The current printout on a TimeSync ping doesn't actually contain the word
greybus and uses the word ping-time in the print string, something that
appears nowhere in our official documentation on this feature. This patch
changes the format string to contain 'greybus' and 'frametime' to bring the
TimeSync printout more in-line with other greybus kernel log strings.
before:
timesync ping-time: ap=8632564 1-svc=8630712 greybus1=8633031 1-8.8=8633026
after:
greybus frametime: ap=8632564 1-svc=8630712 greybus1=8633031 1-8.8=8633026
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
Remember to initialize the TimeSync ping fields to zero so that if a
timesync_get_last_event() returns an error - we display a FrameTime that is
obviously incorrect.
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
We need to make sure we adequately cancel and quiesce any scheduled
TimeSync synchronization operations in the case of greybus.ko being yanked
out of memory i.e. when doing an EHCI runtime suspend or just a plain
rmmod. The scenario is a new TimeSync sync operation has been scheduled.
Next gb_timesync_svc_remove() runs. In this case we should terminate any
scheduled work, terminate our ktime tracking timer and state transition to
GB_TIMESYNC_STATE_INVALID to ensure no other context may schedule any new
TimeSync operations.
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Reviewed-by: Vaibhav Hiremath <vaibhav.hiremath@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
Analysing a backtrace associated with the current EHCI runtime suspend code
has highlighted several places where its perfectly valid to make a
transition to GB_TIMESYNC_STATE_INACTIVE when not already in the
GB_TIMESYNC_STATE_INIT state, for example failure to issue a TimeSync
enable command to the SVC can and should legitimately call
gb_timesync_teardown() - at this point the state will be
GB_TIMESYNC_STATE_WAIT_SVC and it's legitimate and desirable to transition
to the INACTIVE state in this case. This patch fixes by removing the
restrictive and incorrect restriction on the transition to INACTIVE only
being valid when state == GB_TIMESYNC_STATE_INIT.
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Reviewed-by: Vaibhav Hiremath <vaibhav.hiremath@linaro.org>
Tested-by: Vaibhav Hiremath <vaibhav.hiremath@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
There is a scenario where gb_timesync_svc_remove() can run, attain a mutex
and call cancel_delayed_work_sync(). In the meantime a worker may already
be running and trying to attain the same mutex but will never do so as the
gb_timesync_svc_remove() path is holding the mutex and waiting on the
delayed_work_sync() to complete - leading to deadlock. This patch addresses
by calling the cancel_delayed_work_sync() before locking the relevant
mutex.
Reported-by: Vaibhav Agarwal <vaibhav.agarwal@linaro.org>
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Reviewed-by: Vaibhav Hiremath <vaibhav.hiremath@linaro.org>
Tested-by: Vaibhav Hiremath <vaibhav.hiremath@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
We need to use 'do_div()' when doing 64 bit division or modulo division
since the kernel will not pull in the gcc builtins __aeabi_ldivmod and
__aeabi_uldivmod on 32 bit builds.
Reported-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
This patch makes a debugfs entry in
/sys/kernel/debug/greybus/X-svc/frame-ktime that generates a TimeSync ping
event to the system and then subsequently presents that data to user-space
as a ktime/timespec clock-monotonic value rather than as a raw frame-time,
to aid humans in debugging and understanding frame-time and to provide an
example of the converting a frame-time to timespec/ktime to other
developers.
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Acked-by: Alex Elder <elder@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
This patch adds gb_timesync_to_timespec_by_svc() and
gb_timesync_to_timespec_by_interface() respectively. These routines will
convert from a given FrameTime to a ktime/timespec within an envelope of
about 17 seconds. The purpose of this routine is to enable reporting of a
FrameTime from a Module such as a Camera Module and to allow the AP to
then convert this timestamp into a Linux-native timestamp such as ktime.
This is useful and required in the v4l layer.
At 19.2MHz the accuracy of this conversion is about .3 femtoseconds per
count, which means at a 1 second offset from the reference the cumulative
error is about 1.59 nanoseconds. 1.59 nanoseconds is still less than 1
clock's worth of error @ 19.2MHz where each clock is 52.0833~ nanoseconds.
We're aiming for a maximum error rate of 30 nanoseconds which means at the
clock rate we are running at, the conversion from a FrameTime to a Linux
ktime/timespec can be plus-or-minus about 17 seconds from the reference
FrameTime/ktime pair before the routine will refuse to convert.
A realistic use-case for this routine is envisaged to be
- Greybus message received
- Some processing takes place - taking milliseconds
- Call into this routine is made
- Actual time between event in Module and conversion in AP < 1 second
- Error rate in conversion at 1.59 nanoseconds is less than 1 clock
@ 19.2MHz
This routine is not designed to allow for conversions for events with
large gaps between the event time and the current reference time for
conversion. Since FrameTime can be a very large integer we cannot convert
an arbitrarily large FrameTime to ktime, the feeling and objective here is
to make an over-provisioned envelope that in practical terms can never be
exceeded by expected use-cases. To convert longer gaps more work would have
to be done but ultimately some limit needs to be imposed and right now 0.3
femotseconds per clock on MSM8994 is both accurate and generous.
Adds:
- timesync.c::gb_timesync_frame_time_to_timespec_by_svc(
struct gb_svc *,
u64 frame_time,
struct timespec *ts)
- gb_svc is a pointer to a standard greybus SVC data structure
- frame_time is a system FrameTime.
- ts is an output parameter which represents the converted FrameTime
as a CLOCK_MONOTONIC timespec value.
- Returns 0 on success or a negative number indicating the type of
error on failure.
- timesync.c::gb_timesync_frame_time_to_timespec_by_interface(
struct gb_interface *,
u64 frame_time,
struct timespec *ts)
- gb_svc is a pointer to a standard greybus Interface data structure
- frame_time is a system FrameTime.
- ts is an output parameter which represents the converted FrameTime
as a CLOCK_MONOTONIC timespec value.
- Returns 0 on success or a negative number indicating the type of
error on failure.
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Acked-by: Alex Elder <elder@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
This patch adds a tracepoint to the TimeSync ISR, the purpose of which is
to indicate a TimeSync event has happened. This tracepoint can be enabled
by issuing the following command:
echo 1 > /sys/kernel/debug/tracing/events/greybus/gb_timesync_irq/enable
Synchronization looks like this:
TIMESTAMP FUNCTION
| |
147.865788: gb_timesync_irq: strobe 1/4 frame-time 2910076529
147.866781: gb_timesync_irq: strobe 2/4 frame-time 2910095689
147.867777: gb_timesync_irq: strobe 3/4 frame-time 2910114820
147.868791: gb_timesync_irq: strobe 4/4 frame-time 2910134038
A ping can be triggered like this:
cat /sys/kernel/debug/greybus/frame-time
And that ping looks like this:
TIMESTAMP FUNCTION
| |
147.934678: gb_timesync_irq: ping 4/4 frame-time 2911380356
169.280551: gb_timesync_irq: ping 4/4 frame-time 3321221069
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Reviewed-by: Vaibhav Hiremath <vaibhav.hiremath@linaro.org>
Acked-by: Alex Elder <elder@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
This patch adds the core timesync functionality.
0. arche-platform.c/arche-apb-ctrl.c
Modifies the platform layer to hook the incoming TIME_SYNC signal up to
the timesync strobe IRQ handler. If the arche-platform driver can't
satisfy the request for the wake-detect line, it will return -EAGAIN and
the calling work-queue must reschedule the attempt to get exclusive
access to the wake-detect pin logic. A private data field is added to
the arche-platform driver to enable passing of a timesync pointer to the
ISR responsible for synchronizing time.
1. timesync.c
A new file added which contains all of the logic associated with sending
greybus commands to SVC, APBx or Interfaces to enable, disable and
disseminate timing information.
2. timesync_platform.c
Any platform/arch specific code goes into timesync_platform.c.
Originally the idea was to keep the x86 and ARM arch dependencies in a
timesync_platform_arch.c file - however with further refinement that's
currently not necessary however just-in-case it becomes necessary to
resuscitate arch or platform specific methods for accessing timer
resources that access shouldn't be part of the core timesync.c logic and
so for the moment we access these timer resources through a thin access
layer in timesync_platform.c. Expect this to go away long term ideally.
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Acked-by: Alex Elder <elder@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>