| Message ID | 20241021-arm-kasan-vmalloc-crash-v4-0-837d1294344f@linaro.org (mailing list archive) |
|---|---|
| Headers | show |
| Series | Fix KASAN crash when using KASAN_VMALLOC | expand |
On Mon, Oct 21, 2024 at 3:03 PM Linus Walleij <linus.walleij@linaro.org> wrote: > This problem reported by Clement LE GOFFIC manifest when > using CONFIG_KASAN_IN_VMALLOC and VMAP_STACK: > https://lore.kernel.org/linux-arm-kernel/a1a1d062-f3a2-4d05-9836-3b098de9db6d@foss.st.com/ > > After some analysis it seems we are missing to sync the > VMALLOC shadow memory in top level PGD to all CPUs. > > Add some code to perform this sync, and the bug appears > to go away. > > As suggested by Ard, also perform a dummy read from the > shadow memory of the new VMAP_STACK in the low level > assembly. > > Signed-off-by: Linus Walleij <linus.walleij@linaro.org> As these are regressions that need to go in as fixes I'm putting them into Russell's patch tracker now. The 9427/1 patch: https://www.arm.linux.org.uk/developer/patches/viewpatch.php?id=9427/1 Need to be avoided as it causes build regressions. Patch 1/3 supersedes it. Yours, Linus Walleij
This problem reported by Clement LE GOFFIC manifest when using CONFIG_KASAN_IN_VMALLOC and VMAP_STACK: https://lore.kernel.org/linux-arm-kernel/a1a1d062-f3a2-4d05-9836-3b098de9db6d@foss.st.com/ After some analysis it seems we are missing to sync the VMALLOC shadow memory in top level PGD to all CPUs. Add some code to perform this sync, and the bug appears to go away. As suggested by Ard, also perform a dummy read from the shadow memory of the new VMAP_STACK in the low level assembly. Signed-off-by: Linus Walleij <linus.walleij@linaro.org> --- Changes in v4: - Since Kasan is not using header stubs, it is necessary to avoid kasan_*() calls using ifdef when compiling without KASAN. - Lift a line aligning the end of vmalloc from Melon Liu's very similar patch so we have feature parity, credit Melon as co-developer. - Include the atomic_read_acquire() patch in the series due to context requirements. - Verify that the after the patch the kernel still builds and boots without Kasan. - Link to v3: https://lore.kernel.org/r/20241017-arm-kasan-vmalloc-crash-v3-0-d2a34cd5b663@linaro.org Changes in v3: - Collect Mark Rutlands ACK on patch 1 - Change the simplified assembly add r2, ip, lsr #n to the canonical add r2, r2, ip, lsr #n in patch 2. - Link to v2: https://lore.kernel.org/r/20241016-arm-kasan-vmalloc-crash-v2-0-0a52fd086eef@linaro.org Changes in v2: - Implement the two helper functions suggested by Russell making the KASAN PGD copying less messy. - Link to v1: https://lore.kernel.org/r/20241015-arm-kasan-vmalloc-crash-v1-0-dbb23592ca83@linaro.org --- Linus Walleij (3): ARM: ioremap: Sync PGDs for VMALLOC shadow ARM: entry: Do a dummy read from VMAP shadow mm: Pair atomic_set_release() with _read_acquire() arch/arm/kernel/entry-armv.S | 8 ++++++++ arch/arm/mm/ioremap.c | 35 ++++++++++++++++++++++++++++++----- 2 files changed, 38 insertions(+), 5 deletions(-) --- base-commit: 9852d85ec9d492ebef56dc5f229416c925758edc change-id: 20241015-arm-kasan-vmalloc-crash-fcbd51416457 Best regards,