The kmalloc entry points are annotated with __assume_kmalloc_alignment
but return ZERO_SIZE_PTR, currently (void *)16, for zero-size requests.
This violates the annotation when ARCH_KMALLOC_MINALIGN exceeds 16.

This can mislead compiler optimizations. Current compilers retain the
ZERO_OR_NULL_PTR() range check, but GCC and Clang eliminate an exact
ZERO_SIZE_PTR comparison after an annotated allocation. Clang's UBSAN
also reports the failed alignment assumption on armv5.

Set ZERO_SIZE_PTR to ARCH_KMALLOC_MINALIGN when it exceeds 16 and keep
the existing value otherwise. Assert that the alignment does not exceed
128, which keeps the sentinel below PAGE_SIZE and the low pointer
poison values. Deriving the value from the architecture also avoids
changing architectures such as s390, where low absolute addresses are
valid.

Fixes: 94a58c360a45 ("slab.h: sprinkle __assume_aligned attributes")
Assisted-by: Claude:claude-fable-5
Signed-off-by: Karl Mehltretter <kmehltretter@gmail.com>
---

Notes:
    This version is based on the RFC discussion [1].
    
    [1] https://lore.kernel.org/r/20260712120728.96628-1-kmehltretter@gmail.com
    
    A tree-wide audit found no users depending on the numeric value of
    ZERO_SIZE_PTR.
    
    The static assertion follows the ARCH_KMALLOC_MINALIGN fallback, the
    first point where the macro is guaranteed to be defined.
    
    Tested on bcc44b6785f21:
    
      - armv5, armv7, mips64, ppc44x and sh4 with minimum alignments from
        32 to 128
      - x86_64, arm64, riscv64, s390x and m68k as unchanged controls
      - Clang 21 with CONFIG_UBSAN_ALIGNMENT on armv5 no longer shows the
        alignment assumption report
      - no text, data or bss size change on armv5 or arm64
      - the exact ZERO_SIZE_PTR comparison remains in armv5 code
        generation
    
    The sh4 test needed two unrelated workarounds, both reproduced on the
    unpatched tree: compiling tcp_output.c at -O1 because of a GCC 15.2
    hang, and disabling USB_OHCI_HCD because of an sm501-usb boot failure.

 include/linux/slab.h | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/include/linux/slab.h b/include/linux/slab.h
index 32c9f8ed7ae20..0798a714da87f 100644
--- a/include/linux/slab.h
+++ b/include/linux/slab.h
@@ -259,13 +259,16 @@ enum _slab_flag_bits {
 
 /*
  * ZERO_SIZE_PTR will be returned for zero sized kmalloc requests.
+ * It satisfies the alignment promised by __assume_kmalloc_alignment
+ * and keeps the historic value 16 where that is already aligned.
  *
  * Dereferencing ZERO_SIZE_PTR will lead to a distinct access fault.
  *
  * ZERO_SIZE_PTR can be passed to kfree though in the same way that NULL can.
  * Both make kfree a no-op.
  */
-#define ZERO_SIZE_PTR ((void *)16)
+#define ZERO_SIZE_PTR ((void *)(ARCH_KMALLOC_MINALIGN > 16 ? \
+				ARCH_KMALLOC_MINALIGN : 16))
 
 #define ZERO_OR_NULL_PTR(x) ((unsigned long)(x) <= \
 				(unsigned long)ZERO_SIZE_PTR)
@@ -622,6 +625,12 @@ static inline bool kmem_dump_obj(void *object) { return false; }
 #define KMALLOC_SHIFT_LOW ilog2(KMALLOC_MIN_SIZE)
 #endif
 
+/*
+ * Keep ZERO_SIZE_PTR below PAGE_SIZE and the low pointer poison values.
+ * 128 is the largest in-tree ARCH_KMALLOC_MINALIGN.
+ */
+static_assert(ARCH_KMALLOC_MINALIGN <= 128);
+
 /*
  * Setting ARCH_SLAB_MINALIGN in arch headers allows a different alignment.
  * Intended for arches that get misalignment faults even for 64 bit integer
-- 
2.53.0