ping for additional reviewers.
thanks!
On Tue, Aug 22, 2023 at 02:00:39PM -0700, Tyler Retzlaff wrote:
> This series introduces API additions prefixed in the rte namespace that allow
> the optional use of stdatomics.h from C11 using enable_stdatomics=true for
> targets where enable_stdatomics=false no functional change is intended.
>
> Be aware this does not contain all changes to use stdatomics across the DPDK
> tree it only introduces the minimum to allow the option to be used which is
> a pre-requisite for a clean CI (probably using clang) that can be run
> with enable_stdatomics=true enabled.
>
> It is planned that subsequent series will be introduced per lib/driver as
> appropriate to further enable stdatomics use when enable_stdatomics=true.
>
> Notes:
>
> * Additional libraries beyond EAL make visible atomics use across the
> API/ABI surface they will be converted in the subsequent series.
>
> * The eal: add rte atomic qualifier with casts patch needs some discussion
> as to whether or not the legacy rte_atomic APIs should be converted to
> work with enable_stdatomic=true right now some implementation dependent
> casts are used to prevent cascading / having to convert too much in
> the intial series.
>
> * Windows will obviously need complete conversion of libraries including
> atomics that are not crossing API/ABI boundaries. those conversions will
> introduced in separate series as new along side the existing msvc series.
>
> Please keep in mind we would like to prioritize the review / acceptance of
> this patch since it needs to be completed in the 23.11 merge window.
>
> Thank you all for the discussion that lead to the formation of this series.
>
> v6:
> * Adjust checkpatches to warn about use of __rte_atomic_thread_fence
> and suggest use of rte_atomic_thread_fence. Use the existing check
> more generic check for __atomic_xxx to catch use of __atomic_thread_fence
> and recommend rte_atomic_xxx.
>
> v5:
> * Add RTE_ATOMIC to doxygen configuration PREDEFINED macros list to
> fix documentation generation failure
> * Fix two typos in expansion of C11 atomics macros strong -> weak and
> add missing _explicit
> * Adjust devtools/checkpatches messages based on feedback. i have chosen
> not to try and catch use of C11 atomics or _Atomic since using those
> directly will be picked up by existing CI passes where by compilation
> error where enable_stdatomic=false (the default for most platforms)
>
> v4:
> * Move the definition of #define RTE_ATOMIC(type) to patch 1 where it
> belongs (a mistake in v3)
> * Provide comments for both RTE_ATOMIC and __rte_atomic macros indicating
> their use as specified or qualified contexts.
>
> v3:
> * Remove comments from APIs mentioning the mapping to C++ memory model
> memory orders
> * Introduce and use new macro RTE_ATOMIC(type) to be used in contexts
> where _Atomic is used as a type specifier to declare variables. The
> macro allows more clarity about what the atomic type being specified
> is. e.g. _Atomic(T *) vs _Atomic(T) it is easier to understand that
> the former is an atomic pointer type and the latter is an atomic
> type. it also has the benefit of (in the future) being interoperable
> with c++23 syntactically
> note: Morten i have retained your 'reviewed-by' tags if you disagree
> given the changes in the above version please indicate as such but
> i believe the changes are in the spirit of the feedback you provided
>
> v2:
> * Wrap meson_options.txt option description to newline and indent to
> be consistent with other options.
> * Provide separate typedef of rte_memory_order for enable_stdatomic=true
> VS enable_stdatomic=false instead of a single typedef to int
> note: slight tweak to reviewers feedback i've chosen to use a typedef
> for both enable_stdatomic={true,false} (just seemed more consistent)
> * Bring in assert.h and use static_assert macro instead of _Static_assert
> keyword to better interoperate with c/c++
> * Directly include rte_stdatomic.h where into other places it is consumed
> instead of hacking it globally into rte_config.h
> * Provide and use __rte_atomic_thread_fence to allow conditional expansion
> within the body of existing rte_atomic_thread_fence inline function to
> maintain per-arch optimizations when enable_stdatomic=false
>
> Tyler Retzlaff (6):
> eal: provide rte stdatomics optional atomics API
> eal: adapt EAL to present rte optional atomics API
> eal: add rte atomic qualifier with casts
> distributor: adapt for EAL optional atomics API changes
> bpf: adapt for EAL optional atomics API changes
> devtools: forbid new direct use of GCC atomic builtins
>
> app/test/test_mcslock.c | 6 +-
> config/meson.build | 1 +
> devtools/checkpatches.sh | 12 +-
> doc/api/doxy-api.conf.in | 1 +
> lib/bpf/bpf_pkt.c | 6 +-
> lib/distributor/distributor_private.h | 2 +-
> lib/distributor/rte_distributor_single.c | 44 +++----
> lib/eal/arm/include/rte_atomic_32.h | 4 +-
> lib/eal/arm/include/rte_atomic_64.h | 36 +++---
> lib/eal/arm/include/rte_pause_64.h | 26 ++--
> lib/eal/arm/rte_power_intrinsics.c | 8 +-
> lib/eal/common/eal_common_trace.c | 16 +--
> lib/eal/include/generic/rte_atomic.h | 67 +++++++----
> lib/eal/include/generic/rte_pause.h | 50 ++++----
> lib/eal/include/generic/rte_rwlock.h | 48 ++++----
> lib/eal/include/generic/rte_spinlock.h | 20 ++--
> lib/eal/include/meson.build | 1 +
> lib/eal/include/rte_mcslock.h | 51 ++++----
> lib/eal/include/rte_pflock.h | 25 ++--
> lib/eal/include/rte_seqcount.h | 19 +--
> lib/eal/include/rte_stdatomic.h | 198
> +++++++++++++++++++++++++++++++
> lib/eal/include/rte_ticketlock.h | 43 +++----
> lib/eal/include/rte_trace_point.h | 5 +-
> lib/eal/loongarch/include/rte_atomic.h | 4 +-
> lib/eal/ppc/include/rte_atomic.h | 54 ++++-----
> lib/eal/riscv/include/rte_atomic.h | 4 +-
> lib/eal/x86/include/rte_atomic.h | 8 +-
> lib/eal/x86/include/rte_spinlock.h | 2 +-
> lib/eal/x86/rte_power_intrinsics.c | 7 +-
> meson_options.txt | 2 +
> 30 files changed, 501 insertions(+), 269 deletions(-)
> create mode 100644 lib/eal/include/rte_stdatomic.h
>
> --
> 1.8.3.1
