| 1 | /* |
| 2 | * Copyright (C) 2012 David Goulet <dgoulet@efficios.com> |
| 3 | * Copyright (C) 2013 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 4 | * |
| 5 | * SPDX-License-Identifier: GPL-2.0-only |
| 6 | * |
| 7 | */ |
| 8 | |
| 9 | #define _LGPL_SOURCE |
| 10 | #include <assert.h> |
| 11 | #include <inttypes.h> |
| 12 | #include <stdio.h> |
| 13 | #include <stdlib.h> |
| 14 | #include <time.h> |
| 15 | |
| 16 | #include <common/defaults.h> |
| 17 | #include <common/error.h> |
| 18 | #include <common/macros.h> |
| 19 | #include <common/sessiond-comm/inet.h> |
| 20 | |
| 21 | #include <lttng/health-internal.h> |
| 22 | |
| 23 | /* |
| 24 | * An application-specific error state for unregistered thread keeps |
| 25 | * track of thread errors. A thread reporting a health error, normally |
| 26 | * unregisters and quits. This makes the TLS health state not available |
| 27 | * to the health_check_state() call so on unregister we update this |
| 28 | * global error array so we can keep track of which thread was on error |
| 29 | * if the TLS health state has been removed. |
| 30 | */ |
| 31 | struct health_app { |
| 32 | /* List of health state, for each application thread */ |
| 33 | struct cds_list_head list; |
| 34 | /* |
| 35 | * This lock ensures that TLS memory used for the node and its |
| 36 | * container structure don't get reclaimed after the TLS owner |
| 37 | * thread exits until we have finished using it. |
| 38 | */ |
| 39 | pthread_mutex_t lock; |
| 40 | int nr_types; |
| 41 | struct timespec time_delta; |
| 42 | /* Health flags containing thread type error state */ |
| 43 | enum health_flags *flags; |
| 44 | }; |
| 45 | |
| 46 | /* Define TLS health state. */ |
| 47 | DEFINE_URCU_TLS(struct health_state, health_state); |
| 48 | |
| 49 | /* |
| 50 | * Initialize health check subsytem. |
| 51 | */ |
| 52 | static |
| 53 | void health_init(struct health_app *ha) |
| 54 | { |
| 55 | /* |
| 56 | * Get the maximum value between the default delta value and the TCP |
| 57 | * timeout with a safety net of the default health check delta. |
| 58 | */ |
| 59 | ha->time_delta.tv_sec = max_t(unsigned long, |
| 60 | lttcomm_inet_tcp_timeout + DEFAULT_HEALTH_CHECK_DELTA_S, |
| 61 | ha->time_delta.tv_sec); |
| 62 | DBG("Health check time delta in seconds set to %lu", |
| 63 | ha->time_delta.tv_sec); |
| 64 | } |
| 65 | |
| 66 | struct health_app *health_app_create(int nr_types) |
| 67 | { |
| 68 | struct health_app *ha; |
| 69 | |
| 70 | ha = zmalloc(sizeof(*ha)); |
| 71 | if (!ha) { |
| 72 | return NULL; |
| 73 | } |
| 74 | ha->flags = zmalloc(sizeof(*ha->flags) * nr_types); |
| 75 | if (!ha->flags) { |
| 76 | goto error_flags; |
| 77 | } |
| 78 | CDS_INIT_LIST_HEAD(&ha->list); |
| 79 | pthread_mutex_init(&ha->lock, NULL); |
| 80 | ha->nr_types = nr_types; |
| 81 | ha->time_delta.tv_sec = DEFAULT_HEALTH_CHECK_DELTA_S; |
| 82 | ha->time_delta.tv_nsec = DEFAULT_HEALTH_CHECK_DELTA_NS; |
| 83 | health_init(ha); |
| 84 | return ha; |
| 85 | |
| 86 | error_flags: |
| 87 | free(ha); |
| 88 | return NULL; |
| 89 | } |
| 90 | |
| 91 | void health_app_destroy(struct health_app *ha) |
| 92 | { |
| 93 | free(ha->flags); |
| 94 | free(ha); |
| 95 | } |
| 96 | |
| 97 | /* |
| 98 | * Lock health state global list mutex. |
| 99 | */ |
| 100 | static void state_lock(struct health_app *ha) |
| 101 | { |
| 102 | pthread_mutex_lock(&ha->lock); |
| 103 | } |
| 104 | |
| 105 | /* |
| 106 | * Unlock health state global list mutex. |
| 107 | */ |
| 108 | static void state_unlock(struct health_app *ha) |
| 109 | { |
| 110 | pthread_mutex_unlock(&ha->lock); |
| 111 | } |
| 112 | |
| 113 | /* |
| 114 | * Set time difference in res from time_a and time_b. |
| 115 | */ |
| 116 | static void time_diff(const struct timespec *time_a, |
| 117 | const struct timespec *time_b, struct timespec *res) |
| 118 | { |
| 119 | if (time_a->tv_nsec - time_b->tv_nsec < 0) { |
| 120 | res->tv_sec = time_a->tv_sec - time_b->tv_sec - 1; |
| 121 | res->tv_nsec = 1000000000L + time_a->tv_sec - time_b->tv_sec; |
| 122 | } else { |
| 123 | res->tv_sec = time_a->tv_sec - time_b->tv_sec; |
| 124 | res->tv_nsec = time_a->tv_nsec - time_b->tv_nsec; |
| 125 | } |
| 126 | } |
| 127 | |
| 128 | /* |
| 129 | * Return true if time_a - time_b > diff, else false. |
| 130 | */ |
| 131 | static int time_diff_gt(const struct timespec *time_a, |
| 132 | const struct timespec *time_b, const struct timespec *diff) |
| 133 | { |
| 134 | struct timespec res; |
| 135 | |
| 136 | time_diff(time_a, time_b, &res); |
| 137 | time_diff(&res, diff, &res); |
| 138 | |
| 139 | if (res.tv_sec > 0) { |
| 140 | return 1; |
| 141 | } else if (res.tv_sec == 0 && res.tv_nsec > 0) { |
| 142 | return 1; |
| 143 | } |
| 144 | |
| 145 | return 0; |
| 146 | } |
| 147 | |
| 148 | /* |
| 149 | * Validate health state. Checks for the error flag or health conditions. |
| 150 | * |
| 151 | * Return 0 if health is bad or else 1. |
| 152 | */ |
| 153 | static int validate_state(struct health_app *ha, struct health_state *state) |
| 154 | { |
| 155 | int retval = 1, ret; |
| 156 | unsigned long current, last; |
| 157 | struct timespec current_time; |
| 158 | |
| 159 | assert(state); |
| 160 | |
| 161 | last = state->last; |
| 162 | current = uatomic_read(&state->current); |
| 163 | |
| 164 | ret = lttng_clock_gettime(CLOCK_MONOTONIC, ¤t_time); |
| 165 | if (ret < 0) { |
| 166 | PERROR("Error reading time\n"); |
| 167 | /* error */ |
| 168 | retval = 0; |
| 169 | goto end; |
| 170 | } |
| 171 | |
| 172 | /* |
| 173 | * Thread is in bad health if flag HEALTH_ERROR is set. It is also in bad |
| 174 | * health if, after the delta delay has passed, its the progress counter |
| 175 | * has not moved and it has NOT been waiting for a poll() call. |
| 176 | */ |
| 177 | if (uatomic_read(&state->flags) & HEALTH_ERROR) { |
| 178 | retval = 0; |
| 179 | goto end; |
| 180 | } |
| 181 | |
| 182 | /* |
| 183 | * Initial condition need to update the last counter and sample time, but |
| 184 | * should not check health in this initial case, because we don't know how |
| 185 | * much time has passed. |
| 186 | */ |
| 187 | if (state->last_time.tv_sec == 0 && state->last_time.tv_nsec == 0) { |
| 188 | /* update last counter and last sample time */ |
| 189 | state->last = current; |
| 190 | memcpy(&state->last_time, ¤t_time, sizeof(current_time)); |
| 191 | } else { |
| 192 | if (time_diff_gt(¤t_time, &state->last_time, |
| 193 | &ha->time_delta)) { |
| 194 | if (current == last && !HEALTH_IS_IN_POLL(current)) { |
| 195 | /* error */ |
| 196 | retval = 0; |
| 197 | } |
| 198 | /* update last counter and last sample time */ |
| 199 | state->last = current; |
| 200 | memcpy(&state->last_time, ¤t_time, sizeof(current_time)); |
| 201 | |
| 202 | /* On error, stop right now and notify caller. */ |
| 203 | if (retval == 0) { |
| 204 | goto end; |
| 205 | } |
| 206 | } |
| 207 | } |
| 208 | |
| 209 | end: |
| 210 | DBG("Health state current %lu, last %lu, ret %d", |
| 211 | current, last, ret); |
| 212 | return retval; |
| 213 | } |
| 214 | |
| 215 | /* |
| 216 | * Check health of a specific health type. Note that if a thread has not yet |
| 217 | * initialize its health subsystem or has quit, it's considered in a good |
| 218 | * state. |
| 219 | * |
| 220 | * Return 0 if health is bad or else 1. |
| 221 | */ |
| 222 | int health_check_state(struct health_app *ha, int type) |
| 223 | { |
| 224 | int retval = 1; |
| 225 | struct health_state *state; |
| 226 | |
| 227 | assert(type < ha->nr_types); |
| 228 | |
| 229 | state_lock(ha); |
| 230 | |
| 231 | cds_list_for_each_entry(state, &ha->list, node) { |
| 232 | int ret; |
| 233 | |
| 234 | if (state->type != type) { |
| 235 | continue; |
| 236 | } |
| 237 | |
| 238 | ret = validate_state(ha, state); |
| 239 | if (!ret) { |
| 240 | retval = 0; |
| 241 | goto end; |
| 242 | } |
| 243 | } |
| 244 | |
| 245 | /* Check the global state since some state might not be visible anymore. */ |
| 246 | if (ha->flags[type] & HEALTH_ERROR) { |
| 247 | retval = 0; |
| 248 | } |
| 249 | |
| 250 | end: |
| 251 | state_unlock(ha); |
| 252 | |
| 253 | DBG("Health check for type %d is %s", (int) type, |
| 254 | (retval == 0) ? "BAD" : "GOOD"); |
| 255 | return retval; |
| 256 | } |
| 257 | |
| 258 | /* |
| 259 | * Init health state. |
| 260 | */ |
| 261 | void health_register(struct health_app *ha, int type) |
| 262 | { |
| 263 | assert(type < ha->nr_types); |
| 264 | |
| 265 | /* Init TLS state. */ |
| 266 | uatomic_set(&URCU_TLS(health_state).last, 0); |
| 267 | uatomic_set(&URCU_TLS(health_state).last_time.tv_sec, 0); |
| 268 | uatomic_set(&URCU_TLS(health_state).last_time.tv_nsec, 0); |
| 269 | uatomic_set(&URCU_TLS(health_state).current, 0); |
| 270 | uatomic_set(&URCU_TLS(health_state).flags, 0); |
| 271 | uatomic_set(&URCU_TLS(health_state).type, type); |
| 272 | |
| 273 | /* Add it to the global TLS state list. */ |
| 274 | state_lock(ha); |
| 275 | cds_list_add(&URCU_TLS(health_state).node, &ha->list); |
| 276 | state_unlock(ha); |
| 277 | } |
| 278 | |
| 279 | /* |
| 280 | * Remove node from global list. |
| 281 | */ |
| 282 | void health_unregister(struct health_app *ha) |
| 283 | { |
| 284 | state_lock(ha); |
| 285 | /* |
| 286 | * On error, set the global_error_state since we are about to remove |
| 287 | * the node from the global list. |
| 288 | */ |
| 289 | if (uatomic_read(&URCU_TLS(health_state).flags) & HEALTH_ERROR) { |
| 290 | uatomic_set(&ha->flags[URCU_TLS(health_state).type], |
| 291 | HEALTH_ERROR); |
| 292 | } |
| 293 | cds_list_del(&URCU_TLS(health_state).node); |
| 294 | state_unlock(ha); |
| 295 | } |