1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2013 Chris Torek <torek @ torek net> 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 * 28 * $FreeBSD$ 29 */ 30 31 #ifndef _VIRTIO_H_ 32 #define _VIRTIO_H_ 33 34 #include <pthread_np.h> 35 #include <machine/atomic.h> 36 37 /* 38 * These are derived from several virtio specifications. 39 * 40 * Some useful links: 41 * https://github.com/rustyrussell/virtio-spec 42 * http://people.redhat.com/pbonzini/virtio-spec.pdf 43 */ 44 45 /* 46 * A virtual device has zero or more "virtual queues" (virtqueue). 47 * Each virtqueue uses at least two 4096-byte pages, laid out thus: 48 * 49 * +-----------------------------------------------+ 50 * | "desc": <N> descriptors, 16 bytes each | 51 * | ----------------------------------------- | 52 * | "avail": 2 uint16; <N> uint16; 1 uint16 | 53 * | ----------------------------------------- | 54 * | pad to 4k boundary | 55 * +-----------------------------------------------+ 56 * | "used": 2 x uint16; <N> elems; 1 uint16 | 57 * | ----------------------------------------- | 58 * | pad to 4k boundary | 59 * +-----------------------------------------------+ 60 * 61 * The number <N> that appears here is always a power of two and is 62 * limited to no more than 32768 (as it must fit in a 16-bit field). 63 * If <N> is sufficiently large, the above will occupy more than 64 * two pages. In any case, all pages must be physically contiguous 65 * within the guest's physical address space. 66 * 67 * The <N> 16-byte "desc" descriptors consist of a 64-bit guest 68 * physical address <addr>, a 32-bit length <len>, a 16-bit 69 * <flags>, and a 16-bit <next> field (all in guest byte order). 70 * 71 * There are three flags that may be set : 72 * NEXT descriptor is chained, so use its "next" field 73 * WRITE descriptor is for host to write into guest RAM 74 * (else host is to read from guest RAM) 75 * INDIRECT descriptor address field is (guest physical) 76 * address of a linear array of descriptors 77 * 78 * Unless INDIRECT is set, <len> is the number of bytes that may 79 * be read/written from guest physical address <addr>. If 80 * INDIRECT is set, WRITE is ignored and <len> provides the length 81 * of the indirect descriptors (and <len> must be a multiple of 82 * 16). Note that NEXT may still be set in the main descriptor 83 * pointing to the indirect, and should be set in each indirect 84 * descriptor that uses the next descriptor (these should generally 85 * be numbered sequentially). However, INDIRECT must not be set 86 * in the indirect descriptors. Upon reaching an indirect descriptor 87 * without a NEXT bit, control returns to the direct descriptors. 88 * 89 * Except inside an indirect, each <next> value must be in the 90 * range [0 .. N) (i.e., the half-open interval). (Inside an 91 * indirect, each <next> must be in the range [0 .. <len>/16).) 92 * 93 * The "avail" data structures reside in the same pages as the 94 * "desc" structures since both together are used by the device to 95 * pass information to the hypervisor's virtual driver. These 96 * begin with a 16-bit <flags> field and 16-bit index <idx>, then 97 * have <N> 16-bit <ring> values, followed by one final 16-bit 98 * field <used_event>. The <N> <ring> entries are simply indices 99 * indices into the descriptor ring (and thus must meet the same 100 * constraints as each <next> value). However, <idx> is counted 101 * up from 0 (initially) and simply wraps around after 65535; it 102 * is taken mod <N> to find the next available entry. 103 * 104 * The "used" ring occupies a separate page or pages, and contains 105 * values written from the virtual driver back to the guest OS. 106 * This begins with a 16-bit <flags> and 16-bit <idx>, then there 107 * are <N> "vring_used" elements, followed by a 16-bit <avail_event>. 108 * The <N> "vring_used" elements consist of a 32-bit <id> and a 109 * 32-bit <len> (vu_tlen below). The <id> is simply the index of 110 * the head of a descriptor chain the guest made available 111 * earlier, and the <len> is the number of bytes actually written, 112 * e.g., in the case of a network driver that provided a large 113 * receive buffer but received only a small amount of data. 114 * 115 * The two event fields, <used_event> and <avail_event>, in the 116 * avail and used rings (respectively -- note the reversal!), are 117 * always provided, but are used only if the virtual device 118 * negotiates the VIRTIO_RING_F_EVENT_IDX feature during feature 119 * negotiation. Similarly, both rings provide a flag -- 120 * VRING_AVAIL_F_NO_INTERRUPT and VRING_USED_F_NO_NOTIFY -- in 121 * their <flags> field, indicating that the guest does not need an 122 * interrupt, or that the hypervisor driver does not need a 123 * notify, when descriptors are added to the corresponding ring. 124 * (These are provided only for interrupt optimization and need 125 * not be implemented.) 126 */ 127 #define VRING_ALIGN 4096 128 129 #define VRING_DESC_F_NEXT (1 << 0) 130 #define VRING_DESC_F_WRITE (1 << 1) 131 #define VRING_DESC_F_INDIRECT (1 << 2) 132 133 struct virtio_desc { /* AKA vring_desc */ 134 uint64_t vd_addr; /* guest physical address */ 135 uint32_t vd_len; /* length of scatter/gather seg */ 136 uint16_t vd_flags; /* VRING_F_DESC_* */ 137 uint16_t vd_next; /* next desc if F_NEXT */ 138 } __packed; 139 140 struct virtio_used { /* AKA vring_used_elem */ 141 uint32_t vu_idx; /* head of used descriptor chain */ 142 uint32_t vu_tlen; /* length written-to */ 143 } __packed; 144 145 #define VRING_AVAIL_F_NO_INTERRUPT 1 146 147 struct vring_avail { 148 uint16_t va_flags; /* VRING_AVAIL_F_* */ 149 uint16_t va_idx; /* counts to 65535, then cycles */ 150 uint16_t va_ring[]; /* size N, reported in QNUM value */ 151 /* uint16_t va_used_event; -- after N ring entries */ 152 } __packed; 153 154 #define VRING_USED_F_NO_NOTIFY 1 155 struct vring_used { 156 uint16_t vu_flags; /* VRING_USED_F_* */ 157 uint16_t vu_idx; /* counts to 65535, then cycles */ 158 struct virtio_used vu_ring[]; /* size N */ 159 /* uint16_t vu_avail_event; -- after N ring entries */ 160 } __packed; 161 162 /* 163 * The address of any given virtual queue is determined by a single 164 * Page Frame Number register. The guest writes the PFN into the 165 * PCI config space. However, a device that has two or more 166 * virtqueues can have a different PFN, and size, for each queue. 167 * The number of queues is determinable via the PCI config space 168 * VTCFG_R_QSEL register. Writes to QSEL select the queue: 0 means 169 * queue #0, 1 means queue#1, etc. Once a queue is selected, the 170 * remaining PFN and QNUM registers refer to that queue. 171 * 172 * QNUM is a read-only register containing a nonzero power of two 173 * that indicates the (hypervisor's) queue size. Or, if reading it 174 * produces zero, the hypervisor does not have a corresponding 175 * queue. (The number of possible queues depends on the virtual 176 * device. The block device has just one; the network device 177 * provides either two -- 0 = receive, 1 = transmit -- or three, 178 * with 2 = control.) 179 * 180 * PFN is a read/write register giving the physical page address of 181 * the virtqueue in guest memory (the guest must allocate enough space 182 * based on the hypervisor's provided QNUM). 183 * 184 * QNOTIFY is effectively write-only: when the guest writes a queue 185 * number to the register, the hypervisor should scan the specified 186 * virtqueue. (Reading QNOTIFY currently always gets 0). 187 */ 188 189 /* 190 * PFN register shift amount 191 */ 192 #define VRING_PFN 12 193 194 /* 195 * Virtio device types 196 * 197 * XXX Should really be merged with <dev/virtio/virtio.h> defines 198 */ 199 #define VIRTIO_TYPE_NET 1 200 #define VIRTIO_TYPE_BLOCK 2 201 #define VIRTIO_TYPE_CONSOLE 3 202 #define VIRTIO_TYPE_ENTROPY 4 203 #define VIRTIO_TYPE_BALLOON 5 204 #define VIRTIO_TYPE_IOMEMORY 6 205 #define VIRTIO_TYPE_RPMSG 7 206 #define VIRTIO_TYPE_SCSI 8 207 #define VIRTIO_TYPE_9P 9 208 209 /* experimental IDs start at 65535 and work down */ 210 211 /* 212 * PCI vendor/device IDs 213 */ 214 #define VIRTIO_VENDOR 0x1AF4 215 #define VIRTIO_DEV_NET 0x1000 216 #define VIRTIO_DEV_BLOCK 0x1001 217 #define VIRTIO_DEV_CONSOLE 0x1003 218 #define VIRTIO_DEV_RANDOM 0x1005 219 #define VIRTIO_DEV_SCSI 0x1008 220 221 /* 222 * PCI config space constants. 223 * 224 * If MSI-X is enabled, the ISR register is generally not used, 225 * and the configuration vector and queue vector appear at offsets 226 * 20 and 22 with the remaining configuration registers at 24. 227 * If MSI-X is not enabled, those two registers disappear and 228 * the remaining configuration registers start at offset 20. 229 */ 230 #define VTCFG_R_HOSTCAP 0 231 #define VTCFG_R_GUESTCAP 4 232 #define VTCFG_R_PFN 8 233 #define VTCFG_R_QNUM 12 234 #define VTCFG_R_QSEL 14 235 #define VTCFG_R_QNOTIFY 16 236 #define VTCFG_R_STATUS 18 237 #define VTCFG_R_ISR 19 238 #define VTCFG_R_CFGVEC 20 239 #define VTCFG_R_QVEC 22 240 #define VTCFG_R_CFG0 20 /* No MSI-X */ 241 #define VTCFG_R_CFG1 24 /* With MSI-X */ 242 #define VTCFG_R_MSIX 20 243 244 /* 245 * Bits in VTCFG_R_STATUS. Guests need not actually set any of these, 246 * but a guest writing 0 to this register means "please reset". 247 */ 248 #define VTCFG_STATUS_ACK 0x01 /* guest OS has acknowledged dev */ 249 #define VTCFG_STATUS_DRIVER 0x02 /* guest OS driver is loaded */ 250 #define VTCFG_STATUS_DRIVER_OK 0x04 /* guest OS driver ready */ 251 #define VTCFG_STATUS_FAILED 0x80 /* guest has given up on this dev */ 252 253 /* 254 * Bits in VTCFG_R_ISR. These apply only if not using MSI-X. 255 * 256 * (We don't [yet?] ever use CONF_CHANGED.) 257 */ 258 #define VTCFG_ISR_QUEUES 0x01 /* re-scan queues */ 259 #define VTCFG_ISR_CONF_CHANGED 0x80 /* configuration changed */ 260 261 #define VIRTIO_MSI_NO_VECTOR 0xFFFF 262 263 /* 264 * Feature flags. 265 * Note: bits 0 through 23 are reserved to each device type. 266 */ 267 #define VIRTIO_F_NOTIFY_ON_EMPTY (1 << 24) 268 #define VIRTIO_RING_F_INDIRECT_DESC (1 << 28) 269 #define VIRTIO_RING_F_EVENT_IDX (1 << 29) 270 271 /* From section 2.3, "Virtqueue Configuration", of the virtio specification */ 272 static inline size_t 273 vring_size(u_int qsz) 274 { 275 size_t size; 276 277 /* constant 3 below = va_flags, va_idx, va_used_event */ 278 size = sizeof(struct virtio_desc) * qsz + sizeof(uint16_t) * (3 + qsz); 279 size = roundup2(size, VRING_ALIGN); 280 281 /* constant 3 below = vu_flags, vu_idx, vu_avail_event */ 282 size += sizeof(uint16_t) * 3 + sizeof(struct virtio_used) * qsz; 283 size = roundup2(size, VRING_ALIGN); 284 285 return (size); 286 } 287 288 struct vmctx; 289 struct pci_devinst; 290 struct vqueue_info; 291 292 /* 293 * A virtual device, with some number (possibly 0) of virtual 294 * queues and some size (possibly 0) of configuration-space 295 * registers private to the device. The virtio_softc should come 296 * at the front of each "derived class", so that a pointer to the 297 * virtio_softc is also a pointer to the more specific, derived- 298 * from-virtio driver's softc. 299 * 300 * Note: inside each hypervisor virtio driver, changes to these 301 * data structures must be locked against other threads, if any. 302 * Except for PCI config space register read/write, we assume each 303 * driver does the required locking, but we need a pointer to the 304 * lock (if there is one) for PCI config space read/write ops. 305 * 306 * When the guest reads or writes the device's config space, the 307 * generic layer checks for operations on the special registers 308 * described above. If the offset of the register(s) being read 309 * or written is past the CFG area (CFG0 or CFG1), the request is 310 * passed on to the virtual device, after subtracting off the 311 * generic-layer size. (So, drivers can just use the offset as 312 * an offset into "struct config", for instance.) 313 * 314 * (The virtio layer also makes sure that the read or write is to/ 315 * from a "good" config offset, hence vc_cfgsize, and on BAR #0. 316 * However, the driver must verify the read or write size and offset 317 * and that no one is writing a readonly register.) 318 * 319 * The BROKED flag ("this thing done gone and broked") is for future 320 * use. 321 */ 322 #define VIRTIO_USE_MSIX 0x01 323 #define VIRTIO_EVENT_IDX 0x02 /* use the event-index values */ 324 #define VIRTIO_BROKED 0x08 /* ??? */ 325 326 struct virtio_softc { 327 struct virtio_consts *vs_vc; /* constants (see below) */ 328 int vs_flags; /* VIRTIO_* flags from above */ 329 pthread_mutex_t *vs_mtx; /* POSIX mutex, if any */ 330 struct pci_devinst *vs_pi; /* PCI device instance */ 331 uint32_t vs_negotiated_caps; /* negotiated capabilities */ 332 struct vqueue_info *vs_queues; /* one per vc_nvq */ 333 int vs_curq; /* current queue */ 334 uint8_t vs_status; /* value from last status write */ 335 uint8_t vs_isr; /* ISR flags, if not MSI-X */ 336 uint16_t vs_msix_cfg_idx; /* MSI-X vector for config event */ 337 }; 338 339 #define VS_LOCK(vs) \ 340 do { \ 341 if (vs->vs_mtx) \ 342 pthread_mutex_lock(vs->vs_mtx); \ 343 } while (0) 344 345 #define VS_UNLOCK(vs) \ 346 do { \ 347 if (vs->vs_mtx) \ 348 pthread_mutex_unlock(vs->vs_mtx); \ 349 } while (0) 350 351 struct virtio_consts { 352 const char *vc_name; /* name of driver (for diagnostics) */ 353 int vc_nvq; /* number of virtual queues */ 354 size_t vc_cfgsize; /* size of dev-specific config regs */ 355 void (*vc_reset)(void *); /* called on virtual device reset */ 356 void (*vc_qnotify)(void *, struct vqueue_info *); 357 /* called on QNOTIFY if no VQ notify */ 358 int (*vc_cfgread)(void *, int, int, uint32_t *); 359 /* called to read config regs */ 360 int (*vc_cfgwrite)(void *, int, int, uint32_t); 361 /* called to write config regs */ 362 void (*vc_apply_features)(void *, uint64_t); 363 /* called to apply negotiated features */ 364 uint64_t vc_hv_caps; /* hypervisor-provided capabilities */ 365 }; 366 367 /* 368 * Data structure allocated (statically) per virtual queue. 369 * 370 * Drivers may change vq_qsize after a reset. When the guest OS 371 * requests a device reset, the hypervisor first calls 372 * vs->vs_vc->vc_reset(); then the data structure below is 373 * reinitialized (for each virtqueue: vs->vs_vc->vc_nvq). 374 * 375 * The remaining fields should only be fussed-with by the generic 376 * code. 377 * 378 * Note: the addresses of vq_desc, vq_avail, and vq_used are all 379 * computable from each other, but it's a lot simpler if we just 380 * keep a pointer to each one. The event indices are similarly 381 * (but more easily) computable, and this time we'll compute them: 382 * they're just XX_ring[N]. 383 */ 384 #define VQ_ALLOC 0x01 /* set once we have a pfn */ 385 #define VQ_BROKED 0x02 /* ??? */ 386 struct vqueue_info { 387 uint16_t vq_qsize; /* size of this queue (a power of 2) */ 388 void (*vq_notify)(void *, struct vqueue_info *); 389 /* called instead of vc_notify, if not NULL */ 390 391 struct virtio_softc *vq_vs; /* backpointer to softc */ 392 uint16_t vq_num; /* we're the num'th queue in the softc */ 393 394 uint16_t vq_flags; /* flags (see above) */ 395 uint16_t vq_last_avail; /* a recent value of vq_avail->va_idx */ 396 uint16_t vq_next_used; /* index of the next used slot to be filled */ 397 uint16_t vq_save_used; /* saved vq_used->vu_idx; see vq_endchains */ 398 uint16_t vq_msix_idx; /* MSI-X index, or VIRTIO_MSI_NO_VECTOR */ 399 400 uint32_t vq_pfn; /* PFN of virt queue (not shifted!) */ 401 402 volatile struct virtio_desc *vq_desc; /* descriptor array */ 403 volatile struct vring_avail *vq_avail; /* the "avail" ring */ 404 volatile struct vring_used *vq_used; /* the "used" ring */ 405 406 }; 407 /* as noted above, these are sort of backwards, name-wise */ 408 #define VQ_AVAIL_EVENT_IDX(vq) \ 409 (*(volatile uint16_t *)&(vq)->vq_used->vu_ring[(vq)->vq_qsize]) 410 #define VQ_USED_EVENT_IDX(vq) \ 411 ((vq)->vq_avail->va_ring[(vq)->vq_qsize]) 412 413 /* 414 * Is this ring ready for I/O? 415 */ 416 static inline int 417 vq_ring_ready(struct vqueue_info *vq) 418 { 419 420 return (vq->vq_flags & VQ_ALLOC); 421 } 422 423 /* 424 * Are there "available" descriptors? (This does not count 425 * how many, just returns True if there are some.) 426 */ 427 static inline int 428 vq_has_descs(struct vqueue_info *vq) 429 { 430 431 return (vq_ring_ready(vq) && vq->vq_last_avail != 432 vq->vq_avail->va_idx); 433 } 434 435 /* 436 * Deliver an interrupt to guest on the given virtual queue 437 * (if possible, or a generic MSI interrupt if not using MSI-X). 438 */ 439 static inline void 440 vq_interrupt(struct virtio_softc *vs, struct vqueue_info *vq) 441 { 442 443 if (pci_msix_enabled(vs->vs_pi)) 444 pci_generate_msix(vs->vs_pi, vq->vq_msix_idx); 445 else { 446 #ifndef __FreeBSD__ 447 boolean_t unlock = B_FALSE; 448 449 if (vs->vs_mtx && !pthread_mutex_isowned_np(vs->vs_mtx)) { 450 unlock = B_TRUE; 451 pthread_mutex_lock(vs->vs_mtx); 452 } 453 #else 454 VS_LOCK(vs); 455 #endif 456 vs->vs_isr |= VTCFG_ISR_QUEUES; 457 pci_generate_msi(vs->vs_pi, 0); 458 pci_lintr_assert(vs->vs_pi); 459 #ifndef __FreeBSD__ 460 if (unlock) 461 pthread_mutex_unlock(vs->vs_mtx); 462 #else 463 VS_UNLOCK(vs); 464 #endif 465 } 466 } 467 468 static inline void 469 vq_kick_enable(struct vqueue_info *vq) 470 { 471 472 vq->vq_used->vu_flags &= ~VRING_USED_F_NO_NOTIFY; 473 /* 474 * Full memory barrier to make sure the store to vu_flags 475 * happens before the load from va_idx, which results from 476 * a subsequent call to vq_has_descs(). 477 */ 478 atomic_thread_fence_seq_cst(); 479 } 480 481 static inline void 482 vq_kick_disable(struct vqueue_info *vq) 483 { 484 485 vq->vq_used->vu_flags |= VRING_USED_F_NO_NOTIFY; 486 } 487 488 struct iovec; 489 void vi_softc_linkup(struct virtio_softc *vs, struct virtio_consts *vc, 490 void *dev_softc, struct pci_devinst *pi, 491 struct vqueue_info *queues); 492 int vi_intr_init(struct virtio_softc *vs, int barnum, int use_msix); 493 void vi_reset_dev(struct virtio_softc *); 494 void vi_set_io_bar(struct virtio_softc *, int); 495 496 int vq_getchain(struct vqueue_info *vq, uint16_t *pidx, 497 struct iovec *iov, int n_iov, uint16_t *flags); 498 void vq_retchains(struct vqueue_info *vq, uint16_t n_chains); 499 void vq_relchain_prepare(struct vqueue_info *vq, uint16_t idx, 500 uint32_t iolen); 501 void vq_relchain_publish(struct vqueue_info *vq); 502 void vq_relchain(struct vqueue_info *vq, uint16_t idx, uint32_t iolen); 503 void vq_endchains(struct vqueue_info *vq, int used_all_avail); 504 505 uint64_t vi_pci_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi, 506 int baridx, uint64_t offset, int size); 507 void vi_pci_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi, 508 int baridx, uint64_t offset, int size, uint64_t value); 509 #endif /* _VIRTIO_H_ */ 510