[PATCH v2 1/2] x86/mm/ident_map: Add PUD level 1GB page support

[Xunlei Pang]

The current kernel_ident_mapping_init() creates the identity
mapping using 2MB page(PMD level), this patch adds the 1GB
page(PUD level) support.

This is useful on large machines to save some reserved memory
(as paging structures) in the kdump case when kexec setups up
identity mappings before booting into the new kernel.

We will utilize this new support in the following patch.

Signed-off-by: Xunlei Pang <***@redhat.com>
---
v1->v2:
- Rename info.use_pud_page to info.direct_gbpages
- Align PUD_MASK before set_pud()

arch/x86/boot/compressed/pagetable.c | 2 +-
arch/x86/include/asm/init.h | 3 ++-
arch/x86/kernel/machine_kexec_64.c | 2 +-
arch/x86/mm/ident_map.c | 14 +++++++++++++-
arch/x86/power/hibernate_64.c | 2 +-
5 files changed, 18 insertions(+), 5 deletions(-)

diff --git a/arch/x86/boot/compressed/pagetable.c b/arch/x86/boot/compressed/pagetable.c
index 56589d0..1d78f17 100644
--- a/arch/x86/boot/compressed/pagetable.c
+++ b/arch/x86/boot/compressed/pagetable.c
@@ -70,7 +70,7 @@ static void *alloc_pgt_page(void *context)
* Due to relocation, pointers must be assigned at run time not build time.
*/
static struct x86_mapping_info mapping_info = {
- .pmd_flag = __PAGE_KERNEL_LARGE_EXEC,
+ .page_flag = __PAGE_KERNEL_LARGE_EXEC,
};

/* Locates and clears a region for a new top level page table. */
diff --git a/arch/x86/include/asm/init.h b/arch/x86/include/asm/init.h
index 737da62..d6ead7b 100644
--- a/arch/x86/include/asm/init.h
+++ b/arch/x86/include/asm/init.h
@@ -4,8 +4,9 @@
struct x86_mapping_info {
void *(*alloc_pgt_page)(void *); /* allocate buf for page table */
void *context; /* context for alloc_pgt_page */
- unsigned long pmd_flag; /* page flag for PMD entry */
+ unsigned long page_flag; /* page flag for PMD or PUD entry */
unsigned long offset; /* ident mapping offset */
+ bool direct_gbpages; /* PUD level 1GB page support */
};

int kernel_ident_mapping_init(struct x86_mapping_info *info, pgd_t *pgd_page,
diff --git a/arch/x86/kernel/machine_kexec_64.c b/arch/x86/kernel/machine_kexec_64.c
index 085c3b3..1d4f2b0 100644
--- a/arch/x86/kernel/machine_kexec_64.c
+++ b/arch/x86/kernel/machine_kexec_64.c
@@ -113,7 +113,7 @@ static int init_pgtable(struct kimage *image, unsigned long start_pgtable)
struct x86_mapping_info info = {
.alloc_pgt_page = alloc_pgt_page,
.context = image,
- .pmd_flag = __PAGE_KERNEL_LARGE_EXEC,
+ .page_flag = __PAGE_KERNEL_LARGE_EXEC,
};
unsigned long mstart, mend;
pgd_t *level4p;
diff --git a/arch/x86/mm/ident_map.c b/arch/x86/mm/ident_map.c
index 04210a2..adab159 100644
--- a/arch/x86/mm/ident_map.c
+++ b/arch/x86/mm/ident_map.c
@@ -13,7 +13,7 @@ static void ident_pmd_init(struct x86_mapping_info *info, pmd_t *pmd_page,
if (pmd_present(*pmd))
continue;

- set_pmd(pmd, __pmd((addr - info->offset) | info->pmd_flag));
+ set_pmd(pmd, __pmd((addr - info->offset) | info->page_flag));
}
}

@@ -30,6 +30,18 @@ static int ident_pud_init(struct x86_mapping_info *info, pud_t *pud_page,
if (next > end)
next = end;

+ if (info->direct_gbpages) {
+ pud_t pudval;
+
+ if (pud_present(*pud))
+ continue;
+
+ addr &= PUD_MASK;
+ pudval = __pud((addr - info->offset) | info->page_flag);
+ set_pud(pud, pudval);
+ continue;
+ }
+
if (pud_present(*pud)) {
pmd = pmd_offset(pud, 0);
ident_pmd_init(info, pmd, addr, next);
diff --git a/arch/x86/power/hibernate_64.c b/arch/x86/power/hibernate_64.c
index 6a61194..a6e21fe 100644
--- a/arch/x86/power/hibernate_64.c
+++ b/arch/x86/power/hibernate_64.c
@@ -104,7 +104,7 @@ static int set_up_temporary_mappings(void)
{
struct x86_mapping_info info = {
.alloc_pgt_page = alloc_pgt_page,
- .pmd_flag = __PAGE_KERNEL_LARGE_EXEC,
+ .page_flag = __PAGE_KERNEL_LARGE_EXEC,
.offset = __PAGE_OFFSET,
};
unsigned long mstart, mend;

[Xunlei Pang]

Kexec setups all identity mappings before booting into the new
kernel, and this will cause extra memory consumption for paging
structures which is quite considerable on modern machines with
huge memory.

E.g. On one 32TB machine, in kdump case, it could waste around
128MB (around 4MB/TB) from the reserved memory after kexec set
all the identity mappings using the current 2MB page, plus the
loaded kdump kernel, initramfs, etc, it caused kexec syscall
-NOMEM failure. As a result, we had to enlarge reserved memory
via "crashkernel=X".

This causes some trouble for distributions that use policies
to evaluate the proper "crashkernel=X" value for users.

Given that on machines with large number of memory, 1GB feature
is very likely available, and that kernel_ident_mapping_init()
supports PUD level 1GB page, to solve this problem, we use 1GB
size page to create the identity mapping pgtable for kdump if
1GB feature is available.

Signed-off-by: Xunlei Pang <***@redhat.com>
---
arch/x86/kernel/machine_kexec_64.c | 5 +++++
1 file changed, 5 insertions(+)

diff --git a/arch/x86/kernel/machine_kexec_64.c b/arch/x86/kernel/machine_kexec_64.c
index 1d4f2b0..613df62 100644
--- a/arch/x86/kernel/machine_kexec_64.c
+++ b/arch/x86/kernel/machine_kexec_64.c
@@ -122,6 +122,11 @@ static int init_pgtable(struct kimage *image, unsigned long start_pgtable)

level4p = (pgd_t *)__va(start_pgtable);
clear_page(level4p);
+
+ /* Use PUD level page if available, to save crash memory for kdump */
+ if (direct_gbpages)
+ info.direct_gbpages = true;
+
for (i = 0; i < nr_pfn_mapped; i++) {
mstart = pfn_mapped[i].start << PAGE_SHIFT;
mend = pfn_mapped[i].end << PAGE_SHIFT;

[Ingo Molnar]

Post by Xunlei Pang
The current kernel_ident_mapping_init() creates the identity
mapping using 2MB page(PMD level), this patch adds the 1GB
page(PUD level) support.
This is useful on large machines to save some reserved memory
(as paging structures) in the kdump case when kexec setups up
identity mappings before booting into the new kernel.
We will utilize this new support in the following patch.

Well, the primary advantage would be better TLB coverage/performance, because we'd
utilize 1GB TLBs instead of 2MB ones, right?

Any kexec fallout is secondary.

And I'd like to hear more about the primary advantage: what are the effects of
this change on a typical test system you have access to:

- For example what percentage of the identity mapping was 4K mapped (if any) and
2MB mapped - and how did this change due to the patch - how many 2MB mappings
remained and how many 1GB mappings were added?

- Is there anything else we could do to improve the in-RAM layout of kernel data
structures. For example IIRC the CPU breaks up all TLBs under 2MB physical into
4K TLBs. Is this the current limit and could we just reserve all that space and
not use it for anything important? 2MB of RAM wasted is a very small amount of
space, compared to the potential performance advantages.

Post by Xunlei Pang
void *(*alloc_pgt_page)(void *); /* allocate buf for page table */
void *context; /* context for alloc_pgt_page */
- unsigned long pmd_flag; /* page flag for PMD entry */
+ unsigned long page_flag; /* page flag for PMD or PUD entry */
unsigned long offset; /* ident mapping offset */
+ bool direct_gbpages; /* PUD level 1GB page support */

Doesn't follow the existing alignment.

Thanks,

Ingo

[Xunlei Pang]

Post by Ingo Molnar

Post by Xunlei Pang
The current kernel_ident_mapping_init() creates the identity
mapping using 2MB page(PMD level), this patch adds the 1GB
page(PUD level) support.
This is useful on large machines to save some reserved memory
(as paging structures) in the kdump case when kexec setups up
identity mappings before booting into the new kernel.
We will utilize this new support in the following patch.

Well, the primary advantage would be better TLB coverage/performance, because we'd
utilize 1GB TLBs instead of 2MB ones, right?
Any kexec fallout is secondary.
And I'd like to hear more about the primary advantage: what are the effects of
- For example what percentage of the identity mapping was 4K mapped (if any) and
2MB mapped - and how did this change due to the patch - how many 2MB mappings
remained and how many 1GB mappings were added?
- Is there anything else we could do to improve the in-RAM layout of kernel data
structures. For example IIRC the CPU breaks up all TLBs under 2MB physical into
4K TLBs. Is this the current limit and could we just reserve all that space and
not use it for anything important? 2MB of RAM wasted is a very small amount of
space, compared to the potential performance advantages.

Currently kernel_ident_mapping_init() only setups PMD large pages, seems there are no 4KB
mapped or others. Do you mean init_memory_mapping()->kernel_physical_mapping_init()?

The new struct x86_mapping_info::direct_gbpages added has the default value false, it has no
effect on the existing code, after patch 2, there will be 1GB pages(1GB TLBs) for kexec ident
mapping if direct_gbpages is true.

Post by Ingo Molnar

Post by Xunlei Pang
void *(*alloc_pgt_page)(void *); /* allocate buf for page table */
void *context; /* context for alloc_pgt_page */
- unsigned long pmd_flag; /* page flag for PMD entry */
+ unsigned long page_flag; /* page flag for PMD or PUD entry */
unsigned long offset; /* ident mapping offset */
+ bool direct_gbpages; /* PUD level 1GB page support */

Doesn't follow the existing alignment.

Ah, yes, missed the space. Thanks!

Regards,
Xunlei