++ linux-2.6/drivers/pci/pci.c
@@ -436,8 +436,10 @@ static inline int platform_pci_sleep_wak
*/
static int pci_raw_set_power_state(struct pci_dev *dev, pci_power_t state)
{
- u16 pmcsr;
+ u16 pmcsr, pmcsr_r;
+ unsigned int delay;
bool need_restore = false;
+ int error = 0;
/* Check if we're already there */
if (dev->current_state == state)
@@ -488,17 +490,45 @@ static int pci_raw_set_power_state(struc
break;
}
- /* enter specified state */
- pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, pmcsr);
-
- /* Mandatory power management transition delays */
- /* see PCI PM 1.1 5.6.1 table 18 */
+ /*
+ * Mandatory power management transition delays, in microseconds
+ * (see PCI PM 1.1 5.6.1 table 18).
+ */
if (state == PCI_D3hot || dev->current_state == PCI_D3hot)
- msleep(pci_pm_d3_delay);
+ delay = pci_pm_d3_delay * 1000;
else if (state == PCI_D2 || dev->current_state == PCI_D2)
- udelay(PCI_PM_D2_DELAY);
+ delay = PCI_PM_D2_DELAY;
+ else
+ delay = 0;
+
+ /*
+ * Write the new value to the control register, wait as long as needed
+ * and check if the value read back from the register is the same as
+ * the written one. If not, repeat with exponential backoff.
+ */
+ do {
+ pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, pmcsr);
+ if (delay) {
+ if (delay < 1000)
+ udelay(delay);
+ else
+ msleep(DIV_ROUND_UP(delay, 1000));
+ delay <<= 1;
+ }
+ pci_read_config_word(dev, dev->pm_cap + PCI_PM_CTRL, &pmcsr_r);
+ if (pmcsr == pmcsr_r) {
+ dev->current_state = state;
+ break;
+ }
+ } while (delay && delay <= PCI_PM_MAX_DELAY);
- dev->current_state = state;
+ if (pmcsr != pmcsr_r) {
+ dev->current_state = (pmcsr_r & PCI_PM_CTRL_STATE_MASK);
+ dev_warn(&dev->dev,
+ "failed to set power state to D%d, currently in D%d\n",
+ state, dev->current_state);
+ error = -ENODEV;
+ }
/* According to section 5.4.1 of the "PCI BUS POWER MANAGEMENT
* INTERFACE SPECIFICATION, REV. 1.2", a device transitioning
@@ -518,7 +548,7 @@ static int pci_raw_set_power_state(struc
if (dev->bus->self)
pcie_aspm_pm_state_change(dev->bus->self);
- return 0;
+ return error;
}
/**
Index: linux-2.6/include/linux/pci.h
===================================================================
--- linux-2.6.orig/include/linux/pci.h
+++ linux-2.6/include/linux/pci.h
@@ -117,6 +117,7 @@ typedef int __bitwise pci_power_t;
#define PCI_UNKNOWN ((pci_power_t __force) 5)
#define PCI_POWER_ERROR ((pci_power_t __force) -1)
+#define PCI_PM_MAX_DELAY 2000000
#define PCI_PM_D2_DELAY 200
#define PCI_PM_D3_WAIT 10
#define PCI_PM_BUS_WAIT 50
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
