[PATCH] makedumpfile: arm64: Fix page table walk of 1GB section

[PATCH] makedumpfile: arm64: Fix page table walk of 1GB section

[Bradley Bolen]

Using an arm64 board, 2GB of memory, makedumpfile 1.6.2, and the v4.11 kernel,
I ran into an issue where the dumpfile was not excluding any pages thus
creating a huge vmcore.

I found that my board had a 1GB mapping in the PGD of 0xffffffc040000000. This
isn't my area of expertise but I think I've pieced together how the kernel
handles page tables.

cat /proc/kallsyms |grep swapper_pg
ffffff8008a66000 B swapper_pg_dir

00a66800: 7fff9003 // pgd for virtual addr 0xffffffc000000000
00a66804: 00000000
00a66808: 40000711 // pgd for virtual addr 0xffffffc040000000
00a6680c: 00e80000

If I understand everything correctly then the pgd at 0xa66808 is a section and
not a pointer to a PMD, thus it should have been handled as a section instead
of trying to drill down to a PTE by the code in arch/arm64.c

Please find debug below and a proposed (probably not optimal) fix attached.

Thank you.

/usr/sbin/makedumpfile -F -c -d 31 --message-level=31 /proc/vmcore > /dev/null
sadump: unsupported architecture
LOAD (0)
phys_start : 80000
phys_end : a58000
virt_start : ffffff8008080000
virt_end : ffffff8008a58000
LOAD (1)
phys_start : 0
phys_end : 6000000
virt_start : ffffffc000000000
virt_end : ffffffc006000000
LOAD (2)
phys_start : 6800000
phys_end : 8000000
virt_start : ffffffc006800000
virt_end : ffffffc008000000
LOAD (3)
phys_start : a000000
phys_end : 80000000
virt_start : ffffffc00a000000
virt_end : ffffffc080000000
Linux kdump
page_size : 4096
phys_base : 0

max_mapnr : 80000
There is enough free memory to be done in one cycle.

Buffer size for the cyclic mode: 131072
kimage_voffset : ffffff8008000000
max_physmem_bits : 30
section_size_bits: 1e
page_offset : ffffffc000000000

num of NODEs : 1

Memory type : SPARSEMEM_EX

readmem: Can't convert a virtual address(ffffffc07fff6000) to physical address.
readmem: type_addr: 0, addr:ffffffc07fff6000, size:16
section_mem_map_addr: Can't get a struct mem_section(ffffffc07fff6000).
mem_map (0)
mem_map : 0
pfn_start : 0
pfn_end : 40000
readmem: Can't convert a virtual address(ffffffc07fff6010) to physical address.
readmem: type_addr: 0, addr:ffffffc07fff6010, size:16
section_mem_map_addr: Can't get a struct mem_section(ffffffc07fff6010).
mem_map (1)
mem_map : 0
pfn_start : 40000
pfn_end : 80000
mmap() is available on the kernel.
Checking for memory holes : [100.0 %] | STEP [Checking for memory holes ] :
0.017147 seconds
Excluding unnecessary pages : [100.0 %] \ STEP [Excluding unnecessary pages] :
0.000034 seconds

Bradley Bolen (1):
  makedumpfile: arm64: Fix page table walk of 1GB section

 arch/arm64.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

-- 
1.9.3


_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

[PATCH] makedumpfile: arm64: Fix page table walk of 1GB section

[Bradley Bolen]

makedumpfile was generating large (> 500MB) vmcore files for an arm64
board with 2GB of DRAM.  It was not excluding any pages because the
mem_map address was not being converted correctly.

readmem: Can't convert a virtual address(ffffffc07fff6000) to physical
address.
readmem: type_addr: 0, addr:ffffffc07fff6000, size:16
section_mem_map_addr: Can't get a struct mem_section(ffffffc07fff6000).
mem_map (0)
mem_map : 0

makedumpfile was not handling 1GB sections in the PGD and was trying to
drill down to a PTE in which it was trying to dereference invalid
memory.  This patch adds code to check the PGD for a section type and
handle it instead of treating it as a table entry.

Signed-off-by: Bradley Bolen <bradleybolen@gmail.com>
---
 arch/arm64.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/arch/arm64.c b/arch/arm64.c
index 958f57f..cae4b70 100644
--- a/arch/arm64.c
+++ b/arch/arm64.c
@@ -57,6 +57,8 @@ static unsigned long kimage_voffset;
 #define PGDIR_SHIFT		((PAGESHIFT() - 3) * pgtable_level + 3)
 #define PTRS_PER_PGD		(1 << (va_bits - PGDIR_SHIFT))
 #define PUD_SHIFT		get_pud_shift_arm64()
+#define PUD_SIZE		(1UL << PUD_SHIFT)
+#define PUD_MASK		(~(PUD_SIZE - 1))
 #define PTRS_PER_PTE		(1 << (PAGESHIFT() - 3))
 #define PTRS_PER_PUD		PTRS_PER_PTE
 #define PMD_SHIFT		((PAGESHIFT() - 3) * 2 + 3)
@@ -79,6 +81,10 @@ static unsigned long kimage_voffset;
 #define PMD_TYPE_SECT		1
 #define PMD_TYPE_TABLE		3
 
+#define PUD_TYPE_MASK		3
+#define PUD_TYPE_SECT		1
+#define PUD_TYPE_TABLE		3
+
 #define pgd_index(vaddr) 		(((vaddr) >> PGDIR_SHIFT) & (PTRS_PER_PGD - 1))
 #define pgd_offset(pgdir, vaddr)	((pgd_t *)(pgdir) + pgd_index(vaddr))
 
@@ -253,6 +259,13 @@ vaddr_to_paddr_arm64(unsigned long vaddr)
 		return NOT_PADDR;
 	}
 
+	if ((pud_val(pudv) & PUD_TYPE_MASK) == PUD_TYPE_SECT) {
+		/* 1GB section */
+		paddr = (pud_val(pudv) & (PUD_MASK & PMD_SECTION_MASK))
+					+ (vaddr & (PUD_SIZE - 1));
+		return paddr;
+	}
+
 	pmda = pmd_offset(puda, &pudv, vaddr);
 	if (!readmem(PADDR, (unsigned long long)pmda, &pmdv, sizeof(pmdv))) {
 		ERRMSG("Can't read pmd\n");
-- 
1.9.3


_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

RE: [PATCH] makedumpfile: arm64: Fix page table walk of 1GB section

[Atsushi Kumagai]

Hello Bradley,

Sorry for my late reply, I'll merge this patch into v1.6.3.
Thanks for your work.

Regards,
Atsushi Kumagai

>makedumpfile was generating large (> 500MB) vmcore files for an arm64
>board with 2GB of DRAM.  It was not excluding any pages because the
>mem_map address was not being converted correctly.
>
>readmem: Can't convert a virtual address(ffffffc07fff6000) to physical
>address.
>readmem: type_addr: 0, addr:ffffffc07fff6000, size:16
>section_mem_map_addr: Can't get a struct mem_section(ffffffc07fff6000).
>mem_map (0)
>mem_map : 0
>
>makedumpfile was not handling 1GB sections in the PGD and was trying to
>drill down to a PTE in which it was trying to dereference invalid
>memory.  This patch adds code to check the PGD for a section type and
>handle it instead of treating it as a table entry.
>
>Signed-off-by: Bradley Bolen <bradleybolen@gmail.com>
>---
> arch/arm64.c | 13 +++++++++++++
> 1 file changed, 13 insertions(+)
>
>diff --git a/arch/arm64.c b/arch/arm64.c
>index 958f57f..cae4b70 100644
>--- a/arch/arm64.c
>+++ b/arch/arm64.c
>@@ -57,6 +57,8 @@ static unsigned long kimage_voffset;
> #define PGDIR_SHIFT		((PAGESHIFT() - 3) * pgtable_level + 3)
> #define PTRS_PER_PGD		(1 << (va_bits - PGDIR_SHIFT))
> #define PUD_SHIFT		get_pud_shift_arm64()
>+#define PUD_SIZE		(1UL << PUD_SHIFT)
>+#define PUD_MASK		(~(PUD_SIZE - 1))
> #define PTRS_PER_PTE		(1 << (PAGESHIFT() - 3))
> #define PTRS_PER_PUD		PTRS_PER_PTE
> #define PMD_SHIFT		((PAGESHIFT() - 3) * 2 + 3)
>@@ -79,6 +81,10 @@ static unsigned long kimage_voffset;
> #define PMD_TYPE_SECT		1
> #define PMD_TYPE_TABLE		3
>
>+#define PUD_TYPE_MASK		3
>+#define PUD_TYPE_SECT		1
>+#define PUD_TYPE_TABLE		3
>+
> #define pgd_index(vaddr) 		(((vaddr) >> PGDIR_SHIFT) & (PTRS_PER_PGD - 1))
> #define pgd_offset(pgdir, vaddr)	((pgd_t *)(pgdir) + pgd_index(vaddr))
>
>@@ -253,6 +259,13 @@ vaddr_to_paddr_arm64(unsigned long vaddr)
> 		return NOT_PADDR;
> 	}
>
>+	if ((pud_val(pudv) & PUD_TYPE_MASK) == PUD_TYPE_SECT) {
>+		/* 1GB section */
>+		paddr = (pud_val(pudv) & (PUD_MASK & PMD_SECTION_MASK))
>+					+ (vaddr & (PUD_SIZE - 1));
>+		return paddr;
>+	}
>+
> 	pmda = pmd_offset(puda, &pudv, vaddr);
> 	if (!readmem(PADDR, (unsigned long long)pmda, &pmdv, sizeof(pmdv))) {
> 		ERRMSG("Can't read pmd\n");
>--
>1.9.3
>
>
>_______________________________________________
>kexec mailing list
>kexec@lists.infradead.org
>http://lists.infradead.org/mailman/listinfo/kexec



_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH] makedumpfile: arm64: Fix page table walk of 1GB section

[Pratyush Anand]

On Monday 14 August 2017 02:07 PM, Atsushi Kumagai wrote:
> Hello Bradley,
> 
> Sorry for my late reply, I'll merge this patch into v1.6.3.

Sorry, I missed it as well. Patch looks fine to me.However, I think /* 1GB  
Section*/ comment can be improved. Infact, other existing /* 1GB Section*/  
comment for PMD type is wrong.

> Thanks for your work.
> 
> Regards,
> Atsushi Kumagai
> 
>> makedumpfile was generating large (> 500MB) vmcore files for an arm64
>> board with 2GB of DRAM.  It was not excluding any pages because the
>> mem_map address was not being converted correctly.
>>
>> readmem: Can't convert a virtual address(ffffffc07fff6000) to physical
>> address.
>> readmem: type_addr: 0, addr:ffffffc07fff6000, size:16
>> section_mem_map_addr: Can't get a struct mem_section(ffffffc07fff6000).
>> mem_map (0)
>> mem_map : 0
>>
>> makedumpfile was not handling 1GB sections in the PGD and was trying to
>> drill down to a PTE in which it was trying to dereference invalid
>> memory.  This patch adds code to check the PGD for a section type and
>> handle it instead of treating it as a table entry.
>>
>> Signed-off-by: Bradley Bolen <bradleybolen@gmail.com>
>> ---
>> arch/arm64.c | 13 +++++++++++++
>> 1 file changed, 13 insertions(+)
>>
>> diff --git a/arch/arm64.c b/arch/arm64.c
>> index 958f57f..cae4b70 100644
>> --- a/arch/arm64.c
>> +++ b/arch/arm64.c
>> @@ -57,6 +57,8 @@ static unsigned long kimage_voffset;
>> #define PGDIR_SHIFT		((PAGESHIFT() - 3) * pgtable_level + 3)
>> #define PTRS_PER_PGD		(1 << (va_bits - PGDIR_SHIFT))
>> #define PUD_SHIFT		get_pud_shift_arm64()
>> +#define PUD_SIZE		(1UL << PUD_SHIFT)
>> +#define PUD_MASK		(~(PUD_SIZE - 1))
>> #define PTRS_PER_PTE		(1 << (PAGESHIFT() - 3))
>> #define PTRS_PER_PUD		PTRS_PER_PTE
>> #define PMD_SHIFT		((PAGESHIFT() - 3) * 2 + 3)
>> @@ -79,6 +81,10 @@ static unsigned long kimage_voffset;
>> #define PMD_TYPE_SECT		1
>> #define PMD_TYPE_TABLE		3
>>
>> +#define PUD_TYPE_MASK		3
>> +#define PUD_TYPE_SECT		1
>> +#define PUD_TYPE_TABLE		3
>> +
>> #define pgd_index(vaddr) 		(((vaddr) >> PGDIR_SHIFT) & (PTRS_PER_PGD - 1))
>> #define pgd_offset(pgdir, vaddr)	((pgd_t *)(pgdir) + pgd_index(vaddr))
>>
>> @@ -253,6 +259,13 @@ vaddr_to_paddr_arm64(unsigned long vaddr)
>> 		return NOT_PADDR;
>> 	}
>>
>> +	if ((pud_val(pudv) & PUD_TYPE_MASK) == PUD_TYPE_SECT) {
>> +		/* 1GB section */

May be we can write above comment like:
/* 1GB section for Page Table level = 4 and Page Size = 4KB*/

I think the other existing /* 1GB section */ comment for PMD_TYPE should be
/* 512MB section for Page Table level = 3 and Page Size = 64KB*/

>> +		paddr = (pud_val(pudv) & (PUD_MASK & PMD_SECTION_MASK))
>> +					+ (vaddr & (PUD_SIZE - 1));
>> +		return paddr;
>> +	}
>> +
>> 	pmda = pmd_offset(puda, &pudv, vaddr);
>> 	if (!readmem(PADDR, (unsigned long long)pmda, &pmdv, sizeof(pmdv))) {
>> 		ERRMSG("Can't read pmd\n");
>> --
>> 1.9.3
>>
>>
>> _______________________________________________
>> kexec mailing list
>> kexec@lists.infradead.org
>> http://lists.infradead.org/mailman/listinfo/kexec
> 
> 

-- 
Regards
Pratyush

_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec