[PATCH 0/2] kasan: fixes for 5.11-rc

Andrey Konovalov (2):
  kasan, mm: fix conflicts with init_on_alloc/free
  kasan, arm64: fix pointer tags in KASAN reports

 arch/arm64/mm/fault.c | 2 ++
 mm/slub.c             | 7 ++++---
 2 files changed, 6 insertions(+), 3 deletions(-)

-- 
2.30.0.284.gd98b1dd5eaa7-goog


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

A few places where SLUB accesses object's data or metadata were missed in
a previous patch. This leads to false positives with hardware tag-based
KASAN when bulk allocations are used with init_on_alloc/free.

Fix the false-positives by resetting pointer tags during these accesses.

Link: https://linux-review.googlesource.com/id/I50dd32838a666e173fe06c3c5c766f2c36aae901
Fixes: aa1ef4d7b3f67 ("kasan, mm: reset tags when accessing metadata")
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Andrey Konovalov <redacted>
---
 mm/slub.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/mm/slub.c b/mm/slub.c
index dc5b42e700b8..75fb097d990d 100644
--- a/mm/slub.c
+++ b/mm/slub.c

@@ -2791,7 +2791,8 @@ static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
 						   void *obj)
 {
 	if (unlikely(slab_want_init_on_free(s)) && obj)
-		memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
+		memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
+			0, sizeof(void *));
 }
 
 /*

@@ -2883,7 +2884,7 @@ static __always_inline void *slab_alloc_node(struct kmem_cache *s,
 		stat(s, ALLOC_FASTPATH);
 	}
 
-	maybe_wipe_obj_freeptr(s, kasan_reset_tag(object));
+	maybe_wipe_obj_freeptr(s, object);
 
 	if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
 		memset(kasan_reset_tag(object), 0, s->object_size);

@@ -3329,7 +3330,7 @@ int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
 		int j;
 
 		for (j = 0; j < i; j++)
-			memset(p[j], 0, s->object_size);
+			memset(kasan_reset_tag(p[j]), 0, s->object_size);
 	}
 
 	/* memcg and kmem_cache debug support */

-- 
2.30.0.284.gd98b1dd5eaa7-goog


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

As of the "arm64: expose FAR_EL1 tag bits in siginfo" patch, the address
that is passed to report_tag_fault has pointer tags in the format of 0x0X,
while KASAN uses 0xFX format (note the difference in the top 4 bits).

Fix up the pointer tag before calling kasan_report.

Link: https://linux-review.googlesource.com/id/I9ced973866036d8679e8f4ae325de547eb969649
Fixes: dceec3ff7807 ("arm64: expose FAR_EL1 tag bits in siginfo")
Fixes: 4291e9ee6189 ("kasan, arm64: print report from tag fault handler")
Signed-off-by: Andrey Konovalov <redacted>
---
 arch/arm64/mm/fault.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..a218f6f2fdc8 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -304,6 +304,8 @@ static void report_tag_fault(unsigned long addr, unsigned int esr,
 {
 	bool is_write  = ((esr & ESR_ELx_WNR) >> ESR_ELx_WNR_SHIFT) != 0;
 
+	/* The format of KASAN tags is 0xF<x>. */
+	addr |= (0xF0UL << MTE_TAG_SHIFT);
 	/*
 	 * SAS bits aren't set for all faults reported in EL1, so we can't
 	 * find out access size.

-- 
2.30.0.284.gd98b1dd5eaa7-goog


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

On Wed, Jan 13, 2021 at 05:03:30PM +0100, Andrey Konovalov wrote:

Ah, I see, that top 4 bits are zeroed by do_tag_check_fault(). When this
was added, the only tag faults were generated for user addresses.

Anyway, I'd rather fix it in there based on bit 55, something like (only
compile-tested):

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..2b71079d2d32 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -709,10 +709,11 @@ static int do_tag_check_fault(unsigned long far, unsigned int esr,
 			      struct pt_regs *regs)
 {
 	/*
-	 * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN for tag
-	 * check faults. Mask them out now so that userspace doesn't see them.
+	 * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN
+	 * for tag check faults. Set them to the corresponding bits in the
+	 * untagged address.
 	 */
-	far &= (1UL << 60) - 1;
+	far = (untagged_addr(far) & ~MTE_TAG_MASK) | (far & MTE_TAG_MASK) ;
 	do_bad_area(far, esr, regs);
 	return 0;
 }

-- 
Catalin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

On 1/13/21 5:03 PM, Andrey Konovalov wrote:

A few places where SLUB accesses object's data or metadata were missed in
a previous patch. This leads to false positives with hardware tag-based
KASAN when bulk allocations are used with init_on_alloc/free.

Fix the false-positives by resetting pointer tags during these accesses.

Link: https://linux-review.googlesource.com/id/I50dd32838a666e173fe06c3c5c766f2c36aae901
Fixes: aa1ef4d7b3f67 ("kasan, mm: reset tags when accessing metadata")
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Andrey Konovalov <redacted>

Acked-by: Vlastimil Babka <redacted>

And in that case the reset was unnecessary, right. (commit log only mentions
adding missing resets).

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

On Wed, Jan 13, 2021 at 6:25 PM Vlastimil Babka [off-list ref] wrote:

On 1/13/21 5:03 PM, Andrey Konovalov wrote:

quoted

A few places where SLUB accesses object's data or metadata were missed in
a previous patch. This leads to false positives with hardware tag-based
KASAN when bulk allocations are used with init_on_alloc/free.

Fix the false-positives by resetting pointer tags during these accesses.

Link: https://linux-review.googlesource.com/id/I50dd32838a666e173fe06c3c5c766f2c36aae901
Fixes: aa1ef4d7b3f67 ("kasan, mm: reset tags when accessing metadata")
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Andrey Konovalov <redacted>

Acked-by: Vlastimil Babka <redacted>

quoted

---
 mm/slub.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/mm/slub.c b/mm/slub.c
index dc5b42e700b8..75fb097d990d 100644
--- a/mm/slub.c
+++ b/mm/slub.c

@@ -2791,7 +2791,8 @@ static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
                                                 void *obj)
 {
      if (unlikely(slab_want_init_on_free(s)) && obj)
-             memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
+             memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
+                     0, sizeof(void *));
 }

 /*

@@ -2883,7 +2884,7 @@ static __always_inline void *slab_alloc_node(struct kmem_cache *s,
              stat(s, ALLOC_FASTPATH);
      }

-     maybe_wipe_obj_freeptr(s, kasan_reset_tag(object));
+     maybe_wipe_obj_freeptr(s, object);

And in that case the reset was unnecessary, right. (commit log only mentions
adding missing resets).

The reset has been moved into maybe_wipe_obj_freeptr(). I'll mention
it in the changelog in v2.

Thanks!

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

On Wed, Jan 13, 2021 at 5:54 PM Catalin Marinas [off-list ref] wrote:

Sounds good, will do in v3, thanks!

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

On Fri, Jan 15, 2021 at 02:12:24PM +0100, Andrey Konovalov wrote:

On Wed, Jan 13, 2021 at 5:54 PM Catalin Marinas [off-list ref] wrote:

quoted

On Wed, Jan 13, 2021 at 05:03:30PM +0100, Andrey Konovalov wrote:

quoted

As of the "arm64: expose FAR_EL1 tag bits in siginfo" patch, the address
that is passed to report_tag_fault has pointer tags in the format of 0x0X,
while KASAN uses 0xFX format (note the difference in the top 4 bits).

Fix up the pointer tag before calling kasan_report.

Link: https://linux-review.googlesource.com/id/I9ced973866036d8679e8f4ae325de547eb969649
Fixes: dceec3ff7807 ("arm64: expose FAR_EL1 tag bits in siginfo")
Fixes: 4291e9ee6189 ("kasan, arm64: print report from tag fault handler")
Signed-off-by: Andrey Konovalov <redacted>
---
 arch/arm64/mm/fault.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..a218f6f2fdc8 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -304,6 +304,8 @@ static void report_tag_fault(unsigned long addr, unsigned int esr,
 {
      bool is_write  = ((esr & ESR_ELx_WNR) >> ESR_ELx_WNR_SHIFT) != 0;

+     /* The format of KASAN tags is 0xF<x>. */
+     addr |= (0xF0UL << MTE_TAG_SHIFT);

Ah, I see, that top 4 bits are zeroed by do_tag_check_fault(). When this
was added, the only tag faults were generated for user addresses.

Anyway, I'd rather fix it in there based on bit 55, something like (only
compile-tested):

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..2b71079d2d32 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -709,10 +709,11 @@ static int do_tag_check_fault(unsigned long far, unsigned int esr,
                              struct pt_regs *regs)
 {
        /*
-        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN for tag
-        * check faults. Mask them out now so that userspace doesn't see them.
+        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN
+        * for tag check faults. Set them to the corresponding bits in the
+        * untagged address.
         */
-       far &= (1UL << 60) - 1;
+       far = (untagged_addr(far) & ~MTE_TAG_MASK) | (far & MTE_TAG_MASK) ;
        do_bad_area(far, esr, regs);
        return 0;
 }

Sounds good, will do in v3, thanks!

I wonder if this one gives the same result (so please check):

	far = u64_replace_bits(untagged_addr(far), far, MTE_TAG_MASK);

(defined in linux/bitfield.h)

-- 
Catalin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

On Fri, Jan 15, 2021 at 4:07 PM Catalin Marinas [off-list ref] wrote:

On Fri, Jan 15, 2021 at 02:12:24PM +0100, Andrey Konovalov wrote:

quoted

On Wed, Jan 13, 2021 at 5:54 PM Catalin Marinas [off-list ref] wrote:

quoted

On Wed, Jan 13, 2021 at 05:03:30PM +0100, Andrey Konovalov wrote:

quoted

As of the "arm64: expose FAR_EL1 tag bits in siginfo" patch, the address
that is passed to report_tag_fault has pointer tags in the format of 0x0X,
while KASAN uses 0xFX format (note the difference in the top 4 bits).

Fix up the pointer tag before calling kasan_report.

Link: https://linux-review.googlesource.com/id/I9ced973866036d8679e8f4ae325de547eb969649
Fixes: dceec3ff7807 ("arm64: expose FAR_EL1 tag bits in siginfo")
Fixes: 4291e9ee6189 ("kasan, arm64: print report from tag fault handler")
Signed-off-by: Andrey Konovalov <redacted>
---
 arch/arm64/mm/fault.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..a218f6f2fdc8 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -304,6 +304,8 @@ static void report_tag_fault(unsigned long addr, unsigned int esr,
 {
      bool is_write  = ((esr & ESR_ELx_WNR) >> ESR_ELx_WNR_SHIFT) != 0;

+     /* The format of KASAN tags is 0xF<x>. */
+     addr |= (0xF0UL << MTE_TAG_SHIFT);

Ah, I see, that top 4 bits are zeroed by do_tag_check_fault(). When this
was added, the only tag faults were generated for user addresses.

Anyway, I'd rather fix it in there based on bit 55, something like (only
compile-tested):

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..2b71079d2d32 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -709,10 +709,11 @@ static int do_tag_check_fault(unsigned long far, unsigned int esr,
                              struct pt_regs *regs)
 {
        /*
-        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN for tag
-        * check faults. Mask them out now so that userspace doesn't see them.
+        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN
+        * for tag check faults. Set them to the corresponding bits in the
+        * untagged address.
         */
-       far &= (1UL << 60) - 1;
+       far = (untagged_addr(far) & ~MTE_TAG_MASK) | (far & MTE_TAG_MASK) ;
        do_bad_area(far, esr, regs);
        return 0;
 }

Sounds good, will do in v3, thanks!

I wonder if this one gives the same result (so please check):

        far = u64_replace_bits(untagged_addr(far), far, MTE_TAG_MASK);

(defined in linux/bitfield.h)

No, it zeroes out the tag. Not sure why. I took a brief look at the
implementation and didn't get how it's supposed to work - too much bit
trickery.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

On Wed, Jan 13, 2021 at 5:54 PM Catalin Marinas [off-list ref] wrote:

BTW, we can do "untagged_addr(far) | (far & MTE_TAG_MASK)" here, as
untagged_addr() doesn't change kernel pointers.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

On Fri, Jan 15, 2021 at 05:30:40PM +0100, Andrey Konovalov wrote:

On Wed, Jan 13, 2021 at 5:54 PM Catalin Marinas [off-list ref] wrote:

quoted

On Wed, Jan 13, 2021 at 05:03:30PM +0100, Andrey Konovalov wrote:

quoted

As of the "arm64: expose FAR_EL1 tag bits in siginfo" patch, the address
that is passed to report_tag_fault has pointer tags in the format of 0x0X,
while KASAN uses 0xFX format (note the difference in the top 4 bits).

Fix up the pointer tag before calling kasan_report.

Link: https://linux-review.googlesource.com/id/I9ced973866036d8679e8f4ae325de547eb969649
Fixes: dceec3ff7807 ("arm64: expose FAR_EL1 tag bits in siginfo")
Fixes: 4291e9ee6189 ("kasan, arm64: print report from tag fault handler")
Signed-off-by: Andrey Konovalov <redacted>
---
 arch/arm64/mm/fault.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..a218f6f2fdc8 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -304,6 +304,8 @@ static void report_tag_fault(unsigned long addr, unsigned int esr,
 {
      bool is_write  = ((esr & ESR_ELx_WNR) >> ESR_ELx_WNR_SHIFT) != 0;

+     /* The format of KASAN tags is 0xF<x>. */
+     addr |= (0xF0UL << MTE_TAG_SHIFT);

Ah, I see, that top 4 bits are zeroed by do_tag_check_fault(). When this
was added, the only tag faults were generated for user addresses.

Anyway, I'd rather fix it in there based on bit 55, something like (only
compile-tested):

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..2b71079d2d32 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -709,10 +709,11 @@ static int do_tag_check_fault(unsigned long far, unsigned int esr,
                              struct pt_regs *regs)
 {
        /*
-        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN for tag
-        * check faults. Mask them out now so that userspace doesn't see them.
+        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN
+        * for tag check faults. Set them to the corresponding bits in the
+        * untagged address.
         */
-       far &= (1UL << 60) - 1;
+       far = (untagged_addr(far) & ~MTE_TAG_MASK) | (far & MTE_TAG_MASK) ;
        do_bad_area(far, esr, regs);
        return 0;
 }

BTW, we can do "untagged_addr(far) | (far & MTE_TAG_MASK)" here, as
untagged_addr() doesn't change kernel pointers.

untagged_addr() does change tagged kernel pointers, it sign-extends from
bit 55. So the top byte becomes 0xff and you can no longer or the tag
bits in.

-- 
Catalin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

On Fri, Jan 15, 2021 at 5:56 PM Catalin Marinas [off-list ref] wrote:

On Fri, Jan 15, 2021 at 05:30:40PM +0100, Andrey Konovalov wrote:

quoted

On Wed, Jan 13, 2021 at 5:54 PM Catalin Marinas [off-list ref] wrote:

quoted

On Wed, Jan 13, 2021 at 05:03:30PM +0100, Andrey Konovalov wrote:

quoted

As of the "arm64: expose FAR_EL1 tag bits in siginfo" patch, the address
that is passed to report_tag_fault has pointer tags in the format of 0x0X,
while KASAN uses 0xFX format (note the difference in the top 4 bits).

Fix up the pointer tag before calling kasan_report.

Link: https://linux-review.googlesource.com/id/I9ced973866036d8679e8f4ae325de547eb969649
Fixes: dceec3ff7807 ("arm64: expose FAR_EL1 tag bits in siginfo")
Fixes: 4291e9ee6189 ("kasan, arm64: print report from tag fault handler")
Signed-off-by: Andrey Konovalov <redacted>
---
 arch/arm64/mm/fault.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..a218f6f2fdc8 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -304,6 +304,8 @@ static void report_tag_fault(unsigned long addr, unsigned int esr,
 {
      bool is_write  = ((esr & ESR_ELx_WNR) >> ESR_ELx_WNR_SHIFT) != 0;

+     /* The format of KASAN tags is 0xF<x>. */
+     addr |= (0xF0UL << MTE_TAG_SHIFT);

Ah, I see, that top 4 bits are zeroed by do_tag_check_fault(). When this
was added, the only tag faults were generated for user addresses.

Anyway, I'd rather fix it in there based on bit 55, something like (only
compile-tested):

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..2b71079d2d32 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -709,10 +709,11 @@ static int do_tag_check_fault(unsigned long far, unsigned int esr,
                              struct pt_regs *regs)
 {
        /*
-        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN for tag
-        * check faults. Mask them out now so that userspace doesn't see them.
+        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN
+        * for tag check faults. Set them to the corresponding bits in the
+        * untagged address.
         */
-       far &= (1UL << 60) - 1;
+       far = (untagged_addr(far) & ~MTE_TAG_MASK) | (far & MTE_TAG_MASK) ;
        do_bad_area(far, esr, regs);
        return 0;
 }

BTW, we can do "untagged_addr(far) | (far & MTE_TAG_MASK)" here, as
untagged_addr() doesn't change kernel pointers.

untagged_addr() does change tagged kernel pointers, it sign-extends from
bit 55. So the top byte becomes 0xff and you can no longer or the tag
bits in.

That's __untagged_addr(), untagged_addr() keeps the bits for kernel
pointers as of  597399d0cb91.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

On Fri, Jan 15, 2021 at 06:00:36PM +0100, Andrey Konovalov wrote:

On Fri, Jan 15, 2021 at 5:56 PM Catalin Marinas [off-list ref] wrote:

quoted

On Fri, Jan 15, 2021 at 05:30:40PM +0100, Andrey Konovalov wrote:

quoted

On Wed, Jan 13, 2021 at 5:54 PM Catalin Marinas [off-list ref] wrote:

quoted

On Wed, Jan 13, 2021 at 05:03:30PM +0100, Andrey Konovalov wrote:

quoted

As of the "arm64: expose FAR_EL1 tag bits in siginfo" patch, the address
that is passed to report_tag_fault has pointer tags in the format of 0x0X,
while KASAN uses 0xFX format (note the difference in the top 4 bits).

Fix up the pointer tag before calling kasan_report.

Link: https://linux-review.googlesource.com/id/I9ced973866036d8679e8f4ae325de547eb969649
Fixes: dceec3ff7807 ("arm64: expose FAR_EL1 tag bits in siginfo")
Fixes: 4291e9ee6189 ("kasan, arm64: print report from tag fault handler")
Signed-off-by: Andrey Konovalov <redacted>
---
 arch/arm64/mm/fault.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..a218f6f2fdc8 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -304,6 +304,8 @@ static void report_tag_fault(unsigned long addr, unsigned int esr,
 {
      bool is_write  = ((esr & ESR_ELx_WNR) >> ESR_ELx_WNR_SHIFT) != 0;

+     /* The format of KASAN tags is 0xF<x>. */
+     addr |= (0xF0UL << MTE_TAG_SHIFT);

Ah, I see, that top 4 bits are zeroed by do_tag_check_fault(). When this
was added, the only tag faults were generated for user addresses.

Anyway, I'd rather fix it in there based on bit 55, something like (only
compile-tested):

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..2b71079d2d32 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -709,10 +709,11 @@ static int do_tag_check_fault(unsigned long far, unsigned int esr,
                              struct pt_regs *regs)
 {
        /*
-        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN for tag
-        * check faults. Mask them out now so that userspace doesn't see them.
+        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN
+        * for tag check faults. Set them to the corresponding bits in the
+        * untagged address.
         */
-       far &= (1UL << 60) - 1;
+       far = (untagged_addr(far) & ~MTE_TAG_MASK) | (far & MTE_TAG_MASK) ;
        do_bad_area(far, esr, regs);
        return 0;
 }

BTW, we can do "untagged_addr(far) | (far & MTE_TAG_MASK)" here, as
untagged_addr() doesn't change kernel pointers.

untagged_addr() does change tagged kernel pointers, it sign-extends from
bit 55. So the top byte becomes 0xff and you can no longer or the tag
bits in.

That's __untagged_addr(), untagged_addr() keeps the bits for kernel
pointers as of  597399d0cb91.

Ah, you are right. In this case I think we should use __untagged_addr()
above. Even if the tag check fault happened on a kernel address, bits
63:60 are still unknown.

-- 
Catalin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

On Fri, Jan 15, 2021 at 6:06 PM Catalin Marinas [off-list ref] wrote:

On Fri, Jan 15, 2021 at 06:00:36PM +0100, Andrey Konovalov wrote:

quoted

On Fri, Jan 15, 2021 at 5:56 PM Catalin Marinas [off-list ref] wrote:

quoted

On Fri, Jan 15, 2021 at 05:30:40PM +0100, Andrey Konovalov wrote:

quoted

On Wed, Jan 13, 2021 at 5:54 PM Catalin Marinas [off-list ref] wrote:

quoted

On Wed, Jan 13, 2021 at 05:03:30PM +0100, Andrey Konovalov wrote:

quoted

As of the "arm64: expose FAR_EL1 tag bits in siginfo" patch, the address
that is passed to report_tag_fault has pointer tags in the format of 0x0X,
while KASAN uses 0xFX format (note the difference in the top 4 bits).

Fix up the pointer tag before calling kasan_report.

Link: https://linux-review.googlesource.com/id/I9ced973866036d8679e8f4ae325de547eb969649
Fixes: dceec3ff7807 ("arm64: expose FAR_EL1 tag bits in siginfo")
Fixes: 4291e9ee6189 ("kasan, arm64: print report from tag fault handler")
Signed-off-by: Andrey Konovalov <redacted>
---
 arch/arm64/mm/fault.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..a218f6f2fdc8 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -304,6 +304,8 @@ static void report_tag_fault(unsigned long addr, unsigned int esr,
 {
      bool is_write  = ((esr & ESR_ELx_WNR) >> ESR_ELx_WNR_SHIFT) != 0;

+     /* The format of KASAN tags is 0xF<x>. */
+     addr |= (0xF0UL << MTE_TAG_SHIFT);

Ah, I see, that top 4 bits are zeroed by do_tag_check_fault(). When this
was added, the only tag faults were generated for user addresses.

Anyway, I'd rather fix it in there based on bit 55, something like (only
compile-tested):

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 3c40da479899..2b71079d2d32 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c

@@ -709,10 +709,11 @@ static int do_tag_check_fault(unsigned long far, unsigned int esr,
                              struct pt_regs *regs)
 {
        /*
-        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN for tag
-        * check faults. Mask them out now so that userspace doesn't see them.
+        * The architecture specifies that bits 63:60 of FAR_EL1 are UNKNOWN
+        * for tag check faults. Set them to the corresponding bits in the
+        * untagged address.
         */
-       far &= (1UL << 60) - 1;
+       far = (untagged_addr(far) & ~MTE_TAG_MASK) | (far & MTE_TAG_MASK) ;
        do_bad_area(far, esr, regs);
        return 0;
 }

BTW, we can do "untagged_addr(far) | (far & MTE_TAG_MASK)" here, as
untagged_addr() doesn't change kernel pointers.

untagged_addr() does change tagged kernel pointers, it sign-extends from
bit 55. So the top byte becomes 0xff and you can no longer or the tag
bits in.

That's __untagged_addr(), untagged_addr() keeps the bits for kernel
pointers as of  597399d0cb91.

Ah, you are right. In this case I think we should use __untagged_addr()
above. Even if the tag check fault happened on a kernel address, bits
63:60 are still unknown.

Yeah, I keep forgetting about [__]untagged_addr() too. Maybe we need
better names? Like untagged_addr() and untagged_addr_ttbr0()?

Anyway, I'll do the explicit calculation with __untagged_addr() in the
next version.

Thanks!

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel