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