lib/vhost/vhost.h

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2018 Intel Corporation
   3  */
   4
   5 #ifndef _VHOST_NET_CDEV_H_
   6 #define _VHOST_NET_CDEV_H_
   7 #include <stdint.h>
   8 #include <stdio.h>
   9 #include <stdbool.h>
  10 #include <sys/types.h>
  11 #include <sys/queue.h>
  12 #include <unistd.h>
  13 #include <linux/vhost.h>
  14 #include <linux/virtio_net.h>
  15 #include <sys/socket.h>
  16 #include <linux/if.h>
  17
  18 #include <rte_log.h>
  19 #include <rte_ether.h>
  20 #include <rte_malloc.h>
  21 #include <rte_dmadev.h>
  22
  23 #include "rte_vhost.h"
  24 #include "vdpa_driver.h"
  25
  26 #include "rte_vhost_async.h"
  27
  28 /* Used to indicate that the device is running on a data core */
  29 #define VIRTIO_DEV_RUNNING ((uint32_t)1 << 0)
  30 /* Used to indicate that the device is ready to operate */
  31 #define VIRTIO_DEV_READY ((uint32_t)1 << 1)
  32 /* Used to indicate that the built-in vhost net device backend is enabled */
  33 #define VIRTIO_DEV_BUILTIN_VIRTIO_NET ((uint32_t)1 << 2)
  34 /* Used to indicate that the device has its own data path and configured */
  35 #define VIRTIO_DEV_VDPA_CONFIGURED ((uint32_t)1 << 3)
  36 /* Used to indicate that the feature negotiation failed */
  37 #define VIRTIO_DEV_FEATURES_FAILED ((uint32_t)1 << 4)
  38 /* Used to indicate that the virtio_net tx code should fill TX ol_flags */
  39 #define VIRTIO_DEV_LEGACY_OL_FLAGS ((uint32_t)1 << 5)
  40
  41 /* Backend value set by guest. */
  42 #define VIRTIO_DEV_STOPPED -1
  43
  44 #define BUF_VECTOR_MAX 256
  45
  46 #define VHOST_LOG_CACHE_NR 32
  47
  48 #define MAX_PKT_BURST 32
  49
  50 #define VHOST_MAX_ASYNC_IT (MAX_PKT_BURST)
  51 #define VHOST_MAX_ASYNC_VEC 2048
  52 #define VIRTIO_MAX_RX_PKTLEN 9728U
  53 #define VHOST_DMA_MAX_COPY_COMPLETE ((VIRTIO_MAX_RX_PKTLEN / RTE_MBUF_DEFAULT_DATAROOM) \
  54                 * MAX_PKT_BURST)
  55
  56 #define PACKED_DESC_ENQUEUE_USED_FLAG(w)        \
  57         ((w) ? (VRING_DESC_F_AVAIL | VRING_DESC_F_USED | VRING_DESC_F_WRITE) : \
  58                 VRING_DESC_F_WRITE)
  59 #define PACKED_DESC_DEQUEUE_USED_FLAG(w)        \
  60         ((w) ? (VRING_DESC_F_AVAIL | VRING_DESC_F_USED) : 0x0)
  61 #define PACKED_DESC_SINGLE_DEQUEUE_FLAG (VRING_DESC_F_NEXT | \
  62                                          VRING_DESC_F_INDIRECT)
  63
  64 #define PACKED_BATCH_SIZE (RTE_CACHE_LINE_SIZE / \
  65                             sizeof(struct vring_packed_desc))
  66 #define PACKED_BATCH_MASK (PACKED_BATCH_SIZE - 1)
  67
  68 #ifdef VHOST_GCC_UNROLL_PRAGMA
  69 #define vhost_for_each_try_unroll(iter, val, size) _Pragma("GCC unroll 4") \
  70         for (iter = val; iter < size; iter++)
  71 #endif
  72
  73 #ifdef VHOST_CLANG_UNROLL_PRAGMA
  74 #define vhost_for_each_try_unroll(iter, val, size) _Pragma("unroll 4") \
  75         for (iter = val; iter < size; iter++)
  76 #endif
  77
  78 #ifdef VHOST_ICC_UNROLL_PRAGMA
  79 #define vhost_for_each_try_unroll(iter, val, size) _Pragma("unroll (4)") \
  80         for (iter = val; iter < size; iter++)
  81 #endif
  82
  83 #ifndef vhost_for_each_try_unroll
  84 #define vhost_for_each_try_unroll(iter, val, num) \
  85         for (iter = val; iter < num; iter++)
  86 #endif
  87
  88 /**
  89  * Structure contains buffer address, length and descriptor index
  90  * from vring to do scatter RX.
  91  */
  92 struct buf_vector {
  93         uint64_t buf_iova;
  94         uint64_t buf_addr;
  95         uint32_t buf_len;
  96         uint32_t desc_idx;
  97 };
  98
  99 /*
 100  * Structure contains the info for each batched memory copy.
 101  */
 102 struct batch_copy_elem {
 103         void *dst;
 104         void *src;
 105         uint32_t len;
 106         uint64_t log_addr;
 107 };
 108
 109 /*
 110  * Structure that contains the info for batched dirty logging.
 111  */
 112 struct log_cache_entry {
 113         uint32_t offset;
 114         unsigned long val;
 115 };
 116
 117 struct vring_used_elem_packed {
 118         uint16_t id;
 119         uint16_t flags;
 120         uint32_t len;
 121         uint32_t count;
 122 };
 123
 124 /**
 125  * iovec
 126  */
 127 struct vhost_iovec {
 128         void *src_addr;
 129         void *dst_addr;
 130         size_t len;
 131 };
 132
 133 /**
 134  * iovec iterator
 135  */
 136 struct vhost_iov_iter {
 137         /** pointer to the iovec array */
 138         struct vhost_iovec *iov;
 139         /** number of iovec in this iterator */
 140         unsigned long nr_segs;
 141 };
 142
 143 struct async_dma_vchan_info {
 144         /* circular array to track if packet copy completes */
 145         bool **pkts_cmpl_flag_addr;
 146
 147         /* max elements in 'pkts_cmpl_flag_addr' */
 148         uint16_t ring_size;
 149         /* ring index mask for 'pkts_cmpl_flag_addr' */
 150         uint16_t ring_mask;
 151
 152         /**
 153          * DMA virtual channel lock. Although it is able to bind DMA
 154          * virtual channels to data plane threads, vhost control plane
 155          * thread could call data plane functions too, thus causing
 156          * DMA device contention.
 157          *
 158          * For example, in VM exit case, vhost control plane thread needs
 159          * to clear in-flight packets before disable vring, but there could
 160          * be anotther data plane thread is enqueuing packets to the same
 161          * vring with the same DMA virtual channel. As dmadev PMD functions
 162          * are lock-free, the control plane and data plane threads could
 163          * operate the same DMA virtual channel at the same time.
 164          */
 165         rte_spinlock_t dma_lock;
 166 };
 167
 168 struct async_dma_info {
 169         struct async_dma_vchan_info *vchans;
 170         /* number of registered virtual channels */
 171         uint16_t nr_vchans;
 172 };
 173
 174 extern struct async_dma_info dma_copy_track[RTE_DMADEV_DEFAULT_MAX];
 175
 176 /**
 177  * inflight async packet information
 178  */
 179 struct async_inflight_info {
 180         struct rte_mbuf *mbuf;
 181         uint16_t descs; /* num of descs inflight */
 182         uint16_t nr_buffers; /* num of buffers inflight for packed ring */
 183 };
 184
 185 struct vhost_async {
 186         struct vhost_iov_iter iov_iter[VHOST_MAX_ASYNC_IT];
 187         struct vhost_iovec iovec[VHOST_MAX_ASYNC_VEC];
 188         uint16_t iter_idx;
 189         uint16_t iovec_idx;
 190
 191         /* data transfer status */
 192         struct async_inflight_info *pkts_info;
 193         /**
 194          * Packet reorder array. "true" indicates that DMA device
 195          * completes all copies for the packet.
 196          *
 197          * Note that this array could be written by multiple threads
 198          * simultaneously. For example, in the case of thread0 and
 199          * thread1 RX packets from NIC and then enqueue packets to
 200          * vring0 and vring1 with own DMA device DMA0 and DMA1, it's
 201          * possible for thread0 to get completed copies belonging to
 202          * vring1 from DMA0, while thread0 is calling rte_vhost_poll
 203          * _enqueue_completed() for vring0 and thread1 is calling
 204          * rte_vhost_submit_enqueue_burst() for vring1. In this case,
 205          * vq->access_lock cannot protect pkts_cmpl_flag of vring1.
 206          *
 207          * However, since offloading is per-packet basis, each packet
 208          * flag will only be written by one thread. And single byte
 209          * write is atomic, so no lock for pkts_cmpl_flag is needed.
 210          */
 211         bool *pkts_cmpl_flag;
 212         uint16_t pkts_idx;
 213         uint16_t pkts_inflight_n;
 214         union {
 215                 struct vring_used_elem  *descs_split;
 216                 struct vring_used_elem_packed *buffers_packed;
 217         };
 218         union {
 219                 uint16_t desc_idx_split;
 220                 uint16_t buffer_idx_packed;
 221         };
 222         union {
 223                 uint16_t last_desc_idx_split;
 224                 uint16_t last_buffer_idx_packed;
 225         };
 226 };
 227
 228 /**
 229  * Structure contains variables relevant to RX/TX virtqueues.
 230  */
 231 struct vhost_virtqueue {
 232         union {
 233                 struct vring_desc       *desc;
 234                 struct vring_packed_desc   *desc_packed;
 235         };
 236         union {
 237                 struct vring_avail      *avail;
 238                 struct vring_packed_desc_event *driver_event;
 239         };
 240         union {
 241                 struct vring_used       *used;
 242                 struct vring_packed_desc_event *device_event;
 243         };
 244         uint16_t                size;
 245
 246         uint16_t                last_avail_idx;
 247         uint16_t                last_used_idx;
 248         /* Last used index we notify to front end. */
 249         uint16_t                signalled_used;
 250         bool                    signalled_used_valid;
 251 #define VIRTIO_INVALID_EVENTFD          (-1)
 252 #define VIRTIO_UNINITIALIZED_EVENTFD    (-2)
 253
 254         bool                    enabled;
 255         bool                    access_ok;
 256         bool                    ready;
 257
 258         rte_spinlock_t          access_lock;
 259
 260
 261         union {
 262                 struct vring_used_elem  *shadow_used_split;
 263                 struct vring_used_elem_packed *shadow_used_packed;
 264         };
 265         uint16_t                shadow_used_idx;
 266         /* Record packed ring enqueue latest desc cache aligned index */
 267         uint16_t                shadow_aligned_idx;
 268         /* Record packed ring first dequeue desc index */
 269         uint16_t                shadow_last_used_idx;
 270
 271         uint16_t                batch_copy_nb_elems;
 272         struct batch_copy_elem  *batch_copy_elems;
 273         int                     numa_node;
 274         bool                    used_wrap_counter;
 275         bool                    avail_wrap_counter;
 276
 277         /* Physical address of used ring, for logging */
 278         uint16_t                log_cache_nb_elem;
 279         uint64_t                log_guest_addr;
 280         struct log_cache_entry  *log_cache;
 281
 282         rte_rwlock_t    iotlb_lock;
 283         rte_rwlock_t    iotlb_pending_lock;
 284         struct rte_mempool *iotlb_pool;
 285         TAILQ_HEAD(, vhost_iotlb_entry) iotlb_list;
 286         TAILQ_HEAD(, vhost_iotlb_entry) iotlb_pending_list;
 287         int                             iotlb_cache_nr;
 288
 289         /* Used to notify the guest (trigger interrupt) */
 290         int                     callfd;
 291         /* Currently unused as polling mode is enabled */
 292         int                     kickfd;
 293
 294         /* inflight share memory info */
 295         union {
 296                 struct rte_vhost_inflight_info_split *inflight_split;
 297                 struct rte_vhost_inflight_info_packed *inflight_packed;
 298         };
 299         struct rte_vhost_resubmit_info *resubmit_inflight;
 300         uint64_t                global_counter;
 301
 302         struct vhost_async      *async;
 303
 304         int                     notif_enable;
 305 #define VIRTIO_UNINITIALIZED_NOTIF      (-1)
 306
 307         struct vhost_vring_addr ring_addrs;
 308 } __rte_cache_aligned;
 309
 310 /* Virtio device status as per Virtio specification */
 311 #define VIRTIO_DEVICE_STATUS_RESET              0x00
 312 #define VIRTIO_DEVICE_STATUS_ACK                0x01
 313 #define VIRTIO_DEVICE_STATUS_DRIVER             0x02
 314 #define VIRTIO_DEVICE_STATUS_DRIVER_OK          0x04
 315 #define VIRTIO_DEVICE_STATUS_FEATURES_OK        0x08
 316 #define VIRTIO_DEVICE_STATUS_DEV_NEED_RESET     0x40
 317 #define VIRTIO_DEVICE_STATUS_FAILED             0x80
 318
 319 #define VHOST_MAX_VRING                 0x100
 320 #define VHOST_MAX_QUEUE_PAIRS           0x80
 321
 322 /* Declare IOMMU related bits for older kernels */
 323 #ifndef VIRTIO_F_IOMMU_PLATFORM
 324
 325 #define VIRTIO_F_IOMMU_PLATFORM 33
 326
 327 struct vhost_iotlb_msg {
 328         __u64 iova;
 329         __u64 size;
 330         __u64 uaddr;
 331 #define VHOST_ACCESS_RO      0x1
 332 #define VHOST_ACCESS_WO      0x2
 333 #define VHOST_ACCESS_RW      0x3
 334         __u8 perm;
 335 #define VHOST_IOTLB_MISS           1
 336 #define VHOST_IOTLB_UPDATE         2
 337 #define VHOST_IOTLB_INVALIDATE     3
 338 #define VHOST_IOTLB_ACCESS_FAIL    4
 339         __u8 type;
 340 };
 341
 342 #define VHOST_IOTLB_MSG 0x1
 343
 344 struct vhost_msg {
 345         int type;
 346         union {
 347                 struct vhost_iotlb_msg iotlb;
 348                 __u8 padding[64];
 349         };
 350 };
 351 #endif
 352
 353 /*
 354  * Define virtio 1.0 for older kernels
 355  */
 356 #ifndef VIRTIO_F_VERSION_1
 357  #define VIRTIO_F_VERSION_1 32
 358 #endif
 359
 360 /* Declare packed ring related bits for older kernels */
 361 #ifndef VIRTIO_F_RING_PACKED
 362
 363 #define VIRTIO_F_RING_PACKED 34
 364
 365 struct vring_packed_desc {
 366         uint64_t addr;
 367         uint32_t len;
 368         uint16_t id;
 369         uint16_t flags;
 370 };
 371
 372 struct vring_packed_desc_event {
 373         uint16_t off_wrap;
 374         uint16_t flags;
 375 };
 376 #endif
 377
 378 /*
 379  * Declare below packed ring defines unconditionally
 380  * as Kernel header might use different names.
 381  */
 382 #define VRING_DESC_F_AVAIL      (1ULL << 7)
 383 #define VRING_DESC_F_USED       (1ULL << 15)
 384
 385 #define VRING_EVENT_F_ENABLE 0x0
 386 #define VRING_EVENT_F_DISABLE 0x1
 387 #define VRING_EVENT_F_DESC 0x2
 388
 389 /*
 390  * Available and used descs are in same order
 391  */
 392 #ifndef VIRTIO_F_IN_ORDER
 393 #define VIRTIO_F_IN_ORDER      35
 394 #endif
 395
 396 /* Features supported by this builtin vhost-user net driver. */
 397 #define VIRTIO_NET_SUPPORTED_FEATURES ((1ULL << VIRTIO_NET_F_MRG_RXBUF) | \
 398                                 (1ULL << VIRTIO_F_ANY_LAYOUT) | \
 399                                 (1ULL << VIRTIO_NET_F_CTRL_VQ) | \
 400                                 (1ULL << VIRTIO_NET_F_CTRL_RX) | \
 401                                 (1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) | \
 402                                 (1ULL << VIRTIO_NET_F_MQ)      | \
 403                                 (1ULL << VIRTIO_F_VERSION_1)   | \
 404                                 (1ULL << VHOST_F_LOG_ALL)      | \
 405                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES) | \
 406                                 (1ULL << VIRTIO_NET_F_GSO) | \
 407                                 (1ULL << VIRTIO_NET_F_HOST_TSO4) | \
 408                                 (1ULL << VIRTIO_NET_F_HOST_TSO6) | \
 409                                 (1ULL << VIRTIO_NET_F_HOST_UFO) | \
 410                                 (1ULL << VIRTIO_NET_F_HOST_ECN) | \
 411                                 (1ULL << VIRTIO_NET_F_CSUM)    | \
 412                                 (1ULL << VIRTIO_NET_F_GUEST_CSUM) | \
 413                                 (1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
 414                                 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
 415                                 (1ULL << VIRTIO_NET_F_GUEST_UFO) | \
 416                                 (1ULL << VIRTIO_NET_F_GUEST_ECN) | \
 417                                 (1ULL << VIRTIO_RING_F_INDIRECT_DESC) | \
 418                                 (1ULL << VIRTIO_RING_F_EVENT_IDX) | \
 419                                 (1ULL << VIRTIO_NET_F_MTU)  | \
 420                                 (1ULL << VIRTIO_F_IN_ORDER) | \
 421                                 (1ULL << VIRTIO_F_IOMMU_PLATFORM) | \
 422                                 (1ULL << VIRTIO_F_RING_PACKED))
 423
 424
 425 struct guest_page {
 426         uint64_t guest_phys_addr;
 427         uint64_t host_iova;
 428         uint64_t host_user_addr;
 429         uint64_t size;
 430 };
 431
 432 struct inflight_mem_info {
 433         int             fd;
 434         void            *addr;
 435         uint64_t        size;
 436 };
 437
 438 /**
 439  * Device structure contains all configuration information relating
 440  * to the device.
 441  */
 442 struct virtio_net {
 443         /* Frontend (QEMU) memory and memory region information */
 444         struct rte_vhost_memory *mem;
 445         uint64_t                features;
 446         uint64_t                protocol_features;
 447         int                     vid;
 448         uint32_t                flags;
 449         uint16_t                vhost_hlen;
 450         /* to tell if we need broadcast rarp packet */
 451         int16_t                 broadcast_rarp;
 452         uint32_t                nr_vring;
 453         int                     async_copy;
 454
 455         int                     extbuf;
 456         int                     linearbuf;
 457         struct vhost_virtqueue  *virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
 458         struct inflight_mem_info *inflight_info;
 459 #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
 460         char                    ifname[IF_NAME_SZ];
 461         uint64_t                log_size;
 462         uint64_t                log_base;
 463         uint64_t                log_addr;
 464         struct rte_ether_addr   mac;
 465         uint16_t                mtu;
 466         uint8_t                 status;
 467
 468         struct rte_vhost_device_ops const *notify_ops;
 469
 470         uint32_t                nr_guest_pages;
 471         uint32_t                max_guest_pages;
 472         struct guest_page       *guest_pages;
 473
 474         int                     slave_req_fd;
 475         rte_spinlock_t          slave_req_lock;
 476
 477         int                     postcopy_ufd;
 478         int                     postcopy_listening;
 479
 480         struct rte_vdpa_device *vdpa_dev;
 481
 482         /* context data for the external message handlers */
 483         void                    *extern_data;
 484         /* pre and post vhost user message handlers for the device */
 485         struct rte_vhost_user_extern_ops extern_ops;
 486 } __rte_cache_aligned;
 487
 488 static __rte_always_inline bool
 489 vq_is_packed(struct virtio_net *dev)
 490 {
 491         return dev->features & (1ull << VIRTIO_F_RING_PACKED);
 492 }
 493
 494 static inline bool
 495 desc_is_avail(struct vring_packed_desc *desc, bool wrap_counter)
 496 {
 497         uint16_t flags = __atomic_load_n(&desc->flags, __ATOMIC_ACQUIRE);
 498
 499         return wrap_counter == !!(flags & VRING_DESC_F_AVAIL) &&
 500                 wrap_counter != !!(flags & VRING_DESC_F_USED);
 501 }
 502
 503 static inline void
 504 vq_inc_last_used_packed(struct vhost_virtqueue *vq, uint16_t num)
 505 {
 506         vq->last_used_idx += num;
 507         if (vq->last_used_idx >= vq->size) {
 508                 vq->used_wrap_counter ^= 1;
 509                 vq->last_used_idx -= vq->size;
 510         }
 511 }
 512
 513 static inline void
 514 vq_inc_last_avail_packed(struct vhost_virtqueue *vq, uint16_t num)
 515 {
 516         vq->last_avail_idx += num;
 517         if (vq->last_avail_idx >= vq->size) {
 518                 vq->avail_wrap_counter ^= 1;
 519                 vq->last_avail_idx -= vq->size;
 520         }
 521 }
 522
 523 void __vhost_log_cache_write(struct virtio_net *dev,
 524                 struct vhost_virtqueue *vq,
 525                 uint64_t addr, uint64_t len);
 526 void __vhost_log_cache_write_iova(struct virtio_net *dev,
 527                 struct vhost_virtqueue *vq,
 528                 uint64_t iova, uint64_t len);
 529 void __vhost_log_cache_sync(struct virtio_net *dev,
 530                 struct vhost_virtqueue *vq);
 531 void __vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len);
 532 void __vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 533                             uint64_t iova, uint64_t len);
 534
 535 static __rte_always_inline void
 536 vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
 537 {
 538         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL)))
 539                 __vhost_log_write(dev, addr, len);
 540 }
 541
 542 static __rte_always_inline void
 543 vhost_log_cache_sync(struct virtio_net *dev, struct vhost_virtqueue *vq)
 544 {
 545         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL)))
 546                 __vhost_log_cache_sync(dev, vq);
 547 }
 548
 549 static __rte_always_inline void
 550 vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
 551                         uint64_t addr, uint64_t len)
 552 {
 553         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL)))
 554                 __vhost_log_cache_write(dev, vq, addr, len);
 555 }
 556
 557 static __rte_always_inline void
 558 vhost_log_cache_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,
 559                         uint64_t offset, uint64_t len)
 560 {
 561         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL))) {
 562                 if (unlikely(vq->log_guest_addr == 0))
 563                         return;
 564                 __vhost_log_cache_write(dev, vq, vq->log_guest_addr + offset,
 565                                         len);
 566         }
 567 }
 568
 569 static __rte_always_inline void
 570 vhost_log_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,
 571                      uint64_t offset, uint64_t len)
 572 {
 573         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL))) {
 574                 if (unlikely(vq->log_guest_addr == 0))
 575                         return;
 576                 __vhost_log_write(dev, vq->log_guest_addr + offset, len);
 577         }
 578 }
 579
 580 static __rte_always_inline void
 581 vhost_log_cache_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 582                            uint64_t iova, uint64_t len)
 583 {
 584         if (likely(!(dev->features & (1ULL << VHOST_F_LOG_ALL))))
 585                 return;
 586
 587         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 588                 __vhost_log_cache_write_iova(dev, vq, iova, len);
 589         else
 590                 __vhost_log_cache_write(dev, vq, iova, len);
 591 }
 592
 593 static __rte_always_inline void
 594 vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 595                            uint64_t iova, uint64_t len)
 596 {
 597         if (likely(!(dev->features & (1ULL << VHOST_F_LOG_ALL))))
 598                 return;
 599
 600         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 601                 __vhost_log_write_iova(dev, vq, iova, len);
 602         else
 603                 __vhost_log_write(dev, iova, len);
 604 }
 605
 606 extern int vhost_config_log_level;
 607 extern int vhost_data_log_level;
 608
 609 #define VHOST_LOG_CONFIG(level, fmt, args...)                   \
 610         rte_log(RTE_LOG_ ## level, vhost_config_log_level,      \
 611                 "VHOST_CONFIG: " fmt, ##args)
 612
 613 #define VHOST_LOG_DATA(level, fmt, args...) \
 614         (void)((RTE_LOG_ ## level <= RTE_LOG_DP_LEVEL) ?        \
 615          rte_log(RTE_LOG_ ## level,  vhost_data_log_level,      \
 616                 "VHOST_DATA : " fmt, ##args) :                  \
 617          0)
 618
 619 #ifdef RTE_LIBRTE_VHOST_DEBUG
 620 #define VHOST_MAX_PRINT_BUFF 6072
 621 #define PRINT_PACKET(device, addr, size, header) do { \
 622         char *pkt_addr = (char *)(addr); \
 623         unsigned int index; \
 624         char packet[VHOST_MAX_PRINT_BUFF]; \
 625         \
 626         if ((header)) \
 627                 snprintf(packet, VHOST_MAX_PRINT_BUFF, "(%d) Header size %d: ", (device->vid), (size)); \
 628         else \
 629                 snprintf(packet, VHOST_MAX_PRINT_BUFF, "(%d) Packet size %d: ", (device->vid), (size)); \
 630         for (index = 0; index < (size); index++) { \
 631                 snprintf(packet + strnlen(packet, VHOST_MAX_PRINT_BUFF), VHOST_MAX_PRINT_BUFF - strnlen(packet, VHOST_MAX_PRINT_BUFF), \
 632                         "%02hhx ", pkt_addr[index]); \
 633         } \
 634         snprintf(packet + strnlen(packet, VHOST_MAX_PRINT_BUFF), VHOST_MAX_PRINT_BUFF - strnlen(packet, VHOST_MAX_PRINT_BUFF), "\n"); \
 635         \
 636         VHOST_LOG_DATA(DEBUG, "%s", packet); \
 637 } while (0)
 638 #else
 639 #define PRINT_PACKET(device, addr, size, header) do {} while (0)
 640 #endif
 641
 642 extern struct virtio_net *vhost_devices[RTE_MAX_VHOST_DEVICE];
 643
 644 #define VHOST_BINARY_SEARCH_THRESH 256
 645
 646 static __rte_always_inline int guest_page_addrcmp(const void *p1,
 647                                                 const void *p2)
 648 {
 649         const struct guest_page *page1 = (const struct guest_page *)p1;
 650         const struct guest_page *page2 = (const struct guest_page *)p2;
 651
 652         if (page1->guest_phys_addr > page2->guest_phys_addr)
 653                 return 1;
 654         if (page1->guest_phys_addr < page2->guest_phys_addr)
 655                 return -1;
 656
 657         return 0;
 658 }
 659
 660 static __rte_always_inline int guest_page_rangecmp(const void *p1, const void *p2)
 661 {
 662         const struct guest_page *page1 = (const struct guest_page *)p1;
 663         const struct guest_page *page2 = (const struct guest_page *)p2;
 664
 665         if (page1->guest_phys_addr >= page2->guest_phys_addr) {
 666                 if (page1->guest_phys_addr < page2->guest_phys_addr + page2->size)
 667                         return 0;
 668                 else
 669                         return 1;
 670         } else
 671                 return -1;
 672 }
 673
 674 static __rte_always_inline rte_iova_t
 675 gpa_to_first_hpa(struct virtio_net *dev, uint64_t gpa,
 676         uint64_t gpa_size, uint64_t *hpa_size)
 677 {
 678         uint32_t i;
 679         struct guest_page *page;
 680         struct guest_page key;
 681
 682         *hpa_size = gpa_size;
 683         if (dev->nr_guest_pages >= VHOST_BINARY_SEARCH_THRESH) {
 684                 key.guest_phys_addr = gpa;
 685                 page = bsearch(&key, dev->guest_pages, dev->nr_guest_pages,
 686                                sizeof(struct guest_page), guest_page_rangecmp);
 687                 if (page) {
 688                         if (gpa + gpa_size <=
 689                                         page->guest_phys_addr + page->size) {
 690                                 return gpa - page->guest_phys_addr +
 691                                         page->host_iova;
 692                         } else if (gpa < page->guest_phys_addr +
 693                                                 page->size) {
 694                                 *hpa_size = page->guest_phys_addr +
 695                                         page->size - gpa;
 696                                 return gpa - page->guest_phys_addr +
 697                                         page->host_iova;
 698                         }
 699                 }
 700         } else {
 701                 for (i = 0; i < dev->nr_guest_pages; i++) {
 702                         page = &dev->guest_pages[i];
 703
 704                         if (gpa >= page->guest_phys_addr) {
 705                                 if (gpa + gpa_size <=
 706                                         page->guest_phys_addr + page->size) {
 707                                         return gpa - page->guest_phys_addr +
 708                                                 page->host_iova;
 709                                 } else if (gpa < page->guest_phys_addr +
 710                                                         page->size) {
 711                                         *hpa_size = page->guest_phys_addr +
 712                                                 page->size - gpa;
 713                                         return gpa - page->guest_phys_addr +
 714                                                 page->host_iova;
 715                                 }
 716                         }
 717                 }
 718         }
 719
 720         *hpa_size = 0;
 721         return 0;
 722 }
 723
 724 /* Convert guest physical address to host physical address */
 725 static __rte_always_inline rte_iova_t
 726 gpa_to_hpa(struct virtio_net *dev, uint64_t gpa, uint64_t size)
 727 {
 728         rte_iova_t hpa;
 729         uint64_t hpa_size;
 730
 731         hpa = gpa_to_first_hpa(dev, gpa, size, &hpa_size);
 732         return hpa_size == size ? hpa : 0;
 733 }
 734
 735 static __rte_always_inline uint64_t
 736 hva_to_gpa(struct virtio_net *dev, uint64_t vva, uint64_t len)
 737 {
 738         struct rte_vhost_mem_region *r;
 739         uint32_t i;
 740
 741         if (unlikely(!dev || !dev->mem))
 742                 return 0;
 743
 744         for (i = 0; i < dev->mem->nregions; i++) {
 745                 r = &dev->mem->regions[i];
 746
 747                 if (vva >= r->host_user_addr &&
 748                     vva + len <  r->host_user_addr + r->size) {
 749                         return r->guest_phys_addr + vva - r->host_user_addr;
 750                 }
 751         }
 752         return 0;
 753 }
 754
 755 static __rte_always_inline struct virtio_net *
 756 get_device(int vid)
 757 {
 758         struct virtio_net *dev = vhost_devices[vid];
 759
 760         if (unlikely(!dev)) {
 761                 VHOST_LOG_CONFIG(ERR,
 762                         "(%d) device not found.\n", vid);
 763         }
 764
 765         return dev;
 766 }
 767
 768 int vhost_new_device(void);
 769 void cleanup_device(struct virtio_net *dev, int destroy);
 770 void reset_device(struct virtio_net *dev);
 771 void vhost_destroy_device(int);
 772 void vhost_destroy_device_notify(struct virtio_net *dev);
 773
 774 void cleanup_vq(struct vhost_virtqueue *vq, int destroy);
 775 void cleanup_vq_inflight(struct virtio_net *dev, struct vhost_virtqueue *vq);
 776 void free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq);
 777
 778 int alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx);
 779
 780 void vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *dev);
 781
 782 void vhost_set_ifname(int, const char *if_name, unsigned int if_len);
 783 void vhost_setup_virtio_net(int vid, bool enable, bool legacy_ol_flags);
 784 void vhost_enable_extbuf(int vid);
 785 void vhost_enable_linearbuf(int vid);
 786 int vhost_enable_guest_notification(struct virtio_net *dev,
 787                 struct vhost_virtqueue *vq, int enable);
 788
 789 struct rte_vhost_device_ops const *vhost_driver_callback_get(const char *path);
 790
 791 /*
 792  * Backend-specific cleanup.
 793  *
 794  * TODO: fix it; we have one backend now
 795  */
 796 void vhost_backend_cleanup(struct virtio_net *dev);
 797
 798 uint64_t __vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 799                         uint64_t iova, uint64_t *len, uint8_t perm);
 800 void *vhost_alloc_copy_ind_table(struct virtio_net *dev,
 801                         struct vhost_virtqueue *vq,
 802                         uint64_t desc_addr, uint64_t desc_len);
 803 int vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq);
 804 uint64_t translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
 805                 uint64_t log_addr);
 806 void vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq);
 807
 808 static __rte_always_inline uint64_t
 809 vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 810                         uint64_t iova, uint64_t *len, uint8_t perm)
 811 {
 812         if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
 813                 return rte_vhost_va_from_guest_pa(dev->mem, iova, len);
 814
 815         return __vhost_iova_to_vva(dev, vq, iova, len, perm);
 816 }
 817
 818 #define vhost_avail_event(vr) \
 819         (*(volatile uint16_t*)&(vr)->used->ring[(vr)->size])
 820 #define vhost_used_event(vr) \
 821         (*(volatile uint16_t*)&(vr)->avail->ring[(vr)->size])
 822
 823 /*
 824  * The following is used with VIRTIO_RING_F_EVENT_IDX.
 825  * Assuming a given event_idx value from the other size, if we have
 826  * just incremented index from old to new_idx, should we trigger an
 827  * event?
 828  */
 829 static __rte_always_inline int
 830 vhost_need_event(uint16_t event_idx, uint16_t new_idx, uint16_t old)
 831 {
 832         return (uint16_t)(new_idx - event_idx - 1) < (uint16_t)(new_idx - old);
 833 }
 834
 835 static __rte_always_inline void
 836 vhost_vring_call_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
 837 {
 838         /* Flush used->idx update before we read avail->flags. */
 839         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
 840
 841         /* Don't kick guest if we don't reach index specified by guest. */
 842         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
 843                 uint16_t old = vq->signalled_used;
 844                 uint16_t new = vq->last_used_idx;
 845                 bool signalled_used_valid = vq->signalled_used_valid;
 846
 847                 vq->signalled_used = new;
 848                 vq->signalled_used_valid = true;
 849
 850                 VHOST_LOG_DATA(DEBUG, "%s: used_event_idx=%d, old=%d, new=%d\n",
 851                         __func__,
 852                         vhost_used_event(vq),
 853                         old, new);
 854
 855                 if ((vhost_need_event(vhost_used_event(vq), new, old) &&
 856                                         (vq->callfd >= 0)) ||
 857                                 unlikely(!signalled_used_valid)) {
 858                         eventfd_write(vq->callfd, (eventfd_t) 1);
 859                         if (dev->notify_ops->guest_notified)
 860                                 dev->notify_ops->guest_notified(dev->vid);
 861                 }
 862         } else {
 863                 /* Kick the guest if necessary. */
 864                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 865                                 && (vq->callfd >= 0)) {
 866                         eventfd_write(vq->callfd, (eventfd_t)1);
 867                         if (dev->notify_ops->guest_notified)
 868                                 dev->notify_ops->guest_notified(dev->vid);
 869                 }
 870         }
 871 }
 872
 873 static __rte_always_inline void
 874 vhost_vring_call_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
 875 {
 876         uint16_t old, new, off, off_wrap;
 877         bool signalled_used_valid, kick = false;
 878
 879         /* Flush used desc update. */
 880         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
 881
 882         if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
 883                 if (vq->driver_event->flags !=
 884                                 VRING_EVENT_F_DISABLE)
 885                         kick = true;
 886                 goto kick;
 887         }
 888
 889         old = vq->signalled_used;
 890         new = vq->last_used_idx;
 891         vq->signalled_used = new;
 892         signalled_used_valid = vq->signalled_used_valid;
 893         vq->signalled_used_valid = true;
 894
 895         if (vq->driver_event->flags != VRING_EVENT_F_DESC) {
 896                 if (vq->driver_event->flags != VRING_EVENT_F_DISABLE)
 897                         kick = true;
 898                 goto kick;
 899         }
 900
 901         if (unlikely(!signalled_used_valid)) {
 902                 kick = true;
 903                 goto kick;
 904         }
 905
 906         rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
 907
 908         off_wrap = vq->driver_event->off_wrap;
 909         off = off_wrap & ~(1 << 15);
 910
 911         if (new <= old)
 912                 old -= vq->size;
 913
 914         if (vq->used_wrap_counter != off_wrap >> 15)
 915                 off -= vq->size;
 916
 917         if (vhost_need_event(off, new, old))
 918                 kick = true;
 919 kick:
 920         if (kick) {
 921                 eventfd_write(vq->callfd, (eventfd_t)1);
 922                 if (dev->notify_ops->guest_notified)
 923                         dev->notify_ops->guest_notified(dev->vid);
 924         }
 925 }
 926
 927 static __rte_always_inline void
 928 free_ind_table(void *idesc)
 929 {
 930         rte_free(idesc);
 931 }
 932
 933 static __rte_always_inline void
 934 restore_mbuf(struct rte_mbuf *m)
 935 {
 936         uint32_t mbuf_size, priv_size;
 937
 938         while (m) {
 939                 priv_size = rte_pktmbuf_priv_size(m->pool);
 940                 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
 941                 /* start of buffer is after mbuf structure and priv data */
 942
 943                 m->buf_addr = (char *)m + mbuf_size;
 944                 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
 945                 m = m->next;
 946         }
 947 }
 948
 949 static __rte_always_inline bool
 950 mbuf_is_consumed(struct rte_mbuf *m)
 951 {
 952         while (m) {
 953                 if (rte_mbuf_refcnt_read(m) > 1)
 954                         return false;
 955                 m = m->next;
 956         }
 957
 958         return true;
 959 }
 960
 961 #endif /* _VHOST_NET_CDEV_H_ */