Re: [LinuxBIOS] Disabled device help

[Lu, Yinghai]

The Kernel pci_read_bases in drivers/pci/probe.c. has problem when size
if bigger than 4G.

It will first check the 32 bit SPACE_ADDRESS, and pci_size will return
0, So it will skip that resource. And the pre-set values is ignored.

And later, it will try to allocate the value. It will fail.

I will produce one patch for you.

YH

-----Original Message-----
From: [e-mail]
[mailto:[e-mail]] On Behalf Of Lu, Yinghai
Sent: Friday, November 03, 2006 11:46 AM
To: [e-mail]; LinuxBIOS
Subject: Re: [LinuxBIOS] Disabled device help

Boot log?

In LinuxBIOS, there is CONFIG_PCI_64BIT_PREF_MEM, did you try to enable
it in you MB Options.lb

In the kernel, there is pci_assign_unassigned_resources that try to
allocate resource for some devices, and it could cause problem because
Opteron could bave multi ht chain, Aka peer root bus, and every root bus
need to have its own ioport_resource, and iomem_resource. But that only
happen when BIOS is not allocating resource correctly. And LinuxBIOS
does a good job, and every device have be  handled properly.

In case, you have werid pci devices, you still can use pci_quirks in
kernel to do some fix up.

YH

-----Original Message-----
From: [e-mail]
[mailto:[e-mail]] On Behalf Of Myles Watson
Sent: Friday, November 03, 2006 11:02 AM
To: 'LinuxBIOS'
Subject: [LinuxBIOS] Disabled device help

My large device gets disabled somewhere in LinuxBIOS or the kernel. This
is
the encouraging line from LinuxBIOS telling me that the registers are
indeed
set correctly.

PCI: 05:01.0 18 <- [8000000000 - bfffffffff] prefmem64

Unfortunately, Linux sees something different and fails to allocate this
region because it sees it at 00000080-000000bf and that conflicts with
other
things.

Does anyone have a pointer to where I should look for this problem? Has
anyone else had problems with devices larger than 4GB?

Thanks,
Myles

[LinuxBIOS] Disabled device help

[Lu, Yinghai]

Please check the patch.

YH

-----Original Message-----
From: [e-mail] on behalf of Lu, Yinghai
Sent: Fri 11/3/2006 5:01 PM
To: [e-mail]; LinuxBIOS
Subject: Re: [LinuxBIOS] Disabled device help

The Kernel pci_read_bases in drivers/pci/probe.c. has problem when size
if bigger than 4G.

It will first check the 32 bit SPACE_ADDRESS, and pci_size will return
0, So it will skip that resource. And the pre-set values is ignored.

And later, it will try to allocate the value. It will fail.

I will produce one patch for you.

YH

-----Original Message-----
From: [e-mail]
[mailto:[e-mail]] On Behalf Of Lu, Yinghai
Sent: Friday, November 03, 2006 11:46 AM
To: [e-mail]; LinuxBIOS
Subject: Re: [LinuxBIOS] Disabled device help

Boot log?

In LinuxBIOS, there is CONFIG_PCI_64BIT_PREF_MEM, did you try to enable
it in you MB Options.lb

In the kernel, there is pci_assign_unassigned_resources that try to
allocate resource for some devices, and it could cause problem because
Opteron could bave multi ht chain, Aka peer root bus, and every root bus
need to have its own ioport_resource, and iomem_resource. But that only
happen when BIOS is not allocating resource correctly. And LinuxBIOS
does a good job, and every device have be  handled properly.

In case, you have werid pci devices, you still can use pci_quirks in
kernel to do some fix up.

YH

-----Original Message-----
From: [e-mail]
[mailto:[e-mail]] On Behalf Of Myles Watson
Sent: Friday, November 03, 2006 11:02 AM
To: 'LinuxBIOS'
Subject: [LinuxBIOS] Disabled device help

My large device gets disabled somewhere in LinuxBIOS or the kernel. This
is
the encouraging line from LinuxBIOS telling me that the registers are
indeed
set correctly.

PCI: 05:01.0 18 <- [8000000000 - bfffffffff] prefmem64

Unfortunately, Linux sees something different and fails to allocate this
region because it sees it at 00000080-000000bf and that conflicts with
other
things.

Does anyone have a pointer to where I should look for this problem? Has
anyone else had problems with devices larger than 4GB?

Thanks,
Myles

[LinuxBIOS] 64 bit pref mem patch for ck804

[Lu, Yinghai]

Myles,

I can not commit the patch directly, So please apply that in your own tree.

YH

[LinuxBIOS] 64 bit pref mem patch for ck804

[Stefan Reinauer]

* Lu, Yinghai  [061104 03:31]:
> Myles,
>
> I can not commit the patch directly, So please apply that in your own tree.

Is a similar patch required for amd 8xxx?

Content-Description: ck804_64bit_pref_mem.diff
> ck804 64bit pref mem support
>
> Signed-off-by: Yinghai Lu

Acked-by: Stefan Reinauer
> Index: ck804_pci.c
>
> --- ck804_pci.c	(revision 2486)
> +++ ck804_pci.c	(working copy)
> @@ -14,6 +14,10 @@
>  {
>
> uint32_t dword;
> +#if CONFIG_PCI_64BIT_PREF_MEM == 1
> device_t pci_domain_dev;
> struct resource *mem1, *mem2;
>
> +#endif
>
> /* System error enable */
>    dword = pci_read_config32(dev, 0x04);
> @@ -34,8 +38,27 @@
>          pci_write_config32(dev, 0x4c, dword);
>  #endif
>
> +#if CONFIG_PCI_64BIT_PREF_MEM == 1
> pci_domain_dev = dev->bus->dev;
> while(pci_domain_dev) {
> if(pci_domain_dev->path.type == DEVICE_PATH_PCI_DOMAIN) break;
> pci_domain_dev = pci_domain_dev->bus->dev;
> }
> if(!pci_domain_dev) return; // impossiable

It should read "impossible" or better "impossible because we need a PCI domain"

but how can we end up without a pci domain on a ck804 mainboard? I would
think this can only happen if the mainboard config file is wrong?
> mem1 = find_resource(pci_domain_dev, 1); // prefmem, it could be 64bit
> mem2 = find_resource(pci_domain_dev, 2); // mem
> if(mem1->base > mem2->base) {
> dword = mem2->base  & (0xffff0000UL);
> printk_debug("PCI DOMAIN mem2 base = 0x%010Lx\n", mem2->base);
> } else {
> dword = mem1->base  & (0xffff0000UL);
> printk_debug("PCI DOMAIN mem1 (prefmem) base = 0x%010Lx\n", mem1->base);
> }
>
> +#else
>    dword = dev_root.resource[1].base & (0xffff0000UL);
> printk_debug("dev_root mem base = 0x%010Lx\n", dev_root.resource[1].base);
> printk_debug("dev_root mem base = 0x%010Lx\n", dev_root.resource[1].base);
>
> +#endif
>
> printk_debug("[0x50] <-- 0x%08x\n", dword);
>          pci_write_config32(dev, 0x50, dword); //TOM

[LinuxBIOS] 64 bit pref mem patch for ck804

[yhlu]

amd8111, don't need that.

YH

[LinuxBIOS] Disabled device help

[Myles Watson]

Thanks a lot! It worked fine, except it missed the last 4GB of the
allocation. This is an ugly way of getting it back.  I'm not sure what the
elegant way would be. The lower 32-bits need to be filled with ones in order
for the space to be allocated correctly. Otherwise a 4GB allocation would
show up as a one byte allocation.

Now /proc/iomem shows my device at 8000000000-bfffffffff

Myles

--- drivers/pci/probe.c	2006-11-06 09:23:03.000000000 -0700
+++ drivers/pci/probe_new.c	2006-11-06 10:17:08.000000000 -0700
@@ -203,7 +203,7 @@
    		res->end = res->start + sz;
    		if (szhi) {
    			/* This BAR needs > 4GB?  Wow. */
-				res->end |= (unsigned long)szhi<<32;
+				res->end |= ((unsigned
long)szhi<<32|0xffffffffUL);
    		}
 #else
    		if (szhi) {
> Please check the patch.
> YH
> The Kernel pci_read_bases in drivers/pci/probe.c. has problem when size
> if bigger than 4G.
> It will first check the 32 bit SPACE_ADDRESS, and pci_size will return
> 0, So it will skip that resource. And the pre-set values is ignored.
> And later, it will try to allocate the value. It will fail.
> I will produce one patch for you.
> YH
> > My large device gets disabled somewhere in LinuxBIOS or the kernel. This
> > is the encouraging line from LinuxBIOS telling me that the registers are
> > indeed set correctly.
> > PCI: 05:01.0 18 <- [8000000000 - bfffffffff] prefmem64
> > Unfortunately, Linux sees something different and fails to allocate this
> > region because it sees it at 00000080-000000bf and that conflicts with
> > other things.
> > Does anyone have a pointer to where I should look for this problem? Has
> > anyone else had problems with devices larger than 4GB?
> > Thanks,
> > Myles

[LinuxBIOS] Disabled device help

[Segher Boessenkool]

> 		res->end |= (unsigned long)szhi<<32;
> 		res->end |= ((unsigned
>
> long)szhi<<32|0xffffffffUL);

Should read

res->end |= ((uint64_t)szhi << 32) | 0xffffffff;

("long" is 32-bits on many platforms).

Segher