Patchset enabling hardware based cross-timestamps for next gen Intel platforms
Subsystems (7)
v5 cover Patchset enabling hardware based cross-timestamps for next gen Intel platforms
Modern Intel hardware adds an Always Running Timer (ART) that allows the network and audio device clocks to precisely cross timestamp the device clock with the system clock. This allows a precise correlation of the device time and system time.
This patchset adds interfaces to the timekeeping code allowing drivers to translate ART time to system time.
Changelog:
Changes from v4 to v5:
* Changes the history mechanism to interpolate system time using
a single historic system time pair (monotonic raw, realtime) rather than implementing a precise history using shadow timekeeper (see v4 changes). The advantage of this approach is that the history can be arbitrarily long. This approach may also be simpler in terms of coding. The major disadvantage is that the realtime clock can be adjusted. When adjusted, the realtime clock time (when interpolating from history) is always approximate. In general, the longer the interpolation period the larger the potential error. There isn't any error interpolating the monotonic raw clock time.
* This patchset also addresses objections to the previous
patchsets overly complex correlated timestamp structure. This patchset splits that structure into several smaller structures. The correlated timestamp interface is renamed cross timestamp to avoid any confusion with the correlated clocksource.
* The correlated clocksource is separated from the cross
timestamp mechanism.
* Add monotonic raw to the PTP user interface
* Add e1000e driver configuration option that wraps Intel PCH
specific code
Changes v3 to v4:
* Adds a history mechanism to accomodate slower devices. In this
case the response time for timestamp reads to the Intel DSP are too slow to be accomodated by the original correlated time mechanism. The history mechanism turns shadow timekeeper into an array where the history is stored.
Christopher S. Hall (6):
Timekeeping cross timestamp interface for device drivers Always Running Timer (ART) correlated clocksource
Add history to cross timestamp interface supporting slower devices Remove duplicate code from ktime_get_raw_and_real code Add PTP_SYS_OFFSET_PRECISE for driver crosstimestamping Adds hardware supported cross timestamp
Documentation/ptp/testptp.c | 6 +- arch/x86/include/asm/cpufeature.h | 2 +- arch/x86/include/asm/tsc.h | 2 + arch/x86/kernel/tsc.c | 46 ++++++ drivers/net/ethernet/intel/Kconfig | 9 ++ drivers/net/ethernet/intel/e1000e/defines.h | 5 + drivers/net/ethernet/intel/e1000e/ptp.c | 92 +++++++++++++++++++++++++++++ drivers/net/ethernet/intel/e1000e/regs.h | 4 + drivers/ptp/ptp_chardev.c | 31 ++++ include/linux/clocksource.h | 43 +++++ include/linux/ptp_clock_kernel.h | 8 + include/linux/timekeeper_internal.h | 4 + include/linux/timekeeping.h | 42 +++++ include/uapi/linux/ptp_clock.h | 13 +- kernel/time/timekeeping.c | 243 ++++++++++++++++++++++++++-- 15 files changed, 521 insertions(+), 21 deletions(-)
Signature · 1 line
2.1.4
RFC v5 1/6 Timekeeping cross timestamp interface for device drivers
ACKNOWLEDGMENT: The original correlated clock source and cross timestamp code was developed by Thomas Gleixner <[email protected]>. It has changed considerably and any mistakes are mine.
The precision with which events on multiple networked systems can be synchronized using, as an example, PTP (IEEE 1588, 802.1AS) is limited by the precision of the cross timestamps between the system clock and the device (timestamp) clock. Precision here is the degree of simultaneity when capturing the cross timestamp.
Currently the PTP cross timestamp is captured in software using the PTP device driver ioctl PTP_SYS_OFFSET. Reads of the device clock are interleaved with reads of the realtime clock. At best, the precision of this cross timestamp is on the order of several microseconds due to software latencies. Sub-microsecond precision is required for industrial control and some media applications. To achieve this level of precision hardware supported cross timestamping is needed.
Hardware cross timestamps are derived from simultaneously capturing the device clock and the system clock. Applications use timestamps in nanoseconds rather than clock ticks. The device driver can scale device clock ticks to device time in nanoseconds, but cannot transform system clock ticks. Only the kernel timekeeping code can do this. The function get_device_system_crosststamp() allows device drivers to return a cross timestamp with system time properly scaled to nanoseconds.
The cross timestamps contain the realtime and monotonic raw clock times. The realtime value is needed to discipline that clock using PTP and the monotonic raw value is used for applications that don't require a "real" time, but need an unadjusted clock time.
The get_device_system_crosststamp() code calls back into the driver to ensure that the system counter is within the current timekeeping update interval. Because of possible NTP/PTP frequency adjustments, extrapolating a realtime clock time outside the current interval with a potentially different scaling factor can result in a small amount of error.
Modern Intel hardware provides an Always Running Timer (ART) which is exactly related to TSC through a known frequency ratio. The ART is routed to devices on the system and is used to precisely and simultaneously capture the device clock with the ART. The kernel timekeeping code requires a system clock value from the current clock source to calculate the system time. The correlated clocksource adds a means to transform a clock (in this case ART) exactly to the system clock (TSC clocksource) that is used to calculate system time.
--- include/linux/clocksource.h | 27 ++++++++++++++++ include/linux/timekeeping.h | 35 +++++++++++++++++++++ kernel/time/timekeeping.c | 77 ++++++++++++++++++++++++++++++++++++++++++--- 3 files changed, 135 insertions(+), 4 deletions(-)
include/linux/clocksource.h +27 -0
@@ -255,4 +255,31 @@ static inline void clocksource_probe(void) {}#define CLOCKSOURCE_ACPI_DECLARE (name, table_id, fn) \ ACPI_DECLARE_PROBE_ENTRY(clksrc, name, table_id, 0, NULL, 0, fn) +/* +* struct correlated_cs - Descriptor for a clocksource correlated to another +* clocksource +* @related_cs: Pointer to the related timekeeping clocksource +* @convert: Conversion function to convert a timestamp from +* the correlated clocksource to cycles of the related +* timekeeping clocksource +*/ +struct correlated_cs { +struct clocksource *related_cs; +cycle_t (*convert)(struct correlated_cs *cs, +cycle_t cycles); +}; + +/* +* struct raw_system_counterval - system counter value captured in device +* driver used to produce system/device cross-timestamp +* @system: System counter value +* @cs: Clocksource related to system counter value. This is used by +* timekeeping code to verify validity of counter for system time +* conversion +*/ +struct raw_system_counterval { +cycle_t cycles; +struct clocksource *cs; +}; +#endif /* _LINUX_CLOCKSOURCE_H */
include/linux/timekeeping.h +35 -0
@@ -266,6 +266,41 @@ extern void timekeeping_inject_sleeptime64(struct timespec64 *delta);extern void ktime_get_raw_and_real_ts64 (struct timespec64 *ts_raw,struct timespec64 *ts_real); + +/* +* struct system_device_crosststamp - system/device cross-timestamp +* (syncronized capture) +* @device: Device time +* @realtime: Realtime simultaneous with device time +* @monoraw: Monotonic raw simultaneous with device time +*/ +struct system_device_crosststamp { +ktime_t device; +ktime_t sys_realtime; +ktime_t sys_monoraw; +}; + +struct raw_system_counterval ; +/* +* struct get_sync_device_time - Provides method to capture device time +* synchronized with raw system counter value +* @get_time: Callback providing synchronized capture of device time +* and system counter. Returns 0 on success, < 0 on failure +* @ctx: Context provided to callback function +*/ +struct get_sync_device_time { +int (*get_time)(ktime_t *device, +struct raw_system_counterval *system, +void *ctx); +void *ctx; +}; + +/* +* Get cross timestamp between system clock and device clock +*/ +extern int get_device_system_crosststamp (struct system_device_crosststamp *ct, +struct get_sync_device_time *dt); +/* * Persistent clock related interfaces */
kernel/time/timekeeping.c +73 -4
98 lines not on this page · raw message
Signature · 1 line
2.1.4
Re: [RFC v5 1/6] Timekeeping cross timestamp interface for device drivers
Quoted 124 lines · Christopher S. Hall
ACKNOWLEDGMENT: The original correlated clock source and cross timestamp code was developed by Thomas Gleixner <[email protected]>. It has changed considerably and any mistakes are mine.
The precision with which events on multiple networked systems can be synchronized using, as an example, PTP (IEEE 1588, 802.1AS) is limited by the precision of the cross timestamps between the system clock and the device (timestamp) clock. Precision here is the degree of simultaneity when capturing the cross timestamp.
Currently the PTP cross timestamp is captured in software using the PTP device driver ioctl
PTP_SYS_OFFSET. Reads of the device clock are interleaved with reads of the realtime clock. At best, the precision of this cross timestamp is on the order of several microseconds due to software latencies. Sub-microsecond precision is required for industrial control and some media applications. To achieve this level of precision hardware supported cross timestamping is needed.Hardware cross timestamps are derived from simultaneously capturing the device clock and the system clock. Applications use timestamps in nanoseconds rather than clock ticks. The device driver can scale device clock ticks to device time in nanoseconds, but cannot transform system clock ticks. Only the kernel timekeeping code can do this. The function
get_device_system_crosststamp()allows device drivers to return a cross timestamp with system time properly scaled to nanoseconds.The cross timestamps contain the realtime and monotonic raw clock times. The realtime value is needed to discipline that clock using PTP and the monotonic raw value is used for applications that don't require a "real" time, but need an unadjusted clock time.
The
get_device_system_crosststamp()code calls back into the driver to ensure that the system counter is within the current timekeeping update interval. Because of possible NTP/PTP frequency adjustments, extrapolating a realtime clock time outside the current interval with a potentially different scaling factor can result in a small amount of error.Modern Intel hardware provides an Always Running Timer (ART) which is exactly related to TSC through a known frequency ratio. The ART is routed to devices on the system and is used to precisely and simultaneously capture the device clock with the ART. The kernel timekeeping code requires a system clock value from the current clock source to calculate the system time. The correlated clocksource adds a means to transform a clock (in this case ART) exactly to the system clock (TSC clocksource) that is used to calculate system time.
Signed-off-by Christopher S. Hall <[email protected]>--- include/linux/clocksource.h | 27 ++++++++++++++++ include/linux/timekeeping.h | 35 +++++++++++++++++++++ kernel/time/timekeeping.c | 77 ++++++++++++++++++++++++++++++++++++++++++--- 3 files changed, 135 insertions(+), 4 deletions(-)include/linux/clocksource.h +27
-0@@ -255,4 +255,31 @@ static inline void clocksource_probe(void) {}#define CLOCKSOURCE_ACPI_DECLARE (name, table_id, fn) \ ACPI_DECLARE_PROBE_ENTRY(clksrc, name, table_id, 0, NULL, 0, fn) +/* +* struct correlated_cs - Descriptor for a clocksource correlated to another +* clocksource +* @related_cs: Pointer to the related timekeeping clocksource +* @convert: Conversion function to convert a timestamp from +* the correlated clocksource to cycles of the related +* timekeeping clocksource +*/ +struct correlated_cs { +struct clocksource *related_cs; +cycle_t (*convert)(struct correlated_cs *cs, +cycle_t cycles); +}; + +/* +* struct raw_system_counterval - system counter value captured in device +* driver used to produce system/device cross-timestamp +* @system: System counter value +* @cs: Clocksource related to system counter value. This is used by +* timekeeping code to verify validity of counter for system time +* conversion +*/ +struct raw_system_counterval { +cycle_t cycles; +struct clocksource *cs; +}; +#endif /* _LINUX_CLOCKSOURCE_H */ include/linux/timekeeping.h +24
-0@@ -266,6 +266,41 @@ extern void timekeeping_inject_sleeptime64(struct timespec64 *delta);extern void ktime_get_raw_and_real_ts64 (struct timespec64 *ts_raw,struct timespec64 *ts_real); + +/* +* struct system_device_crosststamp - system/device cross-timestamp +* (syncronized capture) +* @device: Device time +* @realtime: Realtime simultaneous with device time +* @monoraw: Monotonic raw simultaneous with device time +*/ +struct system_device_crosststamp { +ktime_t device; +ktime_t sys_realtime; +ktime_t sys_monoraw; +}; + +struct raw_system_counterval ; +/* +* struct get_sync_device_time - Provides method to capture device time +* synchronized with raw system counter value +* @get_time: Callback providing synchronized capture of device time +* and system counter. Returns 0 on success, < 0 on failure +* @ctx: Context provided to callback function +*/ +struct get_sync_device_time { + int (*get_time)(ktime_t *device,
include/linux/timekeeping.h extern void timekeeping_inject_sleeptime64(struct timespec64 *delta); · line 302
Quoted 10 lines
include/linux/timekeeping.h +10
-0+struct raw_system_counterval *system, + void *ctx); +void *ctx; +}; + +/* +* Get cross timestamp between system clock and device clock +*/ +extern int get_device_system_crosststamp (struct system_device_crosststamp *ct, +struct get_sync_device_time *dt);
I feel like this is introducing a lot of very small structures, which to the casual reviewer aren't immediately obvious how they interlink and are used. It also adds to the number of types we have to keep in our head. I dunno, maybe its just the correlated_cs is added but isn't used in this patch, but I keep feeling like these should be more obvious somehow.
That's terribly vague feedback, so I'm sorry. Maybe some concrete suggestions?
* Maybe try renaming get_sync_device_time as just crosststamp_device/struct ?
* raw_system_counterval feels like its of limited value. Maybe just pass the clocksource and cycle value to the functions separately?
Quoted 35 lines
include/linux/timekeeping.h +0
-0/* * Persistent clock related interfaces */ kernel/time/timekeeping.c +10
-4@@ -298,13 +298,11 @@ u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;static inline u32 arch_gettimeoffset (void ) {return 0 ; }#endif -static inline s64 timekeeping_get_ns (struct tk_read_base *tkr) +static inline s64 timekeeping_delta_to_ns (struct tk_read_base *tkr, +cycle_t delta) { -cycle_t delta;s64 nsec; - delta =timekeeping_get_delta (tkr); - nsec = delta * tkr->mult + tkr->xtime_nsec; nsec >>= tkr->shift;@@ -312,6 +310,24 @@ static inline s64 timekeeping_get_ns(struct tk_read_base *tkr)return nsec +arch_gettimeoffset (); } +static inline s64 timekeeping_get_ns (struct tk_read_base *tkr) +{ +cycle_t delta; + + delta =timekeeping_get_delta (tkr); +return timekeeping_delta_to_ns (tkr, delta); +} +
kernel/time/timekeeping.c static inline s64 timekeeping_get_ns(struct tk_read_base *tkr) · line 330
Quoted 10 lines
kernel/time/timekeeping.c +9
-0+static inline s64 timekeeping_cycles_to_ns (struct tk_read_base *tkr, +cycle_t cycles) +{ +cycle_t delta; + +/* calculate the delta since the last update_wall_time */ + delta =clocksource_delta (cycles, tkr->cycle_last, tkr->mask); +return timekeeping_delta_to_ns (tkr, delta); +}
Mind splitting the above out into its own small patch?
Quoted 6 lines
kernel/time/timekeeping.c +0
-0/** * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper. * @tkr: Timekeeping readout base from which we take the update @@ -885,6 +901,59 @@ EXPORT_SYMBOL(ktime_get_raw_and_real_ts64);#endif /* CONFIG_NTP_PPS */
kernel/time/timekeeping.c EXPORT_SYMBOL(ktime_get_raw_and_real_ts64); · line 912
Quoted 10 lines
kernel/time/timekeeping.c +9
-0/** +* get_device_system_crosststamp - Synchronously capture system/device timestamp +* @xtstamp: Receives simultaneously captured system and device time +* @sync_devicetime: Callback to get simultaneous device time and +* system counter from the device driver +* +* Reads a timestamp from a device and correlates it to system time +*/ +int get_device_system_crosststamp (struct system_device_crosststamp *xtstamp, +struct get_sync_device_time *sync_devicetime)
Another nit, but to me something like:
int get_device_system_crosststamp(struct get_sync_device_time *sync_devicetime,
structsystem_device_crosststamp *ret)
Reads better to me. ie: use this device_time -> return that crosststamp.
Otherwise this is looking pretty good.
thanks
-john
Re: [RFC v5 1/6] Timekeeping cross timestamp interface for device drivers
Hi John,
Quoted 18 lines · John Stultz
Quoted 11 lines · Christopher S. Hall
+/*
+ * structcorrelated_cs- Descriptor for a clocksource correlated to another
+ * clocksource
+ * @related_cs: Pointer to the related timekeeping
clocksource
+ * @convert: Conversion function to convert a timestamp from
+ * the correlated clocksource to cycles of the
related
+ * timekeeping clocksource
+ */include/linux/clocksource.h static inline void clocksource_probe(void) {} · line 270
Quoted 5 lines
include/linux/clocksource.h +5
-0+struct correlated_cs { +struct clocksource *related_cs; +cycle_t (*convert)(struct correlated_cs *cs, +cycle_t cycles); +};
How about making this another patch? It's not directly related to the cross timestamp. I would lean toward making it its own patch leaving the ART correlated clocksource patch touching arch/ only.
Quoted 70 lines
Quoted 52 lines
+/*
+ * structraw_system_counterval- system counter value captured in device
+ * driver used to produce system/device cross-timestamp
+ * @system: System counter value
+ * @cs: Clocksource related to system counter value.
This is used by
+ * timekeeping code to verify validity of counter for
system time
+ * conversion
+ */include/linux/clocksource.h +5
-0+struct raw_system_counterval { +cycle_t cycles; +struct clocksource *cs; +}; +#endif /* _LINUX_CLOCKSOURCE_H */ include/linux/timekeeping.h +26
-0@@ -266,6 +266,41 @@ extern void timekeeping_inject_sleeptime64(struct timespec64 *delta);extern void ktime_get_raw_and_real_ts64 (struct timespec64 *ts_raw,struct timespec64 *ts_real); + +/* +* struct system_device_crosststamp - system/device cross-timestamp +* (syncronized capture) +* @device: Device time +* @realtime: Realtime simultaneous with device time +* @monoraw: Monotonic raw simultaneous with device time +*/ +struct system_device_crosststamp { +ktime_t device; +ktime_t sys_realtime; +ktime_t sys_monoraw; +}; + +struct raw_system_counterval ; +/* +* struct get_sync_device_time - Provides method to capture device time +* synchronized with raw system counter value +* @get_time: Callback providing synchronized capture of device time +* and system counter. Returns 0 on success, < 0 on failure +* @ctx: Context provided to callback function +*/ +struct get_sync_device_time { +int (*get_time)(ktime_t *device, +struct raw_system_counterval *system, +void *ctx);include/linux/timekeeping.h extern void timekeeping_inject_sleeptime64(struct · line 305
Quoted 10 lines
include/linux/timekeeping.h +8
-0+void *ctx; +}; + +/* +* Get cross timestamp between system clock and device clock +*/ +extern int get_device_system_crosststamp (struct system_device_crosststamp *ct, +struct get_sync_device_time *dt);I feel like this is introducing a lot of very small structures, which to the casual reviewer aren't immediately obvious how they interlink and are used. It also adds to the number of types we have to keep in our head. I dunno, maybe its just the
correlated_csis added but isn't used in this patch, but I keep feeling like these should be more obvious somehow.
Quoted 5 lines
That's terribly vague feedback, so I'm sorry. Maybe some concrete suggestions?
* Maybe try renaming
get_sync_device_timeas just
crosststamp_device/struct ?
My goal here is to differentiate between the cross timestamp (final result) and the synchronized capture (intermediate value). Maybe it isn't particularly useful (or worse additionally confusing), but when reading the code I found it confusing to call everything a cross timestamp. I think it's the units - cycles and nanoseconds - that are the source of my confusion in this case. The units are obvious from the struct member types, but it seemed more straightforward to differentiate in the struct name. Does this make sense? If I missed the mark maybe it should be changed.
Quoted 2 lines
*
raw_system_countervalfeels like its of limited value. Maybe just pass the clocksource and cycle value to the functions separately?
I agree it is of somewhat limited value, but I think it makes the cross timestamp code more intuitive (the header file, as you pointed out, not as much). Maybe better commenting of the struct declaration would be useful here? To your comment above, I can keep more structs in my head than arguments. Also, to my thinking there isn't any intuitive ordering if these struct members are changed to individual arguments which might make the driver callback confusing.
Quoted 13 lines
Quoted 13 lines
include/linux/timekeeping.h +2
-4-static inline s64 timekeeping_get_ns (struct tk_read_base *tkr) +static inline s64 timekeeping_delta_to_ns (struct tk_read_base *tkr, +cycle_t delta) { -cycle_t delta;s64 nsec; - delta =timekeeping_get_delta (tkr); - nsec = delta * tkr->mult + tkr->xtime_nsec; nsec >>= tkr->shift;
Quoted 6 lines
include/linux/timekeeping.h static inline s64 timekeeping_get_ns(struct · line 312
Quoted 4 lines
include/linux/timekeeping.h +0
-0@@ -312,6 +310,24 @@ static inline s64 timekeeping_get_ns(struct tk_read_base *tkr)return nsec +arch_gettimeoffset (); }Mind splitting the above out into its own small patch?
Makes sense. To be clear your suggestion is to make this patch 1 of X and then 2 of X will be the cross timestamp patch. Is that right?
Quoted 28 lines
Quoted 13 lines
include/linux/timekeeping.h +0
-0/** * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
* @tkr: Timekeeping readout base from which we take the updateinclude/linux/timekeeping.h +1
-0@@ -885,6 +901,59 @@ EXPORT_SYMBOL(ktime_get_raw_and_real_ts64);#endif /* CONFIG_NTP_PPS */ /** +* get_device_system_crosststamp - Synchronously capture system/device timestamp
+ * @xtstamp: Receives simultaneously captured system and
device time
+ * @sync_devicetime: Callback to get simultaneous device time andQuoted 8 lines
+ * system counter from the device driver
+ *
+ * Reads a timestamp from a device and correlates it to system time
+ */
+int get_device_system_crosststamp(structsystem_device_crosststamp*xtstamp,
+ structget_sync_device_time*sync_devicetime)Another nit, but to me something like:
int get_device_system_crosststamp(structget_sync_device_time*sync_devicetime,
structsystem_device_crosststamp*ret)
Quoted 1 line
Reads better to me. ie: use this
device_time-> return that crosststamp.
OK. My brain "works" the opposite way, but I think (after a small amount of research) I'm in the minority:)
Thanks,
Chris
Re: [RFC v5 1/6] Timekeeping cross timestamp interface for device drivers
Quoted 21 lines · Christopher Hall
Hi John,
Quoted 17 lines · John Stultz
Quoted 14 lines · Christopher S. Hall
+/*
+ * structcorrelated_cs- Descriptor for a clocksource correlated to another
+ * clocksource
+ * @related_cs: Pointer to the related timekeeping
clocksource
+ * @convert: Conversion function to convert a timestamp from
+ * the correlated clocksource to cycles of the
related
+ * timekeeping clocksource
+ */include/linux/clocksource.h +2
-0+struct correlated_cs { +struct clocksource *related_cs;
Quoted 8 lines
Quoted 3 lines
include/linux/clocksource.h static inline void clocksource_probe(void) {} · line 270
Quoted 3 lines
include/linux/clocksource.h +3
-0+cycle_t (*convert)(struct correlated_cs *cs, +cycle_t cycles); +};How about making this another patch? It's not directly related to the cross timestamp. I would lean toward making it its own patch leaving the ART correlated clocksource patch touching
arch/only.
Sure. Tying it more closely (via explicit commit message) to the change that uses it would be better then kind of slipping it in here.
Quoted 14 lines
Quoted 7 lines
I feel like this is introducing a lot of very small structures, which to the casual reviewer aren't immediately obvious how they interlink and are used. It also adds to the number of types we have to keep in our head. I dunno, maybe its just the
correlated_csis added but isn't used in this patch, but I keep feeling like these should be more obvious somehow.Quoted 5 lines
That's terribly vague feedback, so I'm sorry. Maybe some concrete suggestions?
* Maybe try renaming
get_sync_device_timeas just
crosststamp_device/struct ?
Quoted 8 lines
My goal here is to differentiate between the cross timestamp (final result) and the synchronized capture (intermediate value). Maybe it isn't particularly useful (or worse additionally confusing), but when reading the code I found it confusing to call everything a cross timestamp. I think it's the units - cycles and nanoseconds - that are the source of my confusion in this case. The units are obvious from the struct member types, but it seemed more straightforward to differentiate in the struct name. Does this make sense? If I missed the mark maybe it should be changed.
Sure. I think having a different type then just cycles here is good, and counterval is fine for an alternate name, but get_sync_device_time looks like a function, not a structure name to me. The fact that the structure is mostly an accessor to the hardware is fine, but maybe counterval_device or something would be better? counterval_source? I dunno.
Quoted 11 lines
Quoted 2 lines
*
raw_system_countervalfeels like its of limited value. Maybe just pass the clocksource and cycle value to the functions separately?I agree it is of somewhat limited value, but I think it makes the cross timestamp code more intuitive (the header file, as you pointed out, not as much). Maybe better commenting of the struct declaration would be useful here? To your comment above, I can keep more structs in my head than arguments. Also, to my thinking there isn't any intuitive ordering if these struct members are changed to individual arguments which might make the driver callback confusing.
Yea. I just feel like lots of structures add extra abstraction that makes the code harder to learn or re-learn. Not only is one trying to remember the base types that are being passed around, but you also have to remember what all the meta-structures are for. Especially for these two-value structures.
And I apologize. I know this sort of nit-picking is sort of the worst. "Its great you named your child that, but I'd prefer to call them..." ;) But for long-term maintenance, under the fog of time, obviousness is really important.
Quoted 21 lines
Quoted 12 lines
Quoted 12 lines
kernel/time/timekeeping.c +2
-4-static inline s64 timekeeping_get_ns (struct tk_read_base *tkr) +static inline s64 timekeeping_delta_to_ns (struct tk_read_base *tkr, +cycle_t delta) { -cycle_t delta;s64 nsec; - delta =timekeeping_get_delta (tkr); - nsec = delta * tkr->mult + tkr->xtime_nsec; nsec >>= tkr->shift;Quoted 7 lines
kernel/time/timekeeping.c static inline s64 timekeeping_get_ns(struct · line 312
Quoted 4 lines
kernel/time/timekeeping.c +0
-0@@ -312,6 +310,24 @@ static inline s64 timekeeping_get_ns(struct tk_read_base *tkr)return nsec +arch_gettimeoffset (); }Mind splitting the above out into its own small patch?
Quoted 2 lines
Makes sense. To be clear your suggestion is to make this patch 1 of X and then 2 of X will be the cross timestamp patch. Is that right?
Yep.
Quoted 29 lines
Quoted 29 lines
Quoted 13 lines
(quoted diff) +0
-0/** * update_fast_timekeeper - Update the fast and NMI safe monotonictimekeeper.
* @tkr: Timekeeping readout base from which we take the updatekernel/time/timekeeping.c +1
-0@@ -885,6 +901,59 @@ EXPORT_SYMBOL(ktime_get_raw_and_real_ts64);#endif /* CONFIG_NTP_PPS */ /** +* get_device_system_crosststamp - Synchronously capture system/device timestamp
+ * @xtstamp: Receives simultaneously captured system and
device time
+ * @sync_devicetime: Callback to get simultaneous device time andQuoted 8 lines
+ * system counter from the device driver
+ *
+ * Reads a timestamp from a device and correlates it to system time
+ */
+int get_device_system_crosststamp(structsystem_device_crosststamp*xtstamp,
+ structget_sync_device_time*sync_devicetime)Another nit, but to me something like:
int get_device_system_crosststamp(structget_sync_device_time*sync_devicetime,
structsystem_device_crosststamp*ret)
Quoted 5 lines
Quoted 1 line
Reads better to me. ie: use this
device_time-> return that crosststamp.OK. My brain "works" the opposite way, but I think (after a small amount of research) I'm in the minority:)
Yea. memcpy and others do work the other way, but my left-to-right bias is pretty strong. :)
thanks
-john
Richard Cochran
maintainerRe: [RFC v5 1/6] Timekeeping cross timestamp interface for device drivers
Quoted 5 lines · John Stultz
Yea. I just feel like lots of structures add extra abstraction that makes the code harder to learn or re-learn. Not only is one trying to remember the base types that are being passed around, but you also have to remember what all the meta-structures are for. Especially for these two-value structures.
FWIW, I had exactly this trouble when reading this series.
Thanks,
Richard
RFC v5 2/6 Always Running Timer (ART) correlated clocksource
In reply to Christopher S. HallOn modern Intel systems TSC is derived from the new Always Running Timer (ART). ART can be captured simultaneous to the capture of audio and network device clocks, allowing a correlation between timebases to be constructed. Upon capture, the driver converts the captured ART value to the appropriate system clock using the correlated clocksource mechanism.
On systems that support ART a new CPUID leaf (0x15) returns parameters ?m? and ?n? such that:
TSC_value = (ART_value * m) / n + k [n >= 2]
[k is an offset that can adjusted by a privileged agent. The IA32_TSC_ADJUST MSR is an example of an interface to adjust k. See 17.14.4 of the Intel SDM for more details]
--- arch/x86/include/asm/cpufeature.h | 2 +- arch/x86/include/asm/tsc.h | 2 ++ arch/x86/kernel/tsc.c | 46 +++++++++++++++++++++++++++++++++++++++ 3 files changed, 49 insertions(+), 1 deletion(-)
arch/x86/include/asm/cpufeature.h +1 -1
@@ -85,7 +85,7 @@#define X86_FEATURE_P4 ( 3*32+ 7)/* "" P4 */ #define X86_FEATURE_CONSTANT_TSC ( 3*32+ 8)/* TSC ticks at a constant rate */ #define X86_FEATURE_UP ( 3*32+ 9)/* smp kernel running on up */ -/* free, was #define X86_FEATURE_FXSAVE_LEAK ( 3*32+10) * "" FXSAVE leaks FOP/FIP/FOP */ +#define X86_FEATURE_ART (3*32+10)/* Platform has always running timer (ART) */ #define X86_FEATURE_ARCH_PERFMON ( 3*32+11)/* Intel Architectural PerfMon */ #define X86_FEATURE_PEBS ( 3*32+12)/* Precise-Event Based Sampling */ #define X86_FEATURE_BTS ( 3*32+13)/* Branch Trace Store */
arch/x86/include/asm/tsc.h +2 -0
@@ -29,6 +29,8 @@ static inline cycles_t get_cycles(void)return rdtsc (); } +extern struct correlated_cs art_timestamper; +extern void tsc_init (void );extern void mark_tsc_unstable (char *reason);extern int unsynchronized_tsc (void );
arch/x86/kernel/tsc.c +46 -0
68 lines not on this page · raw message
Signature · 1 line
2.1.4
RFC v5 3/6 Add history to cross timestamp interface supporting slower devices
In reply to Christopher S. HallAnother representative use case of time sync and the correlated clocksource (in addition to PTP noted above) is PTP synchronized audio.
In a streaming application, as an example, samples will be sent and/or received by multiple devices with a presentation time that is in terms of the PTP master clock. Synchronizing the audio output on these devices requires correlating the audio clock with the PTP master clock. The more precise this correlation is, the better the audio quality (i.e. out of sync audio sounds bad).
From an application standpoint, to correlate the PTP master clock with the audio device clock, the system clock is used as a intermediate timebase. The transforms such an application would perform are:
System Clock <-> Audio clock
System Clock <-> Network Device Clock [<-> PTP Master Clock]Such audio applications make use of some existing ALSA library calls that provide audio/system cross-timestamps (e.g. snd_pcm_status_get_htstamp()). Previous driver implementations capture these cross timestamps by reading the system clock (raw/mono/real) and the device clock with greatest degree of simultaneity possible in software.
Modern Intel platforms can perform a more accurate cross timestamp in hardware (ART,audio device clock). The audio driver requires ART->system time transforms -- the same as required for the network driver. These platforms offload audio processing (including cross-timestamps) to a DSP which to ensure uninterrupted audio processing, communicates and response to the host only once every millsecond. As a result is takes up to a millisecond for the DSP to receive a request, the request is processed by the DSP, the audio output hardware is polled for completion, the result is copied into shared memory, and the host is notified. All of these operation occur on a millisecond cadence. This transaction requires about 2 ms, but under heavier workloads it may take up to 4 ms.
If update_wall_time() is called while waiting for a response within get_device_system_crosststamp() (from previous patch), a retry is attempted. This will occur if the cycle_interval (determined by CONFIG_HZ and mult/shift values) cycles elapse.
Adding a history allows these slow devices the option of providing an ART value outside of the retry loop. In this case, the callback provided is an accessor function for the previously obtained counter value. If get_system_device_crosststamp() receives a counter value previous to cycle_last, it consults the history provided as an argument in history_ref and interpolates the realtime and monotonic raw system time using the provided counter value. If there are any clock discontinuities, e.g. from calling settimeofday(), the monotonic raw time is interpolated in the usual way, but the realtime clock time is adjusted by scaling the monotonic raw adjustment.
When an accessor function is used a history argument *must* be provided. The history is initialized using ktime_get_snapshot() and must be called before the counter values are read.
When the history is used to interpolate timestamp values, the realtime clock time may be inaccurate to some degree. In general, the longer the length of history the larger the interpolation error. If there are discontinuities (large step changes) to the time, the error can be very large.
--- include/linux/clocksource.h | 16 ++++ include/linux/timekeeper_internal.h | 4 + include/linux/timekeeping.h | 9 +- kernel/time/timekeeping.c | 162 ++++++++++++++++++++++++++++++++++-- 4 files changed, 185 insertions(+), 6 deletions(-)
include/linux/clocksource.h +16 -0
@@ -282,4 +282,20 @@ struct raw_system_counterval { struct clocksource *cs; }; +/* +* struct system_time_snapshot - simultaneous raw/real time capture with +* counter value +* @cycles: Clocksource counter value to produce the system times +* @real: Realtime system time +* @raw: Monotonic raw system time +* @cs_seq: Sequence number associated with changed clocksource +*/ +struct system_time_snapshot { +cycles_t cycles; +ktime_t real; +ktime_t raw; +u8 cs_seq; +unsigned int clock_set_seq; +}; +#endif /* _LINUX_CLOCKSOURCE_H */
include/linux/timekeeper_internal.h +4 -0
@@ -13,6 +13,9 @@/** * struct tk_read_base - base structure for timekeeping readout * @clock: Current clocksource used for timekeeping. +* @cs_seq: Clocksource sequence is incremented per clocksource change. +* It's used to determine whether past system time can be related to +* current system time * @read: Read function of @clock * @mask: Bitmask for two's complement subtraction of non 64bit clocks * @cycle_last: @clock cycle value at last update
@@ -29,6 +32,7 @@*/ struct tk_read_base {struct clocksource *clock; +u8 cs_seq;cycle_t (*read)(struct clocksource *cs);cycle_t mask;cycle_t cycle_last;
include/linux/timekeeping.h +8 -1
@@ -295,11 +295,18 @@ struct get_sync_device_time {void *ctx; }; +struct system_time_snapshot ; +/* +* Simultaneously snapshot realtime and monotonic raw clocks +*/ +extern void ktime_get_snapshot (struct system_time_snapshot *systime_snapshot); +/* * Get cross timestamp between system clock and device clock */ extern int get_device_system_crosststamp (struct system_device_crosststamp *ct, -struct get_sync_device_time *dt); +struct get_sync_device_time *dt, +struct system_time_snapshot *history);/* * Persistent clock related interfaces
kernel/time/timekeeping.c +157 -5
214 lines not on this page · raw message
Signature · 1 line
2.1.4
Re: [RFC v5 3/6] Add history to cross timestamp interface supporting slower devices
Quoted 11 lines · Christopher S. Hall
include/linux/timekeeper_internal.h +3
-0@@ -13,6 +13,9 @@/** * struct tk_read_base - base structure for timekeeping readout * @clock: Current clocksource used for timekeeping. +* @cs_seq: Clocksource sequence is incremented per clocksource change. +* It's used to determine whether past system time can be related to +* current system time * @read: Read function of @clock * @mask: Bitmask for two's complement subtraction of non 64bit clocks * @cycle_last: @clock cycle value at last update
include/linux/timekeeper_internal.h line 38
Quoted 7 lines
include/linux/timekeeper_internal.h +1
-0@@ -29,6 +32,7 @@*/ struct tk_read_base {struct clocksource *clock; +u8 cs_seq;cycle_t (*read)(struct clocksource *cs);cycle_t mask;cycle_t cycle_last;
So Thomas optimized the tk_read_bases to fit in a cacheline, and so I worry about the u8 being added there. I'm also cautious about exporting these seq values out further via the system_time_snapshot. But I may just need some more time to warm up to the idea.
Quoted 44 lines
kernel/time/timekeeping.c +24
-0@@ -235,11 +235,13 @@ static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock) old_clock = tk->tkr_mono.clock; tk->tkr_mono.clock = clock; + ++tk->tkr_mono.cs_seq; tk->tkr_mono.read = clock->read; tk->tkr_mono.mask = clock->mask; tk->tkr_mono.cycle_last = tk->tkr_mono.read (clock); tk->tkr_raw.clock = clock; + ++tk->tkr_raw.cs_seq; tk->tkr_raw.read = clock->read; tk->tkr_raw.mask = clock->mask; tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;@@ -862,6 +864,39 @@ time64_t ktime_get_real_seconds(void) }EXPORT_SYMBOL_GPL (ktime_get_real_seconds); +/** +* ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter +* @snapshot: pointer to struct receiving the system time snapshot +*/ +void ktime_get_snapshot (struct system_time_snapshot *systime_snapshot) +{ +struct timekeeper *tk = &tk_core.timekeeper; +unsigned long seq; +ktime_t base_raw; +ktime_t base_real; +s64 nsec_raw; +s64 nsec_real; +cycle_t now; + + do { + seq = read_seqcount_begin(&tk_core.seq); + + now = tk->tkr_mono.read(tk->tkr_mono.clock); + systime_snapshot->cs_seq = tk->tkr_mono.cs_seq; + systime_snapshot->clock_set_seq = tk->clock_was_set_seq; + base_real = ktime_add(tk->tkr_mono.base, + tk_core.timekeeper.offs_real);
kernel/time/timekeeping.c time64_t ktime_get_real_seconds(void) · line 898
Quoted 10 lines
kernel/time/timekeeping.c +10
-0+ base_raw = tk->tkr_raw.base; + nsec_real =timekeeping_cycles_to_ns (&tk->tkr_mono, now); + nsec_raw =timekeeping_cycles_to_ns (&tk->tkr_raw, now); + }while (read_seqcount_retry (&tk_core.seq, seq)); + + systime_snapshot->cycles = now; + systime_snapshot->real =ktime_add_ns (base_real, nsec_real); + systime_snapshot->raw =ktime_add_ns (base_raw, nsec_raw); +} +EXPORT_SYMBOL_GPL (ktime_get_snapshot);
So can you split out this adding of ktime_get_snapshot() (maybe skipping the seqcount bits initially) into a separate patch?
Quoted 38 lines
kernel/time/timekeeping.c +21
-4@@ -936,19 +1044,63 @@ int get_device_system_crosststamp(struct system_device_crosststamp *xtstamp,*/ if (tk->tkr_mono.clock != raw_sys.cs)return -ENODEV ; + cycles = raw_sys.cycles; + +/* +* Check whether the system counter value provided by the +* device driver is on the current interval. +*/ + now = tk->tkr_mono.read (tk->tkr_mono.clock); + interval_start = tk->tkr_mono.cycle_last; +if (!cycle_between (interval_start, cycles, now)) { + cs_seq = tk->tkr_mono.cs_seq; + clock_was_set_seq = tk->clock_was_set_seq; + cycles = interval_start; + do_interp = true; + }else { + do_interp = false; + } base_real =ktime_add (tk->tkr_mono.base, tk_core.timekeeper.offs_real); base_raw = tk->tkr_raw.base; - nsec_real =timekeeping_cycles_to_ns (&tk->tkr_mono, - raw_sys.cycles); - nsec_raw =timekeeping_cycles_to_ns (&tk->tkr_raw, - raw_sys.cycles); + nsec_real =timekeeping_cycles_to_ns (&tk->tkr_mono, cycles); + nsec_raw =timekeeping_cycles_to_ns (&tk->tkr_raw, cycles); }while (read_seqcount_retry (&tk_core.seq, seq)); xtstamp->sys_realtime =ktime_add_ns (base_real, nsec_real); xtstamp->sys_monoraw =ktime_add_ns (base_raw, nsec_raw); + +/* +* Interpolate if necessary, working back from the start of the current
kernel/time/timekeeping.c int get_device_system_crosststamp(struct system_device_crosststamp *xtstamp, · line 1086
Quoted 10 lines
kernel/time/timekeeping.c +10
-0+* interval +*/ +if (do_interp) { +cycle_t total_history_cycles; +ktime_t history_monoraw; +ktime_t history_realtime; +bool discontinuity; +cycle_t partial_history_cycles = cycles - raw_sys.cycles; + + if (!history_ref || history_ref->cs_seq != cs_seq ||
I've not done a super close reading here. But is it very likely the the history_ref->cs_seq is the same as the captured seq? I thought this history_ref was to allow old cross stamps to be used to improve the back-calculation of the time at the given cycle value. So throwing them out if they are older then the last tick seems strange.
thanks
-john
Re: [RFC v5 3/6] Add history to cross timestamp interface supporting slower devices
Quoted 21 lines · John Stultz
Quoted 19 lines · Christopher S. Hall
include/linux/timekeeper_internal.h +1
-0@@ -13,6 +13,9 @@/** * struct tk_read_base - base structure for timekeeping readout * @clock: Current clocksource used for timekeeping. +* @cs_seq: Clocksource sequence is incremented per clocksource change.
+ * It's used to determine whether past system time can be related
to
+ * current system time
* @read: Read function of @clock
* @mask: Bitmask for two's complement subtraction of non 64bit
clocks
* @cycle_last: @clock cycle value at last updateinclude/linux/timekeeper_internal.h +1
-0@@ -29,6 +32,7 @@*/ struct tk_read_base {struct clocksource *clock; +u8 cs_seq;cycle_t (*read)(struct clocksource *cs);
Quoted 8 lines
include/linux/timekeeper_internal.h line 38
Quoted 2 lines
include/linux/timekeeper_internal.h +0
-0cycle_t mask;cycle_t cycle_last;So Thomas optimized the
tk_read_basesto fit in a cacheline, and so I worry about the u8 being added there. I'm also cautious about exporting these seq values out further via thesystem_time_snapshot. But I may just need some more time to warm up to the idea.
I think I missed the reason for the existence tk_read_base. Now I know. The clocksource sequence doesn't belong there. It should be moved timekeeper struct. Below there is more discussion of why the sequence number needs to exist at all. I think it does.
Would it make more sense to take the sequence out of the snapshot struct and make it an additional argument? it makes the snapshot struct more intuitive but the arguments to ktime_get_snapshot() maybe less so.
Quoted 54 lines
Quoted 54 lines
kernel/time/timekeeping.c +2
-0@@ -235,11 +235,13 @@ static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock) old_clock = tk->tkr_mono.clock; tk->tkr_mono.clock = clock; + ++tk->tkr_mono.cs_seq; tk->tkr_mono.read = clock->read; tk->tkr_mono.mask = clock->mask; tk->tkr_mono.cycle_last = tk->tkr_mono.read (clock); tk->tkr_raw.clock = clock; + ++tk->tkr_raw.cs_seq; tk->tkr_raw.read = clock->read; tk->tkr_raw.mask = clock->mask;tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;kernel/time/timekeeping.c +2
-0@@ -862,6 +864,39 @@ time64_t ktime_get_real_seconds(void) }EXPORT_SYMBOL_GPL (ktime_get_real_seconds); +/** +* ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
+ * @snapshot: pointer to struct receiving the system time snapshot
+ */kernel/time/timekeeping.c +20
-0+void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot) +{ + struct timekeeper *tk = &tk_core.timekeeper; + unsigned long seq; + ktime_t base_raw; + ktime_t base_real; + s64 nsec_raw; + s64 nsec_real; + cycle_t now; + + do { + seq = read_seqcount_begin(&tk_core.seq); + + now = tk->tkr_mono.read(tk->tkr_mono.clock); + systime_snapshot->cs_seq = tk->tkr_mono.cs_seq; + systime_snapshot->clock_set_seq = tk->clock_was_set_seq; + base_real = ktime_add(tk->tkr_mono.base, + tk_core.timekeeper.offs_real); + base_raw = tk->tkr_raw.base; + nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,now);
kernel/time/timekeeping.c +3
-0+ nsec_raw =timekeeping_cycles_to_ns (&tk->tkr_raw, now); + }while (read_seqcount_retry (&tk_core.seq, seq)); +
Quoted 8 lines
kernel/time/timekeeping.c time64_t ktime_get_real_seconds(void) · line 898
Quoted 5 lines
kernel/time/timekeeping.c +5
-0+ systime_snapshot->cycles = now; + systime_snapshot->real =ktime_add_ns (base_real, nsec_real); + systime_snapshot->raw =ktime_add_ns (base_raw, nsec_raw); +} +EXPORT_SYMBOL_GPL (ktime_get_snapshot);So can you split out this adding of
ktime_get_snapshot()(maybe skipping the seqcount bits initially) into a separate patch?
Yes. This should be fine.
Quoted 53 lines
Quoted 53 lines
kernel/time/timekeeping.c +4
-0@@ -936,19 +1044,63 @@ int get_device_system_crosststamp(structsystem_device_crosststamp *xtstamp, */ if (tk->tkr_mono.clock != raw_sys.cs)return -ENODEV ; + cycles = raw_sys.cycles; + +/* +* Check whether the system counter value provided by the
kernel/time/timekeeping.c +13
-4+* device driver is on the current interval. +*/ + now = tk->tkr_mono.read (tk->tkr_mono.clock); + interval_start = tk->tkr_mono.cycle_last; +if (!cycle_between (interval_start, cycles, now)) { + cs_seq = tk->tkr_mono.cs_seq; + clock_was_set_seq = tk->clock_was_set_seq; + cycles = interval_start; + do_interp = true; + }else { + do_interp = false; + } base_real =ktime_add (tk->tkr_mono.base, tk_core.timekeeper.offs_real); base_raw = tk->tkr_raw.base; - nsec_real =timekeeping_cycles_to_ns (&tk->tkr_mono, - raw_sys.cycles); - nsec_raw =timekeeping_cycles_to_ns (&tk->tkr_raw, - raw_sys.cycles); + nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,cycles);
+nsec_raw= timekeeping_cycles_to_ns(&tk->tkr_raw,cycles); } while (read_seqcount_retry(&tk_core.seq, seq)); xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real); xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);kernel/time/timekeeping.c +3
-0+ +/* +* Interpolate if necessary, working back from the start of the current
kernel/time/timekeeping.c +8
-0+* interval +*/ +if (do_interp) { +cycle_t total_history_cycles; +ktime_t history_monoraw; +ktime_t history_realtime; +bool discontinuity; + cycle_t partial_history_cycles = cycles -raw_sys.cycles;
Quoted 8 lines
kernel/time/timekeeping.c
Quoted 2 lines
kernel/time/timekeeping.c +2
-0+ + if (!history_ref || history_ref->cs_seq != cs_seq ||I've not done a super close reading here. But is it very likely the the
history_ref->cs_seqis the same as the captured seq? I thought thishistory_refwas to allow old cross stamps to be used to improve the back-calculation of the time at the given cycle value. So throwing them out if they are older then the last tick seems strange.
Maybe this needs more explanation. The clocksource sequence (cs_seq) is incremented for each change in clocksource. I use this to detect a rare corner case where the clocksource is changed from (on x86 anyway) TSC and then back. If the history crosses one of these changes then interpolation shouldn't be attempted (return error). It's not really enough when using the history to just check that the current clocksource is equal to the one used at the start of the history. The clocksource must not have changed at all. To answer your question, it's not at all likely that this would occur.
Quoted 3 lines
thanks
-john
Re: [RFC v5 3/6] Add history to cross timestamp interface supporting slower devices
Quoted 14 lines · Christopher Hall
Quoted 10 lines · John Stultz
On Mon, Jan 4, 2016 at 4:45 AM, Christopher S. Hall
Quoted 2 lines
+ if (!history_ref ||
history_ref->cs_seq!=cs_seq||I've not done a super close reading here. But is it very likely the the
history_ref->cs_seqis the same as the captured seq? I thought thishistory_refwas to allow old cross stamps to be used to improve the back-calculation of the time at the given cycle value. So throwing them out if they are older then the last tick seems strange.
Quoted 8 lines
Maybe this needs more explanation. The clocksource sequence (
cs_seq) is incremented for each change in clocksource. I use this to detect a rare corner case where the clocksource is changed from (on x86 anyway) TSC and then back. If the history crosses one of these changes then interpolation shouldn't be attempted (return error). It's not really enough when using the history to just check that the current clocksource is equal to the one used at the start of the history. The clocksource must not have changed at all. To answer your question, it's not at all likely that this would occur.
Yes. Apologies. I had mis-skimmed the cs_seq increment as happening in the update_wall_time not setup_internals.
Instead of cs_seq, which is easily confused as being related to the seqcount lock, could you instead call it something more explicit like clocksource_changed_count?
And yea, having it as part of the timekeeper structure would be the better place for it.
thanks
-john
Re: [RFC v5 3/6] Add history to cross timestamp interface supporting slower devices
In reply to Christopher HallQuoted 8 lines · Christopher Hall
Quoted 5 lines · John Stultz
I've not done a super close reading here. But is it very likely the the
history_ref->cs_seqis the same as the captured seq? I thought thishistory_refwas to allow old cross stamps to be used to improve the back-calculation of the time at the given cycle value. So throwing them out if they are older then the last tick seems strange.
Quoted 8 lines
Maybe this needs more explanation. The clocksource sequence (
cs_seq) is incremented for each change in clocksource. I use this to detect a rare corner case where the clocksource is changed from (on x86 anyway) TSC and then back. If the history crosses one of these changes then interpolation shouldn't be attempted (return error). It's not really enough when using the history to just check that the current clocksource is equal to the one used at the start of the history. The clocksource must not have changed at all. To answer your question, it's not at all likely that this would occur.
You can flush the history when a clocksource change happens. No need to add extra data to the core structures.
Thanks,
tglx
Re: [RFC v5 3/6] Add history to cross timestamp interface supporting slower devices
Quoted 24 lines · Thomas Gleixner
Quoted 8 lines · Christopher Hall
Quoted 5 lines · John Stultz
I've not done a super close reading here. But is it very likely the the
history_ref->cs_seqis the same as the captured seq? I thought thishistory_refwas to allow old cross stamps to be used to improve the back-calculation of the time at the given cycle value. So throwing them out if they are older then the last tick seems strange.Quoted 14 lines
Maybe this needs more explanation. The clocksource sequence (
cs_seq) is incremented for each change in clocksource. I use this to detect a rare corner
case where the clocksource is changed from (on x86 anyway) TSC and then back.
If the history crosses one of these changes then interpolation shouldn't be
attempted (return error). It's not really enough when using the history to
just check that the current clocksource is equal to the one used at the start
of the history. The clocksource must not have changed at all. To answer your
question, it's not at all likely that this would occur.
Quoted 3 lines
You can flush the history when a clocksource change happens. No need to add
extra data to the core structures.
Hi Thomas,
For slower devices like Intel's audio DSP, it's the responsibility of the driver to manage the history which is supplied as an argument to get_device_system_crosststamp(). The history is a triple (counter value, mono-raw time, real time). The clocksource change sequence number allows the timekeeping code to determine whether the supplied history is stale (returning an error if it is). I'm calling any history stale that crosses a clocksource change boundary.
Logically, the keeping of the history is split between the driver and timekeeping code with most of the burden on the driver. This introduces a small amount of redundancy in the form of the clocksource change sequence, but on balance results in a lot simpler timekeeping code.
Chris
RFC v5 4/6 Remove duplicate code from ktime_get_raw_and_real code
In reply to Christopher S. HallThe code in ktime_get_snapshot() is a superset of the code in ktime_get_raw_and_real() code. Changes the latter to call the former. A side effect of this is that ktime_get_raw_and_real() returns two clock times corresponding to the *exact* same clock tick. Previously, this code read the underlying counter twice.
--- kernel/time/timekeeping.c | 20 ++++---------------- 1 file changed, 4 insertions(+), 16 deletions(-)
kernel/time/timekeeping.c +4 -16
@@ -910,26 +910,14 @@ EXPORT_SYMBOL_GPL(ktime_get_snapshot);*/ void ktime_get_raw_and_real_ts64 (struct timespec64 *ts_raw,struct timespec64 *ts_real) { -struct timekeeper *tk = &tk_core.timekeeper; -unsigned long seq; -s64 nsecs_raw, nsecs_real; +struct system_time_snapshot snap;WARN_ON_ONCE (timekeeping_suspended); -do { - seq =read_seqcount_begin (&tk_core.seq); - - *ts_raw = tk->raw_time; - ts_real->tv_sec = tk->xtime_sec; - ts_real->tv_nsec =0 ; - - nsecs_raw =timekeeping_get_ns (&tk->tkr_raw); - nsecs_real =timekeeping_get_ns (&tk->tkr_mono); - - }while (read_seqcount_retry (&tk_core.seq, seq)); +ktime_get_snapshot (&snap); -timespec64_add_ns (ts_raw, nsecs_raw); -timespec64_add_ns (ts_real, nsecs_real); + *ts_raw =ktime_to_timespec64 (snap.raw); + *ts_real =ktime_to_timespec64 (snap.real); }EXPORT_SYMBOL (ktime_get_raw_and_real_ts64);
Signature · 1 line
2.1.4
Re: [RFC v5 4/6] Remove duplicate code from ktime_get_raw_and_real code
Quoted 35 lines · Christopher S. Hall
The code in
ktime_get_snapshot()is a superset of the code inktime_get_raw_and_real()code. Changes the latter to call the former. A side effect of this is thatktime_get_raw_and_real()returns two clock times corresponding to the *exact* same clock tick. Previously, this code read the underlying counter twice.Signed-off-by Christopher S. Hall <[email protected]>--- kernel/time/timekeeping.c | 20 ++++---------------- 1 file changed, 4 insertions(+), 16 deletions(-)kernel/time/timekeeping.c +1
-12@@ -910,26 +910,14 @@ EXPORT_SYMBOL_GPL(ktime_get_snapshot);*/ void ktime_get_raw_and_real_ts64 (struct timespec64 *ts_raw,struct timespec64 *ts_real) { -struct timekeeper *tk = &tk_core.timekeeper; -unsigned long seq; -s64 nsecs_raw, nsecs_real; +struct system_time_snapshot snap;WARN_ON_ONCE (timekeeping_suspended); - do { - seq = read_seqcount_begin(&tk_core.seq); - - *ts_raw = tk->raw_time; - ts_real->tv_sec = tk->xtime_sec; - ts_real->tv_nsec = 0; - - nsecs_raw = timekeeping_get_ns(&tk->tkr_raw); - nsecs_real = timekeeping_get_ns(&tk->tkr_mono);
kernel/time/timekeeping.c EXPORT_SYMBOL_GPL(ktime_get_snapshot); · line 922
Quoted 10 lines
kernel/time/timekeeping.c +3
-4- - }while (read_seqcount_retry (&tk_core.seq, seq)); +ktime_get_snapshot (&snap); -timespec64_add_ns (ts_raw, nsecs_raw); -timespec64_add_ns (ts_real, nsecs_real); + *ts_raw =ktime_to_timespec64 (snap.raw); + *ts_real =ktime_to_timespec64 (snap.real); }EXPORT_SYMBOL (ktime_get_raw_and_real_ts64);
Looks sane, but you might even be able to get rid of this function entirely and modify the one user (pps_get_ts()) to use ktime_get_snapshot().
thanks
-john
RFC v5 5/6 Add PTP_SYS_OFFSET_PRECISE for driver crosstimestamping
In reply to Christopher S. HallCurrently, network /system cross-timestamping is performed in the PTP_SYS_OFFSET ioctl. The PTP clock driver reads gettimeofday() and the gettime64() callback provided by the driver. The cross-timestamp is best effort where the latency between the capture of system time (getnstimeofday()) and the device time (driver callback) may be significant.
The getsynctime() callback and corresponding PTP_SYS_OFFSET_PRECISE ioctl allows the driver to perform this device/system correlation when for example cross timestamp hardware is available. Modern Intel systems can do this for onboard Ethernet controllers using the ART counter. There is virtually zero latency between captures of the ART and network device clock.
The capabilities ioctl (PTP_CLOCK_GETCAPS), is augmented allowing applications to query whether or not drivers implement the getsynctime callback, providing more precise cross timestamping.
--- Documentation/ptp/testptp.c | 6 ++++-- drivers/ptp/ptp_chardev.c | 31 +++++++++++++++++++++++++++++++ include/linux/ptp_clock_kernel.h | 8 ++++++++ include/uapi/linux/ptp_clock.h | 13 ++++++++++++- 4 files changed, 55 insertions(+), 3 deletions(-)
Documentation/ptp/testptp.c +4 -2
@@ -277,13 +277,15 @@ int main(int argc, char *argv[])" %d external time stamp channels\n" " %d programmable periodic signals\n" " %d pulse per second\n" -" %d programmable pins\n" , +" %d programmable pins\n" +" %d cross timestamping\n" , caps.max_adj, caps.n_alarm, caps.n_ext_ts, caps.n_per_out, caps.pps, - caps.n_pins); + caps.n_pins, + caps.cross_timestamping); } }
drivers/ptp/ptp_chardev.c +31 -0
@@ -22,6 +22,7 @@#include <linux/poll.h> #include <linux/sched.h> #include <linux/slab.h> +#include <linux/timekeeping.h> #include "ptp_private.h"
@@ -120,11 +121,14 @@ long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg)struct ptp_clock_caps caps;struct ptp_clock_request req;struct ptp_sys_offset *sysoff =NULL ; +struct ptp_sys_offset_precise precise_offset;struct ptp_pin_desc pd; struct ptp_clock *ptp = container_of(pc, struct ptp_clock, clock);struct ptp_clock_info *ops = ptp->info;struct ptp_clock_time *pct;struct timespec64 ts; +struct system_device_crosststamp xtstamp; +u32 rem;int enable, err =0 ;unsigned int i, pin_index;
@@ -138,6 +142,7 @@ long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) caps.n_per_out = ptp->info->n_per_out; caps.pps = ptp->info->pps; caps.n_pins = ptp->info->n_pins; + caps.cross_timestamping = ptp->info->getsynctime !=NULL ;if (copy_to_user ((void __user *)arg, &caps,sizeof (caps))) err = -EFAULT ;break ;
@@ -180,6 +185,32 @@ long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) err = ops->enable (ops, &req, enable);break ; +case PTP_SYS_OFFSET_PRECISE : +if (!ptp->info->getsynctime) { + err = -EINVAL ; +break ; + } + err = ptp->info->getsynctime (ptp->info, &xtstamp); +if (err) +break ; + + precise_offset.sys_real.sec = +div_u64_rem (ktime_to_ns (xtstamp.sys_realtime), +NSEC_PER_SEC , &rem); + precise_offset.sys_real.nsec = rem; + precise_offset.sys_raw.sec = +div_u64_rem (ktime_to_ns (xtstamp.sys_monoraw), +NSEC_PER_SEC , &rem); + precise_offset.sys_raw.nsec = rem; + precise_offset.dev.sec = +div_u64_rem (ktime_to_ns (xtstamp.device),NSEC_PER_SEC , + &rem); + precise_offset.dev.nsec = rem; +if (copy_to_user ((void __user *)arg, &precise_offset, +sizeof (precise_offset))) + err = -EFAULT ; +break ; +case PTP_SYS_OFFSET : sysoff =kmalloc (sizeof (*sysoff),GFP_KERNEL );if (!sysoff) {
include/linux/ptp_clock_kernel.h +8 -0
@@ -38,6 +38,7 @@ struct ptp_clock_request { }; }; +struct system_device_crosststamp ;/** * struct ptp_clock_info - decribes a PTP hardware clock *
@@ -67,6 +68,11 @@ struct ptp_clock_request {* @gettime64: Reads the current time from the hardware clock. * parameter ts: Holds the result. * +* @getsynctime: Reads the current time from the hardware clock and +* system clock simultaneously. +* parameter cts: Contains timestamp (device,system) pair, +* where system time is realtime and monotonic. +* * @settime64: Set the current time on the hardware clock. * parameter ts: Time value to set. *
@@ -105,6 +111,8 @@ struct ptp_clock_info {int (*adjfreq)(struct ptp_clock_info *ptp,s32 delta);int (*adjtime)(struct ptp_clock_info *ptp,s64 delta);int (*gettime64)(struct ptp_clock_info *ptp,struct timespec64 *ts); +int (*getsynctime)(struct ptp_clock_info *ptp, +struct system_device_crosststamp *cts);int (*settime64)(struct ptp_clock_info *p,const struct timespec64 *ts);int (*enable)(struct ptp_clock_info *ptp,struct ptp_clock_request *request,int on);
include/uapi/linux/ptp_clock.h +12 -1
@@ -51,7 +51,9 @@ struct ptp_clock_caps {int n_per_out;/* Number of programmable periodic signals. */ int pps;/* Whether the clock supports a PPS callback. */ int n_pins;/* Number of input/output pins. */ -int rsv[14 ];/* Reserved for future use. */ +/* Whether the clock supports precise system-device cross timestamps */ +int cross_timestamping; +int rsv[13 ];/* Reserved for future use. */ };struct ptp_extts_request {
@@ -81,6 +83,13 @@ struct ptp_sys_offset {struct ptp_clock_time ts[2 *PTP_MAX_SAMPLES +1 ]; }; +struct ptp_sys_offset_precise { +struct ptp_clock_time dev; +struct ptp_clock_time sys_real; +struct ptp_clock_time sys_raw; +unsigned int rsv[4 ];/* Reserved for future use. */ +}; +enum ptp_pin_function {PTP_PF_NONE ,PTP_PF_EXTTS ,
@@ -124,6 +133,8 @@ struct ptp_pin_desc {#define PTP_SYS_OFFSET _IOW(PTP_CLK_MAGIC, 5, struct ptp_sys_offset)#define PTP_PIN_GETFUNC _IOWR(PTP_CLK_MAGIC, 6, struct ptp_pin_desc)#define PTP_PIN_SETFUNC _IOW(PTP_CLK_MAGIC, 7, struct ptp_pin_desc) +#define PTP_SYS_OFFSET_PRECISE \ + _IOWR(PTP_CLK_MAGIC, 8, struct ptp_sys_offset_precise)struct ptp_extts_event {struct ptp_clock_time t;/* Time event occured. */
Signature · 1 line
2.1.4
Richard Cochran
maintainerRe: [RFC v5 5/6] Add PTP_SYS_OFFSET_PRECISE for driver crosstimestamping
Quoted 9 lines · Christopher S. Hall
drivers/ptp/ptp_chardev.c +1
-0@@ -138,6 +142,7 @@ long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) caps.n_per_out = ptp->info->n_per_out; caps.pps = ptp->info->pps; caps.n_pins = ptp->info->n_pins; + caps.cross_timestamping = ptp->info->getsynctime !=NULL ;if (copy_to_user ((void __user *)arg, &caps,sizeof (caps))) err = -EFAULT ;break ;
drivers/ptp/ptp_chardev.c long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) · line 190
Quoted 6 lines
drivers/ptp/ptp_chardev.c +3
-0@@ -180,6 +185,32 @@ long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) err = ops->enable (ops, &req, enable);break ; +case PTP_SYS_OFFSET_PRECISE : +if (!ptp->info->getsynctime) { + err = -EINVAL ;
-EOPNOTSUPP would be better here.
drivers/ptp/ptp_chardev.c long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) · line 200
Quoted 10 lines
drivers/ptp/ptp_chardev.c +10
-0+break ; + } + err = ptp->info->getsynctime (ptp->info, &xtstamp); +if (err) +break ; + + precise_offset.sys_real.sec = +div_u64_rem (ktime_to_ns (xtstamp.sys_realtime), +NSEC_PER_SEC , &rem); + precise_offset.sys_real.nsec = rem;
How about this instead:
ts = ktime_to_timespec64(xtstamp.sys_realtime); precise_offset.sys_real.sec = ts.tv_sec; precise_offset.sys_real.nsec = ts.tv_nsec;
drivers/ptp/ptp_chardev.c long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) · line 208
Quoted 8 lines
drivers/ptp/ptp_chardev.c +8
-0+ precise_offset.sys_raw.sec = +div_u64_rem (ktime_to_ns (xtstamp.sys_monoraw), +NSEC_PER_SEC , &rem); + precise_offset.sys_raw.nsec = rem; + precise_offset.dev.sec = +div_u64_rem (ktime_to_ns (xtstamp.device),NSEC_PER_SEC , + &rem); + precise_offset.dev.nsec = rem;
And for these as well.
Thanks,
Richard
Re: [RFC v5 5/6] Add PTP_SYS_OFFSET_PRECISE for driver crosstimestamping
Hi Richard,
This all sounds fine. Thanks for the feedback. I'll roll this into the next patchset.
Chris
Quoted 31 lines · Richard Cochran
drivers/ptp/ptp_chardev.c long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) · line 190
Quoted 3 lines · Christopher S. Hall
drivers/ptp/ptp_chardev.c +3
-0+case PTP_SYS_OFFSET_PRECISE : +if (!ptp->info->getsynctime) { + err = -EINVAL ;
-EOPNOTSUPPwould be better here.drivers/ptp/ptp_chardev.c long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) · line 200
Quoted 4 lines
drivers/ptp/ptp_chardev.c +4
-0+ precise_offset.sys_real.sec = +div_u64_rem (ktime_to_ns (xtstamp.sys_realtime), +NSEC_PER_SEC , &rem); + precise_offset.sys_real.nsec = rem;How about this instead:
ts = ktime_to_timespec64(xtstamp.sys_realtime); precise_offset.sys_real.sec = ts.tv_sec; precise_offset.sys_real.nsec = ts.tv_nsec;drivers/ptp/ptp_chardev.c long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) · line 208
Quoted 8 lines
drivers/ptp/ptp_chardev.c +8
-0+ precise_offset.sys_raw.sec = +div_u64_rem (ktime_to_ns (xtstamp.sys_monoraw), +NSEC_PER_SEC , &rem); + precise_offset.sys_raw.nsec = rem; + precise_offset.dev.sec = +div_u64_rem (ktime_to_ns (xtstamp.device),NSEC_PER_SEC , + &rem); + precise_offset.dev.nsec = rem;And for these as well.
Thanks,
Richard
RFC v5 6/6 Adds hardware supported cross timestamp
In reply to Christopher S. HallModern Intel systems supports cross timestamping of the network device clock and Always Running Timer (ART) in hardware. This allows the device time and system time to be precisely correlated. The timestamp pair is returned through e1000e_phc_get_sync_devicetime() used by get_device_system_crosststamp(). The hardware cross-timestamp result is made available to applications through the PTP_SYS_OFFSET_PRECISE ioctl which call e1000e_phc_getsynctime().
--- drivers/net/ethernet/intel/Kconfig | 9 +++ drivers/net/ethernet/intel/e1000e/defines.h | 5 ++ drivers/net/ethernet/intel/e1000e/ptp.c | 92 +++++++++++++++++++++++++++++ drivers/net/ethernet/intel/e1000e/regs.h | 4 ++ 4 files changed, 110 insertions(+)
drivers/net/ethernet/intel/Kconfig +9 -0
@@ -83,6 +83,15 @@ config E1000E To compile this driver as a module, choose M here. The module will be called e1000e. +config E1000E_HWTS + bool "Support HW cross-timestamp on PCH devices" + default y + depends on E1000E && X86 + ---help--- + Say Y to enable hardware supported cross-timestamping on PCH + devices. The cross-timestamp is available through the PTP clock + driver precise cross-timestamp ioctl (PTP_SYS_OFFSET_PRECISE). + config IGB tristate "Intel(R) 82575/82576 PCI-Express Gigabit Ethernet support" depends on PCI
drivers/net/ethernet/intel/e1000e/defines.h +5 -0
@@ -527,6 +527,11 @@#define E1000_RXCW_C 0x20000000/* Receive config */ #define E1000_RXCW_SYNCH 0x40000000/* Receive config synch */ +/* HH Time Sync */ +#define E1000_TSYNCTXCTL_MAX_ALLOWED_DLY_MASK 0x0000F000/* max delay */ +#define E1000_TSYNCTXCTL_SYNC_COMP 0x40000000/* sync complete */ +#define E1000_TSYNCTXCTL_START_SYNC 0x80000000/* initiate sync */ +#define E1000_TSYNCTXCTL_VALID 0x00000001/* Tx timestamp valid */ #define E1000_TSYNCTXCTL_ENABLED 0x00000010/* enable Tx timestamping */
drivers/net/ethernet/intel/e1000e/ptp.c +92 -0
110 lines not on this page · raw message
drivers/net/ethernet/intel/e1000e/regs.h +4 -0
@@ -245,6 +245,10 @@#define E1000_SYSTIML 0x0B600/* System time register Low - RO */ #define E1000_SYSTIMH 0x0B604/* System time register High - RO */ #define E1000_TIMINCA 0x0B608/* Increment attributes register - RW */ +#define E1000_SYSSTMPL 0x0B648/* HH Timesync system stamp low register */ +#define E1000_SYSSTMPH 0x0B64C/* HH Timesync system stamp hi register */ +#define E1000_PLTSTMPL 0x0B640/* HH Timesync platform stamp low register */ +#define E1000_PLTSTMPH 0x0B644/* HH Timesync platform stamp hi register */ #define E1000_RXMTRL 0x0B634/* Time sync Rx EtherType and Msg Type - RW */ #define E1000_RXUDP 0x0B638/* Time Sync Rx UDP Port - RW */
Signature · 1 line
2.1.4