36 bit DMA address on IBM 440 PPC
I am having a problem doing PCI/memory DMA on the IBM 440 PPC because it has a 36 bit DMA address to/from the PCI bus (3_8000_0000), but the resource structure uses an unsigned long (which is a 32 bit quantity for gcc) to store 'physical/bus/dma' address. There is a kludge to get proper page table entries, which is to remap the 32 bit quatities into their 36 bit counterparts and then sending them on to __ioremap which takes a (36 bit) physical address input. This of course is no help to DMA, which wants not cpu (virtual) address, but physical (translated) address. I have done a similar kludge, making the 32 to 36 bit translation into a function which a driver doing DMA can use. However, it seems to me that the real answer here is to change the resource definition to use a phys_addr_t rather than an unsigned long, for start/end. Does anyone know if this change has been made anywhere, or if it is being planned?
** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/
Quoted 14 lines · On Mon, Mar 10, 2003 at 01:15:57PM +0000, [email protected] wrote:
I am having a problem doing PCI/memory DMA on the IBM 440 PPC because it has a 36 bit DMA address to/from the PCI bus (3_8000_0000), but the resource structure uses an unsigned long (which is a 32 bit quantity for gcc) to store 'physical/bus/dma' address. There is a kludge to get proper page table entries, which is to remap the 32 bit quatities into their 36 bit counterparts and then sending them on to
__ioremapwhich takes a (36 bit) physical address input. This of course is no help to DMA, which wants not cpu (virtual) address, but physical (translated) address. I have done a similar kludge, making the 32 to 36 bit translation into a function which a driver doing DMA can use. However, it seems to me that the real answer here is to change the resource definition to use aphys_addr_trather than an unsigned long, for start/end. Does anyone know if this change has been made anywhere, or if it is being planned?
Maybe in 2.5/2.6. I was promised that resources would become 64-bit on _all_ platforms in 2.5 but it hasn't happened yet. I will have to beat on the people that were working on that. Of course, that's only one piece required to properly handle I/O >4GB on a 32-bit platform.
Regards,
--
Matt Porter
[email protected]
This is Linux Country. On a quiet night, you can hear Windows reboot.
** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/
Is it to be 64 bit, or will the ref be more abstract and come from an /asm/ architecture specific include? It seems to me that in general it would be better to make things that hold or return or take a physical address reference it as such (e.g. phys_addr_t) rather than declare it to be some data size (e.g. unsigned long [long]), since data bus and address bus sizes are not necessarily in sync for size. Thus my vote for a separate architecture specific data type for physical address.
Quoted 22 lines · Matt Porter
Quoted 14 lines · On Mon, Mar 10, 2003 at 01:15:57PM +0000, [email protected] wrote:
I am having a problem doing PCI/memory DMA on the IBM 440 PPC because it has a 36 bit DMA address to/from the PCI bus (3_8000_0000), but the resource structure uses an unsigned long (which is a 32 bit quantity for gcc) to store 'physical/bus/dma' address. There is a kludge to get proper page table entries, which is to remap the 32 bit quatities into their 36 bit counterparts and then sending them on to
__ioremapwhich takes a (36 bit) physical address input. This of course is no help to DMA, which wants not cpu (virtual) address, but physical (translated) address. I have done a similar kludge, making the 32 to 36 bit translation into a function which a driver doing DMA can use. However, it seems to me that the real answer here is to change the resource definition to use aphys_addr_trather than an unsigned long, for start/end. Does anyone know if this change has been made anywhere, or if it is being planned?Maybe in 2.5/2.6. I was promised that resources would become 64-bit on _all_ platforms in 2.5 but it hasn't happened yet. I will have to beat on the people that were working on that. Of course, that's only one piece required to properly handle I/O >4GB on a 32-bit platform.
Quoted 5 lines
Regards,
--
Matt Porter
[email protected]
This is Linux Country. On a quiet night, you can hear Windows reboot.
** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/
Quoted 7 lines · Jeff Boyd
Is it to be 64 bit, or will the ref be more abstract and come from an /asm/ architecture specific include? It seems to me that in general it would be better to make things that hold or return or take a physical address reference it as such (e.g.
phys_addr_t) rather than declare it to be some data size (e.g. unsigned long [long]), since data bus and address bus sizes are not necessarily in sync for size. Thus my vote for a separate architecture specific data type for physical address.
The discussion on lkml suggested that Linus wants resources on all platforms to be 64 bits, period. Of course, the driving factor is to handle some ia32 PAE management in a better manner. One thing to keep in mind is that resources are not necessarily a CPU physical address (although most platforms define them as such), they are defined as an 'ioremapable token'. In the same manner, ioremap produces a token which can be passed to read*/write*, in most implementations that token is a virtual address which dereference directly (though that is cheating the API). One thing that makes a u64 resource convenient is that there is no abstraction necessary to printk info/warnings involving resource values.
Regards,
--
Matt Porter
[email protected]
This is Linux Country. On a quiet night, you can hear Windows reboot.
** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/