ceph/src/seastar/dpdk/lib/librte_vhost/rte_vhost.h

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2017 Intel Corporation
   3  */
   4
   5 #ifndef _RTE_VHOST_H_
   6 #define _RTE_VHOST_H_
   7
   8 /**
   9  * @file
  10  * Interface to vhost-user
  11  */
  12
  13 #include <stdint.h>
  14 #include <sys/eventfd.h>
  15
  16 #include <rte_memory.h>
  17 #include <rte_mempool.h>
  18
  19 #ifdef __cplusplus
  20 extern "C" {
  21 #endif
  22
  23 /* These are not C++-aware. */
  24 #include <linux/vhost.h>
  25 #include <linux/virtio_ring.h>
  26
  27 #define RTE_VHOST_USER_CLIENT           (1ULL << 0)
  28 #define RTE_VHOST_USER_NO_RECONNECT     (1ULL << 1)
  29 #define RTE_VHOST_USER_DEQUEUE_ZERO_COPY        (1ULL << 2)
  30 #define RTE_VHOST_USER_IOMMU_SUPPORT    (1ULL << 3)
  31 #define RTE_VHOST_USER_POSTCOPY_SUPPORT         (1ULL << 4)
  32
  33 /** Protocol features. */
  34 #ifndef VHOST_USER_PROTOCOL_F_MQ
  35 #define VHOST_USER_PROTOCOL_F_MQ        0
  36 #endif
  37
  38 #ifndef VHOST_USER_PROTOCOL_F_LOG_SHMFD
  39 #define VHOST_USER_PROTOCOL_F_LOG_SHMFD 1
  40 #endif
  41
  42 #ifndef VHOST_USER_PROTOCOL_F_RARP
  43 #define VHOST_USER_PROTOCOL_F_RARP      2
  44 #endif
  45
  46 #ifndef VHOST_USER_PROTOCOL_F_REPLY_ACK
  47 #define VHOST_USER_PROTOCOL_F_REPLY_ACK 3
  48 #endif
  49
  50 #ifndef VHOST_USER_PROTOCOL_F_NET_MTU
  51 #define VHOST_USER_PROTOCOL_F_NET_MTU   4
  52 #endif
  53
  54 #ifndef VHOST_USER_PROTOCOL_F_SLAVE_REQ
  55 #define VHOST_USER_PROTOCOL_F_SLAVE_REQ 5
  56 #endif
  57
  58 #ifndef VHOST_USER_PROTOCOL_F_CRYPTO_SESSION
  59 #define VHOST_USER_PROTOCOL_F_CRYPTO_SESSION 7
  60 #endif
  61
  62 #ifndef VHOST_USER_PROTOCOL_F_PAGEFAULT
  63 #define VHOST_USER_PROTOCOL_F_PAGEFAULT 8
  64 #endif
  65
  66 #ifndef VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD
  67 #define VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD 10
  68 #endif
  69
  70 #ifndef VHOST_USER_PROTOCOL_F_HOST_NOTIFIER
  71 #define VHOST_USER_PROTOCOL_F_HOST_NOTIFIER 11
  72 #endif
  73
  74 /** Indicate whether protocol features negotiation is supported. */
  75 #ifndef VHOST_USER_F_PROTOCOL_FEATURES
  76 #define VHOST_USER_F_PROTOCOL_FEATURES  30
  77 #endif
  78
  79 /**
  80  * Information relating to memory regions including offsets to
  81  * addresses in QEMUs memory file.
  82  */
  83 struct rte_vhost_mem_region {
  84         uint64_t guest_phys_addr;
  85         uint64_t guest_user_addr;
  86         uint64_t host_user_addr;
  87         uint64_t size;
  88         void     *mmap_addr;
  89         uint64_t mmap_size;
  90         int fd;
  91 };
  92
  93 /**
  94  * Memory structure includes region and mapping information.
  95  */
  96 struct rte_vhost_memory {
  97         uint32_t nregions;
  98         struct rte_vhost_mem_region regions[];
  99 };
 100
 101 struct rte_vhost_vring {
 102         struct vring_desc       *desc;
 103         struct vring_avail      *avail;
 104         struct vring_used       *used;
 105         uint64_t                log_guest_addr;
 106
 107         /** Deprecated, use rte_vhost_vring_call() instead. */
 108         int                     callfd;
 109
 110         int                     kickfd;
 111         uint16_t                size;
 112 };
 113
 114 /**
 115  * Possible results of the vhost user message handling callbacks
 116  */
 117 enum rte_vhost_msg_result {
 118         /* Message handling failed */
 119         RTE_VHOST_MSG_RESULT_ERR = -1,
 120         /* Message handling successful */
 121         RTE_VHOST_MSG_RESULT_OK =  0,
 122         /* Message handling successful and reply prepared */
 123         RTE_VHOST_MSG_RESULT_REPLY =  1,
 124         /* Message not handled */
 125         RTE_VHOST_MSG_RESULT_NOT_HANDLED,
 126 };
 127
 128 /**
 129  * Function prototype for the vhost backend to handle specific vhost user
 130  * messages.
 131  *
 132  * @param vid
 133  *  vhost device id
 134  * @param msg
 135  *  Message pointer.
 136  * @return
 137  *  RTE_VHOST_MSG_RESULT_OK on success,
 138  *  RTE_VHOST_MSG_RESULT_REPLY on success with reply,
 139  *  RTE_VHOST_MSG_RESULT_ERR on failure,
 140  *  RTE_VHOST_MSG_RESULT_NOT_HANDLED if message was not handled.
 141  */
 142 typedef enum rte_vhost_msg_result (*rte_vhost_msg_handle)(int vid, void *msg);
 143
 144 /**
 145  * Optional vhost user message handlers.
 146  */
 147 struct rte_vhost_user_extern_ops {
 148         /* Called prior to the master message handling. */
 149         rte_vhost_msg_handle pre_msg_handle;
 150         /* Called after the master message handling. */
 151         rte_vhost_msg_handle post_msg_handle;
 152 };
 153
 154 /**
 155  * Device and vring operations.
 156  */
 157 struct vhost_device_ops {
 158         int (*new_device)(int vid);             /**< Add device. */
 159         void (*destroy_device)(int vid);        /**< Remove device. */
 160
 161         int (*vring_state_changed)(int vid, uint16_t queue_id, int enable);     /**< triggered when a vring is enabled or disabled */
 162
 163         /**
 164          * Features could be changed after the feature negotiation.
 165          * For example, VHOST_F_LOG_ALL will be set/cleared at the
 166          * start/end of live migration, respectively. This callback
 167          * is used to inform the application on such change.
 168          */
 169         int (*features_changed)(int vid, uint64_t features);
 170
 171         int (*new_connection)(int vid);
 172         void (*destroy_connection)(int vid);
 173
 174         void *reserved[2]; /**< Reserved for future extension */
 175 };
 176
 177 /**
 178  * Convert guest physical address to host virtual address
 179  *
 180  * This function is deprecated because unsafe.
 181  * New rte_vhost_va_from_guest_pa() should be used instead to ensure
 182  * guest physical ranges are fully and contiguously mapped into
 183  * process virtual address space.
 184  *
 185  * @param mem
 186  *  the guest memory regions
 187  * @param gpa
 188  *  the guest physical address for querying
 189  * @return
 190  *  the host virtual address on success, 0 on failure
 191  */
 192 __rte_deprecated
 193 static __rte_always_inline uint64_t
 194 rte_vhost_gpa_to_vva(struct rte_vhost_memory *mem, uint64_t gpa)
 195 {
 196         struct rte_vhost_mem_region *reg;
 197         uint32_t i;
 198
 199         for (i = 0; i < mem->nregions; i++) {
 200                 reg = &mem->regions[i];
 201                 if (gpa >= reg->guest_phys_addr &&
 202                     gpa <  reg->guest_phys_addr + reg->size) {
 203                         return gpa - reg->guest_phys_addr +
 204                                reg->host_user_addr;
 205                 }
 206         }
 207
 208         return 0;
 209 }
 210
 211 /**
 212  * Convert guest physical address to host virtual address safely
 213  *
 214  * This variant of rte_vhost_gpa_to_vva() takes care all the
 215  * requested length is mapped and contiguous in process address
 216  * space.
 217  *
 218  * @param mem
 219  *  the guest memory regions
 220  * @param gpa
 221  *  the guest physical address for querying
 222  * @param len
 223  *  the size of the requested area to map, updated with actual size mapped
 224  * @return
 225  *  the host virtual address on success, 0 on failure
 226  */
 227 static __rte_always_inline uint64_t
 228 rte_vhost_va_from_guest_pa(struct rte_vhost_memory *mem,
 229                                                    uint64_t gpa, uint64_t *len)
 230 {
 231         struct rte_vhost_mem_region *r;
 232         uint32_t i;
 233
 234         for (i = 0; i < mem->nregions; i++) {
 235                 r = &mem->regions[i];
 236                 if (gpa >= r->guest_phys_addr &&
 237                     gpa <  r->guest_phys_addr + r->size) {
 238
 239                         if (unlikely(*len > r->guest_phys_addr + r->size - gpa))
 240                                 *len = r->guest_phys_addr + r->size - gpa;
 241
 242                         return gpa - r->guest_phys_addr +
 243                                r->host_user_addr;
 244                 }
 245         }
 246         *len = 0;
 247
 248         return 0;
 249 }
 250
 251 #define RTE_VHOST_NEED_LOG(features)    ((features) & (1ULL << VHOST_F_LOG_ALL))
 252
 253 /**
 254  * Log the memory write start with given address.
 255  *
 256  * This function only need be invoked when the live migration starts.
 257  * Therefore, we won't need call it at all in the most of time. For
 258  * making the performance impact be minimum, it's suggested to do a
 259  * check before calling it:
 260  *
 261  *        if (unlikely(RTE_VHOST_NEED_LOG(features)))
 262  *                rte_vhost_log_write(vid, addr, len);
 263  *
 264  * @param vid
 265  *  vhost device ID
 266  * @param addr
 267  *  the starting address for write
 268  * @param len
 269  *  the length to write
 270  */
 271 void rte_vhost_log_write(int vid, uint64_t addr, uint64_t len);
 272
 273 /**
 274  * Log the used ring update start at given offset.
 275  *
 276  * Same as rte_vhost_log_write, it's suggested to do a check before
 277  * calling it:
 278  *
 279  *        if (unlikely(RTE_VHOST_NEED_LOG(features)))
 280  *                rte_vhost_log_used_vring(vid, vring_idx, offset, len);
 281  *
 282  * @param vid
 283  *  vhost device ID
 284  * @param vring_idx
 285  *  the vring index
 286  * @param offset
 287  *  the offset inside the used ring
 288  * @param len
 289  *  the length to write
 290  */
 291 void rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
 292                               uint64_t offset, uint64_t len);
 293
 294 int rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable);
 295
 296 /**
 297  * Register vhost driver. path could be different for multiple
 298  * instance support.
 299  */
 300 int rte_vhost_driver_register(const char *path, uint64_t flags);
 301
 302 /* Unregister vhost driver. This is only meaningful to vhost user. */
 303 int rte_vhost_driver_unregister(const char *path);
 304
 305 /**
 306  * Set the vdpa device id, enforce single connection per socket
 307  *
 308  * @param path
 309  *  The vhost-user socket file path
 310  * @param did
 311  *  Device id
 312  * @return
 313  *  0 on success, -1 on failure
 314  */
 315 int __rte_experimental
 316 rte_vhost_driver_attach_vdpa_device(const char *path, int did);
 317
 318 /**
 319  * Unset the vdpa device id
 320  *
 321  * @param path
 322  *  The vhost-user socket file path
 323  * @return
 324  *  0 on success, -1 on failure
 325  */
 326 int __rte_experimental
 327 rte_vhost_driver_detach_vdpa_device(const char *path);
 328
 329 /**
 330  * Get the device id
 331  *
 332  * @param path
 333  *  The vhost-user socket file path
 334  * @return
 335  *  Device id, -1 on failure
 336  */
 337 int __rte_experimental
 338 rte_vhost_driver_get_vdpa_device_id(const char *path);
 339
 340 /**
 341  * Set the feature bits the vhost-user driver supports.
 342  *
 343  * @param path
 344  *  The vhost-user socket file path
 345  * @param features
 346  *  Supported features
 347  * @return
 348  *  0 on success, -1 on failure
 349  */
 350 int rte_vhost_driver_set_features(const char *path, uint64_t features);
 351
 352 /**
 353  * Enable vhost-user driver features.
 354  *
 355  * Note that
 356  * - the param features should be a subset of the feature bits provided
 357  *   by rte_vhost_driver_set_features().
 358  * - it must be invoked before vhost-user negotiation starts.
 359  *
 360  * @param path
 361  *  The vhost-user socket file path
 362  * @param features
 363  *  Features to enable
 364  * @return
 365  *  0 on success, -1 on failure
 366  */
 367 int rte_vhost_driver_enable_features(const char *path, uint64_t features);
 368
 369 /**
 370  * Disable vhost-user driver features.
 371  *
 372  * The two notes at rte_vhost_driver_enable_features() also apply here.
 373  *
 374  * @param path
 375  *  The vhost-user socket file path
 376  * @param features
 377  *  Features to disable
 378  * @return
 379  *  0 on success, -1 on failure
 380  */
 381 int rte_vhost_driver_disable_features(const char *path, uint64_t features);
 382
 383 /**
 384  * Get the feature bits before feature negotiation.
 385  *
 386  * @param path
 387  *  The vhost-user socket file path
 388  * @param features
 389  *  A pointer to store the queried feature bits
 390  * @return
 391  *  0 on success, -1 on failure
 392  */
 393 int rte_vhost_driver_get_features(const char *path, uint64_t *features);
 394
 395 /**
 396  * Set the protocol feature bits before feature negotiation.
 397  *
 398  * @param path
 399  *  The vhost-user socket file path
 400  * @param protocol_features
 401  *  Supported protocol features
 402  * @return
 403  *  0 on success, -1 on failure
 404  */
 405 int __rte_experimental
 406 rte_vhost_driver_set_protocol_features(const char *path,
 407                 uint64_t protocol_features);
 408
 409 /**
 410  * Get the protocol feature bits before feature negotiation.
 411  *
 412  * @param path
 413  *  The vhost-user socket file path
 414  * @param protocol_features
 415  *  A pointer to store the queried protocol feature bits
 416  * @return
 417  *  0 on success, -1 on failure
 418  */
 419 int __rte_experimental
 420 rte_vhost_driver_get_protocol_features(const char *path,
 421                 uint64_t *protocol_features);
 422
 423 /**
 424  * Get the queue number bits before feature negotiation.
 425  *
 426  * @param path
 427  *  The vhost-user socket file path
 428  * @param queue_num
 429  *  A pointer to store the queried queue number bits
 430  * @return
 431  *  0 on success, -1 on failure
 432  */
 433 int __rte_experimental
 434 rte_vhost_driver_get_queue_num(const char *path, uint32_t *queue_num);
 435
 436 /**
 437  * Get the feature bits after negotiation
 438  *
 439  * @param vid
 440  *  Vhost device ID
 441  * @param features
 442  *  A pointer to store the queried feature bits
 443  * @return
 444  *  0 on success, -1 on failure
 445  */
 446 int rte_vhost_get_negotiated_features(int vid, uint64_t *features);
 447
 448 /* Register callbacks. */
 449 int rte_vhost_driver_callback_register(const char *path,
 450         struct vhost_device_ops const * const ops);
 451
 452 /**
 453  *
 454  * Start the vhost-user driver.
 455  *
 456  * This function triggers the vhost-user negotiation.
 457  *
 458  * @param path
 459  *  The vhost-user socket file path
 460  * @return
 461  *  0 on success, -1 on failure
 462  */
 463 int rte_vhost_driver_start(const char *path);
 464
 465 /**
 466  * Get the MTU value of the device if set in QEMU.
 467  *
 468  * @param vid
 469  *  virtio-net device ID
 470  * @param mtu
 471  *  The variable to store the MTU value
 472  *
 473  * @return
 474  *  0: success
 475  *  -EAGAIN: device not yet started
 476  *  -ENOTSUP: device does not support MTU feature
 477  */
 478 int rte_vhost_get_mtu(int vid, uint16_t *mtu);
 479
 480 /**
 481  * Get the numa node from which the virtio net device's memory
 482  * is allocated.
 483  *
 484  * @param vid
 485  *  vhost device ID
 486  *
 487  * @return
 488  *  The numa node, -1 on failure
 489  */
 490 int rte_vhost_get_numa_node(int vid);
 491
 492 /**
 493  * @deprecated
 494  * Get the number of queues the device supports.
 495  *
 496  * Note this function is deprecated, as it returns a queue pair number,
 497  * which is vhost specific. Instead, rte_vhost_get_vring_num should
 498  * be used.
 499  *
 500  * @param vid
 501  *  vhost device ID
 502  *
 503  * @return
 504  *  The number of queues, 0 on failure
 505  */
 506 __rte_deprecated
 507 uint32_t rte_vhost_get_queue_num(int vid);
 508
 509 /**
 510  * Get the number of vrings the device supports.
 511  *
 512  * @param vid
 513  *  vhost device ID
 514  *
 515  * @return
 516  *  The number of vrings, 0 on failure
 517  */
 518 uint16_t rte_vhost_get_vring_num(int vid);
 519
 520 /**
 521  * Get the virtio net device's ifname, which is the vhost-user socket
 522  * file path.
 523  *
 524  * @param vid
 525  *  vhost device ID
 526  * @param buf
 527  *  The buffer to stored the queried ifname
 528  * @param len
 529  *  The length of buf
 530  *
 531  * @return
 532  *  0 on success, -1 on failure
 533  */
 534 int rte_vhost_get_ifname(int vid, char *buf, size_t len);
 535
 536 /**
 537  * Get how many avail entries are left in the queue
 538  *
 539  * @param vid
 540  *  vhost device ID
 541  * @param queue_id
 542  *  virtio queue index
 543  *
 544  * @return
 545  *  num of avail entries left
 546  */
 547 uint16_t rte_vhost_avail_entries(int vid, uint16_t queue_id);
 548
 549 struct rte_mbuf;
 550 struct rte_mempool;
 551 /**
 552  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 553  * be received from the physical port or from another virtual device. A packet
 554  * count is returned to indicate the number of packets that were successfully
 555  * added to the RX queue.
 556  * @param vid
 557  *  vhost device ID
 558  * @param queue_id
 559  *  virtio queue index in mq case
 560  * @param pkts
 561  *  array to contain packets to be enqueued
 562  * @param count
 563  *  packets num to be enqueued
 564  * @return
 565  *  num of packets enqueued
 566  */
 567 uint16_t rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 568         struct rte_mbuf **pkts, uint16_t count);
 569
 570 /**
 571  * This function gets guest buffers from the virtio device TX virtqueue,
 572  * construct host mbufs, copies guest buffer content to host mbufs and
 573  * store them in pkts to be processed.
 574  * @param vid
 575  *  vhost device ID
 576  * @param queue_id
 577  *  virtio queue index in mq case
 578  * @param mbuf_pool
 579  *  mbuf_pool where host mbuf is allocated.
 580  * @param pkts
 581  *  array to contain packets to be dequeued
 582  * @param count
 583  *  packets num to be dequeued
 584  * @return
 585  *  num of packets dequeued
 586  */
 587 uint16_t rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 588         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count);
 589
 590 /**
 591  * Get guest mem table: a list of memory regions.
 592  *
 593  * An rte_vhost_vhost_memory object will be allocated internally, to hold the
 594  * guest memory regions. Application should free it at destroy_device()
 595  * callback.
 596  *
 597  * @param vid
 598  *  vhost device ID
 599  * @param mem
 600  *  To store the returned mem regions
 601  * @return
 602  *  0 on success, -1 on failure
 603  */
 604 int rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem);
 605
 606 /**
 607  * Get guest vring info, including the vring address, vring size, etc.
 608  *
 609  * @param vid
 610  *  vhost device ID
 611  * @param vring_idx
 612  *  vring index
 613  * @param vring
 614  *  the structure to hold the requested vring info
 615  * @return
 616  *  0 on success, -1 on failure
 617  */
 618 int rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
 619                               struct rte_vhost_vring *vring);
 620
 621 /**
 622  * Notify the guest that used descriptors have been added to the vring.  This
 623  * function acts as a memory barrier.
 624  *
 625  * @param vid
 626  *  vhost device ID
 627  * @param vring_idx
 628  *  vring index
 629  * @return
 630  *  0 on success, -1 on failure
 631  */
 632 int rte_vhost_vring_call(int vid, uint16_t vring_idx);
 633
 634 /**
 635  * Get vhost RX queue avail count.
 636  *
 637  * @param vid
 638  *  vhost device ID
 639  * @param qid
 640  *  virtio queue index in mq case
 641  * @return
 642  *  num of desc available
 643  */
 644 uint32_t rte_vhost_rx_queue_count(int vid, uint16_t qid);
 645
 646 /**
 647  * Get log base and log size of the vhost device
 648  *
 649  * @param vid
 650  *  vhost device ID
 651  * @param log_base
 652  *  vhost log base
 653  * @param log_size
 654  *  vhost log size
 655  * @return
 656  *  0 on success, -1 on failure
 657  */
 658 int __rte_experimental
 659 rte_vhost_get_log_base(int vid, uint64_t *log_base, uint64_t *log_size);
 660
 661 /**
 662  * Get last_avail/used_idx of the vhost virtqueue
 663  *
 664  * @param vid
 665  *  vhost device ID
 666  * @param queue_id
 667  *  vhost queue index
 668  * @param last_avail_idx
 669  *  vhost last_avail_idx to get
 670  * @param last_used_idx
 671  *  vhost last_used_idx to get
 672  * @return
 673  *  0 on success, -1 on failure
 674  */
 675 int __rte_experimental
 676 rte_vhost_get_vring_base(int vid, uint16_t queue_id,
 677                 uint16_t *last_avail_idx, uint16_t *last_used_idx);
 678
 679 /**
 680  * Set last_avail/used_idx of the vhost virtqueue
 681  *
 682  * @param vid
 683  *  vhost device ID
 684  * @param queue_id
 685  *  vhost queue index
 686  * @param last_avail_idx
 687  *  last_avail_idx to set
 688  * @param last_used_idx
 689  *  last_used_idx to set
 690  * @return
 691  *  0 on success, -1 on failure
 692  */
 693 int __rte_experimental
 694 rte_vhost_set_vring_base(int vid, uint16_t queue_id,
 695                 uint16_t last_avail_idx, uint16_t last_used_idx);
 696
 697 /**
 698  * Register external message handling callbacks
 699  *
 700  * @param vid
 701  *  vhost device ID
 702  * @param ops
 703  *  virtio external callbacks to register
 704  * @param ctx
 705  *  additional context passed to the callbacks
 706  * @return
 707  *  0 on success, -1 on failure
 708  */
 709 int __rte_experimental
 710 rte_vhost_extern_callback_register(int vid,
 711                 struct rte_vhost_user_extern_ops const * const ops, void *ctx);
 712
 713 /**
 714  * Get vdpa device id for vhost device.
 715  *
 716  * @param vid
 717  *  vhost device id
 718  * @return
 719  *  device id
 720  */
 721 int __rte_experimental
 722 rte_vhost_get_vdpa_device_id(int vid);
 723
 724 #ifdef __cplusplus
 725 }
 726 #endif
 727
 728 #endif /* _RTE_VHOST_H_ */