Home | History | Annotate | Line # | Download | only in ic
nvme.c revision 1.32
      1 /*	$NetBSD: nvme.c,v 1.32 2018/02/27 12:59:53 christos 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 1.32 2018/02/27 12:59:53 christos 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 		nvme_ccb_put(sc->sc_admin_q, ccb);
    575 		return ENOMEM;
    576 	}
    577 
    578 	memset(&sqe, 0, sizeof(sqe));
    579 	sqe.opcode = NVM_ADMIN_IDENTIFY;
    580 	htolem32(&sqe.nsid, nsid);
    581 	htolem64(&sqe.entry.prp[0], NVME_DMA_DVA(mem));
    582 	htolem32(&sqe.cdw10, 0);
    583 
    584 	ccb->ccb_done = nvme_empty_done;
    585 	ccb->ccb_cookie = &sqe;
    586 
    587 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_PREREAD);
    588 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_IDENT);
    589 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_POSTREAD);
    590 
    591 	nvme_ccb_put(sc->sc_admin_q, ccb);
    592 
    593 	if (rv != 0) {
    594 		rv = EIO;
    595 		goto done;
    596 	}
    597 
    598 	/* commit */
    599 
    600 	identify = kmem_zalloc(sizeof(*identify), KM_SLEEP);
    601 	*identify = *((volatile struct nvm_identify_namespace *)NVME_DMA_KVA(mem));
    602 	//memcpy(identify, NVME_DMA_KVA(mem), sizeof(*identify));
    603 
    604 	ns = nvme_ns_get(sc, nsid);
    605 	KASSERT(ns);
    606 	KASSERT(ns->ident == NULL);
    607 	ns->ident = identify;
    608 
    609 done:
    610 	nvme_dmamem_free(sc, mem);
    611 
    612 	return rv;
    613 }
    614 
    615 int
    616 nvme_ns_dobio(struct nvme_softc *sc, uint16_t nsid, void *cookie,
    617     struct buf *bp, void *data, size_t datasize,
    618     int secsize, daddr_t blkno, int flags, nvme_nnc_done nnc_done)
    619 {
    620 	struct nvme_queue *q = nvme_get_q(sc);
    621 	struct nvme_ccb *ccb;
    622 	bus_dmamap_t dmap;
    623 	int i, error;
    624 
    625 	ccb = nvme_ccb_get(q);
    626 	if (ccb == NULL)
    627 		return EAGAIN;
    628 
    629 	ccb->ccb_done = nvme_ns_io_done;
    630 	ccb->ccb_cookie = cookie;
    631 
    632 	/* namespace context */
    633 	ccb->nnc_nsid = nsid;
    634 	ccb->nnc_flags = flags;
    635 	ccb->nnc_buf = bp;
    636 	ccb->nnc_datasize = datasize;
    637 	ccb->nnc_secsize = secsize;
    638 	ccb->nnc_blkno = blkno;
    639 	ccb->nnc_done = nnc_done;
    640 
    641 	dmap = ccb->ccb_dmamap;
    642 	error = bus_dmamap_load(sc->sc_dmat, dmap, data,
    643 	    datasize, NULL,
    644 	    (ISSET(flags, NVME_NS_CTX_F_POLL) ?
    645 	      BUS_DMA_NOWAIT : BUS_DMA_WAITOK) |
    646 	    (ISSET(flags, NVME_NS_CTX_F_READ) ?
    647 	      BUS_DMA_READ : BUS_DMA_WRITE));
    648 	if (error) {
    649 		nvme_ccb_put(q, ccb);
    650 		return error;
    651 	}
    652 
    653 	bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize,
    654 	    ISSET(flags, NVME_NS_CTX_F_READ) ?
    655 	    BUS_DMASYNC_PREREAD : BUS_DMASYNC_PREWRITE);
    656 
    657 	if (dmap->dm_nsegs > 2) {
    658 		for (i = 1; i < dmap->dm_nsegs; i++) {
    659 			htolem64(&ccb->ccb_prpl[i - 1],
    660 			    dmap->dm_segs[i].ds_addr);
    661 		}
    662 		bus_dmamap_sync(sc->sc_dmat,
    663 		    NVME_DMA_MAP(q->q_ccb_prpls),
    664 		    ccb->ccb_prpl_off,
    665 		    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    666 		    BUS_DMASYNC_PREWRITE);
    667 	}
    668 
    669 	if (ISSET(flags, NVME_NS_CTX_F_POLL)) {
    670 		if (nvme_poll(sc, q, ccb, nvme_ns_io_fill, NVME_TIMO_PT) != 0)
    671 			return EIO;
    672 		return 0;
    673 	}
    674 
    675 	nvme_q_submit(sc, q, ccb, nvme_ns_io_fill);
    676 	return 0;
    677 }
    678 
    679 static void
    680 nvme_ns_io_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    681 {
    682 	struct nvme_sqe_io *sqe = slot;
    683 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    684 
    685 	sqe->opcode = ISSET(ccb->nnc_flags, NVME_NS_CTX_F_READ) ?
    686 	    NVM_CMD_READ : NVM_CMD_WRITE;
    687 	htolem32(&sqe->nsid, ccb->nnc_nsid);
    688 
    689 	htolem64(&sqe->entry.prp[0], dmap->dm_segs[0].ds_addr);
    690 	switch (dmap->dm_nsegs) {
    691 	case 1:
    692 		break;
    693 	case 2:
    694 		htolem64(&sqe->entry.prp[1], dmap->dm_segs[1].ds_addr);
    695 		break;
    696 	default:
    697 		/* the prp list is already set up and synced */
    698 		htolem64(&sqe->entry.prp[1], ccb->ccb_prpl_dva);
    699 		break;
    700 	}
    701 
    702 	htolem64(&sqe->slba, ccb->nnc_blkno);
    703 
    704 	if (ISSET(ccb->nnc_flags, NVME_NS_CTX_F_FUA))
    705 		htolem16(&sqe->ioflags, NVM_SQE_IO_FUA);
    706 
    707 	/* guaranteed by upper layers, but check just in case */
    708 	KASSERT((ccb->nnc_datasize % ccb->nnc_secsize) == 0);
    709 	htolem16(&sqe->nlb, (ccb->nnc_datasize / ccb->nnc_secsize) - 1);
    710 }
    711 
    712 static void
    713 nvme_ns_io_done(struct nvme_queue *q, struct nvme_ccb *ccb,
    714     struct nvme_cqe *cqe)
    715 {
    716 	struct nvme_softc *sc = q->q_sc;
    717 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    718 	void *nnc_cookie = ccb->ccb_cookie;
    719 	nvme_nnc_done nnc_done = ccb->nnc_done;
    720 	struct buf *bp = ccb->nnc_buf;
    721 
    722 	if (dmap->dm_nsegs > 2) {
    723 		bus_dmamap_sync(sc->sc_dmat,
    724 		    NVME_DMA_MAP(q->q_ccb_prpls),
    725 		    ccb->ccb_prpl_off,
    726 		    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    727 		    BUS_DMASYNC_POSTWRITE);
    728 	}
    729 
    730 	bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize,
    731 	    ISSET(ccb->nnc_flags, NVME_NS_CTX_F_READ) ?
    732 	    BUS_DMASYNC_POSTREAD : BUS_DMASYNC_POSTWRITE);
    733 
    734 	bus_dmamap_unload(sc->sc_dmat, dmap);
    735 	nvme_ccb_put(q, ccb);
    736 
    737 	nnc_done(nnc_cookie, bp, lemtoh16(&cqe->flags), lemtoh32(&cqe->cdw0));
    738 }
    739 
    740 /*
    741  * If there is no volatile write cache, it makes no sense to issue
    742  * flush commands or query for the status.
    743  */
    744 bool
    745 nvme_has_volatile_write_cache(struct nvme_softc *sc)
    746 {
    747 	/* sc_identify is filled during attachment */
    748 	return  ((sc->sc_identify.vwc & NVME_ID_CTRLR_VWC_PRESENT) != 0);
    749 }
    750 
    751 int
    752 nvme_ns_sync(struct nvme_softc *sc, uint16_t nsid, void *cookie,
    753     int flags, nvme_nnc_done nnc_done)
    754 {
    755 	struct nvme_queue *q = nvme_get_q(sc);
    756 	struct nvme_ccb *ccb;
    757 
    758 	ccb = nvme_ccb_get(q);
    759 	if (ccb == NULL)
    760 		return EAGAIN;
    761 
    762 	ccb->ccb_done = nvme_ns_sync_done;
    763 	ccb->ccb_cookie = cookie;
    764 
    765 	/* namespace context */
    766 	ccb->nnc_nsid = nsid;
    767 	ccb->nnc_flags = flags;
    768 	ccb->nnc_done = nnc_done;
    769 
    770 	if (ISSET(flags, NVME_NS_CTX_F_POLL)) {
    771 		if (nvme_poll(sc, q, ccb, nvme_ns_sync_fill, NVME_TIMO_SY) != 0)
    772 			return EIO;
    773 		return 0;
    774 	}
    775 
    776 	nvme_q_submit(sc, q, ccb, nvme_ns_sync_fill);
    777 	return 0;
    778 }
    779 
    780 static void
    781 nvme_ns_sync_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    782 {
    783 	struct nvme_sqe *sqe = slot;
    784 
    785 	sqe->opcode = NVM_CMD_FLUSH;
    786 	htolem32(&sqe->nsid, ccb->nnc_nsid);
    787 }
    788 
    789 static void
    790 nvme_ns_sync_done(struct nvme_queue *q, struct nvme_ccb *ccb,
    791     struct nvme_cqe *cqe)
    792 {
    793 	void *cookie = ccb->ccb_cookie;
    794 	nvme_nnc_done nnc_done = ccb->nnc_done;
    795 
    796 	nvme_ccb_put(q, ccb);
    797 
    798 	nnc_done(cookie, NULL, lemtoh16(&cqe->flags), lemtoh32(&cqe->cdw0));
    799 }
    800 
    801 /*
    802  * Get status of volatile write cache. Always asynchronous.
    803  */
    804 int
    805 nvme_admin_getcache(struct nvme_softc *sc, void *cookie, nvme_nnc_done nnc_done)
    806 {
    807 	struct nvme_ccb *ccb;
    808 	struct nvme_queue *q = sc->sc_admin_q;
    809 
    810 	ccb = nvme_ccb_get(q);
    811 	if (ccb == NULL)
    812 		return EAGAIN;
    813 
    814 	ccb->ccb_done = nvme_getcache_done;
    815 	ccb->ccb_cookie = cookie;
    816 
    817 	/* namespace context */
    818 	ccb->nnc_flags = 0;
    819 	ccb->nnc_done = nnc_done;
    820 
    821 	nvme_q_submit(sc, q, ccb, nvme_getcache_fill);
    822 	return 0;
    823 }
    824 
    825 static void
    826 nvme_getcache_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    827 {
    828 	struct nvme_sqe *sqe = slot;
    829 
    830 	sqe->opcode = NVM_ADMIN_GET_FEATURES;
    831 	sqe->cdw10 = NVM_FEATURE_VOLATILE_WRITE_CACHE;
    832 }
    833 
    834 static void
    835 nvme_getcache_done(struct nvme_queue *q, struct nvme_ccb *ccb,
    836     struct nvme_cqe *cqe)
    837 {
    838 	void *cookie = ccb->ccb_cookie;
    839 	nvme_nnc_done nnc_done = ccb->nnc_done;
    840 
    841 	nvme_ccb_put(q, ccb);
    842 
    843 	nnc_done(cookie, NULL, lemtoh16(&cqe->flags), lemtoh32(&cqe->cdw0));
    844 }
    845 
    846 void
    847 nvme_ns_free(struct nvme_softc *sc, uint16_t nsid)
    848 {
    849 	struct nvme_namespace *ns;
    850 	struct nvm_identify_namespace *identify;
    851 
    852 	ns = nvme_ns_get(sc, nsid);
    853 	KASSERT(ns);
    854 
    855 	identify = ns->ident;
    856 	ns->ident = NULL;
    857 	if (identify != NULL)
    858 		kmem_free(identify, sizeof(*identify));
    859 }
    860 
    861 static void
    862 nvme_pt_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
    863 {
    864 	struct nvme_softc *sc = q->q_sc;
    865 	struct nvme_sqe *sqe = slot;
    866 	struct nvme_pt_command *pt = ccb->ccb_cookie;
    867 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    868 	int i;
    869 
    870 	sqe->opcode = pt->cmd.opcode;
    871 	htolem32(&sqe->nsid, pt->cmd.nsid);
    872 
    873 	if (pt->buf != NULL && pt->len > 0) {
    874 		htolem64(&sqe->entry.prp[0], dmap->dm_segs[0].ds_addr);
    875 		switch (dmap->dm_nsegs) {
    876 		case 1:
    877 			break;
    878 		case 2:
    879 			htolem64(&sqe->entry.prp[1], dmap->dm_segs[1].ds_addr);
    880 			break;
    881 		default:
    882 			for (i = 1; i < dmap->dm_nsegs; i++) {
    883 				htolem64(&ccb->ccb_prpl[i - 1],
    884 				    dmap->dm_segs[i].ds_addr);
    885 			}
    886 			bus_dmamap_sync(sc->sc_dmat,
    887 			    NVME_DMA_MAP(q->q_ccb_prpls),
    888 			    ccb->ccb_prpl_off,
    889 			    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    890 			    BUS_DMASYNC_PREWRITE);
    891 			htolem64(&sqe->entry.prp[1], ccb->ccb_prpl_dva);
    892 			break;
    893 		}
    894 	}
    895 
    896 	htolem32(&sqe->cdw10, pt->cmd.cdw10);
    897 	htolem32(&sqe->cdw11, pt->cmd.cdw11);
    898 	htolem32(&sqe->cdw12, pt->cmd.cdw12);
    899 	htolem32(&sqe->cdw13, pt->cmd.cdw13);
    900 	htolem32(&sqe->cdw14, pt->cmd.cdw14);
    901 	htolem32(&sqe->cdw15, pt->cmd.cdw15);
    902 }
    903 
    904 static void
    905 nvme_pt_done(struct nvme_queue *q, struct nvme_ccb *ccb, struct nvme_cqe *cqe)
    906 {
    907 	struct nvme_softc *sc = q->q_sc;
    908 	struct nvme_pt_command *pt = ccb->ccb_cookie;
    909 	bus_dmamap_t dmap = ccb->ccb_dmamap;
    910 
    911 	if (pt->buf != NULL && pt->len > 0) {
    912 		if (dmap->dm_nsegs > 2) {
    913 			bus_dmamap_sync(sc->sc_dmat,
    914 			    NVME_DMA_MAP(q->q_ccb_prpls),
    915 			    ccb->ccb_prpl_off,
    916 			    sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1),
    917 			    BUS_DMASYNC_POSTWRITE);
    918 		}
    919 
    920 		bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize,
    921 		    pt->is_read ? BUS_DMASYNC_POSTREAD : BUS_DMASYNC_POSTWRITE);
    922 		bus_dmamap_unload(sc->sc_dmat, dmap);
    923 	}
    924 
    925 	pt->cpl.cdw0 = lemtoh32(&cqe->cdw0);
    926 	pt->cpl.flags = lemtoh16(&cqe->flags) & ~NVME_CQE_PHASE;
    927 }
    928 
    929 static int
    930 nvme_command_passthrough(struct nvme_softc *sc, struct nvme_pt_command *pt,
    931     uint16_t nsid, struct lwp *l, bool is_adminq)
    932 {
    933 	struct nvme_queue *q;
    934 	struct nvme_ccb *ccb;
    935 	void *buf = NULL;
    936 	int error;
    937 
    938 	/* limit command size to maximum data transfer size */
    939 	if ((pt->buf == NULL && pt->len > 0) ||
    940 	    (pt->buf != NULL && (pt->len == 0 || pt->len > sc->sc_mdts)))
    941 		return EINVAL;
    942 
    943 	q = is_adminq ? sc->sc_admin_q : nvme_get_q(sc);
    944 	ccb = nvme_ccb_get(q);
    945 	if (ccb == NULL)
    946 		return EBUSY;
    947 
    948 	if (pt->buf != NULL) {
    949 		KASSERT(pt->len > 0);
    950 		buf = kmem_alloc(pt->len, KM_SLEEP);
    951 		if (!pt->is_read) {
    952 			error = copyin(pt->buf, buf, pt->len);
    953 			if (error)
    954 				goto kmem_free;
    955 		}
    956 		error = bus_dmamap_load(sc->sc_dmat, ccb->ccb_dmamap, buf,
    957 		    pt->len, NULL,
    958 		    BUS_DMA_WAITOK |
    959 		      (pt->is_read ? BUS_DMA_READ : BUS_DMA_WRITE));
    960 		if (error)
    961 			goto kmem_free;
    962 		bus_dmamap_sync(sc->sc_dmat, ccb->ccb_dmamap,
    963 		    0, ccb->ccb_dmamap->dm_mapsize,
    964 		    pt->is_read ? BUS_DMASYNC_PREREAD : BUS_DMASYNC_PREWRITE);
    965 	}
    966 
    967 	ccb->ccb_done = nvme_pt_done;
    968 	ccb->ccb_cookie = pt;
    969 
    970 	pt->cmd.nsid = nsid;
    971 	if (nvme_poll(sc, q, ccb, nvme_pt_fill, NVME_TIMO_PT)) {
    972 		error = EIO;
    973 		goto out;
    974 	}
    975 
    976 	error = 0;
    977 out:
    978 	if (buf != NULL) {
    979 		if (error == 0 && pt->is_read)
    980 			error = copyout(buf, pt->buf, pt->len);
    981 kmem_free:
    982 		kmem_free(buf, pt->len);
    983 	}
    984 	nvme_ccb_put(q, ccb);
    985 	return error;
    986 }
    987 
    988 static void
    989 nvme_q_submit(struct nvme_softc *sc, struct nvme_queue *q, struct nvme_ccb *ccb,
    990     void (*fill)(struct nvme_queue *, struct nvme_ccb *, void *))
    991 {
    992 	struct nvme_sqe *sqe = NVME_DMA_KVA(q->q_sq_dmamem);
    993 	uint32_t tail;
    994 
    995 	mutex_enter(&q->q_sq_mtx);
    996 	tail = q->q_sq_tail;
    997 	if (++q->q_sq_tail >= q->q_entries)
    998 		q->q_sq_tail = 0;
    999 
   1000 	sqe += tail;
   1001 
   1002 	bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem),
   1003 	    sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_POSTWRITE);
   1004 	memset(sqe, 0, sizeof(*sqe));
   1005 	(*fill)(q, ccb, sqe);
   1006 	sqe->cid = ccb->ccb_id;
   1007 	bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem),
   1008 	    sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_PREWRITE);
   1009 
   1010 	nvme_write4(sc, q->q_sqtdbl, q->q_sq_tail);
   1011 	mutex_exit(&q->q_sq_mtx);
   1012 }
   1013 
   1014 struct nvme_poll_state {
   1015 	struct nvme_sqe s;
   1016 	struct nvme_cqe c;
   1017 };
   1018 
   1019 static int
   1020 nvme_poll(struct nvme_softc *sc, struct nvme_queue *q, struct nvme_ccb *ccb,
   1021     void (*fill)(struct nvme_queue *, struct nvme_ccb *, void *), int timo_sec)
   1022 {
   1023 	struct nvme_poll_state state;
   1024 	void (*done)(struct nvme_queue *, struct nvme_ccb *, struct nvme_cqe *);
   1025 	void *cookie;
   1026 	uint16_t flags;
   1027 	int step = 10;
   1028 	int maxloop = timo_sec * 1000000 / step;
   1029 	int error = 0;
   1030 
   1031 	memset(&state, 0, sizeof(state));
   1032 	(*fill)(q, ccb, &state.s);
   1033 
   1034 	done = ccb->ccb_done;
   1035 	cookie = ccb->ccb_cookie;
   1036 
   1037 	ccb->ccb_done = nvme_poll_done;
   1038 	ccb->ccb_cookie = &state;
   1039 
   1040 	nvme_q_submit(sc, q, ccb, nvme_poll_fill);
   1041 	while (!ISSET(state.c.flags, htole16(NVME_CQE_PHASE))) {
   1042 		if (nvme_q_complete(sc, q) == 0)
   1043 			delay(step);
   1044 
   1045 		if (timo_sec >= 0 && --maxloop <= 0) {
   1046 			error = ETIMEDOUT;
   1047 			break;
   1048 		}
   1049 	}
   1050 
   1051 	ccb->ccb_cookie = cookie;
   1052 	done(q, ccb, &state.c);
   1053 
   1054 	if (error == 0) {
   1055 		flags = lemtoh16(&state.c.flags);
   1056 		return flags & ~NVME_CQE_PHASE;
   1057 	} else {
   1058 		return 1;
   1059 	}
   1060 }
   1061 
   1062 static void
   1063 nvme_poll_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
   1064 {
   1065 	struct nvme_sqe *sqe = slot;
   1066 	struct nvme_poll_state *state = ccb->ccb_cookie;
   1067 
   1068 	*sqe = state->s;
   1069 }
   1070 
   1071 static void
   1072 nvme_poll_done(struct nvme_queue *q, struct nvme_ccb *ccb,
   1073     struct nvme_cqe *cqe)
   1074 {
   1075 	struct nvme_poll_state *state = ccb->ccb_cookie;
   1076 
   1077 	SET(cqe->flags, htole16(NVME_CQE_PHASE));
   1078 	state->c = *cqe;
   1079 }
   1080 
   1081 static void
   1082 nvme_sqe_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
   1083 {
   1084 	struct nvme_sqe *src = ccb->ccb_cookie;
   1085 	struct nvme_sqe *dst = slot;
   1086 
   1087 	*dst = *src;
   1088 }
   1089 
   1090 static void
   1091 nvme_empty_done(struct nvme_queue *q, struct nvme_ccb *ccb,
   1092     struct nvme_cqe *cqe)
   1093 {
   1094 }
   1095 
   1096 static int
   1097 nvme_q_complete(struct nvme_softc *sc, struct nvme_queue *q)
   1098 {
   1099 	struct nvme_ccb *ccb;
   1100 	struct nvme_cqe *ring = NVME_DMA_KVA(q->q_cq_dmamem), *cqe;
   1101 	uint16_t flags;
   1102 	int rv = 0;
   1103 
   1104 	mutex_enter(&q->q_cq_mtx);
   1105 
   1106 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_POSTREAD);
   1107 	for (;;) {
   1108 		cqe = &ring[q->q_cq_head];
   1109 		flags = lemtoh16(&cqe->flags);
   1110 		if ((flags & NVME_CQE_PHASE) != q->q_cq_phase)
   1111 			break;
   1112 
   1113 		ccb = &q->q_ccbs[cqe->cid];
   1114 
   1115 		if (++q->q_cq_head >= q->q_entries) {
   1116 			q->q_cq_head = 0;
   1117 			q->q_cq_phase ^= NVME_CQE_PHASE;
   1118 		}
   1119 
   1120 #ifdef DEBUG
   1121 		/*
   1122 		 * If we get spurious completion notification, something
   1123 		 * is seriously hosed up. Very likely DMA to some random
   1124 		 * memory place happened, so just bail out.
   1125 		 */
   1126 		if ((intptr_t)ccb->ccb_cookie == NVME_CCB_FREE) {
   1127 			panic("%s: invalid ccb detected",
   1128 			    device_xname(sc->sc_dev));
   1129 			/* NOTREACHED */
   1130 		}
   1131 #endif
   1132 
   1133 		rv++;
   1134 
   1135 		/*
   1136 		 * Unlock the mutex before calling the ccb_done callback
   1137 		 * and re-lock afterwards. The callback triggers lddone()
   1138 		 * which schedules another i/o, and also calls nvme_ccb_put().
   1139 		 * Unlock/relock avoids possibility of deadlock.
   1140 		 */
   1141 		mutex_exit(&q->q_cq_mtx);
   1142 		ccb->ccb_done(q, ccb, cqe);
   1143 		mutex_enter(&q->q_cq_mtx);
   1144 	}
   1145 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_PREREAD);
   1146 
   1147 	if (rv)
   1148 		nvme_write4(sc, q->q_cqhdbl, q->q_cq_head);
   1149 
   1150 	mutex_exit(&q->q_cq_mtx);
   1151 
   1152 	if (rv) {
   1153 		mutex_enter(&q->q_ccb_mtx);
   1154 		q->q_nccbs_avail += rv;
   1155 		mutex_exit(&q->q_ccb_mtx);
   1156 	}
   1157 
   1158 	return rv;
   1159 }
   1160 
   1161 static int
   1162 nvme_identify(struct nvme_softc *sc, u_int mps)
   1163 {
   1164 	char sn[41], mn[81], fr[17];
   1165 	struct nvm_identify_controller *identify;
   1166 	struct nvme_dmamem *mem;
   1167 	struct nvme_ccb *ccb;
   1168 	u_int mdts;
   1169 	int rv = 1;
   1170 
   1171 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1172 	KASSERT(ccb != NULL); /* it's a bug if we don't have spare ccb here */
   1173 
   1174 	mem = nvme_dmamem_alloc(sc, sizeof(*identify));
   1175 	if (mem == NULL)
   1176 		return 1;
   1177 
   1178 	ccb->ccb_done = nvme_empty_done;
   1179 	ccb->ccb_cookie = mem;
   1180 
   1181 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_PREREAD);
   1182 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_fill_identify,
   1183 	    NVME_TIMO_IDENT);
   1184 	nvme_dmamem_sync(sc, mem, BUS_DMASYNC_POSTREAD);
   1185 
   1186 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1187 
   1188 	if (rv != 0)
   1189 		goto done;
   1190 
   1191 	identify = NVME_DMA_KVA(mem);
   1192 
   1193 	strnvisx(sn, sizeof(sn), (const char *)identify->sn,
   1194 	    sizeof(identify->sn), VIS_TRIM|VIS_SAFE|VIS_OCTAL);
   1195 	strnvisx(mn, sizeof(mn), (const char *)identify->mn,
   1196 	    sizeof(identify->mn), VIS_TRIM|VIS_SAFE|VIS_OCTAL);
   1197 	strnvisx(fr, sizeof(fr), (const char *)identify->fr,
   1198 	    sizeof(identify->fr), VIS_TRIM|VIS_SAFE|VIS_OCTAL);
   1199 	aprint_normal_dev(sc->sc_dev, "%s, firmware %s, serial %s\n", mn, fr,
   1200 	    sn);
   1201 
   1202 	if (identify->mdts > 0) {
   1203 		mdts = (1 << identify->mdts) * (1 << mps);
   1204 		if (mdts < sc->sc_mdts)
   1205 			sc->sc_mdts = mdts;
   1206 	}
   1207 
   1208 	sc->sc_nn = lemtoh32(&identify->nn);
   1209 
   1210 	memcpy(&sc->sc_identify, identify, sizeof(sc->sc_identify));
   1211 
   1212 done:
   1213 	nvme_dmamem_free(sc, mem);
   1214 
   1215 	return rv;
   1216 }
   1217 
   1218 static int
   1219 nvme_q_create(struct nvme_softc *sc, struct nvme_queue *q)
   1220 {
   1221 	struct nvme_sqe_q sqe;
   1222 	struct nvme_ccb *ccb;
   1223 	int rv;
   1224 
   1225 	if (sc->sc_use_mq && sc->sc_intr_establish(sc, q->q_id, q) != 0)
   1226 		return 1;
   1227 
   1228 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1229 	KASSERT(ccb != NULL);
   1230 
   1231 	ccb->ccb_done = nvme_empty_done;
   1232 	ccb->ccb_cookie = &sqe;
   1233 
   1234 	memset(&sqe, 0, sizeof(sqe));
   1235 	sqe.opcode = NVM_ADMIN_ADD_IOCQ;
   1236 	htolem64(&sqe.prp1, NVME_DMA_DVA(q->q_cq_dmamem));
   1237 	htolem16(&sqe.qsize, q->q_entries - 1);
   1238 	htolem16(&sqe.qid, q->q_id);
   1239 	sqe.qflags = NVM_SQE_CQ_IEN | NVM_SQE_Q_PC;
   1240 	if (sc->sc_use_mq)
   1241 		htolem16(&sqe.cqid, q->q_id);	/* qid == vector */
   1242 
   1243 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1244 	if (rv != 0)
   1245 		goto fail;
   1246 
   1247 	ccb->ccb_done = nvme_empty_done;
   1248 	ccb->ccb_cookie = &sqe;
   1249 
   1250 	memset(&sqe, 0, sizeof(sqe));
   1251 	sqe.opcode = NVM_ADMIN_ADD_IOSQ;
   1252 	htolem64(&sqe.prp1, NVME_DMA_DVA(q->q_sq_dmamem));
   1253 	htolem16(&sqe.qsize, q->q_entries - 1);
   1254 	htolem16(&sqe.qid, q->q_id);
   1255 	htolem16(&sqe.cqid, q->q_id);
   1256 	sqe.qflags = NVM_SQE_Q_PC;
   1257 
   1258 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1259 	if (rv != 0)
   1260 		goto fail;
   1261 
   1262 fail:
   1263 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1264 	return rv;
   1265 }
   1266 
   1267 static int
   1268 nvme_q_delete(struct nvme_softc *sc, struct nvme_queue *q)
   1269 {
   1270 	struct nvme_sqe_q sqe;
   1271 	struct nvme_ccb *ccb;
   1272 	int rv;
   1273 
   1274 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1275 	KASSERT(ccb != NULL);
   1276 
   1277 	ccb->ccb_done = nvme_empty_done;
   1278 	ccb->ccb_cookie = &sqe;
   1279 
   1280 	memset(&sqe, 0, sizeof(sqe));
   1281 	sqe.opcode = NVM_ADMIN_DEL_IOSQ;
   1282 	htolem16(&sqe.qid, q->q_id);
   1283 
   1284 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1285 	if (rv != 0)
   1286 		goto fail;
   1287 
   1288 	ccb->ccb_done = nvme_empty_done;
   1289 	ccb->ccb_cookie = &sqe;
   1290 
   1291 	memset(&sqe, 0, sizeof(sqe));
   1292 	sqe.opcode = NVM_ADMIN_DEL_IOCQ;
   1293 	htolem16(&sqe.qid, q->q_id);
   1294 
   1295 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP);
   1296 	if (rv != 0)
   1297 		goto fail;
   1298 
   1299 fail:
   1300 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1301 
   1302 	if (rv == 0 && sc->sc_use_mq) {
   1303 		if (sc->sc_intr_disestablish(sc, q->q_id))
   1304 			rv = 1;
   1305 	}
   1306 
   1307 	return rv;
   1308 }
   1309 
   1310 static void
   1311 nvme_fill_identify(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot)
   1312 {
   1313 	struct nvme_sqe *sqe = slot;
   1314 	struct nvme_dmamem *mem = ccb->ccb_cookie;
   1315 
   1316 	sqe->opcode = NVM_ADMIN_IDENTIFY;
   1317 	htolem64(&sqe->entry.prp[0], NVME_DMA_DVA(mem));
   1318 	htolem32(&sqe->cdw10, 1);
   1319 }
   1320 
   1321 static int
   1322 nvme_get_number_of_queues(struct nvme_softc *sc, u_int *nqap)
   1323 {
   1324 	struct nvme_pt_command pt;
   1325 	struct nvme_ccb *ccb;
   1326 	uint16_t ncqa, nsqa;
   1327 	int rv;
   1328 
   1329 	ccb = nvme_ccb_get(sc->sc_admin_q);
   1330 	KASSERT(ccb != NULL); /* it's a bug if we don't have spare ccb here */
   1331 
   1332 	memset(&pt, 0, sizeof(pt));
   1333 	pt.cmd.opcode = NVM_ADMIN_GET_FEATURES;
   1334 	pt.cmd.cdw10 = NVM_FEATURE_NUMBER_OF_QUEUES;
   1335 
   1336 	ccb->ccb_done = nvme_pt_done;
   1337 	ccb->ccb_cookie = &pt;
   1338 
   1339 	rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_pt_fill, NVME_TIMO_QOP);
   1340 
   1341 	nvme_ccb_put(sc->sc_admin_q, ccb);
   1342 
   1343 	if (rv != 0) {
   1344 		*nqap = 0;
   1345 		return EIO;
   1346 	}
   1347 
   1348 	ncqa = pt.cpl.cdw0 >> 16;
   1349 	nsqa = pt.cpl.cdw0 & 0xffff;
   1350 	*nqap = MIN(ncqa, nsqa) + 1;
   1351 
   1352 	return 0;
   1353 }
   1354 
   1355 static int
   1356 nvme_ccbs_alloc(struct nvme_queue *q, uint16_t nccbs)
   1357 {
   1358 	struct nvme_softc *sc = q->q_sc;
   1359 	struct nvme_ccb *ccb;
   1360 	bus_addr_t off;
   1361 	uint64_t *prpl;
   1362 	u_int i;
   1363 
   1364 	mutex_init(&q->q_ccb_mtx, MUTEX_DEFAULT, IPL_BIO);
   1365 	SIMPLEQ_INIT(&q->q_ccb_list);
   1366 
   1367 	q->q_ccbs = kmem_alloc(sizeof(*ccb) * nccbs, KM_SLEEP);
   1368 
   1369 	q->q_nccbs = nccbs;
   1370 	q->q_nccbs_avail = nccbs;
   1371 	q->q_ccb_prpls = nvme_dmamem_alloc(sc,
   1372 	    sizeof(*prpl) * sc->sc_max_sgl * nccbs);
   1373 
   1374 	prpl = NVME_DMA_KVA(q->q_ccb_prpls);
   1375 	off = 0;
   1376 
   1377 	for (i = 0; i < nccbs; i++) {
   1378 		ccb = &q->q_ccbs[i];
   1379 
   1380 		if (bus_dmamap_create(sc->sc_dmat, sc->sc_mdts,
   1381 		    sc->sc_max_sgl + 1 /* we get a free prp in the sqe */,
   1382 		    sc->sc_mps, sc->sc_mps, BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW,
   1383 		    &ccb->ccb_dmamap) != 0)
   1384 			goto free_maps;
   1385 
   1386 		ccb->ccb_id = i;
   1387 		ccb->ccb_prpl = prpl;
   1388 		ccb->ccb_prpl_off = off;
   1389 		ccb->ccb_prpl_dva = NVME_DMA_DVA(q->q_ccb_prpls) + off;
   1390 
   1391 		SIMPLEQ_INSERT_TAIL(&q->q_ccb_list, ccb, ccb_entry);
   1392 
   1393 		prpl += sc->sc_max_sgl;
   1394 		off += sizeof(*prpl) * sc->sc_max_sgl;
   1395 	}
   1396 
   1397 	return 0;
   1398 
   1399 free_maps:
   1400 	nvme_ccbs_free(q);
   1401 	return 1;
   1402 }
   1403 
   1404 static struct nvme_ccb *
   1405 nvme_ccb_get(struct nvme_queue *q)
   1406 {
   1407 	struct nvme_ccb *ccb = NULL;
   1408 
   1409 	mutex_enter(&q->q_ccb_mtx);
   1410 	if (q->q_nccbs_avail > 0) {
   1411 		ccb = SIMPLEQ_FIRST(&q->q_ccb_list);
   1412 		KASSERT(ccb != NULL);
   1413 		q->q_nccbs_avail--;
   1414 
   1415 		SIMPLEQ_REMOVE_HEAD(&q->q_ccb_list, ccb_entry);
   1416 #ifdef DEBUG
   1417 		ccb->ccb_cookie = NULL;
   1418 #endif
   1419 	}
   1420 	mutex_exit(&q->q_ccb_mtx);
   1421 
   1422 	return ccb;
   1423 }
   1424 
   1425 static void
   1426 nvme_ccb_put(struct nvme_queue *q, struct nvme_ccb *ccb)
   1427 {
   1428 
   1429 	mutex_enter(&q->q_ccb_mtx);
   1430 #ifdef DEBUG
   1431 	ccb->ccb_cookie = (void *)NVME_CCB_FREE;
   1432 #endif
   1433 	SIMPLEQ_INSERT_HEAD(&q->q_ccb_list, ccb, ccb_entry);
   1434 	mutex_exit(&q->q_ccb_mtx);
   1435 }
   1436 
   1437 static void
   1438 nvme_ccbs_free(struct nvme_queue *q)
   1439 {
   1440 	struct nvme_softc *sc = q->q_sc;
   1441 	struct nvme_ccb *ccb;
   1442 
   1443 	mutex_enter(&q->q_ccb_mtx);
   1444 	while ((ccb = SIMPLEQ_FIRST(&q->q_ccb_list)) != NULL) {
   1445 		SIMPLEQ_REMOVE_HEAD(&q->q_ccb_list, ccb_entry);
   1446 		bus_dmamap_destroy(sc->sc_dmat, ccb->ccb_dmamap);
   1447 	}
   1448 	mutex_exit(&q->q_ccb_mtx);
   1449 
   1450 	nvme_dmamem_free(sc, q->q_ccb_prpls);
   1451 	kmem_free(q->q_ccbs, sizeof(*ccb) * q->q_nccbs);
   1452 	q->q_ccbs = NULL;
   1453 	mutex_destroy(&q->q_ccb_mtx);
   1454 }
   1455 
   1456 static struct nvme_queue *
   1457 nvme_q_alloc(struct nvme_softc *sc, uint16_t id, u_int entries, u_int dstrd)
   1458 {
   1459 	struct nvme_queue *q;
   1460 
   1461 	q = kmem_alloc(sizeof(*q), KM_SLEEP);
   1462 	q->q_sc = sc;
   1463 	q->q_sq_dmamem = nvme_dmamem_alloc(sc,
   1464 	    sizeof(struct nvme_sqe) * entries);
   1465 	if (q->q_sq_dmamem == NULL)
   1466 		goto free;
   1467 
   1468 	q->q_cq_dmamem = nvme_dmamem_alloc(sc,
   1469 	    sizeof(struct nvme_cqe) * entries);
   1470 	if (q->q_cq_dmamem == NULL)
   1471 		goto free_sq;
   1472 
   1473 	memset(NVME_DMA_KVA(q->q_sq_dmamem), 0, NVME_DMA_LEN(q->q_sq_dmamem));
   1474 	memset(NVME_DMA_KVA(q->q_cq_dmamem), 0, NVME_DMA_LEN(q->q_cq_dmamem));
   1475 
   1476 	mutex_init(&q->q_sq_mtx, MUTEX_DEFAULT, IPL_BIO);
   1477 	mutex_init(&q->q_cq_mtx, MUTEX_DEFAULT, IPL_BIO);
   1478 	q->q_sqtdbl = NVME_SQTDBL(id, dstrd);
   1479 	q->q_cqhdbl = NVME_CQHDBL(id, dstrd);
   1480 	q->q_id = id;
   1481 	q->q_entries = entries;
   1482 	q->q_sq_tail = 0;
   1483 	q->q_cq_head = 0;
   1484 	q->q_cq_phase = NVME_CQE_PHASE;
   1485 
   1486 	nvme_dmamem_sync(sc, q->q_sq_dmamem, BUS_DMASYNC_PREWRITE);
   1487 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_PREREAD);
   1488 
   1489 	/*
   1490 	 * Due to definition of full and empty queue (queue is empty
   1491 	 * when head == tail, full when tail is one less then head),
   1492 	 * we can actually only have (entries - 1) in-flight commands.
   1493 	 */
   1494 	if (nvme_ccbs_alloc(q, entries - 1) != 0) {
   1495 		aprint_error_dev(sc->sc_dev, "unable to allocate ccbs\n");
   1496 		goto free_cq;
   1497 	}
   1498 
   1499 	return q;
   1500 
   1501 free_cq:
   1502 	nvme_dmamem_free(sc, q->q_cq_dmamem);
   1503 free_sq:
   1504 	nvme_dmamem_free(sc, q->q_sq_dmamem);
   1505 free:
   1506 	kmem_free(q, sizeof(*q));
   1507 
   1508 	return NULL;
   1509 }
   1510 
   1511 static void
   1512 nvme_q_free(struct nvme_softc *sc, struct nvme_queue *q)
   1513 {
   1514 	nvme_ccbs_free(q);
   1515 	mutex_destroy(&q->q_sq_mtx);
   1516 	mutex_destroy(&q->q_cq_mtx);
   1517 	nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_POSTREAD);
   1518 	nvme_dmamem_sync(sc, q->q_sq_dmamem, BUS_DMASYNC_POSTWRITE);
   1519 	nvme_dmamem_free(sc, q->q_cq_dmamem);
   1520 	nvme_dmamem_free(sc, q->q_sq_dmamem);
   1521 	kmem_free(q, sizeof(*q));
   1522 }
   1523 
   1524 int
   1525 nvme_intr(void *xsc)
   1526 {
   1527 	struct nvme_softc *sc = xsc;
   1528 
   1529 	/*
   1530 	 * INTx is level triggered, controller deasserts the interrupt only
   1531 	 * when we advance command queue head via write to the doorbell.
   1532 	 * Tell the controller to block the interrupts while we process
   1533 	 * the queue(s).
   1534 	 */
   1535 	nvme_write4(sc, NVME_INTMS, 1);
   1536 
   1537 	softint_schedule(sc->sc_softih[0]);
   1538 
   1539 	/* don't know, might not have been for us */
   1540 	return 1;
   1541 }
   1542 
   1543 void
   1544 nvme_softintr_intx(void *xq)
   1545 {
   1546 	struct nvme_queue *q = xq;
   1547 	struct nvme_softc *sc = q->q_sc;
   1548 
   1549 	nvme_q_complete(sc, sc->sc_admin_q);
   1550 	if (sc->sc_q != NULL)
   1551 	        nvme_q_complete(sc, sc->sc_q[0]);
   1552 
   1553 	/*
   1554 	 * Processing done, tell controller to issue interrupts again. There
   1555 	 * is no race, as NVMe spec requires the controller to maintain state,
   1556 	 * and assert the interrupt whenever there are unacknowledged
   1557 	 * completion queue entries.
   1558 	 */
   1559 	nvme_write4(sc, NVME_INTMC, 1);
   1560 }
   1561 
   1562 int
   1563 nvme_intr_msi(void *xq)
   1564 {
   1565 	struct nvme_queue *q = xq;
   1566 
   1567 	KASSERT(q && q->q_sc && q->q_sc->sc_softih
   1568 	    && q->q_sc->sc_softih[q->q_id]);
   1569 
   1570 	/*
   1571 	 * MSI/MSI-X are edge triggered, so can handover processing to softint
   1572 	 * without masking the interrupt.
   1573 	 */
   1574 	softint_schedule(q->q_sc->sc_softih[q->q_id]);
   1575 
   1576 	return 1;
   1577 }
   1578 
   1579 void
   1580 nvme_softintr_msi(void *xq)
   1581 {
   1582 	struct nvme_queue *q = xq;
   1583 	struct nvme_softc *sc = q->q_sc;
   1584 
   1585 	nvme_q_complete(sc, q);
   1586 }
   1587 
   1588 static struct nvme_dmamem *
   1589 nvme_dmamem_alloc(struct nvme_softc *sc, size_t size)
   1590 {
   1591 	struct nvme_dmamem *ndm;
   1592 	int nsegs;
   1593 
   1594 	ndm = kmem_zalloc(sizeof(*ndm), KM_SLEEP);
   1595 	if (ndm == NULL)
   1596 		return NULL;
   1597 
   1598 	ndm->ndm_size = size;
   1599 
   1600 	if (bus_dmamap_create(sc->sc_dmat, size, 1, size, 0,
   1601 	    BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW, &ndm->ndm_map) != 0)
   1602 		goto ndmfree;
   1603 
   1604 	if (bus_dmamem_alloc(sc->sc_dmat, size, sc->sc_mps, 0, &ndm->ndm_seg,
   1605 	    1, &nsegs, BUS_DMA_WAITOK) != 0)
   1606 		goto destroy;
   1607 
   1608 	if (bus_dmamem_map(sc->sc_dmat, &ndm->ndm_seg, nsegs, size,
   1609 	    &ndm->ndm_kva, BUS_DMA_WAITOK) != 0)
   1610 		goto free;
   1611 	memset(ndm->ndm_kva, 0, size);
   1612 
   1613 	if (bus_dmamap_load(sc->sc_dmat, ndm->ndm_map, ndm->ndm_kva, size,
   1614 	    NULL, BUS_DMA_WAITOK) != 0)
   1615 		goto unmap;
   1616 
   1617 	return ndm;
   1618 
   1619 unmap:
   1620 	bus_dmamem_unmap(sc->sc_dmat, ndm->ndm_kva, size);
   1621 free:
   1622 	bus_dmamem_free(sc->sc_dmat, &ndm->ndm_seg, 1);
   1623 destroy:
   1624 	bus_dmamap_destroy(sc->sc_dmat, ndm->ndm_map);
   1625 ndmfree:
   1626 	kmem_free(ndm, sizeof(*ndm));
   1627 	return NULL;
   1628 }
   1629 
   1630 static void
   1631 nvme_dmamem_sync(struct nvme_softc *sc, struct nvme_dmamem *mem, int ops)
   1632 {
   1633 	bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(mem),
   1634 	    0, NVME_DMA_LEN(mem), ops);
   1635 }
   1636 
   1637 void
   1638 nvme_dmamem_free(struct nvme_softc *sc, struct nvme_dmamem *ndm)
   1639 {
   1640 	bus_dmamap_unload(sc->sc_dmat, ndm->ndm_map);
   1641 	bus_dmamem_unmap(sc->sc_dmat, ndm->ndm_kva, ndm->ndm_size);
   1642 	bus_dmamem_free(sc->sc_dmat, &ndm->ndm_seg, 1);
   1643 	bus_dmamap_destroy(sc->sc_dmat, ndm->ndm_map);
   1644 	kmem_free(ndm, sizeof(*ndm));
   1645 }
   1646 
   1647 /*
   1648  * ioctl
   1649  */
   1650 
   1651 dev_type_open(nvmeopen);
   1652 dev_type_close(nvmeclose);
   1653 dev_type_ioctl(nvmeioctl);
   1654 
   1655 const struct cdevsw nvme_cdevsw = {
   1656 	.d_open = nvmeopen,
   1657 	.d_close = nvmeclose,
   1658 	.d_read = noread,
   1659 	.d_write = nowrite,
   1660 	.d_ioctl = nvmeioctl,
   1661 	.d_stop = nostop,
   1662 	.d_tty = notty,
   1663 	.d_poll = nopoll,
   1664 	.d_mmap = nommap,
   1665 	.d_kqfilter = nokqfilter,
   1666 	.d_discard = nodiscard,
   1667 	.d_flag = D_OTHER,
   1668 };
   1669 
   1670 /*
   1671  * Accept an open operation on the control device.
   1672  */
   1673 int
   1674 nvmeopen(dev_t dev, int flag, int mode, struct lwp *l)
   1675 {
   1676 	struct nvme_softc *sc;
   1677 	int unit = minor(dev) / 0x10000;
   1678 	int nsid = minor(dev) & 0xffff;
   1679 	int nsidx;
   1680 
   1681 	if ((sc = device_lookup_private(&nvme_cd, unit)) == NULL)
   1682 		return ENXIO;
   1683 	if ((sc->sc_flags & NVME_F_ATTACHED) == 0)
   1684 		return ENXIO;
   1685 
   1686 	if (nsid == 0) {
   1687 		/* controller */
   1688 		if (ISSET(sc->sc_flags, NVME_F_OPEN))
   1689 			return EBUSY;
   1690 		SET(sc->sc_flags, NVME_F_OPEN);
   1691 	} else {
   1692 		/* namespace */
   1693 		nsidx = nsid - 1;
   1694 		if (nsidx >= sc->sc_nn || sc->sc_namespaces[nsidx].dev == NULL)
   1695 			return ENXIO;
   1696 		if (ISSET(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN))
   1697 			return EBUSY;
   1698 		SET(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN);
   1699 	}
   1700 	return 0;
   1701 }
   1702 
   1703 /*
   1704  * Accept the last close on the control device.
   1705  */
   1706 int
   1707 nvmeclose(dev_t dev, int flag, int mode, struct lwp *l)
   1708 {
   1709 	struct nvme_softc *sc;
   1710 	int unit = minor(dev) / 0x10000;
   1711 	int nsid = minor(dev) & 0xffff;
   1712 	int nsidx;
   1713 
   1714 	sc = device_lookup_private(&nvme_cd, unit);
   1715 	if (sc == NULL)
   1716 		return ENXIO;
   1717 
   1718 	if (nsid == 0) {
   1719 		/* controller */
   1720 		CLR(sc->sc_flags, NVME_F_OPEN);
   1721 	} else {
   1722 		/* namespace */
   1723 		nsidx = nsid - 1;
   1724 		if (nsidx >= sc->sc_nn)
   1725 			return ENXIO;
   1726 		CLR(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN);
   1727 	}
   1728 
   1729 	return 0;
   1730 }
   1731 
   1732 /*
   1733  * Handle control operations.
   1734  */
   1735 int
   1736 nvmeioctl(dev_t dev, u_long cmd, void *data, int flag, struct lwp *l)
   1737 {
   1738 	struct nvme_softc *sc;
   1739 	int unit = minor(dev) / 0x10000;
   1740 	int nsid = minor(dev) & 0xffff;
   1741 	struct nvme_pt_command *pt;
   1742 
   1743 	sc = device_lookup_private(&nvme_cd, unit);
   1744 	if (sc == NULL)
   1745 		return ENXIO;
   1746 
   1747 	switch (cmd) {
   1748 	case NVME_PASSTHROUGH_CMD:
   1749 		pt = data;
   1750 		return nvme_command_passthrough(sc, data,
   1751 		    nsid == 0 ? pt->cmd.nsid : nsid, l, nsid == 0);
   1752 	}
   1753 
   1754 	return ENOTTY;
   1755 }
   1756