On Fri, Apr 08, 2022 at 05:01:50PM +0300, Dmitry Kozlyuk wrote:
> 2022-04-01 06:29 (UTC-0700), Tyler Retzlaff:
> > Implement functions for getting/setting thread affinity.
> > Threads can be pinned to specific cores by setting their
> > affinity attribute.
> >
> > Signed-off-by: Narcisa Vasile <navas...@microsoft.com>
> > Signed-off-by: Tyler Retzlaff <roret...@linux.microsoft.com>
>
> Acked-by: Dmitry Kozlyuk <dmitry.kozl...@gmail.com>
thanks, will address all items mentioned in v2 in addition to some minor
issues detected with CI pipeline.
>
> Please see some small comments below.
>
> > ---
> > lib/eal/include/rte_thread.h | 45 ++++++++++
> > lib/eal/unix/rte_thread.c | 16 ++++
> > lib/eal/version.map | 4 +
> > lib/eal/windows/eal_lcore.c | 173
> > +++++++++++++++++++++++++++++----------
> > lib/eal/windows/eal_windows.h | 10 +++
> > lib/eal/windows/include/rte_os.h | 2 +
> > lib/eal/windows/rte_thread.c | 131 ++++++++++++++++++++++++++++-
> > 7 files changed, 336 insertions(+), 45 deletions(-)
> >
> > diff --git a/lib/eal/include/rte_thread.h b/lib/eal/include/rte_thread.h
> > index 8be8ed8..4eb113f 100644
> > --- a/lib/eal/include/rte_thread.h
> > +++ b/lib/eal/include/rte_thread.h
> > @@ -2,6 +2,8 @@
> > * Copyright(c) 2021 Mellanox Technologies, Ltd
> > */
> >
> > +#include <stdint.h>
> > +
> > #include <rte_os.h>
> > #include <rte_compat.h>
> >
> > @@ -21,6 +23,13 @@
> > #endif
> >
> > /**
> > + * Thread id descriptor.
> > + */
> > +typedef struct rte_thread_tag {
> > + uintptr_t opaque_id; /**< thread identifier */
> > +} rte_thread_t;
> > +
> > +/**
> > * TLS key type, an opaque pointer.
> > */
> > typedef struct eal_tls_key *rte_thread_key;
> > @@ -28,6 +37,42 @@
> > #ifdef RTE_HAS_CPUSET
> >
> > /**
>
> Missing a common part for experimental functions:
>
> * @warning
> * @b EXPERIMENTAL: this API may change without prior notice.
>
> > + * Set the affinity of thread 'thread_id' to the cpu set
> > + * specified by 'cpuset'.
> > + *
> > + * @param thread_id
> > + * Id of the thread for which to set the affinity.
> > + *
> > + * @param cpuset
> > + * Pointer to CPU affinity to set.
> > + *
> > + * @return
> > + * On success, return 0.
> > + * On failure, return a positive errno-style error number.
> > + */
> > +__rte_experimental
> > +int rte_thread_set_affinity_by_id(rte_thread_t thread_id,
> > + const rte_cpuset_t *cpuset);
> > +
> > +/**
>
> Same here.
>
> > + * Get the affinity of thread 'thread_id' and store it
> > + * in 'cpuset'.
> > + *
> > + * @param thread_id
> > + * Id of the thread for which to get the affinity.
> > + *
> > + * @param cpuset
> > + * Pointer for storing the affinity value.
> > + *
> > + * @return
> > + * On success, return 0.
> > + * On failure, return a positive errno-style error number.
> > + */
> > +__rte_experimental
> > +int rte_thread_get_affinity_by_id(rte_thread_t thread_id,
> > + rte_cpuset_t *cpuset);
> > +
> > +/**
> > * Set core affinity of the current thread.
> > * Support both EAL and non-EAL thread and update TLS.
> > *
>
> [...]
>
> > +static int
> > +eal_query_group_affinity(void)
> > +{
> > + SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *infos = NULL;
> > + unsigned int *cpu_count = &cpu_map.cpu_count;
> > + DWORD infos_size = 0;
> > + int ret = 0;
> > + USHORT group_count;
> > + KAFFINITY affinity;
> > + USHORT group_no;
> > + unsigned int i;
> > +
> > + if (!GetLogicalProcessorInformationEx(RelationGroup, NULL,
> > + &infos_size)) {
> > + DWORD error = GetLastError();
> > + if (error != ERROR_INSUFFICIENT_BUFFER) {
> > + log_early("Cannot get group information size, "
> > + "error %lu\n", error);
>
> Please don't break string constants for easy search.
>
> > + rte_errno = EINVAL;
> > + ret = -1;
> > + goto cleanup;
> > + }
> > + }
> [...]
>
> > +static bool
> > +eal_create_lcore_map(const SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *info)
> > +{
> > + const unsigned int node_id = info->NumaNode.NodeNumber;
> > + const GROUP_AFFINITY *cores = &info->NumaNode.GroupMask;
> > + struct lcore_map *lcore;
> > + unsigned int socket_id;
> > + unsigned int i;
> > +
> > + /* NUMA node may be reported multiple times if it includes
> > + * cores from different processor groups, e. g. 80 cores
> > + * of a physical processor comprise one NUMA node, but two
> > + * processor groups, because group size is limited by 32/64.
> > + */
> > + for (socket_id = 0; socket_id < cpu_map.socket_count; socket_id++) {
> > + if (cpu_map.sockets[socket_id].node_id == node_id)
> > + break;
> > + }
>
> Nit: multi-line comments should start with a line containing just "/*",
> and {} are no needed here.
>
> [...]
> > +static int
> > +rte_convert_cpuset_to_affinity(const rte_cpuset_t *cpuset,
> > + PGROUP_AFFINITY affinity)
> > +{
> > + int ret = 0;
> > + PGROUP_AFFINITY cpu_affinity = NULL;
> > + unsigned int cpu_idx;
> > +
> > + memset(affinity, 0, sizeof(GROUP_AFFINITY));
> > + affinity->Group = (USHORT)-1;
> > +
> > + /* Check that all cpus of the set belong to the same processor group and
> > + * accumulate thread affinity to be applied.
> > + */
> > + for (cpu_idx = 0; cpu_idx < CPU_SETSIZE; cpu_idx++) {
> > + if (!CPU_ISSET(cpu_idx, cpuset))
> > + continue;
> > +
> > + cpu_affinity = eal_get_cpu_affinity(cpu_idx);
> > +
> > + if (affinity->Group == (USHORT)-1) {
> > + affinity->Group = cpu_affinity->Group;
> > + } else if (affinity->Group != cpu_affinity->Group) {
> > + ret = EINVAL;
>
> Should it be ENOTSUP, because we don't support this case?
> This limitation is also worth documenting in the commit log.
>
> > + goto cleanup;
> > + }
> > +
> > + affinity->Mask |= cpu_affinity->Mask;
> > + }
> > +
> > + if (affinity->Mask == 0) {
> > + ret = EINVAL;
> > + goto cleanup;
> > + }
> > +
> > +cleanup:
> > + return ret;
> > +}
> [...]
