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
31 #ifndef _BHYVE_VIRTIO_H_
32 #define _BHYVE_VIRTIO_H_
34 #include <machine/atomic.h>
36 #include <dev/virtio/virtio.h>
37 #include <dev/virtio/virtio_ring.h>
38 #include <dev/virtio/pci/virtio_pci_var.h>
41 * These are derived from several virtio specifications.
44 * https://github.com/rustyrussell/virtio-spec
45 * http://people.redhat.com/pbonzini/virtio-spec.pdf
49 * A virtual device has zero or more "virtual queues" (virtqueue).
50 * Each virtqueue uses at least two 4096-byte pages, laid out thus:
52 * +-----------------------------------------------+
53 * | "desc": <N> descriptors, 16 bytes each |
54 * | ----------------------------------------- |
55 * | "avail": 2 uint16; <N> uint16; 1 uint16 |
56 * | ----------------------------------------- |
57 * | pad to 4k boundary |
58 * +-----------------------------------------------+
59 * | "used": 2 x uint16; <N> elems; 1 uint16 |
60 * | ----------------------------------------- |
61 * | pad to 4k boundary |
62 * +-----------------------------------------------+
64 * The number <N> that appears here is always a power of two and is
65 * limited to no more than 32768 (as it must fit in a 16-bit field).
66 * If <N> is sufficiently large, the above will occupy more than
67 * two pages. In any case, all pages must be physically contiguous
68 * within the guest's physical address space.
70 * The <N> 16-byte "desc" descriptors consist of a 64-bit guest
71 * physical address <addr>, a 32-bit length <len>, a 16-bit
72 * <flags>, and a 16-bit <next> field (all in guest byte order).
74 * There are three flags that may be set :
75 * NEXT descriptor is chained, so use its "next" field
76 * WRITE descriptor is for host to write into guest RAM
77 * (else host is to read from guest RAM)
78 * INDIRECT descriptor address field is (guest physical)
79 * address of a linear array of descriptors
81 * Unless INDIRECT is set, <len> is the number of bytes that may
82 * be read/written from guest physical address <addr>. If
83 * INDIRECT is set, WRITE is ignored and <len> provides the length
84 * of the indirect descriptors (and <len> must be a multiple of
85 * 16). Note that NEXT may still be set in the main descriptor
86 * pointing to the indirect, and should be set in each indirect
87 * descriptor that uses the next descriptor (these should generally
88 * be numbered sequentially). However, INDIRECT must not be set
89 * in the indirect descriptors. Upon reaching an indirect descriptor
90 * without a NEXT bit, control returns to the direct descriptors.
92 * Except inside an indirect, each <next> value must be in the
93 * range [0 .. N) (i.e., the half-open interval). (Inside an
94 * indirect, each <next> must be in the range [0 .. <len>/16).)
96 * The "avail" data structures reside in the same pages as the
97 * "desc" structures since both together are used by the device to
98 * pass information to the hypervisor's virtual driver. These
99 * begin with a 16-bit <flags> field and 16-bit index <idx>, then
100 * have <N> 16-bit <ring> values, followed by one final 16-bit
101 * field <used_event>. The <N> <ring> entries are simply indices
102 * indices into the descriptor ring (and thus must meet the same
103 * constraints as each <next> value). However, <idx> is counted
104 * up from 0 (initially) and simply wraps around after 65535; it
105 * is taken mod <N> to find the next available entry.
107 * The "used" ring occupies a separate page or pages, and contains
108 * values written from the virtual driver back to the guest OS.
109 * This begins with a 16-bit <flags> and 16-bit <idx>, then there
110 * are <N> "vring_used" elements, followed by a 16-bit <avail_event>.
111 * The <N> "vring_used" elements consist of a 32-bit <id> and a
112 * 32-bit <len> (vu_tlen below). The <id> is simply the index of
113 * the head of a descriptor chain the guest made available
114 * earlier, and the <len> is the number of bytes actually written,
115 * e.g., in the case of a network driver that provided a large
116 * receive buffer but received only a small amount of data.
118 * The two event fields, <used_event> and <avail_event>, in the
119 * avail and used rings (respectively -- note the reversal!), are
120 * always provided, but are used only if the virtual device
121 * negotiates the VIRTIO_RING_F_EVENT_IDX feature during feature
122 * negotiation. Similarly, both rings provide a flag --
123 * VRING_AVAIL_F_NO_INTERRUPT and VRING_USED_F_NO_NOTIFY -- in
124 * their <flags> field, indicating that the guest does not need an
125 * interrupt, or that the hypervisor driver does not need a
126 * notify, when descriptors are added to the corresponding ring.
127 * (These are provided only for interrupt optimization and need
128 * not be implemented.)
130 #define VRING_ALIGN 4096
133 * The address of any given virtual queue is determined by a single
134 * Page Frame Number register. The guest writes the PFN into the
135 * PCI config space. However, a device that has two or more
136 * virtqueues can have a different PFN, and size, for each queue.
137 * The number of queues is determinable via the PCI config space
138 * VTCFG_R_QSEL register. Writes to QSEL select the queue: 0 means
139 * queue #0, 1 means queue#1, etc. Once a queue is selected, the
140 * remaining PFN and QNUM registers refer to that queue.
142 * QNUM is a read-only register containing a nonzero power of two
143 * that indicates the (hypervisor's) queue size. Or, if reading it
144 * produces zero, the hypervisor does not have a corresponding
145 * queue. (The number of possible queues depends on the virtual
146 * device. The block device has just one; the network device
147 * provides either two -- 0 = receive, 1 = transmit -- or three,
150 * PFN is a read/write register giving the physical page address of
151 * the virtqueue in guest memory (the guest must allocate enough space
152 * based on the hypervisor's provided QNUM).
154 * QNOTIFY is effectively write-only: when the guest writes a queue
155 * number to the register, the hypervisor should scan the specified
156 * virtqueue. (Reading QNOTIFY currently always gets 0).
160 * PFN register shift amount
165 * PCI vendor/device IDs
167 #define VIRTIO_VENDOR 0x1AF4
168 #define VIRTIO_DEV_NET 0x1000
169 #define VIRTIO_DEV_BLOCK 0x1001
170 #define VIRTIO_DEV_CONSOLE 0x1003
171 #define VIRTIO_DEV_SCSI 0x1004
172 #define VIRTIO_DEV_RANDOM 0x1005
173 #define VIRTIO_DEV_9P 0x1009
174 #define VIRTIO_DEV_INPUT 0x1052
179 #define VIRTIO_REV_INPUT 1
184 #define VIRTIO_SUBVEN_INPUT 0x108E
189 #define VIRTIO_SUBDEV_INPUT 0x1100
191 /* From section 2.3, "Virtqueue Configuration", of the virtio specification */
193 vring_size_aligned(u_int qsz)
195 return (roundup2(vring_size(qsz, VRING_ALIGN), VRING_ALIGN));
201 struct vm_snapshot_meta;
204 * A virtual device, with some number (possibly 0) of virtual
205 * queues and some size (possibly 0) of configuration-space
206 * registers private to the device. The virtio_softc should come
207 * at the front of each "derived class", so that a pointer to the
208 * virtio_softc is also a pointer to the more specific, derived-
209 * from-virtio driver's softc.
211 * Note: inside each hypervisor virtio driver, changes to these
212 * data structures must be locked against other threads, if any.
213 * Except for PCI config space register read/write, we assume each
214 * driver does the required locking, but we need a pointer to the
215 * lock (if there is one) for PCI config space read/write ops.
217 * When the guest reads or writes the device's config space, the
218 * generic layer checks for operations on the special registers
219 * described above. If the offset of the register(s) being read
220 * or written is past the CFG area (CFG0 or CFG1), the request is
221 * passed on to the virtual device, after subtracting off the
222 * generic-layer size. (So, drivers can just use the offset as
223 * an offset into "struct config", for instance.)
225 * (The virtio layer also makes sure that the read or write is to/
226 * from a "good" config offset, hence vc_cfgsize, and on BAR #0.
227 * However, the driver must verify the read or write size and offset
228 * and that no one is writing a readonly register.)
230 * The BROKED flag ("this thing done gone and broked") is for future
233 #define VIRTIO_USE_MSIX 0x01
234 #define VIRTIO_EVENT_IDX 0x02 /* use the event-index values */
235 #define VIRTIO_BROKED 0x08 /* ??? */
237 struct virtio_softc {
238 struct virtio_consts *vs_vc; /* constants (see below) */
239 int vs_flags; /* VIRTIO_* flags from above */
240 pthread_mutex_t *vs_mtx; /* POSIX mutex, if any */
241 struct pci_devinst *vs_pi; /* PCI device instance */
242 uint32_t vs_negotiated_caps; /* negotiated capabilities */
243 struct vqueue_info *vs_queues; /* one per vc_nvq */
244 int vs_curq; /* current queue */
245 uint8_t vs_status; /* value from last status write */
246 uint8_t vs_isr; /* ISR flags, if not MSI-X */
247 uint16_t vs_msix_cfg_idx; /* MSI-X vector for config event */
250 #define VS_LOCK(vs) \
253 pthread_mutex_lock(vs->vs_mtx); \
256 #define VS_UNLOCK(vs) \
259 pthread_mutex_unlock(vs->vs_mtx); \
262 struct virtio_consts {
263 const char *vc_name; /* name of driver (for diagnostics) */
264 int vc_nvq; /* number of virtual queues */
265 size_t vc_cfgsize; /* size of dev-specific config regs */
266 void (*vc_reset)(void *); /* called on virtual device reset */
267 void (*vc_qnotify)(void *, struct vqueue_info *);
268 /* called on QNOTIFY if no VQ notify */
269 int (*vc_cfgread)(void *, int, int, uint32_t *);
270 /* called to read config regs */
271 int (*vc_cfgwrite)(void *, int, int, uint32_t);
272 /* called to write config regs */
273 void (*vc_apply_features)(void *, uint64_t);
274 /* called to apply negotiated features */
275 uint64_t vc_hv_caps; /* hypervisor-provided capabilities */
276 void (*vc_pause)(void *); /* called to pause device activity */
277 void (*vc_resume)(void *); /* called to resume device activity */
278 int (*vc_snapshot)(void *, struct vm_snapshot_meta *);
279 /* called to save / restore device state */
283 * Data structure allocated (statically) per virtual queue.
285 * Drivers may change vq_qsize after a reset. When the guest OS
286 * requests a device reset, the hypervisor first calls
287 * vs->vs_vc->vc_reset(); then the data structure below is
288 * reinitialized (for each virtqueue: vs->vs_vc->vc_nvq).
290 * The remaining fields should only be fussed-with by the generic
293 * Note: the addresses of vq_desc, vq_avail, and vq_used are all
294 * computable from each other, but it's a lot simpler if we just
295 * keep a pointer to each one. The event indices are similarly
296 * (but more easily) computable, and this time we'll compute them:
297 * they're just XX_ring[N].
299 #define VQ_ALLOC 0x01 /* set once we have a pfn */
300 #define VQ_BROKED 0x02 /* ??? */
302 uint16_t vq_qsize; /* size of this queue (a power of 2) */
303 void (*vq_notify)(void *, struct vqueue_info *);
304 /* called instead of vc_notify, if not NULL */
306 struct virtio_softc *vq_vs; /* backpointer to softc */
307 uint16_t vq_num; /* we're the num'th queue in the softc */
309 uint16_t vq_flags; /* flags (see above) */
310 uint16_t vq_last_avail; /* a recent value of vq_avail->idx */
311 uint16_t vq_next_used; /* index of the next used slot to be filled */
312 uint16_t vq_save_used; /* saved vq_used->idx; see vq_endchains */
313 uint16_t vq_msix_idx; /* MSI-X index, or VIRTIO_MSI_NO_VECTOR */
315 uint32_t vq_pfn; /* PFN of virt queue (not shifted!) */
317 volatile struct vring_desc *vq_desc; /* descriptor array */
318 volatile struct vring_avail *vq_avail; /* the "avail" ring */
319 volatile struct vring_used *vq_used; /* the "used" ring */
322 /* as noted above, these are sort of backwards, name-wise */
323 #define VQ_AVAIL_EVENT_IDX(vq) \
324 (*(volatile uint16_t *)&(vq)->vq_used->ring[(vq)->vq_qsize])
325 #define VQ_USED_EVENT_IDX(vq) \
326 ((vq)->vq_avail->ring[(vq)->vq_qsize])
329 * Is this ring ready for I/O?
332 vq_ring_ready(struct vqueue_info *vq)
335 return (vq->vq_flags & VQ_ALLOC);
339 * Are there "available" descriptors? (This does not count
340 * how many, just returns True if there are some.)
343 vq_has_descs(struct vqueue_info *vq)
346 return (vq_ring_ready(vq) && vq->vq_last_avail !=
351 * Deliver an interrupt to the guest for a specific MSI-X queue or
355 vi_interrupt(struct virtio_softc *vs, uint8_t isr, uint16_t msix_idx)
358 if (pci_msix_enabled(vs->vs_pi))
359 pci_generate_msix(vs->vs_pi, msix_idx);
363 pci_generate_msi(vs->vs_pi, 0);
364 pci_lintr_assert(vs->vs_pi);
370 * Deliver an interrupt to the guest on the given virtual queue (if
371 * possible, or a generic MSI interrupt if not using MSI-X).
374 vq_interrupt(struct virtio_softc *vs, struct vqueue_info *vq)
377 vi_interrupt(vs, VIRTIO_PCI_ISR_INTR, vq->vq_msix_idx);
381 vq_kick_enable(struct vqueue_info *vq)
384 vq->vq_used->flags &= ~VRING_USED_F_NO_NOTIFY;
386 * Full memory barrier to make sure the store to vq_used->flags
387 * happens before the load from vq_avail->idx, which results from a
388 * subsequent call to vq_has_descs().
390 atomic_thread_fence_seq_cst();
394 vq_kick_disable(struct vqueue_info *vq)
397 vq->vq_used->flags |= VRING_USED_F_NO_NOTIFY;
403 * Request description returned by vq_getchain.
405 * Writable iovecs start at iov[req.readable].
408 int readable; /* num of readable iovecs */
409 int writable; /* num of writable iovecs */
410 unsigned int idx; /* ring index */
413 void vi_softc_linkup(struct virtio_softc *vs, struct virtio_consts *vc,
414 void *dev_softc, struct pci_devinst *pi,
415 struct vqueue_info *queues);
416 int vi_intr_init(struct virtio_softc *vs, int barnum, int use_msix);
417 void vi_reset_dev(struct virtio_softc *);
418 void vi_set_io_bar(struct virtio_softc *, int);
420 int vq_getchain(struct vqueue_info *vq, struct iovec *iov, int niov,
421 struct vi_req *reqp);
422 void vq_retchains(struct vqueue_info *vq, uint16_t n_chains);
423 void vq_relchain_prepare(struct vqueue_info *vq, uint16_t idx,
425 void vq_relchain_publish(struct vqueue_info *vq);
426 void vq_relchain(struct vqueue_info *vq, uint16_t idx, uint32_t iolen);
427 void vq_endchains(struct vqueue_info *vq, int used_all_avail);
429 uint64_t vi_pci_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
430 int baridx, uint64_t offset, int size);
431 void vi_pci_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
432 int baridx, uint64_t offset, int size, uint64_t value);
433 #ifdef BHYVE_SNAPSHOT
434 int vi_pci_snapshot(struct vm_snapshot_meta *meta);
435 int vi_pci_pause(struct vmctx *ctx, struct pci_devinst *pi);
436 int vi_pci_resume(struct vmctx *ctx, struct pci_devinst *pi);
438 #endif /* _BHYVE_VIRTIO_H_ */