drivers/net/virtio/virtio_user/vhost_vdpa.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2020 Red Hat Inc.
   3  */
   4
   5 #include <sys/ioctl.h>
   6 #include <sys/types.h>
   7 #include <sys/stat.h>
   8 #include <fcntl.h>
   9 #include <unistd.h>
  10
  11 #include <rte_memory.h>
  12
  13 #include "vhost.h"
  14 #include "virtio_user_dev.h"
  15
  16 /* vhost kernel & vdpa ioctls */
  17 #define VHOST_VIRTIO 0xAF
  18 #define VHOST_GET_FEATURES _IOR(VHOST_VIRTIO, 0x00, __u64)
  19 #define VHOST_SET_FEATURES _IOW(VHOST_VIRTIO, 0x00, __u64)
  20 #define VHOST_SET_OWNER _IO(VHOST_VIRTIO, 0x01)
  21 #define VHOST_RESET_OWNER _IO(VHOST_VIRTIO, 0x02)
  22 #define VHOST_SET_MEM_TABLE _IOW(VHOST_VIRTIO, 0x03, void *)
  23 #define VHOST_SET_LOG_BASE _IOW(VHOST_VIRTIO, 0x04, __u64)
  24 #define VHOST_SET_LOG_FD _IOW(VHOST_VIRTIO, 0x07, int)
  25 #define VHOST_SET_VRING_NUM _IOW(VHOST_VIRTIO, 0x10, struct vhost_vring_state)
  26 #define VHOST_SET_VRING_ADDR _IOW(VHOST_VIRTIO, 0x11, struct vhost_vring_addr)
  27 #define VHOST_SET_VRING_BASE _IOW(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
  28 #define VHOST_GET_VRING_BASE _IOWR(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
  29 #define VHOST_SET_VRING_KICK _IOW(VHOST_VIRTIO, 0x20, struct vhost_vring_file)
  30 #define VHOST_SET_VRING_CALL _IOW(VHOST_VIRTIO, 0x21, struct vhost_vring_file)
  31 #define VHOST_SET_VRING_ERR _IOW(VHOST_VIRTIO, 0x22, struct vhost_vring_file)
  32 #define VHOST_NET_SET_BACKEND _IOW(VHOST_VIRTIO, 0x30, struct vhost_vring_file)
  33 #define VHOST_VDPA_GET_DEVICE_ID _IOR(VHOST_VIRTIO, 0x70, __u32)
  34 #define VHOST_VDPA_GET_STATUS _IOR(VHOST_VIRTIO, 0x71, __u8)
  35 #define VHOST_VDPA_SET_STATUS _IOW(VHOST_VIRTIO, 0x72, __u8)
  36 #define VHOST_VDPA_SET_VRING_ENABLE     _IOW(VHOST_VIRTIO, 0x75, \
  37                                              struct vhost_vring_state)
  38 #define VHOST_SET_BACKEND_FEATURES _IOW(VHOST_VIRTIO, 0x25, __u64)
  39 #define VHOST_GET_BACKEND_FEATURES _IOR(VHOST_VIRTIO, 0x26, __u64)
  40
  41 static uint64_t vhost_req_user_to_vdpa[] = {
  42         [VHOST_USER_SET_OWNER] = VHOST_SET_OWNER,
  43         [VHOST_USER_RESET_OWNER] = VHOST_RESET_OWNER,
  44         [VHOST_USER_SET_FEATURES] = VHOST_SET_FEATURES,
  45         [VHOST_USER_GET_FEATURES] = VHOST_GET_FEATURES,
  46         [VHOST_USER_SET_VRING_CALL] = VHOST_SET_VRING_CALL,
  47         [VHOST_USER_SET_VRING_NUM] = VHOST_SET_VRING_NUM,
  48         [VHOST_USER_SET_VRING_BASE] = VHOST_SET_VRING_BASE,
  49         [VHOST_USER_GET_VRING_BASE] = VHOST_GET_VRING_BASE,
  50         [VHOST_USER_SET_VRING_ADDR] = VHOST_SET_VRING_ADDR,
  51         [VHOST_USER_SET_VRING_KICK] = VHOST_SET_VRING_KICK,
  52         [VHOST_USER_SET_MEM_TABLE] = VHOST_SET_MEM_TABLE,
  53         [VHOST_USER_SET_STATUS] = VHOST_VDPA_SET_STATUS,
  54         [VHOST_USER_GET_STATUS] = VHOST_VDPA_GET_STATUS,
  55         [VHOST_USER_SET_VRING_ENABLE] = VHOST_VDPA_SET_VRING_ENABLE,
  56         [VHOST_USER_GET_PROTOCOL_FEATURES] = VHOST_GET_BACKEND_FEATURES,
  57         [VHOST_USER_SET_PROTOCOL_FEATURES] = VHOST_SET_BACKEND_FEATURES,
  58 };
  59
  60 /* no alignment requirement */
  61 struct vhost_iotlb_msg {
  62         uint64_t iova;
  63         uint64_t size;
  64         uint64_t uaddr;
  65 #define VHOST_ACCESS_RO      0x1
  66 #define VHOST_ACCESS_WO      0x2
  67 #define VHOST_ACCESS_RW      0x3
  68         uint8_t perm;
  69 #define VHOST_IOTLB_MISS           1
  70 #define VHOST_IOTLB_UPDATE         2
  71 #define VHOST_IOTLB_INVALIDATE     3
  72 #define VHOST_IOTLB_ACCESS_FAIL    4
  73         uint8_t type;
  74 };
  75
  76 #define VHOST_IOTLB_MSG_V2 0x2
  77
  78 struct vhost_msg {
  79         uint32_t type;
  80         uint32_t reserved;
  81         union {
  82                 struct vhost_iotlb_msg iotlb;
  83                 uint8_t padding[64];
  84         };
  85 };
  86
  87 static int
  88 vhost_vdpa_dma_map(struct virtio_user_dev *dev, void *addr,
  89                                   uint64_t iova, size_t len)
  90 {
  91         struct vhost_msg msg = {};
  92
  93         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
  94                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
  95                 return -1;
  96         }
  97
  98         msg.type = VHOST_IOTLB_MSG_V2;
  99         msg.iotlb.type = VHOST_IOTLB_UPDATE;
 100         msg.iotlb.iova = iova;
 101         msg.iotlb.uaddr = (uint64_t)(uintptr_t)addr;
 102         msg.iotlb.size = len;
 103         msg.iotlb.perm = VHOST_ACCESS_RW;
 104
 105         PMD_DRV_LOG(DEBUG, "%s: iova: 0x%" PRIx64 ", addr: %p, len: 0x%zx",
 106                         __func__, iova, addr, len);
 107
 108         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 109                 PMD_DRV_LOG(ERR, "Failed to send IOTLB update (%s)",
 110                                 strerror(errno));
 111                 return -1;
 112         }
 113
 114         return 0;
 115 }
 116
 117 static int
 118 vhost_vdpa_dma_unmap(struct virtio_user_dev *dev, __rte_unused void *addr,
 119                                   uint64_t iova, size_t len)
 120 {
 121         struct vhost_msg msg = {};
 122
 123         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 124                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 125                 return -1;
 126         }
 127
 128         msg.type = VHOST_IOTLB_MSG_V2;
 129         msg.iotlb.type = VHOST_IOTLB_INVALIDATE;
 130         msg.iotlb.iova = iova;
 131         msg.iotlb.size = len;
 132
 133         PMD_DRV_LOG(DEBUG, "%s: iova: 0x%" PRIx64 ", len: 0x%zx",
 134                         __func__, iova, len);
 135
 136         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 137                 PMD_DRV_LOG(ERR, "Failed to send IOTLB invalidate (%s)",
 138                                 strerror(errno));
 139                 return -1;
 140         }
 141
 142         return 0;
 143 }
 144
 145
 146 static int
 147 vhost_vdpa_map_contig(const struct rte_memseg_list *msl,
 148                 const struct rte_memseg *ms, size_t len, void *arg)
 149 {
 150         struct virtio_user_dev *dev = arg;
 151
 152         if (msl->external)
 153                 return 0;
 154
 155         return vhost_vdpa_dma_map(dev, ms->addr, ms->iova, len);
 156 }
 157
 158 static int
 159 vhost_vdpa_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
 160                 void *arg)
 161 {
 162         struct virtio_user_dev *dev = arg;
 163
 164         /* skip external memory that isn't a heap */
 165         if (msl->external && !msl->heap)
 166                 return 0;
 167
 168         /* skip any segments with invalid IOVA addresses */
 169         if (ms->iova == RTE_BAD_IOVA)
 170                 return 0;
 171
 172         /* if IOVA mode is VA, we've already mapped the internal segments */
 173         if (!msl->external && rte_eal_iova_mode() == RTE_IOVA_VA)
 174                 return 0;
 175
 176         return vhost_vdpa_dma_map(dev, ms->addr, ms->iova, ms->len);
 177 }
 178
 179 static int
 180 vhost_vdpa_dma_map_all(struct virtio_user_dev *dev)
 181 {
 182         vhost_vdpa_dma_unmap(dev, NULL, 0, SIZE_MAX);
 183
 184         if (rte_eal_iova_mode() == RTE_IOVA_VA) {
 185                 /* with IOVA as VA mode, we can get away with mapping contiguous
 186                  * chunks rather than going page-by-page.
 187                  */
 188                 int ret = rte_memseg_contig_walk_thread_unsafe(
 189                                 vhost_vdpa_map_contig, dev);
 190                 if (ret)
 191                         return ret;
 192                 /* we have to continue the walk because we've skipped the
 193                  * external segments during the config walk.
 194                  */
 195         }
 196         return rte_memseg_walk_thread_unsafe(vhost_vdpa_map, dev);
 197 }
 198
 199 /* with below features, vhost vdpa does not need to do the checksum and TSO,
 200  * these info will be passed to virtio_user through virtio net header.
 201  */
 202 #define VHOST_VDPA_GUEST_OFFLOADS_MASK  \
 203         ((1ULL << VIRTIO_NET_F_GUEST_CSUM) |    \
 204          (1ULL << VIRTIO_NET_F_GUEST_TSO4) |    \
 205          (1ULL << VIRTIO_NET_F_GUEST_TSO6) |    \
 206          (1ULL << VIRTIO_NET_F_GUEST_ECN)  |    \
 207          (1ULL << VIRTIO_NET_F_GUEST_UFO))
 208
 209 #define VHOST_VDPA_HOST_OFFLOADS_MASK           \
 210         ((1ULL << VIRTIO_NET_F_HOST_TSO4) |     \
 211          (1ULL << VIRTIO_NET_F_HOST_TSO6) |     \
 212          (1ULL << VIRTIO_NET_F_CSUM))
 213
 214 static int
 215 vhost_vdpa_ioctl(struct virtio_user_dev *dev,
 216                    enum vhost_user_request req,
 217                    void *arg)
 218 {
 219         int ret = -1;
 220         uint64_t req_vdpa;
 221
 222         PMD_DRV_LOG(INFO, "%s", vhost_msg_strings[req]);
 223
 224         req_vdpa = vhost_req_user_to_vdpa[req];
 225
 226         if (req_vdpa == VHOST_SET_MEM_TABLE)
 227                 return vhost_vdpa_dma_map_all(dev);
 228
 229         if (req_vdpa == VHOST_SET_FEATURES) {
 230                 /* WORKAROUND */
 231                 *(uint64_t *)arg |= 1ULL << VIRTIO_F_IOMMU_PLATFORM;
 232
 233                 /* Multiqueue not supported for now */
 234                 *(uint64_t *)arg &= ~(1ULL << VIRTIO_NET_F_MQ);
 235         }
 236
 237         switch (req_vdpa) {
 238         case VHOST_SET_VRING_NUM:
 239         case VHOST_SET_VRING_ADDR:
 240         case VHOST_SET_VRING_BASE:
 241         case VHOST_GET_VRING_BASE:
 242         case VHOST_SET_VRING_KICK:
 243         case VHOST_SET_VRING_CALL:
 244                 PMD_DRV_LOG(DEBUG, "vhostfd=%d, index=%u",
 245                             dev->vhostfd, *(unsigned int *)arg);
 246                 break;
 247         default:
 248                 break;
 249         }
 250
 251         ret = ioctl(dev->vhostfd, req_vdpa, arg);
 252         if (ret < 0)
 253                 PMD_DRV_LOG(ERR, "%s failed: %s",
 254                             vhost_msg_strings[req], strerror(errno));
 255
 256         return ret;
 257 }
 258
 259 /**
 260  * Set up environment to talk with a vhost vdpa backend.
 261  *
 262  * @return
 263  *   - (-1) if fail to set up;
 264  *   - (>=0) if successful.
 265  */
 266 static int
 267 vhost_vdpa_setup(struct virtio_user_dev *dev)
 268 {
 269         uint32_t did = (uint32_t)-1;
 270
 271         dev->vhostfd = open(dev->path, O_RDWR);
 272         if (dev->vhostfd < 0) {
 273                 PMD_DRV_LOG(ERR, "Failed to open %s: %s\n",
 274                                 dev->path, strerror(errno));
 275                 return -1;
 276         }
 277
 278         if (ioctl(dev->vhostfd, VHOST_VDPA_GET_DEVICE_ID, &did) < 0 ||
 279                         did != VIRTIO_ID_NETWORK) {
 280                 PMD_DRV_LOG(ERR, "Invalid vdpa device ID: %u\n", did);
 281                 return -1;
 282         }
 283
 284         return 0;
 285 }
 286
 287 static int
 288 vhost_vdpa_enable_queue_pair(struct virtio_user_dev *dev,
 289                                uint16_t pair_idx,
 290                                int enable)
 291 {
 292         int i;
 293
 294         if (dev->qp_enabled[pair_idx] == enable)
 295                 return 0;
 296
 297         for (i = 0; i < 2; ++i) {
 298                 struct vhost_vring_state state = {
 299                         .index = pair_idx * 2 + i,
 300                         .num   = enable,
 301                 };
 302
 303                 if (vhost_vdpa_ioctl(dev, VHOST_USER_SET_VRING_ENABLE, &state))
 304                         return -1;
 305         }
 306
 307         dev->qp_enabled[pair_idx] = enable;
 308
 309         return 0;
 310 }
 311
 312 struct virtio_user_backend_ops virtio_ops_vdpa = {
 313         .setup = vhost_vdpa_setup,
 314         .send_request = vhost_vdpa_ioctl,
 315         .enable_qp = vhost_vdpa_enable_queue_pair,
 316         .dma_map = vhost_vdpa_dma_map,
 317         .dma_unmap = vhost_vdpa_dma_unmap,
 318 };