[PATCH v3 0/3] arm64/mm: migrate swapper_pg_dir

[PATCH v3 0/3] arm64/mm: migrate swapper_pg_dir

[Jun Yao]

Version 3 changes:
	* Fix memory leak problem with CONFIG_ARM64_SW_TTBR0_PAN
	* add comment explaining why the flush is needed and also why
	  __pa_swapper_pg_dir gets placed in the .mmuoff.data.read
	  section.

[v2] https://www.spinics.net/lists/arm-kernel/msg657549.html
[v1] https://www.spinics.net/lists/kernel/msg2819351.html

Jun Yao (3):
  arm64/mm: pass swapper_pg_dir as an argument to __enable_mmu()
  arm64/mm: introduce variable to save new swapper_pg_dir address
  arm64/mm: migrate swapper_pg_dir and tramp_pg_dir

 arch/arm64/include/asm/mmu_context.h |  6 +-
 arch/arm64/include/asm/pgtable.h     |  3 +
 arch/arm64/kernel/cpufeature.c       |  2 +-
 arch/arm64/kernel/head.S             |  6 +-
 arch/arm64/kernel/hibernate.c        |  2 +-
 arch/arm64/kernel/sleep.S            |  1 +
 arch/arm64/mm/kasan_init.c           |  6 +-
 arch/arm64/mm/mmu.c                  | 97 ++++++++++++++++++++--------
 8 files changed, 84 insertions(+), 39 deletions(-)

-- 
2.17.0

[PATCH v3 1/3] arm64/mm: pass swapper_pg_dir as an argument to __enable_mmu()

[Jun Yao]

Introduce __pa_swapper_pg_dir to save physical address of
swapper_pg_dir. And pass it as an argument to __enable_mmu().

Signed-off-by: Jun Yao <yaojun8558363@gmail.com>
---
 arch/arm64/include/asm/mmu_context.h |  4 +---
 arch/arm64/include/asm/pgtable.h     |  1 +
 arch/arm64/kernel/cpufeature.c       |  2 +-
 arch/arm64/kernel/head.S             |  6 ++++--
 arch/arm64/kernel/hibernate.c        |  2 +-
 arch/arm64/kernel/sleep.S            |  1 +
 arch/arm64/mm/kasan_init.c           |  4 ++--
 arch/arm64/mm/mmu.c                  | 20 ++++++++++++++++++--
 8 files changed, 29 insertions(+), 11 deletions(-)

diff --git a/arch/arm64/include/asm/mmu_context.h b/arch/arm64/include/asm/mmu_context.h
index 39ec0b8a689e..3eddb871f251 100644
--- a/arch/arm64/include/asm/mmu_context.h
+++ b/arch/arm64/include/asm/mmu_context.h
@@ -141,14 +141,12 @@ static inline void cpu_install_idmap(void)
  * Atomically replaces the active TTBR1_EL1 PGD with a new VA-compatible PGD,
  * avoiding the possibility of conflicting TLB entries being allocated.
  */
-static inline void cpu_replace_ttbr1(pgd_t *pgdp)
+static inline void cpu_replace_ttbr1(phys_addr_t pgd_phys)
 {
 	typedef void (ttbr_replace_func)(phys_addr_t);
 	extern ttbr_replace_func idmap_cpu_replace_ttbr1;
 	ttbr_replace_func *replace_phys;
 
-	phys_addr_t pgd_phys = virt_to_phys(pgdp);
-
 	replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1);
 
 	cpu_install_idmap();
diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
index 7c4c8f318ba9..519ab5581b08 100644
--- a/arch/arm64/include/asm/pgtable.h
+++ b/arch/arm64/include/asm/pgtable.h
@@ -722,6 +722,7 @@ extern pgd_t swapper_pg_dir[PTRS_PER_PGD];
 extern pgd_t swapper_pg_end[];
 extern pgd_t idmap_pg_dir[PTRS_PER_PGD];
 extern pgd_t tramp_pg_dir[PTRS_PER_PGD];
+extern volatile phys_addr_t __pa_swapper_pg_dir;
 
 /*
  * Encode and decode a swap entry:
diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c
index d2856b129097..e3d76a9dd67a 100644
--- a/arch/arm64/kernel/cpufeature.c
+++ b/arch/arm64/kernel/cpufeature.c
@@ -917,7 +917,7 @@ kpti_install_ng_mappings(const struct arm64_cpu_capabilities *__unused)
 	remap_fn = (void *)__pa_symbol(idmap_kpti_install_ng_mappings);
 
 	cpu_install_idmap();
-	remap_fn(cpu, num_online_cpus(), __pa_symbol(swapper_pg_dir));
+	remap_fn(cpu, num_online_cpus(), __pa_swapper_pg_dir);
 	cpu_uninstall_idmap();
 
 	if (!cpu)
diff --git a/arch/arm64/kernel/head.S b/arch/arm64/kernel/head.S
index b0853069702f..2e871b1cb75f 100644
--- a/arch/arm64/kernel/head.S
+++ b/arch/arm64/kernel/head.S
@@ -706,6 +706,7 @@ secondary_startup:
 	 * Common entry point for secondary CPUs.
 	 */
 	bl	__cpu_setup			// initialise processor
+	ldr_l   x26, __pa_swapper_pg_dir
 	bl	__enable_mmu
 	ldr	x8, =__secondary_switched
 	br	x8
@@ -748,6 +749,7 @@ ENDPROC(__secondary_switched)
  * Enable the MMU.
  *
  *  x0  = SCTLR_EL1 value for turning on the MMU.
+ *  x26 = TTBR1 value for turning on the MMU.
  *
  * Returns to the caller via x30/lr. This requires the caller to be covered
  * by the .idmap.text section.
@@ -762,9 +764,8 @@ ENTRY(__enable_mmu)
 	b.ne	__no_granule_support
 	update_early_cpu_boot_status 0, x1, x2
 	adrp	x1, idmap_pg_dir
-	adrp	x2, swapper_pg_dir
 	phys_to_ttbr x3, x1
-	phys_to_ttbr x4, x2
+	phys_to_ttbr x4, x26
 	msr	ttbr0_el1, x3			// load TTBR0
 	msr	ttbr1_el1, x4			// load TTBR1
 	isb
@@ -823,6 +824,7 @@ __primary_switch:
 	mrs	x20, sctlr_el1			// preserve old SCTLR_EL1 value
 #endif
 
+	adrp    x26, swapper_pg_dir
 	bl	__enable_mmu
 #ifdef CONFIG_RELOCATABLE
 	bl	__relocate_kernel
diff --git a/arch/arm64/kernel/hibernate.c b/arch/arm64/kernel/hibernate.c
index 6b2686d54411..0a0a0ca19f9b 100644
--- a/arch/arm64/kernel/hibernate.c
+++ b/arch/arm64/kernel/hibernate.c
@@ -125,7 +125,7 @@ int arch_hibernation_header_save(void *addr, unsigned int max_size)
 		return -EOVERFLOW;
 
 	arch_hdr_invariants(&hdr->invariants);
-	hdr->ttbr1_el1		= __pa_symbol(swapper_pg_dir);
+	hdr->ttbr1_el1          = __pa_swapper_pg_dir;
 	hdr->reenter_kernel	= _cpu_resume;
 
 	/* We can't use __hyp_get_vectors() because kvm may still be loaded */
diff --git a/arch/arm64/kernel/sleep.S b/arch/arm64/kernel/sleep.S
index bebec8ef9372..03854c329449 100644
--- a/arch/arm64/kernel/sleep.S
+++ b/arch/arm64/kernel/sleep.S
@@ -101,6 +101,7 @@ ENTRY(cpu_resume)
 	bl	el2_setup		// if in EL2 drop to EL1 cleanly
 	bl	__cpu_setup
 	/* enable the MMU early - so we can access sleep_save_stash by va */
+	ldr_l   x26, __pa_swapper_pg_dir
 	bl	__enable_mmu
 	ldr	x8, =_cpu_resume
 	br	x8
diff --git a/arch/arm64/mm/kasan_init.c b/arch/arm64/mm/kasan_init.c
index 12145874c02b..dd4f28c19165 100644
--- a/arch/arm64/mm/kasan_init.c
+++ b/arch/arm64/mm/kasan_init.c
@@ -199,7 +199,7 @@ void __init kasan_init(void)
 	 */
 	memcpy(tmp_pg_dir, swapper_pg_dir, sizeof(tmp_pg_dir));
 	dsb(ishst);
-	cpu_replace_ttbr1(lm_alias(tmp_pg_dir));
+	cpu_replace_ttbr1(__pa_symbol(tmp_pg_dir));
 
 	clear_pgds(KASAN_SHADOW_START, KASAN_SHADOW_END);
 
@@ -236,7 +236,7 @@ void __init kasan_init(void)
 			pfn_pte(sym_to_pfn(kasan_zero_page), PAGE_KERNEL_RO));
 
 	memset(kasan_zero_page, 0, PAGE_SIZE);
-	cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
+	cpu_replace_ttbr1(__pa_swapper_pg_dir);
 
 	/* At this point kasan is fully initialized. Enable error messages */
 	init_task.kasan_depth = 0;
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index 2dbb2c9f1ec1..d518e27792ef 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -55,6 +55,13 @@ u64 idmap_ptrs_per_pgd = PTRS_PER_PGD;
 u64 kimage_voffset __ro_after_init;
 EXPORT_SYMBOL(kimage_voffset);
 
+/*
+ * Ensure that hibernate will re-clean this value to the PoC once
+ * it has restored the memory.
+ */
+volatile phys_addr_t __section(".mmuoff.data.read")
+__pa_swapper_pg_dir;
+
 /*
  * Empty_zero_page is a special page that is used for zero-initialized data
  * and COW.
@@ -631,6 +638,15 @@ void __init paging_init(void)
 	phys_addr_t pgd_phys = early_pgtable_alloc();
 	pgd_t *pgdp = pgd_set_fixmap(pgd_phys);
 
+	__pa_swapper_pg_dir = __pa_symbol(swapper_pg_dir);
+
+	/*
+	 * We need to clean '__pa_swapper_pg_dir' to the PoC, so that
+	 * secondaries will read the new value.
+	 */
+	__flush_dcache_area((void *)&__pa_swapper_pg_dir,
+			    sizeof(__pa_swapper_pg_dir));
+
 	map_kernel(pgdp);
 	map_mem(pgdp);
 
@@ -642,9 +658,9 @@ void __init paging_init(void)
 	 *
 	 * To do this we need to go via a temporary pgd.
 	 */
-	cpu_replace_ttbr1(__va(pgd_phys));
+	cpu_replace_ttbr1(pgd_phys);
 	memcpy(swapper_pg_dir, pgdp, PGD_SIZE);
-	cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
+	cpu_replace_ttbr1(__pa_swapper_pg_dir);
 
 	pgd_clear_fixmap();
 	memblock_free(pgd_phys, PAGE_SIZE);
-- 
2.17.0

[PATCH v3 2/3] arm64/mm: introduce variable to save new swapper_pg_dir address

[Jun Yao]

Prepare for migrating swapper_pg_dir, introduce new_swapper_pg_dir
to save virtual address of swapper_pg_dir.

Signed-off-by: Jun Yao <yaojun8558363@gmail.com>
---
 arch/arm64/include/asm/mmu_context.h | 2 +-
 arch/arm64/include/asm/pgtable.h     | 1 +
 arch/arm64/mm/kasan_init.c           | 2 +-
 arch/arm64/mm/mmu.c                  | 2 ++
 4 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/arch/arm64/include/asm/mmu_context.h b/arch/arm64/include/asm/mmu_context.h
index 3eddb871f251..481c2d16adeb 100644
--- a/arch/arm64/include/asm/mmu_context.h
+++ b/arch/arm64/include/asm/mmu_context.h
@@ -57,7 +57,7 @@ static inline void cpu_set_reserved_ttbr0(void)
 
 static inline void cpu_switch_mm(pgd_t *pgd, struct mm_struct *mm)
 {
-	BUG_ON(pgd == swapper_pg_dir);
+	BUG_ON(pgd == new_swapper_pg_dir);
 	cpu_set_reserved_ttbr0();
 	cpu_do_switch_mm(virt_to_phys(pgd),mm);
 }
diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
index 519ab5581b08..2bda899dcf22 100644
--- a/arch/arm64/include/asm/pgtable.h
+++ b/arch/arm64/include/asm/pgtable.h
@@ -723,6 +723,7 @@ extern pgd_t swapper_pg_end[];
 extern pgd_t idmap_pg_dir[PTRS_PER_PGD];
 extern pgd_t tramp_pg_dir[PTRS_PER_PGD];
 extern volatile phys_addr_t __pa_swapper_pg_dir;
+extern pgd_t *new_swapper_pg_dir;
 
 /*
  * Encode and decode a swap entry:
diff --git a/arch/arm64/mm/kasan_init.c b/arch/arm64/mm/kasan_init.c
index dd4f28c19165..08bcaae4725e 100644
--- a/arch/arm64/mm/kasan_init.c
+++ b/arch/arm64/mm/kasan_init.c
@@ -197,7 +197,7 @@ void __init kasan_init(void)
 	 * tmp_pg_dir used to keep early shadow mapped until full shadow
 	 * setup will be finished.
 	 */
-	memcpy(tmp_pg_dir, swapper_pg_dir, sizeof(tmp_pg_dir));
+	memcpy(tmp_pg_dir, new_swapper_pg_dir, sizeof(tmp_pg_dir));
 	dsb(ishst);
 	cpu_replace_ttbr1(__pa_symbol(tmp_pg_dir));
 
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index d518e27792ef..94056e064c6f 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -62,6 +62,8 @@ EXPORT_SYMBOL(kimage_voffset);
 volatile phys_addr_t __section(".mmuoff.data.read")
 __pa_swapper_pg_dir;
 
+pgd_t *new_swapper_pg_dir = swapper_pg_dir;
+
 /*
  * Empty_zero_page is a special page that is used for zero-initialized data
  * and COW.
-- 
2.17.0

[PATCH v3 3/3] arm64/mm: migrate swapper_pg_dir and tramp_pg_dir

[Jun Yao]

Migrate swapper_pg_dir and tramp_pg_dir. And their virtual addresses
do not correlate with kernel's address.

Signed-off-by: Jun Yao <yaojun8558363@gmail.com>
---
 arch/arm64/include/asm/pgtable.h |  1 +
 arch/arm64/mm/mmu.c              | 79 +++++++++++++++++++++-----------
 2 files changed, 52 insertions(+), 28 deletions(-)

diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
index 2bda899dcf22..b032d6c2e390 100644
--- a/arch/arm64/include/asm/pgtable.h
+++ b/arch/arm64/include/asm/pgtable.h
@@ -722,6 +722,7 @@ extern pgd_t swapper_pg_dir[PTRS_PER_PGD];
 extern pgd_t swapper_pg_end[];
 extern pgd_t idmap_pg_dir[PTRS_PER_PGD];
 extern pgd_t tramp_pg_dir[PTRS_PER_PGD];
+extern pgd_t reserved_ttbr0[PTRS_PER_PGD];
 extern volatile phys_addr_t __pa_swapper_pg_dir;
 extern pgd_t *new_swapper_pg_dir;
 
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index 94056e064c6f..ba0b55158971 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -63,6 +63,9 @@ volatile phys_addr_t __section(".mmuoff.data.read")
 __pa_swapper_pg_dir;
 
 pgd_t *new_swapper_pg_dir = swapper_pg_dir;
+#ifdef CONFIG_UNMAP_KERNEL_AT_EL0
+pgd_t *new_tramp_pg_dir;
+#endif
 
 /*
  * Empty_zero_page is a special page that is used for zero-initialized data
@@ -86,19 +89,14 @@ pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
 }
 EXPORT_SYMBOL(phys_mem_access_prot);
 
-static phys_addr_t __init early_pgtable_alloc(void)
+static void __init clear_page_phys(phys_addr_t phys)
 {
-	phys_addr_t phys;
-	void *ptr;
-
-	phys = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
-
 	/*
 	 * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
 	 * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
 	 * any level of table.
 	 */
-	ptr = pte_set_fixmap(phys);
+	void *ptr = pte_set_fixmap(phys);
 
 	memset(ptr, 0, PAGE_SIZE);
 
@@ -107,6 +105,14 @@ static phys_addr_t __init early_pgtable_alloc(void)
 	 * table walker
 	 */
 	pte_clear_fixmap();
+}
+
+static phys_addr_t __init early_pgtable_alloc(void)
+{
+	phys_addr_t phys;
+
+	phys = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
+	clear_page_phys(phys);
 
 	return phys;
 }
@@ -560,6 +566,10 @@ static int __init map_entry_trampoline(void)
 	__create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, PAGE_SIZE,
 			     prot, pgd_pgtable_alloc, 0);
 
+	memcpy(new_tramp_pg_dir, tramp_pg_dir, PGD_SIZE);
+	memblock_free(__pa_symbol(tramp_pg_dir),
+		__pa_symbol(swapper_pg_dir) - __pa_symbol(tramp_pg_dir));
+
 	/* Map both the text and data into the kernel page table */
 	__set_fixmap(FIX_ENTRY_TRAMP_TEXT, pa_start, prot);
 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
@@ -637,10 +647,29 @@ static void __init map_kernel(pgd_t *pgdp)
  */
 void __init paging_init(void)
 {
-	phys_addr_t pgd_phys = early_pgtable_alloc();
-	pgd_t *pgdp = pgd_set_fixmap(pgd_phys);
+	phys_addr_t pgd_phys;
+	pgd_t *pgdp;
+	phys_addr_t mem_size;
 
-	__pa_swapper_pg_dir = __pa_symbol(swapper_pg_dir);
+	mem_size = __pa_symbol(swapper_pg_dir) + PAGE_SIZE
+			- (__pa_symbol(idmap_pg_dir) + IDMAP_DIR_SIZE);
+
+	if (mem_size == PAGE_SIZE) {
+		pgd_phys = early_pgtable_alloc();
+		__pa_swapper_pg_dir = pgd_phys;
+	} else {
+		phys_addr_t p;
+
+		pgd_phys = memblock_alloc(mem_size, PAGE_SIZE);
+
+		for (p = pgd_phys; p < pgd_phys + mem_size; p += PAGE_SIZE)
+			clear_page_phys(p);
+
+		#ifdef CONFIG_UNMAP_KERNEL_AT_EL0
+		new_tramp_pg_dir = __va(pgd_phys);
+		#endif
+		__pa_swapper_pg_dir = pgd_phys + mem_size - PAGE_SIZE;
+	}
 
 	/*
 	 * We need to clean '__pa_swapper_pg_dir' to the PoC, so that
@@ -649,31 +678,25 @@ void __init paging_init(void)
 	__flush_dcache_area((void *)&__pa_swapper_pg_dir,
 			    sizeof(__pa_swapper_pg_dir));
 
+	new_swapper_pg_dir = __va(__pa_swapper_pg_dir);
+
+	pgdp = pgd_set_fixmap(__pa_swapper_pg_dir);
+
 	map_kernel(pgdp);
 	map_mem(pgdp);
 
-	/*
-	 * We want to reuse the original swapper_pg_dir so we don't have to
-	 * communicate the new address to non-coherent secondaries in
-	 * secondary_entry, and so cpu_switch_mm can generate the address with
-	 * adrp+add rather than a load from some global variable.
-	 *
-	 * To do this we need to go via a temporary pgd.
-	 */
-	cpu_replace_ttbr1(pgd_phys);
-	memcpy(swapper_pg_dir, pgdp, PGD_SIZE);
 	cpu_replace_ttbr1(__pa_swapper_pg_dir);
+	init_mm.pgd = new_swapper_pg_dir;
 
 	pgd_clear_fixmap();
-	memblock_free(pgd_phys, PAGE_SIZE);
 
-	/*
-	 * We only reuse the PGD from the swapper_pg_dir, not the pud + pmd
-	 * allocated with it.
-	 */
-	memblock_free(__pa_symbol(swapper_pg_dir) + PAGE_SIZE,
-		      __pa_symbol(swapper_pg_end) - __pa_symbol(swapper_pg_dir)
-		      - PAGE_SIZE);
+#ifdef CONFIG_ARM64_SW_TTBR0_PAN
+	memblock_free(__pa_symbol(reserved_ttbr0), RESERVED_TTBR0_SIZE);
+#endif
+
+	memblock_free(__pa_symbol(swapper_pg_dir),
+		__pa_symbol(swapper_pg_end) - __pa_symbol(swapper_pg_dir));
+
 }
 
 /*
-- 
2.17.0

Re: [PATCH v3 3/3] arm64/mm: migrate swapper_pg_dir and tramp_pg_dir

[James Morse]

Hi Jun,

On 06/06/18 05:39, Jun Yao wrote:
> Migrate swapper_pg_dir and tramp_pg_dir. And their virtual addresses
> do not correlate with kernel's address.

This is all to make 'KSMA' harder, where an single arbitrary write is used to
add a block mapping to the page-tables, giving the attacker full access to
physical memory.

This series is pretty complicated, and would still be overcome by arbitrary-read
of __pa_swapper_pg_dir. Any attacker already has to defeat KASLR, but I guess
that can be done without having read-access.


> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index 94056e064c6f..ba0b55158971 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -637,10 +647,29 @@ static void __init map_kernel(pgd_t *pgdp)
>   */
>  void __init paging_init(void)
>  {
> -	phys_addr_t pgd_phys = early_pgtable_alloc();
> -	pgd_t *pgdp = pgd_set_fixmap(pgd_phys);
> +	phys_addr_t pgd_phys;
> +	pgd_t *pgdp;
> +	phys_addr_t mem_size;
>
> -	__pa_swapper_pg_dir = __pa_symbol(swapper_pg_dir);
> +	mem_size = __pa_symbol(swapper_pg_dir) + PAGE_SIZE
> +			- (__pa_symbol(idmap_pg_dir) + IDMAP_DIR_SIZE);
> +
> +	if (mem_size == PAGE_SIZE) {
> +		pgd_phys = early_pgtable_alloc();

This is where we migrated swapper_pg_dir to, its the first memblock_alloc()
call, and is guaranteed to be at the top of memory. Moving this allocation later
will still cause a predictable page to be allocated as this setup code is the
only thing running. I don't think we are in a better position after this series.


Ideally we would put these pgd entries in the read-only section. We only modify
swapper_pg_dir for vmalloc()/vmap() once its generated, which we could do
through a fixmap entry.

Like this, an attacker can find where swapper_pg_dir is, but can't write to it.
(the tricky bit will be breaking swapper_pg_dir up, as we free parts of it)


Thanks,

James