hw/virtio/vhost-user.c

   1 /*
   2  * vhost-user
   3  *
   4  * Copyright (c) 2013 Virtual Open Systems Sarl.
   5  *
   6  * This work is licensed under the terms of the GNU GPL, version 2 or later.
   7  * See the COPYING file in the top-level directory.
   8  *
   9  */
  10
  11 #include "hw/virtio/vhost.h"
  12 #include "hw/virtio/vhost-backend.h"
  13 #include "sysemu/char.h"
  14 #include "sysemu/kvm.h"
  15 #include "qemu/error-report.h"
  16 #include "qemu/sockets.h"
  17 #include "exec/ram_addr.h"
  18
  19 #include <fcntl.h>
  20 #include <unistd.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/socket.h>
  23 #include <sys/un.h>
  24 #include <linux/vhost.h>
  25
  26 #define VHOST_MEMORY_MAX_NREGIONS    8
  27 #define VHOST_USER_F_PROTOCOL_FEATURES 30
  28 #define VHOST_USER_PROTOCOL_FEATURE_MASK 0x1ULL
  29
  30 #define VHOST_USER_PROTOCOL_F_MQ    0
  31
  32 typedef enum VhostUserRequest {
  33     VHOST_USER_NONE = 0,
  34     VHOST_USER_GET_FEATURES = 1,
  35     VHOST_USER_SET_FEATURES = 2,
  36     VHOST_USER_SET_OWNER = 3,
  37     VHOST_USER_RESET_DEVICE = 4,
  38     VHOST_USER_SET_MEM_TABLE = 5,
  39     VHOST_USER_SET_LOG_BASE = 6,
  40     VHOST_USER_SET_LOG_FD = 7,
  41     VHOST_USER_SET_VRING_NUM = 8,
  42     VHOST_USER_SET_VRING_ADDR = 9,
  43     VHOST_USER_SET_VRING_BASE = 10,
  44     VHOST_USER_GET_VRING_BASE = 11,
  45     VHOST_USER_SET_VRING_KICK = 12,
  46     VHOST_USER_SET_VRING_CALL = 13,
  47     VHOST_USER_SET_VRING_ERR = 14,
  48     VHOST_USER_GET_PROTOCOL_FEATURES = 15,
  49     VHOST_USER_SET_PROTOCOL_FEATURES = 16,
  50     VHOST_USER_GET_QUEUE_NUM = 17,
  51     VHOST_USER_SET_VRING_ENABLE = 18,
  52     VHOST_USER_MAX
  53 } VhostUserRequest;
  54
  55 typedef struct VhostUserMemoryRegion {
  56     uint64_t guest_phys_addr;
  57     uint64_t memory_size;
  58     uint64_t userspace_addr;
  59     uint64_t mmap_offset;
  60 } VhostUserMemoryRegion;
  61
  62 typedef struct VhostUserMemory {
  63     uint32_t nregions;
  64     uint32_t padding;
  65     VhostUserMemoryRegion regions[VHOST_MEMORY_MAX_NREGIONS];
  66 } VhostUserMemory;
  67
  68 typedef struct VhostUserMsg {
  69     VhostUserRequest request;
  70
  71 #define VHOST_USER_VERSION_MASK     (0x3)
  72 #define VHOST_USER_REPLY_MASK       (0x1<<2)
  73     uint32_t flags;
  74     uint32_t size; /* the following payload size */
  75     union {
  76 #define VHOST_USER_VRING_IDX_MASK   (0xff)
  77 #define VHOST_USER_VRING_NOFD_MASK  (0x1<<8)
  78         uint64_t u64;
  79         struct vhost_vring_state state;
  80         struct vhost_vring_addr addr;
  81         VhostUserMemory memory;
  82     };
  83 } QEMU_PACKED VhostUserMsg;
  84
  85 static VhostUserMsg m __attribute__ ((unused));
  86 #define VHOST_USER_HDR_SIZE (sizeof(m.request) \
  87                             + sizeof(m.flags) \
  88                             + sizeof(m.size))
  89
  90 #define VHOST_USER_PAYLOAD_SIZE (sizeof(m) - VHOST_USER_HDR_SIZE)
  91
  92 /* The version of the protocol we support */
  93 #define VHOST_USER_VERSION    (0x1)
  94
  95 static bool ioeventfd_enabled(void)
  96 {
  97     return kvm_enabled() && kvm_eventfds_enabled();
  98 }
  99
 100 static unsigned long int ioctl_to_vhost_user_request[VHOST_USER_MAX] = {
 101     -1,                     /* VHOST_USER_NONE */
 102     VHOST_GET_FEATURES,     /* VHOST_USER_GET_FEATURES */
 103     VHOST_SET_FEATURES,     /* VHOST_USER_SET_FEATURES */
 104     VHOST_SET_OWNER,        /* VHOST_USER_SET_OWNER */
 105     VHOST_RESET_DEVICE,      /* VHOST_USER_RESET_DEVICE */
 106     VHOST_SET_MEM_TABLE,    /* VHOST_USER_SET_MEM_TABLE */
 107     VHOST_SET_LOG_BASE,     /* VHOST_USER_SET_LOG_BASE */
 108     VHOST_SET_LOG_FD,       /* VHOST_USER_SET_LOG_FD */
 109     VHOST_SET_VRING_NUM,    /* VHOST_USER_SET_VRING_NUM */
 110     VHOST_SET_VRING_ADDR,   /* VHOST_USER_SET_VRING_ADDR */
 111     VHOST_SET_VRING_BASE,   /* VHOST_USER_SET_VRING_BASE */
 112     VHOST_GET_VRING_BASE,   /* VHOST_USER_GET_VRING_BASE */
 113     VHOST_SET_VRING_KICK,   /* VHOST_USER_SET_VRING_KICK */
 114     VHOST_SET_VRING_CALL,   /* VHOST_USER_SET_VRING_CALL */
 115     VHOST_SET_VRING_ERR     /* VHOST_USER_SET_VRING_ERR */
 116 };
 117
 118 static VhostUserRequest vhost_user_request_translate(unsigned long int request)
 119 {
 120     VhostUserRequest idx;
 121
 122     for (idx = 0; idx < VHOST_USER_MAX; idx++) {
 123         if (ioctl_to_vhost_user_request[idx] == request) {
 124             break;
 125         }
 126     }
 127
 128     return (idx == VHOST_USER_MAX) ? VHOST_USER_NONE : idx;
 129 }
 130
 131 static int vhost_user_read(struct vhost_dev *dev, VhostUserMsg *msg)
 132 {
 133     CharDriverState *chr = dev->opaque;
 134     uint8_t *p = (uint8_t *) msg;
 135     int r, size = VHOST_USER_HDR_SIZE;
 136
 137     r = qemu_chr_fe_read_all(chr, p, size);
 138     if (r != size) {
 139         error_report("Failed to read msg header. Read %d instead of %d.", r,
 140                 size);
 141         goto fail;
 142     }
 143
 144     /* validate received flags */
 145     if (msg->flags != (VHOST_USER_REPLY_MASK | VHOST_USER_VERSION)) {
 146         error_report("Failed to read msg header."
 147                 " Flags 0x%x instead of 0x%x.", msg->flags,
 148                 VHOST_USER_REPLY_MASK | VHOST_USER_VERSION);
 149         goto fail;
 150     }
 151
 152     /* validate message size is sane */
 153     if (msg->size > VHOST_USER_PAYLOAD_SIZE) {
 154         error_report("Failed to read msg header."
 155                 " Size %d exceeds the maximum %zu.", msg->size,
 156                 VHOST_USER_PAYLOAD_SIZE);
 157         goto fail;
 158     }
 159
 160     if (msg->size) {
 161         p += VHOST_USER_HDR_SIZE;
 162         size = msg->size;
 163         r = qemu_chr_fe_read_all(chr, p, size);
 164         if (r != size) {
 165             error_report("Failed to read msg payload."
 166                          " Read %d instead of %d.", r, msg->size);
 167             goto fail;
 168         }
 169     }
 170
 171     return 0;
 172
 173 fail:
 174     return -1;
 175 }
 176
 177 static int vhost_user_write(struct vhost_dev *dev, VhostUserMsg *msg,
 178                             int *fds, int fd_num)
 179 {
 180     CharDriverState *chr = dev->opaque;
 181     int size = VHOST_USER_HDR_SIZE + msg->size;
 182
 183     if (fd_num) {
 184         qemu_chr_fe_set_msgfds(chr, fds, fd_num);
 185     }
 186
 187     return qemu_chr_fe_write_all(chr, (const uint8_t *) msg, size) == size ?
 188             0 : -1;
 189 }
 190
 191 static bool vhost_user_one_time_request(VhostUserRequest request)
 192 {
 193     switch (request) {
 194     case VHOST_USER_SET_OWNER:
 195     case VHOST_USER_RESET_DEVICE:
 196     case VHOST_USER_SET_MEM_TABLE:
 197     case VHOST_USER_GET_QUEUE_NUM:
 198         return true;
 199     default:
 200         return false;
 201     }
 202 }
 203
 204 static int vhost_user_call(struct vhost_dev *dev, unsigned long int request,
 205         void *arg)
 206 {
 207     VhostUserMsg msg;
 208     VhostUserRequest msg_request;
 209     struct vhost_vring_file *file = 0;
 210     int need_reply = 0;
 211     int fds[VHOST_MEMORY_MAX_NREGIONS];
 212     int i, fd;
 213     size_t fd_num = 0;
 214
 215     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER);
 216
 217     /* only translate vhost ioctl requests */
 218     if (request > VHOST_USER_MAX) {
 219         msg_request = vhost_user_request_translate(request);
 220     } else {
 221         msg_request = request;
 222     }
 223
 224     /*
 225      * For non-vring specific requests, like VHOST_USER_SET_MEM_TABLE,
 226      * we just need send it once in the first time. For later such
 227      * request, we just ignore it.
 228      */
 229     if (vhost_user_one_time_request(msg_request) && dev->vq_index != 0) {
 230         return 0;
 231     }
 232
 233     msg.request = msg_request;
 234     msg.flags = VHOST_USER_VERSION;
 235     msg.size = 0;
 236
 237     switch (msg_request) {
 238     case VHOST_USER_GET_FEATURES:
 239     case VHOST_USER_GET_PROTOCOL_FEATURES:
 240     case VHOST_USER_GET_QUEUE_NUM:
 241         need_reply = 1;
 242         break;
 243
 244     case VHOST_USER_SET_FEATURES:
 245     case VHOST_USER_SET_LOG_BASE:
 246     case VHOST_USER_SET_PROTOCOL_FEATURES:
 247         msg.u64 = *((__u64 *) arg);
 248         msg.size = sizeof(m.u64);
 249         break;
 250
 251     case VHOST_USER_SET_OWNER:
 252     case VHOST_USER_RESET_DEVICE:
 253         break;
 254
 255     case VHOST_USER_SET_MEM_TABLE:
 256         for (i = 0; i < dev->mem->nregions; ++i) {
 257             struct vhost_memory_region *reg = dev->mem->regions + i;
 258             ram_addr_t ram_addr;
 259
 260             assert((uintptr_t)reg->userspace_addr == reg->userspace_addr);
 261             qemu_ram_addr_from_host((void *)(uintptr_t)reg->userspace_addr, &ram_addr);
 262             fd = qemu_get_ram_fd(ram_addr);
 263             if (fd > 0) {
 264                 msg.memory.regions[fd_num].userspace_addr = reg->userspace_addr;
 265                 msg.memory.regions[fd_num].memory_size  = reg->memory_size;
 266                 msg.memory.regions[fd_num].guest_phys_addr = reg->guest_phys_addr;
 267                 msg.memory.regions[fd_num].mmap_offset = reg->userspace_addr -
 268                     (uintptr_t) qemu_get_ram_block_host_ptr(ram_addr);
 269                 assert(fd_num < VHOST_MEMORY_MAX_NREGIONS);
 270                 fds[fd_num++] = fd;
 271             }
 272         }
 273
 274         msg.memory.nregions = fd_num;
 275
 276         if (!fd_num) {
 277             error_report("Failed initializing vhost-user memory map, "
 278                     "consider using -object memory-backend-file share=on");
 279             return -1;
 280         }
 281
 282         msg.size = sizeof(m.memory.nregions);
 283         msg.size += sizeof(m.memory.padding);
 284         msg.size += fd_num * sizeof(VhostUserMemoryRegion);
 285
 286         break;
 287
 288     case VHOST_USER_SET_LOG_FD:
 289         fds[fd_num++] = *((int *) arg);
 290         break;
 291
 292     case VHOST_USER_SET_VRING_NUM:
 293     case VHOST_USER_SET_VRING_BASE:
 294     case VHOST_USER_SET_VRING_ENABLE:
 295         memcpy(&msg.state, arg, sizeof(struct vhost_vring_state));
 296         msg.size = sizeof(m.state);
 297         break;
 298
 299     case VHOST_USER_GET_VRING_BASE:
 300         memcpy(&msg.state, arg, sizeof(struct vhost_vring_state));
 301         msg.size = sizeof(m.state);
 302         need_reply = 1;
 303         break;
 304
 305     case VHOST_USER_SET_VRING_ADDR:
 306         memcpy(&msg.addr, arg, sizeof(struct vhost_vring_addr));
 307         msg.size = sizeof(m.addr);
 308         break;
 309
 310     case VHOST_USER_SET_VRING_KICK:
 311     case VHOST_USER_SET_VRING_CALL:
 312     case VHOST_USER_SET_VRING_ERR:
 313         file = arg;
 314         msg.u64 = file->index & VHOST_USER_VRING_IDX_MASK;
 315         msg.size = sizeof(m.u64);
 316         if (ioeventfd_enabled() && file->fd > 0) {
 317             fds[fd_num++] = file->fd;
 318         } else {
 319             msg.u64 |= VHOST_USER_VRING_NOFD_MASK;
 320         }
 321         break;
 322     default:
 323         error_report("vhost-user trying to send unhandled ioctl");
 324         return -1;
 325         break;
 326     }
 327
 328     if (vhost_user_write(dev, &msg, fds, fd_num) < 0) {
 329         return 0;
 330     }
 331
 332     if (need_reply) {
 333         if (vhost_user_read(dev, &msg) < 0) {
 334             return 0;
 335         }
 336
 337         if (msg_request != msg.request) {
 338             error_report("Received unexpected msg type."
 339                     " Expected %d received %d", msg_request, msg.request);
 340             return -1;
 341         }
 342
 343         switch (msg_request) {
 344         case VHOST_USER_GET_FEATURES:
 345         case VHOST_USER_GET_PROTOCOL_FEATURES:
 346         case VHOST_USER_GET_QUEUE_NUM:
 347             if (msg.size != sizeof(m.u64)) {
 348                 error_report("Received bad msg size.");
 349                 return -1;
 350             }
 351             *((__u64 *) arg) = msg.u64;
 352             break;
 353         case VHOST_USER_GET_VRING_BASE:
 354             if (msg.size != sizeof(m.state)) {
 355                 error_report("Received bad msg size.");
 356                 return -1;
 357             }
 358             memcpy(arg, &msg.state, sizeof(struct vhost_vring_state));
 359             break;
 360         default:
 361             error_report("Received unexpected msg type.");
 362             return -1;
 363             break;
 364         }
 365     }
 366
 367     return 0;
 368 }
 369
 370 static int vhost_user_init(struct vhost_dev *dev, void *opaque)
 371 {
 372     unsigned long long features;
 373     int err;
 374
 375     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER);
 376
 377     dev->opaque = opaque;
 378
 379     err = vhost_user_call(dev, VHOST_USER_GET_FEATURES, &features);
 380     if (err < 0) {
 381         return err;
 382     }
 383
 384     if (virtio_has_feature(features, VHOST_USER_F_PROTOCOL_FEATURES)) {
 385         dev->backend_features |= 1ULL << VHOST_USER_F_PROTOCOL_FEATURES;
 386
 387         err = vhost_user_call(dev, VHOST_USER_GET_PROTOCOL_FEATURES, &features);
 388         if (err < 0) {
 389             return err;
 390         }
 391
 392         dev->protocol_features = features & VHOST_USER_PROTOCOL_FEATURE_MASK;
 393         err = vhost_user_call(dev, VHOST_USER_SET_PROTOCOL_FEATURES,
 394                               &dev->protocol_features);
 395         if (err < 0) {
 396             return err;
 397         }
 398
 399         /* query the max queues we support if backend supports Multiple Queue */
 400         if (dev->protocol_features & (1ULL << VHOST_USER_PROTOCOL_F_MQ)) {
 401             err = vhost_user_call(dev, VHOST_USER_GET_QUEUE_NUM, &dev->max_queues);
 402             if (err < 0) {
 403                 return err;
 404             }
 405         }
 406     }
 407
 408     return 0;
 409 }
 410
 411 static int vhost_user_set_vring_enable(struct vhost_dev *dev, int enable)
 412 {
 413     struct vhost_vring_state state = {
 414         .index = dev->vq_index,
 415         .num   = enable,
 416     };
 417
 418     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER);
 419
 420     if (!(dev->protocol_features & (1ULL << VHOST_USER_PROTOCOL_F_MQ))) {
 421         return -1;
 422     }
 423
 424     return vhost_user_call(dev, VHOST_USER_SET_VRING_ENABLE, &state);
 425 }
 426
 427 static int vhost_user_cleanup(struct vhost_dev *dev)
 428 {
 429     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER);
 430
 431     dev->opaque = 0;
 432
 433     return 0;
 434 }
 435
 436 static int vhost_user_get_vq_index(struct vhost_dev *dev, int idx)
 437 {
 438     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
 439
 440     return idx;
 441 }
 442
 443 const VhostOps user_ops = {
 444         .backend_type = VHOST_BACKEND_TYPE_USER,
 445         .vhost_call = vhost_user_call,
 446         .vhost_backend_init = vhost_user_init,
 447         .vhost_backend_cleanup = vhost_user_cleanup,
 448         .vhost_backend_get_vq_index = vhost_user_get_vq_index,
 449         .vhost_backend_set_vring_enable = vhost_user_set_vring_enable,
 450 };