Hi all, This patch series reworks the generic REFCOUNT_FULL implementation using atomic_fetch_* operations so that the performance impact of the cmpxchg() loops is mitigated for common refcount operations. The algorithm was heavily inspired by Ard's assembly implementation for arm64:
http://lkml.kernel.org/r/20190619105431.2630-1-ard.biesheu...@linaro.org but I figured we could achieve something similar using atomics in generic code. Although the revised implementation passes all of the lkdtm REFCOUNT tests, there is a race condition introduced by the deferred saturation whereby if INT_MIN + 2 tasks take a reference on a refcount at REFCOUNT_MAX and are each preempted between detecting overflow and writing the saturated value without being rescheduled, then another task may end up erroneously freeing the object when it drops the refcount and sees zero. It doesn't feel like a particularly realistic case to me, but I thought I should mention it in case somebody else knows better. Will Cc: Kees Cook <keesc...@chromium.org> Cc: Ingo Molnar <mi...@kernel.org> Cc: Elena Reshetova <elena.reshet...@intel.com> Cc: Peter Zijlstra <pet...@infradead.org> Cc: Ard Biesheuvel <ard.biesheu...@linaro.org> Cc: Hanjun Guo <guohan...@huawei.com> Cc: Jan Glauber <jglau...@marvell.com> --->8 Will Deacon (6): lib/refcount: Define constants for saturation and max refcount values lib/refcount: Ensure integer operands are treated as signed lib/refcount: Remove unused refcount_*_checked() variants lib/refcount: Move bulk of REFCOUNT_FULL implementation into header lib/refcount: Improve performance of generic REFCOUNT_FULL code lib/refcount: Consolidate REFCOUNT_{MAX,SATURATED} definitions drivers/misc/lkdtm/refcount.c | 8 -- include/linux/refcount.h | 236 +++++++++++++++++++++++++++++++++++++---- lib/refcount.c | 238 +----------------------------------------- 3 files changed, 219 insertions(+), 263 deletions(-) -- 2.11.0
