hwloc: report cpu numbers in the basis we claim, reset NUMA placement counters, honor a qualifier-only --bind-to

[rhc54]

A second review pass over src/hwloc, following #2578. Three substantive
defects, three small ones, and two memory leaks the review turned up
elsewhere. Six commits, each standalone and bisectable.

The cpu numbers we print are not the ones we accept

The bits of an hwloc cpuset are always PU OS indices — the kernel's cpu
numbers. What PRRTE shows a user is a list of cores (or of hwthreads, when
the job treats hwthreads as cpus), and every PRRTE grammar that accepts
such a list — --cpu-set, the rankfile and sequence-file slot lists —
resolves it as an hwloc logical index. Those three numbering schemes
coincide only by luck.

Three renderers short-cut whenever the bits "already were" cores
(npus == ncores) or the job was in hwthreads: they printed the raw bitmap
under a core:L/hwt:L label, and ignored physical outright. So on any
SMT node viewed in hwthreads, or any node whose firmware interleaves cpu
numbers across packages, --display cpus reported numbers that --cpu-set
would resolve to a different set of cpus.

Everything now resolves each set bit to the object that owns it, through one
new primitive, prte_hwloc_base_cpuset2ranges(). prte_hwloc_build_map() is
gone with the short cut — it produced a bitmap of core indices, which a
caller then printed with hwloc_bitmap_list_snprintf(), mixing two index
bases in one string. prte_proc_t.cpuset deliberately stays raw: it is a
wire format read back with hwloc_bitmap_list_sscanf().

The same defect was in the collision messages of rmaps/seq,
rmaps/rank_file and rmaps/lsf, which printed a raw bitmap as the
"available" and "overlapping" cpus beside the user's own logical slot list.
Fixed there too.

--bind-to numa:limit=N works once per DVM

prte_hwloc_base_reset_counters() clears the per-object placement counters
between jobs. It walked only the normal depth hierarchy — and hwloc 2.x keeps
NUMA nodes out of that hierarchy, at HWLOC_TYPE_DEPTH_NUMANODE. Its
HWLOC_OBJ_NUMANODE != type guard could never fire, so a counter attached to
a NUMA node was never reset. In a persistent DVM it accumulated for the life
of the DVM, and the second --bind-to numa:limit=N job found every
domain already at its limit and could not be bound. prterun hides it by
starting a fresh HNP each time.

--bind-to :overload-allowed silently meant --bind-to none

pmix_check_cli_option() compares only min(strlen(a), strlen(b))
characters, so the empty policy word of the :qualifier form matched the
first option it was tested against — none. Binding was disabled, and
because the default mapping policy is derived from the binding, BYCORE
became BYSLOT along with it. Nothing was printed to say so.
--map-by :OVERSUBSCRIBE has always worked correctly; these now agree.

Two leaks

Found with valgrind while checking the above; both pre-existing and outside
src/hwloc, so they are the first two commits.

• prte_relm_start_msg() unloaded the caller's pmix_data_buffer_t into
  a fresh RELM-framed one and never released the emptied container. It owns
  the buffer once it succeeds — the contract prte_rml_send_buffer_nb()
  follows and every PRTE_RML_RELIABLE_SEND caller depends on. This was one
  leaked buffer per inter-daemon reliable message, for the life of the
  DVM, not the one-off the first reproducer suggested.
• prte_parse_locals() has three exits and only the success path
  released temp_argv/env. Any command line whose final segment fails to
  parse leaked both — --display map --display cpus is enough, since a
  repeated option is refused right there.

Smaller things

A core-less topology (PPC64 on old kernels reports only NUMA nodes and PUs)
now renders in hwthreads and says so, instead of resolving nothing and coming
back UNBOUND. A package:core id resolves inside its package rather than
by "objects-per-package × package id". prte_hwloc_print() and its three
callers handle a node whose topology has not arrived. The
hwloc_default_cpu_list MCA value is no longer truncated in place. Plus a
NULL PU deref and an info-array removal loop that skipped the entry shifted
into the vacated slot.

Testing

• make check — test/unit/hwloc extended; the logical-vs-physical case
  needs a hand-written XML topology, because hwloc's synthetic generator
  always makes os_index and logical_index agree. (Note for anyone
  extending it: hwloc's XML importer segfaults rather than diagnoses on an
  object missing complete_cpuset/nodeset/complete_nodeset.)
• make -C test/offline check-offline — 1180/1180.
• contrib/dockerswarm — test_hwloc grew from 23 to 35 cases, including
  the one that needs a persistent DVM (three successive numa:limit=1
  jobs) and the print-then-accept round trip. Full suite: 407 passed, 0
  failed, 2 skipped (both skips pre-existing PMIx-capability gates).
• valgrind on the rendering and launch paths: definitely lost: 0.

Each commit builds warning-free on its own.

🤖 Generated with Claude Code