2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2013 Chris Torek <torek @ torek net>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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.
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
35 * These are derived from several virtio specifications.
38 * https://github.com/rustyrussell/virtio-spec
39 * http://people.redhat.com/pbonzini/virtio-spec.pdf
43 * A virtual device has zero or more "virtual queues" (virtqueue).
44 * Each virtqueue uses at least two 4096-byte pages, laid out thus:
46 * +-----------------------------------------------+
47 * | "desc": <N> descriptors, 16 bytes each |
48 * | ----------------------------------------- |
49 * | "avail": 2 uint16; <N> uint16; 1 uint16 |
50 * | ----------------------------------------- |
51 * | pad to 4k boundary |
52 * +-----------------------------------------------+
53 * | "used": 2 x uint16; <N> elems; 1 uint16 |
54 * | ----------------------------------------- |
55 * | pad to 4k boundary |
56 * +-----------------------------------------------+
58 * The number <N> that appears here is always a power of two and is
59 * limited to no more than 32768 (as it must fit in a 16-bit field).
60 * If <N> is sufficiently large, the above will occupy more than
61 * two pages. In any case, all pages must be physically contiguous
62 * within the guest's physical address space.
64 * The <N> 16-byte "desc" descriptors consist of a 64-bit guest
65 * physical address <addr>, a 32-bit length <len>, a 16-bit
66 * <flags>, and a 16-bit <next> field (all in guest byte order).
68 * There are three flags that may be set :
69 * NEXT descriptor is chained, so use its "next" field
70 * WRITE descriptor is for host to write into guest RAM
71 * (else host is to read from guest RAM)
72 * INDIRECT descriptor address field is (guest physical)
73 * address of a linear array of descriptors
75 * Unless INDIRECT is set, <len> is the number of bytes that may
76 * be read/written from guest physical address <addr>. If
77 * INDIRECT is set, WRITE is ignored and <len> provides the length
78 * of the indirect descriptors (and <len> must be a multiple of
79 * 16). Note that NEXT may still be set in the main descriptor
80 * pointing to the indirect, and should be set in each indirect
81 * descriptor that uses the next descriptor (these should generally
82 * be numbered sequentially). However, INDIRECT must not be set
83 * in the indirect descriptors. Upon reaching an indirect descriptor
84 * without a NEXT bit, control returns to the direct descriptors.
86 * Except inside an indirect, each <next> value must be in the
87 * range [0 .. N) (i.e., the half-open interval). (Inside an
88 * indirect, each <next> must be in the range [0 .. <len>/16).)
90 * The "avail" data structures reside in the same pages as the
91 * "desc" structures since both together are used by the device to
92 * pass information to the hypervisor's virtual driver. These
93 * begin with a 16-bit <flags> field and 16-bit index <idx>, then
94 * have <N> 16-bit <ring> values, followed by one final 16-bit
95 * field <used_event>. The <N> <ring> entries are simply indices
96 * indices into the descriptor ring (and thus must meet the same
97 * constraints as each <next> value). However, <idx> is counted
98 * up from 0 (initially) and simply wraps around after 65535; it
99 * is taken mod <N> to find the next available entry.
101 * The "used" ring occupies a separate page or pages, and contains
102 * values written from the virtual driver back to the guest OS.
103 * This begins with a 16-bit <flags> and 16-bit <idx>, then there
104 * are <N> "vring_used" elements, followed by a 16-bit <avail_event>.
105 * The <N> "vring_used" elements consist of a 32-bit <id> and a
106 * 32-bit <len> (vu_tlen below). The <id> is simply the index of
107 * the head of a descriptor chain the guest made available
108 * earlier, and the <len> is the number of bytes actually written,
109 * e.g., in the case of a network driver that provided a large
110 * receive buffer but received only a small amount of data.
112 * The two event fields, <used_event> and <avail_event>, in the
113 * avail and used rings (respectively -- note the reversal!), are
114 * always provided, but are used only if the virtual device
115 * negotiates the VIRTIO_RING_F_EVENT_IDX feature during feature
116 * negotiation. Similarly, both rings provide a flag --
117 * VRING_AVAIL_F_NO_INTERRUPT and VRING_USED_F_NO_NOTIFY -- in
118 * their <flags> field, indicating that the guest does not need an
119 * interrupt, or that the hypervisor driver does not need a
120 * notify, when descriptors are added to the corresponding ring.
121 * (These are provided only for interrupt optimization and need
122 * not be implemented.)
124 #define VRING_ALIGN 4096
126 #define VRING_DESC_F_NEXT (1 << 0)
127 #define VRING_DESC_F_WRITE (1 << 1)
128 #define VRING_DESC_F_INDIRECT (1 << 2)
130 struct virtio_desc { /* AKA vring_desc */
131 uint64_t vd_addr; /* guest physical address */
132 uint32_t vd_len; /* length of scatter/gather seg */
133 uint16_t vd_flags; /* VRING_F_DESC_* */
134 uint16_t vd_next; /* next desc if F_NEXT */
137 struct virtio_used { /* AKA vring_used_elem */
138 uint32_t vu_idx; /* head of used descriptor chain */
139 uint32_t vu_tlen; /* length written-to */
142 #define VRING_AVAIL_F_NO_INTERRUPT 1
145 uint16_t va_flags; /* VRING_AVAIL_F_* */
146 uint16_t va_idx; /* counts to 65535, then cycles */
147 uint16_t va_ring[]; /* size N, reported in QNUM value */
148 /* uint16_t va_used_event; -- after N ring entries */
151 #define VRING_USED_F_NO_NOTIFY 1
153 uint16_t vu_flags; /* VRING_USED_F_* */
154 uint16_t vu_idx; /* counts to 65535, then cycles */
155 struct virtio_used vu_ring[]; /* size N */
156 /* uint16_t vu_avail_event; -- after N ring entries */
160 * The address of any given virtual queue is determined by a single
161 * Page Frame Number register. The guest writes the PFN into the
162 * PCI config space. However, a device that has two or more
163 * virtqueues can have a different PFN, and size, for each queue.
164 * The number of queues is determinable via the PCI config space
165 * VTCFG_R_QSEL register. Writes to QSEL select the queue: 0 means
166 * queue #0, 1 means queue#1, etc. Once a queue is selected, the
167 * remaining PFN and QNUM registers refer to that queue.
169 * QNUM is a read-only register containing a nonzero power of two
170 * that indicates the (hypervisor's) queue size. Or, if reading it
171 * produces zero, the hypervisor does not have a corresponding
172 * queue. (The number of possible queues depends on the virtual
173 * device. The block device has just one; the network device
174 * provides either two -- 0 = receive, 1 = transmit -- or three,
177 * PFN is a read/write register giving the physical page address of
178 * the virtqueue in guest memory (the guest must allocate enough space
179 * based on the hypervisor's provided QNUM).
181 * QNOTIFY is effectively write-only: when the guest writes a queue
182 * number to the register, the hypervisor should scan the specified
183 * virtqueue. (Reading QNOTIFY currently always gets 0).
187 * PFN register shift amount
192 * Virtio device types
194 * XXX Should really be merged with <dev/virtio/virtio.h> defines
196 #define VIRTIO_TYPE_NET 1
197 #define VIRTIO_TYPE_BLOCK 2
198 #define VIRTIO_TYPE_CONSOLE 3
199 #define VIRTIO_TYPE_ENTROPY 4
200 #define VIRTIO_TYPE_BALLOON 5
201 #define VIRTIO_TYPE_IOMEMORY 6
202 #define VIRTIO_TYPE_RPMSG 7
203 #define VIRTIO_TYPE_SCSI 8
204 #define VIRTIO_TYPE_9P 9
206 /* experimental IDs start at 65535 and work down */
209 * PCI vendor/device IDs
211 #define VIRTIO_VENDOR 0x1AF4
212 #define VIRTIO_DEV_NET 0x1000
213 #define VIRTIO_DEV_BLOCK 0x1001
214 #define VIRTIO_DEV_RANDOM 0x1005
215 #define VIRTIO_DEV_CONSOLE 0x1003
218 * PCI config space constants.
220 * If MSI-X is enabled, the ISR register is generally not used,
221 * and the configuration vector and queue vector appear at offsets
222 * 20 and 22 with the remaining configuration registers at 24.
223 * If MSI-X is not enabled, those two registers disappear and
224 * the remaining configuration registers start at offset 20.
226 #define VTCFG_R_HOSTCAP 0
227 #define VTCFG_R_GUESTCAP 4
228 #define VTCFG_R_PFN 8
229 #define VTCFG_R_QNUM 12
230 #define VTCFG_R_QSEL 14
231 #define VTCFG_R_QNOTIFY 16
232 #define VTCFG_R_STATUS 18
233 #define VTCFG_R_ISR 19
234 #define VTCFG_R_CFGVEC 20
235 #define VTCFG_R_QVEC 22
236 #define VTCFG_R_CFG0 20 /* No MSI-X */
237 #define VTCFG_R_CFG1 24 /* With MSI-X */
238 #define VTCFG_R_MSIX 20
241 * Bits in VTCFG_R_STATUS. Guests need not actually set any of these,
242 * but a guest writing 0 to this register means "please reset".
244 #define VTCFG_STATUS_ACK 0x01 /* guest OS has acknowledged dev */
245 #define VTCFG_STATUS_DRIVER 0x02 /* guest OS driver is loaded */
246 #define VTCFG_STATUS_DRIVER_OK 0x04 /* guest OS driver ready */
247 #define VTCFG_STATUS_FAILED 0x80 /* guest has given up on this dev */
250 * Bits in VTCFG_R_ISR. These apply only if not using MSI-X.
252 * (We don't [yet?] ever use CONF_CHANGED.)
254 #define VTCFG_ISR_QUEUES 0x01 /* re-scan queues */
255 #define VTCFG_ISR_CONF_CHANGED 0x80 /* configuration changed */
257 #define VIRTIO_MSI_NO_VECTOR 0xFFFF
261 * Note: bits 0 through 23 are reserved to each device type.
263 #define VIRTIO_F_NOTIFY_ON_EMPTY (1 << 24)
264 #define VIRTIO_RING_F_INDIRECT_DESC (1 << 28)
265 #define VIRTIO_RING_F_EVENT_IDX (1 << 29)
267 /* From section 2.3, "Virtqueue Configuration", of the virtio specification */
269 vring_size(u_int qsz)
273 /* constant 3 below = va_flags, va_idx, va_used_event */
274 size = sizeof(struct virtio_desc) * qsz + sizeof(uint16_t) * (3 + qsz);
275 size = roundup2(size, VRING_ALIGN);
277 /* constant 3 below = vu_flags, vu_idx, vu_avail_event */
278 size += sizeof(uint16_t) * 3 + sizeof(struct virtio_used) * qsz;
279 size = roundup2(size, VRING_ALIGN);
289 * A virtual device, with some number (possibly 0) of virtual
290 * queues and some size (possibly 0) of configuration-space
291 * registers private to the device. The virtio_softc should come
292 * at the front of each "derived class", so that a pointer to the
293 * virtio_softc is also a pointer to the more specific, derived-
294 * from-virtio driver's softc.
296 * Note: inside each hypervisor virtio driver, changes to these
297 * data structures must be locked against other threads, if any.
298 * Except for PCI config space register read/write, we assume each
299 * driver does the required locking, but we need a pointer to the
300 * lock (if there is one) for PCI config space read/write ops.
302 * When the guest reads or writes the device's config space, the
303 * generic layer checks for operations on the special registers
304 * described above. If the offset of the register(s) being read
305 * or written is past the CFG area (CFG0 or CFG1), the request is
306 * passed on to the virtual device, after subtracting off the
307 * generic-layer size. (So, drivers can just use the offset as
308 * an offset into "struct config", for instance.)
310 * (The virtio layer also makes sure that the read or write is to/
311 * from a "good" config offset, hence vc_cfgsize, and on BAR #0.
312 * However, the driver must verify the read or write size and offset
313 * and that no one is writing a readonly register.)
315 * The BROKED flag ("this thing done gone and broked") is for future
318 #define VIRTIO_USE_MSIX 0x01
319 #define VIRTIO_EVENT_IDX 0x02 /* use the event-index values */
320 #define VIRTIO_BROKED 0x08 /* ??? */
322 struct virtio_softc {
323 struct virtio_consts *vs_vc; /* constants (see below) */
324 int vs_flags; /* VIRTIO_* flags from above */
325 pthread_mutex_t *vs_mtx; /* POSIX mutex, if any */
326 struct pci_devinst *vs_pi; /* PCI device instance */
327 uint32_t vs_negotiated_caps; /* negotiated capabilities */
328 struct vqueue_info *vs_queues; /* one per vc_nvq */
329 int vs_curq; /* current queue */
330 uint8_t vs_status; /* value from last status write */
331 uint8_t vs_isr; /* ISR flags, if not MSI-X */
332 uint16_t vs_msix_cfg_idx; /* MSI-X vector for config event */
335 #define VS_LOCK(vs) \
338 pthread_mutex_lock(vs->vs_mtx); \
341 #define VS_UNLOCK(vs) \
344 pthread_mutex_unlock(vs->vs_mtx); \
347 struct virtio_consts {
348 const char *vc_name; /* name of driver (for diagnostics) */
349 int vc_nvq; /* number of virtual queues */
350 size_t vc_cfgsize; /* size of dev-specific config regs */
351 void (*vc_reset)(void *); /* called on virtual device reset */
352 void (*vc_qnotify)(void *, struct vqueue_info *);
353 /* called on QNOTIFY if no VQ notify */
354 int (*vc_cfgread)(void *, int, int, uint32_t *);
355 /* called to read config regs */
356 int (*vc_cfgwrite)(void *, int, int, uint32_t);
357 /* called to write config regs */
358 void (*vc_apply_features)(void *, uint64_t);
359 /* called to apply negotiated features */
360 uint64_t vc_hv_caps; /* hypervisor-provided capabilities */
364 * Data structure allocated (statically) per virtual queue.
366 * Drivers may change vq_qsize after a reset. When the guest OS
367 * requests a device reset, the hypervisor first calls
368 * vs->vs_vc->vc_reset(); then the data structure below is
369 * reinitialized (for each virtqueue: vs->vs_vc->vc_nvq).
371 * The remaining fields should only be fussed-with by the generic
374 * Note: the addresses of vq_desc, vq_avail, and vq_used are all
375 * computable from each other, but it's a lot simpler if we just
376 * keep a pointer to each one. The event indices are similarly
377 * (but more easily) computable, and this time we'll compute them:
378 * they're just XX_ring[N].
380 #define VQ_ALLOC 0x01 /* set once we have a pfn */
381 #define VQ_BROKED 0x02 /* ??? */
383 uint16_t vq_qsize; /* size of this queue (a power of 2) */
384 void (*vq_notify)(void *, struct vqueue_info *);
385 /* called instead of vc_notify, if not NULL */
387 struct virtio_softc *vq_vs; /* backpointer to softc */
388 uint16_t vq_num; /* we're the num'th queue in the softc */
390 uint16_t vq_flags; /* flags (see above) */
391 uint16_t vq_last_avail; /* a recent value of vq_avail->va_idx */
392 uint16_t vq_save_used; /* saved vq_used->vu_idx; see vq_endchains */
393 uint16_t vq_msix_idx; /* MSI-X index, or VIRTIO_MSI_NO_VECTOR */
395 uint32_t vq_pfn; /* PFN of virt queue (not shifted!) */
397 volatile struct virtio_desc *vq_desc; /* descriptor array */
398 volatile struct vring_avail *vq_avail; /* the "avail" ring */
399 volatile struct vring_used *vq_used; /* the "used" ring */
402 /* as noted above, these are sort of backwards, name-wise */
403 #define VQ_AVAIL_EVENT_IDX(vq) \
404 (*(volatile uint16_t *)&(vq)->vq_used->vu_ring[(vq)->vq_qsize])
405 #define VQ_USED_EVENT_IDX(vq) \
406 ((vq)->vq_avail->va_ring[(vq)->vq_qsize])
409 * Is this ring ready for I/O?
412 vq_ring_ready(struct vqueue_info *vq)
415 return (vq->vq_flags & VQ_ALLOC);
419 * Are there "available" descriptors? (This does not count
420 * how many, just returns True if there are some.)
423 vq_has_descs(struct vqueue_info *vq)
426 return (vq_ring_ready(vq) && vq->vq_last_avail !=
427 vq->vq_avail->va_idx);
431 * Deliver an interrupt to guest on the given virtual queue
432 * (if possible, or a generic MSI interrupt if not using MSI-X).
435 vq_interrupt(struct virtio_softc *vs, struct vqueue_info *vq)
438 if (pci_msix_enabled(vs->vs_pi))
439 pci_generate_msix(vs->vs_pi, vq->vq_msix_idx);
442 vs->vs_isr |= VTCFG_ISR_QUEUES;
443 pci_generate_msi(vs->vs_pi, 0);
444 pci_lintr_assert(vs->vs_pi);
450 void vi_softc_linkup(struct virtio_softc *vs, struct virtio_consts *vc,
451 void *dev_softc, struct pci_devinst *pi,
452 struct vqueue_info *queues);
453 int vi_intr_init(struct virtio_softc *vs, int barnum, int use_msix);
454 void vi_reset_dev(struct virtio_softc *);
455 void vi_set_io_bar(struct virtio_softc *, int);
457 int vq_getchain(struct vqueue_info *vq, uint16_t *pidx,
458 struct iovec *iov, int n_iov, uint16_t *flags);
459 void vq_retchain(struct vqueue_info *vq);
460 void vq_relchain(struct vqueue_info *vq, uint16_t idx, uint32_t iolen);
461 void vq_endchains(struct vqueue_info *vq, int used_all_avail);
463 uint64_t vi_pci_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
464 int baridx, uint64_t offset, int size);
465 void vi_pci_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
466 int baridx, uint64_t offset, int size, uint64_t value);
467 #endif /* _VIRTIO_H_ */