Re: [Qemu-devel] [RFC PATCH V3 4/6] cpu_exec: Add sleeping algorithm

[Sebastian Tanase]

----- Mail original -----
> De: "Paolo Bonzini" <address@hidden>
> À: "Sebastian Tanase" <address@hidden>, address@hidden
> Cc: address@hidden, address@hidden, address@hidden, "peter maydell" 
> <address@hidden>,
> address@hidden, address@hidden, address@hidden, address@hidden, 
> address@hidden,
> address@hidden, address@hidden, address@hidden, address@hidden, 
> address@hidden, address@hidden
> Envoyé: Lundi 30 Juin 2014 18:46:13
> Objet: Re: [RFC PATCH V3 4/6] cpu_exec: Add sleeping algorithm
> 
> Il 30/06/2014 15:59, Sebastian Tanase ha scritto:
> > The goal is to sleep qemu whenever the guest clock
> > is in advance compared to the host clock (we use
> > the monotonic clocks). The amount of time to sleep
> > is calculated in the execution loop in cpu_exec.
> >
> > At first, we tried to approximate at each for loop the real time
> > elapsed
> > while searching for a TB (generating or retrieving from cache) and
> > executing it. We would then approximate the virtual time
> > corresponding
> > to the number of virtual instructions executed. The difference
> > between
> > these 2 values would allow us to know if the guest is in advance or
> > delayed.
> > However, the function used for measuring the real time
> > (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very
> > expensive.
> > We had an added overhead of 13% of the total run time.
> >
> > Therefore, we modified the algorithm and only take into account the
> > difference between the 2 clocks at the begining of the cpu_exec
> > function.
> > During the for loop we try to reduce the advance of the guest only
> > by
> > computing the virtual time elapsed and sleeping if necessary. The
> > overhead
> > is thus reduced to 3%. Even though this method still has a
> > noticeable
> > overhead, it no longer is a bottleneck in trying to achieve a
> > better
> > guest frequency for which the guest clock is faster than the host
> > one.
> >
> > As for the the alignement of the 2 clocks, with the first algorithm
> > the guest clock was oscillating between -1 and 1ms compared to the
> > host clock.
> > Using the second algorithm we notice that the guest is 5ms behind
> > the host, which
> > is still acceptable for our use case.
> >
> > The tests where conducted using fio and stress. The host machine in
> > an i5 CPU at
> > 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is
> > an arm versatile-pb
> > built with buildroot.
> >
> > Currently, on our test machine, the lowest icount we can achieve
> > that is suitable for
> > aligning the 2 clocks is 6. However, we observe that the IO tests
> > (using fio) are
> > slower than the cpu tests (using stress).
> >
> > Signed-off-by: Sebastian Tanase <address@hidden>
> > Tested-by: Camille Bégué <address@hidden>
> > ---
> >  cpu-exec.c | 112
> >  +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> >  1 file changed, 112 insertions(+)
> >
> > diff --git a/cpu-exec.c b/cpu-exec.c
> > index 38e5f02..ac741b7 100644
> > --- a/cpu-exec.c
> > +++ b/cpu-exec.c
> > @@ -22,6 +22,102 @@
> >  #include "tcg.h"
> >  #include "qemu/atomic.h"
> >  #include "sysemu/qtest.h"
> > +#include "qemu/timer.h"
> > +
> > +/* Structs and function pointers for delaying the host */
> > +typedef struct SyncClocks SyncClocks;
> > +typedef void (*init_delay_func)(SyncClocks *sc,
> > +                                const CPUState *cpu);
> > +typedef void (*perform_align_func)(SyncClocks *sc,
> > +                                   const CPUState *cpu);
> > +struct SyncClocks {
> > +    int64_t diff_clk;
> > +    int64_t original_instr_counter;
> > +    init_delay_func init_delay;
> > +    perform_align_func perform_align;
> > +};
> 
> I don't remember exactly what I had in mind :) but if I remove these
> pointers from your patches, the code already looks nice, with no
> CONFIG_USER_ONLY except just below here.
> 

Ok, I will remove the function pointers then :)

> > +#if !defined(CONFIG_USER_ONLY)
> > +/* Allow the guest to have a max 3ms advance.
> > + * The difference between the 2 clocks could therefore
> > + * oscillate around 0.
> > + */
> > +#define VM_CLOCK_ADVANCE 3000000
> 
> How did you tune this?
> 

I computed this value based on the tests run on my machine. Of course,
this value will be different on a different machine running different
tests.

> > +static int64_t delay_host(int64_t diff_clk)
> > +{
> > +    struct timespec sleep_delay, rem_delay;
> > +    if (diff_clk > VM_CLOCK_ADVANCE) {
> > +        sleep_delay.tv_sec = diff_clk / 1000000000LL;
> > +        sleep_delay.tv_nsec = diff_clk % 1000000000LL;
> > +        if (nanosleep(&sleep_delay, &rem_delay) < 0) {
> > +            diff_clk -= (sleep_delay.tv_sec - rem_delay.tv_sec) *
> > 1000000000LL;
> > +            diff_clk -= sleep_delay.tv_nsec - rem_delay.tv_nsec;
> 
> I just remembered that nanosleep doesn't exist on Windows. :(  The
> rem_delay feature of nanosleep is very useful, and I don't think
> there
> is an equivalent.  So for now we shall make this POSIX only.
> 
> Paolo
> 

Should I surround the nanosleep with #ifndef _WIN32 and then add
Sleep for the Windows case ? or just leave out Windows ?

Sebastian


> > +        } else {
> > +            diff_clk = 0;
> > +        }
> > +    }
> > +    return diff_clk;
> > +}
> > +
> > +static int64_t instr_to_vtime(int64_t instr_counter, const
> > CPUState *cpu)
> > +{
> > +    int64_t instr_exec_time;
> > +    instr_exec_time = instr_counter -
> > +                      (cpu->icount_extra +
> > +                       cpu->icount_decr.u16.low);
> > +    instr_exec_time = instr_exec_time << icount_time_shift;
> > +
> > +    return instr_exec_time;
> > +}
> > +
> > +static void align_clocks(SyncClocks *sc, const CPUState *cpu)
> > +{
> > +    if (!icount_align_option) {
> > +        return;
> > +    }
> > +    sc->diff_clk += instr_to_vtime(sc->original_instr_counter,
> > cpu);
> > +    sc->original_instr_counter = cpu->icount_extra +
> > cpu->icount_decr.u16.low;
> > +    sc->diff_clk = delay_host(sc->diff_clk);
> > +}
> > +
> > +static void init_delay_params(SyncClocks *sc,
> > +                              const CPUState *cpu)
> > +{
> > +    static int64_t clocks_offset = -1;
> > +    int64_t realtime_clock_value, virtual_clock_value;
> > +    if (!icount_align_option) {
> > +        return;
> > +    }
> > +    /* On x86 target architecture, the PIT reset function (called
> > +       by qemu_system_reset) will end up calling qemu_clock_warp
> > +       and then icount_warp_rt changing vm_clock_warp_start from 0
> > (initial
> > +       value) to -1. This in turn will make us skip the initial
> > offset
> > +       between the real and virtual clocks (initially virtual
> > clock is 0).
> > +       Therefore we impose that the first time we run the cpu
> > +       the host and virtual clocks should be aligned; we don't
> > alter any of
> > +       the clocks, we just calculate the difference between them.
> > */
> > +    realtime_clock_value = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
> > +    virtual_clock_value = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL);
> > +    if (clocks_offset == -1) {
> > +        clocks_offset = realtime_clock_value -
> > virtual_clock_value;
> > +    }
> > +    sc->diff_clk = virtual_clock_value - realtime_clock_value +
> > clocks_offset;
> > +    sc->original_instr_counter = cpu->icount_extra +
> > cpu->icount_decr.u16.low;
> > +}
> > +#else
> > +/* We don't use the align feature for User emulation
> > +   thus we add empty functions which shall be ignored
> > +   by the compiler */
> > +static void align_clocks(SyncClocks *sc, const CPUState *cpu)
> > +{
> > +}
> > +
> > +static void init_delay_params(SyncClocks *sc,
> > +                              const CPUState *cpu)
> > +{
> > +}
> > +#endif /* CONFIG USER ONLY */
> >
> >  void cpu_loop_exit(CPUState *cpu)
> >  {
> > @@ -227,6 +323,11 @@ int cpu_exec(CPUArchState *env)
> >      TranslationBlock *tb;
> >      uint8_t *tc_ptr;
> >      uintptr_t next_tb;
> > +    /* Delay algorithm */
> > +    static SyncClocks sc = {
> > +        .init_delay = init_delay_params,
> > +        .perform_align = align_clocks
> > +    };
> >      /* This must be volatile so it is not trashed by longjmp() */
> >      volatile bool have_tb_lock = false;
> >
> > @@ -283,6 +384,11 @@ int cpu_exec(CPUArchState *env)
> >  #endif
> >      cpu->exception_index = -1;
> >
> > +    /* Calculate difference between guest clock and host clock.
> > +       This delay includes the delay of the last cycle, so
> > +       what we have to do is sleep until it is 0. As for the
> > +       advance/delay we gain here, we try to fix it next time. */
> > +    sc.init_delay(&sc, cpu);
> >      /* prepare setjmp context for exception handling */
> >      for(;;) {
> >          if (sigsetjmp(cpu->jmp_env, 0) == 0) {
> > @@ -672,6 +778,9 @@ int cpu_exec(CPUArchState *env)
> >                              if (insns_left > 0) {
> >                                  /* Execute remaining instructions.
> >                                   */
> >                                  cpu_exec_nocache(env, insns_left,
> >                                  tb);
> > +                               /* Try to align the host and
> > virtual clocks
> > +                                  if the guest is in advance. */
> > +                                sc.perform_align(&sc, cpu);
> >                              }
> >                              cpu->exception_index = EXCP_INTERRUPT;
> >                              next_tb = 0;
> > @@ -684,6 +793,9 @@ int cpu_exec(CPUArchState *env)
> >                      }
> >                  }
> >                  cpu->current_tb = NULL;
> > +                /* Try to align the host and virtual clocks
> > +                   if the guest is in advance */
> > +                sc.perform_align(&sc, cpu);
> >                  /* reset soft MMU for next block (it can currently
> >                     only be set by a memory fault) */
> >              } /* for(;;) */
> >
> 
>