examples/vhost/vhost-net-cdev.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <errno.h>
  35 #include <fuse/cuse_lowlevel.h>
  36 #include <linux/limits.h>
  37 #include <linux/vhost.h>
  38 #include <stdint.h>
  39 #include <string.h>
  40 #include <unistd.h>
  41
  42 #include <rte_ethdev.h>
  43 #include <rte_log.h>
  44 #include <rte_string_fns.h>
  45
  46 #include "main.h"
  47 #include "vhost-net-cdev.h"
  48
  49 #define FUSE_OPT_DUMMY          "\0\0"
  50 #define FUSE_OPT_FORE           "-f\0\0"
  51 #define FUSE_OPT_NOMULTI        "-s\0\0"
  52
  53 const uint32_t  default_major = 231;
  54 const uint32_t  default_minor = 1;
  55 const char              cuse_device_name[]      = "/dev/cuse";
  56 const char              default_cdev[] = "vhost-net";
  57
  58 static struct fuse_session                      *session;
  59 static struct vhost_net_device_ops      const *ops;
  60
  61 /*
  62  * Returns vhost_device_ctx from given fuse_req_t. The index is populated later when
  63  * the device is added to the device linked list.
  64  */
  65 static struct vhost_device_ctx
  66 fuse_req_to_vhost_ctx(fuse_req_t req, struct fuse_file_info *fi)
  67 {
  68         struct vhost_device_ctx ctx;
  69         struct fuse_ctx const *const req_ctx = fuse_req_ctx(req);
  70
  71         ctx.pid = req_ctx->pid;
  72         ctx.fh = fi->fh;
  73
  74         return ctx;
  75 }
  76
  77 /*
  78  * When the device is created in QEMU it gets initialised here and added to the device linked list.
  79  */
  80 static void
  81 vhost_net_open(fuse_req_t req, struct fuse_file_info *fi)
  82 {
  83         struct vhost_device_ctx ctx = fuse_req_to_vhost_ctx(req, fi);
  84         int err = 0;
  85
  86         err = ops->new_device(ctx);
  87         if (err == -1) {
  88                 fuse_reply_err(req, EPERM);
  89                 return;
  90         }
  91
  92         fi->fh = err;
  93
  94         RTE_LOG(INFO, CONFIG, "(%"PRIu64") Device configuration started\n", fi->fh);
  95         fuse_reply_open(req, fi);
  96 }
  97
  98 /*
  99  * When QEMU is shutdown or killed the device gets released.
 100  */
 101 static void
 102 vhost_net_release(fuse_req_t req, struct fuse_file_info *fi)
 103 {
 104         int err = 0;
 105         struct vhost_device_ctx ctx = fuse_req_to_vhost_ctx(req, fi);
 106
 107         ops->destroy_device(ctx);
 108         RTE_LOG(INFO, CONFIG, "(%"PRIu64") Device released\n", ctx.fh);
 109         fuse_reply_err(req, err);
 110 }
 111
 112 /*
 113  * Boilerplate code for CUSE IOCTL
 114  * Implicit arguments: ctx, req, result.
 115  */
 116 #define VHOST_IOCTL(func) do {                                                          \
 117         result = (func)(ctx);                                                                   \
 118         fuse_reply_ioctl(req, result, NULL, 0);                                 \
 119 } while(0)                                                                                                      \
 120
 121 /*
 122  * Boilerplate IOCTL RETRY
 123  * Implicit arguments: req.
 124  */
 125 #define VHOST_IOCTL_RETRY(size_r, size_w) do {                                                                  \
 126         struct iovec iov_r = { arg, (size_r) };                                                                         \
 127         struct iovec iov_w = { arg, (size_w) };                                                                         \
 128         fuse_reply_ioctl_retry(req, &iov_r, (size_r)?1:0, &iov_w, (size_w)?1:0);        \
 129 } while(0)                                                                                                                                              \
 130
 131 /*
 132  * Boilerplate code for CUSE Read IOCTL
 133  * Implicit arguments: ctx, req, result, in_bufsz, in_buf.
 134  */
 135 #define VHOST_IOCTL_R(type, var, func) do {                             \
 136         if (!in_bufsz) {                                                                        \
 137                 VHOST_IOCTL_RETRY(sizeof(type), 0);                             \
 138         } else {                                                                                        \
 139                 (var) = *(const type * ) in_buf;                                \
 140                 result = func(ctx, &(var));                                             \
 141                 fuse_reply_ioctl(req, result, NULL, 0);                 \
 142         }                                                                                                       \
 143 } while(0)                                                                                              \
 144
 145 /*
 146  *      Boilerplate code for CUSE Write IOCTL
 147  * Implicit arguments: ctx, req, result, out_bufsz.
 148  */
 149 #define VHOST_IOCTL_W(type, var, func) do {                                             \
 150         if (!out_bufsz) {                                                                                       \
 151                 VHOST_IOCTL_RETRY(0, sizeof(type));                                             \
 152         } else {                                                                                                        \
 153                 result = (func)(ctx, &(var));                                                   \
 154                 fuse_reply_ioctl(req, result, &(var), sizeof(type));    \
 155         }                                                                                                                       \
 156 } while(0)                                                                                                              \
 157
 158 /*
 159  * Boilerplate code for CUSE Read/Write IOCTL
 160  * Implicit arguments: ctx, req, result, in_bufsz, in_buf.
 161  */
 162 #define VHOST_IOCTL_RW(type1, var1, type2, var2, func) do {                     \
 163         if (!in_bufsz) {                                                                                                \
 164                 VHOST_IOCTL_RETRY(sizeof(type1), sizeof(type2));                        \
 165         } else {                                                                                                                \
 166                 (var1) = *(const type1* ) (in_buf);                                                     \
 167                 result = (func)(ctx, (var1), &(var2));                                          \
 168                 fuse_reply_ioctl(req, result, &(var2), sizeof(type2));          \
 169         }                                                                                                                               \
 170 } while(0)                                                                                                                      \
 171
 172 /*
 173  * The IOCTLs are handled using CUSE/FUSE in userspace. Depending on
 174  * the type of IOCTL a buffer is requested to read or to write. This
 175  * request is handled by FUSE and the buffer is then given to CUSE.
 176  */
 177 static void
 178 vhost_net_ioctl(fuse_req_t req, int cmd, void *arg,
 179                 struct fuse_file_info *fi, __rte_unused unsigned flags,
 180                 const void *in_buf, size_t in_bufsz, size_t out_bufsz)
 181 {
 182         struct vhost_device_ctx ctx = fuse_req_to_vhost_ctx(req, fi);
 183         struct vhost_vring_file file;
 184         struct vhost_vring_state state;
 185         struct vhost_vring_addr addr;
 186         uint64_t features;
 187         uint32_t index;
 188         int result = 0;
 189
 190         switch(cmd)
 191         {
 192                 case VHOST_NET_SET_BACKEND:
 193                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_NET_SET_BACKEND\n", ctx.fh);
 194                         VHOST_IOCTL_R(struct vhost_vring_file, file, ops->set_backend);
 195                         break;
 196
 197                 case VHOST_GET_FEATURES:
 198                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_GET_FEATURES\n", ctx.fh);
 199                         VHOST_IOCTL_W(uint64_t, features, ops->get_features);
 200                         break;
 201
 202                 case VHOST_SET_FEATURES:
 203                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_SET_FEATURES\n", ctx.fh);
 204                         VHOST_IOCTL_R(uint64_t, features, ops->set_features);
 205                         break;
 206
 207                 case VHOST_RESET_OWNER:
 208                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_RESET_OWNER\n", ctx.fh);
 209                         VHOST_IOCTL(ops->reset_owner);
 210                         break;
 211
 212                 case VHOST_SET_OWNER:
 213                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_SET_OWNER\n", ctx.fh);
 214                         VHOST_IOCTL(ops->set_owner);
 215                         break;
 216
 217                 case VHOST_SET_MEM_TABLE:
 218                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_SET_MEM_TABLE\n", ctx.fh);
 219                         static struct vhost_memory mem_temp;
 220
 221                         switch(in_bufsz){
 222                                 case 0:
 223                                         VHOST_IOCTL_RETRY(sizeof(struct vhost_memory), 0);
 224                                         break;
 225
 226                                 case sizeof(struct vhost_memory):
 227                                         mem_temp = *(const struct vhost_memory *) in_buf;
 228
 229                                         if (mem_temp.nregions > 0) {
 230                                                 VHOST_IOCTL_RETRY(sizeof(struct vhost_memory) + (sizeof(struct vhost_memory_region) * mem_temp.nregions), 0);
 231                                         } else {
 232                                                 result = -1;
 233                                                 fuse_reply_ioctl(req, result, NULL, 0);
 234                                         }
 235                                         break;
 236
 237                                 default:
 238                                         result = ops->set_mem_table(ctx, in_buf, mem_temp.nregions);
 239                                         if (result)
 240                                                 fuse_reply_err(req, EINVAL);
 241                                         else
 242                                                 fuse_reply_ioctl(req, result, NULL, 0);
 243
 244                         }
 245
 246                         break;
 247
 248                 case VHOST_SET_VRING_NUM:
 249                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_SET_VRING_NUM\n", ctx.fh);
 250                         VHOST_IOCTL_R(struct vhost_vring_state, state, ops->set_vring_num);
 251                         break;
 252
 253                 case VHOST_SET_VRING_BASE:
 254                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_SET_VRING_BASE\n", ctx.fh);
 255                         VHOST_IOCTL_R(struct vhost_vring_state, state, ops->set_vring_base);
 256                         break;
 257
 258                 case VHOST_GET_VRING_BASE:
 259                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_GET_VRING_BASE\n", ctx.fh);
 260                         VHOST_IOCTL_RW(uint32_t, index, struct vhost_vring_state, state, ops->get_vring_base);
 261                         break;
 262
 263                 case VHOST_SET_VRING_ADDR:
 264                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_SET_VRING_ADDR\n", ctx.fh);
 265                         VHOST_IOCTL_R(struct vhost_vring_addr, addr, ops->set_vring_addr);
 266                         break;
 267
 268                 case VHOST_SET_VRING_KICK:
 269                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_SET_VRING_KICK\n", ctx.fh);
 270                         VHOST_IOCTL_R(struct vhost_vring_file, file, ops->set_vring_kick);
 271                         break;
 272
 273                 case VHOST_SET_VRING_CALL:
 274                         LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: VHOST_SET_VRING_CALL\n", ctx.fh);
 275                         VHOST_IOCTL_R(struct vhost_vring_file, file, ops->set_vring_call);
 276                         break;
 277
 278                 default:
 279                         RTE_LOG(ERR, CONFIG, "(%"PRIu64") IOCTL: DOESN NOT EXIST\n", ctx.fh);
 280                         result = -1;
 281                         fuse_reply_ioctl(req, result, NULL, 0);
 282         }
 283
 284         if (result < 0) {
 285                 LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: FAIL\n", ctx.fh);
 286         } else {
 287                 LOG_DEBUG(CONFIG, "(%"PRIu64") IOCTL: SUCCESS\n", ctx.fh);
 288         }
 289 }
 290
 291 /*
 292  * Structure handling open, release and ioctl function pointers is populated.
 293  */
 294 static const struct cuse_lowlevel_ops vhost_net_ops = {
 295         .open           = vhost_net_open,
 296         .release        = vhost_net_release,
 297         .ioctl          = vhost_net_ioctl,
 298 };
 299
 300 /*
 301  * cuse_info is populated and used to register the cuse device. vhost_net_device_ops are
 302  * also passed when the device is registered in main.c.
 303  */
 304 int
 305 register_cuse_device(const char *base_name, int index, struct vhost_net_device_ops const * const pops)
 306 {
 307         struct cuse_info cuse_info;
 308         char device_name[PATH_MAX] = "";
 309         char char_device_name[PATH_MAX] = "";
 310         const char *device_argv[] = { device_name };
 311
 312         char fuse_opt_dummy[] = FUSE_OPT_DUMMY;
 313         char fuse_opt_fore[] = FUSE_OPT_FORE;
 314         char fuse_opt_nomulti[] = FUSE_OPT_NOMULTI;
 315         char *fuse_argv[] = {fuse_opt_dummy, fuse_opt_fore, fuse_opt_nomulti};
 316
 317         if (access(cuse_device_name, R_OK | W_OK) < 0) {
 318                 RTE_LOG(ERR, CONFIG, "Character device %s can't be accessed, maybe not exist\n", cuse_device_name);
 319                 return -1;
 320         }
 321
 322         /*
 323          * The device name is created. This is passed to QEMU so that it can register
 324          * the device with our application. The index allows us to have multiple instances
 325          * of userspace vhost which we can then add devices to separately.
 326          */
 327         if (strncmp(base_name, default_cdev, PATH_MAX)!=0) {
 328                 rte_snprintf(device_name, PATH_MAX, "DEVNAME=%s-%d", base_name, index);
 329                 rte_snprintf(char_device_name, PATH_MAX, "/dev/%s-%d", base_name, index);
 330         } else {
 331                 rte_snprintf(device_name, PATH_MAX, "DEVNAME=%s", base_name);
 332                 rte_snprintf(char_device_name, PATH_MAX, "/dev/%s", base_name);
 333         }
 334
 335         /* Check if device already exists. */
 336         if (access(char_device_name, F_OK) != -1) {
 337                 RTE_LOG(ERR, CONFIG, "Character device %s already exists\n", char_device_name);
 338                 return -1;
 339         }
 340
 341         memset(&cuse_info, 0, sizeof(cuse_info));
 342         cuse_info.dev_major = default_major;
 343         cuse_info.dev_minor = default_minor + index;
 344         cuse_info.dev_info_argc = 1;
 345         cuse_info.dev_info_argv = device_argv;
 346         cuse_info.flags = CUSE_UNRESTRICTED_IOCTL;
 347
 348         ops = pops;
 349
 350         session = cuse_lowlevel_setup(3, fuse_argv,
 351                                 &cuse_info, &vhost_net_ops, 0, NULL);
 352         if (session == NULL)
 353                 return -1;
 354
 355         return 0;
 356 }
 357
 358 /*
 359  * The CUSE session is launched allowing the application to receive open, release and ioctl calls.
 360  */
 361 int
 362 start_cuse_session_loop(void)
 363 {
 364         fuse_session_loop(session);
 365
 366         return 0;
 367 }