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