ceph/src/spdk/dpdk/examples/vhost_scsi/vhost_scsi.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2017 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <unistd.h>
   7 #include <stdbool.h>
   8 #include <signal.h>
   9 #include <assert.h>
  10 #include <semaphore.h>
  11 #include <linux/virtio_scsi.h>
  12 #include <linux/virtio_ring.h>
  13
  14 #include <rte_atomic.h>
  15 #include <rte_cycles.h>
  16 #include <rte_log.h>
  17 #include <rte_malloc.h>
  18 #include <rte_vhost.h>
  19
  20 #include "vhost_scsi.h"
  21 #include "scsi_spec.h"
  22
  23 #define VIRTIO_SCSI_FEATURES ((1 << VIRTIO_F_NOTIFY_ON_EMPTY) |\
  24                               (1 << VIRTIO_SCSI_F_INOUT) |\
  25                               (1 << VIRTIO_SCSI_F_CHANGE))
  26
  27 /* Path to folder where character device will be created. Can be set by user. */
  28 static char dev_pathname[PATH_MAX] = "";
  29
  30 static struct vhost_scsi_ctrlr *g_vhost_ctrlr;
  31 static int g_should_stop;
  32 static sem_t exit_sem;
  33
  34 static struct vhost_scsi_ctrlr *
  35 vhost_scsi_ctrlr_find(__rte_unused const char *ctrlr_name)
  36 {
  37         /* currently we only support 1 socket file fd */
  38         return g_vhost_ctrlr;
  39 }
  40
  41 static uint64_t gpa_to_vva(int vid, uint64_t gpa, uint64_t *len)
  42 {
  43         char path[PATH_MAX];
  44         struct vhost_scsi_ctrlr *ctrlr;
  45         int ret = 0;
  46
  47         ret = rte_vhost_get_ifname(vid, path, PATH_MAX);
  48         if (ret) {
  49                 fprintf(stderr, "Cannot get socket name\n");
  50                 assert(ret != 0);
  51         }
  52
  53         ctrlr = vhost_scsi_ctrlr_find(path);
  54         if (!ctrlr) {
  55                 fprintf(stderr, "Controller is not ready\n");
  56                 assert(ctrlr != NULL);
  57         }
  58
  59         assert(ctrlr->mem != NULL);
  60
  61         return rte_vhost_va_from_guest_pa(ctrlr->mem, gpa, len);
  62 }
  63
  64 static struct vring_desc *
  65 descriptor_get_next(struct vring_desc *vq_desc, struct vring_desc *cur_desc)
  66 {
  67         return &vq_desc[cur_desc->next];
  68 }
  69
  70 static bool
  71 descriptor_has_next(struct vring_desc *cur_desc)
  72 {
  73         return !!(cur_desc->flags & VRING_DESC_F_NEXT);
  74 }
  75
  76 static bool
  77 descriptor_is_wr(struct vring_desc *cur_desc)
  78 {
  79         return !!(cur_desc->flags & VRING_DESC_F_WRITE);
  80 }
  81
  82 static void
  83 submit_completion(struct vhost_scsi_task *task, uint32_t q_idx)
  84 {
  85         struct rte_vhost_vring *vq;
  86         struct vring_used *used;
  87
  88         vq = task->vq;
  89         used = vq->used;
  90         /* Fill out the next entry in the "used" ring.  id = the
  91          * index of the descriptor that contained the SCSI request.
  92          * len = the total amount of data transferred for the SCSI
  93          * request. We must report the correct len, for variable
  94          * length SCSI CDBs, where we may return less data than
  95          * allocated by the guest VM.
  96          */
  97         used->ring[used->idx & (vq->size - 1)].id = task->req_idx;
  98         used->ring[used->idx & (vq->size - 1)].len = task->data_len;
  99         used->idx++;
 100
 101         /* Send an interrupt back to the guest VM so that it knows
 102          * a completion is ready to be processed.
 103          */
 104         rte_vhost_vring_call(task->bdev->vid, q_idx);
 105 }
 106
 107 static void
 108 vhost_process_read_payload_chain(struct vhost_scsi_task *task)
 109 {
 110         void *data;
 111         uint64_t chunck_len;
 112
 113         task->iovs_cnt = 0;
 114         chunck_len = task->desc->len;
 115         task->resp = (void *)(uintptr_t)gpa_to_vva(task->bdev->vid,
 116                                                    task->desc->addr,
 117                                                    &chunck_len);
 118         if (!task->resp || chunck_len != task->desc->len) {
 119                 fprintf(stderr, "failed to translate desc address.\n");
 120                 return;
 121         }
 122
 123         while (descriptor_has_next(task->desc)) {
 124                 task->desc = descriptor_get_next(task->vq->desc, task->desc);
 125                 chunck_len = task->desc->len;
 126                 data = (void *)(uintptr_t)gpa_to_vva(task->bdev->vid,
 127                                                      task->desc->addr,
 128                                                          &chunck_len);
 129                 if (!data || chunck_len != task->desc->len) {
 130                         fprintf(stderr, "failed to translate desc address.\n");
 131                         return;
 132                 }
 133
 134                 task->iovs[task->iovs_cnt].iov_base = data;
 135                 task->iovs[task->iovs_cnt].iov_len = task->desc->len;
 136                 task->data_len += task->desc->len;
 137                 task->iovs_cnt++;
 138         }
 139 }
 140
 141 static void
 142 vhost_process_write_payload_chain(struct vhost_scsi_task *task)
 143 {
 144         void *data;
 145         uint64_t chunck_len;
 146
 147         task->iovs_cnt = 0;
 148
 149         do {
 150                 chunck_len = task->desc->len;
 151                 data = (void *)(uintptr_t)gpa_to_vva(task->bdev->vid,
 152                                                      task->desc->addr,
 153                                                          &chunck_len);
 154                 if (!data || chunck_len != task->desc->len) {
 155                         fprintf(stderr, "failed to translate desc address.\n");
 156                         return;
 157                 }
 158
 159                 task->iovs[task->iovs_cnt].iov_base = data;
 160                 task->iovs[task->iovs_cnt].iov_len = task->desc->len;
 161                 task->data_len += task->desc->len;
 162                 task->iovs_cnt++;
 163                 task->desc = descriptor_get_next(task->vq->desc, task->desc);
 164         } while (descriptor_has_next(task->desc));
 165
 166         chunck_len = task->desc->len;
 167         task->resp = (void *)(uintptr_t)gpa_to_vva(task->bdev->vid,
 168                                                    task->desc->addr,
 169                                                    &chunck_len);
 170         if (!task->resp || chunck_len != task->desc->len)
 171                 fprintf(stderr, "failed to translate desc address.\n");
 172 }
 173
 174 static struct vhost_block_dev *
 175 vhost_scsi_bdev_construct(const char *bdev_name, const char *bdev_serial,
 176                           uint32_t blk_size, uint64_t blk_cnt,
 177                           bool wce_enable)
 178 {
 179         struct vhost_block_dev *bdev;
 180
 181         bdev = rte_zmalloc(NULL, sizeof(*bdev), RTE_CACHE_LINE_SIZE);
 182         if (!bdev)
 183                 return NULL;
 184
 185         strncpy(bdev->name, bdev_name, sizeof(bdev->name));
 186         strncpy(bdev->product_name, bdev_serial, sizeof(bdev->product_name));
 187         bdev->blocklen = blk_size;
 188         bdev->blockcnt = blk_cnt;
 189         bdev->write_cache = wce_enable;
 190
 191         /* use memory as disk storage space */
 192         bdev->data = rte_zmalloc(NULL, blk_cnt * blk_size, 0);
 193         if (!bdev->data) {
 194                 fprintf(stderr, "no enough reseverd huge memory for disk\n");
 195                 return NULL;
 196         }
 197
 198         return bdev;
 199 }
 200
 201 static void
 202 process_requestq(struct vhost_scsi_ctrlr *ctrlr, uint32_t q_idx)
 203 {
 204         int ret;
 205         struct vhost_scsi_queue *scsi_vq;
 206         struct rte_vhost_vring *vq;
 207
 208         scsi_vq = &ctrlr->bdev->queues[q_idx];
 209         vq = &scsi_vq->vq;
 210         ret = rte_vhost_get_vhost_vring(ctrlr->bdev->vid, q_idx, vq);
 211         assert(ret == 0);
 212
 213         while (vq->avail->idx != scsi_vq->last_used_idx) {
 214                 int req_idx;
 215                 uint16_t last_idx;
 216                 struct vhost_scsi_task *task;
 217                 uint64_t chunck_len;
 218
 219                 last_idx = scsi_vq->last_used_idx & (vq->size - 1);
 220                 req_idx = vq->avail->ring[last_idx];
 221
 222                 task = rte_zmalloc(NULL, sizeof(*task), 0);
 223                 assert(task != NULL);
 224
 225                 task->ctrlr = ctrlr;
 226                 task->bdev = ctrlr->bdev;
 227                 task->vq = vq;
 228                 task->req_idx = req_idx;
 229                 task->desc = &task->vq->desc[task->req_idx];
 230
 231                 /* does not support indirect descriptors */
 232                 assert((task->desc->flags & VRING_DESC_F_INDIRECT) == 0);
 233                 scsi_vq->last_used_idx++;
 234
 235                 chunck_len = task->desc->len;
 236                 task->req = (void *)(uintptr_t)gpa_to_vva(task->bdev->vid,
 237                                                           task->desc->addr,
 238                                                           &chunck_len);
 239                 if (!task->req || chunck_len != task->desc->len) {
 240                         fprintf(stderr, "failed to translate desc address.\n");
 241                         return;
 242                 }
 243
 244                 task->desc = descriptor_get_next(task->vq->desc, task->desc);
 245                 if (!descriptor_has_next(task->desc)) {
 246                         task->dxfer_dir = SCSI_DIR_NONE;
 247                         chunck_len = task->desc->len;
 248                         task->resp = (void *)(uintptr_t)
 249                                               gpa_to_vva(task->bdev->vid,
 250                                                          task->desc->addr,
 251                                                          &chunck_len);
 252                         if (!task->resp || chunck_len != task->desc->len) {
 253                                 fprintf(stderr, "failed to translate desc address.\n");
 254                                 return;
 255                         }
 256                 } else if (!descriptor_is_wr(task->desc)) {
 257                         task->dxfer_dir = SCSI_DIR_TO_DEV;
 258                         vhost_process_write_payload_chain(task);
 259                 } else {
 260                         task->dxfer_dir = SCSI_DIR_FROM_DEV;
 261                         vhost_process_read_payload_chain(task);
 262                 }
 263
 264                 ret = vhost_bdev_process_scsi_commands(ctrlr->bdev, task);
 265                 if (ret) {
 266                         /* invalid response */
 267                         task->resp->response = VIRTIO_SCSI_S_BAD_TARGET;
 268                 } else {
 269                         /* successfully */
 270                         task->resp->response = VIRTIO_SCSI_S_OK;
 271                         task->resp->status = 0;
 272                         task->resp->resid = 0;
 273                 }
 274                 submit_completion(task, q_idx);
 275                 rte_free(task);
 276         }
 277 }
 278
 279 /* Main framework for processing IOs */
 280 static void *
 281 ctrlr_worker(void *arg)
 282 {
 283         uint32_t idx, num;
 284         struct vhost_scsi_ctrlr *ctrlr = (struct vhost_scsi_ctrlr *)arg;
 285         cpu_set_t cpuset;
 286         pthread_t thread;
 287
 288         thread = pthread_self();
 289         CPU_ZERO(&cpuset);
 290         CPU_SET(0, &cpuset);
 291         pthread_setaffinity_np(thread, sizeof(cpu_set_t), &cpuset);
 292
 293         num =  rte_vhost_get_vring_num(ctrlr->bdev->vid);
 294         fprintf(stdout, "Ctrlr Worker Thread Started with %u Vring\n", num);
 295
 296         if (num != NUM_OF_SCSI_QUEUES) {
 297                 fprintf(stderr, "Only 1 IO queue are supported\n");
 298                 exit(0);
 299         }
 300
 301         while (!g_should_stop && ctrlr->bdev != NULL) {
 302                 /* At least 3 vrings, currently only can support 1 IO queue
 303                  * Queue 2 for IO queue, does not support TMF and hotplug
 304                  * for the example application now
 305                  */
 306                 for (idx = 2; idx < num; idx++)
 307                         process_requestq(ctrlr, idx);
 308         }
 309
 310         fprintf(stdout, "Ctrlr Worker Thread Exiting\n");
 311         sem_post(&exit_sem);
 312         return NULL;
 313 }
 314
 315 static int
 316 new_device(int vid)
 317 {
 318         char path[PATH_MAX];
 319         struct vhost_scsi_ctrlr *ctrlr;
 320         struct vhost_scsi_queue *scsi_vq;
 321         struct rte_vhost_vring *vq;
 322         pthread_t tid;
 323         int i, ret;
 324
 325         ret = rte_vhost_get_ifname(vid, path, PATH_MAX);
 326         if (ret) {
 327                 fprintf(stderr, "Cannot get socket name\n");
 328                 return -1;
 329         }
 330
 331         ctrlr = vhost_scsi_ctrlr_find(path);
 332         if (!ctrlr) {
 333                 fprintf(stderr, "Controller is not ready\n");
 334                 return -1;
 335         }
 336
 337         ret = rte_vhost_get_mem_table(vid, &ctrlr->mem);
 338         if (ret) {
 339                 fprintf(stderr, "Get Controller memory region failed\n");
 340                 return -1;
 341         }
 342         assert(ctrlr->mem != NULL);
 343
 344         /* hardcoded block device information with 128MiB */
 345         ctrlr->bdev = vhost_scsi_bdev_construct("malloc0", "vhost_scsi_malloc0",
 346                                                 4096, 32768, 0);
 347         if (!ctrlr->bdev)
 348                 return -1;
 349
 350         ctrlr->bdev->vid = vid;
 351
 352         /* Disable Notifications */
 353         for (i = 0; i < NUM_OF_SCSI_QUEUES; i++) {
 354                 rte_vhost_enable_guest_notification(vid, i, 0);
 355                 /* restore used index */
 356                 scsi_vq = &ctrlr->bdev->queues[i];
 357                 vq = &scsi_vq->vq;
 358                 ret = rte_vhost_get_vhost_vring(ctrlr->bdev->vid, i, vq);
 359                 assert(ret == 0);
 360                 scsi_vq->last_used_idx = vq->used->idx;
 361                 scsi_vq->last_avail_idx = vq->used->idx;
 362         }
 363
 364         g_should_stop = 0;
 365         fprintf(stdout, "New Device %s, Device ID %d\n", path, vid);
 366         if (pthread_create(&tid, NULL, &ctrlr_worker, ctrlr) < 0) {
 367                 fprintf(stderr, "Worker Thread Started Failed\n");
 368                 return -1;
 369         }
 370         pthread_detach(tid);
 371         return 0;
 372 }
 373
 374 static void
 375 destroy_device(int vid)
 376 {
 377         char path[PATH_MAX];
 378         struct vhost_scsi_ctrlr *ctrlr;
 379
 380         rte_vhost_get_ifname(vid, path, PATH_MAX);
 381         fprintf(stdout, "Destroy %s Device ID %d\n", path, vid);
 382         ctrlr = vhost_scsi_ctrlr_find(path);
 383         if (!ctrlr) {
 384                 fprintf(stderr, "Destroy Ctrlr Failed\n");
 385                 return;
 386         }
 387         ctrlr->bdev = NULL;
 388         g_should_stop = 1;
 389
 390         sem_wait(&exit_sem);
 391 }
 392
 393 static const struct vhost_device_ops vhost_scsi_device_ops = {
 394         .new_device =  new_device,
 395         .destroy_device = destroy_device,
 396 };
 397
 398 static struct vhost_scsi_ctrlr *
 399 vhost_scsi_ctrlr_construct(const char *ctrlr_name)
 400 {
 401         int ret;
 402         struct vhost_scsi_ctrlr *ctrlr;
 403         char *path;
 404         char cwd[PATH_MAX];
 405
 406         /* always use current directory */
 407         path = getcwd(cwd, PATH_MAX);
 408         if (!path) {
 409                 fprintf(stderr, "Cannot get current working directory\n");
 410                 return NULL;
 411         }
 412         snprintf(dev_pathname, sizeof(dev_pathname), "%s/%s", path, ctrlr_name);
 413
 414         if (access(dev_pathname, F_OK) != -1) {
 415                 if (unlink(dev_pathname) != 0)
 416                         rte_exit(EXIT_FAILURE, "Cannot remove %s.\n",
 417                                  dev_pathname);
 418         }
 419
 420         if (rte_vhost_driver_register(dev_pathname, 0) != 0) {
 421                 fprintf(stderr, "socket %s already exists\n", dev_pathname);
 422                 return NULL;
 423         }
 424
 425         fprintf(stdout, "socket file: %s created\n", dev_pathname);
 426
 427         ret = rte_vhost_driver_set_features(dev_pathname, VIRTIO_SCSI_FEATURES);
 428         if (ret != 0) {
 429                 fprintf(stderr, "Set vhost driver features failed\n");
 430                 return NULL;
 431         }
 432
 433         ctrlr = rte_zmalloc(NULL, sizeof(*ctrlr), RTE_CACHE_LINE_SIZE);
 434         if (!ctrlr)
 435                 return NULL;
 436
 437         rte_vhost_driver_callback_register(dev_pathname,
 438                                            &vhost_scsi_device_ops);
 439
 440         return ctrlr;
 441 }
 442
 443 static void
 444 signal_handler(__rte_unused int signum)
 445 {
 446
 447         if (access(dev_pathname, F_OK) == 0)
 448                 unlink(dev_pathname);
 449         exit(0);
 450 }
 451
 452 int main(int argc, char *argv[])
 453 {
 454         int ret;
 455
 456         signal(SIGINT, signal_handler);
 457
 458         /* init EAL */
 459         ret = rte_eal_init(argc, argv);
 460         if (ret < 0)
 461                 rte_exit(EXIT_FAILURE, "Error with EAL initialization\n");
 462
 463         g_vhost_ctrlr = vhost_scsi_ctrlr_construct("vhost.socket");
 464         if (g_vhost_ctrlr == NULL) {
 465                 fprintf(stderr, "Construct vhost scsi controller failed\n");
 466                 return 0;
 467         }
 468
 469         if (sem_init(&exit_sem, 0, 0) < 0) {
 470                 fprintf(stderr, "Error init exit_sem\n");
 471                 return -1;
 472         }
 473
 474         rte_vhost_driver_start(dev_pathname);
 475
 476         /* loop for exit the application */
 477         while (1)
 478                 sleep(1);
 479
 480         return 0;
 481 }
 482