fix PCI-memory access on Alchemy Au15x0
PATCH 1/1 fix PCI-memory access on Alchemy Au15x0
Hello,
the PCI support on Alchemy Au1500 and Au1550 based platforms is actually broken in the current mips-kernel.
The problem was introduced in 2.6.18.3 with the casting of some 36bit-defines (PCI memory) in au1000.h to resource_size_t which may be u32 or u64 depending on the experimental CONFIG_RESOURCES_64BIT.
With unset CONFIG_RESOURCES_64BIT, the pci-memory cannot be accessed because the ioremap in arch/mips/au1000/common/pci.c already used the truncated addresses.
With set CONFIG_RESOURCES_64BIT, things get even worse, because PCI-scan aborts, due to resource conflict: request_resource() in arch/mips/pci/pci.c fails because the maximum iomem-address is 0xffffffff (32bit) but the pci-memory-start-address is 0x440000000 (36bit).
To get pci working again, I propose the following patch:
1. remove the resource_size_t-casting from au1000.h again
2. make the casting in arch/mips/au1000/common/pci.c (it's allowed and necessary here. The 36bit-handling will be done in __fixup_bigphys_addr).
With this patch pci works again like in 2.6.18.2, the gcc-compile warnings in pci.c are gone and it doesn't depend on CONFIG_EXPERIMENTAL.
Only advantages ;-) Nevertheless, I'm open for your comments or even better solutions.
Best regards,
Alexander
arch/mips/au1000/common/pci.c | 8 ++++---- include/asm-mips/mach-au1x00/au1000.h | 12 ++++++------ 2 files changed, 10 insertions(+), 10 deletions(-)
include/asm-mips/mach-au1x00/au1000.h +6 -6
@@ -1665,12 +1665,12 @@* addresses. For PCI IO, it's simpler because we get to do the ioremap * ourselves and then adjust the device's resources. */ -#define Au1500_EXT_CFG ((resource_size_t) 0x600000000ULL) -#define Au1500_EXT_CFG_TYPE1 ((resource_size_t) 0x680000000ULL) -#define Au1500_PCI_IO_START ((resource_size_t) 0x500000000ULL) -#define Au1500_PCI_IO_END ((resource_size_t) 0x5000FFFFFULL) -#define Au1500_PCI_MEM_START ((resource_size_t) 0x440000000ULL) -#define Au1500_PCI_MEM_END ((resource_size_t) 0x44FFFFFFFULL) +#define Au1500_EXT_CFG 0x600000000ULL +#define Au1500_EXT_CFG_TYPE1 0x680000000ULL +#define Au1500_PCI_IO_START 0x500000000ULL +#define Au1500_PCI_IO_END 0x5000FFFFFULL +#define Au1500_PCI_MEM_START 0x440000000ULL +#define Au1500_PCI_MEM_END 0x44FFFFFFFULL#define PCI_IO_START (Au1500_PCI_IO_START + 0x1000)#define PCI_IO_END (Au1500_PCI_IO_END)
arch/mips/au1000/common/pci.c renamed from /arch/mips/au1000/common/pci.c +4 -4
@@ -39,15 +39,15 @@/* TBD */ static struct resource pci_io_resource = { - .start =PCI_IO_START , - .end =PCI_IO_END , + .start = (resource_size_t )PCI_IO_START , + .end = (resource_size_t )PCI_IO_END , .name ="PCI IO space" , .flags =IORESOURCE_IO };static struct resource pci_mem_resource = { - .start =PCI_MEM_START , - .end =PCI_MEM_END , + .start = (resource_size_t )PCI_MEM_START , + .end = (resource_size_t )PCI_MEM_END , .name ="PCI memory space" , .flags =IORESOURCE_MEM };
Signature · 4 lines
Alexander Bigga Tel: +49-4321-55956-25 mycable GmbH Fax: +49-4321-55956-10 Gartenstrasse 10 D-24534 Neumuenster eMail: [email protected]
Quoted 18 lines · Alexander Bigga
The problem was introduced in 2.6.18.3 with the casting of some 36bit-defines (PCI memory) in
au1000.htoresource_size_twhich may be u32 or u64 depending on the experimentalCONFIG_RESOURCES_64BIT.With unset
CONFIG_RESOURCES_64BIT, the pci-memory cannot be accessed because the ioremap inarch/mips/au1000/common/pci.calready used the truncated addresses.
With setCONFIG_RESOURCES_64BIT, things get even worse, because PCI-scan aborts, due to resource conflict:request_resource()inarch/mips/pci/pci.cfails because the maximum iomem-address is0xffffffff(32bit) but the pci-memory-start-address is0x440000000(36bit).To get pci working again, I propose the following patch:
1. remove the resource_size_t-casting from
au1000.hagain
2. make the casting inarch/mips/au1000/common/pci.c(it's allowed and necessary here. The 36bit-handling will be done in__fixup_bigphys_addr).
Quoted 5 lines
With this patch pci works again like in 2.6.18.2, the gcc-compile warnings in
pci.care gone and it doesn't depend onCONFIG_EXPERIMENTAL.Only advantages ;-) Nevertheless, I'm open for your comments or even better solutions.
Looks reasonable, applied.
Ralf