[Linux-ia64] LINUX IA64 Memory Support

[Linux-ia64] LINUX IA64 Memory Support

[Venkatesh Ramamurthy]

Hi,
Is there any way in which one can allocate memory say below 4 GB ( physical
address )in drivers.
TIA
Venkatesh

Re: [Linux-ia64] LINUX IA64 Memory Support

[Jes Sorensen]

>>>>> "Venkatesh" = Venkatesh Ramamurthy <Venkateshr@ami.com> writes:

Venkatesh> Hi, Is there any way in which one can allocate memory say
Venkatesh> below 4 GB ( physical address )in drivers.  TIA Venkatesh

This is not IA64 specific actually, so consider it a general answer
rather than an IA64 one.

A regular GFP_KERNEL should give you memory within the PCI DMA'able
area, which with the old PCI system meant within the 4GB
range. GFP_DMA will give you something within the ISA DMA
space. However, in principle none of the above means that the memory
is within the lower 4GB memory range, if the machine in question has
an IOMMU type device, you may be handed memory which is in the middle
of the memory space, but which is mapped to the 4GB the PCI bus in the
machine can see.

With the new PCI allocation system it is even more flexible, here you
may get addresess anywhere and you then tell the PCI subsystem at run
time to map the memory to the space the PCI bus can address. This
means it will be possible to support systems with for instance 200 PCI
busses (if someone feels like building such) etc.

Jes

Re: [Linux-ia64] LINUX IA64 Memory Support

[David Mosberger]

>>>>> On 15 Feb 2000 22:03:03 +0100, Jes Sorensen <Jes.Sorensen@cern.ch> said:

>>>>> "Venkatesh" = Venkatesh Ramamurthy <Venkateshr@ami.com> writes:
  Venkatesh> Hi, Is there any way in which one can allocate memory say
  Venkatesh> below 4 GB ( physical address )in drivers.  TIA Venkatesh

  Jes> This is not IA64 specific actually, so consider it a general
  Jes> answer rather than an IA64 one.

  Jes> A regular GFP_KERNEL should give you memory within the PCI
  Jes> DMA'able area, which with the old PCI system meant within the
  Jes> 4GB range. GFP_DMA will give you something within the ISA DMA
  Jes> space. However, in principle none of the above means that the
  Jes> memory is within the lower 4GB memory range, if the machine in
  Jes> question has an IOMMU type device, you may be handed memory
  Jes> which is in the middle of the memory space, but which is mapped
  Jes> to the 4GB the PCI bus in the machine can see.

  Jes> With the new PCI allocation system it is even more flexible,
  Jes> here you may get addresess anywhere and you then tell the PCI
  Jes> subsystem at run time to map the memory to the space the PCI
  Jes> bus can address. This means it will be possible to support
  Jes> systems with for instance 200 PCI busses (if someone feels like
  Jes> building such) etc.

Jes's answer is basically correct:

 - If you're writing a driver, use the new PCI DMA API (see pci.h in
   kernel 2.3.44 or later) and you'll be fine.

 - If you need memory below 4GB for some other reason, you're
   presently out of luck.  However, the already agreed upon solution
   is to add a new "zone" which will represent memory below 4GB
   (ZONE_32BIT or some such).  I plan to do this asap as we need this
   zone to properly implement the PCI DMA API (it currently works only
   because no Itanium system we know of/have access to has more than
   1GB of RAM ;-).

	--david