drivers/bus/vdev/vdev.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2016 RehiveTech. All rights reserved.
   3  */
   4
   5 #include <string.h>
   6 #include <inttypes.h>
   7 #include <stdio.h>
   8 #include <stdlib.h>
   9 #include <stdint.h>
  10 #include <stdbool.h>
  11 #include <sys/queue.h>
  12
  13 #include <rte_eal.h>
  14 #include <rte_dev.h>
  15 #include <rte_bus.h>
  16 #include <rte_common.h>
  17 #include <rte_devargs.h>
  18 #include <rte_memory.h>
  19 #include <rte_tailq.h>
  20 #include <rte_spinlock.h>
  21 #include <rte_string_fns.h>
  22 #include <rte_errno.h>
  23
  24 #include "rte_bus_vdev.h"
  25 #include "vdev_logs.h"
  26 #include "vdev_private.h"
  27
  28 #define VDEV_MP_KEY     "bus_vdev_mp"
  29
  30 /* Forward declare to access virtual bus name */
  31 static struct rte_bus rte_vdev_bus;
  32
  33 /** Double linked list of virtual device drivers. */
  34 TAILQ_HEAD(vdev_device_list, rte_vdev_device);
  35
  36 static struct vdev_device_list vdev_device_list =
  37         TAILQ_HEAD_INITIALIZER(vdev_device_list);
  38 /* The lock needs to be recursive because a vdev can manage another vdev. */
  39 static rte_spinlock_recursive_t vdev_device_list_lock =
  40         RTE_SPINLOCK_RECURSIVE_INITIALIZER;
  41
  42 static struct vdev_driver_list vdev_driver_list =
  43         TAILQ_HEAD_INITIALIZER(vdev_driver_list);
  44
  45 struct vdev_custom_scan {
  46         TAILQ_ENTRY(vdev_custom_scan) next;
  47         rte_vdev_scan_callback callback;
  48         void *user_arg;
  49 };
  50 TAILQ_HEAD(vdev_custom_scans, vdev_custom_scan);
  51 static struct vdev_custom_scans vdev_custom_scans =
  52         TAILQ_HEAD_INITIALIZER(vdev_custom_scans);
  53 static rte_spinlock_t vdev_custom_scan_lock = RTE_SPINLOCK_INITIALIZER;
  54
  55 /* register a driver */
  56 void
  57 rte_vdev_register(struct rte_vdev_driver *driver)
  58 {
  59         TAILQ_INSERT_TAIL(&vdev_driver_list, driver, next);
  60 }
  61
  62 /* unregister a driver */
  63 void
  64 rte_vdev_unregister(struct rte_vdev_driver *driver)
  65 {
  66         TAILQ_REMOVE(&vdev_driver_list, driver, next);
  67 }
  68
  69 int
  70 rte_vdev_add_custom_scan(rte_vdev_scan_callback callback, void *user_arg)
  71 {
  72         struct vdev_custom_scan *custom_scan;
  73
  74         rte_spinlock_lock(&vdev_custom_scan_lock);
  75
  76         /* check if already registered */
  77         TAILQ_FOREACH(custom_scan, &vdev_custom_scans, next) {
  78                 if (custom_scan->callback == callback &&
  79                                 custom_scan->user_arg == user_arg)
  80                         break;
  81         }
  82
  83         if (custom_scan == NULL) {
  84                 custom_scan = malloc(sizeof(struct vdev_custom_scan));
  85                 if (custom_scan != NULL) {
  86                         custom_scan->callback = callback;
  87                         custom_scan->user_arg = user_arg;
  88                         TAILQ_INSERT_TAIL(&vdev_custom_scans, custom_scan, next);
  89                 }
  90         }
  91
  92         rte_spinlock_unlock(&vdev_custom_scan_lock);
  93
  94         return (custom_scan == NULL) ? -1 : 0;
  95 }
  96
  97 int
  98 rte_vdev_remove_custom_scan(rte_vdev_scan_callback callback, void *user_arg)
  99 {
 100         struct vdev_custom_scan *custom_scan, *tmp_scan;
 101
 102         rte_spinlock_lock(&vdev_custom_scan_lock);
 103         TAILQ_FOREACH_SAFE(custom_scan, &vdev_custom_scans, next, tmp_scan) {
 104                 if (custom_scan->callback != callback ||
 105                                 (custom_scan->user_arg != (void *)-1 &&
 106                                 custom_scan->user_arg != user_arg))
 107                         continue;
 108                 TAILQ_REMOVE(&vdev_custom_scans, custom_scan, next);
 109                 free(custom_scan);
 110         }
 111         rte_spinlock_unlock(&vdev_custom_scan_lock);
 112
 113         return 0;
 114 }
 115
 116 static int
 117 vdev_parse(const char *name, void *addr)
 118 {
 119         struct rte_vdev_driver **out = addr;
 120         struct rte_vdev_driver *driver = NULL;
 121
 122         TAILQ_FOREACH(driver, &vdev_driver_list, next) {
 123                 if (strncmp(driver->driver.name, name,
 124                             strlen(driver->driver.name)) == 0)
 125                         break;
 126                 if (driver->driver.alias &&
 127                     strncmp(driver->driver.alias, name,
 128                             strlen(driver->driver.alias)) == 0)
 129                         break;
 130         }
 131         if (driver != NULL &&
 132             addr != NULL)
 133                 *out = driver;
 134         return driver == NULL;
 135 }
 136
 137 static int
 138 vdev_probe_all_drivers(struct rte_vdev_device *dev)
 139 {
 140         const char *name;
 141         struct rte_vdev_driver *driver;
 142         int ret;
 143
 144         if (rte_dev_is_probed(&dev->device))
 145                 return -EEXIST;
 146
 147         name = rte_vdev_device_name(dev);
 148         VDEV_LOG(DEBUG, "Search driver to probe device %s", name);
 149
 150         if (vdev_parse(name, &driver))
 151                 return -1;
 152         ret = driver->probe(dev);
 153         if (ret == 0)
 154                 dev->device.driver = &driver->driver;
 155         return ret;
 156 }
 157
 158 /* The caller shall be responsible for thread-safe */
 159 static struct rte_vdev_device *
 160 find_vdev(const char *name)
 161 {
 162         struct rte_vdev_device *dev;
 163
 164         if (!name)
 165                 return NULL;
 166
 167         TAILQ_FOREACH(dev, &vdev_device_list, next) {
 168                 const char *devname = rte_vdev_device_name(dev);
 169
 170                 if (!strcmp(devname, name))
 171                         return dev;
 172         }
 173
 174         return NULL;
 175 }
 176
 177 static struct rte_devargs *
 178 alloc_devargs(const char *name, const char *args)
 179 {
 180         struct rte_devargs *devargs;
 181         int ret;
 182
 183         devargs = calloc(1, sizeof(*devargs));
 184         if (!devargs)
 185                 return NULL;
 186
 187         devargs->bus = &rte_vdev_bus;
 188         if (args)
 189                 devargs->args = strdup(args);
 190         else
 191                 devargs->args = strdup("");
 192
 193         ret = strlcpy(devargs->name, name, sizeof(devargs->name));
 194         if (ret < 0 || ret >= (int)sizeof(devargs->name)) {
 195                 free(devargs->args);
 196                 free(devargs);
 197                 return NULL;
 198         }
 199
 200         return devargs;
 201 }
 202
 203 static int
 204 insert_vdev(const char *name, const char *args,
 205                 struct rte_vdev_device **p_dev,
 206                 bool init)
 207 {
 208         struct rte_vdev_device *dev;
 209         struct rte_devargs *devargs;
 210         int ret;
 211
 212         if (name == NULL)
 213                 return -EINVAL;
 214
 215         devargs = alloc_devargs(name, args);
 216         if (!devargs)
 217                 return -ENOMEM;
 218
 219         dev = calloc(1, sizeof(*dev));
 220         if (!dev) {
 221                 ret = -ENOMEM;
 222                 goto fail;
 223         }
 224
 225         dev->device.bus = &rte_vdev_bus;
 226         dev->device.numa_node = SOCKET_ID_ANY;
 227         dev->device.name = devargs->name;
 228
 229         if (find_vdev(name)) {
 230                 /*
 231                  * A vdev is expected to have only one port.
 232                  * So there is no reason to try probing again,
 233                  * even with new arguments.
 234                  */
 235                 ret = -EEXIST;
 236                 goto fail;
 237         }
 238
 239         if (init)
 240                 rte_devargs_insert(&devargs);
 241         dev->device.devargs = devargs;
 242         TAILQ_INSERT_TAIL(&vdev_device_list, dev, next);
 243
 244         if (p_dev)
 245                 *p_dev = dev;
 246
 247         return 0;
 248 fail:
 249         free(devargs->args);
 250         free(devargs);
 251         free(dev);
 252         return ret;
 253 }
 254
 255 int
 256 rte_vdev_init(const char *name, const char *args)
 257 {
 258         struct rte_vdev_device *dev;
 259         int ret;
 260
 261         rte_spinlock_recursive_lock(&vdev_device_list_lock);
 262         ret = insert_vdev(name, args, &dev, true);
 263         if (ret == 0) {
 264                 ret = vdev_probe_all_drivers(dev);
 265                 if (ret) {
 266                         if (ret > 0)
 267                                 VDEV_LOG(ERR, "no driver found for %s", name);
 268                         /* If fails, remove it from vdev list */
 269                         TAILQ_REMOVE(&vdev_device_list, dev, next);
 270                         rte_devargs_remove(dev->device.devargs);
 271                         free(dev);
 272                 }
 273         }
 274         rte_spinlock_recursive_unlock(&vdev_device_list_lock);
 275         return ret;
 276 }
 277
 278 static int
 279 vdev_remove_driver(struct rte_vdev_device *dev)
 280 {
 281         const char *name = rte_vdev_device_name(dev);
 282         const struct rte_vdev_driver *driver;
 283
 284         if (!dev->device.driver) {
 285                 VDEV_LOG(DEBUG, "no driver attach to device %s", name);
 286                 return 1;
 287         }
 288
 289         driver = container_of(dev->device.driver, const struct rte_vdev_driver,
 290                 driver);
 291         return driver->remove(dev);
 292 }
 293
 294 int
 295 rte_vdev_uninit(const char *name)
 296 {
 297         struct rte_vdev_device *dev;
 298         int ret;
 299
 300         if (name == NULL)
 301                 return -EINVAL;
 302
 303         rte_spinlock_recursive_lock(&vdev_device_list_lock);
 304
 305         dev = find_vdev(name);
 306         if (!dev) {
 307                 ret = -ENOENT;
 308                 goto unlock;
 309         }
 310
 311         ret = vdev_remove_driver(dev);
 312         if (ret)
 313                 goto unlock;
 314
 315         TAILQ_REMOVE(&vdev_device_list, dev, next);
 316         rte_devargs_remove(dev->device.devargs);
 317         free(dev);
 318
 319 unlock:
 320         rte_spinlock_recursive_unlock(&vdev_device_list_lock);
 321         return ret;
 322 }
 323
 324 struct vdev_param {
 325 #define VDEV_SCAN_REQ   1
 326 #define VDEV_SCAN_ONE   2
 327 #define VDEV_SCAN_REP   3
 328         int type;
 329         int num;
 330         char name[RTE_DEV_NAME_MAX_LEN];
 331 };
 332
 333 static int vdev_plug(struct rte_device *dev);
 334
 335 /**
 336  * This function works as the action for both primary and secondary process
 337  * for static vdev discovery when a secondary process is booting.
 338  *
 339  * step 1, secondary process sends a sync request to ask for vdev in primary;
 340  * step 2, primary process receives the request, and send vdevs one by one;
 341  * step 3, primary process sends back reply, which indicates how many vdevs
 342  * are sent.
 343  */
 344 static int
 345 vdev_action(const struct rte_mp_msg *mp_msg, const void *peer)
 346 {
 347         struct rte_vdev_device *dev;
 348         struct rte_mp_msg mp_resp;
 349         struct vdev_param *ou = (struct vdev_param *)&mp_resp.param;
 350         const struct vdev_param *in = (const struct vdev_param *)mp_msg->param;
 351         const char *devname;
 352         int num;
 353         int ret;
 354
 355         strlcpy(mp_resp.name, VDEV_MP_KEY, sizeof(mp_resp.name));
 356         mp_resp.len_param = sizeof(*ou);
 357         mp_resp.num_fds = 0;
 358
 359         switch (in->type) {
 360         case VDEV_SCAN_REQ:
 361                 ou->type = VDEV_SCAN_ONE;
 362                 ou->num = 1;
 363                 num = 0;
 364
 365                 rte_spinlock_recursive_lock(&vdev_device_list_lock);
 366                 TAILQ_FOREACH(dev, &vdev_device_list, next) {
 367                         devname = rte_vdev_device_name(dev);
 368                         if (strlen(devname) == 0) {
 369                                 VDEV_LOG(INFO, "vdev with no name is not sent");
 370                                 continue;
 371                         }
 372                         VDEV_LOG(INFO, "send vdev, %s", devname);
 373                         strlcpy(ou->name, devname, RTE_DEV_NAME_MAX_LEN);
 374                         if (rte_mp_sendmsg(&mp_resp) < 0)
 375                                 VDEV_LOG(ERR, "send vdev, %s, failed, %s",
 376                                          devname, strerror(rte_errno));
 377                         num++;
 378                 }
 379                 rte_spinlock_recursive_unlock(&vdev_device_list_lock);
 380
 381                 ou->type = VDEV_SCAN_REP;
 382                 ou->num = num;
 383                 if (rte_mp_reply(&mp_resp, peer) < 0)
 384                         VDEV_LOG(ERR, "Failed to reply a scan request");
 385                 break;
 386         case VDEV_SCAN_ONE:
 387                 VDEV_LOG(INFO, "receive vdev, %s", in->name);
 388                 ret = insert_vdev(in->name, NULL, NULL, false);
 389                 if (ret == -EEXIST)
 390                         VDEV_LOG(DEBUG, "device already exist, %s", in->name);
 391                 else if (ret < 0)
 392                         VDEV_LOG(ERR, "failed to add vdev, %s", in->name);
 393                 break;
 394         default:
 395                 VDEV_LOG(ERR, "vdev cannot recognize this message");
 396         }
 397
 398         return 0;
 399 }
 400
 401 static int
 402 vdev_scan(void)
 403 {
 404         struct rte_vdev_device *dev;
 405         struct rte_devargs *devargs;
 406         struct vdev_custom_scan *custom_scan;
 407
 408         if (rte_mp_action_register(VDEV_MP_KEY, vdev_action) < 0 &&
 409             rte_errno != EEXIST) {
 410                 /* for primary, unsupported IPC is not an error */
 411                 if (rte_eal_process_type() == RTE_PROC_PRIMARY &&
 412                                 rte_errno == ENOTSUP)
 413                         goto scan;
 414                 VDEV_LOG(ERR, "Failed to add vdev mp action");
 415                 return -1;
 416         }
 417
 418         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 419                 struct rte_mp_msg mp_req, *mp_rep;
 420                 struct rte_mp_reply mp_reply;
 421                 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
 422                 struct vdev_param *req = (struct vdev_param *)mp_req.param;
 423                 struct vdev_param *resp;
 424
 425                 strlcpy(mp_req.name, VDEV_MP_KEY, sizeof(mp_req.name));
 426                 mp_req.len_param = sizeof(*req);
 427                 mp_req.num_fds = 0;
 428                 req->type = VDEV_SCAN_REQ;
 429                 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
 430                     mp_reply.nb_received == 1) {
 431                         mp_rep = &mp_reply.msgs[0];
 432                         resp = (struct vdev_param *)mp_rep->param;
 433                         VDEV_LOG(INFO, "Received %d vdevs", resp->num);
 434                         free(mp_reply.msgs);
 435                 } else
 436                         VDEV_LOG(ERR, "Failed to request vdev from primary");
 437
 438                 /* Fall through to allow private vdevs in secondary process */
 439         }
 440
 441 scan:
 442         /* call custom scan callbacks if any */
 443         rte_spinlock_lock(&vdev_custom_scan_lock);
 444         TAILQ_FOREACH(custom_scan, &vdev_custom_scans, next) {
 445                 if (custom_scan->callback != NULL)
 446                         /*
 447                          * the callback should update devargs list
 448                          * by calling rte_devargs_insert() with
 449                          *     devargs.bus = rte_bus_find_by_name("vdev");
 450                          *     devargs.type = RTE_DEVTYPE_VIRTUAL;
 451                          *     devargs.policy = RTE_DEV_WHITELISTED;
 452                          */
 453                         custom_scan->callback(custom_scan->user_arg);
 454         }
 455         rte_spinlock_unlock(&vdev_custom_scan_lock);
 456
 457         /* for virtual devices we scan the devargs_list populated via cmdline */
 458         RTE_EAL_DEVARGS_FOREACH("vdev", devargs) {
 459
 460                 dev = calloc(1, sizeof(*dev));
 461                 if (!dev)
 462                         return -1;
 463
 464                 rte_spinlock_recursive_lock(&vdev_device_list_lock);
 465
 466                 if (find_vdev(devargs->name)) {
 467                         rte_spinlock_recursive_unlock(&vdev_device_list_lock);
 468                         free(dev);
 469                         continue;
 470                 }
 471
 472                 dev->device.bus = &rte_vdev_bus;
 473                 dev->device.devargs = devargs;
 474                 dev->device.numa_node = SOCKET_ID_ANY;
 475                 dev->device.name = devargs->name;
 476
 477                 TAILQ_INSERT_TAIL(&vdev_device_list, dev, next);
 478
 479                 rte_spinlock_recursive_unlock(&vdev_device_list_lock);
 480         }
 481
 482         return 0;
 483 }
 484
 485 static int
 486 vdev_probe(void)
 487 {
 488         struct rte_vdev_device *dev;
 489         int r, ret = 0;
 490
 491         /* call the init function for each virtual device */
 492         TAILQ_FOREACH(dev, &vdev_device_list, next) {
 493                 /* we don't use the vdev lock here, as it's only used in DPDK
 494                  * initialization; and we don't want to hold such a lock when
 495                  * we call each driver probe.
 496                  */
 497
 498                 r = vdev_probe_all_drivers(dev);
 499                 if (r != 0) {
 500                         if (r == -EEXIST)
 501                                 continue;
 502                         VDEV_LOG(ERR, "failed to initialize %s device",
 503                                 rte_vdev_device_name(dev));
 504                         ret = -1;
 505                 }
 506         }
 507
 508         return ret;
 509 }
 510
 511 struct rte_device *
 512 rte_vdev_find_device(const struct rte_device *start, rte_dev_cmp_t cmp,
 513                      const void *data)
 514 {
 515         const struct rte_vdev_device *vstart;
 516         struct rte_vdev_device *dev;
 517
 518         rte_spinlock_recursive_lock(&vdev_device_list_lock);
 519         if (start != NULL) {
 520                 vstart = RTE_DEV_TO_VDEV_CONST(start);
 521                 dev = TAILQ_NEXT(vstart, next);
 522         } else {
 523                 dev = TAILQ_FIRST(&vdev_device_list);
 524         }
 525         while (dev != NULL) {
 526                 if (cmp(&dev->device, data) == 0)
 527                         break;
 528                 dev = TAILQ_NEXT(dev, next);
 529         }
 530         rte_spinlock_recursive_unlock(&vdev_device_list_lock);
 531
 532         return dev ? &dev->device : NULL;
 533 }
 534
 535 static int
 536 vdev_plug(struct rte_device *dev)
 537 {
 538         return vdev_probe_all_drivers(RTE_DEV_TO_VDEV(dev));
 539 }
 540
 541 static int
 542 vdev_unplug(struct rte_device *dev)
 543 {
 544         return rte_vdev_uninit(dev->name);
 545 }
 546
 547 static struct rte_bus rte_vdev_bus = {
 548         .scan = vdev_scan,
 549         .probe = vdev_probe,
 550         .find_device = rte_vdev_find_device,
 551         .plug = vdev_plug,
 552         .unplug = vdev_unplug,
 553         .parse = vdev_parse,
 554         .dev_iterate = rte_vdev_dev_iterate,
 555 };
 556
 557 RTE_REGISTER_BUS(vdev, rte_vdev_bus);
 558 RTE_LOG_REGISTER(vdev_logtype_bus, bus.vdev, NOTICE);