Hello Ritesh,
Thanks for the review.
On 02/03/25 12:05, Ritesh Harjani (IBM) wrote:
Sourabh Jain <sourabhj...@linux.ibm.com> writes:
The fadump kernel boots with limited memory solely to collect the kernel
core dump. Having gigantic hugepages in the fadump kernel is of no use.
Sure got it.
Many times, the fadump kernel encounters OOM (Out of Memory) issues if
gigantic hugepages are allocated.
To address this, disable gigantic hugepages if fadump is active by
returning early from arch_hugetlb_valid_size() using
hugepages_supported(). When fadump is active, the global variable
hugetlb_disabled is set to true, which is later used by the
PowerPC-specific hugepages_supported() function to determine hugepage
support.
Returning early from arch_hugetlb_vali_size() not only disables
gigantic hugepages but also avoids unnecessary hstate initialization for
every hugepage size supported by the platform.
kernel logs related to hugepages with this patch included:
kernel argument passed: hugepagesz=1G hugepages=1
First kernel: gigantic hugepage got allocated
==============================================
dmesg | grep -i "hugetlb"
-------------------------
HugeTLB: registered 1.00 GiB page size, pre-allocated 1 pages
HugeTLB: 0 KiB vmemmap can be freed for a 1.00 GiB page
HugeTLB: registered 2.00 MiB page size, pre-allocated 0 pages
HugeTLB: 0 KiB vmemmap can be freed for a 2.00 MiB page
$ cat /proc/meminfo | grep -i "hugetlb"
-------------------------------------
Hugetlb: 1048576 kB
Was this tested with patch [1] in your local tree?
[1]:
https://web.git.kernel.org/pub/scm/linux/kernel/git/powerpc/linux.git/commit/?id=d629d7a8efc33
IIUC, this patch [1] disables the boot time allocation of hugepages.
Isn't it also disabling the boot time allocation for gigantic huge pages
passed by the cmdline params like hugepagesz=1G and hugepages=2 ?
Yes, I had the patch [1] in my tree.
My understanding is that gigantic pages are allocated before normal huge
pages.
In hugepages_setup() in hugetlb.c, we have:
if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
hugetlb_hstate_alloc_pages(parsed_hstate);
I believe the above code allocates memory for gigantic pages, and
hugetlb_init() is
called later because it is a subsys_initcall.
So, by the time the kernel reaches hugetlb_init(), the gigantic pages
are already
allocated. Isn't that right?
Please let me know your opinion.
Thanks,
Sourabh Jain
HugeTLB: registered 1.00 GiB page size, pre-allocated 1 pages
This print comes from report_hugepages(). The only place from where
report_hugepages() gets called is hugetlb_init(). hugetlb_init() is what
is responsible for hugepages & gigantic hugepage allocations of the
passed kernel cmdline params.
But hugetlb_init() already checks for hugepages_supported() in the very
beginning. So I am not sure whether we need this extra patch to disable
gigantic hugepages allocation by the kernel cmdline params like
hugepagesz=1G and hugepages=2 type of options.
Hence I was wondering if you had this patch [1] in your tree when you were
testing this?
But I may be missing something. Could you please help clarify on whether
we really need this patch to disable gigantic hugetlb page allocations?
Fadump kernel: gigantic hugepage not allocated
===============================================
dmesg | grep -i "hugetlb"
-------------------------
[ 0.000000] HugeTLB: unsupported hugepagesz=1G
[ 0.000000] HugeTLB: hugepages=1 does not follow a valid hugepagesz, ignoring
[ 0.706375] HugeTLB support is disabled!
[ 0.773530] hugetlbfs: disabling because there are no supported hugepage
sizes
$ cat /proc/meminfo | grep -i "hugetlb"
----------------------------------
<Nothing>
Cc: Hari Bathini <hbath...@linux.ibm.com>
Cc: Madhavan Srinivasan <ma...@linux.ibm.com>
Cc: Mahesh Salgaonkar <mah...@linux.ibm.com>
Cc: Michael Ellerman <m...@ellerman.id.au>
Cc: Ritesh Harjani (IBM)" <ritesh.l...@gmail.com>
I guess the extra " in the above was not adding me in the cc list.
Hence I missed to see this patch early.
Thanks for point it out. I will fix it.
-ritesh
Reviewed-by: Christophe Leroy <christophe.le...@csgroup.eu>
Signed-off-by: Sourabh Jain <sourabhj...@linux.ibm.com>
---
Changelog:
v1:
https://lore.kernel.org/all/20250121150419.1342794-1-sourabhj...@linux.ibm.com/
v2:
https://lore.kernel.org/all/20250124103220.111303-1-sourabhj...@linux.ibm.com/
- disable gigantic hugepage in arch code, arch_hugetlb_valid_size()
v3:
https://lore.kernel.org/all/20250125104928.88881-1-sourabhj...@linux.ibm.com/
- Do not modify the initialization of the shift variable
v4:
- Update commit message to include how hugepages_supported() detects
hugepages support when fadump is active
- Add Reviewed-by tag
- NO functional change
---
arch/powerpc/mm/hugetlbpage.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/arch/powerpc/mm/hugetlbpage.c b/arch/powerpc/mm/hugetlbpage.c
index 6b043180220a..88cfd182db4e 100644
--- a/arch/powerpc/mm/hugetlbpage.c
+++ b/arch/powerpc/mm/hugetlbpage.c
@@ -138,6 +138,9 @@ bool __init arch_hugetlb_valid_size(unsigned long size)
int shift = __ffs(size);
int mmu_psize;
+ if (!hugepages_supported())
+ return false;
+
/* Check that it is a page size supported by the hardware and
* that it fits within pagetable and slice limits. */
if (size <= PAGE_SIZE || !is_power_of_2(size))
--
2.48.1
