diff options
Diffstat (limited to 'collectors/proc.plugin/proc_stat.c')
| -rwxr-xr-x[-rw-r--r--] | collectors/proc.plugin/proc_stat.c | 204 |
1 files changed, 178 insertions, 26 deletions
diff --git a/collectors/proc.plugin/proc_stat.c b/collectors/proc.plugin/proc_stat.c index fb77df647..931b415a5 100644..100755 --- a/collectors/proc.plugin/proc_stat.c +++ b/collectors/proc.plugin/proc_stat.c @@ -12,9 +12,23 @@ struct per_core_single_number_file { RRDDIM *rd; }; +struct last_ticks { + collected_number frequency; + collected_number ticks; +}; + +// This is an extension of struct per_core_single_number_file at CPU_FREQ_INDEX. +// Either scaling_cur_freq or time_in_state file is used at one time. +struct per_core_time_in_state_file { + const char *filename; + procfile *ff; + size_t last_ticks_len; + struct last_ticks *last_ticks; +}; + #define CORE_THROTTLE_COUNT_INDEX 0 #define PACKAGE_THROTTLE_COUNT_INDEX 1 -#define SCALING_CUR_FREQ_INDEX 2 +#define CPU_FREQ_INDEX 2 #define PER_CORE_FILES 3 struct cpu_chart { @@ -33,6 +47,8 @@ struct cpu_chart { RRDDIM *rd_guest_nice; struct per_core_single_number_file files[PER_CORE_FILES]; + + struct per_core_time_in_state_file time_in_state_files; }; static int keep_per_core_fds_open = CONFIG_BOOLEAN_YES; @@ -87,7 +103,6 @@ static int read_per_core_files(struct cpu_chart *all_cpu_charts, size_t len, siz f->found = 1; f->value = str2ll(buf, NULL); - // info("read '%s', parsed as " COLLECTED_NUMBER_FORMAT, buf, f->value); if(likely(f->value != 0)) files_nonzero++; } @@ -101,6 +116,112 @@ static int read_per_core_files(struct cpu_chart *all_cpu_charts, size_t len, siz return (int)files_nonzero; } +static int read_per_core_time_in_state_files(struct cpu_chart *all_cpu_charts, size_t len, size_t index) { + size_t x, files_read = 0, files_nonzero = 0; + + for(x = 0; x < len ; x++) { + struct per_core_single_number_file *f = &all_cpu_charts[x].files[index]; + struct per_core_time_in_state_file *tsf = &all_cpu_charts[x].time_in_state_files; + + f->found = 0; + + if(unlikely(!tsf->filename)) + continue; + + if(unlikely(!tsf->ff)) { + tsf->ff = procfile_open(tsf->filename, " \t:", PROCFILE_FLAG_DEFAULT); + if(unlikely(!tsf->ff)) + { + error("Cannot open file '%s'", tsf->filename); + continue; + } + } + + tsf->ff = procfile_readall(tsf->ff); + if(unlikely(!tsf->ff)) { + error("Cannot read file '%s'", tsf->filename); + procfile_close(tsf->ff); + tsf->ff = NULL; + continue; + } + else { + // successful read + + size_t lines = procfile_lines(tsf->ff), l; + size_t words; + unsigned long long total_ticks_since_last = 0, avg_freq = 0; + + // Check if there is at least one frequency in time_in_state + if (procfile_word(tsf->ff, 0)[0] == '\0') { + if(unlikely(keep_per_core_fds_open != CONFIG_BOOLEAN_YES)) { + procfile_close(tsf->ff); + tsf->ff = NULL; + } + // TODO: Is there a better way to avoid spikes than calculating the average over + // the whole period under schedutil governor? + // freez(tsf->last_ticks); + // tsf->last_ticks = NULL; + // tsf->last_ticks_len = 0; + continue; + } + + if (unlikely(tsf->last_ticks_len < lines || tsf->last_ticks == NULL)) { + tsf->last_ticks = reallocz(tsf->last_ticks, sizeof(struct last_ticks) * lines); + memset(tsf->last_ticks, 0, sizeof(struct last_ticks) * lines); + tsf->last_ticks_len = lines; + } + + f->value = 0; + + for(l = 0; l < lines - 1 ;l++) { + unsigned long long frequency = 0, ticks = 0, ticks_since_last = 0; + + words = procfile_linewords(tsf->ff, l); + if(unlikely(words < 2)) { + error("Cannot read time_in_state line. Expected 2 params, read %zu.", words); + continue; + } + frequency = str2ull(procfile_lineword(tsf->ff, l, 0)); + ticks = str2ull(procfile_lineword(tsf->ff, l, 1)); + + // It is assumed that frequencies are static and sorted + ticks_since_last = ticks - tsf->last_ticks[l].ticks; + tsf->last_ticks[l].frequency = frequency; + tsf->last_ticks[l].ticks = ticks; + + total_ticks_since_last += ticks_since_last; + avg_freq += frequency * ticks_since_last; + + } + + if (likely(total_ticks_since_last)) { + avg_freq /= total_ticks_since_last; + f->value = avg_freq; + } + + if(unlikely(keep_per_core_fds_open != CONFIG_BOOLEAN_YES)) { + procfile_close(tsf->ff); + tsf->ff = NULL; + } + } + + files_read++; + + f->found = 1; + + if(likely(f->value != 0)) + files_nonzero++; + } + + if(unlikely(files_read == 0)) + return -1; + + if(unlikely(files_nonzero == 0)) + return 0; + + return (int)files_nonzero; +} + static void chart_per_core_files(struct cpu_chart *all_cpu_charts, size_t len, size_t index, RRDSET *st, collected_number multiplier, collected_number divisor, RRD_ALGORITHM algorithm) { size_t x; for(x = 0; x < len ; x++) { @@ -122,10 +243,11 @@ int do_proc_stat(int update_every, usec_t dt) { static struct cpu_chart *all_cpu_charts = NULL; static size_t all_cpu_charts_size = 0; static procfile *ff = NULL; - static int do_cpu = -1, do_cpu_cores = -1, do_interrupts = -1, do_context = -1, do_forks = -1, do_processes = -1, do_core_throttle_count = -1, do_package_throttle_count = -1, do_scaling_cur_freq = -1; + static int do_cpu = -1, do_cpu_cores = -1, do_interrupts = -1, do_context = -1, do_forks = -1, do_processes = -1, do_core_throttle_count = -1, do_package_throttle_count = -1, do_cpu_freq = -1; static uint32_t hash_intr, hash_ctxt, hash_processes, hash_procs_running, hash_procs_blocked; - static char *core_throttle_count_filename = NULL, *package_throttle_count_filename = NULL, *scaling_cur_freq_filename = NULL; + static char *core_throttle_count_filename = NULL, *package_throttle_count_filename = NULL, *scaling_cur_freq_filename = NULL, *time_in_state_filename = NULL; static RRDVAR *cpus_var = NULL; + static int accurate_freq_avail = 0, accurate_freq_is_used = 0; size_t cores_found = (size_t)processors; if(unlikely(do_cpu == -1)) { @@ -137,25 +259,25 @@ int do_proc_stat(int update_every, usec_t dt) { do_processes = config_get_boolean("plugin:proc:/proc/stat", "processes running", CONFIG_BOOLEAN_YES); // give sane defaults based on the number of processors - if(processors > 50) { + if(unlikely(processors > 50)) { // the system has too many processors keep_per_core_fds_open = CONFIG_BOOLEAN_NO; do_core_throttle_count = CONFIG_BOOLEAN_NO; do_package_throttle_count = CONFIG_BOOLEAN_NO; - do_scaling_cur_freq = CONFIG_BOOLEAN_NO; + do_cpu_freq = CONFIG_BOOLEAN_NO; } else { // the system has a reasonable number of processors keep_per_core_fds_open = CONFIG_BOOLEAN_YES; do_core_throttle_count = CONFIG_BOOLEAN_AUTO; do_package_throttle_count = CONFIG_BOOLEAN_NO; - do_scaling_cur_freq = CONFIG_BOOLEAN_NO; + do_cpu_freq = CONFIG_BOOLEAN_YES; } keep_per_core_fds_open = config_get_boolean("plugin:proc:/proc/stat", "keep per core files open", keep_per_core_fds_open); do_core_throttle_count = config_get_boolean_ondemand("plugin:proc:/proc/stat", "core_throttle_count", do_core_throttle_count); do_package_throttle_count = config_get_boolean_ondemand("plugin:proc:/proc/stat", "package_throttle_count", do_package_throttle_count); - do_scaling_cur_freq = config_get_boolean_ondemand("plugin:proc:/proc/stat", "scaling_cur_freq", do_scaling_cur_freq); + do_cpu_freq = config_get_boolean_ondemand("plugin:proc:/proc/stat", "cpu frequency", do_cpu_freq); hash_intr = simple_hash("intr"); hash_ctxt = simple_hash("ctxt"); @@ -172,6 +294,9 @@ int do_proc_stat(int update_every, usec_t dt) { snprintfz(filename, FILENAME_MAX, "%s%s", netdata_configured_host_prefix, "/sys/devices/system/cpu/%s/cpufreq/scaling_cur_freq"); scaling_cur_freq_filename = config_get("plugin:proc:/proc/stat", "scaling_cur_freq filename to monitor", filename); + + snprintfz(filename, FILENAME_MAX, "%s%s", netdata_configured_host_prefix, "/sys/devices/system/cpu/%s/cpufreq/stats/time_in_state"); + time_in_state_filename = config_get("plugin:proc:/proc/stat", "time_in_state filename to monitor", filename); } if(unlikely(!ff)) { @@ -202,7 +327,7 @@ int do_proc_stat(int update_every, usec_t dt) { } size_t core = (row_key[3] == '\0') ? 0 : str2ul(&row_key[3]) + 1; - if(core > 0) cores_found = core; + if(likely(core > 0)) cores_found = core; if(likely((core == 0 && do_cpu) || (core > 0 && do_cpu_cores))) { char *id; @@ -227,7 +352,7 @@ int do_proc_stat(int update_every, usec_t dt) { char *title, *type, *context, *family; long priority; - if(core >= all_cpu_charts_size) { + if(unlikely(core >= all_cpu_charts_size)) { size_t old_cpu_charts_size = all_cpu_charts_size; all_cpu_charts_size = core + 1; all_cpu_charts = reallocz(all_cpu_charts, sizeof(struct cpu_chart) * all_cpu_charts_size); @@ -238,7 +363,7 @@ int do_proc_stat(int update_every, usec_t dt) { if(unlikely(!cpu_chart->st)) { cpu_chart->id = strdupz(id); - if(core == 0) { + if(unlikely(core == 0)) { title = "Total CPU utilization"; type = "system"; context = "system.cpu"; @@ -252,9 +377,6 @@ int do_proc_stat(int update_every, usec_t dt) { family = "utilization"; priority = NETDATA_CHART_PRIO_CPU_PER_CORE; - // TODO: check for /sys/devices/system/cpu/cpu*/cpufreq/scaling_cur_freq - // TODO: check for /sys/devices/system/cpu/cpu*/cpufreq/stats/time_in_state - char filename[FILENAME_MAX + 1]; struct stat stbuf; @@ -276,12 +398,23 @@ int do_proc_stat(int update_every, usec_t dt) { } } - if(do_scaling_cur_freq != CONFIG_BOOLEAN_NO) { + if(do_cpu_freq != CONFIG_BOOLEAN_NO) { + snprintfz(filename, FILENAME_MAX, scaling_cur_freq_filename, id); + + if (stat(filename, &stbuf) == 0) { + cpu_chart->files[CPU_FREQ_INDEX].filename = strdupz(filename); + cpu_chart->files[CPU_FREQ_INDEX].fd = -1; + do_cpu_freq = CONFIG_BOOLEAN_YES; + } + + snprintfz(filename, FILENAME_MAX, time_in_state_filename, id); + if (stat(filename, &stbuf) == 0) { - cpu_chart->files[SCALING_CUR_FREQ_INDEX].filename = strdupz(filename); - cpu_chart->files[SCALING_CUR_FREQ_INDEX].fd = -1; - do_scaling_cur_freq = CONFIG_BOOLEAN_YES; + cpu_chart->time_in_state_files.filename = strdupz(filename); + cpu_chart->time_in_state_files.ff = NULL; + do_cpu_freq = CONFIG_BOOLEAN_YES; + accurate_freq_avail = 1; } } } @@ -532,21 +665,40 @@ int do_proc_stat(int update_every, usec_t dt) { } } - if(likely(do_scaling_cur_freq != CONFIG_BOOLEAN_NO)) { - int r = read_per_core_files(&all_cpu_charts[1], all_cpu_charts_size - 1, SCALING_CUR_FREQ_INDEX); - if(likely(r != -1 && (do_scaling_cur_freq == CONFIG_BOOLEAN_YES || r > 0))) { - do_scaling_cur_freq = CONFIG_BOOLEAN_YES; + if(likely(do_cpu_freq != CONFIG_BOOLEAN_NO)) { + char filename[FILENAME_MAX + 1]; + int r = 0; + + if (accurate_freq_avail) { + r = read_per_core_time_in_state_files(&all_cpu_charts[1], all_cpu_charts_size - 1, CPU_FREQ_INDEX); + if(r > 0 && !accurate_freq_is_used) { + accurate_freq_is_used = 1; + snprintfz(filename, FILENAME_MAX, time_in_state_filename, "cpu*"); + info("cpufreq is using %s", filename); + } + } + if (r < 1) { + r = read_per_core_files(&all_cpu_charts[1], all_cpu_charts_size - 1, CPU_FREQ_INDEX); + if(accurate_freq_is_used) { + accurate_freq_is_used = 0; + snprintfz(filename, FILENAME_MAX, scaling_cur_freq_filename, "cpu*"); + info("cpufreq fell back to %s", filename); + } + } + + if(likely(r != -1 && (do_cpu_freq == CONFIG_BOOLEAN_YES || r > 0))) { + do_cpu_freq = CONFIG_BOOLEAN_YES; static RRDSET *st_scaling_cur_freq = NULL; if(unlikely(!st_scaling_cur_freq)) st_scaling_cur_freq = rrdset_create_localhost( "cpu" - , "scaling_cur_freq" + , "cpufreq" , NULL , "cpufreq" - , "cpu.scaling_cur_freq" - , "Per CPU Core, Current CPU Scaling Frequency" + , "cpufreq.cpufreq" + , "Current CPU Frequency" , "MHz" , PLUGIN_PROC_NAME , PLUGIN_PROC_MODULE_STAT_NAME @@ -557,7 +709,7 @@ int do_proc_stat(int update_every, usec_t dt) { else rrdset_next(st_scaling_cur_freq); - chart_per_core_files(&all_cpu_charts[1], all_cpu_charts_size - 1, SCALING_CUR_FREQ_INDEX, st_scaling_cur_freq, 1, 1000, RRD_ALGORITHM_ABSOLUTE); + chart_per_core_files(&all_cpu_charts[1], all_cpu_charts_size - 1, CPU_FREQ_INDEX, st_scaling_cur_freq, 1, 1000, RRD_ALGORITHM_ABSOLUTE); rrdset_done(st_scaling_cur_freq); } } |