* Nadav Amit <[email protected]> wrote:

> > On Apr 25, 2019, at 12:29 PM, Ingo Molnar <[email protected]> wrote:
> > 
> > 
> > * Nadav Amit <[email protected]> wrote:
> > 
> >> Move flush_tlb_info variables off the stack. This allows to align
> >> flush_tlb_info to cache-line and avoid potentially unnecessary cache
> >> line movements. It also allows to have a fixed virtual-to-physical
> >> translation of the variables, which reduces TLB misses.
> >> 
> >> Use per-CPU struct for flush_tlb_mm_range() and
> >> flush_tlb_kernel_range(). Add debug assertions to ensure there are
> >> no nested TLB flushes that might overwrite the per-CPU data. For
> >> arch_tlbbatch_flush() use a const struct.
> >> 
> >> Results when running a microbenchmarks that performs 10^6 MADV_DONTEED
> >> operations and touching a page, in which 3 additional threads run a
> >> busy-wait loop (5 runs, PTI and retpolines are turned off):
> >> 
> >>                    base            off-stack
> >>                    ----            ---------
> >> avg (usec/op)              1.629           1.570   (-3%)
> >> stddev                     0.014           0.009
> >> 
> >> Cc: Peter Zijlstra <[email protected]>
> >> Cc: Andy Lutomirski <[email protected]>
> >> Cc: Dave Hansen <[email protected]>
> >> Cc: Borislav Petkov <[email protected]>
> >> Cc: Thomas Gleixner <[email protected]>
> >> Signed-off-by: Nadav Amit <[email protected]>
> >> 
> >> ---
> >> 
> >> v1->v2:
> >> - Initialize all flush_tlb_info fields [Andy]
> >> ---
> >> arch/x86/mm/tlb.c | 100 ++++++++++++++++++++++++++++++++++------------
> >> 1 file changed, 74 insertions(+), 26 deletions(-)
> >> 
> >> diff --git a/arch/x86/mm/tlb.c b/arch/x86/mm/tlb.c
> >> index 487b8474c01c..aac191eb2b90 100644
> >> --- a/arch/x86/mm/tlb.c
> >> +++ b/arch/x86/mm/tlb.c
> >> @@ -634,7 +634,7 @@ static void flush_tlb_func_common(const struct 
> >> flush_tlb_info *f,
> >>    this_cpu_write(cpu_tlbstate.ctxs[loaded_mm_asid].tlb_gen, mm_tlb_gen);
> >> }
> >> 
> >> -static void flush_tlb_func_local(void *info, enum tlb_flush_reason reason)
> >> +static void flush_tlb_func_local(const void *info, enum tlb_flush_reason 
> >> reason)
> >> {
> >>    const struct flush_tlb_info *f = info;
> >> 
> >> @@ -722,43 +722,81 @@ void native_flush_tlb_others(const struct cpumask 
> >> *cpumask,
> >>  */
> >> unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
> >> 
> >> +static DEFINE_PER_CPU_SHARED_ALIGNED(struct flush_tlb_info, 
> >> flush_tlb_info);
> >> +
> >> +#ifdef CONFIG_DEBUG_VM
> >> +static DEFINE_PER_CPU(unsigned int, flush_tlb_info_idx);
> >> +#endif
> >> +
> >> +static inline struct flush_tlb_info *get_flush_tlb_info(struct mm_struct 
> >> *mm,
> >> +                  unsigned long start, unsigned long end,
> >> +                  unsigned int stride_shift, bool freed_tables,
> >> +                  u64 new_tlb_gen)
> >> +{
> >> +  struct flush_tlb_info *info = this_cpu_ptr(&flush_tlb_info);
> >> +
> >> +#ifdef CONFIG_DEBUG_VM
> >> +  /*
> >> +   * Ensure that the following code is non-reentrant and flush_tlb_info
> >> +   * is not overwritten. This means no TLB flushing is initiated by
> >> +   * interrupt handlers and machine-check exception handlers.
> >> +   */
> >> +  BUG_ON(this_cpu_inc_return(flush_tlb_info_idx) != 1);
> >> +#endif
> > 
> > isn't this effectively VM_BUG_ON()?
> 
> Not exactly. When CONFIG_DEBUG_VM is off we get
> 
>       #define VM_BUG_ON(cond) BUILD_BUG_ON_INVALID(cond)
> 
> This will cause the build to fail since flush_tlb_info_idx is not defined in
> when CONFIG_DEBUG_VM is off.

Ugh, so VM_BUG_ON() should really be named VM_BUILD_BUG_ON()?

Anyway, agreed.

> >> +static inline void put_flush_tlb_info(void)
> >> +{
> >> +#ifdef CONFIG_DEBUG_VM
> >> +  /* Complete reentrency prevention checks */
> >> +  barrier();
> >> +  this_cpu_dec(flush_tlb_info_idx);
> >> +#endif
> > 
> > In principle this_cpu_dec() should imply a compiler barrier?
> 
> this_cpu_dec() is eventually expanded to the macro of percpu_add_op(). And
> the inline assembly does not have a “memory” clobber, so I don’t think so.

I think that's a bug and PeterZ is fixing those.

Thanks,

        Ingo

Reply via email to