Home | History | Annotate | Line # | Download | only in ic
nvme.c revision 1.31
      1 /*	$NetBSD: nvme.c,v 1.31 2017/10/28 04:53:55 riastradh Exp $	*/
      2 /*	$OpenBSD: nvme.c,v 1.49 2016/04/18 05:59:50 dlg Exp $ */
      3 
      4 /*
      5  * Copyright (c) 2014 David Gwynne <dlg (at) openbsd.org>
      6  *
      7  * Permission to use, copy, modify, and distribute this software for any
      8  * purpose with or without fee is hereby granted, provided that the above
      9  * copyright notice and this permission notice appear in all copies.
     10  *
     11  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
     12  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
     13  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
     14  * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
     15  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
     16  * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
     17  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
     18  */
     19 
     20 #include <sys/cdefs.h>
     21 __KERNEL_RCSID(0, "$NetBSD: nvme.c,v 1.31 2017/10/28 04:53:55 riastradh Exp $");
     22 
     23 #include <sys/param.h>
     24 #include <sys/systm.h>
     25 #include <sys/kernel.h>
     26 #include <sys/atomic.h>
     27 #include <sys/bus.h>
     28 #include <sys/buf.h>
     29 #include <sys/conf.h>
     30 #include <sys/device.h>
     31 #include <sys/kmem.h>
     32 #include <sys/once.h>
     33 #include <sys/proc.h>
     34 #include <sys/queue.h>
     35 #include <sys/mutex.h>
     36 
     37 #include <uvm/uvm_extern.h>
     38 
     39 #include <dev/ic/nvmereg.h>
     40 #include <dev/ic/nvmevar.h>
     41 #include <dev/ic/nvmeio.h>
     42 
     43 #include "ioconf.h"
     44 
     45 int nvme_adminq_size = 32;
     46 int nvme_ioq_size = 1024;
     47 
     48 static int	nvme_print(void *, const char *);
     49 
     50 static int	nvme_ready(struct nvme_softc *, uint32_t);
     51 static int	nvme_enable(struct nvme_softc *, u_int);
     52 static int	nvme_disable(struct nvme_softc *);
     53 static int	nvme_shutdown(struct nvme_softc *);
     54 
     55 #ifdef NVME_DEBUG
     56 static void	nvme_dumpregs(struct nvme_softc *);
     57 #endif
     58 static int	nvme_identify(struct nvme_softc *, u_int);
     59 static void	nvme_fill_identify(struct nvme_queue *, struct nvme_ccb *,
     60 		    void *);
     61 
     62 static int	nvme_ccbs_alloc(struct nvme_queue *, uint16_t);
     63 static void	nvme_ccbs_free(struct nvme_queue *);
     64 
     65 static struct nvme_ccb *
     66 		nvme_ccb_get(struct nvme_queue *);
     67 static void	nvme_ccb_put(struct nvme_queue *, struct nvme_ccb *);
     68 
     69 static int	nvme_poll(struct nvme_softc *, struct nvme_queue *,
     70 		    struct nvme_ccb *, void (*)(struct nvme_queue *,
     71 		    struct nvme_ccb *, void *), int);
     72 static void	nvme_poll_fill(struct nvme_queue *, struct nvme_ccb *, void *);
     73 static void	nvme_poll_done(struct nvme_queue *, struct nvme_ccb *,
     74 		    struct nvme_cqe *);
     75 static void	nvme_sqe_fill(struct nvme_queue *, struct nvme_ccb *, void *);
     76 static void	nvme_empty_done(struct nvme_queue *, struct nvme_ccb *,
     77 		    struct nvme_cqe *);
     78 
     79 static struct nvme_queue *
     80 		nvme_q_alloc(struct nvme_softc *, uint16_t, u_int, u_int);
     81 static int	nvme_q_create(struct nvme_softc *, struct nvme_queue *);
     82 static int	nvme_q_delete(struct nvme_softc *, struct nvme_queue *);
     83 static void	nvme_q_submit(struct nvme_softc *, struct nvme_queue *,
     84 		    struct nvme_ccb *, void (*)(struct nvme_queue *,
     85 		    struct nvme_ccb *, void *));
     86 static int	nvme_q_complete(struct nvme_softc *, struct nvme_queue *q);
     87 static void	nvme_q_free(struct nvme_softc *, struct nvme_queue *);
     88 
     89 static struct nvme_dmamem *
     90 		nvme_dmamem_alloc(struct nvme_softc *, size_t);
     91 static void	nvme_dmamem_free(struct nvme_softc *, struct nvme_dmamem *);
     92 static void	nvme_dmamem_sync(struct nvme_softc *, struct nvme_dmamem *,
     93 		    int);
     94 
     95 static void	nvme_ns_io_fill(struct nvme_queue *, struct nvme_ccb *,
     96 		    void *);
     97 static void	nvme_ns_io_done(struct nvme_queue *, struct nvme_ccb *,
     98 		    struct nvme_cqe *);
     99 static void	nvme_ns_sync_fill(struct nvme_queue *, struct nvme_ccb *,
    100 		    void *);
    101 static void	nvme_ns_sync_done(struct nvme_queue *, struct nvme_ccb *,
    102 		    struct nvme_cqe *);
    103 static void	nvme_getcache_fill(struct nvme_queue *, struct nvme_ccb *,
    104 		    void *);
    105 static void	nvme_getcache_done(struct nvme_queue *, struct nvme_ccb *,
    106 		    struct nvme_cqe *);
    107 
    108 static void	nvme_pt_fill(struct nvme_queue *, struct nvme_ccb *,
    109 		    void *);
    110 static void	nvme_pt_done(struct nvme_queue *, struct nvme_ccb *,
    111 		    struct nvme_cqe *);
    112 static int	nvme_command_passthrough(struct nvme_softc *,
    113 		    struct nvme_pt_command *, uint16_t, struct lwp *, bool);
    114 
    115 static int	nvme_get_number_of_queues(struct nvme_softc *, u_int *);
    116 
    117 #define NVME_TIMO_QOP		5	/* queue create and delete timeout */
    118 #define NVME_TIMO_IDENT		10	/* probe identify timeout */
    119 #define NVME_TIMO_PT		-1	/* passthrough cmd timeout */
    120 #define NVME_TIMO_SY		60	/* sync cache timeout */
    121 
    122 #define nvme_read4(_s, _r) \
    123 	bus_space_read_4((_s)->sc_iot, (_s)->sc_ioh, (_r))
    124 #define nvme_write4(_s, _r, _v) \
    125 	bus_space_write_4((_s)->sc_iot, (_s)->sc_ioh, (_r), (_v))
    126 /*
    127  * Some controllers, at least Apple NVMe, always require split
    128  * transfers, so don't use bus_space_{read,write}_8() on LP64.
    129  */
    130 static inline uint64_t
    131 nvme_read8(struct nvme_softc *sc, bus_size_t r)
    132 {
    133 	uint64_t v;
    134 	uint32_t *a = (uint32_t *)&v;
    135 
    136 #if _BYTE_ORDER == _LITTLE_ENDIAN
    137 	a[0] = nvme_read4(sc, r);
    138 	a[1] = nvme_read4(sc, r + 4);
    139 #else /* _BYTE_ORDER == _LITTLE_ENDIAN */
    140 	a[1] = nvme_read4(sc, r);
    141 	a[0] = nvme_read4(sc, r + 4);
    142 #endif
    143 
    144 	return v;
    145 }
    146 
    147 static inline void
    148 nvme_write8(struct nvme_softc *sc, bus_size_t r, uint64_t v)
    149 {
    150 	uint32_t *a = (uint32_t *)&v;
    151 
    152 #if _BYTE_ORDER == _LITTLE_ENDIAN
    153 	nvme_write4(sc, r, a[0]);
    154 	nvme_write4(sc, r + 4, a[1]);
    155 #else /* _BYTE_ORDER == _LITTLE_ENDIAN */
    156 	nvme_write4(sc, r, a[1]);
    157 	nvme_write4(sc, r + 4, a[0]);
    158 #endif
    159 }
    160 #define nvme_barrier(_s, _r, _l, _f) \
    161 	bus_space_barrier((_s)->sc_iot, (_s)->sc_ioh, (_r), (_l), (_f))
    162 
    163 #ifdef NVME_DEBUG
    164 static __used void
    165 nvme_dumpregs(struct nvme_softc *sc)
    166 {
    167 	uint64_t r8;
    168 	uint32_t r4;
    169 
    170 #define	DEVNAME(_sc) device_xname((_sc)->sc_dev)
    171 	r8 = nvme_read8(sc, NVME_CAP);
    172 	printf("%s: cap  0x%016"PRIx64"\n", DEVNAME(sc), nvme_read8(sc, NVME_CAP));
    173 	printf("%s:  mpsmax %u (%u)\n", DEVNAME(sc),
    174 	    (u_int)NVME_CAP_MPSMAX(r8), (1 << NVME_CAP_MPSMAX(r8)));
    175 	printf("%s:  mpsmin %u (%u)\n", DEVNAME(sc),
    176 	    (u_int)NVME_CAP_MPSMIN(r8), (1 << NVME_CAP_MPSMIN(r8)));
    177 	printf("%s:  css %"PRIu64"\n", DEVNAME(sc), NVME_CAP_CSS(r8));
    178 	printf("%s:  nssrs %"PRIu64"\n", DEVNAME(sc), NVME_CAP_NSSRS(r8));
    179 	printf("%s:  dstrd %"PRIu64"\n", DEVNAME(sc), NVME_CAP_DSTRD(r8));
    180 	printf("%s:  to %"PRIu64" msec\n", DEVNAME(sc), NVME_CAP_TO(r8));
    181 	printf("%s:  ams %"PRIu64"\n", DEVNAME(sc), NVME_CAP_AMS(r8));
    182 	printf("%s:  cqr %"PRIu64"\n", DEVNAME(sc), NVME_CAP_CQR(r8));
    183 	printf("%s:  mqes %"PRIu64"\n", DEVNAME(sc), NVME_CAP_MQES(r8));
    184 
    185 	printf("%s: vs   0x%04x\n", DEVNAME(sc), nvme_read4(sc, NVME_VS));
    186 
    187 	r4 = nvme_read4(sc, NVME_CC);
    188 	printf("%s: cc   0x%04x\n", DEVNAME(sc), r4);
    189 	printf("%s:  iocqes %u (%u)\n", DEVNAME(sc), NVME_CC_IOCQES_R(r4),
    190 	    (1 << NVME_CC_IOCQES_R(r4)));
    191 	printf("%s:  iosqes %u (%u)\n", DEVNAME(sc), NVME_CC_IOSQES_R(r4),
    192 	    (1 << NVME_CC_IOSQES_R(r4)));
    193 	printf("%s:  shn %u\n", DEVNAME(sc), NVME_CC_SHN_R(r4));
    194 	printf("%s:  ams %u\n", DEVNAME(sc), NVME_CC_AMS_R(r4));
    195 	printf("%s:  mps %u (%u)\n", DEVNAME(sc), NVME_CC_MPS_R(r4),
    196 	    (1 << NVME_CC_MPS_R(r4)));
    197 	printf("%s:  css %u\n", DEVNAME(sc), NVME_CC_CSS_R(r4));
    198 	printf("%s:  en %u\n", DEVNAME(sc), ISSET(r4, NVME_CC_EN) ? 1 : 0);
    199 
    200 	r4 = nvme_read4(sc, NVME_CSTS);
    201 	printf("%s: csts 0x%08x\n", DEVNAME(sc), r4);
    202 	printf("%s:  rdy %u\n", DEVNAME(sc), r4 & NVME_CSTS_RDY);
    203 	printf("%s:  cfs %u\n", DEVNAME(sc), r4 & NVME_CSTS_CFS);
    204 	printf("%s:  shst %x\n", DEVNAME(sc), r4 & NVME_CSTS_SHST_MASK);
    205 
    206 	r4 = nvme_read4(sc, NVME_AQA);
    207 	printf("%s: aqa  0x%08x\n", DEVNAME(sc), r4);
    208 	printf("%s:  acqs %u\n", DEVNAME(sc), NVME_AQA_ACQS_R(r4));
    209 	printf("%s:  asqs %u\n", DEVNAME(sc), NVME_AQA_ASQS_R(r4));
    210 
    211 	printf("%s: asq  0x%016"PRIx64"\n", DEVNAME(sc), nvme_read8(sc, NVME_ASQ));
    212 	printf("%s: acq  0x%016"PRIx64"\n", DEVNAME(sc), nvme_read8(sc, NVME_ACQ));
    213 #undef	DEVNAME
    214 }
    215 #endif	/* NVME_DEBUG */
    216 
    217 static int
    218 nvme_ready(struct nvme_softc *sc, uint32_t rdy)
    219 {
    220 	u_int i = 0;
    221 	uint32_t cc;
    222 
    223 	cc = nvme_read4(sc, NVME_CC);
    224 	if (((cc & NVME_CC_EN) != 0) != (rdy != 0)) {
    225 		aprint_error_dev(sc->sc_dev,
    226 		    "controller enabled status expected %d, found to be %d\n",
    227 		    (rdy != 0), ((cc & NVME_CC_EN) != 0));
    228 		return ENXIO;
    229 	}
    230 
    231 	while ((nvme_read4(sc, NVME_CSTS) & NVME_CSTS_RDY) != rdy) {
    232 		if (i++ > sc->sc_rdy_to)
    233 			return ENXIO;
    234 
    235 		delay(1000);
    236 		nvme_barrier(sc, NVME_CSTS, 4, BUS_SPACE_BARRIER_READ);
    237 	}
    238 
    239 	return 0;
    240 }
    241 
    242 static int
    243 nvme_enable(struct nvme_softc *sc, u_int mps)
    244 {
    245 	uint32_t cc, csts;
    246 
    247 	cc = nvme_read4(sc, NVME_CC);
    248 	csts = nvme_read4(sc, NVME_CSTS);
    249 
    250 	if (ISSET(cc, NVME_CC_EN)) {
    251 		aprint_error_dev(sc->sc_dev, "controller unexpectedly enabled, failed to stay disabled\n");
    252 
    253 		if (ISSET(csts, NVME_CSTS_RDY))
    254 			return 1;
    255 
    256 		goto waitready;
    257 	}
    258 
    259 	nvme_write8(sc, NVME_ASQ, NVME_DMA_DVA(sc->sc_admin_q->q_sq_dmamem));
    260 	nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE);
    261 	delay(5000);
    262 	nvme_write8(sc, NVME_ACQ, NVME_DMA_DVA(sc->sc_admin_q->q_cq_dmamem));
    263 	nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE);
    264 	delay(5000);
    265 
    266 	nvme_write4(sc, NVME_AQA, NVME_AQA_ACQS(sc->sc_admin_q->q_entries) |
    267 	    NVME_AQA_ASQS(sc->sc_admin_q->q_entries));
    268 	nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE);
    269 	delay(5000);
    270 
    271 	CLR(cc, NVME_CC_IOCQES_MASK | NVME_CC_IOSQES_MASK | NVME_CC_SHN_MASK |
    272 	    NVME_CC_AMS_MASK | NVME_CC_MPS_MASK | NVME_CC_CSS_MASK);
    273 	SET(cc, NVME_CC_IOSQES(ffs(64) - 1) | NVME_CC_IOCQES(ffs(16) - 1));
    274 	SET(cc, NVME_CC_SHN(NVME_CC_SHN_NONE));
    275 	SET(cc, NVME_CC_CSS(NVME_CC_CSS_NVM));
    276 	SET(cc, NVME_CC_AMS(NVME_CC_AMS_RR));
    277 	SET(cc, NVME_CC_MPS(mps));
    278 	SET(cc, NVME_CC_EN);
    279 
    280 	nvme_write4(sc, NVME_CC, cc);
    281 	nvme_barrier(sc, 0, sc->sc_ios,
    282 	    BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
    283 	delay(5000);
    284 
    285     waitready:
    286 	return nvme_ready(sc, NVME_CSTS_RDY);
    287 }
    288 
    289 static int
    290 nvme_disable(struct nvme_softc *sc)
    291 {
    292 	uint32_t cc, csts;
    293 
    294 	cc = nvme_read4(sc, NVME_CC);
    295 	csts = nvme_read4(sc, NVME_CSTS);
    296 
    297 	if (ISSET(cc, NVME_CC_EN) && !ISSET(csts, NVME_CSTS_RDY))
    298 		nvme_ready(sc, NVME_CSTS_RDY);
    299 
    300 	CLR(cc, NVME_CC_EN);
    301 
    302 	nvme_write4(sc, NVME_CC, cc);
    303 	nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_READ);
    304 
    305 	delay(5000);
    306 
    307 	return nvme_ready(sc, 0);
    308 }
    309 
    310 int
    311 nvme_attach(struct nvme_softc *sc)
    312 {
    313 	uint64_t cap;
    314 	uint32_t reg;
    315 	u_int dstrd;
    316 	u_int mps = PAGE_SHIFT;
    317 	u_int ioq_allocated;
    318 	uint16_t adminq_entries = nvme_adminq_size;
    319 	uint16_t ioq_entries = nvme_ioq_size;
    320 	int i;
    321 
    322 	reg = nvme_read4(sc, NVME_VS);
    323 	if (reg == 0xffffffff) {
    324 		aprint_error_dev(sc->sc_dev, "invalid mapping\n");
    325 		return 1;
    326 	}
    327 
    328 	if (NVME_VS_TER(reg) == 0)
    329 		aprint_normal_dev(sc->sc_dev, "NVMe %d.%d\n", NVME_VS_MJR(reg),
    330 		    NVME_VS_MNR(reg));
    331 	else
    332 		aprint_normal_dev(sc->sc_dev, "NVMe %d.%d.%d\n", NVME_VS_MJR(reg),
    333 		    NVME_VS_MNR(reg), NVME_VS_TER(reg));
    334 
    335 	cap = nvme_read8(sc, NVME_CAP);
    336 	dstrd = NVME_CAP_DSTRD(cap);
    337 	if (NVME_CAP_MPSMIN(cap) > PAGE_SHIFT) {
    338 		aprint_error_dev(sc->sc_dev, "NVMe minimum page size %u "
    339 		    "is greater than CPU page size %u\n",
    340 		    1 << NVME_CAP_MPSMIN(cap), 1 << PAGE_SHIFT);
    341 		return 1;
    342 	}
    343 	if (NVME_CAP_MPSMAX(cap) < mps)
    344 		mps = NVME_CAP_MPSMAX(cap);
    345 	if (ioq_entries > NVME_CAP_MQES(cap))
    346 		ioq_entries = NVME_CAP_MQES(cap);
    347 
    348 	/* set initial values to be used for admin queue during probe */
    349 	sc->sc_rdy_to = NVME_CAP_TO(cap);
    350 	sc->sc_mps = 1 << mps;
    351 	sc->sc_mdts = MAXPHYS;
    352 	sc->sc_max_sgl = 2;
    353 
    354 	if (nvme_disable(sc) != 0) {
    355 		aprint_error_dev(sc->sc_dev, "unable to disable controller\n");
    356 		return 1;
    357 	}
    358 
    359 	sc->sc_admin_q = nvme_q_alloc(sc, NVME_ADMIN_Q, adminq_entries, dstrd);
    360 	if (sc->sc_admin_q == NULL) {
    361 		aprint_error_dev(sc->sc_dev,
    362 		    "unable to allocate admin queue\n");
    363 		return 1;
    364 	}
    365 	if (sc->sc_intr_establish(sc, NVME_ADMIN_Q, sc->sc_admin_q))
    366 		goto free_admin_q;
    367 
    368 	if (nvme_enable(sc, mps) != 0) {
    369 		aprint_error_dev(sc->sc_dev, "unable to enable controller\n");
    370 		goto disestablish_admin_q;
    371 	}
    372 
    373 	if (nvme_identify(sc, NVME_CAP_MPSMIN(cap)) != 0) {
    374 		aprint_error_dev(sc->sc_dev, "unable to identify controller\n");
    375 		goto disable;
    376 	}
    377 
    378 	/* we know how big things are now */
    379 	sc->sc_max_sgl = sc->sc_mdts / sc->sc_mps;
    380 
    381 	/* reallocate ccbs of admin queue with new max sgl. */
    382 	nvme_ccbs_free(sc->sc_admin_q);
    383 	nvme_ccbs_alloc(sc->sc_admin_q, sc->sc_admin_q->q_entries);
    384 
    385 	if (sc->sc_use_mq) {
    386 		/* Limit the number of queues to the number allocated in HW */
    387 		if (nvme_get_number_of_queues(sc, &ioq_allocated) != 0) {
    388 			aprint_error_dev(sc->sc_dev,
    389 			    "unable to get number of queues\n");
    390 			goto disable;
    391 		}
    392 		if (sc->sc_nq > ioq_allocated)
    393 			sc->sc_nq = ioq_allocated;
    394 	}
    395 
    396 	sc->sc_q = kmem_zalloc(sizeof(*sc->sc_q) * sc->sc_nq, KM_SLEEP);
    397 	for (i = 0; i < sc->sc_nq; i++) {
    398 		sc->sc_q[i] = nvme_q_alloc(sc, i + 1, ioq_entries, dstrd);
    399 		if (sc->sc_q[i] == NULL) {
    400 			aprint_error_dev(sc->sc_dev,
    401 			    "unable to allocate io queue\n");
    402 			goto free_q;
    403 		}
    404 		if (nvme_q_create(sc, sc->sc_q[i]) != 0) {
    405 			aprint_error_dev(sc->sc_dev,
    406 			    "unable to create io queue\n");
    407 			nvme_q_free(sc, sc->sc_q[i]);
    408 			goto free_q;
    409 		}
    410 	}
    411 
    412 	if (!sc->sc_use_mq)
    413 		nvme_write4(sc, NVME_INTMC, 1);
    414 
    415 	/* probe subdevices */
    416 	sc->sc_namespaces = kmem_zalloc(sizeof(*sc->sc_namespaces) * sc->sc_nn,
    417 	    KM_SLEEP);
    418 	nvme_rescan(sc->sc_dev, "nvme", &i);
    419 
    420 	return 0;
    421 
    422 free_q:
    423 	while (--i >= 0) {
    424 		nvme_q_delete(sc, sc->sc_q[i]);
    425 		nvme_q_free(sc, sc->sc_q[i]);
    426 	}
    427 disable:
    428 	nvme_disable(sc);
    429 disestablish_admin_q:
    430 	sc->sc_intr_disestablish(sc, NVME_ADMIN_Q);
    431 free_admin_q:
    432 	nvme_q_free(sc, sc->sc_admin_q);
    433 
    434 	return 1;
    435 }
    436 
    437 int
    438 nvme_rescan(device_t self, const char *attr, const int *flags)
    439 {
    440 	struct nvme_softc *sc = device_private(self);
    441 	struct nvme_attach_args naa;
    442 	uint64_t cap;
    443 	int ioq_entries = nvme_ioq_size;
    444 	int i;
    445 
    446 	cap = nvme_read8(sc, NVME_CAP);
    447 	if (ioq_entries > NVME_CAP_MQES(cap))
    448 		ioq_entries = NVME_CAP_MQES(cap);
    449 
    450 	for (i = 0; i < sc->sc_nn; i++) {
    451 		if (sc->sc_namespaces[i].dev)
    452 			continue;
    453 		memset(&naa, 0, sizeof(naa));
    454 		naa.naa_nsid = i + 1;
    455 		naa.naa_qentries = (ioq_entries - 1) * sc->sc_nq;
    456 		naa.naa_maxphys = sc->sc_mdts;
    457 		sc->sc_namespaces[i].dev = config_found(sc->sc_dev, &naa,
    458 		    nvme_print);
    459 	}
    460 	return 0;
    461 }
    462 
    463 static int
    464 nvme_print(void *aux, const char *pnp)
    465 {
    466 	struct nvme_attach_args *naa = aux;
    467 
    468 	if (pnp)
    469 		aprint_normal("at %s", pnp);
    470 
    471 	if (naa->naa_nsid > 0)
    472 		aprint_normal(" nsid %d", naa->naa_nsid);
    473 
    474 	return UNCONF;
    475 }
    476 
    477 int
    478 nvme_detach(struct nvme_softc *sc, int flags)
    479 {
    480 	int i, error;
    481 
    482 	error = config_detach_children(sc->sc_dev, flags);
    483 	if (error)
    484 		return error;
    485 
    486 	error = nvme_shutdown(sc);
    487 	if (error)
    488 		return error;
    489 
    490 	/* from now on we are committed to detach, following will never fail */
    491 	for (i = 0; i < sc->sc_nq; i++)
    492 		nvme_q_free(sc, sc->sc_q[i]);
    493 	kmem_free(sc->sc_q, sizeof(*sc->sc_q) * sc->sc_nq);
    494 	nvme_q_free(sc, sc->sc_admin_q);
    495 
    496 	return 0;
    497 }
    498 
    499 static int
    500 nvme_shutdown(struct nvme_softc *sc)
    501 {
    502 	uint32_t cc, csts;
    503 	bool disabled = false;
    504 	int i;
    505 
    506 	if (!sc->sc_use_mq)
    507 		nvme_write4(sc, NVME_INTMS, 1);
    508 
    509 	for (i = 0; i < sc->sc_nq; i++) {
    510 		if (nvme_q_delete(sc, sc->sc_q[i]) != 0) {
    511 			aprint_error_dev(sc->sc_dev,
    512 			    "unable to delete io queue %d, disabling\n", i + 1);
    513 			disabled = true;
    514 		}
    515 	}
    516 	sc->sc_intr_disestablish(sc, NVME_ADMIN_Q);
    517 	if (disabled)
    518 		goto disable;
    519 
    520 	cc = nvme_read4(sc, NVME_CC);
    521 	CLR(cc, NVME_CC_SHN_MASK);
    522 	SET(cc, NVME_CC_SHN(NVME_CC_SHN_NORMAL));
    523 	nvme_write4(sc, NVME_CC, cc);
    524 
    525 	for (i = 0; i < 4000; i++) {
    526 		nvme_barrier(sc, 0, sc->sc_ios,
    527 		    BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
    528 		csts = nvme_read4(sc, NVME_CSTS);
    529 		if ((csts & NVME_CSTS_SHST_MASK) == NVME_CSTS_SHST_DONE)
    530 			return 0;
    531 
    532 		delay(1000);
    533 	}
    534 
    535 	aprint_error_dev(sc->sc_dev, "unable to shudown, disabling\n");
    536 
    537 disable:
    538 	nvme_disable(sc);
    539 	return 0;
    540 }
    541 
    542 void
    543 nvme_childdet(device_t self, device_t child)
    544 {
    545 	struct nvme_softc *sc = device_private(self);
    546 	int i;
    547 
    548 	for (i = 0; i < sc->sc_nn; i++) {
    549 		if (sc->sc_namespaces[i].dev == child) {
    550 			/* Already freed ns->ident. */
    551 			sc->sc_namespaces[i].dev = NULL;
    552 			break;
    553 		}
    554 	}
    555 }
    556 
    557 int
    558 nvme_ns_identify(struct nvme_softc *sc, uint16_t nsid)
    559 {
    560 	struct nvme_sqe sqe;
    561 	struct nvm_identify_namespace *identify;
    562 	struct nvme_dmamem *mem;
    563 	struct nvme_ccb *ccb;
    564 	struct nvme_namespace *ns;
    565 	int rv;
    566 
    567 	KASSERT(nsid > 0);
    568 
    569 	ccb = nvme_ccb_get(sc->sc_admin_q);
    570 	KASSERT(ccb != NULL); /* it's a bug if we don't have spare ccb here */
    571 
    572 	mem = nvme_dmamem_alloc(sc, sizeof(*identify));
    573 	if (mem == NULL)
    574 		return ENOMEM;
    575 
    576 	memset(&sqe, 0, sizeof(sqe));
    577 	sqe.opcode = NVM_ADMIN_IDENTIFY;
    578 	htolem32(&sqe.nsid, nsid);
    579 	htolem64(&sqe.entry.prp[0], NVME_DMA_DVA(mem));
    580 	htolem32(&sqe.cdw10, 0);
    581 
    582 	ccb->ccb_done = nvme_empty_done;
    583 	ccb->ccb_cookie = &sqe;
    584 
    585 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_PREREAD);
    586 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_IDENT);
    587 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_POSTREAD);
    588 
    589 	nvme_ccb_put(sc->sc_admin_q, ccb);
    590 
    591 	if (rv != 0) {
    592 		rv = EIO;
    593 		goto done;
    594 	}
    595 
    596 	/* commit */
    597 
    598 	identify = kmem_zalloc(sizeof(*identify), KM_SLEEP);
    599 	*identify = *((volatile struct nvm_identify_namespace *)NVME_DMA_KVA(mem));
    600 	//memcpy(identify, NVME_DMA_KVA(mem), sizeof(*identify));
    601 
    602 	ns = nvme_ns_get(sc, nsid);
    603 	KASSERT(ns);
    604 	ns->ident = identify;
    605 
    606 done:
    607 	nvme_dmamem_free(sc, mem);
    608 
    609 	return rv;
    610 }
    611 
    612 int
    613 nvme_ns_dobio(struct nvme_softc *sc, uint16_t nsid, void *cookie,
    614     struct buf *bp, void *data, size_t datasize,
    615     int secsize, daddr_t blkno, int flags, nvme_nnc_done nnc_done)
    616 {
    617 	struct nvme_queue *q = nvme_get_q(sc);
    618 	struct nvme_ccb *ccb;
    619 	bus_dmamap_t dmap;
    620 	int i, error;
    621 
    622 	ccb = nvme_ccb_get(q);
    623 	if (ccb == NULL)
    624 		return EAGAIN;
    625 
    626 	ccb->ccb_done = nvme_ns_io_done;
    627 	ccb->ccb_cookie = cookie;
    628 
    629 	/* namespace context */
    630 	ccb->nnc_nsid = nsid;
    631 	ccb->nnc_flags = flags;
    632 	ccb->nnc_buf = bp;
    633 	ccb->nnc_datasize = datasize;
    634 	ccb->nnc_secsize = secsize;
    635 	ccb->nnc_blkno = blkno;
    636 	ccb->nnc_done = nnc_done;
    637 
    638 	dmap = ccb->ccb_dmamap;
    639 	error = bus_dmamap_load(sc->sc_dmat, dmap, data,
    640 	    datasize, NULL,
    641 	    (ISSET(flags, NVME_NS_CTX_F_POLL) ?
    642 	      BUS_DMA_NOWAIT : BUS_DMA_WAITOK) |
    643 	    (ISSET(flags, NVME_NS_CTX_F_READ) ?
    644 	      BUS_DMA_READ : BUS_DMA_WRITE));
    645 	if (error) {
    646 		nvme_ccb_put(q, ccb);
    647 		return error;
    648 	}
    649 
    650 	bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize,
    651 	    ISSET(flags, NVME_NS_CTX_F_READ) ?
    652 	    BUS_DMASYNC_PREREAD : BUS_DMASYNC_PREWRITE);
    653 
    654 	if (dmap->dm_nsegs > 2) {
    655 		for (i = 1; i < dmap->dm_nsegs; i++) {
    656 			htolem64(&ccb->ccb_prpl[i - 1],
    657 			    dmap->dm_segs[i].ds_addr);
    658 		}
    659 		bus_dmamap_sync(sc->sc_dmat,
    660 		    NVME_DMA_MAP(q->q_ccb_prpls),
    661 		    ccb->ccb_prpl_off,
    662 		    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    663 		    BUS_DMASYNC_PREWRITE);
    664 	}
    665 
    666 	if (ISSET(flags, NVME_NS_CTX_F_POLL)) {
    667 		if (nvme_poll(sc, q, ccb, nvme_ns_io_fill, NVME_TIMO_PT) != 0)
    668 			return EIO;
    669 		return 0;
    670 	}
    671 
    672 	nvme_q_submit(sc, q, ccb, nvme_ns_io_fill);
    673 	return 0;
    674 }
    675 
    676 static void
    677 nvme_ns_io_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    678 {
    679 	struct nvme_sqe_io *sqe = slot;
    680 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    681 
    682 	sqe->opcode = ISSET(ccb->nnc_flags, NVME_NS_CTX_F_READ) ?
    683 	    NVM_CMD_READ : NVM_CMD_WRITE;
    684 	htolem32(&sqe->nsid, ccb->nnc_nsid);
    685 
    686 	htolem64(&sqe->entry.prp[0], dmap->dm_segs[0].ds_addr);
    687 	switch (dmap->dm_nsegs) {
    688 	case 1:
    689 		break;
    690 	case 2:
    691 		htolem64(&sqe->entry.prp[1], dmap->dm_segs[1].ds_addr);
    692 		break;
    693 	default:
    694 		/* the prp list is already set up and synced */
    695 		htolem64(&sqe->entry.prp[1], ccb->ccb_prpl_dva);
    696 		break;
    697 	}
    698 
    699 	htolem64(&sqe->slba, ccb->nnc_blkno);
    700 
    701 	if (ISSET(ccb->nnc_flags, NVME_NS_CTX_F_FUA))
    702 		htolem16(&sqe->ioflags, NVM_SQE_IO_FUA);
    703 
    704 	/* guaranteed by upper layers, but check just in case */
    705 	KASSERT((ccb->nnc_datasize % ccb->nnc_secsize) == 0);
    706 	htolem16(&sqe->nlb, (ccb->nnc_datasize / ccb->nnc_secsize) - 1);
    707 }
    708 
    709 static void
    710 nvme_ns_io_done(struct nvme_queue *q, struct nvme_ccb *ccb,
    711     struct nvme_cqe *cqe)
    712 {
    713 	struct nvme_softc *sc = q->q_sc;
    714 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    715 	void *nnc_cookie = ccb->ccb_cookie;
    716 	nvme_nnc_done nnc_done = ccb->nnc_done;
    717 	struct buf *bp = ccb->nnc_buf;
    718 
    719 	if (dmap->dm_nsegs > 2) {
    720 		bus_dmamap_sync(sc->sc_dmat,
    721 		    NVME_DMA_MAP(q->q_ccb_prpls),
    722 		    ccb->ccb_prpl_off,
    723 		    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    724 		    BUS_DMASYNC_POSTWRITE);
    725 	}
    726 
    727 	bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize,
    728 	    ISSET(ccb->nnc_flags, NVME_NS_CTX_F_READ) ?
    729 	    BUS_DMASYNC_POSTREAD : BUS_DMASYNC_POSTWRITE);
    730 
    731 	bus_dmamap_unload(sc->sc_dmat, dmap);
    732 	nvme_ccb_put(q, ccb);
    733 
    734 	nnc_done(nnc_cookie, bp, lemtoh16(&cqe->flags), lemtoh32(&cqe->cdw0));
    735 }
    736 
    737 /*
    738  * If there is no volatile write cache, it makes no sense to issue
    739  * flush commands or query for the status.
    740  */
    741 bool
    742 nvme_has_volatile_write_cache(struct nvme_softc *sc)
    743 {
    744 	/* sc_identify is filled during attachment */
    745 	return  ((sc->sc_identify.vwc & NVME_ID_CTRLR_VWC_PRESENT) != 0);
    746 }
    747 
    748 int
    749 nvme_ns_sync(struct nvme_softc *sc, uint16_t nsid, void *cookie,
    750     int flags, nvme_nnc_done nnc_done)
    751 {
    752 	struct nvme_queue *q = nvme_get_q(sc);
    753 	struct nvme_ccb *ccb;
    754 
    755 	ccb = nvme_ccb_get(q);
    756 	if (ccb == NULL)
    757 		return EAGAIN;
    758 
    759 	ccb->ccb_done = nvme_ns_sync_done;
    760 	ccb->ccb_cookie = cookie;
    761 
    762 	/* namespace context */
    763 	ccb->nnc_nsid = nsid;
    764 	ccb->nnc_flags = flags;
    765 	ccb->nnc_done = nnc_done;
    766 
    767 	if (ISSET(flags, NVME_NS_CTX_F_POLL)) {
    768 		if (nvme_poll(sc, q, ccb, nvme_ns_sync_fill, NVME_TIMO_SY) != 0)
    769 			return EIO;
    770 		return 0;
    771 	}
    772 
    773 	nvme_q_submit(sc, q, ccb, nvme_ns_sync_fill);
    774 	return 0;
    775 }
    776 
    777 static void
    778 nvme_ns_sync_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    779 {
    780 	struct nvme_sqe *sqe = slot;
    781 
    782 	sqe->opcode = NVM_CMD_FLUSH;
    783 	htolem32(&sqe->nsid, ccb->nnc_nsid);
    784 }
    785 
    786 static void
    787 nvme_ns_sync_done(struct nvme_queue *q, struct nvme_ccb *ccb,
    788     struct nvme_cqe *cqe)
    789 {
    790 	void *cookie = ccb->ccb_cookie;
    791 	nvme_nnc_done nnc_done = ccb->nnc_done;
    792 
    793 	nvme_ccb_put(q, ccb);
    794 
    795 	nnc_done(cookie, NULL, lemtoh16(&cqe->flags), lemtoh32(&cqe->cdw0));
    796 }
    797 
    798 /*
    799  * Get status of volatile write cache. Always asynchronous.
    800  */
    801 int
    802 nvme_admin_getcache(struct nvme_softc *sc, void *cookie, nvme_nnc_done nnc_done)
    803 {
    804 	struct nvme_ccb *ccb;
    805 	struct nvme_queue *q = sc->sc_admin_q;
    806 
    807 	ccb = nvme_ccb_get(q);
    808 	if (ccb == NULL)
    809 		return EAGAIN;
    810 
    811 	ccb->ccb_done = nvme_getcache_done;
    812 	ccb->ccb_cookie = cookie;
    813 
    814 	/* namespace context */
    815 	ccb->nnc_flags = 0;
    816 	ccb->nnc_done = nnc_done;
    817 
    818 	nvme_q_submit(sc, q, ccb, nvme_getcache_fill);
    819 	return 0;
    820 }
    821 
    822 static void
    823 nvme_getcache_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    824 {
    825 	struct nvme_sqe *sqe = slot;
    826 
    827 	sqe->opcode = NVM_ADMIN_GET_FEATURES;
    828 	sqe->cdw10 = NVM_FEATURE_VOLATILE_WRITE_CACHE;
    829 }
    830 
    831 static void
    832 nvme_getcache_done(struct nvme_queue *q, struct nvme_ccb *ccb,
    833     struct nvme_cqe *cqe)
    834 {
    835 	void *cookie = ccb->ccb_cookie;
    836 	nvme_nnc_done nnc_done = ccb->nnc_done;
    837 
    838 	nvme_ccb_put(q, ccb);
    839 
    840 	nnc_done(cookie, NULL, lemtoh16(&cqe->flags), lemtoh32(&cqe->cdw0));
    841 }
    842 
    843 void
    844 nvme_ns_free(struct nvme_softc *sc, uint16_t nsid)
    845 {
    846 	struct nvme_namespace *ns;
    847 	struct nvm_identify_namespace *identify;
    848 
    849 	ns = nvme_ns_get(sc, nsid);
    850 	KASSERT(ns);
    851 
    852 	identify = ns->ident;
    853 	ns->ident = NULL;
    854 	if (identify != NULL)
    855 		kmem_free(identify, sizeof(*identify));
    856 }
    857 
    858 static void
    859 nvme_pt_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    860 {
    861 	struct nvme_softc *sc = q->q_sc;
    862 	struct nvme_sqe *sqe = slot;
    863 	struct nvme_pt_command *pt = ccb->ccb_cookie;
    864 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    865 	int i;
    866 
    867 	sqe->opcode = pt->cmd.opcode;
    868 	htolem32(&sqe->nsid, pt->cmd.nsid);
    869 
    870 	if (pt->buf != NULL && pt->len > 0) {
    871 		htolem64(&sqe->entry.prp[0], dmap->dm_segs[0].ds_addr);
    872 		switch (dmap->dm_nsegs) {
    873 		case 1:
    874 			break;
    875 		case 2:
    876 			htolem64(&sqe->entry.prp[1], dmap->dm_segs[1].ds_addr);
    877 			break;
    878 		default:
    879 			for (i = 1; i < dmap->dm_nsegs; i++) {
    880 				htolem64(&ccb->ccb_prpl[i - 1],
    881 				    dmap->dm_segs[i].ds_addr);
    882 			}
    883 			bus_dmamap_sync(sc->sc_dmat,
    884 			    NVME_DMA_MAP(q->q_ccb_prpls),
    885 			    ccb->ccb_prpl_off,
    886 			    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    887 			    BUS_DMASYNC_PREWRITE);
    888 			htolem64(&sqe->entry.prp[1], ccb->ccb_prpl_dva);
    889 			break;
    890 		}
    891 	}
    892 
    893 	htolem32(&sqe->cdw10, pt->cmd.cdw10);
    894 	htolem32(&sqe->cdw11, pt->cmd.cdw11);
    895 	htolem32(&sqe->cdw12, pt->cmd.cdw12);
    896 	htolem32(&sqe->cdw13, pt->cmd.cdw13);
    897 	htolem32(&sqe->cdw14, pt->cmd.cdw14);
    898 	htolem32(&sqe->cdw15, pt->cmd.cdw15);
    899 }
    900 
    901 static void
    902 nvme_pt_done(struct nvme_queue *q, struct nvme_ccb *ccb, struct nvme_cqe *cqe)
    903 {
    904 	struct nvme_softc *sc = q->q_sc;
    905 	struct nvme_pt_command *pt = ccb->ccb_cookie;
    906 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    907 
    908 	if (pt->buf != NULL && pt->len > 0) {
    909 		if (dmap->dm_nsegs > 2) {
    910 			bus_dmamap_sync(sc->sc_dmat,
    911 			    NVME_DMA_MAP(q->q_ccb_prpls),
    912 			    ccb->ccb_prpl_off,
    913 			    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    914 			    BUS_DMASYNC_POSTWRITE);
    915 		}
    916 
    917 		bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize,
    918 		    pt->is_read ? BUS_DMASYNC_POSTREAD : BUS_DMASYNC_POSTWRITE);
    919 		bus_dmamap_unload(sc->sc_dmat, dmap);
    920 	}
    921 
    922 	pt->cpl.cdw0 = lemtoh32(&cqe->cdw0);
    923 	pt->cpl.flags = lemtoh16(&cqe->flags) & ~NVME_CQE_PHASE;
    924 }
    925 
    926 static int
    927 nvme_command_passthrough(struct nvme_softc *sc, struct nvme_pt_command *pt,
    928     uint16_t nsid, struct lwp *l, bool is_adminq)
    929 {
    930 	struct nvme_queue *q;
    931 	struct nvme_ccb *ccb;
    932 	void *buf = NULL;
    933 	int error;
    934 
    935 	/* limit command size to maximum data transfer size */
    936 	if ((pt->buf == NULL && pt->len > 0) ||
    937 	    (pt->buf != NULL && (pt->len == 0 || pt->len > sc->sc_mdts)))
    938 		return EINVAL;
    939 
    940 	q = is_adminq ? sc->sc_admin_q : nvme_get_q(sc);
    941 	ccb = nvme_ccb_get(q);
    942 	if (ccb == NULL)
    943 		return EBUSY;
    944 
    945 	if (pt->buf != NULL) {
    946 		KASSERT(pt->len > 0);
    947 		buf = kmem_alloc(pt->len, KM_SLEEP);
    948 		if (!pt->is_read) {
    949 			error = copyin(pt->buf, buf, pt->len);
    950 			if (error)
    951 				goto kmem_free;
    952 		}
    953 		error = bus_dmamap_load(sc->sc_dmat, ccb->ccb_dmamap, buf,
    954 		    pt->len, NULL,
    955 		    BUS_DMA_WAITOK |
    956 		      (pt->is_read ? BUS_DMA_READ : BUS_DMA_WRITE));
    957 		if (error)
    958 			goto kmem_free;
    959 		bus_dmamap_sync(sc->sc_dmat, ccb->ccb_dmamap,
    960 		    0, ccb->ccb_dmamap->dm_mapsize,
    961 		    pt->is_read ? BUS_DMASYNC_PREREAD : BUS_DMASYNC_PREWRITE);
    962 	}
    963 
    964 	ccb->ccb_done = nvme_pt_done;
    965 	ccb->ccb_cookie = pt;
    966 
    967 	pt->cmd.nsid = nsid;
    968 	if (nvme_poll(sc, q, ccb, nvme_pt_fill, NVME_TIMO_PT)) {
    969 		error = EIO;
    970 		goto out;
    971 	}
    972 
    973 	error = 0;
    974 out:
    975 	if (buf != NULL) {
    976 		if (error == 0 && pt->is_read)
    977 			error = copyout(buf, pt->buf, pt->len);
    978 kmem_free:
    979 		kmem_free(buf, pt->len);
    980 	}
    981 	nvme_ccb_put(q, ccb);
    982 	return error;
    983 }
    984 
    985 static void
    986 nvme_q_submit(struct nvme_softc *sc, struct nvme_queue *q, struct nvme_ccb *ccb,
    987     void (*fill)(struct nvme_queue *, struct nvme_ccb *, void *))
    988 {
    989 	struct nvme_sqe *sqe = NVME_DMA_KVA(q->q_sq_dmamem);
    990 	uint32_t tail;
    991 
    992 	mutex_enter(&q->q_sq_mtx);
    993 	tail = q->q_sq_tail;
    994 	if (++q->q_sq_tail >= q->q_entries)
    995 		q->q_sq_tail = 0;
    996 
    997 	sqe += tail;
    998 
    999 	bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem),
   1000 	    sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_POSTWRITE);
   1001 	memset(sqe, 0, sizeof(*sqe));
   1002 	(*fill)(q, ccb, sqe);
   1003 	sqe->cid = ccb->ccb_id;
   1004 	bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem),
   1005 	    sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_PREWRITE);
   1006 
   1007 	nvme_write4(sc, q->q_sqtdbl, q->q_sq_tail);
   1008 	mutex_exit(&q->q_sq_mtx);
   1009 }
   1010 
   1011 struct nvme_poll_state {
   1012 	struct nvme_sqe s;
   1013 	struct nvme_cqe c;
   1014 };
   1015 
   1016 static int
   1017 nvme_poll(struct nvme_softc *sc, struct nvme_queue *q, struct nvme_ccb *ccb,
   1018     void (*fill)(struct nvme_queue *, struct nvme_ccb *, void *), int timo_sec)
   1019 {
   1020 	struct nvme_poll_state state;
   1021 	void (*done)(struct nvme_queue *, struct nvme_ccb *, struct nvme_cqe *);
   1022 	void *cookie;
   1023 	uint16_t flags;
   1024 	int step = 10;
   1025 	int maxloop = timo_sec * 1000000 / step;
   1026 	int error = 0;
   1027 
   1028 	memset(&state, 0, sizeof(state));
   1029 	(*fill)(q, ccb, &state.s);
   1030 
   1031 	done = ccb->ccb_done;
   1032 	cookie = ccb->ccb_cookie;
   1033 
   1034 	ccb->ccb_done = nvme_poll_done;
   1035 	ccb->ccb_cookie = &state;
   1036 
   1037 	nvme_q_submit(sc, q, ccb, nvme_poll_fill);
   1038 	while (!ISSET(state.c.flags, htole16(NVME_CQE_PHASE))) {
   1039 		if (nvme_q_complete(sc, q) == 0)
   1040 			delay(step);
   1041 
   1042 		if (timo_sec >= 0 && --maxloop <= 0) {
   1043 			error = ETIMEDOUT;
   1044 			break;
   1045 		}
   1046 	}
   1047 
   1048 	ccb->ccb_cookie = cookie;
   1049 	done(q, ccb, &state.c);
   1050 
   1051 	if (error == 0) {
   1052 		flags = lemtoh16(&state.c.flags);
   1053 		return flags & ~NVME_CQE_PHASE;
   1054 	} else {
   1055 		return 1;
   1056 	}
   1057 }
   1058 
   1059 static void
   1060 nvme_poll_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
   1061 {
   1062 	struct nvme_sqe *sqe = slot;
   1063 	struct nvme_poll_state *state = ccb->ccb_cookie;
   1064 
   1065 	*sqe = state->s;
   1066 }
   1067 
   1068 static void
   1069 nvme_poll_done(struct nvme_queue *q, struct nvme_ccb *ccb,
   1070     struct nvme_cqe *cqe)
   1071 {
   1072 	struct nvme_poll_state *state = ccb->ccb_cookie;
   1073 
   1074 	SET(cqe->flags, htole16(NVME_CQE_PHASE));
   1075 	state->c = *cqe;
   1076 }
   1077 
   1078 static void
   1079 nvme_sqe_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
   1080 {
   1081 	struct nvme_sqe *src = ccb->ccb_cookie;
   1082 	struct nvme_sqe *dst = slot;
   1083 
   1084 	*dst = *src;
   1085 }
   1086 
   1087 static void
   1088 nvme_empty_done(struct nvme_queue *q, struct nvme_ccb *ccb,
   1089     struct nvme_cqe *cqe)
   1090 {
   1091 }
   1092 
   1093 static int
   1094 nvme_q_complete(struct nvme_softc *sc, struct nvme_queue *q)
   1095 {
   1096 	struct nvme_ccb *ccb;
   1097 	struct nvme_cqe *ring = NVME_DMA_KVA(q->q_cq_dmamem), *cqe;
   1098 	uint16_t flags;
   1099 	int rv = 0;
   1100 
   1101 	mutex_enter(&q->q_cq_mtx);
   1102 
   1103 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_POSTREAD);
   1104 	for (;;) {
   1105 		cqe = &ring[q->q_cq_head];
   1106 		flags = lemtoh16(&cqe->flags);
   1107 		if ((flags & NVME_CQE_PHASE) != q->q_cq_phase)
   1108 			break;
   1109 
   1110 		ccb = &q->q_ccbs[cqe->cid];
   1111 
   1112 		if (++q->q_cq_head >= q->q_entries) {
   1113 			q->q_cq_head = 0;
   1114 			q->q_cq_phase ^= NVME_CQE_PHASE;
   1115 		}
   1116 
   1117 #ifdef DEBUG
   1118 		/*
   1119 		 * If we get spurious completion notification, something
   1120 		 * is seriously hosed up. Very likely DMA to some random
   1121 		 * memory place happened, so just bail out.
   1122 		 */
   1123 		if ((intptr_t)ccb->ccb_cookie == NVME_CCB_FREE) {
   1124 			panic("%s: invalid ccb detected",
   1125 			    device_xname(sc->sc_dev));
   1126 			/* NOTREACHED */
   1127 		}
   1128 #endif
   1129 
   1130 		rv++;
   1131 
   1132 		/*
   1133 		 * Unlock the mutex before calling the ccb_done callback
   1134 		 * and re-lock afterwards. The callback triggers lddone()
   1135 		 * which schedules another i/o, and also calls nvme_ccb_put().
   1136 		 * Unlock/relock avoids possibility of deadlock.
   1137 		 */
   1138 		mutex_exit(&q->q_cq_mtx);
   1139 		ccb->ccb_done(q, ccb, cqe);
   1140 		mutex_enter(&q->q_cq_mtx);
   1141 	}
   1142 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_PREREAD);
   1143 
   1144 	if (rv)
   1145 		nvme_write4(sc, q->q_cqhdbl, q->q_cq_head);
   1146 
   1147 	mutex_exit(&q->q_cq_mtx);
   1148 
   1149 	if (rv) {
   1150 		mutex_enter(&q->q_ccb_mtx);
   1151 		q->q_nccbs_avail += rv;
   1152 		mutex_exit(&q->q_ccb_mtx);
   1153 	}
   1154 
   1155 	return rv;
   1156 }
   1157 
   1158 static int
   1159 nvme_identify(struct nvme_softc *sc, u_int mps)
   1160 {
   1161 	char sn[41], mn[81], fr[17];
   1162 	struct nvm_identify_controller *identify;
   1163 	struct nvme_dmamem *mem;
   1164 	struct nvme_ccb *ccb;
   1165 	u_int mdts;
   1166 	int rv = 1;
   1167 
   1168 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1169 	KASSERT(ccb != NULL); /* it's a bug if we don't have spare ccb here */
   1170 
   1171 	mem = nvme_dmamem_alloc(sc, sizeof(*identify));
   1172 	if (mem == NULL)
   1173 		return 1;
   1174 
   1175 	ccb->ccb_done = nvme_empty_done;
   1176 	ccb->ccb_cookie = mem;
   1177 
   1178 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_PREREAD);
   1179 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_fill_identify,
   1180 	    NVME_TIMO_IDENT);
   1181 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_POSTREAD);
   1182 
   1183 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1184 
   1185 	if (rv != 0)
   1186 		goto done;
   1187 
   1188 	identify = NVME_DMA_KVA(mem);
   1189 
   1190 	strnvisx(sn, sizeof(sn), (const char *)identify->sn,
   1191 	    sizeof(identify->sn), VIS_TRIM|VIS_SAFE|VIS_OCTAL);
   1192 	strnvisx(mn, sizeof(mn), (const char *)identify->mn,
   1193 	    sizeof(identify->mn), VIS_TRIM|VIS_SAFE|VIS_OCTAL);
   1194 	strnvisx(fr, sizeof(fr), (const char *)identify->fr,
   1195 	    sizeof(identify->fr), VIS_TRIM|VIS_SAFE|VIS_OCTAL);
   1196 	aprint_normal_dev(sc->sc_dev, "%s, firmware %s, serial %s\n", mn, fr,
   1197 	    sn);
   1198 
   1199 	if (identify->mdts > 0) {
   1200 		mdts = (1 << identify->mdts) * (1 << mps);
   1201 		if (mdts < sc->sc_mdts)
   1202 			sc->sc_mdts = mdts;
   1203 	}
   1204 
   1205 	sc->sc_nn = lemtoh32(&identify->nn);
   1206 
   1207 	memcpy(&sc->sc_identify, identify, sizeof(sc->sc_identify));
   1208 
   1209 done:
   1210 	nvme_dmamem_free(sc, mem);
   1211 
   1212 	return rv;
   1213 }
   1214 
   1215 static int
   1216 nvme_q_create(struct nvme_softc *sc, struct nvme_queue *q)
   1217 {
   1218 	struct nvme_sqe_q sqe;
   1219 	struct nvme_ccb *ccb;
   1220 	int rv;
   1221 
   1222 	if (sc->sc_use_mq && sc->sc_intr_establish(sc, q->q_id, q) != 0)
   1223 		return 1;
   1224 
   1225 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1226 	KASSERT(ccb != NULL);
   1227 
   1228 	ccb->ccb_done = nvme_empty_done;
   1229 	ccb->ccb_cookie = &sqe;
   1230 
   1231 	memset(&sqe, 0, sizeof(sqe));
   1232 	sqe.opcode = NVM_ADMIN_ADD_IOCQ;
   1233 	htolem64(&sqe.prp1, NVME_DMA_DVA(q->q_cq_dmamem));
   1234 	htolem16(&sqe.qsize, q->q_entries - 1);
   1235 	htolem16(&sqe.qid, q->q_id);
   1236 	sqe.qflags = NVM_SQE_CQ_IEN | NVM_SQE_Q_PC;
   1237 	if (sc->sc_use_mq)
   1238 		htolem16(&sqe.cqid, q->q_id);	/* qid == vector */
   1239 
   1240 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1241 	if (rv != 0)
   1242 		goto fail;
   1243 
   1244 	ccb->ccb_done = nvme_empty_done;
   1245 	ccb->ccb_cookie = &sqe;
   1246 
   1247 	memset(&sqe, 0, sizeof(sqe));
   1248 	sqe.opcode = NVM_ADMIN_ADD_IOSQ;
   1249 	htolem64(&sqe.prp1, NVME_DMA_DVA(q->q_sq_dmamem));
   1250 	htolem16(&sqe.qsize, q->q_entries - 1);
   1251 	htolem16(&sqe.qid, q->q_id);
   1252 	htolem16(&sqe.cqid, q->q_id);
   1253 	sqe.qflags = NVM_SQE_Q_PC;
   1254 
   1255 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1256 	if (rv != 0)
   1257 		goto fail;
   1258 
   1259 fail:
   1260 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1261 	return rv;
   1262 }
   1263 
   1264 static int
   1265 nvme_q_delete(struct nvme_softc *sc, struct nvme_queue *q)
   1266 {
   1267 	struct nvme_sqe_q sqe;
   1268 	struct nvme_ccb *ccb;
   1269 	int rv;
   1270 
   1271 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1272 	KASSERT(ccb != NULL);
   1273 
   1274 	ccb->ccb_done = nvme_empty_done;
   1275 	ccb->ccb_cookie = &sqe;
   1276 
   1277 	memset(&sqe, 0, sizeof(sqe));
   1278 	sqe.opcode = NVM_ADMIN_DEL_IOSQ;
   1279 	htolem16(&sqe.qid, q->q_id);
   1280 
   1281 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1282 	if (rv != 0)
   1283 		goto fail;
   1284 
   1285 	ccb->ccb_done = nvme_empty_done;
   1286 	ccb->ccb_cookie = &sqe;
   1287 
   1288 	memset(&sqe, 0, sizeof(sqe));
   1289 	sqe.opcode = NVM_ADMIN_DEL_IOCQ;
   1290 	htolem16(&sqe.qid, q->q_id);
   1291 
   1292 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1293 	if (rv != 0)
   1294 		goto fail;
   1295 
   1296 fail:
   1297 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1298 
   1299 	if (rv == 0 && sc->sc_use_mq) {
   1300 		if (sc->sc_intr_disestablish(sc, q->q_id))
   1301 			rv = 1;
   1302 	}
   1303 
   1304 	return rv;
   1305 }
   1306 
   1307 static void
   1308 nvme_fill_identify(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
   1309 {
   1310 	struct nvme_sqe *sqe = slot;
   1311 	struct nvme_dmamem *mem = ccb->ccb_cookie;
   1312 
   1313 	sqe->opcode = NVM_ADMIN_IDENTIFY;
   1314 	htolem64(&sqe->entry.prp[0], NVME_DMA_DVA(mem));
   1315 	htolem32(&sqe->cdw10, 1);
   1316 }
   1317 
   1318 static int
   1319 nvme_get_number_of_queues(struct nvme_softc *sc, u_int *nqap)
   1320 {
   1321 	struct nvme_pt_command pt;
   1322 	struct nvme_ccb *ccb;
   1323 	uint16_t ncqa, nsqa;
   1324 	int rv;
   1325 
   1326 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1327 	KASSERT(ccb != NULL); /* it's a bug if we don't have spare ccb here */
   1328 
   1329 	memset(&pt, 0, sizeof(pt));
   1330 	pt.cmd.opcode = NVM_ADMIN_GET_FEATURES;
   1331 	pt.cmd.cdw10 = NVM_FEATURE_NUMBER_OF_QUEUES;
   1332 
   1333 	ccb->ccb_done = nvme_pt_done;
   1334 	ccb->ccb_cookie = &pt;
   1335 
   1336 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_pt_fill, NVME_TIMO_QOP);
   1337 
   1338 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1339 
   1340 	if (rv != 0) {
   1341 		*nqap = 0;
   1342 		return EIO;
   1343 	}
   1344 
   1345 	ncqa = pt.cpl.cdw0 >> 16;
   1346 	nsqa = pt.cpl.cdw0 & 0xffff;
   1347 	*nqap = MIN(ncqa, nsqa) + 1;
   1348 
   1349 	return 0;
   1350 }
   1351 
   1352 static int
   1353 nvme_ccbs_alloc(struct nvme_queue *q, uint16_t nccbs)
   1354 {
   1355 	struct nvme_softc *sc = q->q_sc;
   1356 	struct nvme_ccb *ccb;
   1357 	bus_addr_t off;
   1358 	uint64_t *prpl;
   1359 	u_int i;
   1360 
   1361 	mutex_init(&q->q_ccb_mtx, MUTEX_DEFAULT, IPL_BIO);
   1362 	SIMPLEQ_INIT(&q->q_ccb_list);
   1363 
   1364 	q->q_ccbs = kmem_alloc(sizeof(*ccb) * nccbs, KM_SLEEP);
   1365 
   1366 	q->q_nccbs = nccbs;
   1367 	q->q_nccbs_avail = nccbs;
   1368 	q->q_ccb_prpls = nvme_dmamem_alloc(sc,
   1369 	    sizeof(*prpl) * sc->sc_max_sgl * nccbs);
   1370 
   1371 	prpl = NVME_DMA_KVA(q->q_ccb_prpls);
   1372 	off = 0;
   1373 
   1374 	for (i = 0; i < nccbs; i++) {
   1375 		ccb = &q->q_ccbs[i];
   1376 
   1377 		if (bus_dmamap_create(sc->sc_dmat, sc->sc_mdts,
   1378 		    sc->sc_max_sgl + 1 /* we get a free prp in the sqe */,
   1379 		    sc->sc_mps, sc->sc_mps, BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW,
   1380 		    &ccb->ccb_dmamap) != 0)
   1381 			goto free_maps;
   1382 
   1383 		ccb->ccb_id = i;
   1384 		ccb->ccb_prpl = prpl;
   1385 		ccb->ccb_prpl_off = off;
   1386 		ccb->ccb_prpl_dva = NVME_DMA_DVA(q->q_ccb_prpls) + off;
   1387 
   1388 		SIMPLEQ_INSERT_TAIL(&q->q_ccb_list, ccb, ccb_entry);
   1389 
   1390 		prpl += sc->sc_max_sgl;
   1391 		off += sizeof(*prpl) * sc->sc_max_sgl;
   1392 	}
   1393 
   1394 	return 0;
   1395 
   1396 free_maps:
   1397 	nvme_ccbs_free(q);
   1398 	return 1;
   1399 }
   1400 
   1401 static struct nvme_ccb *
   1402 nvme_ccb_get(struct nvme_queue *q)
   1403 {
   1404 	struct nvme_ccb *ccb = NULL;
   1405 
   1406 	mutex_enter(&q->q_ccb_mtx);
   1407 	if (q->q_nccbs_avail > 0) {
   1408 		ccb = SIMPLEQ_FIRST(&q->q_ccb_list);
   1409 		KASSERT(ccb != NULL);
   1410 		q->q_nccbs_avail--;
   1411 
   1412 		SIMPLEQ_REMOVE_HEAD(&q->q_ccb_list, ccb_entry);
   1413 #ifdef DEBUG
   1414 		ccb->ccb_cookie = NULL;
   1415 #endif
   1416 	}
   1417 	mutex_exit(&q->q_ccb_mtx);
   1418 
   1419 	return ccb;
   1420 }
   1421 
   1422 static void
   1423 nvme_ccb_put(struct nvme_queue *q, struct nvme_ccb *ccb)
   1424 {
   1425 
   1426 	mutex_enter(&q->q_ccb_mtx);
   1427 #ifdef DEBUG
   1428 	ccb->ccb_cookie = (void *)NVME_CCB_FREE;
   1429 #endif
   1430 	SIMPLEQ_INSERT_HEAD(&q->q_ccb_list, ccb, ccb_entry);
   1431 	mutex_exit(&q->q_ccb_mtx);
   1432 }
   1433 
   1434 static void
   1435 nvme_ccbs_free(struct nvme_queue *q)
   1436 {
   1437 	struct nvme_softc *sc = q->q_sc;
   1438 	struct nvme_ccb *ccb;
   1439 
   1440 	mutex_enter(&q->q_ccb_mtx);
   1441 	while ((ccb = SIMPLEQ_FIRST(&q->q_ccb_list)) != NULL) {
   1442 		SIMPLEQ_REMOVE_HEAD(&q->q_ccb_list, ccb_entry);
   1443 		bus_dmamap_destroy(sc->sc_dmat, ccb->ccb_dmamap);
   1444 	}
   1445 	mutex_exit(&q->q_ccb_mtx);
   1446 
   1447 	nvme_dmamem_free(sc, q->q_ccb_prpls);
   1448 	kmem_free(q->q_ccbs, sizeof(*ccb) * q->q_nccbs);
   1449 	q->q_ccbs = NULL;
   1450 	mutex_destroy(&q->q_ccb_mtx);
   1451 }
   1452 
   1453 static struct nvme_queue *
   1454 nvme_q_alloc(struct nvme_softc *sc, uint16_t id, u_int entries, u_int dstrd)
   1455 {
   1456 	struct nvme_queue *q;
   1457 
   1458 	q = kmem_alloc(sizeof(*q), KM_SLEEP);
   1459 	q->q_sc = sc;
   1460 	q->q_sq_dmamem = nvme_dmamem_alloc(sc,
   1461 	    sizeof(struct nvme_sqe) * entries);
   1462 	if (q->q_sq_dmamem == NULL)
   1463 		goto free;
   1464 
   1465 	q->q_cq_dmamem = nvme_dmamem_alloc(sc,
   1466 	    sizeof(struct nvme_cqe) * entries);
   1467 	if (q->q_cq_dmamem == NULL)
   1468 		goto free_sq;
   1469 
   1470 	memset(NVME_DMA_KVA(q->q_sq_dmamem), 0, NVME_DMA_LEN(q->q_sq_dmamem));
   1471 	memset(NVME_DMA_KVA(q->q_cq_dmamem), 0, NVME_DMA_LEN(q->q_cq_dmamem));
   1472 
   1473 	mutex_init(&q->q_sq_mtx, MUTEX_DEFAULT, IPL_BIO);
   1474 	mutex_init(&q->q_cq_mtx, MUTEX_DEFAULT, IPL_BIO);
   1475 	q->q_sqtdbl = NVME_SQTDBL(id, dstrd);
   1476 	q->q_cqhdbl = NVME_CQHDBL(id, dstrd);
   1477 	q->q_id = id;
   1478 	q->q_entries = entries;
   1479 	q->q_sq_tail = 0;
   1480 	q->q_cq_head = 0;
   1481 	q->q_cq_phase = NVME_CQE_PHASE;
   1482 
   1483 	nvme_dmamem_sync(sc, q->q_sq_dmamem, BUS_DMASYNC_PREWRITE);
   1484 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_PREREAD);
   1485 
   1486 	/*
   1487 	 * Due to definition of full and empty queue (queue is empty
   1488 	 * when head == tail, full when tail is one less then head),
   1489 	 * we can actually only have (entries - 1) in-flight commands.
   1490 	 */
   1491 	if (nvme_ccbs_alloc(q, entries - 1) != 0) {
   1492 		aprint_error_dev(sc->sc_dev, "unable to allocate ccbs\n");
   1493 		goto free_cq;
   1494 	}
   1495 
   1496 	return q;
   1497 
   1498 free_cq:
   1499 	nvme_dmamem_free(sc, q->q_cq_dmamem);
   1500 free_sq:
   1501 	nvme_dmamem_free(sc, q->q_sq_dmamem);
   1502 free:
   1503 	kmem_free(q, sizeof(*q));
   1504 
   1505 	return NULL;
   1506 }
   1507 
   1508 static void
   1509 nvme_q_free(struct nvme_softc *sc, struct nvme_queue *q)
   1510 {
   1511 	nvme_ccbs_free(q);
   1512 	mutex_destroy(&q->q_sq_mtx);
   1513 	mutex_destroy(&q->q_cq_mtx);
   1514 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_POSTREAD);
   1515 	nvme_dmamem_sync(sc, q->q_sq_dmamem, BUS_DMASYNC_POSTWRITE);
   1516 	nvme_dmamem_free(sc, q->q_cq_dmamem);
   1517 	nvme_dmamem_free(sc, q->q_sq_dmamem);
   1518 	kmem_free(q, sizeof(*q));
   1519 }
   1520 
   1521 int
   1522 nvme_intr(void *xsc)
   1523 {
   1524 	struct nvme_softc *sc = xsc;
   1525 
   1526 	/*
   1527 	 * INTx is level triggered, controller deasserts the interrupt only
   1528 	 * when we advance command queue head via write to the doorbell.
   1529 	 * Tell the controller to block the interrupts while we process
   1530 	 * the queue(s).
   1531 	 */
   1532 	nvme_write4(sc, NVME_INTMS, 1);
   1533 
   1534 	softint_schedule(sc->sc_softih[0]);
   1535 
   1536 	/* don't know, might not have been for us */
   1537 	return 1;
   1538 }
   1539 
   1540 void
   1541 nvme_softintr_intx(void *xq)
   1542 {
   1543 	struct nvme_queue *q = xq;
   1544 	struct nvme_softc *sc = q->q_sc;
   1545 
   1546 	nvme_q_complete(sc, sc->sc_admin_q);
   1547 	if (sc->sc_q != NULL)
   1548 	        nvme_q_complete(sc, sc->sc_q[0]);
   1549 
   1550 	/*
   1551 	 * Processing done, tell controller to issue interrupts again. There
   1552 	 * is no race, as NVMe spec requires the controller to maintain state,
   1553 	 * and assert the interrupt whenever there are unacknowledged
   1554 	 * completion queue entries.
   1555 	 */
   1556 	nvme_write4(sc, NVME_INTMC, 1);
   1557 }
   1558 
   1559 int
   1560 nvme_intr_msi(void *xq)
   1561 {
   1562 	struct nvme_queue *q = xq;
   1563 
   1564 	KASSERT(q && q->q_sc && q->q_sc->sc_softih
   1565 	    && q->q_sc->sc_softih[q->q_id]);
   1566 
   1567 	/*
   1568 	 * MSI/MSI-X are edge triggered, so can handover processing to softint
   1569 	 * without masking the interrupt.
   1570 	 */
   1571 	softint_schedule(q->q_sc->sc_softih[q->q_id]);
   1572 
   1573 	return 1;
   1574 }
   1575 
   1576 void
   1577 nvme_softintr_msi(void *xq)
   1578 {
   1579 	struct nvme_queue *q = xq;
   1580 	struct nvme_softc *sc = q->q_sc;
   1581 
   1582 	nvme_q_complete(sc, q);
   1583 }
   1584 
   1585 static struct nvme_dmamem *
   1586 nvme_dmamem_alloc(struct nvme_softc *sc, size_t size)
   1587 {
   1588 	struct nvme_dmamem *ndm;
   1589 	int nsegs;
   1590 
   1591 	ndm = kmem_zalloc(sizeof(*ndm), KM_SLEEP);
   1592 	if (ndm == NULL)
   1593 		return NULL;
   1594 
   1595 	ndm->ndm_size = size;
   1596 
   1597 	if (bus_dmamap_create(sc->sc_dmat, size, 1, size, 0,
   1598 	    BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW, &ndm->ndm_map) != 0)
   1599 		goto ndmfree;
   1600 
   1601 	if (bus_dmamem_alloc(sc->sc_dmat, size, sc->sc_mps, 0, &ndm->ndm_seg,
   1602 	    1, &nsegs, BUS_DMA_WAITOK) != 0)
   1603 		goto destroy;
   1604 
   1605 	if (bus_dmamem_map(sc->sc_dmat, &ndm->ndm_seg, nsegs, size,
   1606 	    &ndm->ndm_kva, BUS_DMA_WAITOK) != 0)
   1607 		goto free;
   1608 	memset(ndm->ndm_kva, 0, size);
   1609 
   1610 	if (bus_dmamap_load(sc->sc_dmat, ndm->ndm_map, ndm->ndm_kva, size,
   1611 	    NULL, BUS_DMA_WAITOK) != 0)
   1612 		goto unmap;
   1613 
   1614 	return ndm;
   1615 
   1616 unmap:
   1617 	bus_dmamem_unmap(sc->sc_dmat, ndm->ndm_kva, size);
   1618 free:
   1619 	bus_dmamem_free(sc->sc_dmat, &ndm->ndm_seg, 1);
   1620 destroy:
   1621 	bus_dmamap_destroy(sc->sc_dmat, ndm->ndm_map);
   1622 ndmfree:
   1623 	kmem_free(ndm, sizeof(*ndm));
   1624 	return NULL;
   1625 }
   1626 
   1627 static void
   1628 nvme_dmamem_sync(struct nvme_softc *sc, struct nvme_dmamem *mem, int ops)
   1629 {
   1630 	bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(mem),
   1631 	    0, NVME_DMA_LEN(mem), ops);
   1632 }
   1633 
   1634 void
   1635 nvme_dmamem_free(struct nvme_softc *sc, struct nvme_dmamem *ndm)
   1636 {
   1637 	bus_dmamap_unload(sc->sc_dmat, ndm->ndm_map);
   1638 	bus_dmamem_unmap(sc->sc_dmat, ndm->ndm_kva, ndm->ndm_size);
   1639 	bus_dmamem_free(sc->sc_dmat, &ndm->ndm_seg, 1);
   1640 	bus_dmamap_destroy(sc->sc_dmat, ndm->ndm_map);
   1641 	kmem_free(ndm, sizeof(*ndm));
   1642 }
   1643 
   1644 /*
   1645  * ioctl
   1646  */
   1647 
   1648 dev_type_open(nvmeopen);
   1649 dev_type_close(nvmeclose);
   1650 dev_type_ioctl(nvmeioctl);
   1651 
   1652 const struct cdevsw nvme_cdevsw = {
   1653 	.d_open = nvmeopen,
   1654 	.d_close = nvmeclose,
   1655 	.d_read = noread,
   1656 	.d_write = nowrite,
   1657 	.d_ioctl = nvmeioctl,
   1658 	.d_stop = nostop,
   1659 	.d_tty = notty,
   1660 	.d_poll = nopoll,
   1661 	.d_mmap = nommap,
   1662 	.d_kqfilter = nokqfilter,
   1663 	.d_discard = nodiscard,
   1664 	.d_flag = D_OTHER,
   1665 };
   1666 
   1667 /*
   1668  * Accept an open operation on the control device.
   1669  */
   1670 int
   1671 nvmeopen(dev_t dev, int flag, int mode, struct lwp *l)
   1672 {
   1673 	struct nvme_softc *sc;
   1674 	int unit = minor(dev) / 0x10000;
   1675 	int nsid = minor(dev) & 0xffff;
   1676 	int nsidx;
   1677 
   1678 	if ((sc = device_lookup_private(&nvme_cd, unit)) == NULL)
   1679 		return ENXIO;
   1680 	if ((sc->sc_flags & NVME_F_ATTACHED) == 0)
   1681 		return ENXIO;
   1682 
   1683 	if (nsid == 0) {
   1684 		/* controller */
   1685 		if (ISSET(sc->sc_flags, NVME_F_OPEN))
   1686 			return EBUSY;
   1687 		SET(sc->sc_flags, NVME_F_OPEN);
   1688 	} else {
   1689 		/* namespace */
   1690 		nsidx = nsid - 1;
   1691 		if (nsidx >= sc->sc_nn || sc->sc_namespaces[nsidx].dev == NULL)
   1692 			return ENXIO;
   1693 		if (ISSET(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN))
   1694 			return EBUSY;
   1695 		SET(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN);
   1696 	}
   1697 	return 0;
   1698 }
   1699 
   1700 /*
   1701  * Accept the last close on the control device.
   1702  */
   1703 int
   1704 nvmeclose(dev_t dev, int flag, int mode, struct lwp *l)
   1705 {
   1706 	struct nvme_softc *sc;
   1707 	int unit = minor(dev) / 0x10000;
   1708 	int nsid = minor(dev) & 0xffff;
   1709 	int nsidx;
   1710 
   1711 	sc = device_lookup_private(&nvme_cd, unit);
   1712 	if (sc == NULL)
   1713 		return ENXIO;
   1714 
   1715 	if (nsid == 0) {
   1716 		/* controller */
   1717 		CLR(sc->sc_flags, NVME_F_OPEN);
   1718 	} else {
   1719 		/* namespace */
   1720 		nsidx = nsid - 1;
   1721 		if (nsidx >= sc->sc_nn)
   1722 			return ENXIO;
   1723 		CLR(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN);
   1724 	}
   1725 
   1726 	return 0;
   1727 }
   1728 
   1729 /*
   1730  * Handle control operations.
   1731  */
   1732 int
   1733 nvmeioctl(dev_t dev, u_long cmd, void *data, int flag, struct lwp *l)
   1734 {
   1735 	struct nvme_softc *sc;
   1736 	int unit = minor(dev) / 0x10000;
   1737 	int nsid = minor(dev) & 0xffff;
   1738 	struct nvme_pt_command *pt;
   1739 
   1740 	sc = device_lookup_private(&nvme_cd, unit);
   1741 	if (sc == NULL)
   1742 		return ENXIO;
   1743 
   1744 	switch (cmd) {
   1745 	case NVME_PASSTHROUGH_CMD:
   1746 		pt = data;
   1747 		return nvme_command_passthrough(sc, data,
   1748 		    nsid == 0 ? pt->cmd.nsid : nsid, l, nsid == 0);
   1749 	}
   1750 
   1751 	return ENOTTY;
   1752 }
   1753