Re: [RFC] [PATCH] tsc_disable_B2 with "soft" rdtsc

[Kasper Dupont]

john stultz wrote:
> 
> Hello all,
> 
>         Here is my next rev of the tsc-disable patch. This one corrects a
> Configure.in typo (Caught by Gabriel Paubert), and probably more
> controversial, implements a soft rdtsc instruction via do_gettimeofday.
> 
> This avoids the earlier "box won't boot" problems with i686 optimized
> glibc's that called rdtsc. The rdtsc instruction will now be caught, and
> faked returning to the user program the same value of gettimeofday. Yes,
> its pretty hackish, but it works, albeit slowly.

+#ifdef CONFIG_X86_NUMA
+       /* "soft" rdtsc implementation */
+       if(!cpu_has_tsc)
+       {
+               char rdtsc_inst[2] = {0x0f, 0x31}; /*rdtsc opcode*/
+               char* inst_ptr = (char*)regs->eip;
+               if((inst_ptr[0] == rdtsc_inst[0])
+                       &&(inst_ptr[1] == rdtsc_inst[1])){

Any particular reason for puting the opcode in an
array and verify against that instead of just
verifying inst_ptr[i] against the constants?

+                       struct timeval tv;
+                       do_gettimeofday(&tv);
+                       regs->eax = tv.tv_usec;
+                       regs->edx = tv.tv_sec;

Looks like the soft tsc is going to be jumping
rather than runing. It is going to be increasing
at a constat rate most of the time, but will make
a big jump ahead every second. Couldn't the jump
easilly be made a lot smaller by using:
   regs->eax = tv.tv_usec<<2;

Of course this is not completely accurate either,
are we willing to pay the price for a more
accurate version?

+                       regs->eip += 2; /*= size of opcode*/
+                       return;
+               }
+       }
+#endif

-- 
Kasper Dupont -- der bruger for meget tid på usenet.
For sending spam use mailto:[e-mail]