xref: /freebsd/usr.sbin/bhyve/pci_nvme.c (revision 3fb8f1272b50cb87cb624b321f7b81e76627c437)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include <sys/errno.h>
58 #include <sys/types.h>
59 #include <sys/crc16.h>
60 #include <net/ieee_oui.h>
61 
62 #include <assert.h>
63 #include <pthread.h>
64 #include <pthread_np.h>
65 #include <semaphore.h>
66 #include <stdbool.h>
67 #include <stddef.h>
68 #include <stdint.h>
69 #include <stdio.h>
70 #include <stdlib.h>
71 #include <string.h>
72 
73 #include <machine/atomic.h>
74 #include <machine/vmm.h>
75 #include <vmmapi.h>
76 
77 #include <dev/nvme/nvme.h>
78 
79 #include "bhyverun.h"
80 #include "block_if.h"
81 #include "config.h"
82 #include "debug.h"
83 #include "pci_emul.h"
84 
85 
86 static int nvme_debug = 0;
87 #define	DPRINTF(fmt, args...) if (nvme_debug) PRINTLN(fmt, ##args)
88 #define	WPRINTF(fmt, args...) PRINTLN(fmt, ##args)
89 
90 /* defaults; can be overridden */
91 #define	NVME_MSIX_BAR		4
92 
93 #define	NVME_IOSLOTS		8
94 
95 /* The NVMe spec defines bits 13:4 in BAR0 as reserved */
96 #define NVME_MMIO_SPACE_MIN	(1 << 14)
97 
98 #define	NVME_QUEUES		16
99 #define	NVME_MAX_QENTRIES	2048
100 /* Memory Page size Minimum reported in CAP register */
101 #define	NVME_MPSMIN		0
102 /* MPSMIN converted to bytes */
103 #define	NVME_MPSMIN_BYTES	(1 << (12 + NVME_MPSMIN))
104 
105 #define	NVME_PRP2_ITEMS		(PAGE_SIZE/sizeof(uint64_t))
106 #define	NVME_MDTS		9
107 /* Note the + 1 allows for the initial descriptor to not be page aligned */
108 #define	NVME_MAX_IOVEC		((1 << NVME_MDTS) + 1)
109 #define	NVME_MAX_DATA_SIZE	((1 << NVME_MDTS) * NVME_MPSMIN_BYTES)
110 
111 /* This is a synthetic status code to indicate there is no status */
112 #define NVME_NO_STATUS		0xffff
113 #define NVME_COMPLETION_VALID(c)	((c).status != NVME_NO_STATUS)
114 
115 /* Reported temperature in Kelvin (i.e. room temperature) */
116 #define NVME_TEMPERATURE 296
117 
118 /* helpers */
119 
120 /* Convert a zero-based value into a one-based value */
121 #define ONE_BASED(zero)		((zero) + 1)
122 /* Convert a one-based value into a zero-based value */
123 #define ZERO_BASED(one)		((one)  - 1)
124 
125 /* Encode number of SQ's and CQ's for Set/Get Features */
126 #define NVME_FEATURE_NUM_QUEUES(sc) \
127 	(ZERO_BASED((sc)->num_squeues) & 0xffff) | \
128 	(ZERO_BASED((sc)->num_cqueues) & 0xffff) << 16
129 
130 #define	NVME_DOORBELL_OFFSET	offsetof(struct nvme_registers, doorbell)
131 
132 enum nvme_controller_register_offsets {
133 	NVME_CR_CAP_LOW = 0x00,
134 	NVME_CR_CAP_HI  = 0x04,
135 	NVME_CR_VS      = 0x08,
136 	NVME_CR_INTMS   = 0x0c,
137 	NVME_CR_INTMC   = 0x10,
138 	NVME_CR_CC      = 0x14,
139 	NVME_CR_CSTS    = 0x1c,
140 	NVME_CR_NSSR    = 0x20,
141 	NVME_CR_AQA     = 0x24,
142 	NVME_CR_ASQ_LOW = 0x28,
143 	NVME_CR_ASQ_HI  = 0x2c,
144 	NVME_CR_ACQ_LOW = 0x30,
145 	NVME_CR_ACQ_HI  = 0x34,
146 };
147 
148 enum nvme_cmd_cdw11 {
149 	NVME_CMD_CDW11_PC  = 0x0001,
150 	NVME_CMD_CDW11_IEN = 0x0002,
151 	NVME_CMD_CDW11_IV  = 0xFFFF0000,
152 };
153 
154 enum nvme_copy_dir {
155 	NVME_COPY_TO_PRP,
156 	NVME_COPY_FROM_PRP,
157 };
158 
159 #define	NVME_CQ_INTEN	0x01
160 #define	NVME_CQ_INTCOAL	0x02
161 
162 struct nvme_completion_queue {
163 	struct nvme_completion *qbase;
164 	pthread_mutex_t	mtx;
165 	uint32_t	size;
166 	uint16_t	tail; /* nvme progress */
167 	uint16_t	head; /* guest progress */
168 	uint16_t	intr_vec;
169 	uint32_t	intr_en;
170 };
171 
172 struct nvme_submission_queue {
173 	struct nvme_command *qbase;
174 	pthread_mutex_t	mtx;
175 	uint32_t	size;
176 	uint16_t	head; /* nvme progress */
177 	uint16_t	tail; /* guest progress */
178 	uint16_t	cqid; /* completion queue id */
179 	int		qpriority;
180 };
181 
182 enum nvme_storage_type {
183 	NVME_STOR_BLOCKIF = 0,
184 	NVME_STOR_RAM = 1,
185 };
186 
187 struct pci_nvme_blockstore {
188 	enum nvme_storage_type type;
189 	void		*ctx;
190 	uint64_t	size;
191 	uint32_t	sectsz;
192 	uint32_t	sectsz_bits;
193 	uint64_t	eui64;
194 	uint32_t	deallocate:1;
195 };
196 
197 /*
198  * Calculate the number of additional page descriptors for guest IO requests
199  * based on the advertised Max Data Transfer (MDTS) and given the number of
200  * default iovec's in a struct blockif_req.
201  */
202 #define MDTS_PAD_SIZE \
203 	( NVME_MAX_IOVEC > BLOCKIF_IOV_MAX ? \
204 	  NVME_MAX_IOVEC - BLOCKIF_IOV_MAX : \
205 	  0 )
206 
207 struct pci_nvme_ioreq {
208 	struct pci_nvme_softc *sc;
209 	STAILQ_ENTRY(pci_nvme_ioreq) link;
210 	struct nvme_submission_queue *nvme_sq;
211 	uint16_t	sqid;
212 
213 	/* command information */
214 	uint16_t	opc;
215 	uint16_t	cid;
216 	uint32_t	nsid;
217 
218 	uint64_t	prev_gpaddr;
219 	size_t		prev_size;
220 	size_t		bytes;
221 
222 	struct blockif_req io_req;
223 
224 	struct iovec	iovpadding[MDTS_PAD_SIZE];
225 };
226 
227 enum nvme_dsm_type {
228 	/* Dataset Management bit in ONCS reflects backing storage capability */
229 	NVME_DATASET_MANAGEMENT_AUTO,
230 	/* Unconditionally set Dataset Management bit in ONCS */
231 	NVME_DATASET_MANAGEMENT_ENABLE,
232 	/* Unconditionally clear Dataset Management bit in ONCS */
233 	NVME_DATASET_MANAGEMENT_DISABLE,
234 };
235 
236 struct pci_nvme_softc;
237 struct nvme_feature_obj;
238 
239 typedef void (*nvme_feature_cb)(struct pci_nvme_softc *,
240     struct nvme_feature_obj *,
241     struct nvme_command *,
242     struct nvme_completion *);
243 
244 struct nvme_feature_obj {
245 	uint32_t	cdw11;
246 	nvme_feature_cb	set;
247 	nvme_feature_cb	get;
248 	bool namespace_specific;
249 };
250 
251 #define NVME_FID_MAX		(NVME_FEAT_ENDURANCE_GROUP_EVENT_CONFIGURATION + 1)
252 
253 typedef enum {
254 	PCI_NVME_AE_TYPE_ERROR = 0,
255 	PCI_NVME_AE_TYPE_SMART,
256 	PCI_NVME_AE_TYPE_NOTICE,
257 	PCI_NVME_AE_TYPE_IO_CMD = 6,
258 	PCI_NVME_AE_TYPE_VENDOR = 7,
259 	PCI_NVME_AE_TYPE_MAX		/* Must be last */
260 } pci_nvme_async_type;
261 
262 /* Asynchronous Event Requests */
263 struct pci_nvme_aer {
264 	STAILQ_ENTRY(pci_nvme_aer) link;
265 	uint16_t	cid;	/* Command ID of the submitted AER */
266 };
267 
268 /** Asynchronous Event Information - Notice */
269 typedef enum {
270 	PCI_NVME_AEI_NOTICE_NS_ATTR_CHANGED = 0,
271 	PCI_NVME_AEI_NOTICE_FW_ACTIVATION,
272 	PCI_NVME_AEI_NOTICE_TELEMETRY_CHANGE,
273 	PCI_NVME_AEI_NOTICE_ANA_CHANGE,
274 	PCI_NVME_AEI_NOTICE_PREDICT_LATENCY_CHANGE,
275 	PCI_NVME_AEI_NOTICE_LBA_STATUS_ALERT,
276 	PCI_NVME_AEI_NOTICE_ENDURANCE_GROUP_CHANGE,
277 	PCI_NVME_AEI_NOTICE_MAX,
278 } pci_nvme_async_event_info_notice;
279 
280 #define PCI_NVME_AEI_NOTICE_SHIFT		8
281 #define PCI_NVME_AEI_NOTICE_MASK(event)	(1 << (event + PCI_NVME_AEI_NOTICE_SHIFT))
282 
283 /* Asynchronous Event Notifications */
284 struct pci_nvme_aen {
285 	pci_nvme_async_type atype;
286 	uint32_t	event_data;
287 	bool		posted;
288 };
289 
290 /*
291  * By default, enable all Asynchrnous Event Notifications:
292  *     SMART / Health Critical Warnings
293  *     Namespace Attribute Notices
294  */
295 #define PCI_NVME_AEN_DEFAULT_MASK	0x11f
296 
297 typedef enum {
298 	NVME_CNTRLTYPE_IO = 1,
299 	NVME_CNTRLTYPE_DISCOVERY = 2,
300 	NVME_CNTRLTYPE_ADMIN = 3,
301 } pci_nvme_cntrl_type;
302 
303 struct pci_nvme_softc {
304 	struct pci_devinst *nsc_pi;
305 
306 	pthread_mutex_t	mtx;
307 
308 	struct nvme_registers regs;
309 
310 	struct nvme_namespace_data  nsdata;
311 	struct nvme_controller_data ctrldata;
312 	struct nvme_error_information_entry err_log;
313 	struct nvme_health_information_page health_log;
314 	struct nvme_firmware_page fw_log;
315 	struct nvme_ns_list ns_log;
316 
317 	struct pci_nvme_blockstore nvstore;
318 
319 	uint16_t	max_qentries;	/* max entries per queue */
320 	uint32_t	max_queues;	/* max number of IO SQ's or CQ's */
321 	uint32_t	num_cqueues;
322 	uint32_t	num_squeues;
323 	bool		num_q_is_set; /* Has host set Number of Queues */
324 
325 	struct pci_nvme_ioreq *ioreqs;
326 	STAILQ_HEAD(, pci_nvme_ioreq) ioreqs_free; /* free list of ioreqs */
327 	uint32_t	pending_ios;
328 	uint32_t	ioslots;
329 	sem_t		iosemlock;
330 
331 	/*
332 	 * Memory mapped Submission and Completion queues
333 	 * Each array includes both Admin and IO queues
334 	 */
335 	struct nvme_completion_queue *compl_queues;
336 	struct nvme_submission_queue *submit_queues;
337 
338 	struct nvme_feature_obj feat[NVME_FID_MAX];
339 
340 	enum nvme_dsm_type dataset_management;
341 
342 	/* Accounting for SMART data */
343 	__uint128_t	read_data_units;
344 	__uint128_t	write_data_units;
345 	__uint128_t	read_commands;
346 	__uint128_t	write_commands;
347 	uint32_t	read_dunits_remainder;
348 	uint32_t	write_dunits_remainder;
349 
350 	STAILQ_HEAD(, pci_nvme_aer) aer_list;
351 	pthread_mutex_t	aer_mtx;
352 	uint32_t	aer_count;
353 	struct pci_nvme_aen aen[PCI_NVME_AE_TYPE_MAX];
354 	pthread_t	aen_tid;
355 	pthread_mutex_t	aen_mtx;
356 	pthread_cond_t	aen_cond;
357 };
358 
359 
360 static void pci_nvme_cq_update(struct pci_nvme_softc *sc,
361     struct nvme_completion_queue *cq,
362     uint32_t cdw0,
363     uint16_t cid,
364     uint16_t sqid,
365     uint16_t status);
366 static struct pci_nvme_ioreq *pci_nvme_get_ioreq(struct pci_nvme_softc *);
367 static void pci_nvme_release_ioreq(struct pci_nvme_softc *, struct pci_nvme_ioreq *);
368 static void pci_nvme_io_done(struct blockif_req *, int);
369 
370 /* Controller Configuration utils */
371 #define	NVME_CC_GET_EN(cc) \
372 	NVMEV(NVME_CC_REG_EN, cc)
373 #define	NVME_CC_GET_CSS(cc) \
374 	NVMEV(NVME_CC_REG_CSS, cc)
375 #define	NVME_CC_GET_SHN(cc) \
376 	NVMEV(NVME_CC_REG_SHN, cc)
377 #define	NVME_CC_GET_IOSQES(cc) \
378 	NVMEV(NVME_CC_REG_IOSQES, cc)
379 #define	NVME_CC_GET_IOCQES(cc) \
380 	NVMEV(NVME_CC_REG_IOCQES, cc)
381 
382 #define	NVME_CC_WRITE_MASK \
383 	(NVMEM(NVME_CC_REG_EN) | \
384 	 NVMEM(NVME_CC_REG_IOSQES) | \
385 	 NVMEM(NVME_CC_REG_IOCQES))
386 
387 #define	NVME_CC_NEN_WRITE_MASK \
388 	(NVMEM(NVME_CC_REG_CSS) | \
389 	 NVMEM(NVME_CC_REG_MPS) | \
390 	 NVMEM(NVME_CC_REG_AMS))
391 
392 /* Controller Status utils */
393 #define	NVME_CSTS_GET_RDY(sts) \
394 	NVMEV(NVME_CSTS_REG_RDY, sts)
395 
396 #define	NVME_CSTS_RDY	(NVMEF(NVME_CSTS_REG_RDY, 1))
397 #define	NVME_CSTS_CFS	(NVMEF(NVME_CSTS_REG_CFS, 1))
398 
399 /* Completion Queue status word utils */
400 #define	NVME_STATUS_P	(NVMEF(NVME_STATUS_P, 1))
401 #define	NVME_STATUS_MASK \
402 	(NVMEM(NVME_STATUS_SCT) | \
403 	 NVMEM(NVME_STATUS_SC))
404 
405 #define NVME_ONCS_DSM	NVMEM(NVME_CTRLR_DATA_ONCS_DSM)
406 
407 static void nvme_feature_invalid_cb(struct pci_nvme_softc *,
408     struct nvme_feature_obj *,
409     struct nvme_command *,
410     struct nvme_completion *);
411 static void nvme_feature_temperature(struct pci_nvme_softc *,
412     struct nvme_feature_obj *,
413     struct nvme_command *,
414     struct nvme_completion *);
415 static void nvme_feature_num_queues(struct pci_nvme_softc *,
416     struct nvme_feature_obj *,
417     struct nvme_command *,
418     struct nvme_completion *);
419 static void nvme_feature_iv_config(struct pci_nvme_softc *,
420     struct nvme_feature_obj *,
421     struct nvme_command *,
422     struct nvme_completion *);
423 static void nvme_feature_async_event(struct pci_nvme_softc *,
424     struct nvme_feature_obj *,
425     struct nvme_command *,
426     struct nvme_completion *);
427 
428 static void *aen_thr(void *arg);
429 
430 static __inline void
431 cpywithpad(char *dst, size_t dst_size, const char *src, char pad)
432 {
433 	size_t len;
434 
435 	len = strnlen(src, dst_size);
436 	memset(dst, pad, dst_size);
437 	memcpy(dst, src, len);
438 }
439 
440 static __inline void
441 pci_nvme_status_tc(uint16_t *status, uint16_t type, uint16_t code)
442 {
443 
444 	*status &= ~NVME_STATUS_MASK;
445 	*status |= NVMEF(NVME_STATUS_SCT, type) | NVMEF(NVME_STATUS_SC, code);
446 }
447 
448 static __inline void
449 pci_nvme_status_genc(uint16_t *status, uint16_t code)
450 {
451 
452 	pci_nvme_status_tc(status, NVME_SCT_GENERIC, code);
453 }
454 
455 /*
456  * Initialize the requested number or IO Submission and Completion Queues.
457  * Admin queues are allocated implicitly.
458  */
459 static void
460 pci_nvme_init_queues(struct pci_nvme_softc *sc, uint32_t nsq, uint32_t ncq)
461 {
462 	uint32_t i;
463 
464 	/*
465 	 * Allocate and initialize the Submission Queues
466 	 */
467 	if (nsq > NVME_QUEUES) {
468 		WPRINTF("%s: clamping number of SQ from %u to %u",
469 					__func__, nsq, NVME_QUEUES);
470 		nsq = NVME_QUEUES;
471 	}
472 
473 	sc->num_squeues = nsq;
474 
475 	sc->submit_queues = calloc(sc->num_squeues + 1,
476 				sizeof(struct nvme_submission_queue));
477 	if (sc->submit_queues == NULL) {
478 		WPRINTF("%s: SQ allocation failed", __func__);
479 		sc->num_squeues = 0;
480 	} else {
481 		struct nvme_submission_queue *sq = sc->submit_queues;
482 
483 		for (i = 0; i < sc->num_squeues + 1; i++)
484 			pthread_mutex_init(&sq[i].mtx, NULL);
485 	}
486 
487 	/*
488 	 * Allocate and initialize the Completion Queues
489 	 */
490 	if (ncq > NVME_QUEUES) {
491 		WPRINTF("%s: clamping number of CQ from %u to %u",
492 					__func__, ncq, NVME_QUEUES);
493 		ncq = NVME_QUEUES;
494 	}
495 
496 	sc->num_cqueues = ncq;
497 
498 	sc->compl_queues = calloc(sc->num_cqueues + 1,
499 				sizeof(struct nvme_completion_queue));
500 	if (sc->compl_queues == NULL) {
501 		WPRINTF("%s: CQ allocation failed", __func__);
502 		sc->num_cqueues = 0;
503 	} else {
504 		struct nvme_completion_queue *cq = sc->compl_queues;
505 
506 		for (i = 0; i < sc->num_cqueues + 1; i++)
507 			pthread_mutex_init(&cq[i].mtx, NULL);
508 	}
509 }
510 
511 static void
512 pci_nvme_init_ctrldata(struct pci_nvme_softc *sc)
513 {
514 	struct nvme_controller_data *cd = &sc->ctrldata;
515 	int ret;
516 
517 	cd->vid = 0xFB5D;
518 	cd->ssvid = 0x0000;
519 
520 	cpywithpad((char *)cd->mn, sizeof(cd->mn), "bhyve-NVMe", ' ');
521 	cpywithpad((char *)cd->fr, sizeof(cd->fr), "1.0", ' ');
522 
523 	/* Num of submission commands that we can handle at a time (2^rab) */
524 	cd->rab   = 4;
525 
526 	/* FreeBSD OUI */
527 	cd->ieee[0] = 0xfc;
528 	cd->ieee[1] = 0x9c;
529 	cd->ieee[2] = 0x58;
530 
531 	cd->mic = 0;
532 
533 	cd->mdts = NVME_MDTS;	/* max data transfer size (2^mdts * CAP.MPSMIN) */
534 
535 	cd->ver = NVME_REV(1,4);
536 
537 	cd->cntrltype = NVME_CNTRLTYPE_IO;
538 	cd->oacs = NVMEF(NVME_CTRLR_DATA_OACS_FORMAT, 1);
539 	cd->oaes = NVMEM(NVME_CTRLR_DATA_OAES_NS_ATTR);
540 	cd->acl = 2;
541 	cd->aerl = 4;
542 
543 	/* Advertise 1, Read-only firmware slot */
544 	cd->frmw = NVMEM(NVME_CTRLR_DATA_FRMW_SLOT1_RO) |
545 	    NVMEF(NVME_CTRLR_DATA_FRMW_NUM_SLOTS, 1);
546 	cd->lpa = 0;	/* TODO: support some simple things like SMART */
547 	cd->elpe = 0;	/* max error log page entries */
548 	/*
549 	 * Report a single power state (zero-based value)
550 	 * power_state[] values are left as zero to indicate "Not reported"
551 	 */
552 	cd->npss = 0;
553 
554 	/* Warning Composite Temperature Threshold */
555 	cd->wctemp = 0x0157;
556 	cd->cctemp = 0x0157;
557 
558 	/* SANICAP must not be 0 for Revision 1.4 and later NVMe Controllers */
559 	cd->sanicap = NVMEF(NVME_CTRLR_DATA_SANICAP_NODMMAS,
560 	    NVME_CTRLR_DATA_SANICAP_NODMMAS_NO);
561 
562 	cd->sqes = NVMEF(NVME_CTRLR_DATA_SQES_MAX, 6) |
563 	    NVMEF(NVME_CTRLR_DATA_SQES_MIN, 6);
564 	cd->cqes = NVMEF(NVME_CTRLR_DATA_CQES_MAX, 4) |
565 	    NVMEF(NVME_CTRLR_DATA_CQES_MIN, 4);
566 	cd->nn = 1;	/* number of namespaces */
567 
568 	cd->oncs = 0;
569 	switch (sc->dataset_management) {
570 	case NVME_DATASET_MANAGEMENT_AUTO:
571 		if (sc->nvstore.deallocate)
572 			cd->oncs |= NVME_ONCS_DSM;
573 		break;
574 	case NVME_DATASET_MANAGEMENT_ENABLE:
575 		cd->oncs |= NVME_ONCS_DSM;
576 		break;
577 	default:
578 		break;
579 	}
580 
581 	cd->fna = NVMEM(NVME_CTRLR_DATA_FNA_FORMAT_ALL);
582 
583 	cd->vwc = NVMEF(NVME_CTRLR_DATA_VWC_ALL, NVME_CTRLR_DATA_VWC_ALL_NO);
584 
585 	ret = snprintf(cd->subnqn, sizeof(cd->subnqn),
586 	    "nqn.2013-12.org.freebsd:bhyve-%s-%u-%u-%u",
587 	    get_config_value("name"), sc->nsc_pi->pi_bus,
588 	    sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
589 	if ((ret < 0) || ((unsigned)ret > sizeof(cd->subnqn)))
590 		EPRINTLN("%s: error setting subnqn (%d)", __func__, ret);
591 }
592 
593 static void
594 pci_nvme_init_nsdata_size(struct pci_nvme_blockstore *nvstore,
595     struct nvme_namespace_data *nd)
596 {
597 
598 	/* Get capacity and block size information from backing store */
599 	nd->nsze = nvstore->size / nvstore->sectsz;
600 	nd->ncap = nd->nsze;
601 	nd->nuse = nd->nsze;
602 }
603 
604 static void
605 pci_nvme_init_nsdata(struct pci_nvme_softc *sc,
606     struct nvme_namespace_data *nd, uint32_t nsid,
607     struct pci_nvme_blockstore *nvstore)
608 {
609 
610 	pci_nvme_init_nsdata_size(nvstore, nd);
611 
612 	if (nvstore->type == NVME_STOR_BLOCKIF)
613 		nvstore->deallocate = blockif_candelete(nvstore->ctx);
614 
615 	nd->nlbaf = 0; /* NLBAF is a 0's based value (i.e. 1 LBA Format) */
616 	nd->flbas = 0;
617 
618 	/* Create an EUI-64 if user did not provide one */
619 	if (nvstore->eui64 == 0) {
620 		char *data = NULL;
621 		uint64_t eui64 = nvstore->eui64;
622 
623 		asprintf(&data, "%s%u%u%u", get_config_value("name"),
624 		    sc->nsc_pi->pi_bus, sc->nsc_pi->pi_slot,
625 		    sc->nsc_pi->pi_func);
626 
627 		if (data != NULL) {
628 			eui64 = OUI_FREEBSD_NVME_LOW | crc16(0, data, strlen(data));
629 			free(data);
630 		}
631 		nvstore->eui64 = (eui64 << 16) | (nsid & 0xffff);
632 	}
633 	be64enc(nd->eui64, nvstore->eui64);
634 
635 	/* LBA data-sz = 2^lbads */
636 	nd->lbaf[0] = NVMEF(NVME_NS_DATA_LBAF_LBADS, nvstore->sectsz_bits);
637 }
638 
639 static void
640 pci_nvme_init_logpages(struct pci_nvme_softc *sc)
641 {
642 	__uint128_t power_cycles = 1;
643 
644 	memset(&sc->err_log, 0, sizeof(sc->err_log));
645 	memset(&sc->health_log, 0, sizeof(sc->health_log));
646 	memset(&sc->fw_log, 0, sizeof(sc->fw_log));
647 	memset(&sc->ns_log, 0, sizeof(sc->ns_log));
648 
649 	/* Set read/write remainder to round up according to spec */
650 	sc->read_dunits_remainder = 999;
651 	sc->write_dunits_remainder = 999;
652 
653 	/* Set nominal Health values checked by implementations */
654 	sc->health_log.temperature = NVME_TEMPERATURE;
655 	sc->health_log.available_spare = 100;
656 	sc->health_log.available_spare_threshold = 10;
657 
658 	/* Set Active Firmware Info to slot 1 */
659 	sc->fw_log.afi = NVMEF(NVME_FIRMWARE_PAGE_AFI_SLOT, 1);
660 	memcpy(&sc->fw_log.revision[0], sc->ctrldata.fr,
661 	    sizeof(sc->fw_log.revision[0]));
662 
663 	memcpy(&sc->health_log.power_cycles, &power_cycles,
664 	    sizeof(sc->health_log.power_cycles));
665 }
666 
667 static void
668 pci_nvme_init_features(struct pci_nvme_softc *sc)
669 {
670 	enum nvme_feature	fid;
671 
672 	for (fid = 0; fid < NVME_FID_MAX; fid++) {
673 		switch (fid) {
674 		case NVME_FEAT_ARBITRATION:
675 		case NVME_FEAT_POWER_MANAGEMENT:
676 		case NVME_FEAT_INTERRUPT_COALESCING: //XXX
677 		case NVME_FEAT_WRITE_ATOMICITY:
678 			/* Mandatory but no special handling required */
679 		//XXX hang - case NVME_FEAT_PREDICTABLE_LATENCY_MODE_CONFIG:
680 		//XXX hang - case NVME_FEAT_HOST_BEHAVIOR_SUPPORT:
681 		//		  this returns a data buffer
682 			break;
683 		case NVME_FEAT_TEMPERATURE_THRESHOLD:
684 			sc->feat[fid].set = nvme_feature_temperature;
685 			break;
686 		case NVME_FEAT_ERROR_RECOVERY:
687 			sc->feat[fid].namespace_specific = true;
688 			break;
689 		case NVME_FEAT_NUMBER_OF_QUEUES:
690 			sc->feat[fid].set = nvme_feature_num_queues;
691 			break;
692 		case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
693 			sc->feat[fid].set = nvme_feature_iv_config;
694 			break;
695 		case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
696 			sc->feat[fid].set = nvme_feature_async_event;
697 			/* Enable all AENs by default */
698 			sc->feat[fid].cdw11 = PCI_NVME_AEN_DEFAULT_MASK;
699 			break;
700 		default:
701 			sc->feat[fid].set = nvme_feature_invalid_cb;
702 			sc->feat[fid].get = nvme_feature_invalid_cb;
703 		}
704 	}
705 }
706 
707 static void
708 pci_nvme_aer_reset(struct pci_nvme_softc *sc)
709 {
710 
711 	STAILQ_INIT(&sc->aer_list);
712 	sc->aer_count = 0;
713 }
714 
715 static void
716 pci_nvme_aer_init(struct pci_nvme_softc *sc)
717 {
718 
719 	pthread_mutex_init(&sc->aer_mtx, NULL);
720 	pci_nvme_aer_reset(sc);
721 }
722 
723 static void
724 pci_nvme_aer_destroy(struct pci_nvme_softc *sc)
725 {
726 	struct pci_nvme_aer *aer = NULL;
727 
728 	pthread_mutex_lock(&sc->aer_mtx);
729 	while (!STAILQ_EMPTY(&sc->aer_list)) {
730 		aer = STAILQ_FIRST(&sc->aer_list);
731 		STAILQ_REMOVE_HEAD(&sc->aer_list, link);
732 		free(aer);
733 	}
734 	pthread_mutex_unlock(&sc->aer_mtx);
735 
736 	pci_nvme_aer_reset(sc);
737 }
738 
739 static bool
740 pci_nvme_aer_available(struct pci_nvme_softc *sc)
741 {
742 
743 	return (sc->aer_count != 0);
744 }
745 
746 static bool
747 pci_nvme_aer_limit_reached(struct pci_nvme_softc *sc)
748 {
749 	struct nvme_controller_data *cd = &sc->ctrldata;
750 
751 	/* AERL is a zero based value while aer_count is one's based */
752 	return (sc->aer_count == (cd->aerl + 1U));
753 }
754 
755 /*
756  * Add an Async Event Request
757  *
758  * Stores an AER to be returned later if the Controller needs to notify the
759  * host of an event.
760  * Note that while the NVMe spec doesn't require Controllers to return AER's
761  * in order, this implementation does preserve the order.
762  */
763 static int
764 pci_nvme_aer_add(struct pci_nvme_softc *sc, uint16_t cid)
765 {
766 	struct pci_nvme_aer *aer = NULL;
767 
768 	aer = calloc(1, sizeof(struct pci_nvme_aer));
769 	if (aer == NULL)
770 		return (-1);
771 
772 	/* Save the Command ID for use in the completion message */
773 	aer->cid = cid;
774 
775 	pthread_mutex_lock(&sc->aer_mtx);
776 	sc->aer_count++;
777 	STAILQ_INSERT_TAIL(&sc->aer_list, aer, link);
778 	pthread_mutex_unlock(&sc->aer_mtx);
779 
780 	return (0);
781 }
782 
783 /*
784  * Get an Async Event Request structure
785  *
786  * Returns a pointer to an AER previously submitted by the host or NULL if
787  * no AER's exist. Caller is responsible for freeing the returned struct.
788  */
789 static struct pci_nvme_aer *
790 pci_nvme_aer_get(struct pci_nvme_softc *sc)
791 {
792 	struct pci_nvme_aer *aer = NULL;
793 
794 	pthread_mutex_lock(&sc->aer_mtx);
795 	aer = STAILQ_FIRST(&sc->aer_list);
796 	if (aer != NULL) {
797 		STAILQ_REMOVE_HEAD(&sc->aer_list, link);
798 		sc->aer_count--;
799 	}
800 	pthread_mutex_unlock(&sc->aer_mtx);
801 
802 	return (aer);
803 }
804 
805 static void
806 pci_nvme_aen_reset(struct pci_nvme_softc *sc)
807 {
808 	uint32_t	atype;
809 
810 	memset(sc->aen, 0, PCI_NVME_AE_TYPE_MAX * sizeof(struct pci_nvme_aen));
811 
812 	for (atype = 0; atype < PCI_NVME_AE_TYPE_MAX; atype++) {
813 		sc->aen[atype].atype = atype;
814 	}
815 }
816 
817 static void
818 pci_nvme_aen_init(struct pci_nvme_softc *sc)
819 {
820 	char nstr[80];
821 
822 	pci_nvme_aen_reset(sc);
823 
824 	pthread_mutex_init(&sc->aen_mtx, NULL);
825 	pthread_create(&sc->aen_tid, NULL, aen_thr, sc);
826 	snprintf(nstr, sizeof(nstr), "nvme-aen-%d:%d", sc->nsc_pi->pi_slot,
827 	    sc->nsc_pi->pi_func);
828 	pthread_set_name_np(sc->aen_tid, nstr);
829 }
830 
831 static void
832 pci_nvme_aen_destroy(struct pci_nvme_softc *sc)
833 {
834 
835 	pci_nvme_aen_reset(sc);
836 }
837 
838 /* Notify the AEN thread of pending work */
839 static void
840 pci_nvme_aen_notify(struct pci_nvme_softc *sc)
841 {
842 
843 	pthread_cond_signal(&sc->aen_cond);
844 }
845 
846 /*
847  * Post an Asynchronous Event Notification
848  */
849 static int32_t
850 pci_nvme_aen_post(struct pci_nvme_softc *sc, pci_nvme_async_type atype,
851 		uint32_t event_data)
852 {
853 	struct pci_nvme_aen *aen;
854 
855 	if (atype >= PCI_NVME_AE_TYPE_MAX) {
856 		return(EINVAL);
857 	}
858 
859 	pthread_mutex_lock(&sc->aen_mtx);
860 	aen = &sc->aen[atype];
861 
862 	/* Has the controller already posted an event of this type? */
863 	if (aen->posted) {
864 		pthread_mutex_unlock(&sc->aen_mtx);
865 		return(EALREADY);
866 	}
867 
868 	aen->event_data = event_data;
869 	aen->posted = true;
870 	pthread_mutex_unlock(&sc->aen_mtx);
871 
872 	pci_nvme_aen_notify(sc);
873 
874 	return(0);
875 }
876 
877 static void
878 pci_nvme_aen_process(struct pci_nvme_softc *sc)
879 {
880 	struct pci_nvme_aer *aer;
881 	struct pci_nvme_aen *aen;
882 	pci_nvme_async_type atype;
883 	uint32_t mask;
884 	uint16_t status;
885 	uint8_t lid;
886 
887 	assert(pthread_mutex_isowned_np(&sc->aen_mtx));
888 	for (atype = 0; atype < PCI_NVME_AE_TYPE_MAX; atype++) {
889 		aen = &sc->aen[atype];
890 		/* Previous iterations may have depleted the available AER's */
891 		if (!pci_nvme_aer_available(sc)) {
892 			DPRINTF("%s: no AER", __func__);
893 			break;
894 		}
895 
896 		if (!aen->posted) {
897 			DPRINTF("%s: no AEN posted for atype=%#x", __func__, atype);
898 			continue;
899 		}
900 
901 		status = NVME_SC_SUCCESS;
902 
903 		/* Is the event masked? */
904 		mask =
905 		    sc->feat[NVME_FEAT_ASYNC_EVENT_CONFIGURATION].cdw11;
906 
907 		DPRINTF("%s: atype=%#x mask=%#x event_data=%#x", __func__, atype, mask, aen->event_data);
908 		switch (atype) {
909 		case PCI_NVME_AE_TYPE_ERROR:
910 			lid = NVME_LOG_ERROR;
911 			break;
912 		case PCI_NVME_AE_TYPE_SMART:
913 			mask &= 0xff;
914 			if ((mask & aen->event_data) == 0)
915 				continue;
916 			lid = NVME_LOG_HEALTH_INFORMATION;
917 			break;
918 		case PCI_NVME_AE_TYPE_NOTICE:
919 			if (aen->event_data >= PCI_NVME_AEI_NOTICE_MAX) {
920 				EPRINTLN("%s unknown AEN notice type %u",
921 				    __func__, aen->event_data);
922 				status = NVME_SC_INTERNAL_DEVICE_ERROR;
923 				lid = 0;
924 				break;
925 			}
926 			if ((PCI_NVME_AEI_NOTICE_MASK(aen->event_data) & mask) == 0)
927 				continue;
928 			switch (aen->event_data) {
929 			case PCI_NVME_AEI_NOTICE_NS_ATTR_CHANGED:
930 				lid = NVME_LOG_CHANGED_NAMESPACE;
931 				break;
932 			case PCI_NVME_AEI_NOTICE_FW_ACTIVATION:
933 				lid = NVME_LOG_FIRMWARE_SLOT;
934 				break;
935 			case PCI_NVME_AEI_NOTICE_TELEMETRY_CHANGE:
936 				lid = NVME_LOG_TELEMETRY_CONTROLLER_INITIATED;
937 				break;
938 			case PCI_NVME_AEI_NOTICE_ANA_CHANGE:
939 				lid = NVME_LOG_ASYMMETRIC_NAMESPACE_ACCESS;
940 				break;
941 			case PCI_NVME_AEI_NOTICE_PREDICT_LATENCY_CHANGE:
942 				lid = NVME_LOG_PREDICTABLE_LATENCY_EVENT_AGGREGATE;
943 				break;
944 			case PCI_NVME_AEI_NOTICE_LBA_STATUS_ALERT:
945 				lid = NVME_LOG_LBA_STATUS_INFORMATION;
946 				break;
947 			case PCI_NVME_AEI_NOTICE_ENDURANCE_GROUP_CHANGE:
948 				lid = NVME_LOG_ENDURANCE_GROUP_EVENT_AGGREGATE;
949 				break;
950 			default:
951 				lid = 0;
952 			}
953 			break;
954 		default:
955 			/* bad type?!? */
956 			EPRINTLN("%s unknown AEN type %u", __func__, atype);
957 			status = NVME_SC_INTERNAL_DEVICE_ERROR;
958 			lid = 0;
959 			break;
960 		}
961 
962 		aer = pci_nvme_aer_get(sc);
963 		assert(aer != NULL);
964 
965 		DPRINTF("%s: CID=%#x CDW0=%#x", __func__, aer->cid, (lid << 16) | (aen->event_data << 8) | atype);
966 		pci_nvme_cq_update(sc, &sc->compl_queues[0],
967 		    (lid << 16) | (aen->event_data << 8) | atype, /* cdw0 */
968 		    aer->cid,
969 		    0,		/* SQID */
970 		    status);
971 
972 		aen->event_data = 0;
973 		aen->posted = false;
974 
975 		pci_generate_msix(sc->nsc_pi, 0);
976 	}
977 }
978 
979 static void *
980 aen_thr(void *arg)
981 {
982 	struct pci_nvme_softc *sc;
983 
984 	sc = arg;
985 
986 	pthread_mutex_lock(&sc->aen_mtx);
987 	for (;;) {
988 		pci_nvme_aen_process(sc);
989 		pthread_cond_wait(&sc->aen_cond, &sc->aen_mtx);
990 	}
991 	pthread_mutex_unlock(&sc->aen_mtx);
992 
993 	pthread_exit(NULL);
994 	return (NULL);
995 }
996 
997 static void
998 pci_nvme_reset_locked(struct pci_nvme_softc *sc)
999 {
1000 	uint32_t i;
1001 
1002 	DPRINTF("%s", __func__);
1003 
1004 	sc->regs.cap_lo = (ZERO_BASED(sc->max_qentries) & NVME_CAP_LO_REG_MQES_MASK) |
1005 	    NVMEF(NVME_CAP_LO_REG_CQR, 1) |
1006 	    NVMEF(NVME_CAP_LO_REG_TO, 60);
1007 
1008 	sc->regs.cap_hi = NVMEF(NVME_CAP_HI_REG_CSS_NVM, 1);
1009 
1010 	sc->regs.vs = NVME_REV(1,4);	/* NVMe v1.4 */
1011 
1012 	sc->regs.cc = 0;
1013 
1014 	assert(sc->submit_queues != NULL);
1015 
1016 	for (i = 0; i < sc->num_squeues + 1; i++) {
1017 		sc->submit_queues[i].qbase = NULL;
1018 		sc->submit_queues[i].size = 0;
1019 		sc->submit_queues[i].cqid = 0;
1020 		sc->submit_queues[i].tail = 0;
1021 		sc->submit_queues[i].head = 0;
1022 	}
1023 
1024 	assert(sc->compl_queues != NULL);
1025 
1026 	for (i = 0; i < sc->num_cqueues + 1; i++) {
1027 		sc->compl_queues[i].qbase = NULL;
1028 		sc->compl_queues[i].size = 0;
1029 		sc->compl_queues[i].tail = 0;
1030 		sc->compl_queues[i].head = 0;
1031 	}
1032 
1033 	sc->num_q_is_set = false;
1034 
1035 	pci_nvme_aer_destroy(sc);
1036 	pci_nvme_aen_destroy(sc);
1037 
1038 	/*
1039 	 * Clear CSTS.RDY last to prevent the host from enabling Controller
1040 	 * before cleanup completes
1041 	 */
1042 	sc->regs.csts = 0;
1043 }
1044 
1045 static void
1046 pci_nvme_reset(struct pci_nvme_softc *sc)
1047 {
1048 	pthread_mutex_lock(&sc->mtx);
1049 	pci_nvme_reset_locked(sc);
1050 	pthread_mutex_unlock(&sc->mtx);
1051 }
1052 
1053 static int
1054 pci_nvme_init_controller(struct pci_nvme_softc *sc)
1055 {
1056 	uint16_t acqs, asqs;
1057 
1058 	DPRINTF("%s", __func__);
1059 
1060 	/*
1061 	 * NVMe 2.0 states that "enabling a controller while this field is
1062 	 * cleared to 0h produces undefined results" for both ACQS and
1063 	 * ASQS. If zero, set CFS and do not become ready.
1064 	 */
1065 	asqs = ONE_BASED(NVMEV(NVME_AQA_REG_ASQS, sc->regs.aqa));
1066 	if (asqs < 2) {
1067 		EPRINTLN("%s: illegal ASQS value %#x (aqa=%#x)", __func__,
1068 		    asqs - 1, sc->regs.aqa);
1069 		sc->regs.csts |= NVME_CSTS_CFS;
1070 		return (-1);
1071 	}
1072 	sc->submit_queues[0].size = asqs;
1073 	sc->submit_queues[0].qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1074 	    sc->regs.asq, sizeof(struct nvme_command) * asqs);
1075 	if (sc->submit_queues[0].qbase == NULL) {
1076 		EPRINTLN("%s: ASQ vm_map_gpa(%lx) failed", __func__,
1077 		    sc->regs.asq);
1078 		sc->regs.csts |= NVME_CSTS_CFS;
1079 		return (-1);
1080 	}
1081 
1082 	DPRINTF("%s mapping Admin-SQ guest 0x%lx, host: %p",
1083 	        __func__, sc->regs.asq, sc->submit_queues[0].qbase);
1084 
1085 	acqs = ONE_BASED(NVMEV(NVME_AQA_REG_ACQS, sc->regs.aqa));
1086 	if (acqs < 2) {
1087 		EPRINTLN("%s: illegal ACQS value %#x (aqa=%#x)", __func__,
1088 		    acqs - 1, sc->regs.aqa);
1089 		sc->regs.csts |= NVME_CSTS_CFS;
1090 		return (-1);
1091 	}
1092 	sc->compl_queues[0].size = acqs;
1093 	sc->compl_queues[0].qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1094 	    sc->regs.acq, sizeof(struct nvme_completion) * acqs);
1095 	if (sc->compl_queues[0].qbase == NULL) {
1096 		EPRINTLN("%s: ACQ vm_map_gpa(%lx) failed", __func__,
1097 		    sc->regs.acq);
1098 		sc->regs.csts |= NVME_CSTS_CFS;
1099 		return (-1);
1100 	}
1101 	sc->compl_queues[0].intr_en = NVME_CQ_INTEN;
1102 
1103 	DPRINTF("%s mapping Admin-CQ guest 0x%lx, host: %p",
1104 	        __func__, sc->regs.acq, sc->compl_queues[0].qbase);
1105 
1106 	return (0);
1107 }
1108 
1109 static int
1110 nvme_prp_memcpy(struct vmctx *ctx, uint64_t prp1, uint64_t prp2, uint8_t *b,
1111 	size_t len, enum nvme_copy_dir dir)
1112 {
1113 	uint8_t *p;
1114 	size_t bytes;
1115 
1116 	if (len > (8 * 1024)) {
1117 		return (-1);
1118 	}
1119 
1120 	/* Copy from the start of prp1 to the end of the physical page */
1121 	bytes = PAGE_SIZE - (prp1 & PAGE_MASK);
1122 	bytes = MIN(bytes, len);
1123 
1124 	p = vm_map_gpa(ctx, prp1, bytes);
1125 	if (p == NULL) {
1126 		return (-1);
1127 	}
1128 
1129 	if (dir == NVME_COPY_TO_PRP)
1130 		memcpy(p, b, bytes);
1131 	else
1132 		memcpy(b, p, bytes);
1133 
1134 	b += bytes;
1135 
1136 	len -= bytes;
1137 	if (len == 0) {
1138 		return (0);
1139 	}
1140 
1141 	len = MIN(len, PAGE_SIZE);
1142 
1143 	p = vm_map_gpa(ctx, prp2, len);
1144 	if (p == NULL) {
1145 		return (-1);
1146 	}
1147 
1148 	if (dir == NVME_COPY_TO_PRP)
1149 		memcpy(p, b, len);
1150 	else
1151 		memcpy(b, p, len);
1152 
1153 	return (0);
1154 }
1155 
1156 /*
1157  * Write a Completion Queue Entry update
1158  *
1159  * Write the completion and update the doorbell value
1160  */
1161 static void
1162 pci_nvme_cq_update(struct pci_nvme_softc *sc,
1163 		struct nvme_completion_queue *cq,
1164 		uint32_t cdw0,
1165 		uint16_t cid,
1166 		uint16_t sqid,
1167 		uint16_t status)
1168 {
1169 	struct nvme_submission_queue *sq = &sc->submit_queues[sqid];
1170 	struct nvme_completion *cqe;
1171 
1172 	assert(cq->qbase != NULL);
1173 
1174 	pthread_mutex_lock(&cq->mtx);
1175 
1176 	cqe = &cq->qbase[cq->tail];
1177 
1178 	/* Flip the phase bit */
1179 	status |= (cqe->status ^ NVME_STATUS_P) & NVME_STATUS_P_MASK;
1180 
1181 	cqe->cdw0 = cdw0;
1182 	cqe->sqhd = sq->head;
1183 	cqe->sqid = sqid;
1184 	cqe->cid = cid;
1185 	cqe->status = status;
1186 
1187 	cq->tail++;
1188 	if (cq->tail >= cq->size) {
1189 		cq->tail = 0;
1190 	}
1191 
1192 	pthread_mutex_unlock(&cq->mtx);
1193 }
1194 
1195 static int
1196 nvme_opc_delete_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
1197 	struct nvme_completion* compl)
1198 {
1199 	uint16_t qid = command->cdw10 & 0xffff;
1200 
1201 	DPRINTF("%s DELETE_IO_SQ %u", __func__, qid);
1202 	if (qid == 0 || qid > sc->num_squeues ||
1203 	    (sc->submit_queues[qid].qbase == NULL)) {
1204 		WPRINTF("%s NOT PERMITTED queue id %u / num_squeues %u",
1205 		        __func__, qid, sc->num_squeues);
1206 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1207 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1208 		return (1);
1209 	}
1210 
1211 	sc->submit_queues[qid].qbase = NULL;
1212 	sc->submit_queues[qid].cqid = 0;
1213 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1214 	return (1);
1215 }
1216 
1217 static int
1218 nvme_opc_create_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
1219 	struct nvme_completion* compl)
1220 {
1221 	if (command->cdw11 & NVME_CMD_CDW11_PC) {
1222 		uint16_t qid = command->cdw10 & 0xffff;
1223 		struct nvme_submission_queue *nsq;
1224 
1225 		if ((qid == 0) || (qid > sc->num_squeues) ||
1226 		    (sc->submit_queues[qid].qbase != NULL)) {
1227 			WPRINTF("%s queue index %u > num_squeues %u",
1228 			        __func__, qid, sc->num_squeues);
1229 			pci_nvme_status_tc(&compl->status,
1230 			    NVME_SCT_COMMAND_SPECIFIC,
1231 			    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1232 			return (1);
1233 		}
1234 
1235 		nsq = &sc->submit_queues[qid];
1236 		nsq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
1237 		DPRINTF("%s size=%u (max=%u)", __func__, nsq->size, sc->max_qentries);
1238 		if ((nsq->size < 2) || (nsq->size > sc->max_qentries)) {
1239 			/*
1240 			 * Queues must specify at least two entries
1241 			 * NOTE: "MAXIMUM QUEUE SIZE EXCEEDED" was renamed to
1242 			 * "INVALID QUEUE SIZE" in the NVM Express 1.3 Spec
1243 			 */
1244 			pci_nvme_status_tc(&compl->status,
1245 			    NVME_SCT_COMMAND_SPECIFIC,
1246 			    NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED);
1247 			return (1);
1248 		}
1249 		nsq->head = nsq->tail = 0;
1250 
1251 		nsq->cqid = (command->cdw11 >> 16) & 0xffff;
1252 		if ((nsq->cqid == 0) || (nsq->cqid > sc->num_cqueues)) {
1253 			pci_nvme_status_tc(&compl->status,
1254 			    NVME_SCT_COMMAND_SPECIFIC,
1255 			    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1256 			return (1);
1257 		}
1258 
1259 		if (sc->compl_queues[nsq->cqid].qbase == NULL) {
1260 			pci_nvme_status_tc(&compl->status,
1261 			    NVME_SCT_COMMAND_SPECIFIC,
1262 			    NVME_SC_COMPLETION_QUEUE_INVALID);
1263 			return (1);
1264 		}
1265 
1266 		nsq->qpriority = (command->cdw11 >> 1) & 0x03;
1267 
1268 		nsq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1269 		              sizeof(struct nvme_command) * (size_t)nsq->size);
1270 
1271 		DPRINTF("%s sq %u size %u gaddr %p cqid %u", __func__,
1272 		        qid, nsq->size, nsq->qbase, nsq->cqid);
1273 
1274 		pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1275 
1276 		DPRINTF("%s completed creating IOSQ qid %u",
1277 		         __func__, qid);
1278 	} else {
1279 		/*
1280 		 * Guest sent non-cont submission queue request.
1281 		 * This setting is unsupported by this emulation.
1282 		 */
1283 		WPRINTF("%s unsupported non-contig (list-based) "
1284 		         "create i/o submission queue", __func__);
1285 
1286 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1287 	}
1288 	return (1);
1289 }
1290 
1291 static int
1292 nvme_opc_delete_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
1293 	struct nvme_completion* compl)
1294 {
1295 	uint16_t qid = command->cdw10 & 0xffff;
1296 	uint16_t sqid;
1297 
1298 	DPRINTF("%s DELETE_IO_CQ %u", __func__, qid);
1299 	if (qid == 0 || qid > sc->num_cqueues ||
1300 	    (sc->compl_queues[qid].qbase == NULL)) {
1301 		WPRINTF("%s queue index %u / num_cqueues %u",
1302 		        __func__, qid, sc->num_cqueues);
1303 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1304 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1305 		return (1);
1306 	}
1307 
1308 	/* Deleting an Active CQ is an error */
1309 	for (sqid = 1; sqid < sc->num_squeues + 1; sqid++)
1310 		if (sc->submit_queues[sqid].cqid == qid) {
1311 			pci_nvme_status_tc(&compl->status,
1312 			    NVME_SCT_COMMAND_SPECIFIC,
1313 			    NVME_SC_INVALID_QUEUE_DELETION);
1314 			return (1);
1315 		}
1316 
1317 	sc->compl_queues[qid].qbase = NULL;
1318 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1319 	return (1);
1320 }
1321 
1322 static int
1323 nvme_opc_create_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
1324 	struct nvme_completion* compl)
1325 {
1326 	struct nvme_completion_queue *ncq;
1327 	uint16_t qid = command->cdw10 & 0xffff;
1328 
1329 	/* Only support Physically Contiguous queues */
1330 	if ((command->cdw11 & NVME_CMD_CDW11_PC) == 0) {
1331 		WPRINTF("%s unsupported non-contig (list-based) "
1332 		         "create i/o completion queue",
1333 		         __func__);
1334 
1335 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1336 		return (1);
1337 	}
1338 
1339 	if ((qid == 0) || (qid > sc->num_cqueues) ||
1340 	    (sc->compl_queues[qid].qbase != NULL)) {
1341 		WPRINTF("%s queue index %u > num_cqueues %u",
1342 			__func__, qid, sc->num_cqueues);
1343 		pci_nvme_status_tc(&compl->status,
1344 		    NVME_SCT_COMMAND_SPECIFIC,
1345 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1346 		return (1);
1347  	}
1348 
1349 	ncq = &sc->compl_queues[qid];
1350 	ncq->intr_en = (command->cdw11 & NVME_CMD_CDW11_IEN) >> 1;
1351 	ncq->intr_vec = (command->cdw11 >> 16) & 0xffff;
1352 	if (ncq->intr_vec > (sc->max_queues + 1)) {
1353 		pci_nvme_status_tc(&compl->status,
1354 		    NVME_SCT_COMMAND_SPECIFIC,
1355 		    NVME_SC_INVALID_INTERRUPT_VECTOR);
1356 		return (1);
1357 	}
1358 
1359 	ncq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
1360 	if ((ncq->size < 2) || (ncq->size > sc->max_qentries))  {
1361 		/*
1362 		 * Queues must specify at least two entries
1363 		 * NOTE: "MAXIMUM QUEUE SIZE EXCEEDED" was renamed to
1364 		 * "INVALID QUEUE SIZE" in the NVM Express 1.3 Spec
1365 		 */
1366 		pci_nvme_status_tc(&compl->status,
1367 		    NVME_SCT_COMMAND_SPECIFIC,
1368 		    NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED);
1369 		return (1);
1370 	}
1371 	ncq->head = ncq->tail = 0;
1372 	ncq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1373 		     command->prp1,
1374 		     sizeof(struct nvme_command) * (size_t)ncq->size);
1375 
1376 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1377 
1378 
1379 	return (1);
1380 }
1381 
1382 static int
1383 nvme_opc_get_log_page(struct pci_nvme_softc* sc, struct nvme_command* command,
1384 	struct nvme_completion* compl)
1385 {
1386 	uint64_t logoff;
1387 	uint32_t logsize;
1388 	uint8_t logpage;
1389 
1390 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1391 
1392 	/*
1393 	 * Command specifies the number of dwords to return in fields NUMDU
1394 	 * and NUMDL. This is a zero-based value.
1395 	 */
1396 	logpage = command->cdw10 & 0xFF;
1397 	logsize = ((command->cdw11 << 16) | (command->cdw10 >> 16)) + 1;
1398 	logsize *= sizeof(uint32_t);
1399 	logoff  = ((uint64_t)(command->cdw13) << 32) | command->cdw12;
1400 
1401 	DPRINTF("%s log page %u len %u", __func__, logpage, logsize);
1402 
1403 	switch (logpage) {
1404 	case NVME_LOG_ERROR:
1405 		if (logoff >= sizeof(sc->err_log)) {
1406 			pci_nvme_status_genc(&compl->status,
1407 			    NVME_SC_INVALID_FIELD);
1408 			break;
1409 		}
1410 
1411 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1412 		    command->prp2, (uint8_t *)&sc->err_log + logoff,
1413 		    MIN(logsize - logoff, sizeof(sc->err_log)),
1414 		    NVME_COPY_TO_PRP);
1415 		break;
1416 	case NVME_LOG_HEALTH_INFORMATION:
1417 		if (logoff >= sizeof(sc->health_log)) {
1418 			pci_nvme_status_genc(&compl->status,
1419 			    NVME_SC_INVALID_FIELD);
1420 			break;
1421 		}
1422 
1423 		pthread_mutex_lock(&sc->mtx);
1424 		memcpy(&sc->health_log.data_units_read, &sc->read_data_units,
1425 		    sizeof(sc->health_log.data_units_read));
1426 		memcpy(&sc->health_log.data_units_written, &sc->write_data_units,
1427 		    sizeof(sc->health_log.data_units_written));
1428 		memcpy(&sc->health_log.host_read_commands, &sc->read_commands,
1429 		    sizeof(sc->health_log.host_read_commands));
1430 		memcpy(&sc->health_log.host_write_commands, &sc->write_commands,
1431 		    sizeof(sc->health_log.host_write_commands));
1432 		pthread_mutex_unlock(&sc->mtx);
1433 
1434 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1435 		    command->prp2, (uint8_t *)&sc->health_log + logoff,
1436 		    MIN(logsize - logoff, sizeof(sc->health_log)),
1437 		    NVME_COPY_TO_PRP);
1438 		break;
1439 	case NVME_LOG_FIRMWARE_SLOT:
1440 		if (logoff >= sizeof(sc->fw_log)) {
1441 			pci_nvme_status_genc(&compl->status,
1442 			    NVME_SC_INVALID_FIELD);
1443 			break;
1444 		}
1445 
1446 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1447 		    command->prp2, (uint8_t *)&sc->fw_log + logoff,
1448 		    MIN(logsize - logoff, sizeof(sc->fw_log)),
1449 		    NVME_COPY_TO_PRP);
1450 		break;
1451 	case NVME_LOG_CHANGED_NAMESPACE:
1452 		if (logoff >= sizeof(sc->ns_log)) {
1453 			pci_nvme_status_genc(&compl->status,
1454 			    NVME_SC_INVALID_FIELD);
1455 			break;
1456 		}
1457 
1458 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1459 		    command->prp2, (uint8_t *)&sc->ns_log + logoff,
1460 		    MIN(logsize - logoff, sizeof(sc->ns_log)),
1461 		    NVME_COPY_TO_PRP);
1462 		memset(&sc->ns_log, 0, sizeof(sc->ns_log));
1463 		break;
1464 	default:
1465 		DPRINTF("%s get log page %x command not supported",
1466 		        __func__, logpage);
1467 
1468 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1469 		    NVME_SC_INVALID_LOG_PAGE);
1470 	}
1471 
1472 	return (1);
1473 }
1474 
1475 static int
1476 nvme_opc_identify(struct pci_nvme_softc* sc, struct nvme_command* command,
1477 	struct nvme_completion* compl)
1478 {
1479 	void *dest;
1480 	uint16_t status;
1481 
1482 	DPRINTF("%s identify 0x%x nsid 0x%x", __func__,
1483 	        command->cdw10 & 0xFF, command->nsid);
1484 
1485 	status = 0;
1486 	pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
1487 
1488 	switch (command->cdw10 & 0xFF) {
1489 	case 0x00: /* return Identify Namespace data structure */
1490 		/* Global NS only valid with NS Management */
1491 		if (command->nsid == NVME_GLOBAL_NAMESPACE_TAG) {
1492 			pci_nvme_status_genc(&status,
1493 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
1494 			break;
1495 		}
1496 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1497 		    command->prp2, (uint8_t *)&sc->nsdata, sizeof(sc->nsdata),
1498 		    NVME_COPY_TO_PRP);
1499 		break;
1500 	case 0x01: /* return Identify Controller data structure */
1501 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1502 		    command->prp2, (uint8_t *)&sc->ctrldata,
1503 		    sizeof(sc->ctrldata),
1504 		    NVME_COPY_TO_PRP);
1505 		break;
1506 	case 0x02: /* list of 1024 active NSIDs > CDW1.NSID */
1507 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1508 		                  sizeof(uint32_t) * 1024);
1509 		/* All unused entries shall be zero */
1510 		memset(dest, 0, sizeof(uint32_t) * 1024);
1511 		((uint32_t *)dest)[0] = 1;
1512 		break;
1513 	case 0x03: /* list of NSID structures in CDW1.NSID, 4096 bytes */
1514 		if (command->nsid != 1) {
1515 			pci_nvme_status_genc(&status,
1516 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
1517 			break;
1518 		}
1519 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1520 		                  sizeof(uint32_t) * 1024);
1521 		/* All bytes after the descriptor shall be zero */
1522 		memset(dest, 0, sizeof(uint32_t) * 1024);
1523 
1524 		/* Return NIDT=1 (i.e. EUI64) descriptor */
1525 		((uint8_t *)dest)[0] = 1;
1526 		((uint8_t *)dest)[1] = sizeof(uint64_t);
1527 		memcpy(((uint8_t *)dest) + 4, sc->nsdata.eui64, sizeof(uint64_t));
1528 		break;
1529 	case 0x13:
1530 		/*
1531 		 * Controller list is optional but used by UNH tests. Return
1532 		 * a valid but empty list.
1533 		 */
1534 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1535 		                  sizeof(uint16_t) * 2048);
1536 		memset(dest, 0, sizeof(uint16_t) * 2048);
1537 		break;
1538 	default:
1539 		DPRINTF("%s unsupported identify command requested 0x%x",
1540 		         __func__, command->cdw10 & 0xFF);
1541 		pci_nvme_status_genc(&status, NVME_SC_INVALID_FIELD);
1542 		break;
1543 	}
1544 
1545 	compl->status = status;
1546 	return (1);
1547 }
1548 
1549 static const char *
1550 nvme_fid_to_name(uint8_t fid)
1551 {
1552 	const char *name;
1553 
1554 	switch (fid) {
1555 	case NVME_FEAT_ARBITRATION:
1556 		name = "Arbitration";
1557 		break;
1558 	case NVME_FEAT_POWER_MANAGEMENT:
1559 		name = "Power Management";
1560 		break;
1561 	case NVME_FEAT_LBA_RANGE_TYPE:
1562 		name = "LBA Range Type";
1563 		break;
1564 	case NVME_FEAT_TEMPERATURE_THRESHOLD:
1565 		name = "Temperature Threshold";
1566 		break;
1567 	case NVME_FEAT_ERROR_RECOVERY:
1568 		name = "Error Recovery";
1569 		break;
1570 	case NVME_FEAT_VOLATILE_WRITE_CACHE:
1571 		name = "Volatile Write Cache";
1572 		break;
1573 	case NVME_FEAT_NUMBER_OF_QUEUES:
1574 		name = "Number of Queues";
1575 		break;
1576 	case NVME_FEAT_INTERRUPT_COALESCING:
1577 		name = "Interrupt Coalescing";
1578 		break;
1579 	case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
1580 		name = "Interrupt Vector Configuration";
1581 		break;
1582 	case NVME_FEAT_WRITE_ATOMICITY:
1583 		name = "Write Atomicity Normal";
1584 		break;
1585 	case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
1586 		name = "Asynchronous Event Configuration";
1587 		break;
1588 	case NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION:
1589 		name = "Autonomous Power State Transition";
1590 		break;
1591 	case NVME_FEAT_HOST_MEMORY_BUFFER:
1592 		name = "Host Memory Buffer";
1593 		break;
1594 	case NVME_FEAT_TIMESTAMP:
1595 		name = "Timestamp";
1596 		break;
1597 	case NVME_FEAT_KEEP_ALIVE_TIMER:
1598 		name = "Keep Alive Timer";
1599 		break;
1600 	case NVME_FEAT_HOST_CONTROLLED_THERMAL_MGMT:
1601 		name = "Host Controlled Thermal Management";
1602 		break;
1603 	case NVME_FEAT_NON_OP_POWER_STATE_CONFIG:
1604 		name = "Non-Operation Power State Config";
1605 		break;
1606 	case NVME_FEAT_READ_RECOVERY_LEVEL_CONFIG:
1607 		name = "Read Recovery Level Config";
1608 		break;
1609 	case NVME_FEAT_PREDICTABLE_LATENCY_MODE_CONFIG:
1610 		name = "Predictable Latency Mode Config";
1611 		break;
1612 	case NVME_FEAT_PREDICTABLE_LATENCY_MODE_WINDOW:
1613 		name = "Predictable Latency Mode Window";
1614 		break;
1615 	case NVME_FEAT_LBA_STATUS_INFORMATION_ATTRIBUTES:
1616 		name = "LBA Status Information Report Interval";
1617 		break;
1618 	case NVME_FEAT_HOST_BEHAVIOR_SUPPORT:
1619 		name = "Host Behavior Support";
1620 		break;
1621 	case NVME_FEAT_SANITIZE_CONFIG:
1622 		name = "Sanitize Config";
1623 		break;
1624 	case NVME_FEAT_ENDURANCE_GROUP_EVENT_CONFIGURATION:
1625 		name = "Endurance Group Event Configuration";
1626 		break;
1627 	case NVME_FEAT_SOFTWARE_PROGRESS_MARKER:
1628 		name = "Software Progress Marker";
1629 		break;
1630 	case NVME_FEAT_HOST_IDENTIFIER:
1631 		name = "Host Identifier";
1632 		break;
1633 	case NVME_FEAT_RESERVATION_NOTIFICATION_MASK:
1634 		name = "Reservation Notification Mask";
1635 		break;
1636 	case NVME_FEAT_RESERVATION_PERSISTENCE:
1637 		name = "Reservation Persistence";
1638 		break;
1639 	case NVME_FEAT_NAMESPACE_WRITE_PROTECTION_CONFIG:
1640 		name = "Namespace Write Protection Config";
1641 		break;
1642 	default:
1643 		name = "Unknown";
1644 		break;
1645 	}
1646 
1647 	return (name);
1648 }
1649 
1650 static void
1651 nvme_feature_invalid_cb(struct pci_nvme_softc *sc __unused,
1652     struct nvme_feature_obj *feat __unused,
1653     struct nvme_command *command __unused,
1654     struct nvme_completion *compl)
1655 {
1656 	pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1657 }
1658 
1659 static void
1660 nvme_feature_iv_config(struct pci_nvme_softc *sc,
1661     struct nvme_feature_obj *feat __unused,
1662     struct nvme_command *command,
1663     struct nvme_completion *compl)
1664 {
1665 	uint32_t i;
1666 	uint32_t cdw11 = command->cdw11;
1667 	uint16_t iv;
1668 	bool cd;
1669 
1670 	pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1671 
1672 	iv = cdw11 & 0xffff;
1673 	cd = cdw11 & (1 << 16);
1674 
1675 	if (iv > (sc->max_queues + 1)) {
1676 		return;
1677 	}
1678 
1679 	/* No Interrupt Coalescing (i.e. not Coalescing Disable) for Admin Q */
1680 	if ((iv == 0) && !cd)
1681 		return;
1682 
1683 	/* Requested Interrupt Vector must be used by a CQ */
1684 	for (i = 0; i < sc->num_cqueues + 1; i++) {
1685 		if (sc->compl_queues[i].intr_vec == iv) {
1686 			pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1687 		}
1688 	}
1689 }
1690 
1691 #define NVME_ASYNC_EVENT_ENDURANCE_GROUP		(0x4000)
1692 static void
1693 nvme_feature_async_event(struct pci_nvme_softc *sc __unused,
1694     struct nvme_feature_obj *feat __unused,
1695     struct nvme_command *command,
1696     struct nvme_completion *compl)
1697 {
1698 	if (command->cdw11 & NVME_ASYNC_EVENT_ENDURANCE_GROUP)
1699 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1700 }
1701 
1702 #define NVME_TEMP_THRESH_OVER	0
1703 #define NVME_TEMP_THRESH_UNDER	1
1704 static void
1705 nvme_feature_temperature(struct pci_nvme_softc *sc,
1706     struct nvme_feature_obj *feat __unused,
1707     struct nvme_command *command,
1708     struct nvme_completion *compl)
1709 {
1710 	uint16_t	tmpth;	/* Temperature Threshold */
1711 	uint8_t		tmpsel; /* Threshold Temperature Select */
1712 	uint8_t		thsel;  /* Threshold Type Select */
1713 	bool		set_crit = false;
1714 	bool		report_crit;
1715 
1716 	tmpth  = command->cdw11 & 0xffff;
1717 	tmpsel = (command->cdw11 >> 16) & 0xf;
1718 	thsel  = (command->cdw11 >> 20) & 0x3;
1719 
1720 	DPRINTF("%s: tmpth=%#x tmpsel=%#x thsel=%#x", __func__, tmpth, tmpsel, thsel);
1721 
1722 	/* Check for unsupported values */
1723 	if (((tmpsel != 0) && (tmpsel != 0xf)) ||
1724 	    (thsel > NVME_TEMP_THRESH_UNDER)) {
1725 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1726 		return;
1727 	}
1728 
1729 	if (((thsel == NVME_TEMP_THRESH_OVER)  && (NVME_TEMPERATURE >= tmpth)) ||
1730 	    ((thsel == NVME_TEMP_THRESH_UNDER) && (NVME_TEMPERATURE <= tmpth)))
1731 		set_crit = true;
1732 
1733 	pthread_mutex_lock(&sc->mtx);
1734 	if (set_crit)
1735 		sc->health_log.critical_warning |=
1736 		    NVME_CRIT_WARN_ST_TEMPERATURE;
1737 	else
1738 		sc->health_log.critical_warning &=
1739 		    ~NVME_CRIT_WARN_ST_TEMPERATURE;
1740 	pthread_mutex_unlock(&sc->mtx);
1741 
1742 	report_crit = sc->feat[NVME_FEAT_ASYNC_EVENT_CONFIGURATION].cdw11 &
1743 	    NVME_CRIT_WARN_ST_TEMPERATURE;
1744 
1745 	if (set_crit && report_crit)
1746 		pci_nvme_aen_post(sc, PCI_NVME_AE_TYPE_SMART,
1747 		    sc->health_log.critical_warning);
1748 
1749 	DPRINTF("%s: set_crit=%c critical_warning=%#x status=%#x", __func__, set_crit ? 'T':'F', sc->health_log.critical_warning, compl->status);
1750 }
1751 
1752 static void
1753 nvme_feature_num_queues(struct pci_nvme_softc *sc,
1754     struct nvme_feature_obj *feat __unused,
1755     struct nvme_command *command,
1756     struct nvme_completion *compl)
1757 {
1758 	uint16_t nqr;	/* Number of Queues Requested */
1759 
1760 	if (sc->num_q_is_set) {
1761 		WPRINTF("%s: Number of Queues already set", __func__);
1762 		pci_nvme_status_genc(&compl->status,
1763 		    NVME_SC_COMMAND_SEQUENCE_ERROR);
1764 		return;
1765 	}
1766 
1767 	nqr = command->cdw11 & 0xFFFF;
1768 	if (nqr == 0xffff) {
1769 		WPRINTF("%s: Illegal NSQR value %#x", __func__, nqr);
1770 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1771 		return;
1772 	}
1773 
1774 	sc->num_squeues = ONE_BASED(nqr);
1775 	if (sc->num_squeues > sc->max_queues) {
1776 		DPRINTF("NSQR=%u is greater than max %u", sc->num_squeues,
1777 					sc->max_queues);
1778 		sc->num_squeues = sc->max_queues;
1779 	}
1780 
1781 	nqr = (command->cdw11 >> 16) & 0xFFFF;
1782 	if (nqr == 0xffff) {
1783 		WPRINTF("%s: Illegal NCQR value %#x", __func__, nqr);
1784 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1785 		return;
1786 	}
1787 
1788 	sc->num_cqueues = ONE_BASED(nqr);
1789 	if (sc->num_cqueues > sc->max_queues) {
1790 		DPRINTF("NCQR=%u is greater than max %u", sc->num_cqueues,
1791 					sc->max_queues);
1792 		sc->num_cqueues = sc->max_queues;
1793 	}
1794 
1795 	/* Patch the command value which will be saved on callback's return */
1796 	command->cdw11 = NVME_FEATURE_NUM_QUEUES(sc);
1797 	compl->cdw0 = NVME_FEATURE_NUM_QUEUES(sc);
1798 
1799 	sc->num_q_is_set = true;
1800 }
1801 
1802 static int
1803 nvme_opc_set_features(struct pci_nvme_softc *sc, struct nvme_command *command,
1804 	struct nvme_completion *compl)
1805 {
1806 	struct nvme_feature_obj *feat;
1807 	uint32_t nsid = command->nsid;
1808 	uint8_t fid = NVMEV(NVME_FEAT_SET_FID, command->cdw10);
1809 	bool sv = NVMEV(NVME_FEAT_SET_SV, command->cdw10);
1810 
1811 	DPRINTF("%s: Feature ID 0x%x (%s)", __func__, fid, nvme_fid_to_name(fid));
1812 
1813 	if (fid >= NVME_FID_MAX) {
1814 		DPRINTF("%s invalid feature 0x%x", __func__, fid);
1815 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1816 		return (1);
1817 	}
1818 
1819 	if (sv) {
1820 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1821 		    NVME_SC_FEATURE_NOT_SAVEABLE);
1822 		return (1);
1823 	}
1824 
1825 	feat = &sc->feat[fid];
1826 
1827 	if (feat->namespace_specific && (nsid == NVME_GLOBAL_NAMESPACE_TAG)) {
1828 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1829 		return (1);
1830 	}
1831 
1832 	if (!feat->namespace_specific &&
1833 	    !((nsid == 0) || (nsid == NVME_GLOBAL_NAMESPACE_TAG))) {
1834 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1835 		    NVME_SC_FEATURE_NOT_NS_SPECIFIC);
1836 		return (1);
1837 	}
1838 
1839 	compl->cdw0 = 0;
1840 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1841 
1842 	if (feat->set)
1843 		feat->set(sc, feat, command, compl);
1844 	else {
1845 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1846 		    NVME_SC_FEATURE_NOT_CHANGEABLE);
1847 		return (1);
1848 	}
1849 
1850 	DPRINTF("%s: status=%#x cdw11=%#x", __func__, compl->status, command->cdw11);
1851 	if (compl->status == NVME_SC_SUCCESS) {
1852 		feat->cdw11 = command->cdw11;
1853 		if ((fid == NVME_FEAT_ASYNC_EVENT_CONFIGURATION) &&
1854 		    (command->cdw11 != 0))
1855 			pci_nvme_aen_notify(sc);
1856 	}
1857 
1858 	return (0);
1859 }
1860 
1861 #define NVME_FEATURES_SEL_SUPPORTED	0x3
1862 #define NVME_FEATURES_NS_SPECIFIC	(1 << 1)
1863 
1864 static int
1865 nvme_opc_get_features(struct pci_nvme_softc* sc, struct nvme_command* command,
1866 	struct nvme_completion* compl)
1867 {
1868 	struct nvme_feature_obj *feat;
1869 	uint8_t fid = command->cdw10 & 0xFF;
1870 	uint8_t sel = (command->cdw10 >> 8) & 0x7;
1871 
1872 	DPRINTF("%s: Feature ID 0x%x (%s)", __func__, fid, nvme_fid_to_name(fid));
1873 
1874 	if (fid >= NVME_FID_MAX) {
1875 		DPRINTF("%s invalid feature 0x%x", __func__, fid);
1876 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1877 		return (1);
1878 	}
1879 
1880 	compl->cdw0 = 0;
1881 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1882 
1883 	feat = &sc->feat[fid];
1884 	if (feat->get) {
1885 		feat->get(sc, feat, command, compl);
1886 	}
1887 
1888 	if (compl->status == NVME_SC_SUCCESS) {
1889 		if ((sel == NVME_FEATURES_SEL_SUPPORTED) && feat->namespace_specific)
1890 			compl->cdw0 = NVME_FEATURES_NS_SPECIFIC;
1891 		else
1892 			compl->cdw0 = feat->cdw11;
1893 	}
1894 
1895 	return (0);
1896 }
1897 
1898 static int
1899 nvme_opc_format_nvm(struct pci_nvme_softc* sc, struct nvme_command* command,
1900 	struct nvme_completion* compl)
1901 {
1902 	uint8_t	ses, lbaf, pi;
1903 
1904 	/* Only supports Secure Erase Setting - User Data Erase */
1905 	ses = (command->cdw10 >> 9) & 0x7;
1906 	if (ses > 0x1) {
1907 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1908 		return (1);
1909 	}
1910 
1911 	/* Only supports a single LBA Format */
1912 	lbaf = command->cdw10 & 0xf;
1913 	if (lbaf != 0) {
1914 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1915 		    NVME_SC_INVALID_FORMAT);
1916 		return (1);
1917 	}
1918 
1919 	/* Doesn't support Protection Information */
1920 	pi = (command->cdw10 >> 5) & 0x7;
1921 	if (pi != 0) {
1922 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1923 		return (1);
1924 	}
1925 
1926 	if (sc->nvstore.type == NVME_STOR_RAM) {
1927 		if (sc->nvstore.ctx)
1928 			free(sc->nvstore.ctx);
1929 		sc->nvstore.ctx = calloc(1, sc->nvstore.size);
1930 		pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1931 	} else {
1932 		struct pci_nvme_ioreq *req;
1933 		int err;
1934 
1935 		req = pci_nvme_get_ioreq(sc);
1936 		if (req == NULL) {
1937 			pci_nvme_status_genc(&compl->status,
1938 			    NVME_SC_INTERNAL_DEVICE_ERROR);
1939 			WPRINTF("%s: unable to allocate IO req", __func__);
1940 			return (1);
1941 		}
1942 		req->nvme_sq = &sc->submit_queues[0];
1943 		req->sqid = 0;
1944 		req->opc = command->opc;
1945 		req->cid = command->cid;
1946 		req->nsid = command->nsid;
1947 
1948 		req->io_req.br_offset = 0;
1949 		req->io_req.br_resid = sc->nvstore.size;
1950 		req->io_req.br_callback = pci_nvme_io_done;
1951 
1952 		err = blockif_delete(sc->nvstore.ctx, &req->io_req);
1953 		if (err) {
1954 			pci_nvme_status_genc(&compl->status,
1955 			    NVME_SC_INTERNAL_DEVICE_ERROR);
1956 			pci_nvme_release_ioreq(sc, req);
1957 		} else
1958 			compl->status = NVME_NO_STATUS;
1959 	}
1960 
1961 	return (1);
1962 }
1963 
1964 static int
1965 nvme_opc_abort(struct pci_nvme_softc *sc __unused, struct nvme_command *command,
1966     struct nvme_completion *compl)
1967 {
1968 	DPRINTF("%s submission queue %u, command ID 0x%x", __func__,
1969 	        command->cdw10 & 0xFFFF, (command->cdw10 >> 16) & 0xFFFF);
1970 
1971 	/* TODO: search for the command ID and abort it */
1972 
1973 	compl->cdw0 = 1;
1974 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1975 	return (1);
1976 }
1977 
1978 static int
1979 nvme_opc_async_event_req(struct pci_nvme_softc* sc,
1980 	struct nvme_command* command, struct nvme_completion* compl)
1981 {
1982 	DPRINTF("%s async event request count=%u aerl=%u cid=%#x", __func__,
1983 	    sc->aer_count, sc->ctrldata.aerl, command->cid);
1984 
1985 	/* Don't exceed the Async Event Request Limit (AERL). */
1986 	if (pci_nvme_aer_limit_reached(sc)) {
1987 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1988 				NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED);
1989 		return (1);
1990 	}
1991 
1992 	if (pci_nvme_aer_add(sc, command->cid)) {
1993 		pci_nvme_status_tc(&compl->status, NVME_SCT_GENERIC,
1994 				NVME_SC_INTERNAL_DEVICE_ERROR);
1995 		return (1);
1996 	}
1997 
1998 	/*
1999 	 * Raise events when they happen based on the Set Features cmd.
2000 	 * These events happen async, so only set completion successful if
2001 	 * there is an event reflective of the request to get event.
2002 	 */
2003 	compl->status = NVME_NO_STATUS;
2004 	pci_nvme_aen_notify(sc);
2005 
2006 	return (0);
2007 }
2008 
2009 static void
2010 pci_nvme_handle_admin_cmd(struct pci_nvme_softc* sc, uint64_t value)
2011 {
2012 	struct nvme_completion compl;
2013 	struct nvme_command *cmd;
2014 	struct nvme_submission_queue *sq;
2015 	struct nvme_completion_queue *cq;
2016 	uint16_t sqhead;
2017 
2018 	DPRINTF("%s index %u", __func__, (uint32_t)value);
2019 
2020 	sq = &sc->submit_queues[0];
2021 	cq = &sc->compl_queues[0];
2022 
2023 	pthread_mutex_lock(&sq->mtx);
2024 
2025 	sqhead = sq->head;
2026 	DPRINTF("sqhead %u, tail %u", sqhead, sq->tail);
2027 
2028 	while (sqhead != atomic_load_acq_short(&sq->tail)) {
2029 		cmd = &(sq->qbase)[sqhead];
2030 		compl.cdw0 = 0;
2031 		compl.status = 0;
2032 
2033 		switch (cmd->opc) {
2034 		case NVME_OPC_DELETE_IO_SQ:
2035 			DPRINTF("%s command DELETE_IO_SQ", __func__);
2036 			nvme_opc_delete_io_sq(sc, cmd, &compl);
2037 			break;
2038 		case NVME_OPC_CREATE_IO_SQ:
2039 			DPRINTF("%s command CREATE_IO_SQ", __func__);
2040 			nvme_opc_create_io_sq(sc, cmd, &compl);
2041 			break;
2042 		case NVME_OPC_DELETE_IO_CQ:
2043 			DPRINTF("%s command DELETE_IO_CQ", __func__);
2044 			nvme_opc_delete_io_cq(sc, cmd, &compl);
2045 			break;
2046 		case NVME_OPC_CREATE_IO_CQ:
2047 			DPRINTF("%s command CREATE_IO_CQ", __func__);
2048 			nvme_opc_create_io_cq(sc, cmd, &compl);
2049 			break;
2050 		case NVME_OPC_GET_LOG_PAGE:
2051 			DPRINTF("%s command GET_LOG_PAGE", __func__);
2052 			nvme_opc_get_log_page(sc, cmd, &compl);
2053 			break;
2054 		case NVME_OPC_IDENTIFY:
2055 			DPRINTF("%s command IDENTIFY", __func__);
2056 			nvme_opc_identify(sc, cmd, &compl);
2057 			break;
2058 		case NVME_OPC_ABORT:
2059 			DPRINTF("%s command ABORT", __func__);
2060 			nvme_opc_abort(sc, cmd, &compl);
2061 			break;
2062 		case NVME_OPC_SET_FEATURES:
2063 			DPRINTF("%s command SET_FEATURES", __func__);
2064 			nvme_opc_set_features(sc, cmd, &compl);
2065 			break;
2066 		case NVME_OPC_GET_FEATURES:
2067 			DPRINTF("%s command GET_FEATURES", __func__);
2068 			nvme_opc_get_features(sc, cmd, &compl);
2069 			break;
2070 		case NVME_OPC_FIRMWARE_ACTIVATE:
2071 			DPRINTF("%s command FIRMWARE_ACTIVATE", __func__);
2072 			pci_nvme_status_tc(&compl.status,
2073 			    NVME_SCT_COMMAND_SPECIFIC,
2074 			    NVME_SC_INVALID_FIRMWARE_SLOT);
2075 			break;
2076 		case NVME_OPC_ASYNC_EVENT_REQUEST:
2077 			DPRINTF("%s command ASYNC_EVENT_REQ", __func__);
2078 			nvme_opc_async_event_req(sc, cmd, &compl);
2079 			break;
2080 		case NVME_OPC_FORMAT_NVM:
2081 			DPRINTF("%s command FORMAT_NVM", __func__);
2082 			if (NVMEV(NVME_CTRLR_DATA_OACS_FORMAT,
2083 			    sc->ctrldata.oacs) == 0) {
2084 				pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_OPCODE);
2085 				break;
2086 			}
2087 			nvme_opc_format_nvm(sc, cmd, &compl);
2088 			break;
2089 		case NVME_OPC_SECURITY_SEND:
2090 		case NVME_OPC_SECURITY_RECEIVE:
2091 		case NVME_OPC_SANITIZE:
2092 		case NVME_OPC_GET_LBA_STATUS:
2093 			DPRINTF("%s command OPC=%#x (unsupported)", __func__,
2094 			    cmd->opc);
2095 			/* Valid but unsupported opcodes */
2096 			pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_FIELD);
2097 			break;
2098 		default:
2099 			DPRINTF("%s command OPC=%#X (not implemented)",
2100 			    __func__,
2101 			    cmd->opc);
2102 			pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_OPCODE);
2103 		}
2104 		sqhead = (sqhead + 1) % sq->size;
2105 
2106 		if (NVME_COMPLETION_VALID(compl)) {
2107 			pci_nvme_cq_update(sc, &sc->compl_queues[0],
2108 			    compl.cdw0,
2109 			    cmd->cid,
2110 			    0,		/* SQID */
2111 			    compl.status);
2112 		}
2113 	}
2114 
2115 	DPRINTF("setting sqhead %u", sqhead);
2116 	sq->head = sqhead;
2117 
2118 	if (cq->head != cq->tail)
2119 		pci_generate_msix(sc->nsc_pi, 0);
2120 
2121 	pthread_mutex_unlock(&sq->mtx);
2122 }
2123 
2124 /*
2125  * Update the Write and Read statistics reported in SMART data
2126  *
2127  * NVMe defines "data unit" as thousand's of 512 byte blocks and is rounded up.
2128  * E.g. 1 data unit is 1 - 1,000 512 byte blocks. 3 data units are 2,001 - 3,000
2129  * 512 byte blocks. Rounding up is achieved by initializing the remainder to 999.
2130  */
2131 static void
2132 pci_nvme_stats_write_read_update(struct pci_nvme_softc *sc, uint8_t opc,
2133     size_t bytes, uint16_t status)
2134 {
2135 
2136 	pthread_mutex_lock(&sc->mtx);
2137 	switch (opc) {
2138 	case NVME_OPC_WRITE:
2139 		sc->write_commands++;
2140 		if (status != NVME_SC_SUCCESS)
2141 			break;
2142 		sc->write_dunits_remainder += (bytes / 512);
2143 		while (sc->write_dunits_remainder >= 1000) {
2144 			sc->write_data_units++;
2145 			sc->write_dunits_remainder -= 1000;
2146 		}
2147 		break;
2148 	case NVME_OPC_READ:
2149 		sc->read_commands++;
2150 		if (status != NVME_SC_SUCCESS)
2151 			break;
2152 		sc->read_dunits_remainder += (bytes / 512);
2153 		while (sc->read_dunits_remainder >= 1000) {
2154 			sc->read_data_units++;
2155 			sc->read_dunits_remainder -= 1000;
2156 		}
2157 		break;
2158 	default:
2159 		DPRINTF("%s: Invalid OPC 0x%02x for stats", __func__, opc);
2160 		break;
2161 	}
2162 	pthread_mutex_unlock(&sc->mtx);
2163 }
2164 
2165 /*
2166  * Check if the combination of Starting LBA (slba) and number of blocks
2167  * exceeds the range of the underlying storage.
2168  *
2169  * Because NVMe specifies the SLBA in blocks as a uint64_t and blockif stores
2170  * the capacity in bytes as a uint64_t, care must be taken to avoid integer
2171  * overflow.
2172  */
2173 static bool
2174 pci_nvme_out_of_range(struct pci_nvme_blockstore *nvstore, uint64_t slba,
2175     uint32_t nblocks)
2176 {
2177 	size_t	offset, bytes;
2178 
2179 	/* Overflow check of multiplying Starting LBA by the sector size */
2180 	if (slba >> (64 - nvstore->sectsz_bits))
2181 		return (true);
2182 
2183 	offset = slba << nvstore->sectsz_bits;
2184 	bytes = nblocks << nvstore->sectsz_bits;
2185 
2186 	/* Overflow check of Number of Logical Blocks */
2187 	if ((nvstore->size <= offset) || ((nvstore->size - offset) < bytes))
2188 		return (true);
2189 
2190 	return (false);
2191 }
2192 
2193 static int
2194 pci_nvme_append_iov_req(struct pci_nvme_softc *sc __unused,
2195     struct pci_nvme_ioreq *req, uint64_t gpaddr, size_t size, uint64_t offset)
2196 {
2197 	int iovidx;
2198 	bool range_is_contiguous;
2199 
2200 	if (req == NULL)
2201 		return (-1);
2202 
2203 	if (req->io_req.br_iovcnt == NVME_MAX_IOVEC) {
2204 		return (-1);
2205 	}
2206 
2207 	/*
2208 	 * Minimize the number of IOVs by concatenating contiguous address
2209 	 * ranges. If the IOV count is zero, there is no previous range to
2210 	 * concatenate.
2211 	 */
2212 	if (req->io_req.br_iovcnt == 0)
2213 		range_is_contiguous = false;
2214 	else
2215 		range_is_contiguous = (req->prev_gpaddr + req->prev_size) == gpaddr;
2216 
2217 	if (range_is_contiguous) {
2218 		iovidx = req->io_req.br_iovcnt - 1;
2219 
2220 		req->io_req.br_iov[iovidx].iov_base =
2221 		    paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
2222 				     req->prev_gpaddr, size);
2223 		if (req->io_req.br_iov[iovidx].iov_base == NULL)
2224 			return (-1);
2225 
2226 		req->prev_size += size;
2227 		req->io_req.br_resid += size;
2228 
2229 		req->io_req.br_iov[iovidx].iov_len = req->prev_size;
2230 	} else {
2231 		iovidx = req->io_req.br_iovcnt;
2232 		if (iovidx == 0) {
2233 			req->io_req.br_offset = offset;
2234 			req->io_req.br_resid = 0;
2235 			req->io_req.br_param = req;
2236 		}
2237 
2238 		req->io_req.br_iov[iovidx].iov_base =
2239 		    paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
2240 				     gpaddr, size);
2241 		if (req->io_req.br_iov[iovidx].iov_base == NULL)
2242 			return (-1);
2243 
2244 		req->io_req.br_iov[iovidx].iov_len = size;
2245 
2246 		req->prev_gpaddr = gpaddr;
2247 		req->prev_size = size;
2248 		req->io_req.br_resid += size;
2249 
2250 		req->io_req.br_iovcnt++;
2251 	}
2252 
2253 	return (0);
2254 }
2255 
2256 static void
2257 pci_nvme_set_completion(struct pci_nvme_softc *sc,
2258     struct nvme_submission_queue *sq, int sqid, uint16_t cid, uint16_t status)
2259 {
2260 	struct nvme_completion_queue *cq = &sc->compl_queues[sq->cqid];
2261 
2262 	DPRINTF("%s sqid %d cqid %u cid %u status: 0x%x 0x%x",
2263 		 __func__, sqid, sq->cqid, cid, NVME_STATUS_GET_SCT(status),
2264 		 NVME_STATUS_GET_SC(status));
2265 
2266 	pci_nvme_cq_update(sc, cq, 0, cid, sqid, status);
2267 
2268 	if (cq->head != cq->tail) {
2269 		if (cq->intr_en & NVME_CQ_INTEN) {
2270 			pci_generate_msix(sc->nsc_pi, cq->intr_vec);
2271 		} else {
2272 			DPRINTF("%s: CQ%u interrupt disabled",
2273 						__func__, sq->cqid);
2274 		}
2275 	}
2276 }
2277 
2278 static void
2279 pci_nvme_release_ioreq(struct pci_nvme_softc *sc, struct pci_nvme_ioreq *req)
2280 {
2281 	req->sc = NULL;
2282 	req->nvme_sq = NULL;
2283 	req->sqid = 0;
2284 
2285 	pthread_mutex_lock(&sc->mtx);
2286 
2287 	STAILQ_INSERT_TAIL(&sc->ioreqs_free, req, link);
2288 	sc->pending_ios--;
2289 
2290 	/* when no more IO pending, can set to ready if device reset/enabled */
2291 	if (sc->pending_ios == 0 &&
2292 	    NVME_CC_GET_EN(sc->regs.cc) && !(NVME_CSTS_GET_RDY(sc->regs.csts)))
2293 		sc->regs.csts |= NVME_CSTS_RDY;
2294 
2295 	pthread_mutex_unlock(&sc->mtx);
2296 
2297 	sem_post(&sc->iosemlock);
2298 }
2299 
2300 static struct pci_nvme_ioreq *
2301 pci_nvme_get_ioreq(struct pci_nvme_softc *sc)
2302 {
2303 	struct pci_nvme_ioreq *req = NULL;
2304 
2305 	sem_wait(&sc->iosemlock);
2306 	pthread_mutex_lock(&sc->mtx);
2307 
2308 	req = STAILQ_FIRST(&sc->ioreqs_free);
2309 	assert(req != NULL);
2310 	STAILQ_REMOVE_HEAD(&sc->ioreqs_free, link);
2311 
2312 	req->sc = sc;
2313 
2314 	sc->pending_ios++;
2315 
2316 	pthread_mutex_unlock(&sc->mtx);
2317 
2318 	req->io_req.br_iovcnt = 0;
2319 	req->io_req.br_offset = 0;
2320 	req->io_req.br_resid = 0;
2321 	req->io_req.br_param = req;
2322 	req->prev_gpaddr = 0;
2323 	req->prev_size = 0;
2324 
2325 	return req;
2326 }
2327 
2328 static void
2329 pci_nvme_io_done(struct blockif_req *br, int err)
2330 {
2331 	struct pci_nvme_ioreq *req = br->br_param;
2332 	struct nvme_submission_queue *sq = req->nvme_sq;
2333 	uint16_t code, status;
2334 
2335 	DPRINTF("%s error %d %s", __func__, err, strerror(err));
2336 
2337 	/* TODO return correct error */
2338 	code = err ? NVME_SC_DATA_TRANSFER_ERROR : NVME_SC_SUCCESS;
2339 	status = 0;
2340 	pci_nvme_status_genc(&status, code);
2341 
2342 	pci_nvme_set_completion(req->sc, sq, req->sqid, req->cid, status);
2343 	pci_nvme_stats_write_read_update(req->sc, req->opc,
2344 	    req->bytes, status);
2345 	pci_nvme_release_ioreq(req->sc, req);
2346 }
2347 
2348 /*
2349  * Implements the Flush command. The specification states:
2350  *    If a volatile write cache is not present, Flush commands complete
2351  *    successfully and have no effect
2352  * in the description of the Volatile Write Cache (VWC) field of the Identify
2353  * Controller data. Therefore, set status to Success if the command is
2354  * not supported (i.e. RAM or as indicated by the blockif).
2355  */
2356 static bool
2357 nvme_opc_flush(struct pci_nvme_softc *sc __unused,
2358     struct nvme_command *cmd __unused,
2359     struct pci_nvme_blockstore *nvstore,
2360     struct pci_nvme_ioreq *req,
2361     uint16_t *status)
2362 {
2363 	bool pending = false;
2364 
2365 	if (nvstore->type == NVME_STOR_RAM) {
2366 		pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2367 	} else {
2368 		int err;
2369 
2370 		req->io_req.br_callback = pci_nvme_io_done;
2371 
2372 		err = blockif_flush(nvstore->ctx, &req->io_req);
2373 		switch (err) {
2374 		case 0:
2375 			pending = true;
2376 			break;
2377 		case EOPNOTSUPP:
2378 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2379 			break;
2380 		default:
2381 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2382 		}
2383 	}
2384 
2385 	return (pending);
2386 }
2387 
2388 static uint16_t
2389 nvme_write_read_ram(struct pci_nvme_softc *sc,
2390     struct pci_nvme_blockstore *nvstore,
2391     uint64_t prp1, uint64_t prp2,
2392     size_t offset, uint64_t bytes,
2393     bool is_write)
2394 {
2395 	uint8_t *buf = nvstore->ctx;
2396 	enum nvme_copy_dir dir;
2397 	uint16_t status;
2398 
2399 	if (is_write)
2400 		dir = NVME_COPY_TO_PRP;
2401 	else
2402 		dir = NVME_COPY_FROM_PRP;
2403 
2404 	status = 0;
2405 	if (nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, prp1, prp2,
2406 	    buf + offset, bytes, dir))
2407 		pci_nvme_status_genc(&status,
2408 		    NVME_SC_DATA_TRANSFER_ERROR);
2409 	else
2410 		pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2411 
2412 	return (status);
2413 }
2414 
2415 static uint16_t
2416 nvme_write_read_blockif(struct pci_nvme_softc *sc,
2417     struct pci_nvme_blockstore *nvstore,
2418     struct pci_nvme_ioreq *req,
2419     uint64_t prp1, uint64_t prp2,
2420     size_t offset, uint64_t bytes,
2421     bool is_write)
2422 {
2423 	uint64_t size;
2424 	int err;
2425 	uint16_t status = NVME_NO_STATUS;
2426 
2427 	size = MIN(PAGE_SIZE - (prp1 % PAGE_SIZE), bytes);
2428 	if (pci_nvme_append_iov_req(sc, req, prp1, size, offset)) {
2429 		err = -1;
2430 		goto out;
2431 	}
2432 
2433 	offset += size;
2434 	bytes  -= size;
2435 
2436 	if (bytes == 0) {
2437 		;
2438 	} else if (bytes <= PAGE_SIZE) {
2439 		size = bytes;
2440 		if (pci_nvme_append_iov_req(sc, req, prp2, size, offset)) {
2441 			err = -1;
2442 			goto out;
2443 		}
2444 	} else {
2445 		void *vmctx = sc->nsc_pi->pi_vmctx;
2446 		uint64_t *prp_list = &prp2;
2447 		uint64_t *last = prp_list;
2448 
2449 		/* PRP2 is pointer to a physical region page list */
2450 		while (bytes) {
2451 			/* Last entry in list points to the next list */
2452 			if ((prp_list == last) && (bytes > PAGE_SIZE)) {
2453 				uint64_t prp = *prp_list;
2454 
2455 				prp_list = paddr_guest2host(vmctx, prp,
2456 				    PAGE_SIZE - (prp % PAGE_SIZE));
2457 				if (prp_list == NULL) {
2458 					err = -1;
2459 					goto out;
2460 				}
2461 				last = prp_list + (NVME_PRP2_ITEMS - 1);
2462 			}
2463 
2464 			size = MIN(bytes, PAGE_SIZE);
2465 
2466 			if (pci_nvme_append_iov_req(sc, req, *prp_list, size,
2467 			    offset)) {
2468 				err = -1;
2469 				goto out;
2470 			}
2471 
2472 			offset += size;
2473 			bytes  -= size;
2474 
2475 			prp_list++;
2476 		}
2477 	}
2478 	req->io_req.br_callback = pci_nvme_io_done;
2479 	if (is_write)
2480 		err = blockif_write(nvstore->ctx, &req->io_req);
2481 	else
2482 		err = blockif_read(nvstore->ctx, &req->io_req);
2483 out:
2484 	if (err)
2485 		pci_nvme_status_genc(&status, NVME_SC_DATA_TRANSFER_ERROR);
2486 
2487 	return (status);
2488 }
2489 
2490 static bool
2491 nvme_opc_write_read(struct pci_nvme_softc *sc,
2492     struct nvme_command *cmd,
2493     struct pci_nvme_blockstore *nvstore,
2494     struct pci_nvme_ioreq *req,
2495     uint16_t *status)
2496 {
2497 	uint64_t lba, nblocks, bytes;
2498 	size_t offset;
2499 	bool is_write = cmd->opc == NVME_OPC_WRITE;
2500 	bool pending = false;
2501 
2502 	lba = ((uint64_t)cmd->cdw11 << 32) | cmd->cdw10;
2503 	nblocks = (cmd->cdw12 & 0xFFFF) + 1;
2504 	bytes = nblocks << nvstore->sectsz_bits;
2505 	if (bytes > NVME_MAX_DATA_SIZE) {
2506 		WPRINTF("%s command would exceed MDTS", __func__);
2507 		pci_nvme_status_genc(status, NVME_SC_INVALID_FIELD);
2508 		goto out;
2509 	}
2510 
2511 	if (pci_nvme_out_of_range(nvstore, lba, nblocks)) {
2512 		WPRINTF("%s command would exceed LBA range(slba=%#lx nblocks=%#lx)",
2513 		    __func__, lba, nblocks);
2514 		pci_nvme_status_genc(status, NVME_SC_LBA_OUT_OF_RANGE);
2515 		goto out;
2516 	}
2517 
2518 	offset = lba << nvstore->sectsz_bits;
2519 
2520 	req->bytes = bytes;
2521 	req->io_req.br_offset = lba;
2522 
2523 	/* PRP bits 1:0 must be zero */
2524 	cmd->prp1 &= ~0x3UL;
2525 	cmd->prp2 &= ~0x3UL;
2526 
2527 	if (nvstore->type == NVME_STOR_RAM) {
2528 		*status = nvme_write_read_ram(sc, nvstore, cmd->prp1,
2529 		    cmd->prp2, offset, bytes, is_write);
2530 	} else {
2531 		*status = nvme_write_read_blockif(sc, nvstore, req,
2532 		    cmd->prp1, cmd->prp2, offset, bytes, is_write);
2533 
2534 		if (*status == NVME_NO_STATUS)
2535 			pending = true;
2536 	}
2537 out:
2538 	if (!pending)
2539 		pci_nvme_stats_write_read_update(sc, cmd->opc, bytes, *status);
2540 
2541 	return (pending);
2542 }
2543 
2544 static void
2545 pci_nvme_dealloc_sm(struct blockif_req *br, int err)
2546 {
2547 	struct pci_nvme_ioreq *req = br->br_param;
2548 	struct pci_nvme_softc *sc = req->sc;
2549 	bool done = true;
2550 	uint16_t status;
2551 
2552 	status = 0;
2553 	if (err) {
2554 		pci_nvme_status_genc(&status, NVME_SC_INTERNAL_DEVICE_ERROR);
2555 	} else if ((req->prev_gpaddr + 1) == (req->prev_size)) {
2556 		pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2557 	} else {
2558 		struct iovec *iov = req->io_req.br_iov;
2559 
2560 		req->prev_gpaddr++;
2561 		iov += req->prev_gpaddr;
2562 
2563 		/* The iov_* values already include the sector size */
2564 		req->io_req.br_offset = (off_t)iov->iov_base;
2565 		req->io_req.br_resid = iov->iov_len;
2566 		if (blockif_delete(sc->nvstore.ctx, &req->io_req)) {
2567 			pci_nvme_status_genc(&status,
2568 			    NVME_SC_INTERNAL_DEVICE_ERROR);
2569 		} else
2570 			done = false;
2571 	}
2572 
2573 	if (done) {
2574 		pci_nvme_set_completion(sc, req->nvme_sq, req->sqid, req->cid,
2575 		    status);
2576 		pci_nvme_release_ioreq(sc, req);
2577 	}
2578 }
2579 
2580 static bool
2581 nvme_opc_dataset_mgmt(struct pci_nvme_softc *sc,
2582     struct nvme_command *cmd,
2583     struct pci_nvme_blockstore *nvstore,
2584     struct pci_nvme_ioreq *req,
2585     uint16_t *status)
2586 {
2587 	struct nvme_dsm_range *range = NULL;
2588 	uint32_t nr, r, non_zero, dr;
2589 	int err;
2590 	bool pending = false;
2591 
2592 	if ((sc->ctrldata.oncs & NVME_ONCS_DSM) == 0) {
2593 		pci_nvme_status_genc(status, NVME_SC_INVALID_OPCODE);
2594 		goto out;
2595 	}
2596 
2597 	nr = cmd->cdw10 & 0xff;
2598 
2599 	/* copy locally because a range entry could straddle PRPs */
2600 	range = calloc(1, NVME_MAX_DSM_TRIM);
2601 	if (range == NULL) {
2602 		pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2603 		goto out;
2604 	}
2605 	nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, cmd->prp1, cmd->prp2,
2606 	    (uint8_t *)range, NVME_MAX_DSM_TRIM, NVME_COPY_FROM_PRP);
2607 
2608 	/* Check for invalid ranges and the number of non-zero lengths */
2609 	non_zero = 0;
2610 	for (r = 0; r <= nr; r++) {
2611 		if (pci_nvme_out_of_range(nvstore,
2612 		    range[r].starting_lba, range[r].length)) {
2613 			pci_nvme_status_genc(status, NVME_SC_LBA_OUT_OF_RANGE);
2614 			goto out;
2615 		}
2616 		if (range[r].length != 0)
2617 			non_zero++;
2618 	}
2619 
2620 	if (cmd->cdw11 & NVME_DSM_ATTR_DEALLOCATE) {
2621 		size_t offset, bytes;
2622 		int sectsz_bits = sc->nvstore.sectsz_bits;
2623 
2624 		/*
2625 		 * DSM calls are advisory only, and compliant controllers
2626 		 * may choose to take no actions (i.e. return Success).
2627 		 */
2628 		if (!nvstore->deallocate) {
2629 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2630 			goto out;
2631 		}
2632 
2633 		/* If all ranges have a zero length, return Success */
2634 		if (non_zero == 0) {
2635 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2636 			goto out;
2637 		}
2638 
2639 		if (req == NULL) {
2640 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2641 			goto out;
2642 		}
2643 
2644 		offset = range[0].starting_lba << sectsz_bits;
2645 		bytes = range[0].length << sectsz_bits;
2646 
2647 		/*
2648 		 * If the request is for more than a single range, store
2649 		 * the ranges in the br_iov. Optimize for the common case
2650 		 * of a single range.
2651 		 *
2652 		 * Note that NVMe Number of Ranges is a zero based value
2653 		 */
2654 		req->io_req.br_iovcnt = 0;
2655 		req->io_req.br_offset = offset;
2656 		req->io_req.br_resid = bytes;
2657 
2658 		if (nr == 0) {
2659 			req->io_req.br_callback = pci_nvme_io_done;
2660 		} else {
2661 			struct iovec *iov = req->io_req.br_iov;
2662 
2663 			for (r = 0, dr = 0; r <= nr; r++) {
2664 				offset = range[r].starting_lba << sectsz_bits;
2665 				bytes = range[r].length << sectsz_bits;
2666 				if (bytes == 0)
2667 					continue;
2668 
2669 				if ((nvstore->size - offset) < bytes) {
2670 					pci_nvme_status_genc(status,
2671 					    NVME_SC_LBA_OUT_OF_RANGE);
2672 					goto out;
2673 				}
2674 				iov[dr].iov_base = (void *)offset;
2675 				iov[dr].iov_len = bytes;
2676 				dr++;
2677 			}
2678 			req->io_req.br_callback = pci_nvme_dealloc_sm;
2679 
2680 			/*
2681 			 * Use prev_gpaddr to track the current entry and
2682 			 * prev_size to track the number of entries
2683 			 */
2684 			req->prev_gpaddr = 0;
2685 			req->prev_size = dr;
2686 		}
2687 
2688 		err = blockif_delete(nvstore->ctx, &req->io_req);
2689 		if (err)
2690 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2691 		else
2692 			pending = true;
2693 	}
2694 out:
2695 	free(range);
2696 	return (pending);
2697 }
2698 
2699 static void
2700 pci_nvme_handle_io_cmd(struct pci_nvme_softc* sc, uint16_t idx)
2701 {
2702 	struct nvme_submission_queue *sq;
2703 	uint16_t status;
2704 	uint16_t sqhead;
2705 
2706 	/* handle all submissions up to sq->tail index */
2707 	sq = &sc->submit_queues[idx];
2708 
2709 	pthread_mutex_lock(&sq->mtx);
2710 
2711 	sqhead = sq->head;
2712 	DPRINTF("nvme_handle_io qid %u head %u tail %u cmdlist %p",
2713 	         idx, sqhead, sq->tail, sq->qbase);
2714 
2715 	while (sqhead != atomic_load_acq_short(&sq->tail)) {
2716 		struct nvme_command *cmd;
2717 		struct pci_nvme_ioreq *req;
2718 		uint32_t nsid;
2719 		bool pending;
2720 
2721 		pending = false;
2722 		req = NULL;
2723 		status = 0;
2724 
2725 		cmd = &sq->qbase[sqhead];
2726 		sqhead = (sqhead + 1) % sq->size;
2727 
2728 		nsid = le32toh(cmd->nsid);
2729 		if ((nsid == 0) || (nsid > sc->ctrldata.nn)) {
2730 			pci_nvme_status_genc(&status,
2731 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
2732 			status |= NVMEM(NVME_STATUS_DNR);
2733 			goto complete;
2734  		}
2735 
2736 		req = pci_nvme_get_ioreq(sc);
2737 		if (req == NULL) {
2738 			pci_nvme_status_genc(&status,
2739 			    NVME_SC_INTERNAL_DEVICE_ERROR);
2740 			WPRINTF("%s: unable to allocate IO req", __func__);
2741 			goto complete;
2742 		}
2743 		req->nvme_sq = sq;
2744 		req->sqid = idx;
2745 		req->opc = cmd->opc;
2746 		req->cid = cmd->cid;
2747 		req->nsid = cmd->nsid;
2748 
2749 		switch (cmd->opc) {
2750 		case NVME_OPC_FLUSH:
2751 			pending = nvme_opc_flush(sc, cmd, &sc->nvstore,
2752 			    req, &status);
2753  			break;
2754 		case NVME_OPC_WRITE:
2755 		case NVME_OPC_READ:
2756 			pending = nvme_opc_write_read(sc, cmd, &sc->nvstore,
2757 			    req, &status);
2758 			break;
2759 		case NVME_OPC_WRITE_ZEROES:
2760 			/* TODO: write zeroes
2761 			WPRINTF("%s write zeroes lba 0x%lx blocks %u",
2762 			        __func__, lba, cmd->cdw12 & 0xFFFF); */
2763 			pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2764 			break;
2765 		case NVME_OPC_DATASET_MANAGEMENT:
2766  			pending = nvme_opc_dataset_mgmt(sc, cmd, &sc->nvstore,
2767 			    req, &status);
2768 			break;
2769  		default:
2770  			WPRINTF("%s unhandled io command 0x%x",
2771 			    __func__, cmd->opc);
2772 			pci_nvme_status_genc(&status, NVME_SC_INVALID_OPCODE);
2773 		}
2774 complete:
2775 		if (!pending) {
2776 			pci_nvme_set_completion(sc, sq, idx, cmd->cid, status);
2777 			if (req != NULL)
2778 				pci_nvme_release_ioreq(sc, req);
2779 		}
2780 	}
2781 
2782 	sq->head = sqhead;
2783 
2784 	pthread_mutex_unlock(&sq->mtx);
2785 }
2786 
2787 static void
2788 pci_nvme_handle_doorbell(struct pci_nvme_softc* sc,
2789 	uint64_t idx, int is_sq, uint64_t value)
2790 {
2791 	DPRINTF("nvme doorbell %lu, %s, val 0x%lx",
2792 	        idx, is_sq ? "SQ" : "CQ", value & 0xFFFF);
2793 
2794 	if (is_sq) {
2795 		if (idx > sc->num_squeues) {
2796 			WPRINTF("%s queue index %lu overflow from "
2797 			         "guest (max %u)",
2798 			         __func__, idx, sc->num_squeues);
2799 			return;
2800 		}
2801 
2802 		atomic_store_short(&sc->submit_queues[idx].tail,
2803 		                   (uint16_t)value);
2804 
2805 		if (idx == 0) {
2806 			pci_nvme_handle_admin_cmd(sc, value);
2807 		} else {
2808 			/* submission queue; handle new entries in SQ */
2809 			if (idx > sc->num_squeues) {
2810 				WPRINTF("%s SQ index %lu overflow from "
2811 				         "guest (max %u)",
2812 				         __func__, idx, sc->num_squeues);
2813 				return;
2814 			}
2815 			pci_nvme_handle_io_cmd(sc, (uint16_t)idx);
2816 		}
2817 	} else {
2818 		if (idx > sc->num_cqueues) {
2819 			WPRINTF("%s queue index %lu overflow from "
2820 			         "guest (max %u)",
2821 			         __func__, idx, sc->num_cqueues);
2822 			return;
2823 		}
2824 
2825 		atomic_store_short(&sc->compl_queues[idx].head,
2826 				(uint16_t)value);
2827 	}
2828 }
2829 
2830 static void
2831 pci_nvme_bar0_reg_dumps(const char *func, uint64_t offset, int iswrite)
2832 {
2833 	const char *s = iswrite ? "WRITE" : "READ";
2834 
2835 	switch (offset) {
2836 	case NVME_CR_CAP_LOW:
2837 		DPRINTF("%s %s NVME_CR_CAP_LOW", func, s);
2838 		break;
2839 	case NVME_CR_CAP_HI:
2840 		DPRINTF("%s %s NVME_CR_CAP_HI", func, s);
2841 		break;
2842 	case NVME_CR_VS:
2843 		DPRINTF("%s %s NVME_CR_VS", func, s);
2844 		break;
2845 	case NVME_CR_INTMS:
2846 		DPRINTF("%s %s NVME_CR_INTMS", func, s);
2847 		break;
2848 	case NVME_CR_INTMC:
2849 		DPRINTF("%s %s NVME_CR_INTMC", func, s);
2850 		break;
2851 	case NVME_CR_CC:
2852 		DPRINTF("%s %s NVME_CR_CC", func, s);
2853 		break;
2854 	case NVME_CR_CSTS:
2855 		DPRINTF("%s %s NVME_CR_CSTS", func, s);
2856 		break;
2857 	case NVME_CR_NSSR:
2858 		DPRINTF("%s %s NVME_CR_NSSR", func, s);
2859 		break;
2860 	case NVME_CR_AQA:
2861 		DPRINTF("%s %s NVME_CR_AQA", func, s);
2862 		break;
2863 	case NVME_CR_ASQ_LOW:
2864 		DPRINTF("%s %s NVME_CR_ASQ_LOW", func, s);
2865 		break;
2866 	case NVME_CR_ASQ_HI:
2867 		DPRINTF("%s %s NVME_CR_ASQ_HI", func, s);
2868 		break;
2869 	case NVME_CR_ACQ_LOW:
2870 		DPRINTF("%s %s NVME_CR_ACQ_LOW", func, s);
2871 		break;
2872 	case NVME_CR_ACQ_HI:
2873 		DPRINTF("%s %s NVME_CR_ACQ_HI", func, s);
2874 		break;
2875 	default:
2876 		DPRINTF("unknown nvme bar-0 offset 0x%lx", offset);
2877 	}
2878 
2879 }
2880 
2881 static void
2882 pci_nvme_write_bar_0(struct pci_nvme_softc *sc, uint64_t offset, int size,
2883     uint64_t value)
2884 {
2885 	uint32_t ccreg;
2886 
2887 	if (offset >= NVME_DOORBELL_OFFSET) {
2888 		uint64_t belloffset = offset - NVME_DOORBELL_OFFSET;
2889 		uint64_t idx = belloffset / 8; /* door bell size = 2*int */
2890 		int is_sq = (belloffset % 8) < 4;
2891 
2892 		if ((sc->regs.csts & NVME_CSTS_RDY) == 0) {
2893 			WPRINTF("doorbell write prior to RDY (offset=%#lx)\n",
2894 			    offset);
2895 			return;
2896 		}
2897 
2898 		if (belloffset > ((sc->max_queues+1) * 8 - 4)) {
2899 			WPRINTF("guest attempted an overflow write offset "
2900 			         "0x%lx, val 0x%lx in %s",
2901 			         offset, value, __func__);
2902 			return;
2903 		}
2904 
2905 		if (is_sq) {
2906 			if (sc->submit_queues[idx].qbase == NULL)
2907 				return;
2908 		} else if (sc->compl_queues[idx].qbase == NULL)
2909 			return;
2910 
2911 		pci_nvme_handle_doorbell(sc, idx, is_sq, value);
2912 		return;
2913 	}
2914 
2915 	DPRINTF("nvme-write offset 0x%lx, size %d, value 0x%lx",
2916 	        offset, size, value);
2917 
2918 	if (size != 4) {
2919 		WPRINTF("guest wrote invalid size %d (offset 0x%lx, "
2920 		         "val 0x%lx) to bar0 in %s",
2921 		         size, offset, value, __func__);
2922 		/* TODO: shutdown device */
2923 		return;
2924 	}
2925 
2926 	pci_nvme_bar0_reg_dumps(__func__, offset, 1);
2927 
2928 	pthread_mutex_lock(&sc->mtx);
2929 
2930 	switch (offset) {
2931 	case NVME_CR_CAP_LOW:
2932 	case NVME_CR_CAP_HI:
2933 		/* readonly */
2934 		break;
2935 	case NVME_CR_VS:
2936 		/* readonly */
2937 		break;
2938 	case NVME_CR_INTMS:
2939 		/* MSI-X, so ignore */
2940 		break;
2941 	case NVME_CR_INTMC:
2942 		/* MSI-X, so ignore */
2943 		break;
2944 	case NVME_CR_CC:
2945 		ccreg = (uint32_t)value;
2946 
2947 		DPRINTF("%s NVME_CR_CC en %x css %x shn %x iosqes %u "
2948 		         "iocqes %u",
2949 		        __func__,
2950 			 NVME_CC_GET_EN(ccreg), NVME_CC_GET_CSS(ccreg),
2951 			 NVME_CC_GET_SHN(ccreg), NVME_CC_GET_IOSQES(ccreg),
2952 			 NVME_CC_GET_IOCQES(ccreg));
2953 
2954 		if (NVME_CC_GET_SHN(ccreg)) {
2955 			/* perform shutdown - flush out data to backend */
2956 			sc->regs.csts &= ~NVMEM(NVME_CSTS_REG_SHST);
2957 			sc->regs.csts |= NVMEF(NVME_CSTS_REG_SHST,
2958 			    NVME_SHST_COMPLETE);
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, "bootindex");
3163 	if (value != NULL) {
3164 		if (pci_emul_add_boot_device(sc->nsc_pi, atoi(value))) {
3165 			EPRINTLN("Invalid bootindex %d", atoi(value));
3166 			return (-1);
3167 		}
3168 	}
3169 
3170 	value = get_config_value_node(nvl, "ram");
3171 	if (value != NULL) {
3172 		uint64_t sz = strtoull(value, NULL, 10);
3173 
3174 		sc->nvstore.type = NVME_STOR_RAM;
3175 		sc->nvstore.size = sz * 1024 * 1024;
3176 		sc->nvstore.ctx = calloc(1, sc->nvstore.size);
3177 		sc->nvstore.sectsz = 4096;
3178 		sc->nvstore.sectsz_bits = 12;
3179 		if (sc->nvstore.ctx == NULL) {
3180 			EPRINTLN("nvme: Unable to allocate RAM");
3181 			return (-1);
3182 		}
3183 	} else {
3184 		snprintf(bident, sizeof(bident), "%u:%u",
3185 		    sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
3186 		sc->nvstore.ctx = blockif_open(nvl, bident);
3187 		if (sc->nvstore.ctx == NULL) {
3188 			EPRINTLN("nvme: Could not open backing file: %s",
3189 			    strerror(errno));
3190 			return (-1);
3191 		}
3192 		sc->nvstore.type = NVME_STOR_BLOCKIF;
3193 		sc->nvstore.size = blockif_size(sc->nvstore.ctx);
3194 	}
3195 
3196 	if (sectsz == 512 || sectsz == 4096 || sectsz == 8192)
3197 		sc->nvstore.sectsz = sectsz;
3198 	else if (sc->nvstore.type != NVME_STOR_RAM)
3199 		sc->nvstore.sectsz = blockif_sectsz(sc->nvstore.ctx);
3200 	for (sc->nvstore.sectsz_bits = 9;
3201 	     (1U << sc->nvstore.sectsz_bits) < sc->nvstore.sectsz;
3202 	     sc->nvstore.sectsz_bits++);
3203 
3204 	if (sc->max_queues <= 0 || sc->max_queues > NVME_QUEUES)
3205 		sc->max_queues = NVME_QUEUES;
3206 
3207 	return (0);
3208 }
3209 
3210 static void
3211 pci_nvme_resized(struct blockif_ctxt *bctxt __unused, void *arg,
3212     size_t new_size)
3213 {
3214 	struct pci_nvme_softc *sc;
3215 	struct pci_nvme_blockstore *nvstore;
3216 	struct nvme_namespace_data *nd;
3217 
3218 	sc = arg;
3219 	nvstore = &sc->nvstore;
3220 	nd = &sc->nsdata;
3221 
3222 	nvstore->size = new_size;
3223 	pci_nvme_init_nsdata_size(nvstore, nd);
3224 
3225 	/* Add changed NSID to list */
3226 	sc->ns_log.ns[0] = 1;
3227 	sc->ns_log.ns[1] = 0;
3228 
3229 	pci_nvme_aen_post(sc, PCI_NVME_AE_TYPE_NOTICE,
3230 	    PCI_NVME_AEI_NOTICE_NS_ATTR_CHANGED);
3231 }
3232 
3233 static int
3234 pci_nvme_init(struct pci_devinst *pi, nvlist_t *nvl)
3235 {
3236 	struct pci_nvme_softc *sc;
3237 	uint32_t pci_membar_sz;
3238 	int	error;
3239 
3240 	error = 0;
3241 
3242 	sc = calloc(1, sizeof(struct pci_nvme_softc));
3243 	pi->pi_arg = sc;
3244 	sc->nsc_pi = pi;
3245 
3246 	error = pci_nvme_parse_config(sc, nvl);
3247 	if (error < 0)
3248 		goto done;
3249 	else
3250 		error = 0;
3251 
3252 	STAILQ_INIT(&sc->ioreqs_free);
3253 	sc->ioreqs = calloc(sc->ioslots, sizeof(struct pci_nvme_ioreq));
3254 	for (uint32_t i = 0; i < sc->ioslots; i++) {
3255 		STAILQ_INSERT_TAIL(&sc->ioreqs_free, &sc->ioreqs[i], link);
3256 	}
3257 
3258 	pci_set_cfgdata16(pi, PCIR_DEVICE, 0x0A0A);
3259 	pci_set_cfgdata16(pi, PCIR_VENDOR, 0xFB5D);
3260 	pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_STORAGE);
3261 	pci_set_cfgdata8(pi, PCIR_SUBCLASS, PCIS_STORAGE_NVM);
3262 	pci_set_cfgdata8(pi, PCIR_PROGIF,
3263 	                 PCIP_STORAGE_NVM_ENTERPRISE_NVMHCI_1_0);
3264 
3265 	/*
3266 	 * Allocate size of NVMe registers + doorbell space for all queues.
3267 	 *
3268 	 * The specification requires a minimum memory I/O window size of 16K.
3269 	 * The Windows driver will refuse to start a device with a smaller
3270 	 * window.
3271 	 */
3272 	pci_membar_sz = sizeof(struct nvme_registers) +
3273 	    2 * sizeof(uint32_t) * (sc->max_queues + 1);
3274 	pci_membar_sz = MAX(pci_membar_sz, NVME_MMIO_SPACE_MIN);
3275 
3276 	DPRINTF("nvme membar size: %u", pci_membar_sz);
3277 
3278 	error = pci_emul_alloc_bar(pi, 0, PCIBAR_MEM64, pci_membar_sz);
3279 	if (error) {
3280 		WPRINTF("%s pci alloc mem bar failed", __func__);
3281 		goto done;
3282 	}
3283 
3284 	error = pci_emul_add_msixcap(pi, sc->max_queues + 1, NVME_MSIX_BAR);
3285 	if (error) {
3286 		WPRINTF("%s pci add msixcap failed", __func__);
3287 		goto done;
3288 	}
3289 
3290 	error = pci_emul_add_pciecap(pi, PCIEM_TYPE_ROOT_INT_EP);
3291 	if (error) {
3292 		WPRINTF("%s pci add Express capability failed", __func__);
3293 		goto done;
3294 	}
3295 
3296 	pthread_mutex_init(&sc->mtx, NULL);
3297 	sem_init(&sc->iosemlock, 0, sc->ioslots);
3298 	blockif_register_resize_callback(sc->nvstore.ctx, pci_nvme_resized, sc);
3299 
3300 	pci_nvme_init_queues(sc, sc->max_queues, sc->max_queues);
3301 	/*
3302 	 * Controller data depends on Namespace data so initialize Namespace
3303 	 * data first.
3304 	 */
3305 	pci_nvme_init_nsdata(sc, &sc->nsdata, 1, &sc->nvstore);
3306 	pci_nvme_init_ctrldata(sc);
3307 	pci_nvme_init_logpages(sc);
3308 	pci_nvme_init_features(sc);
3309 
3310 	pci_nvme_aer_init(sc);
3311 	pci_nvme_aen_init(sc);
3312 
3313 	pci_nvme_reset(sc);
3314 done:
3315 	return (error);
3316 }
3317 
3318 static int
3319 pci_nvme_legacy_config(nvlist_t *nvl, const char *opts)
3320 {
3321 	char *cp, *ram;
3322 
3323 	if (opts == NULL)
3324 		return (0);
3325 
3326 	if (strncmp(opts, "ram=", 4) == 0) {
3327 		cp = strchr(opts, ',');
3328 		if (cp == NULL) {
3329 			set_config_value_node(nvl, "ram", opts + 4);
3330 			return (0);
3331 		}
3332 		ram = strndup(opts + 4, cp - opts - 4);
3333 		set_config_value_node(nvl, "ram", ram);
3334 		free(ram);
3335 		return (pci_parse_legacy_config(nvl, cp + 1));
3336 	} else
3337 		return (blockif_legacy_config(nvl, opts));
3338 }
3339 
3340 static const struct pci_devemu pci_de_nvme = {
3341 	.pe_emu =	"nvme",
3342 	.pe_init =	pci_nvme_init,
3343 	.pe_legacy_config = pci_nvme_legacy_config,
3344 	.pe_barwrite =	pci_nvme_write,
3345 	.pe_barread =	pci_nvme_read
3346 };
3347 PCI_EMUL_SET(pci_de_nvme);
3348