xref: /freebsd/usr.sbin/bhyve/pci_nvme.c (revision 090e9752d7291db0c251a5576892e3bbbaea2479)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2017 Shunsuke Mie
5  * Copyright (c) 2018 Leon Dang
6  * Copyright (c) 2020 Chuck Tuffli
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 /*
31  * bhyve PCIe-NVMe device emulation.
32  *
33  * options:
34  *  -s <n>,nvme,devpath,maxq=#,qsz=#,ioslots=#,sectsz=#,ser=A-Z,eui64=#,dsm=<opt>
35  *
36  *  accepted devpath:
37  *    /dev/blockdev
38  *    /path/to/image
39  *    ram=size_in_MiB
40  *
41  *  maxq    = max number of queues
42  *  qsz     = max elements in each queue
43  *  ioslots = max number of concurrent io requests
44  *  sectsz  = sector size (defaults to blockif sector size)
45  *  ser     = serial number (20-chars max)
46  *  eui64   = IEEE Extended Unique Identifier (8 byte value)
47  *  dsm     = DataSet Management support. Option is one of auto, enable,disable
48  *
49  */
50 
51 /* TODO:
52     - create async event for smart and log
53     - intr coalesce
54  */
55 
56 #include <sys/cdefs.h>
57 __FBSDID("$FreeBSD$");
58 
59 #include <sys/errno.h>
60 #include <sys/types.h>
61 #include <sys/crc16.h>
62 #include <net/ieee_oui.h>
63 
64 #include <assert.h>
65 #include <pthread.h>
66 #include <pthread_np.h>
67 #include <semaphore.h>
68 #include <stdbool.h>
69 #include <stddef.h>
70 #include <stdint.h>
71 #include <stdio.h>
72 #include <stdlib.h>
73 #include <string.h>
74 
75 #include <machine/atomic.h>
76 #include <machine/vmm.h>
77 #include <vmmapi.h>
78 
79 #include <dev/nvme/nvme.h>
80 
81 #include "bhyverun.h"
82 #include "block_if.h"
83 #include "config.h"
84 #include "debug.h"
85 #include "pci_emul.h"
86 
87 
88 static int nvme_debug = 0;
89 #define	DPRINTF(fmt, args...) if (nvme_debug) PRINTLN(fmt, ##args)
90 #define	WPRINTF(fmt, args...) PRINTLN(fmt, ##args)
91 
92 /* defaults; can be overridden */
93 #define	NVME_MSIX_BAR		4
94 
95 #define	NVME_IOSLOTS		8
96 
97 /* The NVMe spec defines bits 13:4 in BAR0 as reserved */
98 #define NVME_MMIO_SPACE_MIN	(1 << 14)
99 
100 #define	NVME_QUEUES		16
101 #define	NVME_MAX_QENTRIES	2048
102 /* Memory Page size Minimum reported in CAP register */
103 #define	NVME_MPSMIN		0
104 /* MPSMIN converted to bytes */
105 #define	NVME_MPSMIN_BYTES	(1 << (12 + NVME_MPSMIN))
106 
107 #define	NVME_PRP2_ITEMS		(PAGE_SIZE/sizeof(uint64_t))
108 #define	NVME_MDTS		9
109 /* Note the + 1 allows for the initial descriptor to not be page aligned */
110 #define	NVME_MAX_IOVEC		((1 << NVME_MDTS) + 1)
111 #define	NVME_MAX_DATA_SIZE	((1 << NVME_MDTS) * NVME_MPSMIN_BYTES)
112 
113 /* This is a synthetic status code to indicate there is no status */
114 #define NVME_NO_STATUS		0xffff
115 #define NVME_COMPLETION_VALID(c)	((c).status != NVME_NO_STATUS)
116 
117 /* Reported temperature in Kelvin (i.e. room temperature) */
118 #define NVME_TEMPERATURE 296
119 
120 /* helpers */
121 
122 /* Convert a zero-based value into a one-based value */
123 #define ONE_BASED(zero)		((zero) + 1)
124 /* Convert a one-based value into a zero-based value */
125 #define ZERO_BASED(one)		((one)  - 1)
126 
127 /* Encode number of SQ's and CQ's for Set/Get Features */
128 #define NVME_FEATURE_NUM_QUEUES(sc) \
129 	(ZERO_BASED((sc)->num_squeues) & 0xffff) | \
130 	(ZERO_BASED((sc)->num_cqueues) & 0xffff) << 16
131 
132 #define	NVME_DOORBELL_OFFSET	offsetof(struct nvme_registers, doorbell)
133 
134 enum nvme_controller_register_offsets {
135 	NVME_CR_CAP_LOW = 0x00,
136 	NVME_CR_CAP_HI  = 0x04,
137 	NVME_CR_VS      = 0x08,
138 	NVME_CR_INTMS   = 0x0c,
139 	NVME_CR_INTMC   = 0x10,
140 	NVME_CR_CC      = 0x14,
141 	NVME_CR_CSTS    = 0x1c,
142 	NVME_CR_NSSR    = 0x20,
143 	NVME_CR_AQA     = 0x24,
144 	NVME_CR_ASQ_LOW = 0x28,
145 	NVME_CR_ASQ_HI  = 0x2c,
146 	NVME_CR_ACQ_LOW = 0x30,
147 	NVME_CR_ACQ_HI  = 0x34,
148 };
149 
150 enum nvme_cmd_cdw11 {
151 	NVME_CMD_CDW11_PC  = 0x0001,
152 	NVME_CMD_CDW11_IEN = 0x0002,
153 	NVME_CMD_CDW11_IV  = 0xFFFF0000,
154 };
155 
156 enum nvme_copy_dir {
157 	NVME_COPY_TO_PRP,
158 	NVME_COPY_FROM_PRP,
159 };
160 
161 #define	NVME_CQ_INTEN	0x01
162 #define	NVME_CQ_INTCOAL	0x02
163 
164 struct nvme_completion_queue {
165 	struct nvme_completion *qbase;
166 	pthread_mutex_t	mtx;
167 	uint32_t	size;
168 	uint16_t	tail; /* nvme progress */
169 	uint16_t	head; /* guest progress */
170 	uint16_t	intr_vec;
171 	uint32_t	intr_en;
172 };
173 
174 struct nvme_submission_queue {
175 	struct nvme_command *qbase;
176 	pthread_mutex_t	mtx;
177 	uint32_t	size;
178 	uint16_t	head; /* nvme progress */
179 	uint16_t	tail; /* guest progress */
180 	uint16_t	cqid; /* completion queue id */
181 	int		qpriority;
182 };
183 
184 enum nvme_storage_type {
185 	NVME_STOR_BLOCKIF = 0,
186 	NVME_STOR_RAM = 1,
187 };
188 
189 struct pci_nvme_blockstore {
190 	enum nvme_storage_type type;
191 	void		*ctx;
192 	uint64_t	size;
193 	uint32_t	sectsz;
194 	uint32_t	sectsz_bits;
195 	uint64_t	eui64;
196 	uint32_t	deallocate:1;
197 };
198 
199 /*
200  * Calculate the number of additional page descriptors for guest IO requests
201  * based on the advertised Max Data Transfer (MDTS) and given the number of
202  * default iovec's in a struct blockif_req.
203  */
204 #define MDTS_PAD_SIZE \
205 	( NVME_MAX_IOVEC > BLOCKIF_IOV_MAX ? \
206 	  NVME_MAX_IOVEC - BLOCKIF_IOV_MAX : \
207 	  0 )
208 
209 struct pci_nvme_ioreq {
210 	struct pci_nvme_softc *sc;
211 	STAILQ_ENTRY(pci_nvme_ioreq) link;
212 	struct nvme_submission_queue *nvme_sq;
213 	uint16_t	sqid;
214 
215 	/* command information */
216 	uint16_t	opc;
217 	uint16_t	cid;
218 	uint32_t	nsid;
219 
220 	uint64_t	prev_gpaddr;
221 	size_t		prev_size;
222 	size_t		bytes;
223 
224 	struct blockif_req io_req;
225 
226 	struct iovec	iovpadding[MDTS_PAD_SIZE];
227 };
228 
229 enum nvme_dsm_type {
230 	/* Dataset Management bit in ONCS reflects backing storage capability */
231 	NVME_DATASET_MANAGEMENT_AUTO,
232 	/* Unconditionally set Dataset Management bit in ONCS */
233 	NVME_DATASET_MANAGEMENT_ENABLE,
234 	/* Unconditionally clear Dataset Management bit in ONCS */
235 	NVME_DATASET_MANAGEMENT_DISABLE,
236 };
237 
238 struct pci_nvme_softc;
239 struct nvme_feature_obj;
240 
241 typedef void (*nvme_feature_cb)(struct pci_nvme_softc *,
242     struct nvme_feature_obj *,
243     struct nvme_command *,
244     struct nvme_completion *);
245 
246 struct nvme_feature_obj {
247 	uint32_t	cdw11;
248 	nvme_feature_cb	set;
249 	nvme_feature_cb	get;
250 	bool namespace_specific;
251 };
252 
253 #define NVME_FID_MAX		(NVME_FEAT_ENDURANCE_GROUP_EVENT_CONFIGURATION + 1)
254 
255 typedef enum {
256 	PCI_NVME_AE_TYPE_ERROR = 0,
257 	PCI_NVME_AE_TYPE_SMART,
258 	PCI_NVME_AE_TYPE_NOTICE,
259 	PCI_NVME_AE_TYPE_IO_CMD = 6,
260 	PCI_NVME_AE_TYPE_VENDOR = 7,
261 	PCI_NVME_AE_TYPE_MAX		/* Must be last */
262 } pci_nvme_async_type;
263 
264 /* Asynchronous Event Requests */
265 struct pci_nvme_aer {
266 	STAILQ_ENTRY(pci_nvme_aer) link;
267 	uint16_t	cid;	/* Command ID of the submitted AER */
268 };
269 
270 /** Asynchronous Event Information - Notice */
271 typedef enum {
272 	PCI_NVME_AEI_NOTICE_NS_ATTR_CHANGED = 0,
273 	PCI_NVME_AEI_NOTICE_FW_ACTIVATION,
274 	PCI_NVME_AEI_NOTICE_TELEMETRY_CHANGE,
275 	PCI_NVME_AEI_NOTICE_ANA_CHANGE,
276 	PCI_NVME_AEI_NOTICE_PREDICT_LATENCY_CHANGE,
277 	PCI_NVME_AEI_NOTICE_LBA_STATUS_ALERT,
278 	PCI_NVME_AEI_NOTICE_ENDURANCE_GROUP_CHANGE,
279 	PCI_NVME_AEI_NOTICE_MAX,
280 } pci_nvme_async_event_info_notice;
281 
282 #define PCI_NVME_AEI_NOTICE_SHIFT		8
283 #define PCI_NVME_AEI_NOTICE_MASK(event)	(1 << (event + PCI_NVME_AEI_NOTICE_SHIFT))
284 
285 /* Asynchronous Event Notifications */
286 struct pci_nvme_aen {
287 	pci_nvme_async_type atype;
288 	uint32_t	event_data;
289 	bool		posted;
290 };
291 
292 /*
293  * By default, enable all Asynchrnous Event Notifications:
294  *     SMART / Health Critical Warnings
295  *     Namespace Attribute Notices
296  */
297 #define PCI_NVME_AEN_DEFAULT_MASK	0x11f
298 
299 typedef enum {
300 	NVME_CNTRLTYPE_IO = 1,
301 	NVME_CNTRLTYPE_DISCOVERY = 2,
302 	NVME_CNTRLTYPE_ADMIN = 3,
303 } pci_nvme_cntrl_type;
304 
305 struct pci_nvme_softc {
306 	struct pci_devinst *nsc_pi;
307 
308 	pthread_mutex_t	mtx;
309 
310 	struct nvme_registers regs;
311 
312 	struct nvme_namespace_data  nsdata;
313 	struct nvme_controller_data ctrldata;
314 	struct nvme_error_information_entry err_log;
315 	struct nvme_health_information_page health_log;
316 	struct nvme_firmware_page fw_log;
317 	struct nvme_ns_list ns_log;
318 
319 	struct pci_nvme_blockstore nvstore;
320 
321 	uint16_t	max_qentries;	/* max entries per queue */
322 	uint32_t	max_queues;	/* max number of IO SQ's or CQ's */
323 	uint32_t	num_cqueues;
324 	uint32_t	num_squeues;
325 	bool		num_q_is_set; /* Has host set Number of Queues */
326 
327 	struct pci_nvme_ioreq *ioreqs;
328 	STAILQ_HEAD(, pci_nvme_ioreq) ioreqs_free; /* free list of ioreqs */
329 	uint32_t	pending_ios;
330 	uint32_t	ioslots;
331 	sem_t		iosemlock;
332 
333 	/*
334 	 * Memory mapped Submission and Completion queues
335 	 * Each array includes both Admin and IO queues
336 	 */
337 	struct nvme_completion_queue *compl_queues;
338 	struct nvme_submission_queue *submit_queues;
339 
340 	struct nvme_feature_obj feat[NVME_FID_MAX];
341 
342 	enum nvme_dsm_type dataset_management;
343 
344 	/* Accounting for SMART data */
345 	__uint128_t	read_data_units;
346 	__uint128_t	write_data_units;
347 	__uint128_t	read_commands;
348 	__uint128_t	write_commands;
349 	uint32_t	read_dunits_remainder;
350 	uint32_t	write_dunits_remainder;
351 
352 	STAILQ_HEAD(, pci_nvme_aer) aer_list;
353 	pthread_mutex_t	aer_mtx;
354 	uint32_t	aer_count;
355 	struct pci_nvme_aen aen[PCI_NVME_AE_TYPE_MAX];
356 	pthread_t	aen_tid;
357 	pthread_mutex_t	aen_mtx;
358 	pthread_cond_t	aen_cond;
359 };
360 
361 
362 static void pci_nvme_cq_update(struct pci_nvme_softc *sc,
363     struct nvme_completion_queue *cq,
364     uint32_t cdw0,
365     uint16_t cid,
366     uint16_t sqid,
367     uint16_t status);
368 static struct pci_nvme_ioreq *pci_nvme_get_ioreq(struct pci_nvme_softc *);
369 static void pci_nvme_release_ioreq(struct pci_nvme_softc *, struct pci_nvme_ioreq *);
370 static void pci_nvme_io_done(struct blockif_req *, int);
371 
372 /* Controller Configuration utils */
373 #define	NVME_CC_GET_EN(cc) \
374 	((cc) >> NVME_CC_REG_EN_SHIFT & NVME_CC_REG_EN_MASK)
375 #define	NVME_CC_GET_CSS(cc) \
376 	((cc) >> NVME_CC_REG_CSS_SHIFT & NVME_CC_REG_CSS_MASK)
377 #define	NVME_CC_GET_SHN(cc) \
378 	((cc) >> NVME_CC_REG_SHN_SHIFT & NVME_CC_REG_SHN_MASK)
379 #define	NVME_CC_GET_IOSQES(cc) \
380 	((cc) >> NVME_CC_REG_IOSQES_SHIFT & NVME_CC_REG_IOSQES_MASK)
381 #define	NVME_CC_GET_IOCQES(cc) \
382 	((cc) >> NVME_CC_REG_IOCQES_SHIFT & NVME_CC_REG_IOCQES_MASK)
383 
384 #define	NVME_CC_WRITE_MASK \
385 	((NVME_CC_REG_EN_MASK << NVME_CC_REG_EN_SHIFT) | \
386 	 (NVME_CC_REG_IOSQES_MASK << NVME_CC_REG_IOSQES_SHIFT) | \
387 	 (NVME_CC_REG_IOCQES_MASK << NVME_CC_REG_IOCQES_SHIFT))
388 
389 #define	NVME_CC_NEN_WRITE_MASK \
390 	((NVME_CC_REG_CSS_MASK << NVME_CC_REG_CSS_SHIFT) | \
391 	 (NVME_CC_REG_MPS_MASK << NVME_CC_REG_MPS_SHIFT) | \
392 	 (NVME_CC_REG_AMS_MASK << NVME_CC_REG_AMS_SHIFT))
393 
394 /* Controller Status utils */
395 #define	NVME_CSTS_GET_RDY(sts) \
396 	((sts) >> NVME_CSTS_REG_RDY_SHIFT & NVME_CSTS_REG_RDY_MASK)
397 
398 #define	NVME_CSTS_RDY	(1 << NVME_CSTS_REG_RDY_SHIFT)
399 #define	NVME_CSTS_CFS	(1 << NVME_CSTS_REG_CFS_SHIFT)
400 
401 /* Completion Queue status word utils */
402 #define	NVME_STATUS_P	(1 << NVME_STATUS_P_SHIFT)
403 #define	NVME_STATUS_MASK \
404 	((NVME_STATUS_SCT_MASK << NVME_STATUS_SCT_SHIFT) |\
405 	 (NVME_STATUS_SC_MASK << NVME_STATUS_SC_SHIFT))
406 
407 #define NVME_ONCS_DSM	(NVME_CTRLR_DATA_ONCS_DSM_MASK << \
408 	NVME_CTRLR_DATA_ONCS_DSM_SHIFT)
409 
410 static void nvme_feature_invalid_cb(struct pci_nvme_softc *,
411     struct nvme_feature_obj *,
412     struct nvme_command *,
413     struct nvme_completion *);
414 static void nvme_feature_temperature(struct pci_nvme_softc *,
415     struct nvme_feature_obj *,
416     struct nvme_command *,
417     struct nvme_completion *);
418 static void nvme_feature_num_queues(struct pci_nvme_softc *,
419     struct nvme_feature_obj *,
420     struct nvme_command *,
421     struct nvme_completion *);
422 static void nvme_feature_iv_config(struct pci_nvme_softc *,
423     struct nvme_feature_obj *,
424     struct nvme_command *,
425     struct nvme_completion *);
426 static void nvme_feature_async_event(struct pci_nvme_softc *,
427     struct nvme_feature_obj *,
428     struct nvme_command *,
429     struct nvme_completion *);
430 
431 static void *aen_thr(void *arg);
432 
433 static __inline void
434 cpywithpad(char *dst, size_t dst_size, const char *src, char pad)
435 {
436 	size_t len;
437 
438 	len = strnlen(src, dst_size);
439 	memset(dst, pad, dst_size);
440 	memcpy(dst, src, len);
441 }
442 
443 static __inline void
444 pci_nvme_status_tc(uint16_t *status, uint16_t type, uint16_t code)
445 {
446 
447 	*status &= ~NVME_STATUS_MASK;
448 	*status |= (type & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT |
449 		(code & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
450 }
451 
452 static __inline void
453 pci_nvme_status_genc(uint16_t *status, uint16_t code)
454 {
455 
456 	pci_nvme_status_tc(status, NVME_SCT_GENERIC, code);
457 }
458 
459 /*
460  * Initialize the requested number or IO Submission and Completion Queues.
461  * Admin queues are allocated implicitly.
462  */
463 static void
464 pci_nvme_init_queues(struct pci_nvme_softc *sc, uint32_t nsq, uint32_t ncq)
465 {
466 	uint32_t i;
467 
468 	/*
469 	 * Allocate and initialize the Submission Queues
470 	 */
471 	if (nsq > NVME_QUEUES) {
472 		WPRINTF("%s: clamping number of SQ from %u to %u",
473 					__func__, nsq, NVME_QUEUES);
474 		nsq = NVME_QUEUES;
475 	}
476 
477 	sc->num_squeues = nsq;
478 
479 	sc->submit_queues = calloc(sc->num_squeues + 1,
480 				sizeof(struct nvme_submission_queue));
481 	if (sc->submit_queues == NULL) {
482 		WPRINTF("%s: SQ allocation failed", __func__);
483 		sc->num_squeues = 0;
484 	} else {
485 		struct nvme_submission_queue *sq = sc->submit_queues;
486 
487 		for (i = 0; i < sc->num_squeues + 1; i++)
488 			pthread_mutex_init(&sq[i].mtx, NULL);
489 	}
490 
491 	/*
492 	 * Allocate and initialize the Completion Queues
493 	 */
494 	if (ncq > NVME_QUEUES) {
495 		WPRINTF("%s: clamping number of CQ from %u to %u",
496 					__func__, ncq, NVME_QUEUES);
497 		ncq = NVME_QUEUES;
498 	}
499 
500 	sc->num_cqueues = ncq;
501 
502 	sc->compl_queues = calloc(sc->num_cqueues + 1,
503 				sizeof(struct nvme_completion_queue));
504 	if (sc->compl_queues == NULL) {
505 		WPRINTF("%s: CQ allocation failed", __func__);
506 		sc->num_cqueues = 0;
507 	} else {
508 		struct nvme_completion_queue *cq = sc->compl_queues;
509 
510 		for (i = 0; i < sc->num_cqueues + 1; i++)
511 			pthread_mutex_init(&cq[i].mtx, NULL);
512 	}
513 }
514 
515 static void
516 pci_nvme_init_ctrldata(struct pci_nvme_softc *sc)
517 {
518 	struct nvme_controller_data *cd = &sc->ctrldata;
519 
520 	cd->vid = 0xFB5D;
521 	cd->ssvid = 0x0000;
522 
523 	cpywithpad((char *)cd->mn, sizeof(cd->mn), "bhyve-NVMe", ' ');
524 	cpywithpad((char *)cd->fr, sizeof(cd->fr), "1.0", ' ');
525 
526 	/* Num of submission commands that we can handle at a time (2^rab) */
527 	cd->rab   = 4;
528 
529 	/* FreeBSD OUI */
530 	cd->ieee[0] = 0x58;
531 	cd->ieee[1] = 0x9c;
532 	cd->ieee[2] = 0xfc;
533 
534 	cd->mic = 0;
535 
536 	cd->mdts = NVME_MDTS;	/* max data transfer size (2^mdts * CAP.MPSMIN) */
537 
538 	cd->ver = NVME_REV(1,4);
539 
540 	cd->cntrltype = NVME_CNTRLTYPE_IO;
541 	cd->oacs = 1 << NVME_CTRLR_DATA_OACS_FORMAT_SHIFT;
542 	cd->oaes = NVMEB(NVME_CTRLR_DATA_OAES_NS_ATTR);
543 	cd->acl = 2;
544 	cd->aerl = 4;
545 
546 	/* Advertise 1, Read-only firmware slot */
547 	cd->frmw = NVMEB(NVME_CTRLR_DATA_FRMW_SLOT1_RO) |
548 	    (1 << NVME_CTRLR_DATA_FRMW_NUM_SLOTS_SHIFT);
549 	cd->lpa = 0;	/* TODO: support some simple things like SMART */
550 	cd->elpe = 0;	/* max error log page entries */
551 	/*
552 	 * Report a single power state (zero-based value)
553 	 * power_state[] values are left as zero to indicate "Not reported"
554 	 */
555 	cd->npss = 0;
556 
557 	/* Warning Composite Temperature Threshold */
558 	cd->wctemp = 0x0157;
559 	cd->cctemp = 0x0157;
560 
561 	/* SANICAP must not be 0 for Revision 1.4 and later NVMe Controllers */
562 	cd->sanicap = (NVME_CTRLR_DATA_SANICAP_NODMMAS_NO <<
563 			NVME_CTRLR_DATA_SANICAP_NODMMAS_SHIFT);
564 
565 	cd->sqes = (6 << NVME_CTRLR_DATA_SQES_MAX_SHIFT) |
566 	    (6 << NVME_CTRLR_DATA_SQES_MIN_SHIFT);
567 	cd->cqes = (4 << NVME_CTRLR_DATA_CQES_MAX_SHIFT) |
568 	    (4 << NVME_CTRLR_DATA_CQES_MIN_SHIFT);
569 	cd->nn = 1;	/* number of namespaces */
570 
571 	cd->oncs = 0;
572 	switch (sc->dataset_management) {
573 	case NVME_DATASET_MANAGEMENT_AUTO:
574 		if (sc->nvstore.deallocate)
575 			cd->oncs |= NVME_ONCS_DSM;
576 		break;
577 	case NVME_DATASET_MANAGEMENT_ENABLE:
578 		cd->oncs |= NVME_ONCS_DSM;
579 		break;
580 	default:
581 		break;
582 	}
583 
584 	cd->fna = NVME_CTRLR_DATA_FNA_FORMAT_ALL_MASK <<
585 	    NVME_CTRLR_DATA_FNA_FORMAT_ALL_SHIFT;
586 
587 	cd->vwc = NVME_CTRLR_DATA_VWC_ALL_NO << NVME_CTRLR_DATA_VWC_ALL_SHIFT;
588 }
589 
590 static void
591 pci_nvme_init_nsdata_size(struct pci_nvme_blockstore *nvstore,
592     struct nvme_namespace_data *nd)
593 {
594 
595 	/* Get capacity and block size information from backing store */
596 	nd->nsze = nvstore->size / nvstore->sectsz;
597 	nd->ncap = nd->nsze;
598 	nd->nuse = nd->nsze;
599 }
600 
601 static void
602 pci_nvme_init_nsdata(struct pci_nvme_softc *sc,
603     struct nvme_namespace_data *nd, uint32_t nsid,
604     struct pci_nvme_blockstore *nvstore)
605 {
606 
607 	pci_nvme_init_nsdata_size(nvstore, nd);
608 
609 	if (nvstore->type == NVME_STOR_BLOCKIF)
610 		nvstore->deallocate = blockif_candelete(nvstore->ctx);
611 
612 	nd->nlbaf = 0; /* NLBAF is a 0's based value (i.e. 1 LBA Format) */
613 	nd->flbas = 0;
614 
615 	/* Create an EUI-64 if user did not provide one */
616 	if (nvstore->eui64 == 0) {
617 		char *data = NULL;
618 		uint64_t eui64 = nvstore->eui64;
619 
620 		asprintf(&data, "%s%u%u%u", get_config_value("name"),
621 		    sc->nsc_pi->pi_bus, sc->nsc_pi->pi_slot,
622 		    sc->nsc_pi->pi_func);
623 
624 		if (data != NULL) {
625 			eui64 = OUI_FREEBSD_NVME_LOW | crc16(0, data, strlen(data));
626 			free(data);
627 		}
628 		nvstore->eui64 = (eui64 << 16) | (nsid & 0xffff);
629 	}
630 	be64enc(nd->eui64, nvstore->eui64);
631 
632 	/* LBA data-sz = 2^lbads */
633 	nd->lbaf[0] = nvstore->sectsz_bits << NVME_NS_DATA_LBAF_LBADS_SHIFT;
634 }
635 
636 static void
637 pci_nvme_init_logpages(struct pci_nvme_softc *sc)
638 {
639 	__uint128_t power_cycles = 1;
640 
641 	memset(&sc->err_log, 0, sizeof(sc->err_log));
642 	memset(&sc->health_log, 0, sizeof(sc->health_log));
643 	memset(&sc->fw_log, 0, sizeof(sc->fw_log));
644 	memset(&sc->ns_log, 0, sizeof(sc->ns_log));
645 
646 	/* Set read/write remainder to round up according to spec */
647 	sc->read_dunits_remainder = 999;
648 	sc->write_dunits_remainder = 999;
649 
650 	/* Set nominal Health values checked by implementations */
651 	sc->health_log.temperature = NVME_TEMPERATURE;
652 	sc->health_log.available_spare = 100;
653 	sc->health_log.available_spare_threshold = 10;
654 
655 	/* Set Active Firmware Info to slot 1 */
656 	sc->fw_log.afi = (1 << NVME_FIRMWARE_PAGE_AFI_SLOT_SHIFT);
657 	memcpy(&sc->fw_log.revision[0], sc->ctrldata.fr,
658 	    sizeof(sc->fw_log.revision[0]));
659 
660 	memcpy(&sc->health_log.power_cycles, &power_cycles,
661 	    sizeof(sc->health_log.power_cycles));
662 }
663 
664 static void
665 pci_nvme_init_features(struct pci_nvme_softc *sc)
666 {
667 	enum nvme_feature	fid;
668 
669 	for (fid = 0; fid < NVME_FID_MAX; fid++) {
670 		switch (fid) {
671 		case NVME_FEAT_ARBITRATION:
672 		case NVME_FEAT_POWER_MANAGEMENT:
673 		case NVME_FEAT_INTERRUPT_COALESCING: //XXX
674 		case NVME_FEAT_WRITE_ATOMICITY:
675 			/* Mandatory but no special handling required */
676 		//XXX hang - case NVME_FEAT_PREDICTABLE_LATENCY_MODE_CONFIG:
677 		//XXX hang - case NVME_FEAT_HOST_BEHAVIOR_SUPPORT:
678 		//		  this returns a data buffer
679 			break;
680 		case NVME_FEAT_TEMPERATURE_THRESHOLD:
681 			sc->feat[fid].set = nvme_feature_temperature;
682 			break;
683 		case NVME_FEAT_ERROR_RECOVERY:
684 			sc->feat[fid].namespace_specific = true;
685 			break;
686 		case NVME_FEAT_NUMBER_OF_QUEUES:
687 			sc->feat[fid].set = nvme_feature_num_queues;
688 			break;
689 		case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
690 			sc->feat[fid].set = nvme_feature_iv_config;
691 			break;
692 		case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
693 			sc->feat[fid].set = nvme_feature_async_event;
694 			/* Enable all AENs by default */
695 			sc->feat[fid].cdw11 = PCI_NVME_AEN_DEFAULT_MASK;
696 			break;
697 		default:
698 			sc->feat[fid].set = nvme_feature_invalid_cb;
699 			sc->feat[fid].get = nvme_feature_invalid_cb;
700 		}
701 	}
702 }
703 
704 static void
705 pci_nvme_aer_reset(struct pci_nvme_softc *sc)
706 {
707 
708 	STAILQ_INIT(&sc->aer_list);
709 	sc->aer_count = 0;
710 }
711 
712 static void
713 pci_nvme_aer_init(struct pci_nvme_softc *sc)
714 {
715 
716 	pthread_mutex_init(&sc->aer_mtx, NULL);
717 	pci_nvme_aer_reset(sc);
718 }
719 
720 static void
721 pci_nvme_aer_destroy(struct pci_nvme_softc *sc)
722 {
723 	struct pci_nvme_aer *aer = NULL;
724 
725 	pthread_mutex_lock(&sc->aer_mtx);
726 	while (!STAILQ_EMPTY(&sc->aer_list)) {
727 		aer = STAILQ_FIRST(&sc->aer_list);
728 		STAILQ_REMOVE_HEAD(&sc->aer_list, link);
729 		free(aer);
730 	}
731 	pthread_mutex_unlock(&sc->aer_mtx);
732 
733 	pci_nvme_aer_reset(sc);
734 }
735 
736 static bool
737 pci_nvme_aer_available(struct pci_nvme_softc *sc)
738 {
739 
740 	return (sc->aer_count != 0);
741 }
742 
743 static bool
744 pci_nvme_aer_limit_reached(struct pci_nvme_softc *sc)
745 {
746 	struct nvme_controller_data *cd = &sc->ctrldata;
747 
748 	/* AERL is a zero based value while aer_count is one's based */
749 	return (sc->aer_count == (cd->aerl + 1U));
750 }
751 
752 /*
753  * Add an Async Event Request
754  *
755  * Stores an AER to be returned later if the Controller needs to notify the
756  * host of an event.
757  * Note that while the NVMe spec doesn't require Controllers to return AER's
758  * in order, this implementation does preserve the order.
759  */
760 static int
761 pci_nvme_aer_add(struct pci_nvme_softc *sc, uint16_t cid)
762 {
763 	struct pci_nvme_aer *aer = NULL;
764 
765 	aer = calloc(1, sizeof(struct pci_nvme_aer));
766 	if (aer == NULL)
767 		return (-1);
768 
769 	/* Save the Command ID for use in the completion message */
770 	aer->cid = cid;
771 
772 	pthread_mutex_lock(&sc->aer_mtx);
773 	sc->aer_count++;
774 	STAILQ_INSERT_TAIL(&sc->aer_list, aer, link);
775 	pthread_mutex_unlock(&sc->aer_mtx);
776 
777 	return (0);
778 }
779 
780 /*
781  * Get an Async Event Request structure
782  *
783  * Returns a pointer to an AER previously submitted by the host or NULL if
784  * no AER's exist. Caller is responsible for freeing the returned struct.
785  */
786 static struct pci_nvme_aer *
787 pci_nvme_aer_get(struct pci_nvme_softc *sc)
788 {
789 	struct pci_nvme_aer *aer = NULL;
790 
791 	pthread_mutex_lock(&sc->aer_mtx);
792 	aer = STAILQ_FIRST(&sc->aer_list);
793 	if (aer != NULL) {
794 		STAILQ_REMOVE_HEAD(&sc->aer_list, link);
795 		sc->aer_count--;
796 	}
797 	pthread_mutex_unlock(&sc->aer_mtx);
798 
799 	return (aer);
800 }
801 
802 static void
803 pci_nvme_aen_reset(struct pci_nvme_softc *sc)
804 {
805 	uint32_t	atype;
806 
807 	memset(sc->aen, 0, PCI_NVME_AE_TYPE_MAX * sizeof(struct pci_nvme_aen));
808 
809 	for (atype = 0; atype < PCI_NVME_AE_TYPE_MAX; atype++) {
810 		sc->aen[atype].atype = atype;
811 	}
812 }
813 
814 static void
815 pci_nvme_aen_init(struct pci_nvme_softc *sc)
816 {
817 	char nstr[80];
818 
819 	pci_nvme_aen_reset(sc);
820 
821 	pthread_mutex_init(&sc->aen_mtx, NULL);
822 	pthread_create(&sc->aen_tid, NULL, aen_thr, sc);
823 	snprintf(nstr, sizeof(nstr), "nvme-aen-%d:%d", sc->nsc_pi->pi_slot,
824 	    sc->nsc_pi->pi_func);
825 	pthread_set_name_np(sc->aen_tid, nstr);
826 }
827 
828 static void
829 pci_nvme_aen_destroy(struct pci_nvme_softc *sc)
830 {
831 
832 	pci_nvme_aen_reset(sc);
833 }
834 
835 /* Notify the AEN thread of pending work */
836 static void
837 pci_nvme_aen_notify(struct pci_nvme_softc *sc)
838 {
839 
840 	pthread_cond_signal(&sc->aen_cond);
841 }
842 
843 /*
844  * Post an Asynchronous Event Notification
845  */
846 static int32_t
847 pci_nvme_aen_post(struct pci_nvme_softc *sc, pci_nvme_async_type atype,
848 		uint32_t event_data)
849 {
850 	struct pci_nvme_aen *aen;
851 
852 	if (atype >= PCI_NVME_AE_TYPE_MAX) {
853 		return(EINVAL);
854 	}
855 
856 	pthread_mutex_lock(&sc->aen_mtx);
857 	aen = &sc->aen[atype];
858 
859 	/* Has the controller already posted an event of this type? */
860 	if (aen->posted) {
861 		pthread_mutex_unlock(&sc->aen_mtx);
862 		return(EALREADY);
863 	}
864 
865 	aen->event_data = event_data;
866 	aen->posted = true;
867 	pthread_mutex_unlock(&sc->aen_mtx);
868 
869 	pci_nvme_aen_notify(sc);
870 
871 	return(0);
872 }
873 
874 static void
875 pci_nvme_aen_process(struct pci_nvme_softc *sc)
876 {
877 	struct pci_nvme_aer *aer;
878 	struct pci_nvme_aen *aen;
879 	pci_nvme_async_type atype;
880 	uint32_t mask;
881 	uint16_t status;
882 	uint8_t lid;
883 
884 	assert(pthread_mutex_isowned_np(&sc->aen_mtx));
885 	for (atype = 0; atype < PCI_NVME_AE_TYPE_MAX; atype++) {
886 		aen = &sc->aen[atype];
887 		/* Previous iterations may have depleted the available AER's */
888 		if (!pci_nvme_aer_available(sc)) {
889 			DPRINTF("%s: no AER", __func__);
890 			break;
891 		}
892 
893 		if (!aen->posted) {
894 			DPRINTF("%s: no AEN posted for atype=%#x", __func__, atype);
895 			continue;
896 		}
897 
898 		status = NVME_SC_SUCCESS;
899 
900 		/* Is the event masked? */
901 		mask =
902 		    sc->feat[NVME_FEAT_ASYNC_EVENT_CONFIGURATION].cdw11;
903 
904 		DPRINTF("%s: atype=%#x mask=%#x event_data=%#x", __func__, atype, mask, aen->event_data);
905 		switch (atype) {
906 		case PCI_NVME_AE_TYPE_ERROR:
907 			lid = NVME_LOG_ERROR;
908 			break;
909 		case PCI_NVME_AE_TYPE_SMART:
910 			mask &= 0xff;
911 			if ((mask & aen->event_data) == 0)
912 				continue;
913 			lid = NVME_LOG_HEALTH_INFORMATION;
914 			break;
915 		case PCI_NVME_AE_TYPE_NOTICE:
916 			if (aen->event_data >= PCI_NVME_AEI_NOTICE_MAX) {
917 				EPRINTLN("%s unknown AEN notice type %u",
918 				    __func__, aen->event_data);
919 				status = NVME_SC_INTERNAL_DEVICE_ERROR;
920 				lid = 0;
921 				break;
922 			}
923 			if ((PCI_NVME_AEI_NOTICE_MASK(aen->event_data) & mask) == 0)
924 				continue;
925 			switch (aen->event_data) {
926 			case PCI_NVME_AEI_NOTICE_NS_ATTR_CHANGED:
927 				lid = NVME_LOG_CHANGED_NAMESPACE;
928 				break;
929 			case PCI_NVME_AEI_NOTICE_FW_ACTIVATION:
930 				lid = NVME_LOG_FIRMWARE_SLOT;
931 				break;
932 			case PCI_NVME_AEI_NOTICE_TELEMETRY_CHANGE:
933 				lid = NVME_LOG_TELEMETRY_CONTROLLER_INITIATED;
934 				break;
935 			case PCI_NVME_AEI_NOTICE_ANA_CHANGE:
936 				lid = NVME_LOG_ASYMMETRIC_NAMESPACE_ACCESS;
937 				break;
938 			case PCI_NVME_AEI_NOTICE_PREDICT_LATENCY_CHANGE:
939 				lid = NVME_LOG_PREDICTABLE_LATENCY_EVENT_AGGREGATE;
940 				break;
941 			case PCI_NVME_AEI_NOTICE_LBA_STATUS_ALERT:
942 				lid = NVME_LOG_LBA_STATUS_INFORMATION;
943 				break;
944 			case PCI_NVME_AEI_NOTICE_ENDURANCE_GROUP_CHANGE:
945 				lid = NVME_LOG_ENDURANCE_GROUP_EVENT_AGGREGATE;
946 				break;
947 			default:
948 				lid = 0;
949 			}
950 			break;
951 		default:
952 			/* bad type?!? */
953 			EPRINTLN("%s unknown AEN type %u", __func__, atype);
954 			status = NVME_SC_INTERNAL_DEVICE_ERROR;
955 			lid = 0;
956 			break;
957 		}
958 
959 		aer = pci_nvme_aer_get(sc);
960 		assert(aer != NULL);
961 
962 		DPRINTF("%s: CID=%#x CDW0=%#x", __func__, aer->cid, (lid << 16) | (aen->event_data << 8) | atype);
963 		pci_nvme_cq_update(sc, &sc->compl_queues[0],
964 		    (lid << 16) | (aen->event_data << 8) | atype, /* cdw0 */
965 		    aer->cid,
966 		    0,		/* SQID */
967 		    status);
968 
969 		aen->event_data = 0;
970 		aen->posted = false;
971 
972 		pci_generate_msix(sc->nsc_pi, 0);
973 	}
974 }
975 
976 static void *
977 aen_thr(void *arg)
978 {
979 	struct pci_nvme_softc *sc;
980 
981 	sc = arg;
982 
983 	pthread_mutex_lock(&sc->aen_mtx);
984 	for (;;) {
985 		pci_nvme_aen_process(sc);
986 		pthread_cond_wait(&sc->aen_cond, &sc->aen_mtx);
987 	}
988 	pthread_mutex_unlock(&sc->aen_mtx);
989 
990 	pthread_exit(NULL);
991 	return (NULL);
992 }
993 
994 static void
995 pci_nvme_reset_locked(struct pci_nvme_softc *sc)
996 {
997 	uint32_t i;
998 
999 	DPRINTF("%s", __func__);
1000 
1001 	sc->regs.cap_lo = (ZERO_BASED(sc->max_qentries) & NVME_CAP_LO_REG_MQES_MASK) |
1002 	    (1 << NVME_CAP_LO_REG_CQR_SHIFT) |
1003 	    (60 << NVME_CAP_LO_REG_TO_SHIFT);
1004 
1005 	sc->regs.cap_hi = 1 << NVME_CAP_HI_REG_CSS_NVM_SHIFT;
1006 
1007 	sc->regs.vs = NVME_REV(1,4);	/* NVMe v1.4 */
1008 
1009 	sc->regs.cc = 0;
1010 
1011 	assert(sc->submit_queues != NULL);
1012 
1013 	for (i = 0; i < sc->num_squeues + 1; i++) {
1014 		sc->submit_queues[i].qbase = NULL;
1015 		sc->submit_queues[i].size = 0;
1016 		sc->submit_queues[i].cqid = 0;
1017 		sc->submit_queues[i].tail = 0;
1018 		sc->submit_queues[i].head = 0;
1019 	}
1020 
1021 	assert(sc->compl_queues != NULL);
1022 
1023 	for (i = 0; i < sc->num_cqueues + 1; i++) {
1024 		sc->compl_queues[i].qbase = NULL;
1025 		sc->compl_queues[i].size = 0;
1026 		sc->compl_queues[i].tail = 0;
1027 		sc->compl_queues[i].head = 0;
1028 	}
1029 
1030 	sc->num_q_is_set = false;
1031 
1032 	pci_nvme_aer_destroy(sc);
1033 	pci_nvme_aen_destroy(sc);
1034 
1035 	/*
1036 	 * Clear CSTS.RDY last to prevent the host from enabling Controller
1037 	 * before cleanup completes
1038 	 */
1039 	sc->regs.csts = 0;
1040 }
1041 
1042 static void
1043 pci_nvme_reset(struct pci_nvme_softc *sc)
1044 {
1045 	pthread_mutex_lock(&sc->mtx);
1046 	pci_nvme_reset_locked(sc);
1047 	pthread_mutex_unlock(&sc->mtx);
1048 }
1049 
1050 static int
1051 pci_nvme_init_controller(struct pci_nvme_softc *sc)
1052 {
1053 	uint16_t acqs, asqs;
1054 
1055 	DPRINTF("%s", __func__);
1056 
1057 	/*
1058 	 * NVMe 2.0 states that "enabling a controller while this field is
1059 	 * cleared to 0h produces undefined results" for both ACQS and
1060 	 * ASQS. If zero, set CFS and do not become ready.
1061 	 */
1062 	asqs = ONE_BASED(sc->regs.aqa & NVME_AQA_REG_ASQS_MASK);
1063 	if (asqs < 2) {
1064 		EPRINTLN("%s: illegal ASQS value %#x (aqa=%#x)", __func__,
1065 		    asqs - 1, sc->regs.aqa);
1066 		sc->regs.csts |= NVME_CSTS_CFS;
1067 		return (-1);
1068 	}
1069 	sc->submit_queues[0].size = asqs;
1070 	sc->submit_queues[0].qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1071 	    sc->regs.asq, sizeof(struct nvme_command) * asqs);
1072 	if (sc->submit_queues[0].qbase == NULL) {
1073 		EPRINTLN("%s: ASQ vm_map_gpa(%lx) failed", __func__,
1074 		    sc->regs.asq);
1075 		sc->regs.csts |= NVME_CSTS_CFS;
1076 		return (-1);
1077 	}
1078 
1079 	DPRINTF("%s mapping Admin-SQ guest 0x%lx, host: %p",
1080 	        __func__, sc->regs.asq, sc->submit_queues[0].qbase);
1081 
1082 	acqs = ONE_BASED((sc->regs.aqa >> NVME_AQA_REG_ACQS_SHIFT) &
1083 	    NVME_AQA_REG_ACQS_MASK);
1084 	if (acqs < 2) {
1085 		EPRINTLN("%s: illegal ACQS value %#x (aqa=%#x)", __func__,
1086 		    acqs - 1, sc->regs.aqa);
1087 		sc->regs.csts |= NVME_CSTS_CFS;
1088 		return (-1);
1089 	}
1090 	sc->compl_queues[0].size = acqs;
1091 	sc->compl_queues[0].qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1092 	    sc->regs.acq, sizeof(struct nvme_completion) * acqs);
1093 	if (sc->compl_queues[0].qbase == NULL) {
1094 		EPRINTLN("%s: ACQ vm_map_gpa(%lx) failed", __func__,
1095 		    sc->regs.acq);
1096 		sc->regs.csts |= NVME_CSTS_CFS;
1097 		return (-1);
1098 	}
1099 	sc->compl_queues[0].intr_en = NVME_CQ_INTEN;
1100 
1101 	DPRINTF("%s mapping Admin-CQ guest 0x%lx, host: %p",
1102 	        __func__, sc->regs.acq, sc->compl_queues[0].qbase);
1103 
1104 	return (0);
1105 }
1106 
1107 static int
1108 nvme_prp_memcpy(struct vmctx *ctx, uint64_t prp1, uint64_t prp2, uint8_t *b,
1109 	size_t len, enum nvme_copy_dir dir)
1110 {
1111 	uint8_t *p;
1112 	size_t bytes;
1113 
1114 	if (len > (8 * 1024)) {
1115 		return (-1);
1116 	}
1117 
1118 	/* Copy from the start of prp1 to the end of the physical page */
1119 	bytes = PAGE_SIZE - (prp1 & PAGE_MASK);
1120 	bytes = MIN(bytes, len);
1121 
1122 	p = vm_map_gpa(ctx, prp1, bytes);
1123 	if (p == NULL) {
1124 		return (-1);
1125 	}
1126 
1127 	if (dir == NVME_COPY_TO_PRP)
1128 		memcpy(p, b, bytes);
1129 	else
1130 		memcpy(b, p, bytes);
1131 
1132 	b += bytes;
1133 
1134 	len -= bytes;
1135 	if (len == 0) {
1136 		return (0);
1137 	}
1138 
1139 	len = MIN(len, PAGE_SIZE);
1140 
1141 	p = vm_map_gpa(ctx, prp2, len);
1142 	if (p == NULL) {
1143 		return (-1);
1144 	}
1145 
1146 	if (dir == NVME_COPY_TO_PRP)
1147 		memcpy(p, b, len);
1148 	else
1149 		memcpy(b, p, len);
1150 
1151 	return (0);
1152 }
1153 
1154 /*
1155  * Write a Completion Queue Entry update
1156  *
1157  * Write the completion and update the doorbell value
1158  */
1159 static void
1160 pci_nvme_cq_update(struct pci_nvme_softc *sc,
1161 		struct nvme_completion_queue *cq,
1162 		uint32_t cdw0,
1163 		uint16_t cid,
1164 		uint16_t sqid,
1165 		uint16_t status)
1166 {
1167 	struct nvme_submission_queue *sq = &sc->submit_queues[sqid];
1168 	struct nvme_completion *cqe;
1169 
1170 	assert(cq->qbase != NULL);
1171 
1172 	pthread_mutex_lock(&cq->mtx);
1173 
1174 	cqe = &cq->qbase[cq->tail];
1175 
1176 	/* Flip the phase bit */
1177 	status |= (cqe->status ^ NVME_STATUS_P) & NVME_STATUS_P_MASK;
1178 
1179 	cqe->cdw0 = cdw0;
1180 	cqe->sqhd = sq->head;
1181 	cqe->sqid = sqid;
1182 	cqe->cid = cid;
1183 	cqe->status = status;
1184 
1185 	cq->tail++;
1186 	if (cq->tail >= cq->size) {
1187 		cq->tail = 0;
1188 	}
1189 
1190 	pthread_mutex_unlock(&cq->mtx);
1191 }
1192 
1193 static int
1194 nvme_opc_delete_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
1195 	struct nvme_completion* compl)
1196 {
1197 	uint16_t qid = command->cdw10 & 0xffff;
1198 
1199 	DPRINTF("%s DELETE_IO_SQ %u", __func__, qid);
1200 	if (qid == 0 || qid > sc->num_squeues ||
1201 	    (sc->submit_queues[qid].qbase == NULL)) {
1202 		WPRINTF("%s NOT PERMITTED queue id %u / num_squeues %u",
1203 		        __func__, qid, sc->num_squeues);
1204 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1205 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1206 		return (1);
1207 	}
1208 
1209 	sc->submit_queues[qid].qbase = NULL;
1210 	sc->submit_queues[qid].cqid = 0;
1211 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1212 	return (1);
1213 }
1214 
1215 static int
1216 nvme_opc_create_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
1217 	struct nvme_completion* compl)
1218 {
1219 	if (command->cdw11 & NVME_CMD_CDW11_PC) {
1220 		uint16_t qid = command->cdw10 & 0xffff;
1221 		struct nvme_submission_queue *nsq;
1222 
1223 		if ((qid == 0) || (qid > sc->num_squeues) ||
1224 		    (sc->submit_queues[qid].qbase != NULL)) {
1225 			WPRINTF("%s queue index %u > num_squeues %u",
1226 			        __func__, qid, sc->num_squeues);
1227 			pci_nvme_status_tc(&compl->status,
1228 			    NVME_SCT_COMMAND_SPECIFIC,
1229 			    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1230 			return (1);
1231 		}
1232 
1233 		nsq = &sc->submit_queues[qid];
1234 		nsq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
1235 		DPRINTF("%s size=%u (max=%u)", __func__, nsq->size, sc->max_qentries);
1236 		if ((nsq->size < 2) || (nsq->size > sc->max_qentries)) {
1237 			/*
1238 			 * Queues must specify at least two entries
1239 			 * NOTE: "MAXIMUM QUEUE SIZE EXCEEDED" was renamed to
1240 			 * "INVALID QUEUE SIZE" in the NVM Express 1.3 Spec
1241 			 */
1242 			pci_nvme_status_tc(&compl->status,
1243 			    NVME_SCT_COMMAND_SPECIFIC,
1244 			    NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED);
1245 			return (1);
1246 		}
1247 		nsq->head = nsq->tail = 0;
1248 
1249 		nsq->cqid = (command->cdw11 >> 16) & 0xffff;
1250 		if ((nsq->cqid == 0) || (nsq->cqid > sc->num_cqueues)) {
1251 			pci_nvme_status_tc(&compl->status,
1252 			    NVME_SCT_COMMAND_SPECIFIC,
1253 			    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1254 			return (1);
1255 		}
1256 
1257 		if (sc->compl_queues[nsq->cqid].qbase == NULL) {
1258 			pci_nvme_status_tc(&compl->status,
1259 			    NVME_SCT_COMMAND_SPECIFIC,
1260 			    NVME_SC_COMPLETION_QUEUE_INVALID);
1261 			return (1);
1262 		}
1263 
1264 		nsq->qpriority = (command->cdw11 >> 1) & 0x03;
1265 
1266 		nsq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1267 		              sizeof(struct nvme_command) * (size_t)nsq->size);
1268 
1269 		DPRINTF("%s sq %u size %u gaddr %p cqid %u", __func__,
1270 		        qid, nsq->size, nsq->qbase, nsq->cqid);
1271 
1272 		pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1273 
1274 		DPRINTF("%s completed creating IOSQ qid %u",
1275 		         __func__, qid);
1276 	} else {
1277 		/*
1278 		 * Guest sent non-cont submission queue request.
1279 		 * This setting is unsupported by this emulation.
1280 		 */
1281 		WPRINTF("%s unsupported non-contig (list-based) "
1282 		         "create i/o submission queue", __func__);
1283 
1284 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1285 	}
1286 	return (1);
1287 }
1288 
1289 static int
1290 nvme_opc_delete_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
1291 	struct nvme_completion* compl)
1292 {
1293 	uint16_t qid = command->cdw10 & 0xffff;
1294 	uint16_t sqid;
1295 
1296 	DPRINTF("%s DELETE_IO_CQ %u", __func__, qid);
1297 	if (qid == 0 || qid > sc->num_cqueues ||
1298 	    (sc->compl_queues[qid].qbase == NULL)) {
1299 		WPRINTF("%s queue index %u / num_cqueues %u",
1300 		        __func__, qid, sc->num_cqueues);
1301 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1302 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1303 		return (1);
1304 	}
1305 
1306 	/* Deleting an Active CQ is an error */
1307 	for (sqid = 1; sqid < sc->num_squeues + 1; sqid++)
1308 		if (sc->submit_queues[sqid].cqid == qid) {
1309 			pci_nvme_status_tc(&compl->status,
1310 			    NVME_SCT_COMMAND_SPECIFIC,
1311 			    NVME_SC_INVALID_QUEUE_DELETION);
1312 			return (1);
1313 		}
1314 
1315 	sc->compl_queues[qid].qbase = NULL;
1316 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1317 	return (1);
1318 }
1319 
1320 static int
1321 nvme_opc_create_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
1322 	struct nvme_completion* compl)
1323 {
1324 	struct nvme_completion_queue *ncq;
1325 	uint16_t qid = command->cdw10 & 0xffff;
1326 
1327 	/* Only support Physically Contiguous queues */
1328 	if ((command->cdw11 & NVME_CMD_CDW11_PC) == 0) {
1329 		WPRINTF("%s unsupported non-contig (list-based) "
1330 		         "create i/o completion queue",
1331 		         __func__);
1332 
1333 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1334 		return (1);
1335 	}
1336 
1337 	if ((qid == 0) || (qid > sc->num_cqueues) ||
1338 	    (sc->compl_queues[qid].qbase != NULL)) {
1339 		WPRINTF("%s queue index %u > num_cqueues %u",
1340 			__func__, qid, sc->num_cqueues);
1341 		pci_nvme_status_tc(&compl->status,
1342 		    NVME_SCT_COMMAND_SPECIFIC,
1343 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1344 		return (1);
1345  	}
1346 
1347 	ncq = &sc->compl_queues[qid];
1348 	ncq->intr_en = (command->cdw11 & NVME_CMD_CDW11_IEN) >> 1;
1349 	ncq->intr_vec = (command->cdw11 >> 16) & 0xffff;
1350 	if (ncq->intr_vec > (sc->max_queues + 1)) {
1351 		pci_nvme_status_tc(&compl->status,
1352 		    NVME_SCT_COMMAND_SPECIFIC,
1353 		    NVME_SC_INVALID_INTERRUPT_VECTOR);
1354 		return (1);
1355 	}
1356 
1357 	ncq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
1358 	if ((ncq->size < 2) || (ncq->size > sc->max_qentries))  {
1359 		/*
1360 		 * Queues must specify at least two entries
1361 		 * NOTE: "MAXIMUM QUEUE SIZE EXCEEDED" was renamed to
1362 		 * "INVALID QUEUE SIZE" in the NVM Express 1.3 Spec
1363 		 */
1364 		pci_nvme_status_tc(&compl->status,
1365 		    NVME_SCT_COMMAND_SPECIFIC,
1366 		    NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED);
1367 		return (1);
1368 	}
1369 	ncq->head = ncq->tail = 0;
1370 	ncq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1371 		     command->prp1,
1372 		     sizeof(struct nvme_command) * (size_t)ncq->size);
1373 
1374 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1375 
1376 
1377 	return (1);
1378 }
1379 
1380 static int
1381 nvme_opc_get_log_page(struct pci_nvme_softc* sc, struct nvme_command* command,
1382 	struct nvme_completion* compl)
1383 {
1384 	uint64_t logoff;
1385 	uint32_t logsize;
1386 	uint8_t logpage;
1387 
1388 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1389 
1390 	/*
1391 	 * Command specifies the number of dwords to return in fields NUMDU
1392 	 * and NUMDL. This is a zero-based value.
1393 	 */
1394 	logpage = command->cdw10 & 0xFF;
1395 	logsize = ((command->cdw11 << 16) | (command->cdw10 >> 16)) + 1;
1396 	logsize *= sizeof(uint32_t);
1397 	logoff  = ((uint64_t)(command->cdw13) << 32) | command->cdw12;
1398 
1399 	DPRINTF("%s log page %u len %u", __func__, logpage, logsize);
1400 
1401 	switch (logpage) {
1402 	case NVME_LOG_ERROR:
1403 		if (logoff >= sizeof(sc->err_log)) {
1404 			pci_nvme_status_genc(&compl->status,
1405 			    NVME_SC_INVALID_FIELD);
1406 			break;
1407 		}
1408 
1409 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1410 		    command->prp2, (uint8_t *)&sc->err_log + logoff,
1411 		    MIN(logsize - logoff, sizeof(sc->err_log)),
1412 		    NVME_COPY_TO_PRP);
1413 		break;
1414 	case NVME_LOG_HEALTH_INFORMATION:
1415 		if (logoff >= sizeof(sc->health_log)) {
1416 			pci_nvme_status_genc(&compl->status,
1417 			    NVME_SC_INVALID_FIELD);
1418 			break;
1419 		}
1420 
1421 		pthread_mutex_lock(&sc->mtx);
1422 		memcpy(&sc->health_log.data_units_read, &sc->read_data_units,
1423 		    sizeof(sc->health_log.data_units_read));
1424 		memcpy(&sc->health_log.data_units_written, &sc->write_data_units,
1425 		    sizeof(sc->health_log.data_units_written));
1426 		memcpy(&sc->health_log.host_read_commands, &sc->read_commands,
1427 		    sizeof(sc->health_log.host_read_commands));
1428 		memcpy(&sc->health_log.host_write_commands, &sc->write_commands,
1429 		    sizeof(sc->health_log.host_write_commands));
1430 		pthread_mutex_unlock(&sc->mtx);
1431 
1432 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1433 		    command->prp2, (uint8_t *)&sc->health_log + logoff,
1434 		    MIN(logsize - logoff, sizeof(sc->health_log)),
1435 		    NVME_COPY_TO_PRP);
1436 		break;
1437 	case NVME_LOG_FIRMWARE_SLOT:
1438 		if (logoff >= sizeof(sc->fw_log)) {
1439 			pci_nvme_status_genc(&compl->status,
1440 			    NVME_SC_INVALID_FIELD);
1441 			break;
1442 		}
1443 
1444 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1445 		    command->prp2, (uint8_t *)&sc->fw_log + logoff,
1446 		    MIN(logsize - logoff, sizeof(sc->fw_log)),
1447 		    NVME_COPY_TO_PRP);
1448 		break;
1449 	case NVME_LOG_CHANGED_NAMESPACE:
1450 		if (logoff >= sizeof(sc->ns_log)) {
1451 			pci_nvme_status_genc(&compl->status,
1452 			    NVME_SC_INVALID_FIELD);
1453 			break;
1454 		}
1455 
1456 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1457 		    command->prp2, (uint8_t *)&sc->ns_log + logoff,
1458 		    MIN(logsize - logoff, sizeof(sc->ns_log)),
1459 		    NVME_COPY_TO_PRP);
1460 		memset(&sc->ns_log, 0, sizeof(sc->ns_log));
1461 		break;
1462 	default:
1463 		DPRINTF("%s get log page %x command not supported",
1464 		        __func__, logpage);
1465 
1466 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1467 		    NVME_SC_INVALID_LOG_PAGE);
1468 	}
1469 
1470 	return (1);
1471 }
1472 
1473 static int
1474 nvme_opc_identify(struct pci_nvme_softc* sc, struct nvme_command* command,
1475 	struct nvme_completion* compl)
1476 {
1477 	void *dest;
1478 	uint16_t status;
1479 
1480 	DPRINTF("%s identify 0x%x nsid 0x%x", __func__,
1481 	        command->cdw10 & 0xFF, command->nsid);
1482 
1483 	status = 0;
1484 	pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
1485 
1486 	switch (command->cdw10 & 0xFF) {
1487 	case 0x00: /* return Identify Namespace data structure */
1488 		/* Global NS only valid with NS Management */
1489 		if (command->nsid == NVME_GLOBAL_NAMESPACE_TAG) {
1490 			pci_nvme_status_genc(&status,
1491 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
1492 			break;
1493 		}
1494 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1495 		    command->prp2, (uint8_t *)&sc->nsdata, sizeof(sc->nsdata),
1496 		    NVME_COPY_TO_PRP);
1497 		break;
1498 	case 0x01: /* return Identify Controller data structure */
1499 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1500 		    command->prp2, (uint8_t *)&sc->ctrldata,
1501 		    sizeof(sc->ctrldata),
1502 		    NVME_COPY_TO_PRP);
1503 		break;
1504 	case 0x02: /* list of 1024 active NSIDs > CDW1.NSID */
1505 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1506 		                  sizeof(uint32_t) * 1024);
1507 		/* All unused entries shall be zero */
1508 		memset(dest, 0, sizeof(uint32_t) * 1024);
1509 		((uint32_t *)dest)[0] = 1;
1510 		break;
1511 	case 0x03: /* list of NSID structures in CDW1.NSID, 4096 bytes */
1512 		if (command->nsid != 1) {
1513 			pci_nvme_status_genc(&status,
1514 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
1515 			break;
1516 		}
1517 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1518 		                  sizeof(uint32_t) * 1024);
1519 		/* All bytes after the descriptor shall be zero */
1520 		memset(dest, 0, sizeof(uint32_t) * 1024);
1521 
1522 		/* Return NIDT=1 (i.e. EUI64) descriptor */
1523 		((uint8_t *)dest)[0] = 1;
1524 		((uint8_t *)dest)[1] = sizeof(uint64_t);
1525 		memcpy(((uint8_t *)dest) + 4, sc->nsdata.eui64, sizeof(uint64_t));
1526 		break;
1527 	case 0x13:
1528 		/*
1529 		 * Controller list is optional but used by UNH tests. Return
1530 		 * a valid but empty list.
1531 		 */
1532 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1533 		                  sizeof(uint16_t) * 2048);
1534 		memset(dest, 0, sizeof(uint16_t) * 2048);
1535 		break;
1536 	default:
1537 		DPRINTF("%s unsupported identify command requested 0x%x",
1538 		         __func__, command->cdw10 & 0xFF);
1539 		pci_nvme_status_genc(&status, NVME_SC_INVALID_FIELD);
1540 		break;
1541 	}
1542 
1543 	compl->status = status;
1544 	return (1);
1545 }
1546 
1547 static const char *
1548 nvme_fid_to_name(uint8_t fid)
1549 {
1550 	const char *name;
1551 
1552 	switch (fid) {
1553 	case NVME_FEAT_ARBITRATION:
1554 		name = "Arbitration";
1555 		break;
1556 	case NVME_FEAT_POWER_MANAGEMENT:
1557 		name = "Power Management";
1558 		break;
1559 	case NVME_FEAT_LBA_RANGE_TYPE:
1560 		name = "LBA Range Type";
1561 		break;
1562 	case NVME_FEAT_TEMPERATURE_THRESHOLD:
1563 		name = "Temperature Threshold";
1564 		break;
1565 	case NVME_FEAT_ERROR_RECOVERY:
1566 		name = "Error Recovery";
1567 		break;
1568 	case NVME_FEAT_VOLATILE_WRITE_CACHE:
1569 		name = "Volatile Write Cache";
1570 		break;
1571 	case NVME_FEAT_NUMBER_OF_QUEUES:
1572 		name = "Number of Queues";
1573 		break;
1574 	case NVME_FEAT_INTERRUPT_COALESCING:
1575 		name = "Interrupt Coalescing";
1576 		break;
1577 	case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
1578 		name = "Interrupt Vector Configuration";
1579 		break;
1580 	case NVME_FEAT_WRITE_ATOMICITY:
1581 		name = "Write Atomicity Normal";
1582 		break;
1583 	case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
1584 		name = "Asynchronous Event Configuration";
1585 		break;
1586 	case NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION:
1587 		name = "Autonomous Power State Transition";
1588 		break;
1589 	case NVME_FEAT_HOST_MEMORY_BUFFER:
1590 		name = "Host Memory Buffer";
1591 		break;
1592 	case NVME_FEAT_TIMESTAMP:
1593 		name = "Timestamp";
1594 		break;
1595 	case NVME_FEAT_KEEP_ALIVE_TIMER:
1596 		name = "Keep Alive Timer";
1597 		break;
1598 	case NVME_FEAT_HOST_CONTROLLED_THERMAL_MGMT:
1599 		name = "Host Controlled Thermal Management";
1600 		break;
1601 	case NVME_FEAT_NON_OP_POWER_STATE_CONFIG:
1602 		name = "Non-Operation Power State Config";
1603 		break;
1604 	case NVME_FEAT_READ_RECOVERY_LEVEL_CONFIG:
1605 		name = "Read Recovery Level Config";
1606 		break;
1607 	case NVME_FEAT_PREDICTABLE_LATENCY_MODE_CONFIG:
1608 		name = "Predictable Latency Mode Config";
1609 		break;
1610 	case NVME_FEAT_PREDICTABLE_LATENCY_MODE_WINDOW:
1611 		name = "Predictable Latency Mode Window";
1612 		break;
1613 	case NVME_FEAT_LBA_STATUS_INFORMATION_ATTRIBUTES:
1614 		name = "LBA Status Information Report Interval";
1615 		break;
1616 	case NVME_FEAT_HOST_BEHAVIOR_SUPPORT:
1617 		name = "Host Behavior Support";
1618 		break;
1619 	case NVME_FEAT_SANITIZE_CONFIG:
1620 		name = "Sanitize Config";
1621 		break;
1622 	case NVME_FEAT_ENDURANCE_GROUP_EVENT_CONFIGURATION:
1623 		name = "Endurance Group Event Configuration";
1624 		break;
1625 	case NVME_FEAT_SOFTWARE_PROGRESS_MARKER:
1626 		name = "Software Progress Marker";
1627 		break;
1628 	case NVME_FEAT_HOST_IDENTIFIER:
1629 		name = "Host Identifier";
1630 		break;
1631 	case NVME_FEAT_RESERVATION_NOTIFICATION_MASK:
1632 		name = "Reservation Notification Mask";
1633 		break;
1634 	case NVME_FEAT_RESERVATION_PERSISTENCE:
1635 		name = "Reservation Persistence";
1636 		break;
1637 	case NVME_FEAT_NAMESPACE_WRITE_PROTECTION_CONFIG:
1638 		name = "Namespace Write Protection Config";
1639 		break;
1640 	default:
1641 		name = "Unknown";
1642 		break;
1643 	}
1644 
1645 	return (name);
1646 }
1647 
1648 static void
1649 nvme_feature_invalid_cb(struct pci_nvme_softc *sc __unused,
1650     struct nvme_feature_obj *feat __unused,
1651     struct nvme_command *command __unused,
1652     struct nvme_completion *compl)
1653 {
1654 	pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1655 }
1656 
1657 static void
1658 nvme_feature_iv_config(struct pci_nvme_softc *sc,
1659     struct nvme_feature_obj *feat __unused,
1660     struct nvme_command *command,
1661     struct nvme_completion *compl)
1662 {
1663 	uint32_t i;
1664 	uint32_t cdw11 = command->cdw11;
1665 	uint16_t iv;
1666 	bool cd;
1667 
1668 	pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1669 
1670 	iv = cdw11 & 0xffff;
1671 	cd = cdw11 & (1 << 16);
1672 
1673 	if (iv > (sc->max_queues + 1)) {
1674 		return;
1675 	}
1676 
1677 	/* No Interrupt Coalescing (i.e. not Coalescing Disable) for Admin Q */
1678 	if ((iv == 0) && !cd)
1679 		return;
1680 
1681 	/* Requested Interrupt Vector must be used by a CQ */
1682 	for (i = 0; i < sc->num_cqueues + 1; i++) {
1683 		if (sc->compl_queues[i].intr_vec == iv) {
1684 			pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1685 		}
1686 	}
1687 }
1688 
1689 #define NVME_ASYNC_EVENT_ENDURANCE_GROUP		(0x4000)
1690 static void
1691 nvme_feature_async_event(struct pci_nvme_softc *sc __unused,
1692     struct nvme_feature_obj *feat __unused,
1693     struct nvme_command *command,
1694     struct nvme_completion *compl)
1695 {
1696 	if (command->cdw11 & NVME_ASYNC_EVENT_ENDURANCE_GROUP)
1697 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1698 }
1699 
1700 #define NVME_TEMP_THRESH_OVER	0
1701 #define NVME_TEMP_THRESH_UNDER	1
1702 static void
1703 nvme_feature_temperature(struct pci_nvme_softc *sc,
1704     struct nvme_feature_obj *feat __unused,
1705     struct nvme_command *command,
1706     struct nvme_completion *compl)
1707 {
1708 	uint16_t	tmpth;	/* Temperature Threshold */
1709 	uint8_t		tmpsel; /* Threshold Temperature Select */
1710 	uint8_t		thsel;  /* Threshold Type Select */
1711 	bool		set_crit = false;
1712 	bool		report_crit;
1713 
1714 	tmpth  = command->cdw11 & 0xffff;
1715 	tmpsel = (command->cdw11 >> 16) & 0xf;
1716 	thsel  = (command->cdw11 >> 20) & 0x3;
1717 
1718 	DPRINTF("%s: tmpth=%#x tmpsel=%#x thsel=%#x", __func__, tmpth, tmpsel, thsel);
1719 
1720 	/* Check for unsupported values */
1721 	if (((tmpsel != 0) && (tmpsel != 0xf)) ||
1722 	    (thsel > NVME_TEMP_THRESH_UNDER)) {
1723 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1724 		return;
1725 	}
1726 
1727 	if (((thsel == NVME_TEMP_THRESH_OVER)  && (NVME_TEMPERATURE >= tmpth)) ||
1728 	    ((thsel == NVME_TEMP_THRESH_UNDER) && (NVME_TEMPERATURE <= tmpth)))
1729 		set_crit = true;
1730 
1731 	pthread_mutex_lock(&sc->mtx);
1732 	if (set_crit)
1733 		sc->health_log.critical_warning |=
1734 		    NVME_CRIT_WARN_ST_TEMPERATURE;
1735 	else
1736 		sc->health_log.critical_warning &=
1737 		    ~NVME_CRIT_WARN_ST_TEMPERATURE;
1738 	pthread_mutex_unlock(&sc->mtx);
1739 
1740 	report_crit = sc->feat[NVME_FEAT_ASYNC_EVENT_CONFIGURATION].cdw11 &
1741 	    NVME_CRIT_WARN_ST_TEMPERATURE;
1742 
1743 	if (set_crit && report_crit)
1744 		pci_nvme_aen_post(sc, PCI_NVME_AE_TYPE_SMART,
1745 		    sc->health_log.critical_warning);
1746 
1747 	DPRINTF("%s: set_crit=%c critical_warning=%#x status=%#x", __func__, set_crit ? 'T':'F', sc->health_log.critical_warning, compl->status);
1748 }
1749 
1750 static void
1751 nvme_feature_num_queues(struct pci_nvme_softc *sc,
1752     struct nvme_feature_obj *feat __unused,
1753     struct nvme_command *command,
1754     struct nvme_completion *compl)
1755 {
1756 	uint16_t nqr;	/* Number of Queues Requested */
1757 
1758 	if (sc->num_q_is_set) {
1759 		WPRINTF("%s: Number of Queues already set", __func__);
1760 		pci_nvme_status_genc(&compl->status,
1761 		    NVME_SC_COMMAND_SEQUENCE_ERROR);
1762 		return;
1763 	}
1764 
1765 	nqr = command->cdw11 & 0xFFFF;
1766 	if (nqr == 0xffff) {
1767 		WPRINTF("%s: Illegal NSQR value %#x", __func__, nqr);
1768 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1769 		return;
1770 	}
1771 
1772 	sc->num_squeues = ONE_BASED(nqr);
1773 	if (sc->num_squeues > sc->max_queues) {
1774 		DPRINTF("NSQR=%u is greater than max %u", sc->num_squeues,
1775 					sc->max_queues);
1776 		sc->num_squeues = sc->max_queues;
1777 	}
1778 
1779 	nqr = (command->cdw11 >> 16) & 0xFFFF;
1780 	if (nqr == 0xffff) {
1781 		WPRINTF("%s: Illegal NCQR value %#x", __func__, nqr);
1782 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1783 		return;
1784 	}
1785 
1786 	sc->num_cqueues = ONE_BASED(nqr);
1787 	if (sc->num_cqueues > sc->max_queues) {
1788 		DPRINTF("NCQR=%u is greater than max %u", sc->num_cqueues,
1789 					sc->max_queues);
1790 		sc->num_cqueues = sc->max_queues;
1791 	}
1792 
1793 	/* Patch the command value which will be saved on callback's return */
1794 	command->cdw11 = NVME_FEATURE_NUM_QUEUES(sc);
1795 	compl->cdw0 = NVME_FEATURE_NUM_QUEUES(sc);
1796 
1797 	sc->num_q_is_set = true;
1798 }
1799 
1800 static int
1801 nvme_opc_set_features(struct pci_nvme_softc *sc, struct nvme_command *command,
1802 	struct nvme_completion *compl)
1803 {
1804 	struct nvme_feature_obj *feat;
1805 	uint32_t nsid = command->nsid;
1806 	uint8_t fid = NVMEV(NVME_FEAT_SET_FID, command->cdw10);
1807 	bool sv = NVMEV(NVME_FEAT_SET_SV, command->cdw10);
1808 
1809 	DPRINTF("%s: Feature ID 0x%x (%s)", __func__, fid, nvme_fid_to_name(fid));
1810 
1811 	if (fid >= NVME_FID_MAX) {
1812 		DPRINTF("%s invalid feature 0x%x", __func__, fid);
1813 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1814 		return (1);
1815 	}
1816 
1817 	if (sv) {
1818 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1819 		    NVME_SC_FEATURE_NOT_SAVEABLE);
1820 		return (1);
1821 	}
1822 
1823 	feat = &sc->feat[fid];
1824 
1825 	if (feat->namespace_specific && (nsid == NVME_GLOBAL_NAMESPACE_TAG)) {
1826 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1827 		return (1);
1828 	}
1829 
1830 	if (!feat->namespace_specific &&
1831 	    !((nsid == 0) || (nsid == NVME_GLOBAL_NAMESPACE_TAG))) {
1832 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1833 		    NVME_SC_FEATURE_NOT_NS_SPECIFIC);
1834 		return (1);
1835 	}
1836 
1837 	compl->cdw0 = 0;
1838 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1839 
1840 	if (feat->set)
1841 		feat->set(sc, feat, command, compl);
1842 	else {
1843 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1844 		    NVME_SC_FEATURE_NOT_CHANGEABLE);
1845 		return (1);
1846 	}
1847 
1848 	DPRINTF("%s: status=%#x cdw11=%#x", __func__, compl->status, command->cdw11);
1849 	if (compl->status == NVME_SC_SUCCESS) {
1850 		feat->cdw11 = command->cdw11;
1851 		if ((fid == NVME_FEAT_ASYNC_EVENT_CONFIGURATION) &&
1852 		    (command->cdw11 != 0))
1853 			pci_nvme_aen_notify(sc);
1854 	}
1855 
1856 	return (0);
1857 }
1858 
1859 #define NVME_FEATURES_SEL_SUPPORTED	0x3
1860 #define NVME_FEATURES_NS_SPECIFIC	(1 << 1)
1861 
1862 static int
1863 nvme_opc_get_features(struct pci_nvme_softc* sc, struct nvme_command* command,
1864 	struct nvme_completion* compl)
1865 {
1866 	struct nvme_feature_obj *feat;
1867 	uint8_t fid = command->cdw10 & 0xFF;
1868 	uint8_t sel = (command->cdw10 >> 8) & 0x7;
1869 
1870 	DPRINTF("%s: Feature ID 0x%x (%s)", __func__, fid, nvme_fid_to_name(fid));
1871 
1872 	if (fid >= NVME_FID_MAX) {
1873 		DPRINTF("%s invalid feature 0x%x", __func__, fid);
1874 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1875 		return (1);
1876 	}
1877 
1878 	compl->cdw0 = 0;
1879 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1880 
1881 	feat = &sc->feat[fid];
1882 	if (feat->get) {
1883 		feat->get(sc, feat, command, compl);
1884 	}
1885 
1886 	if (compl->status == NVME_SC_SUCCESS) {
1887 		if ((sel == NVME_FEATURES_SEL_SUPPORTED) && feat->namespace_specific)
1888 			compl->cdw0 = NVME_FEATURES_NS_SPECIFIC;
1889 		else
1890 			compl->cdw0 = feat->cdw11;
1891 	}
1892 
1893 	return (0);
1894 }
1895 
1896 static int
1897 nvme_opc_format_nvm(struct pci_nvme_softc* sc, struct nvme_command* command,
1898 	struct nvme_completion* compl)
1899 {
1900 	uint8_t	ses, lbaf, pi;
1901 
1902 	/* Only supports Secure Erase Setting - User Data Erase */
1903 	ses = (command->cdw10 >> 9) & 0x7;
1904 	if (ses > 0x1) {
1905 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1906 		return (1);
1907 	}
1908 
1909 	/* Only supports a single LBA Format */
1910 	lbaf = command->cdw10 & 0xf;
1911 	if (lbaf != 0) {
1912 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1913 		    NVME_SC_INVALID_FORMAT);
1914 		return (1);
1915 	}
1916 
1917 	/* Doesn't support Protection Infomation */
1918 	pi = (command->cdw10 >> 5) & 0x7;
1919 	if (pi != 0) {
1920 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1921 		return (1);
1922 	}
1923 
1924 	if (sc->nvstore.type == NVME_STOR_RAM) {
1925 		if (sc->nvstore.ctx)
1926 			free(sc->nvstore.ctx);
1927 		sc->nvstore.ctx = calloc(1, sc->nvstore.size);
1928 		pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1929 	} else {
1930 		struct pci_nvme_ioreq *req;
1931 		int err;
1932 
1933 		req = pci_nvme_get_ioreq(sc);
1934 		if (req == NULL) {
1935 			pci_nvme_status_genc(&compl->status,
1936 			    NVME_SC_INTERNAL_DEVICE_ERROR);
1937 			WPRINTF("%s: unable to allocate IO req", __func__);
1938 			return (1);
1939 		}
1940 		req->nvme_sq = &sc->submit_queues[0];
1941 		req->sqid = 0;
1942 		req->opc = command->opc;
1943 		req->cid = command->cid;
1944 		req->nsid = command->nsid;
1945 
1946 		req->io_req.br_offset = 0;
1947 		req->io_req.br_resid = sc->nvstore.size;
1948 		req->io_req.br_callback = pci_nvme_io_done;
1949 
1950 		err = blockif_delete(sc->nvstore.ctx, &req->io_req);
1951 		if (err) {
1952 			pci_nvme_status_genc(&compl->status,
1953 			    NVME_SC_INTERNAL_DEVICE_ERROR);
1954 			pci_nvme_release_ioreq(sc, req);
1955 		} else
1956 			compl->status = NVME_NO_STATUS;
1957 	}
1958 
1959 	return (1);
1960 }
1961 
1962 static int
1963 nvme_opc_abort(struct pci_nvme_softc *sc __unused, struct nvme_command *command,
1964     struct nvme_completion *compl)
1965 {
1966 	DPRINTF("%s submission queue %u, command ID 0x%x", __func__,
1967 	        command->cdw10 & 0xFFFF, (command->cdw10 >> 16) & 0xFFFF);
1968 
1969 	/* TODO: search for the command ID and abort it */
1970 
1971 	compl->cdw0 = 1;
1972 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1973 	return (1);
1974 }
1975 
1976 static int
1977 nvme_opc_async_event_req(struct pci_nvme_softc* sc,
1978 	struct nvme_command* command, struct nvme_completion* compl)
1979 {
1980 	DPRINTF("%s async event request count=%u aerl=%u cid=%#x", __func__,
1981 	    sc->aer_count, sc->ctrldata.aerl, command->cid);
1982 
1983 	/* Don't exceed the Async Event Request Limit (AERL). */
1984 	if (pci_nvme_aer_limit_reached(sc)) {
1985 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1986 				NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED);
1987 		return (1);
1988 	}
1989 
1990 	if (pci_nvme_aer_add(sc, command->cid)) {
1991 		pci_nvme_status_tc(&compl->status, NVME_SCT_GENERIC,
1992 				NVME_SC_INTERNAL_DEVICE_ERROR);
1993 		return (1);
1994 	}
1995 
1996 	/*
1997 	 * Raise events when they happen based on the Set Features cmd.
1998 	 * These events happen async, so only set completion successful if
1999 	 * there is an event reflective of the request to get event.
2000 	 */
2001 	compl->status = NVME_NO_STATUS;
2002 	pci_nvme_aen_notify(sc);
2003 
2004 	return (0);
2005 }
2006 
2007 static void
2008 pci_nvme_handle_admin_cmd(struct pci_nvme_softc* sc, uint64_t value)
2009 {
2010 	struct nvme_completion compl;
2011 	struct nvme_command *cmd;
2012 	struct nvme_submission_queue *sq;
2013 	struct nvme_completion_queue *cq;
2014 	uint16_t sqhead;
2015 
2016 	DPRINTF("%s index %u", __func__, (uint32_t)value);
2017 
2018 	sq = &sc->submit_queues[0];
2019 	cq = &sc->compl_queues[0];
2020 
2021 	pthread_mutex_lock(&sq->mtx);
2022 
2023 	sqhead = sq->head;
2024 	DPRINTF("sqhead %u, tail %u", sqhead, sq->tail);
2025 
2026 	while (sqhead != atomic_load_acq_short(&sq->tail)) {
2027 		cmd = &(sq->qbase)[sqhead];
2028 		compl.cdw0 = 0;
2029 		compl.status = 0;
2030 
2031 		switch (cmd->opc) {
2032 		case NVME_OPC_DELETE_IO_SQ:
2033 			DPRINTF("%s command DELETE_IO_SQ", __func__);
2034 			nvme_opc_delete_io_sq(sc, cmd, &compl);
2035 			break;
2036 		case NVME_OPC_CREATE_IO_SQ:
2037 			DPRINTF("%s command CREATE_IO_SQ", __func__);
2038 			nvme_opc_create_io_sq(sc, cmd, &compl);
2039 			break;
2040 		case NVME_OPC_DELETE_IO_CQ:
2041 			DPRINTF("%s command DELETE_IO_CQ", __func__);
2042 			nvme_opc_delete_io_cq(sc, cmd, &compl);
2043 			break;
2044 		case NVME_OPC_CREATE_IO_CQ:
2045 			DPRINTF("%s command CREATE_IO_CQ", __func__);
2046 			nvme_opc_create_io_cq(sc, cmd, &compl);
2047 			break;
2048 		case NVME_OPC_GET_LOG_PAGE:
2049 			DPRINTF("%s command GET_LOG_PAGE", __func__);
2050 			nvme_opc_get_log_page(sc, cmd, &compl);
2051 			break;
2052 		case NVME_OPC_IDENTIFY:
2053 			DPRINTF("%s command IDENTIFY", __func__);
2054 			nvme_opc_identify(sc, cmd, &compl);
2055 			break;
2056 		case NVME_OPC_ABORT:
2057 			DPRINTF("%s command ABORT", __func__);
2058 			nvme_opc_abort(sc, cmd, &compl);
2059 			break;
2060 		case NVME_OPC_SET_FEATURES:
2061 			DPRINTF("%s command SET_FEATURES", __func__);
2062 			nvme_opc_set_features(sc, cmd, &compl);
2063 			break;
2064 		case NVME_OPC_GET_FEATURES:
2065 			DPRINTF("%s command GET_FEATURES", __func__);
2066 			nvme_opc_get_features(sc, cmd, &compl);
2067 			break;
2068 		case NVME_OPC_FIRMWARE_ACTIVATE:
2069 			DPRINTF("%s command FIRMWARE_ACTIVATE", __func__);
2070 			pci_nvme_status_tc(&compl.status,
2071 			    NVME_SCT_COMMAND_SPECIFIC,
2072 			    NVME_SC_INVALID_FIRMWARE_SLOT);
2073 			break;
2074 		case NVME_OPC_ASYNC_EVENT_REQUEST:
2075 			DPRINTF("%s command ASYNC_EVENT_REQ", __func__);
2076 			nvme_opc_async_event_req(sc, cmd, &compl);
2077 			break;
2078 		case NVME_OPC_FORMAT_NVM:
2079 			DPRINTF("%s command FORMAT_NVM", __func__);
2080 			if ((sc->ctrldata.oacs &
2081 			    (1 << NVME_CTRLR_DATA_OACS_FORMAT_SHIFT)) == 0) {
2082 				pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_OPCODE);
2083 				break;
2084 			}
2085 			nvme_opc_format_nvm(sc, cmd, &compl);
2086 			break;
2087 		case NVME_OPC_SECURITY_SEND:
2088 		case NVME_OPC_SECURITY_RECEIVE:
2089 		case NVME_OPC_SANITIZE:
2090 		case NVME_OPC_GET_LBA_STATUS:
2091 			DPRINTF("%s command OPC=%#x (unsupported)", __func__,
2092 			    cmd->opc);
2093 			/* Valid but unsupported opcodes */
2094 			pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_FIELD);
2095 			break;
2096 		default:
2097 			DPRINTF("%s command OPC=%#X (not implemented)",
2098 			    __func__,
2099 			    cmd->opc);
2100 			pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_OPCODE);
2101 		}
2102 		sqhead = (sqhead + 1) % sq->size;
2103 
2104 		if (NVME_COMPLETION_VALID(compl)) {
2105 			pci_nvme_cq_update(sc, &sc->compl_queues[0],
2106 			    compl.cdw0,
2107 			    cmd->cid,
2108 			    0,		/* SQID */
2109 			    compl.status);
2110 		}
2111 	}
2112 
2113 	DPRINTF("setting sqhead %u", sqhead);
2114 	sq->head = sqhead;
2115 
2116 	if (cq->head != cq->tail)
2117 		pci_generate_msix(sc->nsc_pi, 0);
2118 
2119 	pthread_mutex_unlock(&sq->mtx);
2120 }
2121 
2122 /*
2123  * Update the Write and Read statistics reported in SMART data
2124  *
2125  * NVMe defines "data unit" as thousand's of 512 byte blocks and is rounded up.
2126  * E.g. 1 data unit is 1 - 1,000 512 byte blocks. 3 data units are 2,001 - 3,000
2127  * 512 byte blocks. Rounding up is acheived by initializing the remainder to 999.
2128  */
2129 static void
2130 pci_nvme_stats_write_read_update(struct pci_nvme_softc *sc, uint8_t opc,
2131     size_t bytes, uint16_t status)
2132 {
2133 
2134 	pthread_mutex_lock(&sc->mtx);
2135 	switch (opc) {
2136 	case NVME_OPC_WRITE:
2137 		sc->write_commands++;
2138 		if (status != NVME_SC_SUCCESS)
2139 			break;
2140 		sc->write_dunits_remainder += (bytes / 512);
2141 		while (sc->write_dunits_remainder >= 1000) {
2142 			sc->write_data_units++;
2143 			sc->write_dunits_remainder -= 1000;
2144 		}
2145 		break;
2146 	case NVME_OPC_READ:
2147 		sc->read_commands++;
2148 		if (status != NVME_SC_SUCCESS)
2149 			break;
2150 		sc->read_dunits_remainder += (bytes / 512);
2151 		while (sc->read_dunits_remainder >= 1000) {
2152 			sc->read_data_units++;
2153 			sc->read_dunits_remainder -= 1000;
2154 		}
2155 		break;
2156 	default:
2157 		DPRINTF("%s: Invalid OPC 0x%02x for stats", __func__, opc);
2158 		break;
2159 	}
2160 	pthread_mutex_unlock(&sc->mtx);
2161 }
2162 
2163 /*
2164  * Check if the combination of Starting LBA (slba) and number of blocks
2165  * exceeds the range of the underlying storage.
2166  *
2167  * Because NVMe specifies the SLBA in blocks as a uint64_t and blockif stores
2168  * the capacity in bytes as a uint64_t, care must be taken to avoid integer
2169  * overflow.
2170  */
2171 static bool
2172 pci_nvme_out_of_range(struct pci_nvme_blockstore *nvstore, uint64_t slba,
2173     uint32_t nblocks)
2174 {
2175 	size_t	offset, bytes;
2176 
2177 	/* Overflow check of multiplying Starting LBA by the sector size */
2178 	if (slba >> (64 - nvstore->sectsz_bits))
2179 		return (true);
2180 
2181 	offset = slba << nvstore->sectsz_bits;
2182 	bytes = nblocks << nvstore->sectsz_bits;
2183 
2184 	/* Overflow check of Number of Logical Blocks */
2185 	if ((nvstore->size <= offset) || ((nvstore->size - offset) < bytes))
2186 		return (true);
2187 
2188 	return (false);
2189 }
2190 
2191 static int
2192 pci_nvme_append_iov_req(struct pci_nvme_softc *sc __unused,
2193     struct pci_nvme_ioreq *req, uint64_t gpaddr, size_t size, uint64_t offset)
2194 {
2195 	int iovidx;
2196 	bool range_is_contiguous;
2197 
2198 	if (req == NULL)
2199 		return (-1);
2200 
2201 	if (req->io_req.br_iovcnt == NVME_MAX_IOVEC) {
2202 		return (-1);
2203 	}
2204 
2205 	/*
2206 	 * Minimize the number of IOVs by concatenating contiguous address
2207 	 * ranges. If the IOV count is zero, there is no previous range to
2208 	 * concatenate.
2209 	 */
2210 	if (req->io_req.br_iovcnt == 0)
2211 		range_is_contiguous = false;
2212 	else
2213 		range_is_contiguous = (req->prev_gpaddr + req->prev_size) == gpaddr;
2214 
2215 	if (range_is_contiguous) {
2216 		iovidx = req->io_req.br_iovcnt - 1;
2217 
2218 		req->io_req.br_iov[iovidx].iov_base =
2219 		    paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
2220 				     req->prev_gpaddr, size);
2221 		if (req->io_req.br_iov[iovidx].iov_base == NULL)
2222 			return (-1);
2223 
2224 		req->prev_size += size;
2225 		req->io_req.br_resid += size;
2226 
2227 		req->io_req.br_iov[iovidx].iov_len = req->prev_size;
2228 	} else {
2229 		iovidx = req->io_req.br_iovcnt;
2230 		if (iovidx == 0) {
2231 			req->io_req.br_offset = offset;
2232 			req->io_req.br_resid = 0;
2233 			req->io_req.br_param = req;
2234 		}
2235 
2236 		req->io_req.br_iov[iovidx].iov_base =
2237 		    paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
2238 				     gpaddr, size);
2239 		if (req->io_req.br_iov[iovidx].iov_base == NULL)
2240 			return (-1);
2241 
2242 		req->io_req.br_iov[iovidx].iov_len = size;
2243 
2244 		req->prev_gpaddr = gpaddr;
2245 		req->prev_size = size;
2246 		req->io_req.br_resid += size;
2247 
2248 		req->io_req.br_iovcnt++;
2249 	}
2250 
2251 	return (0);
2252 }
2253 
2254 static void
2255 pci_nvme_set_completion(struct pci_nvme_softc *sc,
2256     struct nvme_submission_queue *sq, int sqid, uint16_t cid, uint16_t status)
2257 {
2258 	struct nvme_completion_queue *cq = &sc->compl_queues[sq->cqid];
2259 
2260 	DPRINTF("%s sqid %d cqid %u cid %u status: 0x%x 0x%x",
2261 		 __func__, sqid, sq->cqid, cid, NVME_STATUS_GET_SCT(status),
2262 		 NVME_STATUS_GET_SC(status));
2263 
2264 	pci_nvme_cq_update(sc, cq, 0, cid, sqid, status);
2265 
2266 	if (cq->head != cq->tail) {
2267 		if (cq->intr_en & NVME_CQ_INTEN) {
2268 			pci_generate_msix(sc->nsc_pi, cq->intr_vec);
2269 		} else {
2270 			DPRINTF("%s: CQ%u interrupt disabled",
2271 						__func__, sq->cqid);
2272 		}
2273 	}
2274 }
2275 
2276 static void
2277 pci_nvme_release_ioreq(struct pci_nvme_softc *sc, struct pci_nvme_ioreq *req)
2278 {
2279 	req->sc = NULL;
2280 	req->nvme_sq = NULL;
2281 	req->sqid = 0;
2282 
2283 	pthread_mutex_lock(&sc->mtx);
2284 
2285 	STAILQ_INSERT_TAIL(&sc->ioreqs_free, req, link);
2286 	sc->pending_ios--;
2287 
2288 	/* when no more IO pending, can set to ready if device reset/enabled */
2289 	if (sc->pending_ios == 0 &&
2290 	    NVME_CC_GET_EN(sc->regs.cc) && !(NVME_CSTS_GET_RDY(sc->regs.csts)))
2291 		sc->regs.csts |= NVME_CSTS_RDY;
2292 
2293 	pthread_mutex_unlock(&sc->mtx);
2294 
2295 	sem_post(&sc->iosemlock);
2296 }
2297 
2298 static struct pci_nvme_ioreq *
2299 pci_nvme_get_ioreq(struct pci_nvme_softc *sc)
2300 {
2301 	struct pci_nvme_ioreq *req = NULL;
2302 
2303 	sem_wait(&sc->iosemlock);
2304 	pthread_mutex_lock(&sc->mtx);
2305 
2306 	req = STAILQ_FIRST(&sc->ioreqs_free);
2307 	assert(req != NULL);
2308 	STAILQ_REMOVE_HEAD(&sc->ioreqs_free, link);
2309 
2310 	req->sc = sc;
2311 
2312 	sc->pending_ios++;
2313 
2314 	pthread_mutex_unlock(&sc->mtx);
2315 
2316 	req->io_req.br_iovcnt = 0;
2317 	req->io_req.br_offset = 0;
2318 	req->io_req.br_resid = 0;
2319 	req->io_req.br_param = req;
2320 	req->prev_gpaddr = 0;
2321 	req->prev_size = 0;
2322 
2323 	return req;
2324 }
2325 
2326 static void
2327 pci_nvme_io_done(struct blockif_req *br, int err)
2328 {
2329 	struct pci_nvme_ioreq *req = br->br_param;
2330 	struct nvme_submission_queue *sq = req->nvme_sq;
2331 	uint16_t code, status;
2332 
2333 	DPRINTF("%s error %d %s", __func__, err, strerror(err));
2334 
2335 	/* TODO return correct error */
2336 	code = err ? NVME_SC_DATA_TRANSFER_ERROR : NVME_SC_SUCCESS;
2337 	status = 0;
2338 	pci_nvme_status_genc(&status, code);
2339 
2340 	pci_nvme_set_completion(req->sc, sq, req->sqid, req->cid, status);
2341 	pci_nvme_stats_write_read_update(req->sc, req->opc,
2342 	    req->bytes, status);
2343 	pci_nvme_release_ioreq(req->sc, req);
2344 }
2345 
2346 /*
2347  * Implements the Flush command. The specification states:
2348  *    If a volatile write cache is not present, Flush commands complete
2349  *    successfully and have no effect
2350  * in the description of the Volatile Write Cache (VWC) field of the Identify
2351  * Controller data. Therefore, set status to Success if the command is
2352  * not supported (i.e. RAM or as indicated by the blockif).
2353  */
2354 static bool
2355 nvme_opc_flush(struct pci_nvme_softc *sc __unused,
2356     struct nvme_command *cmd __unused,
2357     struct pci_nvme_blockstore *nvstore,
2358     struct pci_nvme_ioreq *req,
2359     uint16_t *status)
2360 {
2361 	bool pending = false;
2362 
2363 	if (nvstore->type == NVME_STOR_RAM) {
2364 		pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2365 	} else {
2366 		int err;
2367 
2368 		req->io_req.br_callback = pci_nvme_io_done;
2369 
2370 		err = blockif_flush(nvstore->ctx, &req->io_req);
2371 		switch (err) {
2372 		case 0:
2373 			pending = true;
2374 			break;
2375 		case EOPNOTSUPP:
2376 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2377 			break;
2378 		default:
2379 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2380 		}
2381 	}
2382 
2383 	return (pending);
2384 }
2385 
2386 static uint16_t
2387 nvme_write_read_ram(struct pci_nvme_softc *sc,
2388     struct pci_nvme_blockstore *nvstore,
2389     uint64_t prp1, uint64_t prp2,
2390     size_t offset, uint64_t bytes,
2391     bool is_write)
2392 {
2393 	uint8_t *buf = nvstore->ctx;
2394 	enum nvme_copy_dir dir;
2395 	uint16_t status;
2396 
2397 	if (is_write)
2398 		dir = NVME_COPY_TO_PRP;
2399 	else
2400 		dir = NVME_COPY_FROM_PRP;
2401 
2402 	status = 0;
2403 	if (nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, prp1, prp2,
2404 	    buf + offset, bytes, dir))
2405 		pci_nvme_status_genc(&status,
2406 		    NVME_SC_DATA_TRANSFER_ERROR);
2407 	else
2408 		pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2409 
2410 	return (status);
2411 }
2412 
2413 static uint16_t
2414 nvme_write_read_blockif(struct pci_nvme_softc *sc,
2415     struct pci_nvme_blockstore *nvstore,
2416     struct pci_nvme_ioreq *req,
2417     uint64_t prp1, uint64_t prp2,
2418     size_t offset, uint64_t bytes,
2419     bool is_write)
2420 {
2421 	uint64_t size;
2422 	int err;
2423 	uint16_t status = NVME_NO_STATUS;
2424 
2425 	size = MIN(PAGE_SIZE - (prp1 % PAGE_SIZE), bytes);
2426 	if (pci_nvme_append_iov_req(sc, req, prp1, size, offset)) {
2427 		err = -1;
2428 		goto out;
2429 	}
2430 
2431 	offset += size;
2432 	bytes  -= size;
2433 
2434 	if (bytes == 0) {
2435 		;
2436 	} else if (bytes <= PAGE_SIZE) {
2437 		size = bytes;
2438 		if (pci_nvme_append_iov_req(sc, req, prp2, size, offset)) {
2439 			err = -1;
2440 			goto out;
2441 		}
2442 	} else {
2443 		void *vmctx = sc->nsc_pi->pi_vmctx;
2444 		uint64_t *prp_list = &prp2;
2445 		uint64_t *last = prp_list;
2446 
2447 		/* PRP2 is pointer to a physical region page list */
2448 		while (bytes) {
2449 			/* Last entry in list points to the next list */
2450 			if ((prp_list == last) && (bytes > PAGE_SIZE)) {
2451 				uint64_t prp = *prp_list;
2452 
2453 				prp_list = paddr_guest2host(vmctx, prp,
2454 				    PAGE_SIZE - (prp % PAGE_SIZE));
2455 				if (prp_list == NULL) {
2456 					err = -1;
2457 					goto out;
2458 				}
2459 				last = prp_list + (NVME_PRP2_ITEMS - 1);
2460 			}
2461 
2462 			size = MIN(bytes, PAGE_SIZE);
2463 
2464 			if (pci_nvme_append_iov_req(sc, req, *prp_list, size,
2465 			    offset)) {
2466 				err = -1;
2467 				goto out;
2468 			}
2469 
2470 			offset += size;
2471 			bytes  -= size;
2472 
2473 			prp_list++;
2474 		}
2475 	}
2476 	req->io_req.br_callback = pci_nvme_io_done;
2477 	if (is_write)
2478 		err = blockif_write(nvstore->ctx, &req->io_req);
2479 	else
2480 		err = blockif_read(nvstore->ctx, &req->io_req);
2481 out:
2482 	if (err)
2483 		pci_nvme_status_genc(&status, NVME_SC_DATA_TRANSFER_ERROR);
2484 
2485 	return (status);
2486 }
2487 
2488 static bool
2489 nvme_opc_write_read(struct pci_nvme_softc *sc,
2490     struct nvme_command *cmd,
2491     struct pci_nvme_blockstore *nvstore,
2492     struct pci_nvme_ioreq *req,
2493     uint16_t *status)
2494 {
2495 	uint64_t lba, nblocks, bytes;
2496 	size_t offset;
2497 	bool is_write = cmd->opc == NVME_OPC_WRITE;
2498 	bool pending = false;
2499 
2500 	lba = ((uint64_t)cmd->cdw11 << 32) | cmd->cdw10;
2501 	nblocks = (cmd->cdw12 & 0xFFFF) + 1;
2502 	bytes = nblocks << nvstore->sectsz_bits;
2503 	if (bytes > NVME_MAX_DATA_SIZE) {
2504 		WPRINTF("%s command would exceed MDTS", __func__);
2505 		pci_nvme_status_genc(status, NVME_SC_INVALID_FIELD);
2506 		goto out;
2507 	}
2508 
2509 	if (pci_nvme_out_of_range(nvstore, lba, nblocks)) {
2510 		WPRINTF("%s command would exceed LBA range(slba=%#lx nblocks=%#lx)",
2511 		    __func__, lba, nblocks);
2512 		pci_nvme_status_genc(status, NVME_SC_LBA_OUT_OF_RANGE);
2513 		goto out;
2514 	}
2515 
2516 	offset = lba << nvstore->sectsz_bits;
2517 
2518 	req->bytes = bytes;
2519 	req->io_req.br_offset = lba;
2520 
2521 	/* PRP bits 1:0 must be zero */
2522 	cmd->prp1 &= ~0x3UL;
2523 	cmd->prp2 &= ~0x3UL;
2524 
2525 	if (nvstore->type == NVME_STOR_RAM) {
2526 		*status = nvme_write_read_ram(sc, nvstore, cmd->prp1,
2527 		    cmd->prp2, offset, bytes, is_write);
2528 	} else {
2529 		*status = nvme_write_read_blockif(sc, nvstore, req,
2530 		    cmd->prp1, cmd->prp2, offset, bytes, is_write);
2531 
2532 		if (*status == NVME_NO_STATUS)
2533 			pending = true;
2534 	}
2535 out:
2536 	if (!pending)
2537 		pci_nvme_stats_write_read_update(sc, cmd->opc, bytes, *status);
2538 
2539 	return (pending);
2540 }
2541 
2542 static void
2543 pci_nvme_dealloc_sm(struct blockif_req *br, int err)
2544 {
2545 	struct pci_nvme_ioreq *req = br->br_param;
2546 	struct pci_nvme_softc *sc = req->sc;
2547 	bool done = true;
2548 	uint16_t status;
2549 
2550 	status = 0;
2551 	if (err) {
2552 		pci_nvme_status_genc(&status, NVME_SC_INTERNAL_DEVICE_ERROR);
2553 	} else if ((req->prev_gpaddr + 1) == (req->prev_size)) {
2554 		pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2555 	} else {
2556 		struct iovec *iov = req->io_req.br_iov;
2557 
2558 		req->prev_gpaddr++;
2559 		iov += req->prev_gpaddr;
2560 
2561 		/* The iov_* values already include the sector size */
2562 		req->io_req.br_offset = (off_t)iov->iov_base;
2563 		req->io_req.br_resid = iov->iov_len;
2564 		if (blockif_delete(sc->nvstore.ctx, &req->io_req)) {
2565 			pci_nvme_status_genc(&status,
2566 			    NVME_SC_INTERNAL_DEVICE_ERROR);
2567 		} else
2568 			done = false;
2569 	}
2570 
2571 	if (done) {
2572 		pci_nvme_set_completion(sc, req->nvme_sq, req->sqid, req->cid,
2573 		    status);
2574 		pci_nvme_release_ioreq(sc, req);
2575 	}
2576 }
2577 
2578 static bool
2579 nvme_opc_dataset_mgmt(struct pci_nvme_softc *sc,
2580     struct nvme_command *cmd,
2581     struct pci_nvme_blockstore *nvstore,
2582     struct pci_nvme_ioreq *req,
2583     uint16_t *status)
2584 {
2585 	struct nvme_dsm_range *range = NULL;
2586 	uint32_t nr, r, non_zero, dr;
2587 	int err;
2588 	bool pending = false;
2589 
2590 	if ((sc->ctrldata.oncs & NVME_ONCS_DSM) == 0) {
2591 		pci_nvme_status_genc(status, NVME_SC_INVALID_OPCODE);
2592 		goto out;
2593 	}
2594 
2595 	nr = cmd->cdw10 & 0xff;
2596 
2597 	/* copy locally because a range entry could straddle PRPs */
2598 	range = calloc(1, NVME_MAX_DSM_TRIM);
2599 	if (range == NULL) {
2600 		pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2601 		goto out;
2602 	}
2603 	nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, cmd->prp1, cmd->prp2,
2604 	    (uint8_t *)range, NVME_MAX_DSM_TRIM, NVME_COPY_FROM_PRP);
2605 
2606 	/* Check for invalid ranges and the number of non-zero lengths */
2607 	non_zero = 0;
2608 	for (r = 0; r <= nr; r++) {
2609 		if (pci_nvme_out_of_range(nvstore,
2610 		    range[r].starting_lba, range[r].length)) {
2611 			pci_nvme_status_genc(status, NVME_SC_LBA_OUT_OF_RANGE);
2612 			goto out;
2613 		}
2614 		if (range[r].length != 0)
2615 			non_zero++;
2616 	}
2617 
2618 	if (cmd->cdw11 & NVME_DSM_ATTR_DEALLOCATE) {
2619 		size_t offset, bytes;
2620 		int sectsz_bits = sc->nvstore.sectsz_bits;
2621 
2622 		/*
2623 		 * DSM calls are advisory only, and compliant controllers
2624 		 * may choose to take no actions (i.e. return Success).
2625 		 */
2626 		if (!nvstore->deallocate) {
2627 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2628 			goto out;
2629 		}
2630 
2631 		/* If all ranges have a zero length, return Success */
2632 		if (non_zero == 0) {
2633 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2634 			goto out;
2635 		}
2636 
2637 		if (req == NULL) {
2638 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2639 			goto out;
2640 		}
2641 
2642 		offset = range[0].starting_lba << sectsz_bits;
2643 		bytes = range[0].length << sectsz_bits;
2644 
2645 		/*
2646 		 * If the request is for more than a single range, store
2647 		 * the ranges in the br_iov. Optimize for the common case
2648 		 * of a single range.
2649 		 *
2650 		 * Note that NVMe Number of Ranges is a zero based value
2651 		 */
2652 		req->io_req.br_iovcnt = 0;
2653 		req->io_req.br_offset = offset;
2654 		req->io_req.br_resid = bytes;
2655 
2656 		if (nr == 0) {
2657 			req->io_req.br_callback = pci_nvme_io_done;
2658 		} else {
2659 			struct iovec *iov = req->io_req.br_iov;
2660 
2661 			for (r = 0, dr = 0; r <= nr; r++) {
2662 				offset = range[r].starting_lba << sectsz_bits;
2663 				bytes = range[r].length << sectsz_bits;
2664 				if (bytes == 0)
2665 					continue;
2666 
2667 				if ((nvstore->size - offset) < bytes) {
2668 					pci_nvme_status_genc(status,
2669 					    NVME_SC_LBA_OUT_OF_RANGE);
2670 					goto out;
2671 				}
2672 				iov[dr].iov_base = (void *)offset;
2673 				iov[dr].iov_len = bytes;
2674 				dr++;
2675 			}
2676 			req->io_req.br_callback = pci_nvme_dealloc_sm;
2677 
2678 			/*
2679 			 * Use prev_gpaddr to track the current entry and
2680 			 * prev_size to track the number of entries
2681 			 */
2682 			req->prev_gpaddr = 0;
2683 			req->prev_size = dr;
2684 		}
2685 
2686 		err = blockif_delete(nvstore->ctx, &req->io_req);
2687 		if (err)
2688 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2689 		else
2690 			pending = true;
2691 	}
2692 out:
2693 	free(range);
2694 	return (pending);
2695 }
2696 
2697 static void
2698 pci_nvme_handle_io_cmd(struct pci_nvme_softc* sc, uint16_t idx)
2699 {
2700 	struct nvme_submission_queue *sq;
2701 	uint16_t status;
2702 	uint16_t sqhead;
2703 
2704 	/* handle all submissions up to sq->tail index */
2705 	sq = &sc->submit_queues[idx];
2706 
2707 	pthread_mutex_lock(&sq->mtx);
2708 
2709 	sqhead = sq->head;
2710 	DPRINTF("nvme_handle_io qid %u head %u tail %u cmdlist %p",
2711 	         idx, sqhead, sq->tail, sq->qbase);
2712 
2713 	while (sqhead != atomic_load_acq_short(&sq->tail)) {
2714 		struct nvme_command *cmd;
2715 		struct pci_nvme_ioreq *req;
2716 		uint32_t nsid;
2717 		bool pending;
2718 
2719 		pending = false;
2720 		req = NULL;
2721 		status = 0;
2722 
2723 		cmd = &sq->qbase[sqhead];
2724 		sqhead = (sqhead + 1) % sq->size;
2725 
2726 		nsid = le32toh(cmd->nsid);
2727 		if ((nsid == 0) || (nsid > sc->ctrldata.nn)) {
2728 			pci_nvme_status_genc(&status,
2729 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
2730 			status |=
2731 			    NVME_STATUS_DNR_MASK << NVME_STATUS_DNR_SHIFT;
2732 			goto complete;
2733  		}
2734 
2735 		req = pci_nvme_get_ioreq(sc);
2736 		if (req == NULL) {
2737 			pci_nvme_status_genc(&status,
2738 			    NVME_SC_INTERNAL_DEVICE_ERROR);
2739 			WPRINTF("%s: unable to allocate IO req", __func__);
2740 			goto complete;
2741 		}
2742 		req->nvme_sq = sq;
2743 		req->sqid = idx;
2744 		req->opc = cmd->opc;
2745 		req->cid = cmd->cid;
2746 		req->nsid = cmd->nsid;
2747 
2748 		switch (cmd->opc) {
2749 		case NVME_OPC_FLUSH:
2750 			pending = nvme_opc_flush(sc, cmd, &sc->nvstore,
2751 			    req, &status);
2752  			break;
2753 		case NVME_OPC_WRITE:
2754 		case NVME_OPC_READ:
2755 			pending = nvme_opc_write_read(sc, cmd, &sc->nvstore,
2756 			    req, &status);
2757 			break;
2758 		case NVME_OPC_WRITE_ZEROES:
2759 			/* TODO: write zeroes
2760 			WPRINTF("%s write zeroes lba 0x%lx blocks %u",
2761 			        __func__, lba, cmd->cdw12 & 0xFFFF); */
2762 			pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2763 			break;
2764 		case NVME_OPC_DATASET_MANAGEMENT:
2765  			pending = nvme_opc_dataset_mgmt(sc, cmd, &sc->nvstore,
2766 			    req, &status);
2767 			break;
2768  		default:
2769  			WPRINTF("%s unhandled io command 0x%x",
2770 			    __func__, cmd->opc);
2771 			pci_nvme_status_genc(&status, NVME_SC_INVALID_OPCODE);
2772 		}
2773 complete:
2774 		if (!pending) {
2775 			pci_nvme_set_completion(sc, sq, idx, cmd->cid, status);
2776 			if (req != NULL)
2777 				pci_nvme_release_ioreq(sc, req);
2778 		}
2779 	}
2780 
2781 	sq->head = sqhead;
2782 
2783 	pthread_mutex_unlock(&sq->mtx);
2784 }
2785 
2786 static void
2787 pci_nvme_handle_doorbell(struct pci_nvme_softc* sc,
2788 	uint64_t idx, int is_sq, uint64_t value)
2789 {
2790 	DPRINTF("nvme doorbell %lu, %s, val 0x%lx",
2791 	        idx, is_sq ? "SQ" : "CQ", value & 0xFFFF);
2792 
2793 	if (is_sq) {
2794 		if (idx > sc->num_squeues) {
2795 			WPRINTF("%s queue index %lu overflow from "
2796 			         "guest (max %u)",
2797 			         __func__, idx, sc->num_squeues);
2798 			return;
2799 		}
2800 
2801 		atomic_store_short(&sc->submit_queues[idx].tail,
2802 		                   (uint16_t)value);
2803 
2804 		if (idx == 0) {
2805 			pci_nvme_handle_admin_cmd(sc, value);
2806 		} else {
2807 			/* submission queue; handle new entries in SQ */
2808 			if (idx > sc->num_squeues) {
2809 				WPRINTF("%s SQ index %lu overflow from "
2810 				         "guest (max %u)",
2811 				         __func__, idx, sc->num_squeues);
2812 				return;
2813 			}
2814 			pci_nvme_handle_io_cmd(sc, (uint16_t)idx);
2815 		}
2816 	} else {
2817 		if (idx > sc->num_cqueues) {
2818 			WPRINTF("%s queue index %lu overflow from "
2819 			         "guest (max %u)",
2820 			         __func__, idx, sc->num_cqueues);
2821 			return;
2822 		}
2823 
2824 		atomic_store_short(&sc->compl_queues[idx].head,
2825 				(uint16_t)value);
2826 	}
2827 }
2828 
2829 static void
2830 pci_nvme_bar0_reg_dumps(const char *func, uint64_t offset, int iswrite)
2831 {
2832 	const char *s = iswrite ? "WRITE" : "READ";
2833 
2834 	switch (offset) {
2835 	case NVME_CR_CAP_LOW:
2836 		DPRINTF("%s %s NVME_CR_CAP_LOW", func, s);
2837 		break;
2838 	case NVME_CR_CAP_HI:
2839 		DPRINTF("%s %s NVME_CR_CAP_HI", func, s);
2840 		break;
2841 	case NVME_CR_VS:
2842 		DPRINTF("%s %s NVME_CR_VS", func, s);
2843 		break;
2844 	case NVME_CR_INTMS:
2845 		DPRINTF("%s %s NVME_CR_INTMS", func, s);
2846 		break;
2847 	case NVME_CR_INTMC:
2848 		DPRINTF("%s %s NVME_CR_INTMC", func, s);
2849 		break;
2850 	case NVME_CR_CC:
2851 		DPRINTF("%s %s NVME_CR_CC", func, s);
2852 		break;
2853 	case NVME_CR_CSTS:
2854 		DPRINTF("%s %s NVME_CR_CSTS", func, s);
2855 		break;
2856 	case NVME_CR_NSSR:
2857 		DPRINTF("%s %s NVME_CR_NSSR", func, s);
2858 		break;
2859 	case NVME_CR_AQA:
2860 		DPRINTF("%s %s NVME_CR_AQA", func, s);
2861 		break;
2862 	case NVME_CR_ASQ_LOW:
2863 		DPRINTF("%s %s NVME_CR_ASQ_LOW", func, s);
2864 		break;
2865 	case NVME_CR_ASQ_HI:
2866 		DPRINTF("%s %s NVME_CR_ASQ_HI", func, s);
2867 		break;
2868 	case NVME_CR_ACQ_LOW:
2869 		DPRINTF("%s %s NVME_CR_ACQ_LOW", func, s);
2870 		break;
2871 	case NVME_CR_ACQ_HI:
2872 		DPRINTF("%s %s NVME_CR_ACQ_HI", func, s);
2873 		break;
2874 	default:
2875 		DPRINTF("unknown nvme bar-0 offset 0x%lx", offset);
2876 	}
2877 
2878 }
2879 
2880 static void
2881 pci_nvme_write_bar_0(struct pci_nvme_softc *sc, uint64_t offset, int size,
2882     uint64_t value)
2883 {
2884 	uint32_t ccreg;
2885 
2886 	if (offset >= NVME_DOORBELL_OFFSET) {
2887 		uint64_t belloffset = offset - NVME_DOORBELL_OFFSET;
2888 		uint64_t idx = belloffset / 8; /* door bell size = 2*int */
2889 		int is_sq = (belloffset % 8) < 4;
2890 
2891 		if ((sc->regs.csts & NVME_CSTS_RDY) == 0) {
2892 			WPRINTF("doorbell write prior to RDY (offset=%#lx)\n",
2893 			    offset);
2894 			return;
2895 		}
2896 
2897 		if (belloffset > ((sc->max_queues+1) * 8 - 4)) {
2898 			WPRINTF("guest attempted an overflow write offset "
2899 			         "0x%lx, val 0x%lx in %s",
2900 			         offset, value, __func__);
2901 			return;
2902 		}
2903 
2904 		if (is_sq) {
2905 			if (sc->submit_queues[idx].qbase == NULL)
2906 				return;
2907 		} else if (sc->compl_queues[idx].qbase == NULL)
2908 			return;
2909 
2910 		pci_nvme_handle_doorbell(sc, idx, is_sq, value);
2911 		return;
2912 	}
2913 
2914 	DPRINTF("nvme-write offset 0x%lx, size %d, value 0x%lx",
2915 	        offset, size, value);
2916 
2917 	if (size != 4) {
2918 		WPRINTF("guest wrote invalid size %d (offset 0x%lx, "
2919 		         "val 0x%lx) to bar0 in %s",
2920 		         size, offset, value, __func__);
2921 		/* TODO: shutdown device */
2922 		return;
2923 	}
2924 
2925 	pci_nvme_bar0_reg_dumps(__func__, offset, 1);
2926 
2927 	pthread_mutex_lock(&sc->mtx);
2928 
2929 	switch (offset) {
2930 	case NVME_CR_CAP_LOW:
2931 	case NVME_CR_CAP_HI:
2932 		/* readonly */
2933 		break;
2934 	case NVME_CR_VS:
2935 		/* readonly */
2936 		break;
2937 	case NVME_CR_INTMS:
2938 		/* MSI-X, so ignore */
2939 		break;
2940 	case NVME_CR_INTMC:
2941 		/* MSI-X, so ignore */
2942 		break;
2943 	case NVME_CR_CC:
2944 		ccreg = (uint32_t)value;
2945 
2946 		DPRINTF("%s NVME_CR_CC en %x css %x shn %x iosqes %u "
2947 		         "iocqes %u",
2948 		        __func__,
2949 			 NVME_CC_GET_EN(ccreg), NVME_CC_GET_CSS(ccreg),
2950 			 NVME_CC_GET_SHN(ccreg), NVME_CC_GET_IOSQES(ccreg),
2951 			 NVME_CC_GET_IOCQES(ccreg));
2952 
2953 		if (NVME_CC_GET_SHN(ccreg)) {
2954 			/* perform shutdown - flush out data to backend */
2955 			sc->regs.csts &= ~(NVME_CSTS_REG_SHST_MASK <<
2956 			    NVME_CSTS_REG_SHST_SHIFT);
2957 			sc->regs.csts |= NVME_SHST_COMPLETE <<
2958 			    NVME_CSTS_REG_SHST_SHIFT;
2959 		}
2960 		if (NVME_CC_GET_EN(ccreg) != NVME_CC_GET_EN(sc->regs.cc)) {
2961 			if (NVME_CC_GET_EN(ccreg) == 0)
2962 				/* transition 1-> causes controller reset */
2963 				pci_nvme_reset_locked(sc);
2964 			else
2965 				pci_nvme_init_controller(sc);
2966 		}
2967 
2968 		/* Insert the iocqes, iosqes and en bits from the write */
2969 		sc->regs.cc &= ~NVME_CC_WRITE_MASK;
2970 		sc->regs.cc |= ccreg & NVME_CC_WRITE_MASK;
2971 		if (NVME_CC_GET_EN(ccreg) == 0) {
2972 			/* Insert the ams, mps and css bit fields */
2973 			sc->regs.cc &= ~NVME_CC_NEN_WRITE_MASK;
2974 			sc->regs.cc |= ccreg & NVME_CC_NEN_WRITE_MASK;
2975 			sc->regs.csts &= ~NVME_CSTS_RDY;
2976 		} else if ((sc->pending_ios == 0) &&
2977 		    !(sc->regs.csts & NVME_CSTS_CFS)) {
2978 			sc->regs.csts |= NVME_CSTS_RDY;
2979 		}
2980 		break;
2981 	case NVME_CR_CSTS:
2982 		break;
2983 	case NVME_CR_NSSR:
2984 		/* ignore writes; don't support subsystem reset */
2985 		break;
2986 	case NVME_CR_AQA:
2987 		sc->regs.aqa = (uint32_t)value;
2988 		break;
2989 	case NVME_CR_ASQ_LOW:
2990 		sc->regs.asq = (sc->regs.asq & (0xFFFFFFFF00000000)) |
2991 		               (0xFFFFF000 & value);
2992 		break;
2993 	case NVME_CR_ASQ_HI:
2994 		sc->regs.asq = (sc->regs.asq & (0x00000000FFFFFFFF)) |
2995 		               (value << 32);
2996 		break;
2997 	case NVME_CR_ACQ_LOW:
2998 		sc->regs.acq = (sc->regs.acq & (0xFFFFFFFF00000000)) |
2999 		               (0xFFFFF000 & value);
3000 		break;
3001 	case NVME_CR_ACQ_HI:
3002 		sc->regs.acq = (sc->regs.acq & (0x00000000FFFFFFFF)) |
3003 		               (value << 32);
3004 		break;
3005 	default:
3006 		DPRINTF("%s unknown offset 0x%lx, value 0x%lx size %d",
3007 		         __func__, offset, value, size);
3008 	}
3009 	pthread_mutex_unlock(&sc->mtx);
3010 }
3011 
3012 static void
3013 pci_nvme_write(struct pci_devinst *pi, int baridx, uint64_t offset, int size,
3014     uint64_t value)
3015 {
3016 	struct pci_nvme_softc* sc = pi->pi_arg;
3017 
3018 	if (baridx == pci_msix_table_bar(pi) ||
3019 	    baridx == pci_msix_pba_bar(pi)) {
3020 		DPRINTF("nvme-write baridx %d, msix: off 0x%lx, size %d, "
3021 		         " value 0x%lx", baridx, offset, size, value);
3022 
3023 		pci_emul_msix_twrite(pi, offset, size, value);
3024 		return;
3025 	}
3026 
3027 	switch (baridx) {
3028 	case 0:
3029 		pci_nvme_write_bar_0(sc, offset, size, value);
3030 		break;
3031 
3032 	default:
3033 		DPRINTF("%s unknown baridx %d, val 0x%lx",
3034 		         __func__, baridx, value);
3035 	}
3036 }
3037 
3038 static uint64_t pci_nvme_read_bar_0(struct pci_nvme_softc* sc,
3039 	uint64_t offset, int size)
3040 {
3041 	uint64_t value;
3042 
3043 	pci_nvme_bar0_reg_dumps(__func__, offset, 0);
3044 
3045 	if (offset < NVME_DOORBELL_OFFSET) {
3046 		void *p = &(sc->regs);
3047 		pthread_mutex_lock(&sc->mtx);
3048 		memcpy(&value, (void *)((uintptr_t)p + offset), size);
3049 		pthread_mutex_unlock(&sc->mtx);
3050 	} else {
3051 		value = 0;
3052                 WPRINTF("pci_nvme: read invalid offset %ld", offset);
3053 	}
3054 
3055 	switch (size) {
3056 	case 1:
3057 		value &= 0xFF;
3058 		break;
3059 	case 2:
3060 		value &= 0xFFFF;
3061 		break;
3062 	case 4:
3063 		value &= 0xFFFFFFFF;
3064 		break;
3065 	}
3066 
3067 	DPRINTF("   nvme-read offset 0x%lx, size %d -> value 0x%x",
3068 	         offset, size, (uint32_t)value);
3069 
3070 	return (value);
3071 }
3072 
3073 
3074 
3075 static uint64_t
3076 pci_nvme_read(struct pci_devinst *pi, int baridx, uint64_t offset, int size)
3077 {
3078 	struct pci_nvme_softc* sc = pi->pi_arg;
3079 
3080 	if (baridx == pci_msix_table_bar(pi) ||
3081 	    baridx == pci_msix_pba_bar(pi)) {
3082 		DPRINTF("nvme-read bar: %d, msix: regoff 0x%lx, size %d",
3083 		        baridx, offset, size);
3084 
3085 		return pci_emul_msix_tread(pi, offset, size);
3086 	}
3087 
3088 	switch (baridx) {
3089 	case 0:
3090        		return pci_nvme_read_bar_0(sc, offset, size);
3091 
3092 	default:
3093 		DPRINTF("unknown bar %d, 0x%lx", baridx, offset);
3094 	}
3095 
3096 	return (0);
3097 }
3098 
3099 static int
3100 pci_nvme_parse_config(struct pci_nvme_softc *sc, nvlist_t *nvl)
3101 {
3102 	char bident[sizeof("XXX:XXX")];
3103 	const char *value;
3104 	uint32_t sectsz;
3105 
3106 	sc->max_queues = NVME_QUEUES;
3107 	sc->max_qentries = NVME_MAX_QENTRIES;
3108 	sc->ioslots = NVME_IOSLOTS;
3109 	sc->num_squeues = sc->max_queues;
3110 	sc->num_cqueues = sc->max_queues;
3111 	sc->dataset_management = NVME_DATASET_MANAGEMENT_AUTO;
3112 	sectsz = 0;
3113 	snprintf(sc->ctrldata.sn, sizeof(sc->ctrldata.sn),
3114 	         "NVME-%d-%d", sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
3115 
3116 	value = get_config_value_node(nvl, "maxq");
3117 	if (value != NULL)
3118 		sc->max_queues = atoi(value);
3119 	value = get_config_value_node(nvl, "qsz");
3120 	if (value != NULL) {
3121 		sc->max_qentries = atoi(value);
3122 		if (sc->max_qentries <= 0) {
3123 			EPRINTLN("nvme: Invalid qsz option %d",
3124 			    sc->max_qentries);
3125 			return (-1);
3126 		}
3127 	}
3128 	value = get_config_value_node(nvl, "ioslots");
3129 	if (value != NULL) {
3130 		sc->ioslots = atoi(value);
3131 		if (sc->ioslots <= 0) {
3132 			EPRINTLN("Invalid ioslots option %d", sc->ioslots);
3133 			return (-1);
3134 		}
3135 	}
3136 	value = get_config_value_node(nvl, "sectsz");
3137 	if (value != NULL)
3138 		sectsz = atoi(value);
3139 	value = get_config_value_node(nvl, "ser");
3140 	if (value != NULL) {
3141 		/*
3142 		 * This field indicates the Product Serial Number in
3143 		 * 7-bit ASCII, unused bytes should be space characters.
3144 		 * Ref: NVMe v1.3c.
3145 		 */
3146 		cpywithpad((char *)sc->ctrldata.sn,
3147 		    sizeof(sc->ctrldata.sn), value, ' ');
3148 	}
3149 	value = get_config_value_node(nvl, "eui64");
3150 	if (value != NULL)
3151 		sc->nvstore.eui64 = htobe64(strtoull(value, NULL, 0));
3152 	value = get_config_value_node(nvl, "dsm");
3153 	if (value != NULL) {
3154 		if (strcmp(value, "auto") == 0)
3155 			sc->dataset_management = NVME_DATASET_MANAGEMENT_AUTO;
3156 		else if (strcmp(value, "enable") == 0)
3157 			sc->dataset_management = NVME_DATASET_MANAGEMENT_ENABLE;
3158 		else if (strcmp(value, "disable") == 0)
3159 			sc->dataset_management = NVME_DATASET_MANAGEMENT_DISABLE;
3160 	}
3161 
3162 	value = get_config_value_node(nvl, "ram");
3163 	if (value != NULL) {
3164 		uint64_t sz = strtoull(value, NULL, 10);
3165 
3166 		sc->nvstore.type = NVME_STOR_RAM;
3167 		sc->nvstore.size = sz * 1024 * 1024;
3168 		sc->nvstore.ctx = calloc(1, sc->nvstore.size);
3169 		sc->nvstore.sectsz = 4096;
3170 		sc->nvstore.sectsz_bits = 12;
3171 		if (sc->nvstore.ctx == NULL) {
3172 			EPRINTLN("nvme: Unable to allocate RAM");
3173 			return (-1);
3174 		}
3175 	} else {
3176 		snprintf(bident, sizeof(bident), "%u:%u",
3177 		    sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
3178 		sc->nvstore.ctx = blockif_open(nvl, bident);
3179 		if (sc->nvstore.ctx == NULL) {
3180 			EPRINTLN("nvme: Could not open backing file: %s",
3181 			    strerror(errno));
3182 			return (-1);
3183 		}
3184 		sc->nvstore.type = NVME_STOR_BLOCKIF;
3185 		sc->nvstore.size = blockif_size(sc->nvstore.ctx);
3186 	}
3187 
3188 	if (sectsz == 512 || sectsz == 4096 || sectsz == 8192)
3189 		sc->nvstore.sectsz = sectsz;
3190 	else if (sc->nvstore.type != NVME_STOR_RAM)
3191 		sc->nvstore.sectsz = blockif_sectsz(sc->nvstore.ctx);
3192 	for (sc->nvstore.sectsz_bits = 9;
3193 	     (1U << sc->nvstore.sectsz_bits) < sc->nvstore.sectsz;
3194 	     sc->nvstore.sectsz_bits++);
3195 
3196 	if (sc->max_queues <= 0 || sc->max_queues > NVME_QUEUES)
3197 		sc->max_queues = NVME_QUEUES;
3198 
3199 	return (0);
3200 }
3201 
3202 static void
3203 pci_nvme_resized(struct blockif_ctxt *bctxt __unused, void *arg,
3204     size_t new_size)
3205 {
3206 	struct pci_nvme_softc *sc;
3207 	struct pci_nvme_blockstore *nvstore;
3208 	struct nvme_namespace_data *nd;
3209 
3210 	sc = arg;
3211 	nvstore = &sc->nvstore;
3212 	nd = &sc->nsdata;
3213 
3214 	nvstore->size = new_size;
3215 	pci_nvme_init_nsdata_size(nvstore, nd);
3216 
3217 	/* Add changed NSID to list */
3218 	sc->ns_log.ns[0] = 1;
3219 	sc->ns_log.ns[1] = 0;
3220 
3221 	pci_nvme_aen_post(sc, PCI_NVME_AE_TYPE_NOTICE,
3222 	    PCI_NVME_AEI_NOTICE_NS_ATTR_CHANGED);
3223 }
3224 
3225 static int
3226 pci_nvme_init(struct pci_devinst *pi, nvlist_t *nvl)
3227 {
3228 	struct pci_nvme_softc *sc;
3229 	uint32_t pci_membar_sz;
3230 	int	error;
3231 
3232 	error = 0;
3233 
3234 	sc = calloc(1, sizeof(struct pci_nvme_softc));
3235 	pi->pi_arg = sc;
3236 	sc->nsc_pi = pi;
3237 
3238 	error = pci_nvme_parse_config(sc, nvl);
3239 	if (error < 0)
3240 		goto done;
3241 	else
3242 		error = 0;
3243 
3244 	STAILQ_INIT(&sc->ioreqs_free);
3245 	sc->ioreqs = calloc(sc->ioslots, sizeof(struct pci_nvme_ioreq));
3246 	for (uint32_t i = 0; i < sc->ioslots; i++) {
3247 		STAILQ_INSERT_TAIL(&sc->ioreqs_free, &sc->ioreqs[i], link);
3248 	}
3249 
3250 	pci_set_cfgdata16(pi, PCIR_DEVICE, 0x0A0A);
3251 	pci_set_cfgdata16(pi, PCIR_VENDOR, 0xFB5D);
3252 	pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_STORAGE);
3253 	pci_set_cfgdata8(pi, PCIR_SUBCLASS, PCIS_STORAGE_NVM);
3254 	pci_set_cfgdata8(pi, PCIR_PROGIF,
3255 	                 PCIP_STORAGE_NVM_ENTERPRISE_NVMHCI_1_0);
3256 
3257 	/*
3258 	 * Allocate size of NVMe registers + doorbell space for all queues.
3259 	 *
3260 	 * The specification requires a minimum memory I/O window size of 16K.
3261 	 * The Windows driver will refuse to start a device with a smaller
3262 	 * window.
3263 	 */
3264 	pci_membar_sz = sizeof(struct nvme_registers) +
3265 	    2 * sizeof(uint32_t) * (sc->max_queues + 1);
3266 	pci_membar_sz = MAX(pci_membar_sz, NVME_MMIO_SPACE_MIN);
3267 
3268 	DPRINTF("nvme membar size: %u", pci_membar_sz);
3269 
3270 	error = pci_emul_alloc_bar(pi, 0, PCIBAR_MEM64, pci_membar_sz);
3271 	if (error) {
3272 		WPRINTF("%s pci alloc mem bar failed", __func__);
3273 		goto done;
3274 	}
3275 
3276 	error = pci_emul_add_msixcap(pi, sc->max_queues + 1, NVME_MSIX_BAR);
3277 	if (error) {
3278 		WPRINTF("%s pci add msixcap failed", __func__);
3279 		goto done;
3280 	}
3281 
3282 	error = pci_emul_add_pciecap(pi, PCIEM_TYPE_ROOT_INT_EP);
3283 	if (error) {
3284 		WPRINTF("%s pci add Express capability failed", __func__);
3285 		goto done;
3286 	}
3287 
3288 	pthread_mutex_init(&sc->mtx, NULL);
3289 	sem_init(&sc->iosemlock, 0, sc->ioslots);
3290 	blockif_register_resize_callback(sc->nvstore.ctx, pci_nvme_resized, sc);
3291 
3292 	pci_nvme_init_queues(sc, sc->max_queues, sc->max_queues);
3293 	/*
3294 	 * Controller data depends on Namespace data so initialize Namespace
3295 	 * data first.
3296 	 */
3297 	pci_nvme_init_nsdata(sc, &sc->nsdata, 1, &sc->nvstore);
3298 	pci_nvme_init_ctrldata(sc);
3299 	pci_nvme_init_logpages(sc);
3300 	pci_nvme_init_features(sc);
3301 
3302 	pci_nvme_aer_init(sc);
3303 	pci_nvme_aen_init(sc);
3304 
3305 	pci_nvme_reset(sc);
3306 
3307 	pci_lintr_request(pi);
3308 
3309 done:
3310 	return (error);
3311 }
3312 
3313 static int
3314 pci_nvme_legacy_config(nvlist_t *nvl, const char *opts)
3315 {
3316 	char *cp, *ram;
3317 
3318 	if (opts == NULL)
3319 		return (0);
3320 
3321 	if (strncmp(opts, "ram=", 4) == 0) {
3322 		cp = strchr(opts, ',');
3323 		if (cp == NULL) {
3324 			set_config_value_node(nvl, "ram", opts + 4);
3325 			return (0);
3326 		}
3327 		ram = strndup(opts + 4, cp - opts - 4);
3328 		set_config_value_node(nvl, "ram", ram);
3329 		free(ram);
3330 		return (pci_parse_legacy_config(nvl, cp + 1));
3331 	} else
3332 		return (blockif_legacy_config(nvl, opts));
3333 }
3334 
3335 static const struct pci_devemu pci_de_nvme = {
3336 	.pe_emu =	"nvme",
3337 	.pe_init =	pci_nvme_init,
3338 	.pe_legacy_config = pci_nvme_legacy_config,
3339 	.pe_barwrite =	pci_nvme_write,
3340 	.pe_barread =	pci_nvme_read
3341 };
3342 PCI_EMUL_SET(pci_de_nvme);
3343