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