From: Jesper Dangaard Brouer hawk@kernel.org
We frequently consult with Jesper's out-of-tree page_pool benchmark to evaluate page_pool changes.
Import the benchmark into the upstream linux kernel tree so that (a) we're all running the same version, (b) pave the way for shared improvements, and (c) maybe one day integrate it with nipa, if possible.
Import bench_page_pool_simple from commit 35b1716d0c30 ("Add page_bench06_walk_all"), from this repository: https://github.com/netoptimizer/prototype-kernel.git
Changes done during upstreaming: - Fix checkpatch issues. - Remove the tasklet logic not needed. - Move under tools/testing - Create ksft for the benchmark. - Changed slightly how the benchmark gets build. Out of tree, time_bench is built as an independent .ko. Here it is included in bench_page_pool.ko
Steps to run:
``` mkdir -p /tmp/run-pp-bench make -C ./tools/testing/selftests/net/bench make -C ./tools/testing/selftests/net/bench install INSTALL_PATH=/tmp/run-pp-bench rsync --delete -avz --progress /tmp/run-pp-bench mina@$SERVER:~/ ssh mina@$SERVER << EOF cd ~/run-pp-bench && sudo ./test_bench_page_pool.sh EOF ```
Output:
``` (benchmrk dmesg logs)
Fast path results: no-softirq-page_pool01 Per elem: 11 cycles(tsc) 4.368 ns
ptr_ring results: no-softirq-page_pool02 Per elem: 527 cycles(tsc) 195.187 ns
slow path results: no-softirq-page_pool03 Per elem: 549 cycles(tsc) 203.466 ns ```
Cc: Jesper Dangaard Brouer hawk@kernel.org Cc: Ilias Apalodimas ilias.apalodimas@linaro.org Cc: Jakub Kicinski kuba@kernel.org Cc: Toke Høiland-Jørgensen toke@toke.dk
Signed-off-by: Mina Almasry almasrymina@google.com Acked-by: Ilias Apalodimas ilias.apalodimas@linaro.org Signed-off-by: Jesper Dangaard Brouer hawk@kernel.org
---
v4: https://lore.kernel.org/netdev/20250614100853.3f2372f2@kernel.org/
- Fix more checkpatch and coccicheck issues (Jakub)
v3: - Non RFC - Collect Signed-off-by from Jesper and Acked-by Ilias. - Move test_bench_page_pool.sh to address nipa complaint. - Remove `static inline` in .c files to address nipa complaint.
v2: - Move under tools/selftests (Jakub) - Create ksft for it. - Remove the tasklet logic no longer needed (Jesper + Toke)
RFC discussion points: - Desirable to import it? - Can the benchmark be imported as-is for an initial version? Or needs lots of modifications? - Code location. I retained the location in Jesper's tree, but a path like net/core/bench/ may make more sense.
--- tools/testing/selftests/net/bench/Makefile | 7 + .../selftests/net/bench/page_pool/Makefile | 17 + .../bench/page_pool/bench_page_pool_simple.c | 276 ++++++++++++ .../net/bench/page_pool/time_bench.c | 394 ++++++++++++++++++ .../net/bench/page_pool/time_bench.h | 238 +++++++++++ .../net/bench/test_bench_page_pool.sh | 32 ++ 6 files changed, 964 insertions(+) create mode 100644 tools/testing/selftests/net/bench/Makefile create mode 100644 tools/testing/selftests/net/bench/page_pool/Makefile create mode 100644 tools/testing/selftests/net/bench/page_pool/bench_page_pool_simple.c create mode 100644 tools/testing/selftests/net/bench/page_pool/time_bench.c create mode 100644 tools/testing/selftests/net/bench/page_pool/time_bench.h create mode 100755 tools/testing/selftests/net/bench/test_bench_page_pool.sh
diff --git a/tools/testing/selftests/net/bench/Makefile b/tools/testing/selftests/net/bench/Makefile new file mode 100644 index 000000000000..2546c45e42f7 --- /dev/null +++ b/tools/testing/selftests/net/bench/Makefile @@ -0,0 +1,7 @@ +# SPDX-License-Identifier: GPL-2.0 + +TEST_GEN_MODS_DIR := page_pool + +TEST_PROGS += test_bench_page_pool.sh + +include ../../lib.mk diff --git a/tools/testing/selftests/net/bench/page_pool/Makefile b/tools/testing/selftests/net/bench/page_pool/Makefile new file mode 100644 index 000000000000..0549a16ba275 --- /dev/null +++ b/tools/testing/selftests/net/bench/page_pool/Makefile @@ -0,0 +1,17 @@ +BENCH_PAGE_POOL_SIMPLE_TEST_DIR := $(realpath $(dir $(abspath $(lastword $(MAKEFILE_LIST))))) +KDIR ?= /lib/modules/$(shell uname -r)/build + +ifeq ($(V),1) +Q = +else +Q = @ +endif + +obj-m += bench_page_pool.o +bench_page_pool-y += bench_page_pool_simple.o time_bench.o + +all: + +$(Q)make -C $(KDIR) M=$(BENCH_PAGE_POOL_SIMPLE_TEST_DIR) modules + +clean: + +$(Q)make -C $(KDIR) M=$(BENCH_PAGE_POOL_SIMPLE_TEST_DIR) clean diff --git a/tools/testing/selftests/net/bench/page_pool/bench_page_pool_simple.c b/tools/testing/selftests/net/bench/page_pool/bench_page_pool_simple.c new file mode 100644 index 000000000000..f183d5e30dc6 --- /dev/null +++ b/tools/testing/selftests/net/bench/page_pool/bench_page_pool_simple.c @@ -0,0 +1,276 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* + * Benchmark module for page_pool. + * + */ +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt + +#include <linux/module.h> +#include <linux/mutex.h> + +#include <linux/version.h> +#include <net/page_pool/helpers.h> + +#include <linux/interrupt.h> +#include <linux/limits.h> + +#include "time_bench.h" + +static int verbose = 1; +#define MY_POOL_SIZE 1024 + +static void _page_pool_put_page(struct page_pool *pool, struct page *page, + bool allow_direct) +{ + page_pool_put_page(pool, page, -1, allow_direct); +} + +/* Makes tests selectable. Useful for perf-record to analyze a single test. + * Hint: Bash shells support writing binary number like: $((2#101010) + * + * # modprobe bench_page_pool_simple run_flags=$((2#100)) + */ +static unsigned long run_flags = 0xFFFFFFFF; +module_param(run_flags, ulong, 0); +MODULE_PARM_DESC(run_flags, "Limit which bench test that runs"); + +/* Count the bit number from the enum */ +enum benchmark_bit { + bit_run_bench_baseline, + bit_run_bench_no_softirq01, + bit_run_bench_no_softirq02, + bit_run_bench_no_softirq03, +}; + +#define bit(b) (1 << (b)) +#define enabled(b) ((run_flags & (bit(b)))) + +/* notice time_bench is limited to U32_MAX nr loops */ +static unsigned long loops = 10000000; +module_param(loops, ulong, 0); +MODULE_PARM_DESC(loops, "Specify loops bench will run"); + +/* Timing at the nanosec level, we need to know the overhead + * introduced by the for loop itself + */ +static int time_bench_for_loop(struct time_bench_record *rec, void *data) +{ + uint64_t loops_cnt = 0; + int i; + + time_bench_start(rec); + /** Loop to measure **/ + for (i = 0; i < rec->loops; i++) { + loops_cnt++; + barrier(); /* avoid compiler to optimize this loop */ + } + time_bench_stop(rec, loops_cnt); + return loops_cnt; +} + +static int time_bench_atomic_inc(struct time_bench_record *rec, void *data) +{ + uint64_t loops_cnt = 0; + atomic_t cnt; + int i; + + atomic_set(&cnt, 0); + + time_bench_start(rec); + /** Loop to measure **/ + for (i = 0; i < rec->loops; i++) { + atomic_inc(&cnt); + barrier(); /* avoid compiler to optimize this loop */ + } + loops_cnt = atomic_read(&cnt); + time_bench_stop(rec, loops_cnt); + return loops_cnt; +} + +/* The ptr_ping in page_pool uses a spinlock. We need to know the minimum + * overhead of taking+releasing a spinlock, to know the cycles that can be saved + * by e.g. amortizing this via bulking. + */ +static int time_bench_lock(struct time_bench_record *rec, void *data) +{ + uint64_t loops_cnt = 0; + spinlock_t lock; + int i; + + spin_lock_init(&lock); + + time_bench_start(rec); + /** Loop to measure **/ + for (i = 0; i < rec->loops; i++) { + spin_lock(&lock); + loops_cnt++; + barrier(); /* avoid compiler to optimize this loop */ + spin_unlock(&lock); + } + time_bench_stop(rec, loops_cnt); + return loops_cnt; +} + +/* Helper for filling some page's into ptr_ring */ +static void pp_fill_ptr_ring(struct page_pool *pp, int elems) +{ + /* GFP_ATOMIC needed when under run softirq */ + gfp_t gfp_mask = GFP_ATOMIC; + struct page **array; + int i; + + array = kcalloc(elems, sizeof(struct page *), gfp_mask); + + for (i = 0; i < elems; i++) + array[i] = page_pool_alloc_pages(pp, gfp_mask); + for (i = 0; i < elems; i++) + _page_pool_put_page(pp, array[i], false); + + kfree(array); +} + +enum test_type { type_fast_path, type_ptr_ring, type_page_allocator }; + +/* Depends on compile optimizing this function */ +static int time_bench_page_pool(struct time_bench_record *rec, void *data, + enum test_type type, const char *func) +{ + uint64_t loops_cnt = 0; + gfp_t gfp_mask = GFP_ATOMIC; /* GFP_ATOMIC is not really needed */ + int i, err; + + struct page_pool *pp; + struct page *page; + + struct page_pool_params pp_params = { + .order = 0, + .flags = 0, + .pool_size = MY_POOL_SIZE, + .nid = NUMA_NO_NODE, + .dev = NULL, /* Only use for DMA mapping */ + .dma_dir = DMA_BIDIRECTIONAL, + }; + + pp = page_pool_create(&pp_params); + if (IS_ERR(pp)) { + err = PTR_ERR(pp); + pr_warn("%s: Error(%d) creating page_pool\n", func, err); + goto out; + } + pp_fill_ptr_ring(pp, 64); + + if (in_serving_softirq()) + pr_warn("%s(): in_serving_softirq fast-path\n", func); + else + pr_warn("%s(): Cannot use page_pool fast-path\n", func); + + time_bench_start(rec); + /** Loop to measure **/ + for (i = 0; i < rec->loops; i++) { + /* Common fast-path alloc that depend on in_serving_softirq() */ + page = page_pool_alloc_pages(pp, gfp_mask); + if (!page) + break; + loops_cnt++; + barrier(); /* avoid compiler to optimize this loop */ + + /* The benchmarks purpose it to test different return paths. + * Compiler should inline optimize other function calls out + */ + if (type == type_fast_path) { + /* Fast-path recycling e.g. XDP_DROP use-case */ + page_pool_recycle_direct(pp, page); + + } else if (type == type_ptr_ring) { + /* Normal return path */ + _page_pool_put_page(pp, page, false); + + } else if (type == type_page_allocator) { + /* Test if not pages are recycled, but instead + * returned back into systems page allocator + */ + get_page(page); /* cause no-recycling */ + _page_pool_put_page(pp, page, false); + put_page(page); + } else { + BUILD_BUG(); + } + } + time_bench_stop(rec, loops_cnt); +out: + page_pool_destroy(pp); + return loops_cnt; +} + +static int time_bench_page_pool01_fast_path(struct time_bench_record *rec, + void *data) +{ + return time_bench_page_pool(rec, data, type_fast_path, __func__); +} + +static int time_bench_page_pool02_ptr_ring(struct time_bench_record *rec, + void *data) +{ + return time_bench_page_pool(rec, data, type_ptr_ring, __func__); +} + +static int time_bench_page_pool03_slow(struct time_bench_record *rec, + void *data) +{ + return time_bench_page_pool(rec, data, type_page_allocator, __func__); +} + +static int run_benchmark_tests(void) +{ + uint32_t nr_loops = loops; + + /* Baseline tests */ + if (enabled(bit_run_bench_baseline)) { + time_bench_loop(nr_loops * 10, 0, "for_loop", NULL, + time_bench_for_loop); + time_bench_loop(nr_loops * 10, 0, "atomic_inc", NULL, + time_bench_atomic_inc); + time_bench_loop(nr_loops, 0, "lock", NULL, time_bench_lock); + } + + /* This test cannot activate correct code path, due to no-softirq ctx */ + if (enabled(bit_run_bench_no_softirq01)) + time_bench_loop(nr_loops, 0, "no-softirq-page_pool01", NULL, + time_bench_page_pool01_fast_path); + if (enabled(bit_run_bench_no_softirq02)) + time_bench_loop(nr_loops, 0, "no-softirq-page_pool02", NULL, + time_bench_page_pool02_ptr_ring); + if (enabled(bit_run_bench_no_softirq03)) + time_bench_loop(nr_loops, 0, "no-softirq-page_pool03", NULL, + time_bench_page_pool03_slow); + + return 0; +} + +static int __init bench_page_pool_simple_module_init(void) +{ + if (verbose) + pr_info("Loaded\n"); + + if (loops > U32_MAX) { + pr_err("Module param loops(%lu) exceeded U32_MAX(%u)\n", loops, + U32_MAX); + return -ECHRNG; + } + + run_benchmark_tests(); + + return 0; +} +module_init(bench_page_pool_simple_module_init); + +static void __exit bench_page_pool_simple_module_exit(void) +{ + if (verbose) + pr_info("Unloaded\n"); +} +module_exit(bench_page_pool_simple_module_exit); + +MODULE_DESCRIPTION("Benchmark of page_pool simple cases"); +MODULE_AUTHOR("Jesper Dangaard Brouer netoptimizer@brouer.com"); +MODULE_LICENSE("GPL"); diff --git a/tools/testing/selftests/net/bench/page_pool/time_bench.c b/tools/testing/selftests/net/bench/page_pool/time_bench.c new file mode 100644 index 000000000000..073bb36ec5f2 --- /dev/null +++ b/tools/testing/selftests/net/bench/page_pool/time_bench.c @@ -0,0 +1,394 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* + * Benchmarking code execution time inside the kernel + * + * Copyright (C) 2014, Red Hat, Inc., Jesper Dangaard Brouer + */ +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt + +#include <linux/module.h> +#include <linux/time.h> + +#include <linux/perf_event.h> /* perf_event_create_kernel_counter() */ + +/* For concurrency testing */ +#include <linux/completion.h> +#include <linux/sched.h> +#include <linux/workqueue.h> +#include <linux/kthread.h> + +#include "time_bench.h" + +static int verbose = 1; + +/** TSC (Time-Stamp Counter) based ** + * See: linux/time_bench.h + * tsc_start_clock() and tsc_stop_clock() + */ + +/** Wall-clock based ** + */ + +/** PMU (Performance Monitor Unit) based ** + */ +#define PERF_FORMAT \ + (PERF_FORMAT_GROUP | PERF_FORMAT_ID | PERF_FORMAT_TOTAL_TIME_ENABLED | \ + PERF_FORMAT_TOTAL_TIME_RUNNING) + +struct raw_perf_event { + uint64_t config; /* event */ + uint64_t config1; /* umask */ + struct perf_event *save; + char *desc; +}; + +/* if HT is enable a maximum of 4 events (5 if one is instructions + * retired can be specified, if HT is disabled a maximum of 8 (9 if + * one is instructions retired) can be specified. + * + * From Table 19-1. Architectural Performance Events + * Architectures Software Developer’s Manual Volume 3: System Programming + * Guide + */ +struct raw_perf_event perf_events[] = { + { 0x3c, 0x00, NULL, "Unhalted CPU Cycles" }, + { 0xc0, 0x00, NULL, "Instruction Retired" } +}; + +#define NUM_EVTS (ARRAY_SIZE(perf_events)) + +/* WARNING: PMU config is currently broken! + */ +bool time_bench_PMU_config(bool enable) +{ + int i; + struct perf_event_attr perf_conf; + struct perf_event *perf_event; + int cpu; + + preempt_disable(); + cpu = smp_processor_id(); + pr_info("DEBUG: cpu:%d\n", cpu); + preempt_enable(); + + memset(&perf_conf, 0, sizeof(struct perf_event_attr)); + perf_conf.type = PERF_TYPE_RAW; + perf_conf.size = sizeof(struct perf_event_attr); + perf_conf.read_format = PERF_FORMAT; + perf_conf.pinned = 1; + perf_conf.exclude_user = 1; /* No userspace events */ + perf_conf.exclude_kernel = 0; /* Only kernel events */ + + for (i = 0; i < NUM_EVTS; i++) { + perf_conf.disabled = enable; + //perf_conf.disabled = (i == 0) ? 1 : 0; + perf_conf.config = perf_events[i].config; + perf_conf.config1 = perf_events[i].config1; + if (verbose) + pr_info("%s() enable PMU counter: %s\n", + __func__, perf_events[i].desc); + perf_event = perf_event_create_kernel_counter(&perf_conf, cpu, + NULL /* task */, + NULL /* overflow_handler*/, + NULL /* context */); + if (perf_event) { + perf_events[i].save = perf_event; + pr_info("%s():DEBUG perf_event success\n", __func__); + + perf_event_enable(perf_event); + } else { + pr_info("%s():DEBUG perf_event is NULL\n", __func__); + } + } + + return true; +} + +/** Generic functions ** + */ + +/* Calculate stats, store results in record */ +bool time_bench_calc_stats(struct time_bench_record *rec) +{ +#define NANOSEC_PER_SEC 1000000000 /* 10^9 */ + uint64_t ns_per_call_tmp_rem = 0; + uint32_t ns_per_call_remainder = 0; + uint64_t pmc_ipc_tmp_rem = 0; + uint32_t pmc_ipc_remainder = 0; + uint32_t pmc_ipc_div = 0; + uint32_t invoked_cnt_precision = 0; + uint32_t invoked_cnt = 0; /* 32-bit due to div_u64_rem() */ + + if (rec->flags & TIME_BENCH_LOOP) { + if (rec->invoked_cnt < 1000) { + pr_err("ERR: need more(>1000) loops(%llu) for timing\n", + rec->invoked_cnt); + return false; + } + if (rec->invoked_cnt > ((1ULL << 32) - 1)) { + /* div_u64_rem() can only support div with 32bit*/ + pr_err("ERR: Invoke cnt(%llu) too big overflow 32bit\n", + rec->invoked_cnt); + return false; + } + invoked_cnt = (uint32_t)rec->invoked_cnt; + } + + /* TSC (Time-Stamp Counter) records */ + if (rec->flags & TIME_BENCH_TSC) { + rec->tsc_interval = rec->tsc_stop - rec->tsc_start; + if (rec->tsc_interval == 0) { + pr_err("ABORT: timing took ZERO TSC time\n"); + return false; + } + /* Calculate stats */ + if (rec->flags & TIME_BENCH_LOOP) + rec->tsc_cycles = rec->tsc_interval / invoked_cnt; + else + rec->tsc_cycles = rec->tsc_interval; + } + + /* Wall-clock time calc */ + if (rec->flags & TIME_BENCH_WALLCLOCK) { + rec->time_start = rec->ts_start.tv_nsec + + (NANOSEC_PER_SEC * rec->ts_start.tv_sec); + rec->time_stop = rec->ts_stop.tv_nsec + + (NANOSEC_PER_SEC * rec->ts_stop.tv_sec); + rec->time_interval = rec->time_stop - rec->time_start; + if (rec->time_interval == 0) { + pr_err("ABORT: timing took ZERO wallclock time\n"); + return false; + } + /* Calculate stats */ + /*** Division in kernel it tricky ***/ + /* Orig: time_sec = (time_interval / NANOSEC_PER_SEC); */ + /* remainder only correct because NANOSEC_PER_SEC is 10^9 */ + rec->time_sec = div_u64_rem(rec->time_interval, NANOSEC_PER_SEC, + &rec->time_sec_remainder); + //TODO: use existing struct timespec records instead of div? + + if (rec->flags & TIME_BENCH_LOOP) { + /*** Division in kernel it tricky ***/ + /* Orig: ns = ((double)time_interval / invoked_cnt); */ + /* First get quotient */ + rec->ns_per_call_quotient = + div_u64_rem(rec->time_interval, invoked_cnt, + &ns_per_call_remainder); + /* Now get decimals .xxx precision (incorrect roundup)*/ + ns_per_call_tmp_rem = ns_per_call_remainder; + invoked_cnt_precision = invoked_cnt / 1000; + if (invoked_cnt_precision > 0) { + rec->ns_per_call_decimal = + div_u64_rem(ns_per_call_tmp_rem, + invoked_cnt_precision, + &ns_per_call_remainder); + } + } + } + + /* Performance Monitor Unit (PMU) counters */ + if (rec->flags & TIME_BENCH_PMU) { + //FIXME: Overflow handling??? + rec->pmc_inst = rec->pmc_inst_stop - rec->pmc_inst_start; + rec->pmc_clk = rec->pmc_clk_stop - rec->pmc_clk_start; + + /* Calc Instruction Per Cycle (IPC) */ + /* First get quotient */ + rec->pmc_ipc_quotient = div_u64_rem(rec->pmc_inst, rec->pmc_clk, + &pmc_ipc_remainder); + /* Now get decimals .xxx precision (incorrect roundup)*/ + pmc_ipc_tmp_rem = pmc_ipc_remainder; + pmc_ipc_div = rec->pmc_clk / 1000; + if (pmc_ipc_div > 0) { + rec->pmc_ipc_decimal = div_u64_rem(pmc_ipc_tmp_rem, + pmc_ipc_div, + &pmc_ipc_remainder); + } + } + + return true; +} + +/* Generic function for invoking a loop function and calculating + * execution time stats. The function being called/timed is assumed + * to perform a tight loop, and update the timing record struct. + */ +bool time_bench_loop(uint32_t loops, int step, char *txt, void *data, + int (*func)(struct time_bench_record *record, void *data)) +{ + struct time_bench_record rec; + + /* Setup record */ + memset(&rec, 0, sizeof(rec)); /* zero func might not update all */ + rec.version_abi = 1; + rec.loops = loops; + rec.step = step; + rec.flags = (TIME_BENCH_LOOP | TIME_BENCH_TSC | TIME_BENCH_WALLCLOCK); + + /*** Loop function being timed ***/ + if (!func(&rec, data)) { + pr_err("ABORT: function being timed failed\n"); + return false; + } + + if (rec.invoked_cnt < loops) + pr_warn("WARNING: Invoke count(%llu) smaller than loops(%d)\n", + rec.invoked_cnt, loops); + + /* Calculate stats */ + time_bench_calc_stats(&rec); + + pr_info("Type:%s Per elem: %llu cycles(tsc) %llu.%03llu ns (step:%d) - (measurement period time:%llu.%09u sec time_interval:%llu) - (invoke count:%llu tsc_interval:%llu)\n", + txt, rec.tsc_cycles, rec.ns_per_call_quotient, + rec.ns_per_call_decimal, rec.step, rec.time_sec, + rec.time_sec_remainder, rec.time_interval, rec.invoked_cnt, + rec.tsc_interval); + if (rec.flags & TIME_BENCH_PMU) + pr_info("Type:%s PMU inst/clock%llu/%llu = %llu.%03llu IPC (inst per cycle)\n", + txt, rec.pmc_inst, rec.pmc_clk, rec.pmc_ipc_quotient, + rec.pmc_ipc_decimal); + return true; +} + +/* Function getting invoked by kthread */ +static int invoke_test_on_cpu_func(void *private) +{ + struct time_bench_cpu *cpu = private; + struct time_bench_sync *sync = cpu->sync; + cpumask_t newmask = CPU_MASK_NONE; + void *data = cpu->data; + + /* Restrict CPU */ + cpumask_set_cpu(cpu->rec.cpu, &newmask); + set_cpus_allowed_ptr(current, &newmask); + + /* Synchronize start of concurrency test */ + atomic_inc(&sync->nr_tests_running); + wait_for_completion(&sync->start_event); + + /* Start benchmark function */ + if (!cpu->bench_func(&cpu->rec, data)) { + pr_err("ERROR: function being timed failed on CPU:%d(%d)\n", + cpu->rec.cpu, smp_processor_id()); + } else { + if (verbose) + pr_info("SUCCESS: ran on CPU:%d(%d)\n", cpu->rec.cpu, + smp_processor_id()); + } + cpu->did_bench_run = true; + + /* End test */ + atomic_dec(&sync->nr_tests_running); + /* Wait for kthread_stop() telling us to stop */ + while (!kthread_should_stop()) { + set_current_state(TASK_INTERRUPTIBLE); + schedule(); + } + __set_current_state(TASK_RUNNING); + return 0; +} + +void time_bench_print_stats_cpumask(const char *desc, + struct time_bench_cpu *cpu_tasks, + const struct cpumask *mask) +{ + uint64_t average = 0; + int cpu; + int step = 0; + struct sum { + uint64_t tsc_cycles; + int records; + } sum = { 0 }; + + /* Get stats */ + for_each_cpu(cpu, mask) { + struct time_bench_cpu *c = &cpu_tasks[cpu]; + struct time_bench_record *rec = &c->rec; + + /* Calculate stats */ + time_bench_calc_stats(rec); + + pr_info("Type:%s CPU(%d) %llu cycles(tsc) %llu.%03llu ns (step:%d) - (measurement period time:%llu.%09u sec time_interval:%llu) - (invoke count:%llu tsc_interval:%llu)\n", + desc, cpu, rec->tsc_cycles, rec->ns_per_call_quotient, + rec->ns_per_call_decimal, rec->step, rec->time_sec, + rec->time_sec_remainder, rec->time_interval, + rec->invoked_cnt, rec->tsc_interval); + + /* Collect average */ + sum.records++; + sum.tsc_cycles += rec->tsc_cycles; + step = rec->step; + } + + if (sum.records) /* avoid div-by-zero */ + average = sum.tsc_cycles / sum.records; + pr_info("Sum Type:%s Average: %llu cycles(tsc) CPUs:%d step:%d\n", desc, + average, sum.records, step); +} + +void time_bench_run_concurrent(uint32_t loops, int step, void *data, + const struct cpumask *mask, /* Support masking outsome CPUs*/ + struct time_bench_sync *sync, + struct time_bench_cpu *cpu_tasks, + int (*func)(struct time_bench_record *record, void *data)) +{ + int cpu, running = 0; + + if (verbose) // DEBUG + pr_warn("%s() Started on CPU:%d\n", __func__, + smp_processor_id()); + + /* Reset sync conditions */ + atomic_set(&sync->nr_tests_running, 0); + init_completion(&sync->start_event); + + /* Spawn off jobs on all CPUs */ + for_each_cpu(cpu, mask) { + struct time_bench_cpu *c = &cpu_tasks[cpu]; + + running++; + c->sync = sync; /* Send sync variable along */ + c->data = data; /* Send opaque along */ + + /* Init benchmark record */ + memset(&c->rec, 0, sizeof(struct time_bench_record)); + c->rec.version_abi = 1; + c->rec.loops = loops; + c->rec.step = step; + c->rec.flags = (TIME_BENCH_LOOP | TIME_BENCH_TSC | + TIME_BENCH_WALLCLOCK); + c->rec.cpu = cpu; + c->bench_func = func; + c->task = kthread_run(invoke_test_on_cpu_func, c, + "time_bench%d", cpu); + if (IS_ERR(c->task)) { + pr_err("%s(): Failed to start test func\n", __func__); + return; /* Argh, what about cleanup?! */ + } + } + + /* Wait until all processes are running */ + while (atomic_read(&sync->nr_tests_running) < running) { + set_current_state(TASK_UNINTERRUPTIBLE); + schedule_timeout(10); + } + /* Kick off all CPU concurrently on completion event */ + complete_all(&sync->start_event); + + /* Wait for CPUs to finish */ + while (atomic_read(&sync->nr_tests_running)) { + set_current_state(TASK_UNINTERRUPTIBLE); + schedule_timeout(10); + } + + /* Stop the kthreads */ + for_each_cpu(cpu, mask) { + struct time_bench_cpu *c = &cpu_tasks[cpu]; + + kthread_stop(c->task); + } + + if (verbose) // DEBUG - happens often, finish on another CPU + pr_warn("%s() Finished on CPU:%d\n", __func__, + smp_processor_id()); +} diff --git a/tools/testing/selftests/net/bench/page_pool/time_bench.h b/tools/testing/selftests/net/bench/page_pool/time_bench.h new file mode 100644 index 000000000000..e113fcf341dc --- /dev/null +++ b/tools/testing/selftests/net/bench/page_pool/time_bench.h @@ -0,0 +1,238 @@ +/* SPDX-License-Identifier: GPL-2.0-only */ +/* + * Benchmarking code execution time inside the kernel + * + * Copyright (C) 2014, Red Hat, Inc., Jesper Dangaard Brouer + * for licensing details see kernel-base/COPYING + */ +#ifndef _LINUX_TIME_BENCH_H +#define _LINUX_TIME_BENCH_H + +/* Main structure used for recording a benchmark run */ +struct time_bench_record { + uint32_t version_abi; + uint32_t loops; /* Requested loop invocations */ + uint32_t step; /* option for e.g. bulk invocations */ + + uint32_t flags; /* Measurements types enabled */ +#define TIME_BENCH_LOOP BIT(0) +#define TIME_BENCH_TSC BIT(1) +#define TIME_BENCH_WALLCLOCK BIT(2) +#define TIME_BENCH_PMU BIT(3) + + uint32_t cpu; /* Used when embedded in time_bench_cpu */ + + /* Records */ + uint64_t invoked_cnt; /* Returned actual invocations */ + uint64_t tsc_start; + uint64_t tsc_stop; + struct timespec64 ts_start; + struct timespec64 ts_stop; + /* PMU counters for instruction and cycles + * instructions counter including pipelined instructions + */ + uint64_t pmc_inst_start; + uint64_t pmc_inst_stop; + /* CPU unhalted clock counter */ + uint64_t pmc_clk_start; + uint64_t pmc_clk_stop; + + /* Result records */ + uint64_t tsc_interval; + uint64_t time_start, time_stop, time_interval; /* in nanosec */ + uint64_t pmc_inst, pmc_clk; + + /* Derived result records */ + uint64_t tsc_cycles; // +decimal? + uint64_t ns_per_call_quotient, ns_per_call_decimal; + uint64_t time_sec; + uint32_t time_sec_remainder; + uint64_t pmc_ipc_quotient, pmc_ipc_decimal; /* inst per cycle */ +}; + +/* For synchronizing parallel CPUs to run concurrently */ +struct time_bench_sync { + atomic_t nr_tests_running; + struct completion start_event; +}; + +/* Keep track of CPUs executing our bench function. + * + * Embed a time_bench_record for storing info per cpu + */ +struct time_bench_cpu { + struct time_bench_record rec; + struct time_bench_sync *sync; /* back ptr */ + struct task_struct *task; + /* "data" opaque could have been placed in time_bench_sync, + * but to avoid any false sharing, place it per CPU + */ + void *data; + /* Support masking outsome CPUs, mark if it ran */ + bool did_bench_run; + /* int cpu; // note CPU stored in time_bench_record */ + int (*bench_func)(struct time_bench_record *record, void *data); +}; + +/* + * Below TSC assembler code is not compatible with other archs, and + * can also fail on guests if cpu-flags are not correct. + * + * The way TSC reading is used, many iterations, does not require as + * high accuracy as described below (in Intel Doc #324264). + * + * Considering changing to use get_cycles() (#include <asm/timex.h>). + */ + +/** TSC (Time-Stamp Counter) based ** + * Recommend reading, to understand details of reading TSC accurately: + * Intel Doc #324264, "How to Benchmark Code Execution Times on Intel" + * + * Consider getting exclusive ownership of CPU by using: + * unsigned long flags; + * preempt_disable(); + * raw_local_irq_save(flags); + * _your_code_ + * raw_local_irq_restore(flags); + * preempt_enable(); + * + * Clobbered registers: "%rax", "%rbx", "%rcx", "%rdx" + * RDTSC only change "%rax" and "%rdx" but + * CPUID clears the high 32-bits of all (rax/rbx/rcx/rdx) + */ +static __always_inline uint64_t tsc_start_clock(void) +{ + /* See: Intel Doc #324264 */ + unsigned int hi, lo; + + asm volatile("CPUID\n\t" + "RDTSC\n\t" + "mov %%edx, %0\n\t" + "mov %%eax, %1\n\t" + : "=r"(hi), "=r"(lo)::"%rax", "%rbx", "%rcx", "%rdx"); + //FIXME: on 32bit use clobbered %eax + %edx + return ((uint64_t)lo) | (((uint64_t)hi) << 32); +} + +static __always_inline uint64_t tsc_stop_clock(void) +{ + /* See: Intel Doc #324264 */ + unsigned int hi, lo; + + asm volatile("RDTSCP\n\t" + "mov %%edx, %0\n\t" + "mov %%eax, %1\n\t" + "CPUID\n\t" + : "=r"(hi), "=r"(lo)::"%rax", "%rbx", "%rcx", "%rdx"); + return ((uint64_t)lo) | (((uint64_t)hi) << 32); +} + +/** Wall-clock based ** + * + * use: getnstimeofday() + * getnstimeofday(&rec->ts_start); + * getnstimeofday(&rec->ts_stop); + * + * API changed see: Documentation/core-api/timekeeping.rst + * https://www.kernel.org/doc/html/latest/core-api/timekeeping.html#c.getnstime... + * + * We should instead use: ktime_get_real_ts64() is a direct + * replacement, but consider using monotonic time (ktime_get_ts64()) + * and/or a ktime_t based interface (ktime_get()/ktime_get_real()). + */ + +/** PMU (Performance Monitor Unit) based ** + * + * Needed for calculating: Instructions Per Cycle (IPC) + * - The IPC number tell how efficient the CPU pipelining were + */ +//lookup: perf_event_create_kernel_counter() + +bool time_bench_PMU_config(bool enable); + +/* Raw reading via rdpmc() using fixed counters + * + * From: https://github.com/andikleen/simple-pmu + */ +enum { + FIXED_SELECT = (1U << 30), /* == 0x40000000 */ + FIXED_INST_RETIRED_ANY = 0, + FIXED_CPU_CLK_UNHALTED_CORE = 1, + FIXED_CPU_CLK_UNHALTED_REF = 2, +}; + +static __always_inline unsigned int long long p_rdpmc(unsigned int in) +{ + unsigned int d, a; + + asm volatile("rdpmc" : "=d"(d), "=a"(a) : "c"(in) : "memory"); + return ((unsigned long long)d << 32) | a; +} + +/* These PMU counter needs to be enabled, but I don't have the + * configure code implemented. My current hack is running: + * sudo perf stat -e cycles:k -e instructions:k insmod lib/ring_queue_test.ko + */ +/* Reading all pipelined instruction */ +static __always_inline unsigned long long pmc_inst(void) +{ + return p_rdpmc(FIXED_SELECT | FIXED_INST_RETIRED_ANY); +} + +/* Reading CPU clock cycles */ +static __always_inline unsigned long long pmc_clk(void) +{ + return p_rdpmc(FIXED_SELECT | FIXED_CPU_CLK_UNHALTED_CORE); +} + +/* Raw reading via MSR rdmsr() is likely wrong + * FIXME: How can I know which raw MSR registers are conf for what? + */ +#define MSR_IA32_PCM0 0x400000C1 /* PERFCTR0 */ +#define MSR_IA32_PCM1 0x400000C2 /* PERFCTR1 */ +#define MSR_IA32_PCM2 0x400000C3 +static inline uint64_t msr_inst(unsigned long long *msr_result) +{ + return rdmsrq_safe(MSR_IA32_PCM0, msr_result); +} + +/** Generic functions ** + */ +bool time_bench_loop(uint32_t loops, int step, char *txt, void *data, + int (*func)(struct time_bench_record *rec, void *data)); +bool time_bench_calc_stats(struct time_bench_record *rec); + +void time_bench_run_concurrent(uint32_t loops, int step, void *data, + const struct cpumask *mask, /* Support masking outsome CPUs*/ + struct time_bench_sync *sync, struct time_bench_cpu *cpu_tasks, + int (*func)(struct time_bench_record *record, void *data)); +void time_bench_print_stats_cpumask(const char *desc, + struct time_bench_cpu *cpu_tasks, + const struct cpumask *mask); + +//FIXME: use rec->flags to select measurement, should be MACRO +static __always_inline void time_bench_start(struct time_bench_record *rec) +{ + //getnstimeofday(&rec->ts_start); + ktime_get_real_ts64(&rec->ts_start); + if (rec->flags & TIME_BENCH_PMU) { + rec->pmc_inst_start = pmc_inst(); + rec->pmc_clk_start = pmc_clk(); + } + rec->tsc_start = tsc_start_clock(); +} + +static __always_inline void time_bench_stop(struct time_bench_record *rec, + uint64_t invoked_cnt) +{ + rec->tsc_stop = tsc_stop_clock(); + if (rec->flags & TIME_BENCH_PMU) { + rec->pmc_inst_stop = pmc_inst(); + rec->pmc_clk_stop = pmc_clk(); + } + //getnstimeofday(&rec->ts_stop); + ktime_get_real_ts64(&rec->ts_stop); + rec->invoked_cnt = invoked_cnt; +} + +#endif /* _LINUX_TIME_BENCH_H */ diff --git a/tools/testing/selftests/net/bench/test_bench_page_pool.sh b/tools/testing/selftests/net/bench/test_bench_page_pool.sh new file mode 100755 index 000000000000..7b8b18cfedce --- /dev/null +++ b/tools/testing/selftests/net/bench/test_bench_page_pool.sh @@ -0,0 +1,32 @@ +#!/bin/bash +# SPDX-License-Identifier: GPL-2.0 +# + +set -e + +DRIVER="./page_pool/bench_page_pool.ko" +result="" + +function run_test() +{ + rmmod "bench_page_pool.ko" || true + insmod $DRIVER > /dev/null 2>&1 + result=$(dmesg | tail -10) + echo "$result" + + echo + echo "Fast path results:" + echo "${result}" | grep -o -E "no-softirq-page_pool01 Per elem: ([0-9]+) cycles(tsc) ([0-9]+.[0-9]+) ns" + + echo + echo "ptr_ring results:" + echo "${result}" | grep -o -E "no-softirq-page_pool02 Per elem: ([0-9]+) cycles(tsc) ([0-9]+.[0-9]+) ns" + + echo + echo "slow path results:" + echo "${result}" | grep -o -E "no-softirq-page_pool03 Per elem: ([0-9]+) cycles(tsc) ([0-9]+.[0-9]+) ns" +} + +run_test + +exit 0
base-commit: 8909f5f4ecd551c2299b28e05254b77424c8c7dc
On 15/06/2025 22.59, Mina Almasry wrote:
From: Jesper Dangaard Brouer hawk@kernel.org
We frequently consult with Jesper's out-of-tree page_pool benchmark to evaluate page_pool changes.
Import the benchmark into the upstream linux kernel tree so that (a) we're all running the same version, (b) pave the way for shared improvements, and (c) maybe one day integrate it with nipa, if possible.
Import bench_page_pool_simple from commit 35b1716d0c30 ("Add page_bench06_walk_all"), from this repository: https://github.com/netoptimizer/prototype-kernel.git
Changes done during upstreaming:
- Fix checkpatch issues.
- Remove the tasklet logic not needed.
- Move under tools/testing
- Create ksft for the benchmark.
- Changed slightly how the benchmark gets build. Out of tree, time_bench is built as an independent .ko. Here it is included in bench_page_pool.ko
Steps to run:
mkdir -p /tmp/run-pp-bench make -C ./tools/testing/selftests/net/bench make -C ./tools/testing/selftests/net/bench install INSTALL_PATH=/tmp/run-pp-bench rsync --delete -avz --progress /tmp/run-pp-bench mina@$SERVER:~/ ssh mina@$SERVER << EOF cd ~/run-pp-bench && sudo ./test_bench_page_pool.sh EOFOutput:
(benchmrk dmesg logs)
Something is off with benchmark numbers compared to the OOT version.
Adding my numbers below, they were run on my testlab with: - CPU E5-1650 v4 @ 3.60GHz - kernel: net.git v6.15-12438-gd9816ec74e6d
Fast path results: no-softirq-page_pool01 Per elem: 11 cycles(tsc) 4.368 ns
Fast-path on your CPU is faster (22 cycles(tsc) 6.128 ns) than my CPU. What CPU is this?
Type:no-softirq-page_pool01 Per elem: 22 cycles(tsc) 6.128 ns (step:0) - (measurement period time:0.061282924 sec time_interval:61282924) - (invoke count:10000000 tsc_interval:220619745)
ptr_ring results: no-softirq-page_pool02 Per elem: 527 cycles(tsc) 195.187 ns
I'm surprised that ptr_ring benchmark is very slow, compared to my result (below) 60 cycles(tsc) 16.853 ns.
Type:no-softirq-page_pool02 Per elem: 60 cycles(tsc) 16.853 ns (step:0) - (measurement period time:0.168535760 sec time_interval:168535760) - (invoke count:10000000 tsc_interval:606734160)
Maybe your kernel is compiled with some CONFIG debug thing that makes it slower?
You can troubleshoot like this: - select the `no-softirq-page_pool02` test via run_flags=$((2#100)).
# perf record -g modprobe bench_page_pool_simple run_flags=$((2#100)) loops=$((100*10**6)) # perf report --no-children
slow path results: no-softirq-page_pool03 Per elem: 549 cycles(tsc) 203.466 ns
Type:no-softirq-page_pool03 Per elem: 265 cycles(tsc) 73.674 ns (step:0) - (measurement period time:0.736740796 sec time_interval:736740796) - (invoke count:10000000 tsc_interval:2652295113)
--Jesper
Cc: Jesper Dangaard Brouer <hawk@kernel.org> Cc: Ilias Apalodimas <ilias.apalodimas@linaro.org> Cc: Jakub Kicinski <kuba@kernel.org> Cc: Toke Høiland-Jørgensen <toke@toke.dk> Signed-off-by: Mina Almasry <almasrymina@google.com> Acked-by: Ilias Apalodimas <ilias.apalodimas@linaro.org> Signed-off-by: Jesper Dangaard Brouer <hawk@kernel.org> --- v4: https://lore.kernel.org/netdev/20250614100853.3f2372f2@kernel.org/ - Fix more checkpatch and coccicheck issues (Jakub) v3: - Non RFC - Collect Signed-off-by from Jesper and Acked-by Ilias. - Move test_bench_page_pool.sh to address nipa complaint. - Remove `static inline` in .c files to address nipa complaint. v2: - Move under tools/selftests (Jakub) - Create ksft for it. - Remove the tasklet logic no longer needed (Jesper + Toke) RFC discussion points: - Desirable to import it? - Can the benchmark be imported as-is for an initial version? Or needs lots of modifications? - Code location. I retained the location in Jesper's tree, but a path like net/core/bench/ may make more sense. --- tools/testing/selftests/net/bench/Makefile | 7 + .../selftests/net/bench/page_pool/Makefile | 17 + .../bench/page_pool/bench_page_pool_simple.c | 276 ++++++++++++ .../net/bench/page_pool/time_bench.c | 394 ++++++++++++++++++ .../net/bench/page_pool/time_bench.h | 238 +++++++++++ .../net/bench/test_bench_page_pool.sh | 32 ++ 6 files changed, 964 insertions(+) create mode 100644 tools/testing/selftests/net/bench/Makefile create mode 100644 tools/testing/selftests/net/bench/page_pool/Makefile create mode 100644 tools/testing/selftests/net/bench/page_pool/bench_page_pool_simple.c create mode 100644 tools/testing/selftests/net/bench/page_pool/time_bench.c create mode 100644 tools/testing/selftests/net/bench/page_pool/time_bench.h create mode 100755 tools/testing/selftests/net/bench/test_bench_page_pool.sh diff --git a/tools/testing/selftests/net/bench/Makefile b/tools/testing/selftests/net/bench/Makefile new file mode 100644 index 000000000000..2546c45e42f7 --- /dev/null +++ b/tools/testing/selftests/net/bench/Makefile @@ -0,0 +1,7 @@ +# SPDX-License-Identifier: GPL-2.0 + +TEST_GEN_MODS_DIR := page_pool + +TEST_PROGS += test_bench_page_pool.sh + +include ../../lib.mk diff --git a/tools/testing/selftests/net/bench/page_pool/Makefile b/tools/testing/selftests/net/bench/page_pool/Makefile new file mode 100644 index 000000000000..0549a16ba275 --- /dev/null +++ b/tools/testing/selftests/net/bench/page_pool/Makefile @@ -0,0 +1,17 @@ +BENCH_PAGE_POOL_SIMPLE_TEST_DIR := $(realpath $(dir $(abspath $(lastword $(MAKEFILE_LIST))))) +KDIR ?= /lib/modules/$(shell uname -r)/build + +ifeq ($(V),1) +Q = +else +Q = @ +endif + +obj-m += bench_page_pool.o +bench_page_pool-y += bench_page_pool_simple.o time_bench.o + +all: + +$(Q)make -C $(KDIR) M=$(BENCH_PAGE_POOL_SIMPLE_TEST_DIR) modules + +clean: + +$(Q)make -C $(KDIR) M=$(BENCH_PAGE_POOL_SIMPLE_TEST_DIR) clean diff --git a/tools/testing/selftests/net/bench/page_pool/bench_page_pool_simple.c b/tools/testing/selftests/net/bench/page_pool/bench_page_pool_simple.c new file mode 100644 index 000000000000..f183d5e30dc6 --- /dev/null +++ b/tools/testing/selftests/net/bench/page_pool/bench_page_pool_simple.c @@ -0,0 +1,276 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* + * Benchmark module for page_pool. + * + */ +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt + +#include <linux/module.h> +#include <linux/mutex.h> + +#include <linux/version.h> +#include <net/page_pool/helpers.h> + +#include <linux/interrupt.h> +#include <linux/limits.h> + +#include "time_bench.h" + +static int verbose = 1; +#define MY_POOL_SIZE 1024 + +static void _page_pool_put_page(struct page_pool *pool, struct page *page, + bool allow_direct) +{ + page_pool_put_page(pool, page, -1, allow_direct); +} + +/* Makes tests selectable. Useful for perf-record to analyze a single test. + * Hint: Bash shells support writing binary number like: $((2#101010) + * + * # modprobe bench_page_pool_simple run_flags=$((2#100)) + */ +static unsigned long run_flags = 0xFFFFFFFF; +module_param(run_flags, ulong, 0); +MODULE_PARM_DESC(run_flags, "Limit which bench test that runs"); + +/* Count the bit number from the enum */ +enum benchmark_bit { + bit_run_bench_baseline, + bit_run_bench_no_softirq01, + bit_run_bench_no_softirq02, + bit_run_bench_no_softirq03, +}; + +#define bit(b) (1 << (b)) +#define enabled(b) ((run_flags & (bit(b)))) + +/* notice time_bench is limited to U32_MAX nr loops */ +static unsigned long loops = 10000000; +module_param(loops, ulong, 0); +MODULE_PARM_DESC(loops, "Specify loops bench will run"); + +/* Timing at the nanosec level, we need to know the overhead + * introduced by the for loop itself + */ +static int time_bench_for_loop(struct time_bench_record *rec, void *data) +{ + uint64_t loops_cnt = 0; + int i; + + time_bench_start(rec); + /** Loop to measure **/ + for (i = 0; i < rec->loops; i++) { + loops_cnt++; + barrier(); /* avoid compiler to optimize this loop */ + } + time_bench_stop(rec, loops_cnt); + return loops_cnt; +} + +static int time_bench_atomic_inc(struct time_bench_record *rec, void *data) +{ + uint64_t loops_cnt = 0; + atomic_t cnt; + int i; + + atomic_set(&cnt, 0); + + time_bench_start(rec); + /** Loop to measure **/ + for (i = 0; i < rec->loops; i++) { + atomic_inc(&cnt); + barrier(); /* avoid compiler to optimize this loop */ + } + loops_cnt = atomic_read(&cnt); + time_bench_stop(rec, loops_cnt); + return loops_cnt; +} + +/* The ptr_ping in page_pool uses a spinlock. We need to know the minimum + * overhead of taking+releasing a spinlock, to know the cycles that can be saved + * by e.g. amortizing this via bulking. + */ +static int time_bench_lock(struct time_bench_record *rec, void *data) +{ + uint64_t loops_cnt = 0; + spinlock_t lock; + int i; + + spin_lock_init(&lock); + + time_bench_start(rec); + /** Loop to measure **/ + for (i = 0; i < rec->loops; i++) { + spin_lock(&lock); + loops_cnt++; + barrier(); /* avoid compiler to optimize this loop */ + spin_unlock(&lock); + } + time_bench_stop(rec, loops_cnt); + return loops_cnt; +} + +/* Helper for filling some page's into ptr_ring */ +static void pp_fill_ptr_ring(struct page_pool *pp, int elems) +{ + /* GFP_ATOMIC needed when under run softirq */ + gfp_t gfp_mask = GFP_ATOMIC; + struct page **array; + int i; + + array = kcalloc(elems, sizeof(struct page *), gfp_mask); + + for (i = 0; i < elems; i++) + array[i] = page_pool_alloc_pages(pp, gfp_mask); + for (i = 0; i < elems; i++) + _page_pool_put_page(pp, array[i], false); + + kfree(array); +} + +enum test_type { type_fast_path, type_ptr_ring, type_page_allocator }; + +/* Depends on compile optimizing this function */ +static int time_bench_page_pool(struct time_bench_record *rec, void *data, + enum test_type type, const char *func) +{ + uint64_t loops_cnt = 0; + gfp_t gfp_mask = GFP_ATOMIC; /* GFP_ATOMIC is not really needed */ + int i, err; + + struct page_pool *pp; + struct page *page; + + struct page_pool_params pp_params = { + .order = 0, + .flags = 0, + .pool_size = MY_POOL_SIZE, + .nid = NUMA_NO_NODE, + .dev = NULL, /* Only use for DMA mapping */ + .dma_dir = DMA_BIDIRECTIONAL, + }; + + pp = page_pool_create(&pp_params); + if (IS_ERR(pp)) { + err = PTR_ERR(pp); + pr_warn("%s: Error(%d) creating page_pool\n", func, err); + goto out; + } + pp_fill_ptr_ring(pp, 64); + + if (in_serving_softirq()) + pr_warn("%s(): in_serving_softirq fast-path\n", func); + else + pr_warn("%s(): Cannot use page_pool fast-path\n", func); + + time_bench_start(rec); + /** Loop to measure **/ + for (i = 0; i < rec->loops; i++) { + /* Common fast-path alloc that depend on in_serving_softirq() */ + page = page_pool_alloc_pages(pp, gfp_mask); + if (!page) + break; + loops_cnt++; + barrier(); /* avoid compiler to optimize this loop */ + + /* The benchmarks purpose it to test different return paths. + * Compiler should inline optimize other function calls out + */ + if (type == type_fast_path) { + /* Fast-path recycling e.g. XDP_DROP use-case */ + page_pool_recycle_direct(pp, page); + + } else if (type == type_ptr_ring) { + /* Normal return path */ + _page_pool_put_page(pp, page, false); + + } else if (type == type_page_allocator) { + /* Test if not pages are recycled, but instead + * returned back into systems page allocator + */ + get_page(page); /* cause no-recycling */ + _page_pool_put_page(pp, page, false); + put_page(page); + } else { + BUILD_BUG(); + } + } + time_bench_stop(rec, loops_cnt); +out: + page_pool_destroy(pp); + return loops_cnt; +} + +static int time_bench_page_pool01_fast_path(struct time_bench_record *rec, + void *data) +{ + return time_bench_page_pool(rec, data, type_fast_path, __func__); +} + +static int time_bench_page_pool02_ptr_ring(struct time_bench_record *rec, + void *data) +{ + return time_bench_page_pool(rec, data, type_ptr_ring, __func__); +} + +static int time_bench_page_pool03_slow(struct time_bench_record *rec, + void *data) +{ + return time_bench_page_pool(rec, data, type_page_allocator, __func__); +} + +static int run_benchmark_tests(void) +{ + uint32_t nr_loops = loops; + + /* Baseline tests */ + if (enabled(bit_run_bench_baseline)) { + time_bench_loop(nr_loops * 10, 0, "for_loop", NULL, + time_bench_for_loop); + time_bench_loop(nr_loops * 10, 0, "atomic_inc", NULL, + time_bench_atomic_inc); + time_bench_loop(nr_loops, 0, "lock", NULL, time_bench_lock); + } + + /* This test cannot activate correct code path, due to no-softirq ctx */ + if (enabled(bit_run_bench_no_softirq01)) + time_bench_loop(nr_loops, 0, "no-softirq-page_pool01", NULL, + time_bench_page_pool01_fast_path); + if (enabled(bit_run_bench_no_softirq02)) + time_bench_loop(nr_loops, 0, "no-softirq-page_pool02", NULL, + time_bench_page_pool02_ptr_ring); + if (enabled(bit_run_bench_no_softirq03)) + time_bench_loop(nr_loops, 0, "no-softirq-page_pool03", NULL, + time_bench_page_pool03_slow); + + return 0; +} + +static int __init bench_page_pool_simple_module_init(void) +{ + if (verbose) + pr_info("Loaded\n"); + + if (loops > U32_MAX) { + pr_err("Module param loops(%lu) exceeded U32_MAX(%u)\n", loops, + U32_MAX); + return -ECHRNG; + } + + run_benchmark_tests(); + + return 0; +} +module_init(bench_page_pool_simple_module_init); + +static void __exit bench_page_pool_simple_module_exit(void) +{ + if (verbose) + pr_info("Unloaded\n"); +} +module_exit(bench_page_pool_simple_module_exit); + +MODULE_DESCRIPTION("Benchmark of page_pool simple cases"); +MODULE_AUTHOR("Jesper Dangaard Brouer <netoptimizer@brouer.com>"); +MODULE_LICENSE("GPL"); diff --git a/tools/testing/selftests/net/bench/page_pool/time_bench.c b/tools/testing/selftests/net/bench/page_pool/time_bench.c new file mode 100644 index 000000000000..073bb36ec5f2 --- /dev/null +++ b/tools/testing/selftests/net/bench/page_pool/time_bench.c @@ -0,0 +1,394 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* + * Benchmarking code execution time inside the kernel + * + * Copyright (C) 2014, Red Hat, Inc., Jesper Dangaard Brouer + */ +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt + +#include <linux/module.h> +#include <linux/time.h> + +#include <linux/perf_event.h> /* perf_event_create_kernel_counter() */ + +/* For concurrency testing */ +#include <linux/completion.h> +#include <linux/sched.h> +#include <linux/workqueue.h> +#include <linux/kthread.h> + +#include "time_bench.h" + +static int verbose = 1; + +/** TSC (Time-Stamp Counter) based ** + * See: linux/time_bench.h + * tsc_start_clock() and tsc_stop_clock() + */ + +/** Wall-clock based ** + */ + +/** PMU (Performance Monitor Unit) based ** + */ +#define PERF_FORMAT \ + (PERF_FORMAT_GROUP | PERF_FORMAT_ID | PERF_FORMAT_TOTAL_TIME_ENABLED | \ + PERF_FORMAT_TOTAL_TIME_RUNNING) + +struct raw_perf_event { + uint64_t config; /* event */ + uint64_t config1; /* umask */ + struct perf_event *save; + char *desc; +}; + +/* if HT is enable a maximum of 4 events (5 if one is instructions + * retired can be specified, if HT is disabled a maximum of 8 (9 if + * one is instructions retired) can be specified. + * + * From Table 19-1. Architectural Performance Events + * Architectures Software Developer’s Manual Volume 3: System Programming + * Guide + */ +struct raw_perf_event perf_events[] = { + { 0x3c, 0x00, NULL, "Unhalted CPU Cycles" }, + { 0xc0, 0x00, NULL, "Instruction Retired" } +}; + +#define NUM_EVTS (ARRAY_SIZE(perf_events)) + +/* WARNING: PMU config is currently broken! + */ +bool time_bench_PMU_config(bool enable) +{ + int i; + struct perf_event_attr perf_conf; + struct perf_event *perf_event; + int cpu; + + preempt_disable(); + cpu = smp_processor_id(); + pr_info("DEBUG: cpu:%d\n", cpu); + preempt_enable(); + + memset(&perf_conf, 0, sizeof(struct perf_event_attr)); + perf_conf.type = PERF_TYPE_RAW; + perf_conf.size = sizeof(struct perf_event_attr); + perf_conf.read_format = PERF_FORMAT; + perf_conf.pinned = 1; + perf_conf.exclude_user = 1; /* No userspace events */ + perf_conf.exclude_kernel = 0; /* Only kernel events */ + + for (i = 0; i < NUM_EVTS; i++) { + perf_conf.disabled = enable; + //perf_conf.disabled = (i == 0) ? 1 : 0; + perf_conf.config = perf_events[i].config; + perf_conf.config1 = perf_events[i].config1; + if (verbose) + pr_info("%s() enable PMU counter: %s\n", + __func__, perf_events[i].desc); + perf_event = perf_event_create_kernel_counter(&perf_conf, cpu, + NULL /* task */, + NULL /* overflow_handler*/, + NULL /* context */); + if (perf_event) { + perf_events[i].save = perf_event; + pr_info("%s():DEBUG perf_event success\n", __func__); + + perf_event_enable(perf_event); + } else { + pr_info("%s():DEBUG perf_event is NULL\n", __func__); + } + } + + return true; +} + +/** Generic functions ** + */ + +/* Calculate stats, store results in record */ +bool time_bench_calc_stats(struct time_bench_record *rec) +{ +#define NANOSEC_PER_SEC 1000000000 /* 10^9 */ + uint64_t ns_per_call_tmp_rem = 0; + uint32_t ns_per_call_remainder = 0; + uint64_t pmc_ipc_tmp_rem = 0; + uint32_t pmc_ipc_remainder = 0; + uint32_t pmc_ipc_div = 0; + uint32_t invoked_cnt_precision = 0; + uint32_t invoked_cnt = 0; /* 32-bit due to div_u64_rem() */ + + if (rec->flags & TIME_BENCH_LOOP) { + if (rec->invoked_cnt < 1000) { + pr_err("ERR: need more(>1000) loops(%llu) for timing\n", + rec->invoked_cnt); + return false; + } + if (rec->invoked_cnt > ((1ULL << 32) - 1)) { + /* div_u64_rem() can only support div with 32bit*/ + pr_err("ERR: Invoke cnt(%llu) too big overflow 32bit\n", + rec->invoked_cnt); + return false; + } + invoked_cnt = (uint32_t)rec->invoked_cnt; + } + + /* TSC (Time-Stamp Counter) records */ + if (rec->flags & TIME_BENCH_TSC) { + rec->tsc_interval = rec->tsc_stop - rec->tsc_start; + if (rec->tsc_interval == 0) { + pr_err("ABORT: timing took ZERO TSC time\n"); + return false; + } + /* Calculate stats */ + if (rec->flags & TIME_BENCH_LOOP) + rec->tsc_cycles = rec->tsc_interval / invoked_cnt; + else + rec->tsc_cycles = rec->tsc_interval; + } + + /* Wall-clock time calc */ + if (rec->flags & TIME_BENCH_WALLCLOCK) { + rec->time_start = rec->ts_start.tv_nsec + + (NANOSEC_PER_SEC * rec->ts_start.tv_sec); + rec->time_stop = rec->ts_stop.tv_nsec + + (NANOSEC_PER_SEC * rec->ts_stop.tv_sec); + rec->time_interval = rec->time_stop - rec->time_start; + if (rec->time_interval == 0) { + pr_err("ABORT: timing took ZERO wallclock time\n"); + return false; + } + /* Calculate stats */ + /*** Division in kernel it tricky ***/ + /* Orig: time_sec = (time_interval / NANOSEC_PER_SEC); */ + /* remainder only correct because NANOSEC_PER_SEC is 10^9 */ + rec->time_sec = div_u64_rem(rec->time_interval, NANOSEC_PER_SEC, + &rec->time_sec_remainder); + //TODO: use existing struct timespec records instead of div? + + if (rec->flags & TIME_BENCH_LOOP) { + /*** Division in kernel it tricky ***/ + /* Orig: ns = ((double)time_interval / invoked_cnt); */ + /* First get quotient */ + rec->ns_per_call_quotient = + div_u64_rem(rec->time_interval, invoked_cnt, + &ns_per_call_remainder); + /* Now get decimals .xxx precision (incorrect roundup)*/ + ns_per_call_tmp_rem = ns_per_call_remainder; + invoked_cnt_precision = invoked_cnt / 1000; + if (invoked_cnt_precision > 0) { + rec->ns_per_call_decimal = + div_u64_rem(ns_per_call_tmp_rem, + invoked_cnt_precision, + &ns_per_call_remainder); + } + } + } + + /* Performance Monitor Unit (PMU) counters */ + if (rec->flags & TIME_BENCH_PMU) { + //FIXME: Overflow handling??? + rec->pmc_inst = rec->pmc_inst_stop - rec->pmc_inst_start; + rec->pmc_clk = rec->pmc_clk_stop - rec->pmc_clk_start; + + /* Calc Instruction Per Cycle (IPC) */ + /* First get quotient */ + rec->pmc_ipc_quotient = div_u64_rem(rec->pmc_inst, rec->pmc_clk, + &pmc_ipc_remainder); + /* Now get decimals .xxx precision (incorrect roundup)*/ + pmc_ipc_tmp_rem = pmc_ipc_remainder; + pmc_ipc_div = rec->pmc_clk / 1000; + if (pmc_ipc_div > 0) { + rec->pmc_ipc_decimal = div_u64_rem(pmc_ipc_tmp_rem, + pmc_ipc_div, + &pmc_ipc_remainder); + } + } + + return true; +} + +/* Generic function for invoking a loop function and calculating + * execution time stats. The function being called/timed is assumed + * to perform a tight loop, and update the timing record struct. + */ +bool time_bench_loop(uint32_t loops, int step, char *txt, void *data, + int (*func)(struct time_bench_record *record, void *data)) +{ + struct time_bench_record rec; + + /* Setup record */ + memset(&rec, 0, sizeof(rec)); /* zero func might not update all */ + rec.version_abi = 1; + rec.loops = loops; + rec.step = step; + rec.flags = (TIME_BENCH_LOOP | TIME_BENCH_TSC | TIME_BENCH_WALLCLOCK); + + /*** Loop function being timed ***/ + if (!func(&rec, data)) { + pr_err("ABORT: function being timed failed\n"); + return false; + } + + if (rec.invoked_cnt < loops) + pr_warn("WARNING: Invoke count(%llu) smaller than loops(%d)\n", + rec.invoked_cnt, loops); + + /* Calculate stats */ + time_bench_calc_stats(&rec); + + pr_info("Type:%s Per elem: %llu cycles(tsc) %llu.%03llu ns (step:%d) - (measurement period time:%llu.%09u sec time_interval:%llu) - (invoke count:%llu tsc_interval:%llu)\n", + txt, rec.tsc_cycles, rec.ns_per_call_quotient, + rec.ns_per_call_decimal, rec.step, rec.time_sec, + rec.time_sec_remainder, rec.time_interval, rec.invoked_cnt, + rec.tsc_interval); + if (rec.flags & TIME_BENCH_PMU) + pr_info("Type:%s PMU inst/clock%llu/%llu = %llu.%03llu IPC (inst per cycle)\n", + txt, rec.pmc_inst, rec.pmc_clk, rec.pmc_ipc_quotient, + rec.pmc_ipc_decimal); + return true; +} + +/* Function getting invoked by kthread */ +static int invoke_test_on_cpu_func(void *private) +{ + struct time_bench_cpu *cpu = private; + struct time_bench_sync *sync = cpu->sync; + cpumask_t newmask = CPU_MASK_NONE; + void *data = cpu->data; + + /* Restrict CPU */ + cpumask_set_cpu(cpu->rec.cpu, &newmask); + set_cpus_allowed_ptr(current, &newmask); + + /* Synchronize start of concurrency test */ + atomic_inc(&sync->nr_tests_running); + wait_for_completion(&sync->start_event); + + /* Start benchmark function */ + if (!cpu->bench_func(&cpu->rec, data)) { + pr_err("ERROR: function being timed failed on CPU:%d(%d)\n", + cpu->rec.cpu, smp_processor_id()); + } else { + if (verbose) + pr_info("SUCCESS: ran on CPU:%d(%d)\n", cpu->rec.cpu, + smp_processor_id()); + } + cpu->did_bench_run = true; + + /* End test */ + atomic_dec(&sync->nr_tests_running); + /* Wait for kthread_stop() telling us to stop */ + while (!kthread_should_stop()) { + set_current_state(TASK_INTERRUPTIBLE); + schedule(); + } + __set_current_state(TASK_RUNNING); + return 0; +} + +void time_bench_print_stats_cpumask(const char *desc, + struct time_bench_cpu *cpu_tasks, + const struct cpumask *mask) +{ + uint64_t average = 0; + int cpu; + int step = 0; + struct sum { + uint64_t tsc_cycles; + int records; + } sum = { 0 }; + + /* Get stats */ + for_each_cpu(cpu, mask) { + struct time_bench_cpu *c = &cpu_tasks[cpu]; + struct time_bench_record *rec = &c->rec; + + /* Calculate stats */ + time_bench_calc_stats(rec); + + pr_info("Type:%s CPU(%d) %llu cycles(tsc) %llu.%03llu ns (step:%d) - (measurement period time:%llu.%09u sec time_interval:%llu) - (invoke count:%llu tsc_interval:%llu)\n", + desc, cpu, rec->tsc_cycles, rec->ns_per_call_quotient, + rec->ns_per_call_decimal, rec->step, rec->time_sec, + rec->time_sec_remainder, rec->time_interval, + rec->invoked_cnt, rec->tsc_interval); + + /* Collect average */ + sum.records++; + sum.tsc_cycles += rec->tsc_cycles; + step = rec->step; + } + + if (sum.records) /* avoid div-by-zero */ + average = sum.tsc_cycles / sum.records; + pr_info("Sum Type:%s Average: %llu cycles(tsc) CPUs:%d step:%d\n", desc, + average, sum.records, step); +} + +void time_bench_run_concurrent(uint32_t loops, int step, void *data, + const struct cpumask *mask, /* Support masking outsome CPUs*/ + struct time_bench_sync *sync, + struct time_bench_cpu *cpu_tasks, + int (*func)(struct time_bench_record *record, void *data)) +{ + int cpu, running = 0; + + if (verbose) // DEBUG + pr_warn("%s() Started on CPU:%d\n", __func__, + smp_processor_id()); + + /* Reset sync conditions */ + atomic_set(&sync->nr_tests_running, 0); + init_completion(&sync->start_event); + + /* Spawn off jobs on all CPUs */ + for_each_cpu(cpu, mask) { + struct time_bench_cpu *c = &cpu_tasks[cpu]; + + running++; + c->sync = sync; /* Send sync variable along */ + c->data = data; /* Send opaque along */ + + /* Init benchmark record */ + memset(&c->rec, 0, sizeof(struct time_bench_record)); + c->rec.version_abi = 1; + c->rec.loops = loops; + c->rec.step = step; + c->rec.flags = (TIME_BENCH_LOOP | TIME_BENCH_TSC | + TIME_BENCH_WALLCLOCK); + c->rec.cpu = cpu; + c->bench_func = func; + c->task = kthread_run(invoke_test_on_cpu_func, c, + "time_bench%d", cpu); + if (IS_ERR(c->task)) { + pr_err("%s(): Failed to start test func\n", __func__); + return; /* Argh, what about cleanup?! */ + } + } + + /* Wait until all processes are running */ + while (atomic_read(&sync->nr_tests_running) < running) { + set_current_state(TASK_UNINTERRUPTIBLE); + schedule_timeout(10); + } + /* Kick off all CPU concurrently on completion event */ + complete_all(&sync->start_event); + + /* Wait for CPUs to finish */ + while (atomic_read(&sync->nr_tests_running)) { + set_current_state(TASK_UNINTERRUPTIBLE); + schedule_timeout(10); + } + + /* Stop the kthreads */ + for_each_cpu(cpu, mask) { + struct time_bench_cpu *c = &cpu_tasks[cpu]; + + kthread_stop(c->task); + } + + if (verbose) // DEBUG - happens often, finish on another CPU + pr_warn("%s() Finished on CPU:%d\n", __func__, + smp_processor_id()); +} diff --git a/tools/testing/selftests/net/bench/page_pool/time_bench.h b/tools/testing/selftests/net/bench/page_pool/time_bench.h new file mode 100644 index 000000000000..e113fcf341dc --- /dev/null +++ b/tools/testing/selftests/net/bench/page_pool/time_bench.h @@ -0,0 +1,238 @@ +/* SPDX-License-Identifier: GPL-2.0-only */ +/* + * Benchmarking code execution time inside the kernel + * + * Copyright (C) 2014, Red Hat, Inc., Jesper Dangaard Brouer + * for licensing details see kernel-base/COPYING + */ +#ifndef _LINUX_TIME_BENCH_H +#define _LINUX_TIME_BENCH_H + +/* Main structure used for recording a benchmark run */ +struct time_bench_record { + uint32_t version_abi; + uint32_t loops; /* Requested loop invocations */ + uint32_t step; /* option for e.g. bulk invocations */ + + uint32_t flags; /* Measurements types enabled */ +#define TIME_BENCH_LOOP BIT(0) +#define TIME_BENCH_TSC BIT(1) +#define TIME_BENCH_WALLCLOCK BIT(2) +#define TIME_BENCH_PMU BIT(3) + + uint32_t cpu; /* Used when embedded in time_bench_cpu */ + + /* Records */ + uint64_t invoked_cnt; /* Returned actual invocations */ + uint64_t tsc_start; + uint64_t tsc_stop; + struct timespec64 ts_start; + struct timespec64 ts_stop; + /* PMU counters for instruction and cycles + * instructions counter including pipelined instructions + */ + uint64_t pmc_inst_start; + uint64_t pmc_inst_stop; + /* CPU unhalted clock counter */ + uint64_t pmc_clk_start; + uint64_t pmc_clk_stop; + + /* Result records */ + uint64_t tsc_interval; + uint64_t time_start, time_stop, time_interval; /* in nanosec */ + uint64_t pmc_inst, pmc_clk; + + /* Derived result records */ + uint64_t tsc_cycles; // +decimal? + uint64_t ns_per_call_quotient, ns_per_call_decimal; + uint64_t time_sec; + uint32_t time_sec_remainder; + uint64_t pmc_ipc_quotient, pmc_ipc_decimal; /* inst per cycle */ +}; + +/* For synchronizing parallel CPUs to run concurrently */ +struct time_bench_sync { + atomic_t nr_tests_running; + struct completion start_event; +}; + +/* Keep track of CPUs executing our bench function. + * + * Embed a time_bench_record for storing info per cpu + */ +struct time_bench_cpu { + struct time_bench_record rec; + struct time_bench_sync *sync; /* back ptr */ + struct task_struct *task; + /* "data" opaque could have been placed in time_bench_sync, + * but to avoid any false sharing, place it per CPU + */ + void *data; + /* Support masking outsome CPUs, mark if it ran */ + bool did_bench_run; + /* int cpu; // note CPU stored in time_bench_record */ + int (*bench_func)(struct time_bench_record *record, void *data); +}; + +/* + * Below TSC assembler code is not compatible with other archs, and + * can also fail on guests if cpu-flags are not correct. + * + * The way TSC reading is used, many iterations, does not require as + * high accuracy as described below (in Intel Doc #324264). + * + * Considering changing to use get_cycles() (#include <asm/timex.h>). + */ + +/** TSC (Time-Stamp Counter) based ** + * Recommend reading, to understand details of reading TSC accurately: + * Intel Doc #324264, "How to Benchmark Code Execution Times on Intel" + * + * Consider getting exclusive ownership of CPU by using: + * unsigned long flags; + * preempt_disable(); + * raw_local_irq_save(flags); + * _your_code_ + * raw_local_irq_restore(flags); + * preempt_enable(); + * + * Clobbered registers: "%rax", "%rbx", "%rcx", "%rdx" + * RDTSC only change "%rax" and "%rdx" but + * CPUID clears the high 32-bits of all (rax/rbx/rcx/rdx) + */ +static __always_inline uint64_t tsc_start_clock(void) +{ + /* See: Intel Doc #324264 */ + unsigned int hi, lo; + + asm volatile("CPUID\n\t" + "RDTSC\n\t" + "mov %%edx, %0\n\t" + "mov %%eax, %1\n\t" + : "=r"(hi), "=r"(lo)::"%rax", "%rbx", "%rcx", "%rdx"); + //FIXME: on 32bit use clobbered %eax + %edx + return ((uint64_t)lo) | (((uint64_t)hi) << 32); +} + +static __always_inline uint64_t tsc_stop_clock(void) +{ + /* See: Intel Doc #324264 */ + unsigned int hi, lo; + + asm volatile("RDTSCP\n\t" + "mov %%edx, %0\n\t" + "mov %%eax, %1\n\t" + "CPUID\n\t" + : "=r"(hi), "=r"(lo)::"%rax", "%rbx", "%rcx", "%rdx"); + return ((uint64_t)lo) | (((uint64_t)hi) << 32); +} + +/** Wall-clock based ** + * + * use: getnstimeofday() + * getnstimeofday(&rec->ts_start); + * getnstimeofday(&rec->ts_stop); + * + * API changed see: Documentation/core-api/timekeeping.rst + * https://www.kernel.org/doc/html/latest/core-api/timekeeping.html#c.getnstimeofday + * + * We should instead use: ktime_get_real_ts64() is a direct + * replacement, but consider using monotonic time (ktime_get_ts64()) + * and/or a ktime_t based interface (ktime_get()/ktime_get_real()). + */ + +/** PMU (Performance Monitor Unit) based ** + * + * Needed for calculating: Instructions Per Cycle (IPC) + * - The IPC number tell how efficient the CPU pipelining were + */ +//lookup: perf_event_create_kernel_counter() + +bool time_bench_PMU_config(bool enable); + +/* Raw reading via rdpmc() using fixed counters + * + * From: https://github.com/andikleen/simple-pmu + */ +enum { + FIXED_SELECT = (1U << 30), /* == 0x40000000 */ + FIXED_INST_RETIRED_ANY = 0, + FIXED_CPU_CLK_UNHALTED_CORE = 1, + FIXED_CPU_CLK_UNHALTED_REF = 2, +}; + +static __always_inline unsigned int long long p_rdpmc(unsigned int in) +{ + unsigned int d, a; + + asm volatile("rdpmc" : "=d"(d), "=a"(a) : "c"(in) : "memory"); + return ((unsigned long long)d << 32) | a; +} + +/* These PMU counter needs to be enabled, but I don't have the + * configure code implemented. My current hack is running: + * sudo perf stat -e cycles:k -e instructions:k insmod lib/ring_queue_test.ko + */ +/* Reading all pipelined instruction */ +static __always_inline unsigned long long pmc_inst(void) +{ + return p_rdpmc(FIXED_SELECT | FIXED_INST_RETIRED_ANY); +} + +/* Reading CPU clock cycles */ +static __always_inline unsigned long long pmc_clk(void) +{ + return p_rdpmc(FIXED_SELECT | FIXED_CPU_CLK_UNHALTED_CORE); +} + +/* Raw reading via MSR rdmsr() is likely wrong + * FIXME: How can I know which raw MSR registers are conf for what? + */ +#define MSR_IA32_PCM0 0x400000C1 /* PERFCTR0 */ +#define MSR_IA32_PCM1 0x400000C2 /* PERFCTR1 */ +#define MSR_IA32_PCM2 0x400000C3 +static inline uint64_t msr_inst(unsigned long long *msr_result) +{ + return rdmsrq_safe(MSR_IA32_PCM0, msr_result); +} + +/** Generic functions ** + */ +bool time_bench_loop(uint32_t loops, int step, char *txt, void *data, + int (*func)(struct time_bench_record *rec, void *data)); +bool time_bench_calc_stats(struct time_bench_record *rec); + +void time_bench_run_concurrent(uint32_t loops, int step, void *data, + const struct cpumask *mask, /* Support masking outsome CPUs*/ + struct time_bench_sync *sync, struct time_bench_cpu *cpu_tasks, + int (*func)(struct time_bench_record *record, void *data)); +void time_bench_print_stats_cpumask(const char *desc, + struct time_bench_cpu *cpu_tasks, + const struct cpumask *mask); + +//FIXME: use rec->flags to select measurement, should be MACRO +static __always_inline void time_bench_start(struct time_bench_record *rec) +{ + //getnstimeofday(&rec->ts_start); + ktime_get_real_ts64(&rec->ts_start); + if (rec->flags & TIME_BENCH_PMU) { + rec->pmc_inst_start = pmc_inst(); + rec->pmc_clk_start = pmc_clk(); + } + rec->tsc_start = tsc_start_clock(); +} + +static __always_inline void time_bench_stop(struct time_bench_record *rec, + uint64_t invoked_cnt) +{ + rec->tsc_stop = tsc_stop_clock(); + if (rec->flags & TIME_BENCH_PMU) { + rec->pmc_inst_stop = pmc_inst(); + rec->pmc_clk_stop = pmc_clk(); + } + //getnstimeofday(&rec->ts_stop); + ktime_get_real_ts64(&rec->ts_stop); + rec->invoked_cnt = invoked_cnt; +} + +#endif /* _LINUX_TIME_BENCH_H */ diff --git a/tools/testing/selftests/net/bench/test_bench_page_pool.sh b/tools/testing/selftests/net/bench/test_bench_page_pool.sh new file mode 100755 index 000000000000..7b8b18cfedce --- /dev/null +++ b/tools/testing/selftests/net/bench/test_bench_page_pool.sh @@ -0,0 +1,32 @@ +#!/bin/bash +# SPDX-License-Identifier: GPL-2.0 +# + +set -e + +DRIVER="./page_pool/bench_page_pool.ko" +result="" + +function run_test() +{ + rmmod "bench_page_pool.ko" || true + insmod $DRIVER > /dev/null 2>&1 + result=$(dmesg | tail -10) + echo "$result" + + echo + echo "Fast path results:" + echo "${result}" | grep -o -E "no-softirq-page_pool01 Per elem: ([0-9]+) cycles\(tsc\) ([0-9]+\.[0-9]+) ns" + + echo + echo "ptr_ring results:" + echo "${result}" | grep -o -E "no-softirq-page_pool02 Per elem: ([0-9]+) cycles\(tsc\) ([0-9]+\.[0-9]+) ns" + + echo + echo "slow path results:" + echo "${result}" | grep -o -E "no-softirq-page_pool03 Per elem: ([0-9]+) cycles\(tsc\) ([0-9]+\.[0-9]+) ns" +} + +run_test + +exit 0 base-commit: 8909f5f4ecd551c2299b28e05254b77424c8c7dc
On Mon, Jun 16, 2025 at 2:29 AM Jesper Dangaard Brouer hawk@kernel.org wrote:
On 15/06/2025 22.59, Mina Almasry wrote:
From: Jesper Dangaard Brouer hawk@kernel.org
We frequently consult with Jesper's out-of-tree page_pool benchmark to evaluate page_pool changes.
Import the benchmark into the upstream linux kernel tree so that (a) we're all running the same version, (b) pave the way for shared improvements, and (c) maybe one day integrate it with nipa, if possible.
Import bench_page_pool_simple from commit 35b1716d0c30 ("Add page_bench06_walk_all"), from this repository: https://github.com/netoptimizer/prototype-kernel.git
Changes done during upstreaming:
- Fix checkpatch issues.
- Remove the tasklet logic not needed.
- Move under tools/testing
- Create ksft for the benchmark.
- Changed slightly how the benchmark gets build. Out of tree, time_bench is built as an independent .ko. Here it is included in bench_page_pool.ko
Steps to run:
mkdir -p /tmp/run-pp-bench make -C ./tools/testing/selftests/net/bench make -C ./tools/testing/selftests/net/bench install INSTALL_PATH=/tmp/run-pp-bench rsync --delete -avz --progress /tmp/run-pp-bench mina@$SERVER:~/ ssh mina@$SERVER << EOF cd ~/run-pp-bench && sudo ./test_bench_page_pool.sh EOFOutput:
(benchmrk dmesg logs)Something is off with benchmark numbers compared to the OOT version.
I assume you're comparing my results (my kernel config + my hardware + upstream benchmark) with your results (your kernel config + your hardware + OOT version). The problem may be in OOT vs upstream but it may be just different code/config/hardware.
Adding my numbers below, they were run on my testlab with:
- CPU E5-1650 v4 @ 3.60GHz
- kernel: net.git v6.15-12438-gd9816ec74e6d
Fast path results: no-softirq-page_pool01 Per elem: 11 cycles(tsc) 4.368 ns
Fast-path on your CPU is faster (22 cycles(tsc) 6.128 ns) than my CPU. What CPU is this?
My test setup is a Gcloud A3 VM (so virtualized). The CPU is:
cat /proc/cpuinfo ... model name : Intel(R) Xeon(R) Platinum 8481C CPU @ 2.70GHz
Type:no-softirq-page_pool01 Per elem: 22 cycles(tsc) 6.128 ns (step:0)
- (measurement period time:0.061282924 sec time_interval:61282924)
- (invoke count:10000000 tsc_interval:220619745)
ptr_ring results: no-softirq-page_pool02 Per elem: 527 cycles(tsc) 195.187 ns
I'm surprised that ptr_ring benchmark is very slow, compared to my result (below) 60 cycles(tsc) 16.853 ns.
Type:no-softirq-page_pool02 Per elem: 60 cycles(tsc) 16.853 ns (step:0)
- (measurement period time:0.168535760 sec time_interval:168535760)
- (invoke count:10000000 tsc_interval:606734160)
Maybe your kernel is compiled with some CONFIG debug thing that makes it slower?
Yeah, I actually just checked and I have CONFIG_DEBUG_NET on in my build, and a lot of other debug configs are turned on.
Let me investigate here. Maybe trimming the debug configs and double checking my tree for debug logs I added would point to the difference.
I could also try to put both the OOT version and upstream version in my tree and do a proper A/B comparison that way.
If you do get chance to run this upstream version from your exact tree and config, that would be a good A/B comparison as well.
You can troubleshoot like this:
- select the `no-softirq-page_pool02` test via run_flags=$((2#100)).
# perf record -g modprobe bench_page_pool_simple run_flags=$((2#100)) loops=$((100*10**6)) # perf report --no-children
Thanks, will do.
On 16/06/2025 23.11, Mina Almasry wrote:
On Mon, Jun 16, 2025 at 2:29 AM Jesper Dangaard Brouer hawk@kernel.org wrote:
On 15/06/2025 22.59, Mina Almasry wrote:
From: Jesper Dangaard Brouer hawk@kernel.org
We frequently consult with Jesper's out-of-tree page_pool benchmark to evaluate page_pool changes.
Import the benchmark into the upstream linux kernel tree so that (a) we're all running the same version, (b) pave the way for shared improvements, and (c) maybe one day integrate it with nipa, if possible.
Import bench_page_pool_simple from commit 35b1716d0c30 ("Add page_bench06_walk_all"), from this repository: https://github.com/netoptimizer/prototype-kernel.git
Changes done during upstreaming:
- Fix checkpatch issues.
- Remove the tasklet logic not needed.
- Move under tools/testing
- Create ksft for the benchmark.
- Changed slightly how the benchmark gets build. Out of tree, time_bench is built as an independent .ko. Here it is included in bench_page_pool.ko
Steps to run:
mkdir -p /tmp/run-pp-bench make -C ./tools/testing/selftests/net/bench make -C ./tools/testing/selftests/net/bench install INSTALL_PATH=/tmp/run-pp-bench rsync --delete -avz --progress /tmp/run-pp-bench mina@$SERVER:~/ ssh mina@$SERVER << EOF cd ~/run-pp-bench && sudo ./test_bench_page_pool.sh EOFOutput:
(benchmrk dmesg logs)Something is off with benchmark numbers compared to the OOT version.
I assume you're comparing my results (my kernel config + my hardware + upstream benchmark) with your results (your kernel config + your hardware + OOT version). The problem may be in OOT vs upstream but it may be just different code/config/hardware.
True I used OOT version.
Just applied this patch, but I get compile error. Because Makefile tries to get kernel headers (net/page_pool/helpers.h) from local Linux installation instead of git tree. This need to be adjusted for patch, such that it builds with src-local/git tree provided headers.
$ make -C ./tools/testing/selftests/net/bench/page_pool/ make: Entering directory '/home/jbrouer/git/kernel/net/tools/testing/selftests/net/bench/page_pool' make[1]: Entering directory '/usr/src/kernels/6.5.12-100.fc37.x86_64' CC [M] /home/jbrouer/git/kernel/net/tools/testing/selftests/net/bench/page_pool/bench_page_pool_simple.o /home/jbrouer/git/kernel/net/tools/testing/selftests/net/bench/page_pool/bench_page_pool_simple.c:12:10: fatal error: net/page_pool/helpers.h: No such file or directory 12 | #include <net/page_pool/helpers.h> | ^~~~~~~~~~~~~~~~~~~~~~~~~ compilation terminated. make[3]: *** [scripts/Makefile.build:243: /home/jbrouer/git/kernel/net/tools/testing/selftests/net/bench/page_pool/bench_page_pool_simple.o] Error 1 make[2]: *** [/usr/src/kernels/6.5.12-100.fc37.x86_64/Makefile:2046: /home/jbrouer/git/kernel/net/tools/testing/selftests/net/bench/page_pool] Error 2 make[1]: *** [Makefile:246: __sub-make] Error 2 make[1]: Leaving directory '/usr/src/kernels/6.5.12-100.fc37.x86_64' make: *** [Makefile:14: all] Error 2 make: Leaving directory '/home/jbrouer/git/kernel/net/tools/testing/selftests/net/bench/page_pool'
Adding my numbers below, they were run on my testlab with:
- CPU E5-1650 v4 @ 3.60GHz
- kernel: net.git v6.15-12438-gd9816ec74e6d
Fast path results: no-softirq-page_pool01 Per elem: 11 cycles(tsc) 4.368 ns
Fast-path on your CPU is faster (22 cycles(tsc) 6.128 ns) than my CPU. What CPU is this?
My test setup is a Gcloud A3 VM (so virtualized). The CPU is:
cat /proc/cpuinfo ... model name : Intel(R) Xeon(R) Platinum 8481C CPU @ 2.70GHz
Type:no-softirq-page_pool01 Per elem: 22 cycles(tsc) 6.128 ns (step:0)
- (measurement period time:0.061282924 sec time_interval:61282924)
- (invoke count:10000000 tsc_interval:220619745)
ptr_ring results: no-softirq-page_pool02 Per elem: 527 cycles(tsc) 195.187 ns
I'm surprised that ptr_ring benchmark is very slow, compared to my result (below) 60 cycles(tsc) 16.853 ns.
Type:no-softirq-page_pool02 Per elem: 60 cycles(tsc) 16.853 ns (step:0)
- (measurement period time:0.168535760 sec time_interval:168535760)
- (invoke count:10000000 tsc_interval:606734160)
Maybe your kernel is compiled with some CONFIG debug thing that makes it slower?
Yeah, I actually just checked and I have CONFIG_DEBUG_NET on in my build, and a lot of other debug configs are turned on.
The CONFIG_DEBUG_NET should be low overhead, so I don't expect this to be the root-cause. Other CONFIG options are more likely the issue.
Let me investigate here. Maybe trimming the debug configs and double checking my tree for debug logs I added would point to the difference.
The perf record trick below will help you identify overhead more rabidly IMHO.
I could also try to put both the OOT version and upstream version in my tree and do a proper A/B comparison that way.
If you do get chance to run this upstream version from your exact tree and config, that would be a good A/B comparison as well.
You can troubleshoot like this:
- select the `no-softirq-page_pool02` test via run_flags=$((2#100)).
# perf record -g modprobe bench_page_pool_simple run_flags=$((2#100)) loops=$((100*10**6)) # perf report --no-children
Thanks, will do.
With a micro-benchmark like this, the perf report should help you to quickly indentify the kernel debug CONFIG that adds the overhead.
--Jesper
On Wed, Jun 18, 2025 at 5:46 AM Jesper Dangaard Brouer hawk@kernel.org wrote:
Something is off with benchmark numbers compared to the OOT version.
I assume you're comparing my results (my kernel config + my hardware + upstream benchmark) with your results (your kernel config + your hardware + OOT version). The problem may be in OOT vs upstream but it may be just different code/config/hardware.
True I used OOT version.
Just applied this patch, but I get compile error. Because Makefile tries to get kernel headers (net/page_pool/helpers.h) from local Linux installation instead of git tree. This need to be adjusted for patch, such that it builds with src-local/git tree provided headers.
I believe the fix to that is to do:
make KDIR=$(pwd) -C ./tools/testing/selftests/net/bench
I.e. the build files assume you're building the test to run it on the current machine, to cross compile it for a different machine under test, we need to pass explicit KDIR. I've kinda copy-pasted what other TEST_GEN_MODS_DIR= makefiles do. In theory we could do something else but I am guessing the way current TEST_GEN_MODS_DIR does it is the way to go. Does it work for you if you do that?
[...]
Yeah, I actually just checked and I have CONFIG_DEBUG_NET on in my build, and a lot of other debug configs are turned on.
The CONFIG_DEBUG_NET should be low overhead, so I don't expect this to be the root-cause. Other CONFIG options are more likely the issue.
Thank you very much for the tips. Perf report showed the locking was taking forever on my kernel... I had locking debug configs enabled in my build... sorry... with those disabled, I get much more sane results:
[ 185.557293] bench_page_pool: time_bench_page_pool01_fast_path(): Cannot use page_pool fast-path [ 185.607873] bench_page_pool: Type:no-softirq-page_pool01 Per elem: 11 cycles(tsc) 4.177 ns (step:0) - (measurement period time:0.041772642 sec time_interval:41772642) - (invoke count:10000000 tsc_interval:112778487) [ 185.627090] bench_page_pool: time_bench_page_pool02_ptr_ring(): Cannot use page_pool fast-path [ 185.826991] bench_page_pool: Type:no-softirq-page_pool02 Per elem: 51 cycles(tsc) 19.117 ns (step:0) - (measurement period time:0.191178107 sec time_interval:191178107) - (invoke count:10000000 tsc_interval:516173586) [ 185.846380] bench_page_pool: time_bench_page_pool03_slow(): Cannot use page_pool fast-path [ 186.479432] bench_page_pool: Type:no-softirq-page_pool03 Per elem: 168 cycles(tsc) 62.469 ns (step:0) - (measurement period time:0.624690697 sec time_interval:624690697) - (invoke count:10000000 tsc_interval:1686656879)
Does this alleviate your concern? Or do you still see an issue here? There is still a delta between our results, on different hardware/configs but results are in a sane range now.
On 19/06/2025 06.19, Mina Almasry wrote:
On Wed, Jun 18, 2025 at 5:46 AM Jesper Dangaard Brouer hawk@kernel.org wrote:
Something is off with benchmark numbers compared to the OOT version.
I assume you're comparing my results (my kernel config + my hardware + upstream benchmark) with your results (your kernel config + your hardware + OOT version). The problem may be in OOT vs upstream but it may be just different code/config/hardware.
True I used OOT version.
Just applied this patch, but I get compile error. Because Makefile tries to get kernel headers (net/page_pool/helpers.h) from local Linux installation instead of git tree. This need to be adjusted for patch, such that it builds with src-local/git tree provided headers.
I believe the fix to that is to do:
make KDIR=$(pwd) -C ./tools/testing/selftests/net/bench
Yes, this worked for me.
I.e. the build files assume you're building the test to run it on the current machine, to cross compile it for a different machine under test, we need to pass explicit KDIR. I've kinda copy-pasted what other TEST_GEN_MODS_DIR= makefiles do. In theory we could do something else but I am guessing the way current TEST_GEN_MODS_DIR does it is the way to go. Does it work for you if you do that?
Yes.
[...]
Yeah, I actually just checked and I have CONFIG_DEBUG_NET on in my build, and a lot of other debug configs are turned on.
The CONFIG_DEBUG_NET should be low overhead, so I don't expect this to be the root-cause. Other CONFIG options are more likely the issue.
Thank you very much for the tips. Perf report showed the locking was taking forever on my kernel... I had locking debug configs enabled in my build... sorry... with those disabled, I get much more sane results:
[ 185.557293] bench_page_pool: time_bench_page_pool01_fast_path(): Cannot use page_pool fast-path [ 185.607873] bench_page_pool: Type:no-softirq-page_pool01 Per elem: 11 cycles(tsc) 4.177 ns (step:0) - (measurement period time:0.041772642 sec time_interval:41772642) - (invoke count:10000000 tsc_interval:112778487) [ 185.627090] bench_page_pool: time_bench_page_pool02_ptr_ring(): Cannot use page_pool fast-path [ 185.826991] bench_page_pool: Type:no-softirq-page_pool02 Per elem: 51 cycles(tsc) 19.117 ns (step:0) - (measurement period time:0.191178107 sec time_interval:191178107) - (invoke count:10000000 tsc_interval:516173586) [ 185.846380] bench_page_pool: time_bench_page_pool03_slow(): Cannot use page_pool fast-path [ 186.479432] bench_page_pool: Type:no-softirq-page_pool03 Per elem: 168 cycles(tsc) 62.469 ns (step:0) - (measurement period time:0.624690697 sec time_interval:624690697) - (invoke count:10000000 tsc_interval:1686656879)
My results with this patch:
$ sudo ./test_bench_page_pool.sh rmmod: ERROR: Module bench_page_pool is not currently loaded [268960.638885] bench_page_pool: Loaded [268960.684603] bench_page_pool: Type:for_loop Per elem: 1 cycles(tsc) 0.420 ns (step:0) - (measurement period time:0.042037752 sec time_interval:42037752) - (invoke count:100000000 tsc_interval:151336355) [268961.203806] bench_page_pool: Type:atomic_inc Per elem: 18 cycles(tsc) 5.010 ns (step:0) - (measurement period time:0.501077936 sec time_interval:501077936) - (invoke count:100000000 tsc_interval:1803899823) [268961.332771] bench_page_pool: Type:lock Per elem: 39 cycles(tsc) 11.041 ns (step:0) - (measurement period time:0.110410468 sec time_interval:110410468) - (invoke count:10000000 tsc_interval:397481145) [268961.350718] bench_page_pool: time_bench_page_pool01_fast_path(): Cannot use page_pool fast-path [268961.425335] bench_page_pool: Type:no-softirq-page_pool01 Per elem: 23 cycles(tsc) 6.571 ns (step:0) - (measurement period time:0.065719390 sec time_interval:65719390) - (invoke count:10000000 tsc_interval:236591475) [268961.444666] bench_page_pool: time_bench_page_pool02_ptr_ring(): Cannot use page_pool fast-path [268961.622103] bench_page_pool: Type:no-softirq-page_pool02 Per elem: 60 cycles(tsc) 16.862 ns (step:0) - (measurement period time:0.168626201 sec time_interval:168626201) - (invoke count:10000000 tsc_interval:607060218) [268961.641608] bench_page_pool: time_bench_page_pool03_slow(): Cannot use page_pool fast-path [268962.387479] bench_page_pool: Type:no-softirq-page_pool03 Per elem: 265 cycles(tsc) 73.739 ns (step:0) - (measurement period time:0.737399722 sec time_interval:737399722) - (invoke count:10000000 tsc_interval:2654665224)
Fast path results: no-softirq-page_pool01 Per elem: 23 cycles(tsc) 6.571 ns
ptr_ring results: no-softirq-page_pool02 Per elem: 60 cycles(tsc) 16.862 ns
slow path results: no-softirq-page_pool03 Per elem: 265 cycles(tsc) 73.739 ns
Does this alleviate your concern? Or do you still see an issue here? There is still a delta between our results, on different hardware/configs but results are in a sane range now.
Now the results a sane and in range :-)
--Jesper
On Thu, Jun 19, 2025 at 3:24 AM Jesper Dangaard Brouer hawk@kernel.org wrote:
Does this alleviate your concern? Or do you still see an issue here? There is still a delta between our results, on different hardware/configs but results are in a sane range now.
Now the results a sane and in range :-)
Thanks!
Jakub, this patch was marked 'changes requested' for good reason, but after investigation we found the patch itself is good and the issues were elsewhere. What do I do here? A resend of v4? Or does this un 'changes requested' as-is?
On Thu, 19 Jun 2025 08:37:05 -0700 Mina Almasry wrote:
Jakub, this patch was marked 'changes requested' for good reason, but after investigation we found the patch itself is good and the issues were elsewhere. What do I do here? A resend of v4? Or does this un 'changes requested' as-is?
Do you want to update the numbers in the commit msg?
On Thu, Jun 19, 2025 at 9:58 AM Jakub Kicinski kuba@kernel.org wrote:
On Thu, 19 Jun 2025 08:37:05 -0700 Mina Almasry wrote:
Jakub, this patch was marked 'changes requested' for good reason, but after investigation we found the patch itself is good and the issues were elsewhere. What do I do here? A resend of v4? Or does this un 'changes requested' as-is?
Do you want to update the numbers in the commit msg?
Oh, I thought it's fine as-is. I can update them if need be, no need to trouble yourself. I'll upload v5 with updated benchmark results + make a note on how to cross build it (specify the KDIR= arg).
linux-kselftest-mirror@lists.linaro.org
-
Jakub Kicinski -
Jesper Dangaard Brouer -
Mina Almasry