[PATCH v7 1/2] badblocks: fix in-place round_up/round_down usage bug

[PATCH v7 1/2] badblocks: fix in-place round_up/round_down usage bug

[Ramesh Adhikari]

rounddown() and roundup() do not modify their first argument in
place; they return the rounded value. _badblocks_set(),
_badblocks_clear() and badblocks_check() were calling them as
bare statements and discarding the result, so 's' (and 'next'/
'target') were never actually rounded. Depending on the caller's
alignment this can leave 'sectors' unchanged or, in the reported
case, produce a range whose end never advances, causing
_badblocks_check()/badblocks_check() to loop with a non-advancing
cursor and stall the CPU (RCU stall) when called through the
nvdimm ioctl path via nvdimm_clear_badblocks_region().

rounddown()/roundup() also do division/modulo on the sector_t
(u64) operand, which requires libgcc helpers (__aeabi_uldivmod,
__umoddi3) that are not linked into the kernel on 32-bit builds,
breaking the build on arm/i386 (reported by kernel test robot).

Switch to round_down()/round_up() (include/linux/math.h), which
are mask-based, assign their result back to the variable being
rounded, and require no 64-bit division, fixing both the
non-rounding bug and the 32-bit build breakage.

Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202604301231.IpPh4AiH-lkp@intel.com/
Fixes: aa511ff8218b ("badblocks: switch to the improved badblock handling code")
Cc: stable@vger.kernel.org
Signed-off-by: Ramesh Adhikari <adhikari.resume@gmail.com>
Reviewed-by: Coly Li <colyli@fygo.io>
---
 block/badblocks.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/block/badblocks.c b/block/badblocks.c
index ece64e76fe8..1f786b193fb 100644
--- a/block/badblocks.c
+++ b/block/badblocks.c
@@ -857,8 +857,8 @@ static bool _badblocks_set(struct badblocks *bb, sector_t s, sector_t sectors,
 		/* round the start down, and the end up */
 		sector_t next = s + sectors;
 
-		rounddown(s, 1 << bb->shift);
-		roundup(next, 1 << bb->shift);
+		s = round_down(s, 1 << bb->shift);
+		next = round_up(next, 1 << bb->shift);
 		sectors = next - s;
 	}
 
@@ -1071,8 +1071,8 @@ static bool _badblocks_clear(struct badblocks *bb, sector_t s, sector_t sectors)
 		 * isn't than to think a block is not bad when it is.
 		 */
 		target = s + sectors;
-		roundup(s, 1 << bb->shift);
-		rounddown(target, 1 << bb->shift);
+		s = round_up(s, 1 << bb->shift);
+		target = round_down(target, 1 << bb->shift);
 		sectors = target - s;
 	}
 
@@ -1307,8 +1307,8 @@ int badblocks_check(struct badblocks *bb, sector_t s, sector_t sectors,
 		/* round the start down, and the end up */
 		sector_t target = s + sectors;
 
-		rounddown(s, 1 << bb->shift);
-		roundup(target, 1 << bb->shift);
+		s = round_down(s, 1 << bb->shift);
+		target = round_up(target, 1 << bb->shift);
 		sectors = target - s;
 	}
 
-- 
2.43.0

[PATCH v7 0/2] badblocks: fix rounding bug and validate input range

[Ramesh Adhikari]

Two related fixes to block/badblocks.c:

Patch 1/2 fixes the actual reported bug: round_up()/round_down() don't
modify their argument in place, they return the rounded value, and the
callers in _badblocks_set()/_badblocks_clear()/badblocks_check() were
discarding that return value. Depending on caller alignment this could
leave sectors unrounded or, in the reported case, stall the CPU with a
non-advancing cursor when reached via the nvdimm ioctl path. It also
fixes a 32-bit build breakage reported by kernel test robot, since the
old rounddown()/roundup() macros pull in 64-bit division helpers not
linked on 32-bit builds. Unchanged since v6.

Patch 2/2 hardens the same three functions against the input-range and
shift edge cases the round_up/round_down fix exposed: s + sectors
overflow, and rounding itself wrapping past ULLONG_MAX. Revised in v7
per Coly's review of v6 -- see the changelog in that patch for details.
The remaining open item, bounding bb->shift where it's populated from
the on-disk MD superblock in drivers/md/md.c, is being sent separately
since it's a different file and a different root cause (missing input
validation, not a rounding bug).

Both patches carry Coly Li's Reviewed-by.

Ramesh Adhikari (2):
  badblocks: fix in-place round_up/round_down usage bug
  badblocks: validate sector range and shift before rounding

 block/badblocks.c         | 52 ++++++++++++++++++++++++++++++++++++++++-------
 include/linux/badblocks.h |  6 +++++-
 2 files changed, 44 insertions(+), 14 deletions(-)

--
2.43.0

[PATCH v7 1/2] badblocks: fix in-place round_up/round_down usage bug

[Ramesh Adhikari]

rounddown() and roundup() do not modify their first argument in
place; they return the rounded value. _badblocks_set(),
_badblocks_clear() and badblocks_check() were calling them as
bare statements and discarding the result, so 's' (and 'next'/
'target') were never actually rounded. Depending on the caller's
alignment this can leave 'sectors' unchanged or, in the reported
case, produce a range whose end never advances, causing
_badblocks_check()/badblocks_check() to loop with a non-advancing
cursor and stall the CPU (RCU stall) when called through the
nvdimm ioctl path via nvdimm_clear_badblocks_region().

rounddown()/roundup() also do division/modulo on the sector_t
(u64) operand, which requires libgcc helpers (__aeabi_uldivmod,
__umoddi3) that are not linked into the kernel on 32-bit builds,
breaking the build on arm/i386 (reported by kernel test robot).

Switch to round_down()/round_up() (include/linux/math.h), which
are mask-based, assign their result back to the variable being
rounded, and require no 64-bit division, fixing both the
non-rounding bug and the 32-bit build breakage.

Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202604301231.IpPh4AiH-lkp@intel.com/
Fixes: aa511ff8218b ("badblocks: switch to the improved badblock handling code")
Cc: stable@vger.kernel.org
Signed-off-by: Ramesh Adhikari <adhikari.resume@gmail.com>
Reviewed-by: Coly Li <colyli@fygo.io>
---
 block/badblocks.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/block/badblocks.c b/block/badblocks.c
index ece64e76fe8..1f786b193fb 100644
--- a/block/badblocks.c
+++ b/block/badblocks.c
@@ -857,8 +857,8 @@ static bool _badblocks_set(struct badblocks *bb, sector_t s, sector_t sectors,
 		/* round the start down, and the end up */
 		sector_t next = s + sectors;

-		rounddown(s, 1 << bb->shift);
-		roundup(next, 1 << bb->shift);
+		s = round_down(s, 1 << bb->shift);
+		next = round_up(next, 1 << bb->shift);
 		sectors = next - s;
 	}

@@ -1071,8 +1071,8 @@ static bool _badblocks_clear(struct badblocks *bb, sector_t s, sector_t sectors)
 		 * isn't than to think a block is not bad when it is.
 		 */
 		target = s + sectors;
-		roundup(s, 1 << bb->shift);
-		rounddown(target, 1 << bb->shift);
+		s = round_up(s, 1 << bb->shift);
+		target = round_down(target, 1 << bb->shift);
 		sectors = target - s;
 	}

@@ -1307,8 +1307,8 @@ int badblocks_check(struct badblocks *bb, sector_t s, sector_t sectors,
 		/* round the start down, and the end up */
 		sector_t target = s + sectors;

-		rounddown(s, 1 << bb->shift);
-		roundup(target, 1 << bb->shift);
+		s = round_down(s, 1 << bb->shift);
+		target = round_up(target, 1 << bb->shift);
 		sectors = target - s;
 	}

--
2.43.0

[PATCH v7 2/2] badblocks: validate sector range and shift before rounding

[Ramesh Adhikari]

_badblocks_set(), _badblocks_clear() and badblocks_check() round
the caller-supplied [s, s+sectors) range to the current bb->shift
block size before touching the bad block table. That rounding
was not defensive against a few cases:

- s + sectors can overflow sector_t (u64), wrapping the range
  end before it is ever compared against s.

- bb->shift is a plain 'int' field, populated in one case
  (drivers/md/md.c, from the on-disk superblock's bblog_shift)
  straight from an unvalidated byte with no upper bound. Shifting
  by an amount >= the width of the shifted type is undefined
  behaviour in C; "1 << bb->shift" was shifting an int literal,
  so this was already undefined for bb->shift >= 32, let alone
  the full 0-255 range bblog_shift allows.

- round_up()/round_down() rounding a value near ULLONG_MAX can
  itself wrap back to a small value, so even with a valid shift
  the rounded end of the range could end up smaller than the
  rounded start, silently turning a small range into a huge one
  (in _badblocks_clear()/badblocks_check(), which round the end
  up) or losing the range entirely.

Add an explicit s+sectors overflow check and cast the shift
operand to sector_t so the shift itself is never performed on a
32-bit int, and detect post-rounding wrap by comparing the
rounded result back against the pre-rounding value.

badblocks.c does not itself bound bb->shift: every caller except
drivers/md/md.c always leaves it at 0, so the one caller that
populates it from untrusted on-disk data is responsible for
bounding it before assigning it, per struct badblocks's shift
field documentation in include/linux/badblocks.h. That md.c-side
bound is being sent as a separate patch.

badblocks_check() returns 0 rather than -EINVAL on the wrap case,
matching its existing "0: no known bad blocks in the range"
return convention instead of introducing a new error path callers
don't expect.

Suggested-by: Coly Li <colyli@fygo.io>
Fixes: aa511ff8218b ("badblocks: switch to the improved badblock handling code")
Cc: stable@vger.kernel.org
Signed-off-by: Ramesh Adhikari <adhikari.resume@gmail.com>
Reviewed-by: Coly Li <colyli@fygo.io>
---
Changes in v7 (per Coly Li's review of v6):
 - Simplify the overflow check in all three call sites from
   "s > ULLONG_MAX - sectors" to "(s + sectors) < s".
 - Drop the "bb->shift >= BITS_PER_LONG_LONG" guards in badblocks.c.
   Only drivers/md/md.c ever sets a nonzero bb->shift, so the bound
   belongs where bblog_shift is read off the on-disk superblock, not
   scattered through the badblocks API. Documented the caller
   requirement on struct badblocks.shift in badblocks.h instead; the
   md.c-side bound follows as a separate patch.
 - badblocks_check() now returns 0 instead of -EINVAL on the wrap
   case, consistent with its existing return convention.

 block/badblocks.c         | 40 ++++++++++++++++++++++++++++++++-------
 include/linux/badblocks.h |  6 +++++-
 2 files changed, 38 insertions(+), 8 deletions(-)

diff --git a/block/badblocks.c b/block/badblocks.c
index 1f786b193fb..728b59a1a5d 100644
--- a/block/badblocks.c
+++ b/block/badblocks.c
@@ -853,12 +853,19 @@ static bool _badblocks_set(struct badblocks *bb, sector_t s, sector_t sectors,
 		/* Invalid sectors number */
 		return false;

+	if ((s + sectors) < s)
+		/* Range wraps past the end of sector_t */
+		return false;
+
 	if (bb->shift) {
 		/* round the start down, and the end up */
 		sector_t next = s + sectors;

-		s = round_down(s, 1 << bb->shift);
-		next = round_up(next, 1 << bb->shift);
+		s = round_down(s, (sector_t)1 << bb->shift);
+		next = round_up(next, (sector_t)1 << bb->shift);
+		if (next <= s)
+			/* Rounding wrapped past the end of sector_t */
+			return false;
 		sectors = next - s;
 	}

@@ -1061,7 +1068,12 @@ static bool _badblocks_clear(struct badblocks *bb, sector_t s, sector_t sectors)
 		/* Invalid sectors number */
 		return false;

+	if ((s + sectors) < s)
+		/* Range wraps past the end of sector_t */
+		return false;
+
 	if (bb->shift) {
+		sector_t orig_s = s;
 		sector_t target;

 		/* When clearing we round the start up and the end down.
@@ -1071,9 +1083,16 @@ static bool _badblocks_clear(struct badblocks *bb, sector_t s, sector_t sectors)
 		 * isn't than to think a block is not bad when it is.
 		 */
 		target = s + sectors;
-		s = round_up(s, 1 << bb->shift);
-		target = round_down(target, 1 << bb->shift);
-		sectors = target - s;
+		s = round_up(s, (sector_t)1 << bb->shift);
+		target = round_down(target, (sector_t)1 << bb->shift);
+		if (s < orig_s || target < s)
+			/* Rounding wrapped, or range collapsed */
+			sectors = 0;
+		else
+			sectors = target - s;
+
+		if (sectors == 0)
+			return false;
 	}

 	write_seqlock_irq(&bb->lock);
@@ -1303,12 +1322,19 @@ int badblocks_check(struct badblocks *bb, sector_t s, sector_t sectors,

 	WARN_ON(bb->shift < 0 || sectors == 0);

+	if ((s + sectors) < s)
+		/* Range wraps past the end of sector_t */
+		return 0;
+
 	if (bb->shift > 0) {
 		/* round the start down, and the end up */
 		sector_t target = s + sectors;

-		s = round_down(s, 1 << bb->shift);
-		target = round_up(target, 1 << bb->shift);
+		s = round_down(s, (sector_t)1 << bb->shift);
+		target = round_up(target, (sector_t)1 << bb->shift);
+		if (target <= s)
+			/* Rounding wrapped past the end of sector_t */
+			return 0;
 		sectors = target - s;
 	}

diff --git a/include/linux/badblocks.h b/include/linux/badblocks.h
index 996493917f3..5d88992b55a 100644
--- a/include/linux/badblocks.h
+++ b/include/linux/badblocks.h
@@ -34,7 +34,11 @@ struct badblocks {
 				 */
 	int shift;		/* shift from sectors to block size
 				 * a -ve shift means badblocks are
-				 * disabled.*/
+				 * disabled. Callers that set this from
+				 * untrusted/on-disk data are responsible
+				 * for bounding it so 1 << shift does not
+				 * overflow a sector_t.
+				 */
 	u64 *page;		/* badblock list */
 	int changed;
 	seqlock_t lock;
--
2.43.0