Home | History | Annotate | Line # | Download | only in pci
virtio.c revision 1.57
      1 /*	$NetBSD: virtio.c,v 1.57 2022/08/12 10:49:57 riastradh Exp $	*/
      2 
      3 /*
      4  * Copyright (c) 2020 The NetBSD Foundation, Inc.
      5  * Copyright (c) 2012 Stefan Fritsch, Alexander Fiveg.
      6  * Copyright (c) 2010 Minoura Makoto.
      7  * All rights reserved.
      8  *
      9  * Redistribution and use in source and binary forms, with or without
     10  * modification, are permitted provided that the following conditions
     11  * are met:
     12  * 1. Redistributions of source code must retain the above copyright
     13  *    notice, this list of conditions and the following disclaimer.
     14  * 2. Redistributions in binary form must reproduce the above copyright
     15  *    notice, this list of conditions and the following disclaimer in the
     16  *    documentation and/or other materials provided with the distribution.
     17  *
     18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
     19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
     20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
     21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
     22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
     23  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
     24  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
     25  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
     26  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
     27  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     28  */
     29 
     30 #include <sys/cdefs.h>
     31 __KERNEL_RCSID(0, "$NetBSD: virtio.c,v 1.57 2022/08/12 10:49:57 riastradh Exp $");
     32 
     33 #include <sys/param.h>
     34 #include <sys/systm.h>
     35 #include <sys/kernel.h>
     36 #include <sys/atomic.h>
     37 #include <sys/bus.h>
     38 #include <sys/device.h>
     39 #include <sys/kmem.h>
     40 #include <sys/module.h>
     41 
     42 #define VIRTIO_PRIVATE
     43 
     44 #include <dev/pci/virtioreg.h> /* XXX: move to non-pci */
     45 #include <dev/pci/virtiovar.h> /* XXX: move to non-pci */
     46 
     47 #define MINSEG_INDIRECT		2 /* use indirect if nsegs >= this value */
     48 
     49 /* incomplete list */
     50 static const char *virtio_device_name[] = {
     51 	"unknown (0)",			/*  0 */
     52 	"network",			/*  1 */
     53 	"block",			/*  2 */
     54 	"console",			/*  3 */
     55 	"entropy",			/*  4 */
     56 	"memory balloon",		/*  5 */
     57 	"I/O memory",			/*  6 */
     58 	"remote processor messaging",	/*  7 */
     59 	"SCSI",				/*  8 */
     60 	"9P transport",			/*  9 */
     61 };
     62 #define NDEVNAMES	__arraycount(virtio_device_name)
     63 
     64 static void	virtio_init_vq(struct virtio_softc *,
     65 		    struct virtqueue *, const bool);
     66 
     67 void
     68 virtio_set_status(struct virtio_softc *sc, int status)
     69 {
     70 	sc->sc_ops->set_status(sc, status);
     71 }
     72 
     73 /*
     74  * Reset the device.
     75  */
     76 /*
     77  * To reset the device to a known state, do following:
     78  *	virtio_reset(sc);	     // this will stop the device activity
     79  *	<dequeue finished requests>; // virtio_dequeue() still can be called
     80  *	<revoke pending requests in the vqs if any>;
     81  *	virtio_reinit_start(sc);     // dequeue prohibitted
     82  *	newfeatures = virtio_negotiate_features(sc, requestedfeatures);
     83  *	<some other initialization>;
     84  *	virtio_reinit_end(sc);	     // device activated; enqueue allowed
     85  * Once attached, feature negotiation can only be allowed after virtio_reset.
     86  */
     87 void
     88 virtio_reset(struct virtio_softc *sc)
     89 {
     90 	virtio_device_reset(sc);
     91 }
     92 
     93 int
     94 virtio_reinit_start(struct virtio_softc *sc)
     95 {
     96 	int i, r;
     97 
     98 	virtio_set_status(sc, VIRTIO_CONFIG_DEVICE_STATUS_ACK);
     99 	virtio_set_status(sc, VIRTIO_CONFIG_DEVICE_STATUS_DRIVER);
    100 	for (i = 0; i < sc->sc_nvqs; i++) {
    101 		int n;
    102 		struct virtqueue *vq = &sc->sc_vqs[i];
    103 		n = sc->sc_ops->read_queue_size(sc, vq->vq_index);
    104 		if (n == 0)	/* vq disappeared */
    105 			continue;
    106 		if (n != vq->vq_num) {
    107 			panic("%s: virtqueue size changed, vq index %d\n",
    108 			      device_xname(sc->sc_dev),
    109 			      vq->vq_index);
    110 		}
    111 		virtio_init_vq(sc, vq, true);
    112 		sc->sc_ops->setup_queue(sc, vq->vq_index,
    113 		    vq->vq_dmamap->dm_segs[0].ds_addr);
    114 	}
    115 
    116 	r = sc->sc_ops->setup_interrupts(sc, 1);
    117 	if (r != 0)
    118 		goto fail;
    119 
    120 	return 0;
    121 
    122 fail:
    123 	virtio_set_status(sc, VIRTIO_CONFIG_DEVICE_STATUS_FAILED);
    124 
    125 	return 1;
    126 }
    127 
    128 void
    129 virtio_reinit_end(struct virtio_softc *sc)
    130 {
    131 	virtio_set_status(sc, VIRTIO_CONFIG_DEVICE_STATUS_DRIVER_OK);
    132 }
    133 
    134 /*
    135  * Feature negotiation.
    136  */
    137 void
    138 virtio_negotiate_features(struct virtio_softc *sc, uint64_t guest_features)
    139 {
    140 	if (!(device_cfdata(sc->sc_dev)->cf_flags & 1) &&
    141 	    !(device_cfdata(sc->sc_child)->cf_flags & 1)) /* XXX */
    142 		guest_features |= VIRTIO_F_RING_INDIRECT_DESC;
    143 	sc->sc_ops->neg_features(sc, guest_features);
    144 	if (sc->sc_active_features & VIRTIO_F_RING_INDIRECT_DESC)
    145 		sc->sc_indirect = true;
    146 	else
    147 		sc->sc_indirect = false;
    148 }
    149 
    150 
    151 /*
    152  * Device configuration registers readers/writers
    153  */
    154 #if 0
    155 #define DPRINTFR(n, fmt, val, index, num) \
    156 	printf("\n%s (", n); \
    157 	for (int i = 0; i < num; i++) \
    158 		printf("%02x ", bus_space_read_1(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, index+i)); \
    159 	printf(") -> "); printf(fmt, val); printf("\n");
    160 #define DPRINTFR2(n, fmt, val_s, val_n) \
    161 	printf("%s ", n); \
    162 	printf("\n        stream "); printf(fmt, val_s); printf(" norm "); printf(fmt, val_n); printf("\n");
    163 #else
    164 #define DPRINTFR(n, fmt, val, index, num)
    165 #define DPRINTFR2(n, fmt, val_s, val_n)
    166 #endif
    167 
    168 
    169 uint8_t
    170 virtio_read_device_config_1(struct virtio_softc *sc, int index) {
    171 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    172 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    173 	uint8_t val;
    174 
    175 	val = bus_space_read_1(iot, ioh, index);
    176 
    177 	DPRINTFR("read_1", "%02x", val, index, 1);
    178 	return val;
    179 }
    180 
    181 uint16_t
    182 virtio_read_device_config_2(struct virtio_softc *sc, int index) {
    183 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    184 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    185 	uint16_t val;
    186 
    187 	val = bus_space_read_2(iot, ioh, index);
    188 	if (BYTE_ORDER != sc->sc_bus_endian)
    189 		val = bswap16(val);
    190 
    191 	DPRINTFR("read_2", "%04x", val, index, 2);
    192 	DPRINTFR2("read_2", "%04x",
    193 		bus_space_read_stream_2(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, index),
    194 		bus_space_read_2(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, index));
    195 	return val;
    196 }
    197 
    198 uint32_t
    199 virtio_read_device_config_4(struct virtio_softc *sc, int index) {
    200 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    201 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    202 	uint32_t val;
    203 
    204 	val = bus_space_read_4(iot, ioh, index);
    205 	if (BYTE_ORDER != sc->sc_bus_endian)
    206 		val = bswap32(val);
    207 
    208 	DPRINTFR("read_4", "%08x", val, index, 4);
    209 	DPRINTFR2("read_4", "%08x",
    210 		bus_space_read_stream_4(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, index),
    211 		bus_space_read_4(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, index));
    212 	return val;
    213 }
    214 
    215 /*
    216  * The Virtio spec explicitly tells that reading and writing 8 bytes are not
    217  * considered atomic and no triggers may be connected to reading or writing
    218  * it. We access it using two 32 reads. See virtio spec 4.1.3.1.
    219  */
    220 uint64_t
    221 virtio_read_device_config_8(struct virtio_softc *sc, int index) {
    222 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    223 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    224 	union {
    225 		uint64_t u64;
    226 		uint32_t l[2];
    227 	} v;
    228 	uint64_t val;
    229 
    230 	v.l[0] = bus_space_read_4(iot, ioh, index);
    231 	v.l[1] = bus_space_read_4(iot, ioh, index + 4);
    232 	if (sc->sc_bus_endian != sc->sc_struct_endian) {
    233 		v.l[0] = bswap32(v.l[0]);
    234 		v.l[1] = bswap32(v.l[1]);
    235 	}
    236 	val = v.u64;
    237 
    238 	if (BYTE_ORDER != sc->sc_struct_endian)
    239 		val = bswap64(val);
    240 
    241 	DPRINTFR("read_8", "%08lx", val, index, 8);
    242 	DPRINTFR2("read_8 low ", "%08x",
    243 		bus_space_read_stream_4(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, index),
    244 		bus_space_read_4(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, index));
    245 	DPRINTFR2("read_8 high ", "%08x",
    246 		bus_space_read_stream_4(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, index + 4),
    247 		bus_space_read_4(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, index + 4));
    248 	return val;
    249 }
    250 
    251 /*
    252  * In the older virtio spec, device config registers are host endian. On newer
    253  * they are little endian. Some newer devices however explicitly specify their
    254  * register to always be little endian. These functions cater for these.
    255  */
    256 uint16_t
    257 virtio_read_device_config_le_2(struct virtio_softc *sc, int index) {
    258 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    259 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    260 	uint16_t val;
    261 
    262 	val = bus_space_read_2(iot, ioh, index);
    263 	if (sc->sc_bus_endian != LITTLE_ENDIAN)
    264 		val = bswap16(val);
    265 
    266 	DPRINTFR("read_le_2", "%04x", val, index, 2);
    267 	DPRINTFR2("read_le_2", "%04x",
    268 		bus_space_read_stream_2(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, 0),
    269 		bus_space_read_2(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, 0));
    270 	return val;
    271 }
    272 
    273 uint32_t
    274 virtio_read_device_config_le_4(struct virtio_softc *sc, int index) {
    275 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    276 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    277 	uint32_t val;
    278 
    279 	val = bus_space_read_4(iot, ioh, index);
    280 	if (sc->sc_bus_endian != LITTLE_ENDIAN)
    281 		val = bswap32(val);
    282 
    283 	DPRINTFR("read_le_4", "%08x", val, index, 4);
    284 	DPRINTFR2("read_le_4", "%08x",
    285 		bus_space_read_stream_4(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, 0),
    286 		bus_space_read_4(sc->sc_devcfg_iot, sc->sc_devcfg_ioh, 0));
    287 	return val;
    288 }
    289 
    290 void
    291 virtio_write_device_config_1(struct virtio_softc *sc, int index, uint8_t value)
    292 {
    293 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    294 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    295 
    296 	bus_space_write_1(iot, ioh, index, value);
    297 }
    298 
    299 void
    300 virtio_write_device_config_2(struct virtio_softc *sc, int index, uint16_t value)
    301 {
    302 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    303 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    304 
    305 	if (BYTE_ORDER != sc->sc_bus_endian)
    306 		value = bswap16(value);
    307 	bus_space_write_2(iot, ioh, index, value);
    308 }
    309 
    310 void
    311 virtio_write_device_config_4(struct virtio_softc *sc, int index, uint32_t value)
    312 {
    313 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    314 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    315 
    316 	if (BYTE_ORDER != sc->sc_bus_endian)
    317 		value = bswap32(value);
    318 	bus_space_write_4(iot, ioh, index, value);
    319 }
    320 
    321 /*
    322  * The Virtio spec explicitly tells that reading and writing 8 bytes are not
    323  * considered atomic and no triggers may be connected to reading or writing
    324  * it. We access it using two 32 bit writes. For good measure it is stated to
    325  * always write lsb first just in case of a hypervisor bug. See See virtio
    326  * spec 4.1.3.1.
    327  */
    328 void
    329 virtio_write_device_config_8(struct virtio_softc *sc, int index, uint64_t value)
    330 {
    331 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    332 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    333 	union {
    334 		uint64_t u64;
    335 		uint32_t l[2];
    336 	} v;
    337 
    338 	if (BYTE_ORDER != sc->sc_struct_endian)
    339 		value = bswap64(value);
    340 
    341 	v.u64 = value;
    342 	if (sc->sc_bus_endian != sc->sc_struct_endian) {
    343 		v.l[0] = bswap32(v.l[0]);
    344 		v.l[1] = bswap32(v.l[1]);
    345 	}
    346 
    347 	if (sc->sc_struct_endian == LITTLE_ENDIAN) {
    348 		bus_space_write_4(iot, ioh, index,     v.l[0]);
    349 		bus_space_write_4(iot, ioh, index + 4, v.l[1]);
    350 	} else {
    351 		bus_space_write_4(iot, ioh, index + 4, v.l[1]);
    352 		bus_space_write_4(iot, ioh, index,     v.l[0]);
    353 	}
    354 }
    355 
    356 /*
    357  * In the older virtio spec, device config registers are host endian. On newer
    358  * they are little endian. Some newer devices however explicitly specify their
    359  * register to always be little endian. These functions cater for these.
    360  */
    361 void
    362 virtio_write_device_config_le_2(struct virtio_softc *sc, int index, uint16_t value)
    363 {
    364 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    365 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    366 
    367 	if (sc->sc_bus_endian != LITTLE_ENDIAN)
    368 		value = bswap16(value);
    369 	bus_space_write_2(iot, ioh, index, value);
    370 }
    371 
    372 void
    373 virtio_write_device_config_le_4(struct virtio_softc *sc, int index, uint32_t value)
    374 {
    375 	bus_space_tag_t	   iot = sc->sc_devcfg_iot;
    376 	bus_space_handle_t ioh = sc->sc_devcfg_ioh;
    377 
    378 	if (sc->sc_bus_endian != LITTLE_ENDIAN)
    379 		value = bswap32(value);
    380 	bus_space_write_4(iot, ioh, index, value);
    381 }
    382 
    383 
    384 /*
    385  * data structures endian helpers
    386  */
    387 uint16_t virtio_rw16(struct virtio_softc *sc, uint16_t val)
    388 {
    389 	KASSERT(sc);
    390 	return BYTE_ORDER != sc->sc_struct_endian ? bswap16(val) : val;
    391 }
    392 
    393 uint32_t virtio_rw32(struct virtio_softc *sc, uint32_t val)
    394 {
    395 	KASSERT(sc);
    396 	return BYTE_ORDER != sc->sc_struct_endian ? bswap32(val) : val;
    397 }
    398 
    399 uint64_t virtio_rw64(struct virtio_softc *sc, uint64_t val)
    400 {
    401 	KASSERT(sc);
    402 	return BYTE_ORDER != sc->sc_struct_endian ? bswap64(val) : val;
    403 }
    404 
    405 
    406 /*
    407  * Interrupt handler.
    408  */
    409 static void
    410 virtio_soft_intr(void *arg)
    411 {
    412 	struct virtio_softc *sc = arg;
    413 
    414 	KASSERT(sc->sc_intrhand != NULL);
    415 
    416 	(*sc->sc_intrhand)(sc);
    417 }
    418 
    419 /*
    420  * dmamap sync operations for a virtqueue.
    421  */
    422 static inline void
    423 vq_sync_descs(struct virtio_softc *sc, struct virtqueue *vq, int ops)
    424 {
    425 
    426 	/* availoffset == sizeof(vring_desc)*vq_num */
    427 	bus_dmamap_sync(sc->sc_dmat, vq->vq_dmamap, 0, vq->vq_availoffset,
    428 	    ops);
    429 }
    430 
    431 static inline void
    432 vq_sync_aring_all(struct virtio_softc *sc, struct virtqueue *vq, int ops)
    433 {
    434 	uint16_t hdrlen = offsetof(struct vring_avail, ring);
    435 	size_t payloadlen = sc->sc_nvqs * sizeof(uint16_t);
    436 	size_t usedlen = 0;
    437 
    438 	if (sc->sc_active_features & VIRTIO_F_RING_EVENT_IDX)
    439 		usedlen = sizeof(uint16_t);
    440 	bus_dmamap_sync(sc->sc_dmat, vq->vq_dmamap,
    441 	    vq->vq_availoffset, hdrlen + payloadlen + usedlen, ops);
    442 }
    443 
    444 static inline void
    445 vq_sync_aring_header(struct virtio_softc *sc, struct virtqueue *vq, int ops)
    446 {
    447 	uint16_t hdrlen = offsetof(struct vring_avail, ring);
    448 
    449 	bus_dmamap_sync(sc->sc_dmat, vq->vq_dmamap,
    450 	    vq->vq_availoffset, hdrlen, ops);
    451 }
    452 
    453 static inline void
    454 vq_sync_aring_payload(struct virtio_softc *sc, struct virtqueue *vq, int ops)
    455 {
    456 	uint16_t hdrlen = offsetof(struct vring_avail, ring);
    457 	size_t payloadlen = sc->sc_nvqs * sizeof(uint16_t);
    458 
    459 	bus_dmamap_sync(sc->sc_dmat, vq->vq_dmamap,
    460 	    vq->vq_availoffset + hdrlen, payloadlen, ops);
    461 }
    462 
    463 static inline void
    464 vq_sync_aring_used(struct virtio_softc *sc, struct virtqueue *vq, int ops)
    465 {
    466 	uint16_t hdrlen = offsetof(struct vring_avail, ring);
    467 	size_t payloadlen = sc->sc_nvqs * sizeof(uint16_t);
    468 	size_t usedlen = sizeof(uint16_t);
    469 
    470 	if ((sc->sc_active_features & VIRTIO_F_RING_EVENT_IDX) == 0)
    471 		return;
    472 	bus_dmamap_sync(sc->sc_dmat, vq->vq_dmamap,
    473 	    vq->vq_availoffset + hdrlen + payloadlen, usedlen, ops);
    474 }
    475 
    476 static inline void
    477 vq_sync_uring_all(struct virtio_softc *sc, struct virtqueue *vq, int ops)
    478 {
    479 	uint16_t hdrlen = offsetof(struct vring_used, ring);
    480 	size_t payloadlen = sc->sc_nvqs * sizeof(struct vring_used_elem);
    481 	size_t availlen = 0;
    482 
    483 	if (sc->sc_active_features & VIRTIO_F_RING_EVENT_IDX)
    484 		availlen = sizeof(uint16_t);
    485 	bus_dmamap_sync(sc->sc_dmat, vq->vq_dmamap,
    486 	    vq->vq_usedoffset, hdrlen + payloadlen + availlen, ops);
    487 }
    488 
    489 static inline void
    490 vq_sync_uring_header(struct virtio_softc *sc, struct virtqueue *vq, int ops)
    491 {
    492 	uint16_t hdrlen = offsetof(struct vring_used, ring);
    493 
    494 	bus_dmamap_sync(sc->sc_dmat, vq->vq_dmamap,
    495 	    vq->vq_usedoffset, hdrlen, ops);
    496 }
    497 
    498 static inline void
    499 vq_sync_uring_payload(struct virtio_softc *sc, struct virtqueue *vq, int ops)
    500 {
    501 	uint16_t hdrlen = offsetof(struct vring_used, ring);
    502 	size_t payloadlen = sc->sc_nvqs * sizeof(struct vring_used_elem);
    503 
    504 	bus_dmamap_sync(sc->sc_dmat, vq->vq_dmamap,
    505 	    vq->vq_usedoffset + hdrlen, payloadlen, ops);
    506 }
    507 
    508 static inline void
    509 vq_sync_uring_avail(struct virtio_softc *sc, struct virtqueue *vq, int ops)
    510 {
    511 	uint16_t hdrlen = offsetof(struct vring_used, ring);
    512 	size_t payloadlen = sc->sc_nvqs * sizeof(struct vring_used_elem);
    513 	size_t availlen = sizeof(uint16_t);
    514 
    515 	if ((sc->sc_active_features & VIRTIO_F_RING_EVENT_IDX) == 0)
    516 		return;
    517 	bus_dmamap_sync(sc->sc_dmat, vq->vq_dmamap,
    518 	    vq->vq_usedoffset + hdrlen + payloadlen, availlen, ops);
    519 }
    520 
    521 static inline void
    522 vq_sync_indirect(struct virtio_softc *sc, struct virtqueue *vq, int slot,
    523     int ops)
    524 {
    525 	int offset = vq->vq_indirectoffset +
    526 	    sizeof(struct vring_desc) * vq->vq_maxnsegs * slot;
    527 
    528 	bus_dmamap_sync(sc->sc_dmat, vq->vq_dmamap,
    529 	    offset, sizeof(struct vring_desc) * vq->vq_maxnsegs, ops);
    530 }
    531 
    532 bool
    533 virtio_vq_is_enqueued(struct virtio_softc *sc, struct virtqueue *vq)
    534 {
    535 
    536 	if (vq->vq_queued) {
    537 		vq->vq_queued = 0;
    538 		vq_sync_aring_all(sc, vq, BUS_DMASYNC_POSTWRITE);
    539 	}
    540 
    541 	vq_sync_uring_header(sc, vq, BUS_DMASYNC_POSTREAD);
    542 	if (vq->vq_used_idx == virtio_rw16(sc, vq->vq_used->idx))
    543 		return 0;
    544 	vq_sync_uring_payload(sc, vq, BUS_DMASYNC_POSTREAD);
    545 	return 1;
    546 }
    547 
    548 /*
    549  * Scan vq, bus_dmamap_sync for the vqs (not for the payload),
    550  * and calls (*vq_done)() if some entries are consumed.
    551  *
    552  * Can be used as sc_intrhand.
    553  */
    554 int
    555 virtio_vq_intr(struct virtio_softc *sc)
    556 {
    557 	struct virtqueue *vq;
    558 	int i, r = 0;
    559 
    560 	for (i = 0; i < sc->sc_nvqs; i++) {
    561 		vq = &sc->sc_vqs[i];
    562 		if (virtio_vq_is_enqueued(sc, vq) == 1) {
    563 			if (vq->vq_done)
    564 				r |= (*vq->vq_done)(vq);
    565 		}
    566 	}
    567 
    568 	return r;
    569 }
    570 
    571 int
    572 virtio_vq_intrhand(struct virtio_softc *sc)
    573 {
    574 	struct virtqueue *vq;
    575 	int i, r = 0;
    576 
    577 	for (i = 0; i < sc->sc_nvqs; i++) {
    578 		vq = &sc->sc_vqs[i];
    579 		r |= (*vq->vq_intrhand)(vq->vq_intrhand_arg);
    580 	}
    581 
    582 	return r;
    583 }
    584 
    585 
    586 /*
    587  * Increase the event index in order to delay interrupts.
    588  */
    589 int
    590 virtio_postpone_intr(struct virtio_softc *sc, struct virtqueue *vq,
    591 		uint16_t nslots)
    592 {
    593 	uint16_t	idx, nused;
    594 
    595 	idx = vq->vq_used_idx + nslots;
    596 
    597 	/* set the new event index: avail_ring->used_event = idx */
    598 	*vq->vq_used_event = virtio_rw16(sc, idx);
    599 	vq_sync_aring_used(vq->vq_owner, vq, BUS_DMASYNC_PREWRITE);
    600 	vq->vq_queued++;
    601 
    602 	nused = (uint16_t)
    603 		(virtio_rw16(sc, vq->vq_used->idx) - vq->vq_used_idx);
    604 	KASSERT(nused <= vq->vq_num);
    605 
    606 	return nslots < nused;
    607 }
    608 
    609 /*
    610  * Postpone interrupt until 3/4 of the available descriptors have been
    611  * consumed.
    612  */
    613 int
    614 virtio_postpone_intr_smart(struct virtio_softc *sc, struct virtqueue *vq)
    615 {
    616 	uint16_t	nslots;
    617 
    618 	nslots = (uint16_t)
    619 		(virtio_rw16(sc, vq->vq_avail->idx) - vq->vq_used_idx) * 3 / 4;
    620 
    621 	return virtio_postpone_intr(sc, vq, nslots);
    622 }
    623 
    624 /*
    625  * Postpone interrupt until all of the available descriptors have been
    626  * consumed.
    627  */
    628 int
    629 virtio_postpone_intr_far(struct virtio_softc *sc, struct virtqueue *vq)
    630 {
    631 	uint16_t	nslots;
    632 
    633 	nslots = (uint16_t)
    634 		(virtio_rw16(sc, vq->vq_avail->idx) - vq->vq_used_idx);
    635 
    636 	return virtio_postpone_intr(sc, vq, nslots);
    637 }
    638 
    639 /*
    640  * Start/stop vq interrupt.  No guarantee.
    641  */
    642 void
    643 virtio_stop_vq_intr(struct virtio_softc *sc, struct virtqueue *vq)
    644 {
    645 
    646 	if (sc->sc_active_features & VIRTIO_F_RING_EVENT_IDX) {
    647 		/*
    648 		 * No way to disable the interrupt completely with
    649 		 * RingEventIdx. Instead advance used_event by half the
    650 		 * possible value. This won't happen soon and is far enough in
    651 		 * the past to not trigger a spurios interrupt.
    652 		 */
    653 		*vq->vq_used_event = virtio_rw16(sc, vq->vq_used_idx + 0x8000);
    654 		vq_sync_aring_used(sc, vq, BUS_DMASYNC_PREWRITE);
    655 	} else {
    656 		vq->vq_avail->flags |=
    657 		    virtio_rw16(sc, VRING_AVAIL_F_NO_INTERRUPT);
    658 		vq_sync_aring_header(sc, vq, BUS_DMASYNC_PREWRITE);
    659 	}
    660 	vq->vq_queued++;
    661 }
    662 
    663 int
    664 virtio_start_vq_intr(struct virtio_softc *sc, struct virtqueue *vq)
    665 {
    666 
    667 	if (sc->sc_active_features & VIRTIO_F_RING_EVENT_IDX) {
    668 		/*
    669 		 * If event index feature is negotiated, enabling interrupts
    670 		 * is done through setting the latest consumed index in the
    671 		 * used_event field
    672 		 */
    673 		*vq->vq_used_event = virtio_rw16(sc, vq->vq_used_idx);
    674 		vq_sync_aring_used(sc, vq, BUS_DMASYNC_PREWRITE);
    675 	} else {
    676 		vq->vq_avail->flags &=
    677 		    ~virtio_rw16(sc, VRING_AVAIL_F_NO_INTERRUPT);
    678 		vq_sync_aring_header(sc, vq, BUS_DMASYNC_PREWRITE);
    679 	}
    680 	vq->vq_queued++;
    681 
    682 	vq_sync_uring_header(sc, vq, BUS_DMASYNC_POSTREAD);
    683 	if (vq->vq_used_idx == virtio_rw16(sc, vq->vq_used->idx))
    684 		return 0;
    685 	vq_sync_uring_payload(sc, vq, BUS_DMASYNC_POSTREAD);
    686 	return 1;
    687 }
    688 
    689 /*
    690  * Initialize vq structure.
    691  */
    692 static void
    693 virtio_init_vq(struct virtio_softc *sc, struct virtqueue *vq,
    694     const bool reinit)
    695 {
    696 	int i, j;
    697 	int vq_size = vq->vq_num;
    698 
    699 	memset(vq->vq_vaddr, 0, vq->vq_bytesize);
    700 
    701 	/* build the indirect descriptor chain */
    702 	if (vq->vq_indirect != NULL) {
    703 		struct vring_desc *vd;
    704 
    705 		for (i = 0; i < vq_size; i++) {
    706 			vd = vq->vq_indirect;
    707 			vd += vq->vq_maxnsegs * i;
    708 			for (j = 0; j < vq->vq_maxnsegs-1; j++) {
    709 				vd[j].next = virtio_rw16(sc, j + 1);
    710 			}
    711 		}
    712 	}
    713 
    714 	/* free slot management */
    715 	SIMPLEQ_INIT(&vq->vq_freelist);
    716 	for (i = 0; i < vq_size; i++) {
    717 		SIMPLEQ_INSERT_TAIL(&vq->vq_freelist,
    718 				    &vq->vq_entries[i], qe_list);
    719 		vq->vq_entries[i].qe_index = i;
    720 	}
    721 	if (!reinit)
    722 		mutex_init(&vq->vq_freelist_lock, MUTEX_SPIN, sc->sc_ipl);
    723 
    724 	/* enqueue/dequeue status */
    725 	vq->vq_avail_idx = 0;
    726 	vq->vq_used_idx = 0;
    727 	vq->vq_queued = 0;
    728 	if (!reinit) {
    729 		mutex_init(&vq->vq_aring_lock, MUTEX_SPIN, sc->sc_ipl);
    730 		mutex_init(&vq->vq_uring_lock, MUTEX_SPIN, sc->sc_ipl);
    731 	}
    732 	vq_sync_uring_all(sc, vq, BUS_DMASYNC_PREREAD);
    733 	vq->vq_queued++;
    734 }
    735 
    736 /*
    737  * Allocate/free a vq.
    738  */
    739 int
    740 virtio_alloc_vq(struct virtio_softc *sc, struct virtqueue *vq, int index,
    741     int maxsegsize, int maxnsegs, const char *name)
    742 {
    743 	int vq_size, allocsize1, allocsize2, allocsize3, allocsize = 0;
    744 	int rsegs, r, hdrlen;
    745 #define VIRTQUEUE_ALIGN(n)	(((n)+(VIRTIO_PAGE_SIZE-1))&	\
    746 				 ~(VIRTIO_PAGE_SIZE-1))
    747 
    748 	/* Make sure callers allocate vqs in order */
    749 	KASSERT(sc->sc_nvqs == index);
    750 
    751 	memset(vq, 0, sizeof(*vq));
    752 
    753 	vq_size = sc->sc_ops->read_queue_size(sc, index);
    754 	if (vq_size == 0) {
    755 		aprint_error_dev(sc->sc_dev,
    756 				 "virtqueue not exist, index %d for %s\n",
    757 				 index, name);
    758 		goto err;
    759 	}
    760 
    761 	hdrlen = sc->sc_active_features & VIRTIO_F_RING_EVENT_IDX ? 3 : 2;
    762 
    763 	/* allocsize1: descriptor table + avail ring + pad */
    764 	allocsize1 = VIRTQUEUE_ALIGN(sizeof(struct vring_desc)*vq_size
    765 			     + sizeof(uint16_t)*(hdrlen + vq_size));
    766 	/* allocsize2: used ring + pad */
    767 	allocsize2 = VIRTQUEUE_ALIGN(sizeof(uint16_t) * hdrlen
    768 			     + sizeof(struct vring_used_elem)*vq_size);
    769 	/* allocsize3: indirect table */
    770 	if (sc->sc_indirect && maxnsegs >= MINSEG_INDIRECT)
    771 		allocsize3 = sizeof(struct vring_desc) * maxnsegs * vq_size;
    772 	else
    773 		allocsize3 = 0;
    774 	allocsize = allocsize1 + allocsize2 + allocsize3;
    775 
    776 	/* alloc and map the memory */
    777 	r = bus_dmamem_alloc(sc->sc_dmat, allocsize, VIRTIO_PAGE_SIZE, 0,
    778 			     &vq->vq_segs[0], 1, &rsegs, BUS_DMA_WAITOK);
    779 	if (r != 0) {
    780 		aprint_error_dev(sc->sc_dev,
    781 				 "virtqueue %d for %s allocation failed, "
    782 				 "error code %d\n", index, name, r);
    783 		goto err;
    784 	}
    785 	r = bus_dmamem_map(sc->sc_dmat, &vq->vq_segs[0], rsegs, allocsize,
    786 			   &vq->vq_vaddr, BUS_DMA_WAITOK);
    787 	if (r != 0) {
    788 		aprint_error_dev(sc->sc_dev,
    789 				 "virtqueue %d for %s map failed, "
    790 				 "error code %d\n", index, name, r);
    791 		goto err;
    792 	}
    793 	r = bus_dmamap_create(sc->sc_dmat, allocsize, 1, allocsize, 0,
    794 			      BUS_DMA_WAITOK, &vq->vq_dmamap);
    795 	if (r != 0) {
    796 		aprint_error_dev(sc->sc_dev,
    797 				 "virtqueue %d for %s dmamap creation failed, "
    798 				 "error code %d\n", index, name, r);
    799 		goto err;
    800 	}
    801 	r = bus_dmamap_load(sc->sc_dmat, vq->vq_dmamap,
    802 			    vq->vq_vaddr, allocsize, NULL, BUS_DMA_WAITOK);
    803 	if (r != 0) {
    804 		aprint_error_dev(sc->sc_dev,
    805 				 "virtqueue %d for %s dmamap load failed, "
    806 				 "error code %d\n", index, name, r);
    807 		goto err;
    808 	}
    809 
    810 	/* remember addresses and offsets for later use */
    811 	vq->vq_owner = sc;
    812 	vq->vq_num = vq_size;
    813 	vq->vq_index = index;
    814 	vq->vq_desc = vq->vq_vaddr;
    815 	vq->vq_availoffset = sizeof(struct vring_desc)*vq_size;
    816 	vq->vq_avail = (void*)(((char*)vq->vq_desc) + vq->vq_availoffset);
    817 	vq->vq_used_event = (uint16_t *) ((char *)vq->vq_avail +
    818 		 offsetof(struct vring_avail, ring[vq->vq_num]));
    819 	vq->vq_usedoffset = allocsize1;
    820 	vq->vq_used = (void*)(((char*)vq->vq_desc) + vq->vq_usedoffset);
    821 	vq->vq_avail_event = (uint16_t *)((char *)vq->vq_used +
    822 		 offsetof(struct vring_used, ring[vq->vq_num]));
    823 
    824 	if (allocsize3 > 0) {
    825 		vq->vq_indirectoffset = allocsize1 + allocsize2;
    826 		vq->vq_indirect = (void*)(((char*)vq->vq_desc)
    827 					  + vq->vq_indirectoffset);
    828 	}
    829 	vq->vq_bytesize = allocsize;
    830 	vq->vq_maxsegsize = maxsegsize;
    831 	vq->vq_maxnsegs = maxnsegs;
    832 
    833 	/* free slot management */
    834 	vq->vq_entries = kmem_zalloc(sizeof(struct vq_entry)*vq_size,
    835 				     KM_SLEEP);
    836 	virtio_init_vq(sc, vq, false);
    837 
    838 	/* set the vq address */
    839 	sc->sc_ops->setup_queue(sc, index,
    840 	    vq->vq_dmamap->dm_segs[0].ds_addr);
    841 
    842 	aprint_verbose_dev(sc->sc_dev,
    843 			   "allocated %u byte for virtqueue %d for %s, "
    844 			   "size %d\n", allocsize, index, name, vq_size);
    845 	if (allocsize3 > 0)
    846 		aprint_verbose_dev(sc->sc_dev,
    847 				   "using %d byte (%d entries) "
    848 				   "indirect descriptors\n",
    849 				   allocsize3, maxnsegs * vq_size);
    850 
    851 	sc->sc_nvqs++;
    852 
    853 	return 0;
    854 
    855 err:
    856 	sc->sc_ops->setup_queue(sc, index, 0);
    857 	if (vq->vq_dmamap)
    858 		bus_dmamap_destroy(sc->sc_dmat, vq->vq_dmamap);
    859 	if (vq->vq_vaddr)
    860 		bus_dmamem_unmap(sc->sc_dmat, vq->vq_vaddr, allocsize);
    861 	if (vq->vq_segs[0].ds_addr)
    862 		bus_dmamem_free(sc->sc_dmat, &vq->vq_segs[0], 1);
    863 	memset(vq, 0, sizeof(*vq));
    864 
    865 	return -1;
    866 }
    867 
    868 int
    869 virtio_free_vq(struct virtio_softc *sc, struct virtqueue *vq)
    870 {
    871 	struct vq_entry *qe;
    872 	int i = 0;
    873 
    874 	/* device must be already deactivated */
    875 	/* confirm the vq is empty */
    876 	SIMPLEQ_FOREACH(qe, &vq->vq_freelist, qe_list) {
    877 		i++;
    878 	}
    879 	if (i != vq->vq_num) {
    880 		printf("%s: freeing non-empty vq, index %d\n",
    881 		       device_xname(sc->sc_dev), vq->vq_index);
    882 		return EBUSY;
    883 	}
    884 
    885 	/* tell device that there's no virtqueue any longer */
    886 	sc->sc_ops->setup_queue(sc, vq->vq_index, 0);
    887 
    888 	vq_sync_aring_all(sc, vq, BUS_DMASYNC_POSTWRITE);
    889 
    890 	kmem_free(vq->vq_entries, sizeof(*vq->vq_entries) * vq->vq_num);
    891 	bus_dmamap_unload(sc->sc_dmat, vq->vq_dmamap);
    892 	bus_dmamap_destroy(sc->sc_dmat, vq->vq_dmamap);
    893 	bus_dmamem_unmap(sc->sc_dmat, vq->vq_vaddr, vq->vq_bytesize);
    894 	bus_dmamem_free(sc->sc_dmat, &vq->vq_segs[0], 1);
    895 	mutex_destroy(&vq->vq_freelist_lock);
    896 	mutex_destroy(&vq->vq_uring_lock);
    897 	mutex_destroy(&vq->vq_aring_lock);
    898 	memset(vq, 0, sizeof(*vq));
    899 
    900 	sc->sc_nvqs--;
    901 
    902 	return 0;
    903 }
    904 
    905 /*
    906  * Free descriptor management.
    907  */
    908 static struct vq_entry *
    909 vq_alloc_entry(struct virtqueue *vq)
    910 {
    911 	struct vq_entry *qe;
    912 
    913 	mutex_enter(&vq->vq_freelist_lock);
    914 	if (SIMPLEQ_EMPTY(&vq->vq_freelist)) {
    915 		mutex_exit(&vq->vq_freelist_lock);
    916 		return NULL;
    917 	}
    918 	qe = SIMPLEQ_FIRST(&vq->vq_freelist);
    919 	SIMPLEQ_REMOVE_HEAD(&vq->vq_freelist, qe_list);
    920 	mutex_exit(&vq->vq_freelist_lock);
    921 
    922 	return qe;
    923 }
    924 
    925 static void
    926 vq_free_entry(struct virtqueue *vq, struct vq_entry *qe)
    927 {
    928 	mutex_enter(&vq->vq_freelist_lock);
    929 	SIMPLEQ_INSERT_TAIL(&vq->vq_freelist, qe, qe_list);
    930 	mutex_exit(&vq->vq_freelist_lock);
    931 
    932 	return;
    933 }
    934 
    935 /*
    936  * Enqueue several dmamaps as a single request.
    937  */
    938 /*
    939  * Typical usage:
    940  *  <queue size> number of followings are stored in arrays
    941  *  - command blocks (in dmamem) should be pre-allocated and mapped
    942  *  - dmamaps for command blocks should be pre-allocated and loaded
    943  *  - dmamaps for payload should be pre-allocated
    944  *      r = virtio_enqueue_prep(sc, vq, &slot);		// allocate a slot
    945  *	if (r)		// currently 0 or EAGAIN
    946  *	  return r;
    947  *	r = bus_dmamap_load(dmat, dmamap_payload[slot], data, count, ..);
    948  *	if (r) {
    949  *	  virtio_enqueue_abort(sc, vq, slot);
    950  *	  return r;
    951  *	}
    952  *	r = virtio_enqueue_reserve(sc, vq, slot,
    953  *				   dmamap_payload[slot]->dm_nsegs+1);
    954  *							// ^ +1 for command
    955  *	if (r) {	// currently 0 or EAGAIN
    956  *	  bus_dmamap_unload(dmat, dmamap_payload[slot]);
    957  *	  return r;					// do not call abort()
    958  *	}
    959  *	<setup and prepare commands>
    960  *	bus_dmamap_sync(dmat, dmamap_cmd[slot],... BUS_DMASYNC_PREWRITE);
    961  *	bus_dmamap_sync(dmat, dmamap_payload[slot],...);
    962  *	virtio_enqueue(sc, vq, slot, dmamap_cmd[slot], false);
    963  *	virtio_enqueue(sc, vq, slot, dmamap_payload[slot], iswrite);
    964  *	virtio_enqueue_commit(sc, vq, slot, true);
    965  */
    966 
    967 /*
    968  * enqueue_prep: allocate a slot number
    969  */
    970 int
    971 virtio_enqueue_prep(struct virtio_softc *sc, struct virtqueue *vq, int *slotp)
    972 {
    973 	struct vq_entry *qe1;
    974 
    975 	KASSERT(slotp != NULL);
    976 
    977 	qe1 = vq_alloc_entry(vq);
    978 	if (qe1 == NULL)
    979 		return EAGAIN;
    980 	/* next slot is not allocated yet */
    981 	qe1->qe_next = -1;
    982 	*slotp = qe1->qe_index;
    983 
    984 	return 0;
    985 }
    986 
    987 /*
    988  * enqueue_reserve: allocate remaining slots and build the descriptor chain.
    989  */
    990 int
    991 virtio_enqueue_reserve(struct virtio_softc *sc, struct virtqueue *vq,
    992 		       int slot, int nsegs)
    993 {
    994 	int indirect;
    995 	struct vq_entry *qe1 = &vq->vq_entries[slot];
    996 
    997 	KASSERT(qe1->qe_next == -1);
    998 	KASSERT(1 <= nsegs && nsegs <= vq->vq_num);
    999 
   1000 	if ((vq->vq_indirect != NULL) &&
   1001 	    (nsegs >= MINSEG_INDIRECT) &&
   1002 	    (nsegs <= vq->vq_maxnsegs))
   1003 		indirect = 1;
   1004 	else
   1005 		indirect = 0;
   1006 	qe1->qe_indirect = indirect;
   1007 
   1008 	if (indirect) {
   1009 		struct vring_desc *vd;
   1010 		uint64_t addr;
   1011 		int i;
   1012 
   1013 		vd = &vq->vq_desc[qe1->qe_index];
   1014 		addr = vq->vq_dmamap->dm_segs[0].ds_addr
   1015 			+ vq->vq_indirectoffset;
   1016 		addr += sizeof(struct vring_desc)
   1017 			* vq->vq_maxnsegs * qe1->qe_index;
   1018 		vd->addr  = virtio_rw64(sc, addr);
   1019 		vd->len   = virtio_rw32(sc, sizeof(struct vring_desc) * nsegs);
   1020 		vd->flags = virtio_rw16(sc, VRING_DESC_F_INDIRECT);
   1021 
   1022 		vd = vq->vq_indirect;
   1023 		vd += vq->vq_maxnsegs * qe1->qe_index;
   1024 		qe1->qe_desc_base = vd;
   1025 
   1026 		for (i = 0; i < nsegs-1; i++) {
   1027 			vd[i].flags = virtio_rw16(sc, VRING_DESC_F_NEXT);
   1028 		}
   1029 		vd[i].flags  = virtio_rw16(sc, 0);
   1030 		qe1->qe_next = 0;
   1031 
   1032 		return 0;
   1033 	} else {
   1034 		struct vring_desc *vd;
   1035 		struct vq_entry *qe;
   1036 		int i, s;
   1037 
   1038 		vd = &vq->vq_desc[0];
   1039 		qe1->qe_desc_base = vd;
   1040 		qe1->qe_next = qe1->qe_index;
   1041 		s = slot;
   1042 		for (i = 0; i < nsegs - 1; i++) {
   1043 			qe = vq_alloc_entry(vq);
   1044 			if (qe == NULL) {
   1045 				vd[s].flags = virtio_rw16(sc, 0);
   1046 				virtio_enqueue_abort(sc, vq, slot);
   1047 				return EAGAIN;
   1048 			}
   1049 			vd[s].flags = virtio_rw16(sc, VRING_DESC_F_NEXT);
   1050 			vd[s].next  = virtio_rw16(sc, qe->qe_index);
   1051 			s = qe->qe_index;
   1052 		}
   1053 		vd[s].flags = virtio_rw16(sc, 0);
   1054 
   1055 		return 0;
   1056 	}
   1057 }
   1058 
   1059 /*
   1060  * enqueue: enqueue a single dmamap.
   1061  */
   1062 int
   1063 virtio_enqueue(struct virtio_softc *sc, struct virtqueue *vq, int slot,
   1064 	       bus_dmamap_t dmamap, bool write)
   1065 {
   1066 	struct vq_entry *qe1 = &vq->vq_entries[slot];
   1067 	struct vring_desc *vd = qe1->qe_desc_base;
   1068 	int i;
   1069 	int s = qe1->qe_next;
   1070 
   1071 	KASSERT(s >= 0);
   1072 	KASSERT(dmamap->dm_nsegs > 0);
   1073 
   1074 	for (i = 0; i < dmamap->dm_nsegs; i++) {
   1075 		vd[s].addr = virtio_rw64(sc, dmamap->dm_segs[i].ds_addr);
   1076 		vd[s].len  = virtio_rw32(sc, dmamap->dm_segs[i].ds_len);
   1077 		if (!write)
   1078 			vd[s].flags |= virtio_rw16(sc, VRING_DESC_F_WRITE);
   1079 		s = virtio_rw16(sc, vd[s].next);
   1080 	}
   1081 	qe1->qe_next = s;
   1082 
   1083 	return 0;
   1084 }
   1085 
   1086 int
   1087 virtio_enqueue_p(struct virtio_softc *sc, struct virtqueue *vq, int slot,
   1088 		 bus_dmamap_t dmamap, bus_addr_t start, bus_size_t len,
   1089 		 bool write)
   1090 {
   1091 	struct vq_entry *qe1 = &vq->vq_entries[slot];
   1092 	struct vring_desc *vd = qe1->qe_desc_base;
   1093 	int s = qe1->qe_next;
   1094 
   1095 	KASSERT(s >= 0);
   1096 	KASSERT(dmamap->dm_nsegs == 1); /* XXX */
   1097 	KASSERT((dmamap->dm_segs[0].ds_len > start) &&
   1098 		(dmamap->dm_segs[0].ds_len >= start + len));
   1099 
   1100 	vd[s].addr = virtio_rw64(sc, dmamap->dm_segs[0].ds_addr + start);
   1101 	vd[s].len  = virtio_rw32(sc, len);
   1102 	if (!write)
   1103 		vd[s].flags |= virtio_rw16(sc, VRING_DESC_F_WRITE);
   1104 	qe1->qe_next = virtio_rw16(sc, vd[s].next);
   1105 
   1106 	return 0;
   1107 }
   1108 
   1109 /*
   1110  * enqueue_commit: add it to the aring.
   1111  */
   1112 int
   1113 virtio_enqueue_commit(struct virtio_softc *sc, struct virtqueue *vq, int slot,
   1114 		      bool notifynow)
   1115 {
   1116 	struct vq_entry *qe1;
   1117 
   1118 	if (slot < 0) {
   1119 		mutex_enter(&vq->vq_aring_lock);
   1120 		goto notify;
   1121 	}
   1122 	vq_sync_descs(sc, vq, BUS_DMASYNC_PREWRITE);
   1123 	qe1 = &vq->vq_entries[slot];
   1124 	if (qe1->qe_indirect)
   1125 		vq_sync_indirect(sc, vq, slot, BUS_DMASYNC_PREWRITE);
   1126 	mutex_enter(&vq->vq_aring_lock);
   1127 	vq->vq_avail->ring[(vq->vq_avail_idx++) % vq->vq_num] =
   1128 	    virtio_rw16(sc, slot);
   1129 
   1130 notify:
   1131 	if (notifynow) {
   1132 		uint16_t o, n, t;
   1133 		uint16_t flags;
   1134 
   1135 		o = virtio_rw16(sc, vq->vq_avail->idx);
   1136 		n = vq->vq_avail_idx;
   1137 
   1138 		/*
   1139 		 * Prepare for `device->CPU' (host->guest) transfer
   1140 		 * into the buffer.  This must happen before we commit
   1141 		 * the vq->vq_avail->idx update to ensure we're not
   1142 		 * still using the buffer in case program-prior loads
   1143 		 * or stores in it get delayed past the store to
   1144 		 * vq->vq_avail->idx.
   1145 		 */
   1146 		vq_sync_uring_all(sc, vq, BUS_DMASYNC_PREREAD);
   1147 
   1148 		/* ensure payload is published, then avail idx */
   1149 		vq_sync_aring_payload(sc, vq, BUS_DMASYNC_PREWRITE);
   1150 		vq->vq_avail->idx = virtio_rw16(sc, vq->vq_avail_idx);
   1151 		vq_sync_aring_header(sc, vq, BUS_DMASYNC_PREWRITE);
   1152 		vq->vq_queued++;
   1153 
   1154 		if (sc->sc_active_features & VIRTIO_F_RING_EVENT_IDX) {
   1155 			vq_sync_uring_avail(sc, vq, BUS_DMASYNC_POSTREAD);
   1156 			t = virtio_rw16(sc, *vq->vq_avail_event) + 1;
   1157 			if ((uint16_t) (n - t) < (uint16_t) (n - o))
   1158 				sc->sc_ops->kick(sc, vq->vq_index);
   1159 		} else {
   1160 			vq_sync_uring_header(sc, vq, BUS_DMASYNC_POSTREAD);
   1161 			flags = virtio_rw16(sc, vq->vq_used->flags);
   1162 			if (!(flags & VRING_USED_F_NO_NOTIFY))
   1163 				sc->sc_ops->kick(sc, vq->vq_index);
   1164 		}
   1165 	}
   1166 	mutex_exit(&vq->vq_aring_lock);
   1167 
   1168 	return 0;
   1169 }
   1170 
   1171 /*
   1172  * enqueue_abort: rollback.
   1173  */
   1174 int
   1175 virtio_enqueue_abort(struct virtio_softc *sc, struct virtqueue *vq, int slot)
   1176 {
   1177 	struct vq_entry *qe = &vq->vq_entries[slot];
   1178 	struct vring_desc *vd;
   1179 	int s;
   1180 
   1181 	if (qe->qe_next < 0) {
   1182 		vq_free_entry(vq, qe);
   1183 		return 0;
   1184 	}
   1185 
   1186 	s = slot;
   1187 	vd = &vq->vq_desc[0];
   1188 	while (virtio_rw16(sc, vd[s].flags) & VRING_DESC_F_NEXT) {
   1189 		s = virtio_rw16(sc, vd[s].next);
   1190 		vq_free_entry(vq, qe);
   1191 		qe = &vq->vq_entries[s];
   1192 	}
   1193 	vq_free_entry(vq, qe);
   1194 	return 0;
   1195 }
   1196 
   1197 /*
   1198  * Dequeue a request.
   1199  */
   1200 /*
   1201  * dequeue: dequeue a request from uring; dmamap_sync for uring is
   1202  *	    already done in the interrupt handler.
   1203  */
   1204 int
   1205 virtio_dequeue(struct virtio_softc *sc, struct virtqueue *vq,
   1206 	       int *slotp, int *lenp)
   1207 {
   1208 	uint16_t slot, usedidx;
   1209 	struct vq_entry *qe;
   1210 
   1211 	if (vq->vq_used_idx == virtio_rw16(sc, vq->vq_used->idx))
   1212 		return ENOENT;
   1213 	mutex_enter(&vq->vq_uring_lock);
   1214 	usedidx = vq->vq_used_idx++;
   1215 	mutex_exit(&vq->vq_uring_lock);
   1216 	usedidx %= vq->vq_num;
   1217 	slot = virtio_rw32(sc, vq->vq_used->ring[usedidx].id);
   1218 	qe = &vq->vq_entries[slot];
   1219 
   1220 	if (qe->qe_indirect)
   1221 		vq_sync_indirect(sc, vq, slot, BUS_DMASYNC_POSTWRITE);
   1222 
   1223 	if (slotp)
   1224 		*slotp = slot;
   1225 	if (lenp)
   1226 		*lenp = virtio_rw32(sc, vq->vq_used->ring[usedidx].len);
   1227 
   1228 	return 0;
   1229 }
   1230 
   1231 /*
   1232  * dequeue_commit: complete dequeue; the slot is recycled for future use.
   1233  *                 if you forget to call this the slot will be leaked.
   1234  */
   1235 int
   1236 virtio_dequeue_commit(struct virtio_softc *sc, struct virtqueue *vq, int slot)
   1237 {
   1238 	struct vq_entry *qe = &vq->vq_entries[slot];
   1239 	struct vring_desc *vd = &vq->vq_desc[0];
   1240 	int s = slot;
   1241 
   1242 	while (virtio_rw16(sc, vd[s].flags) & VRING_DESC_F_NEXT) {
   1243 		s = virtio_rw16(sc, vd[s].next);
   1244 		vq_free_entry(vq, qe);
   1245 		qe = &vq->vq_entries[s];
   1246 	}
   1247 	vq_free_entry(vq, qe);
   1248 
   1249 	return 0;
   1250 }
   1251 
   1252 /*
   1253  * Attach a child, fill all the members.
   1254  */
   1255 void
   1256 virtio_child_attach_start(struct virtio_softc *sc, device_t child, int ipl,
   1257 		    struct virtqueue *vqs,
   1258 		    virtio_callback config_change,
   1259 		    virtio_callback intr_hand,
   1260 		    int req_flags, int req_features, const char *feat_bits)
   1261 {
   1262 	char buf[1024];
   1263 
   1264 	sc->sc_child = child;
   1265 	sc->sc_ipl = ipl;
   1266 	sc->sc_vqs = vqs;
   1267 	sc->sc_config_change = config_change;
   1268 	sc->sc_intrhand = intr_hand;
   1269 	sc->sc_flags = req_flags;
   1270 
   1271 	virtio_negotiate_features(sc, req_features);
   1272 	snprintb(buf, sizeof(buf), feat_bits, sc->sc_active_features);
   1273 	aprint_normal(": features: %s\n", buf);
   1274 	aprint_naive("\n");
   1275 }
   1276 
   1277 void
   1278 virtio_child_attach_set_vqs(struct virtio_softc *sc,
   1279     struct virtqueue *vqs, int nvq_pairs)
   1280 {
   1281 
   1282 	KASSERT(nvq_pairs == 1 ||
   1283 	    (sc->sc_flags & VIRTIO_F_INTR_SOFTINT) == 0);
   1284 	if (nvq_pairs > 1)
   1285 		sc->sc_child_mq = true;
   1286 
   1287 	sc->sc_vqs = vqs;
   1288 }
   1289 
   1290 int
   1291 virtio_child_attach_finish(struct virtio_softc *sc)
   1292 {
   1293 	int r;
   1294 
   1295 	sc->sc_finished_called = true;
   1296 	r = sc->sc_ops->alloc_interrupts(sc);
   1297 	if (r != 0) {
   1298 		aprint_error_dev(sc->sc_dev, "failed to allocate interrupts\n");
   1299 		goto fail;
   1300 	}
   1301 
   1302 	r = sc->sc_ops->setup_interrupts(sc, 0);
   1303 	if (r != 0) {
   1304 		aprint_error_dev(sc->sc_dev, "failed to setup interrupts\n");
   1305 		goto fail;
   1306 	}
   1307 
   1308 	KASSERT(sc->sc_soft_ih == NULL);
   1309 	if (sc->sc_flags & VIRTIO_F_INTR_SOFTINT) {
   1310 		u_int flags = SOFTINT_NET;
   1311 		if (sc->sc_flags & VIRTIO_F_INTR_MPSAFE)
   1312 			flags |= SOFTINT_MPSAFE;
   1313 
   1314 		sc->sc_soft_ih = softint_establish(flags, virtio_soft_intr, sc);
   1315 		if (sc->sc_soft_ih == NULL) {
   1316 			sc->sc_ops->free_interrupts(sc);
   1317 			aprint_error_dev(sc->sc_dev,
   1318 			    "failed to establish soft interrupt\n");
   1319 			goto fail;
   1320 		}
   1321 	}
   1322 
   1323 	virtio_set_status(sc, VIRTIO_CONFIG_DEVICE_STATUS_DRIVER_OK);
   1324 	return 0;
   1325 
   1326 fail:
   1327 	if (sc->sc_soft_ih) {
   1328 		softint_disestablish(sc->sc_soft_ih);
   1329 		sc->sc_soft_ih = NULL;
   1330 	}
   1331 
   1332 	sc->sc_ops->free_interrupts(sc);
   1333 
   1334 	virtio_set_status(sc, VIRTIO_CONFIG_DEVICE_STATUS_FAILED);
   1335 	return 1;
   1336 }
   1337 
   1338 void
   1339 virtio_child_detach(struct virtio_softc *sc)
   1340 {
   1341 	sc->sc_child = NULL;
   1342 	sc->sc_vqs = NULL;
   1343 
   1344 	virtio_device_reset(sc);
   1345 
   1346 	sc->sc_ops->free_interrupts(sc);
   1347 
   1348 	if (sc->sc_soft_ih) {
   1349 		softint_disestablish(sc->sc_soft_ih);
   1350 		sc->sc_soft_ih = NULL;
   1351 	}
   1352 }
   1353 
   1354 void
   1355 virtio_child_attach_failed(struct virtio_softc *sc)
   1356 {
   1357 	virtio_child_detach(sc);
   1358 
   1359 	virtio_set_status(sc, VIRTIO_CONFIG_DEVICE_STATUS_FAILED);
   1360 
   1361 	sc->sc_child = VIRTIO_CHILD_FAILED;
   1362 }
   1363 
   1364 bus_dma_tag_t
   1365 virtio_dmat(struct virtio_softc *sc)
   1366 {
   1367 	return sc->sc_dmat;
   1368 }
   1369 
   1370 device_t
   1371 virtio_child(struct virtio_softc *sc)
   1372 {
   1373 	return sc->sc_child;
   1374 }
   1375 
   1376 int
   1377 virtio_intrhand(struct virtio_softc *sc)
   1378 {
   1379 	return (*sc->sc_intrhand)(sc);
   1380 }
   1381 
   1382 uint64_t
   1383 virtio_features(struct virtio_softc *sc)
   1384 {
   1385 	return sc->sc_active_features;
   1386 }
   1387 
   1388 int
   1389 virtio_attach_failed(struct virtio_softc *sc)
   1390 {
   1391 	device_t self = sc->sc_dev;
   1392 
   1393 	/* no error if its not connected, but its failed */
   1394 	if (sc->sc_childdevid == 0)
   1395 		return 1;
   1396 
   1397 	if (sc->sc_child == NULL) {
   1398 		aprint_error_dev(self,
   1399 			"no matching child driver; not configured\n");
   1400 		return 1;
   1401 	}
   1402 
   1403 	if (sc->sc_child == VIRTIO_CHILD_FAILED) {
   1404 		aprint_error_dev(self, "virtio configuration failed\n");
   1405 		return 1;
   1406 	}
   1407 
   1408 	/* sanity check */
   1409 	if (!sc->sc_finished_called) {
   1410 		aprint_error_dev(self, "virtio internal error, child driver "
   1411 			"signaled OK but didn't initialize interrupts\n");
   1412 		return 1;
   1413 	}
   1414 
   1415 	return 0;
   1416 }
   1417 
   1418 void
   1419 virtio_print_device_type(device_t self, int id, int revision)
   1420 {
   1421 	aprint_normal_dev(self, "%s device (rev. 0x%02x)\n",
   1422 		  (id < NDEVNAMES ? virtio_device_name[id] : "Unknown"),
   1423 		  revision);
   1424 }
   1425 
   1426 
   1427 MODULE(MODULE_CLASS_DRIVER, virtio, NULL);
   1428 
   1429 #ifdef _MODULE
   1430 #include "ioconf.c"
   1431 #endif
   1432 
   1433 static int
   1434 virtio_modcmd(modcmd_t cmd, void *opaque)
   1435 {
   1436 	int error = 0;
   1437 
   1438 #ifdef _MODULE
   1439 	switch (cmd) {
   1440 	case MODULE_CMD_INIT:
   1441 		error = config_init_component(cfdriver_ioconf_virtio,
   1442 		    cfattach_ioconf_virtio, cfdata_ioconf_virtio);
   1443 		break;
   1444 	case MODULE_CMD_FINI:
   1445 		error = config_fini_component(cfdriver_ioconf_virtio,
   1446 		    cfattach_ioconf_virtio, cfdata_ioconf_virtio);
   1447 		break;
   1448 	default:
   1449 		error = ENOTTY;
   1450 		break;
   1451 	}
   1452 #endif
   1453 
   1454 	return error;
   1455 }
   1456