lib: objpool: fix head overrun on RK3588 SBC

objpool overrun stress with test_objpool on OrangePi5+ SBC triggered the
following kernel warnings:

WARNING: CPU: 6 PID: 3115 at lib/objpool.c:168 objpool_push+0xc0/0x100

This message is from objpool.c:168:

WARN_ON_ONCE(tail - head > pool->nr_objs);

The overrun test case is to validate the case that pre-allocated objects
are insufficient: 8 objects are pre-allocated for each node and consumer
thread per node tries to grab 16 objects in a row. The testing system is
OrangePI 5+, with RK3588, a big.LITTLE SOC with 4x A76 and 4x A55. When
disabling either all 4 big or 4 little cores, the overrun tests run well,
and once with big and little cores mixed together, the overrun test would
always cause an overrun loop. It's likely the memory timing differences
of big and little cores cause this trouble. Here are the debugging data
of objpool_try_get_slot after try_cmpxchg_release:

objpool_pop: cpu: 4/0 0:0 head: 278/279 tail:278 last:276/278

The local copies of 'head' and 'last' were 278 and 276, and reloading of
'slot->head' and 'slot->last' got 279 and 278. After try_cmpxchg_release
'slot->head' became 'head + 1', which is correct. But what's wrong here
is the stale value of 'last', and that stale value of 'last' finally led
the overrun of 'head'.

Memory updating of 'last' and 'head' are performed in push() and pop()
independently, which could be the culprit leading this out of order
visibility of 'last' and 'head'. So for objpool_try_get_slot(), it's
not enough only checking the condition of 'head != slot', the implicit
condition 'last - head <= nr_objs' must also be explicitly asserted to
guarantee 'last' is always behind 'head' before the object retrieving.

This patch will check and try reloading of 'head' and 'last' to ensure
'last' is behind 'head' at the time of object retrieving. Performance
testings show the average impact is about 0.1% for X86_64 and 1.12% for
ARM64. Here are the results:

OS: Debian 10 X86_64, Linux 6.6rc
HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s
1T 2T 4T 8T 16T
native: 49543304 99277826 199017659 399070324 795185848
objpool: 29909085 59865637 119692073 239750369 478005250
objpool+: 29879313 59230743 119609856 239067773 478509029
32T 48T 64T 96T 128T
native: 1596927073 2390099988 2929397330 3183875848 3257546602
objpool: 957553042 1435814086 1680872925 2043126796 2165424198
objpool+: 956476281 1434491297 1666055740 2041556569 2157415622

OS: Debian 11 AARCH64, Linux 6.6rc
HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s
1T 2T 4T 8T 16T
native: 30890508 60399915 123111980 242257008 494002946
objpool: 14742531 28883047 57739948 115886644 232455421
objpool+: 14107220 29032998 57286084 113730493 232232850
24T 32T 48T 64T 96T
native: 746406039 1000174750 1493236240 1998318364 2942911180
objpool: 349164852 467284332 702296756 934459713 1387898285
objpool+: 348388180 462750976 696606096 927865887 1368402195

Link: https://lore.kernel.org/all/20231114115148.298821-1-wuqiang.matt@bytedance.com/

Fixes: b4edb8d2d464 ("lib: objpool added: ring-array based lockless MPMC")
Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com>
Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>

lib: objpool added: ring-array based lockless MPMC

objpool is a scalable implementation of high performance queue for
object allocation and reclamation, such as kretprobe instances.

With leveraging percpu ring-array to mitigate hot spots of memory
contention, it delivers near-linear scalability for high parallel
scenarios. The objpool is best suited for the following cases:
1) Memory allocation or reclamation are prohibited or too expensive
2) Consumers are of different priorities, such as irqs and threads

Limitations:
1) Maximum objects (capacity) is fixed after objpool creation
2) All pre-allocated objects are managed in percpu ring array,
which consumes more memory than linked lists

Link: https://lore.kernel.org/all/20231017135654.82270-2-wuqiang.matt@bytedance.com/

Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com>
Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>