On Mon, Oct 20, 2025 at 12:02:01PM +0000, Morten Brørup wrote: > Refactored rte_pktmbuf_prefree_seg() for both performance and readability. > > With the optimized RTE_MBUF_DIRECT() macro, the common likely code path > now fits within one instruction cache line on x86-64 when built with GCC. > > Signed-off-by: Morten Brørup <[email protected]>
Reviewed-by: Bruce Richardson <[email protected]> Comments inline below. > --- > v2: > * Fixed typo in commit description. > * Fixed indentation. > * Added detailed description to the optimized RTE_MBUF_DIRECT() macro. > (Stephen Hemminger) > * Added static_assert() to verify that the optimized RTE_MBUF_DIRECT() > macro is valid, specifically that the tested bits are in the MSB of the > 64-bit field. > --- > lib/mbuf/rte_mbuf.h | 51 +++++++++++++++------------------------- > lib/mbuf/rte_mbuf_core.h | 27 +++++++++++++++++++++ > 2 files changed, 46 insertions(+), 32 deletions(-) > > diff --git a/lib/mbuf/rte_mbuf.h b/lib/mbuf/rte_mbuf.h > index 3df22125de..2004391f57 100644 > --- a/lib/mbuf/rte_mbuf.h > +++ b/lib/mbuf/rte_mbuf.h > @@ -31,6 +31,7 @@ > * http://www.kohala.com/start/tcpipiv2.html > */ > > +#include <stdbool.h> > #include <stdint.h> > > #include <rte_common.h> > @@ -1458,44 +1459,30 @@ static inline int > __rte_pktmbuf_pinned_extbuf_decref(struct rte_mbuf *m) > static __rte_always_inline struct rte_mbuf * > rte_pktmbuf_prefree_seg(struct rte_mbuf *m) > { > - __rte_mbuf_sanity_check(m, 0); > - > - if (likely(rte_mbuf_refcnt_read(m) == 1)) { > - > - if (!RTE_MBUF_DIRECT(m)) { > - rte_pktmbuf_detach(m); > - if (RTE_MBUF_HAS_EXTBUF(m) && > - RTE_MBUF_HAS_PINNED_EXTBUF(m) && > - __rte_pktmbuf_pinned_extbuf_decref(m)) > - return NULL; > - } > - > - if (m->next != NULL) > - m->next = NULL; > - if (m->nb_segs != 1) > - m->nb_segs = 1; > + bool refcnt_not_one; > > - return m; > + __rte_mbuf_sanity_check(m, 0); > > - } else if (__rte_mbuf_refcnt_update(m, -1) == 0) { > + refcnt_not_one = unlikely(rte_mbuf_refcnt_read(m) != 1); > + if (refcnt_not_one && __rte_mbuf_refcnt_update(m, -1) != 0) > + return NULL; > > - if (!RTE_MBUF_DIRECT(m)) { > - rte_pktmbuf_detach(m); > - if (RTE_MBUF_HAS_EXTBUF(m) && > - RTE_MBUF_HAS_PINNED_EXTBUF(m) && > - __rte_pktmbuf_pinned_extbuf_decref(m)) > - return NULL; > - } > + if (unlikely(!RTE_MBUF_DIRECT(m))) { > + rte_pktmbuf_detach(m); > + if (RTE_MBUF_HAS_EXTBUF(m) && > + RTE_MBUF_HAS_PINNED_EXTBUF(m) && > + __rte_pktmbuf_pinned_extbuf_decref(m)) > + return NULL; > + } > > - if (m->next != NULL) > - m->next = NULL; > - if (m->nb_segs != 1) > - m->nb_segs = 1; > + if (refcnt_not_one) > rte_mbuf_refcnt_set(m, 1); > + if (m->nb_segs != 1) > + m->nb_segs = 1; > + if (m->next != NULL) > + m->next = NULL; > > - return m; > - } > - return NULL; > + return m; > } > Nice refactor, much more readable, thanks. > /** > diff --git a/lib/mbuf/rte_mbuf_core.h b/lib/mbuf/rte_mbuf_core.h > index a0df265b5d..41f40e1967 100644 > --- a/lib/mbuf/rte_mbuf_core.h > +++ b/lib/mbuf/rte_mbuf_core.h > @@ -715,6 +715,33 @@ struct rte_mbuf_ext_shared_info { > #define RTE_MBUF_DIRECT(mb) \ > (!((mb)->ol_flags & (RTE_MBUF_F_INDIRECT | RTE_MBUF_F_EXTERNAL))) > > +#if defined(RTE_TOOLCHAIN_GCC) && defined(RTE_ARCH_X86) > +/* Optimization for code size. > + * GCC only optimizes single-bit MSB tests this way, so we do it by hand > with multi-bit. > + * > + * The flags RTE_MBUF_F_INDIRECT and RTE_MBUF_F_EXTERNAL are both in the MSB > of the > + * 64-bit ol_flags field, so we only compare this one byte instead of all 64 > bits. > + * On little endian architecture, the MSB of a 64-bit integer is at byte > offest 7. > + * > + * Note: Tested using GCC version 16.0.0 20251019 (experimental). > + * > + * Without this optimization, GCC generates 17 bytes of instructions: > + * movabs rax,0x6000000000000000 // 10 bytes > + * and rax,QWORD PTR [rdi+0x18] // 4 bytes > + * sete al // 3 bytes > + * With this optimization, GCC generates only 7 bytes of instructions: > + * test BYTE PTR [rdi+0x1f],0x60 // 4 bytes > + * sete al // 3 bytes > + */ > +#undef RTE_MBUF_DIRECT > +#define RTE_MBUF_DIRECT(mb) \ > + (!(((const uint8_t *)(mb))[offsetof(struct rte_mbuf, ol_flags) + 7] & \ > + (uint8_t)((RTE_MBUF_F_INDIRECT | RTE_MBUF_F_EXTERNAL) >> (7 * 8)))) > +static_assert(((RTE_MBUF_F_INDIRECT | RTE_MBUF_F_EXTERNAL) >> (7 * 8)) << (7 > * 8) == > + (RTE_MBUF_F_INDIRECT | RTE_MBUF_F_EXTERNAL), > + "RTE_MBUF_F_INDIRECT and/or RTE_MBUF_F_EXTERNAL are not in MSB."); > +#endif > + Couple of comments/thoughts/questions here. * This looks like a compiler limitation that should be fixed in GCC. IF we put this optimization in, how will we know when/if we can remove it again in future? I'm not sure we want this hanging around forever. * Can the static_assert - which just checks flags are in the MSB - be * simplified to e.g. "((RTE_MBUF_F_INDIRECT | RTE_MBUF_F_EXTERNAL) << CHAR_BIT) == 0" or "__builtin_ctzll(...) > (7 * CHAR_BIT)" * As in prev bullet, I tend to prefer use of CHAR_BIT over hard-coded 8. * Is it necessary to limit this to just GCC and x86? If it leads to the best code on x86, why not include for all compilers? What about non-x86 LE platforms? * Does the actual macro need to be that long and complex? If we simplify a bit, does the compiler go back to generating bad code? For example: using "(mb->ol_flags >> 56) & ((RTE_MBUF_F_INDIRECT | ..) >> 56)" * If the above is true, do we need to actually put this in in assembler to guarantee compiler generates good code in all situations?
