Re: [Qemu-devel] [RFC PATCH v8 11/21] replay: recording and replaying clock ticks

[Pavel Dovgaluk]

> From: Paolo Bonzini [mailto:address@hidden
> On 03/02/2015 11:51, Pavel Dovgaluk wrote:
> >> From: Paolo Bonzini [mailto:address@hidden
> >> On 22/01/2015 09:52, Pavel Dovgalyuk wrote:
> >>> Clock ticks are considered as the sources of non-deterministic data for
> >>> virtual machine. This patch implements saving the clock values when they
> >>> are acquired (virtual, host clock, rdtsc, and some other timers).
> >>> When replaying the execution corresponding values are read from log and
> >>> transfered to the module, which wants to read the values.
> >>> Such a design required the clock polling to be synchronized. Sometimes
> >>> it is not true - e.g. when timeouts for timer lists are checked. In this 
> >>> case
> >>> we use a cached value of the clock, passing it to the client code.
> >>>
> >>> Signed-off-by: Pavel Dovgalyuk <address@hidden>
> >>> ---
> >>>  cpus.c                   |    3 +-
> >>>  include/qemu/timer.h     |   10 +++++
> >>>  qemu-timer.c             |    7 ++--
> >>>  replay/Makefile.objs     |    1 +
> >>>  replay/replay-internal.h |   13 +++++++
> >>>  replay/replay-time.c     |   84 
> >>> ++++++++++++++++++++++++++++++++++++++++++++++
> >>>  replay/replay.h          |   25 ++++++++++++++
> >>>  stubs/replay.c           |    9 +++++
> >>>  8 files changed, 147 insertions(+), 5 deletions(-)
> >>>  create mode 100755 replay/replay-time.c
> >>>
> >>> diff --git a/cpus.c b/cpus.c
> >>> index 8787277..01d89aa 100644
> >>> --- a/cpus.c
> >>> +++ b/cpus.c
> >>> @@ -353,7 +353,8 @@ static void icount_warp_rt(void *opaque)
> >>>
> >>>      seqlock_write_lock(&timers_state.vm_clock_seqlock);
> >>>      if (runstate_is_running()) {
> >>> -        int64_t clock = cpu_get_clock_locked();
> >>> +        int64_t clock = REPLAY_CLOCK(REPLAY_CLOCK_VIRTUAL_RT,
> >>> +                                     cpu_get_clock_locked());
> >>>          int64_t warp_delta;
> >>>
> >>>          warp_delta = clock - vm_clock_warp_start;
> >>> diff --git a/include/qemu/timer.h b/include/qemu/timer.h
> >>> index 0666920..0c2472c 100644
> >>> --- a/include/qemu/timer.h
> >>> +++ b/include/qemu/timer.h
> >>> @@ -4,6 +4,7 @@
> >>>  #include "qemu/typedefs.h"
> >>>  #include "qemu-common.h"
> >>>  #include "qemu/notify.h"
> >>> +#include "replay/replay.h"
> >>>
> >>>  /* timers */
> >>>
> >>> @@ -760,6 +761,8 @@ int64_t cpu_icount_to_ns(int64_t icount);
> >>>  /*******************************************/
> >>>  /* host CPU ticks (if available) */
> >>>
> >>> +#define cpu_get_real_ticks cpu_get_real_ticks_impl
> >>> +
> >>>  #if defined(_ARCH_PPC)
> >>>
> >>>  static inline int64_t cpu_get_real_ticks(void)
> >>> @@ -913,6 +916,13 @@ static inline int64_t cpu_get_real_ticks (void)
> >>>  }
> >>>  #endif
> >>>
> >>> +#undef cpu_get_real_ticks
> >>> +
> >>> +static inline int64_t cpu_get_real_ticks(void)
> >>
> >> cpu_get_real_ticks should never be used.  Please instead wrap
> >> cpu_get_ticks() with REPLAY_CLOCK.
> >
> > I don't quite understand this comment.
> > Do you mean that I should move REPLAY_CLOCK to the cpu_get_real_ticks 
> > usages instead of it's
> implementation?
> 
> Only to the cpu_get_ticks usage.  The others are okay.

cpu_get_ticks cannot call cpu_get_real_ticks in icount mode.
And other functions can. Then we should put REPLAY_CLOCK into those functions?

> 
> >>> +/*! Reads next clock value from the file.
> >>> +    If clock kind read from the file is different from the parameter,
> >>> +    the value is not used.
> >>> +    If the parameter is -1, the clock value is read to the cache anyway. 
> >>> */
> >>
> >> In what case could the clock kind not match?
> >>
> >
> > It was used in full version which had to skip clock from the log while 
> > loading the VM state.
> 
> So can it be removed for now?
> 

I think it can.

Pavel Dovgalyuk