Re: ddb fails to write breakpoint

[Jean-Yves Migeon <jeanyves.migeon%free.fr@localhost>]

Le 30/04/12 23:24, David Laight a écrit :
> I tried to set a kerneld db breakpoint on amd64 current (6.99.6) but the
> system just dumped - couldn't see why since it all scrolled away.
>
> Attempts to write into code space fault in ddb.
> The following makes it work - but is clearly ott.
>
> --- arch/amd64/amd64/db_memrw.c 23 Nov 2011 01:15:02 -0000      1.9
> +++ arch/amd64/amd64/db_memrw.c 30 Apr 2012 20:49:21 -0000
> @@ -146,6 +146,7 @@ db_write_text(vaddr_t addr, size_t size,
>                 pmap_pte_clearbits(ppte, PG_KR);
>                 pmap_pte_setbits(ppte, PG_KW);
>                 pmap_update_pg(pgva);
> +               tlbflushg();
>
>                 /*
>                  * Page is now writable.  Do as much access as we
> @@ -160,6 +161,7 @@ db_write_text(vaddr_t addr, size_t size,
>                 pmap_pte_clearbits(ppte, PG_KW);
>                 pmap_pte_setbits(ppte, PG_KR);
>                 pmap_update_pg(pgva);
> +               tlbflushg();
>
>         } while (size != 0);
>   }
>
> pmap_update_pg(addr) is just the single instruction 'invlpg adddr'
> tlbflushg() is a full tlb zap.
>
> Not looked at what invlpg is supposed to do, or whether it is
> an adequate synchronising instruction.
>
> cpu is an intel i7.

It should handle the invalidation fine, even if it is a global mapping.

Just tested it with a 6.99.4 kernel and worked fine, will test with a 
more recent one.

Do you know whether it happens every time or just for specific addresses 
(like one close to a page boundary or with the PG_PS bit set?)

--
Jean-Yves Migeon
jeanyves.migeon%free.fr@localhost