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 OmniOS Community Edition (OmniOSce) Association.
39 * Copyright 2024 Oxide Computer Company
40 */
41
42 /*
43 * viona - VirtIO-Net, Accelerated
44 *
45 * The purpose of viona is to provide high performance virtio-net devices to
46 * bhyve guests. It does so by sitting directly atop MAC, skipping all of the
47 * DLS/DLD stack.
48 *
49 * --------------------
50 * General Architecture
51 * --------------------
52 *
53 * A single viona instance is comprised of a "link" handle and two "rings".
54 * After opening the viona device, it must be associated with a MAC network
55 * interface and a bhyve (vmm) instance to form its link resource. This is
56 * done with the VNA_IOC_CREATE ioctl, where the datalink ID and vmm fd are
57 * passed in to perform the initialization. With the MAC client opened, and a
58 * driver handle to the vmm instance established, the device is ready to be
59 * configured by the guest.
60 *
61 * The userspace portion of bhyve, which interfaces with the PCI device
62 * emulation framework, is meant to stay out of the datapath if at all
63 * possible. Configuration changes made via PCI are mapped to actions which
64 * will steer the operation of the in-kernel logic.
65 *
66 *
67 * -----------
68 * Ring Basics
69 * -----------
70 *
71 * Each viona link has two viona_vring_t entities, RX and TX, for handling data
72 * transfers to and from the guest. They represent an interface to the
73 * standard virtio ring structures. When initialized and active, each ring is
74 * backed by a kernel worker thread (parented to the bhyve process for the
75 * instance) which handles ring events. The RX worker has the simple task of
76 * watching for ring shutdown conditions. The TX worker does that in addition
77 * to processing all requests to transmit data. Data destined for the guest is
78 * delivered directly by MAC to viona_rx() when the ring is active.
79 *
80 *
81 * -----------
82 * Ring States
83 * -----------
84 *
85 * The viona_vring_t instances follow a simple path through the possible state
86 * values represented in virtio_vring_t`vr_state:
87 *
88 * +<--------------------------------------------+
89 * | |
90 * V ^
91 * +-----------+ This is the initial state when a link is created or
92 * | VRS_RESET | when the ring has been explicitly reset.
93 * +-----------+
94 * | ^
95 * |---* ioctl(VNA_IOC_RING_INIT) issued |
96 * | |
97 * | ^
98 * V
99 * +-----------+ The ring parameters (size, guest physical addresses)
100 * | VRS_SETUP | have been set and start-up of the ring worker thread
101 * +-----------+ has begun.
102 * | ^
103 * | |
104 * |---* ring worker thread begins execution |
105 * | |
106 * +-------------------------------------------->+
107 * | | ^
108 * | |
109 * | * If ring shutdown is requested (by ioctl or impending
110 * | bhyve process death) while the worker thread is
111 * | starting, the worker will transition the ring to
112 * | VRS_RESET and exit.
113 * | ^
114 * | |
115 * |<-------------------------------------------<+
116 * | | |
117 * | | ^
118 * | * If ring is requested to pause (but not stop)from the
119 * | VRS_RUN state, it will return to the VRS_INIT state.
120 * |
121 * | ^
122 * | |
123 * | ^
124 * V
125 * +-----------+ The worker thread associated with the ring has started
126 * | VRS_INIT | executing. It has allocated any extra resources needed
127 * +-----------+ for the ring to operate.
128 * | ^
129 * | |
130 * +-------------------------------------------->+
131 * | | ^
132 * | |
133 * | * If ring shutdown is requested while the worker is
134 * | waiting in VRS_INIT, it will free any extra resources
135 * | and transition to VRS_RESET.
136 * | ^
137 * | |
138 * |--* ioctl(VNA_IOC_RING_KICK) issued |
139 * | ^
140 * V
141 * +-----------+ The worker thread associated with the ring is executing
142 * | VRS_RUN | workload specific to that ring.
143 * +-----------+
144 * | ^
145 * |---* ioctl(VNA_IOC_RING_RESET) issued |
146 * | (or bhyve process begins exit) ^
147 * |
148 * +-----------+ The worker thread associated with the ring is in the
149 * | VRS_STOP | process of exiting. All outstanding TX and RX
150 * +-----------+ requests are allowed to complete, but new requests
151 * | must be ignored.
152 * | ^
153 * | |
154 * +-------------------------------------------->+
155 *
156 *
157 * While the worker thread is not running, changes to vr_state are only made by
158 * viona_ioc_ring_init() under vr_lock. There, it initializes the ring, starts
159 * the worker, and sets the ring state to VRS_SETUP. Once the worker thread
160 * has been started, only it may perform ring state transitions (still under
161 * the protection of vr_lock), when requested by outside consumers via
162 * vr_state_flags or when the containing bhyve process initiates an exit.
163 *
164 *
165 * ----------------------------
166 * Transmission mblk_t Handling
167 * ----------------------------
168 *
169 * For incoming frames destined for a bhyve guest, the data must first land in
170 * a host OS buffer from the physical NIC before it is copied into the awaiting
171 * guest buffer(s). Outbound frames transmitted by the guest are not bound by
172 * this limitation and can avoid extra copying before the buffers are accessed
173 * directly by the NIC. When a guest designates buffers to be transmitted,
174 * viona translates the guest-physical addresses contained in the ring
175 * descriptors to host-virtual addresses via viona_hold_page(). That pointer is
176 * wrapped in an mblk_t using a preallocated viona_desb_t for the desballoc().
177 * Doing so increments vr_xfer_outstanding, preventing the ring from being
178 * reset (allowing the link to drop its vmm handle to the guest) until all
179 * transmit mblks referencing guest memory have been processed. Allocation of
180 * the viona_desb_t entries is done during the VRS_INIT stage of the ring
181 * worker thread. The ring size informs that allocation as the number of
182 * concurrent transmissions is limited by the number of descriptors in the
183 * ring. This minimizes allocation in the transmit hot-path by acquiring those
184 * fixed-size resources during initialization.
185 *
186 * This optimization depends on the underlying NIC driver freeing the mblks in
187 * a timely manner after they have been transmitted by the hardware. Some
188 * drivers have been found to flush TX descriptors only when new transmissions
189 * are initiated. This means that there is no upper bound to the time needed
190 * for an mblk to be flushed and can stall bhyve guests from shutting down
191 * since their memory must be free of viona TX references prior to clean-up.
192 *
193 * This expectation of deterministic mblk_t processing is likely the reason
194 * behind the notable exception to the zero-copy TX path: systems with 'bnxe'
195 * loaded will copy transmit data into fresh buffers rather than passing up
196 * zero-copy mblks. It is a hold-over from the original viona sources provided
197 * by Pluribus and its continued necessity has not been confirmed.
198 *
199 *
200 * ----------------------------
201 * Ring Notification Fast-paths
202 * ----------------------------
203 *
204 * Device operation for viona requires that notifications flow to and from the
205 * guest to indicate certain ring conditions. In order to minimize latency and
206 * processing overhead, the notification procedures are kept in-kernel whenever
207 * possible.
208 *
209 * Guest-to-host notifications, when new available descriptors have been placed
210 * in the ring, are posted via the 'queue notify' address in the virtio BAR.
211 * The vmm_drv_ioport_hook() interface was added to bhyve which allows viona to
212 * install a callback hook on an ioport address. Guest exits for accesses to
213 * viona-hooked ioport addresses will result in direct calls to notify the
214 * appropriate ring worker without a trip to userland.
215 *
216 * Host-to-guest notifications in the form of interrupts enjoy similar
217 * acceleration. Each viona ring can be configured to send MSI notifications
218 * to the guest as virtio conditions dictate. This in-kernel interrupt
219 * configuration is kept synchronized through viona ioctls which are utilized
220 * during writes to the associated PCI config registers or MSI-X BAR.
221 *
222 * Guests which do not utilize MSI-X will result in viona falling back to the
223 * slow path for interrupts. It will poll(2) the viona handle, receiving
224 * notification when ring events necessitate the assertion of an interrupt.
225 *
226 *
227 * ---------------
228 * Nethook Support
229 * ---------------
230 *
231 * Viona provides four nethook events that consumers (e.g. ipf) can hook into
232 * to intercept packets as they go up or down the stack. Unfortunately,
233 * the nethook framework does not understand raw packets, so we can only
234 * generate events (in, out) for IPv4 and IPv6 packets. At driver attach,
235 * we register callbacks with the neti (netinfo) module that will be invoked
236 * for each netstack already present, as well as for any additional netstack
237 * instances created as the system operates. These callbacks will
238 * register/unregister the hooks with the nethook framework for each
239 * netstack instance. This registration occurs prior to creating any
240 * viona instances for a given netstack, and the unregistration for a netstack
241 * instance occurs after all viona instances of the netstack instance have
242 * been deleted.
243 *
244 * ------------------
245 * Metrics/Statistics
246 * -----------------
247 *
248 * During operation, Viona tracks certain metrics as certain events occur.
249 *
250 * One class of metrics, known as the "error stats", refer to abnormal
251 * conditions in ring processing which are likely the fault of a misbehaving
252 * guest. These are tracked on a per-ring basis, and are not formally exposed
253 * to any consumer besides direct memory access through mdb.
254 *
255 * The other class of metrics tracked for an instance are the "transfer stats",
256 * which are the traditional packets/bytes/errors/drops figures. These are
257 * counted per-ring, and then aggregated into link-wide values exposed via
258 * kstats. Atomic operations are used to increment those per-ring stats during
259 * operation, and then when a ring is stopped, the values are consolidated into
260 * the link-wide values (to prevent loss when the ring is zeroed) under the
261 * protection of viona_link`l_stats_lock. When the kstats are being updated,
262 * l_stats_lock is held to protect against a racing consolidation, with the
263 * existing per-ring values being added in at update time to provide an accurate
264 * figure.
265 */
266
267 #include <sys/conf.h>
268 #include <sys/file.h>
269 #include <sys/stat.h>
270
271 #include <sys/dlpi.h>
272 #include <sys/vlan.h>
273
274 #include "viona_impl.h"
275
276
277 #define VIONA_NAME "Virtio Network Accelerator"
278 #define VIONA_CTL_MINOR 0
279 #define VIONA_MODULE_NAME "viona"
280 #define VIONA_KSTAT_CLASS "misc"
281 #define VIONA_KSTAT_NAME "viona_stat"
282
283
284 /*
285 * Host capabilities.
286 */
287 #define VIONA_S_HOSTCAPS ( \
288 VIRTIO_NET_F_GUEST_CSUM | \
289 VIRTIO_NET_F_MAC | \
290 VIRTIO_NET_F_GUEST_TSO4 | \
291 VIRTIO_NET_F_MRG_RXBUF | \
292 VIRTIO_NET_F_STATUS | \
293 VIRTIO_F_RING_NOTIFY_ON_EMPTY | \
294 VIRTIO_F_RING_INDIRECT_DESC)
295
296 /* MAC_CAPAB_HCKSUM specifics of interest */
297 #define VIONA_CAP_HCKSUM_INTEREST \
298 (HCKSUM_INET_PARTIAL | \
299 HCKSUM_INET_FULL_V4 | \
300 HCKSUM_INET_FULL_V6)
301
302 static void *viona_state;
303 static dev_info_t *viona_dip;
304 static id_space_t *viona_minors;
305
306
307 static int viona_info(dev_info_t *dip, ddi_info_cmd_t cmd, void *arg,
308 void **result);
309 static int viona_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
310 static int viona_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
311 static int viona_open(dev_t *devp, int flag, int otype, cred_t *credp);
312 static int viona_close(dev_t dev, int flag, int otype, cred_t *credp);
313 static int viona_ioctl(dev_t dev, int cmd, intptr_t data, int mode,
314 cred_t *credp, int *rval);
315 static int viona_chpoll(dev_t dev, short events, int anyyet, short *reventsp,
316 struct pollhead **phpp);
317
318 static int viona_ioc_create(viona_soft_state_t *, void *, int, cred_t *);
319 static int viona_ioc_delete(viona_soft_state_t *, boolean_t);
320
321 static int viona_ioc_set_notify_ioport(viona_link_t *, uint16_t);
322 static int viona_ioc_set_promisc(viona_link_t *, viona_promisc_t);
323 static int viona_ioc_ring_init(viona_link_t *, void *, int);
324 static int viona_ioc_ring_set_state(viona_link_t *, void *, int);
325 static int viona_ioc_ring_get_state(viona_link_t *, void *, int);
326 static int viona_ioc_ring_reset(viona_link_t *, uint_t);
327 static int viona_ioc_ring_kick(viona_link_t *, uint_t);
328 static int viona_ioc_ring_pause(viona_link_t *, uint_t);
329 static int viona_ioc_ring_set_msi(viona_link_t *, void *, int);
330 static int viona_ioc_ring_intr_clear(viona_link_t *, uint_t);
331 static int viona_ioc_intr_poll(viona_link_t *, void *, int, int *);
332
333 static struct cb_ops viona_cb_ops = {
334 viona_open,
335 viona_close,
336 nodev,
337 nodev,
338 nodev,
339 nodev,
340 nodev,
341 viona_ioctl,
342 nodev,
343 nodev,
344 nodev,
345 viona_chpoll,
346 ddi_prop_op,
347 0,
348 D_MP | D_NEW | D_HOTPLUG,
349 CB_REV,
350 nodev,
351 nodev
352 };
353
354 static struct dev_ops viona_ops = {
355 DEVO_REV,
356 0,
357 viona_info,
358 nulldev,
359 nulldev,
360 viona_attach,
361 viona_detach,
362 nodev,
363 &viona_cb_ops,
364 NULL,
365 ddi_power,
366 ddi_quiesce_not_needed
367 };
368
369 static struct modldrv modldrv = {
370 &mod_driverops,
371 VIONA_NAME,
372 &viona_ops,
373 };
374
375 static struct modlinkage modlinkage = {
376 MODREV_1, &modldrv, NULL
377 };
378
379 int
_init(void)380 _init(void)
381 {
382 int ret;
383
384 ret = ddi_soft_state_init(&viona_state, sizeof (viona_soft_state_t), 0);
385 if (ret != 0) {
386 return (ret);
387 }
388
389 viona_minors = id_space_create("viona_minors",
390 VIONA_CTL_MINOR + 1, UINT16_MAX);
391 viona_rx_init();
392 mutex_init(&viona_force_copy_lock, NULL, MUTEX_DRIVER, NULL);
393
394 ret = mod_install(&modlinkage);
395 if (ret != 0) {
396 ddi_soft_state_fini(&viona_state);
397 id_space_destroy(viona_minors);
398 viona_rx_fini();
399 mutex_destroy(&viona_force_copy_lock);
400 }
401
402 return (ret);
403 }
404
405 int
_fini(void)406 _fini(void)
407 {
408 int ret;
409
410 ret = mod_remove(&modlinkage);
411 if (ret != 0) {
412 return (ret);
413 }
414
415 ddi_soft_state_fini(&viona_state);
416 id_space_destroy(viona_minors);
417 viona_rx_fini();
418 mutex_destroy(&viona_force_copy_lock);
419
420 return (ret);
421 }
422
423 int
_info(struct modinfo * modinfop)424 _info(struct modinfo *modinfop)
425 {
426 return (mod_info(&modlinkage, modinfop));
427 }
428
429 /* ARGSUSED */
430 static int
viona_info(dev_info_t * dip,ddi_info_cmd_t cmd,void * arg,void ** result)431 viona_info(dev_info_t *dip, ddi_info_cmd_t cmd, void *arg, void **result)
432 {
433 int error;
434
435 switch (cmd) {
436 case DDI_INFO_DEVT2DEVINFO:
437 *result = (void *)viona_dip;
438 error = DDI_SUCCESS;
439 break;
440 case DDI_INFO_DEVT2INSTANCE:
441 *result = (void *)0;
442 error = DDI_SUCCESS;
443 break;
444 default:
445 error = DDI_FAILURE;
446 break;
447 }
448 return (error);
449 }
450
451 static int
viona_attach(dev_info_t * dip,ddi_attach_cmd_t cmd)452 viona_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
453 {
454 if (cmd != DDI_ATTACH) {
455 return (DDI_FAILURE);
456 }
457
458 if (ddi_create_minor_node(dip, "viona", S_IFCHR, VIONA_CTL_MINOR,
459 DDI_PSEUDO, 0) != DDI_SUCCESS) {
460 return (DDI_FAILURE);
461 }
462
463 viona_neti_attach();
464
465 viona_dip = dip;
466 ddi_report_dev(viona_dip);
467
468 return (DDI_SUCCESS);
469 }
470
471 static int
viona_detach(dev_info_t * dip,ddi_detach_cmd_t cmd)472 viona_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
473 {
474 dev_info_t *old_dip = viona_dip;
475
476 if (cmd != DDI_DETACH) {
477 return (DDI_FAILURE);
478 }
479
480 VERIFY(old_dip != NULL);
481
482 viona_neti_detach();
483 viona_dip = NULL;
484 ddi_remove_minor_node(old_dip, NULL);
485
486 return (DDI_SUCCESS);
487 }
488
489 static int
viona_open(dev_t * devp,int flag,int otype,cred_t * credp)490 viona_open(dev_t *devp, int flag, int otype, cred_t *credp)
491 {
492 int minor;
493 viona_soft_state_t *ss;
494
495 if (otype != OTYP_CHR) {
496 return (EINVAL);
497 }
498 #if 0
499 /*
500 * XXX-mg: drv_priv() is wrong, but I'm not sure what is right.
501 * Should the check be at open() or ioctl()?
502 */
503 if (drv_priv(credp) != 0) {
504 return (EPERM);
505 }
506 #endif
507 if (getminor(*devp) != VIONA_CTL_MINOR) {
508 return (ENXIO);
509 }
510
511 minor = id_alloc_nosleep(viona_minors);
512 if (minor == -1) {
513 /* All minors are busy */
514 return (EBUSY);
515 }
516 if (ddi_soft_state_zalloc(viona_state, minor) != DDI_SUCCESS) {
517 id_free(viona_minors, minor);
518 return (ENOMEM);
519 }
520
521 ss = ddi_get_soft_state(viona_state, minor);
522 mutex_init(&ss->ss_lock, NULL, MUTEX_DEFAULT, NULL);
523 ss->ss_minor = minor;
524 *devp = makedevice(getmajor(*devp), minor);
525
526 return (0);
527 }
528
529 static int
viona_close(dev_t dev,int flag,int otype,cred_t * credp)530 viona_close(dev_t dev, int flag, int otype, cred_t *credp)
531 {
532 int minor;
533 viona_soft_state_t *ss;
534
535 if (otype != OTYP_CHR) {
536 return (EINVAL);
537 }
538
539 minor = getminor(dev);
540
541 ss = ddi_get_soft_state(viona_state, minor);
542 if (ss == NULL) {
543 return (ENXIO);
544 }
545
546 VERIFY0(viona_ioc_delete(ss, B_TRUE));
547 VERIFY(!list_link_active(&ss->ss_node));
548 ddi_soft_state_free(viona_state, minor);
549 id_free(viona_minors, minor);
550
551 return (0);
552 }
553
554 static int
viona_ioctl(dev_t dev,int cmd,intptr_t data,int md,cred_t * cr,int * rv)555 viona_ioctl(dev_t dev, int cmd, intptr_t data, int md, cred_t *cr, int *rv)
556 {
557 viona_soft_state_t *ss;
558 void *dptr = (void *)data;
559 int err = 0, val;
560 viona_link_t *link;
561
562 ss = ddi_get_soft_state(viona_state, getminor(dev));
563 if (ss == NULL) {
564 return (ENXIO);
565 }
566
567 switch (cmd) {
568 case VNA_IOC_CREATE:
569 return (viona_ioc_create(ss, dptr, md, cr));
570 case VNA_IOC_DELETE:
571 return (viona_ioc_delete(ss, B_FALSE));
572 case VNA_IOC_VERSION:
573 *rv = VIONA_CURRENT_INTERFACE_VERSION;
574 return (0);
575 default:
576 break;
577 }
578
579 mutex_enter(&ss->ss_lock);
580 if ((link = ss->ss_link) == NULL || link->l_destroyed ||
581 vmm_drv_release_reqd(link->l_vm_hold)) {
582 mutex_exit(&ss->ss_lock);
583 return (ENXIO);
584 }
585
586 switch (cmd) {
587 case VNA_IOC_GET_FEATURES:
588 val = VIONA_S_HOSTCAPS | link->l_features_hw;
589 if (ddi_copyout(&val, dptr, sizeof (val), md) != 0) {
590 err = EFAULT;
591 }
592 break;
593 case VNA_IOC_SET_FEATURES:
594 if (ddi_copyin(dptr, &val, sizeof (val), md) != 0) {
595 err = EFAULT;
596 break;
597 }
598 val &= (VIONA_S_HOSTCAPS | link->l_features_hw);
599
600 if ((val & VIRTIO_NET_F_CSUM) == 0)
601 val &= ~VIRTIO_NET_F_HOST_TSO4;
602
603 if ((val & VIRTIO_NET_F_GUEST_CSUM) == 0)
604 val &= ~VIRTIO_NET_F_GUEST_TSO4;
605
606 link->l_features = val;
607 break;
608 case VNA_IOC_RING_INIT:
609 err = viona_ioc_ring_init(link, dptr, md);
610 break;
611 case VNA_IOC_RING_RESET:
612 err = viona_ioc_ring_reset(link, (uint_t)data);
613 break;
614 case VNA_IOC_RING_KICK:
615 err = viona_ioc_ring_kick(link, (uint_t)data);
616 break;
617 case VNA_IOC_RING_SET_MSI:
618 err = viona_ioc_ring_set_msi(link, dptr, md);
619 break;
620 case VNA_IOC_RING_INTR_CLR:
621 err = viona_ioc_ring_intr_clear(link, (uint_t)data);
622 break;
623 case VNA_IOC_RING_SET_STATE:
624 err = viona_ioc_ring_set_state(link, dptr, md);
625 break;
626 case VNA_IOC_RING_GET_STATE:
627 err = viona_ioc_ring_get_state(link, dptr, md);
628 break;
629 case VNA_IOC_RING_PAUSE:
630 err = viona_ioc_ring_pause(link, (uint_t)data);
631 break;
632
633 case VNA_IOC_INTR_POLL:
634 err = viona_ioc_intr_poll(link, dptr, md, rv);
635 break;
636 case VNA_IOC_SET_NOTIFY_IOP:
637 if (data < 0 || data > UINT16_MAX) {
638 err = EINVAL;
639 break;
640 }
641 err = viona_ioc_set_notify_ioport(link, (uint16_t)data);
642 break;
643 case VNA_IOC_SET_PROMISC:
644 err = viona_ioc_set_promisc(link, (viona_promisc_t)data);
645 break;
646 default:
647 err = ENOTTY;
648 break;
649 }
650
651 mutex_exit(&ss->ss_lock);
652 return (err);
653 }
654
655 static int
viona_chpoll(dev_t dev,short events,int anyyet,short * reventsp,struct pollhead ** phpp)656 viona_chpoll(dev_t dev, short events, int anyyet, short *reventsp,
657 struct pollhead **phpp)
658 {
659 viona_soft_state_t *ss;
660 viona_link_t *link;
661
662 ss = ddi_get_soft_state(viona_state, getminor(dev));
663 if (ss == NULL) {
664 return (ENXIO);
665 }
666
667 mutex_enter(&ss->ss_lock);
668 if ((link = ss->ss_link) == NULL || link->l_destroyed) {
669 mutex_exit(&ss->ss_lock);
670 return (ENXIO);
671 }
672
673 *reventsp = 0;
674 if ((events & POLLRDBAND) != 0) {
675 for (uint_t i = 0; i < VIONA_VQ_MAX; i++) {
676 if (link->l_vrings[i].vr_intr_enabled != 0) {
677 *reventsp |= POLLRDBAND;
678 break;
679 }
680 }
681 }
682 if ((*reventsp == 0 && !anyyet) || (events & POLLET)) {
683 *phpp = &link->l_pollhead;
684 }
685 mutex_exit(&ss->ss_lock);
686
687 return (0);
688 }
689
690 static void
viona_get_mac_capab(viona_link_t * link)691 viona_get_mac_capab(viona_link_t *link)
692 {
693 mac_handle_t mh = link->l_mh;
694 uint32_t cap = 0;
695 mac_capab_lso_t lso_cap;
696
697 link->l_features_hw = 0;
698 if (mac_capab_get(mh, MAC_CAPAB_HCKSUM, &cap)) {
699 /*
700 * Only report HW checksum ability if the underlying MAC
701 * resource is capable of populating the L4 header.
702 */
703 if ((cap & VIONA_CAP_HCKSUM_INTEREST) != 0) {
704 link->l_features_hw |= VIRTIO_NET_F_CSUM;
705 }
706 link->l_cap_csum = cap;
707 }
708
709 if ((link->l_features_hw & VIRTIO_NET_F_CSUM) &&
710 mac_capab_get(mh, MAC_CAPAB_LSO, &lso_cap)) {
711 /*
712 * Virtio doesn't allow for negotiating a maximum LSO
713 * packet size. We have to assume that the guest may
714 * send a maximum length IP packet. Make sure the
715 * underlying MAC can handle an LSO of this size.
716 */
717 if ((lso_cap.lso_flags & LSO_TX_BASIC_TCP_IPV4) &&
718 lso_cap.lso_basic_tcp_ipv4.lso_max >= IP_MAXPACKET)
719 link->l_features_hw |= VIRTIO_NET_F_HOST_TSO4;
720 }
721 }
722
723 static int
viona_kstat_update(kstat_t * ksp,int rw)724 viona_kstat_update(kstat_t *ksp, int rw)
725 {
726 viona_link_t *link = ksp->ks_private;
727 viona_kstats_t *vk = ksp->ks_data;
728
729 /*
730 * Avoid the potential for mangled values due to a racing consolidation
731 * of stats for a ring by performing the kstat update with l_stats_lock
732 * held while adding up the central (link) and ring values.
733 */
734 mutex_enter(&link->l_stats_lock);
735
736 const viona_transfer_stats_t *ring_stats =
737 &link->l_vrings[VIONA_VQ_RX].vr_stats;
738 const viona_transfer_stats_t *link_stats = &link->l_stats.vls_rx;
739
740 vk->vk_rx_packets.value.ui64 =
741 link_stats->vts_packets + ring_stats->vts_packets;
742 vk->vk_rx_bytes.value.ui64 =
743 link_stats->vts_bytes + ring_stats->vts_bytes;
744 vk->vk_rx_errors.value.ui64 =
745 link_stats->vts_errors + ring_stats->vts_errors;
746 vk->vk_rx_drops.value.ui64 =
747 link_stats->vts_drops + ring_stats->vts_drops;
748
749 ring_stats = &link->l_vrings[VIONA_VQ_TX].vr_stats;
750 link_stats = &link->l_stats.vls_tx;
751
752 vk->vk_tx_packets.value.ui64 =
753 link_stats->vts_packets + ring_stats->vts_packets;
754 vk->vk_tx_bytes.value.ui64 =
755 link_stats->vts_bytes + ring_stats->vts_bytes;
756 vk->vk_tx_errors.value.ui64 =
757 link_stats->vts_errors + ring_stats->vts_errors;
758 vk->vk_tx_drops.value.ui64 =
759 link_stats->vts_drops + ring_stats->vts_drops;
760
761 mutex_exit(&link->l_stats_lock);
762
763 return (0);
764 }
765
766 static int
viona_kstat_init(viona_soft_state_t * ss,const cred_t * cr)767 viona_kstat_init(viona_soft_state_t *ss, const cred_t *cr)
768 {
769 zoneid_t zid = crgetzoneid(cr);
770 kstat_t *ksp;
771
772 ASSERT(MUTEX_HELD(&ss->ss_lock));
773 ASSERT3P(ss->ss_kstat, ==, NULL);
774
775 ksp = kstat_create_zone(VIONA_MODULE_NAME, ss->ss_minor,
776 VIONA_KSTAT_NAME, VIONA_KSTAT_CLASS, KSTAT_TYPE_NAMED,
777 sizeof (viona_kstats_t) / sizeof (kstat_named_t), 0, zid);
778
779 if (ksp == NULL) {
780 /*
781 * Without detail from kstat_create_zone(), assume that resource
782 * exhaustion is to blame for the failure.
783 */
784 return (ENOMEM);
785 }
786 ss->ss_kstat = ksp;
787
788 /*
789 * If this instance is associated with a non-global zone, make its
790 * kstats visible from the GZ.
791 */
792 if (zid != GLOBAL_ZONEID) {
793 kstat_zone_add(ss->ss_kstat, GLOBAL_ZONEID);
794 }
795
796 viona_kstats_t *vk = ksp->ks_data;
797
798 kstat_named_init(&vk->vk_rx_packets, "rx_packets", KSTAT_DATA_UINT64);
799 kstat_named_init(&vk->vk_rx_bytes, "rx_bytes", KSTAT_DATA_UINT64);
800 kstat_named_init(&vk->vk_rx_errors, "rx_errors", KSTAT_DATA_UINT64);
801 kstat_named_init(&vk->vk_rx_drops, "rx_drops", KSTAT_DATA_UINT64);
802 kstat_named_init(&vk->vk_tx_packets, "tx_packets", KSTAT_DATA_UINT64);
803 kstat_named_init(&vk->vk_tx_bytes, "tx_bytes", KSTAT_DATA_UINT64);
804 kstat_named_init(&vk->vk_tx_errors, "tx_errors", KSTAT_DATA_UINT64);
805 kstat_named_init(&vk->vk_tx_drops, "tx_drops", KSTAT_DATA_UINT64);
806 ksp->ks_private = ss->ss_link;
807 ksp->ks_update = viona_kstat_update;
808
809 kstat_install(ss->ss_kstat);
810 return (0);
811 }
812
813 static void
viona_kstat_fini(viona_soft_state_t * ss)814 viona_kstat_fini(viona_soft_state_t *ss)
815 {
816 ASSERT(MUTEX_HELD(&ss->ss_lock));
817
818 if (ss->ss_kstat != NULL) {
819 kstat_delete(ss->ss_kstat);
820 ss->ss_kstat = NULL;
821 }
822 }
823
824 static int
viona_ioc_create(viona_soft_state_t * ss,void * dptr,int md,cred_t * cr)825 viona_ioc_create(viona_soft_state_t *ss, void *dptr, int md, cred_t *cr)
826 {
827 vioc_create_t kvc;
828 viona_link_t *link = NULL;
829 char cli_name[MAXNAMELEN];
830 int err = 0;
831 file_t *fp;
832 vmm_hold_t *hold = NULL;
833 viona_neti_t *nip = NULL;
834 zoneid_t zid;
835 mac_diag_t mac_diag = MAC_DIAG_NONE;
836 boolean_t rings_allocd = B_FALSE;
837
838 ASSERT(MUTEX_NOT_HELD(&ss->ss_lock));
839
840 if (ddi_copyin(dptr, &kvc, sizeof (kvc), md) != 0) {
841 return (EFAULT);
842 }
843
844 zid = crgetzoneid(cr);
845 nip = viona_neti_lookup_by_zid(zid);
846 if (nip == NULL) {
847 return (EIO);
848 }
849
850 if (!nip->vni_nethook.vnh_hooked) {
851 viona_neti_rele(nip);
852 return (EIO);
853 }
854
855 mutex_enter(&ss->ss_lock);
856 if (ss->ss_link != NULL) {
857 mutex_exit(&ss->ss_lock);
858 viona_neti_rele(nip);
859 return (EEXIST);
860 }
861
862 if ((fp = getf(kvc.c_vmfd)) == NULL) {
863 err = EBADF;
864 goto bail;
865 }
866 err = vmm_drv_hold(fp, cr, &hold);
867 releasef(kvc.c_vmfd);
868 if (err != 0) {
869 goto bail;
870 }
871
872 link = kmem_zalloc(sizeof (viona_link_t), KM_SLEEP);
873 link->l_linkid = kvc.c_linkid;
874 link->l_vm_hold = hold;
875
876 err = mac_open_by_linkid(link->l_linkid, &link->l_mh);
877 if (err != 0) {
878 goto bail;
879 }
880
881 viona_get_mac_capab(link);
882
883 (void) snprintf(cli_name, sizeof (cli_name), "%s-%d", VIONA_MODULE_NAME,
884 link->l_linkid);
885 err = mac_client_open(link->l_mh, &link->l_mch, cli_name, 0);
886 if (err != 0) {
887 goto bail;
888 }
889
890 err = mac_unicast_add(link->l_mch, NULL, MAC_UNICAST_PRIMARY,
891 &link->l_muh, VLAN_ID_NONE, &mac_diag);
892 if (err != 0) {
893 goto bail;
894 }
895
896 viona_ring_alloc(link, &link->l_vrings[VIONA_VQ_RX]);
897 viona_ring_alloc(link, &link->l_vrings[VIONA_VQ_TX]);
898 rings_allocd = B_TRUE;
899
900 /*
901 * Default to passing up all multicast traffic in addition to
902 * classified unicast. Guests which have support will change this
903 * if they need to via the virtio net control queue; guests without
904 * support generally still want to see multicast.
905 */
906 link->l_promisc = VIONA_PROMISC_MULTI;
907 if ((err = viona_rx_set(link, link->l_promisc)) != 0) {
908 goto bail;
909 }
910
911 link->l_neti = nip;
912 ss->ss_link = link;
913
914 if ((err = viona_kstat_init(ss, cr)) != 0) {
915 goto bail;
916 }
917
918 mutex_exit(&ss->ss_lock);
919
920 mutex_enter(&nip->vni_lock);
921 list_insert_tail(&nip->vni_dev_list, ss);
922 mutex_exit(&nip->vni_lock);
923
924 return (0);
925
926 bail:
927 if (link != NULL) {
928 viona_rx_clear(link);
929 if (link->l_mch != NULL) {
930 if (link->l_muh != NULL) {
931 VERIFY0(mac_unicast_remove(link->l_mch,
932 link->l_muh));
933 link->l_muh = NULL;
934 }
935 mac_client_close(link->l_mch, 0);
936 }
937 if (link->l_mh != NULL) {
938 mac_close(link->l_mh);
939 }
940 if (rings_allocd) {
941 viona_ring_free(&link->l_vrings[VIONA_VQ_RX]);
942 viona_ring_free(&link->l_vrings[VIONA_VQ_TX]);
943 }
944 kmem_free(link, sizeof (viona_link_t));
945 ss->ss_link = NULL;
946 }
947 if (hold != NULL) {
948 vmm_drv_rele(hold);
949 }
950 viona_neti_rele(nip);
951
952 mutex_exit(&ss->ss_lock);
953 return (err);
954 }
955
956 static int
viona_ioc_delete(viona_soft_state_t * ss,boolean_t on_close)957 viona_ioc_delete(viona_soft_state_t *ss, boolean_t on_close)
958 {
959 viona_link_t *link;
960 viona_neti_t *nip = NULL;
961
962 mutex_enter(&ss->ss_lock);
963 if ((link = ss->ss_link) == NULL) {
964 /* Link destruction already complete */
965 mutex_exit(&ss->ss_lock);
966 return (0);
967 }
968
969 if (link->l_destroyed) {
970 /*
971 * Link destruction has been started by another thread, but has
972 * not completed. This condition should be impossible to
973 * encounter when performing the on-close destroy of the link,
974 * since racing ioctl accessors must necessarily be absent.
975 */
976 VERIFY(!on_close);
977 mutex_exit(&ss->ss_lock);
978 return (EAGAIN);
979 }
980 /*
981 * The link deletion cannot fail after this point, continuing until its
982 * successful completion is reached.
983 */
984 link->l_destroyed = B_TRUE;
985
986 /*
987 * Tear down the IO port hook so it cannot be used to kick any of the
988 * rings which are about to be reset and stopped.
989 */
990 VERIFY0(viona_ioc_set_notify_ioport(link, 0));
991 mutex_exit(&ss->ss_lock);
992
993 /*
994 * Return the rings to their reset state, ignoring any possible
995 * interruptions from signals.
996 */
997 VERIFY0(viona_ring_reset(&link->l_vrings[VIONA_VQ_RX], B_FALSE));
998 VERIFY0(viona_ring_reset(&link->l_vrings[VIONA_VQ_TX], B_FALSE));
999
1000 mutex_enter(&ss->ss_lock);
1001 viona_kstat_fini(ss);
1002 if (link->l_mch != NULL) {
1003 /* Unhook the receive callbacks and close out the client */
1004 viona_rx_clear(link);
1005 if (link->l_muh != NULL) {
1006 VERIFY0(mac_unicast_remove(link->l_mch, link->l_muh));
1007 link->l_muh = NULL;
1008 }
1009 mac_client_close(link->l_mch, 0);
1010 }
1011 if (link->l_mh != NULL) {
1012 mac_close(link->l_mh);
1013 }
1014 if (link->l_vm_hold != NULL) {
1015 vmm_drv_rele(link->l_vm_hold);
1016 link->l_vm_hold = NULL;
1017 }
1018
1019 nip = link->l_neti;
1020 link->l_neti = NULL;
1021
1022 viona_ring_free(&link->l_vrings[VIONA_VQ_RX]);
1023 viona_ring_free(&link->l_vrings[VIONA_VQ_TX]);
1024 pollhead_clean(&link->l_pollhead);
1025 ss->ss_link = NULL;
1026 mutex_exit(&ss->ss_lock);
1027
1028 mutex_enter(&nip->vni_lock);
1029 list_remove(&nip->vni_dev_list, ss);
1030 mutex_exit(&nip->vni_lock);
1031
1032 viona_neti_rele(nip);
1033
1034 kmem_free(link, sizeof (viona_link_t));
1035 return (0);
1036 }
1037
1038 static int
viona_ioc_ring_init(viona_link_t * link,void * udata,int md)1039 viona_ioc_ring_init(viona_link_t *link, void *udata, int md)
1040 {
1041 vioc_ring_init_t kri;
1042 int err;
1043
1044 if (ddi_copyin(udata, &kri, sizeof (kri), md) != 0) {
1045 return (EFAULT);
1046 }
1047 const struct viona_ring_params params = {
1048 .vrp_pa = kri.ri_qaddr,
1049 .vrp_size = kri.ri_qsize,
1050 .vrp_avail_idx = 0,
1051 .vrp_used_idx = 0,
1052 };
1053
1054 err = viona_ring_init(link, kri.ri_index, ¶ms);
1055
1056 return (err);
1057 }
1058
1059 static int
viona_ioc_ring_set_state(viona_link_t * link,void * udata,int md)1060 viona_ioc_ring_set_state(viona_link_t *link, void *udata, int md)
1061 {
1062 vioc_ring_state_t krs;
1063 int err;
1064
1065 if (ddi_copyin(udata, &krs, sizeof (krs), md) != 0) {
1066 return (EFAULT);
1067 }
1068 const struct viona_ring_params params = {
1069 .vrp_pa = krs.vrs_qaddr,
1070 .vrp_size = krs.vrs_qsize,
1071 .vrp_avail_idx = krs.vrs_avail_idx,
1072 .vrp_used_idx = krs.vrs_used_idx,
1073 };
1074
1075 err = viona_ring_init(link, krs.vrs_index, ¶ms);
1076
1077 return (err);
1078 }
1079
1080 static int
viona_ioc_ring_get_state(viona_link_t * link,void * udata,int md)1081 viona_ioc_ring_get_state(viona_link_t *link, void *udata, int md)
1082 {
1083 vioc_ring_state_t krs;
1084
1085 if (ddi_copyin(udata, &krs, sizeof (krs), md) != 0) {
1086 return (EFAULT);
1087 }
1088
1089 struct viona_ring_params params;
1090 int err = viona_ring_get_state(link, krs.vrs_index, ¶ms);
1091 if (err != 0) {
1092 return (err);
1093 }
1094 krs.vrs_qsize = params.vrp_size;
1095 krs.vrs_qaddr = params.vrp_pa;
1096 krs.vrs_avail_idx = params.vrp_avail_idx;
1097 krs.vrs_used_idx = params.vrp_used_idx;
1098
1099 if (ddi_copyout(&krs, udata, sizeof (krs), md) != 0) {
1100 return (EFAULT);
1101 }
1102 return (0);
1103 }
1104
1105 static int
viona_ioc_ring_reset(viona_link_t * link,uint_t idx)1106 viona_ioc_ring_reset(viona_link_t *link, uint_t idx)
1107 {
1108 viona_vring_t *ring;
1109
1110 if (idx >= VIONA_VQ_MAX) {
1111 return (EINVAL);
1112 }
1113 ring = &link->l_vrings[idx];
1114
1115 return (viona_ring_reset(ring, B_TRUE));
1116 }
1117
1118 static int
viona_ioc_ring_kick(viona_link_t * link,uint_t idx)1119 viona_ioc_ring_kick(viona_link_t *link, uint_t idx)
1120 {
1121 viona_vring_t *ring;
1122 int err;
1123
1124 if (idx >= VIONA_VQ_MAX) {
1125 return (EINVAL);
1126 }
1127 ring = &link->l_vrings[idx];
1128
1129 mutex_enter(&ring->vr_lock);
1130 switch (ring->vr_state) {
1131 case VRS_SETUP:
1132 /*
1133 * An early kick to a ring which is starting its worker thread
1134 * is fine. Once that thread is active, it will process the
1135 * start-up request immediately.
1136 */
1137 /* FALLTHROUGH */
1138 case VRS_INIT:
1139 ring->vr_state_flags |= VRSF_REQ_START;
1140 /* FALLTHROUGH */
1141 case VRS_RUN:
1142 cv_broadcast(&ring->vr_cv);
1143 err = 0;
1144 break;
1145 default:
1146 err = EBUSY;
1147 break;
1148 }
1149 mutex_exit(&ring->vr_lock);
1150
1151 return (err);
1152 }
1153
1154 static int
viona_ioc_ring_pause(viona_link_t * link,uint_t idx)1155 viona_ioc_ring_pause(viona_link_t *link, uint_t idx)
1156 {
1157 if (idx >= VIONA_VQ_MAX) {
1158 return (EINVAL);
1159 }
1160
1161 viona_vring_t *ring = &link->l_vrings[idx];
1162 return (viona_ring_pause(ring));
1163 }
1164
1165 static int
viona_ioc_ring_set_msi(viona_link_t * link,void * data,int md)1166 viona_ioc_ring_set_msi(viona_link_t *link, void *data, int md)
1167 {
1168 vioc_ring_msi_t vrm;
1169 viona_vring_t *ring;
1170
1171 if (ddi_copyin(data, &vrm, sizeof (vrm), md) != 0) {
1172 return (EFAULT);
1173 }
1174 if (vrm.rm_index >= VIONA_VQ_MAX) {
1175 return (EINVAL);
1176 }
1177
1178 ring = &link->l_vrings[vrm.rm_index];
1179 mutex_enter(&ring->vr_lock);
1180 ring->vr_msi_addr = vrm.rm_addr;
1181 ring->vr_msi_msg = vrm.rm_msg;
1182 mutex_exit(&ring->vr_lock);
1183
1184 return (0);
1185 }
1186
1187 static int
viona_notify_iop(void * arg,bool in,uint16_t port,uint8_t bytes,uint32_t * val)1188 viona_notify_iop(void *arg, bool in, uint16_t port, uint8_t bytes,
1189 uint32_t *val)
1190 {
1191 viona_link_t *link = (viona_link_t *)arg;
1192
1193 /*
1194 * If the request is a read (in/ins), or direct at a port other than
1195 * what we expect to be registered on, ignore it.
1196 */
1197 if (in || port != link->l_notify_ioport) {
1198 return (ESRCH);
1199 }
1200
1201 /* Let userspace handle notifications for rings other than RX/TX. */
1202 const uint16_t vq = *val;
1203 if (vq >= VIONA_VQ_MAX) {
1204 return (ESRCH);
1205 }
1206
1207 viona_vring_t *ring = &link->l_vrings[vq];
1208 int res = 0;
1209
1210 mutex_enter(&ring->vr_lock);
1211 if (ring->vr_state == VRS_RUN) {
1212 cv_broadcast(&ring->vr_cv);
1213 } else {
1214 res = ESRCH;
1215 }
1216 mutex_exit(&ring->vr_lock);
1217
1218 return (res);
1219 }
1220
1221 static int
viona_ioc_set_notify_ioport(viona_link_t * link,uint16_t ioport)1222 viona_ioc_set_notify_ioport(viona_link_t *link, uint16_t ioport)
1223 {
1224 int err = 0;
1225
1226 if (link->l_notify_ioport != 0) {
1227 vmm_drv_ioport_unhook(link->l_vm_hold, &link->l_notify_cookie);
1228 link->l_notify_ioport = 0;
1229 }
1230
1231 if (ioport != 0) {
1232 err = vmm_drv_ioport_hook(link->l_vm_hold, ioport,
1233 viona_notify_iop, (void *)link, &link->l_notify_cookie);
1234 if (err == 0) {
1235 link->l_notify_ioport = ioport;
1236 }
1237 }
1238 return (err);
1239 }
1240
1241 static int
viona_ioc_set_promisc(viona_link_t * link,viona_promisc_t mode)1242 viona_ioc_set_promisc(viona_link_t *link, viona_promisc_t mode)
1243 {
1244 int err;
1245
1246 if (mode >= VIONA_PROMISC_MAX) {
1247 return (EINVAL);
1248 }
1249
1250 if (mode == link->l_promisc) {
1251 return (0);
1252 }
1253
1254 if ((err = viona_rx_set(link, mode)) != 0) {
1255 return (err);
1256 }
1257
1258 link->l_promisc = mode;
1259 return (0);
1260 }
1261
1262 static int
viona_ioc_ring_intr_clear(viona_link_t * link,uint_t idx)1263 viona_ioc_ring_intr_clear(viona_link_t *link, uint_t idx)
1264 {
1265 if (idx >= VIONA_VQ_MAX) {
1266 return (EINVAL);
1267 }
1268
1269 link->l_vrings[idx].vr_intr_enabled = 0;
1270 return (0);
1271 }
1272
1273 static int
viona_ioc_intr_poll(viona_link_t * link,void * udata,int md,int * rv)1274 viona_ioc_intr_poll(viona_link_t *link, void *udata, int md, int *rv)
1275 {
1276 uint_t cnt = 0;
1277 vioc_intr_poll_t vip;
1278
1279 for (uint_t i = 0; i < VIONA_VQ_MAX; i++) {
1280 uint_t val = link->l_vrings[i].vr_intr_enabled;
1281
1282 vip.vip_status[i] = val;
1283 if (val != 0) {
1284 cnt++;
1285 }
1286 }
1287
1288 if (ddi_copyout(&vip, udata, sizeof (vip), md) != 0) {
1289 return (EFAULT);
1290 }
1291 *rv = (int)cnt;
1292 return (0);
1293 }
1294