* [PATCH] driver core: fix e1000e ltr bug
@ 2021-06-29 8:21 YeeLi
2021-06-29 14:49 ` [Intel-wired-lan] " Neftin, Sasha
0 siblings, 1 reply; 4+ messages in thread
From: YeeLi @ 2021-06-29 8:21 UTC (permalink / raw)
To: jesse.brandeburg, anthony.l.nguyen, davem, kuba
Cc: intel-wired-lan, linux-kernel, seven.yi.lee
In e1000e driver, a PCIe-like device, the max snoop/no-snoop latency
is the upper limit.So, directly compare the size of lat_enc and
max_ltr_enc is incorrect.
In 1000Mbps, 0x8b9 < 0x1003, 189440 ns < 3145728 ns, correct.
In 100Mbps, 0xc3a < 0x1003, 1900544 ns < 3145728 ns, correct.
In 10Mbps, 0xe40 < 0x1003, 18874368 > 3145728, incorrect.
Decoded the lat_enc and max_ltr_enc before compare them is necessary.
Signed-off-by: YeeLi <seven.yi.lee@gmail.com>
---
drivers/net/ethernet/intel/e1000e/ich8lan.c | 23 ++++++++++++++++++++-
1 file changed, 22 insertions(+), 1 deletion(-)
diff --git a/drivers/net/ethernet/intel/e1000e/ich8lan.c b/drivers/net/ethernet/intel/e1000e/ich8lan.c
index 590ad110d383..3bff1b570b76 100644
--- a/drivers/net/ethernet/intel/e1000e/ich8lan.c
+++ b/drivers/net/ethernet/intel/e1000e/ich8lan.c
@@ -986,6 +986,27 @@ static s32 e1000_k1_workaround_lpt_lp(struct e1000_hw *hw, bool link)
return ret_val;
}
+static u32 convert_e1000e_ltr_scale(u32 val)
+{
+ if (val > 5)
+ return 0;
+
+ return 1U << (5 * val);
+}
+
+static u64 decoded_ltr(u32 val)
+{
+ u64 decoded_latency;
+ u32 value;
+ u32 scale;
+
+ value = val & 0x03FF;
+ scale = (val & 0x1C00) >> 10;
+ decoded_latency = value * convert_e1000e_ltr_scale(scale);
+
+ return decoded_latency;
+}
+
/**
* e1000_platform_pm_pch_lpt - Set platform power management values
* @hw: pointer to the HW structure
@@ -1059,7 +1080,7 @@ static s32 e1000_platform_pm_pch_lpt(struct e1000_hw *hw, bool link)
E1000_PCI_LTR_CAP_LPT + 2, &max_nosnoop);
max_ltr_enc = max_t(u16, max_snoop, max_nosnoop);
- if (lat_enc > max_ltr_enc)
+ if (decoded_ltr(lat_enc) > decoded_ltr(max_ltr_enc))
lat_enc = max_ltr_enc;
}
--
2.25.1
^ permalink raw reply related [flat|nested] 4+ messages in thread
* Re: [Intel-wired-lan] [PATCH] driver core: fix e1000e ltr bug
2021-06-29 8:21 [PATCH] driver core: fix e1000e ltr bug YeeLi
@ 2021-06-29 14:49 ` Neftin, Sasha
[not found] ` <CALX8JfQymbSmCP0xk0C-=v64__uaH=BR0UZjr2yRyLWVwm9dLQ@mail.gmail.com>
0 siblings, 1 reply; 4+ messages in thread
From: Neftin, Sasha @ 2021-06-29 14:49 UTC (permalink / raw)
To: YeeLi, jesse.brandeburg, anthony.l.nguyen, davem, kuba
Cc: intel-wired-lan, linux-kernel, Ruinskiy, Dima
On 6/29/2021 11:21, YeeLi wrote:
Yeeli,
> In e1000e driver, a PCIe-like device, the max snoop/no-snoop latency
> is the upper limit.So, directly compare the size of lat_enc and
> max_ltr_enc is incorrect.
>
why?
> In 1000Mbps, 0x8b9 < 0x1003, 189440 ns < 3145728 ns, correct.
>
> In 100Mbps, 0xc3a < 0x1003, 1900544 ns < 3145728 ns, correct.
>
> In 10Mbps, 0xe40 < 0x1003, 18874368 > 3145728, incorrect.
>
Platform LTR encoded is 0x1003 - right. It is meant 1048576ns x 3 =
3145738ns.
Now,
for 1000M: 0x08b9 => 185ns x 1024 = 189440ns (you are correct)
for 100M: 0x0c3a => 58ns x 32768 = 1900544ns (correct)
for 10M: 0x0e41 => 577ns x 32768 = 18907136ns (ok?)
18907136ns > 3145738ns, (latency encoded is great than maximum LTR
encoded by platform) - so, there is no point to wait more than platform
required, and lat_enc=max_ltr_enc. It is expected and we sent right
value to the power management controller.
What is the problem you try solve?
> Decoded the lat_enc and max_ltr_enc before compare them is necessary.
>
> Signed-off-by: YeeLi <seven.yi.lee@gmail.com>
> ---
> drivers/net/ethernet/intel/e1000e/ich8lan.c | 23 ++++++++++++++++++++-
> 1 file changed, 22 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/net/ethernet/intel/e1000e/ich8lan.c b/drivers/net/ethernet/intel/e1000e/ich8lan.c
> index 590ad110d383..3bff1b570b76 100644
> --- a/drivers/net/ethernet/intel/e1000e/ich8lan.c
> +++ b/drivers/net/ethernet/intel/e1000e/ich8lan.c
> @@ -986,6 +986,27 @@ static s32 e1000_k1_workaround_lpt_lp(struct e1000_hw *hw, bool link)
> return ret_val;
> }
>
> +static u32 convert_e1000e_ltr_scale(u32 val)
> +{
> + if (val > 5)
> + return 0;
> +
> + return 1U << (5 * val);
> +}
> +
> +static u64 decoded_ltr(u32 val)
> +{
> + u64 decoded_latency;
> + u32 value;
> + u32 scale;
> +
> + value = val & 0x03FF;
> + scale = (val & 0x1C00) >> 10;
> + decoded_latency = value * convert_e1000e_ltr_scale(scale);
> +
> + return decoded_latency;
> +}
> +
> /**
> * e1000_platform_pm_pch_lpt - Set platform power management values
> * @hw: pointer to the HW structure
> @@ -1059,7 +1080,7 @@ static s32 e1000_platform_pm_pch_lpt(struct e1000_hw *hw, bool link)
> E1000_PCI_LTR_CAP_LPT + 2, &max_nosnoop);
> max_ltr_enc = max_t(u16, max_snoop, max_nosnoop);
>
> - if (lat_enc > max_ltr_enc)
> + if (decoded_ltr(lat_enc) > decoded_ltr(max_ltr_enc))
> lat_enc = max_ltr_enc;
> }
>
>
sasha
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [Intel-wired-lan] [PATCH] driver core: fix e1000e ltr bug
[not found] ` <CALX8JfQymbSmCP0xk0C-=v64__uaH=BR0UZjr2yRyLWVwm9dLQ@mail.gmail.com>
@ 2021-06-30 6:13 ` Neftin, Sasha
[not found] ` <CALX8JfTOG7bXn7gP+4n1_3CfFFk6aL-RJDY03RCOeBSpUeTEPQ@mail.gmail.com>
0 siblings, 1 reply; 4+ messages in thread
From: Neftin, Sasha @ 2021-06-30 6:13 UTC (permalink / raw)
To: Yee Li
Cc: jesse.brandeburg, anthony.l.nguyen, davem, kuba, intel-wired-lan,
linux-kernel, Ruinskiy, Dima, Edri, Michael, Efrati, Nir
On 6/29/2021 20:33, Yee Li wrote:
>
> Yes, 18874368ns > 3145728ns.
> But, 0xe40 < 0x1003.
I got you. I would agree, direct comparison is error-prone. (10M is
impacted)
I would suggest do not use convert function. lat_en should rather
presented as lat_enc = scale x value:
Introduce two u16 variables, u16 lat_enc_d and u16 max_ltr_enc_d;
lat_enc_d = (lat_enc & 0x0x3ff) x (1U << 5*((max_ltr_enc & 0x1c00) >> 10))
max_ltr_enc_d = (max_ltr_enc & 0x0x3ff) x (1U << 5*((max_ltr_enc &
0x1c00) >> 10))
Then:
if (lat_enc_d > max_ltr_enc_d)
lat_enc = max_ltr_enc;
what do you think?
>
> So, the final lat_enc is 0xe40.
> (Latency encoded is less than maximum LTR encoded by platform)
>
> Neftin, Sasha <sasha.neftin@intel.com <mailto:sasha.neftin@intel.com>>
> 于 2021年6月29日周二 22:49写道:
>
> On 6/29/2021 11:21, YeeLi wrote:
> Yeeli,
> > In e1000e driver, a PCIe-like device, the max snoop/no-snoop latency
> > is the upper limit.So, directly compare the size of lat_enc and
> > max_ltr_enc is incorrect.
> >
> why?
> > In 1000Mbps, 0x8b9 < 0x1003, 189440 ns < 3145728 ns, correct.
> >
> > In 100Mbps, 0xc3a < 0x1003, 1900544 ns < 3145728 ns, correct.
> >
> > In 10Mbps, 0xe40 < 0x1003, 18874368 > 3145728, incorrect.
> >
> Platform LTR encoded is 0x1003 - right. It is meant 1048576ns x 3 =
> 3145738ns.
> Now,
> for 1000M: 0x08b9 => 185ns x 1024 = 189440ns (you are correct)
> for 100M: 0x0c3a => 58ns x 32768 = 1900544ns (correct)
> for 10M: 0x0e41 => 577ns x 32768 = 18907136ns (ok?)
> 18907136ns > 3145738ns, (latency encoded is great than maximum LTR
> encoded by platform) - so, there is no point to wait more than platform
> required, and lat_enc=max_ltr_enc. It is expected and we sent right
> value to the power management controller.
> What is the problem you try solve?
>
> > Decoded the lat_enc and max_ltr_enc before compare them is necessary.
> >
> > Signed-off-by: YeeLi <seven.yi.lee@gmail.com
> <mailto:seven.yi.lee@gmail.com>>
> > ---
> > drivers/net/ethernet/intel/e1000e/ich8lan.c | 23
> ++++++++++++++++++++-
> > 1 file changed, 22 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/net/ethernet/intel/e1000e/ich8lan.c
> b/drivers/net/ethernet/intel/e1000e/ich8lan.c
> > index 590ad110d383..3bff1b570b76 100644
> > --- a/drivers/net/ethernet/intel/e1000e/ich8lan.c
> > +++ b/drivers/net/ethernet/intel/e1000e/ich8lan.c
> > @@ -986,6 +986,27 @@ static s32 e1000_k1_workaround_lpt_lp(struct
> e1000_hw *hw, bool link)
> > return ret_val;
> > }
> >
> > +static u32 convert_e1000e_ltr_scale(u32 val)
> > +{
> > + if (val > 5)
> > + return 0;
> > +
> > + return 1U << (5 * val);
> > +}
> > +
> > +static u64 decoded_ltr(u32 val)
> > +{
> > + u64 decoded_latency;
> > + u32 value;
> > + u32 scale;
> > +
> > + value = val & 0x03FF;
> > + scale = (val & 0x1C00) >> 10;
> > + decoded_latency = value * convert_e1000e_ltr_scale(scale);
> > +
> > + return decoded_latency;
> > +}
> > +
> > /**
> > * e1000_platform_pm_pch_lpt - Set platform power management
> values
> > * @hw: pointer to the HW structure
> > @@ -1059,7 +1080,7 @@ static s32 e1000_platform_pm_pch_lpt(struct
> e1000_hw *hw, bool link)
> > E1000_PCI_LTR_CAP_LPT + 2,
> &max_nosnoop);
> > max_ltr_enc = max_t(u16, max_snoop, max_nosnoop);
> >
> > - if (lat_enc > max_ltr_enc)
> > + if (decoded_ltr(lat_enc) > decoded_ltr(max_ltr_enc))
> > lat_enc = max_ltr_enc;
> > }
> >
> >
> sasha
>
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [Intel-wired-lan] [PATCH] driver core: fix e1000e ltr bug
[not found] ` <CALX8JfTOG7bXn7gP+4n1_3CfFFk6aL-RJDY03RCOeBSpUeTEPQ@mail.gmail.com>
@ 2021-07-01 8:34 ` Neftin, Sasha
0 siblings, 0 replies; 4+ messages in thread
From: Neftin, Sasha @ 2021-07-01 8:34 UTC (permalink / raw)
To: Yee Li
Cc: jesse.brandeburg, anthony.l.nguyen, davem, kuba, intel-wired-lan,
linux-kernel, Ruinskiy, Dima, Edri, Michael, Efrati, Nir, Neftin,
Sasha
On 6/30/2021 09:26, Yee Li wrote:
> I agree with you.
> That's better code.
Thank Yee-Li for discovering this problem. I would suggest we (intel)
will process a patch to address this problem. Can I put you kindly as
"Suggested-by"?
>
> Neftin, Sasha <sasha.neftin@intel.com <mailto:sasha.neftin@intel.com>>
> 于2021年6月30日周三 下午2:13写道:
>
> On 6/29/2021 20:33, Yee Li wrote:
> >
> > Yes, 18874368ns > 3145728ns.
> > But, 0xe40 < 0x1003.
> I got you. I would agree, direct comparison is error-prone. (10M is
> impacted)
> I would suggest do not use convert function. lat_en should rather
> presented as lat_enc = scale x value:
> Introduce two u16 variables, u16 lat_enc_d and u16 max_ltr_enc_d;
>
> lat_enc_d = (lat_enc & 0x0x3ff) x (1U << 5*((max_ltr_enc & 0x1c00)
> >> 10))
>
> max_ltr_enc_d = (max_ltr_enc & 0x0x3ff) x (1U << 5*((max_ltr_enc &
> 0x1c00) >> 10))
>
> Then:
> if (lat_enc_d > max_ltr_enc_d)
> lat_enc = max_ltr_enc;
> what do you think?
>
> >
> > So, the final lat_enc is 0xe40.
> > (Latency encoded is less than maximum LTR encoded by platform)
> >
> > Neftin, Sasha <sasha.neftin@intel.com
> <mailto:sasha.neftin@intel.com> <mailto:sasha.neftin@intel.com
> <mailto:sasha.neftin@intel.com>>>
> > 于 2021年6月29日周二 22:49写道:
> >
> > On 6/29/2021 11:21, YeeLi wrote:
> > Yeeli,
> > > In e1000e driver, a PCIe-like device, the max
> snoop/no-snoop latency
> > > is the upper limit.So, directly compare the size of
> lat_enc and
> > > max_ltr_enc is incorrect.
> > >
> > why?
> > > In 1000Mbps, 0x8b9 < 0x1003, 189440 ns < 3145728 ns,
> correct.
> > >
> > > In 100Mbps, 0xc3a < 0x1003, 1900544 ns < 3145728 ns,
> correct.
> > >
> > > In 10Mbps, 0xe40 < 0x1003, 18874368 > 3145728, incorrect.
> > >
> > Platform LTR encoded is 0x1003 - right. It is meant 1048576ns
> x 3 =
> > 3145738ns.
> > Now,
> > for 1000M: 0x08b9 => 185ns x 1024 = 189440ns (you are correct)
> > for 100M: 0x0c3a => 58ns x 32768 = 1900544ns (correct)
> > for 10M: 0x0e41 => 577ns x 32768 = 18907136ns (ok?)
> > 18907136ns > 3145738ns, (latency encoded is great than
> maximum LTR
> > encoded by platform) - so, there is no point to wait more
> than platform
> > required, and lat_enc=max_ltr_enc. It is expected and we sent
> right
> > value to the power management controller.
> > What is the problem you try solve?
> >
> > > Decoded the lat_enc and max_ltr_enc before compare them is
> necessary.
> > >
> > > Signed-off-by: YeeLi <seven.yi.lee@gmail.com
> <mailto:seven.yi.lee@gmail.com>
> > <mailto:seven.yi.lee@gmail.com <mailto:seven.yi.lee@gmail.com>>>
> > > ---
> > > drivers/net/ethernet/intel/e1000e/ich8lan.c | 23
> > ++++++++++++++++++++-
> > > 1 file changed, 22 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/net/ethernet/intel/e1000e/ich8lan.c
> > b/drivers/net/ethernet/intel/e1000e/ich8lan.c
> > > index 590ad110d383..3bff1b570b76 100644
> > > --- a/drivers/net/ethernet/intel/e1000e/ich8lan.c
> > > +++ b/drivers/net/ethernet/intel/e1000e/ich8lan.c
> > > @@ -986,6 +986,27 @@ static s32
> e1000_k1_workaround_lpt_lp(struct
> > e1000_hw *hw, bool link)
> > > return ret_val;
> > > }
> > >
> > > +static u32 convert_e1000e_ltr_scale(u32 val)
> > > +{
> > > + if (val > 5)
> > > + return 0;
> > > +
> > > + return 1U << (5 * val);
> > > +}
> > > +
> > > +static u64 decoded_ltr(u32 val)
> > > +{
> > > + u64 decoded_latency;
> > > + u32 value;
> > > + u32 scale;
> > > +
> > > + value = val & 0x03FF;
> > > + scale = (val & 0x1C00) >> 10;
> > > + decoded_latency = value *
> convert_e1000e_ltr_scale(scale);
> > > +
> > > + return decoded_latency;
> > > +}
> > > +
> > > /**
> > > * e1000_platform_pm_pch_lpt - Set platform power
> management
> > values
> > > * @hw: pointer to the HW structure
> > > @@ -1059,7 +1080,7 @@ static s32
> e1000_platform_pm_pch_lpt(struct
> > e1000_hw *hw, bool link)
> > > E1000_PCI_LTR_CAP_LPT + 2,
> > &max_nosnoop);
> > > max_ltr_enc = max_t(u16, max_snoop,
> max_nosnoop);
> > >
> > > - if (lat_enc > max_ltr_enc)
> > > + if (decoded_ltr(lat_enc) >
> decoded_ltr(max_ltr_enc))
> > > lat_enc = max_ltr_enc;
> > > }
> > >
> > >
> > sasha
> >
>
sasha
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2021-07-01 8:34 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-06-29 8:21 [PATCH] driver core: fix e1000e ltr bug YeeLi
2021-06-29 14:49 ` [Intel-wired-lan] " Neftin, Sasha
[not found] ` <CALX8JfQymbSmCP0xk0C-=v64__uaH=BR0UZjr2yRyLWVwm9dLQ@mail.gmail.com>
2021-06-30 6:13 ` Neftin, Sasha
[not found] ` <CALX8JfTOG7bXn7gP+4n1_3CfFFk6aL-RJDY03RCOeBSpUeTEPQ@mail.gmail.com>
2021-07-01 8:34 ` Neftin, Sasha
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).