On Tue, May 26, 2026 at 05:38:29PM +0530, Sarthak Sharma wrote: > On 5/26/26 2:31 PM, Mike Rapoport wrote: > > On Mon, May 25, 2026 at 11:59:32AM +0530, Sarthak Sharma wrote: > >> On 5/24/26 10:36 PM, Mike Rapoport wrote: > >>> > >>> and while EXIT_FAILURE == KSFT_FAIL I'm not sure it's robust enough. > >> > >> I used EXIT_FAILURE here because the helper is moving out of selftests > >> and should not include kselftest.h anymore. The helper already > >> terminated the process on these paths, so I tried to preserve that > >> behavior while removing the ksft dependency. > > > > In mm selftests a failure to update a /proc or /sysfs file meant there is > > no point to continue the test. But if we make it a generic helper for > > potentially broader use than mm selftests, exit() on failure is too harsh. > > Okay yeah, this makes sense. > > > > >> We can change this to return errors instead of calling exit() and update > >> the selftest callers to report failures through the ksft_* helpers. I > >> agree this is cleaner, but it would grow the series a bit. > >> > >> If you feel strongly, I can include these changes in v4. Otherwise I > >> feel we can handle it separately later to avoid growing this series. > > > > There are not that many callers of write_file() and write_num(). > > I think a patch that makes them return an error rather than exit() can go > > before moving these functions to lib. > > > > So I will add a patch before the move that makes read_file(), > write_file(), read_num() and write_num() return errors instead of > exiting and update the existing selftest callers to report those > failures via ksft_* helpers. > > For hugepage_settings.c, I’d prefer to keep the existing fail fast > behaviour in this series. After this series, the users are still mm > selftests and the new tools/mm/gup_bench tool and for those users a > failure to read/write THP or HugeTLB state is fatal to the operation > being attempted. > > Converting the full hugepage_settings API to return errors would be a > larger follow-up, because many of its helpers are used throughout mm > selftests. I can handle that in a separate series unless you think it > should be folded into this one as well.
Keeping fail fast in hugepage_settings for now makes sense, they are not as generic as write_file(). -- Sincerely yours, Mike.
