PCI memory-mapping question
Howdy all,
I hope someone here can show me the magic spell I need...
I have a device (Brooktree Bt8230 SAR) that I am trying to port a driver from 2.0 to 2.2 for.
Under 2.0, cat /proc/pci reports:
Bus 0, device 19, function 0:
ATM network controller: Brooktree Bt8230 ATM SAR (rev 3).
Fast devsel. Fast back-to-back capable. IRQ 10. Non-prefetchable 32 bit memory at 0x3b000000.
and the memory-mapping in the driver's pci_probe looks like:
/* Get the board I/O address */ pcibios_read_config_dword(pb,
PCI_DEVICE,
PCI_BASE_ADDRESS_0,
&iobase);/* Map iobase to LINUX virtual address. */
iobase |= 0xC0000000;
(void *)iobase = vremap((unsigned long)0xfb000000,
(unsigned long)256*1024);
.....
why the author hardwired fb000000 instead of passing iobase in vremap, I don't know. Apparently, the remapping is done to prevent caching. The BAT's are setup thusly:
/* Hardwired MMU segments */
/* Segment 0x8XXXXXXX, 0xCXXXXXXX always mapped (for I/O) */ /* Segment 0x9XXXXXXX mapped during init */
BAT BAT0 =
{
{
0x80000000>>17, /* bepi */
BL_256M, /* bl */
1, /* vs */
0, /* vp */
},
{
0x80000000>>17, /* brpn */
1, /* w */
1, /* i (cache disabled) */
0, /* m */
1, /* g */
BPP_RW /* pp */
}
};
BAT BAT1 =
{
{
0xC0000000>>17, /* bepi */
BL_256M, /* bl */
1, /* vs */
0, /* vp */
},
{
0xC0000000>>17, /* brpn */
1, /* w */
1, /* i (cache disabled) */
0, /* m */
1, /* g */
BPP_RW /* pp */
}
};
BAT BAT2 =
{
{
0x00000000>>17, /* bepi */
BL_256M, /* bl */
0, /* vs */
0, /* vp */
},
{
0x00000000>>17, /* brpn */
1, /* w */
1, /* i (cache disabled) */
0, /* m */
0, /* g */
BPP_RW /* pp */
}
};
BAT BAT3 =
{
{
0x00000000>>17, /* bepi */
BL_256M, /* bl */
0, /* vs */
0, /* vp */
},
{
0x00000000>>17, /* brpn */
1, /* w */
1, /* i (cache disabled) */
0, /* m */
0, /* g */
BPP_RW /* pp */
}
};
#ifdef OUR_CUSTOM_HARDWARE
BAT CUSTOM_BAT3 =
{
{
0xF0000000>>17, /* bepi */
BL_256M, /* bl */
1, /* vs */
0, /* vp */
},
{
0xF0000000>>17, /* brpn */
1, /* w */
1, /* i (cache disabled) */
1, /* m */
0, /* g */
BPP_RW /* pp */
}
};
#endif
BAT TMP_BAT2 =
{ /* 0x9XXXXXXX -> 0x0XXXXXXX */
{
0x90000000>>17, /* bepi */
BL_256M, /* bl */
1, /* vs */
1, /* vp */
},
{
0x00000000>>17, /* brpn */
0, /* w */
0, /* i (cache enabled) */
0, /* m */
0, /* g */
BPP_RW /* pp */
}
};
So there is a straight mapping, so far as I can tell.
Now, with 2.2, this is what I have setup:
BATs:
switch (_machine) {
case _MACH_prep:
#ifndef OUR_CUSTOM_HARDWARE
setbat(0, 0x80000000, 0x80000000, 0x10000000, IO_PAGE);
setbat(1, 0xf0000000, 0xc0000000, 0x08000000, IO_PAGE);
#else
setbat(0, 0x80000000, 0x80000000, 0x10000000, IO_PAGE);
setbat(1, 0xf0000000, 0xc0000000, 0x08000000, IO_PAGE);
setbat(3, 0xF8000000, 0xF8000000, 0x08000000, CUSTOM_RAM_PAGE);
#endif
ioremap_base = 0xf8000000;
break;...where CUSTOM_RAM_PAGE is
#define CUSTOM_RAM_PAGE (_PAGE_NO_CACHE | _PAGE_RW | _PAGE_COHERENT)
I didn't want to go messing with the mappings that were already there, so instead of mapping all 0xF0000000-0xFFFFFFFF straight, I wanted to map the upper 128M (since 0xc0000000-0xc7FFFFFF is apparently already mapped to the lower 128M of 0xFxxxxxxx). Therefore, I tried mapping to 0xf8000000...didn't work...kernel panics with:
Jul 24 17:41:00 lyn120 kernel: SAR:: init_module. Jul 24 17:41:00 lyn120 kernel: bt8230_insert_device Jul 24 17:41:00 lyn120 kernel: Initializing SAR driver! Jul 24 17:41:00 lyn120 kernel: Looking for SAR devices... Jul 24 17:41:00 lyn120 kernel: SAR device found: sar0 Jul 24 17:41:00 lyn120 kernel: bt8230_probe Jul 24 17:41:00 lyn120 kernel: bt8230_pci_probe Jul 24 17:41:00 lyn120 kernel: iobase = 0x1000000 Jul 24 17:41:00 lyn120 kernel: iobase now = 0xf8000000 Jul 24 17:41:00 lyn120 kernel: remapped iobase = 0xf8000000 Jul 24 17:41:00 lyn120 kernel: irq = 10 Jul 24 17:41:00 lyn120 kernel: set Bus Mastering ON. Jul 24 17:41:01 lyn120 kernel: check_region passed Jul 24 17:41:01 lyn120 kernel: NIP: CC0148E8 XER: 20000000 LR: CC0148E8 REGS: c4 bb5c80 TRAP: 0300 Jul 24 17:41:01 lyn120 kernel: MSR: 00009032 EE: 1 PR: 0 FP: 0 ME: 1 IR/DR: 11 Jul 24 17:41:01 lyn120 kernel: TASK = c4bb4000[477] 'insmod' mm->pgd c45a8000 La st syscall: 128 Jul 24 17:41:01 lyn120 kernel: last math c4bb4000 Jul 24 17:41:01 lyn120 kernel: GPR00: CC0148E8 C4BB5D30 C4BB4000 00000014 000000 01 00000014 C0120000 C011D010 Jul 24 17:41:01 lyn120 kernel: GPR08: 000F3E8C 00000000 000003F8 C4BB5C80 000000 0D 1001F2F0 00000000 100A4D90 Jul 24 17:41:01 lyn120 kernel: GPR16: 10010000 7FFFDCA8 00000000 C037D5C0 000090 32 00000000 CC3F598C 00000000 Jul 24 17:41:01 lyn120 kernel: GPR24: 00000000 00000000 CC3F0000 CC3F0000 000000 13 0000000A 00000000 F8000000 Jul 24 17:41:01 lyn120 kernel: Call backtrace: Jul 24 17:41:01 lyn120 kernel: CC0148E8 CC0149F0 C0068390 C00B6100 CC0133F0 C001 B078 C0003900 Jul 24 17:41:01 lyn120 kernel: 10010000 100033F0 10003AB4 0FF0B69C 00000000 Jul 24 17:41:01 lyn120 kernel: Kernel panic: kernel access of bad area pc cc0148 e8 lr cc0148e8 address 904A tsk insmod/477 Jul 24 17:46:22 lyn120 syslogd 1.3-3: restart.
So I tried 0xf9000000...same sort of thing happens.
If I try to initialize the driver without mapping, it will complain that the device is in use.
I'm a memory-management virgin, can someone point me in the right direction here? It seems to me that this should be very simple and straightforward, but I must be missing something.
relevant code from the 2.2 version of the driver:
/* Get the board I/O address */
iobase = (pdev->base_address[0] & PCI_BASE_ADDRESS_MEM_MASK);
printk("iobase = 0x%x\n", iobase);
/* Map iobase to LINUX virtual address. */
iobase = 0xF8000000;
printk("iobase now = 0x%x\n", iobase);
(void *)iobase = ioremap((unsigned long)iobase, (unsignedlong)(256*1024));
printk("remapped iobase = 0x%x\n", iobase);--
Regards,
Darin W. Smith
`When you say "I wrote a program that crashed Windows", people just stare at
you blankly and say "Hey, I got those with the system, *for free*".'
--Linus Torvalds
** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/