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