Home | History | Annotate | Line # | Download | only in ic
nvme.c revision 1.26.4.1
      1 /*	$NetBSD: nvme.c,v 1.26.4.1 2017/04/30 10:27:16 pgoyette 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 1.26.4.1 2017/04/30 10:27:16 pgoyette 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 int nvme_adminq_size = 32;
     44 int nvme_ioq_size = 1024;
     45 
     46 static int	nvme_print(void *, const char *);
     47 
     48 static int	nvme_ready(struct nvme_softc *, uint32_t);
     49 static int	nvme_enable(struct nvme_softc *, u_int);
     50 static int	nvme_disable(struct nvme_softc *);
     51 static int	nvme_shutdown(struct nvme_softc *);
     52 
     53 static void	nvme_version(struct nvme_softc *, uint32_t);
     54 #ifdef NVME_DEBUG
     55 static void	nvme_dumpregs(struct nvme_softc *);
     56 #endif
     57 static int	nvme_identify(struct nvme_softc *, u_int);
     58 static void	nvme_fill_identify(struct nvme_queue *, struct nvme_ccb *,
     59 		    void *);
     60 
     61 static int	nvme_ccbs_alloc(struct nvme_queue *, uint16_t);
     62 static void	nvme_ccbs_free(struct nvme_queue *);
     63 
     64 static struct nvme_ccb *
     65 		nvme_ccb_get(struct nvme_queue *);
     66 static void	nvme_ccb_put(struct nvme_queue *, struct nvme_ccb *);
     67 
     68 static int	nvme_poll(struct nvme_softc *, struct nvme_queue *,
     69 		    struct nvme_ccb *, void (*)(struct nvme_queue *,
     70 		    struct nvme_ccb *, void *), int);
     71 static void	nvme_poll_fill(struct nvme_queue *, struct nvme_ccb *, void *);
     72 static void	nvme_poll_done(struct nvme_queue *, struct nvme_ccb *,
     73 		    struct nvme_cqe *);
     74 static void	nvme_sqe_fill(struct nvme_queue *, struct nvme_ccb *, void *);
     75 static void	nvme_empty_done(struct nvme_queue *, struct nvme_ccb *,
     76 		    struct nvme_cqe *);
     77 
     78 static struct nvme_queue *
     79 		nvme_q_alloc(struct nvme_softc *, uint16_t, u_int, u_int);
     80 static int	nvme_q_create(struct nvme_softc *, struct nvme_queue *);
     81 static int	nvme_q_delete(struct nvme_softc *, struct nvme_queue *);
     82 static void	nvme_q_submit(struct nvme_softc *, struct nvme_queue *,
     83 		    struct nvme_ccb *, void (*)(struct nvme_queue *,
     84 		    struct nvme_ccb *, void *));
     85 static int	nvme_q_complete(struct nvme_softc *, struct nvme_queue *q);
     86 static void	nvme_q_free(struct nvme_softc *, struct nvme_queue *);
     87 
     88 static struct nvme_dmamem *
     89 		nvme_dmamem_alloc(struct nvme_softc *, size_t);
     90 static void	nvme_dmamem_free(struct nvme_softc *, struct nvme_dmamem *);
     91 static void	nvme_dmamem_sync(struct nvme_softc *, struct nvme_dmamem *,
     92 		    int);
     93 
     94 static void	nvme_ns_io_fill(struct nvme_queue *, struct nvme_ccb *,
     95 		    void *);
     96 static void	nvme_ns_io_done(struct nvme_queue *, struct nvme_ccb *,
     97 		    struct nvme_cqe *);
     98 static void	nvme_ns_sync_fill(struct nvme_queue *, struct nvme_ccb *,
     99 		    void *);
    100 static void	nvme_ns_sync_done(struct nvme_queue *, struct nvme_ccb *,
    101 		    struct nvme_cqe *);
    102 static void	nvme_getcache_fill(struct nvme_queue *, struct nvme_ccb *,
    103 		    void *);
    104 static void	nvme_getcache_done(struct nvme_queue *, struct nvme_ccb *,
    105 		    struct nvme_cqe *);
    106 
    107 static void	nvme_pt_fill(struct nvme_queue *, struct nvme_ccb *,
    108 		    void *);
    109 static void	nvme_pt_done(struct nvme_queue *, struct nvme_ccb *,
    110 		    struct nvme_cqe *);
    111 static int	nvme_command_passthrough(struct nvme_softc *,
    112 		    struct nvme_pt_command *, uint16_t, struct lwp *, bool);
    113 
    114 static int	nvme_get_number_of_queues(struct nvme_softc *, u_int *);
    115 
    116 #define NVME_TIMO_QOP		5	/* queue create and delete timeout */
    117 #define NVME_TIMO_IDENT		10	/* probe identify timeout */
    118 #define NVME_TIMO_PT		-1	/* passthrough cmd timeout */
    119 #define NVME_TIMO_SY		60	/* sync cache timeout */
    120 
    121 #define nvme_read4(_s, _r) \
    122 	bus_space_read_4((_s)->sc_iot, (_s)->sc_ioh, (_r))
    123 #define nvme_write4(_s, _r, _v) \
    124 	bus_space_write_4((_s)->sc_iot, (_s)->sc_ioh, (_r), (_v))
    125 #ifdef __LP64__
    126 #define nvme_read8(_s, _r) \
    127 	bus_space_read_8((_s)->sc_iot, (_s)->sc_ioh, (_r))
    128 #define nvme_write8(_s, _r, _v) \
    129 	bus_space_write_8((_s)->sc_iot, (_s)->sc_ioh, (_r), (_v))
    130 #else /* __LP64__ */
    131 static inline uint64_t
    132 nvme_read8(struct nvme_softc *sc, bus_size_t r)
    133 {
    134 	uint64_t v;
    135 	uint32_t *a = (uint32_t *)&v;
    136 
    137 #if _BYTE_ORDER == _LITTLE_ENDIAN
    138 	a[0] = nvme_read4(sc, r);
    139 	a[1] = nvme_read4(sc, r + 4);
    140 #else /* _BYTE_ORDER == _LITTLE_ENDIAN */
    141 	a[1] = nvme_read4(sc, r);
    142 	a[0] = nvme_read4(sc, r + 4);
    143 #endif
    144 
    145 	return v;
    146 }
    147 
    148 static inline void
    149 nvme_write8(struct nvme_softc *sc, bus_size_t r, uint64_t v)
    150 {
    151 	uint32_t *a = (uint32_t *)&v;
    152 
    153 #if _BYTE_ORDER == _LITTLE_ENDIAN
    154 	nvme_write4(sc, r, a[0]);
    155 	nvme_write4(sc, r + 4, a[1]);
    156 #else /* _BYTE_ORDER == _LITTLE_ENDIAN */
    157 	nvme_write4(sc, r, a[1]);
    158 	nvme_write4(sc, r + 4, a[0]);
    159 #endif
    160 }
    161 #endif /* __LP64__ */
    162 #define nvme_barrier(_s, _r, _l, _f) \
    163 	bus_space_barrier((_s)->sc_iot, (_s)->sc_ioh, (_r), (_l), (_f))
    164 
    165 static void
    166 nvme_version(struct nvme_softc *sc, uint32_t ver)
    167 {
    168 	const char *v = NULL;
    169 
    170 	switch (ver) {
    171 	case NVME_VS_1_0:
    172 		v = "1.0";
    173 		break;
    174 	case NVME_VS_1_1:
    175 		v = "1.1";
    176 		break;
    177 	case NVME_VS_1_2:
    178 		v = "1.2";
    179 		break;
    180 	case NVME_VS_1_2_1:
    181 		v = "1.2.1";
    182 		break;
    183 	default:
    184 		aprint_error_dev(sc->sc_dev, "unknown version 0x%08x\n", ver);
    185 		return;
    186 	}
    187 
    188 	aprint_normal_dev(sc->sc_dev, "NVMe %s\n", v);
    189 }
    190 
    191 #ifdef NVME_DEBUG
    192 static __used void
    193 nvme_dumpregs(struct nvme_softc *sc)
    194 {
    195 	uint64_t r8;
    196 	uint32_t r4;
    197 
    198 #define	DEVNAME(_sc) device_xname((_sc)->sc_dev)
    199 	r8 = nvme_read8(sc, NVME_CAP);
    200 	printf("%s: cap  0x%016"PRIx64"\n", DEVNAME(sc), nvme_read8(sc, NVME_CAP));
    201 	printf("%s:  mpsmax %u (%u)\n", DEVNAME(sc),
    202 	    (u_int)NVME_CAP_MPSMAX(r8), (1 << NVME_CAP_MPSMAX(r8)));
    203 	printf("%s:  mpsmin %u (%u)\n", DEVNAME(sc),
    204 	    (u_int)NVME_CAP_MPSMIN(r8), (1 << NVME_CAP_MPSMIN(r8)));
    205 	printf("%s:  css %"PRIu64"\n", DEVNAME(sc), NVME_CAP_CSS(r8));
    206 	printf("%s:  nssrs %"PRIu64"\n", DEVNAME(sc), NVME_CAP_NSSRS(r8));
    207 	printf("%s:  dstrd %"PRIu64"\n", DEVNAME(sc), NVME_CAP_DSTRD(r8));
    208 	printf("%s:  to %"PRIu64" msec\n", DEVNAME(sc), NVME_CAP_TO(r8));
    209 	printf("%s:  ams %"PRIu64"\n", DEVNAME(sc), NVME_CAP_AMS(r8));
    210 	printf("%s:  cqr %"PRIu64"\n", DEVNAME(sc), NVME_CAP_CQR(r8));
    211 	printf("%s:  mqes %"PRIu64"\n", DEVNAME(sc), NVME_CAP_MQES(r8));
    212 
    213 	printf("%s: vs   0x%04x\n", DEVNAME(sc), nvme_read4(sc, NVME_VS));
    214 
    215 	r4 = nvme_read4(sc, NVME_CC);
    216 	printf("%s: cc   0x%04x\n", DEVNAME(sc), r4);
    217 	printf("%s:  iocqes %u (%u)\n", DEVNAME(sc), NVME_CC_IOCQES_R(r4),
    218 	    (1 << NVME_CC_IOCQES_R(r4)));
    219 	printf("%s:  iosqes %u (%u)\n", DEVNAME(sc), NVME_CC_IOSQES_R(r4),
    220 	    (1 << NVME_CC_IOSQES_R(r4)));
    221 	printf("%s:  shn %u\n", DEVNAME(sc), NVME_CC_SHN_R(r4));
    222 	printf("%s:  ams %u\n", DEVNAME(sc), NVME_CC_AMS_R(r4));
    223 	printf("%s:  mps %u (%u)\n", DEVNAME(sc), NVME_CC_MPS_R(r4),
    224 	    (1 << NVME_CC_MPS_R(r4)));
    225 	printf("%s:  css %u\n", DEVNAME(sc), NVME_CC_CSS_R(r4));
    226 	printf("%s:  en %u\n", DEVNAME(sc), ISSET(r4, NVME_CC_EN) ? 1 : 0);
    227 
    228 	r4 = nvme_read4(sc, NVME_CSTS);
    229 	printf("%s: csts 0x%08x\n", DEVNAME(sc), r4);
    230 	printf("%s:  rdy %u\n", DEVNAME(sc), r4 & NVME_CSTS_RDY);
    231 	printf("%s:  cfs %u\n", DEVNAME(sc), r4 & NVME_CSTS_CFS);
    232 	printf("%s:  shst %x\n", DEVNAME(sc), r4 & NVME_CSTS_SHST_MASK);
    233 
    234 	r4 = nvme_read4(sc, NVME_AQA);
    235 	printf("%s: aqa  0x%08x\n", DEVNAME(sc), r4);
    236 	printf("%s:  acqs %u\n", DEVNAME(sc), NVME_AQA_ACQS_R(r4));
    237 	printf("%s:  asqs %u\n", DEVNAME(sc), NVME_AQA_ASQS_R(r4));
    238 
    239 	printf("%s: asq  0x%016"PRIx64"\n", DEVNAME(sc), nvme_read8(sc, NVME_ASQ));
    240 	printf("%s: acq  0x%016"PRIx64"\n", DEVNAME(sc), nvme_read8(sc, NVME_ACQ));
    241 #undef	DEVNAME
    242 }
    243 #endif	/* NVME_DEBUG */
    244 
    245 static int
    246 nvme_ready(struct nvme_softc *sc, uint32_t rdy)
    247 {
    248 	u_int i = 0;
    249 	uint32_t cc;
    250 
    251 	cc = nvme_read4(sc, NVME_CC);
    252 	if (((cc & NVME_CC_EN) != 0) != (rdy != 0)) {
    253 		aprint_error_dev(sc->sc_dev,
    254 		    "controller enabled status expected %d, found to be %d\n",
    255 		    (rdy != 0), ((cc & NVME_CC_EN) != 0));
    256 		return ENXIO;
    257 	}
    258 
    259 	while ((nvme_read4(sc, NVME_CSTS) & NVME_CSTS_RDY) != rdy) {
    260 		if (i++ > sc->sc_rdy_to)
    261 			return ENXIO;
    262 
    263 		delay(1000);
    264 		nvme_barrier(sc, NVME_CSTS, 4, BUS_SPACE_BARRIER_READ);
    265 	}
    266 
    267 	return 0;
    268 }
    269 
    270 static int
    271 nvme_enable(struct nvme_softc *sc, u_int mps)
    272 {
    273 	uint32_t cc, csts;
    274 
    275 	cc = nvme_read4(sc, NVME_CC);
    276 	csts = nvme_read4(sc, NVME_CSTS);
    277 
    278 	if (ISSET(cc, NVME_CC_EN)) {
    279 		aprint_error_dev(sc->sc_dev, "controller unexpectedly enabled, failed to stay disabled\n");
    280 
    281 		if (ISSET(csts, NVME_CSTS_RDY))
    282 			return 1;
    283 
    284 		goto waitready;
    285 	}
    286 
    287 	nvme_write8(sc, NVME_ASQ, NVME_DMA_DVA(sc->sc_admin_q->q_sq_dmamem));
    288 	nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE);
    289 	delay(5000);
    290 	nvme_write8(sc, NVME_ACQ, NVME_DMA_DVA(sc->sc_admin_q->q_cq_dmamem));
    291 	nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE);
    292 	delay(5000);
    293 
    294 	nvme_write4(sc, NVME_AQA, NVME_AQA_ACQS(sc->sc_admin_q->q_entries) |
    295 	    NVME_AQA_ASQS(sc->sc_admin_q->q_entries));
    296 	nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE);
    297 	delay(5000);
    298 
    299 	CLR(cc, NVME_CC_IOCQES_MASK | NVME_CC_IOSQES_MASK | NVME_CC_SHN_MASK |
    300 	    NVME_CC_AMS_MASK | NVME_CC_MPS_MASK | NVME_CC_CSS_MASK);
    301 	SET(cc, NVME_CC_IOSQES(ffs(64) - 1) | NVME_CC_IOCQES(ffs(16) - 1));
    302 	SET(cc, NVME_CC_SHN(NVME_CC_SHN_NONE));
    303 	SET(cc, NVME_CC_CSS(NVME_CC_CSS_NVM));
    304 	SET(cc, NVME_CC_AMS(NVME_CC_AMS_RR));
    305 	SET(cc, NVME_CC_MPS(mps));
    306 	SET(cc, NVME_CC_EN);
    307 
    308 	nvme_write4(sc, NVME_CC, cc);
    309 	nvme_barrier(sc, 0, sc->sc_ios,
    310 	    BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
    311 	delay(5000);
    312 
    313     waitready:
    314 	return nvme_ready(sc, NVME_CSTS_RDY);
    315 }
    316 
    317 static int
    318 nvme_disable(struct nvme_softc *sc)
    319 {
    320 	uint32_t cc, csts;
    321 
    322 	cc = nvme_read4(sc, NVME_CC);
    323 	csts = nvme_read4(sc, NVME_CSTS);
    324 
    325 	if (ISSET(cc, NVME_CC_EN) && !ISSET(csts, NVME_CSTS_RDY))
    326 		nvme_ready(sc, NVME_CSTS_RDY);
    327 
    328 	CLR(cc, NVME_CC_EN);
    329 
    330 	nvme_write4(sc, NVME_CC, cc);
    331 	nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_READ);
    332 
    333 	delay(5000);
    334 
    335 	return nvme_ready(sc, 0);
    336 }
    337 
    338 int
    339 nvme_attach(struct nvme_softc *sc)
    340 {
    341 	uint64_t cap;
    342 	uint32_t reg;
    343 	u_int dstrd;
    344 	u_int mps = PAGE_SHIFT;
    345 	u_int ioq_allocated;
    346 	uint16_t adminq_entries = nvme_adminq_size;
    347 	uint16_t ioq_entries = nvme_ioq_size;
    348 	int i;
    349 
    350 	reg = nvme_read4(sc, NVME_VS);
    351 	if (reg == 0xffffffff) {
    352 		aprint_error_dev(sc->sc_dev, "invalid mapping\n");
    353 		return 1;
    354 	}
    355 
    356 	nvme_version(sc, reg);
    357 
    358 	cap = nvme_read8(sc, NVME_CAP);
    359 	dstrd = NVME_CAP_DSTRD(cap);
    360 	if (NVME_CAP_MPSMIN(cap) > PAGE_SHIFT) {
    361 		aprint_error_dev(sc->sc_dev, "NVMe minimum page size %u "
    362 		    "is greater than CPU page size %u\n",
    363 		    1 << NVME_CAP_MPSMIN(cap), 1 << PAGE_SHIFT);
    364 		return 1;
    365 	}
    366 	if (NVME_CAP_MPSMAX(cap) < mps)
    367 		mps = NVME_CAP_MPSMAX(cap);
    368 	if (ioq_entries > NVME_CAP_MQES(cap))
    369 		ioq_entries = NVME_CAP_MQES(cap);
    370 
    371 	/* set initial values to be used for admin queue during probe */
    372 	sc->sc_rdy_to = NVME_CAP_TO(cap);
    373 	sc->sc_mps = 1 << mps;
    374 	sc->sc_mdts = MAXPHYS;
    375 	sc->sc_max_sgl = 2;
    376 
    377 	if (nvme_disable(sc) != 0) {
    378 		aprint_error_dev(sc->sc_dev, "unable to disable controller\n");
    379 		return 1;
    380 	}
    381 
    382 	sc->sc_admin_q = nvme_q_alloc(sc, NVME_ADMIN_Q, adminq_entries, dstrd);
    383 	if (sc->sc_admin_q == NULL) {
    384 		aprint_error_dev(sc->sc_dev,
    385 		    "unable to allocate admin queue\n");
    386 		return 1;
    387 	}
    388 	if (sc->sc_intr_establish(sc, NVME_ADMIN_Q, sc->sc_admin_q))
    389 		goto free_admin_q;
    390 
    391 	if (nvme_enable(sc, mps) != 0) {
    392 		aprint_error_dev(sc->sc_dev, "unable to enable controller\n");
    393 		goto disestablish_admin_q;
    394 	}
    395 
    396 	if (nvme_identify(sc, NVME_CAP_MPSMIN(cap)) != 0) {
    397 		aprint_error_dev(sc->sc_dev, "unable to identify controller\n");
    398 		goto disable;
    399 	}
    400 
    401 	/* we know how big things are now */
    402 	sc->sc_max_sgl = sc->sc_mdts / sc->sc_mps;
    403 
    404 	/* reallocate ccbs of admin queue with new max sgl. */
    405 	nvme_ccbs_free(sc->sc_admin_q);
    406 	nvme_ccbs_alloc(sc->sc_admin_q, sc->sc_admin_q->q_entries);
    407 
    408 	if (sc->sc_use_mq) {
    409 		/* Limit the number of queues to the number allocated in HW */
    410 		if (nvme_get_number_of_queues(sc, &ioq_allocated) != 0) {
    411 			aprint_error_dev(sc->sc_dev,
    412 			    "unable to get number of queues\n");
    413 			goto disable;
    414 		}
    415 		if (sc->sc_nq > ioq_allocated)
    416 			sc->sc_nq = ioq_allocated;
    417 	}
    418 
    419 	sc->sc_q = kmem_zalloc(sizeof(*sc->sc_q) * sc->sc_nq, KM_SLEEP);
    420 	if (sc->sc_q == NULL) {
    421 		aprint_error_dev(sc->sc_dev, "unable to allocate io queue\n");
    422 		goto disable;
    423 	}
    424 	for (i = 0; i < sc->sc_nq; i++) {
    425 		sc->sc_q[i] = nvme_q_alloc(sc, i + 1, ioq_entries, dstrd);
    426 		if (sc->sc_q[i] == NULL) {
    427 			aprint_error_dev(sc->sc_dev,
    428 			    "unable to allocate io queue\n");
    429 			goto free_q;
    430 		}
    431 		if (nvme_q_create(sc, sc->sc_q[i]) != 0) {
    432 			aprint_error_dev(sc->sc_dev,
    433 			    "unable to create io queue\n");
    434 			nvme_q_free(sc, sc->sc_q[i]);
    435 			goto free_q;
    436 		}
    437 	}
    438 
    439 	if (!sc->sc_use_mq)
    440 		nvme_write4(sc, NVME_INTMC, 1);
    441 
    442 	/* probe subdevices */
    443 	sc->sc_namespaces = kmem_zalloc(sizeof(*sc->sc_namespaces) * sc->sc_nn,
    444 	    KM_SLEEP);
    445 	if (sc->sc_namespaces == NULL)
    446 		goto free_q;
    447 	nvme_rescan(sc->sc_dev, "nvme", &i);
    448 
    449 	return 0;
    450 
    451 free_q:
    452 	while (--i >= 0) {
    453 		nvme_q_delete(sc, sc->sc_q[i]);
    454 		nvme_q_free(sc, sc->sc_q[i]);
    455 	}
    456 disable:
    457 	nvme_disable(sc);
    458 disestablish_admin_q:
    459 	sc->sc_intr_disestablish(sc, NVME_ADMIN_Q);
    460 free_admin_q:
    461 	nvme_q_free(sc, sc->sc_admin_q);
    462 
    463 	return 1;
    464 }
    465 
    466 int
    467 nvme_rescan(device_t self, const char *attr, const int *flags)
    468 {
    469 	struct nvme_softc *sc = device_private(self);
    470 	struct nvme_attach_args naa;
    471 	uint64_t cap;
    472 	int ioq_entries = nvme_ioq_size;
    473 	int i;
    474 
    475 	cap = nvme_read8(sc, NVME_CAP);
    476 	if (ioq_entries > NVME_CAP_MQES(cap))
    477 		ioq_entries = NVME_CAP_MQES(cap);
    478 
    479 	for (i = 0; i < sc->sc_nn; i++) {
    480 		if (sc->sc_namespaces[i].dev)
    481 			continue;
    482 		memset(&naa, 0, sizeof(naa));
    483 		naa.naa_nsid = i + 1;
    484 		naa.naa_qentries = (ioq_entries - 1) * sc->sc_nq;
    485 		naa.naa_maxphys = sc->sc_mdts;
    486 		sc->sc_namespaces[i].dev = config_found(sc->sc_dev, &naa,
    487 		    nvme_print);
    488 	}
    489 	return 0;
    490 }
    491 
    492 static int
    493 nvme_print(void *aux, const char *pnp)
    494 {
    495 	struct nvme_attach_args *naa = aux;
    496 
    497 	if (pnp)
    498 		aprint_normal("at %s", pnp);
    499 
    500 	if (naa->naa_nsid > 0)
    501 		aprint_normal(" nsid %d", naa->naa_nsid);
    502 
    503 	return UNCONF;
    504 }
    505 
    506 int
    507 nvme_detach(struct nvme_softc *sc, int flags)
    508 {
    509 	int i, error;
    510 
    511 	error = config_detach_children(sc->sc_dev, flags);
    512 	if (error)
    513 		return error;
    514 
    515 	error = nvme_shutdown(sc);
    516 	if (error)
    517 		return error;
    518 
    519 	/* from now on we are committed to detach, following will never fail */
    520 	for (i = 0; i < sc->sc_nq; i++)
    521 		nvme_q_free(sc, sc->sc_q[i]);
    522 	kmem_free(sc->sc_q, sizeof(*sc->sc_q) * sc->sc_nq);
    523 	nvme_q_free(sc, sc->sc_admin_q);
    524 
    525 	return 0;
    526 }
    527 
    528 static int
    529 nvme_shutdown(struct nvme_softc *sc)
    530 {
    531 	uint32_t cc, csts;
    532 	bool disabled = false;
    533 	int i;
    534 
    535 	if (!sc->sc_use_mq)
    536 		nvme_write4(sc, NVME_INTMS, 1);
    537 
    538 	for (i = 0; i < sc->sc_nq; i++) {
    539 		if (nvme_q_delete(sc, sc->sc_q[i]) != 0) {
    540 			aprint_error_dev(sc->sc_dev,
    541 			    "unable to delete io queue %d, disabling\n", i + 1);
    542 			disabled = true;
    543 		}
    544 	}
    545 	sc->sc_intr_disestablish(sc, NVME_ADMIN_Q);
    546 	if (disabled)
    547 		goto disable;
    548 
    549 	cc = nvme_read4(sc, NVME_CC);
    550 	CLR(cc, NVME_CC_SHN_MASK);
    551 	SET(cc, NVME_CC_SHN(NVME_CC_SHN_NORMAL));
    552 	nvme_write4(sc, NVME_CC, cc);
    553 
    554 	for (i = 0; i < 4000; i++) {
    555 		nvme_barrier(sc, 0, sc->sc_ios,
    556 		    BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
    557 		csts = nvme_read4(sc, NVME_CSTS);
    558 		if ((csts & NVME_CSTS_SHST_MASK) == NVME_CSTS_SHST_DONE)
    559 			return 0;
    560 
    561 		delay(1000);
    562 	}
    563 
    564 	aprint_error_dev(sc->sc_dev, "unable to shudown, disabling\n");
    565 
    566 disable:
    567 	nvme_disable(sc);
    568 	return 0;
    569 }
    570 
    571 void
    572 nvme_childdet(device_t self, device_t child)
    573 {
    574 	struct nvme_softc *sc = device_private(self);
    575 	int i;
    576 
    577 	for (i = 0; i < sc->sc_nn; i++) {
    578 		if (sc->sc_namespaces[i].dev == child) {
    579 			/* Already freed ns->ident. */
    580 			sc->sc_namespaces[i].dev = NULL;
    581 			break;
    582 		}
    583 	}
    584 }
    585 
    586 int
    587 nvme_ns_identify(struct nvme_softc *sc, uint16_t nsid)
    588 {
    589 	struct nvme_sqe sqe;
    590 	struct nvm_identify_namespace *identify;
    591 	struct nvme_dmamem *mem;
    592 	struct nvme_ccb *ccb;
    593 	struct nvme_namespace *ns;
    594 	int rv;
    595 
    596 	KASSERT(nsid > 0);
    597 
    598 	ccb = nvme_ccb_get(sc->sc_admin_q);
    599 	KASSERT(ccb != NULL); /* it's a bug if we don't have spare ccb here */
    600 
    601 	mem = nvme_dmamem_alloc(sc, sizeof(*identify));
    602 	if (mem == NULL)
    603 		return ENOMEM;
    604 
    605 	memset(&sqe, 0, sizeof(sqe));
    606 	sqe.opcode = NVM_ADMIN_IDENTIFY;
    607 	htolem32(&sqe.nsid, nsid);
    608 	htolem64(&sqe.entry.prp[0], NVME_DMA_DVA(mem));
    609 	htolem32(&sqe.cdw10, 0);
    610 
    611 	ccb->ccb_done = nvme_empty_done;
    612 	ccb->ccb_cookie = &sqe;
    613 
    614 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_PREREAD);
    615 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_IDENT);
    616 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_POSTREAD);
    617 
    618 	nvme_ccb_put(sc->sc_admin_q, ccb);
    619 
    620 	if (rv != 0) {
    621 		rv = EIO;
    622 		goto done;
    623 	}
    624 
    625 	/* commit */
    626 
    627 	identify = kmem_zalloc(sizeof(*identify), KM_SLEEP);
    628 	*identify = *((volatile struct nvm_identify_namespace *)NVME_DMA_KVA(mem));
    629 	//memcpy(identify, NVME_DMA_KVA(mem), sizeof(*identify));
    630 
    631 	ns = nvme_ns_get(sc, nsid);
    632 	KASSERT(ns);
    633 	ns->ident = identify;
    634 
    635 done:
    636 	nvme_dmamem_free(sc, mem);
    637 
    638 	return rv;
    639 }
    640 
    641 int
    642 nvme_ns_dobio(struct nvme_softc *sc, uint16_t nsid, void *cookie,
    643     struct buf *bp, void *data, size_t datasize,
    644     int secsize, daddr_t blkno, int flags, nvme_nnc_done nnc_done)
    645 {
    646 	struct nvme_queue *q = nvme_get_q(sc);
    647 	struct nvme_ccb *ccb;
    648 	bus_dmamap_t dmap;
    649 	int i, error;
    650 
    651 	ccb = nvme_ccb_get(q);
    652 	if (ccb == NULL)
    653 		return EAGAIN;
    654 
    655 	ccb->ccb_done = nvme_ns_io_done;
    656 	ccb->ccb_cookie = cookie;
    657 
    658 	/* namespace context */
    659 	ccb->nnc_nsid = nsid;
    660 	ccb->nnc_flags = flags;
    661 	ccb->nnc_buf = bp;
    662 	ccb->nnc_datasize = datasize;
    663 	ccb->nnc_secsize = secsize;
    664 	ccb->nnc_blkno = blkno;
    665 	ccb->nnc_done = nnc_done;
    666 
    667 	dmap = ccb->ccb_dmamap;
    668 	error = bus_dmamap_load(sc->sc_dmat, dmap, data,
    669 	    datasize, NULL,
    670 	    (ISSET(flags, NVME_NS_CTX_F_POLL) ?
    671 	      BUS_DMA_NOWAIT : BUS_DMA_WAITOK) |
    672 	    (ISSET(flags, NVME_NS_CTX_F_READ) ?
    673 	      BUS_DMA_READ : BUS_DMA_WRITE));
    674 	if (error) {
    675 		nvme_ccb_put(q, ccb);
    676 		return error;
    677 	}
    678 
    679 	bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize,
    680 	    ISSET(flags, NVME_NS_CTX_F_READ) ?
    681 	    BUS_DMASYNC_PREREAD : BUS_DMASYNC_PREWRITE);
    682 
    683 	if (dmap->dm_nsegs > 2) {
    684 		for (i = 1; i < dmap->dm_nsegs; i++) {
    685 			htolem64(&ccb->ccb_prpl[i - 1],
    686 			    dmap->dm_segs[i].ds_addr);
    687 		}
    688 		bus_dmamap_sync(sc->sc_dmat,
    689 		    NVME_DMA_MAP(q->q_ccb_prpls),
    690 		    ccb->ccb_prpl_off,
    691 		    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    692 		    BUS_DMASYNC_PREWRITE);
    693 	}
    694 
    695 	if (ISSET(flags, NVME_NS_CTX_F_POLL)) {
    696 		if (nvme_poll(sc, q, ccb, nvme_ns_io_fill, NVME_TIMO_PT) != 0)
    697 			return EIO;
    698 		return 0;
    699 	}
    700 
    701 	nvme_q_submit(sc, q, ccb, nvme_ns_io_fill);
    702 	return 0;
    703 }
    704 
    705 static void
    706 nvme_ns_io_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    707 {
    708 	struct nvme_sqe_io *sqe = slot;
    709 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    710 
    711 	sqe->opcode = ISSET(ccb->nnc_flags, NVME_NS_CTX_F_READ) ?
    712 	    NVM_CMD_READ : NVM_CMD_WRITE;
    713 	htolem32(&sqe->nsid, ccb->nnc_nsid);
    714 
    715 	htolem64(&sqe->entry.prp[0], dmap->dm_segs[0].ds_addr);
    716 	switch (dmap->dm_nsegs) {
    717 	case 1:
    718 		break;
    719 	case 2:
    720 		htolem64(&sqe->entry.prp[1], dmap->dm_segs[1].ds_addr);
    721 		break;
    722 	default:
    723 		/* the prp list is already set up and synced */
    724 		htolem64(&sqe->entry.prp[1], ccb->ccb_prpl_dva);
    725 		break;
    726 	}
    727 
    728 	htolem64(&sqe->slba, ccb->nnc_blkno);
    729 
    730 	if (ISSET(ccb->nnc_flags, NVME_NS_CTX_F_FUA))
    731 		htolem16(&sqe->ioflags, NVM_SQE_IO_FUA);
    732 
    733 	/* guaranteed by upper layers, but check just in case */
    734 	KASSERT((ccb->nnc_datasize % ccb->nnc_secsize) == 0);
    735 	htolem16(&sqe->nlb, (ccb->nnc_datasize / ccb->nnc_secsize) - 1);
    736 }
    737 
    738 static void
    739 nvme_ns_io_done(struct nvme_queue *q, struct nvme_ccb *ccb,
    740     struct nvme_cqe *cqe)
    741 {
    742 	struct nvme_softc *sc = q->q_sc;
    743 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    744 	void *nnc_cookie = ccb->ccb_cookie;
    745 	nvme_nnc_done nnc_done = ccb->nnc_done;
    746 	struct buf *bp = ccb->nnc_buf;
    747 
    748 	if (dmap->dm_nsegs > 2) {
    749 		bus_dmamap_sync(sc->sc_dmat,
    750 		    NVME_DMA_MAP(q->q_ccb_prpls),
    751 		    ccb->ccb_prpl_off,
    752 		    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    753 		    BUS_DMASYNC_POSTWRITE);
    754 	}
    755 
    756 	bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize,
    757 	    ISSET(ccb->nnc_flags, NVME_NS_CTX_F_READ) ?
    758 	    BUS_DMASYNC_POSTREAD : BUS_DMASYNC_POSTWRITE);
    759 
    760 	bus_dmamap_unload(sc->sc_dmat, dmap);
    761 	nvme_ccb_put(q, ccb);
    762 
    763 	nnc_done(nnc_cookie, bp, lemtoh16(&cqe->flags), lemtoh32(&cqe->cdw0));
    764 }
    765 
    766 /*
    767  * If there is no volatile write cache, it makes no sense to issue
    768  * flush commands or query for the status.
    769  */
    770 bool
    771 nvme_has_volatile_write_cache(struct nvme_softc *sc)
    772 {
    773 	/* sc_identify is filled during attachment */
    774 	return  ((sc->sc_identify.vwc & NVME_ID_CTRLR_VWC_PRESENT) != 0);
    775 }
    776 
    777 int
    778 nvme_ns_sync(struct nvme_softc *sc, uint16_t nsid, void *cookie,
    779     int flags, nvme_nnc_done nnc_done)
    780 {
    781 	struct nvme_queue *q = nvme_get_q(sc);
    782 	struct nvme_ccb *ccb;
    783 
    784 	ccb = nvme_ccb_get(q);
    785 	if (ccb == NULL)
    786 		return EAGAIN;
    787 
    788 	ccb->ccb_done = nvme_ns_sync_done;
    789 	ccb->ccb_cookie = cookie;
    790 
    791 	/* namespace context */
    792 	ccb->nnc_nsid = nsid;
    793 	ccb->nnc_flags = flags;
    794 	ccb->nnc_done = nnc_done;
    795 
    796 	if (ISSET(flags, NVME_NS_CTX_F_POLL)) {
    797 		if (nvme_poll(sc, q, ccb, nvme_ns_sync_fill, NVME_TIMO_SY) != 0)
    798 			return EIO;
    799 		return 0;
    800 	}
    801 
    802 	nvme_q_submit(sc, q, ccb, nvme_ns_sync_fill);
    803 	return 0;
    804 }
    805 
    806 static void
    807 nvme_ns_sync_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    808 {
    809 	struct nvme_sqe *sqe = slot;
    810 
    811 	sqe->opcode = NVM_CMD_FLUSH;
    812 	htolem32(&sqe->nsid, ccb->nnc_nsid);
    813 }
    814 
    815 static void
    816 nvme_ns_sync_done(struct nvme_queue *q, struct nvme_ccb *ccb,
    817     struct nvme_cqe *cqe)
    818 {
    819 	void *cookie = ccb->ccb_cookie;
    820 	nvme_nnc_done nnc_done = ccb->nnc_done;
    821 
    822 	nvme_ccb_put(q, ccb);
    823 
    824 	nnc_done(cookie, NULL, lemtoh16(&cqe->flags), lemtoh32(&cqe->cdw0));
    825 }
    826 
    827 /*
    828  * Get status of volatile write cache. Always asynchronous.
    829  */
    830 int
    831 nvme_admin_getcache(struct nvme_softc *sc, void *cookie, nvme_nnc_done nnc_done)
    832 {
    833 	struct nvme_ccb *ccb;
    834 	struct nvme_queue *q = sc->sc_admin_q;
    835 
    836 	ccb = nvme_ccb_get(q);
    837 	if (ccb == NULL)
    838 		return EAGAIN;
    839 
    840 	ccb->ccb_done = nvme_getcache_done;
    841 	ccb->ccb_cookie = cookie;
    842 
    843 	/* namespace context */
    844 	ccb->nnc_flags = 0;
    845 	ccb->nnc_done = nnc_done;
    846 
    847 	nvme_q_submit(sc, q, ccb, nvme_getcache_fill);
    848 	return 0;
    849 }
    850 
    851 static void
    852 nvme_getcache_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    853 {
    854 	struct nvme_sqe *sqe = slot;
    855 
    856 	sqe->opcode = NVM_ADMIN_GET_FEATURES;
    857 	sqe->cdw10 = NVM_FEATURE_VOLATILE_WRITE_CACHE;
    858 }
    859 
    860 static void
    861 nvme_getcache_done(struct nvme_queue *q, struct nvme_ccb *ccb,
    862     struct nvme_cqe *cqe)
    863 {
    864 	void *cookie = ccb->ccb_cookie;
    865 	nvme_nnc_done nnc_done = ccb->nnc_done;
    866 
    867 	nvme_ccb_put(q, ccb);
    868 
    869 	nnc_done(cookie, NULL, lemtoh16(&cqe->flags), lemtoh32(&cqe->cdw0));
    870 }
    871 
    872 void
    873 nvme_ns_free(struct nvme_softc *sc, uint16_t nsid)
    874 {
    875 	struct nvme_namespace *ns;
    876 	struct nvm_identify_namespace *identify;
    877 
    878 	ns = nvme_ns_get(sc, nsid);
    879 	KASSERT(ns);
    880 
    881 	identify = ns->ident;
    882 	ns->ident = NULL;
    883 	if (identify != NULL)
    884 		kmem_free(identify, sizeof(*identify));
    885 }
    886 
    887 static void
    888 nvme_pt_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    889 {
    890 	struct nvme_softc *sc = q->q_sc;
    891 	struct nvme_sqe *sqe = slot;
    892 	struct nvme_pt_command *pt = ccb->ccb_cookie;
    893 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    894 	int i;
    895 
    896 	sqe->opcode = pt->cmd.opcode;
    897 	htolem32(&sqe->nsid, pt->cmd.nsid);
    898 
    899 	if (pt->buf != NULL && pt->len > 0) {
    900 		htolem64(&sqe->entry.prp[0], dmap->dm_segs[0].ds_addr);
    901 		switch (dmap->dm_nsegs) {
    902 		case 1:
    903 			break;
    904 		case 2:
    905 			htolem64(&sqe->entry.prp[1], dmap->dm_segs[1].ds_addr);
    906 			break;
    907 		default:
    908 			for (i = 1; i < dmap->dm_nsegs; i++) {
    909 				htolem64(&ccb->ccb_prpl[i - 1],
    910 				    dmap->dm_segs[i].ds_addr);
    911 			}
    912 			bus_dmamap_sync(sc->sc_dmat,
    913 			    NVME_DMA_MAP(q->q_ccb_prpls),
    914 			    ccb->ccb_prpl_off,
    915 			    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    916 			    BUS_DMASYNC_PREWRITE);
    917 			htolem64(&sqe->entry.prp[1], ccb->ccb_prpl_dva);
    918 			break;
    919 		}
    920 	}
    921 
    922 	htolem32(&sqe->cdw10, pt->cmd.cdw10);
    923 	htolem32(&sqe->cdw11, pt->cmd.cdw11);
    924 	htolem32(&sqe->cdw12, pt->cmd.cdw12);
    925 	htolem32(&sqe->cdw13, pt->cmd.cdw13);
    926 	htolem32(&sqe->cdw14, pt->cmd.cdw14);
    927 	htolem32(&sqe->cdw15, pt->cmd.cdw15);
    928 }
    929 
    930 static void
    931 nvme_pt_done(struct nvme_queue *q, struct nvme_ccb *ccb, struct nvme_cqe *cqe)
    932 {
    933 	struct nvme_softc *sc = q->q_sc;
    934 	struct nvme_pt_command *pt = ccb->ccb_cookie;
    935 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    936 
    937 	if (pt->buf != NULL && pt->len > 0) {
    938 		if (dmap->dm_nsegs > 2) {
    939 			bus_dmamap_sync(sc->sc_dmat,
    940 			    NVME_DMA_MAP(q->q_ccb_prpls),
    941 			    ccb->ccb_prpl_off,
    942 			    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    943 			    BUS_DMASYNC_POSTWRITE);
    944 		}
    945 
    946 		bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize,
    947 		    pt->is_read ? BUS_DMASYNC_POSTREAD : BUS_DMASYNC_POSTWRITE);
    948 		bus_dmamap_unload(sc->sc_dmat, dmap);
    949 	}
    950 
    951 	pt->cpl.cdw0 = lemtoh32(&cqe->cdw0);
    952 	pt->cpl.flags = lemtoh16(&cqe->flags) & ~NVME_CQE_PHASE;
    953 }
    954 
    955 static int
    956 nvme_command_passthrough(struct nvme_softc *sc, struct nvme_pt_command *pt,
    957     uint16_t nsid, struct lwp *l, bool is_adminq)
    958 {
    959 	struct nvme_queue *q;
    960 	struct nvme_ccb *ccb;
    961 	void *buf = NULL;
    962 	int error;
    963 
    964 	/* limit command size to maximum data transfer size */
    965 	if ((pt->buf == NULL && pt->len > 0) ||
    966 	    (pt->buf != NULL && (pt->len == 0 || pt->len > sc->sc_mdts)))
    967 		return EINVAL;
    968 
    969 	q = is_adminq ? sc->sc_admin_q : nvme_get_q(sc);
    970 	ccb = nvme_ccb_get(q);
    971 	if (ccb == NULL)
    972 		return EBUSY;
    973 
    974 	if (pt->buf != NULL) {
    975 		KASSERT(pt->len > 0);
    976 		buf = kmem_alloc(pt->len, KM_SLEEP);
    977 		if (buf == NULL) {
    978 			error = ENOMEM;
    979 			goto ccb_put;
    980 		}
    981 		if (!pt->is_read) {
    982 			error = copyin(pt->buf, buf, pt->len);
    983 			if (error)
    984 				goto kmem_free;
    985 		}
    986 		error = bus_dmamap_load(sc->sc_dmat, ccb->ccb_dmamap, buf,
    987 		    pt->len, NULL,
    988 		    BUS_DMA_WAITOK |
    989 		      (pt->is_read ? BUS_DMA_READ : BUS_DMA_WRITE));
    990 		if (error)
    991 			goto kmem_free;
    992 		bus_dmamap_sync(sc->sc_dmat, ccb->ccb_dmamap,
    993 		    0, ccb->ccb_dmamap->dm_mapsize,
    994 		    pt->is_read ? BUS_DMASYNC_PREREAD : BUS_DMASYNC_PREWRITE);
    995 	}
    996 
    997 	ccb->ccb_done = nvme_pt_done;
    998 	ccb->ccb_cookie = pt;
    999 
   1000 	pt->cmd.nsid = nsid;
   1001 	if (nvme_poll(sc, q, ccb, nvme_pt_fill, NVME_TIMO_PT)) {
   1002 		error = EIO;
   1003 		goto out;
   1004 	}
   1005 
   1006 	error = 0;
   1007 out:
   1008 	if (buf != NULL) {
   1009 		if (error == 0 && pt->is_read)
   1010 			error = copyout(buf, pt->buf, pt->len);
   1011 kmem_free:
   1012 		kmem_free(buf, pt->len);
   1013 	}
   1014 ccb_put:
   1015 	nvme_ccb_put(q, ccb);
   1016 	return error;
   1017 }
   1018 
   1019 static void
   1020 nvme_q_submit(struct nvme_softc *sc, struct nvme_queue *q, struct nvme_ccb *ccb,
   1021     void (*fill)(struct nvme_queue *, struct nvme_ccb *, void *))
   1022 {
   1023 	struct nvme_sqe *sqe = NVME_DMA_KVA(q->q_sq_dmamem);
   1024 	uint32_t tail;
   1025 
   1026 	mutex_enter(&q->q_sq_mtx);
   1027 	tail = q->q_sq_tail;
   1028 	if (++q->q_sq_tail >= q->q_entries)
   1029 		q->q_sq_tail = 0;
   1030 
   1031 	sqe += tail;
   1032 
   1033 	bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem),
   1034 	    sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_POSTWRITE);
   1035 	memset(sqe, 0, sizeof(*sqe));
   1036 	(*fill)(q, ccb, sqe);
   1037 	sqe->cid = ccb->ccb_id;
   1038 	bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem),
   1039 	    sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_PREWRITE);
   1040 
   1041 	nvme_write4(sc, q->q_sqtdbl, q->q_sq_tail);
   1042 	mutex_exit(&q->q_sq_mtx);
   1043 }
   1044 
   1045 struct nvme_poll_state {
   1046 	struct nvme_sqe s;
   1047 	struct nvme_cqe c;
   1048 };
   1049 
   1050 static int
   1051 nvme_poll(struct nvme_softc *sc, struct nvme_queue *q, struct nvme_ccb *ccb,
   1052     void (*fill)(struct nvme_queue *, struct nvme_ccb *, void *), int timo_sec)
   1053 {
   1054 	struct nvme_poll_state state;
   1055 	void (*done)(struct nvme_queue *, struct nvme_ccb *, struct nvme_cqe *);
   1056 	void *cookie;
   1057 	uint16_t flags;
   1058 	int step = 10;
   1059 	int maxloop = timo_sec * 1000000 / step;
   1060 	int error = 0;
   1061 
   1062 	memset(&state, 0, sizeof(state));
   1063 	(*fill)(q, ccb, &state.s);
   1064 
   1065 	done = ccb->ccb_done;
   1066 	cookie = ccb->ccb_cookie;
   1067 
   1068 	ccb->ccb_done = nvme_poll_done;
   1069 	ccb->ccb_cookie = &state;
   1070 
   1071 	nvme_q_submit(sc, q, ccb, nvme_poll_fill);
   1072 	while (!ISSET(state.c.flags, htole16(NVME_CQE_PHASE))) {
   1073 		if (nvme_q_complete(sc, q) == 0)
   1074 			delay(step);
   1075 
   1076 		if (timo_sec >= 0 && --maxloop <= 0) {
   1077 			error = ETIMEDOUT;
   1078 			break;
   1079 		}
   1080 	}
   1081 
   1082 	ccb->ccb_cookie = cookie;
   1083 	done(q, ccb, &state.c);
   1084 
   1085 	if (error == 0) {
   1086 		flags = lemtoh16(&state.c.flags);
   1087 		return flags & ~NVME_CQE_PHASE;
   1088 	} else {
   1089 		return 1;
   1090 	}
   1091 }
   1092 
   1093 static void
   1094 nvme_poll_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
   1095 {
   1096 	struct nvme_sqe *sqe = slot;
   1097 	struct nvme_poll_state *state = ccb->ccb_cookie;
   1098 
   1099 	*sqe = state->s;
   1100 }
   1101 
   1102 static void
   1103 nvme_poll_done(struct nvme_queue *q, struct nvme_ccb *ccb,
   1104     struct nvme_cqe *cqe)
   1105 {
   1106 	struct nvme_poll_state *state = ccb->ccb_cookie;
   1107 
   1108 	SET(cqe->flags, htole16(NVME_CQE_PHASE));
   1109 	state->c = *cqe;
   1110 }
   1111 
   1112 static void
   1113 nvme_sqe_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
   1114 {
   1115 	struct nvme_sqe *src = ccb->ccb_cookie;
   1116 	struct nvme_sqe *dst = slot;
   1117 
   1118 	*dst = *src;
   1119 }
   1120 
   1121 static void
   1122 nvme_empty_done(struct nvme_queue *q, struct nvme_ccb *ccb,
   1123     struct nvme_cqe *cqe)
   1124 {
   1125 }
   1126 
   1127 static int
   1128 nvme_q_complete(struct nvme_softc *sc, struct nvme_queue *q)
   1129 {
   1130 	struct nvme_ccb *ccb;
   1131 	struct nvme_cqe *ring = NVME_DMA_KVA(q->q_cq_dmamem), *cqe;
   1132 	uint16_t flags;
   1133 	int rv = 0;
   1134 
   1135 	mutex_enter(&q->q_cq_mtx);
   1136 
   1137 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_POSTREAD);
   1138 	for (;;) {
   1139 		cqe = &ring[q->q_cq_head];
   1140 		flags = lemtoh16(&cqe->flags);
   1141 		if ((flags & NVME_CQE_PHASE) != q->q_cq_phase)
   1142 			break;
   1143 
   1144 		ccb = &q->q_ccbs[cqe->cid];
   1145 
   1146 		if (++q->q_cq_head >= q->q_entries) {
   1147 			q->q_cq_head = 0;
   1148 			q->q_cq_phase ^= NVME_CQE_PHASE;
   1149 		}
   1150 
   1151 #ifdef DEBUG
   1152 		/*
   1153 		 * If we get spurious completion notification, something
   1154 		 * is seriously hosed up. Very likely DMA to some random
   1155 		 * memory place happened, so just bail out.
   1156 		 */
   1157 		if ((intptr_t)ccb->ccb_cookie == NVME_CCB_FREE) {
   1158 			panic("%s: invalid ccb detected",
   1159 			    device_xname(sc->sc_dev));
   1160 			/* NOTREACHED */
   1161 		}
   1162 #endif
   1163 
   1164 		rv++;
   1165 
   1166 		/*
   1167 		 * Unlock the mutex before calling the ccb_done callback
   1168 		 * and re-lock afterwards. The callback triggers lddone()
   1169 		 * which schedules another i/o, and also calls nvme_ccb_put().
   1170 		 * Unlock/relock avoids possibility of deadlock.
   1171 		 */
   1172 		mutex_exit(&q->q_cq_mtx);
   1173 		ccb->ccb_done(q, ccb, cqe);
   1174 		mutex_enter(&q->q_cq_mtx);
   1175 	}
   1176 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_PREREAD);
   1177 
   1178 	if (rv)
   1179 		nvme_write4(sc, q->q_cqhdbl, q->q_cq_head);
   1180 
   1181 	mutex_exit(&q->q_cq_mtx);
   1182 
   1183 	if (rv) {
   1184 		mutex_enter(&q->q_ccb_mtx);
   1185 		q->q_nccbs_avail += rv;
   1186 		mutex_exit(&q->q_ccb_mtx);
   1187 	}
   1188 
   1189 	return rv;
   1190 }
   1191 
   1192 static int
   1193 nvme_identify(struct nvme_softc *sc, u_int mps)
   1194 {
   1195 	char sn[41], mn[81], fr[17];
   1196 	struct nvm_identify_controller *identify;
   1197 	struct nvme_dmamem *mem;
   1198 	struct nvme_ccb *ccb;
   1199 	u_int mdts;
   1200 	int rv = 1;
   1201 
   1202 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1203 	KASSERT(ccb != NULL); /* it's a bug if we don't have spare ccb here */
   1204 
   1205 	mem = nvme_dmamem_alloc(sc, sizeof(*identify));
   1206 	if (mem == NULL)
   1207 		return 1;
   1208 
   1209 	ccb->ccb_done = nvme_empty_done;
   1210 	ccb->ccb_cookie = mem;
   1211 
   1212 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_PREREAD);
   1213 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_fill_identify,
   1214 	    NVME_TIMO_IDENT);
   1215 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_POSTREAD);
   1216 
   1217 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1218 
   1219 	if (rv != 0)
   1220 		goto done;
   1221 
   1222 	identify = NVME_DMA_KVA(mem);
   1223 
   1224 	strnvisx(sn, sizeof(sn), (const char *)identify->sn,
   1225 	    sizeof(identify->sn), VIS_TRIM|VIS_SAFE|VIS_OCTAL);
   1226 	strnvisx(mn, sizeof(mn), (const char *)identify->mn,
   1227 	    sizeof(identify->mn), VIS_TRIM|VIS_SAFE|VIS_OCTAL);
   1228 	strnvisx(fr, sizeof(fr), (const char *)identify->fr,
   1229 	    sizeof(identify->fr), VIS_TRIM|VIS_SAFE|VIS_OCTAL);
   1230 	aprint_normal_dev(sc->sc_dev, "%s, firmware %s, serial %s\n", mn, fr,
   1231 	    sn);
   1232 
   1233 	if (identify->mdts > 0) {
   1234 		mdts = (1 << identify->mdts) * (1 << mps);
   1235 		if (mdts < sc->sc_mdts)
   1236 			sc->sc_mdts = mdts;
   1237 	}
   1238 
   1239 	sc->sc_nn = lemtoh32(&identify->nn);
   1240 
   1241 	memcpy(&sc->sc_identify, identify, sizeof(sc->sc_identify));
   1242 
   1243 done:
   1244 	nvme_dmamem_free(sc, mem);
   1245 
   1246 	return rv;
   1247 }
   1248 
   1249 static int
   1250 nvme_q_create(struct nvme_softc *sc, struct nvme_queue *q)
   1251 {
   1252 	struct nvme_sqe_q sqe;
   1253 	struct nvme_ccb *ccb;
   1254 	int rv;
   1255 
   1256 	if (sc->sc_use_mq && sc->sc_intr_establish(sc, q->q_id, q) != 0)
   1257 		return 1;
   1258 
   1259 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1260 	KASSERT(ccb != NULL);
   1261 
   1262 	ccb->ccb_done = nvme_empty_done;
   1263 	ccb->ccb_cookie = &sqe;
   1264 
   1265 	memset(&sqe, 0, sizeof(sqe));
   1266 	sqe.opcode = NVM_ADMIN_ADD_IOCQ;
   1267 	htolem64(&sqe.prp1, NVME_DMA_DVA(q->q_cq_dmamem));
   1268 	htolem16(&sqe.qsize, q->q_entries - 1);
   1269 	htolem16(&sqe.qid, q->q_id);
   1270 	sqe.qflags = NVM_SQE_CQ_IEN | NVM_SQE_Q_PC;
   1271 	if (sc->sc_use_mq)
   1272 		htolem16(&sqe.cqid, q->q_id);	/* qid == vector */
   1273 
   1274 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1275 	if (rv != 0)
   1276 		goto fail;
   1277 
   1278 	ccb->ccb_done = nvme_empty_done;
   1279 	ccb->ccb_cookie = &sqe;
   1280 
   1281 	memset(&sqe, 0, sizeof(sqe));
   1282 	sqe.opcode = NVM_ADMIN_ADD_IOSQ;
   1283 	htolem64(&sqe.prp1, NVME_DMA_DVA(q->q_sq_dmamem));
   1284 	htolem16(&sqe.qsize, q->q_entries - 1);
   1285 	htolem16(&sqe.qid, q->q_id);
   1286 	htolem16(&sqe.cqid, q->q_id);
   1287 	sqe.qflags = NVM_SQE_Q_PC;
   1288 
   1289 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1290 	if (rv != 0)
   1291 		goto fail;
   1292 
   1293 fail:
   1294 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1295 	return rv;
   1296 }
   1297 
   1298 static int
   1299 nvme_q_delete(struct nvme_softc *sc, struct nvme_queue *q)
   1300 {
   1301 	struct nvme_sqe_q sqe;
   1302 	struct nvme_ccb *ccb;
   1303 	int rv;
   1304 
   1305 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1306 	KASSERT(ccb != NULL);
   1307 
   1308 	ccb->ccb_done = nvme_empty_done;
   1309 	ccb->ccb_cookie = &sqe;
   1310 
   1311 	memset(&sqe, 0, sizeof(sqe));
   1312 	sqe.opcode = NVM_ADMIN_DEL_IOSQ;
   1313 	htolem16(&sqe.qid, q->q_id);
   1314 
   1315 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1316 	if (rv != 0)
   1317 		goto fail;
   1318 
   1319 	ccb->ccb_done = nvme_empty_done;
   1320 	ccb->ccb_cookie = &sqe;
   1321 
   1322 	memset(&sqe, 0, sizeof(sqe));
   1323 	sqe.opcode = NVM_ADMIN_DEL_IOCQ;
   1324 	htolem64(&sqe.prp1, NVME_DMA_DVA(q->q_sq_dmamem));
   1325 	htolem16(&sqe.qid, q->q_id);
   1326 
   1327 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1328 	if (rv != 0)
   1329 		goto fail;
   1330 
   1331 fail:
   1332 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1333 
   1334 	if (rv == 0 && sc->sc_use_mq) {
   1335 		if (sc->sc_intr_disestablish(sc, q->q_id))
   1336 			rv = 1;
   1337 	}
   1338 
   1339 	return rv;
   1340 }
   1341 
   1342 static void
   1343 nvme_fill_identify(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
   1344 {
   1345 	struct nvme_sqe *sqe = slot;
   1346 	struct nvme_dmamem *mem = ccb->ccb_cookie;
   1347 
   1348 	sqe->opcode = NVM_ADMIN_IDENTIFY;
   1349 	htolem64(&sqe->entry.prp[0], NVME_DMA_DVA(mem));
   1350 	htolem32(&sqe->cdw10, 1);
   1351 }
   1352 
   1353 static int
   1354 nvme_get_number_of_queues(struct nvme_softc *sc, u_int *nqap)
   1355 {
   1356 	struct nvme_pt_command pt;
   1357 	struct nvme_ccb *ccb;
   1358 	uint16_t ncqa, nsqa;
   1359 	int rv;
   1360 
   1361 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1362 	KASSERT(ccb != NULL); /* it's a bug if we don't have spare ccb here */
   1363 
   1364 	memset(&pt, 0, sizeof(pt));
   1365 	pt.cmd.opcode = NVM_ADMIN_GET_FEATURES;
   1366 	pt.cmd.cdw10 = NVM_FEATURE_NUMBER_OF_QUEUES;
   1367 
   1368 	ccb->ccb_done = nvme_pt_done;
   1369 	ccb->ccb_cookie = &pt;
   1370 
   1371 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_pt_fill, NVME_TIMO_QOP);
   1372 
   1373 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1374 
   1375 	if (rv != 0) {
   1376 		*nqap = 0;
   1377 		return EIO;
   1378 	}
   1379 
   1380 	ncqa = pt.cpl.cdw0 >> 16;
   1381 	nsqa = pt.cpl.cdw0 & 0xffff;
   1382 	*nqap = MIN(ncqa, nsqa) + 1;
   1383 
   1384 	return 0;
   1385 }
   1386 
   1387 static int
   1388 nvme_ccbs_alloc(struct nvme_queue *q, uint16_t nccbs)
   1389 {
   1390 	struct nvme_softc *sc = q->q_sc;
   1391 	struct nvme_ccb *ccb;
   1392 	bus_addr_t off;
   1393 	uint64_t *prpl;
   1394 	u_int i;
   1395 
   1396 	mutex_init(&q->q_ccb_mtx, MUTEX_DEFAULT, IPL_BIO);
   1397 	SIMPLEQ_INIT(&q->q_ccb_list);
   1398 
   1399 	q->q_ccbs = kmem_alloc(sizeof(*ccb) * nccbs, KM_SLEEP);
   1400 	if (q->q_ccbs == NULL)
   1401 		return 1;
   1402 
   1403 	q->q_nccbs = nccbs;
   1404 	q->q_nccbs_avail = nccbs;
   1405 	q->q_ccb_prpls = nvme_dmamem_alloc(sc,
   1406 	    sizeof(*prpl) * sc->sc_max_sgl * nccbs);
   1407 
   1408 	prpl = NVME_DMA_KVA(q->q_ccb_prpls);
   1409 	off = 0;
   1410 
   1411 	for (i = 0; i < nccbs; i++) {
   1412 		ccb = &q->q_ccbs[i];
   1413 
   1414 		if (bus_dmamap_create(sc->sc_dmat, sc->sc_mdts,
   1415 		    sc->sc_max_sgl + 1 /* we get a free prp in the sqe */,
   1416 		    sc->sc_mps, sc->sc_mps, BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW,
   1417 		    &ccb->ccb_dmamap) != 0)
   1418 			goto free_maps;
   1419 
   1420 		ccb->ccb_id = i;
   1421 		ccb->ccb_prpl = prpl;
   1422 		ccb->ccb_prpl_off = off;
   1423 		ccb->ccb_prpl_dva = NVME_DMA_DVA(q->q_ccb_prpls) + off;
   1424 
   1425 		SIMPLEQ_INSERT_TAIL(&q->q_ccb_list, ccb, ccb_entry);
   1426 
   1427 		prpl += sc->sc_max_sgl;
   1428 		off += sizeof(*prpl) * sc->sc_max_sgl;
   1429 	}
   1430 
   1431 	return 0;
   1432 
   1433 free_maps:
   1434 	nvme_ccbs_free(q);
   1435 	return 1;
   1436 }
   1437 
   1438 static struct nvme_ccb *
   1439 nvme_ccb_get(struct nvme_queue *q)
   1440 {
   1441 	struct nvme_ccb *ccb = NULL;
   1442 
   1443 	mutex_enter(&q->q_ccb_mtx);
   1444 	if (q->q_nccbs_avail > 0) {
   1445 		ccb = SIMPLEQ_FIRST(&q->q_ccb_list);
   1446 		KASSERT(ccb != NULL);
   1447 		q->q_nccbs_avail--;
   1448 
   1449 		SIMPLEQ_REMOVE_HEAD(&q->q_ccb_list, ccb_entry);
   1450 #ifdef DEBUG
   1451 		ccb->ccb_cookie = NULL;
   1452 #endif
   1453 	}
   1454 	mutex_exit(&q->q_ccb_mtx);
   1455 
   1456 	return ccb;
   1457 }
   1458 
   1459 static void
   1460 nvme_ccb_put(struct nvme_queue *q, struct nvme_ccb *ccb)
   1461 {
   1462 
   1463 	mutex_enter(&q->q_ccb_mtx);
   1464 #ifdef DEBUG
   1465 	ccb->ccb_cookie = (void *)NVME_CCB_FREE;
   1466 #endif
   1467 	SIMPLEQ_INSERT_HEAD(&q->q_ccb_list, ccb, ccb_entry);
   1468 	mutex_exit(&q->q_ccb_mtx);
   1469 }
   1470 
   1471 static void
   1472 nvme_ccbs_free(struct nvme_queue *q)
   1473 {
   1474 	struct nvme_softc *sc = q->q_sc;
   1475 	struct nvme_ccb *ccb;
   1476 
   1477 	mutex_enter(&q->q_ccb_mtx);
   1478 	while ((ccb = SIMPLEQ_FIRST(&q->q_ccb_list)) != NULL) {
   1479 		SIMPLEQ_REMOVE_HEAD(&q->q_ccb_list, ccb_entry);
   1480 		bus_dmamap_destroy(sc->sc_dmat, ccb->ccb_dmamap);
   1481 	}
   1482 	mutex_exit(&q->q_ccb_mtx);
   1483 
   1484 	nvme_dmamem_free(sc, q->q_ccb_prpls);
   1485 	kmem_free(q->q_ccbs, sizeof(*ccb) * q->q_nccbs);
   1486 	q->q_ccbs = NULL;
   1487 	mutex_destroy(&q->q_ccb_mtx);
   1488 }
   1489 
   1490 static struct nvme_queue *
   1491 nvme_q_alloc(struct nvme_softc *sc, uint16_t id, u_int entries, u_int dstrd)
   1492 {
   1493 	struct nvme_queue *q;
   1494 
   1495 	q = kmem_alloc(sizeof(*q), KM_SLEEP);
   1496 	if (q == NULL)
   1497 		return NULL;
   1498 
   1499 	q->q_sc = sc;
   1500 	q->q_sq_dmamem = nvme_dmamem_alloc(sc,
   1501 	    sizeof(struct nvme_sqe) * entries);
   1502 	if (q->q_sq_dmamem == NULL)
   1503 		goto free;
   1504 
   1505 	q->q_cq_dmamem = nvme_dmamem_alloc(sc,
   1506 	    sizeof(struct nvme_cqe) * entries);
   1507 	if (q->q_cq_dmamem == NULL)
   1508 		goto free_sq;
   1509 
   1510 	memset(NVME_DMA_KVA(q->q_sq_dmamem), 0, NVME_DMA_LEN(q->q_sq_dmamem));
   1511 	memset(NVME_DMA_KVA(q->q_cq_dmamem), 0, NVME_DMA_LEN(q->q_cq_dmamem));
   1512 
   1513 	mutex_init(&q->q_sq_mtx, MUTEX_DEFAULT, IPL_BIO);
   1514 	mutex_init(&q->q_cq_mtx, MUTEX_DEFAULT, IPL_BIO);
   1515 	q->q_sqtdbl = NVME_SQTDBL(id, dstrd);
   1516 	q->q_cqhdbl = NVME_CQHDBL(id, dstrd);
   1517 	q->q_id = id;
   1518 	q->q_entries = entries;
   1519 	q->q_sq_tail = 0;
   1520 	q->q_cq_head = 0;
   1521 	q->q_cq_phase = NVME_CQE_PHASE;
   1522 
   1523 	nvme_dmamem_sync(sc, q->q_sq_dmamem, BUS_DMASYNC_PREWRITE);
   1524 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_PREREAD);
   1525 
   1526 	/*
   1527 	 * Due to definition of full and empty queue (queue is empty
   1528 	 * when head == tail, full when tail is one less then head),
   1529 	 * we can actually only have (entries - 1) in-flight commands.
   1530 	 */
   1531 	if (nvme_ccbs_alloc(q, entries - 1) != 0) {
   1532 		aprint_error_dev(sc->sc_dev, "unable to allocate ccbs\n");
   1533 		goto free_cq;
   1534 	}
   1535 
   1536 	return q;
   1537 
   1538 free_cq:
   1539 	nvme_dmamem_free(sc, q->q_cq_dmamem);
   1540 free_sq:
   1541 	nvme_dmamem_free(sc, q->q_sq_dmamem);
   1542 free:
   1543 	kmem_free(q, sizeof(*q));
   1544 
   1545 	return NULL;
   1546 }
   1547 
   1548 static void
   1549 nvme_q_free(struct nvme_softc *sc, struct nvme_queue *q)
   1550 {
   1551 	nvme_ccbs_free(q);
   1552 	mutex_destroy(&q->q_sq_mtx);
   1553 	mutex_destroy(&q->q_cq_mtx);
   1554 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_POSTREAD);
   1555 	nvme_dmamem_sync(sc, q->q_sq_dmamem, BUS_DMASYNC_POSTWRITE);
   1556 	nvme_dmamem_free(sc, q->q_cq_dmamem);
   1557 	nvme_dmamem_free(sc, q->q_sq_dmamem);
   1558 	kmem_free(q, sizeof(*q));
   1559 }
   1560 
   1561 int
   1562 nvme_intr(void *xsc)
   1563 {
   1564 	struct nvme_softc *sc = xsc;
   1565 
   1566 	/*
   1567 	 * INTx is level triggered, controller deasserts the interrupt only
   1568 	 * when we advance command queue head via write to the doorbell.
   1569 	 * Tell the controller to block the interrupts while we process
   1570 	 * the queue(s).
   1571 	 */
   1572 	nvme_write4(sc, NVME_INTMS, 1);
   1573 
   1574 	softint_schedule(sc->sc_softih[0]);
   1575 
   1576 	/* don't know, might not have been for us */
   1577 	return 1;
   1578 }
   1579 
   1580 void
   1581 nvme_softintr_intx(void *xq)
   1582 {
   1583 	struct nvme_queue *q = xq;
   1584 	struct nvme_softc *sc = q->q_sc;
   1585 
   1586 	nvme_q_complete(sc, sc->sc_admin_q);
   1587 	if (sc->sc_q != NULL)
   1588 	        nvme_q_complete(sc, sc->sc_q[0]);
   1589 
   1590 	/*
   1591 	 * Processing done, tell controller to issue interrupts again. There
   1592 	 * is no race, as NVMe spec requires the controller to maintain state,
   1593 	 * and assert the interrupt whenever there are unacknowledged
   1594 	 * completion queue entries.
   1595 	 */
   1596 	nvme_write4(sc, NVME_INTMC, 1);
   1597 }
   1598 
   1599 int
   1600 nvme_intr_msi(void *xq)
   1601 {
   1602 	struct nvme_queue *q = xq;
   1603 
   1604 	KASSERT(q && q->q_sc && q->q_sc->sc_softih
   1605 	    && q->q_sc->sc_softih[q->q_id]);
   1606 
   1607 	/*
   1608 	 * MSI/MSI-X are edge triggered, so can handover processing to softint
   1609 	 * without masking the interrupt.
   1610 	 */
   1611 	softint_schedule(q->q_sc->sc_softih[q->q_id]);
   1612 
   1613 	return 1;
   1614 }
   1615 
   1616 void
   1617 nvme_softintr_msi(void *xq)
   1618 {
   1619 	struct nvme_queue *q = xq;
   1620 	struct nvme_softc *sc = q->q_sc;
   1621 
   1622 	nvme_q_complete(sc, q);
   1623 }
   1624 
   1625 static struct nvme_dmamem *
   1626 nvme_dmamem_alloc(struct nvme_softc *sc, size_t size)
   1627 {
   1628 	struct nvme_dmamem *ndm;
   1629 	int nsegs;
   1630 
   1631 	ndm = kmem_zalloc(sizeof(*ndm), KM_SLEEP);
   1632 	if (ndm == NULL)
   1633 		return NULL;
   1634 
   1635 	ndm->ndm_size = size;
   1636 
   1637 	if (bus_dmamap_create(sc->sc_dmat, size, 1, size, 0,
   1638 	    BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW, &ndm->ndm_map) != 0)
   1639 		goto ndmfree;
   1640 
   1641 	if (bus_dmamem_alloc(sc->sc_dmat, size, sc->sc_mps, 0, &ndm->ndm_seg,
   1642 	    1, &nsegs, BUS_DMA_WAITOK) != 0)
   1643 		goto destroy;
   1644 
   1645 	if (bus_dmamem_map(sc->sc_dmat, &ndm->ndm_seg, nsegs, size,
   1646 	    &ndm->ndm_kva, BUS_DMA_WAITOK) != 0)
   1647 		goto free;
   1648 	memset(ndm->ndm_kva, 0, size);
   1649 
   1650 	if (bus_dmamap_load(sc->sc_dmat, ndm->ndm_map, ndm->ndm_kva, size,
   1651 	    NULL, BUS_DMA_WAITOK) != 0)
   1652 		goto unmap;
   1653 
   1654 	return ndm;
   1655 
   1656 unmap:
   1657 	bus_dmamem_unmap(sc->sc_dmat, ndm->ndm_kva, size);
   1658 free:
   1659 	bus_dmamem_free(sc->sc_dmat, &ndm->ndm_seg, 1);
   1660 destroy:
   1661 	bus_dmamap_destroy(sc->sc_dmat, ndm->ndm_map);
   1662 ndmfree:
   1663 	kmem_free(ndm, sizeof(*ndm));
   1664 	return NULL;
   1665 }
   1666 
   1667 static void
   1668 nvme_dmamem_sync(struct nvme_softc *sc, struct nvme_dmamem *mem, int ops)
   1669 {
   1670 	bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(mem),
   1671 	    0, NVME_DMA_LEN(mem), ops);
   1672 }
   1673 
   1674 void
   1675 nvme_dmamem_free(struct nvme_softc *sc, struct nvme_dmamem *ndm)
   1676 {
   1677 	bus_dmamap_unload(sc->sc_dmat, ndm->ndm_map);
   1678 	bus_dmamem_unmap(sc->sc_dmat, ndm->ndm_kva, ndm->ndm_size);
   1679 	bus_dmamem_free(sc->sc_dmat, &ndm->ndm_seg, 1);
   1680 	bus_dmamap_destroy(sc->sc_dmat, ndm->ndm_map);
   1681 	kmem_free(ndm, sizeof(*ndm));
   1682 }
   1683 
   1684 /*
   1685  * ioctl
   1686  */
   1687 
   1688 dev_type_open(nvmeopen);
   1689 dev_type_close(nvmeclose);
   1690 dev_type_ioctl(nvmeioctl);
   1691 
   1692 const struct cdevsw nvme_cdevsw = {
   1693 	DEVSW_MODULE_INIT
   1694 	.d_open = nvmeopen,
   1695 	.d_close = nvmeclose,
   1696 	.d_read = noread,
   1697 	.d_write = nowrite,
   1698 	.d_ioctl = nvmeioctl,
   1699 	.d_stop = nostop,
   1700 	.d_tty = notty,
   1701 	.d_poll = nopoll,
   1702 	.d_mmap = nommap,
   1703 	.d_kqfilter = nokqfilter,
   1704 	.d_discard = nodiscard,
   1705 	.d_flag = D_OTHER,
   1706 };
   1707 
   1708 extern struct cfdriver nvme_cd;
   1709 
   1710 /*
   1711  * Accept an open operation on the control device.
   1712  */
   1713 int
   1714 nvmeopen(dev_t dev, int flag, int mode, struct lwp *l)
   1715 {
   1716 	struct nvme_softc *sc;
   1717 	int unit = minor(dev) / 0x10000;
   1718 	int nsid = minor(dev) & 0xffff;
   1719 	int nsidx;
   1720 
   1721 	if ((sc = device_lookup_private(&nvme_cd, unit)) == NULL)
   1722 		return ENXIO;
   1723 	if ((sc->sc_flags & NVME_F_ATTACHED) == 0)
   1724 		return ENXIO;
   1725 
   1726 	if (nsid == 0) {
   1727 		/* controller */
   1728 		if (ISSET(sc->sc_flags, NVME_F_OPEN))
   1729 			return EBUSY;
   1730 		SET(sc->sc_flags, NVME_F_OPEN);
   1731 	} else {
   1732 		/* namespace */
   1733 		nsidx = nsid - 1;
   1734 		if (nsidx >= sc->sc_nn || sc->sc_namespaces[nsidx].dev == NULL)
   1735 			return ENXIO;
   1736 		if (ISSET(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN))
   1737 			return EBUSY;
   1738 		SET(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN);
   1739 	}
   1740 	return 0;
   1741 }
   1742 
   1743 /*
   1744  * Accept the last close on the control device.
   1745  */
   1746 int
   1747 nvmeclose(dev_t dev, int flag, int mode, struct lwp *l)
   1748 {
   1749 	struct nvme_softc *sc;
   1750 	int unit = minor(dev) / 0x10000;
   1751 	int nsid = minor(dev) & 0xffff;
   1752 	int nsidx;
   1753 
   1754 	sc = device_lookup_private(&nvme_cd, unit);
   1755 	if (sc == NULL)
   1756 		return ENXIO;
   1757 
   1758 	if (nsid == 0) {
   1759 		/* controller */
   1760 		CLR(sc->sc_flags, NVME_F_OPEN);
   1761 	} else {
   1762 		/* namespace */
   1763 		nsidx = nsid - 1;
   1764 		if (nsidx >= sc->sc_nn)
   1765 			return ENXIO;
   1766 		CLR(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN);
   1767 	}
   1768 
   1769 	return 0;
   1770 }
   1771 
   1772 /*
   1773  * Handle control operations.
   1774  */
   1775 int
   1776 nvmeioctl(dev_t dev, u_long cmd, void *data, int flag, struct lwp *l)
   1777 {
   1778 	struct nvme_softc *sc;
   1779 	int unit = minor(dev) / 0x10000;
   1780 	int nsid = minor(dev) & 0xffff;
   1781 	struct nvme_pt_command *pt;
   1782 
   1783 	sc = device_lookup_private(&nvme_cd, unit);
   1784 	if (sc == NULL)
   1785 		return ENXIO;
   1786 
   1787 	switch (cmd) {
   1788 	case NVME_PASSTHROUGH_CMD:
   1789 		pt = data;
   1790 		return nvme_command_passthrough(sc, data,
   1791 		    nsid == 0 ? pt->cmd.nsid : nsid, l, nsid == 0);
   1792 	}
   1793 
   1794 	return ENOTTY;
   1795 }
   1796