summaryrefslogtreecommitdiffstats
path: root/lib/objpool.c
Commit message (Collapse)AuthorAgeFilesLines
* objpool: cache nr_possible_cpus() and avoid caching nr_cpu_idsAndrii Nakryiko2024-05-011-6/+6
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Profiling shows that calling nr_possible_cpus() in objpool_pop() takes a noticeable amount of CPU (when profiled on 80-core machine), as we need to recalculate number of set bits in a CPU bit mask. This number can't change, so there is no point in paying the price for recalculating it. As such, cache this value in struct objpool_head and use it in objpool_pop(). On the other hand, cached pool->nr_cpus isn't necessary, as it's not used in hot path and is also a pretty trivial value to retrieve. So drop pool->nr_cpus in favor of using nr_cpu_ids everywhere. This way the size of struct objpool_head remains the same, which is a nice bonus. Same BPF selftests benchmarks were used to evaluate the effect. Using changes in previous patch (inlining of objpool_pop/objpool_push) as baseline, here are the differences: BASELINE ======== kretprobe : 9.937 ± 0.174M/s kretprobe-multi: 10.440 ± 0.108M/s AFTER ===== kretprobe : 10.106 ± 0.120M/s (+1.7%) kretprobe-multi: 10.515 ± 0.180M/s (+0.7%) Link: https://lore.kernel.org/all/20240424215214.3956041-3-andrii@kernel.org/ Cc: Matt (Qiang) Wu <wuqiang.matt@bytedance.com> Signed-off-by: Andrii Nakryiko <andrii@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
* objpool: enable inlining objpool_push() and objpool_pop() operationsAndrii Nakryiko2024-05-011-100/+0
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | objpool_push() and objpool_pop() are very performance-critical functions and can be called very frequently in kretprobe triggering path. As such, it makes sense to allow compiler to inline them completely to eliminate function calls overhead. Luckily, their logic is quite well isolated and doesn't have any sprawling dependencies. This patch moves both objpool_push() and objpool_pop() into include/linux/objpool.h and marks them as static inline functions, enabling inlining. To avoid anyone using internal helpers (objpool_try_get_slot, objpool_try_add_slot), rename them to use leading underscores. We used kretprobe microbenchmark from BPF selftests (bench trig-kprobe and trig-kprobe-multi benchmarks) running no-op BPF kretprobe/kretprobe.multi programs in a tight loop to evaluate the effect. BPF own overhead in this case is minimal and it mostly stresses the rest of in-kernel kretprobe infrastructure overhead. Results are in millions of calls per second. This is not super scientific, but shows the trend nevertheless. BEFORE ====== kretprobe : 9.794 ± 0.086M/s kretprobe-multi: 10.219 ± 0.032M/s AFTER ===== kretprobe : 9.937 ± 0.174M/s (+1.5%) kretprobe-multi: 10.440 ± 0.108M/s (+2.2%) Link: https://lore.kernel.org/all/20240424215214.3956041-2-andrii@kernel.org/ Cc: Matt (Qiang) Wu <wuqiang.matt@bytedance.com> Signed-off-by: Andrii Nakryiko <andrii@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
* lib: objpool: fix head overrun on RK3588 SBCwuqiang.matt2023-12-011-0/+17
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | 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 MPMCwuqiang.matt2023-10-181-0/+280
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>