[PATCH 0/1] mm/debug_vm_pgtable: fix alignment for pmd/pud_advanced_tests()

[PATCH 0/1] mm/debug_vm_pgtable: fix alignment for pmd/pud_advanced_tests()

[Gerald Schaefer]

We sometimes see a "BUG task_struct (Not tainted): Padding overwritten"
on s390, directly after running debug_vm_pgtable. This is because of
wrong vaddr alignment in pmd/pud_advanced_tests(), leading to memory
corruption at least on s390, see patch description.

At first glance, other architectures do not seem to care about vaddr in
their xxx_get_and_clear() implementations, so they should not be affected.
One exception is sparc, where the addr is passed over to some tlb_batch
code, but I'm not sure what implication the wrongly aligned vaddr would
have in this case.

Also adding linux-arch, just to make sure.

Gerald Schaefer (1):
  mm/debug_vm_pgtable: fix alignment for pmd/pud_advanced_tests()

 mm/debug_vm_pgtable.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

-- 
2.25.1

[PATCH 1/1] mm/debug_vm_pgtable: fix alignment for pmd/pud_advanced_tests()

[Gerald Schaefer]

In pmd/pud_advanced_tests(), the vaddr is aligned up to the next pmd/pud
entry, and so it does not match the given pmdp/pudp and (aligned down) pfn
any more.

For s390, this results in memory corruption, because the IDTE instruction
used e.g. in xxx_get_and_clear() will take the vaddr for some calculations,
in combination with the given pmdp. It will then end up with a wrong table
origin, ending on ...ff8, and some of those wrongly set low-order bits will
also select a wrong pagetable level for the index addition. IDTE could
therefore invalidate (or 0x20) something outside of the page tables,
depending on the wrongly picked index, which in turn depends on the random
vaddr.

As result, we sometimes see "BUG task_struct (Not tainted): Padding
overwritten" on s390, where one 0x5a padding value got overwritten with
0x7a.

Fix this by aligning down, similar to how the pmd/pud_aligned pfns are
calculated.

Fixes: a5c3b9ffb0f40 ("mm/debug_vm_pgtable: add tests validating advanced arch page table helpers")
Cc: <stable@vger.kernel.org> # v5.9+
Signed-off-by: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
---
 mm/debug_vm_pgtable.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/debug_vm_pgtable.c b/mm/debug_vm_pgtable.c
index 6ff92c8b0a00..f7b23565a04f 100644
--- a/mm/debug_vm_pgtable.c
+++ b/mm/debug_vm_pgtable.c
@@ -193,7 +193,7 @@ static void __init pmd_advanced_tests(struct mm_struct *mm,
 
 	pr_debug("Validating PMD advanced\n");
 	/* Align the address wrt HPAGE_PMD_SIZE */
-	vaddr = (vaddr & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE;
+	vaddr &= HPAGE_PMD_MASK;
 
 	pgtable_trans_huge_deposit(mm, pmdp, pgtable);
 
@@ -318,7 +318,7 @@ static void __init pud_advanced_tests(struct mm_struct *mm,
 
 	pr_debug("Validating PUD advanced\n");
 	/* Align the address wrt HPAGE_PUD_SIZE */
-	vaddr = (vaddr & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE;
+	vaddr &= HPAGE_PUD_MASK;
 
 	pud = pfn_pud(pfn, prot);
 	set_pud_at(mm, vaddr, pudp, pud);
-- 
2.25.1

Re: [PATCH 0/1] mm/debug_vm_pgtable: fix alignment for pmd/pud_advanced_tests()

[Anshuman Khandual]

On 5/25/21 6:30 PM, Gerald Schaefer wrote:
> We sometimes see a "BUG task_struct (Not tainted): Padding overwritten"
> on s390, directly after running debug_vm_pgtable. This is because of
> wrong vaddr alignment in pmd/pud_advanced_tests(), leading to memory
> corruption at least on s390, see patch description.
> 
> At first glance, other architectures do not seem to care about vaddr in
> their xxx_get_and_clear() implementations, so they should not be affected.

IIRC, alignment (regardless up or down) is the only requirement on certain
platforms. Probably it should not affect other platforms as this change
just aligns the virtual address down.

> One exception is sparc, where the addr is passed over to some tlb_batch
> code, but I'm not sure what implication the wrongly aligned vaddr would
> have in this case.
> 
> Also adding linux-arch, just to make sure.

Right. Not sure if this test gets to run on sparc platform for not being
currently supported. But we could take a look if there are any reported
problems because of vaddr.

> 
> Gerald Schaefer (1):
>   mm/debug_vm_pgtable: fix alignment for pmd/pud_advanced_tests()
> 
>  mm/debug_vm_pgtable.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 

Re: [PATCH 1/1] mm/debug_vm_pgtable: fix alignment for pmd/pud_advanced_tests()

[Anshuman Khandual]

On 5/25/21 6:30 PM, Gerald Schaefer wrote:
> In pmd/pud_advanced_tests(), the vaddr is aligned up to the next pmd/pud
> entry, and so it does not match the given pmdp/pudp and (aligned down) pfn
> any more.
> 
> For s390, this results in memory corruption, because the IDTE instruction
> used e.g. in xxx_get_and_clear() will take the vaddr for some calculations,
> in combination with the given pmdp. It will then end up with a wrong table
> origin, ending on ...ff8, and some of those wrongly set low-order bits will
> also select a wrong pagetable level for the index addition. IDTE could
> therefore invalidate (or 0x20) something outside of the page tables,
> depending on the wrongly picked index, which in turn depends on the random
> vaddr.
> 
> As result, we sometimes see "BUG task_struct (Not tainted): Padding
> overwritten" on s390, where one 0x5a padding value got overwritten with
> 0x7a.
> 
> Fix this by aligning down, similar to how the pmd/pud_aligned pfns are
> calculated.
> 
> Fixes: a5c3b9ffb0f40 ("mm/debug_vm_pgtable: add tests validating advanced arch page table helpers")
> Cc: <stable@vger.kernel.org> # v5.9+
> Signed-off-by: Gerald Schaefer <gerald.schaefer@linux.ibm.com>

Did not see any problem on arm64 or x86, builds okay across all
supported platforms. It would be great, if folks could test this
on remaining platforms i.e arc, riscv etc.

+ Vineet Gupta <vgupta@synopsys.com>
+ Palmer Dabbelt <palmer@dabbelt.com>
+ Paul Walmsley <paul.walmsley@sifive.com>

Reviewed-by: Anshuman Khandual <anshuman.khandual@arm.com>

> ---
>  mm/debug_vm_pgtable.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/debug_vm_pgtable.c b/mm/debug_vm_pgtable.c
> index 6ff92c8b0a00..f7b23565a04f 100644
> --- a/mm/debug_vm_pgtable.c
> +++ b/mm/debug_vm_pgtable.c
> @@ -193,7 +193,7 @@ static void __init pmd_advanced_tests(struct mm_struct *mm,
>  
>  	pr_debug("Validating PMD advanced\n");
>  	/* Align the address wrt HPAGE_PMD_SIZE */
> -	vaddr = (vaddr & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE;
> +	vaddr &= HPAGE_PMD_MASK;
>  
>  	pgtable_trans_huge_deposit(mm, pmdp, pgtable);
>  
> @@ -318,7 +318,7 @@ static void __init pud_advanced_tests(struct mm_struct *mm,
>  
>  	pr_debug("Validating PUD advanced\n");
>  	/* Align the address wrt HPAGE_PUD_SIZE */
> -	vaddr = (vaddr & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE;
> +	vaddr &= HPAGE_PUD_MASK;
>  
>  	pud = pfn_pud(pfn, prot);
>  	set_pud_at(mm, vaddr, pudp, pud);
> 

Re: [PATCH 1/1] mm/debug_vm_pgtable: fix alignment for pmd/pud_advanced_tests()

[Anatoly Pugachev]

On Tue, May 25, 2021 at 4:03 PM Gerald Schaefer
<gerald.schaefer@linux.ibm.com> wrote:
>
> In pmd/pud_advanced_tests(), the vaddr is aligned up to the next pmd/pud
> entry, and so it does not match the given pmdp/pudp and (aligned down) pfn
> any more.
>
> For s390, this results in memory corruption, because the IDTE instruction
> used e.g. in xxx_get_and_clear() will take the vaddr for some calculations,
> in combination with the given pmdp. It will then end up with a wrong table
> origin, ending on ...ff8, and some of those wrongly set low-order bits will
> also select a wrong pagetable level for the index addition. IDTE could
> therefore invalidate (or 0x20) something outside of the page tables,
> depending on the wrongly picked index, which in turn depends on the random
> vaddr.
>
> As result, we sometimes see "BUG task_struct (Not tainted): Padding
> overwritten" on s390, where one 0x5a padding value got overwritten with
> 0x7a.
>
> Fix this by aligning down, similar to how the pmd/pud_aligned pfns are
> calculated.
>
> Fixes: a5c3b9ffb0f40 ("mm/debug_vm_pgtable: add tests validating advanced arch page table helpers")
> Cc: <stable@vger.kernel.org> # v5.9+
> Signed-off-by: Gerald Schaefer <gerald.schaefer@linux.ibm.com>

boot tested on sparc64 with quick run of stress-ng ( --class memory
--sequential -1 --timeout 10s -v --pathological --oomable
--metrics-brief )
stress-ng: debug: [371408] system: Linux ttip
5.13.0-rc3-00043-gad9f25d33860-dirty #218 SMP Wed May 26 11:55:54 MSK
2021 sparc64

Tested-by: Anatoly Pugachev <matorola@gmail.com>

Re: [PATCH 1/1] mm/debug_vm_pgtable: fix alignment for pmd/pud_advanced_tests()

[Anshuman Khandual]

On 5/26/21 4:57 PM, Anatoly Pugachev wrote:
> On Tue, May 25, 2021 at 4:03 PM Gerald Schaefer
> <gerald.schaefer@linux.ibm.com> wrote:
>>
>> In pmd/pud_advanced_tests(), the vaddr is aligned up to the next pmd/pud
>> entry, and so it does not match the given pmdp/pudp and (aligned down) pfn
>> any more.
>>
>> For s390, this results in memory corruption, because the IDTE instruction
>> used e.g. in xxx_get_and_clear() will take the vaddr for some calculations,
>> in combination with the given pmdp. It will then end up with a wrong table
>> origin, ending on ...ff8, and some of those wrongly set low-order bits will
>> also select a wrong pagetable level for the index addition. IDTE could
>> therefore invalidate (or 0x20) something outside of the page tables,
>> depending on the wrongly picked index, which in turn depends on the random
>> vaddr.
>>
>> As result, we sometimes see "BUG task_struct (Not tainted): Padding
>> overwritten" on s390, where one 0x5a padding value got overwritten with
>> 0x7a.
>>
>> Fix this by aligning down, similar to how the pmd/pud_aligned pfns are
>> calculated.
>>
>> Fixes: a5c3b9ffb0f40 ("mm/debug_vm_pgtable: add tests validating advanced arch page table helpers")
>> Cc: <stable@vger.kernel.org> # v5.9+
>> Signed-off-by: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
> 
> boot tested on sparc64 with quick run of stress-ng ( --class memory
> --sequential -1 --timeout 10s -v --pathological --oomable
> --metrics-brief )
> stress-ng: debug: [371408] system: Linux ttip
> 5.13.0-rc3-00043-gad9f25d33860-dirty #218 SMP Wed May 26 11:55:54 MSK
> 2021 sparc64
> 
> Tested-by: Anatoly Pugachev <matorola@gmail.com>
> 

spac64 does not enable ARCH_HAS_DEBUG_VM_PGTABLE, did you enable it
before running the test ? Did the entire test debug_vm_pgtable() run
successfully on sparc64 ?

Re: [PATCH 1/1] mm/debug_vm_pgtable: fix alignment for pmd/pud_advanced_tests()

[Anatoly Pugachev]

On Wed, May 26, 2021 at 3:35 PM Anshuman Khandual
<anshuman.khandual@arm.com> wrote:
> spac64 does not enable ARCH_HAS_DEBUG_VM_PGTABLE, did you enable it
> before running the test ? Did the entire test debug_vm_pgtable() run
> successfully on sparc64 ?

Ahh.. Sorry for the noise then... Thought that
CONFIG_TRANSPARENT_HUGEPAGE would be enough...