powerpc/pseries: fix of_read_drc_info_cell() to point at next record
PATCH 1/1 powerpc/pseries: fix of_read_drc_info_cell() to point at next record
The expectation is that when calling of_read_drc_info_cell() repeatedly to parse multiple drc-info records that the in/out curval parameter points at the start of the next record on return. However, the current behavior has curval still pointing at the final value of the record just parsed. The result of which is that if the ibm,drc-info property contains multiple properties the parsed value of the drc_type for any record after the first has the power_domain value of the previous record appended to the type string.
Ex: observed the following 0xffffffff prepended to PHB
[ 69.485037] drc-info: type: \xff\xff\xff\xffPHB, prefix: PHB , index_start: 0x20000001 [ 69.485038] drc-info: suffix_start: 1, sequential_elems: 3072, sequential_inc: 1 [ 69.485038] drc-info: power-domain: 0xffffffff, last_index: 0x20000c00
Fix by incrementing curval past the power_domain value to point at drc_type string of next record.
--- arch/powerpc/platforms/pseries/of_helpers.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/powerpc/platforms/pseries/of_helpers.c +1 -1
@@ -88,7 +88,7 @@ int of_read_drc_info_cell(struct property **prop, const __be32 **curval,return -EINVAL ;/* Should now know end of current entry */ - (*curval) = (void *)p2; + (*curval) = (void *)(++p2); data->last_drc_index = data->drc_index_start + ((data->num_sequential_elems -1 ) * data->sequential_inc);
Signature · 1 line
2.16.4
Quoted 11 lines · Tyrel Datwyler
The expectation is that when calling
of_read_drc_info_cell()repeatedly to parse multiple drc-info records that the in/out curval parameter points at the start of the next record on return. However, the current behavior has curval still pointing at the final value of the record just parsed. The result of which is that if the ibm,drc-info property contains multiple properties the parsed value of thedrc_typefor any record after the first has thepower_domainvalue of the previous record appended to the type string.Ex: observed the following
0xffffffffprepended to PHB
Quoted 8 lines
[ 69.485037] drc-info: type: \xff\xff\xff\xffPHB, prefix: PHB , index_start: 0x20000001 [ 69.485038] drc-info: suffix_start: 1, sequential_elems: 3072, sequential_inc: 1 [ 69.485038] drc-info: power-domain: 0xffffffff, last_index: 0x20000c00Fix by incrementing curval past the
power_domainvalue to point atdrc_typestring of next record.Fixes a29396653b8bf ("pseries/drc-info: Search DRC properties for CPU indexes")
I have a different commit hash for that:e83636ac3334 pseries/drc-info: Search DRC properties for CPU indexes
Quoted 1 line
Signed-off-by Tyrel Datwyler <[email protected]>
Otherwise:
Quoted 16 lines · Nathan Lynch
Quoted 15 lines · Tyrel Datwyler
The expectation is that when calling
of_read_drc_info_cell()repeatedly to parse multiple drc-info records that the in/out curval parameter points at the start of the next record on return. However, the current behavior has curval still pointing at the final value of the record just parsed. The result of which is that if the ibm,drc-info property contains multiple properties the parsed value of thedrc_typefor any record after the first has thepower_domainvalue of the previous record appended to the type string.Ex: observed the following
0xffffffffprepended to PHB[ 69.485037] drc-info: type: \xff\xff\xff\xffPHB, prefix: PHB , index_start: 0x20000001 [ 69.485038] drc-info: suffix_start: 1, sequential_elems: 3072, sequential_inc: 1 [ 69.485038] drc-info: power-domain: 0xffffffff, last_index: 0x20000c00
Quoted 7 lines
Quoted 4 lines
Fix by incrementing curval past the
power_domainvalue to point atdrc_typestring of next record.Fixes a29396653b8bf ("pseries/drc-info: Search DRC properties for CPU indexes")I have a different commit hash for that:
e83636ac3334pseries/drc-info: Search DRC properties for CPU indexes
Oof, looks like I grabbed the commit hash from the SLES 15 SP1 branch. You have the correct upstream commit.
Michael, let me know if you want me to resubmit, or if you will fixup the Fixes tag on your end?
-Tyrel
Quoted 6 lines
Quoted 2 lines
Signed-off-by Tyrel Datwyler <[email protected]>Otherwise:
Acked-by Nathan Lynch <[email protected]>
Michael Ellerman
reviewerQuoted 25 lines · Tyrel Datwyler
Quoted 16 lines · Nathan Lynch
Quoted 15 lines · Tyrel Datwyler
The expectation is that when calling
of_read_drc_info_cell()repeatedly to parse multiple drc-info records that the in/out curval parameter points at the start of the next record on return. However, the current behavior has curval still pointing at the final value of the record just parsed. The result of which is that if the ibm,drc-info property contains multiple properties the parsed value of thedrc_typefor any record after the first has thepower_domainvalue of the previous record appended to the type string.Ex: observed the following
0xffffffffprepended to PHB[ 69.485037] drc-info: type: \xff\xff\xff\xffPHB, prefix: PHB , index_start: 0x20000001 [ 69.485038] drc-info: suffix_start: 1, sequential_elems: 3072, sequential_inc: 1 [ 69.485038] drc-info: power-domain: 0xffffffff, last_index: 0x20000c00Quoted 7 lines
Quoted 4 lines
Fix by incrementing curval past the
power_domainvalue to point atdrc_typestring of next record.Fixes a29396653b8bf ("pseries/drc-info: Search DRC properties for CPU indexes")I have a different commit hash for that:
e83636ac3334pseries/drc-info: Search DRC properties for CPU indexes
Quoted 7 lines
Oof, looks like I grabbed the commit hash from the SLES 15 SP1 branch. You have the correct upstream commit.
Fixes e83636ac3334 ("pseries/drc-info: Search DRC properties for CPU indexes")Michael, let me know if you want me to resubmit, or if you will fixup the Fixes tag on your end?
I can update the Fixes tag.
What's the practical effect of this bug? It seems like it should break all the hotplug stuff, but presumably it doesn't in practice for some reason?
It would also be *really* nice if we had some unit tests for this parsing code, it's demonstrably very bug prone.
cheers
Quoted 3 lines · Michael Ellerman
It would also be *really* nice if we had some unit tests for this parsing code, it's demonstrably very bug prone.
Can you say more about what form unit tests could take? Like some self tests that run at boot time, or is there a way to run the code in a user space test harness?
Michael Ellerman
reviewerQuoted 8 lines · Nathan Lynch
Quoted 3 lines · Michael Ellerman
It would also be *really* nice if we had some unit tests for this parsing code, it's demonstrably very bug prone.
Can you say more about what form unit tests could take? Like some self tests that run at boot time, or is there a way to run the code in a user space test harness?
It depends.
A userspace selftest is ideal as it's the easiest option for people to run, ie. they just have to build a user program. There's also CI systems setup to run the kernel selftests automatically.
See eg. tools/testing/selftests/powerpc/vphn/vphn.c
We also have boot time tests, they are good for code that we want to test in the kernel proper, or that are hard to extract into a userspace program. Most of those are configurable, so testing them requires someone to enable the appropriate CONFIG_FOO and build & boot that kernel. That reduces coverage obviously.
See eg. arch/powerpc/lib/code-patching.c
These days there's also kunit, which is a "proper" way to do kernel unit tests. They get built into the kernel but then there's some support for running those like a user program using UML on x86. On powerpc we'd have to boot the kunit enabled kernel on hardware or in qemu to exercise the tests. So they're essentially just boot time tests but with some nicer infrastructure vs doing it all by hand like we used to.
In this case the actual function we want to test could be trivially built into a userspace program, but the underlying device tree helpers probably can't be.
eg. drivers/of/property.c is quite large and contains quite a few things, we'd need to shim quite a bit to get that building in userspace I suspect, which then becomes a maintenance burden.
So this is probably a good candidate for a kunit test, though I haven't personally written/converted any tests to kunit so I can't say exactly how easy that is.
cheers
Quoted 36 lines · Michael Ellerman
Quoted 25 lines · Tyrel Datwyler
Quoted 16 lines · Nathan Lynch
Quoted 15 lines · Tyrel Datwyler
The expectation is that when calling
of_read_drc_info_cell()repeatedly to parse multiple drc-info records that the in/out curval parameter points at the start of the next record on return. However, the current behavior has curval still pointing at the final value of the record just parsed. The result of which is that if the ibm,drc-info property contains multiple properties the parsed value of thedrc_typefor any record after the first has thepower_domainvalue of the previous record appended to the type string.Ex: observed the following
0xffffffffprepended to PHB[ 69.485037] drc-info: type: \xff\xff\xff\xffPHB, prefix: PHB , index_start: 0x20000001 [ 69.485038] drc-info: suffix_start: 1, sequential_elems: 3072, sequential_inc: 1 [ 69.485038] drc-info: power-domain: 0xffffffff, last_index: 0x20000c00Quoted 7 lines
Quoted 4 lines
Fix by incrementing curval past the
power_domainvalue to point atdrc_typestring of next record.Fixes a29396653b8bf ("pseries/drc-info: Search DRC properties for CPU indexes")I have a different commit hash for that:
e83636ac3334pseries/drc-info: Search DRC properties for CPU indexesQuoted 7 lines
Oof, looks like I grabbed the commit hash from the SLES 15 SP1 branch. You have the correct upstream commit.
Fixes e83636ac3334 ("pseries/drc-info: Search DRC properties for CPU indexes")Michael, let me know if you want me to resubmit, or if you will fixup the Fixes tag on your end?
I can update the Fixes tag.
Quoted 9 lines
What's the practical effect of this bug? It seems like it should break all the hotplug stuff, but presumably it doesn't in practice for some reason?
It would also be *really* nice if we had some unit tests for this parsing code, it's demonstrably very bug prone.
cheers
In practice PHBs are the only type of connector in the ibm,drc-info property that has multiple records. So, it breaks PHB hotplug, but by chance not pci, cpu, slot, or mem because they happen to only ever be a single record.
I have a follow up patch to at least add some pr_debug at the end of this parsing function to dump each record as its read.
-Tyrel
Michael Ellerman
reviewerQuoted 41 lines · Tyrel Datwyler
Quoted 40 lines · Michael Ellerman
Quoted 25 lines · Tyrel Datwyler
Quoted 16 lines · Nathan Lynch
Quoted 15 lines · Tyrel Datwyler
The expectation is that when calling
of_read_drc_info_cell()repeatedly to parse multiple drc-info records that the in/out curval parameter points at the start of the next record on return. However, the current behavior has curval still pointing at the final value of the record just parsed. The result of which is that if the ibm,drc-info property contains multiple properties the parsed value of thedrc_typefor any record after the first has thepower_domainvalue of the previous record appended to the type string.Ex: observed the following
0xffffffffprepended to PHB[ 69.485037] drc-info: type: \xff\xff\xff\xffPHB, prefix: PHB , index_start: 0x20000001 [ 69.485038] drc-info: suffix_start: 1, sequential_elems: 3072, sequential_inc: 1 [ 69.485038] drc-info: power-domain: 0xffffffff, last_index: 0x20000c00Quoted 7 lines
Quoted 4 lines
Fix by incrementing curval past the
power_domainvalue to point atdrc_typestring of next record.Fixes a29396653b8bf ("pseries/drc-info: Search DRC properties for CPU indexes")I have a different commit hash for that:
e83636ac3334pseries/drc-info: Search DRC properties for CPU indexesQuoted 7 lines
Oof, looks like I grabbed the commit hash from the SLES 15 SP1 branch. You have the correct upstream commit.
Fixes e83636ac3334 ("pseries/drc-info: Search DRC properties for CPU indexes")Michael, let me know if you want me to resubmit, or if you will fixup the Fixes tag on your end?
I can update the Fixes tag.
What's the practical effect of this bug? It seems like it should break all the hotplug stuff, but presumably it doesn't in practice for some reason?
Quoted 6 lines
Quoted 2 lines
It would also be *really* nice if we had some unit tests for this parsing code, it's demonstrably very bug prone.
In practice PHBs are the only type of connector in the ibm,drc-info property that has multiple records. So, it breaks PHB hotplug, but by chance not pci, cpu, slot, or mem because they happen to only ever be a single record.
Thanks. I folded that into the change log.
cheers
Michael Ellerman
AppliedQuoted 15 lines · Tyrel Datwyler
The expectation is that when calling
of_read_drc_info_cell()repeatedly to parse multiple drc-info records that the in/out curval parameter points at the start of the next record on return. However, the current behavior has curval still pointing at the final value of the record just parsed. The result of which is that if the ibm,drc-info property contains multiple properties the parsed value of thedrc_typefor any record after the first has thepower_domainvalue of the previous record appended to the type string.Ex: observed the following
0xffffffffprepended to PHB[ 69.485037] drc-info: type: \xff\xff\xff\xffPHB, prefix: PHB , index_start: 0x20000001 [ 69.485038] drc-info: suffix_start: 1, sequential_elems: 3072, sequential_inc: 1 [ 69.485038] drc-info: power-domain: 0xffffffff, last_index: 0x20000c00
Quoted 5 lines
Fix by incrementing curval past the
power_domainvalue to point atdrc_typestring of next record.Fixes a29396653b8bf ("pseries/drc-info: Search DRC properties for CPU indexes")Signed-off-by Tyrel Datwyler <[email protected]>
Applied to powerpc next, thanks.
https://git.kernel.org/powerpc/c/c5e76fa05b2df519b9f08571cc57e623c1569faa
cheers