Bug #21569: [armv7, musl] SIGBUS in ibf_load_object_float due to unaligned VFP double load when reading IBF

• Status: Closed
• Assignee: -
• Target version: -
• ruby -v: ruby 3.3.8 (2025-04-09 revision b200bad6cd) [armv7a-linux-musleabihf]
• Backport: 3.2: WONTFIX, 3.3: DONE, 3.4: DONE

[amacxz (Aleksey Maximov)]

Environment:
CPU: ARMv7-A (NVIDIA Tegra 2), VFPv3-D16, no NEON
OS/libc: Linux, musl (ld-musl-armhf.so.1)
Compiler: GCC 14.3.0
Ruby: 3.3.8 (built from source via Gentoo ebuild)
CFLAGS actually used by system: "-Os -pipe -march=armv7-a -mfpu=vfpv3-d16 -mfloat-abi=hard"

During make install (or Gentoo’s ebuild install phase) Ruby runs:

./miniruby -I./lib -I. -I.ext/common ./tool/runruby.rb --extout=.ext -- \
  --disable-gems -r./armv7a-linux-musleabihf-fake ./tool/rbinstall.rb \
  --make=make --dest-dir="$D" --extout=.ext --ext-build-dir=./ext \
  --mflags="-j1" --make-flags=" V=1" --gnumake --install=all --exclude=doc

This reliably triggers a SIGBUS on armv7 hard-float.

Observed crash:

Thread "ruby33" received signal SIGBUS (Bus error).
#0  ibf_load_object_float () from libruby33.so.3.3
(gdb) bt
#0  ibf_load_object_float
#1  ibf_load_object
#2  rb_ibf_load_iseq_complete
#3  ibf_load_iseq
#4  ...
(gdb) info reg
r0 = 0xb6f508b6  (not 8-byte aligned)
pc = 0xb6cacf78 <ibf_load_object_float+32>
(gdb) x/6i $pc-8
   ...
   0xb6cacf74: vldr d0, [r0]   <-- VFP double load from unaligned addr → SIGBUS

Root cause
In compile.c:

static VALUE
ibf_load_object_float(const struct ibf_load *load, const struct ibf_object_header *header, ibf_offset_t offset)
{
    const double *dblp = IBF_OBJBODY(double, offset);
    return DBL2NUM(*dblp);
}

IBF_OBJBODY(double, ...) may return an unaligned pointer. On ARMv7, VFP vldr with an unaligned double address raises SIGBUS (no kernel fixup). Hence the crash while loading IBF.

Proposed fix
Read into an aligned local with memcpy:

--- a/compile.c
+++ b/compile.c
@@ -12921,10 +12921,12 @@ static VALUE
 ibf_load_object_float(const struct ibf_load *load, const struct ibf_object_header *header, ibf_offset_t offset)
 {
-    const double *dblp = IBF_OBJBODY(double, offset);
-    return DBL2NUM(*dblp);
+    /* IBF buffer may be unaligned on some platforms. On ARMv7, a VFP
+     * double load from an unaligned address causes SIGBUS. */
+    double d;
+    memcpy(&d, IBF_OBJBODY(void, offset), sizeof(d));
+    return DBL2NUM(d);
 }

Notes:
The issue reproduces consistently on Tegra2 (armv7a, vfpv3-d16, no NEON) with musl, but (IMO) conceptually applies to any strict-alignment ARMv7 platform.
A similar audit may be required for other IBF loaders reading 8-byte types.
Please review and merge the fix (or implement an equivalent alignment-safe read for IBF floats).
I can test any proposed patch on this hardware.

[amacxz (Aleksey Maximov)]

• File compile_and_debug_log.txt compile_and_debug_log.txt added

[nobu (Nobuyoshi Nakada)]

Thank you for the report.

amacxz (Aleksey Maximov) wrote:

+    memcpy(&d, IBF_OBJBODY(void, offset), sizeof(d));

I think void should be double since IBF_OBJBODY won't work on void.
At least the fallback definition, offsetof(struct { char _; T t; }, t), is invalid where T is void.

[amacxz (Aleksey Maximov)]

hmmm

Good catch — you’re right that IBF_OBJBODY(void,) is invalid with the fallback definition.
I’ve updated the patch. I’ll test it today and report back.

--- a/compile.c
+++ b/compile.c
@@ -12921,10 +12921,12 @@ static VALUE
 ibf_load_object_float(const struct ibf_load *load, const struct ibf_object_header *header, ibf_offset_t offset)
 {
-    const double *dblp = IBF_OBJBODY(double, offset);
-    return DBL2NUM(*dblp);
+    const double *dblp = IBF_OBJBODY(double, offset);
+    /* IBF buffer may be unaligned; loading a double directly (VFP vldr)
+     * from an unaligned address causes SIGBUS on armv7. */
+    double d;
+    memcpy(&d, dblp, sizeof d);
+    return DBL2NUM(d);
 }

[alanwu (Alan Wu)]

+ const double *dblp = IBF_OBJBODY(double, offset);

This should probably be const void *dblp = .... Merely creating an unaligned pointer triggers undefined behavior.

[nobu (Nobuyoshi Nakada)]

Simply it can be:

memcpy(&d, IBF_OBJBODY(double, offset), sizeof(d));

amacxz: Would you send PR at github, or post the formatted-patch?
Otherwise we can just add "patched by" comment to the commit log.

[amacxz (Aleksey Maximov)]

• File 030-ibf-fix-unaligned-float-load-on-armv7.patch 030-ibf-fix-unaligned-float-load-on-armv7.patch added

Successfully compiled on ARMv7 with this patch (right now); no crash occurred during the project build.
ruby 3.3.8 (2025-04-09 revision b200bad6cd) [armv7a-linux-musleabihf]

Adding a ‘Patched-by: Aleksey Maximov [e-mail]’ tag will be perfectly sufficient.

Thanks for the help. :)

[alanwu (Alan Wu)]

• Backport changed from 3.2: UNKNOWN, 3.3: UNKNOWN, 3.4: UNKNOWN to 3.2: WONTFIX, 3.3: REQUIRED, 3.4: REQUIRED

[amacxz (Aleksey Maximov)]

• Status changed from Open to Closed

Applied in changeset git|354d47ae5bc4edcc94db4a5391ed71a8b9844e57.

---

IBF: Avoid unaligned load on 32 bit platforms

[Bug #21569]

[k0kubun (Takashi Kokubun)]

• Backport changed from 3.2: WONTFIX, 3.3: REQUIRED, 3.4: REQUIRED to 3.2: WONTFIX, 3.3: REQUIRED, 3.4: DONE

ruby_3_4 22c2262b83224f6798d60d539f1f8609d9722766 merged revision(s) 354d47ae5bc4edcc94db4a5391ed71a8b9844e57.

[nagachika (Tomoyuki Chikanaga)]

• Backport changed from 3.2: WONTFIX, 3.3: REQUIRED, 3.4: DONE to 3.2: WONTFIX, 3.3: DONE, 3.4: DONE

ruby_3_3 62ecd47656e0c8c7f308fc798ab6106d738c211e merged revision(s) 354d47ae5bc4edcc94db4a5391ed71a8b9844e57.