How does the kernel map physical to virtual addresses?

How does the kernel map physical to virtual addresses?

[Timur Tabi]

When my driver wants to map virtual to physical (and vice versa) addresses, it
calls virt_to_phys and phys_to_virt. All these macros do is add or subtract a
constant (PAGE_OFFSET) to one address to get the other address.

How does the kernel configure the CPU (x86) to use this mapping?  I was under
the impression that the kernel creates a series of 4MB pages, using the x86's
4MB page feature.  For example, in a 64MB machine, there would be 16 PTEs (PGDs?
PMDs?), each one mapping a consecutive 4MB block of physical memory.  Is this
correct?  Somehow I believe that this is overly simplistic.

The reason I ask is that I'm confused as to what happens when a user process or
tries to allocate memory.  I assume that the VM uses 4KB pages for this
allocatation.  So do we end up with two virtual addresses pointing the same
physical memory?  

What happens if I use ioremap_nocache() on normal memory?  Is that memory
cached or uncached?  If I use the pointer obtained via phys_to_virt(), the
memory is cached.  But if I use the pointer returned from ioremap_nocache(), the
memory is uncached.  My understanding of x86 is that caching is based on
physical, not virtual addresses.  If so, it's not possible for a physical
address to be both cached and uncached at the same.

Could someone please straighten me out?



--
Timur Tabi - ttabi@interactivesi.com
Interactive Silicon - http://www.interactivesi.com

When replying to a mailing-list message, please don't cc: me, because then I'll just get two copies of the same message.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/

Re: How does the kernel map physical to virtual addresses?

[Roman Zippel]

Hi,

> When my driver wants to map virtual to physical (and vice versa) addresses, it
> calls virt_to_phys and phys_to_virt.\x1a All these macros do is add or subtract a
> constant (PAGE_OFFSET) to one address to get the other address.
> 
> How does the kernel configure the CPU (x86) to use this mapping?  I was under
> the impression that the kernel creates a series of 4MB pages, using the x86's
> 4MB page feature.  For example, in a 64MB machine, there would be 16 PTEs (PGDs?
> PMDs?), each one mapping a consecutive 4MB block of physical memory.  Is this
> correct?  Somehow I believe that this is overly simplistic.

Do you want it overly complex? :-) As long as you don't have high memory
(usally more than 1GB) all available memory is mapped in one virtual
mapping, so virt_to_phys/phys_to_virt can simply use an offset.

> The reason I ask is that I'm confused as to what happens when a user process or
> tries to allocate memory.  I assume that the VM uses 4KB pages for this
> allocatation.  So do we end up with two virtual addresses pointing the same
> physical memory?

You shouldn't mix kernel virtual mapping and user virtual mapping and
yes, a page can be in both mappings, but the kernel can't access the
user mapping directly, it has to use copy_(to|from)_user for this.

> What happens if I use ioremap_nocache() on normal memory?  Is that memory
> cached or uncached?  If I use the pointer obtained via phys_to_virt(), the
> memory is cached.  But if I use the pointer returned from ioremap_nocache(), the
> memory is uncached.  My understanding of x86 is that caching is based on
> physical, not virtual addresses.  If so, it's not possible for a physical
> address to be both cached and uncached at the same.

The cache is accessed with physical address, that's correct (at least on
most architectures). But before the cache is accessed a virtual to
physical translation is done and cached seperatly. During this
translation you don't get only the physical address but also some
properties on how it should be accessed, like read/write attributes, but
also cachable/noncachable information. Only after the cpu got this
information, it knows where _and_ how to access the memory and that
might happen through the cache or not.
BTW you might also want to look at __vmalloc() (in 2.4), it has a
protection argument so you can easily create a vmalloc_nocache().
Afterwards you only have to look up the pages that were mapped.

bye, Roman
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/

Re: How does the kernel map physical to virtual addresses?

[Tigran Aivazian]

Hi,

it is interesting to observe that many questions that deal with _details_
are answered quickly but questions related to fundamental concepts related
to how Linux is designed, baffle all of us (since 0 people answered). So,
is there really nobody in the whole world who can answer this? I would
like to know the answer (about global kernel memory layout - i.e. what
goes into PSE pages and what goes into normal ones, and how does PAE mode
change the picture?) myself...

I suppose one could find the answer by looking at each element of
mem_map[] but it is always more comfortable to look for answers when one
already knows the correct answer beforehand.

Regards,
Tigran


 On Fri, 25 Aug 2000, Timur Tabi wrote:

> When my driver wants to map virtual to physical (and vice versa) addresses, it
> calls virt_to_phys and phys_to_virt. All these macros do is add or subtract a
> constant (PAGE_OFFSET) to one address to get the other address.
> 
> How does the kernel configure the CPU (x86) to use this mapping?  I was under
> the impression that the kernel creates a series of 4MB pages, using the x86's
> 4MB page feature.  For example, in a 64MB machine, there would be 16 PTEs (PGDs?
> PMDs?), each one mapping a consecutive 4MB block of physical memory.  Is this
> correct?  Somehow I believe that this is overly simplistic.
> 
> The reason I ask is that I'm confused as to what happens when a user process or
> tries to allocate memory.  I assume that the VM uses 4KB pages for this
> allocatation.  So do we end up with two virtual addresses pointing the same
> physical memory?  
> 
> What happens if I use ioremap_nocache() on normal memory?  Is that memory
> cached or uncached?  If I use the pointer obtained via phys_to_virt(), the
> memory is cached.  But if I use the pointer returned from ioremap_nocache(), the
> memory is uncached.  My understanding of x86 is that caching is based on
> physical, not virtual addresses.  If so, it's not possible for a physical
> address to be both cached and uncached at the same.
> 
> Could someone please straighten me out?
> 
> 
> 
> --
> Timur Tabi - ttabi@interactivesi.com
> Interactive Silicon - http://www.interactivesi.com
> 
> When replying to a mailing-list message, please don't cc: me, because then I'll just get two copies of the same message.
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> Please read the FAQ at http://www.tux.org/lkml/
> 

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/

Re: How does the kernel map physical to virtual addresses?

[Stephen C. Tweedie]

Hi,

On Mon, Aug 28, 2000 at 01:56:34PM +0100, Tigran Aivazian wrote:
> 
> it is interesting to observe that many questions that deal with _details_
> are answered quickly but questions related to fundamental concepts related
> to how Linux is designed, baffle all of us (since 0 people answered). So,
> is there really nobody in the whole world who can answer this? I would
> like to know the answer (about global kernel memory layout - i.e. what
> goes into PSE pages and what goes into normal ones, and how does PAE mode
> change the picture?) myself...

If PSE is available, it is used to map the bits of the kernel's
VA which permanently maps all of physical memory.  As a result, those
pages cannot necessarily be looked up via a normal page table walk.
Anything dynamically mapped --- ie. high pages (if using PAE), or
vmalloc/ioremap pages --- is mapped using normal 4k ptes.  

mem_map[] is completely unaffected by the use of PSE, and continues to
keep one entry per 4k physical page regardless of how the page tables
have been constructed.

--Stephen
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/