Aravind Gopalakrishnan (1):
AMD64_EDAC: Fix incorrect wrap arounds due to left shift beyond 32
bits.
drivers/edac/amd64_edac.c | 9 +++++----
1 file changed, 5 insertions(+), 4 deletions(-)
--
1.7.10.4
Link to the bug report:
http://marc.info/?l=linux-edac&m=137692201732220&w=2
dct_base and dct_limit obtain 32 bit register values when they read their
respective pci config space registers. A left shift beyond 32 bits will
cause them to wrap around. Similar case for chan_addr as can be seen from
the bug report. In the patch, we rectify this by casting chan_addr to u64
and by comparing dct_base and dct_limit against (sys_addr >> 27)
Tested on F15h, M30h with ECC turned on and works fine.
Signed-off-by: Aravind Gopalakrishnan <[email protected]>
diff --git a/drivers/edac/amd64_edac.c b/drivers/edac/amd64_edac.c
index b86228c..eb4793e 100644
--- a/drivers/edac/amd64_edac.c
+++ b/drivers/edac/amd64_edac.c
@@ -1558,11 +1558,12 @@ static int f15_m30h_match_to_this_node(struct amd64_pvt *pvt, unsigned range,
}
/* Verify sys_addr is within DCT Range. */
- dct_base = (dct_sel_baseaddr(pvt) << 27);
- dct_limit = (((dct_cont_limit_reg >> 11) & 0x1FFF) << 27) | 0x7FFFFFF;
+ dct_base = dct_sel_baseaddr(pvt);
+ dct_limit = (dct_cont_limit_reg >> 11) & 0x1FFF;
if (!(dct_cont_base_reg & BIT(0)) &&
- !(dct_base <= sys_addr && dct_limit >= sys_addr))
+ !(dct_base <= (sys_addr >> 27) &&
+ dct_limit >= (sys_addr >> 27)))
return -EINVAL;
/* Verify number of dct's that participate in channel interleaving. */
@@ -1614,7 +1615,7 @@ static int f15_m30h_match_to_this_node(struct amd64_pvt *pvt, unsigned range,
amd64_read_pci_cfg(pvt->F1,
DRAM_CONT_HIGH_OFF + (int) channel * 4,
&tmp);
- chan_addr += ((tmp >> 11) & 0xfff) << 27;
+ chan_addr += (u64) ((tmp >> 11) & 0xfff) << 27;
}
f15h_select_dct(pvt, channel);
--
1.7.10.4
On Mon, Aug 19, 2013 at 07:27:51PM -0500, Aravind Gopalakrishnan wrote:
> Aravind Gopalakrishnan (1):
> AMD64_EDAC: Fix incorrect wrap arounds due to left shift beyond 32
> bits.
>
> drivers/edac/amd64_edac.c | 9 +++++----
> 1 file changed, 5 insertions(+), 4 deletions(-)
Just a minor thing: when you send one patch only, you don't really need
the 0/1 email.
Thanks.
--
Regards/Gruss,
Boris.
Sent from a fat crate under my desk. Formatting is fine.
--
On Mon, Aug 19, 2013 at 07:27:52PM -0500, Aravind Gopalakrishnan wrote:
> Link to the bug report:
> http://marc.info/?l=linux-edac&m=137692201732220&w=2
>
> dct_base and dct_limit obtain 32 bit register values when they read their
> respective pci config space registers. A left shift beyond 32 bits will
> cause them to wrap around. Similar case for chan_addr as can be seen from
> the bug report. In the patch, we rectify this by casting chan_addr to u64
> and by comparing dct_base and dct_limit against (sys_addr >> 27)
>
> Tested on F15h, M30h with ECC turned on and works fine.
>
> Signed-off-by: Aravind Gopalakrishnan <[email protected]>
>
> diff --git a/drivers/edac/amd64_edac.c b/drivers/edac/amd64_edac.c
> index b86228c..eb4793e 100644
> --- a/drivers/edac/amd64_edac.c
> +++ b/drivers/edac/amd64_edac.c
> @@ -1558,11 +1558,12 @@ static int f15_m30h_match_to_this_node(struct amd64_pvt *pvt, unsigned range,
> }
>
> /* Verify sys_addr is within DCT Range. */
> - dct_base = (dct_sel_baseaddr(pvt) << 27);
> - dct_limit = (((dct_cont_limit_reg >> 11) & 0x1FFF) << 27) | 0x7FFFFFF;
> + dct_base = dct_sel_baseaddr(pvt);
This can't be correct.
So the original patch takes the shifted dct_base while your change
doesn't anymore...
> + dct_limit = (dct_cont_limit_reg >> 11) & 0x1FFF;
>
> if (!(dct_cont_base_reg & BIT(0)) &&
> - !(dct_base <= sys_addr && dct_limit >= sys_addr))
> + !(dct_base <= (sys_addr >> 27) &&
> + dct_limit >= (sys_addr >> 27)))
... and while this comparison shifts sys_addr to use the proper bits,
the code does this assignment later:
chan_offset = dct_base;
Now, chan_offset has the << 27 version of dct_base which makes the following
calculation wrong:
chan_addr = sys_addr - chan_offset;
because sys_addr is the full 64-bit, unshifted value.
The right thing to do would be to do:
chan_offset = dct_base << 27;
Or am I missing something?
--
Regards/Gruss,
Boris.
Sent from a fat crate under my desk. Formatting is fine.
--
On 8/23/2013 4:37 PM, Borislav Petkov wrote:
> On Mon, Aug 19, 2013 at 07:27:52PM -0500, Aravind Gopalakrishnan wrote:
>> Link to the bug report:
>> http://marc.info/?l=linux-edac&m=137692201732220&w=2
>>
>> dct_base and dct_limit obtain 32 bit register values when they read their
>> respective pci config space registers. A left shift beyond 32 bits will
>> cause them to wrap around. Similar case for chan_addr as can be seen from
>> the bug report. In the patch, we rectify this by casting chan_addr to u64
>> and by comparing dct_base and dct_limit against (sys_addr >> 27)
>>
>> Tested on F15h, M30h with ECC turned on and works fine.
>>
>> Signed-off-by: Aravind Gopalakrishnan <[email protected]>
>>
>> diff --git a/drivers/edac/amd64_edac.c b/drivers/edac/amd64_edac.c
>> index b86228c..eb4793e 100644
>> --- a/drivers/edac/amd64_edac.c
>> +++ b/drivers/edac/amd64_edac.c
>> @@ -1558,11 +1558,12 @@ static int f15_m30h_match_to_this_node(struct amd64_pvt *pvt, unsigned range,
>> }
>>
>> /* Verify sys_addr is within DCT Range. */
>> - dct_base = (dct_sel_baseaddr(pvt) << 27);
>> - dct_limit = (((dct_cont_limit_reg >> 11) & 0x1FFF) << 27) | 0x7FFFFFF;
>> + dct_base = dct_sel_baseaddr(pvt);
> This can't be correct.
>
> So the original patch takes the shifted dct_base while your change
> doesn't anymore...
>
>> + dct_limit = (dct_cont_limit_reg >> 11) & 0x1FFF;
>>
>> if (!(dct_cont_base_reg & BIT(0)) &&
>> - !(dct_base <= sys_addr && dct_limit >= sys_addr))
>> + !(dct_base <= (sys_addr >> 27) &&
>> + dct_limit >= (sys_addr >> 27)))
> ... and while this comparison shifts sys_addr to use the proper bits,
> the code does this assignment later:
>
> chan_offset = dct_base;
>
> Now, chan_offset has the << 27 version of dct_base which makes the following
> calculation wrong:
>
> chan_addr = sys_addr - chan_offset;
Oops. my apologies.
> because sys_addr is the full 64-bit, unshifted value.
>
> The right thing to do would be to do:
>
> chan_offset = dct_base << 27;
>
> Or am I missing something?
>
No, you are right.
I am re-sending the patch.
Thanks,
-Aravind.