xref: /illumos-gate/usr/src/uts/intel/io/viona/viona_main.c (revision 609febc9a48c79a089214cb5d882759a72a38513)
1 /*
2  * Copyright (c) 2013  Chris Torek <torek @ torek net>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 /*
27  * This file and its contents are supplied under the terms of the
28  * Common Development and Distribution License ("CDDL"), version 1.0.
29  * You may only use this file in accordance with the terms of version
30  * 1.0 of the CDDL.
31  *
32  * A full copy of the text of the CDDL should have accompanied this
33  * source.  A copy of the CDDL is also available via the Internet at
34  * http://www.illumos.org/license/CDDL.
35  *
36  * Copyright 2015 Pluribus Networks Inc.
37  * Copyright 2019 Joyent, Inc.
38  * Copyright 2022 Oxide Computer Company
39  */
40 
41 /*
42  * viona - VirtIO-Net, Accelerated
43  *
44  * The purpose of viona is to provide high performance virtio-net devices to
45  * bhyve guests.  It does so by sitting directly atop MAC, skipping all of the
46  * DLS/DLD stack.
47  *
48  * --------------------
49  * General Architecture
50  * --------------------
51  *
52  * A single viona instance is comprised of a "link" handle and two "rings".
53  * After opening the viona device, it must be associated with a MAC network
54  * interface and a bhyve (vmm) instance to form its link resource.  This is
55  * done with the VNA_IOC_CREATE ioctl, where the datalink ID and vmm fd are
56  * passed in to perform the initialization.  With the MAC client opened, and a
57  * driver handle to the vmm instance established, the device is ready to be
58  * configured by the guest.
59  *
60  * The userspace portion of bhyve, which interfaces with the PCI device
61  * emulation framework, is meant to stay out of the datapath if at all
62  * possible.  Configuration changes made via PCI are mapped to actions which
63  * will steer the operation of the in-kernel logic.
64  *
65  *
66  * -----------
67  * Ring Basics
68  * -----------
69  *
70  * Each viona link has two viona_vring_t entities, RX and TX, for handling data
71  * transfers to and from the guest.  They represent an interface to the
72  * standard virtio ring structures.  When intiailized and active, each ring is
73  * backed by a kernel worker thread (parented to the bhyve process for the
74  * instance) which handles ring events.  The RX worker has the simple task of
75  * watching for ring shutdown conditions.  The TX worker does that in addition
76  * to processing all requests to transmit data.  Data destined for the guest is
77  * delivered directly by MAC to viona_rx() when the ring is active.
78  *
79  *
80  * -----------
81  * Ring States
82  * -----------
83  *
84  * The viona_vring_t instances follow a simple path through the possible state
85  * values represented in virtio_vring_t`vr_state:
86  *
87  *        +<--------------------------------------------+
88  *        |						|
89  *        V						^
90  *  +-----------+	This is the initial state when a link is created or
91  *  | VRS_RESET |	when the ring has been explicitly reset.
92  *  +-----------+
93  *        |						^
94  *        |---* ioctl(VNA_IOC_RING_INIT) issued		|
95  *        |						|
96  *        |						^
97  *        V
98  *  +-----------+	The ring parameters (size, guest physical addresses)
99  *  | VRS_SETUP |	have been set and start-up of the ring worker thread
100  *  +-----------+	has begun.
101  *        |						^
102  *        |						|
103  *        |---* ring worker thread begins execution	|
104  *        |						|
105  *        +-------------------------------------------->+
106  *        |	      |					^
107  *        |	      |
108  *        |	      *	If ring shutdown is requested (by ioctl or impending
109  *        |		bhyve process death) while the worker thread is
110  *        |		starting, the worker will transition the ring to
111  *        |		VRS_RESET and exit.
112  *        |						^
113  *        |						|
114  *        |<-------------------------------------------<+
115  *        |	      |					|
116  *        |	      |					^
117  *        |	      *	If ring is requested to pause (but not stop)from the
118  *        |             VRS_RUN state, it will return to the VRS_INIT state.
119  *        |
120  *        |						^
121  *        |						|
122  *        |						^
123  *        V
124  *  +-----------+	The worker thread associated with the ring has started
125  *  | VRS_INIT  |	executing.  It has allocated any extra resources needed
126  *  +-----------+	for the ring to operate.
127  *        |						^
128  *        |						|
129  *        +-------------------------------------------->+
130  *        |	      |					^
131  *        |	      |
132  *        |	      *	If ring shutdown is requested while the worker is
133  *        |		waiting in VRS_INIT, it will free any extra resources
134  *        |		and transition to VRS_RESET.
135  *        |						^
136  *        |						|
137  *        |--* ioctl(VNA_IOC_RING_KICK) issued		|
138  *        |						^
139  *        V
140  *  +-----------+	The worker thread associated with the ring is executing
141  *  | VRS_RUN   |	workload specific to that ring.
142  *  +-----------+
143  *        |						^
144  *        |---* ioctl(VNA_IOC_RING_RESET) issued	|
145  *        |	(or bhyve process begins exit)		^
146  *        |
147  *  +-----------+	The worker thread associated with the ring is in the
148  *  | VRS_STOP  |	process of exiting. All outstanding TX and RX
149  *  +-----------+	requests are allowed to complete, but new requests
150  *        |		must be ignored.
151  *        |						^
152  *        |						|
153  *        +-------------------------------------------->+
154  *
155  *
156  * While the worker thread is not running, changes to vr_state are only made by
157  * viona_ioc_ring_init() under vr_lock.  There, it initializes the ring, starts
158  * the worker, and sets the ring state to VRS_SETUP.  Once the worker thread
159  * has been started, only it may perform ring state transitions (still under
160  * the protection of vr_lock), when requested by outside consumers via
161  * vr_state_flags or when the containing bhyve process initiates an exit.
162  *
163  *
164  * ----------------------------
165  * Transmission mblk_t Handling
166  * ----------------------------
167  *
168  * For incoming frames destined for a bhyve guest, the data must first land in
169  * a host OS buffer from the physical NIC before it is copied into the awaiting
170  * guest buffer(s).  Outbound frames transmitted by the guest are not bound by
171  * this limitation and can avoid extra copying before the buffers are accessed
172  * directly by the NIC.  When a guest designates buffers to be transmitted,
173  * viona translates the guest-physical addresses contained in the ring
174  * descriptors to host-virtual addresses via viona_hold_page().  That pointer is
175  * wrapped in an mblk_t using a preallocated viona_desb_t for the desballoc().
176  * Doing so increments vr_xfer_outstanding, preventing the ring from being
177  * reset (allowing the link to drop its vmm handle to the guest) until all
178  * transmit mblks referencing guest memory have been processed.  Allocation of
179  * the viona_desb_t entries is done during the VRS_INIT stage of the ring
180  * worker thread.  The ring size informs that allocation as the number of
181  * concurrent transmissions is limited by the number of descriptors in the
182  * ring.  This minimizes allocation in the transmit hot-path by acquiring those
183  * fixed-size resources during initialization.
184  *
185  * This optimization depends on the underlying NIC driver freeing the mblks in
186  * a timely manner after they have been transmitted by the hardware.  Some
187  * drivers have been found to flush TX descriptors only when new transmissions
188  * are initiated.  This means that there is no upper bound to the time needed
189  * for an mblk to be flushed and can stall bhyve guests from shutting down
190  * since their memory must be free of viona TX references prior to clean-up.
191  *
192  * This expectation of deterministic mblk_t processing is likely the reason
193  * behind the notable exception to the zero-copy TX path: systems with 'bnxe'
194  * loaded will copy transmit data into fresh buffers rather than passing up
195  * zero-copy mblks.  It is a hold-over from the original viona sources provided
196  * by Pluribus and its continued necessity has not been confirmed.
197  *
198  *
199  * ----------------------------
200  * Ring Notification Fast-paths
201  * ----------------------------
202  *
203  * Device operation for viona requires that notifications flow to and from the
204  * guest to indicate certain ring conditions.  In order to minimize latency and
205  * processing overhead, the notification procedures are kept in-kernel whenever
206  * possible.
207  *
208  * Guest-to-host notifications, when new available descriptors have been placed
209  * in the ring, are posted via the 'queue notify' address in the virtio BAR.
210  * The vmm_drv_ioport_hook() interface was added to bhyve which allows viona to
211  * install a callback hook on an ioport address.  Guest exits for accesses to
212  * viona-hooked ioport addresses will result in direct calls to notify the
213  * appropriate ring worker without a trip to userland.
214  *
215  * Host-to-guest notifications in the form of interrupts enjoy similar
216  * acceleration.  Each viona ring can be configured to send MSI notifications
217  * to the guest as virtio conditions dictate.  This in-kernel interrupt
218  * configuration is kept synchronized through viona ioctls which are utilized
219  * during writes to the associated PCI config registers or MSI-X BAR.
220  *
221  * Guests which do not utilize MSI-X will result in viona falling back to the
222  * slow path for interrupts.  It will poll(2) the viona handle, receiving
223  * notification when ring events necessitate the assertion of an interrupt.
224  *
225  *
226  * ---------------
227  * Nethook Support
228  * ---------------
229  *
230  * Viona provides four nethook events that consumers (e.g. ipf) can hook into
231  * to intercept packets as they go up or down the stack.  Unfortunately,
232  * the nethook framework does not understand raw packets, so we can only
233  * generate events (in, out) for IPv4 and IPv6 packets.  At driver attach,
234  * we register callbacks with the neti (netinfo) module that will be invoked
235  * for each netstack already present, as well as for any additional netstack
236  * instances created as the system operates.  These callbacks will
237  * register/unregister the hooks with the nethook framework for each
238  * netstack instance.  This registration occurs prior to creating any
239  * viona instances for a given netstack, and the unregistration for a netstack
240  * instance occurs after all viona instances of the netstack instance have
241  * been deleted.
242  */
243 
244 #include <sys/conf.h>
245 #include <sys/file.h>
246 #include <sys/stat.h>
247 
248 #include <sys/dlpi.h>
249 
250 #include "viona_impl.h"
251 
252 
253 #define	VIONA_NAME		"Virtio Network Accelerator"
254 #define	VIONA_CTL_MINOR		0
255 #define	VIONA_CLI_NAME		"viona"		/* MAC client name */
256 
257 
258 /*
259  * Host capabilities.
260  */
261 #define	VIONA_S_HOSTCAPS	(	\
262 	VIRTIO_NET_F_GUEST_CSUM |	\
263 	VIRTIO_NET_F_MAC |		\
264 	VIRTIO_NET_F_GUEST_TSO4 |	\
265 	VIRTIO_NET_F_MRG_RXBUF |	\
266 	VIRTIO_NET_F_STATUS |		\
267 	VIRTIO_F_RING_NOTIFY_ON_EMPTY |	\
268 	VIRTIO_F_RING_INDIRECT_DESC)
269 
270 /* MAC_CAPAB_HCKSUM specifics of interest */
271 #define	VIONA_CAP_HCKSUM_INTEREST	\
272 	(HCKSUM_INET_PARTIAL |		\
273 	HCKSUM_INET_FULL_V4 |		\
274 	HCKSUM_INET_FULL_V6)
275 
276 static void		*viona_state;
277 static dev_info_t	*viona_dip;
278 static id_space_t	*viona_minors;
279 
280 
281 static int viona_info(dev_info_t *dip, ddi_info_cmd_t cmd, void *arg,
282     void **result);
283 static int viona_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
284 static int viona_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
285 static int viona_open(dev_t *devp, int flag, int otype, cred_t *credp);
286 static int viona_close(dev_t dev, int flag, int otype, cred_t *credp);
287 static int viona_ioctl(dev_t dev, int cmd, intptr_t data, int mode,
288     cred_t *credp, int *rval);
289 static int viona_chpoll(dev_t dev, short events, int anyyet, short *reventsp,
290     struct pollhead **phpp);
291 
292 static int viona_ioc_create(viona_soft_state_t *, void *, int, cred_t *);
293 static int viona_ioc_delete(viona_soft_state_t *, boolean_t);
294 
295 static int viona_ioc_set_notify_ioport(viona_link_t *, uint16_t);
296 static int viona_ioc_ring_init(viona_link_t *, void *, int);
297 static int viona_ioc_ring_set_state(viona_link_t *, void *, int);
298 static int viona_ioc_ring_get_state(viona_link_t *, void *, int);
299 static int viona_ioc_ring_reset(viona_link_t *, uint_t);
300 static int viona_ioc_ring_kick(viona_link_t *, uint_t);
301 static int viona_ioc_ring_pause(viona_link_t *, uint_t);
302 static int viona_ioc_ring_set_msi(viona_link_t *, void *, int);
303 static int viona_ioc_ring_intr_clear(viona_link_t *, uint_t);
304 static int viona_ioc_intr_poll(viona_link_t *, void *, int, int *);
305 
306 static struct cb_ops viona_cb_ops = {
307 	viona_open,
308 	viona_close,
309 	nodev,
310 	nodev,
311 	nodev,
312 	nodev,
313 	nodev,
314 	viona_ioctl,
315 	nodev,
316 	nodev,
317 	nodev,
318 	viona_chpoll,
319 	ddi_prop_op,
320 	0,
321 	D_MP | D_NEW | D_HOTPLUG,
322 	CB_REV,
323 	nodev,
324 	nodev
325 };
326 
327 static struct dev_ops viona_ops = {
328 	DEVO_REV,
329 	0,
330 	viona_info,
331 	nulldev,
332 	nulldev,
333 	viona_attach,
334 	viona_detach,
335 	nodev,
336 	&viona_cb_ops,
337 	NULL,
338 	ddi_power,
339 	ddi_quiesce_not_needed
340 };
341 
342 static struct modldrv modldrv = {
343 	&mod_driverops,
344 	VIONA_NAME,
345 	&viona_ops,
346 };
347 
348 static struct modlinkage modlinkage = {
349 	MODREV_1, &modldrv, NULL
350 };
351 
352 int
353 _init(void)
354 {
355 	int ret;
356 
357 	ret = ddi_soft_state_init(&viona_state, sizeof (viona_soft_state_t), 0);
358 	if (ret != 0) {
359 		return (ret);
360 	}
361 
362 	viona_minors = id_space_create("viona_minors",
363 	    VIONA_CTL_MINOR + 1, UINT16_MAX);
364 	viona_rx_init();
365 	mutex_init(&viona_force_copy_lock, NULL, MUTEX_DRIVER, NULL);
366 
367 	ret = mod_install(&modlinkage);
368 	if (ret != 0) {
369 		ddi_soft_state_fini(&viona_state);
370 		id_space_destroy(viona_minors);
371 		viona_rx_fini();
372 		mutex_destroy(&viona_force_copy_lock);
373 	}
374 
375 	return (ret);
376 }
377 
378 int
379 _fini(void)
380 {
381 	int ret;
382 
383 	ret = mod_remove(&modlinkage);
384 	if (ret != 0) {
385 		return (ret);
386 	}
387 
388 	ddi_soft_state_fini(&viona_state);
389 	id_space_destroy(viona_minors);
390 	viona_rx_fini();
391 	mutex_destroy(&viona_force_copy_lock);
392 
393 	return (ret);
394 }
395 
396 int
397 _info(struct modinfo *modinfop)
398 {
399 	return (mod_info(&modlinkage, modinfop));
400 }
401 
402 /* ARGSUSED */
403 static int
404 viona_info(dev_info_t *dip, ddi_info_cmd_t cmd, void *arg, void **result)
405 {
406 	int error;
407 
408 	switch (cmd) {
409 	case DDI_INFO_DEVT2DEVINFO:
410 		*result = (void *)viona_dip;
411 		error = DDI_SUCCESS;
412 		break;
413 	case DDI_INFO_DEVT2INSTANCE:
414 		*result = (void *)0;
415 		error = DDI_SUCCESS;
416 		break;
417 	default:
418 		error = DDI_FAILURE;
419 		break;
420 	}
421 	return (error);
422 }
423 
424 static int
425 viona_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
426 {
427 	if (cmd != DDI_ATTACH) {
428 		return (DDI_FAILURE);
429 	}
430 
431 	if (ddi_create_minor_node(dip, "viona", S_IFCHR, VIONA_CTL_MINOR,
432 	    DDI_PSEUDO, 0) != DDI_SUCCESS) {
433 		return (DDI_FAILURE);
434 	}
435 
436 	viona_neti_attach();
437 
438 	viona_dip = dip;
439 	ddi_report_dev(viona_dip);
440 
441 	return (DDI_SUCCESS);
442 }
443 
444 static int
445 viona_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
446 {
447 	dev_info_t *old_dip = viona_dip;
448 
449 	if (cmd != DDI_DETACH) {
450 		return (DDI_FAILURE);
451 	}
452 
453 	VERIFY(old_dip != NULL);
454 
455 	viona_neti_detach();
456 	viona_dip = NULL;
457 	ddi_remove_minor_node(old_dip, NULL);
458 
459 	return (DDI_SUCCESS);
460 }
461 
462 static int
463 viona_open(dev_t *devp, int flag, int otype, cred_t *credp)
464 {
465 	int	minor;
466 	viona_soft_state_t *ss;
467 
468 	if (otype != OTYP_CHR) {
469 		return (EINVAL);
470 	}
471 #if 0
472 	/*
473 	 * XXX-mg: drv_priv() is wrong, but I'm not sure what is right.
474 	 * Should the check be at open() or ioctl()?
475 	 */
476 	if (drv_priv(credp) != 0) {
477 		return (EPERM);
478 	}
479 #endif
480 	if (getminor(*devp) != VIONA_CTL_MINOR) {
481 		return (ENXIO);
482 	}
483 
484 	minor = id_alloc_nosleep(viona_minors);
485 	if (minor == -1) {
486 		/* All minors are busy */
487 		return (EBUSY);
488 	}
489 	if (ddi_soft_state_zalloc(viona_state, minor) != DDI_SUCCESS) {
490 		id_free(viona_minors, minor);
491 		return (ENOMEM);
492 	}
493 
494 	ss = ddi_get_soft_state(viona_state, minor);
495 	mutex_init(&ss->ss_lock, NULL, MUTEX_DEFAULT, NULL);
496 	*devp = makedevice(getmajor(*devp), minor);
497 
498 	return (0);
499 }
500 
501 static int
502 viona_close(dev_t dev, int flag, int otype, cred_t *credp)
503 {
504 	int			minor;
505 	viona_soft_state_t	*ss;
506 
507 	if (otype != OTYP_CHR) {
508 		return (EINVAL);
509 	}
510 
511 	minor = getminor(dev);
512 
513 	ss = ddi_get_soft_state(viona_state, minor);
514 	if (ss == NULL) {
515 		return (ENXIO);
516 	}
517 
518 	VERIFY0(viona_ioc_delete(ss, B_TRUE));
519 	VERIFY(!list_link_active(&ss->ss_node));
520 	ddi_soft_state_free(viona_state, minor);
521 	id_free(viona_minors, minor);
522 
523 	return (0);
524 }
525 
526 static int
527 viona_ioctl(dev_t dev, int cmd, intptr_t data, int md, cred_t *cr, int *rv)
528 {
529 	viona_soft_state_t *ss;
530 	void *dptr = (void *)data;
531 	int err = 0, val;
532 	viona_link_t *link;
533 
534 	ss = ddi_get_soft_state(viona_state, getminor(dev));
535 	if (ss == NULL) {
536 		return (ENXIO);
537 	}
538 
539 	switch (cmd) {
540 	case VNA_IOC_CREATE:
541 		return (viona_ioc_create(ss, dptr, md, cr));
542 	case VNA_IOC_DELETE:
543 		return (viona_ioc_delete(ss, B_FALSE));
544 	case VNA_IOC_VERSION:
545 		*rv = VIONA_CURRENT_INTERFACE_VERSION;
546 		return (0);
547 	default:
548 		break;
549 	}
550 
551 	mutex_enter(&ss->ss_lock);
552 	if ((link = ss->ss_link) == NULL || link->l_destroyed ||
553 	    vmm_drv_release_reqd(link->l_vm_hold)) {
554 		mutex_exit(&ss->ss_lock);
555 		return (ENXIO);
556 	}
557 
558 	switch (cmd) {
559 	case VNA_IOC_GET_FEATURES:
560 		val = VIONA_S_HOSTCAPS | link->l_features_hw;
561 		if (ddi_copyout(&val, dptr, sizeof (val), md) != 0) {
562 			err = EFAULT;
563 		}
564 		break;
565 	case VNA_IOC_SET_FEATURES:
566 		if (ddi_copyin(dptr, &val, sizeof (val), md) != 0) {
567 			err = EFAULT;
568 			break;
569 		}
570 		val &= (VIONA_S_HOSTCAPS | link->l_features_hw);
571 
572 		if ((val & VIRTIO_NET_F_CSUM) == 0)
573 			val &= ~VIRTIO_NET_F_HOST_TSO4;
574 
575 		if ((val & VIRTIO_NET_F_GUEST_CSUM) == 0)
576 			val &= ~VIRTIO_NET_F_GUEST_TSO4;
577 
578 		link->l_features = val;
579 		break;
580 	case VNA_IOC_RING_INIT:
581 		err = viona_ioc_ring_init(link, dptr, md);
582 		break;
583 	case VNA_IOC_RING_RESET:
584 		err = viona_ioc_ring_reset(link, (uint_t)data);
585 		break;
586 	case VNA_IOC_RING_KICK:
587 		err = viona_ioc_ring_kick(link, (uint_t)data);
588 		break;
589 	case VNA_IOC_RING_SET_MSI:
590 		err = viona_ioc_ring_set_msi(link, dptr, md);
591 		break;
592 	case VNA_IOC_RING_INTR_CLR:
593 		err = viona_ioc_ring_intr_clear(link, (uint_t)data);
594 		break;
595 	case VNA_IOC_RING_SET_STATE:
596 		err = viona_ioc_ring_set_state(link, dptr, md);
597 		break;
598 	case VNA_IOC_RING_GET_STATE:
599 		err = viona_ioc_ring_get_state(link, dptr, md);
600 		break;
601 	case VNA_IOC_RING_PAUSE:
602 		err = viona_ioc_ring_pause(link, (uint_t)data);
603 		break;
604 
605 	case VNA_IOC_INTR_POLL:
606 		err = viona_ioc_intr_poll(link, dptr, md, rv);
607 		break;
608 	case VNA_IOC_SET_NOTIFY_IOP:
609 		if (data < 0 || data > UINT16_MAX) {
610 			err = EINVAL;
611 			break;
612 		}
613 		err = viona_ioc_set_notify_ioport(link, (uint16_t)data);
614 		break;
615 	default:
616 		err = ENOTTY;
617 		break;
618 	}
619 
620 	mutex_exit(&ss->ss_lock);
621 	return (err);
622 }
623 
624 static int
625 viona_chpoll(dev_t dev, short events, int anyyet, short *reventsp,
626     struct pollhead **phpp)
627 {
628 	viona_soft_state_t *ss;
629 	viona_link_t *link;
630 
631 	ss = ddi_get_soft_state(viona_state, getminor(dev));
632 	if (ss == NULL) {
633 		return (ENXIO);
634 	}
635 
636 	mutex_enter(&ss->ss_lock);
637 	if ((link = ss->ss_link) == NULL || link->l_destroyed) {
638 		mutex_exit(&ss->ss_lock);
639 		return (ENXIO);
640 	}
641 
642 	*reventsp = 0;
643 	if ((events & POLLRDBAND) != 0) {
644 		for (uint_t i = 0; i < VIONA_VQ_MAX; i++) {
645 			if (link->l_vrings[i].vr_intr_enabled != 0) {
646 				*reventsp |= POLLRDBAND;
647 				break;
648 			}
649 		}
650 	}
651 	if ((*reventsp == 0 && !anyyet) || (events & POLLET)) {
652 		*phpp = &link->l_pollhead;
653 	}
654 	mutex_exit(&ss->ss_lock);
655 
656 	return (0);
657 }
658 
659 static void
660 viona_get_mac_capab(viona_link_t *link)
661 {
662 	mac_handle_t mh = link->l_mh;
663 	uint32_t cap = 0;
664 	mac_capab_lso_t lso_cap;
665 
666 	link->l_features_hw = 0;
667 	if (mac_capab_get(mh, MAC_CAPAB_HCKSUM, &cap)) {
668 		/*
669 		 * Only report HW checksum ability if the underlying MAC
670 		 * resource is capable of populating the L4 header.
671 		 */
672 		if ((cap & VIONA_CAP_HCKSUM_INTEREST) != 0) {
673 			link->l_features_hw |= VIRTIO_NET_F_CSUM;
674 		}
675 		link->l_cap_csum = cap;
676 	}
677 
678 	if ((link->l_features_hw & VIRTIO_NET_F_CSUM) &&
679 	    mac_capab_get(mh, MAC_CAPAB_LSO, &lso_cap)) {
680 		/*
681 		 * Virtio doesn't allow for negotiating a maximum LSO
682 		 * packet size. We have to assume that the guest may
683 		 * send a maximum length IP packet. Make sure the
684 		 * underlying MAC can handle an LSO of this size.
685 		 */
686 		if ((lso_cap.lso_flags & LSO_TX_BASIC_TCP_IPV4) &&
687 		    lso_cap.lso_basic_tcp_ipv4.lso_max >= IP_MAXPACKET)
688 			link->l_features_hw |= VIRTIO_NET_F_HOST_TSO4;
689 	}
690 }
691 
692 static int
693 viona_ioc_create(viona_soft_state_t *ss, void *dptr, int md, cred_t *cr)
694 {
695 	vioc_create_t	kvc;
696 	viona_link_t	*link = NULL;
697 	char		cli_name[MAXNAMELEN];
698 	int		err = 0;
699 	file_t		*fp;
700 	vmm_hold_t	*hold = NULL;
701 	viona_neti_t	*nip = NULL;
702 	zoneid_t	zid;
703 
704 	ASSERT(MUTEX_NOT_HELD(&ss->ss_lock));
705 
706 	if (ddi_copyin(dptr, &kvc, sizeof (kvc), md) != 0) {
707 		return (EFAULT);
708 	}
709 
710 	zid = crgetzoneid(cr);
711 	nip = viona_neti_lookup_by_zid(zid);
712 	if (nip == NULL) {
713 		return (EIO);
714 	}
715 
716 	if (!nip->vni_nethook.vnh_hooked) {
717 		viona_neti_rele(nip);
718 		return (EIO);
719 	}
720 
721 	mutex_enter(&ss->ss_lock);
722 	if (ss->ss_link != NULL) {
723 		mutex_exit(&ss->ss_lock);
724 		viona_neti_rele(nip);
725 		return (EEXIST);
726 	}
727 
728 	if ((fp = getf(kvc.c_vmfd)) == NULL) {
729 		err = EBADF;
730 		goto bail;
731 	}
732 	err = vmm_drv_hold(fp, cr, &hold);
733 	releasef(kvc.c_vmfd);
734 	if (err != 0) {
735 		goto bail;
736 	}
737 
738 	link = kmem_zalloc(sizeof (viona_link_t), KM_SLEEP);
739 	link->l_linkid = kvc.c_linkid;
740 	link->l_vm_hold = hold;
741 
742 	err = mac_open_by_linkid(link->l_linkid, &link->l_mh);
743 	if (err != 0) {
744 		goto bail;
745 	}
746 
747 	viona_get_mac_capab(link);
748 
749 	(void) snprintf(cli_name, sizeof (cli_name), "%s-%d", VIONA_CLI_NAME,
750 	    link->l_linkid);
751 	err = mac_client_open(link->l_mh, &link->l_mch, cli_name, 0);
752 	if (err != 0) {
753 		goto bail;
754 	}
755 
756 	viona_ring_alloc(link, &link->l_vrings[VIONA_VQ_RX]);
757 	viona_ring_alloc(link, &link->l_vrings[VIONA_VQ_TX]);
758 
759 	if ((err = viona_rx_set(link)) != 0) {
760 		viona_ring_free(&link->l_vrings[VIONA_VQ_RX]);
761 		viona_ring_free(&link->l_vrings[VIONA_VQ_TX]);
762 		goto bail;
763 	}
764 
765 	link->l_neti = nip;
766 	ss->ss_link = link;
767 	mutex_exit(&ss->ss_lock);
768 
769 	mutex_enter(&nip->vni_lock);
770 	list_insert_tail(&nip->vni_dev_list, ss);
771 	mutex_exit(&nip->vni_lock);
772 
773 	return (0);
774 
775 bail:
776 	if (link != NULL) {
777 		if (link->l_mch != NULL) {
778 			mac_client_close(link->l_mch, 0);
779 		}
780 		if (link->l_mh != NULL) {
781 			mac_close(link->l_mh);
782 		}
783 		kmem_free(link, sizeof (viona_link_t));
784 	}
785 	if (hold != NULL) {
786 		vmm_drv_rele(hold);
787 	}
788 	viona_neti_rele(nip);
789 
790 	mutex_exit(&ss->ss_lock);
791 	return (err);
792 }
793 
794 static int
795 viona_ioc_delete(viona_soft_state_t *ss, boolean_t on_close)
796 {
797 	viona_link_t *link;
798 	viona_neti_t *nip = NULL;
799 
800 	mutex_enter(&ss->ss_lock);
801 	if ((link = ss->ss_link) == NULL) {
802 		/* Link destruction already complete */
803 		mutex_exit(&ss->ss_lock);
804 		return (0);
805 	}
806 
807 	if (link->l_destroyed) {
808 		/*
809 		 * Link destruction has been started by another thread, but has
810 		 * not completed.  This condition should be impossible to
811 		 * encounter when performing the on-close destroy of the link,
812 		 * since racing ioctl accessors must necessarily be absent.
813 		 */
814 		VERIFY(!on_close);
815 		mutex_exit(&ss->ss_lock);
816 		return (EAGAIN);
817 	}
818 	/*
819 	 * The link deletion cannot fail after this point, continuing until its
820 	 * successful completion is reached.
821 	 */
822 	link->l_destroyed = B_TRUE;
823 
824 	/*
825 	 * Tear down the IO port hook so it cannot be used to kick any of the
826 	 * rings which are about to be reset and stopped.
827 	 */
828 	VERIFY0(viona_ioc_set_notify_ioport(link, 0));
829 	mutex_exit(&ss->ss_lock);
830 
831 	/*
832 	 * Return the rings to their reset state, ignoring any possible
833 	 * interruptions from signals.
834 	 */
835 	VERIFY0(viona_ring_reset(&link->l_vrings[VIONA_VQ_RX], B_FALSE));
836 	VERIFY0(viona_ring_reset(&link->l_vrings[VIONA_VQ_TX], B_FALSE));
837 
838 	mutex_enter(&ss->ss_lock);
839 	if (link->l_mch != NULL) {
840 		/* Unhook the receive callbacks and close out the client */
841 		viona_rx_clear(link);
842 		mac_client_close(link->l_mch, 0);
843 	}
844 	if (link->l_mh != NULL) {
845 		mac_close(link->l_mh);
846 	}
847 	if (link->l_vm_hold != NULL) {
848 		vmm_drv_rele(link->l_vm_hold);
849 		link->l_vm_hold = NULL;
850 	}
851 
852 	nip = link->l_neti;
853 	link->l_neti = NULL;
854 
855 	viona_ring_free(&link->l_vrings[VIONA_VQ_RX]);
856 	viona_ring_free(&link->l_vrings[VIONA_VQ_TX]);
857 	pollhead_clean(&link->l_pollhead);
858 	ss->ss_link = NULL;
859 	mutex_exit(&ss->ss_lock);
860 
861 	mutex_enter(&nip->vni_lock);
862 	list_remove(&nip->vni_dev_list, ss);
863 	mutex_exit(&nip->vni_lock);
864 
865 	viona_neti_rele(nip);
866 
867 	kmem_free(link, sizeof (viona_link_t));
868 	return (0);
869 }
870 
871 static int
872 viona_ioc_ring_init(viona_link_t *link, void *udata, int md)
873 {
874 	vioc_ring_init_t kri;
875 	int err;
876 
877 	if (ddi_copyin(udata, &kri, sizeof (kri), md) != 0) {
878 		return (EFAULT);
879 	}
880 	const struct viona_ring_params params = {
881 		.vrp_pa = kri.ri_qaddr,
882 		.vrp_size = kri.ri_qsize,
883 		.vrp_avail_idx = 0,
884 		.vrp_used_idx = 0,
885 	};
886 
887 	err = viona_ring_init(link, kri.ri_index, &params);
888 
889 	return (err);
890 }
891 
892 static int
893 viona_ioc_ring_set_state(viona_link_t *link, void *udata, int md)
894 {
895 	vioc_ring_state_t krs;
896 	int err;
897 
898 	if (ddi_copyin(udata, &krs, sizeof (krs), md) != 0) {
899 		return (EFAULT);
900 	}
901 	const struct viona_ring_params params = {
902 		.vrp_pa = krs.vrs_qaddr,
903 		.vrp_size = krs.vrs_qsize,
904 		.vrp_avail_idx = krs.vrs_avail_idx,
905 		.vrp_used_idx = krs.vrs_used_idx,
906 	};
907 
908 	err = viona_ring_init(link, krs.vrs_index, &params);
909 
910 	return (err);
911 }
912 
913 static int
914 viona_ioc_ring_get_state(viona_link_t *link, void *udata, int md)
915 {
916 	vioc_ring_state_t krs;
917 
918 	if (ddi_copyin(udata, &krs, sizeof (krs), md) != 0) {
919 		return (EFAULT);
920 	}
921 
922 	struct viona_ring_params params;
923 	int err = viona_ring_get_state(link, krs.vrs_index, &params);
924 	if (err != 0) {
925 		return (err);
926 	}
927 	krs.vrs_qsize = params.vrp_size;
928 	krs.vrs_qaddr = params.vrp_pa;
929 	krs.vrs_avail_idx = params.vrp_avail_idx;
930 	krs.vrs_used_idx = params.vrp_used_idx;
931 
932 	if (ddi_copyout(&krs, udata, sizeof (krs), md) != 0) {
933 		return (EFAULT);
934 	}
935 	return (0);
936 }
937 
938 static int
939 viona_ioc_ring_reset(viona_link_t *link, uint_t idx)
940 {
941 	viona_vring_t *ring;
942 
943 	if (idx >= VIONA_VQ_MAX) {
944 		return (EINVAL);
945 	}
946 	ring = &link->l_vrings[idx];
947 
948 	return (viona_ring_reset(ring, B_TRUE));
949 }
950 
951 static int
952 viona_ioc_ring_kick(viona_link_t *link, uint_t idx)
953 {
954 	viona_vring_t *ring;
955 	int err;
956 
957 	if (idx >= VIONA_VQ_MAX) {
958 		return (EINVAL);
959 	}
960 	ring = &link->l_vrings[idx];
961 
962 	mutex_enter(&ring->vr_lock);
963 	switch (ring->vr_state) {
964 	case VRS_SETUP:
965 		/*
966 		 * An early kick to a ring which is starting its worker thread
967 		 * is fine.  Once that thread is active, it will process the
968 		 * start-up request immediately.
969 		 */
970 		/* FALLTHROUGH */
971 	case VRS_INIT:
972 		ring->vr_state_flags |= VRSF_REQ_START;
973 		/* FALLTHROUGH */
974 	case VRS_RUN:
975 		cv_broadcast(&ring->vr_cv);
976 		err = 0;
977 		break;
978 	default:
979 		err = EBUSY;
980 		break;
981 	}
982 	mutex_exit(&ring->vr_lock);
983 
984 	return (err);
985 }
986 
987 static int
988 viona_ioc_ring_pause(viona_link_t *link, uint_t idx)
989 {
990 	if (idx >= VIONA_VQ_MAX) {
991 		return (EINVAL);
992 	}
993 
994 	viona_vring_t *ring = &link->l_vrings[idx];
995 	return (viona_ring_pause(ring));
996 }
997 
998 static int
999 viona_ioc_ring_set_msi(viona_link_t *link, void *data, int md)
1000 {
1001 	vioc_ring_msi_t vrm;
1002 	viona_vring_t *ring;
1003 
1004 	if (ddi_copyin(data, &vrm, sizeof (vrm), md) != 0) {
1005 		return (EFAULT);
1006 	}
1007 	if (vrm.rm_index >= VIONA_VQ_MAX) {
1008 		return (EINVAL);
1009 	}
1010 
1011 	ring = &link->l_vrings[vrm.rm_index];
1012 	mutex_enter(&ring->vr_lock);
1013 	ring->vr_msi_addr = vrm.rm_addr;
1014 	ring->vr_msi_msg = vrm.rm_msg;
1015 	mutex_exit(&ring->vr_lock);
1016 
1017 	return (0);
1018 }
1019 
1020 static int
1021 viona_notify_iop(void *arg, bool in, uint16_t port, uint8_t bytes,
1022     uint32_t *val)
1023 {
1024 	viona_link_t *link = (viona_link_t *)arg;
1025 
1026 	/*
1027 	 * If the request is a read (in/ins), or direct at a port other than
1028 	 * what we expect to be registered on, ignore it.
1029 	 */
1030 	if (in || port != link->l_notify_ioport) {
1031 		return (ESRCH);
1032 	}
1033 
1034 	/* Let userspace handle notifications for rings other than RX/TX. */
1035 	const uint16_t vq = *val;
1036 	if (vq >= VIONA_VQ_MAX) {
1037 		return (ESRCH);
1038 	}
1039 
1040 	viona_vring_t *ring = &link->l_vrings[vq];
1041 	int res = 0;
1042 
1043 	mutex_enter(&ring->vr_lock);
1044 	if (ring->vr_state == VRS_RUN) {
1045 		cv_broadcast(&ring->vr_cv);
1046 	} else {
1047 		res = ESRCH;
1048 	}
1049 	mutex_exit(&ring->vr_lock);
1050 
1051 	return (res);
1052 }
1053 
1054 static int
1055 viona_ioc_set_notify_ioport(viona_link_t *link, uint16_t ioport)
1056 {
1057 	int err = 0;
1058 
1059 	if (link->l_notify_ioport != 0) {
1060 		vmm_drv_ioport_unhook(link->l_vm_hold, &link->l_notify_cookie);
1061 		link->l_notify_ioport = 0;
1062 	}
1063 
1064 	if (ioport != 0) {
1065 		err = vmm_drv_ioport_hook(link->l_vm_hold, ioport,
1066 		    viona_notify_iop, (void *)link, &link->l_notify_cookie);
1067 		if (err == 0) {
1068 			link->l_notify_ioport = ioport;
1069 		}
1070 	}
1071 	return (err);
1072 }
1073 
1074 static int
1075 viona_ioc_ring_intr_clear(viona_link_t *link, uint_t idx)
1076 {
1077 	if (idx >= VIONA_VQ_MAX) {
1078 		return (EINVAL);
1079 	}
1080 
1081 	link->l_vrings[idx].vr_intr_enabled = 0;
1082 	return (0);
1083 }
1084 
1085 static int
1086 viona_ioc_intr_poll(viona_link_t *link, void *udata, int md, int *rv)
1087 {
1088 	uint_t cnt = 0;
1089 	vioc_intr_poll_t vip;
1090 
1091 	for (uint_t i = 0; i < VIONA_VQ_MAX; i++) {
1092 		uint_t val = link->l_vrings[i].vr_intr_enabled;
1093 
1094 		vip.vip_status[i] = val;
1095 		if (val != 0) {
1096 			cnt++;
1097 		}
1098 	}
1099 
1100 	if (ddi_copyout(&vip, udata, sizeof (vip), md) != 0) {
1101 		return (EFAULT);
1102 	}
1103 	*rv = (int)cnt;
1104 	return (0);
1105 }
1106