A sampling log could lose most of its samples at close, and
pmcstat -R then read few or none, with no error.

• pmclog_close() wrote the close record before it queued the buffers of the other CPUs. Their samples came after the close record, where pmclog_read() stops.
• pmclog_stop_kthread() cleared PMC_PO_OWNS_LOGFILE before the helper wrote its queue. The helper exited, and pmclog_deconfigure_log() freed the queued buffers.

Queue all buffers, set PMC_PO_SHUTDOWN, queue again, and write the
close record last. After a close, let the helper write its queue
before it exits. Stop waiting after 500 ms without progress, e.g.
when nobody reads a pipe.

Add a test: write records on one CPU, close the log on another, and
check that no record follows the close record.

Sponsored by: AMD

pmcstat -n 1000000 -P instructions -O /tmp/s.log <cpu-bound loop>
pmcstat -R /tmp/s.log -o /dev/stdout | grep -c ^callchain

Before: 486 NMIs, pmcstat -R reads 0-2 samples; most samples are
after the close record.
After: 486 of 486. Same with -S and SIGINT: no record after the close
record. With a pipe that nobody reads, pmcstat still exits in about
1 s on SIGINT or SIGKILL.

tests/sys/pmc/pmc_log_test:close_record_is_last fails before and
passes after.