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