> From: Stephen Hemminger [mailto:[email protected]] > Sent: Friday, 4 September 2026 01.05 > > From: Scott Mitchell <[email protected]> > > RTE_PTR_ADD and RTE_PTR_SUB APIs have a few limitations: > 1. ptr cast to uintptr_t drops pointer provenance and > prevents compiler optimizations > 2. return cast discards qualifiers (const, volatile) > which may hide correctness/concurrency issues. > 3. Accepts both "pointers" and "integers as pointers" which > overloads the use case and constrains the implementation > to address other challenges. > > This patch deprecates support for integer types which allows > addressing each of the challenges above. > > Examples: > 1. Clang is able to optimize and improve __rte_raw_cksum > (which uses RTE_PTR_ADD) by ~40% (100 bytes) to ~8x (1.5k bytes) > TSC cycles/byte. > 2. Refactoring discovered cases that dropped qualifiers (volatile) > that the new API exposes. > > Signed-off-by: Scott Mitchell <[email protected]> > Reviewed-by: Stephen Hemminger <[email protected]> > --- > Changes from the original v24 submission (rebased to 26.11): > > * Rebased on main: release notes moved to release_26_11.rst, > and roc_cpt_debug.c updated for the cpt_cn10k_frag_info_s rename. > > * Release notes: state that RTE_PTR_ADD/SUB reject integer arguments > while RTE_PTR_ALIGN* still compile with them as deprecated usage. > There are still some in-tree uses of integer arguments in mlx5. > > * test_common.c: aligned uint16/32/64 dereference tests now also > require increment to be a multiple of the type size, avoiding > misaligned typed access. > > * ena: BAR NULL check logs at DEBUG rather than ERR since the memory > BAR is legitimately absent on non-LLQ devices. > > * idxd: drop RTE_PTR_UNQUAL and FIXME on portal assignment, and the > redundant (volatile void *) casts, since the fields are already > volatile-qualified upstream. > > * eal_common_memory.c: drop unrelated volatile qualifier on local elem > pointer. > * eal_common_fbarray.c: drop unused rte_debug.h include. > * eal_memalloc.c (linux, windows): move base_va NULL check before > the allocation loop and return -1 instead of goto out. > > * rte_mbuf.h, rte_mempool.h: drop __rte_assume() on pointers that > are dereferenced immediately after; compiler already infers that > form the immediate dereference. keep RTE_ASSERT. >
That's a lot of code. Thank you for working on this, Scott and Stephen. Stronger typing is a valuable improvement. I have reviewed selectively, so it's not a full review, but a big ack. Acked-by: Morten Brørup <[email protected]>
