Back to home page

OSCL-LXR
 
 

     

0001 // SPDX-License-Identifier: GPL-2.0
0002 /*
0003  * numa.c
0004  *
0005  * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
0006  */
0007 
0008 #include <inttypes.h>
0009 /* For the CLR_() macros */
0010 #include <pthread.h>
0011 
0012 #include <subcmd/parse-options.h>
0013 #include "../util/cloexec.h"
0014 
0015 #include "bench.h"
0016 
0017 #include <errno.h>
0018 #include <sched.h>
0019 #include <stdio.h>
0020 #include <assert.h>
0021 #include <malloc.h>
0022 #include <signal.h>
0023 #include <stdlib.h>
0024 #include <string.h>
0025 #include <unistd.h>
0026 #include <sys/mman.h>
0027 #include <sys/time.h>
0028 #include <sys/resource.h>
0029 #include <sys/wait.h>
0030 #include <sys/prctl.h>
0031 #include <sys/types.h>
0032 #include <linux/kernel.h>
0033 #include <linux/time64.h>
0034 #include <linux/numa.h>
0035 #include <linux/zalloc.h>
0036 
0037 #include "../util/header.h"
0038 #include <numa.h>
0039 #include <numaif.h>
0040 
0041 #ifndef RUSAGE_THREAD
0042 # define RUSAGE_THREAD 1
0043 #endif
0044 
0045 /*
0046  * Regular printout to the terminal, suppressed if -q is specified:
0047  */
0048 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
0049 
0050 /*
0051  * Debug printf:
0052  */
0053 #undef dprintf
0054 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
0055 
0056 struct thread_data {
0057     int         curr_cpu;
0058     cpu_set_t       *bind_cpumask;
0059     int         bind_node;
0060     u8          *process_data;
0061     int         process_nr;
0062     int         thread_nr;
0063     int         task_nr;
0064     unsigned int        loops_done;
0065     u64         val;
0066     u64         runtime_ns;
0067     u64         system_time_ns;
0068     u64         user_time_ns;
0069     double          speed_gbs;
0070     pthread_mutex_t     *process_lock;
0071 };
0072 
0073 /* Parameters set by options: */
0074 
0075 struct params {
0076     /* Startup synchronization: */
0077     bool            serialize_startup;
0078 
0079     /* Task hierarchy: */
0080     int         nr_proc;
0081     int         nr_threads;
0082 
0083     /* Working set sizes: */
0084     const char      *mb_global_str;
0085     const char      *mb_proc_str;
0086     const char      *mb_proc_locked_str;
0087     const char      *mb_thread_str;
0088 
0089     double          mb_global;
0090     double          mb_proc;
0091     double          mb_proc_locked;
0092     double          mb_thread;
0093 
0094     /* Access patterns to the working set: */
0095     bool            data_reads;
0096     bool            data_writes;
0097     bool            data_backwards;
0098     bool            data_zero_memset;
0099     bool            data_rand_walk;
0100     u32         nr_loops;
0101     u32         nr_secs;
0102     u32         sleep_usecs;
0103 
0104     /* Working set initialization: */
0105     bool            init_zero;
0106     bool            init_random;
0107     bool            init_cpu0;
0108 
0109     /* Misc options: */
0110     int         show_details;
0111     int         run_all;
0112     int         thp;
0113 
0114     long            bytes_global;
0115     long            bytes_process;
0116     long            bytes_process_locked;
0117     long            bytes_thread;
0118 
0119     int         nr_tasks;
0120     bool            show_quiet;
0121 
0122     bool            show_convergence;
0123     bool            measure_convergence;
0124 
0125     int         perturb_secs;
0126     int         nr_cpus;
0127     int         nr_nodes;
0128 
0129     /* Affinity options -C and -N: */
0130     char            *cpu_list_str;
0131     char            *node_list_str;
0132 };
0133 
0134 
0135 /* Global, read-writable area, accessible to all processes and threads: */
0136 
0137 struct global_info {
0138     u8          *data;
0139 
0140     pthread_mutex_t     startup_mutex;
0141     pthread_cond_t      startup_cond;
0142     int         nr_tasks_started;
0143 
0144     pthread_mutex_t     start_work_mutex;
0145     pthread_cond_t      start_work_cond;
0146     int         nr_tasks_working;
0147     bool            start_work;
0148 
0149     pthread_mutex_t     stop_work_mutex;
0150     u64         bytes_done;
0151 
0152     struct thread_data  *threads;
0153 
0154     /* Convergence latency measurement: */
0155     bool            all_converged;
0156     bool            stop_work;
0157 
0158     int         print_once;
0159 
0160     struct params       p;
0161 };
0162 
0163 static struct global_info   *g = NULL;
0164 
0165 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
0166 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
0167 
0168 struct params p0;
0169 
0170 static const struct option options[] = {
0171     OPT_INTEGER('p', "nr_proc"  , &p0.nr_proc,      "number of processes"),
0172     OPT_INTEGER('t', "nr_threads"   , &p0.nr_threads,   "number of threads per process"),
0173 
0174     OPT_STRING('G', "mb_global" , &p0.mb_global_str,    "MB", "global  memory (MBs)"),
0175     OPT_STRING('P', "mb_proc"   , &p0.mb_proc_str,  "MB", "process memory (MBs)"),
0176     OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
0177     OPT_STRING('T', "mb_thread" , &p0.mb_thread_str,    "MB", "thread  memory (MBs)"),
0178 
0179     OPT_UINTEGER('l', "nr_loops"    , &p0.nr_loops,     "max number of loops to run (default: unlimited)"),
0180     OPT_UINTEGER('s', "nr_secs" , &p0.nr_secs,      "max number of seconds to run (default: 5 secs)"),
0181     OPT_UINTEGER('u', "usleep"  , &p0.sleep_usecs,  "usecs to sleep per loop iteration"),
0182 
0183     OPT_BOOLEAN('R', "data_reads"   , &p0.data_reads,   "access the data via reads (can be mixed with -W)"),
0184     OPT_BOOLEAN('W', "data_writes"  , &p0.data_writes,  "access the data via writes (can be mixed with -R)"),
0185     OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,  "access the data backwards as well"),
0186     OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
0187     OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,  "access the data with random (32bit LFSR) walk"),
0188 
0189 
0190     OPT_BOOLEAN('z', "init_zero"    , &p0.init_zero,    "bzero the initial allocations"),
0191     OPT_BOOLEAN('I', "init_random"  , &p0.init_random,  "randomize the contents of the initial allocations"),
0192     OPT_BOOLEAN('0', "init_cpu0"    , &p0.init_cpu0,    "do the initial allocations on CPU#0"),
0193     OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,  "perturb thread 0/0 every X secs, to test convergence stability"),
0194 
0195     OPT_INCR   ('d', "show_details" , &p0.show_details, "Show details"),
0196     OPT_INCR   ('a', "all"      , &p0.run_all,      "Run all tests in the suite"),
0197     OPT_INTEGER('H', "thp"      , &p0.thp,      "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
0198     OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
0199             "convergence is reached when each process (all its threads) is running on a single NUMA node."),
0200     OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
0201     OPT_BOOLEAN('q', "quiet"    , &p0.show_quiet,   "quiet mode"),
0202     OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
0203 
0204     /* Special option string parsing callbacks: */
0205         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
0206             "bind the first N tasks to these specific cpus (the rest is unbound)",
0207             parse_cpus_opt),
0208         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
0209             "bind the first N tasks to these specific memory nodes (the rest is unbound)",
0210             parse_nodes_opt),
0211     OPT_END()
0212 };
0213 
0214 static const char * const bench_numa_usage[] = {
0215     "perf bench numa <options>",
0216     NULL
0217 };
0218 
0219 static const char * const numa_usage[] = {
0220     "perf bench numa mem [<options>]",
0221     NULL
0222 };
0223 
0224 /*
0225  * To get number of numa nodes present.
0226  */
0227 static int nr_numa_nodes(void)
0228 {
0229     int i, nr_nodes = 0;
0230 
0231     for (i = 0; i < g->p.nr_nodes; i++) {
0232         if (numa_bitmask_isbitset(numa_nodes_ptr, i))
0233             nr_nodes++;
0234     }
0235 
0236     return nr_nodes;
0237 }
0238 
0239 /*
0240  * To check if given numa node is present.
0241  */
0242 static int is_node_present(int node)
0243 {
0244     return numa_bitmask_isbitset(numa_nodes_ptr, node);
0245 }
0246 
0247 /*
0248  * To check given numa node has cpus.
0249  */
0250 static bool node_has_cpus(int node)
0251 {
0252     struct bitmask *cpumask = numa_allocate_cpumask();
0253     bool ret = false; /* fall back to nocpus */
0254     int cpu;
0255 
0256     BUG_ON(!cpumask);
0257     if (!numa_node_to_cpus(node, cpumask)) {
0258         for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
0259             if (numa_bitmask_isbitset(cpumask, cpu)) {
0260                 ret = true;
0261                 break;
0262             }
0263         }
0264     }
0265     numa_free_cpumask(cpumask);
0266 
0267     return ret;
0268 }
0269 
0270 static cpu_set_t *bind_to_cpu(int target_cpu)
0271 {
0272     int nrcpus = numa_num_possible_cpus();
0273     cpu_set_t *orig_mask, *mask;
0274     size_t size;
0275 
0276     orig_mask = CPU_ALLOC(nrcpus);
0277     BUG_ON(!orig_mask);
0278     size = CPU_ALLOC_SIZE(nrcpus);
0279     CPU_ZERO_S(size, orig_mask);
0280 
0281     if (sched_getaffinity(0, size, orig_mask))
0282         goto err_out;
0283 
0284     mask = CPU_ALLOC(nrcpus);
0285     if (!mask)
0286         goto err_out;
0287 
0288     CPU_ZERO_S(size, mask);
0289 
0290     if (target_cpu == -1) {
0291         int cpu;
0292 
0293         for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
0294             CPU_SET_S(cpu, size, mask);
0295     } else {
0296         if (target_cpu < 0 || target_cpu >= g->p.nr_cpus)
0297             goto err;
0298 
0299         CPU_SET_S(target_cpu, size, mask);
0300     }
0301 
0302     if (sched_setaffinity(0, size, mask))
0303         goto err;
0304 
0305     return orig_mask;
0306 
0307 err:
0308     CPU_FREE(mask);
0309 err_out:
0310     CPU_FREE(orig_mask);
0311 
0312     /* BUG_ON due to failure in allocation of orig_mask/mask */
0313     BUG_ON(-1);
0314     return NULL;
0315 }
0316 
0317 static cpu_set_t *bind_to_node(int target_node)
0318 {
0319     int nrcpus = numa_num_possible_cpus();
0320     size_t size;
0321     cpu_set_t *orig_mask, *mask;
0322     int cpu;
0323 
0324     orig_mask = CPU_ALLOC(nrcpus);
0325     BUG_ON(!orig_mask);
0326     size = CPU_ALLOC_SIZE(nrcpus);
0327     CPU_ZERO_S(size, orig_mask);
0328 
0329     if (sched_getaffinity(0, size, orig_mask))
0330         goto err_out;
0331 
0332     mask = CPU_ALLOC(nrcpus);
0333     if (!mask)
0334         goto err_out;
0335 
0336     CPU_ZERO_S(size, mask);
0337 
0338     if (target_node == NUMA_NO_NODE) {
0339         for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
0340             CPU_SET_S(cpu, size, mask);
0341     } else {
0342         struct bitmask *cpumask = numa_allocate_cpumask();
0343 
0344         if (!cpumask)
0345             goto err;
0346 
0347         if (!numa_node_to_cpus(target_node, cpumask)) {
0348             for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
0349                 if (numa_bitmask_isbitset(cpumask, cpu))
0350                     CPU_SET_S(cpu, size, mask);
0351             }
0352         }
0353         numa_free_cpumask(cpumask);
0354     }
0355 
0356     if (sched_setaffinity(0, size, mask))
0357         goto err;
0358 
0359     return orig_mask;
0360 
0361 err:
0362     CPU_FREE(mask);
0363 err_out:
0364     CPU_FREE(orig_mask);
0365 
0366     /* BUG_ON due to failure in allocation of orig_mask/mask */
0367     BUG_ON(-1);
0368     return NULL;
0369 }
0370 
0371 static void bind_to_cpumask(cpu_set_t *mask)
0372 {
0373     int ret;
0374     size_t size = CPU_ALLOC_SIZE(numa_num_possible_cpus());
0375 
0376     ret = sched_setaffinity(0, size, mask);
0377     if (ret) {
0378         CPU_FREE(mask);
0379         BUG_ON(ret);
0380     }
0381 }
0382 
0383 static void mempol_restore(void)
0384 {
0385     int ret;
0386 
0387     ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
0388 
0389     BUG_ON(ret);
0390 }
0391 
0392 static void bind_to_memnode(int node)
0393 {
0394     struct bitmask *node_mask;
0395     int ret;
0396 
0397     if (node == NUMA_NO_NODE)
0398         return;
0399 
0400     node_mask = numa_allocate_nodemask();
0401     BUG_ON(!node_mask);
0402 
0403     numa_bitmask_clearall(node_mask);
0404     numa_bitmask_setbit(node_mask, node);
0405 
0406     ret = set_mempolicy(MPOL_BIND, node_mask->maskp, node_mask->size + 1);
0407     dprintf("binding to node %d, mask: %016lx => %d\n", node, *node_mask->maskp, ret);
0408 
0409     numa_bitmask_free(node_mask);
0410     BUG_ON(ret);
0411 }
0412 
0413 #define HPSIZE (2*1024*1024)
0414 
0415 #define set_taskname(fmt...)                \
0416 do {                            \
0417     char name[20];                  \
0418                             \
0419     snprintf(name, 20, fmt);            \
0420     prctl(PR_SET_NAME, name);           \
0421 } while (0)
0422 
0423 static u8 *alloc_data(ssize_t bytes0, int map_flags,
0424               int init_zero, int init_cpu0, int thp, int init_random)
0425 {
0426     cpu_set_t *orig_mask = NULL;
0427     ssize_t bytes;
0428     u8 *buf;
0429     int ret;
0430 
0431     if (!bytes0)
0432         return NULL;
0433 
0434     /* Allocate and initialize all memory on CPU#0: */
0435     if (init_cpu0) {
0436         int node = numa_node_of_cpu(0);
0437 
0438         orig_mask = bind_to_node(node);
0439         bind_to_memnode(node);
0440     }
0441 
0442     bytes = bytes0 + HPSIZE;
0443 
0444     buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
0445     BUG_ON(buf == (void *)-1);
0446 
0447     if (map_flags == MAP_PRIVATE) {
0448         if (thp > 0) {
0449             ret = madvise(buf, bytes, MADV_HUGEPAGE);
0450             if (ret && !g->print_once) {
0451                 g->print_once = 1;
0452                 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
0453             }
0454         }
0455         if (thp < 0) {
0456             ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
0457             if (ret && !g->print_once) {
0458                 g->print_once = 1;
0459                 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
0460             }
0461         }
0462     }
0463 
0464     if (init_zero) {
0465         bzero(buf, bytes);
0466     } else {
0467         /* Initialize random contents, different in each word: */
0468         if (init_random) {
0469             u64 *wbuf = (void *)buf;
0470             long off = rand();
0471             long i;
0472 
0473             for (i = 0; i < bytes/8; i++)
0474                 wbuf[i] = i + off;
0475         }
0476     }
0477 
0478     /* Align to 2MB boundary: */
0479     buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
0480 
0481     /* Restore affinity: */
0482     if (init_cpu0) {
0483         bind_to_cpumask(orig_mask);
0484         CPU_FREE(orig_mask);
0485         mempol_restore();
0486     }
0487 
0488     return buf;
0489 }
0490 
0491 static void free_data(void *data, ssize_t bytes)
0492 {
0493     int ret;
0494 
0495     if (!data)
0496         return;
0497 
0498     ret = munmap(data, bytes);
0499     BUG_ON(ret);
0500 }
0501 
0502 /*
0503  * Create a shared memory buffer that can be shared between processes, zeroed:
0504  */
0505 static void * zalloc_shared_data(ssize_t bytes)
0506 {
0507     return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
0508 }
0509 
0510 /*
0511  * Create a shared memory buffer that can be shared between processes:
0512  */
0513 static void * setup_shared_data(ssize_t bytes)
0514 {
0515     return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
0516 }
0517 
0518 /*
0519  * Allocate process-local memory - this will either be shared between
0520  * threads of this process, or only be accessed by this thread:
0521  */
0522 static void * setup_private_data(ssize_t bytes)
0523 {
0524     return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
0525 }
0526 
0527 /*
0528  * Return a process-shared (global) mutex:
0529  */
0530 static void init_global_mutex(pthread_mutex_t *mutex)
0531 {
0532     pthread_mutexattr_t attr;
0533 
0534     pthread_mutexattr_init(&attr);
0535     pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
0536     pthread_mutex_init(mutex, &attr);
0537 }
0538 
0539 /*
0540  * Return a process-shared (global) condition variable:
0541  */
0542 static void init_global_cond(pthread_cond_t *cond)
0543 {
0544     pthread_condattr_t attr;
0545 
0546     pthread_condattr_init(&attr);
0547     pthread_condattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
0548     pthread_cond_init(cond, &attr);
0549 }
0550 
0551 static int parse_cpu_list(const char *arg)
0552 {
0553     p0.cpu_list_str = strdup(arg);
0554 
0555     dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
0556 
0557     return 0;
0558 }
0559 
0560 static int parse_setup_cpu_list(void)
0561 {
0562     struct thread_data *td;
0563     char *str0, *str;
0564     int t;
0565 
0566     if (!g->p.cpu_list_str)
0567         return 0;
0568 
0569     dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
0570 
0571     str0 = str = strdup(g->p.cpu_list_str);
0572     t = 0;
0573 
0574     BUG_ON(!str);
0575 
0576     tprintf("# binding tasks to CPUs:\n");
0577     tprintf("#  ");
0578 
0579     while (true) {
0580         int bind_cpu, bind_cpu_0, bind_cpu_1;
0581         char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
0582         int bind_len;
0583         int step;
0584         int mul;
0585 
0586         tok = strsep(&str, ",");
0587         if (!tok)
0588             break;
0589 
0590         tok_end = strstr(tok, "-");
0591 
0592         dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
0593         if (!tok_end) {
0594             /* Single CPU specified: */
0595             bind_cpu_0 = bind_cpu_1 = atol(tok);
0596         } else {
0597             /* CPU range specified (for example: "5-11"): */
0598             bind_cpu_0 = atol(tok);
0599             bind_cpu_1 = atol(tok_end + 1);
0600         }
0601 
0602         step = 1;
0603         tok_step = strstr(tok, "#");
0604         if (tok_step) {
0605             step = atol(tok_step + 1);
0606             BUG_ON(step <= 0 || step >= g->p.nr_cpus);
0607         }
0608 
0609         /*
0610          * Mask length.
0611          * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
0612          * where the _4 means the next 4 CPUs are allowed.
0613          */
0614         bind_len = 1;
0615         tok_len = strstr(tok, "_");
0616         if (tok_len) {
0617             bind_len = atol(tok_len + 1);
0618             BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
0619         }
0620 
0621         /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
0622         mul = 1;
0623         tok_mul = strstr(tok, "x");
0624         if (tok_mul) {
0625             mul = atol(tok_mul + 1);
0626             BUG_ON(mul <= 0);
0627         }
0628 
0629         dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
0630 
0631         if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
0632             printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
0633             return -1;
0634         }
0635 
0636         if (is_cpu_online(bind_cpu_0) != 1 || is_cpu_online(bind_cpu_1) != 1) {
0637             printf("\nTest not applicable, bind_cpu_0 or bind_cpu_1 is offline\n");
0638             return -1;
0639         }
0640 
0641         BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
0642         BUG_ON(bind_cpu_0 > bind_cpu_1);
0643 
0644         for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
0645             size_t size = CPU_ALLOC_SIZE(g->p.nr_cpus);
0646             int i;
0647 
0648             for (i = 0; i < mul; i++) {
0649                 int cpu;
0650 
0651                 if (t >= g->p.nr_tasks) {
0652                     printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
0653                     goto out;
0654                 }
0655                 td = g->threads + t;
0656 
0657                 if (t)
0658                     tprintf(",");
0659                 if (bind_len > 1) {
0660                     tprintf("%2d/%d", bind_cpu, bind_len);
0661                 } else {
0662                     tprintf("%2d", bind_cpu);
0663                 }
0664 
0665                 td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
0666                 BUG_ON(!td->bind_cpumask);
0667                 CPU_ZERO_S(size, td->bind_cpumask);
0668                 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
0669                     if (cpu < 0 || cpu >= g->p.nr_cpus) {
0670                         CPU_FREE(td->bind_cpumask);
0671                         BUG_ON(-1);
0672                     }
0673                     CPU_SET_S(cpu, size, td->bind_cpumask);
0674                 }
0675                 t++;
0676             }
0677         }
0678     }
0679 out:
0680 
0681     tprintf("\n");
0682 
0683     if (t < g->p.nr_tasks)
0684         printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
0685 
0686     free(str0);
0687     return 0;
0688 }
0689 
0690 static int parse_cpus_opt(const struct option *opt __maybe_unused,
0691               const char *arg, int unset __maybe_unused)
0692 {
0693     if (!arg)
0694         return -1;
0695 
0696     return parse_cpu_list(arg);
0697 }
0698 
0699 static int parse_node_list(const char *arg)
0700 {
0701     p0.node_list_str = strdup(arg);
0702 
0703     dprintf("got NODE list: {%s}\n", p0.node_list_str);
0704 
0705     return 0;
0706 }
0707 
0708 static int parse_setup_node_list(void)
0709 {
0710     struct thread_data *td;
0711     char *str0, *str;
0712     int t;
0713 
0714     if (!g->p.node_list_str)
0715         return 0;
0716 
0717     dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
0718 
0719     str0 = str = strdup(g->p.node_list_str);
0720     t = 0;
0721 
0722     BUG_ON(!str);
0723 
0724     tprintf("# binding tasks to NODEs:\n");
0725     tprintf("# ");
0726 
0727     while (true) {
0728         int bind_node, bind_node_0, bind_node_1;
0729         char *tok, *tok_end, *tok_step, *tok_mul;
0730         int step;
0731         int mul;
0732 
0733         tok = strsep(&str, ",");
0734         if (!tok)
0735             break;
0736 
0737         tok_end = strstr(tok, "-");
0738 
0739         dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
0740         if (!tok_end) {
0741             /* Single NODE specified: */
0742             bind_node_0 = bind_node_1 = atol(tok);
0743         } else {
0744             /* NODE range specified (for example: "5-11"): */
0745             bind_node_0 = atol(tok);
0746             bind_node_1 = atol(tok_end + 1);
0747         }
0748 
0749         step = 1;
0750         tok_step = strstr(tok, "#");
0751         if (tok_step) {
0752             step = atol(tok_step + 1);
0753             BUG_ON(step <= 0 || step >= g->p.nr_nodes);
0754         }
0755 
0756         /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
0757         mul = 1;
0758         tok_mul = strstr(tok, "x");
0759         if (tok_mul) {
0760             mul = atol(tok_mul + 1);
0761             BUG_ON(mul <= 0);
0762         }
0763 
0764         dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
0765 
0766         if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
0767             printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
0768             return -1;
0769         }
0770 
0771         BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
0772         BUG_ON(bind_node_0 > bind_node_1);
0773 
0774         for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
0775             int i;
0776 
0777             for (i = 0; i < mul; i++) {
0778                 if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
0779                     printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
0780                     goto out;
0781                 }
0782                 td = g->threads + t;
0783 
0784                 if (!t)
0785                     tprintf(" %2d", bind_node);
0786                 else
0787                     tprintf(",%2d", bind_node);
0788 
0789                 td->bind_node = bind_node;
0790                 t++;
0791             }
0792         }
0793     }
0794 out:
0795 
0796     tprintf("\n");
0797 
0798     if (t < g->p.nr_tasks)
0799         printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
0800 
0801     free(str0);
0802     return 0;
0803 }
0804 
0805 static int parse_nodes_opt(const struct option *opt __maybe_unused,
0806               const char *arg, int unset __maybe_unused)
0807 {
0808     if (!arg)
0809         return -1;
0810 
0811     return parse_node_list(arg);
0812 }
0813 
0814 static inline uint32_t lfsr_32(uint32_t lfsr)
0815 {
0816     const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
0817     return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
0818 }
0819 
0820 /*
0821  * Make sure there's real data dependency to RAM (when read
0822  * accesses are enabled), so the compiler, the CPU and the
0823  * kernel (KSM, zero page, etc.) cannot optimize away RAM
0824  * accesses:
0825  */
0826 static inline u64 access_data(u64 *data, u64 val)
0827 {
0828     if (g->p.data_reads)
0829         val += *data;
0830     if (g->p.data_writes)
0831         *data = val + 1;
0832     return val;
0833 }
0834 
0835 /*
0836  * The worker process does two types of work, a forwards going
0837  * loop and a backwards going loop.
0838  *
0839  * We do this so that on multiprocessor systems we do not create
0840  * a 'train' of processing, with highly synchronized processes,
0841  * skewing the whole benchmark.
0842  */
0843 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
0844 {
0845     long words = bytes/sizeof(u64);
0846     u64 *data = (void *)__data;
0847     long chunk_0, chunk_1;
0848     u64 *d0, *d, *d1;
0849     long off;
0850     long i;
0851 
0852     BUG_ON(!data && words);
0853     BUG_ON(data && !words);
0854 
0855     if (!data)
0856         return val;
0857 
0858     /* Very simple memset() work variant: */
0859     if (g->p.data_zero_memset && !g->p.data_rand_walk) {
0860         bzero(data, bytes);
0861         return val;
0862     }
0863 
0864     /* Spread out by PID/TID nr and by loop nr: */
0865     chunk_0 = words/nr_max;
0866     chunk_1 = words/g->p.nr_loops;
0867     off = nr*chunk_0 + loop*chunk_1;
0868 
0869     while (off >= words)
0870         off -= words;
0871 
0872     if (g->p.data_rand_walk) {
0873         u32 lfsr = nr + loop + val;
0874         int j;
0875 
0876         for (i = 0; i < words/1024; i++) {
0877             long start, end;
0878 
0879             lfsr = lfsr_32(lfsr);
0880 
0881             start = lfsr % words;
0882             end = min(start + 1024, words-1);
0883 
0884             if (g->p.data_zero_memset) {
0885                 bzero(data + start, (end-start) * sizeof(u64));
0886             } else {
0887                 for (j = start; j < end; j++)
0888                     val = access_data(data + j, val);
0889             }
0890         }
0891     } else if (!g->p.data_backwards || (nr + loop) & 1) {
0892         /* Process data forwards: */
0893 
0894         d0 = data + off;
0895         d  = data + off + 1;
0896         d1 = data + words;
0897 
0898         for (;;) {
0899             if (unlikely(d >= d1))
0900                 d = data;
0901             if (unlikely(d == d0))
0902                 break;
0903 
0904             val = access_data(d, val);
0905 
0906             d++;
0907         }
0908     } else {
0909         /* Process data backwards: */
0910 
0911         d0 = data + off;
0912         d  = data + off - 1;
0913         d1 = data + words;
0914 
0915         for (;;) {
0916             if (unlikely(d < data))
0917                 d = data + words-1;
0918             if (unlikely(d == d0))
0919                 break;
0920 
0921             val = access_data(d, val);
0922 
0923             d--;
0924         }
0925     }
0926 
0927     return val;
0928 }
0929 
0930 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
0931 {
0932     unsigned int cpu;
0933 
0934     cpu = sched_getcpu();
0935 
0936     g->threads[task_nr].curr_cpu = cpu;
0937     prctl(0, bytes_worked);
0938 }
0939 
0940 /*
0941  * Count the number of nodes a process's threads
0942  * are spread out on.
0943  *
0944  * A count of 1 means that the process is compressed
0945  * to a single node. A count of g->p.nr_nodes means it's
0946  * spread out on the whole system.
0947  */
0948 static int count_process_nodes(int process_nr)
0949 {
0950     char *node_present;
0951     int nodes;
0952     int n, t;
0953 
0954     node_present = (char *)malloc(g->p.nr_nodes * sizeof(char));
0955     BUG_ON(!node_present);
0956     for (nodes = 0; nodes < g->p.nr_nodes; nodes++)
0957         node_present[nodes] = 0;
0958 
0959     for (t = 0; t < g->p.nr_threads; t++) {
0960         struct thread_data *td;
0961         int task_nr;
0962         int node;
0963 
0964         task_nr = process_nr*g->p.nr_threads + t;
0965         td = g->threads + task_nr;
0966 
0967         node = numa_node_of_cpu(td->curr_cpu);
0968         if (node < 0) /* curr_cpu was likely still -1 */ {
0969             free(node_present);
0970             return 0;
0971         }
0972 
0973         node_present[node] = 1;
0974     }
0975 
0976     nodes = 0;
0977 
0978     for (n = 0; n < g->p.nr_nodes; n++)
0979         nodes += node_present[n];
0980 
0981     free(node_present);
0982     return nodes;
0983 }
0984 
0985 /*
0986  * Count the number of distinct process-threads a node contains.
0987  *
0988  * A count of 1 means that the node contains only a single
0989  * process. If all nodes on the system contain at most one
0990  * process then we are well-converged.
0991  */
0992 static int count_node_processes(int node)
0993 {
0994     int processes = 0;
0995     int t, p;
0996 
0997     for (p = 0; p < g->p.nr_proc; p++) {
0998         for (t = 0; t < g->p.nr_threads; t++) {
0999             struct thread_data *td;
1000             int task_nr;
1001             int n;
1002 
1003             task_nr = p*g->p.nr_threads + t;
1004             td = g->threads + task_nr;
1005 
1006             n = numa_node_of_cpu(td->curr_cpu);
1007             if (n == node) {
1008                 processes++;
1009                 break;
1010             }
1011         }
1012     }
1013 
1014     return processes;
1015 }
1016 
1017 static void calc_convergence_compression(int *strong)
1018 {
1019     unsigned int nodes_min, nodes_max;
1020     int p;
1021 
1022     nodes_min = -1;
1023     nodes_max =  0;
1024 
1025     for (p = 0; p < g->p.nr_proc; p++) {
1026         unsigned int nodes = count_process_nodes(p);
1027 
1028         if (!nodes) {
1029             *strong = 0;
1030             return;
1031         }
1032 
1033         nodes_min = min(nodes, nodes_min);
1034         nodes_max = max(nodes, nodes_max);
1035     }
1036 
1037     /* Strong convergence: all threads compress on a single node: */
1038     if (nodes_min == 1 && nodes_max == 1) {
1039         *strong = 1;
1040     } else {
1041         *strong = 0;
1042         tprintf(" {%d-%d}", nodes_min, nodes_max);
1043     }
1044 }
1045 
1046 static void calc_convergence(double runtime_ns_max, double *convergence)
1047 {
1048     unsigned int loops_done_min, loops_done_max;
1049     int process_groups;
1050     int *nodes;
1051     int distance;
1052     int nr_min;
1053     int nr_max;
1054     int strong;
1055     int sum;
1056     int nr;
1057     int node;
1058     int cpu;
1059     int t;
1060 
1061     if (!g->p.show_convergence && !g->p.measure_convergence)
1062         return;
1063 
1064     nodes = (int *)malloc(g->p.nr_nodes * sizeof(int));
1065     BUG_ON(!nodes);
1066     for (node = 0; node < g->p.nr_nodes; node++)
1067         nodes[node] = 0;
1068 
1069     loops_done_min = -1;
1070     loops_done_max = 0;
1071 
1072     for (t = 0; t < g->p.nr_tasks; t++) {
1073         struct thread_data *td = g->threads + t;
1074         unsigned int loops_done;
1075 
1076         cpu = td->curr_cpu;
1077 
1078         /* Not all threads have written it yet: */
1079         if (cpu < 0)
1080             continue;
1081 
1082         node = numa_node_of_cpu(cpu);
1083 
1084         nodes[node]++;
1085 
1086         loops_done = td->loops_done;
1087         loops_done_min = min(loops_done, loops_done_min);
1088         loops_done_max = max(loops_done, loops_done_max);
1089     }
1090 
1091     nr_max = 0;
1092     nr_min = g->p.nr_tasks;
1093     sum = 0;
1094 
1095     for (node = 0; node < g->p.nr_nodes; node++) {
1096         if (!is_node_present(node))
1097             continue;
1098         nr = nodes[node];
1099         nr_min = min(nr, nr_min);
1100         nr_max = max(nr, nr_max);
1101         sum += nr;
1102     }
1103     BUG_ON(nr_min > nr_max);
1104 
1105     BUG_ON(sum > g->p.nr_tasks);
1106 
1107     if (0 && (sum < g->p.nr_tasks)) {
1108         free(nodes);
1109         return;
1110     }
1111 
1112     /*
1113      * Count the number of distinct process groups present
1114      * on nodes - when we are converged this will decrease
1115      * to g->p.nr_proc:
1116      */
1117     process_groups = 0;
1118 
1119     for (node = 0; node < g->p.nr_nodes; node++) {
1120         int processes;
1121 
1122         if (!is_node_present(node))
1123             continue;
1124         processes = count_node_processes(node);
1125         nr = nodes[node];
1126         tprintf(" %2d/%-2d", nr, processes);
1127 
1128         process_groups += processes;
1129     }
1130 
1131     distance = nr_max - nr_min;
1132 
1133     tprintf(" [%2d/%-2d]", distance, process_groups);
1134 
1135     tprintf(" l:%3d-%-3d (%3d)",
1136         loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1137 
1138     if (loops_done_min && loops_done_max) {
1139         double skew = 1.0 - (double)loops_done_min/loops_done_max;
1140 
1141         tprintf(" [%4.1f%%]", skew * 100.0);
1142     }
1143 
1144     calc_convergence_compression(&strong);
1145 
1146     if (strong && process_groups == g->p.nr_proc) {
1147         if (!*convergence) {
1148             *convergence = runtime_ns_max;
1149             tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1150             if (g->p.measure_convergence) {
1151                 g->all_converged = true;
1152                 g->stop_work = true;
1153             }
1154         }
1155     } else {
1156         if (*convergence) {
1157             tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1158             *convergence = 0;
1159         }
1160         tprintf("\n");
1161     }
1162 
1163     free(nodes);
1164 }
1165 
1166 static void show_summary(double runtime_ns_max, int l, double *convergence)
1167 {
1168     tprintf("\r #  %5.1f%%  [%.1f mins]",
1169         (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1170 
1171     calc_convergence(runtime_ns_max, convergence);
1172 
1173     if (g->p.show_details >= 0)
1174         fflush(stdout);
1175 }
1176 
1177 static void *worker_thread(void *__tdata)
1178 {
1179     struct thread_data *td = __tdata;
1180     struct timeval start0, start, stop, diff;
1181     int process_nr = td->process_nr;
1182     int thread_nr = td->thread_nr;
1183     unsigned long last_perturbance;
1184     int task_nr = td->task_nr;
1185     int details = g->p.show_details;
1186     int first_task, last_task;
1187     double convergence = 0;
1188     u64 val = td->val;
1189     double runtime_ns_max;
1190     u8 *global_data;
1191     u8 *process_data;
1192     u8 *thread_data;
1193     u64 bytes_done, secs;
1194     long work_done;
1195     u32 l;
1196     struct rusage rusage;
1197 
1198     bind_to_cpumask(td->bind_cpumask);
1199     bind_to_memnode(td->bind_node);
1200 
1201     set_taskname("thread %d/%d", process_nr, thread_nr);
1202 
1203     global_data = g->data;
1204     process_data = td->process_data;
1205     thread_data = setup_private_data(g->p.bytes_thread);
1206 
1207     bytes_done = 0;
1208 
1209     last_task = 0;
1210     if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1211         last_task = 1;
1212 
1213     first_task = 0;
1214     if (process_nr == 0 && thread_nr == 0)
1215         first_task = 1;
1216 
1217     if (details >= 2) {
1218         printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1219             process_nr, thread_nr, global_data, process_data, thread_data);
1220     }
1221 
1222     if (g->p.serialize_startup) {
1223         pthread_mutex_lock(&g->startup_mutex);
1224         g->nr_tasks_started++;
1225         /* The last thread wakes the main process. */
1226         if (g->nr_tasks_started == g->p.nr_tasks)
1227             pthread_cond_signal(&g->startup_cond);
1228 
1229         pthread_mutex_unlock(&g->startup_mutex);
1230 
1231         /* Here we will wait for the main process to start us all at once: */
1232         pthread_mutex_lock(&g->start_work_mutex);
1233         g->start_work = false;
1234         g->nr_tasks_working++;
1235         while (!g->start_work)
1236             pthread_cond_wait(&g->start_work_cond, &g->start_work_mutex);
1237 
1238         pthread_mutex_unlock(&g->start_work_mutex);
1239     }
1240 
1241     gettimeofday(&start0, NULL);
1242 
1243     start = stop = start0;
1244     last_perturbance = start.tv_sec;
1245 
1246     for (l = 0; l < g->p.nr_loops; l++) {
1247         start = stop;
1248 
1249         if (g->stop_work)
1250             break;
1251 
1252         val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,  l, val);
1253         val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,   l, val);
1254         val += do_work(thread_data,  g->p.bytes_thread,  0,          1,     l, val);
1255 
1256         if (g->p.sleep_usecs) {
1257             pthread_mutex_lock(td->process_lock);
1258             usleep(g->p.sleep_usecs);
1259             pthread_mutex_unlock(td->process_lock);
1260         }
1261         /*
1262          * Amount of work to be done under a process-global lock:
1263          */
1264         if (g->p.bytes_process_locked) {
1265             pthread_mutex_lock(td->process_lock);
1266             val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,    l, val);
1267             pthread_mutex_unlock(td->process_lock);
1268         }
1269 
1270         work_done = g->p.bytes_global + g->p.bytes_process +
1271                 g->p.bytes_process_locked + g->p.bytes_thread;
1272 
1273         update_curr_cpu(task_nr, work_done);
1274         bytes_done += work_done;
1275 
1276         if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1277             continue;
1278 
1279         td->loops_done = l;
1280 
1281         gettimeofday(&stop, NULL);
1282 
1283         /* Check whether our max runtime timed out: */
1284         if (g->p.nr_secs) {
1285             timersub(&stop, &start0, &diff);
1286             if ((u32)diff.tv_sec >= g->p.nr_secs) {
1287                 g->stop_work = true;
1288                 break;
1289             }
1290         }
1291 
1292         /* Update the summary at most once per second: */
1293         if (start.tv_sec == stop.tv_sec)
1294             continue;
1295 
1296         /*
1297          * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1298          * by migrating to CPU#0:
1299          */
1300         if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1301             cpu_set_t *orig_mask;
1302             int target_cpu;
1303             int this_cpu;
1304 
1305             last_perturbance = stop.tv_sec;
1306 
1307             /*
1308              * Depending on where we are running, move into
1309              * the other half of the system, to create some
1310              * real disturbance:
1311              */
1312             this_cpu = g->threads[task_nr].curr_cpu;
1313             if (this_cpu < g->p.nr_cpus/2)
1314                 target_cpu = g->p.nr_cpus-1;
1315             else
1316                 target_cpu = 0;
1317 
1318             orig_mask = bind_to_cpu(target_cpu);
1319 
1320             /* Here we are running on the target CPU already */
1321             if (details >= 1)
1322                 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1323 
1324             bind_to_cpumask(orig_mask);
1325             CPU_FREE(orig_mask);
1326         }
1327 
1328         if (details >= 3) {
1329             timersub(&stop, &start, &diff);
1330             runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1331             runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1332 
1333             if (details >= 0) {
1334                 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1335                     process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1336             }
1337             fflush(stdout);
1338         }
1339         if (!last_task)
1340             continue;
1341 
1342         timersub(&stop, &start0, &diff);
1343         runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1344         runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1345 
1346         show_summary(runtime_ns_max, l, &convergence);
1347     }
1348 
1349     gettimeofday(&stop, NULL);
1350     timersub(&stop, &start0, &diff);
1351     td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1352     td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1353     secs = td->runtime_ns / NSEC_PER_SEC;
1354     td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1355 
1356     getrusage(RUSAGE_THREAD, &rusage);
1357     td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1358     td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1359     td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1360     td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1361 
1362     free_data(thread_data, g->p.bytes_thread);
1363 
1364     pthread_mutex_lock(&g->stop_work_mutex);
1365     g->bytes_done += bytes_done;
1366     pthread_mutex_unlock(&g->stop_work_mutex);
1367 
1368     return NULL;
1369 }
1370 
1371 /*
1372  * A worker process starts a couple of threads:
1373  */
1374 static void worker_process(int process_nr)
1375 {
1376     pthread_mutex_t process_lock;
1377     struct thread_data *td;
1378     pthread_t *pthreads;
1379     u8 *process_data;
1380     int task_nr;
1381     int ret;
1382     int t;
1383 
1384     pthread_mutex_init(&process_lock, NULL);
1385     set_taskname("process %d", process_nr);
1386 
1387     /*
1388      * Pick up the memory policy and the CPU binding of our first thread,
1389      * so that we initialize memory accordingly:
1390      */
1391     task_nr = process_nr*g->p.nr_threads;
1392     td = g->threads + task_nr;
1393 
1394     bind_to_memnode(td->bind_node);
1395     bind_to_cpumask(td->bind_cpumask);
1396 
1397     pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1398     process_data = setup_private_data(g->p.bytes_process);
1399 
1400     if (g->p.show_details >= 3) {
1401         printf(" # process %2d global mem: %p, process mem: %p\n",
1402             process_nr, g->data, process_data);
1403     }
1404 
1405     for (t = 0; t < g->p.nr_threads; t++) {
1406         task_nr = process_nr*g->p.nr_threads + t;
1407         td = g->threads + task_nr;
1408 
1409         td->process_data = process_data;
1410         td->process_nr   = process_nr;
1411         td->thread_nr    = t;
1412         td->task_nr  = task_nr;
1413         td->val          = rand();
1414         td->curr_cpu     = -1;
1415         td->process_lock = &process_lock;
1416 
1417         ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1418         BUG_ON(ret);
1419     }
1420 
1421     for (t = 0; t < g->p.nr_threads; t++) {
1422                 ret = pthread_join(pthreads[t], NULL);
1423         BUG_ON(ret);
1424     }
1425 
1426     free_data(process_data, g->p.bytes_process);
1427     free(pthreads);
1428 }
1429 
1430 static void print_summary(void)
1431 {
1432     if (g->p.show_details < 0)
1433         return;
1434 
1435     printf("\n ###\n");
1436     printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1437         g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1438     printf(" #      %5dx %5ldMB global  shared mem operations\n",
1439             g->p.nr_loops, g->p.bytes_global/1024/1024);
1440     printf(" #      %5dx %5ldMB process shared mem operations\n",
1441             g->p.nr_loops, g->p.bytes_process/1024/1024);
1442     printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1443             g->p.nr_loops, g->p.bytes_thread/1024/1024);
1444 
1445     printf(" ###\n");
1446 
1447     printf("\n ###\n"); fflush(stdout);
1448 }
1449 
1450 static void init_thread_data(void)
1451 {
1452     ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1453     int t;
1454 
1455     g->threads = zalloc_shared_data(size);
1456 
1457     for (t = 0; t < g->p.nr_tasks; t++) {
1458         struct thread_data *td = g->threads + t;
1459         size_t cpuset_size = CPU_ALLOC_SIZE(g->p.nr_cpus);
1460         int cpu;
1461 
1462         /* Allow all nodes by default: */
1463         td->bind_node = NUMA_NO_NODE;
1464 
1465         /* Allow all CPUs by default: */
1466         td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
1467         BUG_ON(!td->bind_cpumask);
1468         CPU_ZERO_S(cpuset_size, td->bind_cpumask);
1469         for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1470             CPU_SET_S(cpu, cpuset_size, td->bind_cpumask);
1471     }
1472 }
1473 
1474 static void deinit_thread_data(void)
1475 {
1476     ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1477     int t;
1478 
1479     /* Free the bind_cpumask allocated for thread_data */
1480     for (t = 0; t < g->p.nr_tasks; t++) {
1481         struct thread_data *td = g->threads + t;
1482         CPU_FREE(td->bind_cpumask);
1483     }
1484 
1485     free_data(g->threads, size);
1486 }
1487 
1488 static int init(void)
1489 {
1490     g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1491 
1492     /* Copy over options: */
1493     g->p = p0;
1494 
1495     g->p.nr_cpus = numa_num_configured_cpus();
1496 
1497     g->p.nr_nodes = numa_max_node() + 1;
1498 
1499     /* char array in count_process_nodes(): */
1500     BUG_ON(g->p.nr_nodes < 0);
1501 
1502     if (g->p.show_quiet && !g->p.show_details)
1503         g->p.show_details = -1;
1504 
1505     /* Some memory should be specified: */
1506     if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1507         return -1;
1508 
1509     if (g->p.mb_global_str) {
1510         g->p.mb_global = atof(g->p.mb_global_str);
1511         BUG_ON(g->p.mb_global < 0);
1512     }
1513 
1514     if (g->p.mb_proc_str) {
1515         g->p.mb_proc = atof(g->p.mb_proc_str);
1516         BUG_ON(g->p.mb_proc < 0);
1517     }
1518 
1519     if (g->p.mb_proc_locked_str) {
1520         g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1521         BUG_ON(g->p.mb_proc_locked < 0);
1522         BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1523     }
1524 
1525     if (g->p.mb_thread_str) {
1526         g->p.mb_thread = atof(g->p.mb_thread_str);
1527         BUG_ON(g->p.mb_thread < 0);
1528     }
1529 
1530     BUG_ON(g->p.nr_threads <= 0);
1531     BUG_ON(g->p.nr_proc <= 0);
1532 
1533     g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1534 
1535     g->p.bytes_global       = g->p.mb_global    *1024L*1024L;
1536     g->p.bytes_process      = g->p.mb_proc      *1024L*1024L;
1537     g->p.bytes_process_locked   = g->p.mb_proc_locked   *1024L*1024L;
1538     g->p.bytes_thread       = g->p.mb_thread    *1024L*1024L;
1539 
1540     g->data = setup_shared_data(g->p.bytes_global);
1541 
1542     /* Startup serialization: */
1543     init_global_mutex(&g->start_work_mutex);
1544     init_global_cond(&g->start_work_cond);
1545     init_global_mutex(&g->startup_mutex);
1546     init_global_cond(&g->startup_cond);
1547     init_global_mutex(&g->stop_work_mutex);
1548 
1549     init_thread_data();
1550 
1551     tprintf("#\n");
1552     if (parse_setup_cpu_list() || parse_setup_node_list())
1553         return -1;
1554     tprintf("#\n");
1555 
1556     print_summary();
1557 
1558     return 0;
1559 }
1560 
1561 static void deinit(void)
1562 {
1563     free_data(g->data, g->p.bytes_global);
1564     g->data = NULL;
1565 
1566     deinit_thread_data();
1567 
1568     free_data(g, sizeof(*g));
1569     g = NULL;
1570 }
1571 
1572 /*
1573  * Print a short or long result, depending on the verbosity setting:
1574  */
1575 static void print_res(const char *name, double val,
1576               const char *txt_unit, const char *txt_short, const char *txt_long)
1577 {
1578     if (!name)
1579         name = "main,";
1580 
1581     if (!g->p.show_quiet)
1582         printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1583     else
1584         printf(" %14.3f %s\n", val, txt_long);
1585 }
1586 
1587 static int __bench_numa(const char *name)
1588 {
1589     struct timeval start, stop, diff;
1590     u64 runtime_ns_min, runtime_ns_sum;
1591     pid_t *pids, pid, wpid;
1592     double delta_runtime;
1593     double runtime_avg;
1594     double runtime_sec_max;
1595     double runtime_sec_min;
1596     int wait_stat;
1597     double bytes;
1598     int i, t, p;
1599 
1600     if (init())
1601         return -1;
1602 
1603     pids = zalloc(g->p.nr_proc * sizeof(*pids));
1604     pid = -1;
1605 
1606     if (g->p.serialize_startup) {
1607         tprintf(" #\n");
1608         tprintf(" # Startup synchronization: ..."); fflush(stdout);
1609     }
1610 
1611     gettimeofday(&start, NULL);
1612 
1613     for (i = 0; i < g->p.nr_proc; i++) {
1614         pid = fork();
1615         dprintf(" # process %2d: PID %d\n", i, pid);
1616 
1617         BUG_ON(pid < 0);
1618         if (!pid) {
1619             /* Child process: */
1620             worker_process(i);
1621 
1622             exit(0);
1623         }
1624         pids[i] = pid;
1625 
1626     }
1627 
1628     if (g->p.serialize_startup) {
1629         bool threads_ready = false;
1630         double startup_sec;
1631 
1632         /*
1633          * Wait for all the threads to start up. The last thread will
1634          * signal this process.
1635          */
1636         pthread_mutex_lock(&g->startup_mutex);
1637         while (g->nr_tasks_started != g->p.nr_tasks)
1638             pthread_cond_wait(&g->startup_cond, &g->startup_mutex);
1639 
1640         pthread_mutex_unlock(&g->startup_mutex);
1641 
1642         /* Wait for all threads to be at the start_work_cond. */
1643         while (!threads_ready) {
1644             pthread_mutex_lock(&g->start_work_mutex);
1645             threads_ready = (g->nr_tasks_working == g->p.nr_tasks);
1646             pthread_mutex_unlock(&g->start_work_mutex);
1647             if (!threads_ready)
1648                 usleep(1);
1649         }
1650 
1651         gettimeofday(&stop, NULL);
1652 
1653         timersub(&stop, &start, &diff);
1654 
1655         startup_sec = diff.tv_sec * NSEC_PER_SEC;
1656         startup_sec += diff.tv_usec * NSEC_PER_USEC;
1657         startup_sec /= NSEC_PER_SEC;
1658 
1659         tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1660         tprintf(" #\n");
1661 
1662         start = stop;
1663         /* Start all threads running. */
1664         pthread_mutex_lock(&g->start_work_mutex);
1665         g->start_work = true;
1666         pthread_mutex_unlock(&g->start_work_mutex);
1667         pthread_cond_broadcast(&g->start_work_cond);
1668     } else {
1669         gettimeofday(&start, NULL);
1670     }
1671 
1672     /* Parent process: */
1673 
1674 
1675     for (i = 0; i < g->p.nr_proc; i++) {
1676         wpid = waitpid(pids[i], &wait_stat, 0);
1677         BUG_ON(wpid < 0);
1678         BUG_ON(!WIFEXITED(wait_stat));
1679 
1680     }
1681 
1682     runtime_ns_sum = 0;
1683     runtime_ns_min = -1LL;
1684 
1685     for (t = 0; t < g->p.nr_tasks; t++) {
1686         u64 thread_runtime_ns = g->threads[t].runtime_ns;
1687 
1688         runtime_ns_sum += thread_runtime_ns;
1689         runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1690     }
1691 
1692     gettimeofday(&stop, NULL);
1693     timersub(&stop, &start, &diff);
1694 
1695     BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1696 
1697     tprintf("\n ###\n");
1698     tprintf("\n");
1699 
1700     runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1701     runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1702     runtime_sec_max /= NSEC_PER_SEC;
1703 
1704     runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1705 
1706     bytes = g->bytes_done;
1707     runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1708 
1709     if (g->p.measure_convergence) {
1710         print_res(name, runtime_sec_max,
1711             "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1712     }
1713 
1714     print_res(name, runtime_sec_max,
1715         "secs,", "runtime-max/thread",  "secs slowest (max) thread-runtime");
1716 
1717     print_res(name, runtime_sec_min,
1718         "secs,", "runtime-min/thread",  "secs fastest (min) thread-runtime");
1719 
1720     print_res(name, runtime_avg,
1721         "secs,", "runtime-avg/thread",  "secs average thread-runtime");
1722 
1723     delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1724     print_res(name, delta_runtime / runtime_sec_max * 100.0,
1725         "%,", "spread-runtime/thread",  "% difference between max/avg runtime");
1726 
1727     print_res(name, bytes / g->p.nr_tasks / 1e9,
1728         "GB,", "data/thread",       "GB data processed, per thread");
1729 
1730     print_res(name, bytes / 1e9,
1731         "GB,", "data-total",        "GB data processed, total");
1732 
1733     print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1734         "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1735 
1736     print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1737         "GB/sec,", "thread-speed",  "GB/sec/thread speed");
1738 
1739     print_res(name, bytes / runtime_sec_max / 1e9,
1740         "GB/sec,", "total-speed",   "GB/sec total speed");
1741 
1742     if (g->p.show_details >= 2) {
1743         char tname[14 + 2 * 11 + 1];
1744         struct thread_data *td;
1745         for (p = 0; p < g->p.nr_proc; p++) {
1746             for (t = 0; t < g->p.nr_threads; t++) {
1747                 memset(tname, 0, sizeof(tname));
1748                 td = g->threads + p*g->p.nr_threads + t;
1749                 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1750                 print_res(tname, td->speed_gbs,
1751                     "GB/sec",   "thread-speed", "GB/sec/thread speed");
1752                 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1753                     "secs", "thread-system-time", "system CPU time/thread");
1754                 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1755                     "secs", "thread-user-time", "user CPU time/thread");
1756             }
1757         }
1758     }
1759 
1760     free(pids);
1761 
1762     deinit();
1763 
1764     return 0;
1765 }
1766 
1767 #define MAX_ARGS 50
1768 
1769 static int command_size(const char **argv)
1770 {
1771     int size = 0;
1772 
1773     while (*argv) {
1774         size++;
1775         argv++;
1776     }
1777 
1778     BUG_ON(size >= MAX_ARGS);
1779 
1780     return size;
1781 }
1782 
1783 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1784 {
1785     int i;
1786 
1787     printf("\n # Running %s \"perf bench numa", name);
1788 
1789     for (i = 0; i < argc; i++)
1790         printf(" %s", argv[i]);
1791 
1792     printf("\"\n");
1793 
1794     memset(p, 0, sizeof(*p));
1795 
1796     /* Initialize nonzero defaults: */
1797 
1798     p->serialize_startup        = 1;
1799     p->data_reads           = true;
1800     p->data_writes          = true;
1801     p->data_backwards       = true;
1802     p->data_rand_walk       = true;
1803     p->nr_loops         = -1;
1804     p->init_random          = true;
1805     p->mb_global_str        = "1";
1806     p->nr_proc          = 1;
1807     p->nr_threads           = 1;
1808     p->nr_secs          = 5;
1809     p->run_all          = argc == 1;
1810 }
1811 
1812 static int run_bench_numa(const char *name, const char **argv)
1813 {
1814     int argc = command_size(argv);
1815 
1816     init_params(&p0, name, argc, argv);
1817     argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1818     if (argc)
1819         goto err;
1820 
1821     if (__bench_numa(name))
1822         goto err;
1823 
1824     return 0;
1825 
1826 err:
1827     return -1;
1828 }
1829 
1830 #define OPT_BW_RAM      "-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1831 #define OPT_BW_RAM_NOTHP    OPT_BW_RAM,     "--thp", "-1"
1832 
1833 #define OPT_CONV        "-s", "100", "-zZ0qcm", "--thp", " 1"
1834 #define OPT_CONV_NOTHP      OPT_CONV,       "--thp", "-1"
1835 
1836 #define OPT_BW          "-s",  "20", "-zZ0q",   "--thp", " 1"
1837 #define OPT_BW_NOTHP        OPT_BW,         "--thp", "-1"
1838 
1839 /*
1840  * The built-in test-suite executed by "perf bench numa -a".
1841  *
1842  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1843  */
1844 static const char *tests[][MAX_ARGS] = {
1845    /* Basic single-stream NUMA bandwidth measurements: */
1846    { "RAM-bw-local,",     "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1847               "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1848    { "RAM-bw-local-NOTHP,",
1849               "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1850               "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1851    { "RAM-bw-remote,",    "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1852               "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1853 
1854    /* 2-stream NUMA bandwidth measurements: */
1855    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1856                "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1857    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1858                "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1859 
1860    /* Cross-stream NUMA bandwidth measurement: */
1861    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1862                "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1863 
1864    /* Convergence latency measurements: */
1865    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1866    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1867    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1868    { " 2x3-convergence,", "mem",  "-p",  "2", "-t",  "3", "-P", "1020", OPT_CONV },
1869    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1870    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1871    { " 4x4-convergence-NOTHP,",
1872               "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1873    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1874    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1875    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1876    { " 8x4-convergence-NOTHP,",
1877               "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1878    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1879    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1880    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1881    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1882    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1883 
1884    /* Various NUMA process/thread layout bandwidth measurements: */
1885    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1886    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1887    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1888    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1889    { " 8x1-bw-process-NOTHP,",
1890               "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1891    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1892 
1893    { " 1x4-bw-thread,",   "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1894    { " 1x8-bw-thread,",   "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1895    { "1x16-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1896    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1897 
1898    { " 2x3-bw-process,",  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1899    { " 4x4-bw-process,",  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1900    { " 4x6-bw-process,",  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1901    { " 4x8-bw-process,",  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1902    { " 4x8-bw-process-NOTHP,",
1903               "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1904    { " 3x3-bw-process,",  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1905    { " 5x5-bw-process,",  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1906 
1907    { "2x16-bw-process,",  "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1908    { "1x32-bw-process,",  "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1909 
1910    { "numa02-bw,",        "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1911    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1912    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1913    { "numa01-bw-thread-NOTHP,",
1914               "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1915 };
1916 
1917 static int bench_all(void)
1918 {
1919     int nr = ARRAY_SIZE(tests);
1920     int ret;
1921     int i;
1922 
1923     ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1924     BUG_ON(ret < 0);
1925 
1926     for (i = 0; i < nr; i++) {
1927         run_bench_numa(tests[i][0], tests[i] + 1);
1928     }
1929 
1930     printf("\n");
1931 
1932     return 0;
1933 }
1934 
1935 int bench_numa(int argc, const char **argv)
1936 {
1937     init_params(&p0, "main,", argc, argv);
1938     argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1939     if (argc)
1940         goto err;
1941 
1942     if (p0.run_all)
1943         return bench_all();
1944 
1945     if (__bench_numa(NULL))
1946         goto err;
1947 
1948     return 0;
1949 
1950 err:
1951     usage_with_options(numa_usage, options);
1952     return -1;
1953 }
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!