Finding: diffstorm/cb, two defects (fixed upstream)
| Library | diffstorm/cb, MIT |
| Affected | up to commit 547d0812 (2025-07-18) |
| Fixed in | 91ae20b (2026-11-08); checked again at 11e05d6 (v1.3) |
| Defects | 1. the overwrite flag can switch itself on (all platforms); 2. the consumer lacks acquire ordering (multi-core targets) |
| Found by | the verified-libs stress harness and GenMC (RC11 model checking) |
Timeline
| Date | |
|---|---|
| 2026-11-06 | Found: a host reproducer and GenMC show both defects at 547d0812. |
| 2026-11-08 | Reported to the maintainer by email, with the reproducer (demo.c) and the GenMC harness (diffstorm_spsc.c); 90-day notice period. |
| 2026-11-08 | The maintainer pushed a fix the same day (91ae20b). Checked here the same evening: the reproducer gives the correct result, every access to overwrite is a 1-byte load or store on the Cortex-M33, and GenMC passes. |
| 2026-11-08 | Two further upstream commits (17ffdc1, 11e05d6, v1.3) checked the same way: still fixed. |
| 2026-11-09 | The maintainer agreed to publication before the 90 days were over. libs.lock pins 11e05d60; bench and board stress on it: 0 errors, control caught; GenMC run: all checks as expected. |
Thanks to the maintainer for fixing both defects within hours.
Defect 1: overwrite mode can switch itself on (all platforms)
overwrite is a 1-byte atomic_bool (cb.h, inside struct cb).
cb_init_ex() stores it with a 1-byte atomic store, but cb_insert_ex()
(cb.c:373), cb_insert_timeout (cb.c:499) and cb_get_overwrite_ex
(cb.c:597) read it with CB_ATOMIC_LOAD, which always casts to a 32-bit
atomic, and cb_set_overwrite_ex (cb.c:577) writes it with a 32-bit
store. The 3 padding bytes after overwrite are never initialised.
Consequence: if the memory holding a cb object was not zero before
cb_init() (a local variable, a reused buffer, malloc), the flag reads as
true although overwrite mode was never enabled. A full buffer then accepts
the insert, overwrites the oldest item, and the producer moves out, which
belongs to the consumer (not SPSC-safe).
Why CB_ATOMIC_LOAD is always 32-bit: cb_atomic_access.h selects the cast
with #if CB_ATOMIC_INDEX_TYPE == uint32_t. In a preprocessor expression both
type names are unknown identifiers and evaluate to 0, so the first branch is
always taken. The library already has the correct macro,
CB_ATOMIC_LOAD_OW (cb.c:74/82), but the lines above do not use it.
Evidence:
- Host demo (demo.c, x86-64, gcc):
zeroed memory: overwrite flag=0 insert into full buffer=rejected first item=1 -> correct
memory that held 0xAA: overwrite flag=1 insert into full buffer=accepted first item=2 -> WRONG
- Cortex-M33 (arm-none-eabi-gcc 14.2.1 -O2): overwrite is 1 byte at
offset 16 (padding 17–19); cb_insert_ex loads it with
ldr r4, [r0, #16] (32-bit); cb_init_ex stores it with stlb (8-bit).
- GenMC: "Mixed-size accesses detected: tried to read with a 32-bit access".
In our benchmark and stress tests the object is a zero-initialised static, so the defect did not show there.
Defect 2: consumer lacks acquire ordering (multi-core targets)
cb_remove_ex() loads in with a relaxed atomic load and then reads the
item from the buffer; the only barrier comes after the read (release side).
Under the C11 memory model that is a data race on the buffer: on a
weakly ordered multi-core CPU (the library lists aarch64, Apple Silicon and
RISC-V among its targets) the consumer can read the slot before the
producer's write is visible.
Evidence (GenMC 0.20, RC11, 1 producer + 1 consumer, 4 items, 4 slots;
CB_MEMORY_BARRIER mapped to atomic_thread_fence(memory_order_seq_cst),
the usual C11 mapping of dmb; finding 1 neutralised with
CB_ATOMIC_LOAD_OW so it does not mask this one):
| Variant | Result |
|---|---|
| overwrite load fixed only | fail: non-atomic race on the buffer |
+ one CB_MEMORY_BARRIER() after the two index loads in cb_remove_ex |
pass, 686 executions |
Not a problem for interrupt ↔ main on one Cortex-M core (sequential on one core); relevant for cross-core use.
How it was fixed upstream (91ae20b)
- Every access to
overwritenow goes throughCB_ATOMIC_LOAD_OW/CB_ATOMIC_STORE_OW, which use the flag's own 1-byte type. - In
cb_remove_exandcb_peek_exthe consumer loadsoutwith acquire semantics and has aCB_MEMORY_BARRIER()after the index loads, before it reads the item; the producer side storesinwith release semantics.
Reproduce
On lib-lab1 with the pinned toolchain image: bash run.sh here (defect 1,
host and Cortex-M33, against the pre-fix pin diffstorm-cb-547d0812) and
bash ../../genmc/run.sh (both defects at 547d0812; the fixed 11e05d6
must pass).