Re: EFAULT reading /dev/mem... - broken x86info

[Manfred Spraul <manfred@colorfullife.com>]

Petr Vandrovec wrote:

>On Mon, Nov 10, 2003 at 06:17:37PM +0100, Manfred Spraul wrote:
>  
>
>>DEBUG_PAGEALLOC unmaps pages on kmem_cache_free and __free_pages(). The 
>>pages are mapped again during get_free_pages and kmem_cache_alloc.
>>
>>0x86000 looks like a normal page - what guarantees that it's not used by 
>>the kernel?
>>    
>>
>
>With DEBUG_PAGEALLOC there is no problem with page if it is USED by the kernel.
>Problem is if page is NOT USED - in this case kernel does not have it in its
>mapping, and bad thing happen.
>  
>
If the page is used by AGP, then it won't have a mapping either.

>x86info (and other simillar tools for dumping different BIOS structures) just
>scans physical memory for some well known signatures - hoping that kernel did
>not smash these structures.
>  
>
Scanning physical memory is very dangerous: it's undefined what happens 
if a page is mapped multiple times with different cache settings. Athlon 
cpus prefetch whatever they see, and they speculatively set the dirty 
bit in cachelines for WB cacheable pages.

>Up to now it was possible to read whole physical memory from userspace - some 
>pages by reading /dev/mem, some pages by mmaping /dev/mem. Now it is not possible 
>anymore - which I think is bad, as /dev/mem has semantic of it.
>
I think /dev/mem should lie and return 0x00 for pages that are not in 
the linear mapping.

Or /dev/mem users should expect random holes.

--
    Manfred