drivers/infiniband/core/counters.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2019 Mellanox Technologies. All rights reserved.
   4  */
   5 #include <rdma/ib_verbs.h>
   6 #include <rdma/rdma_counter.h>
   7
   8 #include "core_priv.h"
   9 #include "restrack.h"
  10
  11 #define ALL_AUTO_MODE_MASKS (RDMA_COUNTER_MASK_QP_TYPE)
  12
  13 static int __counter_set_mode(struct rdma_counter_mode *curr,
  14                               enum rdma_nl_counter_mode new_mode,
  15                               enum rdma_nl_counter_mask new_mask)
  16 {
  17         if ((new_mode == RDMA_COUNTER_MODE_AUTO) &&
  18             ((new_mask & (~ALL_AUTO_MODE_MASKS)) ||
  19              (curr->mode != RDMA_COUNTER_MODE_NONE)))
  20                 return -EINVAL;
  21
  22         curr->mode = new_mode;
  23         curr->mask = new_mask;
  24         return 0;
  25 }
  26
  27 /**
  28  * rdma_counter_set_auto_mode() - Turn on/off per-port auto mode
  29  *
  30  * When @on is true, the @mask must be set; When @on is false, it goes
  31  * into manual mode if there's any counter, so that the user is able to
  32  * manually access them.
  33  */
  34 int rdma_counter_set_auto_mode(struct ib_device *dev, u8 port,
  35                                bool on, enum rdma_nl_counter_mask mask)
  36 {
  37         struct rdma_port_counter *port_counter;
  38         int ret;
  39
  40         port_counter = &dev->port_data[port].port_counter;
  41         mutex_lock(&port_counter->lock);
  42         if (on) {
  43                 ret = __counter_set_mode(&port_counter->mode,
  44                                          RDMA_COUNTER_MODE_AUTO, mask);
  45         } else {
  46                 if (port_counter->mode.mode != RDMA_COUNTER_MODE_AUTO) {
  47                         ret = -EINVAL;
  48                         goto out;
  49                 }
  50
  51                 if (port_counter->num_counters)
  52                         ret = __counter_set_mode(&port_counter->mode,
  53                                                  RDMA_COUNTER_MODE_MANUAL, 0);
  54                 else
  55                         ret = __counter_set_mode(&port_counter->mode,
  56                                                  RDMA_COUNTER_MODE_NONE, 0);
  57         }
  58
  59 out:
  60         mutex_unlock(&port_counter->lock);
  61         return ret;
  62 }
  63
  64 static struct rdma_counter *rdma_counter_alloc(struct ib_device *dev, u8 port,
  65                                                enum rdma_nl_counter_mode mode)
  66 {
  67         struct rdma_port_counter *port_counter;
  68         struct rdma_counter *counter;
  69         int ret;
  70
  71         if (!dev->ops.counter_dealloc || !dev->ops.counter_alloc_stats)
  72                 return NULL;
  73
  74         counter = kzalloc(sizeof(*counter), GFP_KERNEL);
  75         if (!counter)
  76                 return NULL;
  77
  78         counter->device    = dev;
  79         counter->port      = port;
  80         counter->res.type  = RDMA_RESTRACK_COUNTER;
  81         counter->stats     = dev->ops.counter_alloc_stats(counter);
  82         if (!counter->stats)
  83                 goto err_stats;
  84
  85         port_counter = &dev->port_data[port].port_counter;
  86         mutex_lock(&port_counter->lock);
  87         if (mode == RDMA_COUNTER_MODE_MANUAL) {
  88                 ret = __counter_set_mode(&port_counter->mode,
  89                                          RDMA_COUNTER_MODE_MANUAL, 0);
  90                 if (ret)
  91                         goto err_mode;
  92         }
  93
  94         port_counter->num_counters++;
  95         mutex_unlock(&port_counter->lock);
  96
  97         counter->mode.mode = mode;
  98         kref_init(&counter->kref);
  99         mutex_init(&counter->lock);
 100
 101         return counter;
 102
 103 err_mode:
 104         mutex_unlock(&port_counter->lock);
 105         kfree(counter->stats);
 106 err_stats:
 107         kfree(counter);
 108         return NULL;
 109 }
 110
 111 static void rdma_counter_free(struct rdma_counter *counter)
 112 {
 113         struct rdma_port_counter *port_counter;
 114
 115         port_counter = &counter->device->port_data[counter->port].port_counter;
 116         mutex_lock(&port_counter->lock);
 117         port_counter->num_counters--;
 118         if (!port_counter->num_counters &&
 119             (port_counter->mode.mode == RDMA_COUNTER_MODE_MANUAL))
 120                 __counter_set_mode(&port_counter->mode, RDMA_COUNTER_MODE_NONE,
 121                                    0);
 122
 123         mutex_unlock(&port_counter->lock);
 124
 125         rdma_restrack_del(&counter->res);
 126         kfree(counter->stats);
 127         kfree(counter);
 128 }
 129
 130 static void auto_mode_init_counter(struct rdma_counter *counter,
 131                                    const struct ib_qp *qp,
 132                                    enum rdma_nl_counter_mask new_mask)
 133 {
 134         struct auto_mode_param *param = &counter->mode.param;
 135
 136         counter->mode.mode = RDMA_COUNTER_MODE_AUTO;
 137         counter->mode.mask = new_mask;
 138
 139         if (new_mask & RDMA_COUNTER_MASK_QP_TYPE)
 140                 param->qp_type = qp->qp_type;
 141 }
 142
 143 static bool auto_mode_match(struct ib_qp *qp, struct rdma_counter *counter,
 144                             enum rdma_nl_counter_mask auto_mask)
 145 {
 146         struct auto_mode_param *param = &counter->mode.param;
 147         bool match = true;
 148
 149         if (rdma_is_kernel_res(&counter->res) != rdma_is_kernel_res(&qp->res))
 150                 return false;
 151
 152         /* Ensure that counter belong to right PID */
 153         if (!rdma_is_kernel_res(&counter->res) &&
 154             !rdma_is_kernel_res(&qp->res) &&
 155             (task_pid_vnr(counter->res.task) != current->pid))
 156                 return false;
 157
 158         if (auto_mask & RDMA_COUNTER_MASK_QP_TYPE)
 159                 match &= (param->qp_type == qp->qp_type);
 160
 161         return match;
 162 }
 163
 164 static int __rdma_counter_bind_qp(struct rdma_counter *counter,
 165                                   struct ib_qp *qp)
 166 {
 167         int ret;
 168
 169         if (qp->counter)
 170                 return -EINVAL;
 171
 172         if (!qp->device->ops.counter_bind_qp)
 173                 return -EOPNOTSUPP;
 174
 175         mutex_lock(&counter->lock);
 176         ret = qp->device->ops.counter_bind_qp(counter, qp);
 177         mutex_unlock(&counter->lock);
 178
 179         return ret;
 180 }
 181
 182 static int __rdma_counter_unbind_qp(struct ib_qp *qp)
 183 {
 184         struct rdma_counter *counter = qp->counter;
 185         int ret;
 186
 187         if (!qp->device->ops.counter_unbind_qp)
 188                 return -EOPNOTSUPP;
 189
 190         mutex_lock(&counter->lock);
 191         ret = qp->device->ops.counter_unbind_qp(qp);
 192         mutex_unlock(&counter->lock);
 193
 194         return ret;
 195 }
 196
 197 static void counter_history_stat_update(const struct rdma_counter *counter)
 198 {
 199         struct ib_device *dev = counter->device;
 200         struct rdma_port_counter *port_counter;
 201         int i;
 202
 203         port_counter = &dev->port_data[counter->port].port_counter;
 204         if (!port_counter->hstats)
 205                 return;
 206
 207         for (i = 0; i < counter->stats->num_counters; i++)
 208                 port_counter->hstats->value[i] += counter->stats->value[i];
 209 }
 210
 211 /**
 212  * rdma_get_counter_auto_mode - Find the counter that @qp should be bound
 213  *     with in auto mode
 214  *
 215  * Return: The counter (with ref-count increased) if found
 216  */
 217 static struct rdma_counter *rdma_get_counter_auto_mode(struct ib_qp *qp,
 218                                                        u8 port)
 219 {
 220         struct rdma_port_counter *port_counter;
 221         struct rdma_counter *counter = NULL;
 222         struct ib_device *dev = qp->device;
 223         struct rdma_restrack_entry *res;
 224         struct rdma_restrack_root *rt;
 225         unsigned long id = 0;
 226
 227         port_counter = &dev->port_data[port].port_counter;
 228         rt = &dev->res[RDMA_RESTRACK_COUNTER];
 229         xa_lock(&rt->xa);
 230         xa_for_each(&rt->xa, id, res) {
 231                 if (!rdma_is_visible_in_pid_ns(res))
 232                         continue;
 233
 234                 counter = container_of(res, struct rdma_counter, res);
 235                 if ((counter->device != qp->device) || (counter->port != port))
 236                         goto next;
 237
 238                 if (auto_mode_match(qp, counter, port_counter->mode.mask))
 239                         break;
 240 next:
 241                 counter = NULL;
 242         }
 243
 244         if (counter && !kref_get_unless_zero(&counter->kref))
 245                 counter = NULL;
 246
 247         xa_unlock(&rt->xa);
 248         return counter;
 249 }
 250
 251 static void rdma_counter_res_add(struct rdma_counter *counter,
 252                                  struct ib_qp *qp)
 253 {
 254         if (rdma_is_kernel_res(&qp->res)) {
 255                 rdma_restrack_set_task(&counter->res, qp->res.kern_name);
 256                 rdma_restrack_kadd(&counter->res);
 257         } else {
 258                 rdma_restrack_attach_task(&counter->res, qp->res.task);
 259                 rdma_restrack_uadd(&counter->res);
 260         }
 261 }
 262
 263 static void counter_release(struct kref *kref)
 264 {
 265         struct rdma_counter *counter;
 266
 267         counter = container_of(kref, struct rdma_counter, kref);
 268         counter_history_stat_update(counter);
 269         counter->device->ops.counter_dealloc(counter);
 270         rdma_counter_free(counter);
 271 }
 272
 273 /**
 274  * rdma_counter_bind_qp_auto - Check and bind the QP to a counter base on
 275  *   the auto-mode rule
 276  */
 277 int rdma_counter_bind_qp_auto(struct ib_qp *qp, u8 port)
 278 {
 279         struct rdma_port_counter *port_counter;
 280         struct ib_device *dev = qp->device;
 281         struct rdma_counter *counter;
 282         int ret;
 283
 284         if (!rdma_is_port_valid(dev, port))
 285                 return -EINVAL;
 286
 287         port_counter = &dev->port_data[port].port_counter;
 288         if (port_counter->mode.mode != RDMA_COUNTER_MODE_AUTO)
 289                 return 0;
 290
 291         counter = rdma_get_counter_auto_mode(qp, port);
 292         if (counter) {
 293                 ret = __rdma_counter_bind_qp(counter, qp);
 294                 if (ret) {
 295                         kref_put(&counter->kref, counter_release);
 296                         return ret;
 297                 }
 298         } else {
 299                 counter = rdma_counter_alloc(dev, port, RDMA_COUNTER_MODE_AUTO);
 300                 if (!counter)
 301                         return -ENOMEM;
 302
 303                 auto_mode_init_counter(counter, qp, port_counter->mode.mask);
 304
 305                 ret = __rdma_counter_bind_qp(counter, qp);
 306                 if (ret) {
 307                         rdma_counter_free(counter);
 308                         return ret;
 309                 }
 310
 311                 rdma_counter_res_add(counter, qp);
 312         }
 313
 314         return 0;
 315 }
 316
 317 /**
 318  * rdma_counter_unbind_qp - Unbind a qp from a counter
 319  * @force:
 320  *   true - Decrease the counter ref-count anyway (e.g., qp destroy)
 321  */
 322 int rdma_counter_unbind_qp(struct ib_qp *qp, bool force)
 323 {
 324         struct rdma_counter *counter = qp->counter;
 325         int ret;
 326
 327         if (!counter)
 328                 return -EINVAL;
 329
 330         ret = __rdma_counter_unbind_qp(qp);
 331         if (ret && !force)
 332                 return ret;
 333
 334         kref_put(&counter->kref, counter_release);
 335         return 0;
 336 }
 337
 338 int rdma_counter_query_stats(struct rdma_counter *counter)
 339 {
 340         struct ib_device *dev = counter->device;
 341         int ret;
 342
 343         if (!dev->ops.counter_update_stats)
 344                 return -EINVAL;
 345
 346         mutex_lock(&counter->lock);
 347         ret = dev->ops.counter_update_stats(counter);
 348         mutex_unlock(&counter->lock);
 349
 350         return ret;
 351 }
 352
 353 static u64 get_running_counters_hwstat_sum(struct ib_device *dev,
 354                                            u8 port, u32 index)
 355 {
 356         struct rdma_restrack_entry *res;
 357         struct rdma_restrack_root *rt;
 358         struct rdma_counter *counter;
 359         unsigned long id = 0;
 360         u64 sum = 0;
 361
 362         rt = &dev->res[RDMA_RESTRACK_COUNTER];
 363         xa_lock(&rt->xa);
 364         xa_for_each(&rt->xa, id, res) {
 365                 if (!rdma_restrack_get(res))
 366                         continue;
 367
 368                 xa_unlock(&rt->xa);
 369
 370                 counter = container_of(res, struct rdma_counter, res);
 371                 if ((counter->device != dev) || (counter->port != port) ||
 372                     rdma_counter_query_stats(counter))
 373                         goto next;
 374
 375                 sum += counter->stats->value[index];
 376
 377 next:
 378                 xa_lock(&rt->xa);
 379                 rdma_restrack_put(res);
 380         }
 381
 382         xa_unlock(&rt->xa);
 383         return sum;
 384 }
 385
 386 /**
 387  * rdma_counter_get_hwstat_value() - Get the sum value of all counters on a
 388  *   specific port, including the running ones and history data
 389  */
 390 u64 rdma_counter_get_hwstat_value(struct ib_device *dev, u8 port, u32 index)
 391 {
 392         struct rdma_port_counter *port_counter;
 393         u64 sum;
 394
 395         port_counter = &dev->port_data[port].port_counter;
 396         if (!port_counter->hstats)
 397                 return 0;
 398
 399         sum = get_running_counters_hwstat_sum(dev, port, index);
 400         sum += port_counter->hstats->value[index];
 401
 402         return sum;
 403 }
 404
 405 static struct ib_qp *rdma_counter_get_qp(struct ib_device *dev, u32 qp_num)
 406 {
 407         struct rdma_restrack_entry *res = NULL;
 408         struct ib_qp *qp = NULL;
 409
 410         res = rdma_restrack_get_byid(dev, RDMA_RESTRACK_QP, qp_num);
 411         if (IS_ERR(res))
 412                 return NULL;
 413
 414         if (!rdma_is_visible_in_pid_ns(res))
 415                 goto err;
 416
 417         qp = container_of(res, struct ib_qp, res);
 418         if (qp->qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
 419                 goto err;
 420
 421         return qp;
 422
 423 err:
 424         rdma_restrack_put(&qp->res);
 425         return NULL;
 426 }
 427
 428 static int rdma_counter_bind_qp_manual(struct rdma_counter *counter,
 429                                        struct ib_qp *qp)
 430 {
 431         if ((counter->device != qp->device) || (counter->port != qp->port))
 432                 return -EINVAL;
 433
 434         return __rdma_counter_bind_qp(counter, qp);
 435 }
 436
 437 static struct rdma_counter *rdma_get_counter_by_id(struct ib_device *dev,
 438                                                    u32 counter_id)
 439 {
 440         struct rdma_restrack_entry *res;
 441         struct rdma_counter *counter;
 442
 443         res = rdma_restrack_get_byid(dev, RDMA_RESTRACK_COUNTER, counter_id);
 444         if (IS_ERR(res))
 445                 return NULL;
 446
 447         if (!rdma_is_visible_in_pid_ns(res)) {
 448                 rdma_restrack_put(res);
 449                 return NULL;
 450         }
 451
 452         counter = container_of(res, struct rdma_counter, res);
 453         kref_get(&counter->kref);
 454         rdma_restrack_put(res);
 455
 456         return counter;
 457 }
 458
 459 /**
 460  * rdma_counter_bind_qpn() - Bind QP @qp_num to counter @counter_id
 461  */
 462 int rdma_counter_bind_qpn(struct ib_device *dev, u8 port,
 463                           u32 qp_num, u32 counter_id)
 464 {
 465         struct rdma_counter *counter;
 466         struct ib_qp *qp;
 467         int ret;
 468
 469         qp = rdma_counter_get_qp(dev, qp_num);
 470         if (!qp)
 471                 return -ENOENT;
 472
 473         counter = rdma_get_counter_by_id(dev, counter_id);
 474         if (!counter) {
 475                 ret = -ENOENT;
 476                 goto err;
 477         }
 478
 479         if (counter->res.task != qp->res.task) {
 480                 ret = -EINVAL;
 481                 goto err_task;
 482         }
 483
 484         ret = rdma_counter_bind_qp_manual(counter, qp);
 485         if (ret)
 486                 goto err_task;
 487
 488         rdma_restrack_put(&qp->res);
 489         return 0;
 490
 491 err_task:
 492         kref_put(&counter->kref, counter_release);
 493 err:
 494         rdma_restrack_put(&qp->res);
 495         return ret;
 496 }
 497
 498 /**
 499  * rdma_counter_bind_qpn_alloc() - Alloc a counter and bind QP @qp_num to it
 500  *   The id of new counter is returned in @counter_id
 501  */
 502 int rdma_counter_bind_qpn_alloc(struct ib_device *dev, u8 port,
 503                                 u32 qp_num, u32 *counter_id)
 504 {
 505         struct rdma_counter *counter;
 506         struct ib_qp *qp;
 507         int ret;
 508
 509         if (!rdma_is_port_valid(dev, port))
 510                 return -EINVAL;
 511
 512         qp = rdma_counter_get_qp(dev, qp_num);
 513         if (!qp)
 514                 return -ENOENT;
 515
 516         if (rdma_is_port_valid(dev, qp->port) && (qp->port != port)) {
 517                 ret = -EINVAL;
 518                 goto err;
 519         }
 520
 521         counter = rdma_counter_alloc(dev, port, RDMA_COUNTER_MODE_MANUAL);
 522         if (!counter) {
 523                 ret = -ENOMEM;
 524                 goto err;
 525         }
 526
 527         ret = rdma_counter_bind_qp_manual(counter, qp);
 528         if (ret)
 529                 goto err_bind;
 530
 531         if (counter_id)
 532                 *counter_id = counter->id;
 533
 534         rdma_counter_res_add(counter, qp);
 535
 536         rdma_restrack_put(&qp->res);
 537         return ret;
 538
 539 err_bind:
 540         rdma_counter_free(counter);
 541 err:
 542         rdma_restrack_put(&qp->res);
 543         return ret;
 544 }
 545
 546 /**
 547  * rdma_counter_unbind_qpn() - Unbind QP @qp_num from a counter
 548  */
 549 int rdma_counter_unbind_qpn(struct ib_device *dev, u8 port,
 550                             u32 qp_num, u32 counter_id)
 551 {
 552         struct rdma_port_counter *port_counter;
 553         struct ib_qp *qp;
 554         int ret;
 555
 556         if (!rdma_is_port_valid(dev, port))
 557                 return -EINVAL;
 558
 559         qp = rdma_counter_get_qp(dev, qp_num);
 560         if (!qp)
 561                 return -ENOENT;
 562
 563         if (rdma_is_port_valid(dev, qp->port) && (qp->port != port)) {
 564                 ret = -EINVAL;
 565                 goto out;
 566         }
 567
 568         port_counter = &dev->port_data[port].port_counter;
 569         if (!qp->counter || qp->counter->id != counter_id ||
 570             port_counter->mode.mode != RDMA_COUNTER_MODE_MANUAL) {
 571                 ret = -EINVAL;
 572                 goto out;
 573         }
 574
 575         ret = rdma_counter_unbind_qp(qp, false);
 576
 577 out:
 578         rdma_restrack_put(&qp->res);
 579         return ret;
 580 }
 581
 582 int rdma_counter_get_mode(struct ib_device *dev, u8 port,
 583                           enum rdma_nl_counter_mode *mode,
 584                           enum rdma_nl_counter_mask *mask)
 585 {
 586         struct rdma_port_counter *port_counter;
 587
 588         port_counter = &dev->port_data[port].port_counter;
 589         *mode = port_counter->mode.mode;
 590         *mask = port_counter->mode.mask;
 591
 592         return 0;
 593 }
 594
 595 void rdma_counter_init(struct ib_device *dev)
 596 {
 597         struct rdma_port_counter *port_counter;
 598         u32 port;
 599
 600         if (!dev->port_data)
 601                 return;
 602
 603         rdma_for_each_port(dev, port) {
 604                 port_counter = &dev->port_data[port].port_counter;
 605                 port_counter->mode.mode = RDMA_COUNTER_MODE_NONE;
 606                 mutex_init(&port_counter->lock);
 607
 608                 if (!dev->ops.alloc_hw_stats)
 609                         continue;
 610
 611                 port_counter->hstats = dev->ops.alloc_hw_stats(dev, port);
 612                 if (!port_counter->hstats)
 613                         goto fail;
 614         }
 615
 616         return;
 617
 618 fail:
 619         rdma_for_each_port(dev, port) {
 620                 port_counter = &dev->port_data[port].port_counter;
 621                 kfree(port_counter->hstats);
 622                 port_counter->hstats = NULL;
 623         }
 624
 625         return;
 626 }
 627
 628 void rdma_counter_release(struct ib_device *dev)
 629 {
 630         struct rdma_port_counter *port_counter;
 631         u32 port;
 632
 633         rdma_for_each_port(dev, port) {
 634                 port_counter = &dev->port_data[port].port_counter;
 635                 kfree(port_counter->hstats);
 636         }
 637 }