xref: /freebsd/usr.sbin/bhyve/virtio.h (revision dd41de95a84d979615a2ef11df6850622bf6184e)
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	_BHYVE_VIRTIO_H_
32 #define	_BHYVE_VIRTIO_H_
33 
34 #include <machine/atomic.h>
35 
36 #include <dev/virtio/virtio.h>
37 #include <dev/virtio/virtio_ring.h>
38 #include <dev/virtio/pci/virtio_pci_var.h>
39 
40 /*
41  * These are derived from several virtio specifications.
42  *
43  * Some useful links:
44  *    https://github.com/rustyrussell/virtio-spec
45  *    http://people.redhat.com/pbonzini/virtio-spec.pdf
46  */
47 
48 /*
49  * A virtual device has zero or more "virtual queues" (virtqueue).
50  * Each virtqueue uses at least two 4096-byte pages, laid out thus:
51  *
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  *      +-----------------------------------------------+
63  *
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.
69  *
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).
73  *
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
80  *
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.
91  *
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).)
95  *
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.
106  *
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.
117  *
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.)
129  */
130 #define VRING_ALIGN	4096
131 
132 /*
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.
141  *
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,
148  * with 2 = control.)
149  *
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).
153  *
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).
157  */
158 
159 /*
160  * PFN register shift amount
161  */
162 #define	VRING_PFN		12
163 
164 /*
165  * PCI vendor/device IDs
166  */
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_RANDOM	0x1005
172 #define	VIRTIO_DEV_SCSI		0x1008
173 #define	VIRTIO_DEV_9P		0x1009
174 
175 /* From section 2.3, "Virtqueue Configuration", of the virtio specification */
176 static inline int
177 vring_size_aligned(u_int qsz)
178 {
179 	return (roundup2(vring_size(qsz, VRING_ALIGN), VRING_ALIGN));
180 }
181 
182 struct vmctx;
183 struct pci_devinst;
184 struct vqueue_info;
185 struct vm_snapshot_meta;
186 
187 /*
188  * A virtual device, with some number (possibly 0) of virtual
189  * queues and some size (possibly 0) of configuration-space
190  * registers private to the device.  The virtio_softc should come
191  * at the front of each "derived class", so that a pointer to the
192  * virtio_softc is also a pointer to the more specific, derived-
193  * from-virtio driver's softc.
194  *
195  * Note: inside each hypervisor virtio driver, changes to these
196  * data structures must be locked against other threads, if any.
197  * Except for PCI config space register read/write, we assume each
198  * driver does the required locking, but we need a pointer to the
199  * lock (if there is one) for PCI config space read/write ops.
200  *
201  * When the guest reads or writes the device's config space, the
202  * generic layer checks for operations on the special registers
203  * described above.  If the offset of the register(s) being read
204  * or written is past the CFG area (CFG0 or CFG1), the request is
205  * passed on to the virtual device, after subtracting off the
206  * generic-layer size.  (So, drivers can just use the offset as
207  * an offset into "struct config", for instance.)
208  *
209  * (The virtio layer also makes sure that the read or write is to/
210  * from a "good" config offset, hence vc_cfgsize, and on BAR #0.
211  * However, the driver must verify the read or write size and offset
212  * and that no one is writing a readonly register.)
213  *
214  * The BROKED flag ("this thing done gone and broked") is for future
215  * use.
216  */
217 #define	VIRTIO_USE_MSIX		0x01
218 #define	VIRTIO_EVENT_IDX	0x02	/* use the event-index values */
219 #define	VIRTIO_BROKED		0x08	/* ??? */
220 
221 struct virtio_softc {
222 	struct virtio_consts *vs_vc;	/* constants (see below) */
223 	int	vs_flags;		/* VIRTIO_* flags from above */
224 	pthread_mutex_t *vs_mtx;	/* POSIX mutex, if any */
225 	struct pci_devinst *vs_pi;	/* PCI device instance */
226 	uint32_t vs_negotiated_caps;	/* negotiated capabilities */
227 	struct vqueue_info *vs_queues;	/* one per vc_nvq */
228 	int	vs_curq;		/* current queue */
229 	uint8_t	vs_status;		/* value from last status write */
230 	uint8_t	vs_isr;			/* ISR flags, if not MSI-X */
231 	uint16_t vs_msix_cfg_idx;	/* MSI-X vector for config event */
232 };
233 
234 #define	VS_LOCK(vs)							\
235 do {									\
236 	if (vs->vs_mtx)							\
237 		pthread_mutex_lock(vs->vs_mtx);				\
238 } while (0)
239 
240 #define	VS_UNLOCK(vs)							\
241 do {									\
242 	if (vs->vs_mtx)							\
243 		pthread_mutex_unlock(vs->vs_mtx);			\
244 } while (0)
245 
246 struct virtio_consts {
247 	const char *vc_name;		/* name of driver (for diagnostics) */
248 	int	vc_nvq;			/* number of virtual queues */
249 	size_t	vc_cfgsize;		/* size of dev-specific config regs */
250 	void	(*vc_reset)(void *);	/* called on virtual device reset */
251 	void	(*vc_qnotify)(void *, struct vqueue_info *);
252 					/* called on QNOTIFY if no VQ notify */
253 	int	(*vc_cfgread)(void *, int, int, uint32_t *);
254 					/* called to read config regs */
255 	int	(*vc_cfgwrite)(void *, int, int, uint32_t);
256 					/* called to write config regs */
257 	void    (*vc_apply_features)(void *, uint64_t);
258 				/* called to apply negotiated features */
259 	uint64_t vc_hv_caps;		/* hypervisor-provided capabilities */
260 	void	(*vc_pause)(void *);	/* called to pause device activity */
261 	void	(*vc_resume)(void *);	/* called to resume device activity */
262 	int	(*vc_snapshot)(void *, struct vm_snapshot_meta *);
263 				/* called to save / restore device state */
264 };
265 
266 /*
267  * Data structure allocated (statically) per virtual queue.
268  *
269  * Drivers may change vq_qsize after a reset.  When the guest OS
270  * requests a device reset, the hypervisor first calls
271  * vs->vs_vc->vc_reset(); then the data structure below is
272  * reinitialized (for each virtqueue: vs->vs_vc->vc_nvq).
273  *
274  * The remaining fields should only be fussed-with by the generic
275  * code.
276  *
277  * Note: the addresses of vq_desc, vq_avail, and vq_used are all
278  * computable from each other, but it's a lot simpler if we just
279  * keep a pointer to each one.  The event indices are similarly
280  * (but more easily) computable, and this time we'll compute them:
281  * they're just XX_ring[N].
282  */
283 #define	VQ_ALLOC	0x01	/* set once we have a pfn */
284 #define	VQ_BROKED	0x02	/* ??? */
285 struct vqueue_info {
286 	uint16_t vq_qsize;	/* size of this queue (a power of 2) */
287 	void	(*vq_notify)(void *, struct vqueue_info *);
288 				/* called instead of vc_notify, if not NULL */
289 
290 	struct virtio_softc *vq_vs;	/* backpointer to softc */
291 	uint16_t vq_num;	/* we're the num'th queue in the softc */
292 
293 	uint16_t vq_flags;	/* flags (see above) */
294 	uint16_t vq_last_avail;	/* a recent value of vq_avail->idx */
295 	uint16_t vq_next_used;	/* index of the next used slot to be filled */
296 	uint16_t vq_save_used;	/* saved vq_used->idx; see vq_endchains */
297 	uint16_t vq_msix_idx;	/* MSI-X index, or VIRTIO_MSI_NO_VECTOR */
298 
299 	uint32_t vq_pfn;	/* PFN of virt queue (not shifted!) */
300 
301 	volatile struct vring_desc *vq_desc;	/* descriptor array */
302 	volatile struct vring_avail *vq_avail;	/* the "avail" ring */
303 	volatile struct vring_used *vq_used;	/* the "used" ring */
304 
305 };
306 /* as noted above, these are sort of backwards, name-wise */
307 #define VQ_AVAIL_EVENT_IDX(vq) \
308 	(*(volatile uint16_t *)&(vq)->vq_used->ring[(vq)->vq_qsize])
309 #define VQ_USED_EVENT_IDX(vq) \
310 	((vq)->vq_avail->ring[(vq)->vq_qsize])
311 
312 /*
313  * Is this ring ready for I/O?
314  */
315 static inline int
316 vq_ring_ready(struct vqueue_info *vq)
317 {
318 
319 	return (vq->vq_flags & VQ_ALLOC);
320 }
321 
322 /*
323  * Are there "available" descriptors?  (This does not count
324  * how many, just returns True if there are some.)
325  */
326 static inline int
327 vq_has_descs(struct vqueue_info *vq)
328 {
329 
330 	return (vq_ring_ready(vq) && vq->vq_last_avail !=
331 	    vq->vq_avail->idx);
332 }
333 
334 /*
335  * Deliver an interrupt to guest on the given virtual queue
336  * (if possible, or a generic MSI interrupt if not using MSI-X).
337  */
338 static inline void
339 vq_interrupt(struct virtio_softc *vs, struct vqueue_info *vq)
340 {
341 
342 	if (pci_msix_enabled(vs->vs_pi))
343 		pci_generate_msix(vs->vs_pi, vq->vq_msix_idx);
344 	else {
345 		VS_LOCK(vs);
346 		vs->vs_isr |= VIRTIO_PCI_ISR_INTR;
347 		pci_generate_msi(vs->vs_pi, 0);
348 		pci_lintr_assert(vs->vs_pi);
349 		VS_UNLOCK(vs);
350 	}
351 }
352 
353 static inline void
354 vq_kick_enable(struct vqueue_info *vq)
355 {
356 
357 	vq->vq_used->flags &= ~VRING_USED_F_NO_NOTIFY;
358 	/*
359 	 * Full memory barrier to make sure the store to vq_used->flags
360 	 * happens before the load from vq_avail->idx, which results from a
361 	 * subsequent call to vq_has_descs().
362 	 */
363 	atomic_thread_fence_seq_cst();
364 }
365 
366 static inline void
367 vq_kick_disable(struct vqueue_info *vq)
368 {
369 
370 	vq->vq_used->flags |= VRING_USED_F_NO_NOTIFY;
371 }
372 
373 struct iovec;
374 
375 /*
376  * Request description returned by vq_getchain.
377  *
378  * Writable iovecs start at iov[req.readable].
379  */
380 struct vi_req {
381 	int readable;		/* num of readable iovecs */
382 	int writable;		/* num of writable iovecs */
383 	unsigned int idx;	/* ring index */
384 };
385 
386 void	vi_softc_linkup(struct virtio_softc *vs, struct virtio_consts *vc,
387 			void *dev_softc, struct pci_devinst *pi,
388 			struct vqueue_info *queues);
389 int	vi_intr_init(struct virtio_softc *vs, int barnum, int use_msix);
390 void	vi_reset_dev(struct virtio_softc *);
391 void	vi_set_io_bar(struct virtio_softc *, int);
392 
393 int	vq_getchain(struct vqueue_info *vq, struct iovec *iov, int niov,
394 	    struct vi_req *reqp);
395 void	vq_retchains(struct vqueue_info *vq, uint16_t n_chains);
396 void	vq_relchain_prepare(struct vqueue_info *vq, uint16_t idx,
397 			    uint32_t iolen);
398 void	vq_relchain_publish(struct vqueue_info *vq);
399 void	vq_relchain(struct vqueue_info *vq, uint16_t idx, uint32_t iolen);
400 void	vq_endchains(struct vqueue_info *vq, int used_all_avail);
401 
402 uint64_t vi_pci_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
403 		     int baridx, uint64_t offset, int size);
404 void	vi_pci_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
405 		     int baridx, uint64_t offset, int size, uint64_t value);
406 #ifdef BHYVE_SNAPSHOT
407 int	vi_pci_snapshot(struct vm_snapshot_meta *meta);
408 int	vi_pci_pause(struct vmctx *ctx, struct pci_devinst *pi);
409 int	vi_pci_resume(struct vmctx *ctx, struct pci_devinst *pi);
410 #endif
411 #endif	/* _BHYVE_VIRTIO_H_ */
412