Re: [PATCH v7 0/6] selftests/mm: separate GUP microbenchmarking from functional testing

On 13/08/2026 7:12 pm, Sarthak Sharma wrote:
> gup_test.c currently serves two separate purposes: benchmarking
> (GUP_FAST_BENCHMARK, PIN_FAST_BENCHMARK and PIN_LONGTERM_BENCHMARK) and
> functional testing (GUP_BASIC_TEST, PIN_BASIC_TEST and
> DUMP_USER_PAGES_TEST). Keeping both in one program makes the functional
> tests harder to run and report individually, while run_vmtests.sh has to
> invoke the program repeatedly with different options.
> 
> Separate these roles into tools/mm/gup_bench for benchmarking and
> tools/testing/selftests/mm/gup for functional testing. Move the shared
> file and hugepage helpers to tools/lib/mm/ so both programs can use them
> without duplicating the implementation.
> 
> Patch 1 makes read_file(), write_file(), read_num(), write_num() and
> write_num_ignore_einval() return errors to their callers instead of
> exiting. It also makes read_num() reject negative and malformed values
> and updates the existing callers to handle failures.
> 
> Patch 2 moves these file helpers from vm_util.c to tools/lib/mm/. It keeps
> them available to the mm selftests through vm_util.h and adjusts the
> selftests build accordingly.
> 
> Patch 3 moves hugepage_settings.[ch] from selftests/mm to tools/lib/mm/.
> It also removes its kselftest dependency while preserving TAP-compatible
> diagnostics for selftest users.
> 
> Patch 4 moves the existing gup_test implementation from selftests/mm to
> tools/mm as gup_bench. This keeps the code movement separate from the
> subsequent changes and makes it easier to review.
> 
> Patch 5 removes the functional test modes and kselftest dependency from
> gup_bench. When run without arguments, it performs one GUP_FAST benchmark
> using the existing defaults instead of running the whole matrix. Other
> benchmark configurations can be selected through command-line options.
> 
> Patch 6 adds a new harness-based GUP selftest. It covers THP, non-THP and
> HugeTLB mappings across private/shared and read/write variants. For each
> variant, it tests get_user_pages(), get_user_pages_fast(),
> pin_user_pages(), pin_user_pages_fast(), long-term pinning and both dump
> modes using four batch sizes. It also preserves the existing sparse dump
> coverage for pages 0, 19 and 0x1000. The HugeTLB variants share a one-time
> setup of 128 default-sized huge pages.

I tested the full MM selftest suite before and after this series and found
no regressions. The new GUP selftest and gup_bench also passed.

Tested-by: Muhammad Usama Anjum <[email protected]>

Thanks,
Usama