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 /*  Used to indicate the application has requested statistics collection */
  41 #define VIRTIO_DEV_STATS_ENABLED ((uint32_t)1 << 6)
  42
  43 /* Backend value set by guest. */
  44 #define VIRTIO_DEV_STOPPED -1
  45
  46 #define BUF_VECTOR_MAX 256
  47
  48 #define VHOST_LOG_CACHE_NR 32
  49
  50 #define MAX_PKT_BURST 32
  51
  52 #define VHOST_MAX_ASYNC_IT (MAX_PKT_BURST)
  53 #define VHOST_MAX_ASYNC_VEC 2048
  54 #define VIRTIO_MAX_RX_PKTLEN 9728U
  55 #define VHOST_DMA_MAX_COPY_COMPLETE ((VIRTIO_MAX_RX_PKTLEN / RTE_MBUF_DEFAULT_DATAROOM) \
  56                 * MAX_PKT_BURST)
  57
  58 #define PACKED_DESC_ENQUEUE_USED_FLAG(w)        \
  59         ((w) ? (VRING_DESC_F_AVAIL | VRING_DESC_F_USED | VRING_DESC_F_WRITE) : \
  60                 VRING_DESC_F_WRITE)
  61 #define PACKED_DESC_DEQUEUE_USED_FLAG(w)        \
  62         ((w) ? (VRING_DESC_F_AVAIL | VRING_DESC_F_USED) : 0x0)
  63 #define PACKED_DESC_SINGLE_DEQUEUE_FLAG (VRING_DESC_F_NEXT | \
  64                                          VRING_DESC_F_INDIRECT)
  65
  66 #define PACKED_BATCH_SIZE (RTE_CACHE_LINE_SIZE / \
  67                             sizeof(struct vring_packed_desc))
  68 #define PACKED_BATCH_MASK (PACKED_BATCH_SIZE - 1)
  69
  70 #ifdef VHOST_GCC_UNROLL_PRAGMA
  71 #define vhost_for_each_try_unroll(iter, val, size) _Pragma("GCC unroll 4") \
  72         for (iter = val; iter < size; iter++)
  73 #endif
  74
  75 #ifdef VHOST_CLANG_UNROLL_PRAGMA
  76 #define vhost_for_each_try_unroll(iter, val, size) _Pragma("unroll 4") \
  77         for (iter = val; iter < size; iter++)
  78 #endif
  79
  80 #ifdef VHOST_ICC_UNROLL_PRAGMA
  81 #define vhost_for_each_try_unroll(iter, val, size) _Pragma("unroll (4)") \
  82         for (iter = val; iter < size; iter++)
  83 #endif
  84
  85 #ifndef vhost_for_each_try_unroll
  86 #define vhost_for_each_try_unroll(iter, val, num) \
  87         for (iter = val; iter < num; iter++)
  88 #endif
  89
  90 /**
  91  * Structure contains buffer address, length and descriptor index
  92  * from vring to do scatter RX.
  93  */
  94 struct buf_vector {
  95         uint64_t buf_iova;
  96         uint64_t buf_addr;
  97         uint32_t buf_len;
  98         uint32_t desc_idx;
  99 };
 100
 101 /*
 102  * Structure contains the info for each batched memory copy.
 103  */
 104 struct batch_copy_elem {
 105         void *dst;
 106         void *src;
 107         uint32_t len;
 108         uint64_t log_addr;
 109 };
 110
 111 /*
 112  * Structure that contains the info for batched dirty logging.
 113  */
 114 struct log_cache_entry {
 115         uint32_t offset;
 116         unsigned long val;
 117 };
 118
 119 struct vring_used_elem_packed {
 120         uint16_t id;
 121         uint16_t flags;
 122         uint32_t len;
 123         uint32_t count;
 124 };
 125
 126 /**
 127  * Virtqueue statistics
 128  */
 129 struct virtqueue_stats {
 130         uint64_t packets;
 131         uint64_t bytes;
 132         uint64_t multicast;
 133         uint64_t broadcast;
 134         /* Size bins in array as RFC 2819, undersized [0], 64 [1], etc */
 135         uint64_t size_bins[8];
 136 };
 137
 138 /**
 139  * iovec
 140  */
 141 struct vhost_iovec {
 142         void *src_addr;
 143         void *dst_addr;
 144         size_t len;
 145 };
 146
 147 /**
 148  * iovec iterator
 149  */
 150 struct vhost_iov_iter {
 151         /** pointer to the iovec array */
 152         struct vhost_iovec *iov;
 153         /** number of iovec in this iterator */
 154         unsigned long nr_segs;
 155 };
 156
 157 struct async_dma_vchan_info {
 158         /* circular array to track if packet copy completes */
 159         bool **pkts_cmpl_flag_addr;
 160
 161         /* max elements in 'pkts_cmpl_flag_addr' */
 162         uint16_t ring_size;
 163         /* ring index mask for 'pkts_cmpl_flag_addr' */
 164         uint16_t ring_mask;
 165
 166         /**
 167          * DMA virtual channel lock. Although it is able to bind DMA
 168          * virtual channels to data plane threads, vhost control plane
 169          * thread could call data plane functions too, thus causing
 170          * DMA device contention.
 171          *
 172          * For example, in VM exit case, vhost control plane thread needs
 173          * to clear in-flight packets before disable vring, but there could
 174          * be anotther data plane thread is enqueuing packets to the same
 175          * vring with the same DMA virtual channel. As dmadev PMD functions
 176          * are lock-free, the control plane and data plane threads could
 177          * operate the same DMA virtual channel at the same time.
 178          */
 179         rte_spinlock_t dma_lock;
 180 };
 181
 182 struct async_dma_info {
 183         struct async_dma_vchan_info *vchans;
 184         /* number of registered virtual channels */
 185         uint16_t nr_vchans;
 186 };
 187
 188 extern struct async_dma_info dma_copy_track[RTE_DMADEV_DEFAULT_MAX];
 189
 190 /**
 191  * inflight async packet information
 192  */
 193 struct async_inflight_info {
 194         struct rte_mbuf *mbuf;
 195         uint16_t descs; /* num of descs inflight */
 196         uint16_t nr_buffers; /* num of buffers inflight for packed ring */
 197 };
 198
 199 struct vhost_async {
 200         struct vhost_iov_iter iov_iter[VHOST_MAX_ASYNC_IT];
 201         struct vhost_iovec iovec[VHOST_MAX_ASYNC_VEC];
 202         uint16_t iter_idx;
 203         uint16_t iovec_idx;
 204
 205         /* data transfer status */
 206         struct async_inflight_info *pkts_info;
 207         /**
 208          * Packet reorder array. "true" indicates that DMA device
 209          * completes all copies for the packet.
 210          *
 211          * Note that this array could be written by multiple threads
 212          * simultaneously. For example, in the case of thread0 and
 213          * thread1 RX packets from NIC and then enqueue packets to
 214          * vring0 and vring1 with own DMA device DMA0 and DMA1, it's
 215          * possible for thread0 to get completed copies belonging to
 216          * vring1 from DMA0, while thread0 is calling rte_vhost_poll
 217          * _enqueue_completed() for vring0 and thread1 is calling
 218          * rte_vhost_submit_enqueue_burst() for vring1. In this case,
 219          * vq->access_lock cannot protect pkts_cmpl_flag of vring1.
 220          *
 221          * However, since offloading is per-packet basis, each packet
 222          * flag will only be written by one thread. And single byte
 223          * write is atomic, so no lock for pkts_cmpl_flag is needed.
 224          */
 225         bool *pkts_cmpl_flag;
 226         uint16_t pkts_idx;
 227         uint16_t pkts_inflight_n;
 228         union {
 229                 struct vring_used_elem  *descs_split;
 230                 struct vring_used_elem_packed *buffers_packed;
 231         };
 232         union {
 233                 uint16_t desc_idx_split;
 234                 uint16_t buffer_idx_packed;
 235         };
 236         union {
 237                 uint16_t last_desc_idx_split;
 238                 uint16_t last_buffer_idx_packed;
 239         };
 240 };
 241
 242 /**
 243  * Structure contains variables relevant to RX/TX virtqueues.
 244  */
 245 struct vhost_virtqueue {
 246         union {
 247                 struct vring_desc       *desc;
 248                 struct vring_packed_desc   *desc_packed;
 249         };
 250         union {
 251                 struct vring_avail      *avail;
 252                 struct vring_packed_desc_event *driver_event;
 253         };
 254         union {
 255                 struct vring_used       *used;
 256                 struct vring_packed_desc_event *device_event;
 257         };
 258         uint16_t                size;
 259
 260         uint16_t                last_avail_idx;
 261         uint16_t                last_used_idx;
 262         /* Last used index we notify to front end. */
 263         uint16_t                signalled_used;
 264         bool                    signalled_used_valid;
 265 #define VIRTIO_INVALID_EVENTFD          (-1)
 266 #define VIRTIO_UNINITIALIZED_EVENTFD    (-2)
 267
 268         bool                    enabled;
 269         bool                    access_ok;
 270         bool                    ready;
 271
 272         rte_spinlock_t          access_lock;
 273
 274
 275         union {
 276                 struct vring_used_elem  *shadow_used_split;
 277                 struct vring_used_elem_packed *shadow_used_packed;
 278         };
 279         uint16_t                shadow_used_idx;
 280         /* Record packed ring enqueue latest desc cache aligned index */
 281         uint16_t                shadow_aligned_idx;
 282         /* Record packed ring first dequeue desc index */
 283         uint16_t                shadow_last_used_idx;
 284
 285         uint16_t                batch_copy_nb_elems;
 286         struct batch_copy_elem  *batch_copy_elems;
 287         int                     numa_node;
 288         bool                    used_wrap_counter;
 289         bool                    avail_wrap_counter;
 290
 291         /* Physical address of used ring, for logging */
 292         uint16_t                log_cache_nb_elem;
 293         uint64_t                log_guest_addr;
 294         struct log_cache_entry  *log_cache;
 295
 296         rte_rwlock_t    iotlb_lock;
 297         rte_rwlock_t    iotlb_pending_lock;
 298         struct rte_mempool *iotlb_pool;
 299         TAILQ_HEAD(, vhost_iotlb_entry) iotlb_list;
 300         TAILQ_HEAD(, vhost_iotlb_entry) iotlb_pending_list;
 301         int                             iotlb_cache_nr;
 302
 303         /* Used to notify the guest (trigger interrupt) */
 304         int                     callfd;
 305         /* Currently unused as polling mode is enabled */
 306         int                     kickfd;
 307
 308         /* inflight share memory info */
 309         union {
 310                 struct rte_vhost_inflight_info_split *inflight_split;
 311                 struct rte_vhost_inflight_info_packed *inflight_packed;
 312         };
 313         struct rte_vhost_resubmit_info *resubmit_inflight;
 314         uint64_t                global_counter;
 315
 316         struct vhost_async      *async;
 317
 318         int                     notif_enable;
 319 #define VIRTIO_UNINITIALIZED_NOTIF      (-1)
 320
 321         struct vhost_vring_addr ring_addrs;
 322         struct virtqueue_stats  stats;
 323 } __rte_cache_aligned;
 324
 325 /* Virtio device status as per Virtio specification */
 326 #define VIRTIO_DEVICE_STATUS_RESET              0x00
 327 #define VIRTIO_DEVICE_STATUS_ACK                0x01
 328 #define VIRTIO_DEVICE_STATUS_DRIVER             0x02
 329 #define VIRTIO_DEVICE_STATUS_DRIVER_OK          0x04
 330 #define VIRTIO_DEVICE_STATUS_FEATURES_OK        0x08
 331 #define VIRTIO_DEVICE_STATUS_DEV_NEED_RESET     0x40
 332 #define VIRTIO_DEVICE_STATUS_FAILED             0x80
 333
 334 #define VHOST_MAX_VRING                 0x100
 335 #define VHOST_MAX_QUEUE_PAIRS           0x80
 336
 337 /* Declare IOMMU related bits for older kernels */
 338 #ifndef VIRTIO_F_IOMMU_PLATFORM
 339
 340 #define VIRTIO_F_IOMMU_PLATFORM 33
 341
 342 struct vhost_iotlb_msg {
 343         __u64 iova;
 344         __u64 size;
 345         __u64 uaddr;
 346 #define VHOST_ACCESS_RO      0x1
 347 #define VHOST_ACCESS_WO      0x2
 348 #define VHOST_ACCESS_RW      0x3
 349         __u8 perm;
 350 #define VHOST_IOTLB_MISS           1
 351 #define VHOST_IOTLB_UPDATE         2
 352 #define VHOST_IOTLB_INVALIDATE     3
 353 #define VHOST_IOTLB_ACCESS_FAIL    4
 354         __u8 type;
 355 };
 356
 357 #define VHOST_IOTLB_MSG 0x1
 358
 359 struct vhost_msg {
 360         int type;
 361         union {
 362                 struct vhost_iotlb_msg iotlb;
 363                 __u8 padding[64];
 364         };
 365 };
 366 #endif
 367
 368 /*
 369  * Define virtio 1.0 for older kernels
 370  */
 371 #ifndef VIRTIO_F_VERSION_1
 372  #define VIRTIO_F_VERSION_1 32
 373 #endif
 374
 375 /* Declare packed ring related bits for older kernels */
 376 #ifndef VIRTIO_F_RING_PACKED
 377
 378 #define VIRTIO_F_RING_PACKED 34
 379
 380 struct vring_packed_desc {
 381         uint64_t addr;
 382         uint32_t len;
 383         uint16_t id;
 384         uint16_t flags;
 385 };
 386
 387 struct vring_packed_desc_event {
 388         uint16_t off_wrap;
 389         uint16_t flags;
 390 };
 391 #endif
 392
 393 /*
 394  * Declare below packed ring defines unconditionally
 395  * as Kernel header might use different names.
 396  */
 397 #define VRING_DESC_F_AVAIL      (1ULL << 7)
 398 #define VRING_DESC_F_USED       (1ULL << 15)
 399
 400 #define VRING_EVENT_F_ENABLE 0x0
 401 #define VRING_EVENT_F_DISABLE 0x1
 402 #define VRING_EVENT_F_DESC 0x2
 403
 404 /*
 405  * Available and used descs are in same order
 406  */
 407 #ifndef VIRTIO_F_IN_ORDER
 408 #define VIRTIO_F_IN_ORDER      35
 409 #endif
 410
 411 /* Features supported by this builtin vhost-user net driver. */
 412 #define VIRTIO_NET_SUPPORTED_FEATURES ((1ULL << VIRTIO_NET_F_MRG_RXBUF) | \
 413                                 (1ULL << VIRTIO_F_ANY_LAYOUT) | \
 414                                 (1ULL << VIRTIO_NET_F_CTRL_VQ) | \
 415                                 (1ULL << VIRTIO_NET_F_CTRL_RX) | \
 416                                 (1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) | \
 417                                 (1ULL << VIRTIO_NET_F_MQ)      | \
 418                                 (1ULL << VIRTIO_F_VERSION_1)   | \
 419                                 (1ULL << VHOST_F_LOG_ALL)      | \
 420                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES) | \
 421                                 (1ULL << VIRTIO_NET_F_GSO) | \
 422                                 (1ULL << VIRTIO_NET_F_HOST_TSO4) | \
 423                                 (1ULL << VIRTIO_NET_F_HOST_TSO6) | \
 424                                 (1ULL << VIRTIO_NET_F_HOST_UFO) | \
 425                                 (1ULL << VIRTIO_NET_F_HOST_ECN) | \
 426                                 (1ULL << VIRTIO_NET_F_CSUM)    | \
 427                                 (1ULL << VIRTIO_NET_F_GUEST_CSUM) | \
 428                                 (1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
 429                                 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
 430                                 (1ULL << VIRTIO_NET_F_GUEST_UFO) | \
 431                                 (1ULL << VIRTIO_NET_F_GUEST_ECN) | \
 432                                 (1ULL << VIRTIO_RING_F_INDIRECT_DESC) | \
 433                                 (1ULL << VIRTIO_RING_F_EVENT_IDX) | \
 434                                 (1ULL << VIRTIO_NET_F_MTU)  | \
 435                                 (1ULL << VIRTIO_F_IN_ORDER) | \
 436                                 (1ULL << VIRTIO_F_IOMMU_PLATFORM) | \
 437                                 (1ULL << VIRTIO_F_RING_PACKED))
 438
 439
 440 struct guest_page {
 441         uint64_t guest_phys_addr;
 442         uint64_t host_iova;
 443         uint64_t host_user_addr;
 444         uint64_t size;
 445 };
 446
 447 struct inflight_mem_info {
 448         int             fd;
 449         void            *addr;
 450         uint64_t        size;
 451 };
 452
 453 /**
 454  * Device structure contains all configuration information relating
 455  * to the device.
 456  */
 457 struct virtio_net {
 458         /* Frontend (QEMU) memory and memory region information */
 459         struct rte_vhost_memory *mem;
 460         uint64_t                features;
 461         uint64_t                protocol_features;
 462         int                     vid;
 463         uint32_t                flags;
 464         uint16_t                vhost_hlen;
 465         /* to tell if we need broadcast rarp packet */
 466         int16_t                 broadcast_rarp;
 467         uint32_t                nr_vring;
 468         int                     async_copy;
 469
 470         int                     extbuf;
 471         int                     linearbuf;
 472         struct vhost_virtqueue  *virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
 473         struct inflight_mem_info *inflight_info;
 474 #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
 475         char                    ifname[IF_NAME_SZ];
 476         uint64_t                log_size;
 477         uint64_t                log_base;
 478         uint64_t                log_addr;
 479         struct rte_ether_addr   mac;
 480         uint16_t                mtu;
 481         uint8_t                 status;
 482
 483         struct rte_vhost_device_ops const *notify_ops;
 484
 485         uint32_t                nr_guest_pages;
 486         uint32_t                max_guest_pages;
 487         struct guest_page       *guest_pages;
 488
 489         int                     slave_req_fd;
 490         rte_spinlock_t          slave_req_lock;
 491
 492         int                     postcopy_ufd;
 493         int                     postcopy_listening;
 494
 495         struct rte_vdpa_device *vdpa_dev;
 496
 497         /* context data for the external message handlers */
 498         void                    *extern_data;
 499         /* pre and post vhost user message handlers for the device */
 500         struct rte_vhost_user_extern_ops extern_ops;
 501 } __rte_cache_aligned;
 502
 503 static __rte_always_inline bool
 504 vq_is_packed(struct virtio_net *dev)
 505 {
 506         return dev->features & (1ull << VIRTIO_F_RING_PACKED);
 507 }
 508
 509 static inline bool
 510 desc_is_avail(struct vring_packed_desc *desc, bool wrap_counter)
 511 {
 512         uint16_t flags = __atomic_load_n(&desc->flags, __ATOMIC_ACQUIRE);
 513
 514         return wrap_counter == !!(flags & VRING_DESC_F_AVAIL) &&
 515                 wrap_counter != !!(flags & VRING_DESC_F_USED);
 516 }
 517
 518 static inline void
 519 vq_inc_last_used_packed(struct vhost_virtqueue *vq, uint16_t num)
 520 {
 521         vq->last_used_idx += num;
 522         if (vq->last_used_idx >= vq->size) {
 523                 vq->used_wrap_counter ^= 1;
 524                 vq->last_used_idx -= vq->size;
 525         }
 526 }
 527
 528 static inline void
 529 vq_inc_last_avail_packed(struct vhost_virtqueue *vq, uint16_t num)
 530 {
 531         vq->last_avail_idx += num;
 532         if (vq->last_avail_idx >= vq->size) {
 533                 vq->avail_wrap_counter ^= 1;
 534                 vq->last_avail_idx -= vq->size;
 535         }
 536 }
 537
 538 void __vhost_log_cache_write(struct virtio_net *dev,
 539                 struct vhost_virtqueue *vq,
 540                 uint64_t addr, uint64_t len);
 541 void __vhost_log_cache_write_iova(struct virtio_net *dev,
 542                 struct vhost_virtqueue *vq,
 543                 uint64_t iova, uint64_t len);
 544 void __vhost_log_cache_sync(struct virtio_net *dev,
 545                 struct vhost_virtqueue *vq);
 546 void __vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len);
 547 void __vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 548                             uint64_t iova, uint64_t len);
 549
 550 static __rte_always_inline void
 551 vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
 552 {
 553         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL)))
 554                 __vhost_log_write(dev, addr, len);
 555 }
 556
 557 static __rte_always_inline void
 558 vhost_log_cache_sync(struct virtio_net *dev, struct vhost_virtqueue *vq)
 559 {
 560         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL)))
 561                 __vhost_log_cache_sync(dev, vq);
 562 }
 563
 564 static __rte_always_inline void
 565 vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
 566                         uint64_t addr, uint64_t len)
 567 {
 568         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL)))
 569                 __vhost_log_cache_write(dev, vq, addr, len);
 570 }
 571
 572 static __rte_always_inline void
 573 vhost_log_cache_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,
 574                         uint64_t offset, uint64_t len)
 575 {
 576         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL))) {
 577                 if (unlikely(vq->log_guest_addr == 0))
 578                         return;
 579                 __vhost_log_cache_write(dev, vq, vq->log_guest_addr + offset,
 580                                         len);
 581         }
 582 }
 583
 584 static __rte_always_inline void
 585 vhost_log_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,
 586                      uint64_t offset, uint64_t len)
 587 {
 588         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL))) {
 589                 if (unlikely(vq->log_guest_addr == 0))
 590                         return;
 591                 __vhost_log_write(dev, vq->log_guest_addr + offset, len);
 592         }
 593 }
 594
 595 static __rte_always_inline void
 596 vhost_log_cache_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 597                            uint64_t iova, uint64_t len)
 598 {
 599         if (likely(!(dev->features & (1ULL << VHOST_F_LOG_ALL))))
 600                 return;
 601
 602         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 603                 __vhost_log_cache_write_iova(dev, vq, iova, len);
 604         else
 605                 __vhost_log_cache_write(dev, vq, iova, len);
 606 }
 607
 608 static __rte_always_inline void
 609 vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 610                            uint64_t iova, uint64_t len)
 611 {
 612         if (likely(!(dev->features & (1ULL << VHOST_F_LOG_ALL))))
 613                 return;
 614
 615         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 616                 __vhost_log_write_iova(dev, vq, iova, len);
 617         else
 618                 __vhost_log_write(dev, iova, len);
 619 }
 620
 621 extern int vhost_config_log_level;
 622 extern int vhost_data_log_level;
 623
 624 #define VHOST_LOG_CONFIG(level, fmt, args...)                   \
 625         rte_log(RTE_LOG_ ## level, vhost_config_log_level,      \
 626                 "VHOST_CONFIG: " fmt, ##args)
 627
 628 #define VHOST_LOG_DATA(level, fmt, args...) \
 629         (void)((RTE_LOG_ ## level <= RTE_LOG_DP_LEVEL) ?        \
 630          rte_log(RTE_LOG_ ## level,  vhost_data_log_level,      \
 631                 "VHOST_DATA : " fmt, ##args) :                  \
 632          0)
 633
 634 #ifdef RTE_LIBRTE_VHOST_DEBUG
 635 #define VHOST_MAX_PRINT_BUFF 6072
 636 #define PRINT_PACKET(device, addr, size, header) do { \
 637         char *pkt_addr = (char *)(addr); \
 638         unsigned int index; \
 639         char packet[VHOST_MAX_PRINT_BUFF]; \
 640         \
 641         if ((header)) \
 642                 snprintf(packet, VHOST_MAX_PRINT_BUFF, "(%d) Header size %d: ", (device->vid), (size)); \
 643         else \
 644                 snprintf(packet, VHOST_MAX_PRINT_BUFF, "(%d) Packet size %d: ", (device->vid), (size)); \
 645         for (index = 0; index < (size); index++) { \
 646                 snprintf(packet + strnlen(packet, VHOST_MAX_PRINT_BUFF), VHOST_MAX_PRINT_BUFF - strnlen(packet, VHOST_MAX_PRINT_BUFF), \
 647                         "%02hhx ", pkt_addr[index]); \
 648         } \
 649         snprintf(packet + strnlen(packet, VHOST_MAX_PRINT_BUFF), VHOST_MAX_PRINT_BUFF - strnlen(packet, VHOST_MAX_PRINT_BUFF), "\n"); \
 650         \
 651         VHOST_LOG_DATA(DEBUG, "%s", packet); \
 652 } while (0)
 653 #else
 654 #define PRINT_PACKET(device, addr, size, header) do {} while (0)
 655 #endif
 656
 657 extern struct virtio_net *vhost_devices[RTE_MAX_VHOST_DEVICE];
 658
 659 #define VHOST_BINARY_SEARCH_THRESH 256
 660
 661 static __rte_always_inline int guest_page_addrcmp(const void *p1,
 662                                                 const void *p2)
 663 {
 664         const struct guest_page *page1 = (const struct guest_page *)p1;
 665         const struct guest_page *page2 = (const struct guest_page *)p2;
 666
 667         if (page1->guest_phys_addr > page2->guest_phys_addr)
 668                 return 1;
 669         if (page1->guest_phys_addr < page2->guest_phys_addr)
 670                 return -1;
 671
 672         return 0;
 673 }
 674
 675 static __rte_always_inline int guest_page_rangecmp(const void *p1, const void *p2)
 676 {
 677         const struct guest_page *page1 = (const struct guest_page *)p1;
 678         const struct guest_page *page2 = (const struct guest_page *)p2;
 679
 680         if (page1->guest_phys_addr >= page2->guest_phys_addr) {
 681                 if (page1->guest_phys_addr < page2->guest_phys_addr + page2->size)
 682                         return 0;
 683                 else
 684                         return 1;
 685         } else
 686                 return -1;
 687 }
 688
 689 static __rte_always_inline rte_iova_t
 690 gpa_to_first_hpa(struct virtio_net *dev, uint64_t gpa,
 691         uint64_t gpa_size, uint64_t *hpa_size)
 692 {
 693         uint32_t i;
 694         struct guest_page *page;
 695         struct guest_page key;
 696
 697         *hpa_size = gpa_size;
 698         if (dev->nr_guest_pages >= VHOST_BINARY_SEARCH_THRESH) {
 699                 key.guest_phys_addr = gpa;
 700                 page = bsearch(&key, dev->guest_pages, dev->nr_guest_pages,
 701                                sizeof(struct guest_page), guest_page_rangecmp);
 702                 if (page) {
 703                         if (gpa + gpa_size <=
 704                                         page->guest_phys_addr + page->size) {
 705                                 return gpa - page->guest_phys_addr +
 706                                         page->host_iova;
 707                         } else if (gpa < page->guest_phys_addr +
 708                                                 page->size) {
 709                                 *hpa_size = page->guest_phys_addr +
 710                                         page->size - gpa;
 711                                 return gpa - page->guest_phys_addr +
 712                                         page->host_iova;
 713                         }
 714                 }
 715         } else {
 716                 for (i = 0; i < dev->nr_guest_pages; i++) {
 717                         page = &dev->guest_pages[i];
 718
 719                         if (gpa >= page->guest_phys_addr) {
 720                                 if (gpa + gpa_size <=
 721                                         page->guest_phys_addr + page->size) {
 722                                         return gpa - page->guest_phys_addr +
 723                                                 page->host_iova;
 724                                 } else if (gpa < page->guest_phys_addr +
 725                                                         page->size) {
 726                                         *hpa_size = page->guest_phys_addr +
 727                                                 page->size - gpa;
 728                                         return gpa - page->guest_phys_addr +
 729                                                 page->host_iova;
 730                                 }
 731                         }
 732                 }
 733         }
 734
 735         *hpa_size = 0;
 736         return 0;
 737 }
 738
 739 /* Convert guest physical address to host physical address */
 740 static __rte_always_inline rte_iova_t
 741 gpa_to_hpa(struct virtio_net *dev, uint64_t gpa, uint64_t size)
 742 {
 743         rte_iova_t hpa;
 744         uint64_t hpa_size;
 745
 746         hpa = gpa_to_first_hpa(dev, gpa, size, &hpa_size);
 747         return hpa_size == size ? hpa : 0;
 748 }
 749
 750 static __rte_always_inline uint64_t
 751 hva_to_gpa(struct virtio_net *dev, uint64_t vva, uint64_t len)
 752 {
 753         struct rte_vhost_mem_region *r;
 754         uint32_t i;
 755
 756         if (unlikely(!dev || !dev->mem))
 757                 return 0;
 758
 759         for (i = 0; i < dev->mem->nregions; i++) {
 760                 r = &dev->mem->regions[i];
 761
 762                 if (vva >= r->host_user_addr &&
 763                     vva + len <  r->host_user_addr + r->size) {
 764                         return r->guest_phys_addr + vva - r->host_user_addr;
 765                 }
 766         }
 767         return 0;
 768 }
 769
 770 static __rte_always_inline struct virtio_net *
 771 get_device(int vid)
 772 {
 773         struct virtio_net *dev = vhost_devices[vid];
 774
 775         if (unlikely(!dev)) {
 776                 VHOST_LOG_CONFIG(ERR,
 777                         "(%d) device not found.\n", vid);
 778         }
 779
 780         return dev;
 781 }
 782
 783 int vhost_new_device(void);
 784 void cleanup_device(struct virtio_net *dev, int destroy);
 785 void reset_device(struct virtio_net *dev);
 786 void vhost_destroy_device(int);
 787 void vhost_destroy_device_notify(struct virtio_net *dev);
 788
 789 void cleanup_vq(struct vhost_virtqueue *vq, int destroy);
 790 void cleanup_vq_inflight(struct virtio_net *dev, struct vhost_virtqueue *vq);
 791 void free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq);
 792
 793 int alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx);
 794
 795 void vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *dev);
 796
 797 void vhost_set_ifname(int, const char *if_name, unsigned int if_len);
 798 void vhost_setup_virtio_net(int vid, bool enable, bool legacy_ol_flags, bool stats_enabled);
 799 void vhost_enable_extbuf(int vid);
 800 void vhost_enable_linearbuf(int vid);
 801 int vhost_enable_guest_notification(struct virtio_net *dev,
 802                 struct vhost_virtqueue *vq, int enable);
 803
 804 struct rte_vhost_device_ops const *vhost_driver_callback_get(const char *path);
 805
 806 /*
 807  * Backend-specific cleanup.
 808  *
 809  * TODO: fix it; we have one backend now
 810  */
 811 void vhost_backend_cleanup(struct virtio_net *dev);
 812
 813 uint64_t __vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 814                         uint64_t iova, uint64_t *len, uint8_t perm);
 815 void *vhost_alloc_copy_ind_table(struct virtio_net *dev,
 816                         struct vhost_virtqueue *vq,
 817                         uint64_t desc_addr, uint64_t desc_len);
 818 int vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq);
 819 uint64_t translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
 820                 uint64_t log_addr);
 821 void vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq);
 822
 823 static __rte_always_inline uint64_t
 824 vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 825                         uint64_t iova, uint64_t *len, uint8_t perm)
 826 {
 827         if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
 828                 return rte_vhost_va_from_guest_pa(dev->mem, iova, len);
 829
 830         return __vhost_iova_to_vva(dev, vq, iova, len, perm);
 831 }
 832
 833 #define vhost_avail_event(vr) \
 834         (*(volatile uint16_t*)&(vr)->used->ring[(vr)->size])
 835 #define vhost_used_event(vr) \
 836         (*(volatile uint16_t*)&(vr)->avail->ring[(vr)->size])
 837
 838 /*
 839  * The following is used with VIRTIO_RING_F_EVENT_IDX.
 840  * Assuming a given event_idx value from the other size, if we have
 841  * just incremented index from old to new_idx, should we trigger an
 842  * event?
 843  */
 844 static __rte_always_inline int
 845 vhost_need_event(uint16_t event_idx, uint16_t new_idx, uint16_t old)
 846 {
 847         return (uint16_t)(new_idx - event_idx - 1) < (uint16_t)(new_idx - old);
 848 }
 849
 850 static __rte_always_inline void
 851 vhost_vring_call_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
 852 {
 853         /* Flush used->idx update before we read avail->flags. */
 854         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
 855
 856         /* Don't kick guest if we don't reach index specified by guest. */
 857         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
 858                 uint16_t old = vq->signalled_used;
 859                 uint16_t new = vq->last_used_idx;
 860                 bool signalled_used_valid = vq->signalled_used_valid;
 861
 862                 vq->signalled_used = new;
 863                 vq->signalled_used_valid = true;
 864
 865                 VHOST_LOG_DATA(DEBUG, "%s: used_event_idx=%d, old=%d, new=%d\n",
 866                         __func__,
 867                         vhost_used_event(vq),
 868                         old, new);
 869
 870                 if ((vhost_need_event(vhost_used_event(vq), new, old) &&
 871                                         (vq->callfd >= 0)) ||
 872                                 unlikely(!signalled_used_valid)) {
 873                         eventfd_write(vq->callfd, (eventfd_t) 1);
 874                         if (dev->notify_ops->guest_notified)
 875                                 dev->notify_ops->guest_notified(dev->vid);
 876                 }
 877         } else {
 878                 /* Kick the guest if necessary. */
 879                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 880                                 && (vq->callfd >= 0)) {
 881                         eventfd_write(vq->callfd, (eventfd_t)1);
 882                         if (dev->notify_ops->guest_notified)
 883                                 dev->notify_ops->guest_notified(dev->vid);
 884                 }
 885         }
 886 }
 887
 888 static __rte_always_inline void
 889 vhost_vring_call_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
 890 {
 891         uint16_t old, new, off, off_wrap;
 892         bool signalled_used_valid, kick = false;
 893
 894         /* Flush used desc update. */
 895         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
 896
 897         if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
 898                 if (vq->driver_event->flags !=
 899                                 VRING_EVENT_F_DISABLE)
 900                         kick = true;
 901                 goto kick;
 902         }
 903
 904         old = vq->signalled_used;
 905         new = vq->last_used_idx;
 906         vq->signalled_used = new;
 907         signalled_used_valid = vq->signalled_used_valid;
 908         vq->signalled_used_valid = true;
 909
 910         if (vq->driver_event->flags != VRING_EVENT_F_DESC) {
 911                 if (vq->driver_event->flags != VRING_EVENT_F_DISABLE)
 912                         kick = true;
 913                 goto kick;
 914         }
 915
 916         if (unlikely(!signalled_used_valid)) {
 917                 kick = true;
 918                 goto kick;
 919         }
 920
 921         rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
 922
 923         off_wrap = vq->driver_event->off_wrap;
 924         off = off_wrap & ~(1 << 15);
 925
 926         if (new <= old)
 927                 old -= vq->size;
 928
 929         if (vq->used_wrap_counter != off_wrap >> 15)
 930                 off -= vq->size;
 931
 932         if (vhost_need_event(off, new, old))
 933                 kick = true;
 934 kick:
 935         if (kick) {
 936                 eventfd_write(vq->callfd, (eventfd_t)1);
 937                 if (dev->notify_ops->guest_notified)
 938                         dev->notify_ops->guest_notified(dev->vid);
 939         }
 940 }
 941
 942 static __rte_always_inline void
 943 free_ind_table(void *idesc)
 944 {
 945         rte_free(idesc);
 946 }
 947
 948 static __rte_always_inline void
 949 restore_mbuf(struct rte_mbuf *m)
 950 {
 951         uint32_t mbuf_size, priv_size;
 952
 953         while (m) {
 954                 priv_size = rte_pktmbuf_priv_size(m->pool);
 955                 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
 956                 /* start of buffer is after mbuf structure and priv data */
 957
 958                 m->buf_addr = (char *)m + mbuf_size;
 959                 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
 960                 m = m->next;
 961         }
 962 }
 963
 964 static __rte_always_inline bool
 965 mbuf_is_consumed(struct rte_mbuf *m)
 966 {
 967         while (m) {
 968                 if (rte_mbuf_refcnt_read(m) > 1)
 969                         return false;
 970                 m = m->next;
 971         }
 972
 973         return true;
 974 }
 975 #endif /* _VHOST_NET_CDEV_H_ */