arm64: kaslr: don't pretend KASLR is enabled if offset < MIN_KIMG_ALIGN

Our virtual KASLR displacement is a randomly chosen multiple of
2 MiB plus an offset that is equal to the physical placement modulo 2
MiB. This arrangement ensures that we can always use 2 MiB block
mappings (or contiguous PTE mappings for 16k or 64k pages) to map the
kernel.

This means that a KASLR offset of less than 2 MiB is simply the product
of this physical displacement, and no randomization has actually taken
place. Currently, we use 'kaslr_offset() > 0' to decide whether or not
randomization has occurred, and so we misidentify this case.

If the kernel image placement is not randomized, modules are allocated
from a dedicated region below the kernel mapping, which is only used for
modules and not for other vmalloc() or vmap() calls.

When randomization is enabled, the kernel image is vmap()'ed randomly
inside the vmalloc region, and modules are allocated in the vicinity of
this mapping to ensure that relative references are always in range.
However, unlike the dedicated module region below the vmalloc region,
this region is not reserved exclusively for modules, and so ordinary
vmalloc() calls may end up overlapping with it. This should rarely
happen, given that vmalloc allocates bottom up, although it cannot be
ruled out entirely.

The misidentified case results in a placement of the kernel image within
2 MiB of its default address. However, the logic that randomizes the
module region is still invoked, and this could result in the module
region overlapping with the start of the vmalloc region, instead of
using the dedicated region below it. If this happens, a single large
vmalloc() or vmap() call will use up the entire region, and leave no
space for loading modules after that.

Since commit 82046702 ("efi/libstub/arm64: Replace 'preferred'
offset with alignment check"), this is much more likely to occur on
systems that boot via EFI but lack an implementation of the EFI RNG
protocol, as in that case, the EFI stub will decide to leave the image
where it found it, and the EFI firmware uses 64k alignment only.

Fix this, by correctly identifying the case where the virtual
displacement is a result of the physical displacement only.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Mark Brown <broonie@kernel.org>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lore.kernel.org/r/20230223204101.1500373-1-ardb@kernel.orgSigned-off-by: Catalin Marinas <catalin.marinas@arm.com>

arch/arm64/include/asm/memory.h

@@ -180,6 +180,7 @@
 #include <linux/compiler.h>
 #include <linux/mmdebug.h>
 #include <linux/types.h>
+#include <asm/boot.h>
 #include <asm/bug.h>
 
 #if VA_BITS > 48
@@ -203,6 +204,16 @@ static inline unsigned long kaslr_offset(void)
 	return kimage_vaddr - KIMAGE_VADDR;
 }
 
+static inline bool kaslr_enabled(void)
+{
+	/*
+	 * The KASLR offset modulo MIN_KIMG_ALIGN is taken from the physical
+	 * placement of the image rather than from the seed, so a displacement
+	 * of less than MIN_KIMG_ALIGN means that no seed was provided.
+	 */
+	return kaslr_offset() >= MIN_KIMG_ALIGN;
+}
+
 /*
  * Allow all memory at the discovery stage. We will clip it later.
  */

arch/arm64/kernel/cpufeature.c

@@ -1633,7 +1633,7 @@ bool kaslr_requires_kpti(void)
 			return false;
 	}
 
-	return kaslr_offset() > 0;
+	return kaslr_enabled();
 }
 
 static bool __meltdown_safe = true;

arch/arm64/kernel/kaslr.c

@@ -41,7 +41,7 @@ static int __init kaslr_init(void)
 		return 0;
 	}
 
-	if (!kaslr_offset()) {
+	if (!kaslr_enabled()) {
 		pr_warn("KASLR disabled due to lack of seed\n");
 		return 0;
 	}