On Tue, Sep 23, 2025 at 05:32:16PM +0200, Jan Stancek wrote:
> We've been observing rare non-deterministic kconfig failures during
> olddefconfig, where ARCH_USING_PATCHABLE_FUNCTION_ENTRY was getting
> disabled and with it number of other config options that depend on it.
>
> The reason is that gcc-check-fpatchable-function-entry.sh can fail
> if `grep -q` (or scripts/dummy-tools/gcc) is fast enough to exit while
> there is still someone writing on other side of pipe. `pipefail`
> propagates that error up to kconfig.
>
> This can be seen for example with:
> # (set -e; set -o pipefail; yes | grep -q y); echo $?
> 141
>
> or by running the actual check script in loop extensively:
> ----------------------------- 8< -------------------------------
> function kconfig()
> {
> for i in `seq 1 100`; do
> arch/powerpc/tools/gcc-check-fpatchable-function-entry.sh \
> ./scripts/dummy-tools/gcc -mlittle-endian \
> || { echo "Oops"; exit 1; }
> done
> }
>
> for ((i=0; i<$(nproc); i++)); do kconfig & done
> wait; echo "Done"
> ----------------------------- >8 -------------------------------
>
> Fixes: 0f71dcfb4aef ("powerpc/ftrace: Add support for
> -fpatchable-function-entry")
> Fixes: b71c9ffb1405 ("powerpc: Add arch/powerpc/tools directory")
> Reported-by: Joe Lawrence <[email protected]>
> Acked-by: Joe Lawrence <[email protected]>
> Signed-off-by: Jan Stancek <[email protected]>
Ooh, that is nasty...
The fix looks obviously correct: pipefail isn't needed in these cases
anyway since an earlier pipe failure (e.g., compile error) would result
in EOF getting piped to grep, causing it to returning an error
regardless.
Reviewed-by: Josh Poimboeuf <[email protected]>
--
Josh
