The PMC driver performs a 32-bit read on the sleep s0 residency counter,
followed by a hard-coded multiplication to convert into microseconds.
The maximum value this counter could have would be 0xffffffff*0x64
microseconds, which by my calculations is about 4.9 days. This is well
within a reasonable time period to observe an overflow.
Usermode consumers watching slp_s0_residency_usec need to be aware of
overflows, but have no idea what the maximum value of this counter is,
given the hardcoded multiply of a 32-bit value by
SPT_PMC_SLP_S0_RES_COUNTER_STEP.
Expose a slp_s0_residency_usec_range to usermode as well, which returns
the maximum value this counter could have. Consumers can use this to
manage rollovers.
Signed-off-by: Evan Green <[email protected]>
---
Note: I also looked at a similar bit of functionality in
intel_pmc_s0ix_counter_read(), but noticed it's doing a 64-bit register
access. Is the counter being read here in pmc_core_dev_state_get()
(weird name btw) actually 64-bits long? If so, we can abandon this change
and just create a fix to return the full extended value.
---
drivers/platform/x86/intel_pmc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)
diff --git a/drivers/platform/x86/intel_pmc_core.c b/drivers/platform/x86/intel_pmc_core.c
index f2c621b55f49..bec54be9be93 100644
--- a/drivers/platform/x86/intel_pmc_core.c
+++ b/drivers/platform/x86/intel_pmc_core.c
@@ -396,6 +396,14 @@ static int pmc_core_dev_state_get(void *data, u64 *val)
DEFINE_DEBUGFS_ATTRIBUTE(pmc_core_dev_state, pmc_core_dev_state_get, NULL, "%llu\n");
+static int pmc_core_slp_s0_range_get(void *data, u64 *val)
+{
+ *val = pmc_core_adjust_slp_s0_step(0xffffffff);
+ return 0;
+}
+
+DEFINE_DEBUGFS_ATTRIBUTE(pmc_core_slp_s0_range, pmc_core_slp_s0_range_get, NULL, "%llu\n");
+
static int pmc_core_check_read_lock_bit(void)
{
struct pmc_dev *pmcdev = &pmc;
@@ -764,6 +772,9 @@ static int pmc_core_dbgfs_register(struct pmc_dev *pmcdev)
debugfs_create_file("slp_s0_residency_usec", 0444, dir, pmcdev,
&pmc_core_dev_state);
+ debugfs_create_file("slp_s0_residency_usec_range", 0444, dir, pmcdev,
+ &pmc_core_slp_s0_range);
+
debugfs_create_file("pch_ip_power_gating_status", 0444, dir, pmcdev,
&pmc_core_ppfear_fops);
--
2.20.1
On Mon, Apr 01, 2019 at 11:05:04AM -0700, Evan Green wrote:
> The PMC driver performs a 32-bit read on the sleep s0 residency counter,
> followed by a hard-coded multiplication to convert into microseconds.
> The maximum value this counter could have would be 0xffffffff*0x64
> microseconds, which by my calculations is about 4.9 days. This is well
> within a reasonable time period to observe an overflow.
>
> Usermode consumers watching slp_s0_residency_usec need to be aware of
> overflows, but have no idea what the maximum value of this counter is,
> given the hardcoded multiply of a 32-bit value by
> SPT_PMC_SLP_S0_RES_COUNTER_STEP.
This register is a 32 bit register untill ICL generation and a recent patch
from Rajat fixed the overflow https://patchwork.kernel.org/patch/10816103/
already so i am not sure how this will help userspace. I think the userspace
can still take care of any overflow concerns based on the information
available about this register in EDS so i feel exposing a new debugfs entry
just for the sake of knowing range is probably not needed.
>
> Expose a slp_s0_residency_usec_range to usermode as well, which returns
> the maximum value this counter could have. Consumers can use this to
> manage rollovers.
>
> Signed-off-by: Evan Green <[email protected]>
>
> ---
>
> Note: I also looked at a similar bit of functionality in
> intel_pmc_s0ix_counter_read(), but noticed it's doing a 64-bit register
> access. Is the counter being read here in pmc_core_dev_state_get()
> (weird name btw) actually 64-bits long? If so, we can abandon this change
> and just create a fix to return the full extended value.
>
> ---
> drivers/platform/x86/intel_pmc_core.c | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/drivers/platform/x86/intel_pmc_core.c b/drivers/platform/x86/intel_pmc_core.c
> index f2c621b55f49..bec54be9be93 100644
> --- a/drivers/platform/x86/intel_pmc_core.c
> +++ b/drivers/platform/x86/intel_pmc_core.c
> @@ -396,6 +396,14 @@ static int pmc_core_dev_state_get(void *data, u64 *val)
>
> DEFINE_DEBUGFS_ATTRIBUTE(pmc_core_dev_state, pmc_core_dev_state_get, NULL, "%llu\n");
>
> +static int pmc_core_slp_s0_range_get(void *data, u64 *val)
> +{
> + *val = pmc_core_adjust_slp_s0_step(0xffffffff);
> + return 0;
> +}
> +
> +DEFINE_DEBUGFS_ATTRIBUTE(pmc_core_slp_s0_range, pmc_core_slp_s0_range_get, NULL, "%llu\n");
> +
> static int pmc_core_check_read_lock_bit(void)
> {
> struct pmc_dev *pmcdev = &pmc;
> @@ -764,6 +772,9 @@ static int pmc_core_dbgfs_register(struct pmc_dev *pmcdev)
> debugfs_create_file("slp_s0_residency_usec", 0444, dir, pmcdev,
> &pmc_core_dev_state);
>
> + debugfs_create_file("slp_s0_residency_usec_range", 0444, dir, pmcdev,
> + &pmc_core_slp_s0_range);
> +
> debugfs_create_file("pch_ip_power_gating_status", 0444, dir, pmcdev,
> &pmc_core_ppfear_fops);
>
> --
> 2.20.1
>
--
Best Regards,
Rajneesh
On Fri, Apr 5, 2019 at 12:28 AM Rajneesh Bhardwaj
<[email protected]> wrote:
>
> On Mon, Apr 01, 2019 at 11:05:04AM -0700, Evan Green wrote:
> > The PMC driver performs a 32-bit read on the sleep s0 residency counter,
> > followed by a hard-coded multiplication to convert into microseconds.
> > The maximum value this counter could have would be 0xffffffff*0x64
> > microseconds, which by my calculations is about 4.9 days. This is well
> > within a reasonable time period to observe an overflow.
> >
> > Usermode consumers watching slp_s0_residency_usec need to be aware of
> > overflows, but have no idea what the maximum value of this counter is,
> > given the hardcoded multiply of a 32-bit value by
> > SPT_PMC_SLP_S0_RES_COUNTER_STEP.
>
> This register is a 32 bit register untill ICL generation and a recent patch
> from Rajat fixed the overflow https://patchwork.kernel.org/patch/10816103/
> already so i am not sure how this will help userspace. I think the userspace
> can still take care of any overflow concerns based on the information
> available about this register in EDS so i feel exposing a new debugfs entry
> just for the sake of knowing range is probably not needed.
So you don't anticipate reusing this driver for other models where the
factor of 100 might change, or the counter might be a different width?
We can hardcode the multiply in usermode in order to detect when we've
exceeded the 4.9 day rollover range, but it seemed brittle, and would
break if this hardware implementation detail ever changed.
-Evan
On Fri, Apr 5, 2019 at 8:52 PM Evan Green <[email protected]> wrote:
> On Fri, Apr 5, 2019 at 12:28 AM Rajneesh Bhardwaj
> <[email protected]> wrote:
> > On Mon, Apr 01, 2019 at 11:05:04AM -0700, Evan Green wrote:
> > This register is a 32 bit register untill ICL generation and a recent patch
> > from Rajat fixed the overflow https://patchwork.kernel.org/patch/10816103/
> > already so i am not sure how this will help userspace. I think the userspace
> > can still take care of any overflow concerns based on the information
> > available about this register in EDS so i feel exposing a new debugfs entry
> > just for the sake of knowing range is probably not needed.
>
> So you don't anticipate reusing this driver for other models where the
> factor of 100 might change, or the counter might be a different width?
> We can hardcode the multiply in usermode in order to detect when we've
> exceeded the 4.9 day rollover range, but it seemed brittle, and would
> break if this hardware implementation detail ever changed.
No user space component should ever rely on debugfs interface.
--
With Best Regards,
Andy Shevchenko