Skip to contentSkip to conversation

Time Counter fixes and improvements

PATCH In review Richard Cochran 4 participants 22 messages Dec 21, 2014 – Jan 2, 2015

Subsystems (11)
lore
Share & tools
lore
#0 lore raw

cover Time Counter fixes and improvements

Several PTP Hardware Clock (PHC) drivers implement the clock in software using the timecounter/cyclecounter code. This series adds one simple improvement and one more subtle fix to the shared timecounter facility. Credit for this series goes to Janusz Użycki, who pointed the issues out to me off list.

Patch #1 simply move the timecounter code into its own file. When working on this series, it was really annoying to see half the kernel recompile after every tweak to the timecounter stuff. There is no reason to keep this together with the clocksource code.

Patch #2 implements an improved adjtime() method, and patches 3-10 convert all of the drivers over to the new method.

Patch #11 fixes a subtle but important issue with the timecounter WRT frequency adjustment. As it stands now, a timecounter based PHC will exhibit a variable frequency resolution (and variable time error) depending on how often the clock is read.

In timecounter_read_delta(), the expression

   (delta * cc->mult) >> cc->shift;

can lose resolution from the adjusted value of 'mult'. If the value of 'delta' is too small, then small changes in 'mult' have no effect. However, if the delta value is large enough, then small changes in 'mult' will have an effect.

Reading the clock too often means smaller 'delta' values which in turn will spoil the fine adjustments made to 'mult'. Up until now, this effect did not show up in my testing. The following example explains why.

The CPTS has an input clock of 250 MHz, and the clock source uses mult=0x80000000 and shift=29, making the ticks to nanoseconds conversion like this:

ticks * 2^31
------------

       2^29

Imagine what happens if the clock is read every 10 milliseconds. Ten milliseconds are about 2500000 ticks, which corresponds to about 21 bits. The product in the numerator has then 52 bits. After the shift operation, 23 bits are preserved. This results in a frequency adjustment resolution of about 0.1 ppm (not _too_ bad.)

A frequency resolution of 1 ppm requires 20 bits.
A frequency resolution of 1 ppb requires 30 bits.

For the 250 MHz CPTS clock, reading every 4 seconds yields a 1 ppb resolution (which is the finest that our API allows).

However, the error can be much higher if the clock is read too often or if time stamps occur close in time to read operations. In general it is really not acceptable to allow the rate of clock readings to influence the clock accuracy.

Thanks,
Richard

Richard Cochran (11):
time: move the timecounter/cyclecounter code into its own file.
timecounter: provide a helper function to shift the time.
net: xgbe: convert to timecounter adjtime.
net: bnx2x: convert to timecounter adjtime.
net: fec: convert to timecounter adjtime.
net: e1000e: convert to timecounter adjtime.
net: igb: convert to timecounter adjtime.
net: ixgbe: convert to timecounter adjtime.
net: mlx4: convert to timecounter adjtime.
net: cpts: convert to timecounter adjtime.
timecounter: keep track of accumulated fractional nanoseconds

 drivers/net/ethernet/amd/xgbe/xgbe-ptp.c         |    8 +-
 drivers/net/ethernet/amd/xgbe/xgbe.h             |    2 +-
 drivers/net/ethernet/broadcom/bnx2x/bnx2x.h      |    2 +-
 drivers/net/ethernet/broadcom/bnx2x/bnx2x_main.c |    6 +-
 drivers/net/ethernet/freescale/fec.h             |    1 +
 drivers/net/ethernet/freescale/fec_ptp.c         |   16 +--
 drivers/net/ethernet/intel/e1000e/e1000.h        |    2 +-
 drivers/net/ethernet/intel/e1000e/ptp.c          |    5 +-
 drivers/net/ethernet/intel/igb/igb.h             |    2 +-
 drivers/net/ethernet/intel/igb/igb_ptp.c         |    7 +-
 drivers/net/ethernet/intel/ixgbe/ixgbe.h         |    2 +-
 drivers/net/ethernet/intel/ixgbe/ixgbe_ptp.c     |   11 +-
 drivers/net/ethernet/mellanox/mlx4/en_clock.c    |    9 +-
 drivers/net/ethernet/ti/cpts.c                   |    5 +-
 drivers/net/ethernet/ti/cpts.h                   |    1 +
 include/clocksource/arm_arch_timer.h             |    2 +-
 include/linux/clocksource.h                      |  102 ----------------
 include/linux/mlx4/device.h                      |    2 +-
 include/linux/timecounter.h                      |  136 ++++++++++++++++++++++
 include/linux/types.h                            |    3 +
 kernel/time/Makefile                             |    2 +-
 kernel/time/clocksource.c                        |   76 ------------
 kernel/time/timecounter.c                        |  112 ++++++++++++++++++
 sound/pci/hda/hda_priv.h                         |    2 +-
 virt/kvm/arm/arch_timer.c                        |    3 +-
 25 files changed, 274 insertions(+), 245 deletions(-)
 create mode 100644 include/linux/timecounter.h
 create mode 100644 kernel/time/timecounter.c
Signature · 1 line
1.7.10.4
#1 lore raw

PATCH 1/11 time: move the timecounter/cyclecounter code into its own file.

The timecounter code has almost nothing to do with the clocksource code. Let it live in its own file. This will help isolate the timecounter users from the clocksource users in the source tree.

---
 drivers/net/ethernet/amd/xgbe/xgbe.h        |    2 +-
 drivers/net/ethernet/broadcom/bnx2x/bnx2x.h |    2 +-
 drivers/net/ethernet/freescale/fec.h        |    1 +
 drivers/net/ethernet/intel/e1000e/e1000.h   |    2 +-
 drivers/net/ethernet/intel/igb/igb.h        |    2 +-
 drivers/net/ethernet/intel/ixgbe/ixgbe.h    |    2 +-
 drivers/net/ethernet/ti/cpts.h              |    1 +
 include/clocksource/arm_arch_timer.h        |    2 +-
 include/linux/clocksource.h                 |  102 ----------------------
 include/linux/mlx4/device.h                 |    2 +-
 include/linux/timecounter.h                 |  122 +++++++++++++++++++++++++++
 include/linux/types.h                       |    3 +
 kernel/time/Makefile                        |    2 +-
 kernel/time/clocksource.c                   |   76 -----------------
 kernel/time/timecounter.c                   |   95 +++++++++++++++++++++
 sound/pci/hda/hda_priv.h                    |    2 +-
 16 files changed, 231 insertions(+), 187 deletions(-)
 create mode 100644 include/linux/timecounter.h
 create mode 100644 kernel/time/timecounter.c
drivers/net/ethernet/amd/xgbe/xgbe.h +1 -1
index f9ec762..2af6aff 100644
@@ -124,7 +124,7 @@
 #include <linux/if_vlan.h>
 #include <linux/bitops.h>
 #include <linux/ptp_clock_kernel.h>
-#include <linux/clocksource.h>
+#include <linux/timecounter.h>
 #include <linux/net_tstamp.h>
 #include <net/dcbnl.h>
 
drivers/net/ethernet/broadcom/bnx2x/bnx2x.h +1 -1
index c3a6072..792ba72 100644
@@ -22,7 +22,7 @@
 
 #include <linux/ptp_clock_kernel.h>
 #include <linux/net_tstamp.h>
-#include <linux/clocksource.h>
+#include <linux/timecounter.h>
 
 /* compilation time flags */
 
drivers/net/ethernet/freescale/fec.h +1 -0
index 469691a..df8bbdd 100644
@@ -16,6 +16,7 @@
 #include <linux/clocksource.h>
 #include <linux/net_tstamp.h>
 #include <linux/ptp_clock_kernel.h>
+#include <linux/timecounter.h>
 
 #if defined(CONFIG_M523x) || defined(CONFIG_M527x) || defined(CONFIG_M528x) || \
     defined(CONFIG_M520x) || defined(CONFIG_M532x) || \
drivers/net/ethernet/intel/e1000e/e1000.h +1 -1
index 7785240..9416e5a 100644
@@ -34,7 +34,7 @@
 #include <linux/pci-aspm.h>
 #include <linux/crc32.h>
 #include <linux/if_vlan.h>
-#include <linux/clocksource.h>
+#include <linux/timecounter.h>
 #include <linux/net_tstamp.h>
 #include <linux/ptp_clock_kernel.h>
 #include <linux/ptp_classify.h>
drivers/net/ethernet/intel/igb/igb.h +1 -1
index 82d891e..ee22da3 100644
@@ -29,7 +29,7 @@
 #include "e1000_mac.h"
 #include "e1000_82575.h"
 
-#include <linux/clocksource.h>
+#include <linux/timecounter.h>
 #include <linux/net_tstamp.h>
 #include <linux/ptp_clock_kernel.h>
 #include <linux/bitops.h>
drivers/net/ethernet/intel/ixgbe/ixgbe.h +1 -1
index b6137be..38fc64c 100644
@@ -38,7 +38,7 @@
 #include <linux/if_vlan.h>
 #include <linux/jiffies.h>
 
-#include <linux/clocksource.h>
+#include <linux/timecounter.h>
 #include <linux/net_tstamp.h>
 #include <linux/ptp_clock_kernel.h>
 
drivers/net/ethernet/ti/cpts.h +1 -0
index 1a581ef..69a46b9 100644
@@ -27,6 +27,7 @@
 #include <linux/list.h>
 #include <linux/ptp_clock_kernel.h>
 #include <linux/skbuff.h>
+#include <linux/timecounter.h>
 
 struct cpsw_cpts {
 	u32 idver;                /* Identification and version */
include/clocksource/arm_arch_timer.h +1 -1
index 6d26b40..9916d0e 100644
@@ -16,7 +16,7 @@
 #ifndef __CLKSOURCE_ARM_ARCH_TIMER_H
 #define __CLKSOURCE_ARM_ARCH_TIMER_H
 
-#include <linux/clocksource.h>
+#include <linux/timecounter.h>
 #include <linux/types.h>
 
 #define ARCH_TIMER_CTRL_ENABLE		(1 << 0)
include/linux/clocksource.h +0 -102

114 lines not on this page · raw message

include/linux/mlx4/device.h +1 -1
index 25c791e..f1e41b3 100644
@@ -42,7 +42,7 @@
 
 #include <linux/atomic.h>
 
-#include <linux/clocksource.h>
+#include <linux/timecounter.h>
 
 #define MAX_MSIX_P_PORT		17
 #define MAX_MSIX		64
include/linux/timecounter.h new file +122 -0

122 lines not on this page · raw message

include/linux/types.h +3 -0
index a0bb704..6232382 100644
@@ -213,5 +213,8 @@ struct callback_head {
 };
 #define rcu_head callback_head
 
+/* clocksource cycle base type */
+typedef u64 cycle_t;
+
 #endif /*  __ASSEMBLY__ */
 #endif /* _LINUX_TYPES_H */
kernel/time/Makefile +1 -1
index f622cf2..c09c078 100644
@@ -1,6 +1,6 @@
 obj-y += time.o timer.o hrtimer.o itimer.o posix-timers.o posix-cpu-timers.o
 obj-y += timekeeping.o ntp.o clocksource.o jiffies.o timer_list.o
-obj-y += timeconv.o posix-clock.o alarmtimer.o
+obj-y += timeconv.o timecounter.o posix-clock.o alarmtimer.o
 
 obj-$(CONFIG_GENERIC_CLOCKEVENTS_BUILD)		+= clockevents.o
 obj-$(CONFIG_GENERIC_CLOCKEVENTS)		+= tick-common.o
kernel/time/clocksource.c +0 -76

82 lines not on this page · raw message

kernel/time/timecounter.c new file +95 -0

95 lines not on this page · raw message

sound/pci/hda/hda_priv.h +1 -1
index aa484fd..3fc7cfe 100644
@@ -15,7 +15,7 @@
 #ifndef __SOUND_HDA_PRIV_H
 #define __SOUND_HDA_PRIV_H
 
-#include <linux/clocksource.h>
+#include <linux/timecounter.h>
 #include <sound/core.h>
 #include <sound/pcm.h>
 
Signature · 1 line
1.7.10.4
Acked-by on v1 1/11
#2 lore raw

Re: [PATCH net-next 01/11] time: move the timecounter/cyclecounter code into its own file.

Quoted 5 lines · Richard Cochran

The timecounter code has almost nothing to do with the clocksource code. Let it live in its own file. This will help isolate the timecounter users from the clocksource users in the source tree.

For the Intel driver changes...

Quoted 23 lines
---
 drivers/net/ethernet/amd/xgbe/xgbe.h        |    2 +-
 drivers/net/ethernet/broadcom/bnx2x/bnx2x.h |    2 +-
 drivers/net/ethernet/freescale/fec.h        |    1 +
 drivers/net/ethernet/intel/e1000e/e1000.h   |    2 +-
 drivers/net/ethernet/intel/igb/igb.h        |    2 +-
 drivers/net/ethernet/intel/ixgbe/ixgbe.h    |    2 +-
 drivers/net/ethernet/ti/cpts.h              |    1 +
 include/clocksource/arm_arch_timer.h        |    2 +-
 include/linux/clocksource.h                 |  102
----------------------
 include/linux/mlx4/device.h                 |    2 +-
 include/linux/timecounter.h                 |  122
+++++++++++++++++++++++++++
 include/linux/types.h                       |    3 +
 kernel/time/Makefile                        |    2 +-
 kernel/time/clocksource.c                   |   76 -----------------
 kernel/time/timecounter.c                   |   95
+++++++++++++++++++++
 sound/pci/hda/hda_priv.h                    |    2 +-
 16 files changed, 231 insertions(+), 187 deletions(-)
 create mode 100644 include/linux/timecounter.h
 create mode 100644 kernel/time/timecounter.c
#3 lore raw

PATCH 2/11 timecounter: provide a helper function to shift the time.

In reply to Richard Cochran

Some PTP Hardware Clock drivers use a struct timecounter to represent their clock. To adjust the time by a given offset, these drivers all perform a two step read/write of their timecounter. However, it is better and simpler just to adjust the offset in one step. This patch introduces a little routine to help drivers implement the adjtime method.

---
 include/linux/timecounter.h |    9 +++++++++
 1 file changed, 9 insertions(+)
include/linux/timecounter.h +9 -0
index 146f07a..af3dfa4 100644
@@ -79,6 +79,15 @@ static inline u64 cyclecounter_cyc2ns(const struct cyclecounter *cc,
 }
 
 /**
+ * timecounter_adjtime - Shifts the time of the clock.
+ * @delta:	Desired change in nanoseconds.
+ */
+static inline void timecounter_adjtime(struct timecounter *tc, s64 delta)
+{
+	tc->nsec += delta;
+}
+
+/**
  * timecounter_init - initialize a time counter
  * @tc:			Pointer to time counter which is to be initialized/reset
  * @cc:			A cycle counter, ready to be used.
Signature · 1 line
1.7.10.4
#4 lore raw

PATCH 3/11 net: xgbe: convert to timecounter adjtime.

In reply to Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

Compile tested only.

---
 drivers/net/ethernet/amd/xgbe/xgbe-ptp.c |    8 +-------
 1 file changed, 1 insertion(+), 7 deletions(-)
drivers/net/ethernet/amd/xgbe/xgbe-ptp.c +1 -7
index a1bf9d1c..f5acf4c 100644
@@ -171,15 +171,9 @@ static int xgbe_adjtime(struct ptp_clock_info *info, s64 delta)
 						   struct xgbe_prv_data,
 						   ptp_clock_info);
 	unsigned long flags;
-	u64 nsec;
 
 	spin_lock_irqsave(&pdata->tstamp_lock, flags);
-
-	nsec = timecounter_read(&pdata->tstamp_tc);
-
-	nsec += delta;
-	timecounter_init(&pdata->tstamp_tc, &pdata->tstamp_cc, nsec);
-
+	timecounter_adjtime(&pdata->tstamp_tc, delta);
 	spin_unlock_irqrestore(&pdata->tstamp_lock, flags);
 
 	return 0;
Signature · 1 line
1.7.10.4
#5 lore raw

PATCH 4/11 net: bnx2x: convert to timecounter adjtime.

In reply to Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

Compile tested only.

---
 drivers/net/ethernet/broadcom/bnx2x/bnx2x_main.c |    6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
drivers/net/ethernet/broadcom/bnx2x/bnx2x_main.c +1 -5
index 691f0bf..7b2cc40 100644
@@ -13265,14 +13265,10 @@ static int bnx2x_ptp_adjfreq(struct ptp_clock_info *ptp, s32 ppb)
 static int bnx2x_ptp_adjtime(struct ptp_clock_info *ptp, s64 delta)
 {
 	struct bnx2x *bp = container_of(ptp, struct bnx2x, ptp_clock_info);
-	u64 now;
 
 	DP(BNX2X_MSG_PTP, "PTP adjtime called, delta = %llx\n", delta);
 
-	now = timecounter_read(&bp->timecounter);
-	now += delta;
-	/* Re-init the timecounter */
-	timecounter_init(&bp->timecounter, &bp->cyclecounter, now);
+	timecounter_adjtime(&bp->timecounter, delta);
 
 	return 0;
 }
Signature · 1 line
1.7.10.4
#6 lore raw

Re: [PATCH net-next 04/11] net: bnx2x: convert to timecounter adjtime.

Quoted 2 lines · Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

BTW, I noticed that this driver does not serialize access to the timecounter in any way, but probably it should. Ariel?

Thanks,
Richard

#7 lore raw

PATCH 5/11 net: fec: convert to timecounter adjtime.

In reply to Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

Compile tested only.

---
 drivers/net/ethernet/freescale/fec_ptp.c |   16 +---------------
 1 file changed, 1 insertion(+), 15 deletions(-)
drivers/net/ethernet/freescale/fec_ptp.c +1 -15
index 992c8c3..1f9cf23 100644
@@ -374,23 +374,9 @@ static int fec_ptp_adjtime(struct ptp_clock_info *ptp, s64 delta)
 	struct fec_enet_private *fep =
 	    container_of(ptp, struct fec_enet_private, ptp_caps);
 	unsigned long flags;
-	u64 now;
-	u32 counter;
 
 	spin_lock_irqsave(&fep->tmreg_lock, flags);
-
-	now = timecounter_read(&fep->tc);
-	now += delta;
-
-	/* Get the timer value based on adjusted timestamp.
-	 * Update the counter with the masked value.
-	 */
-	counter = now & fep->cc.mask;
-	writel(counter, fep->hwp + FEC_ATIME);
-
-	/* reset the timecounter */
-	timecounter_init(&fep->tc, &fep->cc, now);
-
+	timecounter_adjtime(&fep->tc, delta);
 	spin_unlock_irqrestore(&fep->tmreg_lock, flags);
 
 	return 0;
Signature · 1 line
1.7.10.4
#8 lore raw

PATCH 6/11 net: e1000e: convert to timecounter adjtime.

In reply to Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

---
 drivers/net/ethernet/intel/e1000e/ptp.c |    5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
drivers/net/ethernet/intel/e1000e/ptp.c +1 -4
index fb1a914..978ef9c 100644
@@ -90,12 +90,9 @@ static int e1000e_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
 	struct e1000_adapter *adapter = container_of(ptp, struct e1000_adapter,
 						     ptp_clock_info);
 	unsigned long flags;
-	s64 now;
 
 	spin_lock_irqsave(&adapter->systim_lock, flags);
-	now = timecounter_read(&adapter->tc);
-	now += delta;
-	timecounter_init(&adapter->tc, &adapter->cc, now);
+	timecounter_adjtime(&adapter->tc, delta);
 	spin_unlock_irqrestore(&adapter->systim_lock, flags);
 
 	return 0;
Signature · 1 line
1.7.10.4
Acked-by on v1 6/11
#9 lore raw

Re: [PATCH net-next 06/11] net: e1000e: convert to timecounter adjtime.

Quoted 4 lines · Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

Quoted 3 lines
---
 drivers/net/ethernet/intel/e1000e/ptp.c |    5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
#10 lore raw

PATCH 7/11 net: igb: convert to timecounter adjtime.

In reply to Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

---
 drivers/net/ethernet/intel/igb/igb_ptp.c |    7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)
drivers/net/ethernet/intel/igb/igb_ptp.c +1 -6
index 794c139..1d27f2d 100644
@@ -256,14 +256,9 @@ static int igb_ptp_adjtime_82576(struct ptp_clock_info *ptp, s64 delta)
 	struct igb_adapter *igb = container_of(ptp, struct igb_adapter,
 					       ptp_caps);
 	unsigned long flags;
-	s64 now;
 
 	spin_lock_irqsave(&igb->tmreg_lock, flags);
-
-	now = timecounter_read(&igb->tc);
-	now += delta;
-	timecounter_init(&igb->tc, &igb->cc, now);
-
+	timecounter_adjtime(&igb->tc, delta);
 	spin_unlock_irqrestore(&igb->tmreg_lock, flags);
 
 	return 0;
Signature · 1 line
1.7.10.4
Acked-by on v1 7/11
#11 lore raw

Re: [PATCH net-next 07/11] net: igb: convert to timecounter adjtime.

Quoted 4 lines · Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

Quoted 3 lines
---
 drivers/net/ethernet/intel/igb/igb_ptp.c |    7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)
#12 lore raw

PATCH 8/11 net: ixgbe: convert to timecounter adjtime.

In reply to Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

Compile tested only.

---
 drivers/net/ethernet/intel/ixgbe/ixgbe_ptp.c |   11 +----------
 1 file changed, 1 insertion(+), 10 deletions(-)
drivers/net/ethernet/intel/ixgbe/ixgbe_ptp.c +1 -10
index 5fd4b52..47c29ea 100644
@@ -261,18 +261,9 @@ static int ixgbe_ptp_adjtime(struct ptp_clock_info *ptp, s64 delta)
 	struct ixgbe_adapter *adapter =
 		container_of(ptp, struct ixgbe_adapter, ptp_caps);
 	unsigned long flags;
-	u64 now;
 
 	spin_lock_irqsave(&adapter->tmreg_lock, flags);
-
-	now = timecounter_read(&adapter->tc);
-	now += delta;
-
-	/* reset the timecounter */
-	timecounter_init(&adapter->tc,
-			 &adapter->cc,
-			 now);
-
+	timecounter_adjtime(&adapter->tc, delta);
 	spin_unlock_irqrestore(&adapter->tmreg_lock, flags);
 
 	ixgbe_ptp_setup_sdp(adapter);
Signature · 1 line
1.7.10.4
Acked-by on v1 8/11
#13 lore raw

Re: [PATCH net-next 08/11] net: ixgbe: convert to timecounter adjtime.

Quoted 6 lines · Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

Compile tested only.

Just for sanity sake, I will have Phillip test the changes, but there is no need to hold this up from being committed in the mean time.

Quoted 3 lines
---
 drivers/net/ethernet/intel/ixgbe/ixgbe_ptp.c |   11 +----------
 1 file changed, 1 insertion(+), 10 deletions(-)
#14 lore raw

Re: [PATCH net-next 08/11] net: ixgbe: convert to timecounter adjtime.

Quoted 2 lines · Jeff Kirsher

Just for sanity sake, I will have Phillip test the changes, but there is no need to hold this up from being committed in the mean time.

FYI, here is a test that may show the improved behavior when comparing before/after.

1. Start ptp4l, let it run long enough to find the frequency offset.

2. Stop ptp4l with ^C.

3. Wait a bit, to let the clock drift away.

4. Restart ptp4l, it will now start with the frequency offset learned in step 1.

5. Notice the the offsets in states s0 -> s1 -> s2. There will likely be less offset error with the patches applied.

Thanks,
Richard

#15 lore raw

PATCH 9/11 net: mlx4: convert to timecounter adjtime.

In reply to Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

Compile tested only.

---
 drivers/net/ethernet/mellanox/mlx4/en_clock.c |    5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
drivers/net/ethernet/mellanox/mlx4/en_clock.c +1 -4
index 9990144..df35d0e 100644
@@ -147,12 +147,9 @@ static int mlx4_en_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
 	struct mlx4_en_dev *mdev = container_of(ptp, struct mlx4_en_dev,
 						ptp_clock_info);
 	unsigned long flags;
-	s64 now;
 
 	write_lock_irqsave(&mdev->clock_lock, flags);
-	now = timecounter_read(&mdev->clock);
-	now += delta;
-	timecounter_init(&mdev->clock, &mdev->cycles, now);
+	timecounter_adjtime(&mdev->clock, delta);
 	write_unlock_irqrestore(&mdev->clock_lock, flags);
 
 	return 0;
Signature · 1 line
1.7.10.4
#16 lore raw

PATCH 10/11 net: cpts: convert to timecounter adjtime.

In reply to Richard Cochran

This patch changes the driver to use the new and improved method for adjusting the offset of a timecounter.

---
 drivers/net/ethernet/ti/cpts.c |    5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
drivers/net/ethernet/ti/cpts.c +1 -4
index 4a4388b..fbe42cb 100644
@@ -157,14 +157,11 @@ static int cpts_ptp_adjfreq(struct ptp_clock_info *ptp, s32 ppb)
 
 static int cpts_ptp_adjtime(struct ptp_clock_info *ptp, s64 delta)
 {
-	s64 now;
 	unsigned long flags;
 	struct cpts *cpts = container_of(ptp, struct cpts, info);
 
 	spin_lock_irqsave(&cpts->lock, flags);
-	now = timecounter_read(&cpts->tc);
-	now += delta;
-	timecounter_init(&cpts->tc, &cpts->cc, now);
+	timecounter_adjtime(&cpts->tc, delta);
 	spin_unlock_irqrestore(&cpts->lock, flags);
 
 	return 0;
Signature · 1 line
1.7.10.4
#17 lore raw

PATCH 11/11 timecounter: keep track of accumulated fractional nanoseconds

In reply to Richard Cochran

The current timecounter implementation will drop a variable amount of resolution, depending on the magnitude of the time delta. In other words, reading the clock too often or too close to a time stamp conversion will introduce errors into the time values. This patch fixes the issue by introducing a fractional nanosecond field that accumulates the low order bits.

---
 drivers/net/ethernet/mellanox/mlx4/en_clock.c |    4 ++--
 include/linux/timecounter.h                   |   19 +++++++++------
 kernel/time/timecounter.c                     |   31 +++++++++++++++++++------
 virt/kvm/arm/arch_timer.c                     |    3 ++-
 4 files changed, 40 insertions(+), 17 deletions(-)
drivers/net/ethernet/mellanox/mlx4/en_clock.c +2 -2
index df35d0e..e9cce4f 100644
@@ -240,7 +240,7 @@ void mlx4_en_init_timestamp(struct mlx4_en_dev *mdev)
 {
 	struct mlx4_dev *dev = mdev->dev;
 	unsigned long flags;
-	u64 ns;
+	u64 ns, zero = 0;
 
 	rwlock_init(&mdev->clock_lock);
 
@@ -265,7 +265,7 @@ void mlx4_en_init_timestamp(struct mlx4_en_dev *mdev)
 	/* Calculate period in seconds to call the overflow watchdog - to make
 	 * sure counter is checked at least once every wrap around.
 	 */
-	ns = cyclecounter_cyc2ns(&mdev->cycles, mdev->cycles.mask);
+	ns = cyclecounter_cyc2ns(&mdev->cycles, mdev->cycles.mask, zero, &zero);
 	do_div(ns, NSEC_PER_SEC / 2 / HZ);
 	mdev->overflow_period = ns;
 
include/linux/timecounter.h +12 -7
index af3dfa4..74f4549 100644
@@ -55,27 +55,32 @@ struct cyclecounter {
  * @cycle_last:		most recent cycle counter value seen by
  *			timecounter_read()
  * @nsec:		continuously increasing count
+ * @mask:		bit mask for maintaining the 'frac' field
+ * @frac:		accumulated fractional nanoseconds
  */
 struct timecounter {
 	const struct cyclecounter *cc;
 	cycle_t cycle_last;
 	u64 nsec;
+	u64 mask;
+	u64 frac;
 };
 
 /**
  * cyclecounter_cyc2ns - converts cycle counter cycles to nanoseconds
  * @cc:		Pointer to cycle counter.
  * @cycles:	Cycles
- *
- * XXX - This could use some mult_lxl_ll() asm optimization. Same code
- * as in cyc2ns, but with unsigned result.
+ * @mask:	bit mask for maintaining the 'frac' field
+ * @frac:	pointer to storage for the fractional nanoseconds.
  */
 static inline u64 cyclecounter_cyc2ns(const struct cyclecounter *cc,
-				      cycle_t cycles)
+				      cycle_t cycles, u64 mask, u64 *frac)
 {
-	u64 ret = (u64)cycles;
-	ret = (ret * cc->mult) >> cc->shift;
-	return ret;
+	u64 ns = (u64) cycles;
+
+	ns = (ns * cc->mult) + *frac;
+	*frac = ns & mask;
+	return ns >> cc->shift;
 }
 
 /**
kernel/time/timecounter.c +24 -7
index 59a1ec3..4687b31 100644
@@ -25,6 +25,8 @@ void timecounter_init(struct timecounter *tc,
 	tc->cc = cc;
 	tc->cycle_last = cc->read(cc);
 	tc->nsec = start_tstamp;
+	tc->mask = (1ULL << cc->shift) - 1;
+	tc->frac = 0;
 }
 EXPORT_SYMBOL_GPL(timecounter_init);
 
@@ -51,7 +53,8 @@ static u64 timecounter_read_delta(struct timecounter *tc)
 	cycle_delta = (cycle_now - tc->cycle_last) & tc->cc->mask;
 
 	/* convert to nanoseconds: */
-	ns_offset = cyclecounter_cyc2ns(tc->cc, cycle_delta);
+	ns_offset = cyclecounter_cyc2ns(tc->cc, cycle_delta,
+					tc->mask, &tc->frac);
 
 	/* update time stamp of timecounter_read_delta() call: */
 	tc->cycle_last = cycle_now;
@@ -72,22 +75,36 @@ u64 timecounter_read(struct timecounter *tc)
 }
 EXPORT_SYMBOL_GPL(timecounter_read);
 
+/*
+ * This is like cyclecounter_cyc2ns(), but it is used for computing a
+ * time previous to the time stored in the cycle counter.
+ */
+static u64 cc_cyc2ns_backwards(const struct cyclecounter *cc,
+			       cycle_t cycles, u64 mask, u64 frac)
+{
+	u64 ns = (u64) cycles;
+
+	ns = ((ns * cc->mult) - frac) >> cc->shift;
+
+	return ns;
+}
+
 u64 timecounter_cyc2time(struct timecounter *tc,
 			 cycle_t cycle_tstamp)
 {
-	u64 cycle_delta = (cycle_tstamp - tc->cycle_last) & tc->cc->mask;
-	u64 nsec;
+	u64 delta = (cycle_tstamp - tc->cycle_last) & tc->cc->mask;
+	u64 nsec = tc->nsec, frac = tc->frac;
 
 	/*
 	 * Instead of always treating cycle_tstamp as more recent
 	 * than tc->cycle_last, detect when it is too far in the
 	 * future and treat it as old time stamp instead.
 	 */
-	if (cycle_delta > tc->cc->mask / 2) {
-		cycle_delta = (tc->cycle_last - cycle_tstamp) & tc->cc->mask;
-		nsec = tc->nsec - cyclecounter_cyc2ns(tc->cc, cycle_delta);
+	if (delta > tc->cc->mask / 2) {
+		delta = (tc->cycle_last - cycle_tstamp) & tc->cc->mask;
+		nsec -= cc_cyc2ns_backwards(tc->cc, delta, tc->mask, frac);
 	} else {
-		nsec = cyclecounter_cyc2ns(tc->cc, cycle_delta) + tc->nsec;
+		nsec += cyclecounter_cyc2ns(tc->cc, delta, tc->mask, &frac);
 	}
 
 	return nsec;
virt/kvm/arm/arch_timer.c +2 -1
index 22fa819..75d9564 100644
@@ -150,7 +150,8 @@ void kvm_timer_sync_hwstate(struct kvm_vcpu *vcpu)
 		return;
 	}
 
-	ns = cyclecounter_cyc2ns(timecounter->cc, cval - now);
+	ns = cyclecounter_cyc2ns(timecounter->cc, cval - now, timecounter->mask,
+				 &timecounter->frac);
 	timer_arm(timer, ns);
 }
 
Signature · 1 line
1.7.10.4
maintainer
#18 lore raw
In reply to Richard Cochran

From: Richard Cochran <[email protected]>
Date: Sun, 21 Dec 2014 19:46:55 +0100

Quoted 5 lines

Several PTP Hardware Clock (PHC) drivers implement the clock in software using the timecounter/cyclecounter code. This series adds one simple improvement and one more subtle fix to the shared timecounter facility. Credit for this series goes to Janusz Użycki, who pointed the issues out to me off list.

Series applied, thanks Richard.

#19 lore raw
Quoted 2 lines · David Miller

Series applied, thanks Richard.

I got an automated email from

kbuild test robot <[email protected]>

showing new warnings and errors introduced by this series. I'll follow up with fixes soon.

Thanks,
Richard

Acked-by on the v1 series
#20 lore raw
In reply to Richard Cochran
Quoted 10 lines · Richard Cochran

Several PTP Hardware Clock (PHC) drivers implement the clock in software using the timecounter/cyclecounter code. This series adds one simple improvement and one more subtle fix to the shared timecounter facility. Credit for this series goes to Janusz Użycki, who pointed the issues out to me off list.

Patch #1 simply move the timecounter code into its own file. When working on this series, it was really annoying to see half the kernel recompile after every tweak to the timecounter stuff. There is no reason to keep this together with the clocksource code.

I did have some faint hope we could merge the cyclecounter and clocksource struct at some point, and while we've gotten closer with much of the time-specific values being kept in the timekeeper, the full cleanup hasn't happened in a few years here, so its clearly not something I've prioritized.

So no objection in concept.

Quoted 39 lines

Patch #2 implements an improved adjtime() method, and patches 3-10 convert all of the drivers over to the new method.

Patch #11 fixes a subtle but important issue with the timecounter WRT frequency adjustment. As it stands now, a timecounter based PHC will exhibit a variable frequency resolution (and variable time error) depending on how often the clock is read.

In timecounter_read_delta(), the expression

   (delta * cc->mult) >> cc->shift;

can lose resolution from the adjusted value of 'mult'. If the value of 'delta' is too small, then small changes in 'mult' have no effect. However, if the delta value is large enough, then small changes in 'mult' will have an effect.

Reading the clock too often means smaller 'delta' values which in turn will spoil the fine adjustments made to 'mult'. Up until now, this effect did not show up in my testing. The following example explains why.

The CPTS has an input clock of 250 MHz, and the clock source uses mult=0x80000000 and shift=29, making the ticks to nanoseconds conversion like this:

   ticks * 2^31
   ------------
       2^29

Imagine what happens if the clock is read every 10 milliseconds. Ten milliseconds are about 2500000 ticks, which corresponds to about 21 bits. The product in the numerator has then 52 bits. After the shift operation, 23 bits are preserved. This results in a frequency adjustment resolution of about 0.1 ppm (not _too_ bad.)

A frequency resolution of 1 ppm requires 20 bits.
A frequency resolution of 1 ppb requires 30 bits.

Quoted 7 lines

For the 250 MHz CPTS clock, reading every 4 seconds yields a 1 ppb resolution (which is the finest that our API allows).

However, the error can be much higher if the clock is read too often or if time stamps occur close in time to read operations. In general it is really not acceptable to allow the rate of clock readings to influence the clock accuracy.

Though, this does start to sound like issues the timekeeping code had to resolve, so while its probably time to let some flowers bloom and see what happens, we should be somewhat watchful for too much logic duplication.

The patch set itself seems fairly straight forward (at least to my back from vacation brain), so..

thanks
-john

#21 lore raw
Quoted 4 lines · John Stultz

Though, this does start to sound like issues the timekeeping code had to resolve, so while its probably time to let some flowers bloom and see what happens, we should be somewhat watchful for too much logic duplication.

You took the words right out of my mouth.

My feeling is that this timecounter code might become more complicated, and if the core time keeping becomes more straightforward, then maybe one day the two will join.

Thanks,
Richard