intel: ptp: convert .adjfreq to .adjfine
Subsystems (2)
cover intel: ptp: convert .adjfreq to .adjfine
Convert all of the Intel drivers with PTP support to the newer .adjfine implementation which uses scaled parts per million.
This improves the precision of the frequency adjustments by taking advantage of the full scaled parts per million input coming from user space.
In addition, all implementations are converted to using the mul_u64_u64_div_u64 function which better handles the intermediate value. This function supports architecture specific instructions where possible to avoid loss of precision if the normal 64-bit multiplication would overflow.
Of note, the i40e implementation is now able to avoid loss of precision on slower link speeds by taking advantage of this to multiply by the link speed factor first. This results in a significantly more precise adjustment by allowing the calculation to impact the lower bits.
This also gets us a step closer to being able to remove the .adjfreq entirely by removing its use from many drivers.
I plan to follow this up with a series to update the drivers from other vendors and drop the .adjfreq implementation entirely.
Jacob Keller (7):
ice: implement adjfine with mul_u64_u64_div_u64
e1000e: remove unnecessary range check in e1000e_phc_adjfreq
e1000e: convert .adjfreq to .adjfine
i40e: use mul_u64_u64_div_u64 for PTP frequency calculation
i40e: convert .adjfreq to .adjfine
ixgbe: convert .adjfreq to .adjfine
igb: convert .adjfreq to .adjfine
drivers/net/ethernet/intel/e1000e/e1000.h | 2 +- drivers/net/ethernet/intel/e1000e/netdev.c | 4 +- drivers/net/ethernet/intel/e1000e/ptp.c | 18 +++-- drivers/net/ethernet/intel/i40e/i40e_ptp.c | 35 ++++------ drivers/net/ethernet/intel/ice/ice_ptp.c | 16 +---- drivers/net/ethernet/intel/igb/igb_ptp.c | 15 ++-- drivers/net/ethernet/intel/ixgbe/ixgbe_ptp.c | 73 +++++++++++--------- 7 files changed, 75 insertions(+), 88 deletions(-)
base-commit: 9f055d4106bdefd5a533316c1d3dc3b5e9821b9a
Signature · 6 lines
2.35.1.456.ga9c7032d4631 _______________________________________________ Intel-wired-lan mailing list [email protected] https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
PATCH 1/7 ice: implement adjfine with mul_u64_u64_div_u64
The PTP frequency adjustment code needs to determine an appropriate adjustment given an input scaled_ppm adjustment.
We calculate the adjustment to the register by multiplying the base (nominal) increment value by the scaled_ppm and then dividing by the scaled one million value.
For very large adjustments, this might overflow. To avoid this, both the scaled_ppm and divisor values are downshifted.
We can avoid that on X86 architectures by using mul_u64_u64_div_u64. This helper function will perform the multiplication and division with 128bit intermediate values. We know that scaled_ppm is never larger than the divisor so this operation will never result in an overflow.
This improves the accuracy of the calculations for large adjustment values on X86. It is likely an improvement on other architectures as well because the default implementation of mul_u64_u64_div_u64 is smarter than the original approach taken in the ice code.
Additionally, this implementation is easier to read, using fewer local variables and lines of code to implement.
--- drivers/net/ethernet/intel/ice/ice_ptp.c | 16 +++------------- 1 file changed, 3 insertions(+), 13 deletions(-)
drivers/net/ethernet/intel/ice/ice_ptp.c +3 -13
@@ -1102,9 +1102,8 @@ static void ice_ptp_reset_phy_timestamping(struct ice_pf *pf)static int ice_ptp_adjfine (struct ptp_clock_info *info,long scaled_ppm) {struct ice_pf *pf =ptp_info_to_pf (info); -u64 freq, divisor =1000000ULL ;struct ice_hw *hw = &pf->hw; -s64 incval, diff; +u64 incval, diff;int neg_adj =0 ;int err;
@@ -1115,17 +1114,8 @@ static int ice_ptp_adjfine(struct ptp_clock_info *info, long scaled_ppm) scaled_ppm = -scaled_ppm; } -while ((u64 )scaled_ppm >div64_u64 (U64_MAX , incval)) { -/* handle overflow by scaling down the scaled_ppm and -* the divisor, losing some precision -*/ - scaled_ppm >>=2 ; - divisor >>=2 ; - } - - freq = (incval * (u64 )scaled_ppm) >>16 ; - diff =div_u64 (freq, divisor); - + diff =mul_u64_u64_div_u64 (incval, (u64 )scaled_ppm, +1000000ULL <<16 );if (neg_adj) incval -= diff;else
Signature · 6 lines
2.35.1.456.ga9c7032d4631 _______________________________________________ Intel-wired-lan mailing list [email protected] https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
Re: [Intel-wired-lan] [net-next PATCH 1/7] ice: implement adjfine with mul_u64_u64_div_u64
Quoted 30 lines
-----Original Message-----
From: Intel-wired-lan <[email protected]> On Behalf Of Jacob Keller
Sent: Friday, July 22, 2022 3:00 AM
To: Intel Wired LAN <[email protected]>
Subject: [Intel-wired-lan] [net-next PATCH 1/7] ice: implement adjfine withmul_u64_u64_div_u64The PTP frequency adjustment code needs to determine an appropriate adjustment given an input
scaled_ppmadjustment.We calculate the adjustment to the register by multiplying the base (nominal) increment value by the
scaled_ppmand then dividing by the scaled one million value.For very large adjustments, this might overflow. To avoid this, both the
scaled_ppmand divisor values are downshifted.We can avoid that on X86 architectures by using
mul_u64_u64_div_u64. This helper function will perform the multiplication and division with 128bit intermediate values. We know thatscaled_ppmis never larger than the divisor so this operation will never result in an overflow.This improves the accuracy of the calculations for large adjustment values on X86. It is likely an improvement on other architectures as well because the default implementation of
mul_u64_u64_div_u64is smarter than the original approach taken in the ice code.
Quoted 8 lines
Additionally, this implementation is easier to read, using fewer local variables and lines of code to implement.
Signed-off-by Jacob Keller <[email protected]>--- drivers/net/ethernet/intel/ice/ice_ptp.c | 16 +++------------- 1 file changed, 3 insertions(+), 13 deletions(-)
_______________________________________________
Intel-wired-lan mailing list
[email protected]
https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
PATCH 2/7 e1000e: remove unnecessary range check in e1000e_phc_adjfreq
In reply to Jacob KellerThe e1000e_phc_adjfreq function validates that the input delta is within the maximum range. This is already handled by the core PTP code and this is a duplicate and thus unnecessary check. It also complicates refactoring to use the newer .adjfine implementation, where the input is no longer specified in parts per billion. Remove the range validation check.
--- drivers/net/ethernet/intel/e1000e/ptp.c | 3 --- 1 file changed, 3 deletions(-)
drivers/net/ethernet/intel/e1000e/ptp.c +0 -3
@@ -33,9 +33,6 @@ static int e1000e_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta)u32 timinca, incvalue;s32 ret_val; -if ((delta > ptp->max_adj) || (delta <=-1000000000 )) -return -EINVAL ; -if (delta <0 ) { neg_adj = true; delta = -delta;
Signature · 6 lines
2.35.1.456.ga9c7032d4631 _______________________________________________ Intel-wired-lan mailing list [email protected] https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
Re: [Intel-wired-lan] [net-next PATCH 2/7] e1000e: remove unnecessary range check in e1000e_phc_adjfreq
Quoted 10 lines · Jacob Keller
The
e1000e_phc_adjfreqfunction validates that the input delta is within the maximum range. This is already handled by the core PTP code and this is a duplicate and thus unnecessary check. It also complicates refactoring to use the newer .adjfine implementation, where the input is no longer specified in parts per billion. Remove the range validation check.Signed-off-by Jacob Keller <[email protected]>--- drivers/net/ethernet/intel/e1000e/ptp.c | 3 --- 1 file changed, 3 deletions(-)
_______________________________________________
Intel-wired-lan mailing list
[email protected]
https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
PATCH 1/1 igb: convert .adjfreq to .adjfine
In reply to Jacob KellerThe 82576 PTP implementation still uses .adjfreq instead of using the newer .adjfine.
This implementation uses a pre-simplified calculation since the base increment value for the 82576 is just 16 * 2^19. Converting this into scaled_ppm is tricky, and makes the intent a bit less clear.
Simply convert to the normal flow of multiplying the base increment value by the scaled_ppm and then dividing by 1000000ULL << 16. This can be implemented using mul_u64_u64_div_u64 which can avoid the possible overflow that might occur for large adjustments.
Use of .adjfine can improve the precision of small adjustments and gets us one driver closer to removing the old implementation from the kernel entirely.
--- drivers/net/ethernet/intel/igb/igb_ptp.c | 15 +++++++-------- 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/net/ethernet/intel/igb/igb_ptp.c +7 -8
@@ -190,7 +190,7 @@ static void igb_ptp_systim_to_hwtstamp(struct igb_adapter *adapter, }/* PTP clock operations */ -static int igb_ptp_adjfreq_82576 (struct ptp_clock_info *ptp,s32 ppb) +static int igb_ptp_adjfine_82576 (struct ptp_clock_info *ptp,long scaled_ppm) { struct igb_adapter *igb = container_of(ptp, struct igb_adapter, ptp_caps);
@@ -199,15 +199,14 @@ static int igb_ptp_adjfreq_82576(struct ptp_clock_info *ptp, s32 ppb)u64 rate;u32 incvalue; -if (ppb <0 ) { +if (scaled_ppm <0 ) { neg_adj =1 ; - ppb = -ppb; + scaled_ppm = -scaled_ppm; } - rate = ppb; - rate <<=14 ; - rate =div_u64 (rate,1953125 ); - incvalue =16 <<IGB_82576_TSYNC_SHIFT ; + incvalue =INCVALUE_82576 ; + rate =mul_u64_u64_div_u64 (incvalue, (u64 )scaled_ppm, +1000000ULL <<16 );if (neg_adj) incvalue -= rate;
@@ -1347,7 +1346,7 @@ void igb_ptp_init(struct igb_adapter *adapter) adapter->ptp_caps.max_adj =999999881 ; adapter->ptp_caps.n_ext_ts =0 ; adapter->ptp_caps.pps =0 ; - adapter->ptp_caps.adjfreq = igb_ptp_adjfreq_82576; + adapter->ptp_caps.adjfine = igb_ptp_adjfine_82576; adapter->ptp_caps.adjtime = igb_ptp_adjtime_82576; adapter->ptp_caps.gettimex64 = igb_ptp_gettimex_82576; adapter->ptp_caps.settime64 = igb_ptp_settime_82576;
Signature · 6 lines
2.35.1.456.ga9c7032d4631 _______________________________________________ Intel-wired-lan mailing list [email protected] https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
PATCH 3/7 e1000e: convert .adjfreq to .adjfine
In reply to Jacob KellerThe PTP implementation for the e1000e driver uses the older .adjfreq method. This method takes an adjustment in parts per billion. The newer .adjfine implementation uses scaled_ppm. The use of scaled_ppm allows for finer grained adjustments and is preferred over using the older implementation.
Make use of mul_u64_u64_div_u64 in order to handle possible overflow of the multiplication used to calculate the desired adjustment to the hardware increment value.
--- drivers/net/ethernet/intel/e1000e/e1000.h | 2 +- drivers/net/ethernet/intel/e1000e/netdev.c | 4 ++-- drivers/net/ethernet/intel/e1000e/ptp.c | 15 ++++++++------- 3 files changed, 11 insertions(+), 10 deletions(-)
drivers/net/ethernet/intel/e1000e/e1000.h +1 -1
@@ -329,7 +329,7 @@ struct e1000_adapter {struct ptp_clock *ptp_clock;struct ptp_clock_info ptp_clock_info;struct pm_qos_request pm_qos_req; -s32 ptp_delta; +long ptp_delta;u16 eee_advert; };
drivers/net/ethernet/intel/e1000e/netdev.c +2 -2
@@ -3922,9 +3922,9 @@ static void e1000e_systim_reset(struct e1000_adapter *adapter)if (!(adapter->flags &FLAG_HAS_HW_TIMESTAMP ))return ; -if (info->adjfreq) { +if (info->adjfine) {/* restore the previous ptp frequency delta */ - ret_val = info->adjfreq (info, adapter->ptp_delta); + ret_val = info->adjfine (info, adapter->ptp_delta); }else {/* set the default base frequency if no adjustment possible */ ret_val =e1000e_get_base_timinca (adapter, &timinca);
drivers/net/ethernet/intel/e1000e/ptp.c +8 -7
@@ -15,14 +15,16 @@#endif /** -* e1000e_phc_adjfreq - adjust the frequency of the hardware clock +* e1000e_phc_adjfine - adjust the frequency of the hardware clock * @ptp: ptp clock structure -* @delta: Desired frequency change in parts per billion +* @delta: Desired frequency chance in scaled parts per million * * Adjust the frequency of the PHC cycle counter by the indicated delta from * the base frequency. +* +* Scaled parts per million is ppm but with a 16 bit binary fractional field. **/ -static int e1000e_phc_adjfreq (struct ptp_clock_info *ptp,s32 delta) +static int e1000e_phc_adjfine (struct ptp_clock_info *ptp,long delta) { struct e1000_adapter *adapter = container_of(ptp, struct e1000_adapter, ptp_clock_info);
@@ -47,9 +49,8 @@ static int e1000e_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta) incvalue = timinca &E1000_TIMINCA_INCVALUE_MASK ; - adjustment = incvalue; - adjustment *= delta; - adjustment =div_u64 (adjustment,1000000000 ); + adjustment =mul_u64_u64_div_u64 (incvalue, (u64 )delta, +1000000ULL <<16 ); incvalue = neg_adj ? (incvalue - adjustment) : (incvalue + adjustment);
@@ -257,7 +258,7 @@ static const struct ptp_clock_info e1000e_ptp_clock_info = { .n_per_out = 0, .n_pins = 0, .pps = 0, - .adjfreq = e1000e_phc_adjfreq, + .adjfine = e1000e_phc_adjfine, .adjtime = e1000e_phc_adjtime, .gettimex64 = e1000e_phc_gettimex, .settime64 = e1000e_phc_settime,
Signature · 6 lines
2.35.1.456.ga9c7032d4631 _______________________________________________ Intel-wired-lan mailing list [email protected] https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
Re: [Intel-wired-lan] [net-next PATCH 3/7] e1000e: convert .adjfreq to .adjfine
Quoted 10 lines · Jacob Keller
The PTP implementation for the e1000e driver uses the older .adjfreq method. This method takes an adjustment in parts per billion. The newer .adjfine implementation uses
scaled_ppm. The use ofscaled_ppmallows for finer grained adjustments and is preferred over using the older implementation.Make use of
mul_u64_u64_div_u64in order to handle possible overflow of the multiplication used to calculate the desired adjustment to the hardware increment value.
Quoted 6 lines
Signed-off-by Jacob Keller <[email protected]>--- drivers/net/ethernet/intel/e1000e/e1000.h | 2 +- drivers/net/ethernet/intel/e1000e/netdev.c | 4 ++-- drivers/net/ethernet/intel/e1000e/ptp.c | 15 ++++++++------- 3 files changed, 11 insertions(+), 10 deletions(-)
_______________________________________________
Intel-wired-lan mailing list
[email protected]
https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
PATCH 4/7 i40e: use mul_u64_u64_div_u64 for PTP frequency calculation
In reply to Jacob KellerThe i40e device has a different clock rate depending on the current link speed. This requires using a different increment rate for the PTP clock registers. For slower link speeds, the base increment value is larger. Directly multiplying the larger increment value by the parts per billion adjustment might overflow.
To avoid this, the i40e implementation defaults to using the lower increment value and then multiplying the adjustment afterwards. This causes a loss of precision for lower link speeds.
We can fix this by using mul_u64_u64_div_u64 instead of performing the multiplications using standard C operations. On X86, this will use special instructions that perform the multiplication and division with 128bit intermediate values. For other architectures, the fallback implementation will limit the loss of precision for large values. Small adjustments don't overflow anyways and won't lose precision at all.
This allows first multiplying the base increment value and then performing the adjustment calculation, since we no longer fear overflowing. It also makes it easier to convert to the even more precise .adjfine implementation in a following change.
--- drivers/net/ethernet/intel/i40e/i40e_ptp.c | 17 ++++------------- 1 file changed, 4 insertions(+), 13 deletions(-)
drivers/net/ethernet/intel/i40e/i40e_ptp.c +4 -13
@@ -353,25 +353,16 @@ static int i40e_ptp_adjfreq(struct ptp_clock_info *ptp, s32 ppb) ppb = -ppb; } - freq =I40E_PTP_40GB_INCVAL ; - freq *= ppb; - diff =div_u64 (freq,1000000000ULL ); +smp_mb ();/* Force any pending update before accessing. */ + freq =I40E_PTP_40GB_INCVAL *READ_ONCE (pf->ptp_adj_mult); + diff =mul_u64_u64_div_u64 (freq, (u64 )ppb, +1000000000ULL );if (neg_adj) adj =I40E_PTP_40GB_INCVAL - diff;else adj =I40E_PTP_40GB_INCVAL + diff; -/* At some link speeds, the base incval is so large that directly -* multiplying by ppb would result in arithmetic overflow even when -* using a u64. Avoid this by instead calculating the new incval -* always in terms of the 40GbE clock rate and then multiplying by the -* link speed factor afterwards. This does result in slightly lower -* precision at lower link speeds, but it is fairly minor. -*/ -smp_mb ();/* Force any pending update before accessing. */ - adj *=READ_ONCE (pf->ptp_adj_mult); -wr32 (hw,I40E_PRTTSYN_INC_L , adj &0xFFFFFFFF );wr32 (hw,I40E_PRTTSYN_INC_H , adj >>32 );
Signature · 6 lines
2.35.1.456.ga9c7032d4631 _______________________________________________ Intel-wired-lan mailing list [email protected] https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
Re: [Intel-wired-lan] [net-next PATCH 4/7] i40e: use mul_u64_u64_div_u64 for PTP frequency calculation
Quoted 32 lines
-----Original Message-----
From: Intel-wired-lan <[email protected]> On Behalf Of Jacob Keller
Sent: Friday, July 22, 2022 3:00 AM
To: Intel Wired LAN <[email protected]>
Subject: [Intel-wired-lan] [net-next PATCH 4/7] i40e: usemul_u64_u64_div_u64for PTP frequency calculationThe i40e device has a different clock rate depending on the current link speed. This requires using a different increment rate for the PTP clock registers. For slower link speeds, the base increment value is larger. Directly multiplying the larger increment value by the parts per billion adjustment might overflow.
To avoid this, the i40e implementation defaults to using the lower increment value and then multiplying the adjustment afterwards. This causes a loss of precision for lower link speeds.
We can fix this by using
mul_u64_u64_div_u64instead of performing the multiplications using standard C operations. On X86, this will use special instructions that perform the multiplication and division with 128bit intermediate values. For other architectures, the fallback implementation will limit the loss of precision for large values. Small adjustments don't overflow anyways and won't lose precision at all.This allows first multiplying the base increment value and then performing the adjustment calculation, since we no longer fear overflowing. It also makes it easier to convert to the even more precise .adjfine implementation in a following change.
Quoted 5 lines
Signed-off-by Jacob Keller <[email protected]>--- drivers/net/ethernet/intel/i40e/i40e_ptp.c | 17 ++++------------- 1 file changed, 4 insertions(+), 13 deletions(-)
_______________________________________________
Intel-wired-lan mailing list
[email protected]
https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
PATCH 5/7 i40e: convert .adjfreq to .adjfine
In reply to Jacob KellerThe i40e driver currently implements the .adjfreq handler for frequency adjustments. This takes the adjustment parameter in parts per billion. The PTP core supports .adjfine which provides an adjustment in scaled parts per million. This has a higher resolution and can result in more precise adjustments for small corrections.
Convert the existing .adjfreq implementation to the newer .adjfine implementation. This is trivial since it just requires changing the divisor from 1000000000ULL to (1000000ULL << 16) in the mul_u64_u64_div_u64 call.
This improves the precision of the adjustments and gets us one driver closer to removing the old .adjfreq support from the kernel.
--- drivers/net/ethernet/intel/i40e/i40e_ptp.c | 22 ++++++++++++---------- 1 file changed, 12 insertions(+), 10 deletions(-)
drivers/net/ethernet/intel/i40e/i40e_ptp.c +12 -10
@@ -334,29 +334,31 @@ static void i40e_ptp_convert_to_hwtstamp(struct skb_shared_hwtstamps *hwtstamps, }/** -* i40e_ptp_adjfreq - Adjust the PHC frequency +* i40e_ptp_adjfine - Adjust the PHC frequency * @ptp: The PTP clock structure -* @ppb: Parts per billion adjustment from the base +* @scaled_ppm: Scaled parts per million adjustment from base * -* Adjust the frequency of the PHC by the indicated parts per billion from the -* base frequency. +* Adjust the frequency of the PHC by the indicated delta from the base +* frequency. +* +* Scaled parts per million is ppm with a 16 bit binary fractional field. **/ -static int i40e_ptp_adjfreq (struct ptp_clock_info *ptp,s32 ppb) +static int i40e_ptp_adjfine (struct ptp_clock_info *ptp,long scaled_ppm) { struct i40e_pf *pf = container_of(ptp, struct i40e_pf, ptp_caps);struct i40e_hw *hw = &pf->hw;u64 adj, freq, diff;int neg_adj =0 ; -if (ppb <0 ) { +if (scaled_ppm <0 ) { neg_adj =1 ; - ppb = -ppb; + scaled_ppm = -scaled_ppm; }smp_mb ();/* Force any pending update before accessing. */ freq =I40E_PTP_40GB_INCVAL *READ_ONCE (pf->ptp_adj_mult); - diff =mul_u64_u64_div_u64 (freq, (u64 )ppb, -1000000000ULL ); + diff =mul_u64_u64_div_u64 (freq, (u64 )scaled_ppm, +1000000ULL <<16 );if (neg_adj) adj =I40E_PTP_40GB_INCVAL - diff;
@@ -1392,7 +1394,7 @@ static long i40e_ptp_create_clock(struct i40e_pf *pf) sizeof(pf->ptp_caps.name) - 1); pf->ptp_caps.owner =THIS_MODULE ; pf->ptp_caps.max_adj =999999999 ; - pf->ptp_caps.adjfreq = i40e_ptp_adjfreq; + pf->ptp_caps.adjfine = i40e_ptp_adjfine; pf->ptp_caps.adjtime = i40e_ptp_adjtime; pf->ptp_caps.gettimex64 = i40e_ptp_gettimex; pf->ptp_caps.settime64 = i40e_ptp_settime;
Signature · 6 lines
2.35.1.456.ga9c7032d4631 _______________________________________________ Intel-wired-lan mailing list [email protected] https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
Re: [Intel-wired-lan] [net-next PATCH 5/7] i40e: convert .adjfreq to .adjfine
Quoted 20 lines
-----Original Message-----
From: Intel-wired-lan <[email protected]> On Behalf Of Jacob Keller
Sent: Friday, July 22, 2022 3:00 AM
To: Intel Wired LAN <[email protected]>
Subject: [Intel-wired-lan] [net-next PATCH 5/7] i40e: convert .adjfreq to .adjfineThe i40e driver currently implements the .adjfreq handler for frequency adjustments. This takes the adjustment parameter in parts per billion. The PTP core supports .adjfine which provides an adjustment in scaled parts per million. This has a higher resolution and can result in more precise adjustments for small corrections.
Convert the existing .adjfreq implementation to the newer .adjfine implementation. This is trivial since it just requires changing the divisor from 1000000000ULL to (1000000ULL << 16) in the
mul_u64_u64_div_u64call.
Quoted 8 lines
This improves the precision of the adjustments and gets us one driver closer to removing the old .adjfreq support from the kernel.
Signed-off-by Jacob Keller <[email protected]>--- drivers/net/ethernet/intel/i40e/i40e_ptp.c | 22 ++++++++++++---------- 1 file changed, 12 insertions(+), 10 deletions(-)
_______________________________________________
Intel-wired-lan mailing list
[email protected]
https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
PATCH 6/7 ixgbe: convert .adjfreq to .adjfine
In reply to Jacob KellerConvert the ixgbe PTP frequency adjustment implementations from .adjfreq to .adjfine. This allows using the scaled parts per million adjustment from the PTP core and results in a more precise adjustment for small corrections.
To avoid overflow, use mul_u64_u64_div_u64 to perform the calculation. On X86 platforms, this will use instructions that perform the operations with 128bit intermediate values. For other architectures, the implementation will limit the loss of precision as much as possible.
This change slightly improves the precision of frequency adjustments for all ixgbe based devices, and gets us one driver closer to being able to remove the older .adjfreq implementation from the kernel.
--- drivers/net/ethernet/intel/ixgbe/ixgbe_ptp.c | 73 +++++++++++--------- 1 file changed, 40 insertions(+), 33 deletions(-)
drivers/net/ethernet/intel/ixgbe/ixgbe_ptp.c +40 -33
@@ -113,12 +113,16 @@* the sign bit. This register enables software to calculate frequency * adjustments and apply them directly to the clock rate. * -* The math for converting ppb into TIMINCA values is fairly straightforward. -* TIMINCA value = ( Base_Frequency * ppb ) / 1000000000ULL +* The math for converting scaled_ppm into TIMINCA values is fairly +* straightforward. * -* This assumes that ppb is never high enough to create a value bigger than -* TIMINCA's 31 bits can store. This is ensured by the stack. Calculating this -* value is also simple. +* TIMINCA value = ( Base_Frequency * scaled_ppm ) / 1000000ULL << 16 +* +* To avoid overflow, we simply use mul_u64_u64_div_u64. +* +* This assumes that scaled_ppm is never high enough to create a value bigger +* than TIMINCA's 31 bits can store. This is ensured by the stack, and is +* measured in parts per billion. Calculating this value is also simple. * Max ppb = ( Max Adjustment / Base Frequency ) / 1000000000ULL * * For the X550, the Max adjustment is +/- 0.5 ns, and the base frequency is
@@ -433,45 +437,45 @@ static void ixgbe_ptp_convert_to_hwtstamp(struct ixgbe_adapter *adapter, }/** -* ixgbe_ptp_adjfreq_82599 +* ixgbe_ptp_adjfine_82599 * @ptp: the ptp clock structure -* @ppb: parts per billion adjustment from base +* @scaled_ppm: scaled parts per million adjustment from base * -* adjust the frequency of the ptp cycle counter by the -* indicated ppb from the base frequency. +* Adjust the frequency of the ptp cycle counter by the +* indicated scaled_ppm from the base frequency. +* +* Scaled parts per million is ppm with a 16-bit binary fractional field. */ -static int ixgbe_ptp_adjfreq_82599 (struct ptp_clock_info *ptp,s32 ppb) +static int ixgbe_ptp_adjfine_82599 (struct ptp_clock_info *ptp,long scaled_ppm) {struct ixgbe_adapter *adapter = container_of(ptp, struct ixgbe_adapter, ptp_caps);struct ixgbe_hw *hw = &adapter->hw; -u64 freq, incval; -u32 diff; +u64 incval, diff;int neg_adj =0 ; -if (ppb <0 ) { +if (scaled_ppm <0 ) { neg_adj =1 ; - ppb = -ppb; + scaled_ppm = -scaled_ppm; }smp_mb (); incval =READ_ONCE (adapter->base_incval); - freq = incval; - freq *= ppb; - diff =div_u64 (freq,1000000000ULL ); + diff =mul_u64_u64_div_u64 (incval, scaled_ppm, +1000000ULL <<16 ); incval = neg_adj ? (incval - diff) : (incval + diff);switch (hw->mac.type) {case ixgbe_mac_X540:if (incval >0xFFFFFFFFULL ) -e_dev_warn ("PTP ppb adjusted SYSTIME rate overflowed!\n" ); +e_dev_warn ("PTP scaled_ppm adjusted SYSTIME rate overflowed!\n" );IXGBE_WRITE_REG (hw,IXGBE_TIMINCA , (u32 )incval);break ;case ixgbe_mac_82599EB:if (incval >0x00FFFFFFULL ) -e_dev_warn ("PTP ppb adjusted SYSTIME rate overflowed!\n" ); +e_dev_warn ("PTP scaled_ppm adjusted SYSTIME rate overflowed!\n" );IXGBE_WRITE_REG (hw,IXGBE_TIMINCA ,BIT (IXGBE_INCPER_SHIFT_82599 ) | ((u32 )incval &0x00FFFFFFUL ));
@@ -484,32 +488,35 @@ static int ixgbe_ptp_adjfreq_82599(struct ptp_clock_info *ptp, s32 ppb) }/** -* ixgbe_ptp_adjfreq_X550 +* ixgbe_ptp_adjfine_X550 * @ptp: the ptp clock structure -* @ppb: parts per billion adjustment from base +* @scaled_ppm: scaled parts per million adjustment from base * -* adjust the frequency of the SYSTIME registers by the indicated ppb from base -* frequency +* Adjust the frequency of the SYSTIME registers by the indicated scaled_ppm +* from base frequency. +* +* Scaled parts per million is ppm with a 16-bit binary fractional field. */ -static int ixgbe_ptp_adjfreq_X550 (struct ptp_clock_info *ptp,s32 ppb) +static int ixgbe_ptp_adjfine_X550 (struct ptp_clock_info *ptp,long scaled_ppm) {struct ixgbe_adapter *adapter = container_of(ptp, struct ixgbe_adapter, ptp_caps);struct ixgbe_hw *hw = &adapter->hw;int neg_adj =0 ; -u64 rate =IXGBE_X550_BASE_PERIOD ; +u64 rate;u32 inca; -if (ppb <0 ) { +if (scaled_ppm <0 ) { neg_adj =1 ; - ppb = -ppb; + scaled_ppm = -scaled_ppm; } - rate *= ppb; - rate =div_u64 (rate,1000000000ULL ); + + rate =mul_u64_u64_div_u64 (IXGBE_X550_BASE_PERIOD , scaled_ppm, +1000000ULL <<16 );/* warn if rate is too large */ if (rate >=INCVALUE_MASK ) -e_dev_warn ("PTP ppb adjusted SYSTIME rate overflowed!\n" ); +e_dev_warn ("PTP scaled_ppm adjusted SYSTIME rate overflowed!\n" ); inca = rate &INCVALUE_MASK ;if (neg_adj)
@@ -1355,7 +1362,7 @@ static long ixgbe_ptp_create_clock(struct ixgbe_adapter *adapter) adapter->ptp_caps.n_ext_ts =0 ; adapter->ptp_caps.n_per_out =0 ; adapter->ptp_caps.pps =1 ; - adapter->ptp_caps.adjfreq = ixgbe_ptp_adjfreq_82599; + adapter->ptp_caps.adjfine = ixgbe_ptp_adjfine_82599; adapter->ptp_caps.adjtime = ixgbe_ptp_adjtime; adapter->ptp_caps.gettimex64 = ixgbe_ptp_gettimex; adapter->ptp_caps.settime64 = ixgbe_ptp_settime;
@@ -1372,7 +1379,7 @@ static long ixgbe_ptp_create_clock(struct ixgbe_adapter *adapter) adapter->ptp_caps.n_ext_ts =0 ; adapter->ptp_caps.n_per_out =0 ; adapter->ptp_caps.pps =0 ; - adapter->ptp_caps.adjfreq = ixgbe_ptp_adjfreq_82599; + adapter->ptp_caps.adjfine = ixgbe_ptp_adjfine_82599; adapter->ptp_caps.adjtime = ixgbe_ptp_adjtime; adapter->ptp_caps.gettimex64 = ixgbe_ptp_gettimex; adapter->ptp_caps.settime64 = ixgbe_ptp_settime;
@@ -1388,7 +1395,7 @@ static long ixgbe_ptp_create_clock(struct ixgbe_adapter *adapter) adapter->ptp_caps.n_ext_ts =0 ; adapter->ptp_caps.n_per_out =0 ; adapter->ptp_caps.pps =1 ; - adapter->ptp_caps.adjfreq = ixgbe_ptp_adjfreq_X550; + adapter->ptp_caps.adjfine = ixgbe_ptp_adjfine_X550; adapter->ptp_caps.adjtime = ixgbe_ptp_adjtime; adapter->ptp_caps.gettimex64 = ixgbe_ptp_gettimex; adapter->ptp_caps.settime64 = ixgbe_ptp_settime;
Signature · 6 lines
2.35.1.456.ga9c7032d4631 _______________________________________________ Intel-wired-lan mailing list [email protected] https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
Re: [Intel-wired-lan] [net-next PATCH 6/7] ixgbe: convert .adjfreq to .adjfine
Quoted 19 lines
-----Original Message-----
From: Intel-wired-lan <[email protected]> On Behalf Of Jacob Keller
Sent: Friday, July 22, 2022 3:00 AM
To: Intel Wired LAN <[email protected]>
Subject: [Intel-wired-lan] [net-next PATCH 6/7] ixgbe: convert .adjfreq to .adjfineConvert the ixgbe PTP frequency adjustment implementations from .adjfreq to .adjfine. This allows using the scaled parts per million adjustment from the PTP core and results in a more precise adjustment for small corrections.
To avoid overflow, use
mul_u64_u64_div_u64to perform the calculation. On X86 platforms, this will use instructions that perform the operations with 128bit intermediate values. For other architectures, the implementation will limit the loss of precision as much as possible.
Quoted 9 lines
This change slightly improves the precision of frequency adjustments for all ixgbe based devices, and gets us one driver closer to being able to remove the older .adjfreq implementation from the kernel.
Signed-off-by Jacob Keller <[email protected]>--- drivers/net/ethernet/intel/ixgbe/ixgbe_ptp.c | 73 +++++++++++--------- 1 file changed, 40 insertions(+), 33 deletions(-)
_______________________________________________
Intel-wired-lan mailing list
[email protected]
https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
PATCH 7/7 igb: convert .adjfreq to .adjfine
In reply to Jacob KellerThe 82576 PTP implementation still uses .adjfreq instead of using the newer .adjfine.
This implementation uses a pre-simplified calculation since the base increment value for the 82576 is just 16 * 2^19. Converting this into scaled_ppm is tricky, and makes the intent a bit less clear.
Simply convert to the normal flow of multiplying the base increment value by the scaled_ppm and then dividing by 1000000ULL << 16. This can be implemented using mul_u64_u64_div_u64 which can avoid the possible overflow that might occur for large adjustments.
Use of .adjfine can improve the precision of small adjustments and gets us one driver closer to removing the old implementation from the kernel entirely.
--- drivers/net/ethernet/intel/igb/igb_ptp.c | 15 +++++++-------- 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/net/ethernet/intel/igb/igb_ptp.c +7 -8
@@ -190,7 +190,7 @@ static void igb_ptp_systim_to_hwtstamp(struct igb_adapter *adapter, }/* PTP clock operations */ -static int igb_ptp_adjfreq_82576 (struct ptp_clock_info *ptp,s32 ppb) +static int igb_ptp_adjfine_82576 (struct ptp_clock_info *ptp,long scaled_ppm) { struct igb_adapter *igb = container_of(ptp, struct igb_adapter, ptp_caps);
@@ -199,15 +199,14 @@ static int igb_ptp_adjfreq_82576(struct ptp_clock_info *ptp, s32 ppb)u64 rate;u32 incvalue; -if (ppb <0 ) { +if (scaled_ppm <0 ) { neg_adj =1 ; - ppb = -ppb; + scaled_ppm = -scaled_ppm; } - rate = ppb; - rate <<=14 ; - rate =div_u64 (rate,1953125 ); - incvalue =16 <<IGB_82576_TSYNC_SHIFT ; + incvalue =INCVALUE_82576 ; + rate =mul_u64_u64_div_u64 (incvalue, (u64 )scaled_ppm, +1000000ULL <<16 );if (neg_adj) incvalue -= rate;
@@ -1347,7 +1346,7 @@ void igb_ptp_init(struct igb_adapter *adapter) adapter->ptp_caps.max_adj =999999881 ; adapter->ptp_caps.n_ext_ts =0 ; adapter->ptp_caps.pps =0 ; - adapter->ptp_caps.adjfreq = igb_ptp_adjfreq_82576; + adapter->ptp_caps.adjfine = igb_ptp_adjfine_82576; adapter->ptp_caps.adjtime = igb_ptp_adjtime_82576; adapter->ptp_caps.gettimex64 = igb_ptp_gettimex_82576; adapter->ptp_caps.settime64 = igb_ptp_settime_82576;
Signature · 6 lines
2.35.1.456.ga9c7032d4631 _______________________________________________ Intel-wired-lan mailing list [email protected] https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
Re: [Intel-wired-lan] [net-next PATCH 7/7] igb: convert .adjfreq to .adjfine
Quoted 22 lines
-----Original Message-----
From: Intel-wired-lan <[email protected]> On Behalf Of Jacob Keller
Sent: Friday, July 22, 2022 3:00 AM
To: Intel Wired LAN <[email protected]>
Subject: [Intel-wired-lan] [net-next PATCH 7/7] igb: convert .adjfreq to .adjfineThe 82576 PTP implementation still uses .adjfreq instead of using the newer .adjfine.
This implementation uses a pre-simplified calculation since the base increment value for the 82576 is just 16 * 2^19. Converting this into
scaled_ppmis tricky, and makes the intent a bit less clear.Simply convert to the normal flow of multiplying the base increment value by the
scaled_ppmand then dividing by 1000000ULL << 16. This can be implemented usingmul_u64_u64_div_u64which can avoid the possible overflow that might occur for large adjustments.
Quoted 9 lines
Use of .adjfine can improve the precision of small adjustments and gets us one driver closer to removing the old implementation from the kernel entirely.
Signed-off-by Jacob Keller <[email protected]>--- drivers/net/ethernet/intel/igb/igb_ptp.c | 15 +++++++-------- 1 file changed, 7 insertions(+), 8 deletions(-)
_______________________________________________
Intel-wired-lan mailing list
[email protected]
https://lists.osuosl.org/mailman/listinfo/intel-wired-lan