* [PATCH] makedumpfile: arm64: Fix page table walk of 1GB section
@ 2017-08-05 22:41 Bradley Bolen
2017-08-05 22:41 ` Bradley Bolen
0 siblings, 1 reply; 4+ messages in thread
From: Bradley Bolen @ 2017-08-05 22:41 UTC (permalink / raw)
To: kexec; +Cc: panand, 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
^ permalink raw reply [flat|nested] 4+ messages in thread
* [PATCH] makedumpfile: arm64: Fix page table walk of 1GB section
2017-08-05 22:41 [PATCH] makedumpfile: arm64: Fix page table walk of 1GB section Bradley Bolen
@ 2017-08-05 22:41 ` Bradley Bolen
2017-08-14 8:37 ` Atsushi Kumagai
0 siblings, 1 reply; 4+ messages in thread
From: Bradley Bolen @ 2017-08-05 22:41 UTC (permalink / raw)
To: kexec; +Cc: panand, 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
^ permalink raw reply related [flat|nested] 4+ messages in thread
* RE: [PATCH] makedumpfile: arm64: Fix page table walk of 1GB section
2017-08-05 22:41 ` Bradley Bolen
@ 2017-08-14 8:37 ` Atsushi Kumagai
2017-08-14 10:35 ` Pratyush Anand
0 siblings, 1 reply; 4+ messages in thread
From: Atsushi Kumagai @ 2017-08-14 8:37 UTC (permalink / raw)
To: Bradley Bolen; +Cc: panand, kexec
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
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [PATCH] makedumpfile: arm64: Fix page table walk of 1GB section
2017-08-14 8:37 ` Atsushi Kumagai
@ 2017-08-14 10:35 ` Pratyush Anand
0 siblings, 0 replies; 4+ messages in thread
From: Pratyush Anand @ 2017-08-14 10:35 UTC (permalink / raw)
To: Atsushi Kumagai, Bradley Bolen; +Cc: kexec
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
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2017-08-14 10:36 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2017-08-05 22:41 [PATCH] makedumpfile: arm64: Fix page table walk of 1GB section Bradley Bolen
2017-08-05 22:41 ` Bradley Bolen
2017-08-14 8:37 ` Atsushi Kumagai
2017-08-14 10:35 ` Pratyush Anand
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.