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         uint64_t guest_notifications;
 137         uint64_t iotlb_hits;
 138         uint64_t iotlb_misses;
 139         uint64_t inflight_submitted;
 140         uint64_t inflight_completed;
 141 };
 142
 143 /**
 144  * iovec
 145  */
 146 struct vhost_iovec {
 147         void *src_addr;
 148         void *dst_addr;
 149         size_t len;
 150 };
 151
 152 /**
 153  * iovec iterator
 154  */
 155 struct vhost_iov_iter {
 156         /** pointer to the iovec array */
 157         struct vhost_iovec *iov;
 158         /** number of iovec in this iterator */
 159         unsigned long nr_segs;
 160 };
 161
 162 struct async_dma_vchan_info {
 163         /* circular array to track if packet copy completes */
 164         bool **pkts_cmpl_flag_addr;
 165
 166         /* max elements in 'pkts_cmpl_flag_addr' */
 167         uint16_t ring_size;
 168         /* ring index mask for 'pkts_cmpl_flag_addr' */
 169         uint16_t ring_mask;
 170
 171         /**
 172          * DMA virtual channel lock. Although it is able to bind DMA
 173          * virtual channels to data plane threads, vhost control plane
 174          * thread could call data plane functions too, thus causing
 175          * DMA device contention.
 176          *
 177          * For example, in VM exit case, vhost control plane thread needs
 178          * to clear in-flight packets before disable vring, but there could
 179          * be anotther data plane thread is enqueuing packets to the same
 180          * vring with the same DMA virtual channel. As dmadev PMD functions
 181          * are lock-free, the control plane and data plane threads could
 182          * operate the same DMA virtual channel at the same time.
 183          */
 184         rte_spinlock_t dma_lock;
 185 };
 186
 187 struct async_dma_info {
 188         struct async_dma_vchan_info *vchans;
 189         /* number of registered virtual channels */
 190         uint16_t nr_vchans;
 191 };
 192
 193 extern struct async_dma_info dma_copy_track[RTE_DMADEV_DEFAULT_MAX];
 194
 195 /**
 196  * inflight async packet information
 197  */
 198 struct async_inflight_info {
 199         struct rte_mbuf *mbuf;
 200         uint16_t descs; /* num of descs inflight */
 201         uint16_t nr_buffers; /* num of buffers inflight for packed ring */
 202         struct virtio_net_hdr nethdr;
 203 };
 204
 205 struct vhost_async {
 206         struct vhost_iov_iter iov_iter[VHOST_MAX_ASYNC_IT];
 207         struct vhost_iovec iovec[VHOST_MAX_ASYNC_VEC];
 208         uint16_t iter_idx;
 209         uint16_t iovec_idx;
 210
 211         /* data transfer status */
 212         struct async_inflight_info *pkts_info;
 213         /**
 214          * Packet reorder array. "true" indicates that DMA device
 215          * completes all copies for the packet.
 216          *
 217          * Note that this array could be written by multiple threads
 218          * simultaneously. For example, in the case of thread0 and
 219          * thread1 RX packets from NIC and then enqueue packets to
 220          * vring0 and vring1 with own DMA device DMA0 and DMA1, it's
 221          * possible for thread0 to get completed copies belonging to
 222          * vring1 from DMA0, while thread0 is calling rte_vhost_poll
 223          * _enqueue_completed() for vring0 and thread1 is calling
 224          * rte_vhost_submit_enqueue_burst() for vring1. In this case,
 225          * vq->access_lock cannot protect pkts_cmpl_flag of vring1.
 226          *
 227          * However, since offloading is per-packet basis, each packet
 228          * flag will only be written by one thread. And single byte
 229          * write is atomic, so no lock for pkts_cmpl_flag is needed.
 230          */
 231         bool *pkts_cmpl_flag;
 232         uint16_t pkts_idx;
 233         uint16_t pkts_inflight_n;
 234         union {
 235                 struct vring_used_elem  *descs_split;
 236                 struct vring_used_elem_packed *buffers_packed;
 237         };
 238         union {
 239                 uint16_t desc_idx_split;
 240                 uint16_t buffer_idx_packed;
 241         };
 242         union {
 243                 uint16_t last_desc_idx_split;
 244                 uint16_t last_buffer_idx_packed;
 245         };
 246 };
 247
 248 /**
 249  * Structure contains variables relevant to RX/TX virtqueues.
 250  */
 251 struct vhost_virtqueue {
 252         union {
 253                 struct vring_desc       *desc;
 254                 struct vring_packed_desc   *desc_packed;
 255         };
 256         union {
 257                 struct vring_avail      *avail;
 258                 struct vring_packed_desc_event *driver_event;
 259         };
 260         union {
 261                 struct vring_used       *used;
 262                 struct vring_packed_desc_event *device_event;
 263         };
 264         uint16_t                size;
 265
 266         uint16_t                last_avail_idx;
 267         uint16_t                last_used_idx;
 268         /* Last used index we notify to front end. */
 269         uint16_t                signalled_used;
 270         bool                    signalled_used_valid;
 271 #define VIRTIO_INVALID_EVENTFD          (-1)
 272 #define VIRTIO_UNINITIALIZED_EVENTFD    (-2)
 273
 274         bool                    enabled;
 275         bool                    access_ok;
 276         bool                    ready;
 277
 278         rte_spinlock_t          access_lock;
 279
 280
 281         union {
 282                 struct vring_used_elem  *shadow_used_split;
 283                 struct vring_used_elem_packed *shadow_used_packed;
 284         };
 285         uint16_t                shadow_used_idx;
 286         /* Record packed ring enqueue latest desc cache aligned index */
 287         uint16_t                shadow_aligned_idx;
 288         /* Record packed ring first dequeue desc index */
 289         uint16_t                shadow_last_used_idx;
 290
 291         uint16_t                batch_copy_nb_elems;
 292         struct batch_copy_elem  *batch_copy_elems;
 293         int                     numa_node;
 294         bool                    used_wrap_counter;
 295         bool                    avail_wrap_counter;
 296
 297         /* Physical address of used ring, for logging */
 298         uint16_t                log_cache_nb_elem;
 299         uint64_t                log_guest_addr;
 300         struct log_cache_entry  *log_cache;
 301
 302         rte_rwlock_t    iotlb_lock;
 303         rte_rwlock_t    iotlb_pending_lock;
 304         struct rte_mempool *iotlb_pool;
 305         TAILQ_HEAD(, vhost_iotlb_entry) iotlb_list;
 306         TAILQ_HEAD(, vhost_iotlb_entry) iotlb_pending_list;
 307         int                             iotlb_cache_nr;
 308
 309         /* Used to notify the guest (trigger interrupt) */
 310         int                     callfd;
 311         /* Currently unused as polling mode is enabled */
 312         int                     kickfd;
 313
 314         /* inflight share memory info */
 315         union {
 316                 struct rte_vhost_inflight_info_split *inflight_split;
 317                 struct rte_vhost_inflight_info_packed *inflight_packed;
 318         };
 319         struct rte_vhost_resubmit_info *resubmit_inflight;
 320         uint64_t                global_counter;
 321
 322         struct vhost_async      *async;
 323
 324         int                     notif_enable;
 325 #define VIRTIO_UNINITIALIZED_NOTIF      (-1)
 326
 327         struct vhost_vring_addr ring_addrs;
 328         struct virtqueue_stats  stats;
 329 } __rte_cache_aligned;
 330
 331 /* Virtio device status as per Virtio specification */
 332 #define VIRTIO_DEVICE_STATUS_RESET              0x00
 333 #define VIRTIO_DEVICE_STATUS_ACK                0x01
 334 #define VIRTIO_DEVICE_STATUS_DRIVER             0x02
 335 #define VIRTIO_DEVICE_STATUS_DRIVER_OK          0x04
 336 #define VIRTIO_DEVICE_STATUS_FEATURES_OK        0x08
 337 #define VIRTIO_DEVICE_STATUS_DEV_NEED_RESET     0x40
 338 #define VIRTIO_DEVICE_STATUS_FAILED             0x80
 339
 340 #define VHOST_MAX_VRING                 0x100
 341 #define VHOST_MAX_QUEUE_PAIRS           0x80
 342
 343 /* Declare IOMMU related bits for older kernels */
 344 #ifndef VIRTIO_F_IOMMU_PLATFORM
 345
 346 #define VIRTIO_F_IOMMU_PLATFORM 33
 347
 348 struct vhost_iotlb_msg {
 349         __u64 iova;
 350         __u64 size;
 351         __u64 uaddr;
 352 #define VHOST_ACCESS_RO      0x1
 353 #define VHOST_ACCESS_WO      0x2
 354 #define VHOST_ACCESS_RW      0x3
 355         __u8 perm;
 356 #define VHOST_IOTLB_MISS           1
 357 #define VHOST_IOTLB_UPDATE         2
 358 #define VHOST_IOTLB_INVALIDATE     3
 359 #define VHOST_IOTLB_ACCESS_FAIL    4
 360         __u8 type;
 361 };
 362
 363 #define VHOST_IOTLB_MSG 0x1
 364
 365 struct vhost_msg {
 366         int type;
 367         union {
 368                 struct vhost_iotlb_msg iotlb;
 369                 __u8 padding[64];
 370         };
 371 };
 372 #endif
 373
 374 /*
 375  * Define virtio 1.0 for older kernels
 376  */
 377 #ifndef VIRTIO_F_VERSION_1
 378  #define VIRTIO_F_VERSION_1 32
 379 #endif
 380
 381 /* Declare packed ring related bits for older kernels */
 382 #ifndef VIRTIO_F_RING_PACKED
 383
 384 #define VIRTIO_F_RING_PACKED 34
 385
 386 struct vring_packed_desc {
 387         uint64_t addr;
 388         uint32_t len;
 389         uint16_t id;
 390         uint16_t flags;
 391 };
 392
 393 struct vring_packed_desc_event {
 394         uint16_t off_wrap;
 395         uint16_t flags;
 396 };
 397 #endif
 398
 399 /*
 400  * Declare below packed ring defines unconditionally
 401  * as Kernel header might use different names.
 402  */
 403 #define VRING_DESC_F_AVAIL      (1ULL << 7)
 404 #define VRING_DESC_F_USED       (1ULL << 15)
 405
 406 #define VRING_EVENT_F_ENABLE 0x0
 407 #define VRING_EVENT_F_DISABLE 0x1
 408 #define VRING_EVENT_F_DESC 0x2
 409
 410 /*
 411  * Available and used descs are in same order
 412  */
 413 #ifndef VIRTIO_F_IN_ORDER
 414 #define VIRTIO_F_IN_ORDER      35
 415 #endif
 416
 417 /* Features supported by this builtin vhost-user net driver. */
 418 #define VIRTIO_NET_SUPPORTED_FEATURES ((1ULL << VIRTIO_NET_F_MRG_RXBUF) | \
 419                                 (1ULL << VIRTIO_F_ANY_LAYOUT) | \
 420                                 (1ULL << VIRTIO_NET_F_CTRL_VQ) | \
 421                                 (1ULL << VIRTIO_NET_F_CTRL_RX) | \
 422                                 (1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) | \
 423                                 (1ULL << VIRTIO_NET_F_MQ)      | \
 424                                 (1ULL << VIRTIO_F_VERSION_1)   | \
 425                                 (1ULL << VHOST_F_LOG_ALL)      | \
 426                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES) | \
 427                                 (1ULL << VIRTIO_NET_F_GSO) | \
 428                                 (1ULL << VIRTIO_NET_F_HOST_TSO4) | \
 429                                 (1ULL << VIRTIO_NET_F_HOST_TSO6) | \
 430                                 (1ULL << VIRTIO_NET_F_HOST_UFO) | \
 431                                 (1ULL << VIRTIO_NET_F_HOST_ECN) | \
 432                                 (1ULL << VIRTIO_NET_F_CSUM)    | \
 433                                 (1ULL << VIRTIO_NET_F_GUEST_CSUM) | \
 434                                 (1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
 435                                 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
 436                                 (1ULL << VIRTIO_NET_F_GUEST_UFO) | \
 437                                 (1ULL << VIRTIO_NET_F_GUEST_ECN) | \
 438                                 (1ULL << VIRTIO_RING_F_INDIRECT_DESC) | \
 439                                 (1ULL << VIRTIO_RING_F_EVENT_IDX) | \
 440                                 (1ULL << VIRTIO_NET_F_MTU)  | \
 441                                 (1ULL << VIRTIO_F_IN_ORDER) | \
 442                                 (1ULL << VIRTIO_F_IOMMU_PLATFORM) | \
 443                                 (1ULL << VIRTIO_F_RING_PACKED))
 444
 445
 446 struct guest_page {
 447         uint64_t guest_phys_addr;
 448         uint64_t host_iova;
 449         uint64_t host_user_addr;
 450         uint64_t size;
 451 };
 452
 453 struct inflight_mem_info {
 454         int             fd;
 455         void            *addr;
 456         uint64_t        size;
 457 };
 458
 459 /**
 460  * Device structure contains all configuration information relating
 461  * to the device.
 462  */
 463 struct virtio_net {
 464         /* Frontend (QEMU) memory and memory region information */
 465         struct rte_vhost_memory *mem;
 466         uint64_t                features;
 467         uint64_t                protocol_features;
 468         int                     vid;
 469         uint32_t                flags;
 470         uint16_t                vhost_hlen;
 471         /* to tell if we need broadcast rarp packet */
 472         int16_t                 broadcast_rarp;
 473         uint32_t                nr_vring;
 474         int                     async_copy;
 475
 476         int                     extbuf;
 477         int                     linearbuf;
 478         struct vhost_virtqueue  *virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
 479         struct inflight_mem_info *inflight_info;
 480 #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
 481         char                    ifname[IF_NAME_SZ];
 482         uint64_t                log_size;
 483         uint64_t                log_base;
 484         uint64_t                log_addr;
 485         struct rte_ether_addr   mac;
 486         uint16_t                mtu;
 487         uint8_t                 status;
 488
 489         struct rte_vhost_device_ops const *notify_ops;
 490
 491         uint32_t                nr_guest_pages;
 492         uint32_t                max_guest_pages;
 493         struct guest_page       *guest_pages;
 494
 495         int                     slave_req_fd;
 496         rte_spinlock_t          slave_req_lock;
 497
 498         int                     postcopy_ufd;
 499         int                     postcopy_listening;
 500
 501         struct rte_vdpa_device *vdpa_dev;
 502
 503         /* context data for the external message handlers */
 504         void                    *extern_data;
 505         /* pre and post vhost user message handlers for the device */
 506         struct rte_vhost_user_extern_ops extern_ops;
 507 } __rte_cache_aligned;
 508
 509 static __rte_always_inline bool
 510 vq_is_packed(struct virtio_net *dev)
 511 {
 512         return dev->features & (1ull << VIRTIO_F_RING_PACKED);
 513 }
 514
 515 static inline bool
 516 desc_is_avail(struct vring_packed_desc *desc, bool wrap_counter)
 517 {
 518         uint16_t flags = __atomic_load_n(&desc->flags, __ATOMIC_ACQUIRE);
 519
 520         return wrap_counter == !!(flags & VRING_DESC_F_AVAIL) &&
 521                 wrap_counter != !!(flags & VRING_DESC_F_USED);
 522 }
 523
 524 static inline void
 525 vq_inc_last_used_packed(struct vhost_virtqueue *vq, uint16_t num)
 526 {
 527         vq->last_used_idx += num;
 528         if (vq->last_used_idx >= vq->size) {
 529                 vq->used_wrap_counter ^= 1;
 530                 vq->last_used_idx -= vq->size;
 531         }
 532 }
 533
 534 static inline void
 535 vq_inc_last_avail_packed(struct vhost_virtqueue *vq, uint16_t num)
 536 {
 537         vq->last_avail_idx += num;
 538         if (vq->last_avail_idx >= vq->size) {
 539                 vq->avail_wrap_counter ^= 1;
 540                 vq->last_avail_idx -= vq->size;
 541         }
 542 }
 543
 544 void __vhost_log_cache_write(struct virtio_net *dev,
 545                 struct vhost_virtqueue *vq,
 546                 uint64_t addr, uint64_t len);
 547 void __vhost_log_cache_write_iova(struct virtio_net *dev,
 548                 struct vhost_virtqueue *vq,
 549                 uint64_t iova, uint64_t len);
 550 void __vhost_log_cache_sync(struct virtio_net *dev,
 551                 struct vhost_virtqueue *vq);
 552 void __vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len);
 553 void __vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 554                             uint64_t iova, uint64_t len);
 555
 556 static __rte_always_inline void
 557 vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
 558 {
 559         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL)))
 560                 __vhost_log_write(dev, addr, len);
 561 }
 562
 563 static __rte_always_inline void
 564 vhost_log_cache_sync(struct virtio_net *dev, struct vhost_virtqueue *vq)
 565 {
 566         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL)))
 567                 __vhost_log_cache_sync(dev, vq);
 568 }
 569
 570 static __rte_always_inline void
 571 vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
 572                         uint64_t addr, uint64_t len)
 573 {
 574         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL)))
 575                 __vhost_log_cache_write(dev, vq, addr, len);
 576 }
 577
 578 static __rte_always_inline void
 579 vhost_log_cache_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,
 580                         uint64_t offset, uint64_t len)
 581 {
 582         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL))) {
 583                 if (unlikely(vq->log_guest_addr == 0))
 584                         return;
 585                 __vhost_log_cache_write(dev, vq, vq->log_guest_addr + offset,
 586                                         len);
 587         }
 588 }
 589
 590 static __rte_always_inline void
 591 vhost_log_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,
 592                      uint64_t offset, uint64_t len)
 593 {
 594         if (unlikely(dev->features & (1ULL << VHOST_F_LOG_ALL))) {
 595                 if (unlikely(vq->log_guest_addr == 0))
 596                         return;
 597                 __vhost_log_write(dev, vq->log_guest_addr + offset, len);
 598         }
 599 }
 600
 601 static __rte_always_inline void
 602 vhost_log_cache_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 603                            uint64_t iova, uint64_t len)
 604 {
 605         if (likely(!(dev->features & (1ULL << VHOST_F_LOG_ALL))))
 606                 return;
 607
 608         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 609                 __vhost_log_cache_write_iova(dev, vq, iova, len);
 610         else
 611                 __vhost_log_cache_write(dev, vq, iova, len);
 612 }
 613
 614 static __rte_always_inline void
 615 vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 616                            uint64_t iova, uint64_t len)
 617 {
 618         if (likely(!(dev->features & (1ULL << VHOST_F_LOG_ALL))))
 619                 return;
 620
 621         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 622                 __vhost_log_write_iova(dev, vq, iova, len);
 623         else
 624                 __vhost_log_write(dev, iova, len);
 625 }
 626
 627 extern int vhost_config_log_level;
 628 extern int vhost_data_log_level;
 629
 630 #define VHOST_LOG_CONFIG(level, fmt, args...)                   \
 631         rte_log(RTE_LOG_ ## level, vhost_config_log_level,      \
 632                 "VHOST_CONFIG: " fmt, ##args)
 633
 634 #define VHOST_LOG_DATA(level, fmt, args...) \
 635         (void)((RTE_LOG_ ## level <= RTE_LOG_DP_LEVEL) ?        \
 636          rte_log(RTE_LOG_ ## level,  vhost_data_log_level,      \
 637                 "VHOST_DATA : " fmt, ##args) :                  \
 638          0)
 639
 640 #ifdef RTE_LIBRTE_VHOST_DEBUG
 641 #define VHOST_MAX_PRINT_BUFF 6072
 642 #define PRINT_PACKET(device, addr, size, header) do { \
 643         char *pkt_addr = (char *)(addr); \
 644         unsigned int index; \
 645         char packet[VHOST_MAX_PRINT_BUFF]; \
 646         \
 647         if ((header)) \
 648                 snprintf(packet, VHOST_MAX_PRINT_BUFF, "(%d) Header size %d: ", (device->vid), (size)); \
 649         else \
 650                 snprintf(packet, VHOST_MAX_PRINT_BUFF, "(%d) Packet size %d: ", (device->vid), (size)); \
 651         for (index = 0; index < (size); index++) { \
 652                 snprintf(packet + strnlen(packet, VHOST_MAX_PRINT_BUFF), VHOST_MAX_PRINT_BUFF - strnlen(packet, VHOST_MAX_PRINT_BUFF), \
 653                         "%02hhx ", pkt_addr[index]); \
 654         } \
 655         snprintf(packet + strnlen(packet, VHOST_MAX_PRINT_BUFF), VHOST_MAX_PRINT_BUFF - strnlen(packet, VHOST_MAX_PRINT_BUFF), "\n"); \
 656         \
 657         VHOST_LOG_DATA(DEBUG, "%s", packet); \
 658 } while (0)
 659 #else
 660 #define PRINT_PACKET(device, addr, size, header) do {} while (0)
 661 #endif
 662
 663 extern struct virtio_net *vhost_devices[RTE_MAX_VHOST_DEVICE];
 664
 665 #define VHOST_BINARY_SEARCH_THRESH 256
 666
 667 static __rte_always_inline int guest_page_addrcmp(const void *p1,
 668                                                 const void *p2)
 669 {
 670         const struct guest_page *page1 = (const struct guest_page *)p1;
 671         const struct guest_page *page2 = (const struct guest_page *)p2;
 672
 673         if (page1->guest_phys_addr > page2->guest_phys_addr)
 674                 return 1;
 675         if (page1->guest_phys_addr < page2->guest_phys_addr)
 676                 return -1;
 677
 678         return 0;
 679 }
 680
 681 static __rte_always_inline int guest_page_rangecmp(const void *p1, const void *p2)
 682 {
 683         const struct guest_page *page1 = (const struct guest_page *)p1;
 684         const struct guest_page *page2 = (const struct guest_page *)p2;
 685
 686         if (page1->guest_phys_addr >= page2->guest_phys_addr) {
 687                 if (page1->guest_phys_addr < page2->guest_phys_addr + page2->size)
 688                         return 0;
 689                 else
 690                         return 1;
 691         } else
 692                 return -1;
 693 }
 694
 695 static __rte_always_inline rte_iova_t
 696 gpa_to_first_hpa(struct virtio_net *dev, uint64_t gpa,
 697         uint64_t gpa_size, uint64_t *hpa_size)
 698 {
 699         uint32_t i;
 700         struct guest_page *page;
 701         struct guest_page key;
 702
 703         *hpa_size = gpa_size;
 704         if (dev->nr_guest_pages >= VHOST_BINARY_SEARCH_THRESH) {
 705                 key.guest_phys_addr = gpa;
 706                 page = bsearch(&key, dev->guest_pages, dev->nr_guest_pages,
 707                                sizeof(struct guest_page), guest_page_rangecmp);
 708                 if (page) {
 709                         if (gpa + gpa_size <=
 710                                         page->guest_phys_addr + page->size) {
 711                                 return gpa - page->guest_phys_addr +
 712                                         page->host_iova;
 713                         } else if (gpa < page->guest_phys_addr +
 714                                                 page->size) {
 715                                 *hpa_size = page->guest_phys_addr +
 716                                         page->size - gpa;
 717                                 return gpa - page->guest_phys_addr +
 718                                         page->host_iova;
 719                         }
 720                 }
 721         } else {
 722                 for (i = 0; i < dev->nr_guest_pages; i++) {
 723                         page = &dev->guest_pages[i];
 724
 725                         if (gpa >= page->guest_phys_addr) {
 726                                 if (gpa + gpa_size <=
 727                                         page->guest_phys_addr + page->size) {
 728                                         return gpa - page->guest_phys_addr +
 729                                                 page->host_iova;
 730                                 } else if (gpa < page->guest_phys_addr +
 731                                                         page->size) {
 732                                         *hpa_size = page->guest_phys_addr +
 733                                                 page->size - gpa;
 734                                         return gpa - page->guest_phys_addr +
 735                                                 page->host_iova;
 736                                 }
 737                         }
 738                 }
 739         }
 740
 741         *hpa_size = 0;
 742         return 0;
 743 }
 744
 745 /* Convert guest physical address to host physical address */
 746 static __rte_always_inline rte_iova_t
 747 gpa_to_hpa(struct virtio_net *dev, uint64_t gpa, uint64_t size)
 748 {
 749         rte_iova_t hpa;
 750         uint64_t hpa_size;
 751
 752         hpa = gpa_to_first_hpa(dev, gpa, size, &hpa_size);
 753         return hpa_size == size ? hpa : 0;
 754 }
 755
 756 static __rte_always_inline uint64_t
 757 hva_to_gpa(struct virtio_net *dev, uint64_t vva, uint64_t len)
 758 {
 759         struct rte_vhost_mem_region *r;
 760         uint32_t i;
 761
 762         if (unlikely(!dev || !dev->mem))
 763                 return 0;
 764
 765         for (i = 0; i < dev->mem->nregions; i++) {
 766                 r = &dev->mem->regions[i];
 767
 768                 if (vva >= r->host_user_addr &&
 769                     vva + len <  r->host_user_addr + r->size) {
 770                         return r->guest_phys_addr + vva - r->host_user_addr;
 771                 }
 772         }
 773         return 0;
 774 }
 775
 776 static __rte_always_inline struct virtio_net *
 777 get_device(int vid)
 778 {
 779         struct virtio_net *dev = vhost_devices[vid];
 780
 781         if (unlikely(!dev)) {
 782                 VHOST_LOG_CONFIG(ERR,
 783                         "(%d) device not found.\n", vid);
 784         }
 785
 786         return dev;
 787 }
 788
 789 int vhost_new_device(void);
 790 void cleanup_device(struct virtio_net *dev, int destroy);
 791 void reset_device(struct virtio_net *dev);
 792 void vhost_destroy_device(int);
 793 void vhost_destroy_device_notify(struct virtio_net *dev);
 794
 795 void cleanup_vq(struct vhost_virtqueue *vq, int destroy);
 796 void cleanup_vq_inflight(struct virtio_net *dev, struct vhost_virtqueue *vq);
 797 void free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq);
 798
 799 int alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx);
 800
 801 void vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *dev);
 802
 803 void vhost_set_ifname(int, const char *if_name, unsigned int if_len);
 804 void vhost_setup_virtio_net(int vid, bool enable, bool legacy_ol_flags, bool stats_enabled);
 805 void vhost_enable_extbuf(int vid);
 806 void vhost_enable_linearbuf(int vid);
 807 int vhost_enable_guest_notification(struct virtio_net *dev,
 808                 struct vhost_virtqueue *vq, int enable);
 809
 810 struct rte_vhost_device_ops const *vhost_driver_callback_get(const char *path);
 811
 812 /*
 813  * Backend-specific cleanup.
 814  *
 815  * TODO: fix it; we have one backend now
 816  */
 817 void vhost_backend_cleanup(struct virtio_net *dev);
 818
 819 uint64_t __vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 820                         uint64_t iova, uint64_t *len, uint8_t perm);
 821 void *vhost_alloc_copy_ind_table(struct virtio_net *dev,
 822                         struct vhost_virtqueue *vq,
 823                         uint64_t desc_addr, uint64_t desc_len);
 824 int vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq);
 825 uint64_t translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
 826                 uint64_t log_addr);
 827 void vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq);
 828
 829 static __rte_always_inline uint64_t
 830 vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 831                         uint64_t iova, uint64_t *len, uint8_t perm)
 832 {
 833         if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
 834                 return rte_vhost_va_from_guest_pa(dev->mem, iova, len);
 835
 836         return __vhost_iova_to_vva(dev, vq, iova, len, perm);
 837 }
 838
 839 #define vhost_avail_event(vr) \
 840         (*(volatile uint16_t*)&(vr)->used->ring[(vr)->size])
 841 #define vhost_used_event(vr) \
 842         (*(volatile uint16_t*)&(vr)->avail->ring[(vr)->size])
 843
 844 /*
 845  * The following is used with VIRTIO_RING_F_EVENT_IDX.
 846  * Assuming a given event_idx value from the other size, if we have
 847  * just incremented index from old to new_idx, should we trigger an
 848  * event?
 849  */
 850 static __rte_always_inline int
 851 vhost_need_event(uint16_t event_idx, uint16_t new_idx, uint16_t old)
 852 {
 853         return (uint16_t)(new_idx - event_idx - 1) < (uint16_t)(new_idx - old);
 854 }
 855
 856 static __rte_always_inline void
 857 vhost_vring_call_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
 858 {
 859         /* Flush used->idx update before we read avail->flags. */
 860         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
 861
 862         /* Don't kick guest if we don't reach index specified by guest. */
 863         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
 864                 uint16_t old = vq->signalled_used;
 865                 uint16_t new = vq->last_used_idx;
 866                 bool signalled_used_valid = vq->signalled_used_valid;
 867
 868                 vq->signalled_used = new;
 869                 vq->signalled_used_valid = true;
 870
 871                 VHOST_LOG_DATA(DEBUG, "%s: used_event_idx=%d, old=%d, new=%d\n",
 872                         __func__,
 873                         vhost_used_event(vq),
 874                         old, new);
 875
 876                 if ((vhost_need_event(vhost_used_event(vq), new, old) &&
 877                                         (vq->callfd >= 0)) ||
 878                                 unlikely(!signalled_used_valid)) {
 879                         eventfd_write(vq->callfd, (eventfd_t) 1);
 880                         if (dev->flags & VIRTIO_DEV_STATS_ENABLED)
 881                                 vq->stats.guest_notifications++;
 882                         if (dev->notify_ops->guest_notified)
 883                                 dev->notify_ops->guest_notified(dev->vid);
 884                 }
 885         } else {
 886                 /* Kick the guest if necessary. */
 887                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 888                                 && (vq->callfd >= 0)) {
 889                         eventfd_write(vq->callfd, (eventfd_t)1);
 890                         if (dev->flags & VIRTIO_DEV_STATS_ENABLED)
 891                                 vq->stats.guest_notifications++;
 892                         if (dev->notify_ops->guest_notified)
 893                                 dev->notify_ops->guest_notified(dev->vid);
 894                 }
 895         }
 896 }
 897
 898 static __rte_always_inline void
 899 vhost_vring_call_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
 900 {
 901         uint16_t old, new, off, off_wrap;
 902         bool signalled_used_valid, kick = false;
 903
 904         /* Flush used desc update. */
 905         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
 906
 907         if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
 908                 if (vq->driver_event->flags !=
 909                                 VRING_EVENT_F_DISABLE)
 910                         kick = true;
 911                 goto kick;
 912         }
 913
 914         old = vq->signalled_used;
 915         new = vq->last_used_idx;
 916         vq->signalled_used = new;
 917         signalled_used_valid = vq->signalled_used_valid;
 918         vq->signalled_used_valid = true;
 919
 920         if (vq->driver_event->flags != VRING_EVENT_F_DESC) {
 921                 if (vq->driver_event->flags != VRING_EVENT_F_DISABLE)
 922                         kick = true;
 923                 goto kick;
 924         }
 925
 926         if (unlikely(!signalled_used_valid)) {
 927                 kick = true;
 928                 goto kick;
 929         }
 930
 931         rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
 932
 933         off_wrap = vq->driver_event->off_wrap;
 934         off = off_wrap & ~(1 << 15);
 935
 936         if (new <= old)
 937                 old -= vq->size;
 938
 939         if (vq->used_wrap_counter != off_wrap >> 15)
 940                 off -= vq->size;
 941
 942         if (vhost_need_event(off, new, old))
 943                 kick = true;
 944 kick:
 945         if (kick) {
 946                 eventfd_write(vq->callfd, (eventfd_t)1);
 947                 if (dev->notify_ops->guest_notified)
 948                         dev->notify_ops->guest_notified(dev->vid);
 949         }
 950 }
 951
 952 static __rte_always_inline void
 953 free_ind_table(void *idesc)
 954 {
 955         rte_free(idesc);
 956 }
 957
 958 static __rte_always_inline void
 959 restore_mbuf(struct rte_mbuf *m)
 960 {
 961         uint32_t mbuf_size, priv_size;
 962
 963         while (m) {
 964                 priv_size = rte_pktmbuf_priv_size(m->pool);
 965                 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
 966                 /* start of buffer is after mbuf structure and priv data */
 967
 968                 m->buf_addr = (char *)m + mbuf_size;
 969                 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
 970                 m = m->next;
 971         }
 972 }
 973
 974 static __rte_always_inline bool
 975 mbuf_is_consumed(struct rte_mbuf *m)
 976 {
 977         while (m) {
 978                 if (rte_mbuf_refcnt_read(m) > 1)
 979                         return false;
 980                 m = m->next;
 981         }
 982
 983         return true;
 984 }
 985 #endif /* _VHOST_NET_CDEV_H_ */