Re: [hatari-devel] Adding cache support for the MegaSTE -> DONE

[Eero Tamminen <oak@xxxxxxxxxxxxxx>]

Hi,

On 4.9.2024 9.47, Nicolas Pomarède wrote:
> Le 04/09/2024 à 01:17, Eero Tamminen a écrit :
> > On 4.9.2024 0.12, Nicolas Pomarède wrote:
> > > Note that the cache needs to be flushed on every transfer when using 
> > > Gemdos HD emulation, or normal FDC / HDC. Else this would transfer 
> > > data to RAM without invalidating the cache (which has no means to 
> > > know RAM has been modified), thus creating crashes in the emulated 
> > > program sooner or later.
> >
> > GEMDOS HD (and NF SCSI driver) are already calling cache invalidation 
> > functions...?
> >
> > Otherwise things would not have worked properly for 030+ caches.
> >
> > See "git grep Flush.*Cache".
>
>
> They were already calling them partly ; through STMemory_WriteLong for 
> example, but STMemory_SafeCopy was missing.
>
> so it's quite likely that STMemory_SafeCopy could have messed with the 
> caches since quite some time, it's just that by luck it never triggered 
> a bug.

Only thing in GEMDOS HD calling that, and SafeClear,
is GemDOS_LoadAndReloc(), which is subset of Pexec().

(NF SCSI driver does not use them.)

Because program is typically loaded to unused Atari memory (especially 
with Hatari autostarting), and there's some code executed between 
loading the data and actual program execution, it's unlikely to have 
been an issue with TT and Falcon, as 030 has only 256 bytes of i/d-cache.

Whereas emulation for 040 and 060 (with 4KB and 8KB caches) is often ran 
with cycle-accuracy i.e. caches disable, for speed reasons as 040/060 
programs expect faster than 32Mhz machine...

(Only issues that I remember for 040/060 with caches enabled have been 
with m68k-linux, and there memory updates are done before emulation starts.)

=> MegaSTE's 16 KB cache is actually largest of these, i.e. most likely 
to trigger issues with bad cache handling!


	- Eero

PS. Wouldn't most correct behavior for SafeCopy/SafeClear would actually 
be to have last pieces of data written to memory, going also to cache, 
instead of cache having been flushed?