VNFs/DPPD-PROX/handle_impair.c

   1 /*
   2 // Copyright (c) 2010-2017 Intel Corporation
   3 //
   4 // Licensed under the Apache License, Version 2.0 (the "License");
   5 // you may not use this file except in compliance with the License.
   6 // You may obtain a copy of the License at
   7 //
   8 //     http://www.apache.org/licenses/LICENSE-2.0
   9 //
  10 // Unless required by applicable law or agreed to in writing, software
  11 // distributed under the License is distributed on an "AS IS" BASIS,
  12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13 // See the License for the specific language governing permissions and
  14 // limitations under the License.
  15 */
  16
  17 #include <string.h>
  18 #include <stdio.h>
  19 #include <rte_cycles.h>
  20 #include <rte_version.h>
  21
  22 #include "prox_malloc.h"
  23 #include "lconf.h"
  24 #include "log.h"
  25 #include "random.h"
  26 #include "handle_impair.h"
  27 #include "prefetch.h"
  28 #include "prox_port_cfg.h"
  29
  30 #if RTE_VERSION < RTE_VERSION_NUM(1,8,0,0)
  31 #define RTE_CACHE_LINE_SIZE CACHE_LINE_SIZE
  32 #endif
  33
  34 #define DELAY_ACCURACY  11              // accuracy of 2048 cycles ~= 1 micro-second
  35 #define DELAY_MAX_MASK  0x1FFFFF        // Maximum 2M * 2K cycles ~1 second
  36
  37 struct queue_elem {
  38         struct rte_mbuf *mbuf;
  39         uint64_t        tsc;
  40 };
  41
  42 struct queue {
  43         struct queue_elem *queue_elem;
  44         unsigned queue_head;
  45         unsigned queue_tail;
  46 };
  47
  48 struct task_impair {
  49         struct task_base base;
  50         struct queue_elem *queue;
  51         uint32_t random_delay_us;
  52         uint32_t delay_us;
  53         uint64_t delay_time;
  54         uint64_t delay_time_mask;
  55         unsigned queue_head;
  56         unsigned queue_tail;
  57         unsigned queue_mask;
  58         int tresh;
  59         unsigned int seed;
  60         struct random state;
  61         uint64_t last_idx;
  62         struct queue *buffer;
  63         uint32_t socket_id;
  64         uint32_t flags;
  65         uint8_t src_mac[6];
  66 };
  67
  68 #define IMPAIR_NEED_UPDATE     1
  69 #define IMPAIR_SET_MAC         2
  70
  71 static int handle_bulk_impair(struct task_base *tbase, struct rte_mbuf **mbufs, uint16_t n_pkts);
  72 static int handle_bulk_impair_random(struct task_base *tbase, struct rte_mbuf **mbufs, uint16_t n_pkts);
  73 static int handle_bulk_random_drop(struct task_base *tbase, struct rte_mbuf **mbufs, uint16_t n_pkts);
  74
  75 void task_impair_set_proba(struct task_base *tbase, float proba)
  76 {
  77         struct task_impair *task = (struct task_impair *)tbase;
  78         task->tresh = ((uint64_t) RAND_MAX) * (uint32_t)(proba * 10000) / 1000000;
  79 }
  80
  81 void task_impair_set_delay_us(struct task_base *tbase, uint32_t delay_us, uint32_t random_delay_us)
  82 {
  83         struct task_impair *task = (struct task_impair *)tbase;
  84         task->flags |= IMPAIR_NEED_UPDATE;
  85         task->random_delay_us = random_delay_us;
  86         task->delay_us = delay_us;
  87 }
  88
  89 static void task_impair_update(struct task_base *tbase)
  90 {
  91         struct task_impair *task = (struct task_impair *)tbase;
  92         uint32_t queue_len = 0;
  93         size_t mem_size;
  94         if ((task->flags & IMPAIR_NEED_UPDATE) == 0)
  95                 return;
  96         task->flags &= ~IMPAIR_NEED_UPDATE;
  97         uint64_t now = rte_rdtsc();
  98         uint8_t out[MAX_PKT_BURST] = {0};
  99         uint64_t now_idx = (now >> DELAY_ACCURACY) & DELAY_MAX_MASK;
 100
 101         if (task->random_delay_us) {
 102                 tbase->handle_bulk = handle_bulk_impair_random;
 103                 task->delay_time = usec_to_tsc(task->random_delay_us);
 104                 task->delay_time_mask = rte_align32pow2(task->delay_time) - 1;
 105                 queue_len = rte_align32pow2((1250L * task->random_delay_us) / 84 / (DELAY_MAX_MASK + 1));
 106         } else if (task->delay_us == 0) {
 107                 tbase->handle_bulk = handle_bulk_random_drop;
 108                 task->delay_time = 0;
 109         } else {
 110                 tbase->handle_bulk = handle_bulk_impair;
 111                 task->delay_time = usec_to_tsc(task->delay_us);
 112                 queue_len = rte_align32pow2(1250 * task->delay_us / 84);
 113         }
 114         if (task->queue) {
 115                 struct rte_mbuf *new_mbufs[MAX_PKT_BURST];
 116                 while (task->queue_tail != task->queue_head) {
 117                         now = rte_rdtsc();
 118                         uint16_t idx = 0;
 119                         while (idx < MAX_PKT_BURST && task->queue_tail != task->queue_head) {
 120                                 if (task->queue[task->queue_tail].tsc <= now) {
 121                                         out[idx] = rand_r(&task->seed) <= task->tresh? 0 : OUT_DISCARD;
 122                                         new_mbufs[idx++] = task->queue[task->queue_tail].mbuf;
 123                                         task->queue_tail = (task->queue_tail + 1) & task->queue_mask;
 124                                 }
 125                                 else {
 126                                         break;
 127                                 }
 128                         }
 129                         if (idx)
 130                                 task->base.tx_pkt(&task->base, new_mbufs, idx, out);
 131                 }
 132                 prox_free(task->queue);
 133                 task->queue = NULL;
 134         }
 135         if (task->buffer) {
 136                 struct rte_mbuf *new_mbufs[MAX_PKT_BURST];
 137                 while (task->last_idx != ((now_idx - 1) & DELAY_MAX_MASK)) {
 138                         now = rte_rdtsc();
 139                         uint16_t pkt_idx = 0;
 140                         while ((pkt_idx < MAX_PKT_BURST) && (task->last_idx != ((now_idx - 1) & DELAY_MAX_MASK))) {
 141                                 struct queue *queue = &task->buffer[task->last_idx];
 142                                 while ((pkt_idx < MAX_PKT_BURST) && (queue->queue_tail != queue->queue_head)) {
 143                                         out[pkt_idx] = rand_r(&task->seed) <= task->tresh? 0 : OUT_DISCARD;
 144                                         new_mbufs[pkt_idx++] = queue->queue_elem[queue->queue_tail].mbuf;
 145                                         queue->queue_tail = (queue->queue_tail + 1) & task->queue_mask;
 146                                 }
 147                                 task->last_idx = (task->last_idx + 1) & DELAY_MAX_MASK;
 148                         }
 149
 150                         if (pkt_idx)
 151                                 task->base.tx_pkt(&task->base, new_mbufs, pkt_idx, out);
 152                 }
 153                 for (int i = 0; i < DELAY_MAX_MASK + 1; i++) {
 154                         if (task->buffer[i].queue_elem)
 155                                 prox_free(task->buffer[i].queue_elem);
 156                 }
 157                 prox_free(task->buffer);
 158                 task->buffer = NULL;
 159         }
 160
 161         if (queue_len < MAX_PKT_BURST)
 162                 queue_len= MAX_PKT_BURST;
 163         task->queue_mask = queue_len - 1;
 164         if (task->queue_mask < MAX_PKT_BURST - 1)
 165                 task->queue_mask = MAX_PKT_BURST - 1;
 166         mem_size = (task->queue_mask + 1) * sizeof(task->queue[0]);
 167
 168         if (task->delay_us) {
 169                 task->queue_head = 0;
 170                 task->queue_tail = 0;
 171                 task->queue = prox_zmalloc(mem_size, task->socket_id);
 172                 if (task->queue == NULL) {
 173                         plog_err("Not enough memory to allocate queue\n");
 174                         task->queue_mask = 0;
 175                 }
 176         } else if (task->random_delay_us) {
 177                 size_t size = (DELAY_MAX_MASK + 1) * sizeof(struct queue);
 178                 plog_info("Allocating %zd bytes\n", size);
 179                 task->buffer = prox_zmalloc(size, task->socket_id);
 180                 PROX_PANIC(task->buffer == NULL, "Not enough memory to allocate buffer\n");
 181                 plog_info("Allocating %d x %zd bytes\n", DELAY_MAX_MASK + 1, mem_size);
 182
 183                 for (int i = 0; i < DELAY_MAX_MASK + 1; i++) {
 184                         task->buffer[i].queue_elem = prox_zmalloc(mem_size, task->socket_id);
 185                         PROX_PANIC(task->buffer[i].queue_elem == NULL, "Not enough memory to allocate buffer elems\n");
 186                 }
 187         }
 188         random_init_seed(&task->state);
 189 }
 190
 191 static int handle_bulk_random_drop(struct task_base *tbase, struct rte_mbuf **mbufs, uint16_t n_pkts)
 192 {
 193         struct task_impair *task = (struct task_impair *)tbase;
 194         uint8_t out[MAX_PKT_BURST];
 195         struct ether_hdr * hdr[MAX_PKT_BURST];
 196         int ret = 0;
 197         for (uint16_t i = 0; i < n_pkts; ++i) {
 198                 PREFETCH0(mbufs[i]);
 199         }
 200         for (uint16_t i = 0; i < n_pkts; ++i) {
 201                 hdr[i] = rte_pktmbuf_mtod(mbufs[i], struct ether_hdr *);
 202                 PREFETCH0(hdr[i]);
 203         }
 204         if (task->flags & IMPAIR_SET_MAC) {
 205                 for (uint16_t i = 0; i < n_pkts; ++i) {
 206                         ether_addr_copy((struct ether_addr *)&task->src_mac[0], &hdr[i]->s_addr);
 207                         out[i] = rand_r(&task->seed) <= task->tresh? 0 : OUT_DISCARD;
 208                 }
 209         } else {
 210                 for (uint16_t i = 0; i < n_pkts; ++i) {
 211                         out[i] = rand_r(&task->seed) <= task->tresh? 0 : OUT_DISCARD;
 212                 }
 213         }
 214         ret = task->base.tx_pkt(&task->base, mbufs, n_pkts, out);
 215         task_impair_update(tbase);
 216         return ret;
 217 }
 218
 219 static int handle_bulk_impair(struct task_base *tbase, struct rte_mbuf **mbufs, uint16_t n_pkts)
 220 {
 221         struct task_impair *task = (struct task_impair *)tbase;
 222         uint64_t now = rte_rdtsc();
 223         uint8_t out[MAX_PKT_BURST] = {0};
 224         uint16_t enqueue_failed;
 225         uint16_t i;
 226         int ret = 0;
 227         struct ether_hdr * hdr[MAX_PKT_BURST];
 228         for (uint16_t i = 0; i < n_pkts; ++i) {
 229                 PREFETCH0(mbufs[i]);
 230         }
 231         for (uint16_t i = 0; i < n_pkts; ++i) {
 232                 hdr[i] = rte_pktmbuf_mtod(mbufs[i], struct ether_hdr *);
 233                 PREFETCH0(hdr[i]);
 234         }
 235
 236         int nb_empty_slots = (task->queue_tail - task->queue_head + task->queue_mask) & task->queue_mask;
 237         if (likely(nb_empty_slots >= n_pkts)) {
 238                 /* We know n_pkts fits, no need to check for every packet */
 239                 for (i = 0; i < n_pkts; ++i) {
 240                         if (task->flags & IMPAIR_SET_MAC)
 241                                 ether_addr_copy((struct ether_addr *)&task->src_mac[0], &hdr[i]->s_addr);
 242                         task->queue[task->queue_head].tsc = now + task->delay_time;
 243                         task->queue[task->queue_head].mbuf = mbufs[i];
 244                         task->queue_head = (task->queue_head + 1) & task->queue_mask;
 245                 }
 246         } else {
 247                 for (i = 0; i < n_pkts; ++i) {
 248                         if (((task->queue_head + 1) & task->queue_mask) != task->queue_tail) {
 249                                 if (task->flags & IMPAIR_SET_MAC)
 250                                         ether_addr_copy((struct ether_addr *)&task->src_mac[0], &hdr[i]->s_addr);
 251                                 task->queue[task->queue_head].tsc = now + task->delay_time;
 252                                 task->queue[task->queue_head].mbuf = mbufs[i];
 253                                 task->queue_head = (task->queue_head + 1) & task->queue_mask;
 254                         }
 255                         else {
 256                                 /* Rest does not fit, need to drop those packets. */
 257                                 enqueue_failed = i;
 258                                 for (;i < n_pkts; ++i) {
 259                                         out[i] = OUT_DISCARD;
 260                                 }
 261                                 ret+= task->base.tx_pkt(&task->base, mbufs + enqueue_failed,
 262                                                 n_pkts - enqueue_failed, out + enqueue_failed);
 263                                 break;
 264                         }
 265                 }
 266         }
 267
 268         struct rte_mbuf *new_mbufs[MAX_PKT_BURST];
 269         uint16_t idx = 0;
 270
 271         if (task->tresh != RAND_MAX) {
 272                 while (idx < MAX_PKT_BURST && task->queue_tail != task->queue_head) {
 273                         if (task->queue[task->queue_tail].tsc <= now) {
 274                                 out[idx] = rand_r(&task->seed) <= task->tresh? 0 : OUT_DISCARD;
 275                                 new_mbufs[idx] = task->queue[task->queue_tail].mbuf;
 276                                 PREFETCH0(new_mbufs[idx]);
 277                                 PREFETCH0(&new_mbufs[idx]->cacheline1);
 278                                 idx++;
 279                                 task->queue_tail = (task->queue_tail + 1) & task->queue_mask;
 280                         }
 281                         else {
 282                                 break;
 283                         }
 284                 }
 285         } else {
 286                 while (idx < MAX_PKT_BURST && task->queue_tail != task->queue_head) {
 287                         if (task->queue[task->queue_tail].tsc <= now) {
 288                                 out[idx] = 0;
 289                                 new_mbufs[idx] = task->queue[task->queue_tail].mbuf;
 290                                 PREFETCH0(new_mbufs[idx]);
 291                                 PREFETCH0(&new_mbufs[idx]->cacheline1);
 292                                 idx++;
 293                                 task->queue_tail = (task->queue_tail + 1) & task->queue_mask;
 294                         }
 295                         else {
 296                                 break;
 297                         }
 298                 }
 299         }
 300
 301         if (idx)
 302                 ret+= task->base.tx_pkt(&task->base, new_mbufs, idx, out);
 303         task_impair_update(tbase);
 304         return ret;
 305 }
 306
 307 /*
 308  * We want to avoid using division and mod for performance reasons.
 309  * We also want to support up to one second delay, and express it in tsc
 310  * So the delay in tsc needs up to 32 bits (supposing procesor freq is less than 4GHz).
 311  * If the max_delay is smaller, we make sure we use less bits.
 312  * Note that we lose the MSB of the xorshift - 64 bits could hold
 313  * two or three delays in TSC - but would probably make implementation more complex
 314  * and not huge gain expected. Maybe room for optimization.
 315  * Using this implementation, we might have to run random more than once for a delay
 316  * but in average this should occur less than 50% of the time.
 317 */
 318
 319 static inline uint64_t random_delay(struct random *state, uint64_t max_delay, uint64_t max_delay_mask)
 320 {
 321         uint64_t val;
 322         while(1) {
 323                 val = random_next(state);
 324                 if ((val & max_delay_mask) < max_delay)
 325                         return (val & max_delay_mask);
 326         }
 327 }
 328
 329 static int handle_bulk_impair_random(struct task_base *tbase, struct rte_mbuf **mbufs, uint16_t n_pkts)
 330 {
 331         struct task_impair *task = (struct task_impair *)tbase;
 332         uint64_t now = rte_rdtsc();
 333         uint8_t out[MAX_PKT_BURST];
 334         uint16_t enqueue_failed;
 335         uint16_t i;
 336         int ret = 0;
 337         uint64_t packet_time, idx;
 338         uint64_t now_idx = (now >> DELAY_ACCURACY) & DELAY_MAX_MASK;
 339         struct ether_hdr * hdr[MAX_PKT_BURST];
 340         for (uint16_t i = 0; i < n_pkts; ++i) {
 341                 PREFETCH0(mbufs[i]);
 342         }
 343         for (uint16_t i = 0; i < n_pkts; ++i) {
 344                 hdr[i] = rte_pktmbuf_mtod(mbufs[i], struct ether_hdr *);
 345                 PREFETCH0(hdr[i]);
 346         }
 347
 348         for (i = 0; i < n_pkts; ++i) {
 349                 packet_time = now + random_delay(&task->state, task->delay_time, task->delay_time_mask);
 350                 idx = (packet_time >> DELAY_ACCURACY) & DELAY_MAX_MASK;
 351                 while (idx != ((now_idx - 1) & DELAY_MAX_MASK)) {
 352                         struct queue *queue = &task->buffer[idx];
 353                         if (((queue->queue_head + 1) & task->queue_mask) != queue->queue_tail) {
 354                                 if (task->flags & IMPAIR_SET_MAC)
 355                                         ether_addr_copy((struct ether_addr *)&task->src_mac[0], &hdr[i]->s_addr);
 356                                 queue->queue_elem[queue->queue_head].mbuf = mbufs[i];
 357                                 queue->queue_head = (queue->queue_head + 1) & task->queue_mask;
 358                                 break;
 359                         } else {
 360                                 idx = (idx + 1) & DELAY_MAX_MASK;
 361                         }
 362                 }
 363                 if (idx == ((now_idx - 1) & DELAY_MAX_MASK)) {
 364                         /* Rest does not fit, need to drop packet. Note that further packets might fit as might want to be sent earlier */
 365                         out[0] = OUT_DISCARD;
 366                         ret+= task->base.tx_pkt(&task->base, mbufs + i, 1, out);
 367                         plog_warn("Unexpectdly dropping packets\n");
 368                 }
 369         }
 370
 371         struct rte_mbuf *new_mbufs[MAX_PKT_BURST];
 372         uint16_t pkt_idx = 0;
 373
 374         while ((pkt_idx < MAX_PKT_BURST) && (task->last_idx != ((now_idx - 1) & DELAY_MAX_MASK))) {
 375                 struct queue *queue = &task->buffer[task->last_idx];
 376                 while ((pkt_idx < MAX_PKT_BURST) && (queue->queue_tail != queue->queue_head)) {
 377                         out[pkt_idx] = rand_r(&task->seed) <= task->tresh? 0 : OUT_DISCARD;
 378                         new_mbufs[pkt_idx] = queue->queue_elem[queue->queue_tail].mbuf;
 379                         PREFETCH0(new_mbufs[pkt_idx]);
 380                         PREFETCH0(&new_mbufs[pkt_idx]->cacheline1);
 381                         pkt_idx++;
 382                         queue->queue_tail = (queue->queue_tail + 1) & task->queue_mask;
 383                 }
 384                 task->last_idx = (task->last_idx + 1) & DELAY_MAX_MASK;
 385         }
 386
 387         if (pkt_idx)
 388                 ret+= task->base.tx_pkt(&task->base, new_mbufs, pkt_idx, out);
 389         task_impair_update(tbase);
 390         return ret;
 391 }
 392
 393 static void init_task(struct task_base *tbase, struct task_args *targ)
 394 {
 395         struct task_impair *task = (struct task_impair *)tbase;
 396         uint32_t queue_len = 0;
 397         size_t mem_size;
 398         unsigned socket_id;
 399         uint64_t delay_us = 0;
 400
 401         task->seed = rte_rdtsc();
 402         if (targ->probability == 0)
 403                 targ->probability = 1000000;
 404
 405         task->tresh = ((uint64_t) RAND_MAX) * targ->probability / 1000000;
 406
 407         if ((targ->delay_us == 0) && (targ->random_delay_us == 0)) {
 408                 tbase->handle_bulk = handle_bulk_random_drop;
 409                 task->delay_time = 0;
 410         } else if (targ->random_delay_us) {
 411                 tbase->handle_bulk = handle_bulk_impair_random;
 412                 task->delay_time = usec_to_tsc(targ->random_delay_us);
 413                 task->delay_time_mask = rte_align32pow2(task->delay_time) - 1;
 414                 delay_us = targ->random_delay_us;
 415                 queue_len = rte_align32pow2((1250L * delay_us) / 84 / (DELAY_MAX_MASK + 1));
 416         } else {
 417                 task->delay_time = usec_to_tsc(targ->delay_us);
 418                 delay_us = targ->delay_us;
 419                 queue_len = rte_align32pow2(1250 * delay_us / 84);
 420         }
 421         /* Assume Line-rate is maximum transmit speed.
 422            TODO: take link speed if tx is port.
 423         */
 424         if (queue_len < MAX_PKT_BURST)
 425                 queue_len= MAX_PKT_BURST;
 426         task->queue_mask = queue_len - 1;
 427         if (task->queue_mask < MAX_PKT_BURST - 1)
 428                 task->queue_mask = MAX_PKT_BURST - 1;
 429
 430         mem_size = (task->queue_mask + 1) * sizeof(task->queue[0]);
 431         socket_id = rte_lcore_to_socket_id(targ->lconf->id);
 432         task->socket_id = rte_lcore_to_socket_id(targ->lconf->id);
 433
 434         if (targ->delay_us) {
 435                 task->queue = prox_zmalloc(mem_size, socket_id);
 436                 PROX_PANIC(task->queue == NULL, "Not enough memory to allocate queue\n");
 437                 task->queue_head = 0;
 438                 task->queue_tail = 0;
 439         } else if (targ->random_delay_us) {
 440                 size_t size = (DELAY_MAX_MASK + 1) * sizeof(struct queue);
 441                 plog_info("Allocating %zd bytes\n", size);
 442                 task->buffer = prox_zmalloc(size, socket_id);
 443                 PROX_PANIC(task->buffer == NULL, "Not enough memory to allocate buffer\n");
 444                 plog_info("Allocating %d x %zd bytes\n", DELAY_MAX_MASK + 1, mem_size);
 445
 446                 for (int i = 0; i < DELAY_MAX_MASK + 1; i++) {
 447                         task->buffer[i].queue_elem = prox_zmalloc(mem_size, socket_id);
 448                         PROX_PANIC(task->buffer[i].queue_elem == NULL, "Not enough memory to allocate buffer elems\n");
 449                 }
 450         }
 451         random_init_seed(&task->state);
 452         if (targ->nb_txports) {
 453                 memcpy(&task->src_mac[0], &prox_port_cfg[tbase->tx_params_hw.tx_port_queue[0].port].eth_addr, sizeof(struct ether_addr));
 454                 task->flags = IMPAIR_SET_MAC;
 455         } else {
 456                 task->flags = 0;
 457         }
 458 }
 459
 460 static struct task_init tinit = {
 461         .mode_str = "impair",
 462         .init = init_task,
 463         .handle = handle_bulk_impair,
 464         .flag_features = TASK_FEATURE_TXQ_FLAGS_NOOFFLOADS | TASK_FEATURE_ZERO_RX,
 465         .size = sizeof(struct task_impair)
 466 };
 467
 468 __attribute__((constructor)) static void ctor(void)
 469 {
 470         reg_task(&tinit);
 471 }