Re: [patch] lazy I/O bitmap copy for i386 (ver 2) ...

[Davide Libenzi <davidel@xmailserver.org>]

On Tue, 24 Aug 2004, Davide Libenzi wrote:

> The following patch implements the lazy I/O bitmap copy for the i386 
> architecture. It uses an invalid bitmap offset inside the TSS to eventually 
> handle the correct bitmap update in the GPF handler. The logic is the same 
> of the first version, plus the usage of get/put_cpu() (thx Brian) and the 
> nesting over the latest Ingo variable bitmap bits.

As followup, I instrumented the switchto code to do something like:

        if (prev->io_bitmap_ptr) {                                                                                      
                if (tss->io_bitmap_base == INVALID_IO_BITMAP_OFFSET_LAZY)                                               
                        no_hit_gpf++;                                                                                   
                else                                                                                                    
                        hit_gpf++;                                                                                      
        }

After a couple of minutes of light X usage I get:

Hit:       1017
NoHit:    29541

This is FC1 up-to-date (yesterday's ~60% rate was due an incorrect test 
performed in the instrumentation code). Did not run any perf test, but for 
sure updating an unsigned long should result faster than memcpy/memset 
memory regions, during context switches.



- Davide