mmap'ed memory cacheable or uncheable

[Teresa Tao]

Hi there,

I got a question regarding the mmap'ed memory. Is the mmap'ed memory cacheable or uncheable? My driver just use the remap_page_range to map a reserved physical memory.

Thanks in advance!

Teresa

[Jun Sun]

Post by Teresa Tao
Hi there,
I got a question regarding the mmap'ed memory. Is the mmap'ed memory cacheable or uncheable? My driver just use the remap_page_range to map a reserved physical memory.

I am pretty much sure it is cached, although I can't pin down exactly
where in the mm subsystem it does so - I have had cache bugs related
to mmap().

Jun

[Teresa Tao]

How about if I specify the following flags in my mmap routine just like what the pgprot_noncached micro did.
pgprot_val(vma->vm_page_prot) &= ~_CACHE_MASK;
pgprot_val(vma->vm_page_prot) |= _CACHE_UNCACHED;

Will this have kernel make the mmap'd memory non-cacheable? Or is there a mmap non-cacheable patch?

Thanks in advance!

Teresa



-----Original Message-----
From: Jun Sun [mailto:***@mvista.com]
Sent: Friday, July 25, 2003 3:02 PM
To: Teresa Tao
Cc: linux-***@linux-mips.org; ***@mvista.com
Subject: Re: mmap'ed memory cacheable or uncheable

Post by Teresa Tao
Hi there,
I got a question regarding the mmap'ed memory. Is the mmap'ed memory cacheable or uncheable? My driver just use the remap_page_range to map a reserved physical memory.

I am pretty much sure it is cached, although I can't pin down exactly
where in the mm subsystem it does so - I have had cache bugs related
to mmap().

Jun

[Jun Sun]

Post by Teresa Tao
How about if I specify the following flags in my mmap routine just like what the pgprot_noncached micro did.
pgprot_val(vma->vm_page_prot) &= ~_CACHE_MASK;
pgprot_val(vma->vm_page_prot) |= _CACHE_UNCACHED;
Will this have kernel make the mmap'd memory non-cacheable? Or is there a mmap non-cacheable patch?

I think this might work. Did you try it? The performance will be bad
though as mmap() is used widely by userland.

Jun

[Maciej W. Rozycki]

Post by Jun Sun

Post by Teresa Tao
How about if I specify the following flags in my mmap routine just like what the pgprot_noncached micro did.
pgprot_val(vma->vm_page_prot) &= ~_CACHE_MASK;
pgprot_val(vma->vm_page_prot) |= _CACHE_UNCACHED;
Will this have kernel make the mmap'd memory non-cacheable? Or is there a mmap non-cacheable patch?

I think this might work. Did you try it? The performance will be bad
though as mmap() is used widely by userland.

See mmap_mem() for how to select between cached and uncached mmap()ing
cleanly.

[Ralf Baechle]

Post by Teresa Tao
I got a question regarding the mmap'ed memory. Is the mmap'ed memory
cacheable or uncheable? My driver just use the remap_page_range to map
a reserved physical memory.

Mmap(2) creates cachable mappings for everything below the highest memory
address and uncached above that. That's a somewhat naive mechanism
which fails a few systems but somehow nobody did complain so far ...

Ralf