Skip to contentSkip to conversation

x86/mm/KASLR: Change the granularity of randomization to PUD size in 5-level

PATCH Superseded Baoquan He 2 participants 13 messages Feb 28, 2019 – Mar 4, 2019

Subsystems (2)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Baoquan He

author
#0 lore raw

v2 cover x86/mm/KASLR: Change the granularity of randomization to PUD size in 5-level

This is v2 post. v1 can be found here:
http://lkml.kernel.org/r/[email protected]

Background:
***
Earlier, during a series of KASLR patch reviewing, Ingo got the current memory region KASLR only has granularity of randomization in PUD size in 4-level paging mode, and P4D size in 5-level paging mode, He suggested me to try to change both of them to be PMD size at granularity:

http://lkml.kernel.org/r/[email protected]

Later, I changed code to support PMD level of randomization for both 4-level and 5-level.

https://github.com/baoquan-he/linux/commits/mm-kaslr-2m-aligned

The test passed on my KVM guest with 1 GB RAM, but failed when I increased the RAM to 4 GB, and failed either on larger RAM.

After analyzing, it's because that 1 GB page mapping need be mapped at 1 GB aligned physical address for intel CPU. The 2 MB level of randomization will break it and cause error. Please check below table in intel IA32 manual.

Table 4-15. Format of an IA-32e Page-Directory-Pointer-Table Entry (PDPTE) that Maps a 1-GByte Page

So PMD level of randomization for mm KASLR is not doable.

However, during investigation and testing above code, it turns out that the current code is misleading to build identity mapping for the real mode trampoline in case KASLR enabled. From code, only a small area (which is smaller than 1 MB) need be identity mapped. Please check below patch which is from above mm-kaslr-2m-aligned patch series. it only builds up 2 MB identity maping for real mode trampoline, and test passed on machines with 32 GB RAM of 4-level and on KVM guest of 5-level.

https://github.com/baoquan-he/linux/commit/e120e67fbf9a5aa818d20084d8dea5b4a27ecf97

Result:
Make a patchset to:
1)change code to only build 1 GB of area for real mode trampoline,

    namely only copy one PUD entry where physical address 0 resides;

2)improve the randomization granularity of 5-level from P4D size to PUD size.

Changelog:
v1->v2:
Improve patch according to Kirill's suggestions:

    *)Add more information to code comment for better understanding;
    *)Improve code to save one low memory page in 4-level;

Baoquan He (2):
x86/mm/KASLR: Only build one PUD entry of area for real mode
trampoline
x86/mm/KASLR: Change the granularity of randomization to PUD size in

    5-level
 arch/x86/mm/kaslr.c | 92 ++++++++++++++++++++-------------------------
 1 file changed, 40 insertions(+), 52 deletions(-)
Signature · 1 line
2.17.2
author
#1 lore raw

PATCH v2 1/2 x86/mm/KASLR: Only build one PUD entry of area for real mode trampoline

The current code builds identity mapping for real mode treampoline by borrowing page tables from the direct mapping section if KASLR is enabled. It will copy present entries of the first PUD table in 4-level paging mode, or the first P4D table in 5-level paging mode.

However, there's only a very small area under low 1 MB reserved for real mode trampoline in reserve_real_mode(). Makes no sense to build up so large area of mapping for it. Since the randomization granularity in 4-level is 1 GB, and 512 GB in 5-level, only copying one PUD entry is enough.

Hence, only copy one PUD entry of area where physical address 0 resides. And this is preparation for later changing the randomization granularity of 5-level paging mode from 512 GB to 1 GB.

---
 arch/x86/mm/kaslr.c | 82 +++++++++++++++++++++------------------------
 1 file changed, 38 insertions(+), 44 deletions(-)
arch/x86/mm/kaslr.c +38 -44
index 754b5da91d43..131e08a10a68 100644
@@ -226,74 +226,68 @@ void __init kernel_randomize_memory(void)
 
 static void __meminit init_trampoline_pud(void)
 {
-	unsigned long paddr, paddr_next;
+	unsigned long paddr, vaddr;
 	pgd_t *pgd;
-	pud_t *pud_page, *pud_page_tramp;
-	int i;
+
+	pud_t *pud_page, *pud_page_tramp, *pud, *pud_tramp;
 
 	pud_page_tramp = alloc_low_page();
 
 	paddr = 0;
+	vaddr = (unsigned long)__va(paddr);
 	pgd = pgd_offset_k((unsigned long)__va(paddr));
-	pud_page = (pud_t *) pgd_page_vaddr(*pgd);
-
-	for (i = pud_index(paddr); i < PTRS_PER_PUD; i++, paddr = paddr_next) {
-		pud_t *pud, *pud_tramp;
-		unsigned long vaddr = (unsigned long)__va(paddr);
 
-		pud_tramp = pud_page_tramp + pud_index(paddr);
-		pud = pud_page + pud_index(vaddr);
-		paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
-
-		*pud_tramp = *pud;
-	}
+	if (pgtable_l5_enabled()) {
+		p4d_t *p4d_page, *p4d_page_tramp, *p4d, *p4d_tramp;
 
+		p4d_page_tramp = alloc_low_page();
 
-	set_pgd(&trampoline_pgd_entry,
-		__pgd(_KERNPG_TABLE | __pa(pud_page_tramp)));
-}
+		p4d_page = (p4d_t *) pgd_page_vaddr(*pgd);
+		p4d = p4d_page + p4d_index(vaddr);
 
-static void __meminit init_trampoline_p4d(void)
-{
-	unsigned long paddr, paddr_next;
-	pgd_t *pgd;
-	p4d_t *p4d_page, *p4d_page_tramp;
-	int i;
+		pud_page = (pud_t *) p4d_page_vaddr(*p4d);
+		pud = pud_page + pud_index(vaddr);
 
-	p4d_page_tramp = alloc_low_page();
+		p4d_tramp = p4d_page_tramp + p4d_index(paddr);
+		pud_tramp = pud_page_tramp + pud_index(paddr);
 
-	paddr = 0;
-	pgd = pgd_offset_k((unsigned long)__va(paddr));
-	p4d_page = (p4d_t *) pgd_page_vaddr(*pgd);
+		*pud_tramp = *pud;
 
-	for (i = p4d_index(paddr); i < PTRS_PER_P4D; i++, paddr = paddr_next) {
-		p4d_t *p4d, *p4d_tramp;
-		unsigned long vaddr = (unsigned long)__va(paddr);
+		set_p4d(p4d_tramp,
+			__p4d(_KERNPG_TABLE | __pa(pud_page_tramp)));
 
-		p4d_tramp = p4d_page_tramp + p4d_index(paddr);
-		p4d = p4d_page + p4d_index(vaddr);
-		paddr_next = (paddr & P4D_MASK) + P4D_SIZE;
+		set_pgd(&trampoline_pgd_entry,
+			__pgd(_KERNPG_TABLE | __pa(p4d_page_tramp)));
+	} else {
+		pud_page = (pud_t *) pgd_page_vaddr(*pgd);
+		pud = pud_page + pud_index(vaddr);
 
-		*p4d_tramp = *p4d;
+		pud_tramp = pud_page_tramp + pud_index(paddr);
+		*pud_tramp = *pud;
+		set_pgd(&trampoline_pgd_entry,
+			__pgd(_KERNPG_TABLE | __pa(pud_page_tramp)));
 	}
-
-	set_pgd(&trampoline_pgd_entry,
-		__pgd(_KERNPG_TABLE | __pa(p4d_page_tramp)));
 }
 
 /*
- * Create PGD aligned trampoline table to allow real mode initialization
- * of additional CPUs. Consume only 1 low memory page.
+ * Real mode trampoline only occupies a small area under low 1 MB
+ * (please check codes in reserve_real_mode() for details). For
+ * APs' booting up, we just borrow as few page tables as possible
+ * from the direct physical mapping to build 1:1 mapping to cover
+ * that area. In case KASLR disabled, the 1st PGD entry of the
+ * direct mapping is copied directly. If KASLR is enabled, only
+ * copy the 1st PUD entry where physical address 0 resides since
+ * the granularity of randomization is PUD size in 4-level, and
+ * P4D size in 5-level.
+ *
+ * This consumes one low memory page in 4-level case, and extra one
+ * in 5-level.
  */
 void __meminit init_trampoline(void)
 {
-
 	if (!kaslr_memory_enabled()) {
 		init_trampoline_default();
 		return;
 	}
 
-	if (pgtable_l5_enabled())
-		init_trampoline_p4d();
-	else
-		init_trampoline_pud();
+	init_trampoline_pud();

}

Signature · 1 line
2.17.2
author
#2 lore raw

PATCH v2 2/2 x86/mm/KASLR: Change the granularity of randomization to PUD size in 5-level

In reply to Baoquan He

The current randomization granularity of 5-level is 512 GB. Improve it to 1 GB. This can add more randomness to memory region KASLR in 5-level paging mode

---
 arch/x86/mm/kaslr.c | 14 ++++----------
 1 file changed, 4 insertions(+), 10 deletions(-)
arch/x86/mm/kaslr.c +4 -10
index 131e08a10a68..99ec75634613 100644
@@ -204,10 +204,7 @@ void __init kernel_randomize_memory(void)
 		 */
 		entropy = remain_entropy / (ARRAY_SIZE(kaslr_regions) - i);
 		prandom_bytes_state(&rand_state, &rand, sizeof(rand));
-		if (pgtable_l5_enabled())
-			entropy = (rand % (entropy + 1)) & P4D_MASK;
-		else
-			entropy = (rand % (entropy + 1)) & PUD_MASK;
+		entropy = (rand % (entropy + 1)) & PUD_MASK;
 		vaddr += entropy;
 		*kaslr_regions[i].base = vaddr;
 
@@ -216,10 +213,7 @@ void __init kernel_randomize_memory(void)
 		 * randomization alignment.
 		 */
 		vaddr += kaslr_regions[i].size_tb << TB_SHIFT;
-		if (pgtable_l5_enabled())
-			vaddr = round_up(vaddr + 1, P4D_SIZE);
-		else
-			vaddr = round_up(vaddr + 1, PUD_SIZE);
+		vaddr = round_up(vaddr + 1, PUD_SIZE);
 		remain_entropy -= entropy;
 	}
 }
@@ -276,8 +270,8 @@ static void __meminit init_trampoline_pud(void)
  * that area. In case KASLR disabled, the 1st PGD entry of the
  * direct mapping is copied directly. If KASLR is enabled, only
  * copy the 1st PUD entry where physical address 0 resides since
- * the granularity of randomization is PUD size in 4-level, and
- * P4D size in 5-level.
+ * the granularity of randomization is PUD size in both 4-level and
+ * 5-level.
  *
  * This consumes one low memory page in 4-level case, and extra one
  * in 5-level.
Signature · 1 line
2.17.2
#3 lore raw
In reply to Baoquan He
Quoted 2 lines · Baoquan He
 arch/x86/mm/kaslr.c | 92 ++++++++++++++++++++-------------------------
 1 file changed, 40 insertions(+), 52 deletions(-)

What tree is it against?

I cannot apply on current Linus' tree or tip/master.

Signature · 1 line
 Kirill A. Shutemov
author
#4 lore raw
Quoted 7 lines · Kirill A. Shutemov
Quoted 2 lines · Baoquan He
 arch/x86/mm/kaslr.c | 92 ++++++++++++++++++++-------------------------
 1 file changed, 40 insertions(+), 52 deletions(-)

What tree is it against?

I cannot apply on current Linus' tree or tip/master.

Oops, I put them on my on my previous kaslr fixing patches:
http://lkml.kernel.org/r/[email protected]
[PATCH v3 0/6] Several patches to fix code bugs, improve documents and clean up

There's a clean up there to open code get_padding(), that cause conflict in patch 0002. Let me rearrange and repost.

Thanks
Baoquan

#5 lore raw
Quoted 7 lines · Baoquan He
Quoted 6 lines · Kirill A. Shutemov
Quoted 2 lines · Baoquan He
 arch/x86/mm/kaslr.c | 92 ++++++++++++++++++++-------------------------
 1 file changed, 40 insertions(+), 52 deletions(-)

What tree is it against?

Quoted 8 lines
Quoted 1 line

I cannot apply on current Linus' tree or tip/master.

Oops, I put them on my on my previous kaslr fixing patches:
http://lkml.kernel.org/r/[email protected]
[PATCH v3 0/6] Several patches to fix code bugs, improve documents and clean up

There's a clean up there to open code get_padding(), that cause conflict in patch 0002. Let me rearrange and repost.

I cannot apply the first patch too. Hm?

Signature · 1 line
 Kirill A. Shutemov
author
#6 lore raw
Quoted 1 line · Kirill A. Shutemov

I cannot apply the first patch too. Hm?

It's weird. I use 'git cherry-pick' and it can be cleanly applied.

And git formatted patch can be applied too. Could you try below one? git am works for me with it, based on this commit on linus's tree. 7d762d69145a afs: Fix manually set volume location server list.

From dcbe8e4846102fdd051deb3c2946125c17b270fb Mon Sep 17 00:00:00 2001
From: Baoquan He <[email protected]>
Date: Thu, 21 Feb 2019 11:46:59 +0800
Subject: [PATCH v3] x86/mm/KASLR: Only build one PUD entry of area for real mode
trampoline

The current code builds identity mapping for real mode treampoline by borrowing page tables from the direct mapping section if KASLR is enabled. It will copy present entries of the first PUD table in 4-level paging mode, or the first P4D table in 5-level paging mode.

However, there's only a very small area under low 1 MB reserved for real mode trampoline in reserve_real_mode(). Makes no sense to build up so large area of mapping for it. Since the randomization granularity in 4-level is 1 GB, and 512 GB in 5-level, only copying one PUD entry is enough.

Hence, only copy one PUD entry of area where physical address 0 resides. And this is preparation for later changing the randomization granularity of 5-level paging mode from 512 GB to 1 GB.

---
 arch/x86/mm/kaslr.c | 82 +++++++++++++++++++++------------------------
 1 file changed, 38 insertions(+), 44 deletions(-)
arch/x86/mm/kaslr.c +38 -44
index 754b5da91d43..131e08a10a68 100644
@@ -226,74 +226,68 @@ void __init kernel_randomize_memory(void)
 
 static void __meminit init_trampoline_pud(void)
 {
-	unsigned long paddr, paddr_next;
+	unsigned long paddr, vaddr;
 	pgd_t *pgd;
-	pud_t *pud_page, *pud_page_tramp;
-	int i;
+
+	pud_t *pud_page, *pud_page_tramp, *pud, *pud_tramp;
 
 	pud_page_tramp = alloc_low_page();
 
 	paddr = 0;
+	vaddr = (unsigned long)__va(paddr);
 	pgd = pgd_offset_k((unsigned long)__va(paddr));
-	pud_page = (pud_t *) pgd_page_vaddr(*pgd);
-
-	for (i = pud_index(paddr); i < PTRS_PER_PUD; i++, paddr = paddr_next) {
-		pud_t *pud, *pud_tramp;
-		unsigned long vaddr = (unsigned long)__va(paddr);
 
-		pud_tramp = pud_page_tramp + pud_index(paddr);
-		pud = pud_page + pud_index(vaddr);
-		paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
-
-		*pud_tramp = *pud;
-	}
+	if (pgtable_l5_enabled()) {
+		p4d_t *p4d_page, *p4d_page_tramp, *p4d, *p4d_tramp;
+		p4d_page_tramp = alloc_low_page();
 
-	set_pgd(&trampoline_pgd_entry,
-		__pgd(_KERNPG_TABLE | __pa(pud_page_tramp)));
-}
+		p4d_page = (p4d_t *) pgd_page_vaddr(*pgd);
+		p4d = p4d_page + p4d_index(vaddr);
 
-static void __meminit init_trampoline_p4d(void)
-{
-	unsigned long paddr, paddr_next;
-	pgd_t *pgd;
-	p4d_t *p4d_page, *p4d_page_tramp;
-	int i;
+		pud_page = (pud_t *) p4d_page_vaddr(*p4d);
+		pud = pud_page + pud_index(vaddr);
 
-	p4d_page_tramp = alloc_low_page();
+		p4d_tramp = p4d_page_tramp + p4d_index(paddr);
+		pud_tramp = pud_page_tramp + pud_index(paddr);
 
-	paddr = 0;
-	pgd = pgd_offset_k((unsigned long)__va(paddr));
-	p4d_page = (p4d_t *) pgd_page_vaddr(*pgd);
+		*pud_tramp = *pud;
 
-	for (i = p4d_index(paddr); i < PTRS_PER_P4D; i++, paddr = paddr_next) {
-		p4d_t *p4d, *p4d_tramp;
-		unsigned long vaddr = (unsigned long)__va(paddr);
+		set_p4d(p4d_tramp,
+			__p4d(_KERNPG_TABLE | __pa(pud_page_tramp)));
 
-		p4d_tramp = p4d_page_tramp + p4d_index(paddr);
-		p4d = p4d_page + p4d_index(vaddr);
-		paddr_next = (paddr & P4D_MASK) + P4D_SIZE;
+		set_pgd(&trampoline_pgd_entry,
+			__pgd(_KERNPG_TABLE | __pa(p4d_page_tramp)));
+	} else {
+		pud_page = (pud_t *) pgd_page_vaddr(*pgd);
+		pud = pud_page + pud_index(vaddr);
 
-		*p4d_tramp = *p4d;
+		pud_tramp = pud_page_tramp + pud_index(paddr);
+		*pud_tramp = *pud;
+		set_pgd(&trampoline_pgd_entry,
+			__pgd(_KERNPG_TABLE | __pa(pud_page_tramp)));
 	}
-
-	set_pgd(&trampoline_pgd_entry,
-		__pgd(_KERNPG_TABLE | __pa(p4d_page_tramp)));
 }
 
 /*
- * Create PGD aligned trampoline table to allow real mode initialization
- * of additional CPUs. Consume only 1 low memory page.
+ * Real mode trampoline only occupies a small area under low 1 MB
+ * (please check codes in reserve_real_mode() for details). For
+ * APs' booting up, we just borrow as few page tables as possible
+ * from the direct physical mapping to build 1:1 mapping to cover
+ * that area. In case KASLR disabled, the 1st PGD entry of the
+ * direct mapping is copied directly. If KASLR is enabled, only
+ * copy the 1st PUD entry where physical address 0 resides since
+ * the granularity of randomization is PUD size in 4-level, and
+ * P4D size in 5-level.
+ *
+ * This consumes one low memory page in 4-level case, and extra one
+ * in 5-level.
  */
 void __meminit init_trampoline(void)
 {
-
 	if (!kaslr_memory_enabled()) {
 		init_trampoline_default();
 		return;
 	}
 
-	if (pgtable_l5_enabled())
-		init_trampoline_p4d();
-	else
-		init_trampoline_pud();
+	init_trampoline_pud();
 }
Signature · 1 line
2.17.2
#7 lore raw
Quoted 41 lines · Baoquan He
Quoted 1 line · Kirill A. Shutemov

I cannot apply the first patch too. Hm?

It's weird. I use 'git cherry-pick' and it can be cleanly applied.

And git formatted patch can be applied too. Could you try below one? git am works for me with it, based on this commit on linus's tree. 7d762d69145a afs: Fix manually set volume location server list.

From dcbe8e4846102fdd051deb3c2946125c17b270fb Mon Sep 17 00:00:00 2001
From: Baoquan He <[email protected]>
Date: Thu, 21 Feb 2019 11:46:59 +0800
Subject: [PATCH v3] x86/mm/KASLR: Only build one PUD entry of area for real mode
trampoline

The current code builds identity mapping for real mode treampoline by borrowing page tables from the direct mapping section if KASLR is enabled. It will copy present entries of the first PUD table in 4-level paging mode, or the first P4D table in 5-level paging mode.

However, there's only a very small area under low 1 MB reserved for real mode trampoline in reserve_real_mode(). Makes no sense to build up so large area of mapping for it. Since the randomization granularity in 4-level is 1 GB, and 512 GB in 5-level, only copying one PUD entry is enough.

Hence, only copy one PUD entry of area where physical address 0 resides. And this is preparation for later changing the randomization granularity of 5-level paging mode from 512 GB to 1 GB.

---
 arch/x86/mm/kaslr.c | 82 +++++++++++++++++++++------------------------
 1 file changed, 38 insertions(+), 44 deletions(-)
arch/x86/mm/kaslr.c +0 -0
index 754b5da91d43..131e08a10a68 100644

arch/x86/mm/kaslr.c void __init kernel_randomize_memory(void) · line 231

Quoted 9 lines
arch/x86/mm/kaslr.c +2 -3
@@ -226,74 +226,68 @@ void __init kernel_randomize_memory(void)
 
 static void __meminit init_trampoline_pud(void)
 {
-	unsigned long paddr, paddr_next;
+	unsigned long paddr, vaddr;
 	pgd_t *pgd;
-	pud_t *pud_page, *pud_page_tramp;
-	int i;
+

Unneeded empty line.

arch/x86/mm/kaslr.c void __init kernel_randomize_memory(void) · line 237

Quoted 6 lines
arch/x86/mm/kaslr.c +2 -0
+	pud_t *pud_page, *pud_page_tramp, *pud, *pud_tramp;
 
 	pud_page_tramp = alloc_low_page();
 
 	paddr = 0;
+	vaddr = (unsigned long)__va(paddr);

Maybe just

	vaddr = PAGE_OFFSET;

?

arch/x86/mm/kaslr.c void __init kernel_randomize_memory(void) · line 238

Quoted 1 line
arch/x86/mm/kaslr.c +0 -0
 	pgd = pgd_offset_k((unsigned long)__va(paddr));

We do have 'vaddr' already.

	pgd = pgd_offset_k(vaddr);
Quoted 24 lines
arch/x86/mm/kaslr.c +5 -16
-	pud_page = (pud_t *) pgd_page_vaddr(*pgd);
-
-	for (i = pud_index(paddr); i < PTRS_PER_PUD; i++, paddr = paddr_next) {
-		pud_t *pud, *pud_tramp;
-		unsigned long vaddr = (unsigned long)__va(paddr);
 
-		pud_tramp = pud_page_tramp + pud_index(paddr);
-		pud = pud_page + pud_index(vaddr);
-		paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
-
-		*pud_tramp = *pud;
-	}
+	if (pgtable_l5_enabled()) {
+		p4d_t *p4d_page, *p4d_page_tramp, *p4d, *p4d_tramp;
+		p4d_page_tramp = alloc_low_page();
 
-	set_pgd(&trampoline_pgd_entry,
-		__pgd(_KERNPG_TABLE | __pa(pud_page_tramp)));
-}
+		p4d_page = (p4d_t *) pgd_page_vaddr(*pgd);
+		p4d = p4d_page + p4d_index(vaddr);
 
-static void __meminit init_trampoline_p4d(void)
-{

arch/x86/mm/kaslr.c void __init kernel_randomize_memory(void) · line 251

Quoted 10 lines
arch/x86/mm/kaslr.c +4 -5
-	unsigned long paddr, paddr_next;
-	pgd_t *pgd;
-	p4d_t *p4d_page, *p4d_page_tramp;
-	int i;
+		pud_page = (pud_t *) p4d_page_vaddr(*p4d);
+		pud = pud_page + pud_index(vaddr);
 
-	p4d_page_tramp = alloc_low_page();
+		p4d_tramp = p4d_page_tramp + p4d_index(paddr);
+		pud_tramp = pud_page_tramp + pud_index(paddr);

p?d_index() has to be called on the virtual address. It shouldn't break anything, but it's wrong from conceptual point of view.

I don't think need paddr at all in the function.

BTW, any reason we cannot use p?d_offset() instead of playing with p?d_index() directly?

Signature · 1 line
 Kirill A. Shutemov
author
#8 lore raw
Quoted 6 lines · Kirill A. Shutemov
Quoted 6 lines
arch/x86/mm/kaslr.c +0 -0
index 754b5da91d43..131e08a10a68 100644
@@ -226,74 +226,68 @@ void __init kernel_randomize_memory(void)
 
Quoted 10 lines

arch/x86/mm/kaslr.c void __init kernel_randomize_memory(void) · line 231

Quoted 8 lines
arch/x86/mm/kaslr.c +2 -3
 static void __meminit init_trampoline_pud(void)
 {
-	unsigned long paddr, paddr_next;
+	unsigned long paddr, vaddr;
 	pgd_t *pgd;
-	pud_t *pud_page, *pud_page_tramp;
-	int i;
+

Unneeded empty line.

Right, I will remove it, and move it to the top since it's the longest line.

Quoted 5 lines
Quoted 5 lines
arch/x86/mm/kaslr.c +1 -0
+	pud_t *pud_page, *pud_page_tramp, *pud, *pud_tramp;
 
 	pud_page_tramp = alloc_low_page();
 
Quoted 8 lines

arch/x86/mm/kaslr.c void __init kernel_randomize_memory(void) · line 237

Quoted 2 lines
arch/x86/mm/kaslr.c +1 -0
 	paddr = 0;
+	vaddr = (unsigned long)__va(paddr);

Maybe just

	vaddr = PAGE_OFFSET;

?

An obvious 0 to PAGE_OFFSET converting can explain the page table borrowing from the direct mapping in some extent. While both is fine to me if you prefer PAGE_OFFSET.

Quoted 4 lines

arch/x86/mm/kaslr.c void __init kernel_randomize_memory(void) · line 238

Quoted 2 lines
arch/x86/mm/kaslr.c +0 -0
 	pgd = pgd_offset_k((unsigned long)__va(paddr));

We do have 'vaddr' already.

Yeah, will replace it.

Quoted 34 lines
	pgd = pgd_offset_k(vaddr);
Quoted 31 lines
arch/x86/mm/kaslr.c +7 -20
-	pud_page = (pud_t *) pgd_page_vaddr(*pgd);
-
-	for (i = pud_index(paddr); i < PTRS_PER_PUD; i++, paddr = paddr_next) {
-		pud_t *pud, *pud_tramp;
-		unsigned long vaddr = (unsigned long)__va(paddr);
 
-		pud_tramp = pud_page_tramp + pud_index(paddr);
-		pud = pud_page + pud_index(vaddr);
-		paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
-
-		*pud_tramp = *pud;
-	}
+	if (pgtable_l5_enabled()) {
+		p4d_t *p4d_page, *p4d_page_tramp, *p4d, *p4d_tramp;
+		p4d_page_tramp = alloc_low_page();
 
-	set_pgd(&trampoline_pgd_entry,
-		__pgd(_KERNPG_TABLE | __pa(pud_page_tramp)));
-}
+		p4d_page = (p4d_t *) pgd_page_vaddr(*pgd);
+		p4d = p4d_page + p4d_index(vaddr);
 
-static void __meminit init_trampoline_p4d(void)
-{
-	unsigned long paddr, paddr_next;
-	pgd_t *pgd;
-	p4d_t *p4d_page, *p4d_page_tramp;
-	int i;
+		pud_page = (pud_t *) p4d_page_vaddr(*p4d);
+		pud = pud_page + pud_index(vaddr);
 
Quoted 7 lines

arch/x86/mm/kaslr.c void __init kernel_randomize_memory(void) · line 251

Quoted 3 lines
arch/x86/mm/kaslr.c +2 -1
-	p4d_page_tramp = alloc_low_page();
+		p4d_tramp = p4d_page_tramp + p4d_index(paddr);
+		pud_tramp = pud_page_tramp + pud_index(paddr);

p?d_index() has to be called on the virtual address. It shouldn't break anything, but it's wrong from conceptual point of view.
I don't think need paddr at all in the function.

Hmm, here the tramp table is for identity mapping real mode. Its paddr equals to its vaddr. Using paddr here can tell it's the identity mapping which is handling.

Quoted 3 lines

BTW, any reason we cannot use p?d_offset() instead of playing with p?d_index() directly?

Sure, p?d_offset() looks better. Just took it from the old code. I will use it instead.

Thanks for careful reviewing.

author
#9 lore raw
In reply to Kirill A. Shutemov

Hi Kirill,

I updated patch per your comments. While I kept the 'paddr' variable and the '0' initilization stuffs. My thought is there are two kinds of mapping in the handling, so keeping these names from old codes can make it more understandable. What do you think?

Kind Physical address Virtual address
----------------------------------------------------------------
Direct mapping paddr vaddr = paddr+PAGE_OFFSET
1:1 mapping paddr paddr

Thanks
Baoquan

From 52c6c80b2e3ffe153902800793c0e732c3c6bf1d Mon Sep 17 00:00:00 2001
From: Baoquan He <[email protected]>
Date: Thu, 21 Feb 2019 11:46:59 +0800
Subject: [PATCH] x86/mm/KASLR: Only build one PUD entry of area for real mode
trampoline

The current code builds identity mapping for real mode treampoline by borrowing page tables from the direct mapping section if KASLR is enabled. It will copy present entries of the first PUD table in 4-level paging mode, or the first P4D table in 5-level paging mode.

However, there's only a very small area under low 1 MB reserved for real mode trampoline in reserve_real_mode(). Makes no sense to build up so large area of mapping for it. Since the randomization granularity in 4-level is 1 GB, and 512 GB in 5-level, only copying one PUD entry is enough.

Hence, only copy one PUD entry of area where physical address 0 resides. And this is preparation for later changing the randomization granularity of 5-level paging mode from 512 GB to 1 GB.

---
 arch/x86/mm/kaslr.c | 79 ++++++++++++++++++---------------------------
 1 file changed, 32 insertions(+), 47 deletions(-)
arch/x86/mm/kaslr.c +32 -47
index 3f452ffed7e9..c6fab7a77439 100644
@@ -147,74 +147,59 @@ void __init kernel_randomize_memory(void)
 
 static void __meminit init_trampoline_pud(void)
 {
-	unsigned long paddr, paddr_next;
+	pud_t *pud_page_tramp, *pud, *pud_tramp;
+	p4d_t *p4d_page_tramp, *p4d, *p4d_tramp;
+	unsigned long paddr, vaddr;
 	pgd_t *pgd;
-	pud_t *pud_page, *pud_page_tramp;
-	int i;
 
 	pud_page_tramp = alloc_low_page();
 
 	paddr = 0;
-	pgd = pgd_offset_k((unsigned long)__va(paddr));
-	pud_page = (pud_t *) pgd_page_vaddr(*pgd);
+	vaddr = (unsigned long)__va(paddr);
+	pgd = pgd_offset_k(vaddr);
 
-	for (i = pud_index(paddr); i < PTRS_PER_PUD; i++, paddr = paddr_next) {
-		pud_t *pud, *pud_tramp;
-		unsigned long vaddr = (unsigned long)__va(paddr);
+	p4d = p4d_offset(pgd, vaddr);
+	pud = pud_offset(p4d, vaddr);
 
-		pud_tramp = pud_page_tramp + pud_index(paddr);
-		pud = pud_page + pud_index(vaddr);
-		paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
+	pud_tramp = pud_page_tramp + pud_index(paddr);
+	*pud_tramp = *pud;
 
-		*pud_tramp = *pud;
-	}
-
-	set_pgd(&trampoline_pgd_entry,
-		__pgd(_KERNPG_TABLE | __pa(pud_page_tramp)));
-}
-
-static void __meminit init_trampoline_p4d(void)
-{
-	unsigned long paddr, paddr_next;
-	pgd_t *pgd;
-	p4d_t *p4d_page, *p4d_page_tramp;
-	int i;
-
-	p4d_page_tramp = alloc_low_page();
-
-	paddr = 0;
-	pgd = pgd_offset_k((unsigned long)__va(paddr));
-	p4d_page = (p4d_t *) pgd_page_vaddr(*pgd);
-
-	for (i = p4d_index(paddr); i < PTRS_PER_P4D; i++, paddr = paddr_next) {
-		p4d_t *p4d, *p4d_tramp;
-		unsigned long vaddr = (unsigned long)__va(paddr);
+	if (pgtable_l5_enabled()) {
+		p4d_page_tramp = alloc_low_page();
 
 		p4d_tramp = p4d_page_tramp + p4d_index(paddr);
-		p4d = p4d_page + p4d_index(vaddr);
-		paddr_next = (paddr & P4D_MASK) + P4D_SIZE;
 
-		*p4d_tramp = *p4d;
-	}
+		set_p4d(p4d_tramp,
+			__p4d(_KERNPG_TABLE | __pa(pud_page_tramp)));
 
-	set_pgd(&trampoline_pgd_entry,
-		__pgd(_KERNPG_TABLE | __pa(p4d_page_tramp)));
+		set_pgd(&trampoline_pgd_entry,
+			__pgd(_KERNPG_TABLE | __pa(p4d_page_tramp)));
+	} else {
+		set_pgd(&trampoline_pgd_entry,
+			__pgd(_KERNPG_TABLE | __pa(pud_page_tramp)));
+	}
 }
 
 /*
- * Create PGD aligned trampoline table to allow real mode initialization
- * of additional CPUs. Consume only 1 low memory page.
+ * Real mode trampoline only occupies a small area under low 1 MB
+ * (please check codes in reserve_real_mode() for details). For
+ * APs' booting up, we just borrow as few page tables as possible
+ * from the direct physical mapping to build 1:1 mapping to cover
+ * that area. In case KASLR disabled, the 1st PGD entry of the
+ * direct mapping is copied directly. If KASLR is enabled, only
+ * copy the 1st PUD entry where physical address 0 resides since
+ * the granularity of randomization is PUD size in 4-level, and
+ * P4D size in 5-level.
+ *
+ * This consumes one low memory page in 4-level case, and extra one
+ * in 5-level.
  */
 void __meminit init_trampoline(void)
 {
-
 	if (!kaslr_memory_enabled()) {
 		init_trampoline_default();
 		return;
 	}
 
-	if (pgtable_l5_enabled())
-		init_trampoline_p4d();
-	else
-		init_trampoline_pud();
+	init_trampoline_pud();
 }
Signature · 1 line
2.17.2
#10 lore raw
Quoted 11 lines · Baoquan He

Hi Kirill,

I updated patch per your comments. While I kept the 'paddr' variable and the '0' initilization stuffs. My thought is there are two kinds of mapping in the handling, so keeping these names from old codes can make it more understandable. What do you think?

Kind Physical address Virtual address
----------------------------------------------------------------
Direct mapping paddr vaddr = paddr+PAGE_OFFSET
1:1 mapping paddr paddr

I agree with your logic and patch looks good to me. But maybe you should exmplain this in the comments too?

Anyway:

Signature · 1 line
 Kirill A. Shutemov
author
#11 lore raw
Quoted 8 lines · Kirill A. Shutemov
Quoted 7 lines · Baoquan He

Hi Kirill,

I updated patch per your comments. While I kept the 'paddr' variable and the '0' initilization stuffs. My thought is there are two kinds of mapping in the handling, so keeping these names from old codes can make it more understandable. What do you think?

Quoted 7 lines
Quoted 4 lines

Kind Physical address Virtual address
----------------------------------------------------------------
Direct mapping paddr vaddr = paddr+PAGE_OFFSET
1:1 mapping paddr paddr

I agree with your logic and patch looks good to me. But maybe you should exmplain this in the comments too?

OK, will add this into code comment, and sent a v3 patchset to wrap all these changes according to your suggestions. Thanks a lot.

Thanks
Baoquan

author
#12 lore raw

v3 cover x86/mm/KASLR: Change the granularity of randomization to PUD size in 5-level

In reply to Kirill A. Shutemov

The current randomization granularity of 5-level is 512 GB. Improve it to 1 GB. This can add more randomness to memory region KASLR in 5-level paging mode

---
v2->v3:
The v2 was based on another patchset. Rebase it on the latest linus's tree.

 arch/x86/mm/kaslr.c | 14 ++++----------
 1 file changed, 4 insertions(+), 10 deletions(-)
arch/x86/mm/kaslr.c +4 -10
index 4e7f07b466c5..4ef9a27a1361 100644
@@ -125,10 +125,7 @@ void __init kernel_randomize_memory(void)
 		 */
 		entropy = remain_entropy / (ARRAY_SIZE(kaslr_regions) - i);
 		prandom_bytes_state(&rand_state, &rand, sizeof(rand));
-		if (pgtable_l5_enabled())
-			entropy = (rand % (entropy + 1)) & P4D_MASK;
-		else
-			entropy = (rand % (entropy + 1)) & PUD_MASK;
+		entropy = (rand % (entropy + 1)) & PUD_MASK;
 		vaddr += entropy;
 		*kaslr_regions[i].base = vaddr;
 
@@ -137,10 +134,7 @@ void __init kernel_randomize_memory(void)
 		 * randomization alignment.
 		 */
 		vaddr += get_padding(&kaslr_regions[i]);
-		if (pgtable_l5_enabled())
-			vaddr = round_up(vaddr + 1, P4D_SIZE);
-		else
-			vaddr = round_up(vaddr + 1, PUD_SIZE);
+		vaddr = round_up(vaddr + 1, PUD_SIZE);
 		remain_entropy -= entropy;
 	}
 }
@@ -197,8 +191,8 @@ static void __meminit init_trampoline_pud(void)
  * that area. In case KASLR disabled, the 1st PGD entry of the
  * direct mapping is copied directly. If KASLR is enabled, only
  * copy the 1st PUD entry where physical address 0 resides since
- * the granularity of randomization is PUD size in 4-level, and
- * P4D size in 5-level.
+ * the granularity of randomization is PUD size in both 4-level and
+ * 5-level.
  *
  * This consumes one low memory page in 4-level case, and extra one
  * in 5-level.
Signature · 1 line
2.17.2